xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision 1587db11)
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 "macsec.h"
51 
52 #define UVERBS_MODULE_NAME mlx5_ib
53 #include <rdma/uverbs_named_ioctl.h>
54 
55 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
56 MODULE_DESCRIPTION("Mellanox 5th generation network adapters (ConnectX series) IB driver");
57 MODULE_LICENSE("Dual BSD/GPL");
58 
59 struct mlx5_ib_event_work {
60 	struct work_struct	work;
61 	union {
62 		struct mlx5_ib_dev	      *dev;
63 		struct mlx5_ib_multiport_info *mpi;
64 	};
65 	bool			is_slave;
66 	unsigned int		event;
67 	void			*param;
68 };
69 
70 enum {
71 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
72 };
73 
74 static struct workqueue_struct *mlx5_ib_event_wq;
75 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
76 static LIST_HEAD(mlx5_ib_dev_list);
77 /*
78  * This mutex should be held when accessing either of the above lists
79  */
80 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
81 
82 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
83 {
84 	struct mlx5_ib_dev *dev;
85 
86 	mutex_lock(&mlx5_ib_multiport_mutex);
87 	dev = mpi->ibdev;
88 	mutex_unlock(&mlx5_ib_multiport_mutex);
89 	return dev;
90 }
91 
92 static enum rdma_link_layer
93 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
94 {
95 	switch (port_type_cap) {
96 	case MLX5_CAP_PORT_TYPE_IB:
97 		return IB_LINK_LAYER_INFINIBAND;
98 	case MLX5_CAP_PORT_TYPE_ETH:
99 		return IB_LINK_LAYER_ETHERNET;
100 	default:
101 		return IB_LINK_LAYER_UNSPECIFIED;
102 	}
103 }
104 
105 static enum rdma_link_layer
106 mlx5_ib_port_link_layer(struct ib_device *device, u32 port_num)
107 {
108 	struct mlx5_ib_dev *dev = to_mdev(device);
109 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
110 
111 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
112 }
113 
114 static int get_port_state(struct ib_device *ibdev,
115 			  u32 port_num,
116 			  enum ib_port_state *state)
117 {
118 	struct ib_port_attr attr;
119 	int ret;
120 
121 	memset(&attr, 0, sizeof(attr));
122 	ret = ibdev->ops.query_port(ibdev, port_num, &attr);
123 	if (!ret)
124 		*state = attr.state;
125 	return ret;
126 }
127 
128 static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev,
129 					   struct net_device *ndev,
130 					   struct net_device *upper,
131 					   u32 *port_num)
132 {
133 	struct net_device *rep_ndev;
134 	struct mlx5_ib_port *port;
135 	int i;
136 
137 	for (i = 0; i < dev->num_ports; i++) {
138 		port  = &dev->port[i];
139 		if (!port->rep)
140 			continue;
141 
142 		if (upper == ndev && port->rep->vport == MLX5_VPORT_UPLINK) {
143 			*port_num = i + 1;
144 			return &port->roce;
145 		}
146 
147 		if (upper && port->rep->vport == MLX5_VPORT_UPLINK)
148 			continue;
149 
150 		read_lock(&port->roce.netdev_lock);
151 		rep_ndev = mlx5_ib_get_rep_netdev(port->rep->esw,
152 						  port->rep->vport);
153 		if (rep_ndev == ndev) {
154 			read_unlock(&port->roce.netdev_lock);
155 			*port_num = i + 1;
156 			return &port->roce;
157 		}
158 		read_unlock(&port->roce.netdev_lock);
159 	}
160 
161 	return NULL;
162 }
163 
164 static int mlx5_netdev_event(struct notifier_block *this,
165 			     unsigned long event, void *ptr)
166 {
167 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
168 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
169 	u32 port_num = roce->native_port_num;
170 	struct mlx5_core_dev *mdev;
171 	struct mlx5_ib_dev *ibdev;
172 
173 	ibdev = roce->dev;
174 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
175 	if (!mdev)
176 		return NOTIFY_DONE;
177 
178 	switch (event) {
179 	case NETDEV_REGISTER:
180 		/* Should already be registered during the load */
181 		if (ibdev->is_rep)
182 			break;
183 		write_lock(&roce->netdev_lock);
184 		if (ndev->dev.parent == mdev->device)
185 			roce->netdev = ndev;
186 		write_unlock(&roce->netdev_lock);
187 		break;
188 
189 	case NETDEV_UNREGISTER:
190 		/* In case of reps, ib device goes away before the netdevs */
191 		write_lock(&roce->netdev_lock);
192 		if (roce->netdev == ndev)
193 			roce->netdev = NULL;
194 		write_unlock(&roce->netdev_lock);
195 		break;
196 
197 	case NETDEV_CHANGE:
198 	case NETDEV_UP:
199 	case NETDEV_DOWN: {
200 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
201 		struct net_device *upper = NULL;
202 
203 		if (lag_ndev) {
204 			upper = netdev_master_upper_dev_get(lag_ndev);
205 			dev_put(lag_ndev);
206 		}
207 
208 		if (ibdev->is_rep)
209 			roce = mlx5_get_rep_roce(ibdev, ndev, upper, &port_num);
210 		if (!roce)
211 			return NOTIFY_DONE;
212 		if ((upper == ndev ||
213 		     ((!upper || ibdev->is_rep) && ndev == roce->netdev)) &&
214 		    ibdev->ib_active) {
215 			struct ib_event ibev = { };
216 			enum ib_port_state port_state;
217 
218 			if (get_port_state(&ibdev->ib_dev, port_num,
219 					   &port_state))
220 				goto done;
221 
222 			if (roce->last_port_state == port_state)
223 				goto done;
224 
225 			roce->last_port_state = port_state;
226 			ibev.device = &ibdev->ib_dev;
227 			if (port_state == IB_PORT_DOWN)
228 				ibev.event = IB_EVENT_PORT_ERR;
229 			else if (port_state == IB_PORT_ACTIVE)
230 				ibev.event = IB_EVENT_PORT_ACTIVE;
231 			else
232 				goto done;
233 
234 			ibev.element.port_num = port_num;
235 			ib_dispatch_event(&ibev);
236 		}
237 		break;
238 	}
239 
240 	default:
241 		break;
242 	}
243 done:
244 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
245 	return NOTIFY_DONE;
246 }
247 
248 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
249 					     u32 port_num)
250 {
251 	struct mlx5_ib_dev *ibdev = to_mdev(device);
252 	struct net_device *ndev;
253 	struct mlx5_core_dev *mdev;
254 
255 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
256 	if (!mdev)
257 		return NULL;
258 
259 	ndev = mlx5_lag_get_roce_netdev(mdev);
260 	if (ndev)
261 		goto out;
262 
263 	/* Ensure ndev does not disappear before we invoke dev_hold()
264 	 */
265 	read_lock(&ibdev->port[port_num - 1].roce.netdev_lock);
266 	ndev = ibdev->port[port_num - 1].roce.netdev;
267 	if (ndev)
268 		dev_hold(ndev);
269 	read_unlock(&ibdev->port[port_num - 1].roce.netdev_lock);
270 
271 out:
272 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
273 	return ndev;
274 }
275 
276 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
277 						   u32 ib_port_num,
278 						   u32 *native_port_num)
279 {
280 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
281 							  ib_port_num);
282 	struct mlx5_core_dev *mdev = NULL;
283 	struct mlx5_ib_multiport_info *mpi;
284 	struct mlx5_ib_port *port;
285 
286 	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
287 	    ll != IB_LINK_LAYER_ETHERNET) {
288 		if (native_port_num)
289 			*native_port_num = ib_port_num;
290 		return ibdev->mdev;
291 	}
292 
293 	if (native_port_num)
294 		*native_port_num = 1;
295 
296 	port = &ibdev->port[ib_port_num - 1];
297 	spin_lock(&port->mp.mpi_lock);
298 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
299 	if (mpi && !mpi->unaffiliate) {
300 		mdev = mpi->mdev;
301 		/* If it's the master no need to refcount, it'll exist
302 		 * as long as the ib_dev exists.
303 		 */
304 		if (!mpi->is_master)
305 			mpi->mdev_refcnt++;
306 	}
307 	spin_unlock(&port->mp.mpi_lock);
308 
309 	return mdev;
310 }
311 
312 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u32 port_num)
313 {
314 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
315 							  port_num);
316 	struct mlx5_ib_multiport_info *mpi;
317 	struct mlx5_ib_port *port;
318 
319 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
320 		return;
321 
322 	port = &ibdev->port[port_num - 1];
323 
324 	spin_lock(&port->mp.mpi_lock);
325 	mpi = ibdev->port[port_num - 1].mp.mpi;
326 	if (mpi->is_master)
327 		goto out;
328 
329 	mpi->mdev_refcnt--;
330 	if (mpi->unaffiliate)
331 		complete(&mpi->unref_comp);
332 out:
333 	spin_unlock(&port->mp.mpi_lock);
334 }
335 
336 static int translate_eth_legacy_proto_oper(u32 eth_proto_oper,
337 					   u16 *active_speed, u8 *active_width)
338 {
339 	switch (eth_proto_oper) {
340 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
341 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
342 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
343 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
344 		*active_width = IB_WIDTH_1X;
345 		*active_speed = IB_SPEED_SDR;
346 		break;
347 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
348 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
349 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
350 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
351 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
352 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
353 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
354 		*active_width = IB_WIDTH_1X;
355 		*active_speed = IB_SPEED_QDR;
356 		break;
357 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
358 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
359 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
360 		*active_width = IB_WIDTH_1X;
361 		*active_speed = IB_SPEED_EDR;
362 		break;
363 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
364 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
365 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
366 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
367 		*active_width = IB_WIDTH_4X;
368 		*active_speed = IB_SPEED_QDR;
369 		break;
370 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
371 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
372 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
373 		*active_width = IB_WIDTH_1X;
374 		*active_speed = IB_SPEED_HDR;
375 		break;
376 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
377 		*active_width = IB_WIDTH_4X;
378 		*active_speed = IB_SPEED_FDR;
379 		break;
380 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
381 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
382 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
383 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
384 		*active_width = IB_WIDTH_4X;
385 		*active_speed = IB_SPEED_EDR;
386 		break;
387 	default:
388 		return -EINVAL;
389 	}
390 
391 	return 0;
392 }
393 
394 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u16 *active_speed,
395 					u8 *active_width)
396 {
397 	switch (eth_proto_oper) {
398 	case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
399 	case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
400 		*active_width = IB_WIDTH_1X;
401 		*active_speed = IB_SPEED_SDR;
402 		break;
403 	case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
404 		*active_width = IB_WIDTH_1X;
405 		*active_speed = IB_SPEED_DDR;
406 		break;
407 	case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
408 		*active_width = IB_WIDTH_1X;
409 		*active_speed = IB_SPEED_QDR;
410 		break;
411 	case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
412 		*active_width = IB_WIDTH_4X;
413 		*active_speed = IB_SPEED_QDR;
414 		break;
415 	case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
416 		*active_width = IB_WIDTH_1X;
417 		*active_speed = IB_SPEED_EDR;
418 		break;
419 	case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
420 		*active_width = IB_WIDTH_2X;
421 		*active_speed = IB_SPEED_EDR;
422 		break;
423 	case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
424 		*active_width = IB_WIDTH_1X;
425 		*active_speed = IB_SPEED_HDR;
426 		break;
427 	case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
428 		*active_width = IB_WIDTH_4X;
429 		*active_speed = IB_SPEED_EDR;
430 		break;
431 	case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
432 		*active_width = IB_WIDTH_2X;
433 		*active_speed = IB_SPEED_HDR;
434 		break;
435 	case MLX5E_PROT_MASK(MLX5E_100GAUI_1_100GBASE_CR_KR):
436 		*active_width = IB_WIDTH_1X;
437 		*active_speed = IB_SPEED_NDR;
438 		break;
439 	case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4):
440 		*active_width = IB_WIDTH_4X;
441 		*active_speed = IB_SPEED_HDR;
442 		break;
443 	case MLX5E_PROT_MASK(MLX5E_200GAUI_2_200GBASE_CR2_KR2):
444 		*active_width = IB_WIDTH_2X;
445 		*active_speed = IB_SPEED_NDR;
446 		break;
447 	case MLX5E_PROT_MASK(MLX5E_400GAUI_8_400GBASE_CR8):
448 		*active_width = IB_WIDTH_8X;
449 		*active_speed = IB_SPEED_HDR;
450 		break;
451 	case MLX5E_PROT_MASK(MLX5E_400GAUI_4_400GBASE_CR4_KR4):
452 		*active_width = IB_WIDTH_4X;
453 		*active_speed = IB_SPEED_NDR;
454 		break;
455 	default:
456 		return -EINVAL;
457 	}
458 
459 	return 0;
460 }
461 
462 static int translate_eth_proto_oper(u32 eth_proto_oper, u16 *active_speed,
463 				    u8 *active_width, bool ext)
464 {
465 	return ext ?
466 		translate_eth_ext_proto_oper(eth_proto_oper, active_speed,
467 					     active_width) :
468 		translate_eth_legacy_proto_oper(eth_proto_oper, active_speed,
469 						active_width);
470 }
471 
472 static int mlx5_query_port_roce(struct ib_device *device, u32 port_num,
473 				struct ib_port_attr *props)
474 {
475 	struct mlx5_ib_dev *dev = to_mdev(device);
476 	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0};
477 	struct mlx5_core_dev *mdev;
478 	struct net_device *ndev, *upper;
479 	enum ib_mtu ndev_ib_mtu;
480 	bool put_mdev = true;
481 	u32 eth_prot_oper;
482 	u32 mdev_port_num;
483 	bool ext;
484 	int err;
485 
486 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
487 	if (!mdev) {
488 		/* This means the port isn't affiliated yet. Get the
489 		 * info for the master port instead.
490 		 */
491 		put_mdev = false;
492 		mdev = dev->mdev;
493 		mdev_port_num = 1;
494 		port_num = 1;
495 	}
496 
497 	/* Possible bad flows are checked before filling out props so in case
498 	 * of an error it will still be zeroed out.
499 	 * Use native port in case of reps
500 	 */
501 	if (dev->is_rep)
502 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
503 					   1);
504 	else
505 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
506 					   mdev_port_num);
507 	if (err)
508 		goto out;
509 	ext = !!MLX5_GET_ETH_PROTO(ptys_reg, out, true, eth_proto_capability);
510 	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
511 
512 	props->active_width     = IB_WIDTH_4X;
513 	props->active_speed     = IB_SPEED_QDR;
514 
515 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
516 				 &props->active_width, ext);
517 
518 	if (!dev->is_rep && dev->mdev->roce.roce_en) {
519 		u16 qkey_viol_cntr;
520 
521 		props->port_cap_flags |= IB_PORT_CM_SUP;
522 		props->ip_gids = true;
523 		props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev,
524 						   roce_address_table_size);
525 		mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
526 		props->qkey_viol_cntr = qkey_viol_cntr;
527 	}
528 	props->max_mtu          = IB_MTU_4096;
529 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
530 	props->pkey_tbl_len     = 1;
531 	props->state            = IB_PORT_DOWN;
532 	props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
533 
534 	/* If this is a stub query for an unaffiliated port stop here */
535 	if (!put_mdev)
536 		goto out;
537 
538 	ndev = mlx5_ib_get_netdev(device, port_num);
539 	if (!ndev)
540 		goto out;
541 
542 	if (mlx5_lag_is_roce(mdev) || mlx5_lag_is_sriov(mdev)) {
543 		rcu_read_lock();
544 		upper = netdev_master_upper_dev_get_rcu(ndev);
545 		if (upper) {
546 			dev_put(ndev);
547 			ndev = upper;
548 			dev_hold(ndev);
549 		}
550 		rcu_read_unlock();
551 	}
552 
553 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
554 		props->state      = IB_PORT_ACTIVE;
555 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
556 	}
557 
558 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
559 
560 	dev_put(ndev);
561 
562 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
563 out:
564 	if (put_mdev)
565 		mlx5_ib_put_native_port_mdev(dev, port_num);
566 	return err;
567 }
568 
569 int set_roce_addr(struct mlx5_ib_dev *dev, u32 port_num,
570 		  unsigned int index, const union ib_gid *gid,
571 		  const struct ib_gid_attr *attr)
572 {
573 	enum ib_gid_type gid_type;
574 	u16 vlan_id = 0xffff;
575 	u8 roce_version = 0;
576 	u8 roce_l3_type = 0;
577 	u8 mac[ETH_ALEN];
578 	int ret;
579 
580 	gid_type = attr->gid_type;
581 	if (gid) {
582 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
583 		if (ret)
584 			return ret;
585 	}
586 
587 	switch (gid_type) {
588 	case IB_GID_TYPE_ROCE:
589 		roce_version = MLX5_ROCE_VERSION_1;
590 		break;
591 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
592 		roce_version = MLX5_ROCE_VERSION_2;
593 		if (gid && ipv6_addr_v4mapped((void *)gid))
594 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
595 		else
596 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
597 		break;
598 
599 	default:
600 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
601 	}
602 
603 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
604 				      roce_l3_type, gid->raw, mac,
605 				      vlan_id < VLAN_CFI_MASK, vlan_id,
606 				      port_num);
607 }
608 
609 static int mlx5_ib_add_gid(const struct ib_gid_attr *attr,
610 			   __always_unused void **context)
611 {
612 	int ret;
613 
614 	ret = mlx5r_add_gid_macsec_operations(attr);
615 	if (ret)
616 		return ret;
617 
618 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
619 			     attr->index, &attr->gid, attr);
620 }
621 
622 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
623 			   __always_unused void **context)
624 {
625 	int ret;
626 
627 	ret = set_roce_addr(to_mdev(attr->device), attr->port_num,
628 			    attr->index, NULL, attr);
629 	if (ret)
630 		return ret;
631 
632 	mlx5r_del_gid_macsec_operations(attr);
633 	return 0;
634 }
635 
636 __be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev,
637 				   const struct ib_gid_attr *attr)
638 {
639 	if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
640 		return 0;
641 
642 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
643 }
644 
645 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
646 {
647 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
648 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
649 	return 0;
650 }
651 
652 enum {
653 	MLX5_VPORT_ACCESS_METHOD_MAD,
654 	MLX5_VPORT_ACCESS_METHOD_HCA,
655 	MLX5_VPORT_ACCESS_METHOD_NIC,
656 };
657 
658 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
659 {
660 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
661 		return MLX5_VPORT_ACCESS_METHOD_MAD;
662 
663 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
664 	    IB_LINK_LAYER_ETHERNET)
665 		return MLX5_VPORT_ACCESS_METHOD_NIC;
666 
667 	return MLX5_VPORT_ACCESS_METHOD_HCA;
668 }
669 
670 static void get_atomic_caps(struct mlx5_ib_dev *dev,
671 			    u8 atomic_size_qp,
672 			    struct ib_device_attr *props)
673 {
674 	u8 tmp;
675 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
676 	u8 atomic_req_8B_endianness_mode =
677 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
678 
679 	/* Check if HW supports 8 bytes standard atomic operations and capable
680 	 * of host endianness respond
681 	 */
682 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
683 	if (((atomic_operations & tmp) == tmp) &&
684 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
685 	    (atomic_req_8B_endianness_mode)) {
686 		props->atomic_cap = IB_ATOMIC_HCA;
687 	} else {
688 		props->atomic_cap = IB_ATOMIC_NONE;
689 	}
690 }
691 
692 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
693 			       struct ib_device_attr *props)
694 {
695 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
696 
697 	get_atomic_caps(dev, atomic_size_qp, props);
698 }
699 
700 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
701 					__be64 *sys_image_guid)
702 {
703 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
704 	struct mlx5_core_dev *mdev = dev->mdev;
705 	u64 tmp;
706 	int err;
707 
708 	switch (mlx5_get_vport_access_method(ibdev)) {
709 	case MLX5_VPORT_ACCESS_METHOD_MAD:
710 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
711 							    sys_image_guid);
712 
713 	case MLX5_VPORT_ACCESS_METHOD_HCA:
714 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
715 		break;
716 
717 	case MLX5_VPORT_ACCESS_METHOD_NIC:
718 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
719 		break;
720 
721 	default:
722 		return -EINVAL;
723 	}
724 
725 	if (!err)
726 		*sys_image_guid = cpu_to_be64(tmp);
727 
728 	return err;
729 
730 }
731 
732 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
733 				u16 *max_pkeys)
734 {
735 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
736 	struct mlx5_core_dev *mdev = dev->mdev;
737 
738 	switch (mlx5_get_vport_access_method(ibdev)) {
739 	case MLX5_VPORT_ACCESS_METHOD_MAD:
740 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
741 
742 	case MLX5_VPORT_ACCESS_METHOD_HCA:
743 	case MLX5_VPORT_ACCESS_METHOD_NIC:
744 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
745 						pkey_table_size));
746 		return 0;
747 
748 	default:
749 		return -EINVAL;
750 	}
751 }
752 
753 static int mlx5_query_vendor_id(struct ib_device *ibdev,
754 				u32 *vendor_id)
755 {
756 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
757 
758 	switch (mlx5_get_vport_access_method(ibdev)) {
759 	case MLX5_VPORT_ACCESS_METHOD_MAD:
760 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
761 
762 	case MLX5_VPORT_ACCESS_METHOD_HCA:
763 	case MLX5_VPORT_ACCESS_METHOD_NIC:
764 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
765 
766 	default:
767 		return -EINVAL;
768 	}
769 }
770 
771 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
772 				__be64 *node_guid)
773 {
774 	u64 tmp;
775 	int err;
776 
777 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
778 	case MLX5_VPORT_ACCESS_METHOD_MAD:
779 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
780 
781 	case MLX5_VPORT_ACCESS_METHOD_HCA:
782 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
783 		break;
784 
785 	case MLX5_VPORT_ACCESS_METHOD_NIC:
786 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
787 		break;
788 
789 	default:
790 		return -EINVAL;
791 	}
792 
793 	if (!err)
794 		*node_guid = cpu_to_be64(tmp);
795 
796 	return err;
797 }
798 
799 struct mlx5_reg_node_desc {
800 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
801 };
802 
803 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
804 {
805 	struct mlx5_reg_node_desc in;
806 
807 	if (mlx5_use_mad_ifc(dev))
808 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
809 
810 	memset(&in, 0, sizeof(in));
811 
812 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
813 				    sizeof(struct mlx5_reg_node_desc),
814 				    MLX5_REG_NODE_DESC, 0, 0);
815 }
816 
817 static int mlx5_ib_query_device(struct ib_device *ibdev,
818 				struct ib_device_attr *props,
819 				struct ib_udata *uhw)
820 {
821 	size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
822 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
823 	struct mlx5_core_dev *mdev = dev->mdev;
824 	int err = -ENOMEM;
825 	int max_sq_desc;
826 	int max_rq_sg;
827 	int max_sq_sg;
828 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
829 	bool raw_support = !mlx5_core_mp_enabled(mdev);
830 	struct mlx5_ib_query_device_resp resp = {};
831 	size_t resp_len;
832 	u64 max_tso;
833 
834 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
835 	if (uhw_outlen && uhw_outlen < resp_len)
836 		return -EINVAL;
837 
838 	resp.response_length = resp_len;
839 
840 	if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
841 		return -EINVAL;
842 
843 	memset(props, 0, sizeof(*props));
844 	err = mlx5_query_system_image_guid(ibdev,
845 					   &props->sys_image_guid);
846 	if (err)
847 		return err;
848 
849 	props->max_pkeys = dev->pkey_table_len;
850 
851 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
852 	if (err)
853 		return err;
854 
855 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
856 		(fw_rev_min(dev->mdev) << 16) |
857 		fw_rev_sub(dev->mdev);
858 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
859 		IB_DEVICE_PORT_ACTIVE_EVENT		|
860 		IB_DEVICE_SYS_IMAGE_GUID		|
861 		IB_DEVICE_RC_RNR_NAK_GEN;
862 
863 	if (MLX5_CAP_GEN(mdev, pkv))
864 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
865 	if (MLX5_CAP_GEN(mdev, qkv))
866 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
867 	if (MLX5_CAP_GEN(mdev, apm))
868 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
869 	if (MLX5_CAP_GEN(mdev, xrc))
870 		props->device_cap_flags |= IB_DEVICE_XRC;
871 	if (MLX5_CAP_GEN(mdev, imaicl)) {
872 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
873 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
874 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
875 		/* We support 'Gappy' memory registration too */
876 		props->kernel_cap_flags |= IBK_SG_GAPS_REG;
877 	}
878 	/* IB_WR_REG_MR always requires changing the entity size with UMR */
879 	if (!MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled))
880 		props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
881 	if (MLX5_CAP_GEN(mdev, sho)) {
882 		props->kernel_cap_flags |= IBK_INTEGRITY_HANDOVER;
883 		/* At this stage no support for signature handover */
884 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
885 				      IB_PROT_T10DIF_TYPE_2 |
886 				      IB_PROT_T10DIF_TYPE_3;
887 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
888 				       IB_GUARD_T10DIF_CSUM;
889 	}
890 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
891 		props->kernel_cap_flags |= IBK_BLOCK_MULTICAST_LOOPBACK;
892 
893 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
894 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
895 			/* Legacy bit to support old userspace libraries */
896 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
897 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
898 		}
899 
900 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
901 			props->raw_packet_caps |=
902 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
903 
904 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) {
905 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
906 			if (max_tso) {
907 				resp.tso_caps.max_tso = 1 << max_tso;
908 				resp.tso_caps.supported_qpts |=
909 					1 << IB_QPT_RAW_PACKET;
910 				resp.response_length += sizeof(resp.tso_caps);
911 			}
912 		}
913 
914 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) {
915 			resp.rss_caps.rx_hash_function =
916 						MLX5_RX_HASH_FUNC_TOEPLITZ;
917 			resp.rss_caps.rx_hash_fields_mask =
918 						MLX5_RX_HASH_SRC_IPV4 |
919 						MLX5_RX_HASH_DST_IPV4 |
920 						MLX5_RX_HASH_SRC_IPV6 |
921 						MLX5_RX_HASH_DST_IPV6 |
922 						MLX5_RX_HASH_SRC_PORT_TCP |
923 						MLX5_RX_HASH_DST_PORT_TCP |
924 						MLX5_RX_HASH_SRC_PORT_UDP |
925 						MLX5_RX_HASH_DST_PORT_UDP |
926 						MLX5_RX_HASH_INNER;
927 			resp.response_length += sizeof(resp.rss_caps);
928 		}
929 	} else {
930 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen)
931 			resp.response_length += sizeof(resp.tso_caps);
932 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen)
933 			resp.response_length += sizeof(resp.rss_caps);
934 	}
935 
936 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
937 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
938 		props->kernel_cap_flags |= IBK_UD_TSO;
939 	}
940 
941 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
942 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
943 	    raw_support)
944 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
945 
946 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
947 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
948 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
949 
950 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
951 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
952 	    raw_support) {
953 		/* Legacy bit to support old userspace libraries */
954 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
955 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
956 	}
957 
958 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
959 		props->max_dm_size =
960 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
961 	}
962 
963 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
964 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
965 
966 	if (MLX5_CAP_GEN(mdev, end_pad))
967 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
968 
969 	props->vendor_part_id	   = mdev->pdev->device;
970 	props->hw_ver		   = mdev->pdev->revision;
971 
972 	props->max_mr_size	   = ~0ull;
973 	props->page_size_cap	   = ~(min_page_size - 1);
974 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
975 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
976 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
977 		     sizeof(struct mlx5_wqe_data_seg);
978 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
979 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
980 		     sizeof(struct mlx5_wqe_raddr_seg)) /
981 		sizeof(struct mlx5_wqe_data_seg);
982 	props->max_send_sge = max_sq_sg;
983 	props->max_recv_sge = max_rq_sg;
984 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
985 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
986 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
987 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
988 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
989 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
990 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
991 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
992 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
993 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
994 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
995 	props->max_srq_sge	   = max_rq_sg - 1;
996 	props->max_fast_reg_page_list_len =
997 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
998 	props->max_pi_fast_reg_page_list_len =
999 		props->max_fast_reg_page_list_len / 2;
1000 	props->max_sgl_rd =
1001 		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
1002 	get_atomic_caps_qp(dev, props);
1003 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
1004 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
1005 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
1006 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1007 					   props->max_mcast_grp;
1008 	props->max_ah = INT_MAX;
1009 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
1010 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
1011 
1012 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
1013 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
1014 			props->kernel_cap_flags |= IBK_ON_DEMAND_PAGING;
1015 		props->odp_caps = dev->odp_caps;
1016 		if (!uhw) {
1017 			/* ODP for kernel QPs is not implemented for receive
1018 			 * WQEs and SRQ WQEs
1019 			 */
1020 			props->odp_caps.per_transport_caps.rc_odp_caps &=
1021 				~(IB_ODP_SUPPORT_READ |
1022 				  IB_ODP_SUPPORT_SRQ_RECV);
1023 			props->odp_caps.per_transport_caps.uc_odp_caps &=
1024 				~(IB_ODP_SUPPORT_READ |
1025 				  IB_ODP_SUPPORT_SRQ_RECV);
1026 			props->odp_caps.per_transport_caps.ud_odp_caps &=
1027 				~(IB_ODP_SUPPORT_READ |
1028 				  IB_ODP_SUPPORT_SRQ_RECV);
1029 			props->odp_caps.per_transport_caps.xrc_odp_caps &=
1030 				~(IB_ODP_SUPPORT_READ |
1031 				  IB_ODP_SUPPORT_SRQ_RECV);
1032 		}
1033 	}
1034 
1035 	if (mlx5_core_is_vf(mdev))
1036 		props->kernel_cap_flags |= IBK_VIRTUAL_FUNCTION;
1037 
1038 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1039 	    IB_LINK_LAYER_ETHERNET && raw_support) {
1040 		props->rss_caps.max_rwq_indirection_tables =
1041 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1042 		props->rss_caps.max_rwq_indirection_table_size =
1043 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1044 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1045 		props->max_wq_type_rq =
1046 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1047 	}
1048 
1049 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1050 		props->tm_caps.max_num_tags =
1051 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1052 		props->tm_caps.max_ops =
1053 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1054 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1055 	}
1056 
1057 	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1058 	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1059 		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1060 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1061 	}
1062 
1063 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1064 		props->cq_caps.max_cq_moderation_count =
1065 						MLX5_MAX_CQ_COUNT;
1066 		props->cq_caps.max_cq_moderation_period =
1067 						MLX5_MAX_CQ_PERIOD;
1068 	}
1069 
1070 	if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) {
1071 		resp.response_length += sizeof(resp.cqe_comp_caps);
1072 
1073 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1074 			resp.cqe_comp_caps.max_num =
1075 				MLX5_CAP_GEN(dev->mdev,
1076 					     cqe_compression_max_num);
1077 
1078 			resp.cqe_comp_caps.supported_format =
1079 				MLX5_IB_CQE_RES_FORMAT_HASH |
1080 				MLX5_IB_CQE_RES_FORMAT_CSUM;
1081 
1082 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1083 				resp.cqe_comp_caps.supported_format |=
1084 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1085 		}
1086 	}
1087 
1088 	if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen &&
1089 	    raw_support) {
1090 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1091 		    MLX5_CAP_GEN(mdev, qos)) {
1092 			resp.packet_pacing_caps.qp_rate_limit_max =
1093 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1094 			resp.packet_pacing_caps.qp_rate_limit_min =
1095 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1096 			resp.packet_pacing_caps.supported_qpts |=
1097 				1 << IB_QPT_RAW_PACKET;
1098 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1099 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1100 				resp.packet_pacing_caps.cap_flags |=
1101 					MLX5_IB_PP_SUPPORT_BURST;
1102 		}
1103 		resp.response_length += sizeof(resp.packet_pacing_caps);
1104 	}
1105 
1106 	if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <=
1107 	    uhw_outlen) {
1108 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1109 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1110 				MLX5_IB_ALLOW_MPW;
1111 
1112 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1113 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1114 				MLX5_IB_SUPPORT_EMPW;
1115 
1116 		resp.response_length +=
1117 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1118 	}
1119 
1120 	if (offsetofend(typeof(resp), flags) <= uhw_outlen) {
1121 		resp.response_length += sizeof(resp.flags);
1122 
1123 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1124 			resp.flags |=
1125 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1126 
1127 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1128 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1129 		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1130 			resp.flags |=
1131 				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1132 
1133 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1134 	}
1135 
1136 	if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) {
1137 		resp.response_length += sizeof(resp.sw_parsing_caps);
1138 		if (MLX5_CAP_ETH(mdev, swp)) {
1139 			resp.sw_parsing_caps.sw_parsing_offloads |=
1140 				MLX5_IB_SW_PARSING;
1141 
1142 			if (MLX5_CAP_ETH(mdev, swp_csum))
1143 				resp.sw_parsing_caps.sw_parsing_offloads |=
1144 					MLX5_IB_SW_PARSING_CSUM;
1145 
1146 			if (MLX5_CAP_ETH(mdev, swp_lso))
1147 				resp.sw_parsing_caps.sw_parsing_offloads |=
1148 					MLX5_IB_SW_PARSING_LSO;
1149 
1150 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1151 				resp.sw_parsing_caps.supported_qpts =
1152 					BIT(IB_QPT_RAW_PACKET);
1153 		}
1154 	}
1155 
1156 	if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen &&
1157 	    raw_support) {
1158 		resp.response_length += sizeof(resp.striding_rq_caps);
1159 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1160 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1161 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1162 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1163 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1164 			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1165 				resp.striding_rq_caps
1166 					.min_single_wqe_log_num_of_strides =
1167 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1168 			else
1169 				resp.striding_rq_caps
1170 					.min_single_wqe_log_num_of_strides =
1171 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1172 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1173 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1174 			resp.striding_rq_caps.supported_qpts =
1175 				BIT(IB_QPT_RAW_PACKET);
1176 		}
1177 	}
1178 
1179 	if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) {
1180 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1181 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1182 			resp.tunnel_offloads_caps |=
1183 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1184 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1185 			resp.tunnel_offloads_caps |=
1186 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1187 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1188 			resp.tunnel_offloads_caps |=
1189 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1190 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre))
1191 			resp.tunnel_offloads_caps |=
1192 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1193 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp))
1194 			resp.tunnel_offloads_caps |=
1195 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1196 	}
1197 
1198 	if (offsetofend(typeof(resp), dci_streams_caps) <= uhw_outlen) {
1199 		resp.response_length += sizeof(resp.dci_streams_caps);
1200 
1201 		resp.dci_streams_caps.max_log_num_concurent =
1202 			MLX5_CAP_GEN(mdev, log_max_dci_stream_channels);
1203 
1204 		resp.dci_streams_caps.max_log_num_errored =
1205 			MLX5_CAP_GEN(mdev, log_max_dci_errored_streams);
1206 	}
1207 
1208 	if (uhw_outlen) {
1209 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1210 
1211 		if (err)
1212 			return err;
1213 	}
1214 
1215 	return 0;
1216 }
1217 
1218 static void translate_active_width(struct ib_device *ibdev, u16 active_width,
1219 				   u8 *ib_width)
1220 {
1221 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1222 
1223 	if (active_width & MLX5_PTYS_WIDTH_1X)
1224 		*ib_width = IB_WIDTH_1X;
1225 	else if (active_width & MLX5_PTYS_WIDTH_2X)
1226 		*ib_width = IB_WIDTH_2X;
1227 	else if (active_width & MLX5_PTYS_WIDTH_4X)
1228 		*ib_width = IB_WIDTH_4X;
1229 	else if (active_width & MLX5_PTYS_WIDTH_8X)
1230 		*ib_width = IB_WIDTH_8X;
1231 	else if (active_width & MLX5_PTYS_WIDTH_12X)
1232 		*ib_width = IB_WIDTH_12X;
1233 	else {
1234 		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1235 			    active_width);
1236 		*ib_width = IB_WIDTH_4X;
1237 	}
1238 
1239 	return;
1240 }
1241 
1242 static int mlx5_mtu_to_ib_mtu(int mtu)
1243 {
1244 	switch (mtu) {
1245 	case 256: return 1;
1246 	case 512: return 2;
1247 	case 1024: return 3;
1248 	case 2048: return 4;
1249 	case 4096: return 5;
1250 	default:
1251 		pr_warn("invalid mtu\n");
1252 		return -1;
1253 	}
1254 }
1255 
1256 enum ib_max_vl_num {
1257 	__IB_MAX_VL_0		= 1,
1258 	__IB_MAX_VL_0_1		= 2,
1259 	__IB_MAX_VL_0_3		= 3,
1260 	__IB_MAX_VL_0_7		= 4,
1261 	__IB_MAX_VL_0_14	= 5,
1262 };
1263 
1264 enum mlx5_vl_hw_cap {
1265 	MLX5_VL_HW_0	= 1,
1266 	MLX5_VL_HW_0_1	= 2,
1267 	MLX5_VL_HW_0_2	= 3,
1268 	MLX5_VL_HW_0_3	= 4,
1269 	MLX5_VL_HW_0_4	= 5,
1270 	MLX5_VL_HW_0_5	= 6,
1271 	MLX5_VL_HW_0_6	= 7,
1272 	MLX5_VL_HW_0_7	= 8,
1273 	MLX5_VL_HW_0_14	= 15
1274 };
1275 
1276 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1277 				u8 *max_vl_num)
1278 {
1279 	switch (vl_hw_cap) {
1280 	case MLX5_VL_HW_0:
1281 		*max_vl_num = __IB_MAX_VL_0;
1282 		break;
1283 	case MLX5_VL_HW_0_1:
1284 		*max_vl_num = __IB_MAX_VL_0_1;
1285 		break;
1286 	case MLX5_VL_HW_0_3:
1287 		*max_vl_num = __IB_MAX_VL_0_3;
1288 		break;
1289 	case MLX5_VL_HW_0_7:
1290 		*max_vl_num = __IB_MAX_VL_0_7;
1291 		break;
1292 	case MLX5_VL_HW_0_14:
1293 		*max_vl_num = __IB_MAX_VL_0_14;
1294 		break;
1295 
1296 	default:
1297 		return -EINVAL;
1298 	}
1299 
1300 	return 0;
1301 }
1302 
1303 static int mlx5_query_hca_port(struct ib_device *ibdev, u32 port,
1304 			       struct ib_port_attr *props)
1305 {
1306 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1307 	struct mlx5_core_dev *mdev = dev->mdev;
1308 	struct mlx5_hca_vport_context *rep;
1309 	u16 max_mtu;
1310 	u16 oper_mtu;
1311 	int err;
1312 	u16 ib_link_width_oper;
1313 	u8 vl_hw_cap;
1314 
1315 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1316 	if (!rep) {
1317 		err = -ENOMEM;
1318 		goto out;
1319 	}
1320 
1321 	/* props being zeroed by the caller, avoid zeroing it here */
1322 
1323 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1324 	if (err)
1325 		goto out;
1326 
1327 	props->lid		= rep->lid;
1328 	props->lmc		= rep->lmc;
1329 	props->sm_lid		= rep->sm_lid;
1330 	props->sm_sl		= rep->sm_sl;
1331 	props->state		= rep->vport_state;
1332 	props->phys_state	= rep->port_physical_state;
1333 	props->port_cap_flags	= rep->cap_mask1;
1334 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1335 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1336 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1337 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1338 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1339 	props->subnet_timeout	= rep->subnet_timeout;
1340 	props->init_type_reply	= rep->init_type_reply;
1341 
1342 	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1343 		props->port_cap_flags2 = rep->cap_mask2;
1344 
1345 	err = mlx5_query_ib_port_oper(mdev, &ib_link_width_oper,
1346 				      &props->active_speed, port);
1347 	if (err)
1348 		goto out;
1349 
1350 	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1351 
1352 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1353 
1354 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1355 
1356 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1357 
1358 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1359 
1360 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1361 	if (err)
1362 		goto out;
1363 
1364 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1365 				   &props->max_vl_num);
1366 out:
1367 	kfree(rep);
1368 	return err;
1369 }
1370 
1371 int mlx5_ib_query_port(struct ib_device *ibdev, u32 port,
1372 		       struct ib_port_attr *props)
1373 {
1374 	unsigned int count;
1375 	int ret;
1376 
1377 	switch (mlx5_get_vport_access_method(ibdev)) {
1378 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1379 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1380 		break;
1381 
1382 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1383 		ret = mlx5_query_hca_port(ibdev, port, props);
1384 		break;
1385 
1386 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1387 		ret = mlx5_query_port_roce(ibdev, port, props);
1388 		break;
1389 
1390 	default:
1391 		ret = -EINVAL;
1392 	}
1393 
1394 	if (!ret && props) {
1395 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1396 		struct mlx5_core_dev *mdev;
1397 		bool put_mdev = true;
1398 
1399 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1400 		if (!mdev) {
1401 			/* If the port isn't affiliated yet query the master.
1402 			 * The master and slave will have the same values.
1403 			 */
1404 			mdev = dev->mdev;
1405 			port = 1;
1406 			put_mdev = false;
1407 		}
1408 		count = mlx5_core_reserved_gids_count(mdev);
1409 		if (put_mdev)
1410 			mlx5_ib_put_native_port_mdev(dev, port);
1411 		props->gid_tbl_len -= count;
1412 	}
1413 	return ret;
1414 }
1415 
1416 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u32 port,
1417 				  struct ib_port_attr *props)
1418 {
1419 	return mlx5_query_port_roce(ibdev, port, props);
1420 }
1421 
1422 static int mlx5_ib_rep_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1423 				  u16 *pkey)
1424 {
1425 	/* Default special Pkey for representor device port as per the
1426 	 * IB specification 1.3 section 10.9.1.2.
1427 	 */
1428 	*pkey = 0xffff;
1429 	return 0;
1430 }
1431 
1432 static int mlx5_ib_query_gid(struct ib_device *ibdev, u32 port, int index,
1433 			     union ib_gid *gid)
1434 {
1435 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1436 	struct mlx5_core_dev *mdev = dev->mdev;
1437 
1438 	switch (mlx5_get_vport_access_method(ibdev)) {
1439 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1440 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1441 
1442 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1443 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1444 
1445 	default:
1446 		return -EINVAL;
1447 	}
1448 
1449 }
1450 
1451 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u32 port,
1452 				   u16 index, u16 *pkey)
1453 {
1454 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1455 	struct mlx5_core_dev *mdev;
1456 	bool put_mdev = true;
1457 	u32 mdev_port_num;
1458 	int err;
1459 
1460 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1461 	if (!mdev) {
1462 		/* The port isn't affiliated yet, get the PKey from the master
1463 		 * port. For RoCE the PKey tables will be the same.
1464 		 */
1465 		put_mdev = false;
1466 		mdev = dev->mdev;
1467 		mdev_port_num = 1;
1468 	}
1469 
1470 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1471 					index, pkey);
1472 	if (put_mdev)
1473 		mlx5_ib_put_native_port_mdev(dev, port);
1474 
1475 	return err;
1476 }
1477 
1478 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1479 			      u16 *pkey)
1480 {
1481 	switch (mlx5_get_vport_access_method(ibdev)) {
1482 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1483 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1484 
1485 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1486 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1487 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1488 	default:
1489 		return -EINVAL;
1490 	}
1491 }
1492 
1493 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1494 				 struct ib_device_modify *props)
1495 {
1496 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1497 	struct mlx5_reg_node_desc in;
1498 	struct mlx5_reg_node_desc out;
1499 	int err;
1500 
1501 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1502 		return -EOPNOTSUPP;
1503 
1504 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1505 		return 0;
1506 
1507 	/*
1508 	 * If possible, pass node desc to FW, so it can generate
1509 	 * a 144 trap.  If cmd fails, just ignore.
1510 	 */
1511 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1512 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1513 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1514 	if (err)
1515 		return err;
1516 
1517 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1518 
1519 	return err;
1520 }
1521 
1522 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u32 port_num, u32 mask,
1523 				u32 value)
1524 {
1525 	struct mlx5_hca_vport_context ctx = {};
1526 	struct mlx5_core_dev *mdev;
1527 	u32 mdev_port_num;
1528 	int err;
1529 
1530 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1531 	if (!mdev)
1532 		return -ENODEV;
1533 
1534 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1535 	if (err)
1536 		goto out;
1537 
1538 	if (~ctx.cap_mask1_perm & mask) {
1539 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1540 			     mask, ctx.cap_mask1_perm);
1541 		err = -EINVAL;
1542 		goto out;
1543 	}
1544 
1545 	ctx.cap_mask1 = value;
1546 	ctx.cap_mask1_perm = mask;
1547 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1548 						 0, &ctx);
1549 
1550 out:
1551 	mlx5_ib_put_native_port_mdev(dev, port_num);
1552 
1553 	return err;
1554 }
1555 
1556 static int mlx5_ib_modify_port(struct ib_device *ibdev, u32 port, int mask,
1557 			       struct ib_port_modify *props)
1558 {
1559 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1560 	struct ib_port_attr attr;
1561 	u32 tmp;
1562 	int err;
1563 	u32 change_mask;
1564 	u32 value;
1565 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1566 		      IB_LINK_LAYER_INFINIBAND);
1567 
1568 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1569 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1570 	 */
1571 	if (!is_ib)
1572 		return 0;
1573 
1574 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1575 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1576 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1577 		return set_port_caps_atomic(dev, port, change_mask, value);
1578 	}
1579 
1580 	mutex_lock(&dev->cap_mask_mutex);
1581 
1582 	err = ib_query_port(ibdev, port, &attr);
1583 	if (err)
1584 		goto out;
1585 
1586 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1587 		~props->clr_port_cap_mask;
1588 
1589 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1590 
1591 out:
1592 	mutex_unlock(&dev->cap_mask_mutex);
1593 	return err;
1594 }
1595 
1596 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1597 {
1598 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1599 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1600 }
1601 
1602 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1603 {
1604 	/* Large page with non 4k uar support might limit the dynamic size */
1605 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1606 		return MLX5_MIN_DYN_BFREGS;
1607 
1608 	return MLX5_MAX_DYN_BFREGS;
1609 }
1610 
1611 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1612 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1613 			     struct mlx5_bfreg_info *bfregi)
1614 {
1615 	int uars_per_sys_page;
1616 	int bfregs_per_sys_page;
1617 	int ref_bfregs = req->total_num_bfregs;
1618 
1619 	if (req->total_num_bfregs == 0)
1620 		return -EINVAL;
1621 
1622 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1623 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1624 
1625 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1626 		return -ENOMEM;
1627 
1628 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1629 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1630 	/* This holds the required static allocation asked by the user */
1631 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1632 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1633 		return -EINVAL;
1634 
1635 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1636 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1637 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1638 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1639 
1640 	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",
1641 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1642 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1643 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1644 		    bfregi->num_sys_pages);
1645 
1646 	return 0;
1647 }
1648 
1649 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1650 {
1651 	struct mlx5_bfreg_info *bfregi;
1652 	int err;
1653 	int i;
1654 
1655 	bfregi = &context->bfregi;
1656 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1657 		err = mlx5_cmd_uar_alloc(dev->mdev, &bfregi->sys_pages[i],
1658 					 context->devx_uid);
1659 		if (err)
1660 			goto error;
1661 
1662 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1663 	}
1664 
1665 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1666 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1667 
1668 	return 0;
1669 
1670 error:
1671 	for (--i; i >= 0; i--)
1672 		if (mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i],
1673 					 context->devx_uid))
1674 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1675 
1676 	return err;
1677 }
1678 
1679 static void deallocate_uars(struct mlx5_ib_dev *dev,
1680 			    struct mlx5_ib_ucontext *context)
1681 {
1682 	struct mlx5_bfreg_info *bfregi;
1683 	int i;
1684 
1685 	bfregi = &context->bfregi;
1686 	for (i = 0; i < bfregi->num_sys_pages; i++)
1687 		if (i < bfregi->num_static_sys_pages ||
1688 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1689 			mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i],
1690 					     context->devx_uid);
1691 }
1692 
1693 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1694 {
1695 	int err = 0;
1696 
1697 	mutex_lock(&dev->lb.mutex);
1698 	if (td)
1699 		dev->lb.user_td++;
1700 	if (qp)
1701 		dev->lb.qps++;
1702 
1703 	if (dev->lb.user_td == 2 ||
1704 	    dev->lb.qps == 1) {
1705 		if (!dev->lb.enabled) {
1706 			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1707 			dev->lb.enabled = true;
1708 		}
1709 	}
1710 
1711 	mutex_unlock(&dev->lb.mutex);
1712 
1713 	return err;
1714 }
1715 
1716 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1717 {
1718 	mutex_lock(&dev->lb.mutex);
1719 	if (td)
1720 		dev->lb.user_td--;
1721 	if (qp)
1722 		dev->lb.qps--;
1723 
1724 	if (dev->lb.user_td == 1 &&
1725 	    dev->lb.qps == 0) {
1726 		if (dev->lb.enabled) {
1727 			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1728 			dev->lb.enabled = false;
1729 		}
1730 	}
1731 
1732 	mutex_unlock(&dev->lb.mutex);
1733 }
1734 
1735 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1736 					  u16 uid)
1737 {
1738 	int err;
1739 
1740 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1741 		return 0;
1742 
1743 	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1744 	if (err)
1745 		return err;
1746 
1747 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1748 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1749 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1750 		return err;
1751 
1752 	return mlx5_ib_enable_lb(dev, true, false);
1753 }
1754 
1755 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1756 					     u16 uid)
1757 {
1758 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1759 		return;
1760 
1761 	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1762 
1763 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1764 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1765 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1766 		return;
1767 
1768 	mlx5_ib_disable_lb(dev, true, false);
1769 }
1770 
1771 static int set_ucontext_resp(struct ib_ucontext *uctx,
1772 			     struct mlx5_ib_alloc_ucontext_resp *resp)
1773 {
1774 	struct ib_device *ibdev = uctx->device;
1775 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1776 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1777 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1778 
1779 	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1780 		resp->dump_fill_mkey = dev->mkeys.dump_fill_mkey;
1781 		resp->comp_mask |=
1782 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1783 	}
1784 
1785 	resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1786 	if (dev->wc_support)
1787 		resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev,
1788 						      log_bf_reg_size);
1789 	resp->cache_line_size = cache_line_size();
1790 	resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1791 	resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1792 	resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1793 	resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1794 	resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1795 	resp->cqe_version = context->cqe_version;
1796 	resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1797 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1798 	resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1799 					MLX5_CAP_GEN(dev->mdev,
1800 						     num_of_uars_per_page) : 1;
1801 	resp->tot_bfregs = bfregi->lib_uar_dyn ? 0 :
1802 			bfregi->total_num_bfregs - bfregi->num_dyn_bfregs;
1803 	resp->num_ports = dev->num_ports;
1804 	resp->cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1805 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1806 
1807 	if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1808 		mlx5_query_min_inline(dev->mdev, &resp->eth_min_inline);
1809 		resp->eth_min_inline++;
1810 	}
1811 
1812 	if (dev->mdev->clock_info)
1813 		resp->clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1814 
1815 	/*
1816 	 * We don't want to expose information from the PCI bar that is located
1817 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1818 	 * pretend we don't support reading the HCA's core clock. This is also
1819 	 * forced by mmap function.
1820 	 */
1821 	if (PAGE_SIZE <= 4096) {
1822 		resp->comp_mask |=
1823 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1824 		resp->hca_core_clock_offset =
1825 			offsetof(struct mlx5_init_seg,
1826 				 internal_timer_h) % PAGE_SIZE;
1827 	}
1828 
1829 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
1830 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE;
1831 
1832 	if (rt_supported(MLX5_CAP_GEN(dev->mdev, sq_ts_format)) &&
1833 	    rt_supported(MLX5_CAP_GEN(dev->mdev, rq_ts_format)) &&
1834 	    rt_supported(MLX5_CAP_ROCE(dev->mdev, qp_ts_format)))
1835 		resp->comp_mask |=
1836 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_REAL_TIME_TS;
1837 
1838 	resp->num_dyn_bfregs = bfregi->num_dyn_bfregs;
1839 
1840 	if (MLX5_CAP_GEN(dev->mdev, drain_sigerr))
1841 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_SQD2RTS;
1842 
1843 	resp->comp_mask |=
1844 		MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_MKEY_UPDATE_TAG;
1845 
1846 	return 0;
1847 }
1848 
1849 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1850 				  struct ib_udata *udata)
1851 {
1852 	struct ib_device *ibdev = uctx->device;
1853 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1854 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1855 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1856 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1857 	struct mlx5_bfreg_info *bfregi;
1858 	int ver;
1859 	int err;
1860 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1861 				     max_cqe_version);
1862 	bool lib_uar_4k;
1863 	bool lib_uar_dyn;
1864 
1865 	if (!dev->ib_active)
1866 		return -EAGAIN;
1867 
1868 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1869 		ver = 0;
1870 	else if (udata->inlen >= min_req_v2)
1871 		ver = 2;
1872 	else
1873 		return -EINVAL;
1874 
1875 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1876 	if (err)
1877 		return err;
1878 
1879 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1880 		return -EOPNOTSUPP;
1881 
1882 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1883 		return -EOPNOTSUPP;
1884 
1885 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1886 				    MLX5_NON_FP_BFREGS_PER_UAR);
1887 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1888 		return -EINVAL;
1889 
1890 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1891 		err = mlx5_ib_devx_create(dev, true);
1892 		if (err < 0)
1893 			goto out_ctx;
1894 		context->devx_uid = err;
1895 	}
1896 
1897 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1898 	lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
1899 	bfregi = &context->bfregi;
1900 
1901 	if (lib_uar_dyn) {
1902 		bfregi->lib_uar_dyn = lib_uar_dyn;
1903 		goto uar_done;
1904 	}
1905 
1906 	/* updates req->total_num_bfregs */
1907 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1908 	if (err)
1909 		goto out_devx;
1910 
1911 	mutex_init(&bfregi->lock);
1912 	bfregi->lib_uar_4k = lib_uar_4k;
1913 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1914 				GFP_KERNEL);
1915 	if (!bfregi->count) {
1916 		err = -ENOMEM;
1917 		goto out_devx;
1918 	}
1919 
1920 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1921 				    sizeof(*bfregi->sys_pages),
1922 				    GFP_KERNEL);
1923 	if (!bfregi->sys_pages) {
1924 		err = -ENOMEM;
1925 		goto out_count;
1926 	}
1927 
1928 	err = allocate_uars(dev, context);
1929 	if (err)
1930 		goto out_sys_pages;
1931 
1932 uar_done:
1933 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1934 					     context->devx_uid);
1935 	if (err)
1936 		goto out_uars;
1937 
1938 	INIT_LIST_HEAD(&context->db_page_list);
1939 	mutex_init(&context->db_page_mutex);
1940 
1941 	context->cqe_version = min_t(__u8,
1942 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1943 				 req.max_cqe_version);
1944 
1945 	err = set_ucontext_resp(uctx, &resp);
1946 	if (err)
1947 		goto out_mdev;
1948 
1949 	resp.response_length = min(udata->outlen, sizeof(resp));
1950 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1951 	if (err)
1952 		goto out_mdev;
1953 
1954 	bfregi->ver = ver;
1955 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1956 	context->lib_caps = req.lib_caps;
1957 	print_lib_caps(dev, context->lib_caps);
1958 
1959 	if (mlx5_ib_lag_should_assign_affinity(dev)) {
1960 		u32 port = mlx5_core_native_port_num(dev->mdev) - 1;
1961 
1962 		atomic_set(&context->tx_port_affinity,
1963 			   atomic_add_return(
1964 				   1, &dev->port[port].roce.tx_port_affinity));
1965 	}
1966 
1967 	return 0;
1968 
1969 out_mdev:
1970 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1971 
1972 out_uars:
1973 	deallocate_uars(dev, context);
1974 
1975 out_sys_pages:
1976 	kfree(bfregi->sys_pages);
1977 
1978 out_count:
1979 	kfree(bfregi->count);
1980 
1981 out_devx:
1982 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1983 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1984 
1985 out_ctx:
1986 	return err;
1987 }
1988 
1989 static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext,
1990 				  struct uverbs_attr_bundle *attrs)
1991 {
1992 	struct mlx5_ib_alloc_ucontext_resp uctx_resp = {};
1993 	int ret;
1994 
1995 	ret = set_ucontext_resp(ibcontext, &uctx_resp);
1996 	if (ret)
1997 		return ret;
1998 
1999 	uctx_resp.response_length =
2000 		min_t(size_t,
2001 		      uverbs_attr_get_len(attrs,
2002 				MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX),
2003 		      sizeof(uctx_resp));
2004 
2005 	ret = uverbs_copy_to_struct_or_zero(attrs,
2006 					MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
2007 					&uctx_resp,
2008 					sizeof(uctx_resp));
2009 	return ret;
2010 }
2011 
2012 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
2013 {
2014 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2015 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2016 	struct mlx5_bfreg_info *bfregi;
2017 
2018 	bfregi = &context->bfregi;
2019 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
2020 
2021 	deallocate_uars(dev, context);
2022 	kfree(bfregi->sys_pages);
2023 	kfree(bfregi->count);
2024 
2025 	if (context->devx_uid)
2026 		mlx5_ib_devx_destroy(dev, context->devx_uid);
2027 }
2028 
2029 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2030 				 int uar_idx)
2031 {
2032 	int fw_uars_per_page;
2033 
2034 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2035 
2036 	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2037 }
2038 
2039 static u64 uar_index2paddress(struct mlx5_ib_dev *dev,
2040 				 int uar_idx)
2041 {
2042 	unsigned int fw_uars_per_page;
2043 
2044 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
2045 				MLX5_UARS_IN_PAGE : 1;
2046 
2047 	return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE);
2048 }
2049 
2050 static int get_command(unsigned long offset)
2051 {
2052 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2053 }
2054 
2055 static int get_arg(unsigned long offset)
2056 {
2057 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2058 }
2059 
2060 static int get_index(unsigned long offset)
2061 {
2062 	return get_arg(offset);
2063 }
2064 
2065 /* Index resides in an extra byte to enable larger values than 255 */
2066 static int get_extended_index(unsigned long offset)
2067 {
2068 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2069 }
2070 
2071 
2072 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2073 {
2074 }
2075 
2076 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2077 {
2078 	switch (cmd) {
2079 	case MLX5_IB_MMAP_WC_PAGE:
2080 		return "WC";
2081 	case MLX5_IB_MMAP_REGULAR_PAGE:
2082 		return "best effort WC";
2083 	case MLX5_IB_MMAP_NC_PAGE:
2084 		return "NC";
2085 	case MLX5_IB_MMAP_DEVICE_MEM:
2086 		return "Device Memory";
2087 	default:
2088 		return "Unknown";
2089 	}
2090 }
2091 
2092 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2093 					struct vm_area_struct *vma,
2094 					struct mlx5_ib_ucontext *context)
2095 {
2096 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2097 	    !(vma->vm_flags & VM_SHARED))
2098 		return -EINVAL;
2099 
2100 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2101 		return -EOPNOTSUPP;
2102 
2103 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2104 		return -EPERM;
2105 	vm_flags_clear(vma, VM_MAYWRITE);
2106 
2107 	if (!dev->mdev->clock_info)
2108 		return -EOPNOTSUPP;
2109 
2110 	return vm_insert_page(vma, vma->vm_start,
2111 			      virt_to_page(dev->mdev->clock_info));
2112 }
2113 
2114 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2115 {
2116 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2117 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2118 	struct mlx5_var_table *var_table = &dev->var_table;
2119 	struct mlx5_ib_ucontext *context = to_mucontext(entry->ucontext);
2120 
2121 	switch (mentry->mmap_flag) {
2122 	case MLX5_IB_MMAP_TYPE_MEMIC:
2123 	case MLX5_IB_MMAP_TYPE_MEMIC_OP:
2124 		mlx5_ib_dm_mmap_free(dev, mentry);
2125 		break;
2126 	case MLX5_IB_MMAP_TYPE_VAR:
2127 		mutex_lock(&var_table->bitmap_lock);
2128 		clear_bit(mentry->page_idx, var_table->bitmap);
2129 		mutex_unlock(&var_table->bitmap_lock);
2130 		kfree(mentry);
2131 		break;
2132 	case MLX5_IB_MMAP_TYPE_UAR_WC:
2133 	case MLX5_IB_MMAP_TYPE_UAR_NC:
2134 		mlx5_cmd_uar_dealloc(dev->mdev, mentry->page_idx,
2135 				     context->devx_uid);
2136 		kfree(mentry);
2137 		break;
2138 	default:
2139 		WARN_ON(true);
2140 	}
2141 }
2142 
2143 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2144 		    struct vm_area_struct *vma,
2145 		    struct mlx5_ib_ucontext *context)
2146 {
2147 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2148 	int err;
2149 	unsigned long idx;
2150 	phys_addr_t pfn;
2151 	pgprot_t prot;
2152 	u32 bfreg_dyn_idx = 0;
2153 	u32 uar_index;
2154 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2155 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2156 				bfregi->num_static_sys_pages;
2157 
2158 	if (bfregi->lib_uar_dyn)
2159 		return -EINVAL;
2160 
2161 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2162 		return -EINVAL;
2163 
2164 	if (dyn_uar)
2165 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2166 	else
2167 		idx = get_index(vma->vm_pgoff);
2168 
2169 	if (idx >= max_valid_idx) {
2170 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2171 			     idx, max_valid_idx);
2172 		return -EINVAL;
2173 	}
2174 
2175 	switch (cmd) {
2176 	case MLX5_IB_MMAP_WC_PAGE:
2177 	case MLX5_IB_MMAP_ALLOC_WC:
2178 	case MLX5_IB_MMAP_REGULAR_PAGE:
2179 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2180 		prot = pgprot_writecombine(vma->vm_page_prot);
2181 		break;
2182 	case MLX5_IB_MMAP_NC_PAGE:
2183 		prot = pgprot_noncached(vma->vm_page_prot);
2184 		break;
2185 	default:
2186 		return -EINVAL;
2187 	}
2188 
2189 	if (dyn_uar) {
2190 		int uars_per_page;
2191 
2192 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2193 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2194 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2195 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2196 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2197 			return -EINVAL;
2198 		}
2199 
2200 		mutex_lock(&bfregi->lock);
2201 		/* Fail if uar already allocated, first bfreg index of each
2202 		 * page holds its count.
2203 		 */
2204 		if (bfregi->count[bfreg_dyn_idx]) {
2205 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2206 			mutex_unlock(&bfregi->lock);
2207 			return -EINVAL;
2208 		}
2209 
2210 		bfregi->count[bfreg_dyn_idx]++;
2211 		mutex_unlock(&bfregi->lock);
2212 
2213 		err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index,
2214 					 context->devx_uid);
2215 		if (err) {
2216 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2217 			goto free_bfreg;
2218 		}
2219 	} else {
2220 		uar_index = bfregi->sys_pages[idx];
2221 	}
2222 
2223 	pfn = uar_index2pfn(dev, uar_index);
2224 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2225 
2226 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2227 				prot, NULL);
2228 	if (err) {
2229 		mlx5_ib_err(dev,
2230 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2231 			    err, mmap_cmd2str(cmd));
2232 		goto err;
2233 	}
2234 
2235 	if (dyn_uar)
2236 		bfregi->sys_pages[idx] = uar_index;
2237 	return 0;
2238 
2239 err:
2240 	if (!dyn_uar)
2241 		return err;
2242 
2243 	mlx5_cmd_uar_dealloc(dev->mdev, idx, context->devx_uid);
2244 
2245 free_bfreg:
2246 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2247 
2248 	return err;
2249 }
2250 
2251 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2252 {
2253 	unsigned long idx;
2254 	u8 command;
2255 
2256 	command = get_command(vma->vm_pgoff);
2257 	idx = get_extended_index(vma->vm_pgoff);
2258 
2259 	return (command << 16 | idx);
2260 }
2261 
2262 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2263 			       struct vm_area_struct *vma,
2264 			       struct ib_ucontext *ucontext)
2265 {
2266 	struct mlx5_user_mmap_entry *mentry;
2267 	struct rdma_user_mmap_entry *entry;
2268 	unsigned long pgoff;
2269 	pgprot_t prot;
2270 	phys_addr_t pfn;
2271 	int ret;
2272 
2273 	pgoff = mlx5_vma_to_pgoff(vma);
2274 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2275 	if (!entry)
2276 		return -EINVAL;
2277 
2278 	mentry = to_mmmap(entry);
2279 	pfn = (mentry->address >> PAGE_SHIFT);
2280 	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
2281 	    mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
2282 		prot = pgprot_noncached(vma->vm_page_prot);
2283 	else
2284 		prot = pgprot_writecombine(vma->vm_page_prot);
2285 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2286 				entry->npages * PAGE_SIZE,
2287 				prot,
2288 				entry);
2289 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2290 	return ret;
2291 }
2292 
2293 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2294 {
2295 	u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2296 	u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2297 
2298 	return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2299 		(index & 0xFF)) << PAGE_SHIFT;
2300 }
2301 
2302 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2303 {
2304 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2305 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2306 	unsigned long command;
2307 	phys_addr_t pfn;
2308 
2309 	command = get_command(vma->vm_pgoff);
2310 	switch (command) {
2311 	case MLX5_IB_MMAP_WC_PAGE:
2312 	case MLX5_IB_MMAP_ALLOC_WC:
2313 		if (!dev->wc_support)
2314 			return -EPERM;
2315 		fallthrough;
2316 	case MLX5_IB_MMAP_NC_PAGE:
2317 	case MLX5_IB_MMAP_REGULAR_PAGE:
2318 		return uar_mmap(dev, command, vma, context);
2319 
2320 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2321 		return -ENOSYS;
2322 
2323 	case MLX5_IB_MMAP_CORE_CLOCK:
2324 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2325 			return -EINVAL;
2326 
2327 		if (vma->vm_flags & VM_WRITE)
2328 			return -EPERM;
2329 		vm_flags_clear(vma, VM_MAYWRITE);
2330 
2331 		/* Don't expose to user-space information it shouldn't have */
2332 		if (PAGE_SIZE > 4096)
2333 			return -EOPNOTSUPP;
2334 
2335 		pfn = (dev->mdev->iseg_base +
2336 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2337 			PAGE_SHIFT;
2338 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2339 					 PAGE_SIZE,
2340 					 pgprot_noncached(vma->vm_page_prot),
2341 					 NULL);
2342 	case MLX5_IB_MMAP_CLOCK_INFO:
2343 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2344 
2345 	default:
2346 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2347 	}
2348 
2349 	return 0;
2350 }
2351 
2352 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2353 {
2354 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2355 	struct ib_device *ibdev = ibpd->device;
2356 	struct mlx5_ib_alloc_pd_resp resp;
2357 	int err;
2358 	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2359 	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {};
2360 	u16 uid = 0;
2361 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2362 		udata, struct mlx5_ib_ucontext, ibucontext);
2363 
2364 	uid = context ? context->devx_uid : 0;
2365 	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2366 	MLX5_SET(alloc_pd_in, in, uid, uid);
2367 	err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out);
2368 	if (err)
2369 		return err;
2370 
2371 	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2372 	pd->uid = uid;
2373 	if (udata) {
2374 		resp.pdn = pd->pdn;
2375 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2376 			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2377 			return -EFAULT;
2378 		}
2379 	}
2380 
2381 	return 0;
2382 }
2383 
2384 static int mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2385 {
2386 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2387 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2388 
2389 	return mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2390 }
2391 
2392 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2393 {
2394 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2395 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
2396 	int err;
2397 	u16 uid;
2398 
2399 	uid = ibqp->pd ?
2400 		to_mpd(ibqp->pd)->uid : 0;
2401 
2402 	if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) {
2403 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
2404 		return -EOPNOTSUPP;
2405 	}
2406 
2407 	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2408 	if (err)
2409 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2410 			     ibqp->qp_num, gid->raw);
2411 
2412 	return err;
2413 }
2414 
2415 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2416 {
2417 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2418 	int err;
2419 	u16 uid;
2420 
2421 	uid = ibqp->pd ?
2422 		to_mpd(ibqp->pd)->uid : 0;
2423 	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2424 	if (err)
2425 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2426 			     ibqp->qp_num, gid->raw);
2427 
2428 	return err;
2429 }
2430 
2431 static int init_node_data(struct mlx5_ib_dev *dev)
2432 {
2433 	int err;
2434 
2435 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2436 	if (err)
2437 		return err;
2438 
2439 	dev->mdev->rev_id = dev->mdev->pdev->revision;
2440 
2441 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2442 }
2443 
2444 static ssize_t fw_pages_show(struct device *device,
2445 			     struct device_attribute *attr, char *buf)
2446 {
2447 	struct mlx5_ib_dev *dev =
2448 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2449 
2450 	return sysfs_emit(buf, "%d\n", dev->mdev->priv.fw_pages);
2451 }
2452 static DEVICE_ATTR_RO(fw_pages);
2453 
2454 static ssize_t reg_pages_show(struct device *device,
2455 			      struct device_attribute *attr, char *buf)
2456 {
2457 	struct mlx5_ib_dev *dev =
2458 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2459 
2460 	return sysfs_emit(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2461 }
2462 static DEVICE_ATTR_RO(reg_pages);
2463 
2464 static ssize_t hca_type_show(struct device *device,
2465 			     struct device_attribute *attr, char *buf)
2466 {
2467 	struct mlx5_ib_dev *dev =
2468 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2469 
2470 	return sysfs_emit(buf, "MT%d\n", dev->mdev->pdev->device);
2471 }
2472 static DEVICE_ATTR_RO(hca_type);
2473 
2474 static ssize_t hw_rev_show(struct device *device,
2475 			   struct device_attribute *attr, char *buf)
2476 {
2477 	struct mlx5_ib_dev *dev =
2478 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2479 
2480 	return sysfs_emit(buf, "%x\n", dev->mdev->rev_id);
2481 }
2482 static DEVICE_ATTR_RO(hw_rev);
2483 
2484 static ssize_t board_id_show(struct device *device,
2485 			     struct device_attribute *attr, char *buf)
2486 {
2487 	struct mlx5_ib_dev *dev =
2488 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2489 
2490 	return sysfs_emit(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2491 			  dev->mdev->board_id);
2492 }
2493 static DEVICE_ATTR_RO(board_id);
2494 
2495 static struct attribute *mlx5_class_attributes[] = {
2496 	&dev_attr_hw_rev.attr,
2497 	&dev_attr_hca_type.attr,
2498 	&dev_attr_board_id.attr,
2499 	&dev_attr_fw_pages.attr,
2500 	&dev_attr_reg_pages.attr,
2501 	NULL,
2502 };
2503 
2504 static const struct attribute_group mlx5_attr_group = {
2505 	.attrs = mlx5_class_attributes,
2506 };
2507 
2508 static void pkey_change_handler(struct work_struct *work)
2509 {
2510 	struct mlx5_ib_port_resources *ports =
2511 		container_of(work, struct mlx5_ib_port_resources,
2512 			     pkey_change_work);
2513 
2514 	if (!ports->gsi)
2515 		/*
2516 		 * We got this event before device was fully configured
2517 		 * and MAD registration code wasn't called/finished yet.
2518 		 */
2519 		return;
2520 
2521 	mlx5_ib_gsi_pkey_change(ports->gsi);
2522 }
2523 
2524 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2525 {
2526 	struct mlx5_ib_qp *mqp;
2527 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2528 	struct mlx5_core_cq *mcq;
2529 	struct list_head cq_armed_list;
2530 	unsigned long flags_qp;
2531 	unsigned long flags_cq;
2532 	unsigned long flags;
2533 
2534 	INIT_LIST_HEAD(&cq_armed_list);
2535 
2536 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2537 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2538 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2539 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2540 		if (mqp->sq.tail != mqp->sq.head) {
2541 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2542 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2543 			if (send_mcq->mcq.comp &&
2544 			    mqp->ibqp.send_cq->comp_handler) {
2545 				if (!send_mcq->mcq.reset_notify_added) {
2546 					send_mcq->mcq.reset_notify_added = 1;
2547 					list_add_tail(&send_mcq->mcq.reset_notify,
2548 						      &cq_armed_list);
2549 				}
2550 			}
2551 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2552 		}
2553 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2554 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2555 		/* no handling is needed for SRQ */
2556 		if (!mqp->ibqp.srq) {
2557 			if (mqp->rq.tail != mqp->rq.head) {
2558 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2559 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2560 				if (recv_mcq->mcq.comp &&
2561 				    mqp->ibqp.recv_cq->comp_handler) {
2562 					if (!recv_mcq->mcq.reset_notify_added) {
2563 						recv_mcq->mcq.reset_notify_added = 1;
2564 						list_add_tail(&recv_mcq->mcq.reset_notify,
2565 							      &cq_armed_list);
2566 					}
2567 				}
2568 				spin_unlock_irqrestore(&recv_mcq->lock,
2569 						       flags_cq);
2570 			}
2571 		}
2572 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2573 	}
2574 	/*At that point all inflight post send were put to be executed as of we
2575 	 * lock/unlock above locks Now need to arm all involved CQs.
2576 	 */
2577 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2578 		mcq->comp(mcq, NULL);
2579 	}
2580 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2581 }
2582 
2583 static void delay_drop_handler(struct work_struct *work)
2584 {
2585 	int err;
2586 	struct mlx5_ib_delay_drop *delay_drop =
2587 		container_of(work, struct mlx5_ib_delay_drop,
2588 			     delay_drop_work);
2589 
2590 	atomic_inc(&delay_drop->events_cnt);
2591 
2592 	mutex_lock(&delay_drop->lock);
2593 	err = mlx5_core_set_delay_drop(delay_drop->dev, delay_drop->timeout);
2594 	if (err) {
2595 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
2596 			     delay_drop->timeout);
2597 		delay_drop->activate = false;
2598 	}
2599 	mutex_unlock(&delay_drop->lock);
2600 }
2601 
2602 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2603 				 struct ib_event *ibev)
2604 {
2605 	u32 port = (eqe->data.port.port >> 4) & 0xf;
2606 
2607 	switch (eqe->sub_type) {
2608 	case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
2609 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2610 					    IB_LINK_LAYER_ETHERNET)
2611 			schedule_work(&ibdev->delay_drop.delay_drop_work);
2612 		break;
2613 	default: /* do nothing */
2614 		return;
2615 	}
2616 }
2617 
2618 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2619 			      struct ib_event *ibev)
2620 {
2621 	u32 port = (eqe->data.port.port >> 4) & 0xf;
2622 
2623 	ibev->element.port_num = port;
2624 
2625 	switch (eqe->sub_type) {
2626 	case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
2627 	case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
2628 	case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
2629 		/* In RoCE, port up/down events are handled in
2630 		 * mlx5_netdev_event().
2631 		 */
2632 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2633 					    IB_LINK_LAYER_ETHERNET)
2634 			return -EINVAL;
2635 
2636 		ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
2637 				IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2638 		break;
2639 
2640 	case MLX5_PORT_CHANGE_SUBTYPE_LID:
2641 		ibev->event = IB_EVENT_LID_CHANGE;
2642 		break;
2643 
2644 	case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
2645 		ibev->event = IB_EVENT_PKEY_CHANGE;
2646 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2647 		break;
2648 
2649 	case MLX5_PORT_CHANGE_SUBTYPE_GUID:
2650 		ibev->event = IB_EVENT_GID_CHANGE;
2651 		break;
2652 
2653 	case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
2654 		ibev->event = IB_EVENT_CLIENT_REREGISTER;
2655 		break;
2656 	default:
2657 		return -EINVAL;
2658 	}
2659 
2660 	return 0;
2661 }
2662 
2663 static void mlx5_ib_handle_event(struct work_struct *_work)
2664 {
2665 	struct mlx5_ib_event_work *work =
2666 		container_of(_work, struct mlx5_ib_event_work, work);
2667 	struct mlx5_ib_dev *ibdev;
2668 	struct ib_event ibev;
2669 	bool fatal = false;
2670 
2671 	if (work->is_slave) {
2672 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
2673 		if (!ibdev)
2674 			goto out;
2675 	} else {
2676 		ibdev = work->dev;
2677 	}
2678 
2679 	switch (work->event) {
2680 	case MLX5_DEV_EVENT_SYS_ERROR:
2681 		ibev.event = IB_EVENT_DEVICE_FATAL;
2682 		mlx5_ib_handle_internal_error(ibdev);
2683 		ibev.element.port_num  = (u8)(unsigned long)work->param;
2684 		fatal = true;
2685 		break;
2686 	case MLX5_EVENT_TYPE_PORT_CHANGE:
2687 		if (handle_port_change(ibdev, work->param, &ibev))
2688 			goto out;
2689 		break;
2690 	case MLX5_EVENT_TYPE_GENERAL_EVENT:
2691 		handle_general_event(ibdev, work->param, &ibev);
2692 		fallthrough;
2693 	default:
2694 		goto out;
2695 	}
2696 
2697 	ibev.device = &ibdev->ib_dev;
2698 
2699 	if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
2700 		mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
2701 		goto out;
2702 	}
2703 
2704 	if (ibdev->ib_active)
2705 		ib_dispatch_event(&ibev);
2706 
2707 	if (fatal)
2708 		ibdev->ib_active = false;
2709 out:
2710 	kfree(work);
2711 }
2712 
2713 static int mlx5_ib_event(struct notifier_block *nb,
2714 			 unsigned long event, void *param)
2715 {
2716 	struct mlx5_ib_event_work *work;
2717 
2718 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2719 	if (!work)
2720 		return NOTIFY_DONE;
2721 
2722 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2723 	work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
2724 	work->is_slave = false;
2725 	work->param = param;
2726 	work->event = event;
2727 
2728 	queue_work(mlx5_ib_event_wq, &work->work);
2729 
2730 	return NOTIFY_OK;
2731 }
2732 
2733 static int mlx5_ib_event_slave_port(struct notifier_block *nb,
2734 				    unsigned long event, void *param)
2735 {
2736 	struct mlx5_ib_event_work *work;
2737 
2738 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2739 	if (!work)
2740 		return NOTIFY_DONE;
2741 
2742 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2743 	work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
2744 	work->is_slave = true;
2745 	work->param = param;
2746 	work->event = event;
2747 	queue_work(mlx5_ib_event_wq, &work->work);
2748 
2749 	return NOTIFY_OK;
2750 }
2751 
2752 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2753 {
2754 	struct mlx5_hca_vport_context vport_ctx;
2755 	int err;
2756 	int port;
2757 
2758 	if (MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_IB)
2759 		return 0;
2760 
2761 	for (port = 1; port <= dev->num_ports; port++) {
2762 		if (!MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2763 			dev->port_caps[port - 1].has_smi = true;
2764 			continue;
2765 		}
2766 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port, 0,
2767 						   &vport_ctx);
2768 		if (err) {
2769 			mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2770 				    port, err);
2771 			return err;
2772 		}
2773 		dev->port_caps[port - 1].has_smi = vport_ctx.has_smi;
2774 	}
2775 
2776 	return 0;
2777 }
2778 
2779 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2780 {
2781 	unsigned int port;
2782 
2783 	rdma_for_each_port (&dev->ib_dev, port)
2784 		mlx5_query_ext_port_caps(dev, port);
2785 }
2786 
2787 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
2788 {
2789 	switch (umr_fence_cap) {
2790 	case MLX5_CAP_UMR_FENCE_NONE:
2791 		return MLX5_FENCE_MODE_NONE;
2792 	case MLX5_CAP_UMR_FENCE_SMALL:
2793 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
2794 	default:
2795 		return MLX5_FENCE_MODE_STRONG_ORDERING;
2796 	}
2797 }
2798 
2799 static int mlx5_ib_dev_res_init(struct mlx5_ib_dev *dev)
2800 {
2801 	struct mlx5_ib_resources *devr = &dev->devr;
2802 	struct ib_srq_init_attr attr;
2803 	struct ib_device *ibdev;
2804 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2805 	int port;
2806 	int ret = 0;
2807 
2808 	ibdev = &dev->ib_dev;
2809 
2810 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
2811 		return -EOPNOTSUPP;
2812 
2813 	devr->p0 = ib_alloc_pd(ibdev, 0);
2814 	if (IS_ERR(devr->p0))
2815 		return PTR_ERR(devr->p0);
2816 
2817 	devr->c0 = ib_create_cq(ibdev, NULL, NULL, NULL, &cq_attr);
2818 	if (IS_ERR(devr->c0)) {
2819 		ret = PTR_ERR(devr->c0);
2820 		goto error1;
2821 	}
2822 
2823 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0);
2824 	if (ret)
2825 		goto error2;
2826 
2827 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0);
2828 	if (ret)
2829 		goto error3;
2830 
2831 	memset(&attr, 0, sizeof(attr));
2832 	attr.attr.max_sge = 1;
2833 	attr.attr.max_wr = 1;
2834 	attr.srq_type = IB_SRQT_XRC;
2835 	attr.ext.cq = devr->c0;
2836 
2837 	devr->s0 = ib_create_srq(devr->p0, &attr);
2838 	if (IS_ERR(devr->s0)) {
2839 		ret = PTR_ERR(devr->s0);
2840 		goto err_create;
2841 	}
2842 
2843 	memset(&attr, 0, sizeof(attr));
2844 	attr.attr.max_sge = 1;
2845 	attr.attr.max_wr = 1;
2846 	attr.srq_type = IB_SRQT_BASIC;
2847 
2848 	devr->s1 = ib_create_srq(devr->p0, &attr);
2849 	if (IS_ERR(devr->s1)) {
2850 		ret = PTR_ERR(devr->s1);
2851 		goto error6;
2852 	}
2853 
2854 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2855 		INIT_WORK(&devr->ports[port].pkey_change_work,
2856 			  pkey_change_handler);
2857 
2858 	return 0;
2859 
2860 error6:
2861 	ib_destroy_srq(devr->s0);
2862 err_create:
2863 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2864 error3:
2865 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2866 error2:
2867 	ib_destroy_cq(devr->c0);
2868 error1:
2869 	ib_dealloc_pd(devr->p0);
2870 	return ret;
2871 }
2872 
2873 static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev)
2874 {
2875 	struct mlx5_ib_resources *devr = &dev->devr;
2876 	int port;
2877 
2878 	/*
2879 	 * Make sure no change P_Key work items are still executing.
2880 	 *
2881 	 * At this stage, the mlx5_ib_event should be unregistered
2882 	 * and it ensures that no new works are added.
2883 	 */
2884 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2885 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2886 
2887 	ib_destroy_srq(devr->s1);
2888 	ib_destroy_srq(devr->s0);
2889 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2890 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2891 	ib_destroy_cq(devr->c0);
2892 	ib_dealloc_pd(devr->p0);
2893 }
2894 
2895 static u32 get_core_cap_flags(struct ib_device *ibdev,
2896 			      struct mlx5_hca_vport_context *rep)
2897 {
2898 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2899 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2900 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2901 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2902 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
2903 	u32 ret = 0;
2904 
2905 	if (rep->grh_required)
2906 		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
2907 
2908 	if (ll == IB_LINK_LAYER_INFINIBAND)
2909 		return ret | RDMA_CORE_PORT_IBA_IB;
2910 
2911 	if (raw_support)
2912 		ret |= RDMA_CORE_PORT_RAW_PACKET;
2913 
2914 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2915 		return ret;
2916 
2917 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2918 		return ret;
2919 
2920 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2921 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2922 
2923 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2924 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2925 
2926 	return ret;
2927 }
2928 
2929 static int mlx5_port_immutable(struct ib_device *ibdev, u32 port_num,
2930 			       struct ib_port_immutable *immutable)
2931 {
2932 	struct ib_port_attr attr;
2933 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2934 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
2935 	struct mlx5_hca_vport_context rep = {0};
2936 	int err;
2937 
2938 	err = ib_query_port(ibdev, port_num, &attr);
2939 	if (err)
2940 		return err;
2941 
2942 	if (ll == IB_LINK_LAYER_INFINIBAND) {
2943 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
2944 						   &rep);
2945 		if (err)
2946 			return err;
2947 	}
2948 
2949 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2950 	immutable->gid_tbl_len = attr.gid_tbl_len;
2951 	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
2952 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2953 
2954 	return 0;
2955 }
2956 
2957 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u32 port_num,
2958 				   struct ib_port_immutable *immutable)
2959 {
2960 	struct ib_port_attr attr;
2961 	int err;
2962 
2963 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2964 
2965 	err = ib_query_port(ibdev, port_num, &attr);
2966 	if (err)
2967 		return err;
2968 
2969 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2970 	immutable->gid_tbl_len = attr.gid_tbl_len;
2971 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2972 
2973 	return 0;
2974 }
2975 
2976 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
2977 {
2978 	struct mlx5_ib_dev *dev =
2979 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2980 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
2981 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
2982 		 fw_rev_sub(dev->mdev));
2983 }
2984 
2985 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
2986 {
2987 	struct mlx5_core_dev *mdev = dev->mdev;
2988 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
2989 								 MLX5_FLOW_NAMESPACE_LAG);
2990 	struct mlx5_flow_table *ft;
2991 	int err;
2992 
2993 	if (!ns || !mlx5_lag_is_active(mdev))
2994 		return 0;
2995 
2996 	err = mlx5_cmd_create_vport_lag(mdev);
2997 	if (err)
2998 		return err;
2999 
3000 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3001 	if (IS_ERR(ft)) {
3002 		err = PTR_ERR(ft);
3003 		goto err_destroy_vport_lag;
3004 	}
3005 
3006 	dev->flow_db->lag_demux_ft = ft;
3007 	dev->lag_ports = mlx5_lag_get_num_ports(mdev);
3008 	dev->lag_active = true;
3009 	return 0;
3010 
3011 err_destroy_vport_lag:
3012 	mlx5_cmd_destroy_vport_lag(mdev);
3013 	return err;
3014 }
3015 
3016 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3017 {
3018 	struct mlx5_core_dev *mdev = dev->mdev;
3019 
3020 	if (dev->lag_active) {
3021 		dev->lag_active = false;
3022 
3023 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
3024 		dev->flow_db->lag_demux_ft = NULL;
3025 
3026 		mlx5_cmd_destroy_vport_lag(mdev);
3027 	}
3028 }
3029 
3030 static void mlx5_netdev_notifier_register(struct mlx5_roce *roce,
3031 					  struct net_device *netdev)
3032 {
3033 	int err;
3034 
3035 	if (roce->tracking_netdev)
3036 		return;
3037 	roce->tracking_netdev = netdev;
3038 	roce->nb.notifier_call = mlx5_netdev_event;
3039 	err = register_netdevice_notifier_dev_net(netdev, &roce->nb, &roce->nn);
3040 	WARN_ON(err);
3041 }
3042 
3043 static void mlx5_netdev_notifier_unregister(struct mlx5_roce *roce)
3044 {
3045 	if (!roce->tracking_netdev)
3046 		return;
3047 	unregister_netdevice_notifier_dev_net(roce->tracking_netdev, &roce->nb,
3048 					      &roce->nn);
3049 	roce->tracking_netdev = NULL;
3050 }
3051 
3052 static int mlx5e_mdev_notifier_event(struct notifier_block *nb,
3053 				     unsigned long event, void *data)
3054 {
3055 	struct mlx5_roce *roce = container_of(nb, struct mlx5_roce, mdev_nb);
3056 	struct net_device *netdev = data;
3057 
3058 	switch (event) {
3059 	case MLX5_DRIVER_EVENT_UPLINK_NETDEV:
3060 		if (netdev)
3061 			mlx5_netdev_notifier_register(roce, netdev);
3062 		else
3063 			mlx5_netdev_notifier_unregister(roce);
3064 		break;
3065 	default:
3066 		return NOTIFY_DONE;
3067 	}
3068 
3069 	return NOTIFY_OK;
3070 }
3071 
3072 static void mlx5_mdev_netdev_track(struct mlx5_ib_dev *dev, u32 port_num)
3073 {
3074 	struct mlx5_roce *roce = &dev->port[port_num].roce;
3075 
3076 	roce->mdev_nb.notifier_call = mlx5e_mdev_notifier_event;
3077 	mlx5_blocking_notifier_register(dev->mdev, &roce->mdev_nb);
3078 	mlx5_core_uplink_netdev_event_replay(dev->mdev);
3079 }
3080 
3081 static void mlx5_mdev_netdev_untrack(struct mlx5_ib_dev *dev, u32 port_num)
3082 {
3083 	struct mlx5_roce *roce = &dev->port[port_num].roce;
3084 
3085 	mlx5_blocking_notifier_unregister(dev->mdev, &roce->mdev_nb);
3086 	mlx5_netdev_notifier_unregister(roce);
3087 }
3088 
3089 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3090 {
3091 	int err;
3092 
3093 	if (!dev->is_rep && dev->profile != &raw_eth_profile) {
3094 		err = mlx5_nic_vport_enable_roce(dev->mdev);
3095 		if (err)
3096 			return err;
3097 	}
3098 
3099 	err = mlx5_eth_lag_init(dev);
3100 	if (err)
3101 		goto err_disable_roce;
3102 
3103 	return 0;
3104 
3105 err_disable_roce:
3106 	if (!dev->is_rep && dev->profile != &raw_eth_profile)
3107 		mlx5_nic_vport_disable_roce(dev->mdev);
3108 
3109 	return err;
3110 }
3111 
3112 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3113 {
3114 	mlx5_eth_lag_cleanup(dev);
3115 	if (!dev->is_rep && dev->profile != &raw_eth_profile)
3116 		mlx5_nic_vport_disable_roce(dev->mdev);
3117 }
3118 
3119 static int mlx5_ib_rn_get_params(struct ib_device *device, u32 port_num,
3120 				 enum rdma_netdev_t type,
3121 				 struct rdma_netdev_alloc_params *params)
3122 {
3123 	if (type != RDMA_NETDEV_IPOIB)
3124 		return -EOPNOTSUPP;
3125 
3126 	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
3127 }
3128 
3129 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
3130 				       size_t count, loff_t *pos)
3131 {
3132 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3133 	char lbuf[20];
3134 	int len;
3135 
3136 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
3137 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
3138 }
3139 
3140 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
3141 					size_t count, loff_t *pos)
3142 {
3143 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3144 	u32 timeout;
3145 	u32 var;
3146 
3147 	if (kstrtouint_from_user(buf, count, 0, &var))
3148 		return -EFAULT;
3149 
3150 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
3151 			1000);
3152 	if (timeout != var)
3153 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
3154 			    timeout);
3155 
3156 	delay_drop->timeout = timeout;
3157 
3158 	return count;
3159 }
3160 
3161 static const struct file_operations fops_delay_drop_timeout = {
3162 	.owner	= THIS_MODULE,
3163 	.open	= simple_open,
3164 	.write	= delay_drop_timeout_write,
3165 	.read	= delay_drop_timeout_read,
3166 };
3167 
3168 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
3169 				      struct mlx5_ib_multiport_info *mpi)
3170 {
3171 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3172 	struct mlx5_ib_port *port = &ibdev->port[port_num];
3173 	int comps;
3174 	int err;
3175 	int i;
3176 
3177 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3178 
3179 	mlx5_core_mp_event_replay(ibdev->mdev,
3180 				  MLX5_DRIVER_EVENT_AFFILIATION_REMOVED,
3181 				  NULL);
3182 	mlx5_core_mp_event_replay(mpi->mdev,
3183 				  MLX5_DRIVER_EVENT_AFFILIATION_REMOVED,
3184 				  NULL);
3185 
3186 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
3187 
3188 	spin_lock(&port->mp.mpi_lock);
3189 	if (!mpi->ibdev) {
3190 		spin_unlock(&port->mp.mpi_lock);
3191 		return;
3192 	}
3193 
3194 	mpi->ibdev = NULL;
3195 
3196 	spin_unlock(&port->mp.mpi_lock);
3197 	if (mpi->mdev_events.notifier_call)
3198 		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
3199 	mpi->mdev_events.notifier_call = NULL;
3200 	mlx5_mdev_netdev_untrack(ibdev, port_num);
3201 	spin_lock(&port->mp.mpi_lock);
3202 
3203 	comps = mpi->mdev_refcnt;
3204 	if (comps) {
3205 		mpi->unaffiliate = true;
3206 		init_completion(&mpi->unref_comp);
3207 		spin_unlock(&port->mp.mpi_lock);
3208 
3209 		for (i = 0; i < comps; i++)
3210 			wait_for_completion(&mpi->unref_comp);
3211 
3212 		spin_lock(&port->mp.mpi_lock);
3213 		mpi->unaffiliate = false;
3214 	}
3215 
3216 	port->mp.mpi = NULL;
3217 
3218 	spin_unlock(&port->mp.mpi_lock);
3219 
3220 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
3221 
3222 	mlx5_ib_dbg(ibdev, "unaffiliated port %u\n", port_num + 1);
3223 	/* Log an error, still needed to cleanup the pointers and add
3224 	 * it back to the list.
3225 	 */
3226 	if (err)
3227 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
3228 			    port_num + 1);
3229 
3230 	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
3231 }
3232 
3233 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
3234 				    struct mlx5_ib_multiport_info *mpi)
3235 {
3236 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3237 	u64 key;
3238 	int err;
3239 
3240 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3241 
3242 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
3243 	if (ibdev->port[port_num].mp.mpi) {
3244 		mlx5_ib_dbg(ibdev, "port %u already affiliated.\n",
3245 			    port_num + 1);
3246 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3247 		return false;
3248 	}
3249 
3250 	ibdev->port[port_num].mp.mpi = mpi;
3251 	mpi->ibdev = ibdev;
3252 	mpi->mdev_events.notifier_call = NULL;
3253 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3254 
3255 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
3256 	if (err)
3257 		goto unbind;
3258 
3259 	mlx5_mdev_netdev_track(ibdev, port_num);
3260 
3261 	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
3262 	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
3263 
3264 	mlx5_ib_init_cong_debugfs(ibdev, port_num);
3265 
3266 	key = mpi->mdev->priv.adev_idx;
3267 	mlx5_core_mp_event_replay(mpi->mdev,
3268 				  MLX5_DRIVER_EVENT_AFFILIATION_DONE,
3269 				  &key);
3270 	mlx5_core_mp_event_replay(ibdev->mdev,
3271 				  MLX5_DRIVER_EVENT_AFFILIATION_DONE,
3272 				  &key);
3273 
3274 	return true;
3275 
3276 unbind:
3277 	mlx5_ib_unbind_slave_port(ibdev, mpi);
3278 	return false;
3279 }
3280 
3281 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
3282 {
3283 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3284 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3285 							  port_num + 1);
3286 	struct mlx5_ib_multiport_info *mpi;
3287 	int err;
3288 	u32 i;
3289 
3290 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3291 		return 0;
3292 
3293 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
3294 						     &dev->sys_image_guid);
3295 	if (err)
3296 		return err;
3297 
3298 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3299 	if (err)
3300 		return err;
3301 
3302 	mutex_lock(&mlx5_ib_multiport_mutex);
3303 	for (i = 0; i < dev->num_ports; i++) {
3304 		bool bound = false;
3305 
3306 		/* build a stub multiport info struct for the native port. */
3307 		if (i == port_num) {
3308 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
3309 			if (!mpi) {
3310 				mutex_unlock(&mlx5_ib_multiport_mutex);
3311 				mlx5_nic_vport_disable_roce(dev->mdev);
3312 				return -ENOMEM;
3313 			}
3314 
3315 			mpi->is_master = true;
3316 			mpi->mdev = dev->mdev;
3317 			mpi->sys_image_guid = dev->sys_image_guid;
3318 			dev->port[i].mp.mpi = mpi;
3319 			mpi->ibdev = dev;
3320 			mpi = NULL;
3321 			continue;
3322 		}
3323 
3324 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
3325 				    list) {
3326 			if (dev->sys_image_guid == mpi->sys_image_guid &&
3327 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
3328 				bound = mlx5_ib_bind_slave_port(dev, mpi);
3329 			}
3330 
3331 			if (bound) {
3332 				dev_dbg(mpi->mdev->device,
3333 					"removing port from unaffiliated list.\n");
3334 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
3335 				list_del(&mpi->list);
3336 				break;
3337 			}
3338 		}
3339 		if (!bound)
3340 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
3341 				    i + 1);
3342 	}
3343 
3344 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
3345 	mutex_unlock(&mlx5_ib_multiport_mutex);
3346 	return err;
3347 }
3348 
3349 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
3350 {
3351 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3352 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3353 							  port_num + 1);
3354 	u32 i;
3355 
3356 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3357 		return;
3358 
3359 	mutex_lock(&mlx5_ib_multiport_mutex);
3360 	for (i = 0; i < dev->num_ports; i++) {
3361 		if (dev->port[i].mp.mpi) {
3362 			/* Destroy the native port stub */
3363 			if (i == port_num) {
3364 				kfree(dev->port[i].mp.mpi);
3365 				dev->port[i].mp.mpi = NULL;
3366 			} else {
3367 				mlx5_ib_dbg(dev, "unbinding port_num: %u\n",
3368 					    i + 1);
3369 				list_add_tail(&dev->port[i].mp.mpi->list,
3370 					      &mlx5_ib_unaffiliated_port_list);
3371 				mlx5_ib_unbind_slave_port(dev,
3372 							  dev->port[i].mp.mpi);
3373 			}
3374 		}
3375 	}
3376 
3377 	mlx5_ib_dbg(dev, "removing from devlist\n");
3378 	list_del(&dev->ib_dev_list);
3379 	mutex_unlock(&mlx5_ib_multiport_mutex);
3380 
3381 	mlx5_nic_vport_disable_roce(dev->mdev);
3382 }
3383 
3384 static int mmap_obj_cleanup(struct ib_uobject *uobject,
3385 			    enum rdma_remove_reason why,
3386 			    struct uverbs_attr_bundle *attrs)
3387 {
3388 	struct mlx5_user_mmap_entry *obj = uobject->object;
3389 
3390 	rdma_user_mmap_entry_remove(&obj->rdma_entry);
3391 	return 0;
3392 }
3393 
3394 static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c,
3395 					    struct mlx5_user_mmap_entry *entry,
3396 					    size_t length)
3397 {
3398 	return rdma_user_mmap_entry_insert_range(
3399 		&c->ibucontext, &entry->rdma_entry, length,
3400 		(MLX5_IB_MMAP_OFFSET_START << 16),
3401 		((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1));
3402 }
3403 
3404 static struct mlx5_user_mmap_entry *
3405 alloc_var_entry(struct mlx5_ib_ucontext *c)
3406 {
3407 	struct mlx5_user_mmap_entry *entry;
3408 	struct mlx5_var_table *var_table;
3409 	u32 page_idx;
3410 	int err;
3411 
3412 	var_table = &to_mdev(c->ibucontext.device)->var_table;
3413 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3414 	if (!entry)
3415 		return ERR_PTR(-ENOMEM);
3416 
3417 	mutex_lock(&var_table->bitmap_lock);
3418 	page_idx = find_first_zero_bit(var_table->bitmap,
3419 				       var_table->num_var_hw_entries);
3420 	if (page_idx >= var_table->num_var_hw_entries) {
3421 		err = -ENOSPC;
3422 		mutex_unlock(&var_table->bitmap_lock);
3423 		goto end;
3424 	}
3425 
3426 	set_bit(page_idx, var_table->bitmap);
3427 	mutex_unlock(&var_table->bitmap_lock);
3428 
3429 	entry->address = var_table->hw_start_addr +
3430 				(page_idx * var_table->stride_size);
3431 	entry->page_idx = page_idx;
3432 	entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
3433 
3434 	err = mlx5_rdma_user_mmap_entry_insert(c, entry,
3435 					       var_table->stride_size);
3436 	if (err)
3437 		goto err_insert;
3438 
3439 	return entry;
3440 
3441 err_insert:
3442 	mutex_lock(&var_table->bitmap_lock);
3443 	clear_bit(page_idx, var_table->bitmap);
3444 	mutex_unlock(&var_table->bitmap_lock);
3445 end:
3446 	kfree(entry);
3447 	return ERR_PTR(err);
3448 }
3449 
3450 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
3451 	struct uverbs_attr_bundle *attrs)
3452 {
3453 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3454 		attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3455 	struct mlx5_ib_ucontext *c;
3456 	struct mlx5_user_mmap_entry *entry;
3457 	u64 mmap_offset;
3458 	u32 length;
3459 	int err;
3460 
3461 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3462 	if (IS_ERR(c))
3463 		return PTR_ERR(c);
3464 
3465 	entry = alloc_var_entry(c);
3466 	if (IS_ERR(entry))
3467 		return PTR_ERR(entry);
3468 
3469 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3470 	length = entry->rdma_entry.npages * PAGE_SIZE;
3471 	uobj->object = entry;
3472 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3473 
3474 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3475 			     &mmap_offset, sizeof(mmap_offset));
3476 	if (err)
3477 		return err;
3478 
3479 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3480 			     &entry->page_idx, sizeof(entry->page_idx));
3481 	if (err)
3482 		return err;
3483 
3484 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3485 			     &length, sizeof(length));
3486 	return err;
3487 }
3488 
3489 DECLARE_UVERBS_NAMED_METHOD(
3490 	MLX5_IB_METHOD_VAR_OBJ_ALLOC,
3491 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
3492 			MLX5_IB_OBJECT_VAR,
3493 			UVERBS_ACCESS_NEW,
3494 			UA_MANDATORY),
3495 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3496 			   UVERBS_ATTR_TYPE(u32),
3497 			   UA_MANDATORY),
3498 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3499 			   UVERBS_ATTR_TYPE(u32),
3500 			   UA_MANDATORY),
3501 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3502 			    UVERBS_ATTR_TYPE(u64),
3503 			    UA_MANDATORY));
3504 
3505 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3506 	MLX5_IB_METHOD_VAR_OBJ_DESTROY,
3507 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
3508 			MLX5_IB_OBJECT_VAR,
3509 			UVERBS_ACCESS_DESTROY,
3510 			UA_MANDATORY));
3511 
3512 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
3513 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3514 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
3515 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
3516 
3517 static bool var_is_supported(struct ib_device *device)
3518 {
3519 	struct mlx5_ib_dev *dev = to_mdev(device);
3520 
3521 	return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3522 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
3523 }
3524 
3525 static struct mlx5_user_mmap_entry *
3526 alloc_uar_entry(struct mlx5_ib_ucontext *c,
3527 		enum mlx5_ib_uapi_uar_alloc_type alloc_type)
3528 {
3529 	struct mlx5_user_mmap_entry *entry;
3530 	struct mlx5_ib_dev *dev;
3531 	u32 uar_index;
3532 	int err;
3533 
3534 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3535 	if (!entry)
3536 		return ERR_PTR(-ENOMEM);
3537 
3538 	dev = to_mdev(c->ibucontext.device);
3539 	err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index, c->devx_uid);
3540 	if (err)
3541 		goto end;
3542 
3543 	entry->page_idx = uar_index;
3544 	entry->address = uar_index2paddress(dev, uar_index);
3545 	if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3546 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC;
3547 	else
3548 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC;
3549 
3550 	err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE);
3551 	if (err)
3552 		goto err_insert;
3553 
3554 	return entry;
3555 
3556 err_insert:
3557 	mlx5_cmd_uar_dealloc(dev->mdev, uar_index, c->devx_uid);
3558 end:
3559 	kfree(entry);
3560 	return ERR_PTR(err);
3561 }
3562 
3563 static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)(
3564 	struct uverbs_attr_bundle *attrs)
3565 {
3566 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3567 		attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3568 	enum mlx5_ib_uapi_uar_alloc_type alloc_type;
3569 	struct mlx5_ib_ucontext *c;
3570 	struct mlx5_user_mmap_entry *entry;
3571 	u64 mmap_offset;
3572 	u32 length;
3573 	int err;
3574 
3575 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3576 	if (IS_ERR(c))
3577 		return PTR_ERR(c);
3578 
3579 	err = uverbs_get_const(&alloc_type, attrs,
3580 			       MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE);
3581 	if (err)
3582 		return err;
3583 
3584 	if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF &&
3585 	    alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC)
3586 		return -EOPNOTSUPP;
3587 
3588 	if (!to_mdev(c->ibucontext.device)->wc_support &&
3589 	    alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3590 		return -EOPNOTSUPP;
3591 
3592 	entry = alloc_uar_entry(c, alloc_type);
3593 	if (IS_ERR(entry))
3594 		return PTR_ERR(entry);
3595 
3596 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3597 	length = entry->rdma_entry.npages * PAGE_SIZE;
3598 	uobj->object = entry;
3599 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3600 
3601 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3602 			     &mmap_offset, sizeof(mmap_offset));
3603 	if (err)
3604 		return err;
3605 
3606 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3607 			     &entry->page_idx, sizeof(entry->page_idx));
3608 	if (err)
3609 		return err;
3610 
3611 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3612 			     &length, sizeof(length));
3613 	return err;
3614 }
3615 
3616 DECLARE_UVERBS_NAMED_METHOD(
3617 	MLX5_IB_METHOD_UAR_OBJ_ALLOC,
3618 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE,
3619 			MLX5_IB_OBJECT_UAR,
3620 			UVERBS_ACCESS_NEW,
3621 			UA_MANDATORY),
3622 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE,
3623 			     enum mlx5_ib_uapi_uar_alloc_type,
3624 			     UA_MANDATORY),
3625 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3626 			   UVERBS_ATTR_TYPE(u32),
3627 			   UA_MANDATORY),
3628 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3629 			   UVERBS_ATTR_TYPE(u32),
3630 			   UA_MANDATORY),
3631 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3632 			    UVERBS_ATTR_TYPE(u64),
3633 			    UA_MANDATORY));
3634 
3635 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3636 	MLX5_IB_METHOD_UAR_OBJ_DESTROY,
3637 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE,
3638 			MLX5_IB_OBJECT_UAR,
3639 			UVERBS_ACCESS_DESTROY,
3640 			UA_MANDATORY));
3641 
3642 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR,
3643 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3644 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC),
3645 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY));
3646 
3647 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3648 	mlx5_ib_query_context,
3649 	UVERBS_OBJECT_DEVICE,
3650 	UVERBS_METHOD_QUERY_CONTEXT,
3651 	UVERBS_ATTR_PTR_OUT(
3652 		MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
3653 		UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp,
3654 				   dump_fill_mkey),
3655 		UA_MANDATORY));
3656 
3657 static const struct uapi_definition mlx5_ib_defs[] = {
3658 	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
3659 	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
3660 	UAPI_DEF_CHAIN(mlx5_ib_qos_defs),
3661 	UAPI_DEF_CHAIN(mlx5_ib_std_types_defs),
3662 	UAPI_DEF_CHAIN(mlx5_ib_dm_defs),
3663 
3664 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context),
3665 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
3666 				UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
3667 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR),
3668 	{}
3669 };
3670 
3671 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
3672 {
3673 	mlx5_ib_cleanup_multiport_master(dev);
3674 	WARN_ON(!xa_empty(&dev->odp_mkeys));
3675 	mutex_destroy(&dev->cap_mask_mutex);
3676 	WARN_ON(!xa_empty(&dev->sig_mrs));
3677 	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
3678 	mlx5r_macsec_dealloc_gids(dev);
3679 }
3680 
3681 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
3682 {
3683 	struct mlx5_core_dev *mdev = dev->mdev;
3684 	int err, i;
3685 
3686 	dev->ib_dev.node_type = RDMA_NODE_IB_CA;
3687 	dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
3688 	dev->ib_dev.phys_port_cnt = dev->num_ports;
3689 	dev->ib_dev.dev.parent = mdev->device;
3690 	dev->ib_dev.lag_flags = RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
3691 
3692 	for (i = 0; i < dev->num_ports; i++) {
3693 		spin_lock_init(&dev->port[i].mp.mpi_lock);
3694 		rwlock_init(&dev->port[i].roce.netdev_lock);
3695 		dev->port[i].roce.dev = dev;
3696 		dev->port[i].roce.native_port_num = i + 1;
3697 		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
3698 	}
3699 
3700 	err = mlx5r_cmd_query_special_mkeys(dev);
3701 	if (err)
3702 		return err;
3703 
3704 	err = mlx5r_macsec_init_gids_and_devlist(dev);
3705 	if (err)
3706 		return err;
3707 
3708 	err = mlx5_ib_init_multiport_master(dev);
3709 	if (err)
3710 		goto err;
3711 
3712 	err = set_has_smi_cap(dev);
3713 	if (err)
3714 		goto err_mp;
3715 
3716 	err = mlx5_query_max_pkeys(&dev->ib_dev, &dev->pkey_table_len);
3717 	if (err)
3718 		goto err_mp;
3719 
3720 	if (mlx5_use_mad_ifc(dev))
3721 		get_ext_port_caps(dev);
3722 
3723 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_max(mdev);
3724 
3725 	mutex_init(&dev->cap_mask_mutex);
3726 	INIT_LIST_HEAD(&dev->qp_list);
3727 	spin_lock_init(&dev->reset_flow_resource_lock);
3728 	xa_init(&dev->odp_mkeys);
3729 	xa_init(&dev->sig_mrs);
3730 	atomic_set(&dev->mkey_var, 0);
3731 
3732 	spin_lock_init(&dev->dm.lock);
3733 	dev->dm.dev = mdev;
3734 	return 0;
3735 err_mp:
3736 	mlx5_ib_cleanup_multiport_master(dev);
3737 err:
3738 	mlx5r_macsec_dealloc_gids(dev);
3739 	return err;
3740 }
3741 
3742 static int mlx5_ib_enable_driver(struct ib_device *dev)
3743 {
3744 	struct mlx5_ib_dev *mdev = to_mdev(dev);
3745 	int ret;
3746 
3747 	ret = mlx5_ib_test_wc(mdev);
3748 	mlx5_ib_dbg(mdev, "Write-Combining %s",
3749 		    mdev->wc_support ? "supported" : "not supported");
3750 
3751 	return ret;
3752 }
3753 
3754 static const struct ib_device_ops mlx5_ib_dev_ops = {
3755 	.owner = THIS_MODULE,
3756 	.driver_id = RDMA_DRIVER_MLX5,
3757 	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
3758 
3759 	.add_gid = mlx5_ib_add_gid,
3760 	.alloc_mr = mlx5_ib_alloc_mr,
3761 	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
3762 	.alloc_pd = mlx5_ib_alloc_pd,
3763 	.alloc_ucontext = mlx5_ib_alloc_ucontext,
3764 	.attach_mcast = mlx5_ib_mcg_attach,
3765 	.check_mr_status = mlx5_ib_check_mr_status,
3766 	.create_ah = mlx5_ib_create_ah,
3767 	.create_cq = mlx5_ib_create_cq,
3768 	.create_qp = mlx5_ib_create_qp,
3769 	.create_srq = mlx5_ib_create_srq,
3770 	.create_user_ah = mlx5_ib_create_ah,
3771 	.dealloc_pd = mlx5_ib_dealloc_pd,
3772 	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
3773 	.del_gid = mlx5_ib_del_gid,
3774 	.dereg_mr = mlx5_ib_dereg_mr,
3775 	.destroy_ah = mlx5_ib_destroy_ah,
3776 	.destroy_cq = mlx5_ib_destroy_cq,
3777 	.destroy_qp = mlx5_ib_destroy_qp,
3778 	.destroy_srq = mlx5_ib_destroy_srq,
3779 	.detach_mcast = mlx5_ib_mcg_detach,
3780 	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
3781 	.drain_rq = mlx5_ib_drain_rq,
3782 	.drain_sq = mlx5_ib_drain_sq,
3783 	.device_group = &mlx5_attr_group,
3784 	.enable_driver = mlx5_ib_enable_driver,
3785 	.get_dev_fw_str = get_dev_fw_str,
3786 	.get_dma_mr = mlx5_ib_get_dma_mr,
3787 	.get_link_layer = mlx5_ib_port_link_layer,
3788 	.map_mr_sg = mlx5_ib_map_mr_sg,
3789 	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
3790 	.mmap = mlx5_ib_mmap,
3791 	.mmap_free = mlx5_ib_mmap_free,
3792 	.modify_cq = mlx5_ib_modify_cq,
3793 	.modify_device = mlx5_ib_modify_device,
3794 	.modify_port = mlx5_ib_modify_port,
3795 	.modify_qp = mlx5_ib_modify_qp,
3796 	.modify_srq = mlx5_ib_modify_srq,
3797 	.poll_cq = mlx5_ib_poll_cq,
3798 	.post_recv = mlx5_ib_post_recv_nodrain,
3799 	.post_send = mlx5_ib_post_send_nodrain,
3800 	.post_srq_recv = mlx5_ib_post_srq_recv,
3801 	.process_mad = mlx5_ib_process_mad,
3802 	.query_ah = mlx5_ib_query_ah,
3803 	.query_device = mlx5_ib_query_device,
3804 	.query_gid = mlx5_ib_query_gid,
3805 	.query_pkey = mlx5_ib_query_pkey,
3806 	.query_qp = mlx5_ib_query_qp,
3807 	.query_srq = mlx5_ib_query_srq,
3808 	.query_ucontext = mlx5_ib_query_ucontext,
3809 	.reg_user_mr = mlx5_ib_reg_user_mr,
3810 	.reg_user_mr_dmabuf = mlx5_ib_reg_user_mr_dmabuf,
3811 	.req_notify_cq = mlx5_ib_arm_cq,
3812 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
3813 	.resize_cq = mlx5_ib_resize_cq,
3814 
3815 	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
3816 	INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs),
3817 	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
3818 	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
3819 	INIT_RDMA_OBJ_SIZE(ib_qp, mlx5_ib_qp, ibqp),
3820 	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
3821 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
3822 };
3823 
3824 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
3825 	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
3826 };
3827 
3828 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
3829 	.get_vf_config = mlx5_ib_get_vf_config,
3830 	.get_vf_guid = mlx5_ib_get_vf_guid,
3831 	.get_vf_stats = mlx5_ib_get_vf_stats,
3832 	.set_vf_guid = mlx5_ib_set_vf_guid,
3833 	.set_vf_link_state = mlx5_ib_set_vf_link_state,
3834 };
3835 
3836 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
3837 	.alloc_mw = mlx5_ib_alloc_mw,
3838 	.dealloc_mw = mlx5_ib_dealloc_mw,
3839 
3840 	INIT_RDMA_OBJ_SIZE(ib_mw, mlx5_ib_mw, ibmw),
3841 };
3842 
3843 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
3844 	.alloc_xrcd = mlx5_ib_alloc_xrcd,
3845 	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
3846 
3847 	INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd),
3848 };
3849 
3850 static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
3851 {
3852 	struct mlx5_core_dev *mdev = dev->mdev;
3853 	struct mlx5_var_table *var_table = &dev->var_table;
3854 	u8 log_doorbell_bar_size;
3855 	u8 log_doorbell_stride;
3856 	u64 bar_size;
3857 
3858 	log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3859 					log_doorbell_bar_size);
3860 	log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3861 					log_doorbell_stride);
3862 	var_table->hw_start_addr = dev->mdev->bar_addr +
3863 				MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
3864 					doorbell_bar_offset);
3865 	bar_size = (1ULL << log_doorbell_bar_size) * 4096;
3866 	var_table->stride_size = 1ULL << log_doorbell_stride;
3867 	var_table->num_var_hw_entries = div_u64(bar_size,
3868 						var_table->stride_size);
3869 	mutex_init(&var_table->bitmap_lock);
3870 	var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
3871 					  GFP_KERNEL);
3872 	return (var_table->bitmap) ? 0 : -ENOMEM;
3873 }
3874 
3875 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
3876 {
3877 	bitmap_free(dev->var_table.bitmap);
3878 }
3879 
3880 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
3881 {
3882 	struct mlx5_core_dev *mdev = dev->mdev;
3883 	int err;
3884 
3885 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
3886 	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
3887 		ib_set_device_ops(&dev->ib_dev,
3888 				  &mlx5_ib_dev_ipoib_enhanced_ops);
3889 
3890 	if (mlx5_core_is_pf(mdev))
3891 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
3892 
3893 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
3894 
3895 	if (MLX5_CAP_GEN(mdev, imaicl))
3896 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
3897 
3898 	if (MLX5_CAP_GEN(mdev, xrc))
3899 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
3900 
3901 	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
3902 	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3903 	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
3904 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
3905 
3906 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
3907 
3908 	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
3909 		dev->ib_dev.driver_def = mlx5_ib_defs;
3910 
3911 	err = init_node_data(dev);
3912 	if (err)
3913 		return err;
3914 
3915 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
3916 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
3917 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
3918 		mutex_init(&dev->lb.mutex);
3919 
3920 	if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3921 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
3922 		err = mlx5_ib_init_var_table(dev);
3923 		if (err)
3924 			return err;
3925 	}
3926 
3927 	dev->ib_dev.use_cq_dim = true;
3928 
3929 	return 0;
3930 }
3931 
3932 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
3933 	.get_port_immutable = mlx5_port_immutable,
3934 	.query_port = mlx5_ib_query_port,
3935 };
3936 
3937 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
3938 {
3939 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
3940 	return 0;
3941 }
3942 
3943 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
3944 	.get_port_immutable = mlx5_port_rep_immutable,
3945 	.query_port = mlx5_ib_rep_query_port,
3946 	.query_pkey = mlx5_ib_rep_query_pkey,
3947 };
3948 
3949 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
3950 {
3951 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
3952 	return 0;
3953 }
3954 
3955 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
3956 	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
3957 	.create_wq = mlx5_ib_create_wq,
3958 	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
3959 	.destroy_wq = mlx5_ib_destroy_wq,
3960 	.get_netdev = mlx5_ib_get_netdev,
3961 	.modify_wq = mlx5_ib_modify_wq,
3962 
3963 	INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table,
3964 			   ib_rwq_ind_tbl),
3965 };
3966 
3967 static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev)
3968 {
3969 	struct mlx5_core_dev *mdev = dev->mdev;
3970 	enum rdma_link_layer ll;
3971 	int port_type_cap;
3972 	u32 port_num = 0;
3973 	int err;
3974 
3975 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3976 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3977 
3978 	if (ll == IB_LINK_LAYER_ETHERNET) {
3979 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
3980 
3981 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3982 
3983 		/* Register only for native ports */
3984 		mlx5_mdev_netdev_track(dev, port_num);
3985 
3986 		err = mlx5_enable_eth(dev);
3987 		if (err)
3988 			goto cleanup;
3989 	}
3990 
3991 	return 0;
3992 cleanup:
3993 	mlx5_mdev_netdev_untrack(dev, port_num);
3994 	return err;
3995 }
3996 
3997 static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev)
3998 {
3999 	struct mlx5_core_dev *mdev = dev->mdev;
4000 	enum rdma_link_layer ll;
4001 	int port_type_cap;
4002 	u32 port_num;
4003 
4004 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4005 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4006 
4007 	if (ll == IB_LINK_LAYER_ETHERNET) {
4008 		mlx5_disable_eth(dev);
4009 
4010 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4011 		mlx5_mdev_netdev_untrack(dev, port_num);
4012 	}
4013 }
4014 
4015 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
4016 {
4017 	mlx5_ib_init_cong_debugfs(dev,
4018 				  mlx5_core_native_port_num(dev->mdev) - 1);
4019 	return 0;
4020 }
4021 
4022 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
4023 {
4024 	mlx5_ib_cleanup_cong_debugfs(dev,
4025 				     mlx5_core_native_port_num(dev->mdev) - 1);
4026 }
4027 
4028 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
4029 {
4030 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
4031 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
4032 }
4033 
4034 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
4035 {
4036 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
4037 }
4038 
4039 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
4040 {
4041 	int err;
4042 
4043 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
4044 	if (err)
4045 		return err;
4046 
4047 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
4048 	if (err)
4049 		mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4050 
4051 	return err;
4052 }
4053 
4054 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
4055 {
4056 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4057 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4058 }
4059 
4060 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
4061 {
4062 	const char *name;
4063 
4064 	if (!mlx5_lag_is_active(dev->mdev))
4065 		name = "mlx5_%d";
4066 	else
4067 		name = "mlx5_bond_%d";
4068 	return ib_register_device(&dev->ib_dev, name, &dev->mdev->pdev->dev);
4069 }
4070 
4071 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
4072 {
4073 	mlx5_mkey_cache_cleanup(dev);
4074 	mlx5r_umr_resource_cleanup(dev);
4075 }
4076 
4077 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
4078 {
4079 	ib_unregister_device(&dev->ib_dev);
4080 }
4081 
4082 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
4083 {
4084 	int ret;
4085 
4086 	ret = mlx5r_umr_resource_init(dev);
4087 	if (ret)
4088 		return ret;
4089 
4090 	ret = mlx5_mkey_cache_init(dev);
4091 	if (ret)
4092 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4093 	return ret;
4094 }
4095 
4096 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
4097 {
4098 	struct dentry *root;
4099 
4100 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4101 		return 0;
4102 
4103 	mutex_init(&dev->delay_drop.lock);
4104 	dev->delay_drop.dev = dev;
4105 	dev->delay_drop.activate = false;
4106 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4107 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4108 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4109 	atomic_set(&dev->delay_drop.events_cnt, 0);
4110 
4111 	if (!mlx5_debugfs_root)
4112 		return 0;
4113 
4114 	root = debugfs_create_dir("delay_drop", mlx5_debugfs_get_dev_root(dev->mdev));
4115 	dev->delay_drop.dir_debugfs = root;
4116 
4117 	debugfs_create_atomic_t("num_timeout_events", 0400, root,
4118 				&dev->delay_drop.events_cnt);
4119 	debugfs_create_atomic_t("num_rqs", 0400, root,
4120 				&dev->delay_drop.rqs_cnt);
4121 	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
4122 			    &fops_delay_drop_timeout);
4123 	return 0;
4124 }
4125 
4126 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
4127 {
4128 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4129 		return;
4130 
4131 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4132 	if (!dev->delay_drop.dir_debugfs)
4133 		return;
4134 
4135 	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
4136 	dev->delay_drop.dir_debugfs = NULL;
4137 }
4138 
4139 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
4140 {
4141 	dev->mdev_events.notifier_call = mlx5_ib_event;
4142 	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
4143 
4144 	mlx5r_macsec_event_register(dev);
4145 
4146 	return 0;
4147 }
4148 
4149 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
4150 {
4151 	mlx5r_macsec_event_unregister(dev);
4152 	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
4153 }
4154 
4155 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
4156 		      const struct mlx5_ib_profile *profile,
4157 		      int stage)
4158 {
4159 	dev->ib_active = false;
4160 
4161 	/* Number of stages to cleanup */
4162 	while (stage) {
4163 		stage--;
4164 		if (profile->stage[stage].cleanup)
4165 			profile->stage[stage].cleanup(dev);
4166 	}
4167 
4168 	kfree(dev->port);
4169 	ib_dealloc_device(&dev->ib_dev);
4170 }
4171 
4172 int __mlx5_ib_add(struct mlx5_ib_dev *dev,
4173 		  const struct mlx5_ib_profile *profile)
4174 {
4175 	int err;
4176 	int i;
4177 
4178 	dev->profile = profile;
4179 
4180 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
4181 		if (profile->stage[i].init) {
4182 			err = profile->stage[i].init(dev);
4183 			if (err)
4184 				goto err_out;
4185 		}
4186 	}
4187 
4188 	dev->ib_active = true;
4189 	return 0;
4190 
4191 err_out:
4192 	/* Clean up stages which were initialized */
4193 	while (i) {
4194 		i--;
4195 		if (profile->stage[i].cleanup)
4196 			profile->stage[i].cleanup(dev);
4197 	}
4198 	return -ENOMEM;
4199 }
4200 
4201 static const struct mlx5_ib_profile pf_profile = {
4202 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4203 		     mlx5_ib_stage_init_init,
4204 		     mlx5_ib_stage_init_cleanup),
4205 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4206 		     mlx5_ib_fs_init,
4207 		     mlx5_ib_fs_cleanup),
4208 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4209 		     mlx5_ib_stage_caps_init,
4210 		     mlx5_ib_stage_caps_cleanup),
4211 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4212 		     mlx5_ib_stage_non_default_cb,
4213 		     NULL),
4214 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4215 		     mlx5_ib_roce_init,
4216 		     mlx5_ib_roce_cleanup),
4217 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4218 		     mlx5_init_qp_table,
4219 		     mlx5_cleanup_qp_table),
4220 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4221 		     mlx5_init_srq_table,
4222 		     mlx5_cleanup_srq_table),
4223 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4224 		     mlx5_ib_dev_res_init,
4225 		     mlx5_ib_dev_res_cleanup),
4226 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4227 		     mlx5_ib_stage_dev_notifier_init,
4228 		     mlx5_ib_stage_dev_notifier_cleanup),
4229 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
4230 		     mlx5_ib_odp_init_one,
4231 		     mlx5_ib_odp_cleanup_one),
4232 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4233 		     mlx5_ib_counters_init,
4234 		     mlx5_ib_counters_cleanup),
4235 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4236 		     mlx5_ib_stage_cong_debugfs_init,
4237 		     mlx5_ib_stage_cong_debugfs_cleanup),
4238 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4239 		     mlx5_ib_stage_uar_init,
4240 		     mlx5_ib_stage_uar_cleanup),
4241 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4242 		     mlx5_ib_stage_bfrag_init,
4243 		     mlx5_ib_stage_bfrag_cleanup),
4244 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4245 		     NULL,
4246 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4247 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4248 		     mlx5_ib_devx_init,
4249 		     mlx5_ib_devx_cleanup),
4250 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4251 		     mlx5_ib_stage_ib_reg_init,
4252 		     mlx5_ib_stage_ib_reg_cleanup),
4253 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4254 		     mlx5_ib_stage_post_ib_reg_umr_init,
4255 		     NULL),
4256 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4257 		     mlx5_ib_stage_delay_drop_init,
4258 		     mlx5_ib_stage_delay_drop_cleanup),
4259 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4260 		     mlx5_ib_restrack_init,
4261 		     NULL),
4262 };
4263 
4264 const struct mlx5_ib_profile raw_eth_profile = {
4265 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4266 		     mlx5_ib_stage_init_init,
4267 		     mlx5_ib_stage_init_cleanup),
4268 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4269 		     mlx5_ib_fs_init,
4270 		     mlx5_ib_fs_cleanup),
4271 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4272 		     mlx5_ib_stage_caps_init,
4273 		     mlx5_ib_stage_caps_cleanup),
4274 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4275 		     mlx5_ib_stage_raw_eth_non_default_cb,
4276 		     NULL),
4277 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4278 		     mlx5_ib_roce_init,
4279 		     mlx5_ib_roce_cleanup),
4280 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4281 		     mlx5_init_qp_table,
4282 		     mlx5_cleanup_qp_table),
4283 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4284 		     mlx5_init_srq_table,
4285 		     mlx5_cleanup_srq_table),
4286 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4287 		     mlx5_ib_dev_res_init,
4288 		     mlx5_ib_dev_res_cleanup),
4289 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4290 		     mlx5_ib_stage_dev_notifier_init,
4291 		     mlx5_ib_stage_dev_notifier_cleanup),
4292 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4293 		     mlx5_ib_counters_init,
4294 		     mlx5_ib_counters_cleanup),
4295 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4296 		     mlx5_ib_stage_cong_debugfs_init,
4297 		     mlx5_ib_stage_cong_debugfs_cleanup),
4298 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4299 		     mlx5_ib_stage_uar_init,
4300 		     mlx5_ib_stage_uar_cleanup),
4301 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4302 		     mlx5_ib_stage_bfrag_init,
4303 		     mlx5_ib_stage_bfrag_cleanup),
4304 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4305 		     NULL,
4306 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4307 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4308 		     mlx5_ib_devx_init,
4309 		     mlx5_ib_devx_cleanup),
4310 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4311 		     mlx5_ib_stage_ib_reg_init,
4312 		     mlx5_ib_stage_ib_reg_cleanup),
4313 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4314 		     mlx5_ib_stage_post_ib_reg_umr_init,
4315 		     NULL),
4316 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4317 		     mlx5_ib_stage_delay_drop_init,
4318 		     mlx5_ib_stage_delay_drop_cleanup),
4319 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4320 		     mlx5_ib_restrack_init,
4321 		     NULL),
4322 };
4323 
4324 static int mlx5r_mp_probe(struct auxiliary_device *adev,
4325 			  const struct auxiliary_device_id *id)
4326 {
4327 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4328 	struct mlx5_core_dev *mdev = idev->mdev;
4329 	struct mlx5_ib_multiport_info *mpi;
4330 	struct mlx5_ib_dev *dev;
4331 	bool bound = false;
4332 	int err;
4333 
4334 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4335 	if (!mpi)
4336 		return -ENOMEM;
4337 
4338 	mpi->mdev = mdev;
4339 	err = mlx5_query_nic_vport_system_image_guid(mdev,
4340 						     &mpi->sys_image_guid);
4341 	if (err) {
4342 		kfree(mpi);
4343 		return err;
4344 	}
4345 
4346 	mutex_lock(&mlx5_ib_multiport_mutex);
4347 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
4348 		if (dev->sys_image_guid == mpi->sys_image_guid)
4349 			bound = mlx5_ib_bind_slave_port(dev, mpi);
4350 
4351 		if (bound) {
4352 			rdma_roce_rescan_device(&dev->ib_dev);
4353 			mpi->ibdev->ib_active = true;
4354 			break;
4355 		}
4356 	}
4357 
4358 	if (!bound) {
4359 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4360 		dev_dbg(mdev->device,
4361 			"no suitable IB device found to bind to, added to unaffiliated list.\n");
4362 	}
4363 	mutex_unlock(&mlx5_ib_multiport_mutex);
4364 
4365 	auxiliary_set_drvdata(adev, mpi);
4366 	return 0;
4367 }
4368 
4369 static void mlx5r_mp_remove(struct auxiliary_device *adev)
4370 {
4371 	struct mlx5_ib_multiport_info *mpi;
4372 
4373 	mpi = auxiliary_get_drvdata(adev);
4374 	mutex_lock(&mlx5_ib_multiport_mutex);
4375 	if (mpi->ibdev)
4376 		mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
4377 	else
4378 		list_del(&mpi->list);
4379 	mutex_unlock(&mlx5_ib_multiport_mutex);
4380 	kfree(mpi);
4381 }
4382 
4383 static int mlx5r_probe(struct auxiliary_device *adev,
4384 		       const struct auxiliary_device_id *id)
4385 {
4386 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4387 	struct mlx5_core_dev *mdev = idev->mdev;
4388 	const struct mlx5_ib_profile *profile;
4389 	int port_type_cap, num_ports, ret;
4390 	enum rdma_link_layer ll;
4391 	struct mlx5_ib_dev *dev;
4392 
4393 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4394 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4395 
4396 	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
4397 			MLX5_CAP_GEN(mdev, num_vhca_ports));
4398 	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
4399 	if (!dev)
4400 		return -ENOMEM;
4401 	dev->port = kcalloc(num_ports, sizeof(*dev->port),
4402 			     GFP_KERNEL);
4403 	if (!dev->port) {
4404 		ib_dealloc_device(&dev->ib_dev);
4405 		return -ENOMEM;
4406 	}
4407 
4408 	dev->mdev = mdev;
4409 	dev->num_ports = num_ports;
4410 
4411 	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_get_roce_state(mdev))
4412 		profile = &raw_eth_profile;
4413 	else
4414 		profile = &pf_profile;
4415 
4416 	ret = __mlx5_ib_add(dev, profile);
4417 	if (ret) {
4418 		kfree(dev->port);
4419 		ib_dealloc_device(&dev->ib_dev);
4420 		return ret;
4421 	}
4422 
4423 	auxiliary_set_drvdata(adev, dev);
4424 	return 0;
4425 }
4426 
4427 static void mlx5r_remove(struct auxiliary_device *adev)
4428 {
4429 	struct mlx5_ib_dev *dev;
4430 
4431 	dev = auxiliary_get_drvdata(adev);
4432 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
4433 }
4434 
4435 static const struct auxiliary_device_id mlx5r_mp_id_table[] = {
4436 	{ .name = MLX5_ADEV_NAME ".multiport", },
4437 	{},
4438 };
4439 
4440 static const struct auxiliary_device_id mlx5r_id_table[] = {
4441 	{ .name = MLX5_ADEV_NAME ".rdma", },
4442 	{},
4443 };
4444 
4445 MODULE_DEVICE_TABLE(auxiliary, mlx5r_mp_id_table);
4446 MODULE_DEVICE_TABLE(auxiliary, mlx5r_id_table);
4447 
4448 static struct auxiliary_driver mlx5r_mp_driver = {
4449 	.name = "multiport",
4450 	.probe = mlx5r_mp_probe,
4451 	.remove = mlx5r_mp_remove,
4452 	.id_table = mlx5r_mp_id_table,
4453 };
4454 
4455 static struct auxiliary_driver mlx5r_driver = {
4456 	.name = "rdma",
4457 	.probe = mlx5r_probe,
4458 	.remove = mlx5r_remove,
4459 	.id_table = mlx5r_id_table,
4460 };
4461 
4462 static int __init mlx5_ib_init(void)
4463 {
4464 	int ret;
4465 
4466 	xlt_emergency_page = (void *)__get_free_page(GFP_KERNEL);
4467 	if (!xlt_emergency_page)
4468 		return -ENOMEM;
4469 
4470 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
4471 	if (!mlx5_ib_event_wq) {
4472 		free_page((unsigned long)xlt_emergency_page);
4473 		return -ENOMEM;
4474 	}
4475 
4476 	ret = mlx5_ib_qp_event_init();
4477 	if (ret)
4478 		goto qp_event_err;
4479 
4480 	mlx5_ib_odp_init();
4481 	ret = mlx5r_rep_init();
4482 	if (ret)
4483 		goto rep_err;
4484 	ret = auxiliary_driver_register(&mlx5r_mp_driver);
4485 	if (ret)
4486 		goto mp_err;
4487 	ret = auxiliary_driver_register(&mlx5r_driver);
4488 	if (ret)
4489 		goto drv_err;
4490 	return 0;
4491 
4492 drv_err:
4493 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4494 mp_err:
4495 	mlx5r_rep_cleanup();
4496 rep_err:
4497 	mlx5_ib_qp_event_cleanup();
4498 qp_event_err:
4499 	destroy_workqueue(mlx5_ib_event_wq);
4500 	free_page((unsigned long)xlt_emergency_page);
4501 	return ret;
4502 }
4503 
4504 static void __exit mlx5_ib_cleanup(void)
4505 {
4506 	auxiliary_driver_unregister(&mlx5r_driver);
4507 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4508 	mlx5r_rep_cleanup();
4509 
4510 	mlx5_ib_qp_event_cleanup();
4511 	destroy_workqueue(mlx5_ib_event_wq);
4512 	free_page((unsigned long)xlt_emergency_page);
4513 }
4514 
4515 module_init(mlx5_ib_init);
4516 module_exit(mlx5_ib_cleanup);
4517