xref: /linux/drivers/infiniband/hw/mlx5/main.c (revision f2745dc0ba3dadd8fa2b2c33f48253d78e133a12)
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 "umr.h"
45 #include <rdma/uverbs_std_types.h>
46 #include <rdma/uverbs_ioctl.h>
47 #include <rdma/mlx5_user_ioctl_verbs.h>
48 #include <rdma/mlx5_user_ioctl_cmds.h>
49 #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->kernel_cap_flags |= IBK_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->kernel_cap_flags |= IBK_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->kernel_cap_flags |= IBK_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 			resp.response_length += sizeof(resp.rss_caps);
910 		}
911 	} else {
912 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen)
913 			resp.response_length += sizeof(resp.tso_caps);
914 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen)
915 			resp.response_length += sizeof(resp.rss_caps);
916 	}
917 
918 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
919 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
920 		props->kernel_cap_flags |= IBK_UD_TSO;
921 	}
922 
923 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
924 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
925 	    raw_support)
926 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
927 
928 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
929 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
930 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
931 
932 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
933 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
934 	    raw_support) {
935 		/* Legacy bit to support old userspace libraries */
936 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
937 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
938 	}
939 
940 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
941 		props->max_dm_size =
942 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
943 	}
944 
945 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
946 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
947 
948 	if (MLX5_CAP_GEN(mdev, end_pad))
949 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
950 
951 	props->vendor_part_id	   = mdev->pdev->device;
952 	props->hw_ver		   = mdev->pdev->revision;
953 
954 	props->max_mr_size	   = ~0ull;
955 	props->page_size_cap	   = ~(min_page_size - 1);
956 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
957 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
958 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
959 		     sizeof(struct mlx5_wqe_data_seg);
960 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
961 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
962 		     sizeof(struct mlx5_wqe_raddr_seg)) /
963 		sizeof(struct mlx5_wqe_data_seg);
964 	props->max_send_sge = max_sq_sg;
965 	props->max_recv_sge = max_rq_sg;
966 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
967 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
968 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
969 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
970 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
971 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
972 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
973 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
974 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
975 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
976 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
977 	props->max_srq_sge	   = max_rq_sg - 1;
978 	props->max_fast_reg_page_list_len =
979 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
980 	props->max_pi_fast_reg_page_list_len =
981 		props->max_fast_reg_page_list_len / 2;
982 	props->max_sgl_rd =
983 		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
984 	get_atomic_caps_qp(dev, props);
985 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
986 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
987 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
988 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
989 					   props->max_mcast_grp;
990 	props->max_ah = INT_MAX;
991 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
992 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
993 
994 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
995 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
996 			props->kernel_cap_flags |= IBK_ON_DEMAND_PAGING;
997 		props->odp_caps = dev->odp_caps;
998 		if (!uhw) {
999 			/* ODP for kernel QPs is not implemented for receive
1000 			 * WQEs and SRQ WQEs
1001 			 */
1002 			props->odp_caps.per_transport_caps.rc_odp_caps &=
1003 				~(IB_ODP_SUPPORT_READ |
1004 				  IB_ODP_SUPPORT_SRQ_RECV);
1005 			props->odp_caps.per_transport_caps.uc_odp_caps &=
1006 				~(IB_ODP_SUPPORT_READ |
1007 				  IB_ODP_SUPPORT_SRQ_RECV);
1008 			props->odp_caps.per_transport_caps.ud_odp_caps &=
1009 				~(IB_ODP_SUPPORT_READ |
1010 				  IB_ODP_SUPPORT_SRQ_RECV);
1011 			props->odp_caps.per_transport_caps.xrc_odp_caps &=
1012 				~(IB_ODP_SUPPORT_READ |
1013 				  IB_ODP_SUPPORT_SRQ_RECV);
1014 		}
1015 	}
1016 
1017 	if (mlx5_core_is_vf(mdev))
1018 		props->kernel_cap_flags |= IBK_VIRTUAL_FUNCTION;
1019 
1020 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1021 	    IB_LINK_LAYER_ETHERNET && raw_support) {
1022 		props->rss_caps.max_rwq_indirection_tables =
1023 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1024 		props->rss_caps.max_rwq_indirection_table_size =
1025 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1026 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1027 		props->max_wq_type_rq =
1028 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1029 	}
1030 
1031 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1032 		props->tm_caps.max_num_tags =
1033 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1034 		props->tm_caps.max_ops =
1035 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1036 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1037 	}
1038 
1039 	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1040 	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1041 		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1042 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1043 	}
1044 
1045 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1046 		props->cq_caps.max_cq_moderation_count =
1047 						MLX5_MAX_CQ_COUNT;
1048 		props->cq_caps.max_cq_moderation_period =
1049 						MLX5_MAX_CQ_PERIOD;
1050 	}
1051 
1052 	if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) {
1053 		resp.response_length += sizeof(resp.cqe_comp_caps);
1054 
1055 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1056 			resp.cqe_comp_caps.max_num =
1057 				MLX5_CAP_GEN(dev->mdev,
1058 					     cqe_compression_max_num);
1059 
1060 			resp.cqe_comp_caps.supported_format =
1061 				MLX5_IB_CQE_RES_FORMAT_HASH |
1062 				MLX5_IB_CQE_RES_FORMAT_CSUM;
1063 
1064 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1065 				resp.cqe_comp_caps.supported_format |=
1066 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1067 		}
1068 	}
1069 
1070 	if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen &&
1071 	    raw_support) {
1072 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1073 		    MLX5_CAP_GEN(mdev, qos)) {
1074 			resp.packet_pacing_caps.qp_rate_limit_max =
1075 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1076 			resp.packet_pacing_caps.qp_rate_limit_min =
1077 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1078 			resp.packet_pacing_caps.supported_qpts |=
1079 				1 << IB_QPT_RAW_PACKET;
1080 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1081 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1082 				resp.packet_pacing_caps.cap_flags |=
1083 					MLX5_IB_PP_SUPPORT_BURST;
1084 		}
1085 		resp.response_length += sizeof(resp.packet_pacing_caps);
1086 	}
1087 
1088 	if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <=
1089 	    uhw_outlen) {
1090 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1091 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1092 				MLX5_IB_ALLOW_MPW;
1093 
1094 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1095 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1096 				MLX5_IB_SUPPORT_EMPW;
1097 
1098 		resp.response_length +=
1099 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1100 	}
1101 
1102 	if (offsetofend(typeof(resp), flags) <= uhw_outlen) {
1103 		resp.response_length += sizeof(resp.flags);
1104 
1105 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1106 			resp.flags |=
1107 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1108 
1109 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1110 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1111 		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1112 			resp.flags |=
1113 				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1114 
1115 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1116 	}
1117 
1118 	if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) {
1119 		resp.response_length += sizeof(resp.sw_parsing_caps);
1120 		if (MLX5_CAP_ETH(mdev, swp)) {
1121 			resp.sw_parsing_caps.sw_parsing_offloads |=
1122 				MLX5_IB_SW_PARSING;
1123 
1124 			if (MLX5_CAP_ETH(mdev, swp_csum))
1125 				resp.sw_parsing_caps.sw_parsing_offloads |=
1126 					MLX5_IB_SW_PARSING_CSUM;
1127 
1128 			if (MLX5_CAP_ETH(mdev, swp_lso))
1129 				resp.sw_parsing_caps.sw_parsing_offloads |=
1130 					MLX5_IB_SW_PARSING_LSO;
1131 
1132 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1133 				resp.sw_parsing_caps.supported_qpts =
1134 					BIT(IB_QPT_RAW_PACKET);
1135 		}
1136 	}
1137 
1138 	if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen &&
1139 	    raw_support) {
1140 		resp.response_length += sizeof(resp.striding_rq_caps);
1141 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1142 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1143 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1144 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1145 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1146 			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1147 				resp.striding_rq_caps
1148 					.min_single_wqe_log_num_of_strides =
1149 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1150 			else
1151 				resp.striding_rq_caps
1152 					.min_single_wqe_log_num_of_strides =
1153 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1154 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1155 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1156 			resp.striding_rq_caps.supported_qpts =
1157 				BIT(IB_QPT_RAW_PACKET);
1158 		}
1159 	}
1160 
1161 	if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) {
1162 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1163 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1164 			resp.tunnel_offloads_caps |=
1165 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1166 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1167 			resp.tunnel_offloads_caps |=
1168 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1169 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1170 			resp.tunnel_offloads_caps |=
1171 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1172 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre))
1173 			resp.tunnel_offloads_caps |=
1174 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1175 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp))
1176 			resp.tunnel_offloads_caps |=
1177 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1178 	}
1179 
1180 	if (offsetofend(typeof(resp), dci_streams_caps) <= uhw_outlen) {
1181 		resp.response_length += sizeof(resp.dci_streams_caps);
1182 
1183 		resp.dci_streams_caps.max_log_num_concurent =
1184 			MLX5_CAP_GEN(mdev, log_max_dci_stream_channels);
1185 
1186 		resp.dci_streams_caps.max_log_num_errored =
1187 			MLX5_CAP_GEN(mdev, log_max_dci_errored_streams);
1188 	}
1189 
1190 	if (uhw_outlen) {
1191 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1192 
1193 		if (err)
1194 			return err;
1195 	}
1196 
1197 	return 0;
1198 }
1199 
1200 static void translate_active_width(struct ib_device *ibdev, u16 active_width,
1201 				   u8 *ib_width)
1202 {
1203 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1204 
1205 	if (active_width & MLX5_PTYS_WIDTH_1X)
1206 		*ib_width = IB_WIDTH_1X;
1207 	else if (active_width & MLX5_PTYS_WIDTH_2X)
1208 		*ib_width = IB_WIDTH_2X;
1209 	else if (active_width & MLX5_PTYS_WIDTH_4X)
1210 		*ib_width = IB_WIDTH_4X;
1211 	else if (active_width & MLX5_PTYS_WIDTH_8X)
1212 		*ib_width = IB_WIDTH_8X;
1213 	else if (active_width & MLX5_PTYS_WIDTH_12X)
1214 		*ib_width = IB_WIDTH_12X;
1215 	else {
1216 		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1217 			    active_width);
1218 		*ib_width = IB_WIDTH_4X;
1219 	}
1220 
1221 	return;
1222 }
1223 
1224 static int mlx5_mtu_to_ib_mtu(int mtu)
1225 {
1226 	switch (mtu) {
1227 	case 256: return 1;
1228 	case 512: return 2;
1229 	case 1024: return 3;
1230 	case 2048: return 4;
1231 	case 4096: return 5;
1232 	default:
1233 		pr_warn("invalid mtu\n");
1234 		return -1;
1235 	}
1236 }
1237 
1238 enum ib_max_vl_num {
1239 	__IB_MAX_VL_0		= 1,
1240 	__IB_MAX_VL_0_1		= 2,
1241 	__IB_MAX_VL_0_3		= 3,
1242 	__IB_MAX_VL_0_7		= 4,
1243 	__IB_MAX_VL_0_14	= 5,
1244 };
1245 
1246 enum mlx5_vl_hw_cap {
1247 	MLX5_VL_HW_0	= 1,
1248 	MLX5_VL_HW_0_1	= 2,
1249 	MLX5_VL_HW_0_2	= 3,
1250 	MLX5_VL_HW_0_3	= 4,
1251 	MLX5_VL_HW_0_4	= 5,
1252 	MLX5_VL_HW_0_5	= 6,
1253 	MLX5_VL_HW_0_6	= 7,
1254 	MLX5_VL_HW_0_7	= 8,
1255 	MLX5_VL_HW_0_14	= 15
1256 };
1257 
1258 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1259 				u8 *max_vl_num)
1260 {
1261 	switch (vl_hw_cap) {
1262 	case MLX5_VL_HW_0:
1263 		*max_vl_num = __IB_MAX_VL_0;
1264 		break;
1265 	case MLX5_VL_HW_0_1:
1266 		*max_vl_num = __IB_MAX_VL_0_1;
1267 		break;
1268 	case MLX5_VL_HW_0_3:
1269 		*max_vl_num = __IB_MAX_VL_0_3;
1270 		break;
1271 	case MLX5_VL_HW_0_7:
1272 		*max_vl_num = __IB_MAX_VL_0_7;
1273 		break;
1274 	case MLX5_VL_HW_0_14:
1275 		*max_vl_num = __IB_MAX_VL_0_14;
1276 		break;
1277 
1278 	default:
1279 		return -EINVAL;
1280 	}
1281 
1282 	return 0;
1283 }
1284 
1285 static int mlx5_query_hca_port(struct ib_device *ibdev, u32 port,
1286 			       struct ib_port_attr *props)
1287 {
1288 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1289 	struct mlx5_core_dev *mdev = dev->mdev;
1290 	struct mlx5_hca_vport_context *rep;
1291 	u16 max_mtu;
1292 	u16 oper_mtu;
1293 	int err;
1294 	u16 ib_link_width_oper;
1295 	u8 vl_hw_cap;
1296 
1297 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1298 	if (!rep) {
1299 		err = -ENOMEM;
1300 		goto out;
1301 	}
1302 
1303 	/* props being zeroed by the caller, avoid zeroing it here */
1304 
1305 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1306 	if (err)
1307 		goto out;
1308 
1309 	props->lid		= rep->lid;
1310 	props->lmc		= rep->lmc;
1311 	props->sm_lid		= rep->sm_lid;
1312 	props->sm_sl		= rep->sm_sl;
1313 	props->state		= rep->vport_state;
1314 	props->phys_state	= rep->port_physical_state;
1315 	props->port_cap_flags	= rep->cap_mask1;
1316 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1317 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1318 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1319 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1320 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1321 	props->subnet_timeout	= rep->subnet_timeout;
1322 	props->init_type_reply	= rep->init_type_reply;
1323 
1324 	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1325 		props->port_cap_flags2 = rep->cap_mask2;
1326 
1327 	err = mlx5_query_ib_port_oper(mdev, &ib_link_width_oper,
1328 				      &props->active_speed, port);
1329 	if (err)
1330 		goto out;
1331 
1332 	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1333 
1334 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1335 
1336 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1337 
1338 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1339 
1340 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1341 
1342 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1343 	if (err)
1344 		goto out;
1345 
1346 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1347 				   &props->max_vl_num);
1348 out:
1349 	kfree(rep);
1350 	return err;
1351 }
1352 
1353 int mlx5_ib_query_port(struct ib_device *ibdev, u32 port,
1354 		       struct ib_port_attr *props)
1355 {
1356 	unsigned int count;
1357 	int ret;
1358 
1359 	switch (mlx5_get_vport_access_method(ibdev)) {
1360 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1361 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1362 		break;
1363 
1364 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1365 		ret = mlx5_query_hca_port(ibdev, port, props);
1366 		break;
1367 
1368 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1369 		ret = mlx5_query_port_roce(ibdev, port, props);
1370 		break;
1371 
1372 	default:
1373 		ret = -EINVAL;
1374 	}
1375 
1376 	if (!ret && props) {
1377 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1378 		struct mlx5_core_dev *mdev;
1379 		bool put_mdev = true;
1380 
1381 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1382 		if (!mdev) {
1383 			/* If the port isn't affiliated yet query the master.
1384 			 * The master and slave will have the same values.
1385 			 */
1386 			mdev = dev->mdev;
1387 			port = 1;
1388 			put_mdev = false;
1389 		}
1390 		count = mlx5_core_reserved_gids_count(mdev);
1391 		if (put_mdev)
1392 			mlx5_ib_put_native_port_mdev(dev, port);
1393 		props->gid_tbl_len -= count;
1394 	}
1395 	return ret;
1396 }
1397 
1398 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u32 port,
1399 				  struct ib_port_attr *props)
1400 {
1401 	return mlx5_query_port_roce(ibdev, port, props);
1402 }
1403 
1404 static int mlx5_ib_rep_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1405 				  u16 *pkey)
1406 {
1407 	/* Default special Pkey for representor device port as per the
1408 	 * IB specification 1.3 section 10.9.1.2.
1409 	 */
1410 	*pkey = 0xffff;
1411 	return 0;
1412 }
1413 
1414 static int mlx5_ib_query_gid(struct ib_device *ibdev, u32 port, int index,
1415 			     union ib_gid *gid)
1416 {
1417 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1418 	struct mlx5_core_dev *mdev = dev->mdev;
1419 
1420 	switch (mlx5_get_vport_access_method(ibdev)) {
1421 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1422 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1423 
1424 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1425 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1426 
1427 	default:
1428 		return -EINVAL;
1429 	}
1430 
1431 }
1432 
1433 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u32 port,
1434 				   u16 index, u16 *pkey)
1435 {
1436 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1437 	struct mlx5_core_dev *mdev;
1438 	bool put_mdev = true;
1439 	u32 mdev_port_num;
1440 	int err;
1441 
1442 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1443 	if (!mdev) {
1444 		/* The port isn't affiliated yet, get the PKey from the master
1445 		 * port. For RoCE the PKey tables will be the same.
1446 		 */
1447 		put_mdev = false;
1448 		mdev = dev->mdev;
1449 		mdev_port_num = 1;
1450 	}
1451 
1452 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1453 					index, pkey);
1454 	if (put_mdev)
1455 		mlx5_ib_put_native_port_mdev(dev, port);
1456 
1457 	return err;
1458 }
1459 
1460 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1461 			      u16 *pkey)
1462 {
1463 	switch (mlx5_get_vport_access_method(ibdev)) {
1464 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1465 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1466 
1467 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1468 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1469 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1470 	default:
1471 		return -EINVAL;
1472 	}
1473 }
1474 
1475 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1476 				 struct ib_device_modify *props)
1477 {
1478 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1479 	struct mlx5_reg_node_desc in;
1480 	struct mlx5_reg_node_desc out;
1481 	int err;
1482 
1483 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1484 		return -EOPNOTSUPP;
1485 
1486 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1487 		return 0;
1488 
1489 	/*
1490 	 * If possible, pass node desc to FW, so it can generate
1491 	 * a 144 trap.  If cmd fails, just ignore.
1492 	 */
1493 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1494 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1495 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1496 	if (err)
1497 		return err;
1498 
1499 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1500 
1501 	return err;
1502 }
1503 
1504 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u32 port_num, u32 mask,
1505 				u32 value)
1506 {
1507 	struct mlx5_hca_vport_context ctx = {};
1508 	struct mlx5_core_dev *mdev;
1509 	u32 mdev_port_num;
1510 	int err;
1511 
1512 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1513 	if (!mdev)
1514 		return -ENODEV;
1515 
1516 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1517 	if (err)
1518 		goto out;
1519 
1520 	if (~ctx.cap_mask1_perm & mask) {
1521 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1522 			     mask, ctx.cap_mask1_perm);
1523 		err = -EINVAL;
1524 		goto out;
1525 	}
1526 
1527 	ctx.cap_mask1 = value;
1528 	ctx.cap_mask1_perm = mask;
1529 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1530 						 0, &ctx);
1531 
1532 out:
1533 	mlx5_ib_put_native_port_mdev(dev, port_num);
1534 
1535 	return err;
1536 }
1537 
1538 static int mlx5_ib_modify_port(struct ib_device *ibdev, u32 port, int mask,
1539 			       struct ib_port_modify *props)
1540 {
1541 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1542 	struct ib_port_attr attr;
1543 	u32 tmp;
1544 	int err;
1545 	u32 change_mask;
1546 	u32 value;
1547 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1548 		      IB_LINK_LAYER_INFINIBAND);
1549 
1550 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1551 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1552 	 */
1553 	if (!is_ib)
1554 		return 0;
1555 
1556 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1557 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1558 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1559 		return set_port_caps_atomic(dev, port, change_mask, value);
1560 	}
1561 
1562 	mutex_lock(&dev->cap_mask_mutex);
1563 
1564 	err = ib_query_port(ibdev, port, &attr);
1565 	if (err)
1566 		goto out;
1567 
1568 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1569 		~props->clr_port_cap_mask;
1570 
1571 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1572 
1573 out:
1574 	mutex_unlock(&dev->cap_mask_mutex);
1575 	return err;
1576 }
1577 
1578 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1579 {
1580 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1581 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1582 }
1583 
1584 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1585 {
1586 	/* Large page with non 4k uar support might limit the dynamic size */
1587 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1588 		return MLX5_MIN_DYN_BFREGS;
1589 
1590 	return MLX5_MAX_DYN_BFREGS;
1591 }
1592 
1593 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1594 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1595 			     struct mlx5_bfreg_info *bfregi)
1596 {
1597 	int uars_per_sys_page;
1598 	int bfregs_per_sys_page;
1599 	int ref_bfregs = req->total_num_bfregs;
1600 
1601 	if (req->total_num_bfregs == 0)
1602 		return -EINVAL;
1603 
1604 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1605 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1606 
1607 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1608 		return -ENOMEM;
1609 
1610 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1611 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1612 	/* This holds the required static allocation asked by the user */
1613 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1614 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1615 		return -EINVAL;
1616 
1617 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1618 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1619 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1620 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1621 
1622 	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",
1623 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1624 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1625 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1626 		    bfregi->num_sys_pages);
1627 
1628 	return 0;
1629 }
1630 
1631 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1632 {
1633 	struct mlx5_bfreg_info *bfregi;
1634 	int err;
1635 	int i;
1636 
1637 	bfregi = &context->bfregi;
1638 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1639 		err = mlx5_cmd_uar_alloc(dev->mdev, &bfregi->sys_pages[i],
1640 					 context->devx_uid);
1641 		if (err)
1642 			goto error;
1643 
1644 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1645 	}
1646 
1647 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1648 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1649 
1650 	return 0;
1651 
1652 error:
1653 	for (--i; i >= 0; i--)
1654 		if (mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i],
1655 					 context->devx_uid))
1656 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1657 
1658 	return err;
1659 }
1660 
1661 static void deallocate_uars(struct mlx5_ib_dev *dev,
1662 			    struct mlx5_ib_ucontext *context)
1663 {
1664 	struct mlx5_bfreg_info *bfregi;
1665 	int i;
1666 
1667 	bfregi = &context->bfregi;
1668 	for (i = 0; i < bfregi->num_sys_pages; i++)
1669 		if (i < bfregi->num_static_sys_pages ||
1670 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1671 			mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i],
1672 					     context->devx_uid);
1673 }
1674 
1675 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1676 {
1677 	int err = 0;
1678 
1679 	mutex_lock(&dev->lb.mutex);
1680 	if (td)
1681 		dev->lb.user_td++;
1682 	if (qp)
1683 		dev->lb.qps++;
1684 
1685 	if (dev->lb.user_td == 2 ||
1686 	    dev->lb.qps == 1) {
1687 		if (!dev->lb.enabled) {
1688 			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1689 			dev->lb.enabled = true;
1690 		}
1691 	}
1692 
1693 	mutex_unlock(&dev->lb.mutex);
1694 
1695 	return err;
1696 }
1697 
1698 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1699 {
1700 	mutex_lock(&dev->lb.mutex);
1701 	if (td)
1702 		dev->lb.user_td--;
1703 	if (qp)
1704 		dev->lb.qps--;
1705 
1706 	if (dev->lb.user_td == 1 &&
1707 	    dev->lb.qps == 0) {
1708 		if (dev->lb.enabled) {
1709 			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1710 			dev->lb.enabled = false;
1711 		}
1712 	}
1713 
1714 	mutex_unlock(&dev->lb.mutex);
1715 }
1716 
1717 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1718 					  u16 uid)
1719 {
1720 	int err;
1721 
1722 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1723 		return 0;
1724 
1725 	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1726 	if (err)
1727 		return err;
1728 
1729 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1730 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1731 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1732 		return err;
1733 
1734 	return mlx5_ib_enable_lb(dev, true, false);
1735 }
1736 
1737 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1738 					     u16 uid)
1739 {
1740 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1741 		return;
1742 
1743 	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1744 
1745 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1746 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1747 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1748 		return;
1749 
1750 	mlx5_ib_disable_lb(dev, true, false);
1751 }
1752 
1753 static int set_ucontext_resp(struct ib_ucontext *uctx,
1754 			     struct mlx5_ib_alloc_ucontext_resp *resp)
1755 {
1756 	struct ib_device *ibdev = uctx->device;
1757 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1758 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1759 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1760 	int err;
1761 
1762 	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1763 		err = mlx5_cmd_dump_fill_mkey(dev->mdev,
1764 					      &resp->dump_fill_mkey);
1765 		if (err)
1766 			return err;
1767 		resp->comp_mask |=
1768 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1769 	}
1770 
1771 	resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1772 	if (dev->wc_support)
1773 		resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev,
1774 						      log_bf_reg_size);
1775 	resp->cache_line_size = cache_line_size();
1776 	resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1777 	resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1778 	resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1779 	resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1780 	resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1781 	resp->cqe_version = context->cqe_version;
1782 	resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1783 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1784 	resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1785 					MLX5_CAP_GEN(dev->mdev,
1786 						     num_of_uars_per_page) : 1;
1787 	resp->tot_bfregs = bfregi->lib_uar_dyn ? 0 :
1788 			bfregi->total_num_bfregs - bfregi->num_dyn_bfregs;
1789 	resp->num_ports = dev->num_ports;
1790 	resp->cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1791 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1792 
1793 	if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1794 		mlx5_query_min_inline(dev->mdev, &resp->eth_min_inline);
1795 		resp->eth_min_inline++;
1796 	}
1797 
1798 	if (dev->mdev->clock_info)
1799 		resp->clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1800 
1801 	/*
1802 	 * We don't want to expose information from the PCI bar that is located
1803 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1804 	 * pretend we don't support reading the HCA's core clock. This is also
1805 	 * forced by mmap function.
1806 	 */
1807 	if (PAGE_SIZE <= 4096) {
1808 		resp->comp_mask |=
1809 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1810 		resp->hca_core_clock_offset =
1811 			offsetof(struct mlx5_init_seg,
1812 				 internal_timer_h) % PAGE_SIZE;
1813 	}
1814 
1815 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
1816 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE;
1817 
1818 	if (rt_supported(MLX5_CAP_GEN(dev->mdev, sq_ts_format)) &&
1819 	    rt_supported(MLX5_CAP_GEN(dev->mdev, rq_ts_format)) &&
1820 	    rt_supported(MLX5_CAP_ROCE(dev->mdev, qp_ts_format)))
1821 		resp->comp_mask |=
1822 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_REAL_TIME_TS;
1823 
1824 	resp->num_dyn_bfregs = bfregi->num_dyn_bfregs;
1825 
1826 	if (MLX5_CAP_GEN(dev->mdev, drain_sigerr))
1827 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_SQD2RTS;
1828 
1829 	return 0;
1830 }
1831 
1832 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1833 				  struct ib_udata *udata)
1834 {
1835 	struct ib_device *ibdev = uctx->device;
1836 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1837 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1838 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1839 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1840 	struct mlx5_bfreg_info *bfregi;
1841 	int ver;
1842 	int err;
1843 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1844 				     max_cqe_version);
1845 	bool lib_uar_4k;
1846 	bool lib_uar_dyn;
1847 
1848 	if (!dev->ib_active)
1849 		return -EAGAIN;
1850 
1851 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1852 		ver = 0;
1853 	else if (udata->inlen >= min_req_v2)
1854 		ver = 2;
1855 	else
1856 		return -EINVAL;
1857 
1858 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1859 	if (err)
1860 		return err;
1861 
1862 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1863 		return -EOPNOTSUPP;
1864 
1865 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1866 		return -EOPNOTSUPP;
1867 
1868 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1869 				    MLX5_NON_FP_BFREGS_PER_UAR);
1870 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1871 		return -EINVAL;
1872 
1873 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1874 		err = mlx5_ib_devx_create(dev, true);
1875 		if (err < 0)
1876 			goto out_ctx;
1877 		context->devx_uid = err;
1878 	}
1879 
1880 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1881 	lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
1882 	bfregi = &context->bfregi;
1883 
1884 	if (lib_uar_dyn) {
1885 		bfregi->lib_uar_dyn = lib_uar_dyn;
1886 		goto uar_done;
1887 	}
1888 
1889 	/* updates req->total_num_bfregs */
1890 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1891 	if (err)
1892 		goto out_devx;
1893 
1894 	mutex_init(&bfregi->lock);
1895 	bfregi->lib_uar_4k = lib_uar_4k;
1896 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1897 				GFP_KERNEL);
1898 	if (!bfregi->count) {
1899 		err = -ENOMEM;
1900 		goto out_devx;
1901 	}
1902 
1903 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1904 				    sizeof(*bfregi->sys_pages),
1905 				    GFP_KERNEL);
1906 	if (!bfregi->sys_pages) {
1907 		err = -ENOMEM;
1908 		goto out_count;
1909 	}
1910 
1911 	err = allocate_uars(dev, context);
1912 	if (err)
1913 		goto out_sys_pages;
1914 
1915 uar_done:
1916 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1917 					     context->devx_uid);
1918 	if (err)
1919 		goto out_uars;
1920 
1921 	INIT_LIST_HEAD(&context->db_page_list);
1922 	mutex_init(&context->db_page_mutex);
1923 
1924 	context->cqe_version = min_t(__u8,
1925 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1926 				 req.max_cqe_version);
1927 
1928 	err = set_ucontext_resp(uctx, &resp);
1929 	if (err)
1930 		goto out_mdev;
1931 
1932 	resp.response_length = min(udata->outlen, sizeof(resp));
1933 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1934 	if (err)
1935 		goto out_mdev;
1936 
1937 	bfregi->ver = ver;
1938 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1939 	context->lib_caps = req.lib_caps;
1940 	print_lib_caps(dev, context->lib_caps);
1941 
1942 	if (mlx5_ib_lag_should_assign_affinity(dev)) {
1943 		u32 port = mlx5_core_native_port_num(dev->mdev) - 1;
1944 
1945 		atomic_set(&context->tx_port_affinity,
1946 			   atomic_add_return(
1947 				   1, &dev->port[port].roce.tx_port_affinity));
1948 	}
1949 
1950 	return 0;
1951 
1952 out_mdev:
1953 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1954 
1955 out_uars:
1956 	deallocate_uars(dev, context);
1957 
1958 out_sys_pages:
1959 	kfree(bfregi->sys_pages);
1960 
1961 out_count:
1962 	kfree(bfregi->count);
1963 
1964 out_devx:
1965 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1966 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1967 
1968 out_ctx:
1969 	return err;
1970 }
1971 
1972 static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext,
1973 				  struct uverbs_attr_bundle *attrs)
1974 {
1975 	struct mlx5_ib_alloc_ucontext_resp uctx_resp = {};
1976 	int ret;
1977 
1978 	ret = set_ucontext_resp(ibcontext, &uctx_resp);
1979 	if (ret)
1980 		return ret;
1981 
1982 	uctx_resp.response_length =
1983 		min_t(size_t,
1984 		      uverbs_attr_get_len(attrs,
1985 				MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX),
1986 		      sizeof(uctx_resp));
1987 
1988 	ret = uverbs_copy_to_struct_or_zero(attrs,
1989 					MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
1990 					&uctx_resp,
1991 					sizeof(uctx_resp));
1992 	return ret;
1993 }
1994 
1995 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1996 {
1997 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1998 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1999 	struct mlx5_bfreg_info *bfregi;
2000 
2001 	bfregi = &context->bfregi;
2002 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
2003 
2004 	deallocate_uars(dev, context);
2005 	kfree(bfregi->sys_pages);
2006 	kfree(bfregi->count);
2007 
2008 	if (context->devx_uid)
2009 		mlx5_ib_devx_destroy(dev, context->devx_uid);
2010 }
2011 
2012 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2013 				 int uar_idx)
2014 {
2015 	int fw_uars_per_page;
2016 
2017 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2018 
2019 	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2020 }
2021 
2022 static u64 uar_index2paddress(struct mlx5_ib_dev *dev,
2023 				 int uar_idx)
2024 {
2025 	unsigned int fw_uars_per_page;
2026 
2027 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
2028 				MLX5_UARS_IN_PAGE : 1;
2029 
2030 	return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE);
2031 }
2032 
2033 static int get_command(unsigned long offset)
2034 {
2035 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2036 }
2037 
2038 static int get_arg(unsigned long offset)
2039 {
2040 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2041 }
2042 
2043 static int get_index(unsigned long offset)
2044 {
2045 	return get_arg(offset);
2046 }
2047 
2048 /* Index resides in an extra byte to enable larger values than 255 */
2049 static int get_extended_index(unsigned long offset)
2050 {
2051 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2052 }
2053 
2054 
2055 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2056 {
2057 }
2058 
2059 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2060 {
2061 	switch (cmd) {
2062 	case MLX5_IB_MMAP_WC_PAGE:
2063 		return "WC";
2064 	case MLX5_IB_MMAP_REGULAR_PAGE:
2065 		return "best effort WC";
2066 	case MLX5_IB_MMAP_NC_PAGE:
2067 		return "NC";
2068 	case MLX5_IB_MMAP_DEVICE_MEM:
2069 		return "Device Memory";
2070 	default:
2071 		return NULL;
2072 	}
2073 }
2074 
2075 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2076 					struct vm_area_struct *vma,
2077 					struct mlx5_ib_ucontext *context)
2078 {
2079 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2080 	    !(vma->vm_flags & VM_SHARED))
2081 		return -EINVAL;
2082 
2083 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2084 		return -EOPNOTSUPP;
2085 
2086 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2087 		return -EPERM;
2088 	vma->vm_flags &= ~VM_MAYWRITE;
2089 
2090 	if (!dev->mdev->clock_info)
2091 		return -EOPNOTSUPP;
2092 
2093 	return vm_insert_page(vma, vma->vm_start,
2094 			      virt_to_page(dev->mdev->clock_info));
2095 }
2096 
2097 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2098 {
2099 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2100 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2101 	struct mlx5_var_table *var_table = &dev->var_table;
2102 	struct mlx5_ib_ucontext *context = to_mucontext(entry->ucontext);
2103 
2104 	switch (mentry->mmap_flag) {
2105 	case MLX5_IB_MMAP_TYPE_MEMIC:
2106 	case MLX5_IB_MMAP_TYPE_MEMIC_OP:
2107 		mlx5_ib_dm_mmap_free(dev, mentry);
2108 		break;
2109 	case MLX5_IB_MMAP_TYPE_VAR:
2110 		mutex_lock(&var_table->bitmap_lock);
2111 		clear_bit(mentry->page_idx, var_table->bitmap);
2112 		mutex_unlock(&var_table->bitmap_lock);
2113 		kfree(mentry);
2114 		break;
2115 	case MLX5_IB_MMAP_TYPE_UAR_WC:
2116 	case MLX5_IB_MMAP_TYPE_UAR_NC:
2117 		mlx5_cmd_uar_dealloc(dev->mdev, mentry->page_idx,
2118 				     context->devx_uid);
2119 		kfree(mentry);
2120 		break;
2121 	default:
2122 		WARN_ON(true);
2123 	}
2124 }
2125 
2126 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2127 		    struct vm_area_struct *vma,
2128 		    struct mlx5_ib_ucontext *context)
2129 {
2130 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2131 	int err;
2132 	unsigned long idx;
2133 	phys_addr_t pfn;
2134 	pgprot_t prot;
2135 	u32 bfreg_dyn_idx = 0;
2136 	u32 uar_index;
2137 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2138 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2139 				bfregi->num_static_sys_pages;
2140 
2141 	if (bfregi->lib_uar_dyn)
2142 		return -EINVAL;
2143 
2144 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2145 		return -EINVAL;
2146 
2147 	if (dyn_uar)
2148 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2149 	else
2150 		idx = get_index(vma->vm_pgoff);
2151 
2152 	if (idx >= max_valid_idx) {
2153 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2154 			     idx, max_valid_idx);
2155 		return -EINVAL;
2156 	}
2157 
2158 	switch (cmd) {
2159 	case MLX5_IB_MMAP_WC_PAGE:
2160 	case MLX5_IB_MMAP_ALLOC_WC:
2161 	case MLX5_IB_MMAP_REGULAR_PAGE:
2162 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2163 		prot = pgprot_writecombine(vma->vm_page_prot);
2164 		break;
2165 	case MLX5_IB_MMAP_NC_PAGE:
2166 		prot = pgprot_noncached(vma->vm_page_prot);
2167 		break;
2168 	default:
2169 		return -EINVAL;
2170 	}
2171 
2172 	if (dyn_uar) {
2173 		int uars_per_page;
2174 
2175 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2176 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2177 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2178 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2179 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2180 			return -EINVAL;
2181 		}
2182 
2183 		mutex_lock(&bfregi->lock);
2184 		/* Fail if uar already allocated, first bfreg index of each
2185 		 * page holds its count.
2186 		 */
2187 		if (bfregi->count[bfreg_dyn_idx]) {
2188 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2189 			mutex_unlock(&bfregi->lock);
2190 			return -EINVAL;
2191 		}
2192 
2193 		bfregi->count[bfreg_dyn_idx]++;
2194 		mutex_unlock(&bfregi->lock);
2195 
2196 		err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index,
2197 					 context->devx_uid);
2198 		if (err) {
2199 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2200 			goto free_bfreg;
2201 		}
2202 	} else {
2203 		uar_index = bfregi->sys_pages[idx];
2204 	}
2205 
2206 	pfn = uar_index2pfn(dev, uar_index);
2207 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2208 
2209 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2210 				prot, NULL);
2211 	if (err) {
2212 		mlx5_ib_err(dev,
2213 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2214 			    err, mmap_cmd2str(cmd));
2215 		goto err;
2216 	}
2217 
2218 	if (dyn_uar)
2219 		bfregi->sys_pages[idx] = uar_index;
2220 	return 0;
2221 
2222 err:
2223 	if (!dyn_uar)
2224 		return err;
2225 
2226 	mlx5_cmd_uar_dealloc(dev->mdev, idx, context->devx_uid);
2227 
2228 free_bfreg:
2229 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2230 
2231 	return err;
2232 }
2233 
2234 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2235 {
2236 	unsigned long idx;
2237 	u8 command;
2238 
2239 	command = get_command(vma->vm_pgoff);
2240 	idx = get_extended_index(vma->vm_pgoff);
2241 
2242 	return (command << 16 | idx);
2243 }
2244 
2245 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2246 			       struct vm_area_struct *vma,
2247 			       struct ib_ucontext *ucontext)
2248 {
2249 	struct mlx5_user_mmap_entry *mentry;
2250 	struct rdma_user_mmap_entry *entry;
2251 	unsigned long pgoff;
2252 	pgprot_t prot;
2253 	phys_addr_t pfn;
2254 	int ret;
2255 
2256 	pgoff = mlx5_vma_to_pgoff(vma);
2257 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2258 	if (!entry)
2259 		return -EINVAL;
2260 
2261 	mentry = to_mmmap(entry);
2262 	pfn = (mentry->address >> PAGE_SHIFT);
2263 	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
2264 	    mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
2265 		prot = pgprot_noncached(vma->vm_page_prot);
2266 	else
2267 		prot = pgprot_writecombine(vma->vm_page_prot);
2268 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2269 				entry->npages * PAGE_SIZE,
2270 				prot,
2271 				entry);
2272 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2273 	return ret;
2274 }
2275 
2276 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2277 {
2278 	u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2279 	u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2280 
2281 	return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2282 		(index & 0xFF)) << PAGE_SHIFT;
2283 }
2284 
2285 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2286 {
2287 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2288 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2289 	unsigned long command;
2290 	phys_addr_t pfn;
2291 
2292 	command = get_command(vma->vm_pgoff);
2293 	switch (command) {
2294 	case MLX5_IB_MMAP_WC_PAGE:
2295 	case MLX5_IB_MMAP_ALLOC_WC:
2296 		if (!dev->wc_support)
2297 			return -EPERM;
2298 		fallthrough;
2299 	case MLX5_IB_MMAP_NC_PAGE:
2300 	case MLX5_IB_MMAP_REGULAR_PAGE:
2301 		return uar_mmap(dev, command, vma, context);
2302 
2303 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2304 		return -ENOSYS;
2305 
2306 	case MLX5_IB_MMAP_CORE_CLOCK:
2307 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2308 			return -EINVAL;
2309 
2310 		if (vma->vm_flags & VM_WRITE)
2311 			return -EPERM;
2312 		vma->vm_flags &= ~VM_MAYWRITE;
2313 
2314 		/* Don't expose to user-space information it shouldn't have */
2315 		if (PAGE_SIZE > 4096)
2316 			return -EOPNOTSUPP;
2317 
2318 		pfn = (dev->mdev->iseg_base +
2319 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2320 			PAGE_SHIFT;
2321 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2322 					 PAGE_SIZE,
2323 					 pgprot_noncached(vma->vm_page_prot),
2324 					 NULL);
2325 	case MLX5_IB_MMAP_CLOCK_INFO:
2326 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2327 
2328 	default:
2329 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2330 	}
2331 
2332 	return 0;
2333 }
2334 
2335 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2336 {
2337 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2338 	struct ib_device *ibdev = ibpd->device;
2339 	struct mlx5_ib_alloc_pd_resp resp;
2340 	int err;
2341 	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2342 	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {};
2343 	u16 uid = 0;
2344 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2345 		udata, struct mlx5_ib_ucontext, ibucontext);
2346 
2347 	uid = context ? context->devx_uid : 0;
2348 	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2349 	MLX5_SET(alloc_pd_in, in, uid, uid);
2350 	err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out);
2351 	if (err)
2352 		return err;
2353 
2354 	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2355 	pd->uid = uid;
2356 	if (udata) {
2357 		resp.pdn = pd->pdn;
2358 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2359 			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2360 			return -EFAULT;
2361 		}
2362 	}
2363 
2364 	return 0;
2365 }
2366 
2367 static int mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2368 {
2369 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2370 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2371 
2372 	return mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2373 }
2374 
2375 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2376 {
2377 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2378 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
2379 	int err;
2380 	u16 uid;
2381 
2382 	uid = ibqp->pd ?
2383 		to_mpd(ibqp->pd)->uid : 0;
2384 
2385 	if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) {
2386 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
2387 		return -EOPNOTSUPP;
2388 	}
2389 
2390 	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2391 	if (err)
2392 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2393 			     ibqp->qp_num, gid->raw);
2394 
2395 	return err;
2396 }
2397 
2398 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2399 {
2400 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2401 	int err;
2402 	u16 uid;
2403 
2404 	uid = ibqp->pd ?
2405 		to_mpd(ibqp->pd)->uid : 0;
2406 	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2407 	if (err)
2408 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2409 			     ibqp->qp_num, gid->raw);
2410 
2411 	return err;
2412 }
2413 
2414 static int init_node_data(struct mlx5_ib_dev *dev)
2415 {
2416 	int err;
2417 
2418 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2419 	if (err)
2420 		return err;
2421 
2422 	dev->mdev->rev_id = dev->mdev->pdev->revision;
2423 
2424 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2425 }
2426 
2427 static ssize_t fw_pages_show(struct device *device,
2428 			     struct device_attribute *attr, char *buf)
2429 {
2430 	struct mlx5_ib_dev *dev =
2431 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2432 
2433 	return sysfs_emit(buf, "%d\n", dev->mdev->priv.fw_pages);
2434 }
2435 static DEVICE_ATTR_RO(fw_pages);
2436 
2437 static ssize_t reg_pages_show(struct device *device,
2438 			      struct device_attribute *attr, char *buf)
2439 {
2440 	struct mlx5_ib_dev *dev =
2441 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2442 
2443 	return sysfs_emit(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2444 }
2445 static DEVICE_ATTR_RO(reg_pages);
2446 
2447 static ssize_t hca_type_show(struct device *device,
2448 			     struct device_attribute *attr, char *buf)
2449 {
2450 	struct mlx5_ib_dev *dev =
2451 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2452 
2453 	return sysfs_emit(buf, "MT%d\n", dev->mdev->pdev->device);
2454 }
2455 static DEVICE_ATTR_RO(hca_type);
2456 
2457 static ssize_t hw_rev_show(struct device *device,
2458 			   struct device_attribute *attr, char *buf)
2459 {
2460 	struct mlx5_ib_dev *dev =
2461 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2462 
2463 	return sysfs_emit(buf, "%x\n", dev->mdev->rev_id);
2464 }
2465 static DEVICE_ATTR_RO(hw_rev);
2466 
2467 static ssize_t board_id_show(struct device *device,
2468 			     struct device_attribute *attr, char *buf)
2469 {
2470 	struct mlx5_ib_dev *dev =
2471 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2472 
2473 	return sysfs_emit(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2474 			  dev->mdev->board_id);
2475 }
2476 static DEVICE_ATTR_RO(board_id);
2477 
2478 static struct attribute *mlx5_class_attributes[] = {
2479 	&dev_attr_hw_rev.attr,
2480 	&dev_attr_hca_type.attr,
2481 	&dev_attr_board_id.attr,
2482 	&dev_attr_fw_pages.attr,
2483 	&dev_attr_reg_pages.attr,
2484 	NULL,
2485 };
2486 
2487 static const struct attribute_group mlx5_attr_group = {
2488 	.attrs = mlx5_class_attributes,
2489 };
2490 
2491 static void pkey_change_handler(struct work_struct *work)
2492 {
2493 	struct mlx5_ib_port_resources *ports =
2494 		container_of(work, struct mlx5_ib_port_resources,
2495 			     pkey_change_work);
2496 
2497 	if (!ports->gsi)
2498 		/*
2499 		 * We got this event before device was fully configured
2500 		 * and MAD registration code wasn't called/finished yet.
2501 		 */
2502 		return;
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 	devr->p0 = ib_alloc_pd(ibdev, 0);
2799 	if (IS_ERR(devr->p0))
2800 		return PTR_ERR(devr->p0);
2801 
2802 	devr->c0 = ib_create_cq(ibdev, NULL, NULL, NULL, &cq_attr);
2803 	if (IS_ERR(devr->c0)) {
2804 		ret = PTR_ERR(devr->c0);
2805 		goto error1;
2806 	}
2807 
2808 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0);
2809 	if (ret)
2810 		goto error2;
2811 
2812 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0);
2813 	if (ret)
2814 		goto error3;
2815 
2816 	memset(&attr, 0, sizeof(attr));
2817 	attr.attr.max_sge = 1;
2818 	attr.attr.max_wr = 1;
2819 	attr.srq_type = IB_SRQT_XRC;
2820 	attr.ext.cq = devr->c0;
2821 
2822 	devr->s0 = ib_create_srq(devr->p0, &attr);
2823 	if (IS_ERR(devr->s0)) {
2824 		ret = PTR_ERR(devr->s0);
2825 		goto err_create;
2826 	}
2827 
2828 	memset(&attr, 0, sizeof(attr));
2829 	attr.attr.max_sge = 1;
2830 	attr.attr.max_wr = 1;
2831 	attr.srq_type = IB_SRQT_BASIC;
2832 
2833 	devr->s1 = ib_create_srq(devr->p0, &attr);
2834 	if (IS_ERR(devr->s1)) {
2835 		ret = PTR_ERR(devr->s1);
2836 		goto error6;
2837 	}
2838 
2839 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2840 		INIT_WORK(&devr->ports[port].pkey_change_work,
2841 			  pkey_change_handler);
2842 
2843 	return 0;
2844 
2845 error6:
2846 	ib_destroy_srq(devr->s0);
2847 err_create:
2848 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2849 error3:
2850 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2851 error2:
2852 	ib_destroy_cq(devr->c0);
2853 error1:
2854 	ib_dealloc_pd(devr->p0);
2855 	return ret;
2856 }
2857 
2858 static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev)
2859 {
2860 	struct mlx5_ib_resources *devr = &dev->devr;
2861 	int port;
2862 
2863 	/*
2864 	 * Make sure no change P_Key work items are still executing.
2865 	 *
2866 	 * At this stage, the mlx5_ib_event should be unregistered
2867 	 * and it ensures that no new works are added.
2868 	 */
2869 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2870 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2871 
2872 	ib_destroy_srq(devr->s1);
2873 	ib_destroy_srq(devr->s0);
2874 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2875 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2876 	ib_destroy_cq(devr->c0);
2877 	ib_dealloc_pd(devr->p0);
2878 }
2879 
2880 static u32 get_core_cap_flags(struct ib_device *ibdev,
2881 			      struct mlx5_hca_vport_context *rep)
2882 {
2883 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2884 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2885 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2886 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2887 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
2888 	u32 ret = 0;
2889 
2890 	if (rep->grh_required)
2891 		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
2892 
2893 	if (ll == IB_LINK_LAYER_INFINIBAND)
2894 		return ret | RDMA_CORE_PORT_IBA_IB;
2895 
2896 	if (raw_support)
2897 		ret |= RDMA_CORE_PORT_RAW_PACKET;
2898 
2899 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2900 		return ret;
2901 
2902 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2903 		return ret;
2904 
2905 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2906 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2907 
2908 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2909 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2910 
2911 	return ret;
2912 }
2913 
2914 static int mlx5_port_immutable(struct ib_device *ibdev, u32 port_num,
2915 			       struct ib_port_immutable *immutable)
2916 {
2917 	struct ib_port_attr attr;
2918 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2919 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
2920 	struct mlx5_hca_vport_context rep = {0};
2921 	int err;
2922 
2923 	err = ib_query_port(ibdev, port_num, &attr);
2924 	if (err)
2925 		return err;
2926 
2927 	if (ll == IB_LINK_LAYER_INFINIBAND) {
2928 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
2929 						   &rep);
2930 		if (err)
2931 			return err;
2932 	}
2933 
2934 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2935 	immutable->gid_tbl_len = attr.gid_tbl_len;
2936 	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
2937 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2938 
2939 	return 0;
2940 }
2941 
2942 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u32 port_num,
2943 				   struct ib_port_immutable *immutable)
2944 {
2945 	struct ib_port_attr attr;
2946 	int err;
2947 
2948 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2949 
2950 	err = ib_query_port(ibdev, port_num, &attr);
2951 	if (err)
2952 		return err;
2953 
2954 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2955 	immutable->gid_tbl_len = attr.gid_tbl_len;
2956 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2957 
2958 	return 0;
2959 }
2960 
2961 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
2962 {
2963 	struct mlx5_ib_dev *dev =
2964 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2965 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
2966 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
2967 		 fw_rev_sub(dev->mdev));
2968 }
2969 
2970 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
2971 {
2972 	struct mlx5_core_dev *mdev = dev->mdev;
2973 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
2974 								 MLX5_FLOW_NAMESPACE_LAG);
2975 	struct mlx5_flow_table *ft;
2976 	int err;
2977 
2978 	if (!ns || !mlx5_lag_is_active(mdev))
2979 		return 0;
2980 
2981 	err = mlx5_cmd_create_vport_lag(mdev);
2982 	if (err)
2983 		return err;
2984 
2985 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
2986 	if (IS_ERR(ft)) {
2987 		err = PTR_ERR(ft);
2988 		goto err_destroy_vport_lag;
2989 	}
2990 
2991 	dev->flow_db->lag_demux_ft = ft;
2992 	dev->lag_ports = mlx5_lag_get_num_ports(mdev);
2993 	dev->lag_active = true;
2994 	return 0;
2995 
2996 err_destroy_vport_lag:
2997 	mlx5_cmd_destroy_vport_lag(mdev);
2998 	return err;
2999 }
3000 
3001 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3002 {
3003 	struct mlx5_core_dev *mdev = dev->mdev;
3004 
3005 	if (dev->lag_active) {
3006 		dev->lag_active = false;
3007 
3008 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
3009 		dev->flow_db->lag_demux_ft = NULL;
3010 
3011 		mlx5_cmd_destroy_vport_lag(mdev);
3012 	}
3013 }
3014 
3015 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u32 port_num)
3016 {
3017 	int err;
3018 
3019 	dev->port[port_num].roce.nb.notifier_call = mlx5_netdev_event;
3020 	err = register_netdevice_notifier(&dev->port[port_num].roce.nb);
3021 	if (err) {
3022 		dev->port[port_num].roce.nb.notifier_call = NULL;
3023 		return err;
3024 	}
3025 
3026 	return 0;
3027 }
3028 
3029 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u32 port_num)
3030 {
3031 	if (dev->port[port_num].roce.nb.notifier_call) {
3032 		unregister_netdevice_notifier(&dev->port[port_num].roce.nb);
3033 		dev->port[port_num].roce.nb.notifier_call = NULL;
3034 	}
3035 }
3036 
3037 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3038 {
3039 	int err;
3040 
3041 	if (!dev->is_rep && dev->profile != &raw_eth_profile) {
3042 		err = mlx5_nic_vport_enable_roce(dev->mdev);
3043 		if (err)
3044 			return err;
3045 	}
3046 
3047 	err = mlx5_eth_lag_init(dev);
3048 	if (err)
3049 		goto err_disable_roce;
3050 
3051 	return 0;
3052 
3053 err_disable_roce:
3054 	if (!dev->is_rep && dev->profile != &raw_eth_profile)
3055 		mlx5_nic_vport_disable_roce(dev->mdev);
3056 
3057 	return err;
3058 }
3059 
3060 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3061 {
3062 	mlx5_eth_lag_cleanup(dev);
3063 	if (!dev->is_rep && dev->profile != &raw_eth_profile)
3064 		mlx5_nic_vport_disable_roce(dev->mdev);
3065 }
3066 
3067 static int mlx5_ib_rn_get_params(struct ib_device *device, u32 port_num,
3068 				 enum rdma_netdev_t type,
3069 				 struct rdma_netdev_alloc_params *params)
3070 {
3071 	if (type != RDMA_NETDEV_IPOIB)
3072 		return -EOPNOTSUPP;
3073 
3074 	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
3075 }
3076 
3077 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
3078 				       size_t count, loff_t *pos)
3079 {
3080 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3081 	char lbuf[20];
3082 	int len;
3083 
3084 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
3085 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
3086 }
3087 
3088 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
3089 					size_t count, loff_t *pos)
3090 {
3091 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3092 	u32 timeout;
3093 	u32 var;
3094 
3095 	if (kstrtouint_from_user(buf, count, 0, &var))
3096 		return -EFAULT;
3097 
3098 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
3099 			1000);
3100 	if (timeout != var)
3101 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
3102 			    timeout);
3103 
3104 	delay_drop->timeout = timeout;
3105 
3106 	return count;
3107 }
3108 
3109 static const struct file_operations fops_delay_drop_timeout = {
3110 	.owner	= THIS_MODULE,
3111 	.open	= simple_open,
3112 	.write	= delay_drop_timeout_write,
3113 	.read	= delay_drop_timeout_read,
3114 };
3115 
3116 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
3117 				      struct mlx5_ib_multiport_info *mpi)
3118 {
3119 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3120 	struct mlx5_ib_port *port = &ibdev->port[port_num];
3121 	int comps;
3122 	int err;
3123 	int i;
3124 
3125 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3126 
3127 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
3128 
3129 	spin_lock(&port->mp.mpi_lock);
3130 	if (!mpi->ibdev) {
3131 		spin_unlock(&port->mp.mpi_lock);
3132 		return;
3133 	}
3134 
3135 	mpi->ibdev = NULL;
3136 
3137 	spin_unlock(&port->mp.mpi_lock);
3138 	if (mpi->mdev_events.notifier_call)
3139 		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
3140 	mpi->mdev_events.notifier_call = NULL;
3141 	mlx5_remove_netdev_notifier(ibdev, port_num);
3142 	spin_lock(&port->mp.mpi_lock);
3143 
3144 	comps = mpi->mdev_refcnt;
3145 	if (comps) {
3146 		mpi->unaffiliate = true;
3147 		init_completion(&mpi->unref_comp);
3148 		spin_unlock(&port->mp.mpi_lock);
3149 
3150 		for (i = 0; i < comps; i++)
3151 			wait_for_completion(&mpi->unref_comp);
3152 
3153 		spin_lock(&port->mp.mpi_lock);
3154 		mpi->unaffiliate = false;
3155 	}
3156 
3157 	port->mp.mpi = NULL;
3158 
3159 	spin_unlock(&port->mp.mpi_lock);
3160 
3161 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
3162 
3163 	mlx5_ib_dbg(ibdev, "unaffiliated port %u\n", port_num + 1);
3164 	/* Log an error, still needed to cleanup the pointers and add
3165 	 * it back to the list.
3166 	 */
3167 	if (err)
3168 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
3169 			    port_num + 1);
3170 
3171 	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
3172 }
3173 
3174 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
3175 				    struct mlx5_ib_multiport_info *mpi)
3176 {
3177 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3178 	int err;
3179 
3180 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3181 
3182 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
3183 	if (ibdev->port[port_num].mp.mpi) {
3184 		mlx5_ib_dbg(ibdev, "port %u already affiliated.\n",
3185 			    port_num + 1);
3186 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3187 		return false;
3188 	}
3189 
3190 	ibdev->port[port_num].mp.mpi = mpi;
3191 	mpi->ibdev = ibdev;
3192 	mpi->mdev_events.notifier_call = NULL;
3193 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3194 
3195 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
3196 	if (err)
3197 		goto unbind;
3198 
3199 	err = mlx5_add_netdev_notifier(ibdev, port_num);
3200 	if (err) {
3201 		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
3202 			    port_num + 1);
3203 		goto unbind;
3204 	}
3205 
3206 	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
3207 	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
3208 
3209 	mlx5_ib_init_cong_debugfs(ibdev, port_num);
3210 
3211 	return true;
3212 
3213 unbind:
3214 	mlx5_ib_unbind_slave_port(ibdev, mpi);
3215 	return false;
3216 }
3217 
3218 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
3219 {
3220 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3221 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3222 							  port_num + 1);
3223 	struct mlx5_ib_multiport_info *mpi;
3224 	int err;
3225 	u32 i;
3226 
3227 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3228 		return 0;
3229 
3230 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
3231 						     &dev->sys_image_guid);
3232 	if (err)
3233 		return err;
3234 
3235 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3236 	if (err)
3237 		return err;
3238 
3239 	mutex_lock(&mlx5_ib_multiport_mutex);
3240 	for (i = 0; i < dev->num_ports; i++) {
3241 		bool bound = false;
3242 
3243 		/* build a stub multiport info struct for the native port. */
3244 		if (i == port_num) {
3245 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
3246 			if (!mpi) {
3247 				mutex_unlock(&mlx5_ib_multiport_mutex);
3248 				mlx5_nic_vport_disable_roce(dev->mdev);
3249 				return -ENOMEM;
3250 			}
3251 
3252 			mpi->is_master = true;
3253 			mpi->mdev = dev->mdev;
3254 			mpi->sys_image_guid = dev->sys_image_guid;
3255 			dev->port[i].mp.mpi = mpi;
3256 			mpi->ibdev = dev;
3257 			mpi = NULL;
3258 			continue;
3259 		}
3260 
3261 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
3262 				    list) {
3263 			if (dev->sys_image_guid == mpi->sys_image_guid &&
3264 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
3265 				bound = mlx5_ib_bind_slave_port(dev, mpi);
3266 			}
3267 
3268 			if (bound) {
3269 				dev_dbg(mpi->mdev->device,
3270 					"removing port from unaffiliated list.\n");
3271 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
3272 				list_del(&mpi->list);
3273 				break;
3274 			}
3275 		}
3276 		if (!bound)
3277 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
3278 				    i + 1);
3279 	}
3280 
3281 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
3282 	mutex_unlock(&mlx5_ib_multiport_mutex);
3283 	return err;
3284 }
3285 
3286 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
3287 {
3288 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3289 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3290 							  port_num + 1);
3291 	u32 i;
3292 
3293 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3294 		return;
3295 
3296 	mutex_lock(&mlx5_ib_multiport_mutex);
3297 	for (i = 0; i < dev->num_ports; i++) {
3298 		if (dev->port[i].mp.mpi) {
3299 			/* Destroy the native port stub */
3300 			if (i == port_num) {
3301 				kfree(dev->port[i].mp.mpi);
3302 				dev->port[i].mp.mpi = NULL;
3303 			} else {
3304 				mlx5_ib_dbg(dev, "unbinding port_num: %u\n",
3305 					    i + 1);
3306 				list_add_tail(&dev->port[i].mp.mpi->list,
3307 					      &mlx5_ib_unaffiliated_port_list);
3308 				mlx5_ib_unbind_slave_port(dev,
3309 							  dev->port[i].mp.mpi);
3310 			}
3311 		}
3312 	}
3313 
3314 	mlx5_ib_dbg(dev, "removing from devlist\n");
3315 	list_del(&dev->ib_dev_list);
3316 	mutex_unlock(&mlx5_ib_multiport_mutex);
3317 
3318 	mlx5_nic_vport_disable_roce(dev->mdev);
3319 }
3320 
3321 static int mmap_obj_cleanup(struct ib_uobject *uobject,
3322 			    enum rdma_remove_reason why,
3323 			    struct uverbs_attr_bundle *attrs)
3324 {
3325 	struct mlx5_user_mmap_entry *obj = uobject->object;
3326 
3327 	rdma_user_mmap_entry_remove(&obj->rdma_entry);
3328 	return 0;
3329 }
3330 
3331 static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c,
3332 					    struct mlx5_user_mmap_entry *entry,
3333 					    size_t length)
3334 {
3335 	return rdma_user_mmap_entry_insert_range(
3336 		&c->ibucontext, &entry->rdma_entry, length,
3337 		(MLX5_IB_MMAP_OFFSET_START << 16),
3338 		((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1));
3339 }
3340 
3341 static struct mlx5_user_mmap_entry *
3342 alloc_var_entry(struct mlx5_ib_ucontext *c)
3343 {
3344 	struct mlx5_user_mmap_entry *entry;
3345 	struct mlx5_var_table *var_table;
3346 	u32 page_idx;
3347 	int err;
3348 
3349 	var_table = &to_mdev(c->ibucontext.device)->var_table;
3350 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3351 	if (!entry)
3352 		return ERR_PTR(-ENOMEM);
3353 
3354 	mutex_lock(&var_table->bitmap_lock);
3355 	page_idx = find_first_zero_bit(var_table->bitmap,
3356 				       var_table->num_var_hw_entries);
3357 	if (page_idx >= var_table->num_var_hw_entries) {
3358 		err = -ENOSPC;
3359 		mutex_unlock(&var_table->bitmap_lock);
3360 		goto end;
3361 	}
3362 
3363 	set_bit(page_idx, var_table->bitmap);
3364 	mutex_unlock(&var_table->bitmap_lock);
3365 
3366 	entry->address = var_table->hw_start_addr +
3367 				(page_idx * var_table->stride_size);
3368 	entry->page_idx = page_idx;
3369 	entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
3370 
3371 	err = mlx5_rdma_user_mmap_entry_insert(c, entry,
3372 					       var_table->stride_size);
3373 	if (err)
3374 		goto err_insert;
3375 
3376 	return entry;
3377 
3378 err_insert:
3379 	mutex_lock(&var_table->bitmap_lock);
3380 	clear_bit(page_idx, var_table->bitmap);
3381 	mutex_unlock(&var_table->bitmap_lock);
3382 end:
3383 	kfree(entry);
3384 	return ERR_PTR(err);
3385 }
3386 
3387 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
3388 	struct uverbs_attr_bundle *attrs)
3389 {
3390 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3391 		attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3392 	struct mlx5_ib_ucontext *c;
3393 	struct mlx5_user_mmap_entry *entry;
3394 	u64 mmap_offset;
3395 	u32 length;
3396 	int err;
3397 
3398 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3399 	if (IS_ERR(c))
3400 		return PTR_ERR(c);
3401 
3402 	entry = alloc_var_entry(c);
3403 	if (IS_ERR(entry))
3404 		return PTR_ERR(entry);
3405 
3406 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3407 	length = entry->rdma_entry.npages * PAGE_SIZE;
3408 	uobj->object = entry;
3409 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3410 
3411 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3412 			     &mmap_offset, sizeof(mmap_offset));
3413 	if (err)
3414 		return err;
3415 
3416 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3417 			     &entry->page_idx, sizeof(entry->page_idx));
3418 	if (err)
3419 		return err;
3420 
3421 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3422 			     &length, sizeof(length));
3423 	return err;
3424 }
3425 
3426 DECLARE_UVERBS_NAMED_METHOD(
3427 	MLX5_IB_METHOD_VAR_OBJ_ALLOC,
3428 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
3429 			MLX5_IB_OBJECT_VAR,
3430 			UVERBS_ACCESS_NEW,
3431 			UA_MANDATORY),
3432 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3433 			   UVERBS_ATTR_TYPE(u32),
3434 			   UA_MANDATORY),
3435 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3436 			   UVERBS_ATTR_TYPE(u32),
3437 			   UA_MANDATORY),
3438 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3439 			    UVERBS_ATTR_TYPE(u64),
3440 			    UA_MANDATORY));
3441 
3442 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3443 	MLX5_IB_METHOD_VAR_OBJ_DESTROY,
3444 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
3445 			MLX5_IB_OBJECT_VAR,
3446 			UVERBS_ACCESS_DESTROY,
3447 			UA_MANDATORY));
3448 
3449 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
3450 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3451 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
3452 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
3453 
3454 static bool var_is_supported(struct ib_device *device)
3455 {
3456 	struct mlx5_ib_dev *dev = to_mdev(device);
3457 
3458 	return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3459 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
3460 }
3461 
3462 static struct mlx5_user_mmap_entry *
3463 alloc_uar_entry(struct mlx5_ib_ucontext *c,
3464 		enum mlx5_ib_uapi_uar_alloc_type alloc_type)
3465 {
3466 	struct mlx5_user_mmap_entry *entry;
3467 	struct mlx5_ib_dev *dev;
3468 	u32 uar_index;
3469 	int err;
3470 
3471 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3472 	if (!entry)
3473 		return ERR_PTR(-ENOMEM);
3474 
3475 	dev = to_mdev(c->ibucontext.device);
3476 	err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index, c->devx_uid);
3477 	if (err)
3478 		goto end;
3479 
3480 	entry->page_idx = uar_index;
3481 	entry->address = uar_index2paddress(dev, uar_index);
3482 	if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3483 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC;
3484 	else
3485 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC;
3486 
3487 	err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE);
3488 	if (err)
3489 		goto err_insert;
3490 
3491 	return entry;
3492 
3493 err_insert:
3494 	mlx5_cmd_uar_dealloc(dev->mdev, uar_index, c->devx_uid);
3495 end:
3496 	kfree(entry);
3497 	return ERR_PTR(err);
3498 }
3499 
3500 static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)(
3501 	struct uverbs_attr_bundle *attrs)
3502 {
3503 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3504 		attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3505 	enum mlx5_ib_uapi_uar_alloc_type alloc_type;
3506 	struct mlx5_ib_ucontext *c;
3507 	struct mlx5_user_mmap_entry *entry;
3508 	u64 mmap_offset;
3509 	u32 length;
3510 	int err;
3511 
3512 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3513 	if (IS_ERR(c))
3514 		return PTR_ERR(c);
3515 
3516 	err = uverbs_get_const(&alloc_type, attrs,
3517 			       MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE);
3518 	if (err)
3519 		return err;
3520 
3521 	if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF &&
3522 	    alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC)
3523 		return -EOPNOTSUPP;
3524 
3525 	if (!to_mdev(c->ibucontext.device)->wc_support &&
3526 	    alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3527 		return -EOPNOTSUPP;
3528 
3529 	entry = alloc_uar_entry(c, alloc_type);
3530 	if (IS_ERR(entry))
3531 		return PTR_ERR(entry);
3532 
3533 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3534 	length = entry->rdma_entry.npages * PAGE_SIZE;
3535 	uobj->object = entry;
3536 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3537 
3538 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3539 			     &mmap_offset, sizeof(mmap_offset));
3540 	if (err)
3541 		return err;
3542 
3543 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3544 			     &entry->page_idx, sizeof(entry->page_idx));
3545 	if (err)
3546 		return err;
3547 
3548 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3549 			     &length, sizeof(length));
3550 	return err;
3551 }
3552 
3553 DECLARE_UVERBS_NAMED_METHOD(
3554 	MLX5_IB_METHOD_UAR_OBJ_ALLOC,
3555 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE,
3556 			MLX5_IB_OBJECT_UAR,
3557 			UVERBS_ACCESS_NEW,
3558 			UA_MANDATORY),
3559 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE,
3560 			     enum mlx5_ib_uapi_uar_alloc_type,
3561 			     UA_MANDATORY),
3562 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3563 			   UVERBS_ATTR_TYPE(u32),
3564 			   UA_MANDATORY),
3565 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3566 			   UVERBS_ATTR_TYPE(u32),
3567 			   UA_MANDATORY),
3568 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3569 			    UVERBS_ATTR_TYPE(u64),
3570 			    UA_MANDATORY));
3571 
3572 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3573 	MLX5_IB_METHOD_UAR_OBJ_DESTROY,
3574 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE,
3575 			MLX5_IB_OBJECT_UAR,
3576 			UVERBS_ACCESS_DESTROY,
3577 			UA_MANDATORY));
3578 
3579 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR,
3580 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3581 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC),
3582 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY));
3583 
3584 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3585 	mlx5_ib_query_context,
3586 	UVERBS_OBJECT_DEVICE,
3587 	UVERBS_METHOD_QUERY_CONTEXT,
3588 	UVERBS_ATTR_PTR_OUT(
3589 		MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
3590 		UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp,
3591 				   dump_fill_mkey),
3592 		UA_MANDATORY));
3593 
3594 static const struct uapi_definition mlx5_ib_defs[] = {
3595 	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
3596 	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
3597 	UAPI_DEF_CHAIN(mlx5_ib_qos_defs),
3598 	UAPI_DEF_CHAIN(mlx5_ib_std_types_defs),
3599 	UAPI_DEF_CHAIN(mlx5_ib_dm_defs),
3600 
3601 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context),
3602 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
3603 				UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
3604 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR),
3605 	{}
3606 };
3607 
3608 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
3609 {
3610 	mlx5_ib_cleanup_multiport_master(dev);
3611 	WARN_ON(!xa_empty(&dev->odp_mkeys));
3612 	mutex_destroy(&dev->cap_mask_mutex);
3613 	WARN_ON(!xa_empty(&dev->sig_mrs));
3614 	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
3615 }
3616 
3617 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
3618 {
3619 	struct mlx5_core_dev *mdev = dev->mdev;
3620 	int err;
3621 	int i;
3622 
3623 	dev->ib_dev.node_type = RDMA_NODE_IB_CA;
3624 	dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
3625 	dev->ib_dev.phys_port_cnt = dev->num_ports;
3626 	dev->ib_dev.dev.parent = mdev->device;
3627 	dev->ib_dev.lag_flags = RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
3628 
3629 	for (i = 0; i < dev->num_ports; i++) {
3630 		spin_lock_init(&dev->port[i].mp.mpi_lock);
3631 		rwlock_init(&dev->port[i].roce.netdev_lock);
3632 		dev->port[i].roce.dev = dev;
3633 		dev->port[i].roce.native_port_num = i + 1;
3634 		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
3635 	}
3636 
3637 	err = mlx5_ib_init_multiport_master(dev);
3638 	if (err)
3639 		return err;
3640 
3641 	err = set_has_smi_cap(dev);
3642 	if (err)
3643 		goto err_mp;
3644 
3645 	err = mlx5_query_max_pkeys(&dev->ib_dev, &dev->pkey_table_len);
3646 	if (err)
3647 		goto err_mp;
3648 
3649 	if (mlx5_use_mad_ifc(dev))
3650 		get_ext_port_caps(dev);
3651 
3652 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_count(mdev);
3653 
3654 	mutex_init(&dev->cap_mask_mutex);
3655 	INIT_LIST_HEAD(&dev->qp_list);
3656 	spin_lock_init(&dev->reset_flow_resource_lock);
3657 	xa_init(&dev->odp_mkeys);
3658 	xa_init(&dev->sig_mrs);
3659 	atomic_set(&dev->mkey_var, 0);
3660 
3661 	spin_lock_init(&dev->dm.lock);
3662 	dev->dm.dev = mdev;
3663 	return 0;
3664 
3665 err_mp:
3666 	mlx5_ib_cleanup_multiport_master(dev);
3667 	return err;
3668 }
3669 
3670 static int mlx5_ib_enable_driver(struct ib_device *dev)
3671 {
3672 	struct mlx5_ib_dev *mdev = to_mdev(dev);
3673 	int ret;
3674 
3675 	ret = mlx5_ib_test_wc(mdev);
3676 	mlx5_ib_dbg(mdev, "Write-Combining %s",
3677 		    mdev->wc_support ? "supported" : "not supported");
3678 
3679 	return ret;
3680 }
3681 
3682 static const struct ib_device_ops mlx5_ib_dev_ops = {
3683 	.owner = THIS_MODULE,
3684 	.driver_id = RDMA_DRIVER_MLX5,
3685 	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
3686 
3687 	.add_gid = mlx5_ib_add_gid,
3688 	.alloc_mr = mlx5_ib_alloc_mr,
3689 	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
3690 	.alloc_pd = mlx5_ib_alloc_pd,
3691 	.alloc_ucontext = mlx5_ib_alloc_ucontext,
3692 	.attach_mcast = mlx5_ib_mcg_attach,
3693 	.check_mr_status = mlx5_ib_check_mr_status,
3694 	.create_ah = mlx5_ib_create_ah,
3695 	.create_cq = mlx5_ib_create_cq,
3696 	.create_qp = mlx5_ib_create_qp,
3697 	.create_srq = mlx5_ib_create_srq,
3698 	.create_user_ah = mlx5_ib_create_ah,
3699 	.dealloc_pd = mlx5_ib_dealloc_pd,
3700 	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
3701 	.del_gid = mlx5_ib_del_gid,
3702 	.dereg_mr = mlx5_ib_dereg_mr,
3703 	.destroy_ah = mlx5_ib_destroy_ah,
3704 	.destroy_cq = mlx5_ib_destroy_cq,
3705 	.destroy_qp = mlx5_ib_destroy_qp,
3706 	.destroy_srq = mlx5_ib_destroy_srq,
3707 	.detach_mcast = mlx5_ib_mcg_detach,
3708 	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
3709 	.drain_rq = mlx5_ib_drain_rq,
3710 	.drain_sq = mlx5_ib_drain_sq,
3711 	.device_group = &mlx5_attr_group,
3712 	.enable_driver = mlx5_ib_enable_driver,
3713 	.get_dev_fw_str = get_dev_fw_str,
3714 	.get_dma_mr = mlx5_ib_get_dma_mr,
3715 	.get_link_layer = mlx5_ib_port_link_layer,
3716 	.map_mr_sg = mlx5_ib_map_mr_sg,
3717 	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
3718 	.mmap = mlx5_ib_mmap,
3719 	.mmap_free = mlx5_ib_mmap_free,
3720 	.modify_cq = mlx5_ib_modify_cq,
3721 	.modify_device = mlx5_ib_modify_device,
3722 	.modify_port = mlx5_ib_modify_port,
3723 	.modify_qp = mlx5_ib_modify_qp,
3724 	.modify_srq = mlx5_ib_modify_srq,
3725 	.poll_cq = mlx5_ib_poll_cq,
3726 	.post_recv = mlx5_ib_post_recv_nodrain,
3727 	.post_send = mlx5_ib_post_send_nodrain,
3728 	.post_srq_recv = mlx5_ib_post_srq_recv,
3729 	.process_mad = mlx5_ib_process_mad,
3730 	.query_ah = mlx5_ib_query_ah,
3731 	.query_device = mlx5_ib_query_device,
3732 	.query_gid = mlx5_ib_query_gid,
3733 	.query_pkey = mlx5_ib_query_pkey,
3734 	.query_qp = mlx5_ib_query_qp,
3735 	.query_srq = mlx5_ib_query_srq,
3736 	.query_ucontext = mlx5_ib_query_ucontext,
3737 	.reg_user_mr = mlx5_ib_reg_user_mr,
3738 	.reg_user_mr_dmabuf = mlx5_ib_reg_user_mr_dmabuf,
3739 	.req_notify_cq = mlx5_ib_arm_cq,
3740 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
3741 	.resize_cq = mlx5_ib_resize_cq,
3742 
3743 	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
3744 	INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs),
3745 	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
3746 	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
3747 	INIT_RDMA_OBJ_SIZE(ib_qp, mlx5_ib_qp, ibqp),
3748 	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
3749 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
3750 };
3751 
3752 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
3753 	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
3754 };
3755 
3756 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
3757 	.get_vf_config = mlx5_ib_get_vf_config,
3758 	.get_vf_guid = mlx5_ib_get_vf_guid,
3759 	.get_vf_stats = mlx5_ib_get_vf_stats,
3760 	.set_vf_guid = mlx5_ib_set_vf_guid,
3761 	.set_vf_link_state = mlx5_ib_set_vf_link_state,
3762 };
3763 
3764 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
3765 	.alloc_mw = mlx5_ib_alloc_mw,
3766 	.dealloc_mw = mlx5_ib_dealloc_mw,
3767 
3768 	INIT_RDMA_OBJ_SIZE(ib_mw, mlx5_ib_mw, ibmw),
3769 };
3770 
3771 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
3772 	.alloc_xrcd = mlx5_ib_alloc_xrcd,
3773 	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
3774 
3775 	INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd),
3776 };
3777 
3778 static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
3779 {
3780 	struct mlx5_core_dev *mdev = dev->mdev;
3781 	struct mlx5_var_table *var_table = &dev->var_table;
3782 	u8 log_doorbell_bar_size;
3783 	u8 log_doorbell_stride;
3784 	u64 bar_size;
3785 
3786 	log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3787 					log_doorbell_bar_size);
3788 	log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3789 					log_doorbell_stride);
3790 	var_table->hw_start_addr = dev->mdev->bar_addr +
3791 				MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
3792 					doorbell_bar_offset);
3793 	bar_size = (1ULL << log_doorbell_bar_size) * 4096;
3794 	var_table->stride_size = 1ULL << log_doorbell_stride;
3795 	var_table->num_var_hw_entries = div_u64(bar_size,
3796 						var_table->stride_size);
3797 	mutex_init(&var_table->bitmap_lock);
3798 	var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
3799 					  GFP_KERNEL);
3800 	return (var_table->bitmap) ? 0 : -ENOMEM;
3801 }
3802 
3803 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
3804 {
3805 	bitmap_free(dev->var_table.bitmap);
3806 }
3807 
3808 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
3809 {
3810 	struct mlx5_core_dev *mdev = dev->mdev;
3811 	int err;
3812 
3813 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
3814 	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
3815 		ib_set_device_ops(&dev->ib_dev,
3816 				  &mlx5_ib_dev_ipoib_enhanced_ops);
3817 
3818 	if (mlx5_core_is_pf(mdev))
3819 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
3820 
3821 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
3822 
3823 	if (MLX5_CAP_GEN(mdev, imaicl))
3824 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
3825 
3826 	if (MLX5_CAP_GEN(mdev, xrc))
3827 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
3828 
3829 	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
3830 	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3831 	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
3832 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
3833 
3834 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
3835 
3836 	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
3837 		dev->ib_dev.driver_def = mlx5_ib_defs;
3838 
3839 	err = init_node_data(dev);
3840 	if (err)
3841 		return err;
3842 
3843 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
3844 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
3845 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
3846 		mutex_init(&dev->lb.mutex);
3847 
3848 	if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3849 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
3850 		err = mlx5_ib_init_var_table(dev);
3851 		if (err)
3852 			return err;
3853 	}
3854 
3855 	dev->ib_dev.use_cq_dim = true;
3856 
3857 	return 0;
3858 }
3859 
3860 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
3861 	.get_port_immutable = mlx5_port_immutable,
3862 	.query_port = mlx5_ib_query_port,
3863 };
3864 
3865 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
3866 {
3867 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
3868 	return 0;
3869 }
3870 
3871 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
3872 	.get_port_immutable = mlx5_port_rep_immutable,
3873 	.query_port = mlx5_ib_rep_query_port,
3874 	.query_pkey = mlx5_ib_rep_query_pkey,
3875 };
3876 
3877 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
3878 {
3879 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
3880 	return 0;
3881 }
3882 
3883 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
3884 	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
3885 	.create_wq = mlx5_ib_create_wq,
3886 	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
3887 	.destroy_wq = mlx5_ib_destroy_wq,
3888 	.get_netdev = mlx5_ib_get_netdev,
3889 	.modify_wq = mlx5_ib_modify_wq,
3890 
3891 	INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table,
3892 			   ib_rwq_ind_tbl),
3893 };
3894 
3895 static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev)
3896 {
3897 	struct mlx5_core_dev *mdev = dev->mdev;
3898 	enum rdma_link_layer ll;
3899 	int port_type_cap;
3900 	u32 port_num = 0;
3901 	int err;
3902 
3903 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3904 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3905 
3906 	if (ll == IB_LINK_LAYER_ETHERNET) {
3907 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
3908 
3909 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3910 
3911 		/* Register only for native ports */
3912 		err = mlx5_add_netdev_notifier(dev, port_num);
3913 		if (err)
3914 			return err;
3915 
3916 		err = mlx5_enable_eth(dev);
3917 		if (err)
3918 			goto cleanup;
3919 	}
3920 
3921 	return 0;
3922 cleanup:
3923 	mlx5_remove_netdev_notifier(dev, port_num);
3924 	return err;
3925 }
3926 
3927 static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev)
3928 {
3929 	struct mlx5_core_dev *mdev = dev->mdev;
3930 	enum rdma_link_layer ll;
3931 	int port_type_cap;
3932 	u32 port_num;
3933 
3934 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3935 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3936 
3937 	if (ll == IB_LINK_LAYER_ETHERNET) {
3938 		mlx5_disable_eth(dev);
3939 
3940 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3941 		mlx5_remove_netdev_notifier(dev, port_num);
3942 	}
3943 }
3944 
3945 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
3946 {
3947 	mlx5_ib_init_cong_debugfs(dev,
3948 				  mlx5_core_native_port_num(dev->mdev) - 1);
3949 	return 0;
3950 }
3951 
3952 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
3953 {
3954 	mlx5_ib_cleanup_cong_debugfs(dev,
3955 				     mlx5_core_native_port_num(dev->mdev) - 1);
3956 }
3957 
3958 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
3959 {
3960 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
3961 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
3962 }
3963 
3964 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
3965 {
3966 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
3967 }
3968 
3969 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
3970 {
3971 	int err;
3972 
3973 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
3974 	if (err)
3975 		return err;
3976 
3977 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
3978 	if (err)
3979 		mlx5_free_bfreg(dev->mdev, &dev->bfreg);
3980 
3981 	return err;
3982 }
3983 
3984 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
3985 {
3986 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
3987 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
3988 }
3989 
3990 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
3991 {
3992 	const char *name;
3993 
3994 	if (!mlx5_lag_is_active(dev->mdev))
3995 		name = "mlx5_%d";
3996 	else
3997 		name = "mlx5_bond_%d";
3998 	return ib_register_device(&dev->ib_dev, name, &dev->mdev->pdev->dev);
3999 }
4000 
4001 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
4002 {
4003 	int err;
4004 
4005 	err = mlx5_mr_cache_cleanup(dev);
4006 	if (err)
4007 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
4008 
4009 	mlx5r_umr_resource_cleanup(dev);
4010 }
4011 
4012 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
4013 {
4014 	ib_unregister_device(&dev->ib_dev);
4015 }
4016 
4017 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
4018 {
4019 	int ret;
4020 
4021 	ret = mlx5r_umr_resource_init(dev);
4022 	if (ret)
4023 		return ret;
4024 
4025 	ret = mlx5_mr_cache_init(dev);
4026 	if (ret) {
4027 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4028 		mlx5r_umr_resource_cleanup(dev);
4029 	}
4030 	return ret;
4031 }
4032 
4033 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
4034 {
4035 	struct dentry *root;
4036 
4037 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4038 		return 0;
4039 
4040 	mutex_init(&dev->delay_drop.lock);
4041 	dev->delay_drop.dev = dev;
4042 	dev->delay_drop.activate = false;
4043 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4044 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4045 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4046 	atomic_set(&dev->delay_drop.events_cnt, 0);
4047 
4048 	if (!mlx5_debugfs_root)
4049 		return 0;
4050 
4051 	root = debugfs_create_dir("delay_drop", mlx5_debugfs_get_dev_root(dev->mdev));
4052 	dev->delay_drop.dir_debugfs = root;
4053 
4054 	debugfs_create_atomic_t("num_timeout_events", 0400, root,
4055 				&dev->delay_drop.events_cnt);
4056 	debugfs_create_atomic_t("num_rqs", 0400, root,
4057 				&dev->delay_drop.rqs_cnt);
4058 	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
4059 			    &fops_delay_drop_timeout);
4060 	return 0;
4061 }
4062 
4063 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
4064 {
4065 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4066 		return;
4067 
4068 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4069 	if (!dev->delay_drop.dir_debugfs)
4070 		return;
4071 
4072 	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
4073 	dev->delay_drop.dir_debugfs = NULL;
4074 }
4075 
4076 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
4077 {
4078 	dev->mdev_events.notifier_call = mlx5_ib_event;
4079 	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
4080 	return 0;
4081 }
4082 
4083 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
4084 {
4085 	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
4086 }
4087 
4088 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
4089 		      const struct mlx5_ib_profile *profile,
4090 		      int stage)
4091 {
4092 	dev->ib_active = false;
4093 
4094 	/* Number of stages to cleanup */
4095 	while (stage) {
4096 		stage--;
4097 		if (profile->stage[stage].cleanup)
4098 			profile->stage[stage].cleanup(dev);
4099 	}
4100 
4101 	kfree(dev->port);
4102 	ib_dealloc_device(&dev->ib_dev);
4103 }
4104 
4105 int __mlx5_ib_add(struct mlx5_ib_dev *dev,
4106 		  const struct mlx5_ib_profile *profile)
4107 {
4108 	int err;
4109 	int i;
4110 
4111 	dev->profile = profile;
4112 
4113 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
4114 		if (profile->stage[i].init) {
4115 			err = profile->stage[i].init(dev);
4116 			if (err)
4117 				goto err_out;
4118 		}
4119 	}
4120 
4121 	dev->ib_active = true;
4122 	return 0;
4123 
4124 err_out:
4125 	/* Clean up stages which were initialized */
4126 	while (i) {
4127 		i--;
4128 		if (profile->stage[i].cleanup)
4129 			profile->stage[i].cleanup(dev);
4130 	}
4131 	return -ENOMEM;
4132 }
4133 
4134 static const struct mlx5_ib_profile pf_profile = {
4135 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4136 		     mlx5_ib_stage_init_init,
4137 		     mlx5_ib_stage_init_cleanup),
4138 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4139 		     mlx5_ib_fs_init,
4140 		     mlx5_ib_fs_cleanup),
4141 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4142 		     mlx5_ib_stage_caps_init,
4143 		     mlx5_ib_stage_caps_cleanup),
4144 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4145 		     mlx5_ib_stage_non_default_cb,
4146 		     NULL),
4147 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4148 		     mlx5_ib_roce_init,
4149 		     mlx5_ib_roce_cleanup),
4150 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4151 		     mlx5_init_qp_table,
4152 		     mlx5_cleanup_qp_table),
4153 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4154 		     mlx5_init_srq_table,
4155 		     mlx5_cleanup_srq_table),
4156 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4157 		     mlx5_ib_dev_res_init,
4158 		     mlx5_ib_dev_res_cleanup),
4159 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4160 		     mlx5_ib_stage_dev_notifier_init,
4161 		     mlx5_ib_stage_dev_notifier_cleanup),
4162 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
4163 		     mlx5_ib_odp_init_one,
4164 		     mlx5_ib_odp_cleanup_one),
4165 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4166 		     mlx5_ib_counters_init,
4167 		     mlx5_ib_counters_cleanup),
4168 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4169 		     mlx5_ib_stage_cong_debugfs_init,
4170 		     mlx5_ib_stage_cong_debugfs_cleanup),
4171 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4172 		     mlx5_ib_stage_uar_init,
4173 		     mlx5_ib_stage_uar_cleanup),
4174 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4175 		     mlx5_ib_stage_bfrag_init,
4176 		     mlx5_ib_stage_bfrag_cleanup),
4177 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4178 		     NULL,
4179 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4180 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4181 		     mlx5_ib_devx_init,
4182 		     mlx5_ib_devx_cleanup),
4183 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4184 		     mlx5_ib_stage_ib_reg_init,
4185 		     mlx5_ib_stage_ib_reg_cleanup),
4186 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4187 		     mlx5_ib_stage_post_ib_reg_umr_init,
4188 		     NULL),
4189 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4190 		     mlx5_ib_stage_delay_drop_init,
4191 		     mlx5_ib_stage_delay_drop_cleanup),
4192 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4193 		     mlx5_ib_restrack_init,
4194 		     NULL),
4195 };
4196 
4197 const struct mlx5_ib_profile raw_eth_profile = {
4198 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4199 		     mlx5_ib_stage_init_init,
4200 		     mlx5_ib_stage_init_cleanup),
4201 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4202 		     mlx5_ib_fs_init,
4203 		     mlx5_ib_fs_cleanup),
4204 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4205 		     mlx5_ib_stage_caps_init,
4206 		     mlx5_ib_stage_caps_cleanup),
4207 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4208 		     mlx5_ib_stage_raw_eth_non_default_cb,
4209 		     NULL),
4210 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4211 		     mlx5_ib_roce_init,
4212 		     mlx5_ib_roce_cleanup),
4213 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4214 		     mlx5_init_qp_table,
4215 		     mlx5_cleanup_qp_table),
4216 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4217 		     mlx5_init_srq_table,
4218 		     mlx5_cleanup_srq_table),
4219 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4220 		     mlx5_ib_dev_res_init,
4221 		     mlx5_ib_dev_res_cleanup),
4222 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4223 		     mlx5_ib_stage_dev_notifier_init,
4224 		     mlx5_ib_stage_dev_notifier_cleanup),
4225 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4226 		     mlx5_ib_counters_init,
4227 		     mlx5_ib_counters_cleanup),
4228 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4229 		     mlx5_ib_stage_cong_debugfs_init,
4230 		     mlx5_ib_stage_cong_debugfs_cleanup),
4231 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4232 		     mlx5_ib_stage_uar_init,
4233 		     mlx5_ib_stage_uar_cleanup),
4234 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4235 		     mlx5_ib_stage_bfrag_init,
4236 		     mlx5_ib_stage_bfrag_cleanup),
4237 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4238 		     NULL,
4239 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4240 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4241 		     mlx5_ib_devx_init,
4242 		     mlx5_ib_devx_cleanup),
4243 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4244 		     mlx5_ib_stage_ib_reg_init,
4245 		     mlx5_ib_stage_ib_reg_cleanup),
4246 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4247 		     mlx5_ib_stage_post_ib_reg_umr_init,
4248 		     NULL),
4249 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4250 		     mlx5_ib_restrack_init,
4251 		     NULL),
4252 };
4253 
4254 static int mlx5r_mp_probe(struct auxiliary_device *adev,
4255 			  const struct auxiliary_device_id *id)
4256 {
4257 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4258 	struct mlx5_core_dev *mdev = idev->mdev;
4259 	struct mlx5_ib_multiport_info *mpi;
4260 	struct mlx5_ib_dev *dev;
4261 	bool bound = false;
4262 	int err;
4263 
4264 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4265 	if (!mpi)
4266 		return -ENOMEM;
4267 
4268 	mpi->mdev = mdev;
4269 	err = mlx5_query_nic_vport_system_image_guid(mdev,
4270 						     &mpi->sys_image_guid);
4271 	if (err) {
4272 		kfree(mpi);
4273 		return err;
4274 	}
4275 
4276 	mutex_lock(&mlx5_ib_multiport_mutex);
4277 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
4278 		if (dev->sys_image_guid == mpi->sys_image_guid)
4279 			bound = mlx5_ib_bind_slave_port(dev, mpi);
4280 
4281 		if (bound) {
4282 			rdma_roce_rescan_device(&dev->ib_dev);
4283 			mpi->ibdev->ib_active = true;
4284 			break;
4285 		}
4286 	}
4287 
4288 	if (!bound) {
4289 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4290 		dev_dbg(mdev->device,
4291 			"no suitable IB device found to bind to, added to unaffiliated list.\n");
4292 	}
4293 	mutex_unlock(&mlx5_ib_multiport_mutex);
4294 
4295 	auxiliary_set_drvdata(adev, mpi);
4296 	return 0;
4297 }
4298 
4299 static void mlx5r_mp_remove(struct auxiliary_device *adev)
4300 {
4301 	struct mlx5_ib_multiport_info *mpi;
4302 
4303 	mpi = auxiliary_get_drvdata(adev);
4304 	mutex_lock(&mlx5_ib_multiport_mutex);
4305 	if (mpi->ibdev)
4306 		mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
4307 	else
4308 		list_del(&mpi->list);
4309 	mutex_unlock(&mlx5_ib_multiport_mutex);
4310 	kfree(mpi);
4311 }
4312 
4313 static int mlx5r_probe(struct auxiliary_device *adev,
4314 		       const struct auxiliary_device_id *id)
4315 {
4316 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4317 	struct mlx5_core_dev *mdev = idev->mdev;
4318 	const struct mlx5_ib_profile *profile;
4319 	int port_type_cap, num_ports, ret;
4320 	enum rdma_link_layer ll;
4321 	struct mlx5_ib_dev *dev;
4322 
4323 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4324 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4325 
4326 	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
4327 			MLX5_CAP_GEN(mdev, num_vhca_ports));
4328 	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
4329 	if (!dev)
4330 		return -ENOMEM;
4331 	dev->port = kcalloc(num_ports, sizeof(*dev->port),
4332 			     GFP_KERNEL);
4333 	if (!dev->port) {
4334 		ib_dealloc_device(&dev->ib_dev);
4335 		return -ENOMEM;
4336 	}
4337 
4338 	dev->mdev = mdev;
4339 	dev->num_ports = num_ports;
4340 
4341 	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_is_roce_init_enabled(mdev))
4342 		profile = &raw_eth_profile;
4343 	else
4344 		profile = &pf_profile;
4345 
4346 	ret = __mlx5_ib_add(dev, profile);
4347 	if (ret) {
4348 		kfree(dev->port);
4349 		ib_dealloc_device(&dev->ib_dev);
4350 		return ret;
4351 	}
4352 
4353 	auxiliary_set_drvdata(adev, dev);
4354 	return 0;
4355 }
4356 
4357 static void mlx5r_remove(struct auxiliary_device *adev)
4358 {
4359 	struct mlx5_ib_dev *dev;
4360 
4361 	dev = auxiliary_get_drvdata(adev);
4362 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
4363 }
4364 
4365 static const struct auxiliary_device_id mlx5r_mp_id_table[] = {
4366 	{ .name = MLX5_ADEV_NAME ".multiport", },
4367 	{},
4368 };
4369 
4370 static const struct auxiliary_device_id mlx5r_id_table[] = {
4371 	{ .name = MLX5_ADEV_NAME ".rdma", },
4372 	{},
4373 };
4374 
4375 MODULE_DEVICE_TABLE(auxiliary, mlx5r_mp_id_table);
4376 MODULE_DEVICE_TABLE(auxiliary, mlx5r_id_table);
4377 
4378 static struct auxiliary_driver mlx5r_mp_driver = {
4379 	.name = "multiport",
4380 	.probe = mlx5r_mp_probe,
4381 	.remove = mlx5r_mp_remove,
4382 	.id_table = mlx5r_mp_id_table,
4383 };
4384 
4385 static struct auxiliary_driver mlx5r_driver = {
4386 	.name = "rdma",
4387 	.probe = mlx5r_probe,
4388 	.remove = mlx5r_remove,
4389 	.id_table = mlx5r_id_table,
4390 };
4391 
4392 static int __init mlx5_ib_init(void)
4393 {
4394 	int ret;
4395 
4396 	xlt_emergency_page = (void *)__get_free_page(GFP_KERNEL);
4397 	if (!xlt_emergency_page)
4398 		return -ENOMEM;
4399 
4400 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
4401 	if (!mlx5_ib_event_wq) {
4402 		free_page((unsigned long)xlt_emergency_page);
4403 		return -ENOMEM;
4404 	}
4405 
4406 	mlx5_ib_odp_init();
4407 	ret = mlx5r_rep_init();
4408 	if (ret)
4409 		goto rep_err;
4410 	ret = auxiliary_driver_register(&mlx5r_mp_driver);
4411 	if (ret)
4412 		goto mp_err;
4413 	ret = auxiliary_driver_register(&mlx5r_driver);
4414 	if (ret)
4415 		goto drv_err;
4416 	return 0;
4417 
4418 drv_err:
4419 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4420 mp_err:
4421 	mlx5r_rep_cleanup();
4422 rep_err:
4423 	destroy_workqueue(mlx5_ib_event_wq);
4424 	free_page((unsigned long)xlt_emergency_page);
4425 	return ret;
4426 }
4427 
4428 static void __exit mlx5_ib_cleanup(void)
4429 {
4430 	auxiliary_driver_unregister(&mlx5r_driver);
4431 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4432 	mlx5r_rep_cleanup();
4433 
4434 	destroy_workqueue(mlx5_ib_event_wq);
4435 	free_page((unsigned long)xlt_emergency_page);
4436 }
4437 
4438 module_init(mlx5_ib_init);
4439 module_exit(mlx5_ib_cleanup);
4440