xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision 464a5728)
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2013-2020, Mellanox Technologies inc. All rights reserved.
4  * Copyright (c) 2020, Intel Corporation. All rights reserved.
5  */
6 
7 #include <linux/debugfs.h>
8 #include <linux/highmem.h>
9 #include <linux/module.h>
10 #include <linux/init.h>
11 #include <linux/errno.h>
12 #include <linux/pci.h>
13 #include <linux/dma-mapping.h>
14 #include <linux/slab.h>
15 #include <linux/bitmap.h>
16 #include <linux/sched.h>
17 #include <linux/sched/mm.h>
18 #include <linux/sched/task.h>
19 #include <linux/delay.h>
20 #include <rdma/ib_user_verbs.h>
21 #include <rdma/ib_addr.h>
22 #include <rdma/ib_cache.h>
23 #include <linux/mlx5/port.h>
24 #include <linux/mlx5/vport.h>
25 #include <linux/mlx5/fs.h>
26 #include <linux/mlx5/eswitch.h>
27 #include <linux/list.h>
28 #include <rdma/ib_smi.h>
29 #include <rdma/ib_umem.h>
30 #include <rdma/lag.h>
31 #include <linux/in.h>
32 #include <linux/etherdevice.h>
33 #include "mlx5_ib.h"
34 #include "ib_rep.h"
35 #include "cmd.h"
36 #include "devx.h"
37 #include "dm.h"
38 #include "fs.h"
39 #include "srq.h"
40 #include "qp.h"
41 #include "wr.h"
42 #include "restrack.h"
43 #include "counters.h"
44 #include <linux/mlx5/accel.h>
45 #include <rdma/uverbs_std_types.h>
46 #include <rdma/uverbs_ioctl.h>
47 #include <rdma/mlx5_user_ioctl_verbs.h>
48 #include <rdma/mlx5_user_ioctl_cmds.h>
49 #include <rdma/ib_umem_odp.h>
50 
51 #define UVERBS_MODULE_NAME mlx5_ib
52 #include <rdma/uverbs_named_ioctl.h>
53 
54 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
55 MODULE_DESCRIPTION("Mellanox 5th generation network adapters (ConnectX series) IB driver");
56 MODULE_LICENSE("Dual BSD/GPL");
57 
58 struct mlx5_ib_event_work {
59 	struct work_struct	work;
60 	union {
61 		struct mlx5_ib_dev	      *dev;
62 		struct mlx5_ib_multiport_info *mpi;
63 	};
64 	bool			is_slave;
65 	unsigned int		event;
66 	void			*param;
67 };
68 
69 enum {
70 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
71 };
72 
73 static struct workqueue_struct *mlx5_ib_event_wq;
74 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
75 static LIST_HEAD(mlx5_ib_dev_list);
76 /*
77  * This mutex should be held when accessing either of the above lists
78  */
79 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
80 
81 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
82 {
83 	struct mlx5_ib_dev *dev;
84 
85 	mutex_lock(&mlx5_ib_multiport_mutex);
86 	dev = mpi->ibdev;
87 	mutex_unlock(&mlx5_ib_multiport_mutex);
88 	return dev;
89 }
90 
91 static enum rdma_link_layer
92 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
93 {
94 	switch (port_type_cap) {
95 	case MLX5_CAP_PORT_TYPE_IB:
96 		return IB_LINK_LAYER_INFINIBAND;
97 	case MLX5_CAP_PORT_TYPE_ETH:
98 		return IB_LINK_LAYER_ETHERNET;
99 	default:
100 		return IB_LINK_LAYER_UNSPECIFIED;
101 	}
102 }
103 
104 static enum rdma_link_layer
105 mlx5_ib_port_link_layer(struct ib_device *device, u32 port_num)
106 {
107 	struct mlx5_ib_dev *dev = to_mdev(device);
108 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
109 
110 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
111 }
112 
113 static int get_port_state(struct ib_device *ibdev,
114 			  u32 port_num,
115 			  enum ib_port_state *state)
116 {
117 	struct ib_port_attr attr;
118 	int ret;
119 
120 	memset(&attr, 0, sizeof(attr));
121 	ret = ibdev->ops.query_port(ibdev, port_num, &attr);
122 	if (!ret)
123 		*state = attr.state;
124 	return ret;
125 }
126 
127 static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev,
128 					   struct net_device *ndev,
129 					   struct net_device *upper,
130 					   u32 *port_num)
131 {
132 	struct net_device *rep_ndev;
133 	struct mlx5_ib_port *port;
134 	int i;
135 
136 	for (i = 0; i < dev->num_ports; i++) {
137 		port  = &dev->port[i];
138 		if (!port->rep)
139 			continue;
140 
141 		if (upper == ndev && port->rep->vport == MLX5_VPORT_UPLINK) {
142 			*port_num = i + 1;
143 			return &port->roce;
144 		}
145 
146 		if (upper && port->rep->vport == MLX5_VPORT_UPLINK)
147 			continue;
148 
149 		read_lock(&port->roce.netdev_lock);
150 		rep_ndev = mlx5_ib_get_rep_netdev(port->rep->esw,
151 						  port->rep->vport);
152 		if (rep_ndev == ndev) {
153 			read_unlock(&port->roce.netdev_lock);
154 			*port_num = i + 1;
155 			return &port->roce;
156 		}
157 		read_unlock(&port->roce.netdev_lock);
158 	}
159 
160 	return NULL;
161 }
162 
163 static int mlx5_netdev_event(struct notifier_block *this,
164 			     unsigned long event, void *ptr)
165 {
166 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
167 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
168 	u32 port_num = roce->native_port_num;
169 	struct mlx5_core_dev *mdev;
170 	struct mlx5_ib_dev *ibdev;
171 
172 	ibdev = roce->dev;
173 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
174 	if (!mdev)
175 		return NOTIFY_DONE;
176 
177 	switch (event) {
178 	case NETDEV_REGISTER:
179 		/* Should already be registered during the load */
180 		if (ibdev->is_rep)
181 			break;
182 		write_lock(&roce->netdev_lock);
183 		if (ndev->dev.parent == mdev->device)
184 			roce->netdev = ndev;
185 		write_unlock(&roce->netdev_lock);
186 		break;
187 
188 	case NETDEV_UNREGISTER:
189 		/* In case of reps, ib device goes away before the netdevs */
190 		write_lock(&roce->netdev_lock);
191 		if (roce->netdev == ndev)
192 			roce->netdev = NULL;
193 		write_unlock(&roce->netdev_lock);
194 		break;
195 
196 	case NETDEV_CHANGE:
197 	case NETDEV_UP:
198 	case NETDEV_DOWN: {
199 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
200 		struct net_device *upper = NULL;
201 
202 		if (lag_ndev) {
203 			upper = netdev_master_upper_dev_get(lag_ndev);
204 			dev_put(lag_ndev);
205 		}
206 
207 		if (ibdev->is_rep)
208 			roce = mlx5_get_rep_roce(ibdev, ndev, upper, &port_num);
209 		if (!roce)
210 			return NOTIFY_DONE;
211 		if ((upper == ndev ||
212 		     ((!upper || ibdev->is_rep) && ndev == roce->netdev)) &&
213 		    ibdev->ib_active) {
214 			struct ib_event ibev = { };
215 			enum ib_port_state port_state;
216 
217 			if (get_port_state(&ibdev->ib_dev, port_num,
218 					   &port_state))
219 				goto done;
220 
221 			if (roce->last_port_state == port_state)
222 				goto done;
223 
224 			roce->last_port_state = port_state;
225 			ibev.device = &ibdev->ib_dev;
226 			if (port_state == IB_PORT_DOWN)
227 				ibev.event = IB_EVENT_PORT_ERR;
228 			else if (port_state == IB_PORT_ACTIVE)
229 				ibev.event = IB_EVENT_PORT_ACTIVE;
230 			else
231 				goto done;
232 
233 			ibev.element.port_num = port_num;
234 			ib_dispatch_event(&ibev);
235 		}
236 		break;
237 	}
238 
239 	default:
240 		break;
241 	}
242 done:
243 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
244 	return NOTIFY_DONE;
245 }
246 
247 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
248 					     u32 port_num)
249 {
250 	struct mlx5_ib_dev *ibdev = to_mdev(device);
251 	struct net_device *ndev;
252 	struct mlx5_core_dev *mdev;
253 
254 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
255 	if (!mdev)
256 		return NULL;
257 
258 	ndev = mlx5_lag_get_roce_netdev(mdev);
259 	if (ndev)
260 		goto out;
261 
262 	/* Ensure ndev does not disappear before we invoke dev_hold()
263 	 */
264 	read_lock(&ibdev->port[port_num - 1].roce.netdev_lock);
265 	ndev = ibdev->port[port_num - 1].roce.netdev;
266 	if (ndev)
267 		dev_hold(ndev);
268 	read_unlock(&ibdev->port[port_num - 1].roce.netdev_lock);
269 
270 out:
271 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
272 	return ndev;
273 }
274 
275 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
276 						   u32 ib_port_num,
277 						   u32 *native_port_num)
278 {
279 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
280 							  ib_port_num);
281 	struct mlx5_core_dev *mdev = NULL;
282 	struct mlx5_ib_multiport_info *mpi;
283 	struct mlx5_ib_port *port;
284 
285 	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
286 	    ll != IB_LINK_LAYER_ETHERNET) {
287 		if (native_port_num)
288 			*native_port_num = ib_port_num;
289 		return ibdev->mdev;
290 	}
291 
292 	if (native_port_num)
293 		*native_port_num = 1;
294 
295 	port = &ibdev->port[ib_port_num - 1];
296 	spin_lock(&port->mp.mpi_lock);
297 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
298 	if (mpi && !mpi->unaffiliate) {
299 		mdev = mpi->mdev;
300 		/* If it's the master no need to refcount, it'll exist
301 		 * as long as the ib_dev exists.
302 		 */
303 		if (!mpi->is_master)
304 			mpi->mdev_refcnt++;
305 	}
306 	spin_unlock(&port->mp.mpi_lock);
307 
308 	return mdev;
309 }
310 
311 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u32 port_num)
312 {
313 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
314 							  port_num);
315 	struct mlx5_ib_multiport_info *mpi;
316 	struct mlx5_ib_port *port;
317 
318 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
319 		return;
320 
321 	port = &ibdev->port[port_num - 1];
322 
323 	spin_lock(&port->mp.mpi_lock);
324 	mpi = ibdev->port[port_num - 1].mp.mpi;
325 	if (mpi->is_master)
326 		goto out;
327 
328 	mpi->mdev_refcnt--;
329 	if (mpi->unaffiliate)
330 		complete(&mpi->unref_comp);
331 out:
332 	spin_unlock(&port->mp.mpi_lock);
333 }
334 
335 static int translate_eth_legacy_proto_oper(u32 eth_proto_oper,
336 					   u16 *active_speed, u8 *active_width)
337 {
338 	switch (eth_proto_oper) {
339 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
340 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
341 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
342 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
343 		*active_width = IB_WIDTH_1X;
344 		*active_speed = IB_SPEED_SDR;
345 		break;
346 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
347 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
348 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
349 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
350 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
351 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
352 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
353 		*active_width = IB_WIDTH_1X;
354 		*active_speed = IB_SPEED_QDR;
355 		break;
356 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
357 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
358 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
359 		*active_width = IB_WIDTH_1X;
360 		*active_speed = IB_SPEED_EDR;
361 		break;
362 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
363 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
364 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
365 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
366 		*active_width = IB_WIDTH_4X;
367 		*active_speed = IB_SPEED_QDR;
368 		break;
369 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
370 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
371 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
372 		*active_width = IB_WIDTH_1X;
373 		*active_speed = IB_SPEED_HDR;
374 		break;
375 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
376 		*active_width = IB_WIDTH_4X;
377 		*active_speed = IB_SPEED_FDR;
378 		break;
379 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
380 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
381 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
382 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
383 		*active_width = IB_WIDTH_4X;
384 		*active_speed = IB_SPEED_EDR;
385 		break;
386 	default:
387 		return -EINVAL;
388 	}
389 
390 	return 0;
391 }
392 
393 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u16 *active_speed,
394 					u8 *active_width)
395 {
396 	switch (eth_proto_oper) {
397 	case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
398 	case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
399 		*active_width = IB_WIDTH_1X;
400 		*active_speed = IB_SPEED_SDR;
401 		break;
402 	case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
403 		*active_width = IB_WIDTH_1X;
404 		*active_speed = IB_SPEED_DDR;
405 		break;
406 	case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
407 		*active_width = IB_WIDTH_1X;
408 		*active_speed = IB_SPEED_QDR;
409 		break;
410 	case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
411 		*active_width = IB_WIDTH_4X;
412 		*active_speed = IB_SPEED_QDR;
413 		break;
414 	case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
415 		*active_width = IB_WIDTH_1X;
416 		*active_speed = IB_SPEED_EDR;
417 		break;
418 	case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
419 		*active_width = IB_WIDTH_2X;
420 		*active_speed = IB_SPEED_EDR;
421 		break;
422 	case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
423 		*active_width = IB_WIDTH_1X;
424 		*active_speed = IB_SPEED_HDR;
425 		break;
426 	case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
427 		*active_width = IB_WIDTH_4X;
428 		*active_speed = IB_SPEED_EDR;
429 		break;
430 	case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
431 		*active_width = IB_WIDTH_2X;
432 		*active_speed = IB_SPEED_HDR;
433 		break;
434 	case MLX5E_PROT_MASK(MLX5E_100GAUI_1_100GBASE_CR_KR):
435 		*active_width = IB_WIDTH_1X;
436 		*active_speed = IB_SPEED_NDR;
437 		break;
438 	case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4):
439 		*active_width = IB_WIDTH_4X;
440 		*active_speed = IB_SPEED_HDR;
441 		break;
442 	case MLX5E_PROT_MASK(MLX5E_200GAUI_2_200GBASE_CR2_KR2):
443 		*active_width = IB_WIDTH_2X;
444 		*active_speed = IB_SPEED_NDR;
445 		break;
446 	case MLX5E_PROT_MASK(MLX5E_400GAUI_4_400GBASE_CR4_KR4):
447 		*active_width = IB_WIDTH_4X;
448 		*active_speed = IB_SPEED_NDR;
449 		break;
450 	default:
451 		return -EINVAL;
452 	}
453 
454 	return 0;
455 }
456 
457 static int translate_eth_proto_oper(u32 eth_proto_oper, u16 *active_speed,
458 				    u8 *active_width, bool ext)
459 {
460 	return ext ?
461 		translate_eth_ext_proto_oper(eth_proto_oper, active_speed,
462 					     active_width) :
463 		translate_eth_legacy_proto_oper(eth_proto_oper, active_speed,
464 						active_width);
465 }
466 
467 static int mlx5_query_port_roce(struct ib_device *device, u32 port_num,
468 				struct ib_port_attr *props)
469 {
470 	struct mlx5_ib_dev *dev = to_mdev(device);
471 	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0};
472 	struct mlx5_core_dev *mdev;
473 	struct net_device *ndev, *upper;
474 	enum ib_mtu ndev_ib_mtu;
475 	bool put_mdev = true;
476 	u32 eth_prot_oper;
477 	u32 mdev_port_num;
478 	bool ext;
479 	int err;
480 
481 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
482 	if (!mdev) {
483 		/* This means the port isn't affiliated yet. Get the
484 		 * info for the master port instead.
485 		 */
486 		put_mdev = false;
487 		mdev = dev->mdev;
488 		mdev_port_num = 1;
489 		port_num = 1;
490 	}
491 
492 	/* Possible bad flows are checked before filling out props so in case
493 	 * of an error it will still be zeroed out.
494 	 * Use native port in case of reps
495 	 */
496 	if (dev->is_rep)
497 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
498 					   1);
499 	else
500 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
501 					   mdev_port_num);
502 	if (err)
503 		goto out;
504 	ext = !!MLX5_GET_ETH_PROTO(ptys_reg, out, true, eth_proto_capability);
505 	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
506 
507 	props->active_width     = IB_WIDTH_4X;
508 	props->active_speed     = IB_SPEED_QDR;
509 
510 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
511 				 &props->active_width, ext);
512 
513 	if (!dev->is_rep && dev->mdev->roce.roce_en) {
514 		u16 qkey_viol_cntr;
515 
516 		props->port_cap_flags |= IB_PORT_CM_SUP;
517 		props->ip_gids = true;
518 		props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev,
519 						   roce_address_table_size);
520 		mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
521 		props->qkey_viol_cntr = qkey_viol_cntr;
522 	}
523 	props->max_mtu          = IB_MTU_4096;
524 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
525 	props->pkey_tbl_len     = 1;
526 	props->state            = IB_PORT_DOWN;
527 	props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
528 
529 	/* If this is a stub query for an unaffiliated port stop here */
530 	if (!put_mdev)
531 		goto out;
532 
533 	ndev = mlx5_ib_get_netdev(device, port_num);
534 	if (!ndev)
535 		goto out;
536 
537 	if (dev->lag_active) {
538 		rcu_read_lock();
539 		upper = netdev_master_upper_dev_get_rcu(ndev);
540 		if (upper) {
541 			dev_put(ndev);
542 			ndev = upper;
543 			dev_hold(ndev);
544 		}
545 		rcu_read_unlock();
546 	}
547 
548 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
549 		props->state      = IB_PORT_ACTIVE;
550 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
551 	}
552 
553 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
554 
555 	dev_put(ndev);
556 
557 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
558 out:
559 	if (put_mdev)
560 		mlx5_ib_put_native_port_mdev(dev, port_num);
561 	return err;
562 }
563 
564 static int set_roce_addr(struct mlx5_ib_dev *dev, u32 port_num,
565 			 unsigned int index, const union ib_gid *gid,
566 			 const struct ib_gid_attr *attr)
567 {
568 	enum ib_gid_type gid_type;
569 	u16 vlan_id = 0xffff;
570 	u8 roce_version = 0;
571 	u8 roce_l3_type = 0;
572 	u8 mac[ETH_ALEN];
573 	int ret;
574 
575 	gid_type = attr->gid_type;
576 	if (gid) {
577 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
578 		if (ret)
579 			return ret;
580 	}
581 
582 	switch (gid_type) {
583 	case IB_GID_TYPE_ROCE:
584 		roce_version = MLX5_ROCE_VERSION_1;
585 		break;
586 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
587 		roce_version = MLX5_ROCE_VERSION_2;
588 		if (gid && ipv6_addr_v4mapped((void *)gid))
589 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
590 		else
591 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
592 		break;
593 
594 	default:
595 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
596 	}
597 
598 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
599 				      roce_l3_type, gid->raw, mac,
600 				      vlan_id < VLAN_CFI_MASK, vlan_id,
601 				      port_num);
602 }
603 
604 static int mlx5_ib_add_gid(const struct ib_gid_attr *attr,
605 			   __always_unused void **context)
606 {
607 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
608 			     attr->index, &attr->gid, attr);
609 }
610 
611 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
612 			   __always_unused void **context)
613 {
614 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
615 			     attr->index, NULL, attr);
616 }
617 
618 __be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev,
619 				   const struct ib_gid_attr *attr)
620 {
621 	if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
622 		return 0;
623 
624 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
625 }
626 
627 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
628 {
629 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
630 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
631 	return 0;
632 }
633 
634 enum {
635 	MLX5_VPORT_ACCESS_METHOD_MAD,
636 	MLX5_VPORT_ACCESS_METHOD_HCA,
637 	MLX5_VPORT_ACCESS_METHOD_NIC,
638 };
639 
640 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
641 {
642 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
643 		return MLX5_VPORT_ACCESS_METHOD_MAD;
644 
645 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
646 	    IB_LINK_LAYER_ETHERNET)
647 		return MLX5_VPORT_ACCESS_METHOD_NIC;
648 
649 	return MLX5_VPORT_ACCESS_METHOD_HCA;
650 }
651 
652 static void get_atomic_caps(struct mlx5_ib_dev *dev,
653 			    u8 atomic_size_qp,
654 			    struct ib_device_attr *props)
655 {
656 	u8 tmp;
657 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
658 	u8 atomic_req_8B_endianness_mode =
659 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
660 
661 	/* Check if HW supports 8 bytes standard atomic operations and capable
662 	 * of host endianness respond
663 	 */
664 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
665 	if (((atomic_operations & tmp) == tmp) &&
666 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
667 	    (atomic_req_8B_endianness_mode)) {
668 		props->atomic_cap = IB_ATOMIC_HCA;
669 	} else {
670 		props->atomic_cap = IB_ATOMIC_NONE;
671 	}
672 }
673 
674 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
675 			       struct ib_device_attr *props)
676 {
677 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
678 
679 	get_atomic_caps(dev, atomic_size_qp, props);
680 }
681 
682 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
683 					__be64 *sys_image_guid)
684 {
685 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
686 	struct mlx5_core_dev *mdev = dev->mdev;
687 	u64 tmp;
688 	int err;
689 
690 	switch (mlx5_get_vport_access_method(ibdev)) {
691 	case MLX5_VPORT_ACCESS_METHOD_MAD:
692 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
693 							    sys_image_guid);
694 
695 	case MLX5_VPORT_ACCESS_METHOD_HCA:
696 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
697 		break;
698 
699 	case MLX5_VPORT_ACCESS_METHOD_NIC:
700 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
701 		break;
702 
703 	default:
704 		return -EINVAL;
705 	}
706 
707 	if (!err)
708 		*sys_image_guid = cpu_to_be64(tmp);
709 
710 	return err;
711 
712 }
713 
714 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
715 				u16 *max_pkeys)
716 {
717 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
718 	struct mlx5_core_dev *mdev = dev->mdev;
719 
720 	switch (mlx5_get_vport_access_method(ibdev)) {
721 	case MLX5_VPORT_ACCESS_METHOD_MAD:
722 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
723 
724 	case MLX5_VPORT_ACCESS_METHOD_HCA:
725 	case MLX5_VPORT_ACCESS_METHOD_NIC:
726 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
727 						pkey_table_size));
728 		return 0;
729 
730 	default:
731 		return -EINVAL;
732 	}
733 }
734 
735 static int mlx5_query_vendor_id(struct ib_device *ibdev,
736 				u32 *vendor_id)
737 {
738 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
739 
740 	switch (mlx5_get_vport_access_method(ibdev)) {
741 	case MLX5_VPORT_ACCESS_METHOD_MAD:
742 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
743 
744 	case MLX5_VPORT_ACCESS_METHOD_HCA:
745 	case MLX5_VPORT_ACCESS_METHOD_NIC:
746 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
747 
748 	default:
749 		return -EINVAL;
750 	}
751 }
752 
753 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
754 				__be64 *node_guid)
755 {
756 	u64 tmp;
757 	int err;
758 
759 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
760 	case MLX5_VPORT_ACCESS_METHOD_MAD:
761 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
762 
763 	case MLX5_VPORT_ACCESS_METHOD_HCA:
764 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
765 		break;
766 
767 	case MLX5_VPORT_ACCESS_METHOD_NIC:
768 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
769 		break;
770 
771 	default:
772 		return -EINVAL;
773 	}
774 
775 	if (!err)
776 		*node_guid = cpu_to_be64(tmp);
777 
778 	return err;
779 }
780 
781 struct mlx5_reg_node_desc {
782 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
783 };
784 
785 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
786 {
787 	struct mlx5_reg_node_desc in;
788 
789 	if (mlx5_use_mad_ifc(dev))
790 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
791 
792 	memset(&in, 0, sizeof(in));
793 
794 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
795 				    sizeof(struct mlx5_reg_node_desc),
796 				    MLX5_REG_NODE_DESC, 0, 0);
797 }
798 
799 static int mlx5_ib_query_device(struct ib_device *ibdev,
800 				struct ib_device_attr *props,
801 				struct ib_udata *uhw)
802 {
803 	size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
804 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
805 	struct mlx5_core_dev *mdev = dev->mdev;
806 	int err = -ENOMEM;
807 	int max_sq_desc;
808 	int max_rq_sg;
809 	int max_sq_sg;
810 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
811 	bool raw_support = !mlx5_core_mp_enabled(mdev);
812 	struct mlx5_ib_query_device_resp resp = {};
813 	size_t resp_len;
814 	u64 max_tso;
815 
816 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
817 	if (uhw_outlen && uhw_outlen < resp_len)
818 		return -EINVAL;
819 
820 	resp.response_length = resp_len;
821 
822 	if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
823 		return -EINVAL;
824 
825 	memset(props, 0, sizeof(*props));
826 	err = mlx5_query_system_image_guid(ibdev,
827 					   &props->sys_image_guid);
828 	if (err)
829 		return err;
830 
831 	props->max_pkeys = dev->pkey_table_len;
832 
833 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
834 	if (err)
835 		return err;
836 
837 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
838 		(fw_rev_min(dev->mdev) << 16) |
839 		fw_rev_sub(dev->mdev);
840 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
841 		IB_DEVICE_PORT_ACTIVE_EVENT		|
842 		IB_DEVICE_SYS_IMAGE_GUID		|
843 		IB_DEVICE_RC_RNR_NAK_GEN;
844 
845 	if (MLX5_CAP_GEN(mdev, pkv))
846 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
847 	if (MLX5_CAP_GEN(mdev, qkv))
848 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
849 	if (MLX5_CAP_GEN(mdev, apm))
850 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
851 	if (MLX5_CAP_GEN(mdev, xrc))
852 		props->device_cap_flags |= IB_DEVICE_XRC;
853 	if (MLX5_CAP_GEN(mdev, imaicl)) {
854 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
855 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
856 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
857 		/* We support 'Gappy' memory registration too */
858 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
859 	}
860 	/* IB_WR_REG_MR always requires changing the entity size with UMR */
861 	if (!MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled))
862 		props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
863 	if (MLX5_CAP_GEN(mdev, sho)) {
864 		props->device_cap_flags |= IB_DEVICE_INTEGRITY_HANDOVER;
865 		/* At this stage no support for signature handover */
866 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
867 				      IB_PROT_T10DIF_TYPE_2 |
868 				      IB_PROT_T10DIF_TYPE_3;
869 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
870 				       IB_GUARD_T10DIF_CSUM;
871 	}
872 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
873 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
874 
875 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
876 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
877 			/* Legacy bit to support old userspace libraries */
878 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
879 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
880 		}
881 
882 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
883 			props->raw_packet_caps |=
884 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
885 
886 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) {
887 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
888 			if (max_tso) {
889 				resp.tso_caps.max_tso = 1 << max_tso;
890 				resp.tso_caps.supported_qpts |=
891 					1 << IB_QPT_RAW_PACKET;
892 				resp.response_length += sizeof(resp.tso_caps);
893 			}
894 		}
895 
896 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) {
897 			resp.rss_caps.rx_hash_function =
898 						MLX5_RX_HASH_FUNC_TOEPLITZ;
899 			resp.rss_caps.rx_hash_fields_mask =
900 						MLX5_RX_HASH_SRC_IPV4 |
901 						MLX5_RX_HASH_DST_IPV4 |
902 						MLX5_RX_HASH_SRC_IPV6 |
903 						MLX5_RX_HASH_DST_IPV6 |
904 						MLX5_RX_HASH_SRC_PORT_TCP |
905 						MLX5_RX_HASH_DST_PORT_TCP |
906 						MLX5_RX_HASH_SRC_PORT_UDP |
907 						MLX5_RX_HASH_DST_PORT_UDP |
908 						MLX5_RX_HASH_INNER;
909 			if (mlx5_accel_ipsec_device_caps(dev->mdev) &
910 			    MLX5_ACCEL_IPSEC_CAP_DEVICE)
911 				resp.rss_caps.rx_hash_fields_mask |=
912 					MLX5_RX_HASH_IPSEC_SPI;
913 			resp.response_length += sizeof(resp.rss_caps);
914 		}
915 	} else {
916 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen)
917 			resp.response_length += sizeof(resp.tso_caps);
918 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen)
919 			resp.response_length += sizeof(resp.rss_caps);
920 	}
921 
922 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
923 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
924 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
925 	}
926 
927 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
928 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
929 	    raw_support)
930 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
931 
932 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
933 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
934 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
935 
936 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
937 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
938 	    raw_support) {
939 		/* Legacy bit to support old userspace libraries */
940 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
941 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
942 	}
943 
944 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
945 		props->max_dm_size =
946 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
947 	}
948 
949 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
950 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
951 
952 	if (MLX5_CAP_GEN(mdev, end_pad))
953 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
954 
955 	props->vendor_part_id	   = mdev->pdev->device;
956 	props->hw_ver		   = mdev->pdev->revision;
957 
958 	props->max_mr_size	   = ~0ull;
959 	props->page_size_cap	   = ~(min_page_size - 1);
960 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
961 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
962 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
963 		     sizeof(struct mlx5_wqe_data_seg);
964 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
965 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
966 		     sizeof(struct mlx5_wqe_raddr_seg)) /
967 		sizeof(struct mlx5_wqe_data_seg);
968 	props->max_send_sge = max_sq_sg;
969 	props->max_recv_sge = max_rq_sg;
970 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
971 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
972 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
973 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
974 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
975 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
976 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
977 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
978 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
979 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
980 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
981 	props->max_srq_sge	   = max_rq_sg - 1;
982 	props->max_fast_reg_page_list_len =
983 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
984 	props->max_pi_fast_reg_page_list_len =
985 		props->max_fast_reg_page_list_len / 2;
986 	props->max_sgl_rd =
987 		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
988 	get_atomic_caps_qp(dev, props);
989 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
990 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
991 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
992 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
993 					   props->max_mcast_grp;
994 	props->max_ah = INT_MAX;
995 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
996 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
997 
998 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
999 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
1000 			props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
1001 		props->odp_caps = dev->odp_caps;
1002 		if (!uhw) {
1003 			/* ODP for kernel QPs is not implemented for receive
1004 			 * WQEs and SRQ WQEs
1005 			 */
1006 			props->odp_caps.per_transport_caps.rc_odp_caps &=
1007 				~(IB_ODP_SUPPORT_READ |
1008 				  IB_ODP_SUPPORT_SRQ_RECV);
1009 			props->odp_caps.per_transport_caps.uc_odp_caps &=
1010 				~(IB_ODP_SUPPORT_READ |
1011 				  IB_ODP_SUPPORT_SRQ_RECV);
1012 			props->odp_caps.per_transport_caps.ud_odp_caps &=
1013 				~(IB_ODP_SUPPORT_READ |
1014 				  IB_ODP_SUPPORT_SRQ_RECV);
1015 			props->odp_caps.per_transport_caps.xrc_odp_caps &=
1016 				~(IB_ODP_SUPPORT_READ |
1017 				  IB_ODP_SUPPORT_SRQ_RECV);
1018 		}
1019 	}
1020 
1021 	if (MLX5_CAP_GEN(mdev, cd))
1022 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
1023 
1024 	if (mlx5_core_is_vf(mdev))
1025 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
1026 
1027 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1028 	    IB_LINK_LAYER_ETHERNET && raw_support) {
1029 		props->rss_caps.max_rwq_indirection_tables =
1030 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1031 		props->rss_caps.max_rwq_indirection_table_size =
1032 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1033 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1034 		props->max_wq_type_rq =
1035 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1036 	}
1037 
1038 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1039 		props->tm_caps.max_num_tags =
1040 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1041 		props->tm_caps.max_ops =
1042 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1043 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1044 	}
1045 
1046 	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1047 	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1048 		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1049 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1050 	}
1051 
1052 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1053 		props->cq_caps.max_cq_moderation_count =
1054 						MLX5_MAX_CQ_COUNT;
1055 		props->cq_caps.max_cq_moderation_period =
1056 						MLX5_MAX_CQ_PERIOD;
1057 	}
1058 
1059 	if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) {
1060 		resp.response_length += sizeof(resp.cqe_comp_caps);
1061 
1062 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1063 			resp.cqe_comp_caps.max_num =
1064 				MLX5_CAP_GEN(dev->mdev,
1065 					     cqe_compression_max_num);
1066 
1067 			resp.cqe_comp_caps.supported_format =
1068 				MLX5_IB_CQE_RES_FORMAT_HASH |
1069 				MLX5_IB_CQE_RES_FORMAT_CSUM;
1070 
1071 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1072 				resp.cqe_comp_caps.supported_format |=
1073 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1074 		}
1075 	}
1076 
1077 	if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen &&
1078 	    raw_support) {
1079 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1080 		    MLX5_CAP_GEN(mdev, qos)) {
1081 			resp.packet_pacing_caps.qp_rate_limit_max =
1082 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1083 			resp.packet_pacing_caps.qp_rate_limit_min =
1084 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1085 			resp.packet_pacing_caps.supported_qpts |=
1086 				1 << IB_QPT_RAW_PACKET;
1087 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1088 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1089 				resp.packet_pacing_caps.cap_flags |=
1090 					MLX5_IB_PP_SUPPORT_BURST;
1091 		}
1092 		resp.response_length += sizeof(resp.packet_pacing_caps);
1093 	}
1094 
1095 	if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <=
1096 	    uhw_outlen) {
1097 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1098 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1099 				MLX5_IB_ALLOW_MPW;
1100 
1101 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1102 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1103 				MLX5_IB_SUPPORT_EMPW;
1104 
1105 		resp.response_length +=
1106 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1107 	}
1108 
1109 	if (offsetofend(typeof(resp), flags) <= uhw_outlen) {
1110 		resp.response_length += sizeof(resp.flags);
1111 
1112 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1113 			resp.flags |=
1114 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1115 
1116 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1117 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1118 		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1119 			resp.flags |=
1120 				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1121 
1122 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1123 	}
1124 
1125 	if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) {
1126 		resp.response_length += sizeof(resp.sw_parsing_caps);
1127 		if (MLX5_CAP_ETH(mdev, swp)) {
1128 			resp.sw_parsing_caps.sw_parsing_offloads |=
1129 				MLX5_IB_SW_PARSING;
1130 
1131 			if (MLX5_CAP_ETH(mdev, swp_csum))
1132 				resp.sw_parsing_caps.sw_parsing_offloads |=
1133 					MLX5_IB_SW_PARSING_CSUM;
1134 
1135 			if (MLX5_CAP_ETH(mdev, swp_lso))
1136 				resp.sw_parsing_caps.sw_parsing_offloads |=
1137 					MLX5_IB_SW_PARSING_LSO;
1138 
1139 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1140 				resp.sw_parsing_caps.supported_qpts =
1141 					BIT(IB_QPT_RAW_PACKET);
1142 		}
1143 	}
1144 
1145 	if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen &&
1146 	    raw_support) {
1147 		resp.response_length += sizeof(resp.striding_rq_caps);
1148 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1149 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1150 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1151 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1152 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1153 			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1154 				resp.striding_rq_caps
1155 					.min_single_wqe_log_num_of_strides =
1156 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1157 			else
1158 				resp.striding_rq_caps
1159 					.min_single_wqe_log_num_of_strides =
1160 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1161 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1162 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1163 			resp.striding_rq_caps.supported_qpts =
1164 				BIT(IB_QPT_RAW_PACKET);
1165 		}
1166 	}
1167 
1168 	if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) {
1169 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1170 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1171 			resp.tunnel_offloads_caps |=
1172 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1173 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1174 			resp.tunnel_offloads_caps |=
1175 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1176 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1177 			resp.tunnel_offloads_caps |=
1178 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1179 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre))
1180 			resp.tunnel_offloads_caps |=
1181 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1182 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp))
1183 			resp.tunnel_offloads_caps |=
1184 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1185 	}
1186 
1187 	if (uhw_outlen) {
1188 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1189 
1190 		if (err)
1191 			return err;
1192 	}
1193 
1194 	return 0;
1195 }
1196 
1197 static void translate_active_width(struct ib_device *ibdev, u16 active_width,
1198 				   u8 *ib_width)
1199 {
1200 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1201 
1202 	if (active_width & MLX5_PTYS_WIDTH_1X)
1203 		*ib_width = IB_WIDTH_1X;
1204 	else if (active_width & MLX5_PTYS_WIDTH_2X)
1205 		*ib_width = IB_WIDTH_2X;
1206 	else if (active_width & MLX5_PTYS_WIDTH_4X)
1207 		*ib_width = IB_WIDTH_4X;
1208 	else if (active_width & MLX5_PTYS_WIDTH_8X)
1209 		*ib_width = IB_WIDTH_8X;
1210 	else if (active_width & MLX5_PTYS_WIDTH_12X)
1211 		*ib_width = IB_WIDTH_12X;
1212 	else {
1213 		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1214 			    active_width);
1215 		*ib_width = IB_WIDTH_4X;
1216 	}
1217 
1218 	return;
1219 }
1220 
1221 static int mlx5_mtu_to_ib_mtu(int mtu)
1222 {
1223 	switch (mtu) {
1224 	case 256: return 1;
1225 	case 512: return 2;
1226 	case 1024: return 3;
1227 	case 2048: return 4;
1228 	case 4096: return 5;
1229 	default:
1230 		pr_warn("invalid mtu\n");
1231 		return -1;
1232 	}
1233 }
1234 
1235 enum ib_max_vl_num {
1236 	__IB_MAX_VL_0		= 1,
1237 	__IB_MAX_VL_0_1		= 2,
1238 	__IB_MAX_VL_0_3		= 3,
1239 	__IB_MAX_VL_0_7		= 4,
1240 	__IB_MAX_VL_0_14	= 5,
1241 };
1242 
1243 enum mlx5_vl_hw_cap {
1244 	MLX5_VL_HW_0	= 1,
1245 	MLX5_VL_HW_0_1	= 2,
1246 	MLX5_VL_HW_0_2	= 3,
1247 	MLX5_VL_HW_0_3	= 4,
1248 	MLX5_VL_HW_0_4	= 5,
1249 	MLX5_VL_HW_0_5	= 6,
1250 	MLX5_VL_HW_0_6	= 7,
1251 	MLX5_VL_HW_0_7	= 8,
1252 	MLX5_VL_HW_0_14	= 15
1253 };
1254 
1255 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1256 				u8 *max_vl_num)
1257 {
1258 	switch (vl_hw_cap) {
1259 	case MLX5_VL_HW_0:
1260 		*max_vl_num = __IB_MAX_VL_0;
1261 		break;
1262 	case MLX5_VL_HW_0_1:
1263 		*max_vl_num = __IB_MAX_VL_0_1;
1264 		break;
1265 	case MLX5_VL_HW_0_3:
1266 		*max_vl_num = __IB_MAX_VL_0_3;
1267 		break;
1268 	case MLX5_VL_HW_0_7:
1269 		*max_vl_num = __IB_MAX_VL_0_7;
1270 		break;
1271 	case MLX5_VL_HW_0_14:
1272 		*max_vl_num = __IB_MAX_VL_0_14;
1273 		break;
1274 
1275 	default:
1276 		return -EINVAL;
1277 	}
1278 
1279 	return 0;
1280 }
1281 
1282 static int mlx5_query_hca_port(struct ib_device *ibdev, u32 port,
1283 			       struct ib_port_attr *props)
1284 {
1285 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1286 	struct mlx5_core_dev *mdev = dev->mdev;
1287 	struct mlx5_hca_vport_context *rep;
1288 	u16 max_mtu;
1289 	u16 oper_mtu;
1290 	int err;
1291 	u16 ib_link_width_oper;
1292 	u8 vl_hw_cap;
1293 
1294 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1295 	if (!rep) {
1296 		err = -ENOMEM;
1297 		goto out;
1298 	}
1299 
1300 	/* props being zeroed by the caller, avoid zeroing it here */
1301 
1302 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1303 	if (err)
1304 		goto out;
1305 
1306 	props->lid		= rep->lid;
1307 	props->lmc		= rep->lmc;
1308 	props->sm_lid		= rep->sm_lid;
1309 	props->sm_sl		= rep->sm_sl;
1310 	props->state		= rep->vport_state;
1311 	props->phys_state	= rep->port_physical_state;
1312 	props->port_cap_flags	= rep->cap_mask1;
1313 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1314 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1315 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1316 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1317 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1318 	props->subnet_timeout	= rep->subnet_timeout;
1319 	props->init_type_reply	= rep->init_type_reply;
1320 
1321 	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1322 		props->port_cap_flags2 = rep->cap_mask2;
1323 
1324 	err = mlx5_query_ib_port_oper(mdev, &ib_link_width_oper,
1325 				      &props->active_speed, port);
1326 	if (err)
1327 		goto out;
1328 
1329 	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1330 
1331 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1332 
1333 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1334 
1335 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1336 
1337 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1338 
1339 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1340 	if (err)
1341 		goto out;
1342 
1343 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1344 				   &props->max_vl_num);
1345 out:
1346 	kfree(rep);
1347 	return err;
1348 }
1349 
1350 int mlx5_ib_query_port(struct ib_device *ibdev, u32 port,
1351 		       struct ib_port_attr *props)
1352 {
1353 	unsigned int count;
1354 	int ret;
1355 
1356 	switch (mlx5_get_vport_access_method(ibdev)) {
1357 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1358 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1359 		break;
1360 
1361 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1362 		ret = mlx5_query_hca_port(ibdev, port, props);
1363 		break;
1364 
1365 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1366 		ret = mlx5_query_port_roce(ibdev, port, props);
1367 		break;
1368 
1369 	default:
1370 		ret = -EINVAL;
1371 	}
1372 
1373 	if (!ret && props) {
1374 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1375 		struct mlx5_core_dev *mdev;
1376 		bool put_mdev = true;
1377 
1378 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1379 		if (!mdev) {
1380 			/* If the port isn't affiliated yet query the master.
1381 			 * The master and slave will have the same values.
1382 			 */
1383 			mdev = dev->mdev;
1384 			port = 1;
1385 			put_mdev = false;
1386 		}
1387 		count = mlx5_core_reserved_gids_count(mdev);
1388 		if (put_mdev)
1389 			mlx5_ib_put_native_port_mdev(dev, port);
1390 		props->gid_tbl_len -= count;
1391 	}
1392 	return ret;
1393 }
1394 
1395 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u32 port,
1396 				  struct ib_port_attr *props)
1397 {
1398 	return mlx5_query_port_roce(ibdev, port, props);
1399 }
1400 
1401 static int mlx5_ib_rep_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1402 				  u16 *pkey)
1403 {
1404 	/* Default special Pkey for representor device port as per the
1405 	 * IB specification 1.3 section 10.9.1.2.
1406 	 */
1407 	*pkey = 0xffff;
1408 	return 0;
1409 }
1410 
1411 static int mlx5_ib_query_gid(struct ib_device *ibdev, u32 port, int index,
1412 			     union ib_gid *gid)
1413 {
1414 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1415 	struct mlx5_core_dev *mdev = dev->mdev;
1416 
1417 	switch (mlx5_get_vport_access_method(ibdev)) {
1418 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1419 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1420 
1421 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1422 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1423 
1424 	default:
1425 		return -EINVAL;
1426 	}
1427 
1428 }
1429 
1430 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u32 port,
1431 				   u16 index, u16 *pkey)
1432 {
1433 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1434 	struct mlx5_core_dev *mdev;
1435 	bool put_mdev = true;
1436 	u32 mdev_port_num;
1437 	int err;
1438 
1439 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1440 	if (!mdev) {
1441 		/* The port isn't affiliated yet, get the PKey from the master
1442 		 * port. For RoCE the PKey tables will be the same.
1443 		 */
1444 		put_mdev = false;
1445 		mdev = dev->mdev;
1446 		mdev_port_num = 1;
1447 	}
1448 
1449 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1450 					index, pkey);
1451 	if (put_mdev)
1452 		mlx5_ib_put_native_port_mdev(dev, port);
1453 
1454 	return err;
1455 }
1456 
1457 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1458 			      u16 *pkey)
1459 {
1460 	switch (mlx5_get_vport_access_method(ibdev)) {
1461 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1462 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1463 
1464 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1465 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1466 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1467 	default:
1468 		return -EINVAL;
1469 	}
1470 }
1471 
1472 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1473 				 struct ib_device_modify *props)
1474 {
1475 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1476 	struct mlx5_reg_node_desc in;
1477 	struct mlx5_reg_node_desc out;
1478 	int err;
1479 
1480 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1481 		return -EOPNOTSUPP;
1482 
1483 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1484 		return 0;
1485 
1486 	/*
1487 	 * If possible, pass node desc to FW, so it can generate
1488 	 * a 144 trap.  If cmd fails, just ignore.
1489 	 */
1490 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1491 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1492 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1493 	if (err)
1494 		return err;
1495 
1496 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1497 
1498 	return err;
1499 }
1500 
1501 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u32 port_num, u32 mask,
1502 				u32 value)
1503 {
1504 	struct mlx5_hca_vport_context ctx = {};
1505 	struct mlx5_core_dev *mdev;
1506 	u32 mdev_port_num;
1507 	int err;
1508 
1509 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1510 	if (!mdev)
1511 		return -ENODEV;
1512 
1513 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1514 	if (err)
1515 		goto out;
1516 
1517 	if (~ctx.cap_mask1_perm & mask) {
1518 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1519 			     mask, ctx.cap_mask1_perm);
1520 		err = -EINVAL;
1521 		goto out;
1522 	}
1523 
1524 	ctx.cap_mask1 = value;
1525 	ctx.cap_mask1_perm = mask;
1526 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1527 						 0, &ctx);
1528 
1529 out:
1530 	mlx5_ib_put_native_port_mdev(dev, port_num);
1531 
1532 	return err;
1533 }
1534 
1535 static int mlx5_ib_modify_port(struct ib_device *ibdev, u32 port, int mask,
1536 			       struct ib_port_modify *props)
1537 {
1538 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1539 	struct ib_port_attr attr;
1540 	u32 tmp;
1541 	int err;
1542 	u32 change_mask;
1543 	u32 value;
1544 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1545 		      IB_LINK_LAYER_INFINIBAND);
1546 
1547 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1548 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1549 	 */
1550 	if (!is_ib)
1551 		return 0;
1552 
1553 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1554 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1555 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1556 		return set_port_caps_atomic(dev, port, change_mask, value);
1557 	}
1558 
1559 	mutex_lock(&dev->cap_mask_mutex);
1560 
1561 	err = ib_query_port(ibdev, port, &attr);
1562 	if (err)
1563 		goto out;
1564 
1565 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1566 		~props->clr_port_cap_mask;
1567 
1568 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1569 
1570 out:
1571 	mutex_unlock(&dev->cap_mask_mutex);
1572 	return err;
1573 }
1574 
1575 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1576 {
1577 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1578 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1579 }
1580 
1581 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1582 {
1583 	/* Large page with non 4k uar support might limit the dynamic size */
1584 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1585 		return MLX5_MIN_DYN_BFREGS;
1586 
1587 	return MLX5_MAX_DYN_BFREGS;
1588 }
1589 
1590 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1591 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1592 			     struct mlx5_bfreg_info *bfregi)
1593 {
1594 	int uars_per_sys_page;
1595 	int bfregs_per_sys_page;
1596 	int ref_bfregs = req->total_num_bfregs;
1597 
1598 	if (req->total_num_bfregs == 0)
1599 		return -EINVAL;
1600 
1601 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1602 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1603 
1604 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1605 		return -ENOMEM;
1606 
1607 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1608 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1609 	/* This holds the required static allocation asked by the user */
1610 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1611 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1612 		return -EINVAL;
1613 
1614 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1615 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1616 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1617 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1618 
1619 	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",
1620 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1621 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1622 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1623 		    bfregi->num_sys_pages);
1624 
1625 	return 0;
1626 }
1627 
1628 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1629 {
1630 	struct mlx5_bfreg_info *bfregi;
1631 	int err;
1632 	int i;
1633 
1634 	bfregi = &context->bfregi;
1635 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1636 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1637 		if (err)
1638 			goto error;
1639 
1640 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1641 	}
1642 
1643 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1644 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1645 
1646 	return 0;
1647 
1648 error:
1649 	for (--i; i >= 0; i--)
1650 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1651 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1652 
1653 	return err;
1654 }
1655 
1656 static void deallocate_uars(struct mlx5_ib_dev *dev,
1657 			    struct mlx5_ib_ucontext *context)
1658 {
1659 	struct mlx5_bfreg_info *bfregi;
1660 	int i;
1661 
1662 	bfregi = &context->bfregi;
1663 	for (i = 0; i < bfregi->num_sys_pages; i++)
1664 		if (i < bfregi->num_static_sys_pages ||
1665 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1666 			mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1667 }
1668 
1669 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1670 {
1671 	int err = 0;
1672 
1673 	mutex_lock(&dev->lb.mutex);
1674 	if (td)
1675 		dev->lb.user_td++;
1676 	if (qp)
1677 		dev->lb.qps++;
1678 
1679 	if (dev->lb.user_td == 2 ||
1680 	    dev->lb.qps == 1) {
1681 		if (!dev->lb.enabled) {
1682 			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1683 			dev->lb.enabled = true;
1684 		}
1685 	}
1686 
1687 	mutex_unlock(&dev->lb.mutex);
1688 
1689 	return err;
1690 }
1691 
1692 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1693 {
1694 	mutex_lock(&dev->lb.mutex);
1695 	if (td)
1696 		dev->lb.user_td--;
1697 	if (qp)
1698 		dev->lb.qps--;
1699 
1700 	if (dev->lb.user_td == 1 &&
1701 	    dev->lb.qps == 0) {
1702 		if (dev->lb.enabled) {
1703 			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1704 			dev->lb.enabled = false;
1705 		}
1706 	}
1707 
1708 	mutex_unlock(&dev->lb.mutex);
1709 }
1710 
1711 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1712 					  u16 uid)
1713 {
1714 	int err;
1715 
1716 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1717 		return 0;
1718 
1719 	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1720 	if (err)
1721 		return err;
1722 
1723 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1724 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1725 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1726 		return err;
1727 
1728 	return mlx5_ib_enable_lb(dev, true, false);
1729 }
1730 
1731 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1732 					     u16 uid)
1733 {
1734 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1735 		return;
1736 
1737 	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1738 
1739 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1740 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1741 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1742 		return;
1743 
1744 	mlx5_ib_disable_lb(dev, true, false);
1745 }
1746 
1747 static int set_ucontext_resp(struct ib_ucontext *uctx,
1748 			     struct mlx5_ib_alloc_ucontext_resp *resp)
1749 {
1750 	struct ib_device *ibdev = uctx->device;
1751 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1752 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1753 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1754 	int err;
1755 
1756 	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1757 		err = mlx5_cmd_dump_fill_mkey(dev->mdev,
1758 					      &resp->dump_fill_mkey);
1759 		if (err)
1760 			return err;
1761 		resp->comp_mask |=
1762 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1763 	}
1764 
1765 	resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1766 	if (dev->wc_support)
1767 		resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev,
1768 						      log_bf_reg_size);
1769 	resp->cache_line_size = cache_line_size();
1770 	resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1771 	resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1772 	resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1773 	resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1774 	resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1775 	resp->cqe_version = context->cqe_version;
1776 	resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1777 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1778 	resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1779 					MLX5_CAP_GEN(dev->mdev,
1780 						     num_of_uars_per_page) : 1;
1781 
1782 	if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1783 				MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1784 		if (mlx5_get_flow_namespace(dev->mdev,
1785 				MLX5_FLOW_NAMESPACE_EGRESS))
1786 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1787 		if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1788 				MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1789 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1790 		if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1791 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1792 		if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1793 				MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1794 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1795 		/* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1796 	}
1797 
1798 	resp->tot_bfregs = bfregi->lib_uar_dyn ? 0 :
1799 			bfregi->total_num_bfregs - bfregi->num_dyn_bfregs;
1800 	resp->num_ports = dev->num_ports;
1801 	resp->cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1802 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1803 
1804 	if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1805 		mlx5_query_min_inline(dev->mdev, &resp->eth_min_inline);
1806 		resp->eth_min_inline++;
1807 	}
1808 
1809 	if (dev->mdev->clock_info)
1810 		resp->clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1811 
1812 	/*
1813 	 * We don't want to expose information from the PCI bar that is located
1814 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1815 	 * pretend we don't support reading the HCA's core clock. This is also
1816 	 * forced by mmap function.
1817 	 */
1818 	if (PAGE_SIZE <= 4096) {
1819 		resp->comp_mask |=
1820 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1821 		resp->hca_core_clock_offset =
1822 			offsetof(struct mlx5_init_seg,
1823 				 internal_timer_h) % PAGE_SIZE;
1824 	}
1825 
1826 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
1827 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE;
1828 
1829 	if (rt_supported(MLX5_CAP_GEN(dev->mdev, sq_ts_format)) &&
1830 	    rt_supported(MLX5_CAP_GEN(dev->mdev, rq_ts_format)) &&
1831 	    rt_supported(MLX5_CAP_ROCE(dev->mdev, qp_ts_format)))
1832 		resp->comp_mask |=
1833 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_REAL_TIME_TS;
1834 
1835 	resp->num_dyn_bfregs = bfregi->num_dyn_bfregs;
1836 
1837 	if (MLX5_CAP_GEN(dev->mdev, drain_sigerr))
1838 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_SQD2RTS;
1839 
1840 	return 0;
1841 }
1842 
1843 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1844 				  struct ib_udata *udata)
1845 {
1846 	struct ib_device *ibdev = uctx->device;
1847 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1848 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1849 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1850 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1851 	struct mlx5_bfreg_info *bfregi;
1852 	int ver;
1853 	int err;
1854 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1855 				     max_cqe_version);
1856 	bool lib_uar_4k;
1857 	bool lib_uar_dyn;
1858 
1859 	if (!dev->ib_active)
1860 		return -EAGAIN;
1861 
1862 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1863 		ver = 0;
1864 	else if (udata->inlen >= min_req_v2)
1865 		ver = 2;
1866 	else
1867 		return -EINVAL;
1868 
1869 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1870 	if (err)
1871 		return err;
1872 
1873 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1874 		return -EOPNOTSUPP;
1875 
1876 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1877 		return -EOPNOTSUPP;
1878 
1879 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1880 				    MLX5_NON_FP_BFREGS_PER_UAR);
1881 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1882 		return -EINVAL;
1883 
1884 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1885 	lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
1886 	bfregi = &context->bfregi;
1887 
1888 	if (lib_uar_dyn) {
1889 		bfregi->lib_uar_dyn = lib_uar_dyn;
1890 		goto uar_done;
1891 	}
1892 
1893 	/* updates req->total_num_bfregs */
1894 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1895 	if (err)
1896 		goto out_ctx;
1897 
1898 	mutex_init(&bfregi->lock);
1899 	bfregi->lib_uar_4k = lib_uar_4k;
1900 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1901 				GFP_KERNEL);
1902 	if (!bfregi->count) {
1903 		err = -ENOMEM;
1904 		goto out_ctx;
1905 	}
1906 
1907 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1908 				    sizeof(*bfregi->sys_pages),
1909 				    GFP_KERNEL);
1910 	if (!bfregi->sys_pages) {
1911 		err = -ENOMEM;
1912 		goto out_count;
1913 	}
1914 
1915 	err = allocate_uars(dev, context);
1916 	if (err)
1917 		goto out_sys_pages;
1918 
1919 uar_done:
1920 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1921 		err = mlx5_ib_devx_create(dev, true);
1922 		if (err < 0)
1923 			goto out_uars;
1924 		context->devx_uid = err;
1925 	}
1926 
1927 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1928 					     context->devx_uid);
1929 	if (err)
1930 		goto out_devx;
1931 
1932 	INIT_LIST_HEAD(&context->db_page_list);
1933 	mutex_init(&context->db_page_mutex);
1934 
1935 	context->cqe_version = min_t(__u8,
1936 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1937 				 req.max_cqe_version);
1938 
1939 	err = set_ucontext_resp(uctx, &resp);
1940 	if (err)
1941 		goto out_mdev;
1942 
1943 	resp.response_length = min(udata->outlen, sizeof(resp));
1944 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1945 	if (err)
1946 		goto out_mdev;
1947 
1948 	bfregi->ver = ver;
1949 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1950 	context->lib_caps = req.lib_caps;
1951 	print_lib_caps(dev, context->lib_caps);
1952 
1953 	if (mlx5_ib_lag_should_assign_affinity(dev)) {
1954 		u32 port = mlx5_core_native_port_num(dev->mdev) - 1;
1955 
1956 		atomic_set(&context->tx_port_affinity,
1957 			   atomic_add_return(
1958 				   1, &dev->port[port].roce.tx_port_affinity));
1959 	}
1960 
1961 	return 0;
1962 
1963 out_mdev:
1964 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1965 out_devx:
1966 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1967 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1968 
1969 out_uars:
1970 	deallocate_uars(dev, context);
1971 
1972 out_sys_pages:
1973 	kfree(bfregi->sys_pages);
1974 
1975 out_count:
1976 	kfree(bfregi->count);
1977 
1978 out_ctx:
1979 	return err;
1980 }
1981 
1982 static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext,
1983 				  struct uverbs_attr_bundle *attrs)
1984 {
1985 	struct mlx5_ib_alloc_ucontext_resp uctx_resp = {};
1986 	int ret;
1987 
1988 	ret = set_ucontext_resp(ibcontext, &uctx_resp);
1989 	if (ret)
1990 		return ret;
1991 
1992 	uctx_resp.response_length =
1993 		min_t(size_t,
1994 		      uverbs_attr_get_len(attrs,
1995 				MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX),
1996 		      sizeof(uctx_resp));
1997 
1998 	ret = uverbs_copy_to_struct_or_zero(attrs,
1999 					MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
2000 					&uctx_resp,
2001 					sizeof(uctx_resp));
2002 	return ret;
2003 }
2004 
2005 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
2006 {
2007 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2008 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2009 	struct mlx5_bfreg_info *bfregi;
2010 
2011 	bfregi = &context->bfregi;
2012 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
2013 
2014 	if (context->devx_uid)
2015 		mlx5_ib_devx_destroy(dev, context->devx_uid);
2016 
2017 	deallocate_uars(dev, context);
2018 	kfree(bfregi->sys_pages);
2019 	kfree(bfregi->count);
2020 }
2021 
2022 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2023 				 int uar_idx)
2024 {
2025 	int fw_uars_per_page;
2026 
2027 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2028 
2029 	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2030 }
2031 
2032 static u64 uar_index2paddress(struct mlx5_ib_dev *dev,
2033 				 int uar_idx)
2034 {
2035 	unsigned int fw_uars_per_page;
2036 
2037 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
2038 				MLX5_UARS_IN_PAGE : 1;
2039 
2040 	return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE);
2041 }
2042 
2043 static int get_command(unsigned long offset)
2044 {
2045 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2046 }
2047 
2048 static int get_arg(unsigned long offset)
2049 {
2050 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2051 }
2052 
2053 static int get_index(unsigned long offset)
2054 {
2055 	return get_arg(offset);
2056 }
2057 
2058 /* Index resides in an extra byte to enable larger values than 255 */
2059 static int get_extended_index(unsigned long offset)
2060 {
2061 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2062 }
2063 
2064 
2065 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2066 {
2067 }
2068 
2069 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2070 {
2071 	switch (cmd) {
2072 	case MLX5_IB_MMAP_WC_PAGE:
2073 		return "WC";
2074 	case MLX5_IB_MMAP_REGULAR_PAGE:
2075 		return "best effort WC";
2076 	case MLX5_IB_MMAP_NC_PAGE:
2077 		return "NC";
2078 	case MLX5_IB_MMAP_DEVICE_MEM:
2079 		return "Device Memory";
2080 	default:
2081 		return NULL;
2082 	}
2083 }
2084 
2085 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2086 					struct vm_area_struct *vma,
2087 					struct mlx5_ib_ucontext *context)
2088 {
2089 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2090 	    !(vma->vm_flags & VM_SHARED))
2091 		return -EINVAL;
2092 
2093 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2094 		return -EOPNOTSUPP;
2095 
2096 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2097 		return -EPERM;
2098 	vma->vm_flags &= ~VM_MAYWRITE;
2099 
2100 	if (!dev->mdev->clock_info)
2101 		return -EOPNOTSUPP;
2102 
2103 	return vm_insert_page(vma, vma->vm_start,
2104 			      virt_to_page(dev->mdev->clock_info));
2105 }
2106 
2107 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2108 {
2109 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2110 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2111 	struct mlx5_var_table *var_table = &dev->var_table;
2112 
2113 	switch (mentry->mmap_flag) {
2114 	case MLX5_IB_MMAP_TYPE_MEMIC:
2115 	case MLX5_IB_MMAP_TYPE_MEMIC_OP:
2116 		mlx5_ib_dm_mmap_free(dev, mentry);
2117 		break;
2118 	case MLX5_IB_MMAP_TYPE_VAR:
2119 		mutex_lock(&var_table->bitmap_lock);
2120 		clear_bit(mentry->page_idx, var_table->bitmap);
2121 		mutex_unlock(&var_table->bitmap_lock);
2122 		kfree(mentry);
2123 		break;
2124 	case MLX5_IB_MMAP_TYPE_UAR_WC:
2125 	case MLX5_IB_MMAP_TYPE_UAR_NC:
2126 		mlx5_cmd_free_uar(dev->mdev, mentry->page_idx);
2127 		kfree(mentry);
2128 		break;
2129 	default:
2130 		WARN_ON(true);
2131 	}
2132 }
2133 
2134 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2135 		    struct vm_area_struct *vma,
2136 		    struct mlx5_ib_ucontext *context)
2137 {
2138 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2139 	int err;
2140 	unsigned long idx;
2141 	phys_addr_t pfn;
2142 	pgprot_t prot;
2143 	u32 bfreg_dyn_idx = 0;
2144 	u32 uar_index;
2145 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2146 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2147 				bfregi->num_static_sys_pages;
2148 
2149 	if (bfregi->lib_uar_dyn)
2150 		return -EINVAL;
2151 
2152 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2153 		return -EINVAL;
2154 
2155 	if (dyn_uar)
2156 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2157 	else
2158 		idx = get_index(vma->vm_pgoff);
2159 
2160 	if (idx >= max_valid_idx) {
2161 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2162 			     idx, max_valid_idx);
2163 		return -EINVAL;
2164 	}
2165 
2166 	switch (cmd) {
2167 	case MLX5_IB_MMAP_WC_PAGE:
2168 	case MLX5_IB_MMAP_ALLOC_WC:
2169 	case MLX5_IB_MMAP_REGULAR_PAGE:
2170 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2171 		prot = pgprot_writecombine(vma->vm_page_prot);
2172 		break;
2173 	case MLX5_IB_MMAP_NC_PAGE:
2174 		prot = pgprot_noncached(vma->vm_page_prot);
2175 		break;
2176 	default:
2177 		return -EINVAL;
2178 	}
2179 
2180 	if (dyn_uar) {
2181 		int uars_per_page;
2182 
2183 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2184 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2185 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2186 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2187 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2188 			return -EINVAL;
2189 		}
2190 
2191 		mutex_lock(&bfregi->lock);
2192 		/* Fail if uar already allocated, first bfreg index of each
2193 		 * page holds its count.
2194 		 */
2195 		if (bfregi->count[bfreg_dyn_idx]) {
2196 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2197 			mutex_unlock(&bfregi->lock);
2198 			return -EINVAL;
2199 		}
2200 
2201 		bfregi->count[bfreg_dyn_idx]++;
2202 		mutex_unlock(&bfregi->lock);
2203 
2204 		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2205 		if (err) {
2206 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2207 			goto free_bfreg;
2208 		}
2209 	} else {
2210 		uar_index = bfregi->sys_pages[idx];
2211 	}
2212 
2213 	pfn = uar_index2pfn(dev, uar_index);
2214 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2215 
2216 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2217 				prot, NULL);
2218 	if (err) {
2219 		mlx5_ib_err(dev,
2220 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2221 			    err, mmap_cmd2str(cmd));
2222 		goto err;
2223 	}
2224 
2225 	if (dyn_uar)
2226 		bfregi->sys_pages[idx] = uar_index;
2227 	return 0;
2228 
2229 err:
2230 	if (!dyn_uar)
2231 		return err;
2232 
2233 	mlx5_cmd_free_uar(dev->mdev, idx);
2234 
2235 free_bfreg:
2236 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2237 
2238 	return err;
2239 }
2240 
2241 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2242 {
2243 	unsigned long idx;
2244 	u8 command;
2245 
2246 	command = get_command(vma->vm_pgoff);
2247 	idx = get_extended_index(vma->vm_pgoff);
2248 
2249 	return (command << 16 | idx);
2250 }
2251 
2252 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2253 			       struct vm_area_struct *vma,
2254 			       struct ib_ucontext *ucontext)
2255 {
2256 	struct mlx5_user_mmap_entry *mentry;
2257 	struct rdma_user_mmap_entry *entry;
2258 	unsigned long pgoff;
2259 	pgprot_t prot;
2260 	phys_addr_t pfn;
2261 	int ret;
2262 
2263 	pgoff = mlx5_vma_to_pgoff(vma);
2264 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2265 	if (!entry)
2266 		return -EINVAL;
2267 
2268 	mentry = to_mmmap(entry);
2269 	pfn = (mentry->address >> PAGE_SHIFT);
2270 	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
2271 	    mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
2272 		prot = pgprot_noncached(vma->vm_page_prot);
2273 	else
2274 		prot = pgprot_writecombine(vma->vm_page_prot);
2275 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2276 				entry->npages * PAGE_SIZE,
2277 				prot,
2278 				entry);
2279 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2280 	return ret;
2281 }
2282 
2283 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2284 {
2285 	u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2286 	u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2287 
2288 	return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2289 		(index & 0xFF)) << PAGE_SHIFT;
2290 }
2291 
2292 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2293 {
2294 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2295 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2296 	unsigned long command;
2297 	phys_addr_t pfn;
2298 
2299 	command = get_command(vma->vm_pgoff);
2300 	switch (command) {
2301 	case MLX5_IB_MMAP_WC_PAGE:
2302 	case MLX5_IB_MMAP_ALLOC_WC:
2303 		if (!dev->wc_support)
2304 			return -EPERM;
2305 		fallthrough;
2306 	case MLX5_IB_MMAP_NC_PAGE:
2307 	case MLX5_IB_MMAP_REGULAR_PAGE:
2308 		return uar_mmap(dev, command, vma, context);
2309 
2310 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2311 		return -ENOSYS;
2312 
2313 	case MLX5_IB_MMAP_CORE_CLOCK:
2314 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2315 			return -EINVAL;
2316 
2317 		if (vma->vm_flags & VM_WRITE)
2318 			return -EPERM;
2319 		vma->vm_flags &= ~VM_MAYWRITE;
2320 
2321 		/* Don't expose to user-space information it shouldn't have */
2322 		if (PAGE_SIZE > 4096)
2323 			return -EOPNOTSUPP;
2324 
2325 		pfn = (dev->mdev->iseg_base +
2326 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2327 			PAGE_SHIFT;
2328 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2329 					 PAGE_SIZE,
2330 					 pgprot_noncached(vma->vm_page_prot),
2331 					 NULL);
2332 	case MLX5_IB_MMAP_CLOCK_INFO:
2333 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2334 
2335 	default:
2336 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2337 	}
2338 
2339 	return 0;
2340 }
2341 
2342 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2343 {
2344 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2345 	struct ib_device *ibdev = ibpd->device;
2346 	struct mlx5_ib_alloc_pd_resp resp;
2347 	int err;
2348 	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2349 	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {};
2350 	u16 uid = 0;
2351 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2352 		udata, struct mlx5_ib_ucontext, ibucontext);
2353 
2354 	uid = context ? context->devx_uid : 0;
2355 	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2356 	MLX5_SET(alloc_pd_in, in, uid, uid);
2357 	err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out);
2358 	if (err)
2359 		return err;
2360 
2361 	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2362 	pd->uid = uid;
2363 	if (udata) {
2364 		resp.pdn = pd->pdn;
2365 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2366 			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2367 			return -EFAULT;
2368 		}
2369 	}
2370 
2371 	return 0;
2372 }
2373 
2374 static int mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2375 {
2376 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2377 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2378 
2379 	return mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2380 }
2381 
2382 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2383 {
2384 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2385 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
2386 	int err;
2387 	u16 uid;
2388 
2389 	uid = ibqp->pd ?
2390 		to_mpd(ibqp->pd)->uid : 0;
2391 
2392 	if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) {
2393 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
2394 		return -EOPNOTSUPP;
2395 	}
2396 
2397 	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2398 	if (err)
2399 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2400 			     ibqp->qp_num, gid->raw);
2401 
2402 	return err;
2403 }
2404 
2405 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2406 {
2407 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2408 	int err;
2409 	u16 uid;
2410 
2411 	uid = ibqp->pd ?
2412 		to_mpd(ibqp->pd)->uid : 0;
2413 	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2414 	if (err)
2415 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2416 			     ibqp->qp_num, gid->raw);
2417 
2418 	return err;
2419 }
2420 
2421 static int init_node_data(struct mlx5_ib_dev *dev)
2422 {
2423 	int err;
2424 
2425 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2426 	if (err)
2427 		return err;
2428 
2429 	dev->mdev->rev_id = dev->mdev->pdev->revision;
2430 
2431 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2432 }
2433 
2434 static ssize_t fw_pages_show(struct device *device,
2435 			     struct device_attribute *attr, char *buf)
2436 {
2437 	struct mlx5_ib_dev *dev =
2438 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2439 
2440 	return sysfs_emit(buf, "%d\n", dev->mdev->priv.fw_pages);
2441 }
2442 static DEVICE_ATTR_RO(fw_pages);
2443 
2444 static ssize_t reg_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", atomic_read(&dev->mdev->priv.reg_pages));
2451 }
2452 static DEVICE_ATTR_RO(reg_pages);
2453 
2454 static ssize_t hca_type_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, "MT%d\n", dev->mdev->pdev->device);
2461 }
2462 static DEVICE_ATTR_RO(hca_type);
2463 
2464 static ssize_t hw_rev_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, "%x\n", dev->mdev->rev_id);
2471 }
2472 static DEVICE_ATTR_RO(hw_rev);
2473 
2474 static ssize_t board_id_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, "%.*s\n", MLX5_BOARD_ID_LEN,
2481 			  dev->mdev->board_id);
2482 }
2483 static DEVICE_ATTR_RO(board_id);
2484 
2485 static struct attribute *mlx5_class_attributes[] = {
2486 	&dev_attr_hw_rev.attr,
2487 	&dev_attr_hca_type.attr,
2488 	&dev_attr_board_id.attr,
2489 	&dev_attr_fw_pages.attr,
2490 	&dev_attr_reg_pages.attr,
2491 	NULL,
2492 };
2493 
2494 static const struct attribute_group mlx5_attr_group = {
2495 	.attrs = mlx5_class_attributes,
2496 };
2497 
2498 static void pkey_change_handler(struct work_struct *work)
2499 {
2500 	struct mlx5_ib_port_resources *ports =
2501 		container_of(work, struct mlx5_ib_port_resources,
2502 			     pkey_change_work);
2503 
2504 	mlx5_ib_gsi_pkey_change(ports->gsi);
2505 }
2506 
2507 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2508 {
2509 	struct mlx5_ib_qp *mqp;
2510 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2511 	struct mlx5_core_cq *mcq;
2512 	struct list_head cq_armed_list;
2513 	unsigned long flags_qp;
2514 	unsigned long flags_cq;
2515 	unsigned long flags;
2516 
2517 	INIT_LIST_HEAD(&cq_armed_list);
2518 
2519 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2520 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2521 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2522 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2523 		if (mqp->sq.tail != mqp->sq.head) {
2524 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2525 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2526 			if (send_mcq->mcq.comp &&
2527 			    mqp->ibqp.send_cq->comp_handler) {
2528 				if (!send_mcq->mcq.reset_notify_added) {
2529 					send_mcq->mcq.reset_notify_added = 1;
2530 					list_add_tail(&send_mcq->mcq.reset_notify,
2531 						      &cq_armed_list);
2532 				}
2533 			}
2534 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2535 		}
2536 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2537 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2538 		/* no handling is needed for SRQ */
2539 		if (!mqp->ibqp.srq) {
2540 			if (mqp->rq.tail != mqp->rq.head) {
2541 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2542 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2543 				if (recv_mcq->mcq.comp &&
2544 				    mqp->ibqp.recv_cq->comp_handler) {
2545 					if (!recv_mcq->mcq.reset_notify_added) {
2546 						recv_mcq->mcq.reset_notify_added = 1;
2547 						list_add_tail(&recv_mcq->mcq.reset_notify,
2548 							      &cq_armed_list);
2549 					}
2550 				}
2551 				spin_unlock_irqrestore(&recv_mcq->lock,
2552 						       flags_cq);
2553 			}
2554 		}
2555 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2556 	}
2557 	/*At that point all inflight post send were put to be executed as of we
2558 	 * lock/unlock above locks Now need to arm all involved CQs.
2559 	 */
2560 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2561 		mcq->comp(mcq, NULL);
2562 	}
2563 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2564 }
2565 
2566 static void delay_drop_handler(struct work_struct *work)
2567 {
2568 	int err;
2569 	struct mlx5_ib_delay_drop *delay_drop =
2570 		container_of(work, struct mlx5_ib_delay_drop,
2571 			     delay_drop_work);
2572 
2573 	atomic_inc(&delay_drop->events_cnt);
2574 
2575 	mutex_lock(&delay_drop->lock);
2576 	err = mlx5_core_set_delay_drop(delay_drop->dev, delay_drop->timeout);
2577 	if (err) {
2578 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
2579 			     delay_drop->timeout);
2580 		delay_drop->activate = false;
2581 	}
2582 	mutex_unlock(&delay_drop->lock);
2583 }
2584 
2585 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2586 				 struct ib_event *ibev)
2587 {
2588 	u32 port = (eqe->data.port.port >> 4) & 0xf;
2589 
2590 	switch (eqe->sub_type) {
2591 	case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
2592 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2593 					    IB_LINK_LAYER_ETHERNET)
2594 			schedule_work(&ibdev->delay_drop.delay_drop_work);
2595 		break;
2596 	default: /* do nothing */
2597 		return;
2598 	}
2599 }
2600 
2601 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2602 			      struct ib_event *ibev)
2603 {
2604 	u32 port = (eqe->data.port.port >> 4) & 0xf;
2605 
2606 	ibev->element.port_num = port;
2607 
2608 	switch (eqe->sub_type) {
2609 	case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
2610 	case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
2611 	case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
2612 		/* In RoCE, port up/down events are handled in
2613 		 * mlx5_netdev_event().
2614 		 */
2615 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2616 					    IB_LINK_LAYER_ETHERNET)
2617 			return -EINVAL;
2618 
2619 		ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
2620 				IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2621 		break;
2622 
2623 	case MLX5_PORT_CHANGE_SUBTYPE_LID:
2624 		ibev->event = IB_EVENT_LID_CHANGE;
2625 		break;
2626 
2627 	case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
2628 		ibev->event = IB_EVENT_PKEY_CHANGE;
2629 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2630 		break;
2631 
2632 	case MLX5_PORT_CHANGE_SUBTYPE_GUID:
2633 		ibev->event = IB_EVENT_GID_CHANGE;
2634 		break;
2635 
2636 	case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
2637 		ibev->event = IB_EVENT_CLIENT_REREGISTER;
2638 		break;
2639 	default:
2640 		return -EINVAL;
2641 	}
2642 
2643 	return 0;
2644 }
2645 
2646 static void mlx5_ib_handle_event(struct work_struct *_work)
2647 {
2648 	struct mlx5_ib_event_work *work =
2649 		container_of(_work, struct mlx5_ib_event_work, work);
2650 	struct mlx5_ib_dev *ibdev;
2651 	struct ib_event ibev;
2652 	bool fatal = false;
2653 
2654 	if (work->is_slave) {
2655 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
2656 		if (!ibdev)
2657 			goto out;
2658 	} else {
2659 		ibdev = work->dev;
2660 	}
2661 
2662 	switch (work->event) {
2663 	case MLX5_DEV_EVENT_SYS_ERROR:
2664 		ibev.event = IB_EVENT_DEVICE_FATAL;
2665 		mlx5_ib_handle_internal_error(ibdev);
2666 		ibev.element.port_num  = (u8)(unsigned long)work->param;
2667 		fatal = true;
2668 		break;
2669 	case MLX5_EVENT_TYPE_PORT_CHANGE:
2670 		if (handle_port_change(ibdev, work->param, &ibev))
2671 			goto out;
2672 		break;
2673 	case MLX5_EVENT_TYPE_GENERAL_EVENT:
2674 		handle_general_event(ibdev, work->param, &ibev);
2675 		fallthrough;
2676 	default:
2677 		goto out;
2678 	}
2679 
2680 	ibev.device = &ibdev->ib_dev;
2681 
2682 	if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
2683 		mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
2684 		goto out;
2685 	}
2686 
2687 	if (ibdev->ib_active)
2688 		ib_dispatch_event(&ibev);
2689 
2690 	if (fatal)
2691 		ibdev->ib_active = false;
2692 out:
2693 	kfree(work);
2694 }
2695 
2696 static int mlx5_ib_event(struct notifier_block *nb,
2697 			 unsigned long event, void *param)
2698 {
2699 	struct mlx5_ib_event_work *work;
2700 
2701 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2702 	if (!work)
2703 		return NOTIFY_DONE;
2704 
2705 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2706 	work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
2707 	work->is_slave = false;
2708 	work->param = param;
2709 	work->event = event;
2710 
2711 	queue_work(mlx5_ib_event_wq, &work->work);
2712 
2713 	return NOTIFY_OK;
2714 }
2715 
2716 static int mlx5_ib_event_slave_port(struct notifier_block *nb,
2717 				    unsigned long event, void *param)
2718 {
2719 	struct mlx5_ib_event_work *work;
2720 
2721 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2722 	if (!work)
2723 		return NOTIFY_DONE;
2724 
2725 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2726 	work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
2727 	work->is_slave = true;
2728 	work->param = param;
2729 	work->event = event;
2730 	queue_work(mlx5_ib_event_wq, &work->work);
2731 
2732 	return NOTIFY_OK;
2733 }
2734 
2735 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2736 {
2737 	struct mlx5_hca_vport_context vport_ctx;
2738 	int err;
2739 	int port;
2740 
2741 	for (port = 1; port <= ARRAY_SIZE(dev->port_caps); port++) {
2742 		dev->port_caps[port - 1].has_smi = false;
2743 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
2744 		    MLX5_CAP_PORT_TYPE_IB) {
2745 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2746 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
2747 								   port, 0,
2748 								   &vport_ctx);
2749 				if (err) {
2750 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2751 						    port, err);
2752 					return err;
2753 				}
2754 				dev->port_caps[port - 1].has_smi =
2755 					vport_ctx.has_smi;
2756 			} else {
2757 				dev->port_caps[port - 1].has_smi = true;
2758 			}
2759 		}
2760 	}
2761 	return 0;
2762 }
2763 
2764 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2765 {
2766 	unsigned int port;
2767 
2768 	rdma_for_each_port (&dev->ib_dev, port)
2769 		mlx5_query_ext_port_caps(dev, port);
2770 }
2771 
2772 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
2773 {
2774 	switch (umr_fence_cap) {
2775 	case MLX5_CAP_UMR_FENCE_NONE:
2776 		return MLX5_FENCE_MODE_NONE;
2777 	case MLX5_CAP_UMR_FENCE_SMALL:
2778 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
2779 	default:
2780 		return MLX5_FENCE_MODE_STRONG_ORDERING;
2781 	}
2782 }
2783 
2784 static int mlx5_ib_dev_res_init(struct mlx5_ib_dev *dev)
2785 {
2786 	struct mlx5_ib_resources *devr = &dev->devr;
2787 	struct ib_srq_init_attr attr;
2788 	struct ib_device *ibdev;
2789 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2790 	int port;
2791 	int ret = 0;
2792 
2793 	ibdev = &dev->ib_dev;
2794 
2795 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
2796 		return -EOPNOTSUPP;
2797 
2798 	mutex_init(&devr->mutex);
2799 
2800 	devr->p0 = rdma_zalloc_drv_obj(ibdev, ib_pd);
2801 	if (!devr->p0)
2802 		return -ENOMEM;
2803 
2804 	devr->p0->device  = ibdev;
2805 	devr->p0->uobject = NULL;
2806 	atomic_set(&devr->p0->usecnt, 0);
2807 
2808 	ret = mlx5_ib_alloc_pd(devr->p0, NULL);
2809 	if (ret)
2810 		goto error0;
2811 
2812 	devr->c0 = rdma_zalloc_drv_obj(ibdev, ib_cq);
2813 	if (!devr->c0) {
2814 		ret = -ENOMEM;
2815 		goto error1;
2816 	}
2817 
2818 	devr->c0->device = &dev->ib_dev;
2819 	atomic_set(&devr->c0->usecnt, 0);
2820 
2821 	ret = mlx5_ib_create_cq(devr->c0, &cq_attr, NULL);
2822 	if (ret)
2823 		goto err_create_cq;
2824 
2825 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0);
2826 	if (ret)
2827 		goto error2;
2828 
2829 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0);
2830 	if (ret)
2831 		goto error3;
2832 
2833 	memset(&attr, 0, sizeof(attr));
2834 	attr.attr.max_sge = 1;
2835 	attr.attr.max_wr = 1;
2836 	attr.srq_type = IB_SRQT_XRC;
2837 	attr.ext.cq = devr->c0;
2838 
2839 	devr->s0 = rdma_zalloc_drv_obj(ibdev, ib_srq);
2840 	if (!devr->s0) {
2841 		ret = -ENOMEM;
2842 		goto error4;
2843 	}
2844 
2845 	devr->s0->device	= &dev->ib_dev;
2846 	devr->s0->pd		= devr->p0;
2847 	devr->s0->srq_type      = IB_SRQT_XRC;
2848 	devr->s0->ext.cq	= devr->c0;
2849 	ret = mlx5_ib_create_srq(devr->s0, &attr, NULL);
2850 	if (ret)
2851 		goto err_create;
2852 
2853 	atomic_inc(&devr->s0->ext.cq->usecnt);
2854 	atomic_inc(&devr->p0->usecnt);
2855 	atomic_set(&devr->s0->usecnt, 0);
2856 
2857 	memset(&attr, 0, sizeof(attr));
2858 	attr.attr.max_sge = 1;
2859 	attr.attr.max_wr = 1;
2860 	attr.srq_type = IB_SRQT_BASIC;
2861 	devr->s1 = rdma_zalloc_drv_obj(ibdev, ib_srq);
2862 	if (!devr->s1) {
2863 		ret = -ENOMEM;
2864 		goto error5;
2865 	}
2866 
2867 	devr->s1->device	= &dev->ib_dev;
2868 	devr->s1->pd		= devr->p0;
2869 	devr->s1->srq_type      = IB_SRQT_BASIC;
2870 	devr->s1->ext.cq	= devr->c0;
2871 
2872 	ret = mlx5_ib_create_srq(devr->s1, &attr, NULL);
2873 	if (ret)
2874 		goto error6;
2875 
2876 	atomic_inc(&devr->p0->usecnt);
2877 	atomic_set(&devr->s1->usecnt, 0);
2878 
2879 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2880 		INIT_WORK(&devr->ports[port].pkey_change_work,
2881 			  pkey_change_handler);
2882 
2883 	return 0;
2884 
2885 error6:
2886 	kfree(devr->s1);
2887 error5:
2888 	mlx5_ib_destroy_srq(devr->s0, NULL);
2889 err_create:
2890 	kfree(devr->s0);
2891 error4:
2892 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2893 error3:
2894 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2895 error2:
2896 	mlx5_ib_destroy_cq(devr->c0, NULL);
2897 err_create_cq:
2898 	kfree(devr->c0);
2899 error1:
2900 	mlx5_ib_dealloc_pd(devr->p0, NULL);
2901 error0:
2902 	kfree(devr->p0);
2903 	return ret;
2904 }
2905 
2906 static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev)
2907 {
2908 	struct mlx5_ib_resources *devr = &dev->devr;
2909 	int port;
2910 
2911 	mlx5_ib_destroy_srq(devr->s1, NULL);
2912 	kfree(devr->s1);
2913 	mlx5_ib_destroy_srq(devr->s0, NULL);
2914 	kfree(devr->s0);
2915 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2916 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2917 	mlx5_ib_destroy_cq(devr->c0, NULL);
2918 	kfree(devr->c0);
2919 	mlx5_ib_dealloc_pd(devr->p0, NULL);
2920 	kfree(devr->p0);
2921 
2922 	/* Make sure no change P_Key work items are still executing */
2923 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2924 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2925 }
2926 
2927 static u32 get_core_cap_flags(struct ib_device *ibdev,
2928 			      struct mlx5_hca_vport_context *rep)
2929 {
2930 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2931 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2932 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2933 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2934 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
2935 	u32 ret = 0;
2936 
2937 	if (rep->grh_required)
2938 		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
2939 
2940 	if (ll == IB_LINK_LAYER_INFINIBAND)
2941 		return ret | RDMA_CORE_PORT_IBA_IB;
2942 
2943 	if (raw_support)
2944 		ret |= RDMA_CORE_PORT_RAW_PACKET;
2945 
2946 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2947 		return ret;
2948 
2949 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2950 		return ret;
2951 
2952 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2953 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2954 
2955 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2956 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2957 
2958 	return ret;
2959 }
2960 
2961 static int mlx5_port_immutable(struct ib_device *ibdev, u32 port_num,
2962 			       struct ib_port_immutable *immutable)
2963 {
2964 	struct ib_port_attr attr;
2965 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2966 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
2967 	struct mlx5_hca_vport_context rep = {0};
2968 	int err;
2969 
2970 	err = ib_query_port(ibdev, port_num, &attr);
2971 	if (err)
2972 		return err;
2973 
2974 	if (ll == IB_LINK_LAYER_INFINIBAND) {
2975 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
2976 						   &rep);
2977 		if (err)
2978 			return err;
2979 	}
2980 
2981 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2982 	immutable->gid_tbl_len = attr.gid_tbl_len;
2983 	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
2984 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2985 
2986 	return 0;
2987 }
2988 
2989 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u32 port_num,
2990 				   struct ib_port_immutable *immutable)
2991 {
2992 	struct ib_port_attr attr;
2993 	int err;
2994 
2995 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2996 
2997 	err = ib_query_port(ibdev, port_num, &attr);
2998 	if (err)
2999 		return err;
3000 
3001 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
3002 	immutable->gid_tbl_len = attr.gid_tbl_len;
3003 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
3004 
3005 	return 0;
3006 }
3007 
3008 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
3009 {
3010 	struct mlx5_ib_dev *dev =
3011 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
3012 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
3013 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
3014 		 fw_rev_sub(dev->mdev));
3015 }
3016 
3017 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
3018 {
3019 	struct mlx5_core_dev *mdev = dev->mdev;
3020 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
3021 								 MLX5_FLOW_NAMESPACE_LAG);
3022 	struct mlx5_flow_table *ft;
3023 	int err;
3024 
3025 	if (!ns || !mlx5_lag_is_active(mdev))
3026 		return 0;
3027 
3028 	err = mlx5_cmd_create_vport_lag(mdev);
3029 	if (err)
3030 		return err;
3031 
3032 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3033 	if (IS_ERR(ft)) {
3034 		err = PTR_ERR(ft);
3035 		goto err_destroy_vport_lag;
3036 	}
3037 
3038 	dev->flow_db->lag_demux_ft = ft;
3039 	dev->lag_active = true;
3040 	return 0;
3041 
3042 err_destroy_vport_lag:
3043 	mlx5_cmd_destroy_vport_lag(mdev);
3044 	return err;
3045 }
3046 
3047 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3048 {
3049 	struct mlx5_core_dev *mdev = dev->mdev;
3050 
3051 	if (dev->lag_active) {
3052 		dev->lag_active = false;
3053 
3054 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
3055 		dev->flow_db->lag_demux_ft = NULL;
3056 
3057 		mlx5_cmd_destroy_vport_lag(mdev);
3058 	}
3059 }
3060 
3061 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u32 port_num)
3062 {
3063 	int err;
3064 
3065 	dev->port[port_num].roce.nb.notifier_call = mlx5_netdev_event;
3066 	err = register_netdevice_notifier(&dev->port[port_num].roce.nb);
3067 	if (err) {
3068 		dev->port[port_num].roce.nb.notifier_call = NULL;
3069 		return err;
3070 	}
3071 
3072 	return 0;
3073 }
3074 
3075 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u32 port_num)
3076 {
3077 	if (dev->port[port_num].roce.nb.notifier_call) {
3078 		unregister_netdevice_notifier(&dev->port[port_num].roce.nb);
3079 		dev->port[port_num].roce.nb.notifier_call = NULL;
3080 	}
3081 }
3082 
3083 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3084 {
3085 	int err;
3086 
3087 	if (!dev->is_rep && dev->profile != &raw_eth_profile) {
3088 		err = mlx5_nic_vport_enable_roce(dev->mdev);
3089 		if (err)
3090 			return err;
3091 	}
3092 
3093 	err = mlx5_eth_lag_init(dev);
3094 	if (err)
3095 		goto err_disable_roce;
3096 
3097 	return 0;
3098 
3099 err_disable_roce:
3100 	if (!dev->is_rep && dev->profile != &raw_eth_profile)
3101 		mlx5_nic_vport_disable_roce(dev->mdev);
3102 
3103 	return err;
3104 }
3105 
3106 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3107 {
3108 	mlx5_eth_lag_cleanup(dev);
3109 	if (!dev->is_rep && dev->profile != &raw_eth_profile)
3110 		mlx5_nic_vport_disable_roce(dev->mdev);
3111 }
3112 
3113 static int mlx5_ib_rn_get_params(struct ib_device *device, u32 port_num,
3114 				 enum rdma_netdev_t type,
3115 				 struct rdma_netdev_alloc_params *params)
3116 {
3117 	if (type != RDMA_NETDEV_IPOIB)
3118 		return -EOPNOTSUPP;
3119 
3120 	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
3121 }
3122 
3123 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
3124 				       size_t count, loff_t *pos)
3125 {
3126 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3127 	char lbuf[20];
3128 	int len;
3129 
3130 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
3131 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
3132 }
3133 
3134 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
3135 					size_t count, loff_t *pos)
3136 {
3137 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3138 	u32 timeout;
3139 	u32 var;
3140 
3141 	if (kstrtouint_from_user(buf, count, 0, &var))
3142 		return -EFAULT;
3143 
3144 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
3145 			1000);
3146 	if (timeout != var)
3147 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
3148 			    timeout);
3149 
3150 	delay_drop->timeout = timeout;
3151 
3152 	return count;
3153 }
3154 
3155 static const struct file_operations fops_delay_drop_timeout = {
3156 	.owner	= THIS_MODULE,
3157 	.open	= simple_open,
3158 	.write	= delay_drop_timeout_write,
3159 	.read	= delay_drop_timeout_read,
3160 };
3161 
3162 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
3163 				      struct mlx5_ib_multiport_info *mpi)
3164 {
3165 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3166 	struct mlx5_ib_port *port = &ibdev->port[port_num];
3167 	int comps;
3168 	int err;
3169 	int i;
3170 
3171 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3172 
3173 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
3174 
3175 	spin_lock(&port->mp.mpi_lock);
3176 	if (!mpi->ibdev) {
3177 		spin_unlock(&port->mp.mpi_lock);
3178 		return;
3179 	}
3180 
3181 	mpi->ibdev = NULL;
3182 
3183 	spin_unlock(&port->mp.mpi_lock);
3184 	if (mpi->mdev_events.notifier_call)
3185 		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
3186 	mpi->mdev_events.notifier_call = NULL;
3187 	mlx5_remove_netdev_notifier(ibdev, port_num);
3188 	spin_lock(&port->mp.mpi_lock);
3189 
3190 	comps = mpi->mdev_refcnt;
3191 	if (comps) {
3192 		mpi->unaffiliate = true;
3193 		init_completion(&mpi->unref_comp);
3194 		spin_unlock(&port->mp.mpi_lock);
3195 
3196 		for (i = 0; i < comps; i++)
3197 			wait_for_completion(&mpi->unref_comp);
3198 
3199 		spin_lock(&port->mp.mpi_lock);
3200 		mpi->unaffiliate = false;
3201 	}
3202 
3203 	port->mp.mpi = NULL;
3204 
3205 	spin_unlock(&port->mp.mpi_lock);
3206 
3207 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
3208 
3209 	mlx5_ib_dbg(ibdev, "unaffiliated port %u\n", port_num + 1);
3210 	/* Log an error, still needed to cleanup the pointers and add
3211 	 * it back to the list.
3212 	 */
3213 	if (err)
3214 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
3215 			    port_num + 1);
3216 
3217 	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
3218 }
3219 
3220 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
3221 				    struct mlx5_ib_multiport_info *mpi)
3222 {
3223 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3224 	int err;
3225 
3226 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3227 
3228 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
3229 	if (ibdev->port[port_num].mp.mpi) {
3230 		mlx5_ib_dbg(ibdev, "port %u already affiliated.\n",
3231 			    port_num + 1);
3232 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3233 		return false;
3234 	}
3235 
3236 	ibdev->port[port_num].mp.mpi = mpi;
3237 	mpi->ibdev = ibdev;
3238 	mpi->mdev_events.notifier_call = NULL;
3239 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3240 
3241 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
3242 	if (err)
3243 		goto unbind;
3244 
3245 	err = mlx5_add_netdev_notifier(ibdev, port_num);
3246 	if (err) {
3247 		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
3248 			    port_num + 1);
3249 		goto unbind;
3250 	}
3251 
3252 	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
3253 	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
3254 
3255 	mlx5_ib_init_cong_debugfs(ibdev, port_num);
3256 
3257 	return true;
3258 
3259 unbind:
3260 	mlx5_ib_unbind_slave_port(ibdev, mpi);
3261 	return false;
3262 }
3263 
3264 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
3265 {
3266 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3267 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3268 							  port_num + 1);
3269 	struct mlx5_ib_multiport_info *mpi;
3270 	int err;
3271 	u32 i;
3272 
3273 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3274 		return 0;
3275 
3276 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
3277 						     &dev->sys_image_guid);
3278 	if (err)
3279 		return err;
3280 
3281 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3282 	if (err)
3283 		return err;
3284 
3285 	mutex_lock(&mlx5_ib_multiport_mutex);
3286 	for (i = 0; i < dev->num_ports; i++) {
3287 		bool bound = false;
3288 
3289 		/* build a stub multiport info struct for the native port. */
3290 		if (i == port_num) {
3291 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
3292 			if (!mpi) {
3293 				mutex_unlock(&mlx5_ib_multiport_mutex);
3294 				mlx5_nic_vport_disable_roce(dev->mdev);
3295 				return -ENOMEM;
3296 			}
3297 
3298 			mpi->is_master = true;
3299 			mpi->mdev = dev->mdev;
3300 			mpi->sys_image_guid = dev->sys_image_guid;
3301 			dev->port[i].mp.mpi = mpi;
3302 			mpi->ibdev = dev;
3303 			mpi = NULL;
3304 			continue;
3305 		}
3306 
3307 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
3308 				    list) {
3309 			if (dev->sys_image_guid == mpi->sys_image_guid &&
3310 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
3311 				bound = mlx5_ib_bind_slave_port(dev, mpi);
3312 			}
3313 
3314 			if (bound) {
3315 				dev_dbg(mpi->mdev->device,
3316 					"removing port from unaffiliated list.\n");
3317 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
3318 				list_del(&mpi->list);
3319 				break;
3320 			}
3321 		}
3322 		if (!bound)
3323 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
3324 				    i + 1);
3325 	}
3326 
3327 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
3328 	mutex_unlock(&mlx5_ib_multiport_mutex);
3329 	return err;
3330 }
3331 
3332 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
3333 {
3334 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3335 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3336 							  port_num + 1);
3337 	u32 i;
3338 
3339 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3340 		return;
3341 
3342 	mutex_lock(&mlx5_ib_multiport_mutex);
3343 	for (i = 0; i < dev->num_ports; i++) {
3344 		if (dev->port[i].mp.mpi) {
3345 			/* Destroy the native port stub */
3346 			if (i == port_num) {
3347 				kfree(dev->port[i].mp.mpi);
3348 				dev->port[i].mp.mpi = NULL;
3349 			} else {
3350 				mlx5_ib_dbg(dev, "unbinding port_num: %u\n",
3351 					    i + 1);
3352 				list_add_tail(&dev->port[i].mp.mpi->list,
3353 					      &mlx5_ib_unaffiliated_port_list);
3354 				mlx5_ib_unbind_slave_port(dev,
3355 							  dev->port[i].mp.mpi);
3356 			}
3357 		}
3358 	}
3359 
3360 	mlx5_ib_dbg(dev, "removing from devlist\n");
3361 	list_del(&dev->ib_dev_list);
3362 	mutex_unlock(&mlx5_ib_multiport_mutex);
3363 
3364 	mlx5_nic_vport_disable_roce(dev->mdev);
3365 }
3366 
3367 static int mmap_obj_cleanup(struct ib_uobject *uobject,
3368 			    enum rdma_remove_reason why,
3369 			    struct uverbs_attr_bundle *attrs)
3370 {
3371 	struct mlx5_user_mmap_entry *obj = uobject->object;
3372 
3373 	rdma_user_mmap_entry_remove(&obj->rdma_entry);
3374 	return 0;
3375 }
3376 
3377 static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c,
3378 					    struct mlx5_user_mmap_entry *entry,
3379 					    size_t length)
3380 {
3381 	return rdma_user_mmap_entry_insert_range(
3382 		&c->ibucontext, &entry->rdma_entry, length,
3383 		(MLX5_IB_MMAP_OFFSET_START << 16),
3384 		((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1));
3385 }
3386 
3387 static struct mlx5_user_mmap_entry *
3388 alloc_var_entry(struct mlx5_ib_ucontext *c)
3389 {
3390 	struct mlx5_user_mmap_entry *entry;
3391 	struct mlx5_var_table *var_table;
3392 	u32 page_idx;
3393 	int err;
3394 
3395 	var_table = &to_mdev(c->ibucontext.device)->var_table;
3396 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3397 	if (!entry)
3398 		return ERR_PTR(-ENOMEM);
3399 
3400 	mutex_lock(&var_table->bitmap_lock);
3401 	page_idx = find_first_zero_bit(var_table->bitmap,
3402 				       var_table->num_var_hw_entries);
3403 	if (page_idx >= var_table->num_var_hw_entries) {
3404 		err = -ENOSPC;
3405 		mutex_unlock(&var_table->bitmap_lock);
3406 		goto end;
3407 	}
3408 
3409 	set_bit(page_idx, var_table->bitmap);
3410 	mutex_unlock(&var_table->bitmap_lock);
3411 
3412 	entry->address = var_table->hw_start_addr +
3413 				(page_idx * var_table->stride_size);
3414 	entry->page_idx = page_idx;
3415 	entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
3416 
3417 	err = mlx5_rdma_user_mmap_entry_insert(c, entry,
3418 					       var_table->stride_size);
3419 	if (err)
3420 		goto err_insert;
3421 
3422 	return entry;
3423 
3424 err_insert:
3425 	mutex_lock(&var_table->bitmap_lock);
3426 	clear_bit(page_idx, var_table->bitmap);
3427 	mutex_unlock(&var_table->bitmap_lock);
3428 end:
3429 	kfree(entry);
3430 	return ERR_PTR(err);
3431 }
3432 
3433 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
3434 	struct uverbs_attr_bundle *attrs)
3435 {
3436 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3437 		attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3438 	struct mlx5_ib_ucontext *c;
3439 	struct mlx5_user_mmap_entry *entry;
3440 	u64 mmap_offset;
3441 	u32 length;
3442 	int err;
3443 
3444 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3445 	if (IS_ERR(c))
3446 		return PTR_ERR(c);
3447 
3448 	entry = alloc_var_entry(c);
3449 	if (IS_ERR(entry))
3450 		return PTR_ERR(entry);
3451 
3452 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3453 	length = entry->rdma_entry.npages * PAGE_SIZE;
3454 	uobj->object = entry;
3455 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3456 
3457 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3458 			     &mmap_offset, sizeof(mmap_offset));
3459 	if (err)
3460 		return err;
3461 
3462 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3463 			     &entry->page_idx, sizeof(entry->page_idx));
3464 	if (err)
3465 		return err;
3466 
3467 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3468 			     &length, sizeof(length));
3469 	return err;
3470 }
3471 
3472 DECLARE_UVERBS_NAMED_METHOD(
3473 	MLX5_IB_METHOD_VAR_OBJ_ALLOC,
3474 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
3475 			MLX5_IB_OBJECT_VAR,
3476 			UVERBS_ACCESS_NEW,
3477 			UA_MANDATORY),
3478 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3479 			   UVERBS_ATTR_TYPE(u32),
3480 			   UA_MANDATORY),
3481 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3482 			   UVERBS_ATTR_TYPE(u32),
3483 			   UA_MANDATORY),
3484 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3485 			    UVERBS_ATTR_TYPE(u64),
3486 			    UA_MANDATORY));
3487 
3488 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3489 	MLX5_IB_METHOD_VAR_OBJ_DESTROY,
3490 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
3491 			MLX5_IB_OBJECT_VAR,
3492 			UVERBS_ACCESS_DESTROY,
3493 			UA_MANDATORY));
3494 
3495 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
3496 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3497 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
3498 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
3499 
3500 static bool var_is_supported(struct ib_device *device)
3501 {
3502 	struct mlx5_ib_dev *dev = to_mdev(device);
3503 
3504 	return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3505 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
3506 }
3507 
3508 static struct mlx5_user_mmap_entry *
3509 alloc_uar_entry(struct mlx5_ib_ucontext *c,
3510 		enum mlx5_ib_uapi_uar_alloc_type alloc_type)
3511 {
3512 	struct mlx5_user_mmap_entry *entry;
3513 	struct mlx5_ib_dev *dev;
3514 	u32 uar_index;
3515 	int err;
3516 
3517 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3518 	if (!entry)
3519 		return ERR_PTR(-ENOMEM);
3520 
3521 	dev = to_mdev(c->ibucontext.device);
3522 	err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
3523 	if (err)
3524 		goto end;
3525 
3526 	entry->page_idx = uar_index;
3527 	entry->address = uar_index2paddress(dev, uar_index);
3528 	if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3529 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC;
3530 	else
3531 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC;
3532 
3533 	err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE);
3534 	if (err)
3535 		goto err_insert;
3536 
3537 	return entry;
3538 
3539 err_insert:
3540 	mlx5_cmd_free_uar(dev->mdev, uar_index);
3541 end:
3542 	kfree(entry);
3543 	return ERR_PTR(err);
3544 }
3545 
3546 static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)(
3547 	struct uverbs_attr_bundle *attrs)
3548 {
3549 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3550 		attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3551 	enum mlx5_ib_uapi_uar_alloc_type alloc_type;
3552 	struct mlx5_ib_ucontext *c;
3553 	struct mlx5_user_mmap_entry *entry;
3554 	u64 mmap_offset;
3555 	u32 length;
3556 	int err;
3557 
3558 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3559 	if (IS_ERR(c))
3560 		return PTR_ERR(c);
3561 
3562 	err = uverbs_get_const(&alloc_type, attrs,
3563 			       MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE);
3564 	if (err)
3565 		return err;
3566 
3567 	if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF &&
3568 	    alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC)
3569 		return -EOPNOTSUPP;
3570 
3571 	if (!to_mdev(c->ibucontext.device)->wc_support &&
3572 	    alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3573 		return -EOPNOTSUPP;
3574 
3575 	entry = alloc_uar_entry(c, alloc_type);
3576 	if (IS_ERR(entry))
3577 		return PTR_ERR(entry);
3578 
3579 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3580 	length = entry->rdma_entry.npages * PAGE_SIZE;
3581 	uobj->object = entry;
3582 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3583 
3584 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3585 			     &mmap_offset, sizeof(mmap_offset));
3586 	if (err)
3587 		return err;
3588 
3589 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3590 			     &entry->page_idx, sizeof(entry->page_idx));
3591 	if (err)
3592 		return err;
3593 
3594 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3595 			     &length, sizeof(length));
3596 	return err;
3597 }
3598 
3599 DECLARE_UVERBS_NAMED_METHOD(
3600 	MLX5_IB_METHOD_UAR_OBJ_ALLOC,
3601 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE,
3602 			MLX5_IB_OBJECT_UAR,
3603 			UVERBS_ACCESS_NEW,
3604 			UA_MANDATORY),
3605 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE,
3606 			     enum mlx5_ib_uapi_uar_alloc_type,
3607 			     UA_MANDATORY),
3608 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3609 			   UVERBS_ATTR_TYPE(u32),
3610 			   UA_MANDATORY),
3611 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3612 			   UVERBS_ATTR_TYPE(u32),
3613 			   UA_MANDATORY),
3614 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3615 			    UVERBS_ATTR_TYPE(u64),
3616 			    UA_MANDATORY));
3617 
3618 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3619 	MLX5_IB_METHOD_UAR_OBJ_DESTROY,
3620 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE,
3621 			MLX5_IB_OBJECT_UAR,
3622 			UVERBS_ACCESS_DESTROY,
3623 			UA_MANDATORY));
3624 
3625 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR,
3626 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3627 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC),
3628 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY));
3629 
3630 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3631 	mlx5_ib_flow_action,
3632 	UVERBS_OBJECT_FLOW_ACTION,
3633 	UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
3634 	UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
3635 			     enum mlx5_ib_uapi_flow_action_flags));
3636 
3637 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3638 	mlx5_ib_query_context,
3639 	UVERBS_OBJECT_DEVICE,
3640 	UVERBS_METHOD_QUERY_CONTEXT,
3641 	UVERBS_ATTR_PTR_OUT(
3642 		MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
3643 		UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp,
3644 				   dump_fill_mkey),
3645 		UA_MANDATORY));
3646 
3647 static const struct uapi_definition mlx5_ib_defs[] = {
3648 	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
3649 	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
3650 	UAPI_DEF_CHAIN(mlx5_ib_qos_defs),
3651 	UAPI_DEF_CHAIN(mlx5_ib_std_types_defs),
3652 	UAPI_DEF_CHAIN(mlx5_ib_dm_defs),
3653 
3654 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_FLOW_ACTION,
3655 				&mlx5_ib_flow_action),
3656 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context),
3657 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
3658 				UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
3659 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR),
3660 	{}
3661 };
3662 
3663 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
3664 {
3665 	mlx5_ib_cleanup_multiport_master(dev);
3666 	WARN_ON(!xa_empty(&dev->odp_mkeys));
3667 	mutex_destroy(&dev->cap_mask_mutex);
3668 	WARN_ON(!xa_empty(&dev->sig_mrs));
3669 	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
3670 }
3671 
3672 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
3673 {
3674 	struct mlx5_core_dev *mdev = dev->mdev;
3675 	int err;
3676 	int i;
3677 
3678 	dev->ib_dev.node_type = RDMA_NODE_IB_CA;
3679 	dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
3680 	dev->ib_dev.phys_port_cnt = dev->num_ports;
3681 	dev->ib_dev.dev.parent = mdev->device;
3682 	dev->ib_dev.lag_flags = RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
3683 
3684 	for (i = 0; i < dev->num_ports; i++) {
3685 		spin_lock_init(&dev->port[i].mp.mpi_lock);
3686 		rwlock_init(&dev->port[i].roce.netdev_lock);
3687 		dev->port[i].roce.dev = dev;
3688 		dev->port[i].roce.native_port_num = i + 1;
3689 		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
3690 	}
3691 
3692 	err = mlx5_ib_init_multiport_master(dev);
3693 	if (err)
3694 		return err;
3695 
3696 	err = set_has_smi_cap(dev);
3697 	if (err)
3698 		goto err_mp;
3699 
3700 	err = mlx5_query_max_pkeys(&dev->ib_dev, &dev->pkey_table_len);
3701 	if (err)
3702 		goto err_mp;
3703 
3704 	if (mlx5_use_mad_ifc(dev))
3705 		get_ext_port_caps(dev);
3706 
3707 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_count(mdev);
3708 
3709 	mutex_init(&dev->cap_mask_mutex);
3710 	INIT_LIST_HEAD(&dev->qp_list);
3711 	spin_lock_init(&dev->reset_flow_resource_lock);
3712 	xa_init(&dev->odp_mkeys);
3713 	xa_init(&dev->sig_mrs);
3714 	atomic_set(&dev->mkey_var, 0);
3715 
3716 	spin_lock_init(&dev->dm.lock);
3717 	dev->dm.dev = mdev;
3718 	return 0;
3719 
3720 err_mp:
3721 	mlx5_ib_cleanup_multiport_master(dev);
3722 	return err;
3723 }
3724 
3725 static int mlx5_ib_enable_driver(struct ib_device *dev)
3726 {
3727 	struct mlx5_ib_dev *mdev = to_mdev(dev);
3728 	int ret;
3729 
3730 	ret = mlx5_ib_test_wc(mdev);
3731 	mlx5_ib_dbg(mdev, "Write-Combining %s",
3732 		    mdev->wc_support ? "supported" : "not supported");
3733 
3734 	return ret;
3735 }
3736 
3737 static const struct ib_device_ops mlx5_ib_dev_ops = {
3738 	.owner = THIS_MODULE,
3739 	.driver_id = RDMA_DRIVER_MLX5,
3740 	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
3741 
3742 	.add_gid = mlx5_ib_add_gid,
3743 	.alloc_mr = mlx5_ib_alloc_mr,
3744 	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
3745 	.alloc_pd = mlx5_ib_alloc_pd,
3746 	.alloc_ucontext = mlx5_ib_alloc_ucontext,
3747 	.attach_mcast = mlx5_ib_mcg_attach,
3748 	.check_mr_status = mlx5_ib_check_mr_status,
3749 	.create_ah = mlx5_ib_create_ah,
3750 	.create_cq = mlx5_ib_create_cq,
3751 	.create_qp = mlx5_ib_create_qp,
3752 	.create_srq = mlx5_ib_create_srq,
3753 	.create_user_ah = mlx5_ib_create_ah,
3754 	.dealloc_pd = mlx5_ib_dealloc_pd,
3755 	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
3756 	.del_gid = mlx5_ib_del_gid,
3757 	.dereg_mr = mlx5_ib_dereg_mr,
3758 	.destroy_ah = mlx5_ib_destroy_ah,
3759 	.destroy_cq = mlx5_ib_destroy_cq,
3760 	.destroy_qp = mlx5_ib_destroy_qp,
3761 	.destroy_srq = mlx5_ib_destroy_srq,
3762 	.detach_mcast = mlx5_ib_mcg_detach,
3763 	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
3764 	.drain_rq = mlx5_ib_drain_rq,
3765 	.drain_sq = mlx5_ib_drain_sq,
3766 	.device_group = &mlx5_attr_group,
3767 	.enable_driver = mlx5_ib_enable_driver,
3768 	.get_dev_fw_str = get_dev_fw_str,
3769 	.get_dma_mr = mlx5_ib_get_dma_mr,
3770 	.get_link_layer = mlx5_ib_port_link_layer,
3771 	.map_mr_sg = mlx5_ib_map_mr_sg,
3772 	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
3773 	.mmap = mlx5_ib_mmap,
3774 	.mmap_free = mlx5_ib_mmap_free,
3775 	.modify_cq = mlx5_ib_modify_cq,
3776 	.modify_device = mlx5_ib_modify_device,
3777 	.modify_port = mlx5_ib_modify_port,
3778 	.modify_qp = mlx5_ib_modify_qp,
3779 	.modify_srq = mlx5_ib_modify_srq,
3780 	.poll_cq = mlx5_ib_poll_cq,
3781 	.post_recv = mlx5_ib_post_recv_nodrain,
3782 	.post_send = mlx5_ib_post_send_nodrain,
3783 	.post_srq_recv = mlx5_ib_post_srq_recv,
3784 	.process_mad = mlx5_ib_process_mad,
3785 	.query_ah = mlx5_ib_query_ah,
3786 	.query_device = mlx5_ib_query_device,
3787 	.query_gid = mlx5_ib_query_gid,
3788 	.query_pkey = mlx5_ib_query_pkey,
3789 	.query_qp = mlx5_ib_query_qp,
3790 	.query_srq = mlx5_ib_query_srq,
3791 	.query_ucontext = mlx5_ib_query_ucontext,
3792 	.reg_user_mr = mlx5_ib_reg_user_mr,
3793 	.reg_user_mr_dmabuf = mlx5_ib_reg_user_mr_dmabuf,
3794 	.req_notify_cq = mlx5_ib_arm_cq,
3795 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
3796 	.resize_cq = mlx5_ib_resize_cq,
3797 
3798 	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
3799 	INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs),
3800 	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
3801 	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
3802 	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
3803 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
3804 };
3805 
3806 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
3807 	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
3808 };
3809 
3810 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
3811 	.get_vf_config = mlx5_ib_get_vf_config,
3812 	.get_vf_guid = mlx5_ib_get_vf_guid,
3813 	.get_vf_stats = mlx5_ib_get_vf_stats,
3814 	.set_vf_guid = mlx5_ib_set_vf_guid,
3815 	.set_vf_link_state = mlx5_ib_set_vf_link_state,
3816 };
3817 
3818 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
3819 	.alloc_mw = mlx5_ib_alloc_mw,
3820 	.dealloc_mw = mlx5_ib_dealloc_mw,
3821 
3822 	INIT_RDMA_OBJ_SIZE(ib_mw, mlx5_ib_mw, ibmw),
3823 };
3824 
3825 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
3826 	.alloc_xrcd = mlx5_ib_alloc_xrcd,
3827 	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
3828 
3829 	INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd),
3830 };
3831 
3832 static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
3833 {
3834 	struct mlx5_core_dev *mdev = dev->mdev;
3835 	struct mlx5_var_table *var_table = &dev->var_table;
3836 	u8 log_doorbell_bar_size;
3837 	u8 log_doorbell_stride;
3838 	u64 bar_size;
3839 
3840 	log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3841 					log_doorbell_bar_size);
3842 	log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3843 					log_doorbell_stride);
3844 	var_table->hw_start_addr = dev->mdev->bar_addr +
3845 				MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
3846 					doorbell_bar_offset);
3847 	bar_size = (1ULL << log_doorbell_bar_size) * 4096;
3848 	var_table->stride_size = 1ULL << log_doorbell_stride;
3849 	var_table->num_var_hw_entries = div_u64(bar_size,
3850 						var_table->stride_size);
3851 	mutex_init(&var_table->bitmap_lock);
3852 	var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
3853 					  GFP_KERNEL);
3854 	return (var_table->bitmap) ? 0 : -ENOMEM;
3855 }
3856 
3857 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
3858 {
3859 	bitmap_free(dev->var_table.bitmap);
3860 }
3861 
3862 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
3863 {
3864 	struct mlx5_core_dev *mdev = dev->mdev;
3865 	int err;
3866 
3867 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
3868 	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
3869 		ib_set_device_ops(&dev->ib_dev,
3870 				  &mlx5_ib_dev_ipoib_enhanced_ops);
3871 
3872 	if (mlx5_core_is_pf(mdev))
3873 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
3874 
3875 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
3876 
3877 	if (MLX5_CAP_GEN(mdev, imaicl))
3878 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
3879 
3880 	if (MLX5_CAP_GEN(mdev, xrc))
3881 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
3882 
3883 	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
3884 	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3885 	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
3886 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
3887 
3888 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
3889 
3890 	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
3891 		dev->ib_dev.driver_def = mlx5_ib_defs;
3892 
3893 	err = init_node_data(dev);
3894 	if (err)
3895 		return err;
3896 
3897 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
3898 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
3899 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
3900 		mutex_init(&dev->lb.mutex);
3901 
3902 	if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3903 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
3904 		err = mlx5_ib_init_var_table(dev);
3905 		if (err)
3906 			return err;
3907 	}
3908 
3909 	dev->ib_dev.use_cq_dim = true;
3910 
3911 	return 0;
3912 }
3913 
3914 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
3915 	.get_port_immutable = mlx5_port_immutable,
3916 	.query_port = mlx5_ib_query_port,
3917 };
3918 
3919 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
3920 {
3921 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
3922 	return 0;
3923 }
3924 
3925 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
3926 	.get_port_immutable = mlx5_port_rep_immutable,
3927 	.query_port = mlx5_ib_rep_query_port,
3928 	.query_pkey = mlx5_ib_rep_query_pkey,
3929 };
3930 
3931 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
3932 {
3933 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
3934 	return 0;
3935 }
3936 
3937 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
3938 	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
3939 	.create_wq = mlx5_ib_create_wq,
3940 	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
3941 	.destroy_wq = mlx5_ib_destroy_wq,
3942 	.get_netdev = mlx5_ib_get_netdev,
3943 	.modify_wq = mlx5_ib_modify_wq,
3944 
3945 	INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table,
3946 			   ib_rwq_ind_tbl),
3947 };
3948 
3949 static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev)
3950 {
3951 	struct mlx5_core_dev *mdev = dev->mdev;
3952 	enum rdma_link_layer ll;
3953 	int port_type_cap;
3954 	u32 port_num = 0;
3955 	int err;
3956 
3957 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3958 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3959 
3960 	if (ll == IB_LINK_LAYER_ETHERNET) {
3961 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
3962 
3963 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3964 
3965 		/* Register only for native ports */
3966 		err = mlx5_add_netdev_notifier(dev, port_num);
3967 		if (err)
3968 			return err;
3969 
3970 		err = mlx5_enable_eth(dev);
3971 		if (err)
3972 			goto cleanup;
3973 	}
3974 
3975 	return 0;
3976 cleanup:
3977 	mlx5_remove_netdev_notifier(dev, port_num);
3978 	return err;
3979 }
3980 
3981 static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev)
3982 {
3983 	struct mlx5_core_dev *mdev = dev->mdev;
3984 	enum rdma_link_layer ll;
3985 	int port_type_cap;
3986 	u32 port_num;
3987 
3988 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3989 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3990 
3991 	if (ll == IB_LINK_LAYER_ETHERNET) {
3992 		mlx5_disable_eth(dev);
3993 
3994 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3995 		mlx5_remove_netdev_notifier(dev, port_num);
3996 	}
3997 }
3998 
3999 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
4000 {
4001 	mlx5_ib_init_cong_debugfs(dev,
4002 				  mlx5_core_native_port_num(dev->mdev) - 1);
4003 	return 0;
4004 }
4005 
4006 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
4007 {
4008 	mlx5_ib_cleanup_cong_debugfs(dev,
4009 				     mlx5_core_native_port_num(dev->mdev) - 1);
4010 }
4011 
4012 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
4013 {
4014 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
4015 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
4016 }
4017 
4018 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
4019 {
4020 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
4021 }
4022 
4023 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
4024 {
4025 	int err;
4026 
4027 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
4028 	if (err)
4029 		return err;
4030 
4031 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
4032 	if (err)
4033 		mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4034 
4035 	return err;
4036 }
4037 
4038 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
4039 {
4040 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4041 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4042 }
4043 
4044 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
4045 {
4046 	const char *name;
4047 
4048 	if (!mlx5_lag_is_active(dev->mdev))
4049 		name = "mlx5_%d";
4050 	else
4051 		name = "mlx5_bond_%d";
4052 	return ib_register_device(&dev->ib_dev, name, &dev->mdev->pdev->dev);
4053 }
4054 
4055 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
4056 {
4057 	int err;
4058 
4059 	err = mlx5_mr_cache_cleanup(dev);
4060 	if (err)
4061 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
4062 
4063 	if (dev->umrc.qp)
4064 		mlx5_ib_destroy_qp(dev->umrc.qp, NULL);
4065 	if (dev->umrc.cq)
4066 		ib_free_cq(dev->umrc.cq);
4067 	if (dev->umrc.pd)
4068 		ib_dealloc_pd(dev->umrc.pd);
4069 }
4070 
4071 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
4072 {
4073 	ib_unregister_device(&dev->ib_dev);
4074 }
4075 
4076 enum {
4077 	MAX_UMR_WR = 128,
4078 };
4079 
4080 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
4081 {
4082 	struct ib_qp_init_attr *init_attr = NULL;
4083 	struct ib_qp_attr *attr = NULL;
4084 	struct ib_pd *pd;
4085 	struct ib_cq *cq;
4086 	struct ib_qp *qp;
4087 	int ret;
4088 
4089 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
4090 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
4091 	if (!attr || !init_attr) {
4092 		ret = -ENOMEM;
4093 		goto error_0;
4094 	}
4095 
4096 	pd = ib_alloc_pd(&dev->ib_dev, 0);
4097 	if (IS_ERR(pd)) {
4098 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
4099 		ret = PTR_ERR(pd);
4100 		goto error_0;
4101 	}
4102 
4103 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
4104 	if (IS_ERR(cq)) {
4105 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
4106 		ret = PTR_ERR(cq);
4107 		goto error_2;
4108 	}
4109 
4110 	init_attr->send_cq = cq;
4111 	init_attr->recv_cq = cq;
4112 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
4113 	init_attr->cap.max_send_wr = MAX_UMR_WR;
4114 	init_attr->cap.max_send_sge = 1;
4115 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
4116 	init_attr->port_num = 1;
4117 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
4118 	if (IS_ERR(qp)) {
4119 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
4120 		ret = PTR_ERR(qp);
4121 		goto error_3;
4122 	}
4123 	qp->device     = &dev->ib_dev;
4124 	qp->real_qp    = qp;
4125 	qp->uobject    = NULL;
4126 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
4127 	qp->send_cq    = init_attr->send_cq;
4128 	qp->recv_cq    = init_attr->recv_cq;
4129 
4130 	attr->qp_state = IB_QPS_INIT;
4131 	attr->port_num = 1;
4132 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
4133 				IB_QP_PORT, NULL);
4134 	if (ret) {
4135 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
4136 		goto error_4;
4137 	}
4138 
4139 	memset(attr, 0, sizeof(*attr));
4140 	attr->qp_state = IB_QPS_RTR;
4141 	attr->path_mtu = IB_MTU_256;
4142 
4143 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4144 	if (ret) {
4145 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
4146 		goto error_4;
4147 	}
4148 
4149 	memset(attr, 0, sizeof(*attr));
4150 	attr->qp_state = IB_QPS_RTS;
4151 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4152 	if (ret) {
4153 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
4154 		goto error_4;
4155 	}
4156 
4157 	dev->umrc.qp = qp;
4158 	dev->umrc.cq = cq;
4159 	dev->umrc.pd = pd;
4160 
4161 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
4162 	ret = mlx5_mr_cache_init(dev);
4163 	if (ret) {
4164 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4165 		goto error_4;
4166 	}
4167 
4168 	kfree(attr);
4169 	kfree(init_attr);
4170 
4171 	return 0;
4172 
4173 error_4:
4174 	mlx5_ib_destroy_qp(qp, NULL);
4175 	dev->umrc.qp = NULL;
4176 
4177 error_3:
4178 	ib_free_cq(cq);
4179 	dev->umrc.cq = NULL;
4180 
4181 error_2:
4182 	ib_dealloc_pd(pd);
4183 	dev->umrc.pd = NULL;
4184 
4185 error_0:
4186 	kfree(attr);
4187 	kfree(init_attr);
4188 	return ret;
4189 }
4190 
4191 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
4192 {
4193 	struct dentry *root;
4194 
4195 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4196 		return 0;
4197 
4198 	mutex_init(&dev->delay_drop.lock);
4199 	dev->delay_drop.dev = dev;
4200 	dev->delay_drop.activate = false;
4201 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4202 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4203 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4204 	atomic_set(&dev->delay_drop.events_cnt, 0);
4205 
4206 	if (!mlx5_debugfs_root)
4207 		return 0;
4208 
4209 	root = debugfs_create_dir("delay_drop", dev->mdev->priv.dbg_root);
4210 	dev->delay_drop.dir_debugfs = root;
4211 
4212 	debugfs_create_atomic_t("num_timeout_events", 0400, root,
4213 				&dev->delay_drop.events_cnt);
4214 	debugfs_create_atomic_t("num_rqs", 0400, root,
4215 				&dev->delay_drop.rqs_cnt);
4216 	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
4217 			    &fops_delay_drop_timeout);
4218 	return 0;
4219 }
4220 
4221 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
4222 {
4223 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4224 		return;
4225 
4226 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4227 	if (!dev->delay_drop.dir_debugfs)
4228 		return;
4229 
4230 	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
4231 	dev->delay_drop.dir_debugfs = NULL;
4232 }
4233 
4234 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
4235 {
4236 	dev->mdev_events.notifier_call = mlx5_ib_event;
4237 	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
4238 	return 0;
4239 }
4240 
4241 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
4242 {
4243 	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
4244 }
4245 
4246 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
4247 		      const struct mlx5_ib_profile *profile,
4248 		      int stage)
4249 {
4250 	dev->ib_active = false;
4251 
4252 	/* Number of stages to cleanup */
4253 	while (stage) {
4254 		stage--;
4255 		if (profile->stage[stage].cleanup)
4256 			profile->stage[stage].cleanup(dev);
4257 	}
4258 
4259 	kfree(dev->port);
4260 	ib_dealloc_device(&dev->ib_dev);
4261 }
4262 
4263 int __mlx5_ib_add(struct mlx5_ib_dev *dev,
4264 		  const struct mlx5_ib_profile *profile)
4265 {
4266 	int err;
4267 	int i;
4268 
4269 	dev->profile = profile;
4270 
4271 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
4272 		if (profile->stage[i].init) {
4273 			err = profile->stage[i].init(dev);
4274 			if (err)
4275 				goto err_out;
4276 		}
4277 	}
4278 
4279 	dev->ib_active = true;
4280 	return 0;
4281 
4282 err_out:
4283 	/* Clean up stages which were initialized */
4284 	while (i) {
4285 		i--;
4286 		if (profile->stage[i].cleanup)
4287 			profile->stage[i].cleanup(dev);
4288 	}
4289 	return -ENOMEM;
4290 }
4291 
4292 static const struct mlx5_ib_profile pf_profile = {
4293 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4294 		     mlx5_ib_stage_init_init,
4295 		     mlx5_ib_stage_init_cleanup),
4296 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4297 		     mlx5_ib_fs_init,
4298 		     mlx5_ib_fs_cleanup),
4299 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4300 		     mlx5_ib_stage_caps_init,
4301 		     mlx5_ib_stage_caps_cleanup),
4302 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4303 		     mlx5_ib_stage_non_default_cb,
4304 		     NULL),
4305 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4306 		     mlx5_ib_roce_init,
4307 		     mlx5_ib_roce_cleanup),
4308 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4309 		     mlx5_init_qp_table,
4310 		     mlx5_cleanup_qp_table),
4311 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4312 		     mlx5_init_srq_table,
4313 		     mlx5_cleanup_srq_table),
4314 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4315 		     mlx5_ib_dev_res_init,
4316 		     mlx5_ib_dev_res_cleanup),
4317 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4318 		     mlx5_ib_stage_dev_notifier_init,
4319 		     mlx5_ib_stage_dev_notifier_cleanup),
4320 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
4321 		     mlx5_ib_odp_init_one,
4322 		     mlx5_ib_odp_cleanup_one),
4323 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4324 		     mlx5_ib_counters_init,
4325 		     mlx5_ib_counters_cleanup),
4326 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4327 		     mlx5_ib_stage_cong_debugfs_init,
4328 		     mlx5_ib_stage_cong_debugfs_cleanup),
4329 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4330 		     mlx5_ib_stage_uar_init,
4331 		     mlx5_ib_stage_uar_cleanup),
4332 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4333 		     mlx5_ib_stage_bfrag_init,
4334 		     mlx5_ib_stage_bfrag_cleanup),
4335 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4336 		     NULL,
4337 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4338 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4339 		     mlx5_ib_devx_init,
4340 		     mlx5_ib_devx_cleanup),
4341 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4342 		     mlx5_ib_stage_ib_reg_init,
4343 		     mlx5_ib_stage_ib_reg_cleanup),
4344 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4345 		     mlx5_ib_stage_post_ib_reg_umr_init,
4346 		     NULL),
4347 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4348 		     mlx5_ib_stage_delay_drop_init,
4349 		     mlx5_ib_stage_delay_drop_cleanup),
4350 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4351 		     mlx5_ib_restrack_init,
4352 		     NULL),
4353 };
4354 
4355 const struct mlx5_ib_profile raw_eth_profile = {
4356 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4357 		     mlx5_ib_stage_init_init,
4358 		     mlx5_ib_stage_init_cleanup),
4359 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4360 		     mlx5_ib_fs_init,
4361 		     mlx5_ib_fs_cleanup),
4362 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4363 		     mlx5_ib_stage_caps_init,
4364 		     mlx5_ib_stage_caps_cleanup),
4365 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4366 		     mlx5_ib_stage_raw_eth_non_default_cb,
4367 		     NULL),
4368 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4369 		     mlx5_ib_roce_init,
4370 		     mlx5_ib_roce_cleanup),
4371 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4372 		     mlx5_init_qp_table,
4373 		     mlx5_cleanup_qp_table),
4374 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4375 		     mlx5_init_srq_table,
4376 		     mlx5_cleanup_srq_table),
4377 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4378 		     mlx5_ib_dev_res_init,
4379 		     mlx5_ib_dev_res_cleanup),
4380 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4381 		     mlx5_ib_stage_dev_notifier_init,
4382 		     mlx5_ib_stage_dev_notifier_cleanup),
4383 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4384 		     mlx5_ib_counters_init,
4385 		     mlx5_ib_counters_cleanup),
4386 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4387 		     mlx5_ib_stage_cong_debugfs_init,
4388 		     mlx5_ib_stage_cong_debugfs_cleanup),
4389 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4390 		     mlx5_ib_stage_uar_init,
4391 		     mlx5_ib_stage_uar_cleanup),
4392 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4393 		     mlx5_ib_stage_bfrag_init,
4394 		     mlx5_ib_stage_bfrag_cleanup),
4395 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4396 		     NULL,
4397 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4398 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4399 		     mlx5_ib_devx_init,
4400 		     mlx5_ib_devx_cleanup),
4401 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4402 		     mlx5_ib_stage_ib_reg_init,
4403 		     mlx5_ib_stage_ib_reg_cleanup),
4404 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4405 		     mlx5_ib_stage_post_ib_reg_umr_init,
4406 		     NULL),
4407 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4408 		     mlx5_ib_restrack_init,
4409 		     NULL),
4410 };
4411 
4412 static int mlx5r_mp_probe(struct auxiliary_device *adev,
4413 			  const struct auxiliary_device_id *id)
4414 {
4415 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4416 	struct mlx5_core_dev *mdev = idev->mdev;
4417 	struct mlx5_ib_multiport_info *mpi;
4418 	struct mlx5_ib_dev *dev;
4419 	bool bound = false;
4420 	int err;
4421 
4422 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4423 	if (!mpi)
4424 		return -ENOMEM;
4425 
4426 	mpi->mdev = mdev;
4427 	err = mlx5_query_nic_vport_system_image_guid(mdev,
4428 						     &mpi->sys_image_guid);
4429 	if (err) {
4430 		kfree(mpi);
4431 		return err;
4432 	}
4433 
4434 	mutex_lock(&mlx5_ib_multiport_mutex);
4435 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
4436 		if (dev->sys_image_guid == mpi->sys_image_guid)
4437 			bound = mlx5_ib_bind_slave_port(dev, mpi);
4438 
4439 		if (bound) {
4440 			rdma_roce_rescan_device(&dev->ib_dev);
4441 			mpi->ibdev->ib_active = true;
4442 			break;
4443 		}
4444 	}
4445 
4446 	if (!bound) {
4447 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4448 		dev_dbg(mdev->device,
4449 			"no suitable IB device found to bind to, added to unaffiliated list.\n");
4450 	}
4451 	mutex_unlock(&mlx5_ib_multiport_mutex);
4452 
4453 	dev_set_drvdata(&adev->dev, mpi);
4454 	return 0;
4455 }
4456 
4457 static void mlx5r_mp_remove(struct auxiliary_device *adev)
4458 {
4459 	struct mlx5_ib_multiport_info *mpi;
4460 
4461 	mpi = dev_get_drvdata(&adev->dev);
4462 	mutex_lock(&mlx5_ib_multiport_mutex);
4463 	if (mpi->ibdev)
4464 		mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
4465 	else
4466 		list_del(&mpi->list);
4467 	mutex_unlock(&mlx5_ib_multiport_mutex);
4468 	kfree(mpi);
4469 }
4470 
4471 static int mlx5r_probe(struct auxiliary_device *adev,
4472 		       const struct auxiliary_device_id *id)
4473 {
4474 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4475 	struct mlx5_core_dev *mdev = idev->mdev;
4476 	const struct mlx5_ib_profile *profile;
4477 	int port_type_cap, num_ports, ret;
4478 	enum rdma_link_layer ll;
4479 	struct mlx5_ib_dev *dev;
4480 
4481 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4482 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4483 
4484 	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
4485 			MLX5_CAP_GEN(mdev, num_vhca_ports));
4486 	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
4487 	if (!dev)
4488 		return -ENOMEM;
4489 	dev->port = kcalloc(num_ports, sizeof(*dev->port),
4490 			     GFP_KERNEL);
4491 	if (!dev->port) {
4492 		ib_dealloc_device(&dev->ib_dev);
4493 		return -ENOMEM;
4494 	}
4495 
4496 	dev->mdev = mdev;
4497 	dev->num_ports = num_ports;
4498 
4499 	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_is_roce_init_enabled(mdev))
4500 		profile = &raw_eth_profile;
4501 	else
4502 		profile = &pf_profile;
4503 
4504 	ret = __mlx5_ib_add(dev, profile);
4505 	if (ret) {
4506 		kfree(dev->port);
4507 		ib_dealloc_device(&dev->ib_dev);
4508 		return ret;
4509 	}
4510 
4511 	dev_set_drvdata(&adev->dev, dev);
4512 	return 0;
4513 }
4514 
4515 static void mlx5r_remove(struct auxiliary_device *adev)
4516 {
4517 	struct mlx5_ib_dev *dev;
4518 
4519 	dev = dev_get_drvdata(&adev->dev);
4520 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
4521 }
4522 
4523 static const struct auxiliary_device_id mlx5r_mp_id_table[] = {
4524 	{ .name = MLX5_ADEV_NAME ".multiport", },
4525 	{},
4526 };
4527 
4528 static const struct auxiliary_device_id mlx5r_id_table[] = {
4529 	{ .name = MLX5_ADEV_NAME ".rdma", },
4530 	{},
4531 };
4532 
4533 MODULE_DEVICE_TABLE(auxiliary, mlx5r_mp_id_table);
4534 MODULE_DEVICE_TABLE(auxiliary, mlx5r_id_table);
4535 
4536 static struct auxiliary_driver mlx5r_mp_driver = {
4537 	.name = "multiport",
4538 	.probe = mlx5r_mp_probe,
4539 	.remove = mlx5r_mp_remove,
4540 	.id_table = mlx5r_mp_id_table,
4541 };
4542 
4543 static struct auxiliary_driver mlx5r_driver = {
4544 	.name = "rdma",
4545 	.probe = mlx5r_probe,
4546 	.remove = mlx5r_remove,
4547 	.id_table = mlx5r_id_table,
4548 };
4549 
4550 static int __init mlx5_ib_init(void)
4551 {
4552 	int ret;
4553 
4554 	xlt_emergency_page = (void *)__get_free_page(GFP_KERNEL);
4555 	if (!xlt_emergency_page)
4556 		return -ENOMEM;
4557 
4558 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
4559 	if (!mlx5_ib_event_wq) {
4560 		free_page((unsigned long)xlt_emergency_page);
4561 		return -ENOMEM;
4562 	}
4563 
4564 	mlx5_ib_odp_init();
4565 	ret = mlx5r_rep_init();
4566 	if (ret)
4567 		goto rep_err;
4568 	ret = auxiliary_driver_register(&mlx5r_mp_driver);
4569 	if (ret)
4570 		goto mp_err;
4571 	ret = auxiliary_driver_register(&mlx5r_driver);
4572 	if (ret)
4573 		goto drv_err;
4574 	return 0;
4575 
4576 drv_err:
4577 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4578 mp_err:
4579 	mlx5r_rep_cleanup();
4580 rep_err:
4581 	destroy_workqueue(mlx5_ib_event_wq);
4582 	free_page((unsigned long)xlt_emergency_page);
4583 	return ret;
4584 }
4585 
4586 static void __exit mlx5_ib_cleanup(void)
4587 {
4588 	auxiliary_driver_unregister(&mlx5r_driver);
4589 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4590 	mlx5r_rep_cleanup();
4591 
4592 	destroy_workqueue(mlx5_ib_event_wq);
4593 	free_page((unsigned long)xlt_emergency_page);
4594 }
4595 
4596 module_init(mlx5_ib_init);
4597 module_exit(mlx5_ib_cleanup);
4598