xref: /linux/drivers/infiniband/hw/mlx5/main.c (revision 17ade5366345656e1a7f4e9da16863a7499da21b)
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/mlx5/driver.h>
28 #include <linux/list.h>
29 #include <rdma/ib_smi.h>
30 #include <rdma/ib_umem_odp.h>
31 #include <rdma/lag.h>
32 #include <linux/in.h>
33 #include <linux/etherdevice.h>
34 #include "mlx5_ib.h"
35 #include "ib_rep.h"
36 #include "cmd.h"
37 #include "devx.h"
38 #include "dm.h"
39 #include "fs.h"
40 #include "srq.h"
41 #include "qp.h"
42 #include "wr.h"
43 #include "restrack.h"
44 #include "counters.h"
45 #include "umr.h"
46 #include <rdma/uverbs_std_types.h>
47 #include <rdma/uverbs_ioctl.h>
48 #include <rdma/mlx5_user_ioctl_verbs.h>
49 #include <rdma/mlx5_user_ioctl_cmds.h>
50 #include <rdma/ib_ucaps.h>
51 #include "macsec.h"
52 #include "data_direct.h"
53 
54 #define UVERBS_MODULE_NAME mlx5_ib
55 #include <rdma/uverbs_named_ioctl.h>
56 
57 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
58 MODULE_DESCRIPTION("Mellanox 5th generation network adapters (ConnectX series) IB driver");
59 MODULE_LICENSE("Dual BSD/GPL");
60 
61 struct mlx5_ib_event_work {
62 	struct work_struct	work;
63 	union {
64 		struct mlx5_ib_dev	      *dev;
65 		struct mlx5_ib_multiport_info *mpi;
66 	};
67 	bool			is_slave;
68 	unsigned int		event;
69 	void			*param;
70 };
71 
72 enum {
73 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
74 };
75 
76 static struct workqueue_struct *mlx5_ib_event_wq;
77 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
78 static LIST_HEAD(mlx5_ib_dev_list);
79 /*
80  * This mutex should be held when accessing either of the above lists
81  */
82 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
83 
84 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
85 {
86 	struct mlx5_ib_dev *dev;
87 
88 	mutex_lock(&mlx5_ib_multiport_mutex);
89 	dev = mpi->ibdev;
90 	mutex_unlock(&mlx5_ib_multiport_mutex);
91 	return dev;
92 }
93 
94 static enum rdma_link_layer
95 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
96 {
97 	switch (port_type_cap) {
98 	case MLX5_CAP_PORT_TYPE_IB:
99 		return IB_LINK_LAYER_INFINIBAND;
100 	case MLX5_CAP_PORT_TYPE_ETH:
101 		return IB_LINK_LAYER_ETHERNET;
102 	default:
103 		return IB_LINK_LAYER_UNSPECIFIED;
104 	}
105 }
106 
107 static enum rdma_link_layer
108 mlx5_ib_port_link_layer(struct ib_device *device, u32 port_num)
109 {
110 	struct mlx5_ib_dev *dev = to_mdev(device);
111 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
112 
113 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
114 }
115 
116 static int get_port_state(struct ib_device *ibdev,
117 			  u32 port_num,
118 			  enum ib_port_state *state)
119 {
120 	struct ib_port_attr attr;
121 	int ret;
122 
123 	memset(&attr, 0, sizeof(attr));
124 	ret = ibdev->ops.query_port(ibdev, port_num, &attr);
125 	if (!ret)
126 		*state = attr.state;
127 	return ret;
128 }
129 
130 static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev,
131 					   struct net_device *ndev,
132 					   struct net_device *upper,
133 					   u32 *port_num)
134 {
135 	struct net_device *rep_ndev;
136 	struct mlx5_ib_port *port;
137 	int i;
138 
139 	for (i = 0; i < dev->num_ports; i++) {
140 		port  = &dev->port[i];
141 		if (!port->rep)
142 			continue;
143 
144 		if (upper == ndev && port->rep->vport == MLX5_VPORT_UPLINK) {
145 			*port_num = i + 1;
146 			return &port->roce;
147 		}
148 
149 		if (upper && port->rep->vport == MLX5_VPORT_UPLINK)
150 			continue;
151 		rep_ndev = ib_device_get_netdev(&dev->ib_dev, i + 1);
152 		if (rep_ndev && rep_ndev == ndev) {
153 			dev_put(rep_ndev);
154 			*port_num = i + 1;
155 			return &port->roce;
156 		}
157 
158 		dev_put(rep_ndev);
159 	}
160 
161 	return NULL;
162 }
163 
164 static bool mlx5_netdev_send_event(struct mlx5_ib_dev *dev,
165 				   struct net_device *ndev,
166 				   struct net_device *upper,
167 				   struct net_device *ib_ndev)
168 {
169 	if (!dev->ib_active)
170 		return false;
171 
172 	/* Event is about our upper device */
173 	if (upper == ndev)
174 		return true;
175 
176 	/* RDMA device is not in lag and not in switchdev */
177 	if (!dev->is_rep && !upper && ndev == ib_ndev)
178 		return true;
179 
180 	/* RDMA devie is in switchdev */
181 	if (dev->is_rep && ndev == ib_ndev)
182 		return true;
183 
184 	return false;
185 }
186 
187 static struct net_device *mlx5_ib_get_rep_uplink_netdev(struct mlx5_ib_dev *ibdev)
188 {
189 	struct mlx5_ib_port *port;
190 	int i;
191 
192 	for (i = 0; i < ibdev->num_ports; i++) {
193 		port = &ibdev->port[i];
194 		if (port->rep && port->rep->vport == MLX5_VPORT_UPLINK) {
195 			return ib_device_get_netdev(&ibdev->ib_dev, i + 1);
196 		}
197 	}
198 
199 	return NULL;
200 }
201 
202 static int mlx5_netdev_event(struct notifier_block *this,
203 			     unsigned long event, void *ptr)
204 {
205 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
206 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
207 	u32 port_num = roce->native_port_num;
208 	struct net_device *ib_ndev = NULL;
209 	struct mlx5_core_dev *mdev;
210 	struct mlx5_ib_dev *ibdev;
211 
212 	ibdev = roce->dev;
213 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
214 	if (!mdev)
215 		return NOTIFY_DONE;
216 
217 	switch (event) {
218 	case NETDEV_REGISTER:
219 		/* Should already be registered during the load */
220 		if (ibdev->is_rep)
221 			break;
222 
223 		ib_ndev = ib_device_get_netdev(&ibdev->ib_dev, port_num);
224 		/* Exit if already registered */
225 		if (ib_ndev)
226 			goto put_ndev;
227 
228 		if (ndev->dev.parent == mdev->device)
229 			ib_device_set_netdev(&ibdev->ib_dev, ndev, port_num);
230 		break;
231 
232 	case NETDEV_UNREGISTER:
233 		/* In case of reps, ib device goes away before the netdevs */
234 		if (ibdev->is_rep)
235 			break;
236 		ib_ndev = ib_device_get_netdev(&ibdev->ib_dev, port_num);
237 		if (ib_ndev == ndev)
238 			ib_device_set_netdev(&ibdev->ib_dev, NULL, port_num);
239 		goto put_ndev;
240 
241 	case NETDEV_CHANGE:
242 	case NETDEV_UP:
243 	case NETDEV_DOWN: {
244 		struct net_device *upper = NULL;
245 
246 		if (!netif_is_lag_master(ndev) && !netif_is_lag_port(ndev) &&
247 		    !mlx5_core_mp_enabled(mdev))
248 			return NOTIFY_DONE;
249 
250 		if (mlx5_lag_is_roce(mdev) || mlx5_lag_is_sriov(mdev)) {
251 			struct net_device *lag_ndev;
252 
253 			if(mlx5_lag_is_roce(mdev))
254 				lag_ndev = ib_device_get_netdev(&ibdev->ib_dev, 1);
255 			else /* sriov lag */
256 				lag_ndev = mlx5_ib_get_rep_uplink_netdev(ibdev);
257 
258 			if (lag_ndev) {
259 				upper = netdev_master_upper_dev_get(lag_ndev);
260 				dev_put(lag_ndev);
261 			} else {
262 				goto done;
263 			}
264 		}
265 
266 		if (ibdev->is_rep)
267 			roce = mlx5_get_rep_roce(ibdev, ndev, upper, &port_num);
268 		if (!roce)
269 			return NOTIFY_DONE;
270 
271 		ib_ndev = ib_device_get_netdev(&ibdev->ib_dev, port_num);
272 
273 		if (mlx5_netdev_send_event(ibdev, ndev, upper, ib_ndev)) {
274 			struct ib_event ibev = { };
275 			enum ib_port_state port_state;
276 
277 			if (get_port_state(&ibdev->ib_dev, port_num,
278 					   &port_state))
279 				goto put_ndev;
280 
281 			if (roce->last_port_state == port_state)
282 				goto put_ndev;
283 
284 			roce->last_port_state = port_state;
285 			ibev.device = &ibdev->ib_dev;
286 			if (port_state == IB_PORT_DOWN)
287 				ibev.event = IB_EVENT_PORT_ERR;
288 			else if (port_state == IB_PORT_ACTIVE)
289 				ibev.event = IB_EVENT_PORT_ACTIVE;
290 			else
291 				goto put_ndev;
292 
293 			ibev.element.port_num = port_num;
294 			ib_dispatch_event(&ibev);
295 		}
296 		break;
297 	}
298 
299 	default:
300 		break;
301 	}
302 put_ndev:
303 	dev_put(ib_ndev);
304 done:
305 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
306 	return NOTIFY_DONE;
307 }
308 
309 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
310 						   u32 ib_port_num,
311 						   u32 *native_port_num)
312 {
313 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
314 							  ib_port_num);
315 	struct mlx5_core_dev *mdev = NULL;
316 	struct mlx5_ib_multiport_info *mpi;
317 	struct mlx5_ib_port *port;
318 
319 	if (ibdev->ib_dev.type == RDMA_DEVICE_TYPE_SMI) {
320 		if (native_port_num)
321 			*native_port_num = smi_to_native_portnum(ibdev,
322 								 ib_port_num);
323 		return ibdev->mdev;
324 
325 	}
326 
327 	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
328 	    ll != IB_LINK_LAYER_ETHERNET) {
329 		if (native_port_num)
330 			*native_port_num = ib_port_num;
331 		return ibdev->mdev;
332 	}
333 
334 	if (native_port_num)
335 		*native_port_num = 1;
336 
337 	port = &ibdev->port[ib_port_num - 1];
338 	spin_lock(&port->mp.mpi_lock);
339 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
340 	if (mpi && !mpi->unaffiliate) {
341 		mdev = mpi->mdev;
342 		/* If it's the master no need to refcount, it'll exist
343 		 * as long as the ib_dev exists.
344 		 */
345 		if (!mpi->is_master)
346 			mpi->mdev_refcnt++;
347 	}
348 	spin_unlock(&port->mp.mpi_lock);
349 
350 	return mdev;
351 }
352 
353 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u32 port_num)
354 {
355 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
356 							  port_num);
357 	struct mlx5_ib_multiport_info *mpi;
358 	struct mlx5_ib_port *port;
359 
360 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
361 		return;
362 
363 	port = &ibdev->port[port_num - 1];
364 
365 	spin_lock(&port->mp.mpi_lock);
366 	mpi = ibdev->port[port_num - 1].mp.mpi;
367 	if (mpi->is_master)
368 		goto out;
369 
370 	mpi->mdev_refcnt--;
371 	if (mpi->unaffiliate)
372 		complete(&mpi->unref_comp);
373 out:
374 	spin_unlock(&port->mp.mpi_lock);
375 }
376 
377 static int translate_eth_legacy_proto_oper(u32 eth_proto_oper,
378 					   u16 *active_speed, u8 *active_width)
379 {
380 	switch (eth_proto_oper) {
381 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
382 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
383 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
384 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
385 		*active_width = IB_WIDTH_1X;
386 		*active_speed = IB_SPEED_SDR;
387 		break;
388 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
389 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
390 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
391 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
392 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
393 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
394 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
395 		*active_width = IB_WIDTH_1X;
396 		*active_speed = IB_SPEED_QDR;
397 		break;
398 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
399 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
400 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
401 		*active_width = IB_WIDTH_1X;
402 		*active_speed = IB_SPEED_EDR;
403 		break;
404 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
405 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
406 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
407 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
408 		*active_width = IB_WIDTH_4X;
409 		*active_speed = IB_SPEED_QDR;
410 		break;
411 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
412 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
413 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
414 		*active_width = IB_WIDTH_1X;
415 		*active_speed = IB_SPEED_HDR;
416 		break;
417 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
418 		*active_width = IB_WIDTH_4X;
419 		*active_speed = IB_SPEED_FDR;
420 		break;
421 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
422 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
423 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
424 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
425 		*active_width = IB_WIDTH_4X;
426 		*active_speed = IB_SPEED_EDR;
427 		break;
428 	default:
429 		return -EINVAL;
430 	}
431 
432 	return 0;
433 }
434 
435 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u16 *active_speed,
436 					u8 *active_width)
437 {
438 	switch (eth_proto_oper) {
439 	case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
440 	case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
441 		*active_width = IB_WIDTH_1X;
442 		*active_speed = IB_SPEED_SDR;
443 		break;
444 	case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
445 		*active_width = IB_WIDTH_1X;
446 		*active_speed = IB_SPEED_DDR;
447 		break;
448 	case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
449 		*active_width = IB_WIDTH_1X;
450 		*active_speed = IB_SPEED_QDR;
451 		break;
452 	case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
453 		*active_width = IB_WIDTH_4X;
454 		*active_speed = IB_SPEED_QDR;
455 		break;
456 	case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
457 		*active_width = IB_WIDTH_1X;
458 		*active_speed = IB_SPEED_EDR;
459 		break;
460 	case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
461 		*active_width = IB_WIDTH_2X;
462 		*active_speed = IB_SPEED_EDR;
463 		break;
464 	case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
465 		*active_width = IB_WIDTH_1X;
466 		*active_speed = IB_SPEED_HDR;
467 		break;
468 	case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
469 		*active_width = IB_WIDTH_4X;
470 		*active_speed = IB_SPEED_EDR;
471 		break;
472 	case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
473 		*active_width = IB_WIDTH_2X;
474 		*active_speed = IB_SPEED_HDR;
475 		break;
476 	case MLX5E_PROT_MASK(MLX5E_100GAUI_1_100GBASE_CR_KR):
477 		*active_width = IB_WIDTH_1X;
478 		*active_speed = IB_SPEED_NDR;
479 		break;
480 	case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4):
481 		*active_width = IB_WIDTH_4X;
482 		*active_speed = IB_SPEED_HDR;
483 		break;
484 	case MLX5E_PROT_MASK(MLX5E_200GAUI_2_200GBASE_CR2_KR2):
485 		*active_width = IB_WIDTH_2X;
486 		*active_speed = IB_SPEED_NDR;
487 		break;
488 	case MLX5E_PROT_MASK(MLX5E_400GAUI_8_400GBASE_CR8):
489 		*active_width = IB_WIDTH_8X;
490 		*active_speed = IB_SPEED_HDR;
491 		break;
492 	case MLX5E_PROT_MASK(MLX5E_400GAUI_4_400GBASE_CR4_KR4):
493 		*active_width = IB_WIDTH_4X;
494 		*active_speed = IB_SPEED_NDR;
495 		break;
496 	case MLX5E_PROT_MASK(MLX5E_800GAUI_8_800GBASE_CR8_KR8):
497 		*active_width = IB_WIDTH_8X;
498 		*active_speed = IB_SPEED_NDR;
499 		break;
500 	default:
501 		return -EINVAL;
502 	}
503 
504 	return 0;
505 }
506 
507 static int translate_eth_proto_oper(u32 eth_proto_oper, u16 *active_speed,
508 				    u8 *active_width, bool ext)
509 {
510 	return ext ?
511 		translate_eth_ext_proto_oper(eth_proto_oper, active_speed,
512 					     active_width) :
513 		translate_eth_legacy_proto_oper(eth_proto_oper, active_speed,
514 						active_width);
515 }
516 
517 static int mlx5_query_port_roce(struct ib_device *device, u32 port_num,
518 				struct ib_port_attr *props)
519 {
520 	struct mlx5_ib_dev *dev = to_mdev(device);
521 	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0};
522 	struct mlx5_core_dev *mdev;
523 	struct net_device *ndev, *upper;
524 	enum ib_mtu ndev_ib_mtu;
525 	bool put_mdev = true;
526 	u32 eth_prot_oper;
527 	u32 mdev_port_num;
528 	bool ext;
529 	int err;
530 
531 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
532 	if (!mdev) {
533 		/* This means the port isn't affiliated yet. Get the
534 		 * info for the master port instead.
535 		 */
536 		put_mdev = false;
537 		mdev = dev->mdev;
538 		mdev_port_num = 1;
539 		port_num = 1;
540 	}
541 
542 	/* Possible bad flows are checked before filling out props so in case
543 	 * of an error it will still be zeroed out.
544 	 * Use native port in case of reps
545 	 */
546 	if (dev->is_rep)
547 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
548 					   1, 0);
549 	else
550 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
551 					   mdev_port_num, 0);
552 	if (err)
553 		goto out;
554 	ext = !!MLX5_GET_ETH_PROTO(ptys_reg, out, true, eth_proto_capability);
555 	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
556 
557 	props->active_width     = IB_WIDTH_4X;
558 	props->active_speed     = IB_SPEED_QDR;
559 
560 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
561 				 &props->active_width, ext);
562 
563 	if (!dev->is_rep && dev->mdev->roce.roce_en) {
564 		u16 qkey_viol_cntr;
565 
566 		props->port_cap_flags |= IB_PORT_CM_SUP;
567 		props->ip_gids = true;
568 		props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev,
569 						   roce_address_table_size);
570 		mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
571 		props->qkey_viol_cntr = qkey_viol_cntr;
572 	}
573 	props->max_mtu          = IB_MTU_4096;
574 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
575 	props->pkey_tbl_len     = 1;
576 	props->state            = IB_PORT_DOWN;
577 	props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
578 
579 	/* If this is a stub query for an unaffiliated port stop here */
580 	if (!put_mdev)
581 		goto out;
582 
583 	ndev = ib_device_get_netdev(device, port_num);
584 	if (!ndev)
585 		goto out;
586 
587 	if (mlx5_lag_is_roce(mdev) || mlx5_lag_is_sriov(mdev)) {
588 		rcu_read_lock();
589 		upper = netdev_master_upper_dev_get_rcu(ndev);
590 		if (upper) {
591 			dev_put(ndev);
592 			ndev = upper;
593 			dev_hold(ndev);
594 		}
595 		rcu_read_unlock();
596 	}
597 
598 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
599 		props->state      = IB_PORT_ACTIVE;
600 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
601 	}
602 
603 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
604 
605 	dev_put(ndev);
606 
607 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
608 out:
609 	if (put_mdev)
610 		mlx5_ib_put_native_port_mdev(dev, port_num);
611 	return err;
612 }
613 
614 int set_roce_addr(struct mlx5_ib_dev *dev, u32 port_num,
615 		  unsigned int index, const union ib_gid *gid,
616 		  const struct ib_gid_attr *attr)
617 {
618 	enum ib_gid_type gid_type;
619 	u16 vlan_id = 0xffff;
620 	u8 roce_version = 0;
621 	u8 roce_l3_type = 0;
622 	u8 mac[ETH_ALEN];
623 	int ret;
624 
625 	gid_type = attr->gid_type;
626 	if (gid) {
627 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
628 		if (ret)
629 			return ret;
630 	}
631 
632 	switch (gid_type) {
633 	case IB_GID_TYPE_ROCE:
634 		roce_version = MLX5_ROCE_VERSION_1;
635 		break;
636 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
637 		roce_version = MLX5_ROCE_VERSION_2;
638 		if (gid && ipv6_addr_v4mapped((void *)gid))
639 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
640 		else
641 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
642 		break;
643 
644 	default:
645 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
646 	}
647 
648 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
649 				      roce_l3_type, gid->raw, mac,
650 				      vlan_id < VLAN_CFI_MASK, vlan_id,
651 				      port_num);
652 }
653 
654 static int mlx5_ib_add_gid(const struct ib_gid_attr *attr,
655 			   __always_unused void **context)
656 {
657 	int ret;
658 
659 	ret = mlx5r_add_gid_macsec_operations(attr);
660 	if (ret)
661 		return ret;
662 
663 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
664 			     attr->index, &attr->gid, attr);
665 }
666 
667 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
668 			   __always_unused void **context)
669 {
670 	int ret;
671 
672 	ret = set_roce_addr(to_mdev(attr->device), attr->port_num,
673 			    attr->index, NULL, attr);
674 	if (ret)
675 		return ret;
676 
677 	mlx5r_del_gid_macsec_operations(attr);
678 	return 0;
679 }
680 
681 __be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev,
682 				   const struct ib_gid_attr *attr)
683 {
684 	if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
685 		return 0;
686 
687 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
688 }
689 
690 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
691 {
692 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
693 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
694 	return 0;
695 }
696 
697 enum {
698 	MLX5_VPORT_ACCESS_METHOD_MAD,
699 	MLX5_VPORT_ACCESS_METHOD_HCA,
700 	MLX5_VPORT_ACCESS_METHOD_NIC,
701 };
702 
703 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
704 {
705 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
706 		return MLX5_VPORT_ACCESS_METHOD_MAD;
707 
708 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
709 	    IB_LINK_LAYER_ETHERNET)
710 		return MLX5_VPORT_ACCESS_METHOD_NIC;
711 
712 	return MLX5_VPORT_ACCESS_METHOD_HCA;
713 }
714 
715 static void get_atomic_caps(struct mlx5_ib_dev *dev,
716 			    u8 atomic_size_qp,
717 			    struct ib_device_attr *props)
718 {
719 	u8 tmp;
720 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
721 	u8 atomic_req_8B_endianness_mode =
722 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
723 
724 	/* Check if HW supports 8 bytes standard atomic operations and capable
725 	 * of host endianness respond
726 	 */
727 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
728 	if (((atomic_operations & tmp) == tmp) &&
729 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
730 	    (atomic_req_8B_endianness_mode)) {
731 		props->atomic_cap = IB_ATOMIC_HCA;
732 	} else {
733 		props->atomic_cap = IB_ATOMIC_NONE;
734 	}
735 }
736 
737 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
738 			       struct ib_device_attr *props)
739 {
740 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
741 
742 	get_atomic_caps(dev, atomic_size_qp, props);
743 }
744 
745 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
746 					__be64 *sys_image_guid)
747 {
748 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
749 	struct mlx5_core_dev *mdev = dev->mdev;
750 	u64 tmp;
751 	int err;
752 
753 	switch (mlx5_get_vport_access_method(ibdev)) {
754 	case MLX5_VPORT_ACCESS_METHOD_MAD:
755 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
756 							    sys_image_guid);
757 
758 	case MLX5_VPORT_ACCESS_METHOD_HCA:
759 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
760 		break;
761 
762 	case MLX5_VPORT_ACCESS_METHOD_NIC:
763 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
764 		break;
765 
766 	default:
767 		return -EINVAL;
768 	}
769 
770 	if (!err)
771 		*sys_image_guid = cpu_to_be64(tmp);
772 
773 	return err;
774 
775 }
776 
777 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
778 				u16 *max_pkeys)
779 {
780 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
781 	struct mlx5_core_dev *mdev = dev->mdev;
782 
783 	switch (mlx5_get_vport_access_method(ibdev)) {
784 	case MLX5_VPORT_ACCESS_METHOD_MAD:
785 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
786 
787 	case MLX5_VPORT_ACCESS_METHOD_HCA:
788 	case MLX5_VPORT_ACCESS_METHOD_NIC:
789 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
790 						pkey_table_size));
791 		return 0;
792 
793 	default:
794 		return -EINVAL;
795 	}
796 }
797 
798 static int mlx5_query_vendor_id(struct ib_device *ibdev,
799 				u32 *vendor_id)
800 {
801 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
802 
803 	switch (mlx5_get_vport_access_method(ibdev)) {
804 	case MLX5_VPORT_ACCESS_METHOD_MAD:
805 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
806 
807 	case MLX5_VPORT_ACCESS_METHOD_HCA:
808 	case MLX5_VPORT_ACCESS_METHOD_NIC:
809 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
810 
811 	default:
812 		return -EINVAL;
813 	}
814 }
815 
816 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
817 				__be64 *node_guid)
818 {
819 	u64 tmp;
820 	int err;
821 
822 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
823 	case MLX5_VPORT_ACCESS_METHOD_MAD:
824 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
825 
826 	case MLX5_VPORT_ACCESS_METHOD_HCA:
827 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
828 		break;
829 
830 	case MLX5_VPORT_ACCESS_METHOD_NIC:
831 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
832 		break;
833 
834 	default:
835 		return -EINVAL;
836 	}
837 
838 	if (!err)
839 		*node_guid = cpu_to_be64(tmp);
840 
841 	return err;
842 }
843 
844 struct mlx5_reg_node_desc {
845 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
846 };
847 
848 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
849 {
850 	struct mlx5_reg_node_desc in;
851 
852 	if (mlx5_use_mad_ifc(dev))
853 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
854 
855 	memset(&in, 0, sizeof(in));
856 
857 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
858 				    sizeof(struct mlx5_reg_node_desc),
859 				    MLX5_REG_NODE_DESC, 0, 0);
860 }
861 
862 static void fill_esw_mgr_reg_c0(struct mlx5_core_dev *mdev,
863 				struct mlx5_ib_query_device_resp *resp)
864 {
865 	struct mlx5_eswitch *esw = mdev->priv.eswitch;
866 	u16 vport = mlx5_eswitch_manager_vport(mdev);
867 
868 	resp->reg_c0.value = mlx5_eswitch_get_vport_metadata_for_match(esw,
869 								      vport);
870 	resp->reg_c0.mask = mlx5_eswitch_get_vport_metadata_mask();
871 }
872 
873 static int mlx5_ib_query_device(struct ib_device *ibdev,
874 				struct ib_device_attr *props,
875 				struct ib_udata *uhw)
876 {
877 	size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
878 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
879 	struct mlx5_core_dev *mdev = dev->mdev;
880 	int err = -ENOMEM;
881 	int max_sq_desc;
882 	int max_rq_sg;
883 	int max_sq_sg;
884 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
885 	bool raw_support = !mlx5_core_mp_enabled(mdev);
886 	struct mlx5_ib_query_device_resp resp = {};
887 	size_t resp_len;
888 	u64 max_tso;
889 
890 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
891 	if (uhw_outlen && uhw_outlen < resp_len)
892 		return -EINVAL;
893 
894 	resp.response_length = resp_len;
895 
896 	if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
897 		return -EINVAL;
898 
899 	memset(props, 0, sizeof(*props));
900 	err = mlx5_query_system_image_guid(ibdev,
901 					   &props->sys_image_guid);
902 	if (err)
903 		return err;
904 
905 	props->max_pkeys = dev->pkey_table_len;
906 
907 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
908 	if (err)
909 		return err;
910 
911 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
912 		(fw_rev_min(dev->mdev) << 16) |
913 		fw_rev_sub(dev->mdev);
914 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
915 		IB_DEVICE_PORT_ACTIVE_EVENT		|
916 		IB_DEVICE_SYS_IMAGE_GUID		|
917 		IB_DEVICE_RC_RNR_NAK_GEN;
918 
919 	if (MLX5_CAP_GEN(mdev, pkv))
920 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
921 	if (MLX5_CAP_GEN(mdev, qkv))
922 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
923 	if (MLX5_CAP_GEN(mdev, apm))
924 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
925 	if (MLX5_CAP_GEN(mdev, xrc))
926 		props->device_cap_flags |= IB_DEVICE_XRC;
927 	if (MLX5_CAP_GEN(mdev, imaicl)) {
928 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
929 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
930 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
931 		/* We support 'Gappy' memory registration too */
932 		props->kernel_cap_flags |= IBK_SG_GAPS_REG;
933 	}
934 	/* IB_WR_REG_MR always requires changing the entity size with UMR */
935 	if (!MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled))
936 		props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
937 	if (MLX5_CAP_GEN(mdev, sho)) {
938 		props->kernel_cap_flags |= IBK_INTEGRITY_HANDOVER;
939 		/* At this stage no support for signature handover */
940 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
941 				      IB_PROT_T10DIF_TYPE_2 |
942 				      IB_PROT_T10DIF_TYPE_3;
943 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
944 				       IB_GUARD_T10DIF_CSUM;
945 	}
946 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
947 		props->kernel_cap_flags |= IBK_BLOCK_MULTICAST_LOOPBACK;
948 
949 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
950 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
951 			/* Legacy bit to support old userspace libraries */
952 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
953 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
954 		}
955 
956 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
957 			props->raw_packet_caps |=
958 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
959 
960 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) {
961 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
962 			if (max_tso) {
963 				resp.tso_caps.max_tso = 1 << max_tso;
964 				resp.tso_caps.supported_qpts |=
965 					1 << IB_QPT_RAW_PACKET;
966 				resp.response_length += sizeof(resp.tso_caps);
967 			}
968 		}
969 
970 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) {
971 			resp.rss_caps.rx_hash_function =
972 						MLX5_RX_HASH_FUNC_TOEPLITZ;
973 			resp.rss_caps.rx_hash_fields_mask =
974 						MLX5_RX_HASH_SRC_IPV4 |
975 						MLX5_RX_HASH_DST_IPV4 |
976 						MLX5_RX_HASH_SRC_IPV6 |
977 						MLX5_RX_HASH_DST_IPV6 |
978 						MLX5_RX_HASH_SRC_PORT_TCP |
979 						MLX5_RX_HASH_DST_PORT_TCP |
980 						MLX5_RX_HASH_SRC_PORT_UDP |
981 						MLX5_RX_HASH_DST_PORT_UDP |
982 						MLX5_RX_HASH_INNER;
983 			resp.response_length += sizeof(resp.rss_caps);
984 		}
985 	} else {
986 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen)
987 			resp.response_length += sizeof(resp.tso_caps);
988 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen)
989 			resp.response_length += sizeof(resp.rss_caps);
990 	}
991 
992 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
993 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
994 		props->kernel_cap_flags |= IBK_UD_TSO;
995 	}
996 
997 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
998 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
999 	    raw_support)
1000 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
1001 
1002 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
1003 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
1004 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
1005 
1006 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
1007 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
1008 	    raw_support) {
1009 		/* Legacy bit to support old userspace libraries */
1010 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
1011 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
1012 	}
1013 
1014 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
1015 		props->max_dm_size =
1016 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
1017 	}
1018 
1019 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
1020 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
1021 
1022 	if (MLX5_CAP_GEN(mdev, end_pad))
1023 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
1024 
1025 	props->vendor_part_id	   = mdev->pdev->device;
1026 	props->hw_ver		   = mdev->pdev->revision;
1027 
1028 	props->max_mr_size	   = ~0ull;
1029 	props->page_size_cap	   = ~(min_page_size - 1);
1030 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
1031 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1032 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
1033 		     sizeof(struct mlx5_wqe_data_seg);
1034 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
1035 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
1036 		     sizeof(struct mlx5_wqe_raddr_seg)) /
1037 		sizeof(struct mlx5_wqe_data_seg);
1038 	props->max_send_sge = max_sq_sg;
1039 	props->max_recv_sge = max_rq_sg;
1040 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
1041 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
1042 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
1043 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
1044 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
1045 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
1046 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
1047 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
1048 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
1049 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
1050 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
1051 	props->max_srq_sge	   = max_rq_sg - 1;
1052 	props->max_fast_reg_page_list_len =
1053 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
1054 	props->max_pi_fast_reg_page_list_len =
1055 		props->max_fast_reg_page_list_len / 2;
1056 	props->max_sgl_rd =
1057 		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
1058 	get_atomic_caps_qp(dev, props);
1059 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
1060 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
1061 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
1062 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1063 					   props->max_mcast_grp;
1064 	props->max_ah = INT_MAX;
1065 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
1066 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
1067 
1068 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
1069 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
1070 			props->kernel_cap_flags |= IBK_ON_DEMAND_PAGING;
1071 		props->odp_caps = dev->odp_caps;
1072 		if (!uhw) {
1073 			/* ODP for kernel QPs is not implemented for receive
1074 			 * WQEs and SRQ WQEs
1075 			 */
1076 			props->odp_caps.per_transport_caps.rc_odp_caps &=
1077 				~(IB_ODP_SUPPORT_READ |
1078 				  IB_ODP_SUPPORT_SRQ_RECV);
1079 			props->odp_caps.per_transport_caps.uc_odp_caps &=
1080 				~(IB_ODP_SUPPORT_READ |
1081 				  IB_ODP_SUPPORT_SRQ_RECV);
1082 			props->odp_caps.per_transport_caps.ud_odp_caps &=
1083 				~(IB_ODP_SUPPORT_READ |
1084 				  IB_ODP_SUPPORT_SRQ_RECV);
1085 			props->odp_caps.per_transport_caps.xrc_odp_caps &=
1086 				~(IB_ODP_SUPPORT_READ |
1087 				  IB_ODP_SUPPORT_SRQ_RECV);
1088 		}
1089 	}
1090 
1091 	if (mlx5_core_is_vf(mdev))
1092 		props->kernel_cap_flags |= IBK_VIRTUAL_FUNCTION;
1093 
1094 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1095 	    IB_LINK_LAYER_ETHERNET && raw_support) {
1096 		props->rss_caps.max_rwq_indirection_tables =
1097 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1098 		props->rss_caps.max_rwq_indirection_table_size =
1099 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1100 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1101 		props->max_wq_type_rq =
1102 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1103 	}
1104 
1105 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1106 		props->tm_caps.max_num_tags =
1107 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1108 		props->tm_caps.max_ops =
1109 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1110 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1111 	}
1112 
1113 	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1114 	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1115 		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1116 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1117 	}
1118 
1119 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1120 		props->cq_caps.max_cq_moderation_count =
1121 						MLX5_MAX_CQ_COUNT;
1122 		props->cq_caps.max_cq_moderation_period =
1123 						MLX5_MAX_CQ_PERIOD;
1124 	}
1125 
1126 	if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) {
1127 		resp.response_length += sizeof(resp.cqe_comp_caps);
1128 
1129 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1130 			resp.cqe_comp_caps.max_num =
1131 				MLX5_CAP_GEN(dev->mdev,
1132 					     cqe_compression_max_num);
1133 
1134 			resp.cqe_comp_caps.supported_format =
1135 				MLX5_IB_CQE_RES_FORMAT_HASH |
1136 				MLX5_IB_CQE_RES_FORMAT_CSUM;
1137 
1138 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1139 				resp.cqe_comp_caps.supported_format |=
1140 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1141 		}
1142 	}
1143 
1144 	if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen &&
1145 	    raw_support) {
1146 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1147 		    MLX5_CAP_GEN(mdev, qos)) {
1148 			resp.packet_pacing_caps.qp_rate_limit_max =
1149 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1150 			resp.packet_pacing_caps.qp_rate_limit_min =
1151 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1152 			resp.packet_pacing_caps.supported_qpts |=
1153 				1 << IB_QPT_RAW_PACKET;
1154 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1155 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1156 				resp.packet_pacing_caps.cap_flags |=
1157 					MLX5_IB_PP_SUPPORT_BURST;
1158 		}
1159 		resp.response_length += sizeof(resp.packet_pacing_caps);
1160 	}
1161 
1162 	if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <=
1163 	    uhw_outlen) {
1164 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1165 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1166 				MLX5_IB_ALLOW_MPW;
1167 
1168 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1169 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1170 				MLX5_IB_SUPPORT_EMPW;
1171 
1172 		resp.response_length +=
1173 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1174 	}
1175 
1176 	if (offsetofend(typeof(resp), flags) <= uhw_outlen) {
1177 		resp.response_length += sizeof(resp.flags);
1178 
1179 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1180 			resp.flags |=
1181 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1182 
1183 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1184 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1185 		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1186 			resp.flags |=
1187 				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1188 
1189 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1190 
1191 		if (MLX5_CAP_GEN_2(mdev, dp_ordering_force) &&
1192 		    (MLX5_CAP_GEN(mdev, dp_ordering_ooo_all_xrc) ||
1193 		    MLX5_CAP_GEN(mdev, dp_ordering_ooo_all_dc) ||
1194 		    MLX5_CAP_GEN(mdev, dp_ordering_ooo_all_rc) ||
1195 		    MLX5_CAP_GEN(mdev, dp_ordering_ooo_all_ud) ||
1196 		    MLX5_CAP_GEN(mdev, dp_ordering_ooo_all_uc)))
1197 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_OOO_DP;
1198 	}
1199 
1200 	if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) {
1201 		resp.response_length += sizeof(resp.sw_parsing_caps);
1202 		if (MLX5_CAP_ETH(mdev, swp)) {
1203 			resp.sw_parsing_caps.sw_parsing_offloads |=
1204 				MLX5_IB_SW_PARSING;
1205 
1206 			if (MLX5_CAP_ETH(mdev, swp_csum))
1207 				resp.sw_parsing_caps.sw_parsing_offloads |=
1208 					MLX5_IB_SW_PARSING_CSUM;
1209 
1210 			if (MLX5_CAP_ETH(mdev, swp_lso))
1211 				resp.sw_parsing_caps.sw_parsing_offloads |=
1212 					MLX5_IB_SW_PARSING_LSO;
1213 
1214 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1215 				resp.sw_parsing_caps.supported_qpts =
1216 					BIT(IB_QPT_RAW_PACKET);
1217 		}
1218 	}
1219 
1220 	if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen &&
1221 	    raw_support) {
1222 		resp.response_length += sizeof(resp.striding_rq_caps);
1223 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1224 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1225 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1226 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1227 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1228 			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1229 				resp.striding_rq_caps
1230 					.min_single_wqe_log_num_of_strides =
1231 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1232 			else
1233 				resp.striding_rq_caps
1234 					.min_single_wqe_log_num_of_strides =
1235 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1236 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1237 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1238 			resp.striding_rq_caps.supported_qpts =
1239 				BIT(IB_QPT_RAW_PACKET);
1240 		}
1241 	}
1242 
1243 	if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) {
1244 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1245 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1246 			resp.tunnel_offloads_caps |=
1247 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1248 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1249 			resp.tunnel_offloads_caps |=
1250 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1251 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1252 			resp.tunnel_offloads_caps |=
1253 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1254 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre))
1255 			resp.tunnel_offloads_caps |=
1256 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1257 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp))
1258 			resp.tunnel_offloads_caps |=
1259 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1260 	}
1261 
1262 	if (offsetofend(typeof(resp), dci_streams_caps) <= uhw_outlen) {
1263 		resp.response_length += sizeof(resp.dci_streams_caps);
1264 
1265 		resp.dci_streams_caps.max_log_num_concurent =
1266 			MLX5_CAP_GEN(mdev, log_max_dci_stream_channels);
1267 
1268 		resp.dci_streams_caps.max_log_num_errored =
1269 			MLX5_CAP_GEN(mdev, log_max_dci_errored_streams);
1270 	}
1271 
1272 	if (offsetofend(typeof(resp), reserved) <= uhw_outlen)
1273 		resp.response_length += sizeof(resp.reserved);
1274 
1275 	if (offsetofend(typeof(resp), reg_c0) <= uhw_outlen) {
1276 		struct mlx5_eswitch *esw = mdev->priv.eswitch;
1277 
1278 		resp.response_length += sizeof(resp.reg_c0);
1279 
1280 		if (mlx5_eswitch_mode(mdev) == MLX5_ESWITCH_OFFLOADS &&
1281 		    mlx5_eswitch_vport_match_metadata_enabled(esw))
1282 			fill_esw_mgr_reg_c0(mdev, &resp);
1283 	}
1284 
1285 	if (uhw_outlen) {
1286 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1287 
1288 		if (err)
1289 			return err;
1290 	}
1291 
1292 	return 0;
1293 }
1294 
1295 static void translate_active_width(struct ib_device *ibdev, u16 active_width,
1296 				   u8 *ib_width)
1297 {
1298 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1299 
1300 	if (active_width & MLX5_PTYS_WIDTH_1X)
1301 		*ib_width = IB_WIDTH_1X;
1302 	else if (active_width & MLX5_PTYS_WIDTH_2X)
1303 		*ib_width = IB_WIDTH_2X;
1304 	else if (active_width & MLX5_PTYS_WIDTH_4X)
1305 		*ib_width = IB_WIDTH_4X;
1306 	else if (active_width & MLX5_PTYS_WIDTH_8X)
1307 		*ib_width = IB_WIDTH_8X;
1308 	else if (active_width & MLX5_PTYS_WIDTH_12X)
1309 		*ib_width = IB_WIDTH_12X;
1310 	else {
1311 		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1312 			    active_width);
1313 		*ib_width = IB_WIDTH_4X;
1314 	}
1315 
1316 	return;
1317 }
1318 
1319 static int mlx5_mtu_to_ib_mtu(int mtu)
1320 {
1321 	switch (mtu) {
1322 	case 256: return 1;
1323 	case 512: return 2;
1324 	case 1024: return 3;
1325 	case 2048: return 4;
1326 	case 4096: return 5;
1327 	default:
1328 		pr_warn("invalid mtu\n");
1329 		return -1;
1330 	}
1331 }
1332 
1333 enum ib_max_vl_num {
1334 	__IB_MAX_VL_0		= 1,
1335 	__IB_MAX_VL_0_1		= 2,
1336 	__IB_MAX_VL_0_3		= 3,
1337 	__IB_MAX_VL_0_7		= 4,
1338 	__IB_MAX_VL_0_14	= 5,
1339 };
1340 
1341 enum mlx5_vl_hw_cap {
1342 	MLX5_VL_HW_0	= 1,
1343 	MLX5_VL_HW_0_1	= 2,
1344 	MLX5_VL_HW_0_2	= 3,
1345 	MLX5_VL_HW_0_3	= 4,
1346 	MLX5_VL_HW_0_4	= 5,
1347 	MLX5_VL_HW_0_5	= 6,
1348 	MLX5_VL_HW_0_6	= 7,
1349 	MLX5_VL_HW_0_7	= 8,
1350 	MLX5_VL_HW_0_14	= 15
1351 };
1352 
1353 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1354 				u8 *max_vl_num)
1355 {
1356 	switch (vl_hw_cap) {
1357 	case MLX5_VL_HW_0:
1358 		*max_vl_num = __IB_MAX_VL_0;
1359 		break;
1360 	case MLX5_VL_HW_0_1:
1361 		*max_vl_num = __IB_MAX_VL_0_1;
1362 		break;
1363 	case MLX5_VL_HW_0_3:
1364 		*max_vl_num = __IB_MAX_VL_0_3;
1365 		break;
1366 	case MLX5_VL_HW_0_7:
1367 		*max_vl_num = __IB_MAX_VL_0_7;
1368 		break;
1369 	case MLX5_VL_HW_0_14:
1370 		*max_vl_num = __IB_MAX_VL_0_14;
1371 		break;
1372 
1373 	default:
1374 		return -EINVAL;
1375 	}
1376 
1377 	return 0;
1378 }
1379 
1380 static int mlx5_query_hca_port(struct ib_device *ibdev, u32 port,
1381 			       struct ib_port_attr *props)
1382 {
1383 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1384 	struct mlx5_core_dev *mdev = dev->mdev;
1385 	struct mlx5_hca_vport_context *rep;
1386 	u8 vl_hw_cap, plane_index = 0;
1387 	u16 max_mtu;
1388 	u16 oper_mtu;
1389 	int err;
1390 	u16 ib_link_width_oper;
1391 
1392 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1393 	if (!rep) {
1394 		err = -ENOMEM;
1395 		goto out;
1396 	}
1397 
1398 	/* props being zeroed by the caller, avoid zeroing it here */
1399 
1400 	if (ibdev->type == RDMA_DEVICE_TYPE_SMI) {
1401 		plane_index = port;
1402 		port = smi_to_native_portnum(dev, port);
1403 	}
1404 
1405 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1406 	if (err)
1407 		goto out;
1408 
1409 	props->lid		= rep->lid;
1410 	props->lmc		= rep->lmc;
1411 	props->sm_lid		= rep->sm_lid;
1412 	props->sm_sl		= rep->sm_sl;
1413 	props->state		= rep->vport_state;
1414 	props->phys_state	= rep->port_physical_state;
1415 
1416 	props->port_cap_flags = rep->cap_mask1;
1417 	if (dev->num_plane) {
1418 		props->port_cap_flags |= IB_PORT_SM_DISABLED;
1419 		props->port_cap_flags &= ~IB_PORT_SM;
1420 	} else if (ibdev->type == RDMA_DEVICE_TYPE_SMI)
1421 		props->port_cap_flags &= ~IB_PORT_CM_SUP;
1422 
1423 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1424 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1425 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1426 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1427 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1428 	props->subnet_timeout	= rep->subnet_timeout;
1429 	props->init_type_reply	= rep->init_type_reply;
1430 
1431 	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1432 		props->port_cap_flags2 = rep->cap_mask2;
1433 
1434 	err = mlx5_query_ib_port_oper(mdev, &ib_link_width_oper,
1435 				      &props->active_speed, port, plane_index);
1436 	if (err)
1437 		goto out;
1438 
1439 	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1440 
1441 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1442 
1443 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1444 
1445 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1446 
1447 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1448 
1449 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1450 	if (err)
1451 		goto out;
1452 
1453 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1454 				   &props->max_vl_num);
1455 out:
1456 	kfree(rep);
1457 	return err;
1458 }
1459 
1460 int mlx5_ib_query_port(struct ib_device *ibdev, u32 port,
1461 		       struct ib_port_attr *props)
1462 {
1463 	unsigned int count;
1464 	int ret;
1465 
1466 	switch (mlx5_get_vport_access_method(ibdev)) {
1467 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1468 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1469 		break;
1470 
1471 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1472 		ret = mlx5_query_hca_port(ibdev, port, props);
1473 		break;
1474 
1475 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1476 		ret = mlx5_query_port_roce(ibdev, port, props);
1477 		break;
1478 
1479 	default:
1480 		ret = -EINVAL;
1481 	}
1482 
1483 	if (!ret && props) {
1484 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1485 		struct mlx5_core_dev *mdev;
1486 		bool put_mdev = true;
1487 
1488 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1489 		if (!mdev) {
1490 			/* If the port isn't affiliated yet query the master.
1491 			 * The master and slave will have the same values.
1492 			 */
1493 			mdev = dev->mdev;
1494 			port = 1;
1495 			put_mdev = false;
1496 		}
1497 		count = mlx5_core_reserved_gids_count(mdev);
1498 		if (put_mdev)
1499 			mlx5_ib_put_native_port_mdev(dev, port);
1500 		props->gid_tbl_len -= count;
1501 	}
1502 	return ret;
1503 }
1504 
1505 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u32 port,
1506 				  struct ib_port_attr *props)
1507 {
1508 	return mlx5_query_port_roce(ibdev, port, props);
1509 }
1510 
1511 static int mlx5_ib_rep_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1512 				  u16 *pkey)
1513 {
1514 	/* Default special Pkey for representor device port as per the
1515 	 * IB specification 1.3 section 10.9.1.2.
1516 	 */
1517 	*pkey = 0xffff;
1518 	return 0;
1519 }
1520 
1521 static int mlx5_ib_query_gid(struct ib_device *ibdev, u32 port, int index,
1522 			     union ib_gid *gid)
1523 {
1524 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1525 	struct mlx5_core_dev *mdev = dev->mdev;
1526 
1527 	switch (mlx5_get_vport_access_method(ibdev)) {
1528 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1529 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1530 
1531 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1532 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1533 
1534 	default:
1535 		return -EINVAL;
1536 	}
1537 
1538 }
1539 
1540 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u32 port,
1541 				   u16 index, u16 *pkey)
1542 {
1543 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1544 	struct mlx5_core_dev *mdev;
1545 	bool put_mdev = true;
1546 	u32 mdev_port_num;
1547 	int err;
1548 
1549 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1550 	if (!mdev) {
1551 		/* The port isn't affiliated yet, get the PKey from the master
1552 		 * port. For RoCE the PKey tables will be the same.
1553 		 */
1554 		put_mdev = false;
1555 		mdev = dev->mdev;
1556 		mdev_port_num = 1;
1557 	}
1558 
1559 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1560 					index, pkey);
1561 	if (put_mdev)
1562 		mlx5_ib_put_native_port_mdev(dev, port);
1563 
1564 	return err;
1565 }
1566 
1567 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1568 			      u16 *pkey)
1569 {
1570 	switch (mlx5_get_vport_access_method(ibdev)) {
1571 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1572 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1573 
1574 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1575 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1576 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1577 	default:
1578 		return -EINVAL;
1579 	}
1580 }
1581 
1582 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1583 				 struct ib_device_modify *props)
1584 {
1585 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1586 	struct mlx5_reg_node_desc in;
1587 	struct mlx5_reg_node_desc out;
1588 	int err;
1589 
1590 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1591 		return -EOPNOTSUPP;
1592 
1593 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1594 		return 0;
1595 
1596 	/*
1597 	 * If possible, pass node desc to FW, so it can generate
1598 	 * a 144 trap.  If cmd fails, just ignore.
1599 	 */
1600 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1601 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1602 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1603 	if (err)
1604 		return err;
1605 
1606 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1607 
1608 	return err;
1609 }
1610 
1611 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u32 port_num, u32 mask,
1612 				u32 value)
1613 {
1614 	struct mlx5_hca_vport_context ctx = {};
1615 	struct mlx5_core_dev *mdev;
1616 	u32 mdev_port_num;
1617 	int err;
1618 
1619 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1620 	if (!mdev)
1621 		return -ENODEV;
1622 
1623 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1624 	if (err)
1625 		goto out;
1626 
1627 	if (~ctx.cap_mask1_perm & mask) {
1628 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1629 			     mask, ctx.cap_mask1_perm);
1630 		err = -EINVAL;
1631 		goto out;
1632 	}
1633 
1634 	ctx.cap_mask1 = value;
1635 	ctx.cap_mask1_perm = mask;
1636 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1637 						 0, &ctx);
1638 
1639 out:
1640 	mlx5_ib_put_native_port_mdev(dev, port_num);
1641 
1642 	return err;
1643 }
1644 
1645 static int mlx5_ib_modify_port(struct ib_device *ibdev, u32 port, int mask,
1646 			       struct ib_port_modify *props)
1647 {
1648 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1649 	struct ib_port_attr attr;
1650 	u32 tmp;
1651 	int err;
1652 	u32 change_mask;
1653 	u32 value;
1654 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1655 		      IB_LINK_LAYER_INFINIBAND);
1656 
1657 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1658 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1659 	 */
1660 	if (!is_ib)
1661 		return 0;
1662 
1663 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1664 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1665 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1666 		return set_port_caps_atomic(dev, port, change_mask, value);
1667 	}
1668 
1669 	mutex_lock(&dev->cap_mask_mutex);
1670 
1671 	err = ib_query_port(ibdev, port, &attr);
1672 	if (err)
1673 		goto out;
1674 
1675 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1676 		~props->clr_port_cap_mask;
1677 
1678 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1679 
1680 out:
1681 	mutex_unlock(&dev->cap_mask_mutex);
1682 	return err;
1683 }
1684 
1685 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1686 {
1687 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1688 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1689 }
1690 
1691 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1692 {
1693 	/* Large page with non 4k uar support might limit the dynamic size */
1694 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1695 		return MLX5_MIN_DYN_BFREGS;
1696 
1697 	return MLX5_MAX_DYN_BFREGS;
1698 }
1699 
1700 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1701 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1702 			     struct mlx5_bfreg_info *bfregi)
1703 {
1704 	int uars_per_sys_page;
1705 	int bfregs_per_sys_page;
1706 	int ref_bfregs = req->total_num_bfregs;
1707 
1708 	if (req->total_num_bfregs == 0)
1709 		return -EINVAL;
1710 
1711 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1712 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1713 
1714 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1715 		return -ENOMEM;
1716 
1717 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1718 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1719 	/* This holds the required static allocation asked by the user */
1720 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1721 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1722 		return -EINVAL;
1723 
1724 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1725 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1726 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1727 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1728 
1729 	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",
1730 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1731 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1732 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1733 		    bfregi->num_sys_pages);
1734 
1735 	return 0;
1736 }
1737 
1738 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1739 {
1740 	struct mlx5_bfreg_info *bfregi;
1741 	int err;
1742 	int i;
1743 
1744 	bfregi = &context->bfregi;
1745 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1746 		err = mlx5_cmd_uar_alloc(dev->mdev, &bfregi->sys_pages[i],
1747 					 context->devx_uid);
1748 		if (err)
1749 			goto error;
1750 
1751 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1752 	}
1753 
1754 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1755 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1756 
1757 	return 0;
1758 
1759 error:
1760 	for (--i; i >= 0; i--)
1761 		if (mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i],
1762 					 context->devx_uid))
1763 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1764 
1765 	return err;
1766 }
1767 
1768 static void deallocate_uars(struct mlx5_ib_dev *dev,
1769 			    struct mlx5_ib_ucontext *context)
1770 {
1771 	struct mlx5_bfreg_info *bfregi;
1772 	int i;
1773 
1774 	bfregi = &context->bfregi;
1775 	for (i = 0; i < bfregi->num_sys_pages; i++)
1776 		if (i < bfregi->num_static_sys_pages ||
1777 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1778 			mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i],
1779 					     context->devx_uid);
1780 }
1781 
1782 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1783 {
1784 	int err = 0;
1785 
1786 	mutex_lock(&dev->lb.mutex);
1787 	if (td)
1788 		dev->lb.user_td++;
1789 	if (qp)
1790 		dev->lb.qps++;
1791 
1792 	if (dev->lb.user_td == 2 ||
1793 	    dev->lb.qps == 1) {
1794 		if (!dev->lb.enabled) {
1795 			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1796 			dev->lb.enabled = true;
1797 		}
1798 	}
1799 
1800 	mutex_unlock(&dev->lb.mutex);
1801 
1802 	return err;
1803 }
1804 
1805 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1806 {
1807 	mutex_lock(&dev->lb.mutex);
1808 	if (td)
1809 		dev->lb.user_td--;
1810 	if (qp)
1811 		dev->lb.qps--;
1812 
1813 	if (dev->lb.user_td == 1 &&
1814 	    dev->lb.qps == 0) {
1815 		if (dev->lb.enabled) {
1816 			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1817 			dev->lb.enabled = false;
1818 		}
1819 	}
1820 
1821 	mutex_unlock(&dev->lb.mutex);
1822 }
1823 
1824 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1825 					  u16 uid)
1826 {
1827 	int err;
1828 
1829 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1830 		return 0;
1831 
1832 	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1833 	if (err)
1834 		return err;
1835 
1836 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1837 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1838 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1839 		return err;
1840 
1841 	return mlx5_ib_enable_lb(dev, true, false);
1842 }
1843 
1844 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1845 					     u16 uid)
1846 {
1847 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1848 		return;
1849 
1850 	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1851 
1852 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1853 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1854 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1855 		return;
1856 
1857 	mlx5_ib_disable_lb(dev, true, false);
1858 }
1859 
1860 static int set_ucontext_resp(struct ib_ucontext *uctx,
1861 			     struct mlx5_ib_alloc_ucontext_resp *resp)
1862 {
1863 	struct ib_device *ibdev = uctx->device;
1864 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1865 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1866 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1867 
1868 	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1869 		resp->dump_fill_mkey = dev->mkeys.dump_fill_mkey;
1870 		resp->comp_mask |=
1871 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1872 	}
1873 
1874 	resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1875 	if (mlx5_wc_support_get(dev->mdev))
1876 		resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev,
1877 						      log_bf_reg_size);
1878 	resp->cache_line_size = cache_line_size();
1879 	resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1880 	resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1881 	resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1882 	resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1883 	resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1884 	resp->cqe_version = context->cqe_version;
1885 	resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1886 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1887 	resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1888 					MLX5_CAP_GEN(dev->mdev,
1889 						     num_of_uars_per_page) : 1;
1890 	resp->tot_bfregs = bfregi->lib_uar_dyn ? 0 :
1891 			bfregi->total_num_bfregs - bfregi->num_dyn_bfregs;
1892 	resp->num_ports = dev->num_ports;
1893 	resp->cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1894 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1895 
1896 	if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1897 		mlx5_query_min_inline(dev->mdev, &resp->eth_min_inline);
1898 		resp->eth_min_inline++;
1899 	}
1900 
1901 	if (dev->mdev->clock_info)
1902 		resp->clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1903 
1904 	/*
1905 	 * We don't want to expose information from the PCI bar that is located
1906 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1907 	 * pretend we don't support reading the HCA's core clock. This is also
1908 	 * forced by mmap function.
1909 	 */
1910 	if (PAGE_SIZE <= 4096) {
1911 		resp->comp_mask |=
1912 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1913 		resp->hca_core_clock_offset =
1914 			offsetof(struct mlx5_init_seg,
1915 				 internal_timer_h) % PAGE_SIZE;
1916 	}
1917 
1918 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
1919 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE;
1920 
1921 	if (rt_supported(MLX5_CAP_GEN(dev->mdev, sq_ts_format)) &&
1922 	    rt_supported(MLX5_CAP_GEN(dev->mdev, rq_ts_format)) &&
1923 	    rt_supported(MLX5_CAP_ROCE(dev->mdev, qp_ts_format)))
1924 		resp->comp_mask |=
1925 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_REAL_TIME_TS;
1926 
1927 	resp->num_dyn_bfregs = bfregi->num_dyn_bfregs;
1928 
1929 	if (MLX5_CAP_GEN(dev->mdev, drain_sigerr))
1930 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_SQD2RTS;
1931 
1932 	resp->comp_mask |=
1933 		MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_MKEY_UPDATE_TAG;
1934 
1935 	return 0;
1936 }
1937 
1938 static bool uctx_rdma_ctrl_is_enabled(u64 enabled_caps)
1939 {
1940 	return UCAP_ENABLED(enabled_caps, RDMA_UCAP_MLX5_CTRL_LOCAL) ||
1941 	       UCAP_ENABLED(enabled_caps, RDMA_UCAP_MLX5_CTRL_OTHER_VHCA);
1942 }
1943 
1944 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1945 				  struct ib_udata *udata)
1946 {
1947 	struct ib_device *ibdev = uctx->device;
1948 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1949 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1950 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1951 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1952 	struct mlx5_bfreg_info *bfregi;
1953 	int ver;
1954 	int err;
1955 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1956 				     max_cqe_version);
1957 	bool lib_uar_4k;
1958 	bool lib_uar_dyn;
1959 
1960 	if (!dev->ib_active)
1961 		return -EAGAIN;
1962 
1963 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1964 		ver = 0;
1965 	else if (udata->inlen >= min_req_v2)
1966 		ver = 2;
1967 	else
1968 		return -EINVAL;
1969 
1970 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1971 	if (err)
1972 		return err;
1973 
1974 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1975 		return -EOPNOTSUPP;
1976 
1977 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1978 		return -EOPNOTSUPP;
1979 
1980 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1981 				    MLX5_NON_FP_BFREGS_PER_UAR);
1982 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1983 		return -EINVAL;
1984 
1985 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1986 		err = mlx5_ib_devx_create(dev, true, uctx->enabled_caps);
1987 		if (err < 0)
1988 			goto out_ctx;
1989 		context->devx_uid = err;
1990 
1991 		if (uctx_rdma_ctrl_is_enabled(uctx->enabled_caps)) {
1992 			err = mlx5_cmd_add_privileged_uid(dev->mdev,
1993 							  context->devx_uid);
1994 			if (err)
1995 				goto out_devx;
1996 		}
1997 	}
1998 
1999 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
2000 	lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
2001 	bfregi = &context->bfregi;
2002 
2003 	if (lib_uar_dyn) {
2004 		bfregi->lib_uar_dyn = lib_uar_dyn;
2005 		goto uar_done;
2006 	}
2007 
2008 	/* updates req->total_num_bfregs */
2009 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
2010 	if (err)
2011 		goto out_ucap;
2012 
2013 	mutex_init(&bfregi->lock);
2014 	bfregi->lib_uar_4k = lib_uar_4k;
2015 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
2016 				GFP_KERNEL);
2017 	if (!bfregi->count) {
2018 		err = -ENOMEM;
2019 		goto out_ucap;
2020 	}
2021 
2022 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
2023 				    sizeof(*bfregi->sys_pages),
2024 				    GFP_KERNEL);
2025 	if (!bfregi->sys_pages) {
2026 		err = -ENOMEM;
2027 		goto out_count;
2028 	}
2029 
2030 	err = allocate_uars(dev, context);
2031 	if (err)
2032 		goto out_sys_pages;
2033 
2034 uar_done:
2035 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
2036 					     context->devx_uid);
2037 	if (err)
2038 		goto out_uars;
2039 
2040 	INIT_LIST_HEAD(&context->db_page_list);
2041 	mutex_init(&context->db_page_mutex);
2042 
2043 	context->cqe_version = min_t(__u8,
2044 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
2045 				 req.max_cqe_version);
2046 
2047 	err = set_ucontext_resp(uctx, &resp);
2048 	if (err)
2049 		goto out_mdev;
2050 
2051 	resp.response_length = min(udata->outlen, sizeof(resp));
2052 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
2053 	if (err)
2054 		goto out_mdev;
2055 
2056 	bfregi->ver = ver;
2057 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
2058 	context->lib_caps = req.lib_caps;
2059 	print_lib_caps(dev, context->lib_caps);
2060 
2061 	if (mlx5_ib_lag_should_assign_affinity(dev)) {
2062 		u32 port = mlx5_core_native_port_num(dev->mdev) - 1;
2063 
2064 		atomic_set(&context->tx_port_affinity,
2065 			   atomic_add_return(
2066 				   1, &dev->port[port].roce.tx_port_affinity));
2067 	}
2068 
2069 	return 0;
2070 
2071 out_mdev:
2072 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
2073 
2074 out_uars:
2075 	deallocate_uars(dev, context);
2076 
2077 out_sys_pages:
2078 	kfree(bfregi->sys_pages);
2079 
2080 out_count:
2081 	kfree(bfregi->count);
2082 
2083 out_ucap:
2084 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX &&
2085 	    uctx_rdma_ctrl_is_enabled(uctx->enabled_caps))
2086 		mlx5_cmd_remove_privileged_uid(dev->mdev, context->devx_uid);
2087 
2088 out_devx:
2089 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
2090 		mlx5_ib_devx_destroy(dev, context->devx_uid);
2091 
2092 out_ctx:
2093 	return err;
2094 }
2095 
2096 static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext,
2097 				  struct uverbs_attr_bundle *attrs)
2098 {
2099 	struct mlx5_ib_alloc_ucontext_resp uctx_resp = {};
2100 	int ret;
2101 
2102 	ret = set_ucontext_resp(ibcontext, &uctx_resp);
2103 	if (ret)
2104 		return ret;
2105 
2106 	uctx_resp.response_length =
2107 		min_t(size_t,
2108 		      uverbs_attr_get_len(attrs,
2109 				MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX),
2110 		      sizeof(uctx_resp));
2111 
2112 	ret = uverbs_copy_to_struct_or_zero(attrs,
2113 					MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
2114 					&uctx_resp,
2115 					sizeof(uctx_resp));
2116 	return ret;
2117 }
2118 
2119 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
2120 {
2121 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2122 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2123 	struct mlx5_bfreg_info *bfregi;
2124 
2125 	bfregi = &context->bfregi;
2126 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
2127 
2128 	deallocate_uars(dev, context);
2129 	kfree(bfregi->sys_pages);
2130 	kfree(bfregi->count);
2131 
2132 	if (context->devx_uid) {
2133 		if (uctx_rdma_ctrl_is_enabled(ibcontext->enabled_caps))
2134 			mlx5_cmd_remove_privileged_uid(dev->mdev,
2135 						       context->devx_uid);
2136 		mlx5_ib_devx_destroy(dev, context->devx_uid);
2137 	}
2138 }
2139 
2140 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2141 				 int uar_idx)
2142 {
2143 	int fw_uars_per_page;
2144 
2145 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2146 
2147 	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2148 }
2149 
2150 static u64 uar_index2paddress(struct mlx5_ib_dev *dev,
2151 				 int uar_idx)
2152 {
2153 	unsigned int fw_uars_per_page;
2154 
2155 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
2156 				MLX5_UARS_IN_PAGE : 1;
2157 
2158 	return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE);
2159 }
2160 
2161 static int get_command(unsigned long offset)
2162 {
2163 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2164 }
2165 
2166 static int get_arg(unsigned long offset)
2167 {
2168 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2169 }
2170 
2171 static int get_index(unsigned long offset)
2172 {
2173 	return get_arg(offset);
2174 }
2175 
2176 /* Index resides in an extra byte to enable larger values than 255 */
2177 static int get_extended_index(unsigned long offset)
2178 {
2179 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2180 }
2181 
2182 
2183 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2184 {
2185 }
2186 
2187 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2188 {
2189 	switch (cmd) {
2190 	case MLX5_IB_MMAP_WC_PAGE:
2191 		return "WC";
2192 	case MLX5_IB_MMAP_REGULAR_PAGE:
2193 		return "best effort WC";
2194 	case MLX5_IB_MMAP_NC_PAGE:
2195 		return "NC";
2196 	case MLX5_IB_MMAP_DEVICE_MEM:
2197 		return "Device Memory";
2198 	default:
2199 		return "Unknown";
2200 	}
2201 }
2202 
2203 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2204 					struct vm_area_struct *vma,
2205 					struct mlx5_ib_ucontext *context)
2206 {
2207 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2208 	    !(vma->vm_flags & VM_SHARED))
2209 		return -EINVAL;
2210 
2211 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2212 		return -EOPNOTSUPP;
2213 
2214 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2215 		return -EPERM;
2216 	vm_flags_clear(vma, VM_MAYWRITE);
2217 
2218 	if (!dev->mdev->clock_info)
2219 		return -EOPNOTSUPP;
2220 
2221 	return vm_insert_page(vma, vma->vm_start,
2222 			      virt_to_page(dev->mdev->clock_info));
2223 }
2224 
2225 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2226 {
2227 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2228 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2229 	struct mlx5_var_table *var_table = &dev->var_table;
2230 	struct mlx5_ib_ucontext *context = to_mucontext(entry->ucontext);
2231 
2232 	switch (mentry->mmap_flag) {
2233 	case MLX5_IB_MMAP_TYPE_MEMIC:
2234 	case MLX5_IB_MMAP_TYPE_MEMIC_OP:
2235 		mlx5_ib_dm_mmap_free(dev, mentry);
2236 		break;
2237 	case MLX5_IB_MMAP_TYPE_VAR:
2238 		mutex_lock(&var_table->bitmap_lock);
2239 		clear_bit(mentry->page_idx, var_table->bitmap);
2240 		mutex_unlock(&var_table->bitmap_lock);
2241 		kfree(mentry);
2242 		break;
2243 	case MLX5_IB_MMAP_TYPE_UAR_WC:
2244 	case MLX5_IB_MMAP_TYPE_UAR_NC:
2245 		mlx5_cmd_uar_dealloc(dev->mdev, mentry->page_idx,
2246 				     context->devx_uid);
2247 		kfree(mentry);
2248 		break;
2249 	default:
2250 		WARN_ON(true);
2251 	}
2252 }
2253 
2254 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2255 		    struct vm_area_struct *vma,
2256 		    struct mlx5_ib_ucontext *context)
2257 {
2258 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2259 	int err;
2260 	unsigned long idx;
2261 	phys_addr_t pfn;
2262 	pgprot_t prot;
2263 	u32 bfreg_dyn_idx = 0;
2264 	u32 uar_index;
2265 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2266 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2267 				bfregi->num_static_sys_pages;
2268 
2269 	if (bfregi->lib_uar_dyn)
2270 		return -EINVAL;
2271 
2272 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2273 		return -EINVAL;
2274 
2275 	if (dyn_uar)
2276 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2277 	else
2278 		idx = get_index(vma->vm_pgoff);
2279 
2280 	if (idx >= max_valid_idx) {
2281 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2282 			     idx, max_valid_idx);
2283 		return -EINVAL;
2284 	}
2285 
2286 	switch (cmd) {
2287 	case MLX5_IB_MMAP_WC_PAGE:
2288 	case MLX5_IB_MMAP_ALLOC_WC:
2289 	case MLX5_IB_MMAP_REGULAR_PAGE:
2290 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2291 		prot = pgprot_writecombine(vma->vm_page_prot);
2292 		break;
2293 	case MLX5_IB_MMAP_NC_PAGE:
2294 		prot = pgprot_noncached(vma->vm_page_prot);
2295 		break;
2296 	default:
2297 		return -EINVAL;
2298 	}
2299 
2300 	if (dyn_uar) {
2301 		int uars_per_page;
2302 
2303 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2304 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2305 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2306 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2307 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2308 			return -EINVAL;
2309 		}
2310 
2311 		mutex_lock(&bfregi->lock);
2312 		/* Fail if uar already allocated, first bfreg index of each
2313 		 * page holds its count.
2314 		 */
2315 		if (bfregi->count[bfreg_dyn_idx]) {
2316 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2317 			mutex_unlock(&bfregi->lock);
2318 			return -EINVAL;
2319 		}
2320 
2321 		bfregi->count[bfreg_dyn_idx]++;
2322 		mutex_unlock(&bfregi->lock);
2323 
2324 		err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index,
2325 					 context->devx_uid);
2326 		if (err) {
2327 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2328 			goto free_bfreg;
2329 		}
2330 	} else {
2331 		uar_index = bfregi->sys_pages[idx];
2332 	}
2333 
2334 	pfn = uar_index2pfn(dev, uar_index);
2335 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2336 
2337 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2338 				prot, NULL);
2339 	if (err) {
2340 		mlx5_ib_err(dev,
2341 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2342 			    err, mmap_cmd2str(cmd));
2343 		goto err;
2344 	}
2345 
2346 	if (dyn_uar)
2347 		bfregi->sys_pages[idx] = uar_index;
2348 	return 0;
2349 
2350 err:
2351 	if (!dyn_uar)
2352 		return err;
2353 
2354 	mlx5_cmd_uar_dealloc(dev->mdev, idx, context->devx_uid);
2355 
2356 free_bfreg:
2357 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2358 
2359 	return err;
2360 }
2361 
2362 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2363 {
2364 	unsigned long idx;
2365 	u8 command;
2366 
2367 	command = get_command(vma->vm_pgoff);
2368 	idx = get_extended_index(vma->vm_pgoff);
2369 
2370 	return (command << 16 | idx);
2371 }
2372 
2373 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2374 			       struct vm_area_struct *vma,
2375 			       struct ib_ucontext *ucontext)
2376 {
2377 	struct mlx5_user_mmap_entry *mentry;
2378 	struct rdma_user_mmap_entry *entry;
2379 	unsigned long pgoff;
2380 	pgprot_t prot;
2381 	phys_addr_t pfn;
2382 	int ret;
2383 
2384 	pgoff = mlx5_vma_to_pgoff(vma);
2385 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2386 	if (!entry)
2387 		return -EINVAL;
2388 
2389 	mentry = to_mmmap(entry);
2390 	pfn = (mentry->address >> PAGE_SHIFT);
2391 	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
2392 	    mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
2393 		prot = pgprot_noncached(vma->vm_page_prot);
2394 	else
2395 		prot = pgprot_writecombine(vma->vm_page_prot);
2396 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2397 				entry->npages * PAGE_SIZE,
2398 				prot,
2399 				entry);
2400 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2401 	return ret;
2402 }
2403 
2404 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2405 {
2406 	u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2407 	u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2408 
2409 	return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2410 		(index & 0xFF)) << PAGE_SHIFT;
2411 }
2412 
2413 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2414 {
2415 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2416 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2417 	unsigned long command;
2418 	phys_addr_t pfn;
2419 
2420 	command = get_command(vma->vm_pgoff);
2421 	switch (command) {
2422 	case MLX5_IB_MMAP_WC_PAGE:
2423 	case MLX5_IB_MMAP_ALLOC_WC:
2424 		if (!mlx5_wc_support_get(dev->mdev))
2425 			return -EPERM;
2426 		fallthrough;
2427 	case MLX5_IB_MMAP_NC_PAGE:
2428 	case MLX5_IB_MMAP_REGULAR_PAGE:
2429 		return uar_mmap(dev, command, vma, context);
2430 
2431 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2432 		return -ENOSYS;
2433 
2434 	case MLX5_IB_MMAP_CORE_CLOCK:
2435 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2436 			return -EINVAL;
2437 
2438 		if (vma->vm_flags & VM_WRITE)
2439 			return -EPERM;
2440 		vm_flags_clear(vma, VM_MAYWRITE);
2441 
2442 		/* Don't expose to user-space information it shouldn't have */
2443 		if (PAGE_SIZE > 4096)
2444 			return -EOPNOTSUPP;
2445 
2446 		pfn = (dev->mdev->iseg_base +
2447 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2448 			PAGE_SHIFT;
2449 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2450 					 PAGE_SIZE,
2451 					 pgprot_noncached(vma->vm_page_prot),
2452 					 NULL);
2453 	case MLX5_IB_MMAP_CLOCK_INFO:
2454 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2455 
2456 	default:
2457 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2458 	}
2459 
2460 	return 0;
2461 }
2462 
2463 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2464 {
2465 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2466 	struct ib_device *ibdev = ibpd->device;
2467 	struct mlx5_ib_alloc_pd_resp resp;
2468 	int err;
2469 	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2470 	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {};
2471 	u16 uid = 0;
2472 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2473 		udata, struct mlx5_ib_ucontext, ibucontext);
2474 
2475 	uid = context ? context->devx_uid : 0;
2476 	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2477 	MLX5_SET(alloc_pd_in, in, uid, uid);
2478 	err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out);
2479 	if (err)
2480 		return err;
2481 
2482 	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2483 	pd->uid = uid;
2484 	if (udata) {
2485 		resp.pdn = pd->pdn;
2486 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2487 			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2488 			return -EFAULT;
2489 		}
2490 	}
2491 
2492 	return 0;
2493 }
2494 
2495 static int mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2496 {
2497 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2498 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2499 
2500 	return mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2501 }
2502 
2503 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2504 {
2505 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2506 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
2507 	int err;
2508 	u16 uid;
2509 
2510 	uid = ibqp->pd ?
2511 		to_mpd(ibqp->pd)->uid : 0;
2512 
2513 	if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) {
2514 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
2515 		return -EOPNOTSUPP;
2516 	}
2517 
2518 	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2519 	if (err)
2520 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2521 			     ibqp->qp_num, gid->raw);
2522 
2523 	return err;
2524 }
2525 
2526 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2527 {
2528 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2529 	int err;
2530 	u16 uid;
2531 
2532 	uid = ibqp->pd ?
2533 		to_mpd(ibqp->pd)->uid : 0;
2534 	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2535 	if (err)
2536 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2537 			     ibqp->qp_num, gid->raw);
2538 
2539 	return err;
2540 }
2541 
2542 static int init_node_data(struct mlx5_ib_dev *dev)
2543 {
2544 	int err;
2545 
2546 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2547 	if (err)
2548 		return err;
2549 
2550 	dev->mdev->rev_id = dev->mdev->pdev->revision;
2551 
2552 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2553 }
2554 
2555 static ssize_t fw_pages_show(struct device *device,
2556 			     struct device_attribute *attr, char *buf)
2557 {
2558 	struct mlx5_ib_dev *dev =
2559 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2560 
2561 	return sysfs_emit(buf, "%d\n", dev->mdev->priv.fw_pages);
2562 }
2563 static DEVICE_ATTR_RO(fw_pages);
2564 
2565 static ssize_t reg_pages_show(struct device *device,
2566 			      struct device_attribute *attr, char *buf)
2567 {
2568 	struct mlx5_ib_dev *dev =
2569 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2570 
2571 	return sysfs_emit(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2572 }
2573 static DEVICE_ATTR_RO(reg_pages);
2574 
2575 static ssize_t hca_type_show(struct device *device,
2576 			     struct device_attribute *attr, char *buf)
2577 {
2578 	struct mlx5_ib_dev *dev =
2579 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2580 
2581 	return sysfs_emit(buf, "MT%d\n", dev->mdev->pdev->device);
2582 }
2583 static DEVICE_ATTR_RO(hca_type);
2584 
2585 static ssize_t hw_rev_show(struct device *device,
2586 			   struct device_attribute *attr, char *buf)
2587 {
2588 	struct mlx5_ib_dev *dev =
2589 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2590 
2591 	return sysfs_emit(buf, "%x\n", dev->mdev->rev_id);
2592 }
2593 static DEVICE_ATTR_RO(hw_rev);
2594 
2595 static ssize_t board_id_show(struct device *device,
2596 			     struct device_attribute *attr, char *buf)
2597 {
2598 	struct mlx5_ib_dev *dev =
2599 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2600 
2601 	return sysfs_emit(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2602 			  dev->mdev->board_id);
2603 }
2604 static DEVICE_ATTR_RO(board_id);
2605 
2606 static struct attribute *mlx5_class_attributes[] = {
2607 	&dev_attr_hw_rev.attr,
2608 	&dev_attr_hca_type.attr,
2609 	&dev_attr_board_id.attr,
2610 	&dev_attr_fw_pages.attr,
2611 	&dev_attr_reg_pages.attr,
2612 	NULL,
2613 };
2614 
2615 static const struct attribute_group mlx5_attr_group = {
2616 	.attrs = mlx5_class_attributes,
2617 };
2618 
2619 static void pkey_change_handler(struct work_struct *work)
2620 {
2621 	struct mlx5_ib_port_resources *ports =
2622 		container_of(work, struct mlx5_ib_port_resources,
2623 			     pkey_change_work);
2624 
2625 	if (!ports->gsi)
2626 		/*
2627 		 * We got this event before device was fully configured
2628 		 * and MAD registration code wasn't called/finished yet.
2629 		 */
2630 		return;
2631 
2632 	mlx5_ib_gsi_pkey_change(ports->gsi);
2633 }
2634 
2635 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2636 {
2637 	struct mlx5_ib_qp *mqp;
2638 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2639 	struct mlx5_core_cq *mcq;
2640 	struct list_head cq_armed_list;
2641 	unsigned long flags_qp;
2642 	unsigned long flags_cq;
2643 	unsigned long flags;
2644 
2645 	INIT_LIST_HEAD(&cq_armed_list);
2646 
2647 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2648 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2649 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2650 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2651 		if (mqp->sq.tail != mqp->sq.head) {
2652 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2653 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2654 			if (send_mcq->mcq.comp &&
2655 			    mqp->ibqp.send_cq->comp_handler) {
2656 				if (!send_mcq->mcq.reset_notify_added) {
2657 					send_mcq->mcq.reset_notify_added = 1;
2658 					list_add_tail(&send_mcq->mcq.reset_notify,
2659 						      &cq_armed_list);
2660 				}
2661 			}
2662 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2663 		}
2664 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2665 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2666 		/* no handling is needed for SRQ */
2667 		if (!mqp->ibqp.srq) {
2668 			if (mqp->rq.tail != mqp->rq.head) {
2669 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2670 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2671 				if (recv_mcq->mcq.comp &&
2672 				    mqp->ibqp.recv_cq->comp_handler) {
2673 					if (!recv_mcq->mcq.reset_notify_added) {
2674 						recv_mcq->mcq.reset_notify_added = 1;
2675 						list_add_tail(&recv_mcq->mcq.reset_notify,
2676 							      &cq_armed_list);
2677 					}
2678 				}
2679 				spin_unlock_irqrestore(&recv_mcq->lock,
2680 						       flags_cq);
2681 			}
2682 		}
2683 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2684 	}
2685 	/*At that point all inflight post send were put to be executed as of we
2686 	 * lock/unlock above locks Now need to arm all involved CQs.
2687 	 */
2688 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2689 		mcq->comp(mcq, NULL);
2690 	}
2691 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2692 }
2693 
2694 static void delay_drop_handler(struct work_struct *work)
2695 {
2696 	int err;
2697 	struct mlx5_ib_delay_drop *delay_drop =
2698 		container_of(work, struct mlx5_ib_delay_drop,
2699 			     delay_drop_work);
2700 
2701 	atomic_inc(&delay_drop->events_cnt);
2702 
2703 	mutex_lock(&delay_drop->lock);
2704 	err = mlx5_core_set_delay_drop(delay_drop->dev, delay_drop->timeout);
2705 	if (err) {
2706 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
2707 			     delay_drop->timeout);
2708 		delay_drop->activate = false;
2709 	}
2710 	mutex_unlock(&delay_drop->lock);
2711 }
2712 
2713 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2714 				 struct ib_event *ibev)
2715 {
2716 	u32 port = (eqe->data.port.port >> 4) & 0xf;
2717 
2718 	switch (eqe->sub_type) {
2719 	case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
2720 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2721 					    IB_LINK_LAYER_ETHERNET)
2722 			schedule_work(&ibdev->delay_drop.delay_drop_work);
2723 		break;
2724 	default: /* do nothing */
2725 		return;
2726 	}
2727 }
2728 
2729 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2730 			      struct ib_event *ibev)
2731 {
2732 	u32 port = (eqe->data.port.port >> 4) & 0xf;
2733 
2734 	ibev->element.port_num = port;
2735 
2736 	switch (eqe->sub_type) {
2737 	case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
2738 	case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
2739 	case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
2740 		/* In RoCE, port up/down events are handled in
2741 		 * mlx5_netdev_event().
2742 		 */
2743 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2744 					    IB_LINK_LAYER_ETHERNET)
2745 			return -EINVAL;
2746 
2747 		ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
2748 				IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2749 		break;
2750 
2751 	case MLX5_PORT_CHANGE_SUBTYPE_LID:
2752 		ibev->event = IB_EVENT_LID_CHANGE;
2753 		break;
2754 
2755 	case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
2756 		ibev->event = IB_EVENT_PKEY_CHANGE;
2757 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2758 		break;
2759 
2760 	case MLX5_PORT_CHANGE_SUBTYPE_GUID:
2761 		ibev->event = IB_EVENT_GID_CHANGE;
2762 		break;
2763 
2764 	case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
2765 		ibev->event = IB_EVENT_CLIENT_REREGISTER;
2766 		break;
2767 	default:
2768 		return -EINVAL;
2769 	}
2770 
2771 	return 0;
2772 }
2773 
2774 static void mlx5_ib_handle_event(struct work_struct *_work)
2775 {
2776 	struct mlx5_ib_event_work *work =
2777 		container_of(_work, struct mlx5_ib_event_work, work);
2778 	struct mlx5_ib_dev *ibdev;
2779 	struct ib_event ibev;
2780 	bool fatal = false;
2781 
2782 	if (work->is_slave) {
2783 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
2784 		if (!ibdev)
2785 			goto out;
2786 	} else {
2787 		ibdev = work->dev;
2788 	}
2789 
2790 	switch (work->event) {
2791 	case MLX5_DEV_EVENT_SYS_ERROR:
2792 		ibev.event = IB_EVENT_DEVICE_FATAL;
2793 		mlx5_ib_handle_internal_error(ibdev);
2794 		ibev.element.port_num  = (u8)(unsigned long)work->param;
2795 		fatal = true;
2796 		break;
2797 	case MLX5_EVENT_TYPE_PORT_CHANGE:
2798 		if (handle_port_change(ibdev, work->param, &ibev))
2799 			goto out;
2800 		break;
2801 	case MLX5_EVENT_TYPE_GENERAL_EVENT:
2802 		handle_general_event(ibdev, work->param, &ibev);
2803 		fallthrough;
2804 	default:
2805 		goto out;
2806 	}
2807 
2808 	ibev.device = &ibdev->ib_dev;
2809 
2810 	if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
2811 		mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
2812 		goto out;
2813 	}
2814 
2815 	if (ibdev->ib_active)
2816 		ib_dispatch_event(&ibev);
2817 
2818 	if (fatal)
2819 		ibdev->ib_active = false;
2820 out:
2821 	kfree(work);
2822 }
2823 
2824 static int mlx5_ib_event(struct notifier_block *nb,
2825 			 unsigned long event, void *param)
2826 {
2827 	struct mlx5_ib_event_work *work;
2828 
2829 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2830 	if (!work)
2831 		return NOTIFY_DONE;
2832 
2833 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2834 	work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
2835 	work->is_slave = false;
2836 	work->param = param;
2837 	work->event = event;
2838 
2839 	queue_work(mlx5_ib_event_wq, &work->work);
2840 
2841 	return NOTIFY_OK;
2842 }
2843 
2844 static int mlx5_ib_event_slave_port(struct notifier_block *nb,
2845 				    unsigned long event, void *param)
2846 {
2847 	struct mlx5_ib_event_work *work;
2848 
2849 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2850 	if (!work)
2851 		return NOTIFY_DONE;
2852 
2853 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2854 	work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
2855 	work->is_slave = true;
2856 	work->param = param;
2857 	work->event = event;
2858 	queue_work(mlx5_ib_event_wq, &work->work);
2859 
2860 	return NOTIFY_OK;
2861 }
2862 
2863 static int mlx5_ib_get_plane_num(struct mlx5_core_dev *mdev, u8 *num_plane)
2864 {
2865 	struct mlx5_hca_vport_context vport_ctx;
2866 	int err;
2867 
2868 	*num_plane = 0;
2869 	if (!MLX5_CAP_GEN(mdev, ib_virt) || !MLX5_CAP_GEN_2(mdev, multiplane))
2870 		return 0;
2871 
2872 	err = mlx5_query_hca_vport_context(mdev, 0, 1, 0, &vport_ctx);
2873 	if (err)
2874 		return err;
2875 
2876 	*num_plane = vport_ctx.num_plane;
2877 	return 0;
2878 }
2879 
2880 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2881 {
2882 	struct mlx5_hca_vport_context vport_ctx;
2883 	int err;
2884 	int port;
2885 
2886 	if (MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_IB)
2887 		return 0;
2888 
2889 	for (port = 1; port <= dev->num_ports; port++) {
2890 		if (dev->num_plane) {
2891 			dev->port_caps[port - 1].has_smi = false;
2892 			continue;
2893 		} else if (!MLX5_CAP_GEN(dev->mdev, ib_virt) ||
2894 			dev->ib_dev.type == RDMA_DEVICE_TYPE_SMI) {
2895 			dev->port_caps[port - 1].has_smi = true;
2896 			continue;
2897 		}
2898 
2899 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port, 0,
2900 						   &vport_ctx);
2901 		if (err) {
2902 			mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2903 				    port, err);
2904 			return err;
2905 		}
2906 		dev->port_caps[port - 1].has_smi = vport_ctx.has_smi;
2907 	}
2908 
2909 	return 0;
2910 }
2911 
2912 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2913 {
2914 	unsigned int port;
2915 
2916 	rdma_for_each_port (&dev->ib_dev, port)
2917 		mlx5_query_ext_port_caps(dev, port);
2918 }
2919 
2920 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
2921 {
2922 	switch (umr_fence_cap) {
2923 	case MLX5_CAP_UMR_FENCE_NONE:
2924 		return MLX5_FENCE_MODE_NONE;
2925 	case MLX5_CAP_UMR_FENCE_SMALL:
2926 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
2927 	default:
2928 		return MLX5_FENCE_MODE_STRONG_ORDERING;
2929 	}
2930 }
2931 
2932 int mlx5_ib_dev_res_cq_init(struct mlx5_ib_dev *dev)
2933 {
2934 	struct mlx5_ib_resources *devr = &dev->devr;
2935 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2936 	struct ib_device *ibdev;
2937 	struct ib_pd *pd;
2938 	struct ib_cq *cq;
2939 	int ret = 0;
2940 
2941 
2942 	/*
2943 	 * devr->c0 is set once, never changed until device unload.
2944 	 * Avoid taking the mutex if initialization is already done.
2945 	 */
2946 	if (devr->c0)
2947 		return 0;
2948 
2949 	mutex_lock(&devr->cq_lock);
2950 	if (devr->c0)
2951 		goto unlock;
2952 
2953 	ibdev = &dev->ib_dev;
2954 	pd = ib_alloc_pd(ibdev, 0);
2955 	if (IS_ERR(pd)) {
2956 		ret = PTR_ERR(pd);
2957 		mlx5_ib_err(dev, "Couldn't allocate PD for res init, err=%d\n", ret);
2958 		goto unlock;
2959 	}
2960 
2961 	cq = ib_create_cq(ibdev, NULL, NULL, NULL, &cq_attr);
2962 	if (IS_ERR(cq)) {
2963 		ret = PTR_ERR(cq);
2964 		mlx5_ib_err(dev, "Couldn't create CQ for res init, err=%d\n", ret);
2965 		ib_dealloc_pd(pd);
2966 		goto unlock;
2967 	}
2968 
2969 	devr->p0 = pd;
2970 	devr->c0 = cq;
2971 
2972 unlock:
2973 	mutex_unlock(&devr->cq_lock);
2974 	return ret;
2975 }
2976 
2977 int mlx5_ib_dev_res_srq_init(struct mlx5_ib_dev *dev)
2978 {
2979 	struct mlx5_ib_resources *devr = &dev->devr;
2980 	struct ib_srq_init_attr attr;
2981 	struct ib_srq *s0, *s1;
2982 	int ret = 0;
2983 
2984 	/*
2985 	 * devr->s1 is set once, never changed until device unload.
2986 	 * Avoid taking the mutex if initialization is already done.
2987 	 */
2988 	if (devr->s1)
2989 		return 0;
2990 
2991 	mutex_lock(&devr->srq_lock);
2992 	if (devr->s1)
2993 		goto unlock;
2994 
2995 	ret = mlx5_ib_dev_res_cq_init(dev);
2996 	if (ret)
2997 		goto unlock;
2998 
2999 	memset(&attr, 0, sizeof(attr));
3000 	attr.attr.max_sge = 1;
3001 	attr.attr.max_wr = 1;
3002 	attr.srq_type = IB_SRQT_XRC;
3003 	attr.ext.cq = devr->c0;
3004 
3005 	s0 = ib_create_srq(devr->p0, &attr);
3006 	if (IS_ERR(s0)) {
3007 		ret = PTR_ERR(s0);
3008 		mlx5_ib_err(dev, "Couldn't create SRQ 0 for res init, err=%d\n", ret);
3009 		goto unlock;
3010 	}
3011 
3012 	memset(&attr, 0, sizeof(attr));
3013 	attr.attr.max_sge = 1;
3014 	attr.attr.max_wr = 1;
3015 	attr.srq_type = IB_SRQT_BASIC;
3016 
3017 	s1 = ib_create_srq(devr->p0, &attr);
3018 	if (IS_ERR(s1)) {
3019 		ret = PTR_ERR(s1);
3020 		mlx5_ib_err(dev, "Couldn't create SRQ 1 for res init, err=%d\n", ret);
3021 		ib_destroy_srq(s0);
3022 	}
3023 
3024 	devr->s0 = s0;
3025 	devr->s1 = s1;
3026 
3027 unlock:
3028 	mutex_unlock(&devr->srq_lock);
3029 	return ret;
3030 }
3031 
3032 static int mlx5_ib_dev_res_init(struct mlx5_ib_dev *dev)
3033 {
3034 	struct mlx5_ib_resources *devr = &dev->devr;
3035 	int ret;
3036 
3037 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
3038 		return -EOPNOTSUPP;
3039 
3040 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0);
3041 	if (ret)
3042 		return ret;
3043 
3044 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0);
3045 	if (ret) {
3046 		mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
3047 		return ret;
3048 	}
3049 
3050 	mutex_init(&devr->cq_lock);
3051 	mutex_init(&devr->srq_lock);
3052 
3053 	return 0;
3054 }
3055 
3056 static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev)
3057 {
3058 	struct mlx5_ib_resources *devr = &dev->devr;
3059 
3060 	/* After s0/s1 init, they are not unset during the device lifetime. */
3061 	if (devr->s1) {
3062 		ib_destroy_srq(devr->s1);
3063 		ib_destroy_srq(devr->s0);
3064 	}
3065 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
3066 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
3067 	/* After p0/c0 init, they are not unset during the device lifetime. */
3068 	if (devr->c0) {
3069 		ib_destroy_cq(devr->c0);
3070 		ib_dealloc_pd(devr->p0);
3071 	}
3072 	mutex_destroy(&devr->cq_lock);
3073 	mutex_destroy(&devr->srq_lock);
3074 }
3075 
3076 static int
3077 mlx5_ib_create_data_direct_resources(struct mlx5_ib_dev *dev)
3078 {
3079 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
3080 	struct mlx5_core_dev *mdev = dev->mdev;
3081 	void *mkc;
3082 	u32 mkey;
3083 	u32 pdn;
3084 	u32 *in;
3085 	int err;
3086 
3087 	err = mlx5_core_alloc_pd(mdev, &pdn);
3088 	if (err)
3089 		return err;
3090 
3091 	in = kvzalloc(inlen, GFP_KERNEL);
3092 	if (!in) {
3093 		err = -ENOMEM;
3094 		goto err;
3095 	}
3096 
3097 	MLX5_SET(create_mkey_in, in, data_direct, 1);
3098 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
3099 	MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_PA);
3100 	MLX5_SET(mkc, mkc, lw, 1);
3101 	MLX5_SET(mkc, mkc, lr, 1);
3102 	MLX5_SET(mkc, mkc, rw, 1);
3103 	MLX5_SET(mkc, mkc, rr, 1);
3104 	MLX5_SET(mkc, mkc, a, 1);
3105 	MLX5_SET(mkc, mkc, pd, pdn);
3106 	MLX5_SET(mkc, mkc, length64, 1);
3107 	MLX5_SET(mkc, mkc, qpn, 0xffffff);
3108 	err = mlx5_core_create_mkey(mdev, &mkey, in, inlen);
3109 	kvfree(in);
3110 	if (err)
3111 		goto err;
3112 
3113 	dev->ddr.mkey = mkey;
3114 	dev->ddr.pdn = pdn;
3115 	return 0;
3116 
3117 err:
3118 	mlx5_core_dealloc_pd(mdev, pdn);
3119 	return err;
3120 }
3121 
3122 static void
3123 mlx5_ib_free_data_direct_resources(struct mlx5_ib_dev *dev)
3124 {
3125 	mlx5_core_destroy_mkey(dev->mdev, dev->ddr.mkey);
3126 	mlx5_core_dealloc_pd(dev->mdev, dev->ddr.pdn);
3127 }
3128 
3129 static u32 get_core_cap_flags(struct ib_device *ibdev,
3130 			      struct mlx5_hca_vport_context *rep)
3131 {
3132 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3133 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
3134 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
3135 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
3136 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
3137 	u32 ret = 0;
3138 
3139 	if (rep->grh_required)
3140 		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
3141 
3142 	if (dev->num_plane)
3143 		return ret | RDMA_CORE_CAP_PROT_IB | RDMA_CORE_CAP_IB_MAD |
3144 			RDMA_CORE_CAP_IB_CM | RDMA_CORE_CAP_IB_SA |
3145 			RDMA_CORE_CAP_AF_IB;
3146 	else if (ibdev->type == RDMA_DEVICE_TYPE_SMI)
3147 		return ret | RDMA_CORE_CAP_IB_MAD | RDMA_CORE_CAP_IB_SMI;
3148 
3149 	if (ll == IB_LINK_LAYER_INFINIBAND)
3150 		return ret | RDMA_CORE_PORT_IBA_IB;
3151 
3152 	if (raw_support)
3153 		ret |= RDMA_CORE_PORT_RAW_PACKET;
3154 
3155 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
3156 		return ret;
3157 
3158 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
3159 		return ret;
3160 
3161 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
3162 		ret |= RDMA_CORE_PORT_IBA_ROCE;
3163 
3164 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
3165 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3166 
3167 	return ret;
3168 }
3169 
3170 static int mlx5_port_immutable(struct ib_device *ibdev, u32 port_num,
3171 			       struct ib_port_immutable *immutable)
3172 {
3173 	struct ib_port_attr attr;
3174 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3175 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
3176 	struct mlx5_hca_vport_context rep = {0};
3177 	int err;
3178 
3179 	err = ib_query_port(ibdev, port_num, &attr);
3180 	if (err)
3181 		return err;
3182 
3183 	if (ll == IB_LINK_LAYER_INFINIBAND) {
3184 		if (ibdev->type == RDMA_DEVICE_TYPE_SMI)
3185 			port_num = smi_to_native_portnum(dev, port_num);
3186 
3187 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
3188 						   &rep);
3189 		if (err)
3190 			return err;
3191 	}
3192 
3193 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
3194 	immutable->gid_tbl_len = attr.gid_tbl_len;
3195 	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
3196 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3197 
3198 	return 0;
3199 }
3200 
3201 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u32 port_num,
3202 				   struct ib_port_immutable *immutable)
3203 {
3204 	struct ib_port_attr attr;
3205 	int err;
3206 
3207 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
3208 
3209 	err = ib_query_port(ibdev, port_num, &attr);
3210 	if (err)
3211 		return err;
3212 
3213 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
3214 	immutable->gid_tbl_len = attr.gid_tbl_len;
3215 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
3216 
3217 	return 0;
3218 }
3219 
3220 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
3221 {
3222 	struct mlx5_ib_dev *dev =
3223 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
3224 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
3225 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
3226 		 fw_rev_sub(dev->mdev));
3227 }
3228 
3229 static int lag_event(struct notifier_block *nb, unsigned long event, void *data)
3230 {
3231 	struct mlx5_ib_dev *dev = container_of(nb, struct mlx5_ib_dev,
3232 					       lag_events);
3233 	struct mlx5_core_dev *mdev = dev->mdev;
3234 	struct ib_device *ibdev = &dev->ib_dev;
3235 	struct net_device *old_ndev = NULL;
3236 	struct mlx5_ib_port *port;
3237 	struct net_device *ndev;
3238 	u32 portnum = 0;
3239 	int ret = 0;
3240 	int i;
3241 
3242 	switch (event) {
3243 	case MLX5_DRIVER_EVENT_ACTIVE_BACKUP_LAG_CHANGE_LOWERSTATE:
3244 		ndev = data;
3245 		if (ndev) {
3246 			if (!mlx5_lag_is_roce(mdev)) {
3247 				// sriov lag
3248 				for (i = 0; i < dev->num_ports; i++) {
3249 					port = &dev->port[i];
3250 					if (port->rep && port->rep->vport ==
3251 					    MLX5_VPORT_UPLINK) {
3252 						portnum = i;
3253 						break;
3254 					}
3255 				}
3256 			}
3257 			old_ndev = ib_device_get_netdev(ibdev, portnum + 1);
3258 			ret = ib_device_set_netdev(ibdev, ndev, portnum + 1);
3259 			if (ret)
3260 				goto out;
3261 
3262 			if (old_ndev)
3263 				roce_del_all_netdev_gids(ibdev, portnum + 1,
3264 							 old_ndev);
3265 			rdma_roce_rescan_port(ibdev, portnum + 1);
3266 		}
3267 		break;
3268 	default:
3269 		return NOTIFY_DONE;
3270 	}
3271 
3272 out:
3273 	dev_put(old_ndev);
3274 	return notifier_from_errno(ret);
3275 }
3276 
3277 static void mlx5e_lag_event_register(struct mlx5_ib_dev *dev)
3278 {
3279 	dev->lag_events.notifier_call = lag_event;
3280 	blocking_notifier_chain_register(&dev->mdev->priv.lag_nh,
3281 					 &dev->lag_events);
3282 }
3283 
3284 static void mlx5e_lag_event_unregister(struct mlx5_ib_dev *dev)
3285 {
3286 	blocking_notifier_chain_unregister(&dev->mdev->priv.lag_nh,
3287 					   &dev->lag_events);
3288 }
3289 
3290 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
3291 {
3292 	struct mlx5_core_dev *mdev = dev->mdev;
3293 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
3294 								 MLX5_FLOW_NAMESPACE_LAG);
3295 	struct mlx5_flow_table *ft;
3296 	int err;
3297 
3298 	if (!ns || !mlx5_lag_is_active(mdev))
3299 		return 0;
3300 
3301 	err = mlx5_cmd_create_vport_lag(mdev);
3302 	if (err)
3303 		return err;
3304 
3305 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3306 	if (IS_ERR(ft)) {
3307 		err = PTR_ERR(ft);
3308 		goto err_destroy_vport_lag;
3309 	}
3310 
3311 	mlx5e_lag_event_register(dev);
3312 	dev->flow_db->lag_demux_ft = ft;
3313 	dev->lag_ports = mlx5_lag_get_num_ports(mdev);
3314 	dev->lag_active = true;
3315 	return 0;
3316 
3317 err_destroy_vport_lag:
3318 	mlx5_cmd_destroy_vport_lag(mdev);
3319 	return err;
3320 }
3321 
3322 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3323 {
3324 	struct mlx5_core_dev *mdev = dev->mdev;
3325 
3326 	if (dev->lag_active) {
3327 		dev->lag_active = false;
3328 
3329 		mlx5e_lag_event_unregister(dev);
3330 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
3331 		dev->flow_db->lag_demux_ft = NULL;
3332 
3333 		mlx5_cmd_destroy_vport_lag(mdev);
3334 	}
3335 }
3336 
3337 static void mlx5_netdev_notifier_register(struct mlx5_roce *roce,
3338 					  struct net_device *netdev)
3339 {
3340 	int err;
3341 
3342 	if (roce->tracking_netdev)
3343 		return;
3344 	roce->tracking_netdev = netdev;
3345 	roce->nb.notifier_call = mlx5_netdev_event;
3346 	err = register_netdevice_notifier_dev_net(netdev, &roce->nb, &roce->nn);
3347 	WARN_ON(err);
3348 }
3349 
3350 static void mlx5_netdev_notifier_unregister(struct mlx5_roce *roce)
3351 {
3352 	if (!roce->tracking_netdev)
3353 		return;
3354 	unregister_netdevice_notifier_dev_net(roce->tracking_netdev, &roce->nb,
3355 					      &roce->nn);
3356 	roce->tracking_netdev = NULL;
3357 }
3358 
3359 static int mlx5e_mdev_notifier_event(struct notifier_block *nb,
3360 				     unsigned long event, void *data)
3361 {
3362 	struct mlx5_roce *roce = container_of(nb, struct mlx5_roce, mdev_nb);
3363 	struct net_device *netdev = data;
3364 
3365 	switch (event) {
3366 	case MLX5_DRIVER_EVENT_UPLINK_NETDEV:
3367 		if (netdev)
3368 			mlx5_netdev_notifier_register(roce, netdev);
3369 		else
3370 			mlx5_netdev_notifier_unregister(roce);
3371 		break;
3372 	default:
3373 		return NOTIFY_DONE;
3374 	}
3375 
3376 	return NOTIFY_OK;
3377 }
3378 
3379 static void mlx5_mdev_netdev_track(struct mlx5_ib_dev *dev, u32 port_num)
3380 {
3381 	struct mlx5_roce *roce = &dev->port[port_num].roce;
3382 
3383 	roce->mdev_nb.notifier_call = mlx5e_mdev_notifier_event;
3384 	mlx5_blocking_notifier_register(dev->mdev, &roce->mdev_nb);
3385 	mlx5_core_uplink_netdev_event_replay(dev->mdev);
3386 }
3387 
3388 static void mlx5_mdev_netdev_untrack(struct mlx5_ib_dev *dev, u32 port_num)
3389 {
3390 	struct mlx5_roce *roce = &dev->port[port_num].roce;
3391 
3392 	mlx5_blocking_notifier_unregister(dev->mdev, &roce->mdev_nb);
3393 	mlx5_netdev_notifier_unregister(roce);
3394 }
3395 
3396 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3397 {
3398 	int err;
3399 
3400 	if (!dev->is_rep && dev->profile != &raw_eth_profile) {
3401 		err = mlx5_nic_vport_enable_roce(dev->mdev);
3402 		if (err)
3403 			return err;
3404 	}
3405 
3406 	err = mlx5_eth_lag_init(dev);
3407 	if (err)
3408 		goto err_disable_roce;
3409 
3410 	return 0;
3411 
3412 err_disable_roce:
3413 	if (!dev->is_rep && dev->profile != &raw_eth_profile)
3414 		mlx5_nic_vport_disable_roce(dev->mdev);
3415 
3416 	return err;
3417 }
3418 
3419 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3420 {
3421 	mlx5_eth_lag_cleanup(dev);
3422 	if (!dev->is_rep && dev->profile != &raw_eth_profile)
3423 		mlx5_nic_vport_disable_roce(dev->mdev);
3424 }
3425 
3426 static int mlx5_ib_rn_get_params(struct ib_device *device, u32 port_num,
3427 				 enum rdma_netdev_t type,
3428 				 struct rdma_netdev_alloc_params *params)
3429 {
3430 	if (type != RDMA_NETDEV_IPOIB)
3431 		return -EOPNOTSUPP;
3432 
3433 	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
3434 }
3435 
3436 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
3437 				       size_t count, loff_t *pos)
3438 {
3439 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3440 	char lbuf[20];
3441 	int len;
3442 
3443 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
3444 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
3445 }
3446 
3447 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
3448 					size_t count, loff_t *pos)
3449 {
3450 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3451 	u32 timeout;
3452 	u32 var;
3453 
3454 	if (kstrtouint_from_user(buf, count, 0, &var))
3455 		return -EFAULT;
3456 
3457 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
3458 			1000);
3459 	if (timeout != var)
3460 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
3461 			    timeout);
3462 
3463 	delay_drop->timeout = timeout;
3464 
3465 	return count;
3466 }
3467 
3468 static const struct file_operations fops_delay_drop_timeout = {
3469 	.owner	= THIS_MODULE,
3470 	.open	= simple_open,
3471 	.write	= delay_drop_timeout_write,
3472 	.read	= delay_drop_timeout_read,
3473 };
3474 
3475 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
3476 				      struct mlx5_ib_multiport_info *mpi)
3477 {
3478 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3479 	struct mlx5_ib_port *port = &ibdev->port[port_num];
3480 	int comps;
3481 	int err;
3482 	int i;
3483 
3484 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3485 
3486 	mlx5_core_mp_event_replay(ibdev->mdev,
3487 				  MLX5_DRIVER_EVENT_AFFILIATION_REMOVED,
3488 				  NULL);
3489 	mlx5_core_mp_event_replay(mpi->mdev,
3490 				  MLX5_DRIVER_EVENT_AFFILIATION_REMOVED,
3491 				  NULL);
3492 
3493 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
3494 
3495 	spin_lock(&port->mp.mpi_lock);
3496 	if (!mpi->ibdev) {
3497 		spin_unlock(&port->mp.mpi_lock);
3498 		return;
3499 	}
3500 
3501 	mpi->ibdev = NULL;
3502 
3503 	spin_unlock(&port->mp.mpi_lock);
3504 	if (mpi->mdev_events.notifier_call)
3505 		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
3506 	mpi->mdev_events.notifier_call = NULL;
3507 	mlx5_mdev_netdev_untrack(ibdev, port_num);
3508 	spin_lock(&port->mp.mpi_lock);
3509 
3510 	comps = mpi->mdev_refcnt;
3511 	if (comps) {
3512 		mpi->unaffiliate = true;
3513 		init_completion(&mpi->unref_comp);
3514 		spin_unlock(&port->mp.mpi_lock);
3515 
3516 		for (i = 0; i < comps; i++)
3517 			wait_for_completion(&mpi->unref_comp);
3518 
3519 		spin_lock(&port->mp.mpi_lock);
3520 		mpi->unaffiliate = false;
3521 	}
3522 
3523 	port->mp.mpi = NULL;
3524 
3525 	spin_unlock(&port->mp.mpi_lock);
3526 
3527 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
3528 
3529 	mlx5_ib_dbg(ibdev, "unaffiliated port %u\n", port_num + 1);
3530 	/* Log an error, still needed to cleanup the pointers and add
3531 	 * it back to the list.
3532 	 */
3533 	if (err)
3534 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
3535 			    port_num + 1);
3536 
3537 	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
3538 }
3539 
3540 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
3541 				    struct mlx5_ib_multiport_info *mpi)
3542 {
3543 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3544 	u64 key;
3545 	int err;
3546 
3547 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3548 
3549 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
3550 	if (ibdev->port[port_num].mp.mpi) {
3551 		mlx5_ib_dbg(ibdev, "port %u already affiliated.\n",
3552 			    port_num + 1);
3553 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3554 		return false;
3555 	}
3556 
3557 	ibdev->port[port_num].mp.mpi = mpi;
3558 	mpi->ibdev = ibdev;
3559 	mpi->mdev_events.notifier_call = NULL;
3560 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3561 
3562 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
3563 	if (err)
3564 		goto unbind;
3565 
3566 	mlx5_mdev_netdev_track(ibdev, port_num);
3567 
3568 	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
3569 	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
3570 
3571 	mlx5_ib_init_cong_debugfs(ibdev, port_num);
3572 
3573 	key = mpi->mdev->priv.adev_idx;
3574 	mlx5_core_mp_event_replay(mpi->mdev,
3575 				  MLX5_DRIVER_EVENT_AFFILIATION_DONE,
3576 				  &key);
3577 	mlx5_core_mp_event_replay(ibdev->mdev,
3578 				  MLX5_DRIVER_EVENT_AFFILIATION_DONE,
3579 				  &key);
3580 
3581 	return true;
3582 
3583 unbind:
3584 	mlx5_ib_unbind_slave_port(ibdev, mpi);
3585 	return false;
3586 }
3587 
3588 static int mlx5_ib_data_direct_init(struct mlx5_ib_dev *dev)
3589 {
3590 	char vuid[MLX5_ST_SZ_BYTES(array1024_auto) + 1] = {};
3591 	int ret;
3592 
3593 	if (!MLX5_CAP_GEN(dev->mdev, data_direct) ||
3594 	    !MLX5_CAP_GEN_2(dev->mdev, query_vuid))
3595 		return 0;
3596 
3597 	ret = mlx5_cmd_query_vuid(dev->mdev, true, vuid);
3598 	if (ret)
3599 		return ret;
3600 
3601 	ret = mlx5_ib_create_data_direct_resources(dev);
3602 	if (ret)
3603 		return ret;
3604 
3605 	INIT_LIST_HEAD(&dev->data_direct_mr_list);
3606 	ret = mlx5_data_direct_ib_reg(dev, vuid);
3607 	if (ret)
3608 		mlx5_ib_free_data_direct_resources(dev);
3609 
3610 	return ret;
3611 }
3612 
3613 static void mlx5_ib_data_direct_cleanup(struct mlx5_ib_dev *dev)
3614 {
3615 	if (!MLX5_CAP_GEN(dev->mdev, data_direct) ||
3616 	    !MLX5_CAP_GEN_2(dev->mdev, query_vuid))
3617 		return;
3618 
3619 	mlx5_data_direct_ib_unreg(dev);
3620 	mlx5_ib_free_data_direct_resources(dev);
3621 }
3622 
3623 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
3624 {
3625 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3626 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3627 							  port_num + 1);
3628 	struct mlx5_ib_multiport_info *mpi;
3629 	int err;
3630 	u32 i;
3631 
3632 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3633 		return 0;
3634 
3635 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
3636 						     &dev->sys_image_guid);
3637 	if (err)
3638 		return err;
3639 
3640 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3641 	if (err)
3642 		return err;
3643 
3644 	mutex_lock(&mlx5_ib_multiport_mutex);
3645 	for (i = 0; i < dev->num_ports; i++) {
3646 		bool bound = false;
3647 
3648 		/* build a stub multiport info struct for the native port. */
3649 		if (i == port_num) {
3650 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
3651 			if (!mpi) {
3652 				mutex_unlock(&mlx5_ib_multiport_mutex);
3653 				mlx5_nic_vport_disable_roce(dev->mdev);
3654 				return -ENOMEM;
3655 			}
3656 
3657 			mpi->is_master = true;
3658 			mpi->mdev = dev->mdev;
3659 			mpi->sys_image_guid = dev->sys_image_guid;
3660 			dev->port[i].mp.mpi = mpi;
3661 			mpi->ibdev = dev;
3662 			mpi = NULL;
3663 			continue;
3664 		}
3665 
3666 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
3667 				    list) {
3668 			if (dev->sys_image_guid == mpi->sys_image_guid &&
3669 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i &&
3670 			    mlx5_core_same_coredev_type(dev->mdev, mpi->mdev)) {
3671 				bound = mlx5_ib_bind_slave_port(dev, mpi);
3672 			}
3673 
3674 			if (bound) {
3675 				dev_dbg(mpi->mdev->device,
3676 					"removing port from unaffiliated list.\n");
3677 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
3678 				list_del(&mpi->list);
3679 				break;
3680 			}
3681 		}
3682 		if (!bound)
3683 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
3684 				    i + 1);
3685 	}
3686 
3687 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
3688 	mutex_unlock(&mlx5_ib_multiport_mutex);
3689 	return err;
3690 }
3691 
3692 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
3693 {
3694 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3695 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3696 							  port_num + 1);
3697 	u32 i;
3698 
3699 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3700 		return;
3701 
3702 	mutex_lock(&mlx5_ib_multiport_mutex);
3703 	for (i = 0; i < dev->num_ports; i++) {
3704 		if (dev->port[i].mp.mpi) {
3705 			/* Destroy the native port stub */
3706 			if (i == port_num) {
3707 				kfree(dev->port[i].mp.mpi);
3708 				dev->port[i].mp.mpi = NULL;
3709 			} else {
3710 				mlx5_ib_dbg(dev, "unbinding port_num: %u\n",
3711 					    i + 1);
3712 				list_add_tail(&dev->port[i].mp.mpi->list,
3713 					      &mlx5_ib_unaffiliated_port_list);
3714 				mlx5_ib_unbind_slave_port(dev,
3715 							  dev->port[i].mp.mpi);
3716 			}
3717 		}
3718 	}
3719 
3720 	mlx5_ib_dbg(dev, "removing from devlist\n");
3721 	list_del(&dev->ib_dev_list);
3722 	mutex_unlock(&mlx5_ib_multiport_mutex);
3723 
3724 	mlx5_nic_vport_disable_roce(dev->mdev);
3725 }
3726 
3727 static int mmap_obj_cleanup(struct ib_uobject *uobject,
3728 			    enum rdma_remove_reason why,
3729 			    struct uverbs_attr_bundle *attrs)
3730 {
3731 	struct mlx5_user_mmap_entry *obj = uobject->object;
3732 
3733 	rdma_user_mmap_entry_remove(&obj->rdma_entry);
3734 	return 0;
3735 }
3736 
3737 static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c,
3738 					    struct mlx5_user_mmap_entry *entry,
3739 					    size_t length)
3740 {
3741 	return rdma_user_mmap_entry_insert_range(
3742 		&c->ibucontext, &entry->rdma_entry, length,
3743 		(MLX5_IB_MMAP_OFFSET_START << 16),
3744 		((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1));
3745 }
3746 
3747 static struct mlx5_user_mmap_entry *
3748 alloc_var_entry(struct mlx5_ib_ucontext *c)
3749 {
3750 	struct mlx5_user_mmap_entry *entry;
3751 	struct mlx5_var_table *var_table;
3752 	u32 page_idx;
3753 	int err;
3754 
3755 	var_table = &to_mdev(c->ibucontext.device)->var_table;
3756 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3757 	if (!entry)
3758 		return ERR_PTR(-ENOMEM);
3759 
3760 	mutex_lock(&var_table->bitmap_lock);
3761 	page_idx = find_first_zero_bit(var_table->bitmap,
3762 				       var_table->num_var_hw_entries);
3763 	if (page_idx >= var_table->num_var_hw_entries) {
3764 		err = -ENOSPC;
3765 		mutex_unlock(&var_table->bitmap_lock);
3766 		goto end;
3767 	}
3768 
3769 	set_bit(page_idx, var_table->bitmap);
3770 	mutex_unlock(&var_table->bitmap_lock);
3771 
3772 	entry->address = var_table->hw_start_addr +
3773 				(page_idx * var_table->stride_size);
3774 	entry->page_idx = page_idx;
3775 	entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
3776 
3777 	err = mlx5_rdma_user_mmap_entry_insert(c, entry,
3778 					       var_table->stride_size);
3779 	if (err)
3780 		goto err_insert;
3781 
3782 	return entry;
3783 
3784 err_insert:
3785 	mutex_lock(&var_table->bitmap_lock);
3786 	clear_bit(page_idx, var_table->bitmap);
3787 	mutex_unlock(&var_table->bitmap_lock);
3788 end:
3789 	kfree(entry);
3790 	return ERR_PTR(err);
3791 }
3792 
3793 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
3794 	struct uverbs_attr_bundle *attrs)
3795 {
3796 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3797 		attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3798 	struct mlx5_ib_ucontext *c;
3799 	struct mlx5_user_mmap_entry *entry;
3800 	u64 mmap_offset;
3801 	u32 length;
3802 	int err;
3803 
3804 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3805 	if (IS_ERR(c))
3806 		return PTR_ERR(c);
3807 
3808 	entry = alloc_var_entry(c);
3809 	if (IS_ERR(entry))
3810 		return PTR_ERR(entry);
3811 
3812 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3813 	length = entry->rdma_entry.npages * PAGE_SIZE;
3814 	uobj->object = entry;
3815 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3816 
3817 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3818 			     &mmap_offset, sizeof(mmap_offset));
3819 	if (err)
3820 		return err;
3821 
3822 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3823 			     &entry->page_idx, sizeof(entry->page_idx));
3824 	if (err)
3825 		return err;
3826 
3827 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3828 			     &length, sizeof(length));
3829 	return err;
3830 }
3831 
3832 DECLARE_UVERBS_NAMED_METHOD(
3833 	MLX5_IB_METHOD_VAR_OBJ_ALLOC,
3834 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
3835 			MLX5_IB_OBJECT_VAR,
3836 			UVERBS_ACCESS_NEW,
3837 			UA_MANDATORY),
3838 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3839 			   UVERBS_ATTR_TYPE(u32),
3840 			   UA_MANDATORY),
3841 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3842 			   UVERBS_ATTR_TYPE(u32),
3843 			   UA_MANDATORY),
3844 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3845 			    UVERBS_ATTR_TYPE(u64),
3846 			    UA_MANDATORY));
3847 
3848 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3849 	MLX5_IB_METHOD_VAR_OBJ_DESTROY,
3850 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
3851 			MLX5_IB_OBJECT_VAR,
3852 			UVERBS_ACCESS_DESTROY,
3853 			UA_MANDATORY));
3854 
3855 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
3856 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3857 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
3858 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
3859 
3860 static bool var_is_supported(struct ib_device *device)
3861 {
3862 	struct mlx5_ib_dev *dev = to_mdev(device);
3863 
3864 	return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3865 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
3866 }
3867 
3868 static struct mlx5_user_mmap_entry *
3869 alloc_uar_entry(struct mlx5_ib_ucontext *c,
3870 		enum mlx5_ib_uapi_uar_alloc_type alloc_type)
3871 {
3872 	struct mlx5_user_mmap_entry *entry;
3873 	struct mlx5_ib_dev *dev;
3874 	u32 uar_index;
3875 	int err;
3876 
3877 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3878 	if (!entry)
3879 		return ERR_PTR(-ENOMEM);
3880 
3881 	dev = to_mdev(c->ibucontext.device);
3882 	err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index, c->devx_uid);
3883 	if (err)
3884 		goto end;
3885 
3886 	entry->page_idx = uar_index;
3887 	entry->address = uar_index2paddress(dev, uar_index);
3888 	if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3889 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC;
3890 	else
3891 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC;
3892 
3893 	err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE);
3894 	if (err)
3895 		goto err_insert;
3896 
3897 	return entry;
3898 
3899 err_insert:
3900 	mlx5_cmd_uar_dealloc(dev->mdev, uar_index, c->devx_uid);
3901 end:
3902 	kfree(entry);
3903 	return ERR_PTR(err);
3904 }
3905 
3906 static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)(
3907 	struct uverbs_attr_bundle *attrs)
3908 {
3909 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3910 		attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3911 	enum mlx5_ib_uapi_uar_alloc_type alloc_type;
3912 	struct mlx5_ib_ucontext *c;
3913 	struct mlx5_user_mmap_entry *entry;
3914 	u64 mmap_offset;
3915 	u32 length;
3916 	int err;
3917 
3918 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3919 	if (IS_ERR(c))
3920 		return PTR_ERR(c);
3921 
3922 	err = uverbs_get_const(&alloc_type, attrs,
3923 			       MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE);
3924 	if (err)
3925 		return err;
3926 
3927 	if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF &&
3928 	    alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC)
3929 		return -EOPNOTSUPP;
3930 
3931 	if (!mlx5_wc_support_get(to_mdev(c->ibucontext.device)->mdev) &&
3932 	    alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3933 		return -EOPNOTSUPP;
3934 
3935 	entry = alloc_uar_entry(c, alloc_type);
3936 	if (IS_ERR(entry))
3937 		return PTR_ERR(entry);
3938 
3939 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3940 	length = entry->rdma_entry.npages * PAGE_SIZE;
3941 	uobj->object = entry;
3942 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3943 
3944 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3945 			     &mmap_offset, sizeof(mmap_offset));
3946 	if (err)
3947 		return err;
3948 
3949 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3950 			     &entry->page_idx, sizeof(entry->page_idx));
3951 	if (err)
3952 		return err;
3953 
3954 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3955 			     &length, sizeof(length));
3956 	return err;
3957 }
3958 
3959 DECLARE_UVERBS_NAMED_METHOD(
3960 	MLX5_IB_METHOD_UAR_OBJ_ALLOC,
3961 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE,
3962 			MLX5_IB_OBJECT_UAR,
3963 			UVERBS_ACCESS_NEW,
3964 			UA_MANDATORY),
3965 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE,
3966 			     enum mlx5_ib_uapi_uar_alloc_type,
3967 			     UA_MANDATORY),
3968 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3969 			   UVERBS_ATTR_TYPE(u32),
3970 			   UA_MANDATORY),
3971 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3972 			   UVERBS_ATTR_TYPE(u32),
3973 			   UA_MANDATORY),
3974 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3975 			    UVERBS_ATTR_TYPE(u64),
3976 			    UA_MANDATORY));
3977 
3978 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3979 	MLX5_IB_METHOD_UAR_OBJ_DESTROY,
3980 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE,
3981 			MLX5_IB_OBJECT_UAR,
3982 			UVERBS_ACCESS_DESTROY,
3983 			UA_MANDATORY));
3984 
3985 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR,
3986 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3987 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC),
3988 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY));
3989 
3990 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3991 	mlx5_ib_query_context,
3992 	UVERBS_OBJECT_DEVICE,
3993 	UVERBS_METHOD_QUERY_CONTEXT,
3994 	UVERBS_ATTR_PTR_OUT(
3995 		MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
3996 		UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp,
3997 				   dump_fill_mkey),
3998 		UA_MANDATORY));
3999 
4000 ADD_UVERBS_ATTRIBUTES_SIMPLE(
4001 	mlx5_ib_reg_dmabuf_mr,
4002 	UVERBS_OBJECT_MR,
4003 	UVERBS_METHOD_REG_DMABUF_MR,
4004 	UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_REG_DMABUF_MR_ACCESS_FLAGS,
4005 			     enum mlx5_ib_uapi_reg_dmabuf_flags,
4006 			     UA_OPTIONAL));
4007 
4008 static const struct uapi_definition mlx5_ib_defs[] = {
4009 	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
4010 	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
4011 	UAPI_DEF_CHAIN(mlx5_ib_qos_defs),
4012 	UAPI_DEF_CHAIN(mlx5_ib_std_types_defs),
4013 	UAPI_DEF_CHAIN(mlx5_ib_dm_defs),
4014 	UAPI_DEF_CHAIN(mlx5_ib_create_cq_defs),
4015 
4016 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context),
4017 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_MR, &mlx5_ib_reg_dmabuf_mr),
4018 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
4019 				UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
4020 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR),
4021 	{}
4022 };
4023 
4024 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
4025 {
4026 	mlx5_ib_data_direct_cleanup(dev);
4027 	mlx5_ib_cleanup_multiport_master(dev);
4028 	WARN_ON(!xa_empty(&dev->odp_mkeys));
4029 	mutex_destroy(&dev->cap_mask_mutex);
4030 	WARN_ON(!xa_empty(&dev->sig_mrs));
4031 	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
4032 	mlx5r_macsec_dealloc_gids(dev);
4033 }
4034 
4035 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
4036 {
4037 	struct mlx5_core_dev *mdev = dev->mdev;
4038 	int err, i;
4039 
4040 	dev->ib_dev.node_type = RDMA_NODE_IB_CA;
4041 	dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
4042 	dev->ib_dev.dev.parent = mdev->device;
4043 	dev->ib_dev.lag_flags = RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
4044 
4045 	for (i = 0; i < dev->num_ports; i++) {
4046 		spin_lock_init(&dev->port[i].mp.mpi_lock);
4047 		dev->port[i].roce.dev = dev;
4048 		dev->port[i].roce.native_port_num = i + 1;
4049 		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
4050 	}
4051 
4052 	err = mlx5r_cmd_query_special_mkeys(dev);
4053 	if (err)
4054 		return err;
4055 
4056 	err = mlx5r_macsec_init_gids_and_devlist(dev);
4057 	if (err)
4058 		return err;
4059 
4060 	err = mlx5_ib_init_multiport_master(dev);
4061 	if (err)
4062 		goto err;
4063 
4064 	err = set_has_smi_cap(dev);
4065 	if (err)
4066 		goto err_mp;
4067 
4068 	err = mlx5_query_max_pkeys(&dev->ib_dev, &dev->pkey_table_len);
4069 	if (err)
4070 		goto err_mp;
4071 
4072 	if (mlx5_use_mad_ifc(dev))
4073 		get_ext_port_caps(dev);
4074 
4075 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_max(mdev);
4076 
4077 	mutex_init(&dev->cap_mask_mutex);
4078 	mutex_init(&dev->data_direct_lock);
4079 	INIT_LIST_HEAD(&dev->qp_list);
4080 	spin_lock_init(&dev->reset_flow_resource_lock);
4081 	xa_init(&dev->odp_mkeys);
4082 	xa_init(&dev->sig_mrs);
4083 	atomic_set(&dev->mkey_var, 0);
4084 
4085 	spin_lock_init(&dev->dm.lock);
4086 	dev->dm.dev = mdev;
4087 	err = mlx5_ib_data_direct_init(dev);
4088 	if (err)
4089 		goto err_mp;
4090 
4091 	return 0;
4092 err_mp:
4093 	mlx5_ib_cleanup_multiport_master(dev);
4094 err:
4095 	mlx5r_macsec_dealloc_gids(dev);
4096 	return err;
4097 }
4098 
4099 static struct ib_device *mlx5_ib_add_sub_dev(struct ib_device *parent,
4100 					     enum rdma_nl_dev_type type,
4101 					     const char *name);
4102 static void mlx5_ib_del_sub_dev(struct ib_device *sub_dev);
4103 
4104 static const struct ib_device_ops mlx5_ib_dev_ops = {
4105 	.owner = THIS_MODULE,
4106 	.driver_id = RDMA_DRIVER_MLX5,
4107 	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
4108 
4109 	.add_gid = mlx5_ib_add_gid,
4110 	.add_sub_dev = mlx5_ib_add_sub_dev,
4111 	.alloc_mr = mlx5_ib_alloc_mr,
4112 	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
4113 	.alloc_pd = mlx5_ib_alloc_pd,
4114 	.alloc_ucontext = mlx5_ib_alloc_ucontext,
4115 	.attach_mcast = mlx5_ib_mcg_attach,
4116 	.check_mr_status = mlx5_ib_check_mr_status,
4117 	.create_ah = mlx5_ib_create_ah,
4118 	.create_cq = mlx5_ib_create_cq,
4119 	.create_qp = mlx5_ib_create_qp,
4120 	.create_srq = mlx5_ib_create_srq,
4121 	.create_user_ah = mlx5_ib_create_ah,
4122 	.dealloc_pd = mlx5_ib_dealloc_pd,
4123 	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
4124 	.del_gid = mlx5_ib_del_gid,
4125 	.del_sub_dev = mlx5_ib_del_sub_dev,
4126 	.dereg_mr = mlx5_ib_dereg_mr,
4127 	.destroy_ah = mlx5_ib_destroy_ah,
4128 	.destroy_cq = mlx5_ib_destroy_cq,
4129 	.destroy_qp = mlx5_ib_destroy_qp,
4130 	.destroy_srq = mlx5_ib_destroy_srq,
4131 	.detach_mcast = mlx5_ib_mcg_detach,
4132 	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
4133 	.drain_rq = mlx5_ib_drain_rq,
4134 	.drain_sq = mlx5_ib_drain_sq,
4135 	.device_group = &mlx5_attr_group,
4136 	.get_dev_fw_str = get_dev_fw_str,
4137 	.get_dma_mr = mlx5_ib_get_dma_mr,
4138 	.get_link_layer = mlx5_ib_port_link_layer,
4139 	.map_mr_sg = mlx5_ib_map_mr_sg,
4140 	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
4141 	.mmap = mlx5_ib_mmap,
4142 	.mmap_free = mlx5_ib_mmap_free,
4143 	.modify_cq = mlx5_ib_modify_cq,
4144 	.modify_device = mlx5_ib_modify_device,
4145 	.modify_port = mlx5_ib_modify_port,
4146 	.modify_qp = mlx5_ib_modify_qp,
4147 	.modify_srq = mlx5_ib_modify_srq,
4148 	.poll_cq = mlx5_ib_poll_cq,
4149 	.post_recv = mlx5_ib_post_recv_nodrain,
4150 	.post_send = mlx5_ib_post_send_nodrain,
4151 	.post_srq_recv = mlx5_ib_post_srq_recv,
4152 	.process_mad = mlx5_ib_process_mad,
4153 	.query_ah = mlx5_ib_query_ah,
4154 	.query_device = mlx5_ib_query_device,
4155 	.query_gid = mlx5_ib_query_gid,
4156 	.query_pkey = mlx5_ib_query_pkey,
4157 	.query_qp = mlx5_ib_query_qp,
4158 	.query_srq = mlx5_ib_query_srq,
4159 	.query_ucontext = mlx5_ib_query_ucontext,
4160 	.reg_user_mr = mlx5_ib_reg_user_mr,
4161 	.reg_user_mr_dmabuf = mlx5_ib_reg_user_mr_dmabuf,
4162 	.req_notify_cq = mlx5_ib_arm_cq,
4163 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
4164 	.resize_cq = mlx5_ib_resize_cq,
4165 	.ufile_hw_cleanup = mlx5_ib_ufile_hw_cleanup,
4166 
4167 	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
4168 	INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs),
4169 	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
4170 	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
4171 	INIT_RDMA_OBJ_SIZE(ib_qp, mlx5_ib_qp, ibqp),
4172 	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
4173 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
4174 };
4175 
4176 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
4177 	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
4178 };
4179 
4180 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
4181 	.get_vf_config = mlx5_ib_get_vf_config,
4182 	.get_vf_guid = mlx5_ib_get_vf_guid,
4183 	.get_vf_stats = mlx5_ib_get_vf_stats,
4184 	.set_vf_guid = mlx5_ib_set_vf_guid,
4185 	.set_vf_link_state = mlx5_ib_set_vf_link_state,
4186 };
4187 
4188 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
4189 	.alloc_mw = mlx5_ib_alloc_mw,
4190 	.dealloc_mw = mlx5_ib_dealloc_mw,
4191 
4192 	INIT_RDMA_OBJ_SIZE(ib_mw, mlx5_ib_mw, ibmw),
4193 };
4194 
4195 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
4196 	.alloc_xrcd = mlx5_ib_alloc_xrcd,
4197 	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
4198 
4199 	INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd),
4200 };
4201 
4202 static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
4203 {
4204 	struct mlx5_core_dev *mdev = dev->mdev;
4205 	struct mlx5_var_table *var_table = &dev->var_table;
4206 	u8 log_doorbell_bar_size;
4207 	u8 log_doorbell_stride;
4208 	u64 bar_size;
4209 
4210 	log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
4211 					log_doorbell_bar_size);
4212 	log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
4213 					log_doorbell_stride);
4214 	var_table->hw_start_addr = dev->mdev->bar_addr +
4215 				MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
4216 					doorbell_bar_offset);
4217 	bar_size = (1ULL << log_doorbell_bar_size) * 4096;
4218 	var_table->stride_size = 1ULL << log_doorbell_stride;
4219 	var_table->num_var_hw_entries = div_u64(bar_size,
4220 						var_table->stride_size);
4221 	mutex_init(&var_table->bitmap_lock);
4222 	var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
4223 					  GFP_KERNEL);
4224 	return (var_table->bitmap) ? 0 : -ENOMEM;
4225 }
4226 
4227 static void mlx5_ib_cleanup_ucaps(struct mlx5_ib_dev *dev)
4228 {
4229 	if (MLX5_CAP_GEN(dev->mdev, uctx_cap) & MLX5_UCTX_CAP_RDMA_CTRL)
4230 		ib_remove_ucap(RDMA_UCAP_MLX5_CTRL_LOCAL);
4231 
4232 	if (MLX5_CAP_GEN(dev->mdev, uctx_cap) &
4233 	    MLX5_UCTX_CAP_RDMA_CTRL_OTHER_VHCA)
4234 		ib_remove_ucap(RDMA_UCAP_MLX5_CTRL_OTHER_VHCA);
4235 }
4236 
4237 static int mlx5_ib_init_ucaps(struct mlx5_ib_dev *dev)
4238 {
4239 	int ret;
4240 
4241 	if (MLX5_CAP_GEN(dev->mdev, uctx_cap) & MLX5_UCTX_CAP_RDMA_CTRL) {
4242 		ret = ib_create_ucap(RDMA_UCAP_MLX5_CTRL_LOCAL);
4243 		if (ret)
4244 			return ret;
4245 	}
4246 
4247 	if (MLX5_CAP_GEN(dev->mdev, uctx_cap) &
4248 	    MLX5_UCTX_CAP_RDMA_CTRL_OTHER_VHCA) {
4249 		ret = ib_create_ucap(RDMA_UCAP_MLX5_CTRL_OTHER_VHCA);
4250 		if (ret)
4251 			goto remove_local;
4252 	}
4253 
4254 	return 0;
4255 
4256 remove_local:
4257 	if (MLX5_CAP_GEN(dev->mdev, uctx_cap) & MLX5_UCTX_CAP_RDMA_CTRL)
4258 		ib_remove_ucap(RDMA_UCAP_MLX5_CTRL_LOCAL);
4259 	return ret;
4260 }
4261 
4262 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
4263 {
4264 	if (MLX5_CAP_GEN_2_64(dev->mdev, general_obj_types_127_64) &
4265 	    MLX5_HCA_CAP_2_GENERAL_OBJECT_TYPES_RDMA_CTRL)
4266 		mlx5_ib_cleanup_ucaps(dev);
4267 
4268 	bitmap_free(dev->var_table.bitmap);
4269 }
4270 
4271 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
4272 {
4273 	struct mlx5_core_dev *mdev = dev->mdev;
4274 	int err;
4275 
4276 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
4277 	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
4278 		ib_set_device_ops(&dev->ib_dev,
4279 				  &mlx5_ib_dev_ipoib_enhanced_ops);
4280 
4281 	if (mlx5_core_is_pf(mdev))
4282 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
4283 
4284 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
4285 
4286 	if (MLX5_CAP_GEN(mdev, imaicl))
4287 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
4288 
4289 	if (MLX5_CAP_GEN(mdev, xrc))
4290 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
4291 
4292 	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
4293 	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
4294 	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
4295 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
4296 
4297 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
4298 
4299 	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
4300 		dev->ib_dev.driver_def = mlx5_ib_defs;
4301 
4302 	err = init_node_data(dev);
4303 	if (err)
4304 		return err;
4305 
4306 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
4307 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
4308 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
4309 		mutex_init(&dev->lb.mutex);
4310 
4311 	if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
4312 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
4313 		err = mlx5_ib_init_var_table(dev);
4314 		if (err)
4315 			return err;
4316 	}
4317 
4318 	if (MLX5_CAP_GEN_2_64(dev->mdev, general_obj_types_127_64) &
4319 	    MLX5_HCA_CAP_2_GENERAL_OBJECT_TYPES_RDMA_CTRL) {
4320 		err = mlx5_ib_init_ucaps(dev);
4321 		if (err)
4322 			return err;
4323 	}
4324 
4325 	dev->ib_dev.use_cq_dim = true;
4326 
4327 	return 0;
4328 }
4329 
4330 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
4331 	.get_port_immutable = mlx5_port_immutable,
4332 	.query_port = mlx5_ib_query_port,
4333 };
4334 
4335 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
4336 {
4337 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
4338 	return 0;
4339 }
4340 
4341 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
4342 	.get_port_immutable = mlx5_port_rep_immutable,
4343 	.query_port = mlx5_ib_rep_query_port,
4344 	.query_pkey = mlx5_ib_rep_query_pkey,
4345 };
4346 
4347 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
4348 {
4349 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
4350 	return 0;
4351 }
4352 
4353 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
4354 	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
4355 	.create_wq = mlx5_ib_create_wq,
4356 	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
4357 	.destroy_wq = mlx5_ib_destroy_wq,
4358 	.modify_wq = mlx5_ib_modify_wq,
4359 
4360 	INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table,
4361 			   ib_rwq_ind_tbl),
4362 };
4363 
4364 static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev)
4365 {
4366 	struct mlx5_core_dev *mdev = dev->mdev;
4367 	enum rdma_link_layer ll;
4368 	int port_type_cap;
4369 	u32 port_num = 0;
4370 	int err;
4371 
4372 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4373 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4374 
4375 	if (ll == IB_LINK_LAYER_ETHERNET) {
4376 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
4377 
4378 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4379 
4380 		/* Register only for native ports */
4381 		mlx5_mdev_netdev_track(dev, port_num);
4382 
4383 		err = mlx5_enable_eth(dev);
4384 		if (err)
4385 			goto cleanup;
4386 	}
4387 
4388 	return 0;
4389 cleanup:
4390 	mlx5_mdev_netdev_untrack(dev, port_num);
4391 	return err;
4392 }
4393 
4394 static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev)
4395 {
4396 	struct mlx5_core_dev *mdev = dev->mdev;
4397 	enum rdma_link_layer ll;
4398 	int port_type_cap;
4399 	u32 port_num;
4400 
4401 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4402 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4403 
4404 	if (ll == IB_LINK_LAYER_ETHERNET) {
4405 		mlx5_disable_eth(dev);
4406 
4407 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4408 		mlx5_mdev_netdev_untrack(dev, port_num);
4409 	}
4410 }
4411 
4412 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
4413 {
4414 	mlx5_ib_init_cong_debugfs(dev,
4415 				  mlx5_core_native_port_num(dev->mdev) - 1);
4416 	return 0;
4417 }
4418 
4419 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
4420 {
4421 	mlx5_ib_cleanup_cong_debugfs(dev,
4422 				     mlx5_core_native_port_num(dev->mdev) - 1);
4423 }
4424 
4425 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
4426 {
4427 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
4428 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
4429 }
4430 
4431 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
4432 {
4433 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
4434 }
4435 
4436 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
4437 {
4438 	int err;
4439 
4440 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
4441 	if (err)
4442 		return err;
4443 
4444 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
4445 	if (err)
4446 		mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4447 
4448 	return err;
4449 }
4450 
4451 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
4452 {
4453 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4454 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4455 }
4456 
4457 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
4458 {
4459 	const char *name;
4460 
4461 	if (dev->sub_dev_name) {
4462 		name = dev->sub_dev_name;
4463 		ib_mark_name_assigned_by_user(&dev->ib_dev);
4464 	} else if (!mlx5_lag_is_active(dev->mdev))
4465 		name = "mlx5_%d";
4466 	else
4467 		name = "mlx5_bond_%d";
4468 	return ib_register_device(&dev->ib_dev, name, &dev->mdev->pdev->dev);
4469 }
4470 
4471 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
4472 {
4473 	mlx5_mkey_cache_cleanup(dev);
4474 	mlx5r_umr_resource_cleanup(dev);
4475 	mlx5r_umr_cleanup(dev);
4476 }
4477 
4478 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
4479 {
4480 	ib_unregister_device(&dev->ib_dev);
4481 }
4482 
4483 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
4484 {
4485 	int ret;
4486 
4487 	ret = mlx5r_umr_init(dev);
4488 	if (ret)
4489 		return ret;
4490 
4491 	ret = mlx5_mkey_cache_init(dev);
4492 	if (ret)
4493 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4494 	return ret;
4495 }
4496 
4497 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
4498 {
4499 	struct dentry *root;
4500 
4501 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4502 		return 0;
4503 
4504 	mutex_init(&dev->delay_drop.lock);
4505 	dev->delay_drop.dev = dev;
4506 	dev->delay_drop.activate = false;
4507 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4508 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4509 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4510 	atomic_set(&dev->delay_drop.events_cnt, 0);
4511 
4512 	if (!mlx5_debugfs_root)
4513 		return 0;
4514 
4515 	root = debugfs_create_dir("delay_drop", mlx5_debugfs_get_dev_root(dev->mdev));
4516 	dev->delay_drop.dir_debugfs = root;
4517 
4518 	debugfs_create_atomic_t("num_timeout_events", 0400, root,
4519 				&dev->delay_drop.events_cnt);
4520 	debugfs_create_atomic_t("num_rqs", 0400, root,
4521 				&dev->delay_drop.rqs_cnt);
4522 	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
4523 			    &fops_delay_drop_timeout);
4524 	return 0;
4525 }
4526 
4527 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
4528 {
4529 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4530 		return;
4531 
4532 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4533 	if (!dev->delay_drop.dir_debugfs)
4534 		return;
4535 
4536 	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
4537 	dev->delay_drop.dir_debugfs = NULL;
4538 }
4539 
4540 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
4541 {
4542 	struct mlx5_ib_resources *devr = &dev->devr;
4543 	int port;
4544 
4545 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
4546 		INIT_WORK(&devr->ports[port].pkey_change_work,
4547 			  pkey_change_handler);
4548 
4549 	dev->mdev_events.notifier_call = mlx5_ib_event;
4550 	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
4551 
4552 	mlx5r_macsec_event_register(dev);
4553 
4554 	return 0;
4555 }
4556 
4557 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
4558 {
4559 	struct mlx5_ib_resources *devr = &dev->devr;
4560 	int port;
4561 
4562 	mlx5r_macsec_event_unregister(dev);
4563 	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
4564 
4565 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
4566 		cancel_work_sync(&devr->ports[port].pkey_change_work);
4567 }
4568 
4569 void mlx5_ib_data_direct_bind(struct mlx5_ib_dev *ibdev,
4570 			      struct mlx5_data_direct_dev *dev)
4571 {
4572 	mutex_lock(&ibdev->data_direct_lock);
4573 	ibdev->data_direct_dev = dev;
4574 	mutex_unlock(&ibdev->data_direct_lock);
4575 }
4576 
4577 void mlx5_ib_data_direct_unbind(struct mlx5_ib_dev *ibdev)
4578 {
4579 	mutex_lock(&ibdev->data_direct_lock);
4580 	mlx5_ib_revoke_data_direct_mrs(ibdev);
4581 	ibdev->data_direct_dev = NULL;
4582 	mutex_unlock(&ibdev->data_direct_lock);
4583 }
4584 
4585 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
4586 		      const struct mlx5_ib_profile *profile,
4587 		      int stage)
4588 {
4589 	dev->ib_active = false;
4590 
4591 	/* Number of stages to cleanup */
4592 	while (stage) {
4593 		stage--;
4594 		if (profile->stage[stage].cleanup)
4595 			profile->stage[stage].cleanup(dev);
4596 	}
4597 
4598 	kfree(dev->port);
4599 	ib_dealloc_device(&dev->ib_dev);
4600 }
4601 
4602 int __mlx5_ib_add(struct mlx5_ib_dev *dev,
4603 		  const struct mlx5_ib_profile *profile)
4604 {
4605 	int err;
4606 	int i;
4607 
4608 	dev->profile = profile;
4609 
4610 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
4611 		if (profile->stage[i].init) {
4612 			err = profile->stage[i].init(dev);
4613 			if (err)
4614 				goto err_out;
4615 		}
4616 	}
4617 
4618 	dev->ib_active = true;
4619 	return 0;
4620 
4621 err_out:
4622 	/* Clean up stages which were initialized */
4623 	while (i) {
4624 		i--;
4625 		if (profile->stage[i].cleanup)
4626 			profile->stage[i].cleanup(dev);
4627 	}
4628 	return -ENOMEM;
4629 }
4630 
4631 static const struct mlx5_ib_profile pf_profile = {
4632 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4633 		     mlx5_ib_stage_init_init,
4634 		     mlx5_ib_stage_init_cleanup),
4635 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4636 		     mlx5_ib_fs_init,
4637 		     mlx5_ib_fs_cleanup),
4638 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4639 		     mlx5_ib_stage_caps_init,
4640 		     mlx5_ib_stage_caps_cleanup),
4641 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4642 		     mlx5_ib_stage_non_default_cb,
4643 		     NULL),
4644 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4645 		     mlx5_ib_roce_init,
4646 		     mlx5_ib_roce_cleanup),
4647 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4648 		     mlx5_init_qp_table,
4649 		     mlx5_cleanup_qp_table),
4650 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4651 		     mlx5_init_srq_table,
4652 		     mlx5_cleanup_srq_table),
4653 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4654 		     mlx5_ib_dev_res_init,
4655 		     mlx5_ib_dev_res_cleanup),
4656 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
4657 		     mlx5_ib_odp_init_one,
4658 		     mlx5_ib_odp_cleanup_one),
4659 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4660 		     mlx5_ib_counters_init,
4661 		     mlx5_ib_counters_cleanup),
4662 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4663 		     mlx5_ib_stage_cong_debugfs_init,
4664 		     mlx5_ib_stage_cong_debugfs_cleanup),
4665 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4666 		     mlx5_ib_stage_uar_init,
4667 		     mlx5_ib_stage_uar_cleanup),
4668 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4669 		     mlx5_ib_stage_bfrag_init,
4670 		     mlx5_ib_stage_bfrag_cleanup),
4671 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4672 		     NULL,
4673 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4674 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4675 		     mlx5_ib_devx_init,
4676 		     mlx5_ib_devx_cleanup),
4677 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4678 		     mlx5_ib_stage_ib_reg_init,
4679 		     mlx5_ib_stage_ib_reg_cleanup),
4680 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4681 		     mlx5_ib_stage_dev_notifier_init,
4682 		     mlx5_ib_stage_dev_notifier_cleanup),
4683 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4684 		     mlx5_ib_stage_post_ib_reg_umr_init,
4685 		     NULL),
4686 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4687 		     mlx5_ib_stage_delay_drop_init,
4688 		     mlx5_ib_stage_delay_drop_cleanup),
4689 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4690 		     mlx5_ib_restrack_init,
4691 		     NULL),
4692 };
4693 
4694 const struct mlx5_ib_profile raw_eth_profile = {
4695 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4696 		     mlx5_ib_stage_init_init,
4697 		     mlx5_ib_stage_init_cleanup),
4698 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4699 		     mlx5_ib_fs_init,
4700 		     mlx5_ib_fs_cleanup),
4701 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4702 		     mlx5_ib_stage_caps_init,
4703 		     mlx5_ib_stage_caps_cleanup),
4704 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4705 		     mlx5_ib_stage_raw_eth_non_default_cb,
4706 		     NULL),
4707 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4708 		     mlx5_ib_roce_init,
4709 		     mlx5_ib_roce_cleanup),
4710 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4711 		     mlx5_init_qp_table,
4712 		     mlx5_cleanup_qp_table),
4713 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4714 		     mlx5_init_srq_table,
4715 		     mlx5_cleanup_srq_table),
4716 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4717 		     mlx5_ib_dev_res_init,
4718 		     mlx5_ib_dev_res_cleanup),
4719 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4720 		     mlx5_ib_counters_init,
4721 		     mlx5_ib_counters_cleanup),
4722 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4723 		     mlx5_ib_stage_cong_debugfs_init,
4724 		     mlx5_ib_stage_cong_debugfs_cleanup),
4725 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4726 		     mlx5_ib_stage_uar_init,
4727 		     mlx5_ib_stage_uar_cleanup),
4728 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4729 		     mlx5_ib_stage_bfrag_init,
4730 		     mlx5_ib_stage_bfrag_cleanup),
4731 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4732 		     NULL,
4733 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4734 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4735 		     mlx5_ib_devx_init,
4736 		     mlx5_ib_devx_cleanup),
4737 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4738 		     mlx5_ib_stage_ib_reg_init,
4739 		     mlx5_ib_stage_ib_reg_cleanup),
4740 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4741 		     mlx5_ib_stage_dev_notifier_init,
4742 		     mlx5_ib_stage_dev_notifier_cleanup),
4743 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4744 		     mlx5_ib_stage_post_ib_reg_umr_init,
4745 		     NULL),
4746 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4747 		     mlx5_ib_stage_delay_drop_init,
4748 		     mlx5_ib_stage_delay_drop_cleanup),
4749 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4750 		     mlx5_ib_restrack_init,
4751 		     NULL),
4752 };
4753 
4754 static const struct mlx5_ib_profile plane_profile = {
4755 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4756 		     mlx5_ib_stage_init_init,
4757 		     mlx5_ib_stage_init_cleanup),
4758 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4759 		     mlx5_ib_stage_caps_init,
4760 		     mlx5_ib_stage_caps_cleanup),
4761 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4762 		     mlx5_ib_stage_non_default_cb,
4763 		     NULL),
4764 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4765 		     mlx5_init_qp_table,
4766 		     mlx5_cleanup_qp_table),
4767 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4768 		     mlx5_init_srq_table,
4769 		     mlx5_cleanup_srq_table),
4770 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4771 		     mlx5_ib_dev_res_init,
4772 		     mlx5_ib_dev_res_cleanup),
4773 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4774 		     mlx5_ib_stage_bfrag_init,
4775 		     mlx5_ib_stage_bfrag_cleanup),
4776 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4777 		     mlx5_ib_stage_ib_reg_init,
4778 		     mlx5_ib_stage_ib_reg_cleanup),
4779 };
4780 
4781 static struct ib_device *mlx5_ib_add_sub_dev(struct ib_device *parent,
4782 					     enum rdma_nl_dev_type type,
4783 					     const char *name)
4784 {
4785 	struct mlx5_ib_dev *mparent = to_mdev(parent), *mplane;
4786 	enum rdma_link_layer ll;
4787 	int ret;
4788 
4789 	if (mparent->smi_dev)
4790 		return ERR_PTR(-EEXIST);
4791 
4792 	ll = mlx5_port_type_cap_to_rdma_ll(MLX5_CAP_GEN(mparent->mdev,
4793 							port_type));
4794 	if (type != RDMA_DEVICE_TYPE_SMI || !mparent->num_plane ||
4795 	    ll != IB_LINK_LAYER_INFINIBAND ||
4796 	    !MLX5_CAP_GEN_2(mparent->mdev, multiplane_qp_ud))
4797 		return ERR_PTR(-EOPNOTSUPP);
4798 
4799 	mplane = ib_alloc_device(mlx5_ib_dev, ib_dev);
4800 	if (!mplane)
4801 		return ERR_PTR(-ENOMEM);
4802 
4803 	mplane->port = kcalloc(mparent->num_plane * mparent->num_ports,
4804 			       sizeof(*mplane->port), GFP_KERNEL);
4805 	if (!mplane->port) {
4806 		ret = -ENOMEM;
4807 		goto fail_kcalloc;
4808 	}
4809 
4810 	mplane->ib_dev.type = type;
4811 	mplane->mdev = mparent->mdev;
4812 	mplane->num_ports = mparent->num_plane;
4813 	mplane->sub_dev_name = name;
4814 	mplane->ib_dev.phys_port_cnt = mplane->num_ports;
4815 
4816 	ret = __mlx5_ib_add(mplane, &plane_profile);
4817 	if (ret)
4818 		goto fail_ib_add;
4819 
4820 	mparent->smi_dev = mplane;
4821 	return &mplane->ib_dev;
4822 
4823 fail_ib_add:
4824 	kfree(mplane->port);
4825 fail_kcalloc:
4826 	ib_dealloc_device(&mplane->ib_dev);
4827 	return ERR_PTR(ret);
4828 }
4829 
4830 static void mlx5_ib_del_sub_dev(struct ib_device *sub_dev)
4831 {
4832 	struct mlx5_ib_dev *mdev = to_mdev(sub_dev);
4833 
4834 	to_mdev(sub_dev->parent)->smi_dev = NULL;
4835 	__mlx5_ib_remove(mdev, mdev->profile, MLX5_IB_STAGE_MAX);
4836 }
4837 
4838 static int mlx5r_mp_probe(struct auxiliary_device *adev,
4839 			  const struct auxiliary_device_id *id)
4840 {
4841 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4842 	struct mlx5_core_dev *mdev = idev->mdev;
4843 	struct mlx5_ib_multiport_info *mpi;
4844 	struct mlx5_ib_dev *dev;
4845 	bool bound = false;
4846 	int err;
4847 
4848 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4849 	if (!mpi)
4850 		return -ENOMEM;
4851 
4852 	mpi->mdev = mdev;
4853 	err = mlx5_query_nic_vport_system_image_guid(mdev,
4854 						     &mpi->sys_image_guid);
4855 	if (err) {
4856 		kfree(mpi);
4857 		return err;
4858 	}
4859 
4860 	mutex_lock(&mlx5_ib_multiport_mutex);
4861 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
4862 		if (dev->sys_image_guid == mpi->sys_image_guid &&
4863 		    mlx5_core_same_coredev_type(dev->mdev, mpi->mdev))
4864 			bound = mlx5_ib_bind_slave_port(dev, mpi);
4865 
4866 		if (bound) {
4867 			rdma_roce_rescan_device(&dev->ib_dev);
4868 			mpi->ibdev->ib_active = true;
4869 			break;
4870 		}
4871 	}
4872 
4873 	if (!bound) {
4874 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4875 		dev_dbg(mdev->device,
4876 			"no suitable IB device found to bind to, added to unaffiliated list.\n");
4877 	}
4878 	mutex_unlock(&mlx5_ib_multiport_mutex);
4879 
4880 	auxiliary_set_drvdata(adev, mpi);
4881 	return 0;
4882 }
4883 
4884 static void mlx5r_mp_remove(struct auxiliary_device *adev)
4885 {
4886 	struct mlx5_ib_multiport_info *mpi;
4887 
4888 	mpi = auxiliary_get_drvdata(adev);
4889 	mutex_lock(&mlx5_ib_multiport_mutex);
4890 	if (mpi->ibdev)
4891 		mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
4892 	else
4893 		list_del(&mpi->list);
4894 	mutex_unlock(&mlx5_ib_multiport_mutex);
4895 	kfree(mpi);
4896 }
4897 
4898 static int mlx5r_probe(struct auxiliary_device *adev,
4899 		       const struct auxiliary_device_id *id)
4900 {
4901 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4902 	struct mlx5_core_dev *mdev = idev->mdev;
4903 	const struct mlx5_ib_profile *profile;
4904 	int port_type_cap, num_ports, ret;
4905 	enum rdma_link_layer ll;
4906 	struct mlx5_ib_dev *dev;
4907 
4908 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4909 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4910 
4911 	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
4912 			MLX5_CAP_GEN(mdev, num_vhca_ports));
4913 	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
4914 	if (!dev)
4915 		return -ENOMEM;
4916 
4917 	if (ll == IB_LINK_LAYER_INFINIBAND) {
4918 		ret = mlx5_ib_get_plane_num(mdev, &dev->num_plane);
4919 		if (ret)
4920 			goto fail;
4921 	}
4922 
4923 	dev->port = kcalloc(num_ports, sizeof(*dev->port),
4924 			     GFP_KERNEL);
4925 	if (!dev->port) {
4926 		ret = -ENOMEM;
4927 		goto fail;
4928 	}
4929 
4930 	dev->mdev = mdev;
4931 	dev->num_ports = num_ports;
4932 	dev->ib_dev.phys_port_cnt = num_ports;
4933 
4934 	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_get_roce_state(mdev))
4935 		profile = &raw_eth_profile;
4936 	else
4937 		profile = &pf_profile;
4938 
4939 	ret = __mlx5_ib_add(dev, profile);
4940 	if (ret)
4941 		goto fail_ib_add;
4942 
4943 	auxiliary_set_drvdata(adev, dev);
4944 	return 0;
4945 
4946 fail_ib_add:
4947 	kfree(dev->port);
4948 fail:
4949 	ib_dealloc_device(&dev->ib_dev);
4950 	return ret;
4951 }
4952 
4953 static void mlx5r_remove(struct auxiliary_device *adev)
4954 {
4955 	struct mlx5_ib_dev *dev;
4956 
4957 	dev = auxiliary_get_drvdata(adev);
4958 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
4959 }
4960 
4961 static const struct auxiliary_device_id mlx5r_mp_id_table[] = {
4962 	{ .name = MLX5_ADEV_NAME ".multiport", },
4963 	{},
4964 };
4965 
4966 static const struct auxiliary_device_id mlx5r_id_table[] = {
4967 	{ .name = MLX5_ADEV_NAME ".rdma", },
4968 	{},
4969 };
4970 
4971 MODULE_DEVICE_TABLE(auxiliary, mlx5r_mp_id_table);
4972 MODULE_DEVICE_TABLE(auxiliary, mlx5r_id_table);
4973 
4974 static struct auxiliary_driver mlx5r_mp_driver = {
4975 	.name = "multiport",
4976 	.probe = mlx5r_mp_probe,
4977 	.remove = mlx5r_mp_remove,
4978 	.id_table = mlx5r_mp_id_table,
4979 };
4980 
4981 static struct auxiliary_driver mlx5r_driver = {
4982 	.name = "rdma",
4983 	.probe = mlx5r_probe,
4984 	.remove = mlx5r_remove,
4985 	.id_table = mlx5r_id_table,
4986 };
4987 
4988 static int __init mlx5_ib_init(void)
4989 {
4990 	int ret;
4991 
4992 	xlt_emergency_page = (void *)__get_free_page(GFP_KERNEL);
4993 	if (!xlt_emergency_page)
4994 		return -ENOMEM;
4995 
4996 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
4997 	if (!mlx5_ib_event_wq) {
4998 		free_page((unsigned long)xlt_emergency_page);
4999 		return -ENOMEM;
5000 	}
5001 
5002 	ret = mlx5_ib_qp_event_init();
5003 	if (ret)
5004 		goto qp_event_err;
5005 
5006 	mlx5_ib_odp_init();
5007 	ret = mlx5r_rep_init();
5008 	if (ret)
5009 		goto rep_err;
5010 	ret = mlx5_data_direct_driver_register();
5011 	if (ret)
5012 		goto dd_err;
5013 	ret = auxiliary_driver_register(&mlx5r_mp_driver);
5014 	if (ret)
5015 		goto mp_err;
5016 	ret = auxiliary_driver_register(&mlx5r_driver);
5017 	if (ret)
5018 		goto drv_err;
5019 
5020 	return 0;
5021 
5022 drv_err:
5023 	auxiliary_driver_unregister(&mlx5r_mp_driver);
5024 mp_err:
5025 	mlx5_data_direct_driver_unregister();
5026 dd_err:
5027 	mlx5r_rep_cleanup();
5028 rep_err:
5029 	mlx5_ib_qp_event_cleanup();
5030 qp_event_err:
5031 	destroy_workqueue(mlx5_ib_event_wq);
5032 	free_page((unsigned long)xlt_emergency_page);
5033 	return ret;
5034 }
5035 
5036 static void __exit mlx5_ib_cleanup(void)
5037 {
5038 	mlx5_data_direct_driver_unregister();
5039 	auxiliary_driver_unregister(&mlx5r_driver);
5040 	auxiliary_driver_unregister(&mlx5r_mp_driver);
5041 	mlx5r_rep_cleanup();
5042 
5043 	mlx5_ib_qp_event_cleanup();
5044 	destroy_workqueue(mlx5_ib_event_wq);
5045 	free_page((unsigned long)xlt_emergency_page);
5046 }
5047 
5048 module_init(mlx5_ib_init);
5049 module_exit(mlx5_ib_cleanup);
5050