xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision 32daa5d7)
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 "fs.h"
38 #include "srq.h"
39 #include "qp.h"
40 #include "wr.h"
41 #include "restrack.h"
42 #include "counters.h"
43 #include <linux/mlx5/accel.h>
44 #include <rdma/uverbs_std_types.h>
45 #include <rdma/mlx5_user_ioctl_verbs.h>
46 #include <rdma/mlx5_user_ioctl_cmds.h>
47 #include <rdma/ib_umem_odp.h>
48 
49 #define UVERBS_MODULE_NAME mlx5_ib
50 #include <rdma/uverbs_named_ioctl.h>
51 
52 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
53 MODULE_DESCRIPTION("Mellanox 5th generation network adapters (ConnectX series) IB driver");
54 MODULE_LICENSE("Dual BSD/GPL");
55 
56 struct mlx5_ib_event_work {
57 	struct work_struct	work;
58 	union {
59 		struct mlx5_ib_dev	      *dev;
60 		struct mlx5_ib_multiport_info *mpi;
61 	};
62 	bool			is_slave;
63 	unsigned int		event;
64 	void			*param;
65 };
66 
67 enum {
68 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
69 };
70 
71 static struct workqueue_struct *mlx5_ib_event_wq;
72 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
73 static LIST_HEAD(mlx5_ib_dev_list);
74 /*
75  * This mutex should be held when accessing either of the above lists
76  */
77 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
78 
79 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
80 {
81 	struct mlx5_ib_dev *dev;
82 
83 	mutex_lock(&mlx5_ib_multiport_mutex);
84 	dev = mpi->ibdev;
85 	mutex_unlock(&mlx5_ib_multiport_mutex);
86 	return dev;
87 }
88 
89 static enum rdma_link_layer
90 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
91 {
92 	switch (port_type_cap) {
93 	case MLX5_CAP_PORT_TYPE_IB:
94 		return IB_LINK_LAYER_INFINIBAND;
95 	case MLX5_CAP_PORT_TYPE_ETH:
96 		return IB_LINK_LAYER_ETHERNET;
97 	default:
98 		return IB_LINK_LAYER_UNSPECIFIED;
99 	}
100 }
101 
102 static enum rdma_link_layer
103 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
104 {
105 	struct mlx5_ib_dev *dev = to_mdev(device);
106 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
107 
108 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
109 }
110 
111 static int get_port_state(struct ib_device *ibdev,
112 			  u8 port_num,
113 			  enum ib_port_state *state)
114 {
115 	struct ib_port_attr attr;
116 	int ret;
117 
118 	memset(&attr, 0, sizeof(attr));
119 	ret = ibdev->ops.query_port(ibdev, port_num, &attr);
120 	if (!ret)
121 		*state = attr.state;
122 	return ret;
123 }
124 
125 static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev,
126 					   struct net_device *ndev,
127 					   u8 *port_num)
128 {
129 	struct mlx5_eswitch *esw = dev->mdev->priv.eswitch;
130 	struct net_device *rep_ndev;
131 	struct mlx5_ib_port *port;
132 	int i;
133 
134 	for (i = 0; i < dev->num_ports; i++) {
135 		port  = &dev->port[i];
136 		if (!port->rep)
137 			continue;
138 
139 		read_lock(&port->roce.netdev_lock);
140 		rep_ndev = mlx5_ib_get_rep_netdev(esw,
141 						  port->rep->vport);
142 		if (rep_ndev == ndev) {
143 			read_unlock(&port->roce.netdev_lock);
144 			*port_num = i + 1;
145 			return &port->roce;
146 		}
147 		read_unlock(&port->roce.netdev_lock);
148 	}
149 
150 	return NULL;
151 }
152 
153 static int mlx5_netdev_event(struct notifier_block *this,
154 			     unsigned long event, void *ptr)
155 {
156 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
157 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
158 	u8 port_num = roce->native_port_num;
159 	struct mlx5_core_dev *mdev;
160 	struct mlx5_ib_dev *ibdev;
161 
162 	ibdev = roce->dev;
163 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
164 	if (!mdev)
165 		return NOTIFY_DONE;
166 
167 	switch (event) {
168 	case NETDEV_REGISTER:
169 		/* Should already be registered during the load */
170 		if (ibdev->is_rep)
171 			break;
172 		write_lock(&roce->netdev_lock);
173 		if (ndev->dev.parent == mdev->device)
174 			roce->netdev = ndev;
175 		write_unlock(&roce->netdev_lock);
176 		break;
177 
178 	case NETDEV_UNREGISTER:
179 		/* In case of reps, ib device goes away before the netdevs */
180 		write_lock(&roce->netdev_lock);
181 		if (roce->netdev == ndev)
182 			roce->netdev = NULL;
183 		write_unlock(&roce->netdev_lock);
184 		break;
185 
186 	case NETDEV_CHANGE:
187 	case NETDEV_UP:
188 	case NETDEV_DOWN: {
189 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
190 		struct net_device *upper = NULL;
191 
192 		if (lag_ndev) {
193 			upper = netdev_master_upper_dev_get(lag_ndev);
194 			dev_put(lag_ndev);
195 		}
196 
197 		if (ibdev->is_rep)
198 			roce = mlx5_get_rep_roce(ibdev, ndev, &port_num);
199 		if (!roce)
200 			return NOTIFY_DONE;
201 		if ((upper == ndev || (!upper && ndev == roce->netdev))
202 		    && ibdev->ib_active) {
203 			struct ib_event ibev = { };
204 			enum ib_port_state port_state;
205 
206 			if (get_port_state(&ibdev->ib_dev, port_num,
207 					   &port_state))
208 				goto done;
209 
210 			if (roce->last_port_state == port_state)
211 				goto done;
212 
213 			roce->last_port_state = port_state;
214 			ibev.device = &ibdev->ib_dev;
215 			if (port_state == IB_PORT_DOWN)
216 				ibev.event = IB_EVENT_PORT_ERR;
217 			else if (port_state == IB_PORT_ACTIVE)
218 				ibev.event = IB_EVENT_PORT_ACTIVE;
219 			else
220 				goto done;
221 
222 			ibev.element.port_num = port_num;
223 			ib_dispatch_event(&ibev);
224 		}
225 		break;
226 	}
227 
228 	default:
229 		break;
230 	}
231 done:
232 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
233 	return NOTIFY_DONE;
234 }
235 
236 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
237 					     u8 port_num)
238 {
239 	struct mlx5_ib_dev *ibdev = to_mdev(device);
240 	struct net_device *ndev;
241 	struct mlx5_core_dev *mdev;
242 
243 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
244 	if (!mdev)
245 		return NULL;
246 
247 	ndev = mlx5_lag_get_roce_netdev(mdev);
248 	if (ndev)
249 		goto out;
250 
251 	/* Ensure ndev does not disappear before we invoke dev_hold()
252 	 */
253 	read_lock(&ibdev->port[port_num - 1].roce.netdev_lock);
254 	ndev = ibdev->port[port_num - 1].roce.netdev;
255 	if (ndev)
256 		dev_hold(ndev);
257 	read_unlock(&ibdev->port[port_num - 1].roce.netdev_lock);
258 
259 out:
260 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
261 	return ndev;
262 }
263 
264 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
265 						   u8 ib_port_num,
266 						   u8 *native_port_num)
267 {
268 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
269 							  ib_port_num);
270 	struct mlx5_core_dev *mdev = NULL;
271 	struct mlx5_ib_multiport_info *mpi;
272 	struct mlx5_ib_port *port;
273 
274 	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
275 	    ll != IB_LINK_LAYER_ETHERNET) {
276 		if (native_port_num)
277 			*native_port_num = ib_port_num;
278 		return ibdev->mdev;
279 	}
280 
281 	if (native_port_num)
282 		*native_port_num = 1;
283 
284 	port = &ibdev->port[ib_port_num - 1];
285 	spin_lock(&port->mp.mpi_lock);
286 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
287 	if (mpi && !mpi->unaffiliate) {
288 		mdev = mpi->mdev;
289 		/* If it's the master no need to refcount, it'll exist
290 		 * as long as the ib_dev exists.
291 		 */
292 		if (!mpi->is_master)
293 			mpi->mdev_refcnt++;
294 	}
295 	spin_unlock(&port->mp.mpi_lock);
296 
297 	return mdev;
298 }
299 
300 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u8 port_num)
301 {
302 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
303 							  port_num);
304 	struct mlx5_ib_multiport_info *mpi;
305 	struct mlx5_ib_port *port;
306 
307 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
308 		return;
309 
310 	port = &ibdev->port[port_num - 1];
311 
312 	spin_lock(&port->mp.mpi_lock);
313 	mpi = ibdev->port[port_num - 1].mp.mpi;
314 	if (mpi->is_master)
315 		goto out;
316 
317 	mpi->mdev_refcnt--;
318 	if (mpi->unaffiliate)
319 		complete(&mpi->unref_comp);
320 out:
321 	spin_unlock(&port->mp.mpi_lock);
322 }
323 
324 static int translate_eth_legacy_proto_oper(u32 eth_proto_oper,
325 					   u16 *active_speed, u8 *active_width)
326 {
327 	switch (eth_proto_oper) {
328 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
329 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
330 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
331 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
332 		*active_width = IB_WIDTH_1X;
333 		*active_speed = IB_SPEED_SDR;
334 		break;
335 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
336 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
337 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
338 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
339 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
340 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
341 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
342 		*active_width = IB_WIDTH_1X;
343 		*active_speed = IB_SPEED_QDR;
344 		break;
345 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
346 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
347 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
348 		*active_width = IB_WIDTH_1X;
349 		*active_speed = IB_SPEED_EDR;
350 		break;
351 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
352 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
353 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
354 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
355 		*active_width = IB_WIDTH_4X;
356 		*active_speed = IB_SPEED_QDR;
357 		break;
358 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
359 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
360 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
361 		*active_width = IB_WIDTH_1X;
362 		*active_speed = IB_SPEED_HDR;
363 		break;
364 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
365 		*active_width = IB_WIDTH_4X;
366 		*active_speed = IB_SPEED_FDR;
367 		break;
368 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
369 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
370 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
371 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
372 		*active_width = IB_WIDTH_4X;
373 		*active_speed = IB_SPEED_EDR;
374 		break;
375 	default:
376 		return -EINVAL;
377 	}
378 
379 	return 0;
380 }
381 
382 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u16 *active_speed,
383 					u8 *active_width)
384 {
385 	switch (eth_proto_oper) {
386 	case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
387 	case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
388 		*active_width = IB_WIDTH_1X;
389 		*active_speed = IB_SPEED_SDR;
390 		break;
391 	case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
392 		*active_width = IB_WIDTH_1X;
393 		*active_speed = IB_SPEED_DDR;
394 		break;
395 	case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
396 		*active_width = IB_WIDTH_1X;
397 		*active_speed = IB_SPEED_QDR;
398 		break;
399 	case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
400 		*active_width = IB_WIDTH_4X;
401 		*active_speed = IB_SPEED_QDR;
402 		break;
403 	case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
404 		*active_width = IB_WIDTH_1X;
405 		*active_speed = IB_SPEED_EDR;
406 		break;
407 	case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
408 		*active_width = IB_WIDTH_2X;
409 		*active_speed = IB_SPEED_EDR;
410 		break;
411 	case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
412 		*active_width = IB_WIDTH_1X;
413 		*active_speed = IB_SPEED_HDR;
414 		break;
415 	case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
416 		*active_width = IB_WIDTH_4X;
417 		*active_speed = IB_SPEED_EDR;
418 		break;
419 	case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
420 		*active_width = IB_WIDTH_2X;
421 		*active_speed = IB_SPEED_HDR;
422 		break;
423 	case MLX5E_PROT_MASK(MLX5E_100GAUI_1_100GBASE_CR_KR):
424 		*active_width = IB_WIDTH_1X;
425 		*active_speed = IB_SPEED_NDR;
426 		break;
427 	case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4):
428 		*active_width = IB_WIDTH_4X;
429 		*active_speed = IB_SPEED_HDR;
430 		break;
431 	case MLX5E_PROT_MASK(MLX5E_200GAUI_2_200GBASE_CR2_KR2):
432 		*active_width = IB_WIDTH_2X;
433 		*active_speed = IB_SPEED_NDR;
434 		break;
435 	case MLX5E_PROT_MASK(MLX5E_400GAUI_4_400GBASE_CR4_KR4):
436 		*active_width = IB_WIDTH_4X;
437 		*active_speed = IB_SPEED_NDR;
438 		break;
439 	default:
440 		return -EINVAL;
441 	}
442 
443 	return 0;
444 }
445 
446 static int translate_eth_proto_oper(u32 eth_proto_oper, u16 *active_speed,
447 				    u8 *active_width, bool ext)
448 {
449 	return ext ?
450 		translate_eth_ext_proto_oper(eth_proto_oper, active_speed,
451 					     active_width) :
452 		translate_eth_legacy_proto_oper(eth_proto_oper, active_speed,
453 						active_width);
454 }
455 
456 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
457 				struct ib_port_attr *props)
458 {
459 	struct mlx5_ib_dev *dev = to_mdev(device);
460 	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0};
461 	struct mlx5_core_dev *mdev;
462 	struct net_device *ndev, *upper;
463 	enum ib_mtu ndev_ib_mtu;
464 	bool put_mdev = true;
465 	u32 eth_prot_oper;
466 	u8 mdev_port_num;
467 	bool ext;
468 	int err;
469 
470 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
471 	if (!mdev) {
472 		/* This means the port isn't affiliated yet. Get the
473 		 * info for the master port instead.
474 		 */
475 		put_mdev = false;
476 		mdev = dev->mdev;
477 		mdev_port_num = 1;
478 		port_num = 1;
479 	}
480 
481 	/* Possible bad flows are checked before filling out props so in case
482 	 * of an error it will still be zeroed out.
483 	 * Use native port in case of reps
484 	 */
485 	if (dev->is_rep)
486 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
487 					   1);
488 	else
489 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
490 					   mdev_port_num);
491 	if (err)
492 		goto out;
493 	ext = !!MLX5_GET_ETH_PROTO(ptys_reg, out, true, eth_proto_capability);
494 	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
495 
496 	props->active_width     = IB_WIDTH_4X;
497 	props->active_speed     = IB_SPEED_QDR;
498 
499 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
500 				 &props->active_width, ext);
501 
502 	if (!dev->is_rep && mlx5_is_roce_enabled(mdev)) {
503 		u16 qkey_viol_cntr;
504 
505 		props->port_cap_flags |= IB_PORT_CM_SUP;
506 		props->ip_gids = true;
507 		props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev,
508 						   roce_address_table_size);
509 		mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
510 		props->qkey_viol_cntr = qkey_viol_cntr;
511 	}
512 	props->max_mtu          = IB_MTU_4096;
513 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
514 	props->pkey_tbl_len     = 1;
515 	props->state            = IB_PORT_DOWN;
516 	props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
517 
518 	/* If this is a stub query for an unaffiliated port stop here */
519 	if (!put_mdev)
520 		goto out;
521 
522 	ndev = mlx5_ib_get_netdev(device, port_num);
523 	if (!ndev)
524 		goto out;
525 
526 	if (dev->lag_active) {
527 		rcu_read_lock();
528 		upper = netdev_master_upper_dev_get_rcu(ndev);
529 		if (upper) {
530 			dev_put(ndev);
531 			ndev = upper;
532 			dev_hold(ndev);
533 		}
534 		rcu_read_unlock();
535 	}
536 
537 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
538 		props->state      = IB_PORT_ACTIVE;
539 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
540 	}
541 
542 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
543 
544 	dev_put(ndev);
545 
546 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
547 out:
548 	if (put_mdev)
549 		mlx5_ib_put_native_port_mdev(dev, port_num);
550 	return err;
551 }
552 
553 static int set_roce_addr(struct mlx5_ib_dev *dev, u8 port_num,
554 			 unsigned int index, const union ib_gid *gid,
555 			 const struct ib_gid_attr *attr)
556 {
557 	enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
558 	u16 vlan_id = 0xffff;
559 	u8 roce_version = 0;
560 	u8 roce_l3_type = 0;
561 	u8 mac[ETH_ALEN];
562 	int ret;
563 
564 	if (gid) {
565 		gid_type = attr->gid_type;
566 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
567 		if (ret)
568 			return ret;
569 	}
570 
571 	switch (gid_type) {
572 	case IB_GID_TYPE_ROCE:
573 		roce_version = MLX5_ROCE_VERSION_1;
574 		break;
575 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
576 		roce_version = MLX5_ROCE_VERSION_2;
577 		if (ipv6_addr_v4mapped((void *)gid))
578 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
579 		else
580 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
581 		break;
582 
583 	default:
584 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
585 	}
586 
587 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
588 				      roce_l3_type, gid->raw, mac,
589 				      vlan_id < VLAN_CFI_MASK, vlan_id,
590 				      port_num);
591 }
592 
593 static int mlx5_ib_add_gid(const struct ib_gid_attr *attr,
594 			   __always_unused void **context)
595 {
596 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
597 			     attr->index, &attr->gid, attr);
598 }
599 
600 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
601 			   __always_unused void **context)
602 {
603 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
604 			     attr->index, NULL, NULL);
605 }
606 
607 __be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev,
608 				   const struct ib_gid_attr *attr)
609 {
610 	if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
611 		return 0;
612 
613 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
614 }
615 
616 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
617 {
618 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
619 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
620 	return 0;
621 }
622 
623 enum {
624 	MLX5_VPORT_ACCESS_METHOD_MAD,
625 	MLX5_VPORT_ACCESS_METHOD_HCA,
626 	MLX5_VPORT_ACCESS_METHOD_NIC,
627 };
628 
629 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
630 {
631 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
632 		return MLX5_VPORT_ACCESS_METHOD_MAD;
633 
634 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
635 	    IB_LINK_LAYER_ETHERNET)
636 		return MLX5_VPORT_ACCESS_METHOD_NIC;
637 
638 	return MLX5_VPORT_ACCESS_METHOD_HCA;
639 }
640 
641 static void get_atomic_caps(struct mlx5_ib_dev *dev,
642 			    u8 atomic_size_qp,
643 			    struct ib_device_attr *props)
644 {
645 	u8 tmp;
646 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
647 	u8 atomic_req_8B_endianness_mode =
648 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
649 
650 	/* Check if HW supports 8 bytes standard atomic operations and capable
651 	 * of host endianness respond
652 	 */
653 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
654 	if (((atomic_operations & tmp) == tmp) &&
655 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
656 	    (atomic_req_8B_endianness_mode)) {
657 		props->atomic_cap = IB_ATOMIC_HCA;
658 	} else {
659 		props->atomic_cap = IB_ATOMIC_NONE;
660 	}
661 }
662 
663 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
664 			       struct ib_device_attr *props)
665 {
666 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
667 
668 	get_atomic_caps(dev, atomic_size_qp, props);
669 }
670 
671 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
672 					__be64 *sys_image_guid)
673 {
674 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
675 	struct mlx5_core_dev *mdev = dev->mdev;
676 	u64 tmp;
677 	int err;
678 
679 	switch (mlx5_get_vport_access_method(ibdev)) {
680 	case MLX5_VPORT_ACCESS_METHOD_MAD:
681 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
682 							    sys_image_guid);
683 
684 	case MLX5_VPORT_ACCESS_METHOD_HCA:
685 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
686 		break;
687 
688 	case MLX5_VPORT_ACCESS_METHOD_NIC:
689 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
690 		break;
691 
692 	default:
693 		return -EINVAL;
694 	}
695 
696 	if (!err)
697 		*sys_image_guid = cpu_to_be64(tmp);
698 
699 	return err;
700 
701 }
702 
703 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
704 				u16 *max_pkeys)
705 {
706 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
707 	struct mlx5_core_dev *mdev = dev->mdev;
708 
709 	switch (mlx5_get_vport_access_method(ibdev)) {
710 	case MLX5_VPORT_ACCESS_METHOD_MAD:
711 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
712 
713 	case MLX5_VPORT_ACCESS_METHOD_HCA:
714 	case MLX5_VPORT_ACCESS_METHOD_NIC:
715 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
716 						pkey_table_size));
717 		return 0;
718 
719 	default:
720 		return -EINVAL;
721 	}
722 }
723 
724 static int mlx5_query_vendor_id(struct ib_device *ibdev,
725 				u32 *vendor_id)
726 {
727 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
728 
729 	switch (mlx5_get_vport_access_method(ibdev)) {
730 	case MLX5_VPORT_ACCESS_METHOD_MAD:
731 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
732 
733 	case MLX5_VPORT_ACCESS_METHOD_HCA:
734 	case MLX5_VPORT_ACCESS_METHOD_NIC:
735 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
736 
737 	default:
738 		return -EINVAL;
739 	}
740 }
741 
742 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
743 				__be64 *node_guid)
744 {
745 	u64 tmp;
746 	int err;
747 
748 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
749 	case MLX5_VPORT_ACCESS_METHOD_MAD:
750 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
751 
752 	case MLX5_VPORT_ACCESS_METHOD_HCA:
753 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
754 		break;
755 
756 	case MLX5_VPORT_ACCESS_METHOD_NIC:
757 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
758 		break;
759 
760 	default:
761 		return -EINVAL;
762 	}
763 
764 	if (!err)
765 		*node_guid = cpu_to_be64(tmp);
766 
767 	return err;
768 }
769 
770 struct mlx5_reg_node_desc {
771 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
772 };
773 
774 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
775 {
776 	struct mlx5_reg_node_desc in;
777 
778 	if (mlx5_use_mad_ifc(dev))
779 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
780 
781 	memset(&in, 0, sizeof(in));
782 
783 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
784 				    sizeof(struct mlx5_reg_node_desc),
785 				    MLX5_REG_NODE_DESC, 0, 0);
786 }
787 
788 static int mlx5_ib_query_device(struct ib_device *ibdev,
789 				struct ib_device_attr *props,
790 				struct ib_udata *uhw)
791 {
792 	size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
793 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
794 	struct mlx5_core_dev *mdev = dev->mdev;
795 	int err = -ENOMEM;
796 	int max_sq_desc;
797 	int max_rq_sg;
798 	int max_sq_sg;
799 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
800 	bool raw_support = !mlx5_core_mp_enabled(mdev);
801 	struct mlx5_ib_query_device_resp resp = {};
802 	size_t resp_len;
803 	u64 max_tso;
804 
805 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
806 	if (uhw_outlen && uhw_outlen < resp_len)
807 		return -EINVAL;
808 
809 	resp.response_length = resp_len;
810 
811 	if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
812 		return -EINVAL;
813 
814 	memset(props, 0, sizeof(*props));
815 	err = mlx5_query_system_image_guid(ibdev,
816 					   &props->sys_image_guid);
817 	if (err)
818 		return err;
819 
820 	props->max_pkeys = dev->pkey_table_len;
821 
822 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
823 	if (err)
824 		return err;
825 
826 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
827 		(fw_rev_min(dev->mdev) << 16) |
828 		fw_rev_sub(dev->mdev);
829 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
830 		IB_DEVICE_PORT_ACTIVE_EVENT		|
831 		IB_DEVICE_SYS_IMAGE_GUID		|
832 		IB_DEVICE_RC_RNR_NAK_GEN;
833 
834 	if (MLX5_CAP_GEN(mdev, pkv))
835 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
836 	if (MLX5_CAP_GEN(mdev, qkv))
837 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
838 	if (MLX5_CAP_GEN(mdev, apm))
839 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
840 	if (MLX5_CAP_GEN(mdev, xrc))
841 		props->device_cap_flags |= IB_DEVICE_XRC;
842 	if (MLX5_CAP_GEN(mdev, imaicl)) {
843 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
844 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
845 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
846 		/* We support 'Gappy' memory registration too */
847 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
848 	}
849 	/* IB_WR_REG_MR always requires changing the entity size with UMR */
850 	if (!MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled))
851 		props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
852 	if (MLX5_CAP_GEN(mdev, sho)) {
853 		props->device_cap_flags |= IB_DEVICE_INTEGRITY_HANDOVER;
854 		/* At this stage no support for signature handover */
855 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
856 				      IB_PROT_T10DIF_TYPE_2 |
857 				      IB_PROT_T10DIF_TYPE_3;
858 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
859 				       IB_GUARD_T10DIF_CSUM;
860 	}
861 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
862 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
863 
864 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
865 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
866 			/* Legacy bit to support old userspace libraries */
867 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
868 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
869 		}
870 
871 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
872 			props->raw_packet_caps |=
873 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
874 
875 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) {
876 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
877 			if (max_tso) {
878 				resp.tso_caps.max_tso = 1 << max_tso;
879 				resp.tso_caps.supported_qpts |=
880 					1 << IB_QPT_RAW_PACKET;
881 				resp.response_length += sizeof(resp.tso_caps);
882 			}
883 		}
884 
885 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) {
886 			resp.rss_caps.rx_hash_function =
887 						MLX5_RX_HASH_FUNC_TOEPLITZ;
888 			resp.rss_caps.rx_hash_fields_mask =
889 						MLX5_RX_HASH_SRC_IPV4 |
890 						MLX5_RX_HASH_DST_IPV4 |
891 						MLX5_RX_HASH_SRC_IPV6 |
892 						MLX5_RX_HASH_DST_IPV6 |
893 						MLX5_RX_HASH_SRC_PORT_TCP |
894 						MLX5_RX_HASH_DST_PORT_TCP |
895 						MLX5_RX_HASH_SRC_PORT_UDP |
896 						MLX5_RX_HASH_DST_PORT_UDP |
897 						MLX5_RX_HASH_INNER;
898 			if (mlx5_accel_ipsec_device_caps(dev->mdev) &
899 			    MLX5_ACCEL_IPSEC_CAP_DEVICE)
900 				resp.rss_caps.rx_hash_fields_mask |=
901 					MLX5_RX_HASH_IPSEC_SPI;
902 			resp.response_length += sizeof(resp.rss_caps);
903 		}
904 	} else {
905 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen)
906 			resp.response_length += sizeof(resp.tso_caps);
907 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen)
908 			resp.response_length += sizeof(resp.rss_caps);
909 	}
910 
911 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
912 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
913 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
914 	}
915 
916 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
917 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
918 	    raw_support)
919 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
920 
921 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
922 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
923 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
924 
925 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
926 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
927 	    raw_support) {
928 		/* Legacy bit to support old userspace libraries */
929 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
930 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
931 	}
932 
933 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
934 		props->max_dm_size =
935 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
936 	}
937 
938 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
939 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
940 
941 	if (MLX5_CAP_GEN(mdev, end_pad))
942 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
943 
944 	props->vendor_part_id	   = mdev->pdev->device;
945 	props->hw_ver		   = mdev->pdev->revision;
946 
947 	props->max_mr_size	   = ~0ull;
948 	props->page_size_cap	   = ~(min_page_size - 1);
949 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
950 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
951 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
952 		     sizeof(struct mlx5_wqe_data_seg);
953 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
954 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
955 		     sizeof(struct mlx5_wqe_raddr_seg)) /
956 		sizeof(struct mlx5_wqe_data_seg);
957 	props->max_send_sge = max_sq_sg;
958 	props->max_recv_sge = max_rq_sg;
959 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
960 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
961 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
962 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
963 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
964 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
965 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
966 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
967 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
968 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
969 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
970 	props->max_srq_sge	   = max_rq_sg - 1;
971 	props->max_fast_reg_page_list_len =
972 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
973 	props->max_pi_fast_reg_page_list_len =
974 		props->max_fast_reg_page_list_len / 2;
975 	props->max_sgl_rd =
976 		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
977 	get_atomic_caps_qp(dev, props);
978 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
979 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
980 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
981 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
982 					   props->max_mcast_grp;
983 	props->max_ah = INT_MAX;
984 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
985 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
986 
987 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
988 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
989 			props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
990 		props->odp_caps = dev->odp_caps;
991 		if (!uhw) {
992 			/* ODP for kernel QPs is not implemented for receive
993 			 * WQEs and SRQ WQEs
994 			 */
995 			props->odp_caps.per_transport_caps.rc_odp_caps &=
996 				~(IB_ODP_SUPPORT_READ |
997 				  IB_ODP_SUPPORT_SRQ_RECV);
998 			props->odp_caps.per_transport_caps.uc_odp_caps &=
999 				~(IB_ODP_SUPPORT_READ |
1000 				  IB_ODP_SUPPORT_SRQ_RECV);
1001 			props->odp_caps.per_transport_caps.ud_odp_caps &=
1002 				~(IB_ODP_SUPPORT_READ |
1003 				  IB_ODP_SUPPORT_SRQ_RECV);
1004 			props->odp_caps.per_transport_caps.xrc_odp_caps &=
1005 				~(IB_ODP_SUPPORT_READ |
1006 				  IB_ODP_SUPPORT_SRQ_RECV);
1007 		}
1008 	}
1009 
1010 	if (MLX5_CAP_GEN(mdev, cd))
1011 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
1012 
1013 	if (mlx5_core_is_vf(mdev))
1014 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
1015 
1016 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1017 	    IB_LINK_LAYER_ETHERNET && raw_support) {
1018 		props->rss_caps.max_rwq_indirection_tables =
1019 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1020 		props->rss_caps.max_rwq_indirection_table_size =
1021 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1022 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1023 		props->max_wq_type_rq =
1024 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1025 	}
1026 
1027 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1028 		props->tm_caps.max_num_tags =
1029 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1030 		props->tm_caps.max_ops =
1031 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1032 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1033 	}
1034 
1035 	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1036 	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1037 		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1038 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1039 	}
1040 
1041 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1042 		props->cq_caps.max_cq_moderation_count =
1043 						MLX5_MAX_CQ_COUNT;
1044 		props->cq_caps.max_cq_moderation_period =
1045 						MLX5_MAX_CQ_PERIOD;
1046 	}
1047 
1048 	if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) {
1049 		resp.response_length += sizeof(resp.cqe_comp_caps);
1050 
1051 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1052 			resp.cqe_comp_caps.max_num =
1053 				MLX5_CAP_GEN(dev->mdev,
1054 					     cqe_compression_max_num);
1055 
1056 			resp.cqe_comp_caps.supported_format =
1057 				MLX5_IB_CQE_RES_FORMAT_HASH |
1058 				MLX5_IB_CQE_RES_FORMAT_CSUM;
1059 
1060 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1061 				resp.cqe_comp_caps.supported_format |=
1062 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1063 		}
1064 	}
1065 
1066 	if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen &&
1067 	    raw_support) {
1068 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1069 		    MLX5_CAP_GEN(mdev, qos)) {
1070 			resp.packet_pacing_caps.qp_rate_limit_max =
1071 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1072 			resp.packet_pacing_caps.qp_rate_limit_min =
1073 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1074 			resp.packet_pacing_caps.supported_qpts |=
1075 				1 << IB_QPT_RAW_PACKET;
1076 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1077 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1078 				resp.packet_pacing_caps.cap_flags |=
1079 					MLX5_IB_PP_SUPPORT_BURST;
1080 		}
1081 		resp.response_length += sizeof(resp.packet_pacing_caps);
1082 	}
1083 
1084 	if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <=
1085 	    uhw_outlen) {
1086 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1087 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1088 				MLX5_IB_ALLOW_MPW;
1089 
1090 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1091 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1092 				MLX5_IB_SUPPORT_EMPW;
1093 
1094 		resp.response_length +=
1095 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1096 	}
1097 
1098 	if (offsetofend(typeof(resp), flags) <= uhw_outlen) {
1099 		resp.response_length += sizeof(resp.flags);
1100 
1101 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1102 			resp.flags |=
1103 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1104 
1105 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1106 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1107 		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1108 			resp.flags |=
1109 				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1110 
1111 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1112 	}
1113 
1114 	if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) {
1115 		resp.response_length += sizeof(resp.sw_parsing_caps);
1116 		if (MLX5_CAP_ETH(mdev, swp)) {
1117 			resp.sw_parsing_caps.sw_parsing_offloads |=
1118 				MLX5_IB_SW_PARSING;
1119 
1120 			if (MLX5_CAP_ETH(mdev, swp_csum))
1121 				resp.sw_parsing_caps.sw_parsing_offloads |=
1122 					MLX5_IB_SW_PARSING_CSUM;
1123 
1124 			if (MLX5_CAP_ETH(mdev, swp_lso))
1125 				resp.sw_parsing_caps.sw_parsing_offloads |=
1126 					MLX5_IB_SW_PARSING_LSO;
1127 
1128 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1129 				resp.sw_parsing_caps.supported_qpts =
1130 					BIT(IB_QPT_RAW_PACKET);
1131 		}
1132 	}
1133 
1134 	if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen &&
1135 	    raw_support) {
1136 		resp.response_length += sizeof(resp.striding_rq_caps);
1137 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1138 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1139 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1140 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1141 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1142 			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1143 				resp.striding_rq_caps
1144 					.min_single_wqe_log_num_of_strides =
1145 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1146 			else
1147 				resp.striding_rq_caps
1148 					.min_single_wqe_log_num_of_strides =
1149 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1150 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1151 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1152 			resp.striding_rq_caps.supported_qpts =
1153 				BIT(IB_QPT_RAW_PACKET);
1154 		}
1155 	}
1156 
1157 	if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) {
1158 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1159 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1160 			resp.tunnel_offloads_caps |=
1161 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1162 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1163 			resp.tunnel_offloads_caps |=
1164 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1165 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1166 			resp.tunnel_offloads_caps |=
1167 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1168 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre))
1169 			resp.tunnel_offloads_caps |=
1170 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1171 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp))
1172 			resp.tunnel_offloads_caps |=
1173 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1174 	}
1175 
1176 	if (uhw_outlen) {
1177 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1178 
1179 		if (err)
1180 			return err;
1181 	}
1182 
1183 	return 0;
1184 }
1185 
1186 static void translate_active_width(struct ib_device *ibdev, u16 active_width,
1187 				   u8 *ib_width)
1188 {
1189 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1190 
1191 	if (active_width & MLX5_PTYS_WIDTH_1X)
1192 		*ib_width = IB_WIDTH_1X;
1193 	else if (active_width & MLX5_PTYS_WIDTH_2X)
1194 		*ib_width = IB_WIDTH_2X;
1195 	else if (active_width & MLX5_PTYS_WIDTH_4X)
1196 		*ib_width = IB_WIDTH_4X;
1197 	else if (active_width & MLX5_PTYS_WIDTH_8X)
1198 		*ib_width = IB_WIDTH_8X;
1199 	else if (active_width & MLX5_PTYS_WIDTH_12X)
1200 		*ib_width = IB_WIDTH_12X;
1201 	else {
1202 		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1203 			    active_width);
1204 		*ib_width = IB_WIDTH_4X;
1205 	}
1206 
1207 	return;
1208 }
1209 
1210 static int mlx5_mtu_to_ib_mtu(int mtu)
1211 {
1212 	switch (mtu) {
1213 	case 256: return 1;
1214 	case 512: return 2;
1215 	case 1024: return 3;
1216 	case 2048: return 4;
1217 	case 4096: return 5;
1218 	default:
1219 		pr_warn("invalid mtu\n");
1220 		return -1;
1221 	}
1222 }
1223 
1224 enum ib_max_vl_num {
1225 	__IB_MAX_VL_0		= 1,
1226 	__IB_MAX_VL_0_1		= 2,
1227 	__IB_MAX_VL_0_3		= 3,
1228 	__IB_MAX_VL_0_7		= 4,
1229 	__IB_MAX_VL_0_14	= 5,
1230 };
1231 
1232 enum mlx5_vl_hw_cap {
1233 	MLX5_VL_HW_0	= 1,
1234 	MLX5_VL_HW_0_1	= 2,
1235 	MLX5_VL_HW_0_2	= 3,
1236 	MLX5_VL_HW_0_3	= 4,
1237 	MLX5_VL_HW_0_4	= 5,
1238 	MLX5_VL_HW_0_5	= 6,
1239 	MLX5_VL_HW_0_6	= 7,
1240 	MLX5_VL_HW_0_7	= 8,
1241 	MLX5_VL_HW_0_14	= 15
1242 };
1243 
1244 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1245 				u8 *max_vl_num)
1246 {
1247 	switch (vl_hw_cap) {
1248 	case MLX5_VL_HW_0:
1249 		*max_vl_num = __IB_MAX_VL_0;
1250 		break;
1251 	case MLX5_VL_HW_0_1:
1252 		*max_vl_num = __IB_MAX_VL_0_1;
1253 		break;
1254 	case MLX5_VL_HW_0_3:
1255 		*max_vl_num = __IB_MAX_VL_0_3;
1256 		break;
1257 	case MLX5_VL_HW_0_7:
1258 		*max_vl_num = __IB_MAX_VL_0_7;
1259 		break;
1260 	case MLX5_VL_HW_0_14:
1261 		*max_vl_num = __IB_MAX_VL_0_14;
1262 		break;
1263 
1264 	default:
1265 		return -EINVAL;
1266 	}
1267 
1268 	return 0;
1269 }
1270 
1271 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
1272 			       struct ib_port_attr *props)
1273 {
1274 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1275 	struct mlx5_core_dev *mdev = dev->mdev;
1276 	struct mlx5_hca_vport_context *rep;
1277 	u16 max_mtu;
1278 	u16 oper_mtu;
1279 	int err;
1280 	u16 ib_link_width_oper;
1281 	u8 vl_hw_cap;
1282 
1283 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1284 	if (!rep) {
1285 		err = -ENOMEM;
1286 		goto out;
1287 	}
1288 
1289 	/* props being zeroed by the caller, avoid zeroing it here */
1290 
1291 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1292 	if (err)
1293 		goto out;
1294 
1295 	props->lid		= rep->lid;
1296 	props->lmc		= rep->lmc;
1297 	props->sm_lid		= rep->sm_lid;
1298 	props->sm_sl		= rep->sm_sl;
1299 	props->state		= rep->vport_state;
1300 	props->phys_state	= rep->port_physical_state;
1301 	props->port_cap_flags	= rep->cap_mask1;
1302 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1303 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1304 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1305 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1306 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1307 	props->subnet_timeout	= rep->subnet_timeout;
1308 	props->init_type_reply	= rep->init_type_reply;
1309 
1310 	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1311 		props->port_cap_flags2 = rep->cap_mask2;
1312 
1313 	err = mlx5_query_ib_port_oper(mdev, &ib_link_width_oper,
1314 				      &props->active_speed, port);
1315 	if (err)
1316 		goto out;
1317 
1318 	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1319 
1320 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1321 
1322 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1323 
1324 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1325 
1326 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1327 
1328 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1329 	if (err)
1330 		goto out;
1331 
1332 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1333 				   &props->max_vl_num);
1334 out:
1335 	kfree(rep);
1336 	return err;
1337 }
1338 
1339 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
1340 		       struct ib_port_attr *props)
1341 {
1342 	unsigned int count;
1343 	int ret;
1344 
1345 	switch (mlx5_get_vport_access_method(ibdev)) {
1346 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1347 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1348 		break;
1349 
1350 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1351 		ret = mlx5_query_hca_port(ibdev, port, props);
1352 		break;
1353 
1354 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1355 		ret = mlx5_query_port_roce(ibdev, port, props);
1356 		break;
1357 
1358 	default:
1359 		ret = -EINVAL;
1360 	}
1361 
1362 	if (!ret && props) {
1363 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1364 		struct mlx5_core_dev *mdev;
1365 		bool put_mdev = true;
1366 
1367 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1368 		if (!mdev) {
1369 			/* If the port isn't affiliated yet query the master.
1370 			 * The master and slave will have the same values.
1371 			 */
1372 			mdev = dev->mdev;
1373 			port = 1;
1374 			put_mdev = false;
1375 		}
1376 		count = mlx5_core_reserved_gids_count(mdev);
1377 		if (put_mdev)
1378 			mlx5_ib_put_native_port_mdev(dev, port);
1379 		props->gid_tbl_len -= count;
1380 	}
1381 	return ret;
1382 }
1383 
1384 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u8 port,
1385 				  struct ib_port_attr *props)
1386 {
1387 	return mlx5_query_port_roce(ibdev, port, props);
1388 }
1389 
1390 static int mlx5_ib_rep_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1391 				  u16 *pkey)
1392 {
1393 	/* Default special Pkey for representor device port as per the
1394 	 * IB specification 1.3 section 10.9.1.2.
1395 	 */
1396 	*pkey = 0xffff;
1397 	return 0;
1398 }
1399 
1400 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1401 			     union ib_gid *gid)
1402 {
1403 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1404 	struct mlx5_core_dev *mdev = dev->mdev;
1405 
1406 	switch (mlx5_get_vport_access_method(ibdev)) {
1407 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1408 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1409 
1410 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1411 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1412 
1413 	default:
1414 		return -EINVAL;
1415 	}
1416 
1417 }
1418 
1419 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u8 port,
1420 				   u16 index, u16 *pkey)
1421 {
1422 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1423 	struct mlx5_core_dev *mdev;
1424 	bool put_mdev = true;
1425 	u8 mdev_port_num;
1426 	int err;
1427 
1428 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1429 	if (!mdev) {
1430 		/* The port isn't affiliated yet, get the PKey from the master
1431 		 * port. For RoCE the PKey tables will be the same.
1432 		 */
1433 		put_mdev = false;
1434 		mdev = dev->mdev;
1435 		mdev_port_num = 1;
1436 	}
1437 
1438 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1439 					index, pkey);
1440 	if (put_mdev)
1441 		mlx5_ib_put_native_port_mdev(dev, port);
1442 
1443 	return err;
1444 }
1445 
1446 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1447 			      u16 *pkey)
1448 {
1449 	switch (mlx5_get_vport_access_method(ibdev)) {
1450 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1451 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1452 
1453 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1454 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1455 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1456 	default:
1457 		return -EINVAL;
1458 	}
1459 }
1460 
1461 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1462 				 struct ib_device_modify *props)
1463 {
1464 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1465 	struct mlx5_reg_node_desc in;
1466 	struct mlx5_reg_node_desc out;
1467 	int err;
1468 
1469 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1470 		return -EOPNOTSUPP;
1471 
1472 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1473 		return 0;
1474 
1475 	/*
1476 	 * If possible, pass node desc to FW, so it can generate
1477 	 * a 144 trap.  If cmd fails, just ignore.
1478 	 */
1479 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1480 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1481 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1482 	if (err)
1483 		return err;
1484 
1485 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1486 
1487 	return err;
1488 }
1489 
1490 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1491 				u32 value)
1492 {
1493 	struct mlx5_hca_vport_context ctx = {};
1494 	struct mlx5_core_dev *mdev;
1495 	u8 mdev_port_num;
1496 	int err;
1497 
1498 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1499 	if (!mdev)
1500 		return -ENODEV;
1501 
1502 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1503 	if (err)
1504 		goto out;
1505 
1506 	if (~ctx.cap_mask1_perm & mask) {
1507 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1508 			     mask, ctx.cap_mask1_perm);
1509 		err = -EINVAL;
1510 		goto out;
1511 	}
1512 
1513 	ctx.cap_mask1 = value;
1514 	ctx.cap_mask1_perm = mask;
1515 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1516 						 0, &ctx);
1517 
1518 out:
1519 	mlx5_ib_put_native_port_mdev(dev, port_num);
1520 
1521 	return err;
1522 }
1523 
1524 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1525 			       struct ib_port_modify *props)
1526 {
1527 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1528 	struct ib_port_attr attr;
1529 	u32 tmp;
1530 	int err;
1531 	u32 change_mask;
1532 	u32 value;
1533 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1534 		      IB_LINK_LAYER_INFINIBAND);
1535 
1536 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1537 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1538 	 */
1539 	if (!is_ib)
1540 		return 0;
1541 
1542 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1543 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1544 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1545 		return set_port_caps_atomic(dev, port, change_mask, value);
1546 	}
1547 
1548 	mutex_lock(&dev->cap_mask_mutex);
1549 
1550 	err = ib_query_port(ibdev, port, &attr);
1551 	if (err)
1552 		goto out;
1553 
1554 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1555 		~props->clr_port_cap_mask;
1556 
1557 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1558 
1559 out:
1560 	mutex_unlock(&dev->cap_mask_mutex);
1561 	return err;
1562 }
1563 
1564 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1565 {
1566 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1567 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1568 }
1569 
1570 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1571 {
1572 	/* Large page with non 4k uar support might limit the dynamic size */
1573 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1574 		return MLX5_MIN_DYN_BFREGS;
1575 
1576 	return MLX5_MAX_DYN_BFREGS;
1577 }
1578 
1579 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1580 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1581 			     struct mlx5_bfreg_info *bfregi)
1582 {
1583 	int uars_per_sys_page;
1584 	int bfregs_per_sys_page;
1585 	int ref_bfregs = req->total_num_bfregs;
1586 
1587 	if (req->total_num_bfregs == 0)
1588 		return -EINVAL;
1589 
1590 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1591 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1592 
1593 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1594 		return -ENOMEM;
1595 
1596 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1597 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1598 	/* This holds the required static allocation asked by the user */
1599 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1600 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1601 		return -EINVAL;
1602 
1603 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1604 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1605 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1606 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1607 
1608 	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",
1609 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1610 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1611 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1612 		    bfregi->num_sys_pages);
1613 
1614 	return 0;
1615 }
1616 
1617 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1618 {
1619 	struct mlx5_bfreg_info *bfregi;
1620 	int err;
1621 	int i;
1622 
1623 	bfregi = &context->bfregi;
1624 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1625 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1626 		if (err)
1627 			goto error;
1628 
1629 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1630 	}
1631 
1632 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1633 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1634 
1635 	return 0;
1636 
1637 error:
1638 	for (--i; i >= 0; i--)
1639 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1640 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1641 
1642 	return err;
1643 }
1644 
1645 static void deallocate_uars(struct mlx5_ib_dev *dev,
1646 			    struct mlx5_ib_ucontext *context)
1647 {
1648 	struct mlx5_bfreg_info *bfregi;
1649 	int i;
1650 
1651 	bfregi = &context->bfregi;
1652 	for (i = 0; i < bfregi->num_sys_pages; i++)
1653 		if (i < bfregi->num_static_sys_pages ||
1654 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1655 			mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1656 }
1657 
1658 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1659 {
1660 	int err = 0;
1661 
1662 	mutex_lock(&dev->lb.mutex);
1663 	if (td)
1664 		dev->lb.user_td++;
1665 	if (qp)
1666 		dev->lb.qps++;
1667 
1668 	if (dev->lb.user_td == 2 ||
1669 	    dev->lb.qps == 1) {
1670 		if (!dev->lb.enabled) {
1671 			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1672 			dev->lb.enabled = true;
1673 		}
1674 	}
1675 
1676 	mutex_unlock(&dev->lb.mutex);
1677 
1678 	return err;
1679 }
1680 
1681 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1682 {
1683 	mutex_lock(&dev->lb.mutex);
1684 	if (td)
1685 		dev->lb.user_td--;
1686 	if (qp)
1687 		dev->lb.qps--;
1688 
1689 	if (dev->lb.user_td == 1 &&
1690 	    dev->lb.qps == 0) {
1691 		if (dev->lb.enabled) {
1692 			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1693 			dev->lb.enabled = false;
1694 		}
1695 	}
1696 
1697 	mutex_unlock(&dev->lb.mutex);
1698 }
1699 
1700 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1701 					  u16 uid)
1702 {
1703 	int err;
1704 
1705 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1706 		return 0;
1707 
1708 	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1709 	if (err)
1710 		return err;
1711 
1712 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1713 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1714 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1715 		return err;
1716 
1717 	return mlx5_ib_enable_lb(dev, true, false);
1718 }
1719 
1720 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1721 					     u16 uid)
1722 {
1723 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1724 		return;
1725 
1726 	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1727 
1728 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1729 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1730 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1731 		return;
1732 
1733 	mlx5_ib_disable_lb(dev, true, false);
1734 }
1735 
1736 static int set_ucontext_resp(struct ib_ucontext *uctx,
1737 			     struct mlx5_ib_alloc_ucontext_resp *resp)
1738 {
1739 	struct ib_device *ibdev = uctx->device;
1740 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1741 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1742 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1743 	int err;
1744 
1745 	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1746 		err = mlx5_cmd_dump_fill_mkey(dev->mdev,
1747 					      &resp->dump_fill_mkey);
1748 		if (err)
1749 			return err;
1750 		resp->comp_mask |=
1751 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1752 	}
1753 
1754 	resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1755 	if (dev->wc_support)
1756 		resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev,
1757 						      log_bf_reg_size);
1758 	resp->cache_line_size = cache_line_size();
1759 	resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1760 	resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1761 	resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1762 	resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1763 	resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1764 	resp->cqe_version = context->cqe_version;
1765 	resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1766 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1767 	resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1768 					MLX5_CAP_GEN(dev->mdev,
1769 						     num_of_uars_per_page) : 1;
1770 
1771 	if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1772 				MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1773 		if (mlx5_get_flow_namespace(dev->mdev,
1774 				MLX5_FLOW_NAMESPACE_EGRESS))
1775 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1776 		if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1777 				MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1778 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1779 		if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1780 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1781 		if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1782 				MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1783 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1784 		/* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1785 	}
1786 
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 	resp->num_dyn_bfregs = bfregi->num_dyn_bfregs;
1819 	return 0;
1820 }
1821 
1822 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1823 				  struct ib_udata *udata)
1824 {
1825 	struct ib_device *ibdev = uctx->device;
1826 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1827 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1828 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1829 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1830 	struct mlx5_bfreg_info *bfregi;
1831 	int ver;
1832 	int err;
1833 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1834 				     max_cqe_version);
1835 	bool lib_uar_4k;
1836 	bool lib_uar_dyn;
1837 
1838 	if (!dev->ib_active)
1839 		return -EAGAIN;
1840 
1841 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1842 		ver = 0;
1843 	else if (udata->inlen >= min_req_v2)
1844 		ver = 2;
1845 	else
1846 		return -EINVAL;
1847 
1848 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1849 	if (err)
1850 		return err;
1851 
1852 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1853 		return -EOPNOTSUPP;
1854 
1855 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1856 		return -EOPNOTSUPP;
1857 
1858 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1859 				    MLX5_NON_FP_BFREGS_PER_UAR);
1860 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1861 		return -EINVAL;
1862 
1863 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1864 	lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
1865 	bfregi = &context->bfregi;
1866 
1867 	if (lib_uar_dyn) {
1868 		bfregi->lib_uar_dyn = lib_uar_dyn;
1869 		goto uar_done;
1870 	}
1871 
1872 	/* updates req->total_num_bfregs */
1873 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1874 	if (err)
1875 		goto out_ctx;
1876 
1877 	mutex_init(&bfregi->lock);
1878 	bfregi->lib_uar_4k = lib_uar_4k;
1879 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1880 				GFP_KERNEL);
1881 	if (!bfregi->count) {
1882 		err = -ENOMEM;
1883 		goto out_ctx;
1884 	}
1885 
1886 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1887 				    sizeof(*bfregi->sys_pages),
1888 				    GFP_KERNEL);
1889 	if (!bfregi->sys_pages) {
1890 		err = -ENOMEM;
1891 		goto out_count;
1892 	}
1893 
1894 	err = allocate_uars(dev, context);
1895 	if (err)
1896 		goto out_sys_pages;
1897 
1898 uar_done:
1899 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1900 		err = mlx5_ib_devx_create(dev, true);
1901 		if (err < 0)
1902 			goto out_uars;
1903 		context->devx_uid = err;
1904 	}
1905 
1906 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1907 					     context->devx_uid);
1908 	if (err)
1909 		goto out_devx;
1910 
1911 	INIT_LIST_HEAD(&context->db_page_list);
1912 	mutex_init(&context->db_page_mutex);
1913 
1914 	context->cqe_version = min_t(__u8,
1915 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1916 				 req.max_cqe_version);
1917 
1918 	err = set_ucontext_resp(uctx, &resp);
1919 	if (err)
1920 		goto out_mdev;
1921 
1922 	resp.response_length = min(udata->outlen, sizeof(resp));
1923 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1924 	if (err)
1925 		goto out_mdev;
1926 
1927 	bfregi->ver = ver;
1928 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1929 	context->lib_caps = req.lib_caps;
1930 	print_lib_caps(dev, context->lib_caps);
1931 
1932 	if (mlx5_ib_lag_should_assign_affinity(dev)) {
1933 		u8 port = mlx5_core_native_port_num(dev->mdev) - 1;
1934 
1935 		atomic_set(&context->tx_port_affinity,
1936 			   atomic_add_return(
1937 				   1, &dev->port[port].roce.tx_port_affinity));
1938 	}
1939 
1940 	return 0;
1941 
1942 out_mdev:
1943 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1944 out_devx:
1945 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1946 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1947 
1948 out_uars:
1949 	deallocate_uars(dev, context);
1950 
1951 out_sys_pages:
1952 	kfree(bfregi->sys_pages);
1953 
1954 out_count:
1955 	kfree(bfregi->count);
1956 
1957 out_ctx:
1958 	return err;
1959 }
1960 
1961 static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext,
1962 				  struct uverbs_attr_bundle *attrs)
1963 {
1964 	struct mlx5_ib_alloc_ucontext_resp uctx_resp = {};
1965 	int ret;
1966 
1967 	ret = set_ucontext_resp(ibcontext, &uctx_resp);
1968 	if (ret)
1969 		return ret;
1970 
1971 	uctx_resp.response_length =
1972 		min_t(size_t,
1973 		      uverbs_attr_get_len(attrs,
1974 				MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX),
1975 		      sizeof(uctx_resp));
1976 
1977 	ret = uverbs_copy_to_struct_or_zero(attrs,
1978 					MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
1979 					&uctx_resp,
1980 					sizeof(uctx_resp));
1981 	return ret;
1982 }
1983 
1984 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1985 {
1986 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1987 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1988 	struct mlx5_bfreg_info *bfregi;
1989 
1990 	bfregi = &context->bfregi;
1991 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1992 
1993 	if (context->devx_uid)
1994 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1995 
1996 	deallocate_uars(dev, context);
1997 	kfree(bfregi->sys_pages);
1998 	kfree(bfregi->count);
1999 }
2000 
2001 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2002 				 int uar_idx)
2003 {
2004 	int fw_uars_per_page;
2005 
2006 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2007 
2008 	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2009 }
2010 
2011 static u64 uar_index2paddress(struct mlx5_ib_dev *dev,
2012 				 int uar_idx)
2013 {
2014 	unsigned int fw_uars_per_page;
2015 
2016 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
2017 				MLX5_UARS_IN_PAGE : 1;
2018 
2019 	return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE);
2020 }
2021 
2022 static int get_command(unsigned long offset)
2023 {
2024 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2025 }
2026 
2027 static int get_arg(unsigned long offset)
2028 {
2029 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2030 }
2031 
2032 static int get_index(unsigned long offset)
2033 {
2034 	return get_arg(offset);
2035 }
2036 
2037 /* Index resides in an extra byte to enable larger values than 255 */
2038 static int get_extended_index(unsigned long offset)
2039 {
2040 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2041 }
2042 
2043 
2044 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2045 {
2046 }
2047 
2048 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2049 {
2050 	switch (cmd) {
2051 	case MLX5_IB_MMAP_WC_PAGE:
2052 		return "WC";
2053 	case MLX5_IB_MMAP_REGULAR_PAGE:
2054 		return "best effort WC";
2055 	case MLX5_IB_MMAP_NC_PAGE:
2056 		return "NC";
2057 	case MLX5_IB_MMAP_DEVICE_MEM:
2058 		return "Device Memory";
2059 	default:
2060 		return NULL;
2061 	}
2062 }
2063 
2064 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2065 					struct vm_area_struct *vma,
2066 					struct mlx5_ib_ucontext *context)
2067 {
2068 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2069 	    !(vma->vm_flags & VM_SHARED))
2070 		return -EINVAL;
2071 
2072 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2073 		return -EOPNOTSUPP;
2074 
2075 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2076 		return -EPERM;
2077 	vma->vm_flags &= ~VM_MAYWRITE;
2078 
2079 	if (!dev->mdev->clock_info)
2080 		return -EOPNOTSUPP;
2081 
2082 	return vm_insert_page(vma, vma->vm_start,
2083 			      virt_to_page(dev->mdev->clock_info));
2084 }
2085 
2086 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2087 {
2088 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2089 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2090 	struct mlx5_var_table *var_table = &dev->var_table;
2091 	struct mlx5_ib_dm *mdm;
2092 
2093 	switch (mentry->mmap_flag) {
2094 	case MLX5_IB_MMAP_TYPE_MEMIC:
2095 		mdm = container_of(mentry, struct mlx5_ib_dm, mentry);
2096 		mlx5_cmd_dealloc_memic(&dev->dm, mdm->dev_addr,
2097 				       mdm->size);
2098 		kfree(mdm);
2099 		break;
2100 	case MLX5_IB_MMAP_TYPE_VAR:
2101 		mutex_lock(&var_table->bitmap_lock);
2102 		clear_bit(mentry->page_idx, var_table->bitmap);
2103 		mutex_unlock(&var_table->bitmap_lock);
2104 		kfree(mentry);
2105 		break;
2106 	case MLX5_IB_MMAP_TYPE_UAR_WC:
2107 	case MLX5_IB_MMAP_TYPE_UAR_NC:
2108 		mlx5_cmd_free_uar(dev->mdev, mentry->page_idx);
2109 		kfree(mentry);
2110 		break;
2111 	default:
2112 		WARN_ON(true);
2113 	}
2114 }
2115 
2116 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2117 		    struct vm_area_struct *vma,
2118 		    struct mlx5_ib_ucontext *context)
2119 {
2120 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2121 	int err;
2122 	unsigned long idx;
2123 	phys_addr_t pfn;
2124 	pgprot_t prot;
2125 	u32 bfreg_dyn_idx = 0;
2126 	u32 uar_index;
2127 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2128 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2129 				bfregi->num_static_sys_pages;
2130 
2131 	if (bfregi->lib_uar_dyn)
2132 		return -EINVAL;
2133 
2134 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2135 		return -EINVAL;
2136 
2137 	if (dyn_uar)
2138 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2139 	else
2140 		idx = get_index(vma->vm_pgoff);
2141 
2142 	if (idx >= max_valid_idx) {
2143 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2144 			     idx, max_valid_idx);
2145 		return -EINVAL;
2146 	}
2147 
2148 	switch (cmd) {
2149 	case MLX5_IB_MMAP_WC_PAGE:
2150 	case MLX5_IB_MMAP_ALLOC_WC:
2151 	case MLX5_IB_MMAP_REGULAR_PAGE:
2152 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2153 		prot = pgprot_writecombine(vma->vm_page_prot);
2154 		break;
2155 	case MLX5_IB_MMAP_NC_PAGE:
2156 		prot = pgprot_noncached(vma->vm_page_prot);
2157 		break;
2158 	default:
2159 		return -EINVAL;
2160 	}
2161 
2162 	if (dyn_uar) {
2163 		int uars_per_page;
2164 
2165 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2166 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2167 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2168 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2169 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2170 			return -EINVAL;
2171 		}
2172 
2173 		mutex_lock(&bfregi->lock);
2174 		/* Fail if uar already allocated, first bfreg index of each
2175 		 * page holds its count.
2176 		 */
2177 		if (bfregi->count[bfreg_dyn_idx]) {
2178 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2179 			mutex_unlock(&bfregi->lock);
2180 			return -EINVAL;
2181 		}
2182 
2183 		bfregi->count[bfreg_dyn_idx]++;
2184 		mutex_unlock(&bfregi->lock);
2185 
2186 		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2187 		if (err) {
2188 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2189 			goto free_bfreg;
2190 		}
2191 	} else {
2192 		uar_index = bfregi->sys_pages[idx];
2193 	}
2194 
2195 	pfn = uar_index2pfn(dev, uar_index);
2196 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2197 
2198 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2199 				prot, NULL);
2200 	if (err) {
2201 		mlx5_ib_err(dev,
2202 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2203 			    err, mmap_cmd2str(cmd));
2204 		goto err;
2205 	}
2206 
2207 	if (dyn_uar)
2208 		bfregi->sys_pages[idx] = uar_index;
2209 	return 0;
2210 
2211 err:
2212 	if (!dyn_uar)
2213 		return err;
2214 
2215 	mlx5_cmd_free_uar(dev->mdev, idx);
2216 
2217 free_bfreg:
2218 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2219 
2220 	return err;
2221 }
2222 
2223 static int add_dm_mmap_entry(struct ib_ucontext *context,
2224 			     struct mlx5_ib_dm *mdm,
2225 			     u64 address)
2226 {
2227 	mdm->mentry.mmap_flag = MLX5_IB_MMAP_TYPE_MEMIC;
2228 	mdm->mentry.address = address;
2229 	return rdma_user_mmap_entry_insert_range(
2230 			context, &mdm->mentry.rdma_entry,
2231 			mdm->size,
2232 			MLX5_IB_MMAP_DEVICE_MEM << 16,
2233 			(MLX5_IB_MMAP_DEVICE_MEM << 16) + (1UL << 16) - 1);
2234 }
2235 
2236 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2237 {
2238 	unsigned long idx;
2239 	u8 command;
2240 
2241 	command = get_command(vma->vm_pgoff);
2242 	idx = get_extended_index(vma->vm_pgoff);
2243 
2244 	return (command << 16 | idx);
2245 }
2246 
2247 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2248 			       struct vm_area_struct *vma,
2249 			       struct ib_ucontext *ucontext)
2250 {
2251 	struct mlx5_user_mmap_entry *mentry;
2252 	struct rdma_user_mmap_entry *entry;
2253 	unsigned long pgoff;
2254 	pgprot_t prot;
2255 	phys_addr_t pfn;
2256 	int ret;
2257 
2258 	pgoff = mlx5_vma_to_pgoff(vma);
2259 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2260 	if (!entry)
2261 		return -EINVAL;
2262 
2263 	mentry = to_mmmap(entry);
2264 	pfn = (mentry->address >> PAGE_SHIFT);
2265 	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
2266 	    mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
2267 		prot = pgprot_noncached(vma->vm_page_prot);
2268 	else
2269 		prot = pgprot_writecombine(vma->vm_page_prot);
2270 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2271 				entry->npages * PAGE_SIZE,
2272 				prot,
2273 				entry);
2274 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2275 	return ret;
2276 }
2277 
2278 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2279 {
2280 	u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2281 	u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2282 
2283 	return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2284 		(index & 0xFF)) << PAGE_SHIFT;
2285 }
2286 
2287 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2288 {
2289 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2290 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2291 	unsigned long command;
2292 	phys_addr_t pfn;
2293 
2294 	command = get_command(vma->vm_pgoff);
2295 	switch (command) {
2296 	case MLX5_IB_MMAP_WC_PAGE:
2297 	case MLX5_IB_MMAP_ALLOC_WC:
2298 		if (!dev->wc_support)
2299 			return -EPERM;
2300 		fallthrough;
2301 	case MLX5_IB_MMAP_NC_PAGE:
2302 	case MLX5_IB_MMAP_REGULAR_PAGE:
2303 		return uar_mmap(dev, command, vma, context);
2304 
2305 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2306 		return -ENOSYS;
2307 
2308 	case MLX5_IB_MMAP_CORE_CLOCK:
2309 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2310 			return -EINVAL;
2311 
2312 		if (vma->vm_flags & VM_WRITE)
2313 			return -EPERM;
2314 		vma->vm_flags &= ~VM_MAYWRITE;
2315 
2316 		/* Don't expose to user-space information it shouldn't have */
2317 		if (PAGE_SIZE > 4096)
2318 			return -EOPNOTSUPP;
2319 
2320 		pfn = (dev->mdev->iseg_base +
2321 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2322 			PAGE_SHIFT;
2323 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2324 					 PAGE_SIZE,
2325 					 pgprot_noncached(vma->vm_page_prot),
2326 					 NULL);
2327 	case MLX5_IB_MMAP_CLOCK_INFO:
2328 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2329 
2330 	default:
2331 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2332 	}
2333 
2334 	return 0;
2335 }
2336 
2337 static inline int check_dm_type_support(struct mlx5_ib_dev *dev,
2338 					u32 type)
2339 {
2340 	switch (type) {
2341 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2342 		if (!MLX5_CAP_DEV_MEM(dev->mdev, memic))
2343 			return -EOPNOTSUPP;
2344 		break;
2345 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2346 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2347 		if (!capable(CAP_SYS_RAWIO) ||
2348 		    !capable(CAP_NET_RAW))
2349 			return -EPERM;
2350 
2351 		if (!(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
2352 		      MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, sw_owner) ||
2353 		      MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner_v2) ||
2354 		      MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, sw_owner_v2)))
2355 			return -EOPNOTSUPP;
2356 		break;
2357 	}
2358 
2359 	return 0;
2360 }
2361 
2362 static int handle_alloc_dm_memic(struct ib_ucontext *ctx,
2363 				 struct mlx5_ib_dm *dm,
2364 				 struct ib_dm_alloc_attr *attr,
2365 				 struct uverbs_attr_bundle *attrs)
2366 {
2367 	struct mlx5_dm *dm_db = &to_mdev(ctx->device)->dm;
2368 	u64 start_offset;
2369 	u16 page_idx;
2370 	int err;
2371 	u64 address;
2372 
2373 	dm->size = roundup(attr->length, MLX5_MEMIC_BASE_SIZE);
2374 
2375 	err = mlx5_cmd_alloc_memic(dm_db, &dm->dev_addr,
2376 				   dm->size, attr->alignment);
2377 	if (err)
2378 		return err;
2379 
2380 	address = dm->dev_addr & PAGE_MASK;
2381 	err = add_dm_mmap_entry(ctx, dm, address);
2382 	if (err)
2383 		goto err_dealloc;
2384 
2385 	page_idx = dm->mentry.rdma_entry.start_pgoff & 0xFFFF;
2386 	err = uverbs_copy_to(attrs,
2387 			     MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
2388 			     &page_idx,
2389 			     sizeof(page_idx));
2390 	if (err)
2391 		goto err_copy;
2392 
2393 	start_offset = dm->dev_addr & ~PAGE_MASK;
2394 	err = uverbs_copy_to(attrs,
2395 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2396 			     &start_offset, sizeof(start_offset));
2397 	if (err)
2398 		goto err_copy;
2399 
2400 	return 0;
2401 
2402 err_copy:
2403 	rdma_user_mmap_entry_remove(&dm->mentry.rdma_entry);
2404 err_dealloc:
2405 	mlx5_cmd_dealloc_memic(dm_db, dm->dev_addr, dm->size);
2406 
2407 	return err;
2408 }
2409 
2410 static int handle_alloc_dm_sw_icm(struct ib_ucontext *ctx,
2411 				  struct mlx5_ib_dm *dm,
2412 				  struct ib_dm_alloc_attr *attr,
2413 				  struct uverbs_attr_bundle *attrs,
2414 				  int type)
2415 {
2416 	struct mlx5_core_dev *dev = to_mdev(ctx->device)->mdev;
2417 	u64 act_size;
2418 	int err;
2419 
2420 	/* Allocation size must a multiple of the basic block size
2421 	 * and a power of 2.
2422 	 */
2423 	act_size = round_up(attr->length, MLX5_SW_ICM_BLOCK_SIZE(dev));
2424 	act_size = roundup_pow_of_two(act_size);
2425 
2426 	dm->size = act_size;
2427 	err = mlx5_dm_sw_icm_alloc(dev, type, act_size, attr->alignment,
2428 				   to_mucontext(ctx)->devx_uid, &dm->dev_addr,
2429 				   &dm->icm_dm.obj_id);
2430 	if (err)
2431 		return err;
2432 
2433 	err = uverbs_copy_to(attrs,
2434 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2435 			     &dm->dev_addr, sizeof(dm->dev_addr));
2436 	if (err)
2437 		mlx5_dm_sw_icm_dealloc(dev, type, dm->size,
2438 				       to_mucontext(ctx)->devx_uid, dm->dev_addr,
2439 				       dm->icm_dm.obj_id);
2440 
2441 	return err;
2442 }
2443 
2444 struct ib_dm *mlx5_ib_alloc_dm(struct ib_device *ibdev,
2445 			       struct ib_ucontext *context,
2446 			       struct ib_dm_alloc_attr *attr,
2447 			       struct uverbs_attr_bundle *attrs)
2448 {
2449 	struct mlx5_ib_dm *dm;
2450 	enum mlx5_ib_uapi_dm_type type;
2451 	int err;
2452 
2453 	err = uverbs_get_const_default(&type, attrs,
2454 				       MLX5_IB_ATTR_ALLOC_DM_REQ_TYPE,
2455 				       MLX5_IB_UAPI_DM_TYPE_MEMIC);
2456 	if (err)
2457 		return ERR_PTR(err);
2458 
2459 	mlx5_ib_dbg(to_mdev(ibdev), "alloc_dm req: dm_type=%d user_length=0x%llx log_alignment=%d\n",
2460 		    type, attr->length, attr->alignment);
2461 
2462 	err = check_dm_type_support(to_mdev(ibdev), type);
2463 	if (err)
2464 		return ERR_PTR(err);
2465 
2466 	dm = kzalloc(sizeof(*dm), GFP_KERNEL);
2467 	if (!dm)
2468 		return ERR_PTR(-ENOMEM);
2469 
2470 	dm->type = type;
2471 
2472 	switch (type) {
2473 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2474 		err = handle_alloc_dm_memic(context, dm,
2475 					    attr,
2476 					    attrs);
2477 		break;
2478 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2479 		err = handle_alloc_dm_sw_icm(context, dm,
2480 					     attr, attrs,
2481 					     MLX5_SW_ICM_TYPE_STEERING);
2482 		break;
2483 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2484 		err = handle_alloc_dm_sw_icm(context, dm,
2485 					     attr, attrs,
2486 					     MLX5_SW_ICM_TYPE_HEADER_MODIFY);
2487 		break;
2488 	default:
2489 		err = -EOPNOTSUPP;
2490 	}
2491 
2492 	if (err)
2493 		goto err_free;
2494 
2495 	return &dm->ibdm;
2496 
2497 err_free:
2498 	kfree(dm);
2499 	return ERR_PTR(err);
2500 }
2501 
2502 int mlx5_ib_dealloc_dm(struct ib_dm *ibdm, struct uverbs_attr_bundle *attrs)
2503 {
2504 	struct mlx5_ib_ucontext *ctx = rdma_udata_to_drv_context(
2505 		&attrs->driver_udata, struct mlx5_ib_ucontext, ibucontext);
2506 	struct mlx5_core_dev *dev = to_mdev(ibdm->device)->mdev;
2507 	struct mlx5_ib_dm *dm = to_mdm(ibdm);
2508 	int ret;
2509 
2510 	switch (dm->type) {
2511 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2512 		rdma_user_mmap_entry_remove(&dm->mentry.rdma_entry);
2513 		return 0;
2514 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2515 		ret = mlx5_dm_sw_icm_dealloc(dev, MLX5_SW_ICM_TYPE_STEERING,
2516 					     dm->size, ctx->devx_uid, dm->dev_addr,
2517 					     dm->icm_dm.obj_id);
2518 		if (ret)
2519 			return ret;
2520 		break;
2521 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2522 		ret = mlx5_dm_sw_icm_dealloc(dev, MLX5_SW_ICM_TYPE_HEADER_MODIFY,
2523 					     dm->size, ctx->devx_uid, dm->dev_addr,
2524 					     dm->icm_dm.obj_id);
2525 		if (ret)
2526 			return ret;
2527 		break;
2528 	default:
2529 		return -EOPNOTSUPP;
2530 	}
2531 
2532 	kfree(dm);
2533 
2534 	return 0;
2535 }
2536 
2537 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2538 {
2539 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2540 	struct ib_device *ibdev = ibpd->device;
2541 	struct mlx5_ib_alloc_pd_resp resp;
2542 	int err;
2543 	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2544 	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {};
2545 	u16 uid = 0;
2546 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2547 		udata, struct mlx5_ib_ucontext, ibucontext);
2548 
2549 	uid = context ? context->devx_uid : 0;
2550 	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2551 	MLX5_SET(alloc_pd_in, in, uid, uid);
2552 	err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out);
2553 	if (err)
2554 		return err;
2555 
2556 	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2557 	pd->uid = uid;
2558 	if (udata) {
2559 		resp.pdn = pd->pdn;
2560 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2561 			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2562 			return -EFAULT;
2563 		}
2564 	}
2565 
2566 	return 0;
2567 }
2568 
2569 static int mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2570 {
2571 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2572 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2573 
2574 	return mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2575 }
2576 
2577 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2578 {
2579 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2580 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
2581 	int err;
2582 	u16 uid;
2583 
2584 	uid = ibqp->pd ?
2585 		to_mpd(ibqp->pd)->uid : 0;
2586 
2587 	if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) {
2588 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
2589 		return -EOPNOTSUPP;
2590 	}
2591 
2592 	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2593 	if (err)
2594 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2595 			     ibqp->qp_num, gid->raw);
2596 
2597 	return err;
2598 }
2599 
2600 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2601 {
2602 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2603 	int err;
2604 	u16 uid;
2605 
2606 	uid = ibqp->pd ?
2607 		to_mpd(ibqp->pd)->uid : 0;
2608 	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2609 	if (err)
2610 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2611 			     ibqp->qp_num, gid->raw);
2612 
2613 	return err;
2614 }
2615 
2616 static int init_node_data(struct mlx5_ib_dev *dev)
2617 {
2618 	int err;
2619 
2620 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2621 	if (err)
2622 		return err;
2623 
2624 	dev->mdev->rev_id = dev->mdev->pdev->revision;
2625 
2626 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2627 }
2628 
2629 static ssize_t fw_pages_show(struct device *device,
2630 			     struct device_attribute *attr, char *buf)
2631 {
2632 	struct mlx5_ib_dev *dev =
2633 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2634 
2635 	return sysfs_emit(buf, "%d\n", dev->mdev->priv.fw_pages);
2636 }
2637 static DEVICE_ATTR_RO(fw_pages);
2638 
2639 static ssize_t reg_pages_show(struct device *device,
2640 			      struct device_attribute *attr, char *buf)
2641 {
2642 	struct mlx5_ib_dev *dev =
2643 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2644 
2645 	return sysfs_emit(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2646 }
2647 static DEVICE_ATTR_RO(reg_pages);
2648 
2649 static ssize_t hca_type_show(struct device *device,
2650 			     struct device_attribute *attr, char *buf)
2651 {
2652 	struct mlx5_ib_dev *dev =
2653 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2654 
2655 	return sysfs_emit(buf, "MT%d\n", dev->mdev->pdev->device);
2656 }
2657 static DEVICE_ATTR_RO(hca_type);
2658 
2659 static ssize_t hw_rev_show(struct device *device,
2660 			   struct device_attribute *attr, char *buf)
2661 {
2662 	struct mlx5_ib_dev *dev =
2663 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2664 
2665 	return sysfs_emit(buf, "%x\n", dev->mdev->rev_id);
2666 }
2667 static DEVICE_ATTR_RO(hw_rev);
2668 
2669 static ssize_t board_id_show(struct device *device,
2670 			     struct device_attribute *attr, char *buf)
2671 {
2672 	struct mlx5_ib_dev *dev =
2673 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2674 
2675 	return sysfs_emit(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2676 			  dev->mdev->board_id);
2677 }
2678 static DEVICE_ATTR_RO(board_id);
2679 
2680 static struct attribute *mlx5_class_attributes[] = {
2681 	&dev_attr_hw_rev.attr,
2682 	&dev_attr_hca_type.attr,
2683 	&dev_attr_board_id.attr,
2684 	&dev_attr_fw_pages.attr,
2685 	&dev_attr_reg_pages.attr,
2686 	NULL,
2687 };
2688 
2689 static const struct attribute_group mlx5_attr_group = {
2690 	.attrs = mlx5_class_attributes,
2691 };
2692 
2693 static void pkey_change_handler(struct work_struct *work)
2694 {
2695 	struct mlx5_ib_port_resources *ports =
2696 		container_of(work, struct mlx5_ib_port_resources,
2697 			     pkey_change_work);
2698 
2699 	mlx5_ib_gsi_pkey_change(ports->gsi);
2700 }
2701 
2702 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2703 {
2704 	struct mlx5_ib_qp *mqp;
2705 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2706 	struct mlx5_core_cq *mcq;
2707 	struct list_head cq_armed_list;
2708 	unsigned long flags_qp;
2709 	unsigned long flags_cq;
2710 	unsigned long flags;
2711 
2712 	INIT_LIST_HEAD(&cq_armed_list);
2713 
2714 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2715 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2716 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2717 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2718 		if (mqp->sq.tail != mqp->sq.head) {
2719 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2720 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2721 			if (send_mcq->mcq.comp &&
2722 			    mqp->ibqp.send_cq->comp_handler) {
2723 				if (!send_mcq->mcq.reset_notify_added) {
2724 					send_mcq->mcq.reset_notify_added = 1;
2725 					list_add_tail(&send_mcq->mcq.reset_notify,
2726 						      &cq_armed_list);
2727 				}
2728 			}
2729 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2730 		}
2731 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2732 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2733 		/* no handling is needed for SRQ */
2734 		if (!mqp->ibqp.srq) {
2735 			if (mqp->rq.tail != mqp->rq.head) {
2736 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2737 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2738 				if (recv_mcq->mcq.comp &&
2739 				    mqp->ibqp.recv_cq->comp_handler) {
2740 					if (!recv_mcq->mcq.reset_notify_added) {
2741 						recv_mcq->mcq.reset_notify_added = 1;
2742 						list_add_tail(&recv_mcq->mcq.reset_notify,
2743 							      &cq_armed_list);
2744 					}
2745 				}
2746 				spin_unlock_irqrestore(&recv_mcq->lock,
2747 						       flags_cq);
2748 			}
2749 		}
2750 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2751 	}
2752 	/*At that point all inflight post send were put to be executed as of we
2753 	 * lock/unlock above locks Now need to arm all involved CQs.
2754 	 */
2755 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2756 		mcq->comp(mcq, NULL);
2757 	}
2758 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2759 }
2760 
2761 static void delay_drop_handler(struct work_struct *work)
2762 {
2763 	int err;
2764 	struct mlx5_ib_delay_drop *delay_drop =
2765 		container_of(work, struct mlx5_ib_delay_drop,
2766 			     delay_drop_work);
2767 
2768 	atomic_inc(&delay_drop->events_cnt);
2769 
2770 	mutex_lock(&delay_drop->lock);
2771 	err = mlx5_core_set_delay_drop(delay_drop->dev, delay_drop->timeout);
2772 	if (err) {
2773 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
2774 			     delay_drop->timeout);
2775 		delay_drop->activate = false;
2776 	}
2777 	mutex_unlock(&delay_drop->lock);
2778 }
2779 
2780 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2781 				 struct ib_event *ibev)
2782 {
2783 	u8 port = (eqe->data.port.port >> 4) & 0xf;
2784 
2785 	switch (eqe->sub_type) {
2786 	case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
2787 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2788 					    IB_LINK_LAYER_ETHERNET)
2789 			schedule_work(&ibdev->delay_drop.delay_drop_work);
2790 		break;
2791 	default: /* do nothing */
2792 		return;
2793 	}
2794 }
2795 
2796 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2797 			      struct ib_event *ibev)
2798 {
2799 	u8 port = (eqe->data.port.port >> 4) & 0xf;
2800 
2801 	ibev->element.port_num = port;
2802 
2803 	switch (eqe->sub_type) {
2804 	case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
2805 	case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
2806 	case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
2807 		/* In RoCE, port up/down events are handled in
2808 		 * mlx5_netdev_event().
2809 		 */
2810 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2811 					    IB_LINK_LAYER_ETHERNET)
2812 			return -EINVAL;
2813 
2814 		ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
2815 				IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2816 		break;
2817 
2818 	case MLX5_PORT_CHANGE_SUBTYPE_LID:
2819 		ibev->event = IB_EVENT_LID_CHANGE;
2820 		break;
2821 
2822 	case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
2823 		ibev->event = IB_EVENT_PKEY_CHANGE;
2824 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2825 		break;
2826 
2827 	case MLX5_PORT_CHANGE_SUBTYPE_GUID:
2828 		ibev->event = IB_EVENT_GID_CHANGE;
2829 		break;
2830 
2831 	case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
2832 		ibev->event = IB_EVENT_CLIENT_REREGISTER;
2833 		break;
2834 	default:
2835 		return -EINVAL;
2836 	}
2837 
2838 	return 0;
2839 }
2840 
2841 static void mlx5_ib_handle_event(struct work_struct *_work)
2842 {
2843 	struct mlx5_ib_event_work *work =
2844 		container_of(_work, struct mlx5_ib_event_work, work);
2845 	struct mlx5_ib_dev *ibdev;
2846 	struct ib_event ibev;
2847 	bool fatal = false;
2848 
2849 	if (work->is_slave) {
2850 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
2851 		if (!ibdev)
2852 			goto out;
2853 	} else {
2854 		ibdev = work->dev;
2855 	}
2856 
2857 	switch (work->event) {
2858 	case MLX5_DEV_EVENT_SYS_ERROR:
2859 		ibev.event = IB_EVENT_DEVICE_FATAL;
2860 		mlx5_ib_handle_internal_error(ibdev);
2861 		ibev.element.port_num  = (u8)(unsigned long)work->param;
2862 		fatal = true;
2863 		break;
2864 	case MLX5_EVENT_TYPE_PORT_CHANGE:
2865 		if (handle_port_change(ibdev, work->param, &ibev))
2866 			goto out;
2867 		break;
2868 	case MLX5_EVENT_TYPE_GENERAL_EVENT:
2869 		handle_general_event(ibdev, work->param, &ibev);
2870 		fallthrough;
2871 	default:
2872 		goto out;
2873 	}
2874 
2875 	ibev.device = &ibdev->ib_dev;
2876 
2877 	if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
2878 		mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
2879 		goto out;
2880 	}
2881 
2882 	if (ibdev->ib_active)
2883 		ib_dispatch_event(&ibev);
2884 
2885 	if (fatal)
2886 		ibdev->ib_active = false;
2887 out:
2888 	kfree(work);
2889 }
2890 
2891 static int mlx5_ib_event(struct notifier_block *nb,
2892 			 unsigned long event, void *param)
2893 {
2894 	struct mlx5_ib_event_work *work;
2895 
2896 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2897 	if (!work)
2898 		return NOTIFY_DONE;
2899 
2900 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2901 	work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
2902 	work->is_slave = false;
2903 	work->param = param;
2904 	work->event = event;
2905 
2906 	queue_work(mlx5_ib_event_wq, &work->work);
2907 
2908 	return NOTIFY_OK;
2909 }
2910 
2911 static int mlx5_ib_event_slave_port(struct notifier_block *nb,
2912 				    unsigned long event, void *param)
2913 {
2914 	struct mlx5_ib_event_work *work;
2915 
2916 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2917 	if (!work)
2918 		return NOTIFY_DONE;
2919 
2920 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2921 	work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
2922 	work->is_slave = true;
2923 	work->param = param;
2924 	work->event = event;
2925 	queue_work(mlx5_ib_event_wq, &work->work);
2926 
2927 	return NOTIFY_OK;
2928 }
2929 
2930 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2931 {
2932 	struct mlx5_hca_vport_context vport_ctx;
2933 	int err;
2934 	int port;
2935 
2936 	for (port = 1; port <= ARRAY_SIZE(dev->port_caps); port++) {
2937 		dev->port_caps[port - 1].has_smi = false;
2938 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
2939 		    MLX5_CAP_PORT_TYPE_IB) {
2940 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2941 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
2942 								   port, 0,
2943 								   &vport_ctx);
2944 				if (err) {
2945 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2946 						    port, err);
2947 					return err;
2948 				}
2949 				dev->port_caps[port - 1].has_smi =
2950 					vport_ctx.has_smi;
2951 			} else {
2952 				dev->port_caps[port - 1].has_smi = true;
2953 			}
2954 		}
2955 	}
2956 	return 0;
2957 }
2958 
2959 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2960 {
2961 	unsigned int port;
2962 
2963 	rdma_for_each_port (&dev->ib_dev, port)
2964 		mlx5_query_ext_port_caps(dev, port);
2965 }
2966 
2967 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
2968 {
2969 	switch (umr_fence_cap) {
2970 	case MLX5_CAP_UMR_FENCE_NONE:
2971 		return MLX5_FENCE_MODE_NONE;
2972 	case MLX5_CAP_UMR_FENCE_SMALL:
2973 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
2974 	default:
2975 		return MLX5_FENCE_MODE_STRONG_ORDERING;
2976 	}
2977 }
2978 
2979 static int mlx5_ib_dev_res_init(struct mlx5_ib_dev *dev)
2980 {
2981 	struct mlx5_ib_resources *devr = &dev->devr;
2982 	struct ib_srq_init_attr attr;
2983 	struct ib_device *ibdev;
2984 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2985 	int port;
2986 	int ret = 0;
2987 
2988 	ibdev = &dev->ib_dev;
2989 
2990 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
2991 		return -EOPNOTSUPP;
2992 
2993 	mutex_init(&devr->mutex);
2994 
2995 	devr->p0 = rdma_zalloc_drv_obj(ibdev, ib_pd);
2996 	if (!devr->p0)
2997 		return -ENOMEM;
2998 
2999 	devr->p0->device  = ibdev;
3000 	devr->p0->uobject = NULL;
3001 	atomic_set(&devr->p0->usecnt, 0);
3002 
3003 	ret = mlx5_ib_alloc_pd(devr->p0, NULL);
3004 	if (ret)
3005 		goto error0;
3006 
3007 	devr->c0 = rdma_zalloc_drv_obj(ibdev, ib_cq);
3008 	if (!devr->c0) {
3009 		ret = -ENOMEM;
3010 		goto error1;
3011 	}
3012 
3013 	devr->c0->device = &dev->ib_dev;
3014 	atomic_set(&devr->c0->usecnt, 0);
3015 
3016 	ret = mlx5_ib_create_cq(devr->c0, &cq_attr, NULL);
3017 	if (ret)
3018 		goto err_create_cq;
3019 
3020 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0);
3021 	if (ret)
3022 		goto error2;
3023 
3024 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0);
3025 	if (ret)
3026 		goto error3;
3027 
3028 	memset(&attr, 0, sizeof(attr));
3029 	attr.attr.max_sge = 1;
3030 	attr.attr.max_wr = 1;
3031 	attr.srq_type = IB_SRQT_XRC;
3032 	attr.ext.cq = devr->c0;
3033 
3034 	devr->s0 = rdma_zalloc_drv_obj(ibdev, ib_srq);
3035 	if (!devr->s0) {
3036 		ret = -ENOMEM;
3037 		goto error4;
3038 	}
3039 
3040 	devr->s0->device	= &dev->ib_dev;
3041 	devr->s0->pd		= devr->p0;
3042 	devr->s0->srq_type      = IB_SRQT_XRC;
3043 	devr->s0->ext.cq	= devr->c0;
3044 	ret = mlx5_ib_create_srq(devr->s0, &attr, NULL);
3045 	if (ret)
3046 		goto err_create;
3047 
3048 	atomic_inc(&devr->s0->ext.cq->usecnt);
3049 	atomic_inc(&devr->p0->usecnt);
3050 	atomic_set(&devr->s0->usecnt, 0);
3051 
3052 	memset(&attr, 0, sizeof(attr));
3053 	attr.attr.max_sge = 1;
3054 	attr.attr.max_wr = 1;
3055 	attr.srq_type = IB_SRQT_BASIC;
3056 	devr->s1 = rdma_zalloc_drv_obj(ibdev, ib_srq);
3057 	if (!devr->s1) {
3058 		ret = -ENOMEM;
3059 		goto error5;
3060 	}
3061 
3062 	devr->s1->device	= &dev->ib_dev;
3063 	devr->s1->pd		= devr->p0;
3064 	devr->s1->srq_type      = IB_SRQT_BASIC;
3065 	devr->s1->ext.cq	= devr->c0;
3066 
3067 	ret = mlx5_ib_create_srq(devr->s1, &attr, NULL);
3068 	if (ret)
3069 		goto error6;
3070 
3071 	atomic_inc(&devr->p0->usecnt);
3072 	atomic_set(&devr->s1->usecnt, 0);
3073 
3074 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
3075 		INIT_WORK(&devr->ports[port].pkey_change_work,
3076 			  pkey_change_handler);
3077 
3078 	return 0;
3079 
3080 error6:
3081 	kfree(devr->s1);
3082 error5:
3083 	mlx5_ib_destroy_srq(devr->s0, NULL);
3084 err_create:
3085 	kfree(devr->s0);
3086 error4:
3087 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
3088 error3:
3089 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
3090 error2:
3091 	mlx5_ib_destroy_cq(devr->c0, NULL);
3092 err_create_cq:
3093 	kfree(devr->c0);
3094 error1:
3095 	mlx5_ib_dealloc_pd(devr->p0, NULL);
3096 error0:
3097 	kfree(devr->p0);
3098 	return ret;
3099 }
3100 
3101 static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev)
3102 {
3103 	struct mlx5_ib_resources *devr = &dev->devr;
3104 	int port;
3105 
3106 	mlx5_ib_destroy_srq(devr->s1, NULL);
3107 	kfree(devr->s1);
3108 	mlx5_ib_destroy_srq(devr->s0, NULL);
3109 	kfree(devr->s0);
3110 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
3111 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
3112 	mlx5_ib_destroy_cq(devr->c0, NULL);
3113 	kfree(devr->c0);
3114 	mlx5_ib_dealloc_pd(devr->p0, NULL);
3115 	kfree(devr->p0);
3116 
3117 	/* Make sure no change P_Key work items are still executing */
3118 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
3119 		cancel_work_sync(&devr->ports[port].pkey_change_work);
3120 }
3121 
3122 static u32 get_core_cap_flags(struct ib_device *ibdev,
3123 			      struct mlx5_hca_vport_context *rep)
3124 {
3125 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3126 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
3127 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
3128 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
3129 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
3130 	u32 ret = 0;
3131 
3132 	if (rep->grh_required)
3133 		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
3134 
3135 	if (ll == IB_LINK_LAYER_INFINIBAND)
3136 		return ret | RDMA_CORE_PORT_IBA_IB;
3137 
3138 	if (raw_support)
3139 		ret |= RDMA_CORE_PORT_RAW_PACKET;
3140 
3141 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
3142 		return ret;
3143 
3144 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
3145 		return ret;
3146 
3147 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
3148 		ret |= RDMA_CORE_PORT_IBA_ROCE;
3149 
3150 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
3151 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3152 
3153 	return ret;
3154 }
3155 
3156 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
3157 			       struct ib_port_immutable *immutable)
3158 {
3159 	struct ib_port_attr attr;
3160 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3161 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
3162 	struct mlx5_hca_vport_context rep = {0};
3163 	int err;
3164 
3165 	err = ib_query_port(ibdev, port_num, &attr);
3166 	if (err)
3167 		return err;
3168 
3169 	if (ll == IB_LINK_LAYER_INFINIBAND) {
3170 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
3171 						   &rep);
3172 		if (err)
3173 			return err;
3174 	}
3175 
3176 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
3177 	immutable->gid_tbl_len = attr.gid_tbl_len;
3178 	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
3179 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3180 
3181 	return 0;
3182 }
3183 
3184 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u8 port_num,
3185 				   struct ib_port_immutable *immutable)
3186 {
3187 	struct ib_port_attr attr;
3188 	int err;
3189 
3190 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
3191 
3192 	err = ib_query_port(ibdev, port_num, &attr);
3193 	if (err)
3194 		return err;
3195 
3196 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
3197 	immutable->gid_tbl_len = attr.gid_tbl_len;
3198 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
3199 
3200 	return 0;
3201 }
3202 
3203 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
3204 {
3205 	struct mlx5_ib_dev *dev =
3206 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
3207 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
3208 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
3209 		 fw_rev_sub(dev->mdev));
3210 }
3211 
3212 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
3213 {
3214 	struct mlx5_core_dev *mdev = dev->mdev;
3215 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
3216 								 MLX5_FLOW_NAMESPACE_LAG);
3217 	struct mlx5_flow_table *ft;
3218 	int err;
3219 
3220 	if (!ns || !mlx5_lag_is_roce(mdev))
3221 		return 0;
3222 
3223 	err = mlx5_cmd_create_vport_lag(mdev);
3224 	if (err)
3225 		return err;
3226 
3227 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3228 	if (IS_ERR(ft)) {
3229 		err = PTR_ERR(ft);
3230 		goto err_destroy_vport_lag;
3231 	}
3232 
3233 	dev->flow_db->lag_demux_ft = ft;
3234 	dev->lag_active = true;
3235 	return 0;
3236 
3237 err_destroy_vport_lag:
3238 	mlx5_cmd_destroy_vport_lag(mdev);
3239 	return err;
3240 }
3241 
3242 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3243 {
3244 	struct mlx5_core_dev *mdev = dev->mdev;
3245 
3246 	if (dev->lag_active) {
3247 		dev->lag_active = false;
3248 
3249 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
3250 		dev->flow_db->lag_demux_ft = NULL;
3251 
3252 		mlx5_cmd_destroy_vport_lag(mdev);
3253 	}
3254 }
3255 
3256 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
3257 {
3258 	int err;
3259 
3260 	dev->port[port_num].roce.nb.notifier_call = mlx5_netdev_event;
3261 	err = register_netdevice_notifier(&dev->port[port_num].roce.nb);
3262 	if (err) {
3263 		dev->port[port_num].roce.nb.notifier_call = NULL;
3264 		return err;
3265 	}
3266 
3267 	return 0;
3268 }
3269 
3270 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
3271 {
3272 	if (dev->port[port_num].roce.nb.notifier_call) {
3273 		unregister_netdevice_notifier(&dev->port[port_num].roce.nb);
3274 		dev->port[port_num].roce.nb.notifier_call = NULL;
3275 	}
3276 }
3277 
3278 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3279 {
3280 	int err;
3281 
3282 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3283 	if (err)
3284 		return err;
3285 
3286 	err = mlx5_eth_lag_init(dev);
3287 	if (err)
3288 		goto err_disable_roce;
3289 
3290 	return 0;
3291 
3292 err_disable_roce:
3293 	mlx5_nic_vport_disable_roce(dev->mdev);
3294 
3295 	return err;
3296 }
3297 
3298 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3299 {
3300 	mlx5_eth_lag_cleanup(dev);
3301 	mlx5_nic_vport_disable_roce(dev->mdev);
3302 }
3303 
3304 static int mlx5_ib_rn_get_params(struct ib_device *device, u8 port_num,
3305 				 enum rdma_netdev_t type,
3306 				 struct rdma_netdev_alloc_params *params)
3307 {
3308 	if (type != RDMA_NETDEV_IPOIB)
3309 		return -EOPNOTSUPP;
3310 
3311 	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
3312 }
3313 
3314 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
3315 				       size_t count, loff_t *pos)
3316 {
3317 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3318 	char lbuf[20];
3319 	int len;
3320 
3321 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
3322 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
3323 }
3324 
3325 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
3326 					size_t count, loff_t *pos)
3327 {
3328 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3329 	u32 timeout;
3330 	u32 var;
3331 
3332 	if (kstrtouint_from_user(buf, count, 0, &var))
3333 		return -EFAULT;
3334 
3335 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
3336 			1000);
3337 	if (timeout != var)
3338 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
3339 			    timeout);
3340 
3341 	delay_drop->timeout = timeout;
3342 
3343 	return count;
3344 }
3345 
3346 static const struct file_operations fops_delay_drop_timeout = {
3347 	.owner	= THIS_MODULE,
3348 	.open	= simple_open,
3349 	.write	= delay_drop_timeout_write,
3350 	.read	= delay_drop_timeout_read,
3351 };
3352 
3353 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
3354 				      struct mlx5_ib_multiport_info *mpi)
3355 {
3356 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3357 	struct mlx5_ib_port *port = &ibdev->port[port_num];
3358 	int comps;
3359 	int err;
3360 	int i;
3361 
3362 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3363 
3364 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
3365 
3366 	spin_lock(&port->mp.mpi_lock);
3367 	if (!mpi->ibdev) {
3368 		spin_unlock(&port->mp.mpi_lock);
3369 		return;
3370 	}
3371 
3372 	mpi->ibdev = NULL;
3373 
3374 	spin_unlock(&port->mp.mpi_lock);
3375 	if (mpi->mdev_events.notifier_call)
3376 		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
3377 	mpi->mdev_events.notifier_call = NULL;
3378 	mlx5_remove_netdev_notifier(ibdev, port_num);
3379 	spin_lock(&port->mp.mpi_lock);
3380 
3381 	comps = mpi->mdev_refcnt;
3382 	if (comps) {
3383 		mpi->unaffiliate = true;
3384 		init_completion(&mpi->unref_comp);
3385 		spin_unlock(&port->mp.mpi_lock);
3386 
3387 		for (i = 0; i < comps; i++)
3388 			wait_for_completion(&mpi->unref_comp);
3389 
3390 		spin_lock(&port->mp.mpi_lock);
3391 		mpi->unaffiliate = false;
3392 	}
3393 
3394 	port->mp.mpi = NULL;
3395 
3396 	list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
3397 
3398 	spin_unlock(&port->mp.mpi_lock);
3399 
3400 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
3401 
3402 	mlx5_ib_dbg(ibdev, "unaffiliated port %d\n", port_num + 1);
3403 	/* Log an error, still needed to cleanup the pointers and add
3404 	 * it back to the list.
3405 	 */
3406 	if (err)
3407 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
3408 			    port_num + 1);
3409 
3410 	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
3411 }
3412 
3413 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
3414 				    struct mlx5_ib_multiport_info *mpi)
3415 {
3416 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3417 	int err;
3418 
3419 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3420 
3421 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
3422 	if (ibdev->port[port_num].mp.mpi) {
3423 		mlx5_ib_dbg(ibdev, "port %d already affiliated.\n",
3424 			    port_num + 1);
3425 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3426 		return false;
3427 	}
3428 
3429 	ibdev->port[port_num].mp.mpi = mpi;
3430 	mpi->ibdev = ibdev;
3431 	mpi->mdev_events.notifier_call = NULL;
3432 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3433 
3434 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
3435 	if (err)
3436 		goto unbind;
3437 
3438 	err = mlx5_add_netdev_notifier(ibdev, port_num);
3439 	if (err) {
3440 		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
3441 			    port_num + 1);
3442 		goto unbind;
3443 	}
3444 
3445 	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
3446 	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
3447 
3448 	mlx5_ib_init_cong_debugfs(ibdev, port_num);
3449 
3450 	return true;
3451 
3452 unbind:
3453 	mlx5_ib_unbind_slave_port(ibdev, mpi);
3454 	return false;
3455 }
3456 
3457 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
3458 {
3459 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3460 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3461 							  port_num + 1);
3462 	struct mlx5_ib_multiport_info *mpi;
3463 	int err;
3464 	int i;
3465 
3466 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3467 		return 0;
3468 
3469 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
3470 						     &dev->sys_image_guid);
3471 	if (err)
3472 		return err;
3473 
3474 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3475 	if (err)
3476 		return err;
3477 
3478 	mutex_lock(&mlx5_ib_multiport_mutex);
3479 	for (i = 0; i < dev->num_ports; i++) {
3480 		bool bound = false;
3481 
3482 		/* build a stub multiport info struct for the native port. */
3483 		if (i == port_num) {
3484 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
3485 			if (!mpi) {
3486 				mutex_unlock(&mlx5_ib_multiport_mutex);
3487 				mlx5_nic_vport_disable_roce(dev->mdev);
3488 				return -ENOMEM;
3489 			}
3490 
3491 			mpi->is_master = true;
3492 			mpi->mdev = dev->mdev;
3493 			mpi->sys_image_guid = dev->sys_image_guid;
3494 			dev->port[i].mp.mpi = mpi;
3495 			mpi->ibdev = dev;
3496 			mpi = NULL;
3497 			continue;
3498 		}
3499 
3500 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
3501 				    list) {
3502 			if (dev->sys_image_guid == mpi->sys_image_guid &&
3503 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
3504 				bound = mlx5_ib_bind_slave_port(dev, mpi);
3505 			}
3506 
3507 			if (bound) {
3508 				dev_dbg(mpi->mdev->device,
3509 					"removing port from unaffiliated list.\n");
3510 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
3511 				list_del(&mpi->list);
3512 				break;
3513 			}
3514 		}
3515 		if (!bound)
3516 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
3517 				    i + 1);
3518 	}
3519 
3520 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
3521 	mutex_unlock(&mlx5_ib_multiport_mutex);
3522 	return err;
3523 }
3524 
3525 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
3526 {
3527 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3528 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3529 							  port_num + 1);
3530 	int i;
3531 
3532 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3533 		return;
3534 
3535 	mutex_lock(&mlx5_ib_multiport_mutex);
3536 	for (i = 0; i < dev->num_ports; i++) {
3537 		if (dev->port[i].mp.mpi) {
3538 			/* Destroy the native port stub */
3539 			if (i == port_num) {
3540 				kfree(dev->port[i].mp.mpi);
3541 				dev->port[i].mp.mpi = NULL;
3542 			} else {
3543 				mlx5_ib_dbg(dev, "unbinding port_num: %d\n", i + 1);
3544 				mlx5_ib_unbind_slave_port(dev, dev->port[i].mp.mpi);
3545 			}
3546 		}
3547 	}
3548 
3549 	mlx5_ib_dbg(dev, "removing from devlist\n");
3550 	list_del(&dev->ib_dev_list);
3551 	mutex_unlock(&mlx5_ib_multiport_mutex);
3552 
3553 	mlx5_nic_vport_disable_roce(dev->mdev);
3554 }
3555 
3556 static int mmap_obj_cleanup(struct ib_uobject *uobject,
3557 			    enum rdma_remove_reason why,
3558 			    struct uverbs_attr_bundle *attrs)
3559 {
3560 	struct mlx5_user_mmap_entry *obj = uobject->object;
3561 
3562 	rdma_user_mmap_entry_remove(&obj->rdma_entry);
3563 	return 0;
3564 }
3565 
3566 static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c,
3567 					    struct mlx5_user_mmap_entry *entry,
3568 					    size_t length)
3569 {
3570 	return rdma_user_mmap_entry_insert_range(
3571 		&c->ibucontext, &entry->rdma_entry, length,
3572 		(MLX5_IB_MMAP_OFFSET_START << 16),
3573 		((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1));
3574 }
3575 
3576 static struct mlx5_user_mmap_entry *
3577 alloc_var_entry(struct mlx5_ib_ucontext *c)
3578 {
3579 	struct mlx5_user_mmap_entry *entry;
3580 	struct mlx5_var_table *var_table;
3581 	u32 page_idx;
3582 	int err;
3583 
3584 	var_table = &to_mdev(c->ibucontext.device)->var_table;
3585 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3586 	if (!entry)
3587 		return ERR_PTR(-ENOMEM);
3588 
3589 	mutex_lock(&var_table->bitmap_lock);
3590 	page_idx = find_first_zero_bit(var_table->bitmap,
3591 				       var_table->num_var_hw_entries);
3592 	if (page_idx >= var_table->num_var_hw_entries) {
3593 		err = -ENOSPC;
3594 		mutex_unlock(&var_table->bitmap_lock);
3595 		goto end;
3596 	}
3597 
3598 	set_bit(page_idx, var_table->bitmap);
3599 	mutex_unlock(&var_table->bitmap_lock);
3600 
3601 	entry->address = var_table->hw_start_addr +
3602 				(page_idx * var_table->stride_size);
3603 	entry->page_idx = page_idx;
3604 	entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
3605 
3606 	err = mlx5_rdma_user_mmap_entry_insert(c, entry,
3607 					       var_table->stride_size);
3608 	if (err)
3609 		goto err_insert;
3610 
3611 	return entry;
3612 
3613 err_insert:
3614 	mutex_lock(&var_table->bitmap_lock);
3615 	clear_bit(page_idx, var_table->bitmap);
3616 	mutex_unlock(&var_table->bitmap_lock);
3617 end:
3618 	kfree(entry);
3619 	return ERR_PTR(err);
3620 }
3621 
3622 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
3623 	struct uverbs_attr_bundle *attrs)
3624 {
3625 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3626 		attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3627 	struct mlx5_ib_ucontext *c;
3628 	struct mlx5_user_mmap_entry *entry;
3629 	u64 mmap_offset;
3630 	u32 length;
3631 	int err;
3632 
3633 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3634 	if (IS_ERR(c))
3635 		return PTR_ERR(c);
3636 
3637 	entry = alloc_var_entry(c);
3638 	if (IS_ERR(entry))
3639 		return PTR_ERR(entry);
3640 
3641 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3642 	length = entry->rdma_entry.npages * PAGE_SIZE;
3643 	uobj->object = entry;
3644 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3645 
3646 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3647 			     &mmap_offset, sizeof(mmap_offset));
3648 	if (err)
3649 		return err;
3650 
3651 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3652 			     &entry->page_idx, sizeof(entry->page_idx));
3653 	if (err)
3654 		return err;
3655 
3656 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3657 			     &length, sizeof(length));
3658 	return err;
3659 }
3660 
3661 DECLARE_UVERBS_NAMED_METHOD(
3662 	MLX5_IB_METHOD_VAR_OBJ_ALLOC,
3663 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
3664 			MLX5_IB_OBJECT_VAR,
3665 			UVERBS_ACCESS_NEW,
3666 			UA_MANDATORY),
3667 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3668 			   UVERBS_ATTR_TYPE(u32),
3669 			   UA_MANDATORY),
3670 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3671 			   UVERBS_ATTR_TYPE(u32),
3672 			   UA_MANDATORY),
3673 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3674 			    UVERBS_ATTR_TYPE(u64),
3675 			    UA_MANDATORY));
3676 
3677 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3678 	MLX5_IB_METHOD_VAR_OBJ_DESTROY,
3679 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
3680 			MLX5_IB_OBJECT_VAR,
3681 			UVERBS_ACCESS_DESTROY,
3682 			UA_MANDATORY));
3683 
3684 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
3685 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3686 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
3687 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
3688 
3689 static bool var_is_supported(struct ib_device *device)
3690 {
3691 	struct mlx5_ib_dev *dev = to_mdev(device);
3692 
3693 	return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3694 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
3695 }
3696 
3697 static struct mlx5_user_mmap_entry *
3698 alloc_uar_entry(struct mlx5_ib_ucontext *c,
3699 		enum mlx5_ib_uapi_uar_alloc_type alloc_type)
3700 {
3701 	struct mlx5_user_mmap_entry *entry;
3702 	struct mlx5_ib_dev *dev;
3703 	u32 uar_index;
3704 	int err;
3705 
3706 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3707 	if (!entry)
3708 		return ERR_PTR(-ENOMEM);
3709 
3710 	dev = to_mdev(c->ibucontext.device);
3711 	err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
3712 	if (err)
3713 		goto end;
3714 
3715 	entry->page_idx = uar_index;
3716 	entry->address = uar_index2paddress(dev, uar_index);
3717 	if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3718 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC;
3719 	else
3720 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC;
3721 
3722 	err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE);
3723 	if (err)
3724 		goto err_insert;
3725 
3726 	return entry;
3727 
3728 err_insert:
3729 	mlx5_cmd_free_uar(dev->mdev, uar_index);
3730 end:
3731 	kfree(entry);
3732 	return ERR_PTR(err);
3733 }
3734 
3735 static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)(
3736 	struct uverbs_attr_bundle *attrs)
3737 {
3738 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3739 		attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3740 	enum mlx5_ib_uapi_uar_alloc_type alloc_type;
3741 	struct mlx5_ib_ucontext *c;
3742 	struct mlx5_user_mmap_entry *entry;
3743 	u64 mmap_offset;
3744 	u32 length;
3745 	int err;
3746 
3747 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3748 	if (IS_ERR(c))
3749 		return PTR_ERR(c);
3750 
3751 	err = uverbs_get_const(&alloc_type, attrs,
3752 			       MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE);
3753 	if (err)
3754 		return err;
3755 
3756 	if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF &&
3757 	    alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC)
3758 		return -EOPNOTSUPP;
3759 
3760 	if (!to_mdev(c->ibucontext.device)->wc_support &&
3761 	    alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3762 		return -EOPNOTSUPP;
3763 
3764 	entry = alloc_uar_entry(c, alloc_type);
3765 	if (IS_ERR(entry))
3766 		return PTR_ERR(entry);
3767 
3768 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3769 	length = entry->rdma_entry.npages * PAGE_SIZE;
3770 	uobj->object = entry;
3771 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3772 
3773 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3774 			     &mmap_offset, sizeof(mmap_offset));
3775 	if (err)
3776 		return err;
3777 
3778 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3779 			     &entry->page_idx, sizeof(entry->page_idx));
3780 	if (err)
3781 		return err;
3782 
3783 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3784 			     &length, sizeof(length));
3785 	return err;
3786 }
3787 
3788 DECLARE_UVERBS_NAMED_METHOD(
3789 	MLX5_IB_METHOD_UAR_OBJ_ALLOC,
3790 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE,
3791 			MLX5_IB_OBJECT_UAR,
3792 			UVERBS_ACCESS_NEW,
3793 			UA_MANDATORY),
3794 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE,
3795 			     enum mlx5_ib_uapi_uar_alloc_type,
3796 			     UA_MANDATORY),
3797 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3798 			   UVERBS_ATTR_TYPE(u32),
3799 			   UA_MANDATORY),
3800 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3801 			   UVERBS_ATTR_TYPE(u32),
3802 			   UA_MANDATORY),
3803 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3804 			    UVERBS_ATTR_TYPE(u64),
3805 			    UA_MANDATORY));
3806 
3807 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3808 	MLX5_IB_METHOD_UAR_OBJ_DESTROY,
3809 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE,
3810 			MLX5_IB_OBJECT_UAR,
3811 			UVERBS_ACCESS_DESTROY,
3812 			UA_MANDATORY));
3813 
3814 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR,
3815 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3816 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC),
3817 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY));
3818 
3819 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3820 	mlx5_ib_dm,
3821 	UVERBS_OBJECT_DM,
3822 	UVERBS_METHOD_DM_ALLOC,
3823 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
3824 			    UVERBS_ATTR_TYPE(u64),
3825 			    UA_MANDATORY),
3826 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
3827 			    UVERBS_ATTR_TYPE(u16),
3828 			    UA_OPTIONAL),
3829 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_ALLOC_DM_REQ_TYPE,
3830 			     enum mlx5_ib_uapi_dm_type,
3831 			     UA_OPTIONAL));
3832 
3833 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3834 	mlx5_ib_flow_action,
3835 	UVERBS_OBJECT_FLOW_ACTION,
3836 	UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
3837 	UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
3838 			     enum mlx5_ib_uapi_flow_action_flags));
3839 
3840 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3841 	mlx5_ib_query_context,
3842 	UVERBS_OBJECT_DEVICE,
3843 	UVERBS_METHOD_QUERY_CONTEXT,
3844 	UVERBS_ATTR_PTR_OUT(
3845 		MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
3846 		UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp,
3847 				   dump_fill_mkey),
3848 		UA_MANDATORY));
3849 
3850 static const struct uapi_definition mlx5_ib_defs[] = {
3851 	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
3852 	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
3853 	UAPI_DEF_CHAIN(mlx5_ib_qos_defs),
3854 	UAPI_DEF_CHAIN(mlx5_ib_std_types_defs),
3855 
3856 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_FLOW_ACTION,
3857 				&mlx5_ib_flow_action),
3858 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DM, &mlx5_ib_dm),
3859 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context),
3860 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
3861 				UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
3862 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR),
3863 	{}
3864 };
3865 
3866 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
3867 {
3868 	mlx5_ib_cleanup_multiport_master(dev);
3869 	WARN_ON(!xa_empty(&dev->odp_mkeys));
3870 	mutex_destroy(&dev->cap_mask_mutex);
3871 	WARN_ON(!xa_empty(&dev->sig_mrs));
3872 	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
3873 }
3874 
3875 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
3876 {
3877 	struct mlx5_core_dev *mdev = dev->mdev;
3878 	int err;
3879 	int i;
3880 
3881 	dev->ib_dev.node_type = RDMA_NODE_IB_CA;
3882 	dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
3883 	dev->ib_dev.phys_port_cnt = dev->num_ports;
3884 	dev->ib_dev.dev.parent = mdev->device;
3885 	dev->ib_dev.lag_flags = RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
3886 
3887 	for (i = 0; i < dev->num_ports; i++) {
3888 		spin_lock_init(&dev->port[i].mp.mpi_lock);
3889 		rwlock_init(&dev->port[i].roce.netdev_lock);
3890 		dev->port[i].roce.dev = dev;
3891 		dev->port[i].roce.native_port_num = i + 1;
3892 		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
3893 	}
3894 
3895 	mlx5_ib_internal_fill_odp_caps(dev);
3896 
3897 	err = mlx5_ib_init_multiport_master(dev);
3898 	if (err)
3899 		return err;
3900 
3901 	err = set_has_smi_cap(dev);
3902 	if (err)
3903 		goto err_mp;
3904 
3905 	err = mlx5_query_max_pkeys(&dev->ib_dev, &dev->pkey_table_len);
3906 	if (err)
3907 		goto err_mp;
3908 
3909 	if (mlx5_use_mad_ifc(dev))
3910 		get_ext_port_caps(dev);
3911 
3912 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_count(mdev);
3913 
3914 	mutex_init(&dev->cap_mask_mutex);
3915 	INIT_LIST_HEAD(&dev->qp_list);
3916 	spin_lock_init(&dev->reset_flow_resource_lock);
3917 	xa_init(&dev->odp_mkeys);
3918 	xa_init(&dev->sig_mrs);
3919 	atomic_set(&dev->mkey_var, 0);
3920 
3921 	spin_lock_init(&dev->dm.lock);
3922 	dev->dm.dev = mdev;
3923 	return 0;
3924 
3925 err_mp:
3926 	mlx5_ib_cleanup_multiport_master(dev);
3927 	return err;
3928 }
3929 
3930 static int mlx5_ib_enable_driver(struct ib_device *dev)
3931 {
3932 	struct mlx5_ib_dev *mdev = to_mdev(dev);
3933 	int ret;
3934 
3935 	ret = mlx5_ib_test_wc(mdev);
3936 	mlx5_ib_dbg(mdev, "Write-Combining %s",
3937 		    mdev->wc_support ? "supported" : "not supported");
3938 
3939 	return ret;
3940 }
3941 
3942 static const struct ib_device_ops mlx5_ib_dev_ops = {
3943 	.owner = THIS_MODULE,
3944 	.driver_id = RDMA_DRIVER_MLX5,
3945 	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
3946 
3947 	.add_gid = mlx5_ib_add_gid,
3948 	.alloc_mr = mlx5_ib_alloc_mr,
3949 	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
3950 	.alloc_pd = mlx5_ib_alloc_pd,
3951 	.alloc_ucontext = mlx5_ib_alloc_ucontext,
3952 	.attach_mcast = mlx5_ib_mcg_attach,
3953 	.check_mr_status = mlx5_ib_check_mr_status,
3954 	.create_ah = mlx5_ib_create_ah,
3955 	.create_cq = mlx5_ib_create_cq,
3956 	.create_qp = mlx5_ib_create_qp,
3957 	.create_srq = mlx5_ib_create_srq,
3958 	.create_user_ah = mlx5_ib_create_ah,
3959 	.dealloc_pd = mlx5_ib_dealloc_pd,
3960 	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
3961 	.del_gid = mlx5_ib_del_gid,
3962 	.dereg_mr = mlx5_ib_dereg_mr,
3963 	.destroy_ah = mlx5_ib_destroy_ah,
3964 	.destroy_cq = mlx5_ib_destroy_cq,
3965 	.destroy_qp = mlx5_ib_destroy_qp,
3966 	.destroy_srq = mlx5_ib_destroy_srq,
3967 	.detach_mcast = mlx5_ib_mcg_detach,
3968 	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
3969 	.drain_rq = mlx5_ib_drain_rq,
3970 	.drain_sq = mlx5_ib_drain_sq,
3971 	.enable_driver = mlx5_ib_enable_driver,
3972 	.get_dev_fw_str = get_dev_fw_str,
3973 	.get_dma_mr = mlx5_ib_get_dma_mr,
3974 	.get_link_layer = mlx5_ib_port_link_layer,
3975 	.map_mr_sg = mlx5_ib_map_mr_sg,
3976 	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
3977 	.mmap = mlx5_ib_mmap,
3978 	.mmap_free = mlx5_ib_mmap_free,
3979 	.modify_cq = mlx5_ib_modify_cq,
3980 	.modify_device = mlx5_ib_modify_device,
3981 	.modify_port = mlx5_ib_modify_port,
3982 	.modify_qp = mlx5_ib_modify_qp,
3983 	.modify_srq = mlx5_ib_modify_srq,
3984 	.poll_cq = mlx5_ib_poll_cq,
3985 	.post_recv = mlx5_ib_post_recv_nodrain,
3986 	.post_send = mlx5_ib_post_send_nodrain,
3987 	.post_srq_recv = mlx5_ib_post_srq_recv,
3988 	.process_mad = mlx5_ib_process_mad,
3989 	.query_ah = mlx5_ib_query_ah,
3990 	.query_device = mlx5_ib_query_device,
3991 	.query_gid = mlx5_ib_query_gid,
3992 	.query_pkey = mlx5_ib_query_pkey,
3993 	.query_qp = mlx5_ib_query_qp,
3994 	.query_srq = mlx5_ib_query_srq,
3995 	.query_ucontext = mlx5_ib_query_ucontext,
3996 	.reg_user_mr = mlx5_ib_reg_user_mr,
3997 	.reg_user_mr_dmabuf = mlx5_ib_reg_user_mr_dmabuf,
3998 	.req_notify_cq = mlx5_ib_arm_cq,
3999 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
4000 	.resize_cq = mlx5_ib_resize_cq,
4001 
4002 	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
4003 	INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs),
4004 	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
4005 	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
4006 	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
4007 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
4008 };
4009 
4010 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
4011 	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
4012 };
4013 
4014 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
4015 	.get_vf_config = mlx5_ib_get_vf_config,
4016 	.get_vf_guid = mlx5_ib_get_vf_guid,
4017 	.get_vf_stats = mlx5_ib_get_vf_stats,
4018 	.set_vf_guid = mlx5_ib_set_vf_guid,
4019 	.set_vf_link_state = mlx5_ib_set_vf_link_state,
4020 };
4021 
4022 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
4023 	.alloc_mw = mlx5_ib_alloc_mw,
4024 	.dealloc_mw = mlx5_ib_dealloc_mw,
4025 
4026 	INIT_RDMA_OBJ_SIZE(ib_mw, mlx5_ib_mw, ibmw),
4027 };
4028 
4029 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
4030 	.alloc_xrcd = mlx5_ib_alloc_xrcd,
4031 	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
4032 
4033 	INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd),
4034 };
4035 
4036 static const struct ib_device_ops mlx5_ib_dev_dm_ops = {
4037 	.alloc_dm = mlx5_ib_alloc_dm,
4038 	.dealloc_dm = mlx5_ib_dealloc_dm,
4039 	.reg_dm_mr = mlx5_ib_reg_dm_mr,
4040 };
4041 
4042 static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
4043 {
4044 	struct mlx5_core_dev *mdev = dev->mdev;
4045 	struct mlx5_var_table *var_table = &dev->var_table;
4046 	u8 log_doorbell_bar_size;
4047 	u8 log_doorbell_stride;
4048 	u64 bar_size;
4049 
4050 	log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
4051 					log_doorbell_bar_size);
4052 	log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
4053 					log_doorbell_stride);
4054 	var_table->hw_start_addr = dev->mdev->bar_addr +
4055 				MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
4056 					doorbell_bar_offset);
4057 	bar_size = (1ULL << log_doorbell_bar_size) * 4096;
4058 	var_table->stride_size = 1ULL << log_doorbell_stride;
4059 	var_table->num_var_hw_entries = div_u64(bar_size,
4060 						var_table->stride_size);
4061 	mutex_init(&var_table->bitmap_lock);
4062 	var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
4063 					  GFP_KERNEL);
4064 	return (var_table->bitmap) ? 0 : -ENOMEM;
4065 }
4066 
4067 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
4068 {
4069 	bitmap_free(dev->var_table.bitmap);
4070 }
4071 
4072 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
4073 {
4074 	struct mlx5_core_dev *mdev = dev->mdev;
4075 	int err;
4076 
4077 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
4078 	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
4079 		ib_set_device_ops(&dev->ib_dev,
4080 				  &mlx5_ib_dev_ipoib_enhanced_ops);
4081 
4082 	if (mlx5_core_is_pf(mdev))
4083 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
4084 
4085 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
4086 
4087 	if (MLX5_CAP_GEN(mdev, imaicl))
4088 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
4089 
4090 	if (MLX5_CAP_GEN(mdev, xrc))
4091 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
4092 
4093 	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
4094 	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
4095 	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
4096 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
4097 
4098 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
4099 
4100 	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
4101 		dev->ib_dev.driver_def = mlx5_ib_defs;
4102 
4103 	err = init_node_data(dev);
4104 	if (err)
4105 		return err;
4106 
4107 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
4108 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
4109 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
4110 		mutex_init(&dev->lb.mutex);
4111 
4112 	if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
4113 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
4114 		err = mlx5_ib_init_var_table(dev);
4115 		if (err)
4116 			return err;
4117 	}
4118 
4119 	dev->ib_dev.use_cq_dim = true;
4120 
4121 	return 0;
4122 }
4123 
4124 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
4125 	.get_port_immutable = mlx5_port_immutable,
4126 	.query_port = mlx5_ib_query_port,
4127 };
4128 
4129 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
4130 {
4131 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
4132 	return 0;
4133 }
4134 
4135 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
4136 	.get_port_immutable = mlx5_port_rep_immutable,
4137 	.query_port = mlx5_ib_rep_query_port,
4138 	.query_pkey = mlx5_ib_rep_query_pkey,
4139 };
4140 
4141 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
4142 {
4143 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
4144 	return 0;
4145 }
4146 
4147 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
4148 	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
4149 	.create_wq = mlx5_ib_create_wq,
4150 	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
4151 	.destroy_wq = mlx5_ib_destroy_wq,
4152 	.get_netdev = mlx5_ib_get_netdev,
4153 	.modify_wq = mlx5_ib_modify_wq,
4154 
4155 	INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table,
4156 			   ib_rwq_ind_tbl),
4157 };
4158 
4159 static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev)
4160 {
4161 	struct mlx5_core_dev *mdev = dev->mdev;
4162 	enum rdma_link_layer ll;
4163 	int port_type_cap;
4164 	u8 port_num = 0;
4165 	int err;
4166 
4167 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4168 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4169 
4170 	if (ll == IB_LINK_LAYER_ETHERNET) {
4171 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
4172 
4173 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4174 
4175 		/* Register only for native ports */
4176 		err = mlx5_add_netdev_notifier(dev, port_num);
4177 		if (err || dev->is_rep || !mlx5_is_roce_enabled(mdev))
4178 			/*
4179 			 * We don't enable ETH interface for
4180 			 * 1. IB representors
4181 			 * 2. User disabled ROCE through devlink interface
4182 			 */
4183 			return err;
4184 
4185 		err = mlx5_enable_eth(dev);
4186 		if (err)
4187 			goto cleanup;
4188 	}
4189 
4190 	return 0;
4191 cleanup:
4192 	mlx5_remove_netdev_notifier(dev, port_num);
4193 	return err;
4194 }
4195 
4196 static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev)
4197 {
4198 	struct mlx5_core_dev *mdev = dev->mdev;
4199 	enum rdma_link_layer ll;
4200 	int port_type_cap;
4201 	u8 port_num;
4202 
4203 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4204 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4205 
4206 	if (ll == IB_LINK_LAYER_ETHERNET) {
4207 		if (!dev->is_rep)
4208 			mlx5_disable_eth(dev);
4209 
4210 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4211 		mlx5_remove_netdev_notifier(dev, port_num);
4212 	}
4213 }
4214 
4215 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
4216 {
4217 	mlx5_ib_init_cong_debugfs(dev,
4218 				  mlx5_core_native_port_num(dev->mdev) - 1);
4219 	return 0;
4220 }
4221 
4222 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
4223 {
4224 	mlx5_ib_cleanup_cong_debugfs(dev,
4225 				     mlx5_core_native_port_num(dev->mdev) - 1);
4226 }
4227 
4228 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
4229 {
4230 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
4231 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
4232 }
4233 
4234 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
4235 {
4236 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
4237 }
4238 
4239 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
4240 {
4241 	int err;
4242 
4243 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
4244 	if (err)
4245 		return err;
4246 
4247 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
4248 	if (err)
4249 		mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4250 
4251 	return err;
4252 }
4253 
4254 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
4255 {
4256 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4257 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4258 }
4259 
4260 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
4261 {
4262 	const char *name;
4263 
4264 	rdma_set_device_sysfs_group(&dev->ib_dev, &mlx5_attr_group);
4265 	if (!mlx5_lag_is_roce(dev->mdev))
4266 		name = "mlx5_%d";
4267 	else
4268 		name = "mlx5_bond_%d";
4269 	return ib_register_device(&dev->ib_dev, name, &dev->mdev->pdev->dev);
4270 }
4271 
4272 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
4273 {
4274 	int err;
4275 
4276 	err = mlx5_mr_cache_cleanup(dev);
4277 	if (err)
4278 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
4279 
4280 	if (dev->umrc.qp)
4281 		mlx5_ib_destroy_qp(dev->umrc.qp, NULL);
4282 	if (dev->umrc.cq)
4283 		ib_free_cq(dev->umrc.cq);
4284 	if (dev->umrc.pd)
4285 		ib_dealloc_pd(dev->umrc.pd);
4286 }
4287 
4288 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
4289 {
4290 	ib_unregister_device(&dev->ib_dev);
4291 }
4292 
4293 enum {
4294 	MAX_UMR_WR = 128,
4295 };
4296 
4297 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
4298 {
4299 	struct ib_qp_init_attr *init_attr = NULL;
4300 	struct ib_qp_attr *attr = NULL;
4301 	struct ib_pd *pd;
4302 	struct ib_cq *cq;
4303 	struct ib_qp *qp;
4304 	int ret;
4305 
4306 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
4307 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
4308 	if (!attr || !init_attr) {
4309 		ret = -ENOMEM;
4310 		goto error_0;
4311 	}
4312 
4313 	pd = ib_alloc_pd(&dev->ib_dev, 0);
4314 	if (IS_ERR(pd)) {
4315 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
4316 		ret = PTR_ERR(pd);
4317 		goto error_0;
4318 	}
4319 
4320 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
4321 	if (IS_ERR(cq)) {
4322 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
4323 		ret = PTR_ERR(cq);
4324 		goto error_2;
4325 	}
4326 
4327 	init_attr->send_cq = cq;
4328 	init_attr->recv_cq = cq;
4329 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
4330 	init_attr->cap.max_send_wr = MAX_UMR_WR;
4331 	init_attr->cap.max_send_sge = 1;
4332 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
4333 	init_attr->port_num = 1;
4334 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
4335 	if (IS_ERR(qp)) {
4336 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
4337 		ret = PTR_ERR(qp);
4338 		goto error_3;
4339 	}
4340 	qp->device     = &dev->ib_dev;
4341 	qp->real_qp    = qp;
4342 	qp->uobject    = NULL;
4343 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
4344 	qp->send_cq    = init_attr->send_cq;
4345 	qp->recv_cq    = init_attr->recv_cq;
4346 
4347 	attr->qp_state = IB_QPS_INIT;
4348 	attr->port_num = 1;
4349 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
4350 				IB_QP_PORT, NULL);
4351 	if (ret) {
4352 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
4353 		goto error_4;
4354 	}
4355 
4356 	memset(attr, 0, sizeof(*attr));
4357 	attr->qp_state = IB_QPS_RTR;
4358 	attr->path_mtu = IB_MTU_256;
4359 
4360 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4361 	if (ret) {
4362 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
4363 		goto error_4;
4364 	}
4365 
4366 	memset(attr, 0, sizeof(*attr));
4367 	attr->qp_state = IB_QPS_RTS;
4368 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4369 	if (ret) {
4370 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
4371 		goto error_4;
4372 	}
4373 
4374 	dev->umrc.qp = qp;
4375 	dev->umrc.cq = cq;
4376 	dev->umrc.pd = pd;
4377 
4378 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
4379 	ret = mlx5_mr_cache_init(dev);
4380 	if (ret) {
4381 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4382 		goto error_4;
4383 	}
4384 
4385 	kfree(attr);
4386 	kfree(init_attr);
4387 
4388 	return 0;
4389 
4390 error_4:
4391 	mlx5_ib_destroy_qp(qp, NULL);
4392 	dev->umrc.qp = NULL;
4393 
4394 error_3:
4395 	ib_free_cq(cq);
4396 	dev->umrc.cq = NULL;
4397 
4398 error_2:
4399 	ib_dealloc_pd(pd);
4400 	dev->umrc.pd = NULL;
4401 
4402 error_0:
4403 	kfree(attr);
4404 	kfree(init_attr);
4405 	return ret;
4406 }
4407 
4408 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
4409 {
4410 	struct dentry *root;
4411 
4412 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4413 		return 0;
4414 
4415 	mutex_init(&dev->delay_drop.lock);
4416 	dev->delay_drop.dev = dev;
4417 	dev->delay_drop.activate = false;
4418 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4419 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4420 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4421 	atomic_set(&dev->delay_drop.events_cnt, 0);
4422 
4423 	if (!mlx5_debugfs_root)
4424 		return 0;
4425 
4426 	root = debugfs_create_dir("delay_drop", dev->mdev->priv.dbg_root);
4427 	dev->delay_drop.dir_debugfs = root;
4428 
4429 	debugfs_create_atomic_t("num_timeout_events", 0400, root,
4430 				&dev->delay_drop.events_cnt);
4431 	debugfs_create_atomic_t("num_rqs", 0400, root,
4432 				&dev->delay_drop.rqs_cnt);
4433 	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
4434 			    &fops_delay_drop_timeout);
4435 	return 0;
4436 }
4437 
4438 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
4439 {
4440 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4441 		return;
4442 
4443 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4444 	if (!dev->delay_drop.dir_debugfs)
4445 		return;
4446 
4447 	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
4448 	dev->delay_drop.dir_debugfs = NULL;
4449 }
4450 
4451 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
4452 {
4453 	dev->mdev_events.notifier_call = mlx5_ib_event;
4454 	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
4455 	return 0;
4456 }
4457 
4458 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
4459 {
4460 	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
4461 }
4462 
4463 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
4464 		      const struct mlx5_ib_profile *profile,
4465 		      int stage)
4466 {
4467 	dev->ib_active = false;
4468 
4469 	/* Number of stages to cleanup */
4470 	while (stage) {
4471 		stage--;
4472 		if (profile->stage[stage].cleanup)
4473 			profile->stage[stage].cleanup(dev);
4474 	}
4475 
4476 	kfree(dev->port);
4477 	ib_dealloc_device(&dev->ib_dev);
4478 }
4479 
4480 int __mlx5_ib_add(struct mlx5_ib_dev *dev,
4481 		  const struct mlx5_ib_profile *profile)
4482 {
4483 	int err;
4484 	int i;
4485 
4486 	dev->profile = profile;
4487 
4488 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
4489 		if (profile->stage[i].init) {
4490 			err = profile->stage[i].init(dev);
4491 			if (err)
4492 				goto err_out;
4493 		}
4494 	}
4495 
4496 	dev->ib_active = true;
4497 	return 0;
4498 
4499 err_out:
4500 	/* Clean up stages which were initialized */
4501 	while (i) {
4502 		i--;
4503 		if (profile->stage[i].cleanup)
4504 			profile->stage[i].cleanup(dev);
4505 	}
4506 	return -ENOMEM;
4507 }
4508 
4509 static const struct mlx5_ib_profile pf_profile = {
4510 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4511 		     mlx5_ib_stage_init_init,
4512 		     mlx5_ib_stage_init_cleanup),
4513 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4514 		     mlx5_ib_fs_init,
4515 		     mlx5_ib_fs_cleanup),
4516 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4517 		     mlx5_ib_stage_caps_init,
4518 		     mlx5_ib_stage_caps_cleanup),
4519 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4520 		     mlx5_ib_stage_non_default_cb,
4521 		     NULL),
4522 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4523 		     mlx5_ib_roce_init,
4524 		     mlx5_ib_roce_cleanup),
4525 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4526 		     mlx5_init_qp_table,
4527 		     mlx5_cleanup_qp_table),
4528 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4529 		     mlx5_init_srq_table,
4530 		     mlx5_cleanup_srq_table),
4531 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4532 		     mlx5_ib_dev_res_init,
4533 		     mlx5_ib_dev_res_cleanup),
4534 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4535 		     mlx5_ib_stage_dev_notifier_init,
4536 		     mlx5_ib_stage_dev_notifier_cleanup),
4537 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
4538 		     mlx5_ib_odp_init_one,
4539 		     mlx5_ib_odp_cleanup_one),
4540 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4541 		     mlx5_ib_counters_init,
4542 		     mlx5_ib_counters_cleanup),
4543 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4544 		     mlx5_ib_stage_cong_debugfs_init,
4545 		     mlx5_ib_stage_cong_debugfs_cleanup),
4546 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4547 		     mlx5_ib_stage_uar_init,
4548 		     mlx5_ib_stage_uar_cleanup),
4549 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4550 		     mlx5_ib_stage_bfrag_init,
4551 		     mlx5_ib_stage_bfrag_cleanup),
4552 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4553 		     NULL,
4554 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4555 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4556 		     mlx5_ib_devx_init,
4557 		     mlx5_ib_devx_cleanup),
4558 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4559 		     mlx5_ib_stage_ib_reg_init,
4560 		     mlx5_ib_stage_ib_reg_cleanup),
4561 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4562 		     mlx5_ib_stage_post_ib_reg_umr_init,
4563 		     NULL),
4564 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4565 		     mlx5_ib_stage_delay_drop_init,
4566 		     mlx5_ib_stage_delay_drop_cleanup),
4567 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4568 		     mlx5_ib_restrack_init,
4569 		     NULL),
4570 };
4571 
4572 const struct mlx5_ib_profile raw_eth_profile = {
4573 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4574 		     mlx5_ib_stage_init_init,
4575 		     mlx5_ib_stage_init_cleanup),
4576 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4577 		     mlx5_ib_fs_init,
4578 		     mlx5_ib_fs_cleanup),
4579 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4580 		     mlx5_ib_stage_caps_init,
4581 		     mlx5_ib_stage_caps_cleanup),
4582 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4583 		     mlx5_ib_stage_raw_eth_non_default_cb,
4584 		     NULL),
4585 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4586 		     mlx5_ib_roce_init,
4587 		     mlx5_ib_roce_cleanup),
4588 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4589 		     mlx5_init_qp_table,
4590 		     mlx5_cleanup_qp_table),
4591 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4592 		     mlx5_init_srq_table,
4593 		     mlx5_cleanup_srq_table),
4594 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4595 		     mlx5_ib_dev_res_init,
4596 		     mlx5_ib_dev_res_cleanup),
4597 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4598 		     mlx5_ib_stage_dev_notifier_init,
4599 		     mlx5_ib_stage_dev_notifier_cleanup),
4600 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4601 		     mlx5_ib_counters_init,
4602 		     mlx5_ib_counters_cleanup),
4603 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4604 		     mlx5_ib_stage_cong_debugfs_init,
4605 		     mlx5_ib_stage_cong_debugfs_cleanup),
4606 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4607 		     mlx5_ib_stage_uar_init,
4608 		     mlx5_ib_stage_uar_cleanup),
4609 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4610 		     mlx5_ib_stage_bfrag_init,
4611 		     mlx5_ib_stage_bfrag_cleanup),
4612 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4613 		     NULL,
4614 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4615 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4616 		     mlx5_ib_devx_init,
4617 		     mlx5_ib_devx_cleanup),
4618 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4619 		     mlx5_ib_stage_ib_reg_init,
4620 		     mlx5_ib_stage_ib_reg_cleanup),
4621 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4622 		     mlx5_ib_stage_post_ib_reg_umr_init,
4623 		     NULL),
4624 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4625 		     mlx5_ib_restrack_init,
4626 		     NULL),
4627 };
4628 
4629 static int mlx5r_mp_probe(struct auxiliary_device *adev,
4630 			  const struct auxiliary_device_id *id)
4631 {
4632 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4633 	struct mlx5_core_dev *mdev = idev->mdev;
4634 	struct mlx5_ib_multiport_info *mpi;
4635 	struct mlx5_ib_dev *dev;
4636 	bool bound = false;
4637 	int err;
4638 
4639 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4640 	if (!mpi)
4641 		return -ENOMEM;
4642 
4643 	mpi->mdev = mdev;
4644 	err = mlx5_query_nic_vport_system_image_guid(mdev,
4645 						     &mpi->sys_image_guid);
4646 	if (err) {
4647 		kfree(mpi);
4648 		return err;
4649 	}
4650 
4651 	mutex_lock(&mlx5_ib_multiport_mutex);
4652 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
4653 		if (dev->sys_image_guid == mpi->sys_image_guid)
4654 			bound = mlx5_ib_bind_slave_port(dev, mpi);
4655 
4656 		if (bound) {
4657 			rdma_roce_rescan_device(&dev->ib_dev);
4658 			break;
4659 		}
4660 	}
4661 
4662 	if (!bound) {
4663 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4664 		dev_dbg(mdev->device,
4665 			"no suitable IB device found to bind to, added to unaffiliated list.\n");
4666 	}
4667 	mutex_unlock(&mlx5_ib_multiport_mutex);
4668 
4669 	dev_set_drvdata(&adev->dev, mpi);
4670 	return 0;
4671 }
4672 
4673 static void mlx5r_mp_remove(struct auxiliary_device *adev)
4674 {
4675 	struct mlx5_ib_multiport_info *mpi;
4676 
4677 	mpi = dev_get_drvdata(&adev->dev);
4678 	mutex_lock(&mlx5_ib_multiport_mutex);
4679 	if (mpi->ibdev)
4680 		mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
4681 	list_del(&mpi->list);
4682 	mutex_unlock(&mlx5_ib_multiport_mutex);
4683 	kfree(mpi);
4684 }
4685 
4686 static int mlx5r_probe(struct auxiliary_device *adev,
4687 		       const struct auxiliary_device_id *id)
4688 {
4689 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4690 	struct mlx5_core_dev *mdev = idev->mdev;
4691 	const struct mlx5_ib_profile *profile;
4692 	int port_type_cap, num_ports, ret;
4693 	enum rdma_link_layer ll;
4694 	struct mlx5_ib_dev *dev;
4695 
4696 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4697 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4698 
4699 	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
4700 			MLX5_CAP_GEN(mdev, num_vhca_ports));
4701 	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
4702 	if (!dev)
4703 		return -ENOMEM;
4704 	dev->port = kcalloc(num_ports, sizeof(*dev->port),
4705 			     GFP_KERNEL);
4706 	if (!dev->port) {
4707 		ib_dealloc_device(&dev->ib_dev);
4708 		return -ENOMEM;
4709 	}
4710 
4711 	dev->mdev = mdev;
4712 	dev->num_ports = num_ports;
4713 
4714 	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_is_roce_enabled(mdev))
4715 		profile = &raw_eth_profile;
4716 	else
4717 		profile = &pf_profile;
4718 
4719 	ret = __mlx5_ib_add(dev, profile);
4720 	if (ret) {
4721 		kfree(dev->port);
4722 		ib_dealloc_device(&dev->ib_dev);
4723 		return ret;
4724 	}
4725 
4726 	dev_set_drvdata(&adev->dev, dev);
4727 	return 0;
4728 }
4729 
4730 static void mlx5r_remove(struct auxiliary_device *adev)
4731 {
4732 	struct mlx5_ib_dev *dev;
4733 
4734 	dev = dev_get_drvdata(&adev->dev);
4735 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
4736 }
4737 
4738 static const struct auxiliary_device_id mlx5r_mp_id_table[] = {
4739 	{ .name = MLX5_ADEV_NAME ".multiport", },
4740 	{},
4741 };
4742 
4743 static const struct auxiliary_device_id mlx5r_id_table[] = {
4744 	{ .name = MLX5_ADEV_NAME ".rdma", },
4745 	{},
4746 };
4747 
4748 MODULE_DEVICE_TABLE(auxiliary, mlx5r_mp_id_table);
4749 MODULE_DEVICE_TABLE(auxiliary, mlx5r_id_table);
4750 
4751 static struct auxiliary_driver mlx5r_mp_driver = {
4752 	.name = "multiport",
4753 	.probe = mlx5r_mp_probe,
4754 	.remove = mlx5r_mp_remove,
4755 	.id_table = mlx5r_mp_id_table,
4756 };
4757 
4758 static struct auxiliary_driver mlx5r_driver = {
4759 	.name = "rdma",
4760 	.probe = mlx5r_probe,
4761 	.remove = mlx5r_remove,
4762 	.id_table = mlx5r_id_table,
4763 };
4764 
4765 static int __init mlx5_ib_init(void)
4766 {
4767 	int ret;
4768 
4769 	xlt_emergency_page = (void *)__get_free_page(GFP_KERNEL);
4770 	if (!xlt_emergency_page)
4771 		return -ENOMEM;
4772 
4773 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
4774 	if (!mlx5_ib_event_wq) {
4775 		free_page((unsigned long)xlt_emergency_page);
4776 		return -ENOMEM;
4777 	}
4778 
4779 	mlx5_ib_odp_init();
4780 	ret = mlx5r_rep_init();
4781 	if (ret)
4782 		goto rep_err;
4783 	ret = auxiliary_driver_register(&mlx5r_mp_driver);
4784 	if (ret)
4785 		goto mp_err;
4786 	ret = auxiliary_driver_register(&mlx5r_driver);
4787 	if (ret)
4788 		goto drv_err;
4789 	return 0;
4790 
4791 drv_err:
4792 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4793 mp_err:
4794 	mlx5r_rep_cleanup();
4795 rep_err:
4796 	destroy_workqueue(mlx5_ib_event_wq);
4797 	free_page((unsigned long)xlt_emergency_page);
4798 	return ret;
4799 }
4800 
4801 static void __exit mlx5_ib_cleanup(void)
4802 {
4803 	auxiliary_driver_unregister(&mlx5r_driver);
4804 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4805 	mlx5r_rep_cleanup();
4806 
4807 	destroy_workqueue(mlx5_ib_event_wq);
4808 	free_page((unsigned long)xlt_emergency_page);
4809 }
4810 
4811 module_init(mlx5_ib_init);
4812 module_exit(mlx5_ib_cleanup);
4813