xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision 43ffe817)
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2013-2020, Mellanox Technologies inc. All rights reserved.
4  */
5 
6 #include <linux/debugfs.h>
7 #include <linux/highmem.h>
8 #include <linux/module.h>
9 #include <linux/init.h>
10 #include <linux/errno.h>
11 #include <linux/pci.h>
12 #include <linux/dma-mapping.h>
13 #include <linux/slab.h>
14 #include <linux/bitmap.h>
15 #include <linux/sched.h>
16 #include <linux/sched/mm.h>
17 #include <linux/sched/task.h>
18 #include <linux/delay.h>
19 #include <rdma/ib_user_verbs.h>
20 #include <rdma/ib_addr.h>
21 #include <rdma/ib_cache.h>
22 #include <linux/mlx5/port.h>
23 #include <linux/mlx5/vport.h>
24 #include <linux/mlx5/fs.h>
25 #include <linux/mlx5/eswitch.h>
26 #include <linux/list.h>
27 #include <rdma/ib_smi.h>
28 #include <rdma/ib_umem.h>
29 #include <rdma/lag.h>
30 #include <linux/in.h>
31 #include <linux/etherdevice.h>
32 #include "mlx5_ib.h"
33 #include "ib_rep.h"
34 #include "cmd.h"
35 #include "devx.h"
36 #include "fs.h"
37 #include "srq.h"
38 #include "qp.h"
39 #include "wr.h"
40 #include "restrack.h"
41 #include "counters.h"
42 #include <linux/mlx5/accel.h>
43 #include <rdma/uverbs_std_types.h>
44 #include <rdma/mlx5_user_ioctl_verbs.h>
45 #include <rdma/mlx5_user_ioctl_cmds.h>
46 #include <rdma/ib_umem_odp.h>
47 
48 #define UVERBS_MODULE_NAME mlx5_ib
49 #include <rdma/uverbs_named_ioctl.h>
50 
51 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
52 MODULE_DESCRIPTION("Mellanox 5th generation network adapters (ConnectX series) IB driver");
53 MODULE_LICENSE("Dual BSD/GPL");
54 
55 struct mlx5_ib_event_work {
56 	struct work_struct	work;
57 	union {
58 		struct mlx5_ib_dev	      *dev;
59 		struct mlx5_ib_multiport_info *mpi;
60 	};
61 	bool			is_slave;
62 	unsigned int		event;
63 	void			*param;
64 };
65 
66 enum {
67 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
68 };
69 
70 static struct workqueue_struct *mlx5_ib_event_wq;
71 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
72 static LIST_HEAD(mlx5_ib_dev_list);
73 /*
74  * This mutex should be held when accessing either of the above lists
75  */
76 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
77 
78 /* We can't use an array for xlt_emergency_page because dma_map_single
79  * doesn't work on kernel modules memory
80  */
81 static unsigned long xlt_emergency_page;
82 static struct mutex xlt_emergency_page_mutex;
83 
84 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
85 {
86 	struct mlx5_ib_dev *dev;
87 
88 	mutex_lock(&mlx5_ib_multiport_mutex);
89 	dev = mpi->ibdev;
90 	mutex_unlock(&mlx5_ib_multiport_mutex);
91 	return dev;
92 }
93 
94 static enum rdma_link_layer
95 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
96 {
97 	switch (port_type_cap) {
98 	case MLX5_CAP_PORT_TYPE_IB:
99 		return IB_LINK_LAYER_INFINIBAND;
100 	case MLX5_CAP_PORT_TYPE_ETH:
101 		return IB_LINK_LAYER_ETHERNET;
102 	default:
103 		return IB_LINK_LAYER_UNSPECIFIED;
104 	}
105 }
106 
107 static enum rdma_link_layer
108 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
109 {
110 	struct mlx5_ib_dev *dev = to_mdev(device);
111 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
112 
113 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
114 }
115 
116 static int get_port_state(struct ib_device *ibdev,
117 			  u8 port_num,
118 			  enum ib_port_state *state)
119 {
120 	struct ib_port_attr attr;
121 	int ret;
122 
123 	memset(&attr, 0, sizeof(attr));
124 	ret = ibdev->ops.query_port(ibdev, port_num, &attr);
125 	if (!ret)
126 		*state = attr.state;
127 	return ret;
128 }
129 
130 static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev,
131 					   struct net_device *ndev,
132 					   u8 *port_num)
133 {
134 	struct mlx5_eswitch *esw = dev->mdev->priv.eswitch;
135 	struct net_device *rep_ndev;
136 	struct mlx5_ib_port *port;
137 	int i;
138 
139 	for (i = 0; i < dev->num_ports; i++) {
140 		port  = &dev->port[i];
141 		if (!port->rep)
142 			continue;
143 
144 		read_lock(&port->roce.netdev_lock);
145 		rep_ndev = mlx5_ib_get_rep_netdev(esw,
146 						  port->rep->vport);
147 		if (rep_ndev == ndev) {
148 			read_unlock(&port->roce.netdev_lock);
149 			*port_num = i + 1;
150 			return &port->roce;
151 		}
152 		read_unlock(&port->roce.netdev_lock);
153 	}
154 
155 	return NULL;
156 }
157 
158 static int mlx5_netdev_event(struct notifier_block *this,
159 			     unsigned long event, void *ptr)
160 {
161 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
162 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
163 	u8 port_num = roce->native_port_num;
164 	struct mlx5_core_dev *mdev;
165 	struct mlx5_ib_dev *ibdev;
166 
167 	ibdev = roce->dev;
168 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
169 	if (!mdev)
170 		return NOTIFY_DONE;
171 
172 	switch (event) {
173 	case NETDEV_REGISTER:
174 		/* Should already be registered during the load */
175 		if (ibdev->is_rep)
176 			break;
177 		write_lock(&roce->netdev_lock);
178 		if (ndev->dev.parent == mdev->device)
179 			roce->netdev = ndev;
180 		write_unlock(&roce->netdev_lock);
181 		break;
182 
183 	case NETDEV_UNREGISTER:
184 		/* In case of reps, ib device goes away before the netdevs */
185 		write_lock(&roce->netdev_lock);
186 		if (roce->netdev == ndev)
187 			roce->netdev = NULL;
188 		write_unlock(&roce->netdev_lock);
189 		break;
190 
191 	case NETDEV_CHANGE:
192 	case NETDEV_UP:
193 	case NETDEV_DOWN: {
194 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
195 		struct net_device *upper = NULL;
196 
197 		if (lag_ndev) {
198 			upper = netdev_master_upper_dev_get(lag_ndev);
199 			dev_put(lag_ndev);
200 		}
201 
202 		if (ibdev->is_rep)
203 			roce = mlx5_get_rep_roce(ibdev, ndev, &port_num);
204 		if (!roce)
205 			return NOTIFY_DONE;
206 		if ((upper == ndev || (!upper && ndev == roce->netdev))
207 		    && ibdev->ib_active) {
208 			struct ib_event ibev = { };
209 			enum ib_port_state port_state;
210 
211 			if (get_port_state(&ibdev->ib_dev, port_num,
212 					   &port_state))
213 				goto done;
214 
215 			if (roce->last_port_state == port_state)
216 				goto done;
217 
218 			roce->last_port_state = port_state;
219 			ibev.device = &ibdev->ib_dev;
220 			if (port_state == IB_PORT_DOWN)
221 				ibev.event = IB_EVENT_PORT_ERR;
222 			else if (port_state == IB_PORT_ACTIVE)
223 				ibev.event = IB_EVENT_PORT_ACTIVE;
224 			else
225 				goto done;
226 
227 			ibev.element.port_num = port_num;
228 			ib_dispatch_event(&ibev);
229 		}
230 		break;
231 	}
232 
233 	default:
234 		break;
235 	}
236 done:
237 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
238 	return NOTIFY_DONE;
239 }
240 
241 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
242 					     u8 port_num)
243 {
244 	struct mlx5_ib_dev *ibdev = to_mdev(device);
245 	struct net_device *ndev;
246 	struct mlx5_core_dev *mdev;
247 
248 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
249 	if (!mdev)
250 		return NULL;
251 
252 	ndev = mlx5_lag_get_roce_netdev(mdev);
253 	if (ndev)
254 		goto out;
255 
256 	/* Ensure ndev does not disappear before we invoke dev_hold()
257 	 */
258 	read_lock(&ibdev->port[port_num - 1].roce.netdev_lock);
259 	ndev = ibdev->port[port_num - 1].roce.netdev;
260 	if (ndev)
261 		dev_hold(ndev);
262 	read_unlock(&ibdev->port[port_num - 1].roce.netdev_lock);
263 
264 out:
265 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
266 	return ndev;
267 }
268 
269 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
270 						   u8 ib_port_num,
271 						   u8 *native_port_num)
272 {
273 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
274 							  ib_port_num);
275 	struct mlx5_core_dev *mdev = NULL;
276 	struct mlx5_ib_multiport_info *mpi;
277 	struct mlx5_ib_port *port;
278 
279 	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
280 	    ll != IB_LINK_LAYER_ETHERNET) {
281 		if (native_port_num)
282 			*native_port_num = ib_port_num;
283 		return ibdev->mdev;
284 	}
285 
286 	if (native_port_num)
287 		*native_port_num = 1;
288 
289 	port = &ibdev->port[ib_port_num - 1];
290 	spin_lock(&port->mp.mpi_lock);
291 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
292 	if (mpi && !mpi->unaffiliate) {
293 		mdev = mpi->mdev;
294 		/* If it's the master no need to refcount, it'll exist
295 		 * as long as the ib_dev exists.
296 		 */
297 		if (!mpi->is_master)
298 			mpi->mdev_refcnt++;
299 	}
300 	spin_unlock(&port->mp.mpi_lock);
301 
302 	return mdev;
303 }
304 
305 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u8 port_num)
306 {
307 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
308 							  port_num);
309 	struct mlx5_ib_multiport_info *mpi;
310 	struct mlx5_ib_port *port;
311 
312 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
313 		return;
314 
315 	port = &ibdev->port[port_num - 1];
316 
317 	spin_lock(&port->mp.mpi_lock);
318 	mpi = ibdev->port[port_num - 1].mp.mpi;
319 	if (mpi->is_master)
320 		goto out;
321 
322 	mpi->mdev_refcnt--;
323 	if (mpi->unaffiliate)
324 		complete(&mpi->unref_comp);
325 out:
326 	spin_unlock(&port->mp.mpi_lock);
327 }
328 
329 static int translate_eth_legacy_proto_oper(u32 eth_proto_oper,
330 					   u16 *active_speed, u8 *active_width)
331 {
332 	switch (eth_proto_oper) {
333 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
334 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
335 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
336 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
337 		*active_width = IB_WIDTH_1X;
338 		*active_speed = IB_SPEED_SDR;
339 		break;
340 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
341 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
342 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
343 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
344 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
345 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
346 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
347 		*active_width = IB_WIDTH_1X;
348 		*active_speed = IB_SPEED_QDR;
349 		break;
350 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
351 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
352 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
353 		*active_width = IB_WIDTH_1X;
354 		*active_speed = IB_SPEED_EDR;
355 		break;
356 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
357 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
358 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
359 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
360 		*active_width = IB_WIDTH_4X;
361 		*active_speed = IB_SPEED_QDR;
362 		break;
363 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
364 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
365 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
366 		*active_width = IB_WIDTH_1X;
367 		*active_speed = IB_SPEED_HDR;
368 		break;
369 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
370 		*active_width = IB_WIDTH_4X;
371 		*active_speed = IB_SPEED_FDR;
372 		break;
373 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
374 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
375 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
376 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
377 		*active_width = IB_WIDTH_4X;
378 		*active_speed = IB_SPEED_EDR;
379 		break;
380 	default:
381 		return -EINVAL;
382 	}
383 
384 	return 0;
385 }
386 
387 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u16 *active_speed,
388 					u8 *active_width)
389 {
390 	switch (eth_proto_oper) {
391 	case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
392 	case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
393 		*active_width = IB_WIDTH_1X;
394 		*active_speed = IB_SPEED_SDR;
395 		break;
396 	case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
397 		*active_width = IB_WIDTH_1X;
398 		*active_speed = IB_SPEED_DDR;
399 		break;
400 	case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
401 		*active_width = IB_WIDTH_1X;
402 		*active_speed = IB_SPEED_QDR;
403 		break;
404 	case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
405 		*active_width = IB_WIDTH_4X;
406 		*active_speed = IB_SPEED_QDR;
407 		break;
408 	case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
409 		*active_width = IB_WIDTH_1X;
410 		*active_speed = IB_SPEED_EDR;
411 		break;
412 	case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
413 		*active_width = IB_WIDTH_2X;
414 		*active_speed = IB_SPEED_EDR;
415 		break;
416 	case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
417 		*active_width = IB_WIDTH_1X;
418 		*active_speed = IB_SPEED_HDR;
419 		break;
420 	case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
421 		*active_width = IB_WIDTH_4X;
422 		*active_speed = IB_SPEED_EDR;
423 		break;
424 	case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
425 		*active_width = IB_WIDTH_2X;
426 		*active_speed = IB_SPEED_HDR;
427 		break;
428 	case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4):
429 		*active_width = IB_WIDTH_4X;
430 		*active_speed = IB_SPEED_HDR;
431 		break;
432 	default:
433 		return -EINVAL;
434 	}
435 
436 	return 0;
437 }
438 
439 static int translate_eth_proto_oper(u32 eth_proto_oper, u16 *active_speed,
440 				    u8 *active_width, bool ext)
441 {
442 	return ext ?
443 		translate_eth_ext_proto_oper(eth_proto_oper, active_speed,
444 					     active_width) :
445 		translate_eth_legacy_proto_oper(eth_proto_oper, active_speed,
446 						active_width);
447 }
448 
449 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
450 				struct ib_port_attr *props)
451 {
452 	struct mlx5_ib_dev *dev = to_mdev(device);
453 	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0};
454 	struct mlx5_core_dev *mdev;
455 	struct net_device *ndev, *upper;
456 	enum ib_mtu ndev_ib_mtu;
457 	bool put_mdev = true;
458 	u16 qkey_viol_cntr;
459 	u32 eth_prot_oper;
460 	u8 mdev_port_num;
461 	bool ext;
462 	int err;
463 
464 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
465 	if (!mdev) {
466 		/* This means the port isn't affiliated yet. Get the
467 		 * info for the master port instead.
468 		 */
469 		put_mdev = false;
470 		mdev = dev->mdev;
471 		mdev_port_num = 1;
472 		port_num = 1;
473 	}
474 
475 	/* Possible bad flows are checked before filling out props so in case
476 	 * of an error it will still be zeroed out.
477 	 * Use native port in case of reps
478 	 */
479 	if (dev->is_rep)
480 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
481 					   1);
482 	else
483 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
484 					   mdev_port_num);
485 	if (err)
486 		goto out;
487 	ext = !!MLX5_GET_ETH_PROTO(ptys_reg, out, true, eth_proto_capability);
488 	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
489 
490 	props->active_width     = IB_WIDTH_4X;
491 	props->active_speed     = IB_SPEED_QDR;
492 
493 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
494 				 &props->active_width, ext);
495 
496 	props->port_cap_flags |= IB_PORT_CM_SUP;
497 	props->ip_gids = true;
498 
499 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
500 						roce_address_table_size);
501 	props->max_mtu          = IB_MTU_4096;
502 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
503 	props->pkey_tbl_len     = 1;
504 	props->state            = IB_PORT_DOWN;
505 	props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
506 
507 	mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
508 	props->qkey_viol_cntr = qkey_viol_cntr;
509 
510 	/* If this is a stub query for an unaffiliated port stop here */
511 	if (!put_mdev)
512 		goto out;
513 
514 	ndev = mlx5_ib_get_netdev(device, port_num);
515 	if (!ndev)
516 		goto out;
517 
518 	if (dev->lag_active) {
519 		rcu_read_lock();
520 		upper = netdev_master_upper_dev_get_rcu(ndev);
521 		if (upper) {
522 			dev_put(ndev);
523 			ndev = upper;
524 			dev_hold(ndev);
525 		}
526 		rcu_read_unlock();
527 	}
528 
529 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
530 		props->state      = IB_PORT_ACTIVE;
531 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
532 	}
533 
534 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
535 
536 	dev_put(ndev);
537 
538 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
539 out:
540 	if (put_mdev)
541 		mlx5_ib_put_native_port_mdev(dev, port_num);
542 	return err;
543 }
544 
545 static int set_roce_addr(struct mlx5_ib_dev *dev, u8 port_num,
546 			 unsigned int index, const union ib_gid *gid,
547 			 const struct ib_gid_attr *attr)
548 {
549 	enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
550 	u16 vlan_id = 0xffff;
551 	u8 roce_version = 0;
552 	u8 roce_l3_type = 0;
553 	u8 mac[ETH_ALEN];
554 	int ret;
555 
556 	if (gid) {
557 		gid_type = attr->gid_type;
558 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
559 		if (ret)
560 			return ret;
561 	}
562 
563 	switch (gid_type) {
564 	case IB_GID_TYPE_ROCE:
565 		roce_version = MLX5_ROCE_VERSION_1;
566 		break;
567 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
568 		roce_version = MLX5_ROCE_VERSION_2;
569 		if (ipv6_addr_v4mapped((void *)gid))
570 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
571 		else
572 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
573 		break;
574 
575 	default:
576 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
577 	}
578 
579 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
580 				      roce_l3_type, gid->raw, mac,
581 				      vlan_id < VLAN_CFI_MASK, vlan_id,
582 				      port_num);
583 }
584 
585 static int mlx5_ib_add_gid(const struct ib_gid_attr *attr,
586 			   __always_unused void **context)
587 {
588 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
589 			     attr->index, &attr->gid, attr);
590 }
591 
592 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
593 			   __always_unused void **context)
594 {
595 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
596 			     attr->index, NULL, NULL);
597 }
598 
599 __be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev,
600 				   const struct ib_gid_attr *attr)
601 {
602 	if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
603 		return 0;
604 
605 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
606 }
607 
608 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
609 {
610 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
611 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
612 	return 0;
613 }
614 
615 enum {
616 	MLX5_VPORT_ACCESS_METHOD_MAD,
617 	MLX5_VPORT_ACCESS_METHOD_HCA,
618 	MLX5_VPORT_ACCESS_METHOD_NIC,
619 };
620 
621 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
622 {
623 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
624 		return MLX5_VPORT_ACCESS_METHOD_MAD;
625 
626 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
627 	    IB_LINK_LAYER_ETHERNET)
628 		return MLX5_VPORT_ACCESS_METHOD_NIC;
629 
630 	return MLX5_VPORT_ACCESS_METHOD_HCA;
631 }
632 
633 static void get_atomic_caps(struct mlx5_ib_dev *dev,
634 			    u8 atomic_size_qp,
635 			    struct ib_device_attr *props)
636 {
637 	u8 tmp;
638 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
639 	u8 atomic_req_8B_endianness_mode =
640 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
641 
642 	/* Check if HW supports 8 bytes standard atomic operations and capable
643 	 * of host endianness respond
644 	 */
645 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
646 	if (((atomic_operations & tmp) == tmp) &&
647 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
648 	    (atomic_req_8B_endianness_mode)) {
649 		props->atomic_cap = IB_ATOMIC_HCA;
650 	} else {
651 		props->atomic_cap = IB_ATOMIC_NONE;
652 	}
653 }
654 
655 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
656 			       struct ib_device_attr *props)
657 {
658 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
659 
660 	get_atomic_caps(dev, atomic_size_qp, props);
661 }
662 
663 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
664 					__be64 *sys_image_guid)
665 {
666 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
667 	struct mlx5_core_dev *mdev = dev->mdev;
668 	u64 tmp;
669 	int err;
670 
671 	switch (mlx5_get_vport_access_method(ibdev)) {
672 	case MLX5_VPORT_ACCESS_METHOD_MAD:
673 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
674 							    sys_image_guid);
675 
676 	case MLX5_VPORT_ACCESS_METHOD_HCA:
677 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
678 		break;
679 
680 	case MLX5_VPORT_ACCESS_METHOD_NIC:
681 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
682 		break;
683 
684 	default:
685 		return -EINVAL;
686 	}
687 
688 	if (!err)
689 		*sys_image_guid = cpu_to_be64(tmp);
690 
691 	return err;
692 
693 }
694 
695 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
696 				u16 *max_pkeys)
697 {
698 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
699 	struct mlx5_core_dev *mdev = dev->mdev;
700 
701 	switch (mlx5_get_vport_access_method(ibdev)) {
702 	case MLX5_VPORT_ACCESS_METHOD_MAD:
703 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
704 
705 	case MLX5_VPORT_ACCESS_METHOD_HCA:
706 	case MLX5_VPORT_ACCESS_METHOD_NIC:
707 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
708 						pkey_table_size));
709 		return 0;
710 
711 	default:
712 		return -EINVAL;
713 	}
714 }
715 
716 static int mlx5_query_vendor_id(struct ib_device *ibdev,
717 				u32 *vendor_id)
718 {
719 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
720 
721 	switch (mlx5_get_vport_access_method(ibdev)) {
722 	case MLX5_VPORT_ACCESS_METHOD_MAD:
723 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
724 
725 	case MLX5_VPORT_ACCESS_METHOD_HCA:
726 	case MLX5_VPORT_ACCESS_METHOD_NIC:
727 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
728 
729 	default:
730 		return -EINVAL;
731 	}
732 }
733 
734 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
735 				__be64 *node_guid)
736 {
737 	u64 tmp;
738 	int err;
739 
740 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
741 	case MLX5_VPORT_ACCESS_METHOD_MAD:
742 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
743 
744 	case MLX5_VPORT_ACCESS_METHOD_HCA:
745 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
746 		break;
747 
748 	case MLX5_VPORT_ACCESS_METHOD_NIC:
749 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
750 		break;
751 
752 	default:
753 		return -EINVAL;
754 	}
755 
756 	if (!err)
757 		*node_guid = cpu_to_be64(tmp);
758 
759 	return err;
760 }
761 
762 struct mlx5_reg_node_desc {
763 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
764 };
765 
766 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
767 {
768 	struct mlx5_reg_node_desc in;
769 
770 	if (mlx5_use_mad_ifc(dev))
771 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
772 
773 	memset(&in, 0, sizeof(in));
774 
775 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
776 				    sizeof(struct mlx5_reg_node_desc),
777 				    MLX5_REG_NODE_DESC, 0, 0);
778 }
779 
780 static int mlx5_ib_query_device(struct ib_device *ibdev,
781 				struct ib_device_attr *props,
782 				struct ib_udata *uhw)
783 {
784 	size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
785 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
786 	struct mlx5_core_dev *mdev = dev->mdev;
787 	int err = -ENOMEM;
788 	int max_sq_desc;
789 	int max_rq_sg;
790 	int max_sq_sg;
791 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
792 	bool raw_support = !mlx5_core_mp_enabled(mdev);
793 	struct mlx5_ib_query_device_resp resp = {};
794 	size_t resp_len;
795 	u64 max_tso;
796 
797 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
798 	if (uhw_outlen && uhw_outlen < resp_len)
799 		return -EINVAL;
800 
801 	resp.response_length = resp_len;
802 
803 	if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
804 		return -EINVAL;
805 
806 	memset(props, 0, sizeof(*props));
807 	err = mlx5_query_system_image_guid(ibdev,
808 					   &props->sys_image_guid);
809 	if (err)
810 		return err;
811 
812 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
813 	if (err)
814 		return err;
815 
816 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
817 	if (err)
818 		return err;
819 
820 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
821 		(fw_rev_min(dev->mdev) << 16) |
822 		fw_rev_sub(dev->mdev);
823 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
824 		IB_DEVICE_PORT_ACTIVE_EVENT		|
825 		IB_DEVICE_SYS_IMAGE_GUID		|
826 		IB_DEVICE_RC_RNR_NAK_GEN;
827 
828 	if (MLX5_CAP_GEN(mdev, pkv))
829 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
830 	if (MLX5_CAP_GEN(mdev, qkv))
831 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
832 	if (MLX5_CAP_GEN(mdev, apm))
833 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
834 	if (MLX5_CAP_GEN(mdev, xrc))
835 		props->device_cap_flags |= IB_DEVICE_XRC;
836 	if (MLX5_CAP_GEN(mdev, imaicl)) {
837 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
838 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
839 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
840 		/* We support 'Gappy' memory registration too */
841 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
842 	}
843 	/* IB_WR_REG_MR always requires changing the entity size with UMR */
844 	if (!MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled))
845 		props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
846 	if (MLX5_CAP_GEN(mdev, sho)) {
847 		props->device_cap_flags |= IB_DEVICE_INTEGRITY_HANDOVER;
848 		/* At this stage no support for signature handover */
849 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
850 				      IB_PROT_T10DIF_TYPE_2 |
851 				      IB_PROT_T10DIF_TYPE_3;
852 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
853 				       IB_GUARD_T10DIF_CSUM;
854 	}
855 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
856 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
857 
858 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
859 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
860 			/* Legacy bit to support old userspace libraries */
861 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
862 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
863 		}
864 
865 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
866 			props->raw_packet_caps |=
867 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
868 
869 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) {
870 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
871 			if (max_tso) {
872 				resp.tso_caps.max_tso = 1 << max_tso;
873 				resp.tso_caps.supported_qpts |=
874 					1 << IB_QPT_RAW_PACKET;
875 				resp.response_length += sizeof(resp.tso_caps);
876 			}
877 		}
878 
879 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) {
880 			resp.rss_caps.rx_hash_function =
881 						MLX5_RX_HASH_FUNC_TOEPLITZ;
882 			resp.rss_caps.rx_hash_fields_mask =
883 						MLX5_RX_HASH_SRC_IPV4 |
884 						MLX5_RX_HASH_DST_IPV4 |
885 						MLX5_RX_HASH_SRC_IPV6 |
886 						MLX5_RX_HASH_DST_IPV6 |
887 						MLX5_RX_HASH_SRC_PORT_TCP |
888 						MLX5_RX_HASH_DST_PORT_TCP |
889 						MLX5_RX_HASH_SRC_PORT_UDP |
890 						MLX5_RX_HASH_DST_PORT_UDP |
891 						MLX5_RX_HASH_INNER;
892 			if (mlx5_accel_ipsec_device_caps(dev->mdev) &
893 			    MLX5_ACCEL_IPSEC_CAP_DEVICE)
894 				resp.rss_caps.rx_hash_fields_mask |=
895 					MLX5_RX_HASH_IPSEC_SPI;
896 			resp.response_length += sizeof(resp.rss_caps);
897 		}
898 	} else {
899 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen)
900 			resp.response_length += sizeof(resp.tso_caps);
901 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen)
902 			resp.response_length += sizeof(resp.rss_caps);
903 	}
904 
905 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
906 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
907 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
908 	}
909 
910 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
911 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
912 	    raw_support)
913 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
914 
915 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
916 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
917 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
918 
919 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
920 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
921 	    raw_support) {
922 		/* Legacy bit to support old userspace libraries */
923 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
924 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
925 	}
926 
927 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
928 		props->max_dm_size =
929 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
930 	}
931 
932 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
933 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
934 
935 	if (MLX5_CAP_GEN(mdev, end_pad))
936 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
937 
938 	props->vendor_part_id	   = mdev->pdev->device;
939 	props->hw_ver		   = mdev->pdev->revision;
940 
941 	props->max_mr_size	   = ~0ull;
942 	props->page_size_cap	   = ~(min_page_size - 1);
943 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
944 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
945 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
946 		     sizeof(struct mlx5_wqe_data_seg);
947 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
948 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
949 		     sizeof(struct mlx5_wqe_raddr_seg)) /
950 		sizeof(struct mlx5_wqe_data_seg);
951 	props->max_send_sge = max_sq_sg;
952 	props->max_recv_sge = max_rq_sg;
953 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
954 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
955 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
956 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
957 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
958 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
959 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
960 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
961 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
962 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
963 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
964 	props->max_srq_sge	   = max_rq_sg - 1;
965 	props->max_fast_reg_page_list_len =
966 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
967 	props->max_pi_fast_reg_page_list_len =
968 		props->max_fast_reg_page_list_len / 2;
969 	props->max_sgl_rd =
970 		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
971 	get_atomic_caps_qp(dev, props);
972 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
973 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
974 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
975 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
976 					   props->max_mcast_grp;
977 	props->max_ah = INT_MAX;
978 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
979 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
980 
981 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
982 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
983 			props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
984 		props->odp_caps = dev->odp_caps;
985 		if (!uhw) {
986 			/* ODP for kernel QPs is not implemented for receive
987 			 * WQEs and SRQ WQEs
988 			 */
989 			props->odp_caps.per_transport_caps.rc_odp_caps &=
990 				~(IB_ODP_SUPPORT_READ |
991 				  IB_ODP_SUPPORT_SRQ_RECV);
992 			props->odp_caps.per_transport_caps.uc_odp_caps &=
993 				~(IB_ODP_SUPPORT_READ |
994 				  IB_ODP_SUPPORT_SRQ_RECV);
995 			props->odp_caps.per_transport_caps.ud_odp_caps &=
996 				~(IB_ODP_SUPPORT_READ |
997 				  IB_ODP_SUPPORT_SRQ_RECV);
998 			props->odp_caps.per_transport_caps.xrc_odp_caps &=
999 				~(IB_ODP_SUPPORT_READ |
1000 				  IB_ODP_SUPPORT_SRQ_RECV);
1001 		}
1002 	}
1003 
1004 	if (MLX5_CAP_GEN(mdev, cd))
1005 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
1006 
1007 	if (mlx5_core_is_vf(mdev))
1008 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
1009 
1010 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1011 	    IB_LINK_LAYER_ETHERNET && raw_support) {
1012 		props->rss_caps.max_rwq_indirection_tables =
1013 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1014 		props->rss_caps.max_rwq_indirection_table_size =
1015 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1016 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1017 		props->max_wq_type_rq =
1018 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1019 	}
1020 
1021 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1022 		props->tm_caps.max_num_tags =
1023 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1024 		props->tm_caps.max_ops =
1025 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1026 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1027 	}
1028 
1029 	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1030 	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1031 		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1032 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1033 	}
1034 
1035 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1036 		props->cq_caps.max_cq_moderation_count =
1037 						MLX5_MAX_CQ_COUNT;
1038 		props->cq_caps.max_cq_moderation_period =
1039 						MLX5_MAX_CQ_PERIOD;
1040 	}
1041 
1042 	if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) {
1043 		resp.response_length += sizeof(resp.cqe_comp_caps);
1044 
1045 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1046 			resp.cqe_comp_caps.max_num =
1047 				MLX5_CAP_GEN(dev->mdev,
1048 					     cqe_compression_max_num);
1049 
1050 			resp.cqe_comp_caps.supported_format =
1051 				MLX5_IB_CQE_RES_FORMAT_HASH |
1052 				MLX5_IB_CQE_RES_FORMAT_CSUM;
1053 
1054 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1055 				resp.cqe_comp_caps.supported_format |=
1056 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1057 		}
1058 	}
1059 
1060 	if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen &&
1061 	    raw_support) {
1062 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1063 		    MLX5_CAP_GEN(mdev, qos)) {
1064 			resp.packet_pacing_caps.qp_rate_limit_max =
1065 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1066 			resp.packet_pacing_caps.qp_rate_limit_min =
1067 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1068 			resp.packet_pacing_caps.supported_qpts |=
1069 				1 << IB_QPT_RAW_PACKET;
1070 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1071 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1072 				resp.packet_pacing_caps.cap_flags |=
1073 					MLX5_IB_PP_SUPPORT_BURST;
1074 		}
1075 		resp.response_length += sizeof(resp.packet_pacing_caps);
1076 	}
1077 
1078 	if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <=
1079 	    uhw_outlen) {
1080 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1081 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1082 				MLX5_IB_ALLOW_MPW;
1083 
1084 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1085 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1086 				MLX5_IB_SUPPORT_EMPW;
1087 
1088 		resp.response_length +=
1089 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1090 	}
1091 
1092 	if (offsetofend(typeof(resp), flags) <= uhw_outlen) {
1093 		resp.response_length += sizeof(resp.flags);
1094 
1095 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1096 			resp.flags |=
1097 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1098 
1099 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1100 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1101 		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1102 			resp.flags |=
1103 				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1104 
1105 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1106 	}
1107 
1108 	if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) {
1109 		resp.response_length += sizeof(resp.sw_parsing_caps);
1110 		if (MLX5_CAP_ETH(mdev, swp)) {
1111 			resp.sw_parsing_caps.sw_parsing_offloads |=
1112 				MLX5_IB_SW_PARSING;
1113 
1114 			if (MLX5_CAP_ETH(mdev, swp_csum))
1115 				resp.sw_parsing_caps.sw_parsing_offloads |=
1116 					MLX5_IB_SW_PARSING_CSUM;
1117 
1118 			if (MLX5_CAP_ETH(mdev, swp_lso))
1119 				resp.sw_parsing_caps.sw_parsing_offloads |=
1120 					MLX5_IB_SW_PARSING_LSO;
1121 
1122 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1123 				resp.sw_parsing_caps.supported_qpts =
1124 					BIT(IB_QPT_RAW_PACKET);
1125 		}
1126 	}
1127 
1128 	if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen &&
1129 	    raw_support) {
1130 		resp.response_length += sizeof(resp.striding_rq_caps);
1131 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1132 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1133 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1134 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1135 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1136 			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1137 				resp.striding_rq_caps
1138 					.min_single_wqe_log_num_of_strides =
1139 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1140 			else
1141 				resp.striding_rq_caps
1142 					.min_single_wqe_log_num_of_strides =
1143 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1144 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1145 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1146 			resp.striding_rq_caps.supported_qpts =
1147 				BIT(IB_QPT_RAW_PACKET);
1148 		}
1149 	}
1150 
1151 	if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) {
1152 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1153 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1154 			resp.tunnel_offloads_caps |=
1155 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1156 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1157 			resp.tunnel_offloads_caps |=
1158 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1159 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1160 			resp.tunnel_offloads_caps |=
1161 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1162 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre))
1163 			resp.tunnel_offloads_caps |=
1164 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1165 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp))
1166 			resp.tunnel_offloads_caps |=
1167 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1168 	}
1169 
1170 	if (uhw_outlen) {
1171 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1172 
1173 		if (err)
1174 			return err;
1175 	}
1176 
1177 	return 0;
1178 }
1179 
1180 static void translate_active_width(struct ib_device *ibdev, u16 active_width,
1181 				   u8 *ib_width)
1182 {
1183 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1184 
1185 	if (active_width & MLX5_PTYS_WIDTH_1X)
1186 		*ib_width = IB_WIDTH_1X;
1187 	else if (active_width & MLX5_PTYS_WIDTH_2X)
1188 		*ib_width = IB_WIDTH_2X;
1189 	else if (active_width & MLX5_PTYS_WIDTH_4X)
1190 		*ib_width = IB_WIDTH_4X;
1191 	else if (active_width & MLX5_PTYS_WIDTH_8X)
1192 		*ib_width = IB_WIDTH_8X;
1193 	else if (active_width & MLX5_PTYS_WIDTH_12X)
1194 		*ib_width = IB_WIDTH_12X;
1195 	else {
1196 		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1197 			    active_width);
1198 		*ib_width = IB_WIDTH_4X;
1199 	}
1200 
1201 	return;
1202 }
1203 
1204 static int mlx5_mtu_to_ib_mtu(int mtu)
1205 {
1206 	switch (mtu) {
1207 	case 256: return 1;
1208 	case 512: return 2;
1209 	case 1024: return 3;
1210 	case 2048: return 4;
1211 	case 4096: return 5;
1212 	default:
1213 		pr_warn("invalid mtu\n");
1214 		return -1;
1215 	}
1216 }
1217 
1218 enum ib_max_vl_num {
1219 	__IB_MAX_VL_0		= 1,
1220 	__IB_MAX_VL_0_1		= 2,
1221 	__IB_MAX_VL_0_3		= 3,
1222 	__IB_MAX_VL_0_7		= 4,
1223 	__IB_MAX_VL_0_14	= 5,
1224 };
1225 
1226 enum mlx5_vl_hw_cap {
1227 	MLX5_VL_HW_0	= 1,
1228 	MLX5_VL_HW_0_1	= 2,
1229 	MLX5_VL_HW_0_2	= 3,
1230 	MLX5_VL_HW_0_3	= 4,
1231 	MLX5_VL_HW_0_4	= 5,
1232 	MLX5_VL_HW_0_5	= 6,
1233 	MLX5_VL_HW_0_6	= 7,
1234 	MLX5_VL_HW_0_7	= 8,
1235 	MLX5_VL_HW_0_14	= 15
1236 };
1237 
1238 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1239 				u8 *max_vl_num)
1240 {
1241 	switch (vl_hw_cap) {
1242 	case MLX5_VL_HW_0:
1243 		*max_vl_num = __IB_MAX_VL_0;
1244 		break;
1245 	case MLX5_VL_HW_0_1:
1246 		*max_vl_num = __IB_MAX_VL_0_1;
1247 		break;
1248 	case MLX5_VL_HW_0_3:
1249 		*max_vl_num = __IB_MAX_VL_0_3;
1250 		break;
1251 	case MLX5_VL_HW_0_7:
1252 		*max_vl_num = __IB_MAX_VL_0_7;
1253 		break;
1254 	case MLX5_VL_HW_0_14:
1255 		*max_vl_num = __IB_MAX_VL_0_14;
1256 		break;
1257 
1258 	default:
1259 		return -EINVAL;
1260 	}
1261 
1262 	return 0;
1263 }
1264 
1265 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
1266 			       struct ib_port_attr *props)
1267 {
1268 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1269 	struct mlx5_core_dev *mdev = dev->mdev;
1270 	struct mlx5_hca_vport_context *rep;
1271 	u16 max_mtu;
1272 	u16 oper_mtu;
1273 	int err;
1274 	u16 ib_link_width_oper;
1275 	u8 vl_hw_cap;
1276 
1277 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1278 	if (!rep) {
1279 		err = -ENOMEM;
1280 		goto out;
1281 	}
1282 
1283 	/* props being zeroed by the caller, avoid zeroing it here */
1284 
1285 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1286 	if (err)
1287 		goto out;
1288 
1289 	props->lid		= rep->lid;
1290 	props->lmc		= rep->lmc;
1291 	props->sm_lid		= rep->sm_lid;
1292 	props->sm_sl		= rep->sm_sl;
1293 	props->state		= rep->vport_state;
1294 	props->phys_state	= rep->port_physical_state;
1295 	props->port_cap_flags	= rep->cap_mask1;
1296 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1297 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1298 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1299 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1300 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1301 	props->subnet_timeout	= rep->subnet_timeout;
1302 	props->init_type_reply	= rep->init_type_reply;
1303 
1304 	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1305 		props->port_cap_flags2 = rep->cap_mask2;
1306 
1307 	err = mlx5_query_ib_port_oper(mdev, &ib_link_width_oper,
1308 				      &props->active_speed, port);
1309 	if (err)
1310 		goto out;
1311 
1312 	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1313 
1314 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1315 
1316 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1317 
1318 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1319 
1320 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1321 
1322 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1323 	if (err)
1324 		goto out;
1325 
1326 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1327 				   &props->max_vl_num);
1328 out:
1329 	kfree(rep);
1330 	return err;
1331 }
1332 
1333 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
1334 		       struct ib_port_attr *props)
1335 {
1336 	unsigned int count;
1337 	int ret;
1338 
1339 	switch (mlx5_get_vport_access_method(ibdev)) {
1340 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1341 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1342 		break;
1343 
1344 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1345 		ret = mlx5_query_hca_port(ibdev, port, props);
1346 		break;
1347 
1348 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1349 		ret = mlx5_query_port_roce(ibdev, port, props);
1350 		break;
1351 
1352 	default:
1353 		ret = -EINVAL;
1354 	}
1355 
1356 	if (!ret && props) {
1357 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1358 		struct mlx5_core_dev *mdev;
1359 		bool put_mdev = true;
1360 
1361 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1362 		if (!mdev) {
1363 			/* If the port isn't affiliated yet query the master.
1364 			 * The master and slave will have the same values.
1365 			 */
1366 			mdev = dev->mdev;
1367 			port = 1;
1368 			put_mdev = false;
1369 		}
1370 		count = mlx5_core_reserved_gids_count(mdev);
1371 		if (put_mdev)
1372 			mlx5_ib_put_native_port_mdev(dev, port);
1373 		props->gid_tbl_len -= count;
1374 	}
1375 	return ret;
1376 }
1377 
1378 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u8 port,
1379 				  struct ib_port_attr *props)
1380 {
1381 	int ret;
1382 
1383 	/* Only link layer == ethernet is valid for representors
1384 	 * and we always use port 1
1385 	 */
1386 	ret = mlx5_query_port_roce(ibdev, port, props);
1387 	if (ret || !props)
1388 		return ret;
1389 
1390 	/* We don't support GIDS */
1391 	props->gid_tbl_len = 0;
1392 
1393 	return ret;
1394 }
1395 
1396 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1397 			     union ib_gid *gid)
1398 {
1399 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1400 	struct mlx5_core_dev *mdev = dev->mdev;
1401 
1402 	switch (mlx5_get_vport_access_method(ibdev)) {
1403 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1404 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1405 
1406 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1407 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1408 
1409 	default:
1410 		return -EINVAL;
1411 	}
1412 
1413 }
1414 
1415 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u8 port,
1416 				   u16 index, u16 *pkey)
1417 {
1418 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1419 	struct mlx5_core_dev *mdev;
1420 	bool put_mdev = true;
1421 	u8 mdev_port_num;
1422 	int err;
1423 
1424 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1425 	if (!mdev) {
1426 		/* The port isn't affiliated yet, get the PKey from the master
1427 		 * port. For RoCE the PKey tables will be the same.
1428 		 */
1429 		put_mdev = false;
1430 		mdev = dev->mdev;
1431 		mdev_port_num = 1;
1432 	}
1433 
1434 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1435 					index, pkey);
1436 	if (put_mdev)
1437 		mlx5_ib_put_native_port_mdev(dev, port);
1438 
1439 	return err;
1440 }
1441 
1442 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1443 			      u16 *pkey)
1444 {
1445 	switch (mlx5_get_vport_access_method(ibdev)) {
1446 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1447 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1448 
1449 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1450 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1451 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1452 	default:
1453 		return -EINVAL;
1454 	}
1455 }
1456 
1457 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1458 				 struct ib_device_modify *props)
1459 {
1460 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1461 	struct mlx5_reg_node_desc in;
1462 	struct mlx5_reg_node_desc out;
1463 	int err;
1464 
1465 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1466 		return -EOPNOTSUPP;
1467 
1468 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1469 		return 0;
1470 
1471 	/*
1472 	 * If possible, pass node desc to FW, so it can generate
1473 	 * a 144 trap.  If cmd fails, just ignore.
1474 	 */
1475 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1476 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1477 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1478 	if (err)
1479 		return err;
1480 
1481 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1482 
1483 	return err;
1484 }
1485 
1486 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1487 				u32 value)
1488 {
1489 	struct mlx5_hca_vport_context ctx = {};
1490 	struct mlx5_core_dev *mdev;
1491 	u8 mdev_port_num;
1492 	int err;
1493 
1494 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1495 	if (!mdev)
1496 		return -ENODEV;
1497 
1498 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1499 	if (err)
1500 		goto out;
1501 
1502 	if (~ctx.cap_mask1_perm & mask) {
1503 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1504 			     mask, ctx.cap_mask1_perm);
1505 		err = -EINVAL;
1506 		goto out;
1507 	}
1508 
1509 	ctx.cap_mask1 = value;
1510 	ctx.cap_mask1_perm = mask;
1511 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1512 						 0, &ctx);
1513 
1514 out:
1515 	mlx5_ib_put_native_port_mdev(dev, port_num);
1516 
1517 	return err;
1518 }
1519 
1520 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1521 			       struct ib_port_modify *props)
1522 {
1523 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1524 	struct ib_port_attr attr;
1525 	u32 tmp;
1526 	int err;
1527 	u32 change_mask;
1528 	u32 value;
1529 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1530 		      IB_LINK_LAYER_INFINIBAND);
1531 
1532 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1533 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1534 	 */
1535 	if (!is_ib)
1536 		return 0;
1537 
1538 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1539 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1540 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1541 		return set_port_caps_atomic(dev, port, change_mask, value);
1542 	}
1543 
1544 	mutex_lock(&dev->cap_mask_mutex);
1545 
1546 	err = ib_query_port(ibdev, port, &attr);
1547 	if (err)
1548 		goto out;
1549 
1550 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1551 		~props->clr_port_cap_mask;
1552 
1553 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1554 
1555 out:
1556 	mutex_unlock(&dev->cap_mask_mutex);
1557 	return err;
1558 }
1559 
1560 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1561 {
1562 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1563 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1564 }
1565 
1566 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1567 {
1568 	/* Large page with non 4k uar support might limit the dynamic size */
1569 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1570 		return MLX5_MIN_DYN_BFREGS;
1571 
1572 	return MLX5_MAX_DYN_BFREGS;
1573 }
1574 
1575 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1576 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1577 			     struct mlx5_bfreg_info *bfregi)
1578 {
1579 	int uars_per_sys_page;
1580 	int bfregs_per_sys_page;
1581 	int ref_bfregs = req->total_num_bfregs;
1582 
1583 	if (req->total_num_bfregs == 0)
1584 		return -EINVAL;
1585 
1586 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1587 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1588 
1589 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1590 		return -ENOMEM;
1591 
1592 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1593 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1594 	/* This holds the required static allocation asked by the user */
1595 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1596 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1597 		return -EINVAL;
1598 
1599 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1600 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1601 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1602 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1603 
1604 	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",
1605 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1606 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1607 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1608 		    bfregi->num_sys_pages);
1609 
1610 	return 0;
1611 }
1612 
1613 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1614 {
1615 	struct mlx5_bfreg_info *bfregi;
1616 	int err;
1617 	int i;
1618 
1619 	bfregi = &context->bfregi;
1620 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1621 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1622 		if (err)
1623 			goto error;
1624 
1625 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1626 	}
1627 
1628 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1629 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1630 
1631 	return 0;
1632 
1633 error:
1634 	for (--i; i >= 0; i--)
1635 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1636 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1637 
1638 	return err;
1639 }
1640 
1641 static void deallocate_uars(struct mlx5_ib_dev *dev,
1642 			    struct mlx5_ib_ucontext *context)
1643 {
1644 	struct mlx5_bfreg_info *bfregi;
1645 	int i;
1646 
1647 	bfregi = &context->bfregi;
1648 	for (i = 0; i < bfregi->num_sys_pages; i++)
1649 		if (i < bfregi->num_static_sys_pages ||
1650 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1651 			mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1652 }
1653 
1654 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1655 {
1656 	int err = 0;
1657 
1658 	mutex_lock(&dev->lb.mutex);
1659 	if (td)
1660 		dev->lb.user_td++;
1661 	if (qp)
1662 		dev->lb.qps++;
1663 
1664 	if (dev->lb.user_td == 2 ||
1665 	    dev->lb.qps == 1) {
1666 		if (!dev->lb.enabled) {
1667 			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1668 			dev->lb.enabled = true;
1669 		}
1670 	}
1671 
1672 	mutex_unlock(&dev->lb.mutex);
1673 
1674 	return err;
1675 }
1676 
1677 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1678 {
1679 	mutex_lock(&dev->lb.mutex);
1680 	if (td)
1681 		dev->lb.user_td--;
1682 	if (qp)
1683 		dev->lb.qps--;
1684 
1685 	if (dev->lb.user_td == 1 &&
1686 	    dev->lb.qps == 0) {
1687 		if (dev->lb.enabled) {
1688 			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1689 			dev->lb.enabled = false;
1690 		}
1691 	}
1692 
1693 	mutex_unlock(&dev->lb.mutex);
1694 }
1695 
1696 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1697 					  u16 uid)
1698 {
1699 	int err;
1700 
1701 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1702 		return 0;
1703 
1704 	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1705 	if (err)
1706 		return err;
1707 
1708 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1709 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1710 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1711 		return err;
1712 
1713 	return mlx5_ib_enable_lb(dev, true, false);
1714 }
1715 
1716 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1717 					     u16 uid)
1718 {
1719 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1720 		return;
1721 
1722 	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1723 
1724 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1725 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1726 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1727 		return;
1728 
1729 	mlx5_ib_disable_lb(dev, true, false);
1730 }
1731 
1732 static int set_ucontext_resp(struct ib_ucontext *uctx,
1733 			     struct mlx5_ib_alloc_ucontext_resp *resp)
1734 {
1735 	struct ib_device *ibdev = uctx->device;
1736 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1737 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1738 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1739 	int err;
1740 
1741 	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1742 		err = mlx5_cmd_dump_fill_mkey(dev->mdev,
1743 					      &resp->dump_fill_mkey);
1744 		if (err)
1745 			return err;
1746 		resp->comp_mask |=
1747 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1748 	}
1749 
1750 	resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1751 	if (dev->wc_support)
1752 		resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev,
1753 						      log_bf_reg_size);
1754 	resp->cache_line_size = cache_line_size();
1755 	resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1756 	resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1757 	resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1758 	resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1759 	resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1760 	resp->cqe_version = context->cqe_version;
1761 	resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1762 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1763 	resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1764 					MLX5_CAP_GEN(dev->mdev,
1765 						     num_of_uars_per_page) : 1;
1766 
1767 	if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1768 				MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1769 		if (mlx5_get_flow_namespace(dev->mdev,
1770 				MLX5_FLOW_NAMESPACE_EGRESS))
1771 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1772 		if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1773 				MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1774 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1775 		if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1776 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1777 		if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1778 				MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1779 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1780 		/* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1781 	}
1782 
1783 	resp->tot_bfregs = bfregi->lib_uar_dyn ? 0 :
1784 			bfregi->total_num_bfregs - bfregi->num_dyn_bfregs;
1785 	resp->num_ports = dev->num_ports;
1786 	resp->cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1787 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1788 
1789 	if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1790 		mlx5_query_min_inline(dev->mdev, &resp->eth_min_inline);
1791 		resp->eth_min_inline++;
1792 	}
1793 
1794 	if (dev->mdev->clock_info)
1795 		resp->clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1796 
1797 	/*
1798 	 * We don't want to expose information from the PCI bar that is located
1799 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1800 	 * pretend we don't support reading the HCA's core clock. This is also
1801 	 * forced by mmap function.
1802 	 */
1803 	if (PAGE_SIZE <= 4096) {
1804 		resp->comp_mask |=
1805 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1806 		resp->hca_core_clock_offset =
1807 			offsetof(struct mlx5_init_seg,
1808 				 internal_timer_h) % PAGE_SIZE;
1809 	}
1810 
1811 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
1812 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE;
1813 
1814 	resp->num_dyn_bfregs = bfregi->num_dyn_bfregs;
1815 	return 0;
1816 }
1817 
1818 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1819 				  struct ib_udata *udata)
1820 {
1821 	struct ib_device *ibdev = uctx->device;
1822 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1823 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1824 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1825 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1826 	struct mlx5_bfreg_info *bfregi;
1827 	int ver;
1828 	int err;
1829 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1830 				     max_cqe_version);
1831 	bool lib_uar_4k;
1832 	bool lib_uar_dyn;
1833 
1834 	if (!dev->ib_active)
1835 		return -EAGAIN;
1836 
1837 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1838 		ver = 0;
1839 	else if (udata->inlen >= min_req_v2)
1840 		ver = 2;
1841 	else
1842 		return -EINVAL;
1843 
1844 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1845 	if (err)
1846 		return err;
1847 
1848 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1849 		return -EOPNOTSUPP;
1850 
1851 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1852 		return -EOPNOTSUPP;
1853 
1854 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1855 				    MLX5_NON_FP_BFREGS_PER_UAR);
1856 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1857 		return -EINVAL;
1858 
1859 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1860 	lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
1861 	bfregi = &context->bfregi;
1862 
1863 	if (lib_uar_dyn) {
1864 		bfregi->lib_uar_dyn = lib_uar_dyn;
1865 		goto uar_done;
1866 	}
1867 
1868 	/* updates req->total_num_bfregs */
1869 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1870 	if (err)
1871 		goto out_ctx;
1872 
1873 	mutex_init(&bfregi->lock);
1874 	bfregi->lib_uar_4k = lib_uar_4k;
1875 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1876 				GFP_KERNEL);
1877 	if (!bfregi->count) {
1878 		err = -ENOMEM;
1879 		goto out_ctx;
1880 	}
1881 
1882 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1883 				    sizeof(*bfregi->sys_pages),
1884 				    GFP_KERNEL);
1885 	if (!bfregi->sys_pages) {
1886 		err = -ENOMEM;
1887 		goto out_count;
1888 	}
1889 
1890 	err = allocate_uars(dev, context);
1891 	if (err)
1892 		goto out_sys_pages;
1893 
1894 uar_done:
1895 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1896 		err = mlx5_ib_devx_create(dev, true);
1897 		if (err < 0)
1898 			goto out_uars;
1899 		context->devx_uid = err;
1900 	}
1901 
1902 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1903 					     context->devx_uid);
1904 	if (err)
1905 		goto out_devx;
1906 
1907 	INIT_LIST_HEAD(&context->db_page_list);
1908 	mutex_init(&context->db_page_mutex);
1909 
1910 	context->cqe_version = min_t(__u8,
1911 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1912 				 req.max_cqe_version);
1913 
1914 	err = set_ucontext_resp(uctx, &resp);
1915 	if (err)
1916 		goto out_mdev;
1917 
1918 	resp.response_length = min(udata->outlen, sizeof(resp));
1919 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1920 	if (err)
1921 		goto out_mdev;
1922 
1923 	bfregi->ver = ver;
1924 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1925 	context->lib_caps = req.lib_caps;
1926 	print_lib_caps(dev, context->lib_caps);
1927 
1928 	if (mlx5_ib_lag_should_assign_affinity(dev)) {
1929 		u8 port = mlx5_core_native_port_num(dev->mdev) - 1;
1930 
1931 		atomic_set(&context->tx_port_affinity,
1932 			   atomic_add_return(
1933 				   1, &dev->port[port].roce.tx_port_affinity));
1934 	}
1935 
1936 	return 0;
1937 
1938 out_mdev:
1939 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1940 out_devx:
1941 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1942 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1943 
1944 out_uars:
1945 	deallocate_uars(dev, context);
1946 
1947 out_sys_pages:
1948 	kfree(bfregi->sys_pages);
1949 
1950 out_count:
1951 	kfree(bfregi->count);
1952 
1953 out_ctx:
1954 	return err;
1955 }
1956 
1957 static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext,
1958 				  struct uverbs_attr_bundle *attrs)
1959 {
1960 	struct mlx5_ib_alloc_ucontext_resp uctx_resp = {};
1961 	int ret;
1962 
1963 	ret = set_ucontext_resp(ibcontext, &uctx_resp);
1964 	if (ret)
1965 		return ret;
1966 
1967 	uctx_resp.response_length =
1968 		min_t(size_t,
1969 		      uverbs_attr_get_len(attrs,
1970 				MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX),
1971 		      sizeof(uctx_resp));
1972 
1973 	ret = uverbs_copy_to_struct_or_zero(attrs,
1974 					MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
1975 					&uctx_resp,
1976 					sizeof(uctx_resp));
1977 	return ret;
1978 }
1979 
1980 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1981 {
1982 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1983 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1984 	struct mlx5_bfreg_info *bfregi;
1985 
1986 	bfregi = &context->bfregi;
1987 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1988 
1989 	if (context->devx_uid)
1990 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1991 
1992 	deallocate_uars(dev, context);
1993 	kfree(bfregi->sys_pages);
1994 	kfree(bfregi->count);
1995 }
1996 
1997 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
1998 				 int uar_idx)
1999 {
2000 	int fw_uars_per_page;
2001 
2002 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2003 
2004 	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2005 }
2006 
2007 static u64 uar_index2paddress(struct mlx5_ib_dev *dev,
2008 				 int uar_idx)
2009 {
2010 	unsigned int fw_uars_per_page;
2011 
2012 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
2013 				MLX5_UARS_IN_PAGE : 1;
2014 
2015 	return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE);
2016 }
2017 
2018 static int get_command(unsigned long offset)
2019 {
2020 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2021 }
2022 
2023 static int get_arg(unsigned long offset)
2024 {
2025 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2026 }
2027 
2028 static int get_index(unsigned long offset)
2029 {
2030 	return get_arg(offset);
2031 }
2032 
2033 /* Index resides in an extra byte to enable larger values than 255 */
2034 static int get_extended_index(unsigned long offset)
2035 {
2036 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2037 }
2038 
2039 
2040 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2041 {
2042 }
2043 
2044 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2045 {
2046 	switch (cmd) {
2047 	case MLX5_IB_MMAP_WC_PAGE:
2048 		return "WC";
2049 	case MLX5_IB_MMAP_REGULAR_PAGE:
2050 		return "best effort WC";
2051 	case MLX5_IB_MMAP_NC_PAGE:
2052 		return "NC";
2053 	case MLX5_IB_MMAP_DEVICE_MEM:
2054 		return "Device Memory";
2055 	default:
2056 		return NULL;
2057 	}
2058 }
2059 
2060 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2061 					struct vm_area_struct *vma,
2062 					struct mlx5_ib_ucontext *context)
2063 {
2064 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2065 	    !(vma->vm_flags & VM_SHARED))
2066 		return -EINVAL;
2067 
2068 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2069 		return -EOPNOTSUPP;
2070 
2071 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2072 		return -EPERM;
2073 	vma->vm_flags &= ~VM_MAYWRITE;
2074 
2075 	if (!dev->mdev->clock_info)
2076 		return -EOPNOTSUPP;
2077 
2078 	return vm_insert_page(vma, vma->vm_start,
2079 			      virt_to_page(dev->mdev->clock_info));
2080 }
2081 
2082 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2083 {
2084 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2085 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2086 	struct mlx5_var_table *var_table = &dev->var_table;
2087 	struct mlx5_ib_dm *mdm;
2088 
2089 	switch (mentry->mmap_flag) {
2090 	case MLX5_IB_MMAP_TYPE_MEMIC:
2091 		mdm = container_of(mentry, struct mlx5_ib_dm, mentry);
2092 		mlx5_cmd_dealloc_memic(&dev->dm, mdm->dev_addr,
2093 				       mdm->size);
2094 		kfree(mdm);
2095 		break;
2096 	case MLX5_IB_MMAP_TYPE_VAR:
2097 		mutex_lock(&var_table->bitmap_lock);
2098 		clear_bit(mentry->page_idx, var_table->bitmap);
2099 		mutex_unlock(&var_table->bitmap_lock);
2100 		kfree(mentry);
2101 		break;
2102 	case MLX5_IB_MMAP_TYPE_UAR_WC:
2103 	case MLX5_IB_MMAP_TYPE_UAR_NC:
2104 		mlx5_cmd_free_uar(dev->mdev, mentry->page_idx);
2105 		kfree(mentry);
2106 		break;
2107 	default:
2108 		WARN_ON(true);
2109 	}
2110 }
2111 
2112 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2113 		    struct vm_area_struct *vma,
2114 		    struct mlx5_ib_ucontext *context)
2115 {
2116 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2117 	int err;
2118 	unsigned long idx;
2119 	phys_addr_t pfn;
2120 	pgprot_t prot;
2121 	u32 bfreg_dyn_idx = 0;
2122 	u32 uar_index;
2123 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2124 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2125 				bfregi->num_static_sys_pages;
2126 
2127 	if (bfregi->lib_uar_dyn)
2128 		return -EINVAL;
2129 
2130 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2131 		return -EINVAL;
2132 
2133 	if (dyn_uar)
2134 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2135 	else
2136 		idx = get_index(vma->vm_pgoff);
2137 
2138 	if (idx >= max_valid_idx) {
2139 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2140 			     idx, max_valid_idx);
2141 		return -EINVAL;
2142 	}
2143 
2144 	switch (cmd) {
2145 	case MLX5_IB_MMAP_WC_PAGE:
2146 	case MLX5_IB_MMAP_ALLOC_WC:
2147 	case MLX5_IB_MMAP_REGULAR_PAGE:
2148 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2149 		prot = pgprot_writecombine(vma->vm_page_prot);
2150 		break;
2151 	case MLX5_IB_MMAP_NC_PAGE:
2152 		prot = pgprot_noncached(vma->vm_page_prot);
2153 		break;
2154 	default:
2155 		return -EINVAL;
2156 	}
2157 
2158 	if (dyn_uar) {
2159 		int uars_per_page;
2160 
2161 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2162 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2163 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2164 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2165 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2166 			return -EINVAL;
2167 		}
2168 
2169 		mutex_lock(&bfregi->lock);
2170 		/* Fail if uar already allocated, first bfreg index of each
2171 		 * page holds its count.
2172 		 */
2173 		if (bfregi->count[bfreg_dyn_idx]) {
2174 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2175 			mutex_unlock(&bfregi->lock);
2176 			return -EINVAL;
2177 		}
2178 
2179 		bfregi->count[bfreg_dyn_idx]++;
2180 		mutex_unlock(&bfregi->lock);
2181 
2182 		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2183 		if (err) {
2184 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2185 			goto free_bfreg;
2186 		}
2187 	} else {
2188 		uar_index = bfregi->sys_pages[idx];
2189 	}
2190 
2191 	pfn = uar_index2pfn(dev, uar_index);
2192 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2193 
2194 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2195 				prot, NULL);
2196 	if (err) {
2197 		mlx5_ib_err(dev,
2198 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2199 			    err, mmap_cmd2str(cmd));
2200 		goto err;
2201 	}
2202 
2203 	if (dyn_uar)
2204 		bfregi->sys_pages[idx] = uar_index;
2205 	return 0;
2206 
2207 err:
2208 	if (!dyn_uar)
2209 		return err;
2210 
2211 	mlx5_cmd_free_uar(dev->mdev, idx);
2212 
2213 free_bfreg:
2214 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2215 
2216 	return err;
2217 }
2218 
2219 static int add_dm_mmap_entry(struct ib_ucontext *context,
2220 			     struct mlx5_ib_dm *mdm,
2221 			     u64 address)
2222 {
2223 	mdm->mentry.mmap_flag = MLX5_IB_MMAP_TYPE_MEMIC;
2224 	mdm->mentry.address = address;
2225 	return rdma_user_mmap_entry_insert_range(
2226 			context, &mdm->mentry.rdma_entry,
2227 			mdm->size,
2228 			MLX5_IB_MMAP_DEVICE_MEM << 16,
2229 			(MLX5_IB_MMAP_DEVICE_MEM << 16) + (1UL << 16) - 1);
2230 }
2231 
2232 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2233 {
2234 	unsigned long idx;
2235 	u8 command;
2236 
2237 	command = get_command(vma->vm_pgoff);
2238 	idx = get_extended_index(vma->vm_pgoff);
2239 
2240 	return (command << 16 | idx);
2241 }
2242 
2243 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2244 			       struct vm_area_struct *vma,
2245 			       struct ib_ucontext *ucontext)
2246 {
2247 	struct mlx5_user_mmap_entry *mentry;
2248 	struct rdma_user_mmap_entry *entry;
2249 	unsigned long pgoff;
2250 	pgprot_t prot;
2251 	phys_addr_t pfn;
2252 	int ret;
2253 
2254 	pgoff = mlx5_vma_to_pgoff(vma);
2255 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2256 	if (!entry)
2257 		return -EINVAL;
2258 
2259 	mentry = to_mmmap(entry);
2260 	pfn = (mentry->address >> PAGE_SHIFT);
2261 	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
2262 	    mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
2263 		prot = pgprot_noncached(vma->vm_page_prot);
2264 	else
2265 		prot = pgprot_writecombine(vma->vm_page_prot);
2266 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2267 				entry->npages * PAGE_SIZE,
2268 				prot,
2269 				entry);
2270 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2271 	return ret;
2272 }
2273 
2274 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2275 {
2276 	u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2277 	u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2278 
2279 	return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2280 		(index & 0xFF)) << PAGE_SHIFT;
2281 }
2282 
2283 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2284 {
2285 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2286 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2287 	unsigned long command;
2288 	phys_addr_t pfn;
2289 
2290 	command = get_command(vma->vm_pgoff);
2291 	switch (command) {
2292 	case MLX5_IB_MMAP_WC_PAGE:
2293 	case MLX5_IB_MMAP_ALLOC_WC:
2294 		if (!dev->wc_support)
2295 			return -EPERM;
2296 		fallthrough;
2297 	case MLX5_IB_MMAP_NC_PAGE:
2298 	case MLX5_IB_MMAP_REGULAR_PAGE:
2299 		return uar_mmap(dev, command, vma, context);
2300 
2301 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2302 		return -ENOSYS;
2303 
2304 	case MLX5_IB_MMAP_CORE_CLOCK:
2305 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2306 			return -EINVAL;
2307 
2308 		if (vma->vm_flags & VM_WRITE)
2309 			return -EPERM;
2310 		vma->vm_flags &= ~VM_MAYWRITE;
2311 
2312 		/* Don't expose to user-space information it shouldn't have */
2313 		if (PAGE_SIZE > 4096)
2314 			return -EOPNOTSUPP;
2315 
2316 		pfn = (dev->mdev->iseg_base +
2317 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2318 			PAGE_SHIFT;
2319 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2320 					 PAGE_SIZE,
2321 					 pgprot_noncached(vma->vm_page_prot),
2322 					 NULL);
2323 	case MLX5_IB_MMAP_CLOCK_INFO:
2324 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2325 
2326 	default:
2327 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2328 	}
2329 
2330 	return 0;
2331 }
2332 
2333 static inline int check_dm_type_support(struct mlx5_ib_dev *dev,
2334 					u32 type)
2335 {
2336 	switch (type) {
2337 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2338 		if (!MLX5_CAP_DEV_MEM(dev->mdev, memic))
2339 			return -EOPNOTSUPP;
2340 		break;
2341 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2342 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2343 		if (!capable(CAP_SYS_RAWIO) ||
2344 		    !capable(CAP_NET_RAW))
2345 			return -EPERM;
2346 
2347 		if (!(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
2348 		      MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, sw_owner) ||
2349 		      MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner_v2) ||
2350 		      MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, sw_owner_v2)))
2351 			return -EOPNOTSUPP;
2352 		break;
2353 	}
2354 
2355 	return 0;
2356 }
2357 
2358 static int handle_alloc_dm_memic(struct ib_ucontext *ctx,
2359 				 struct mlx5_ib_dm *dm,
2360 				 struct ib_dm_alloc_attr *attr,
2361 				 struct uverbs_attr_bundle *attrs)
2362 {
2363 	struct mlx5_dm *dm_db = &to_mdev(ctx->device)->dm;
2364 	u64 start_offset;
2365 	u16 page_idx;
2366 	int err;
2367 	u64 address;
2368 
2369 	dm->size = roundup(attr->length, MLX5_MEMIC_BASE_SIZE);
2370 
2371 	err = mlx5_cmd_alloc_memic(dm_db, &dm->dev_addr,
2372 				   dm->size, attr->alignment);
2373 	if (err)
2374 		return err;
2375 
2376 	address = dm->dev_addr & PAGE_MASK;
2377 	err = add_dm_mmap_entry(ctx, dm, address);
2378 	if (err)
2379 		goto err_dealloc;
2380 
2381 	page_idx = dm->mentry.rdma_entry.start_pgoff & 0xFFFF;
2382 	err = uverbs_copy_to(attrs,
2383 			     MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
2384 			     &page_idx,
2385 			     sizeof(page_idx));
2386 	if (err)
2387 		goto err_copy;
2388 
2389 	start_offset = dm->dev_addr & ~PAGE_MASK;
2390 	err = uverbs_copy_to(attrs,
2391 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2392 			     &start_offset, sizeof(start_offset));
2393 	if (err)
2394 		goto err_copy;
2395 
2396 	return 0;
2397 
2398 err_copy:
2399 	rdma_user_mmap_entry_remove(&dm->mentry.rdma_entry);
2400 err_dealloc:
2401 	mlx5_cmd_dealloc_memic(dm_db, dm->dev_addr, dm->size);
2402 
2403 	return err;
2404 }
2405 
2406 static int handle_alloc_dm_sw_icm(struct ib_ucontext *ctx,
2407 				  struct mlx5_ib_dm *dm,
2408 				  struct ib_dm_alloc_attr *attr,
2409 				  struct uverbs_attr_bundle *attrs,
2410 				  int type)
2411 {
2412 	struct mlx5_core_dev *dev = to_mdev(ctx->device)->mdev;
2413 	u64 act_size;
2414 	int err;
2415 
2416 	/* Allocation size must a multiple of the basic block size
2417 	 * and a power of 2.
2418 	 */
2419 	act_size = round_up(attr->length, MLX5_SW_ICM_BLOCK_SIZE(dev));
2420 	act_size = roundup_pow_of_two(act_size);
2421 
2422 	dm->size = act_size;
2423 	err = mlx5_dm_sw_icm_alloc(dev, type, act_size, attr->alignment,
2424 				   to_mucontext(ctx)->devx_uid, &dm->dev_addr,
2425 				   &dm->icm_dm.obj_id);
2426 	if (err)
2427 		return err;
2428 
2429 	err = uverbs_copy_to(attrs,
2430 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2431 			     &dm->dev_addr, sizeof(dm->dev_addr));
2432 	if (err)
2433 		mlx5_dm_sw_icm_dealloc(dev, type, dm->size,
2434 				       to_mucontext(ctx)->devx_uid, dm->dev_addr,
2435 				       dm->icm_dm.obj_id);
2436 
2437 	return err;
2438 }
2439 
2440 struct ib_dm *mlx5_ib_alloc_dm(struct ib_device *ibdev,
2441 			       struct ib_ucontext *context,
2442 			       struct ib_dm_alloc_attr *attr,
2443 			       struct uverbs_attr_bundle *attrs)
2444 {
2445 	struct mlx5_ib_dm *dm;
2446 	enum mlx5_ib_uapi_dm_type type;
2447 	int err;
2448 
2449 	err = uverbs_get_const_default(&type, attrs,
2450 				       MLX5_IB_ATTR_ALLOC_DM_REQ_TYPE,
2451 				       MLX5_IB_UAPI_DM_TYPE_MEMIC);
2452 	if (err)
2453 		return ERR_PTR(err);
2454 
2455 	mlx5_ib_dbg(to_mdev(ibdev), "alloc_dm req: dm_type=%d user_length=0x%llx log_alignment=%d\n",
2456 		    type, attr->length, attr->alignment);
2457 
2458 	err = check_dm_type_support(to_mdev(ibdev), type);
2459 	if (err)
2460 		return ERR_PTR(err);
2461 
2462 	dm = kzalloc(sizeof(*dm), GFP_KERNEL);
2463 	if (!dm)
2464 		return ERR_PTR(-ENOMEM);
2465 
2466 	dm->type = type;
2467 
2468 	switch (type) {
2469 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2470 		err = handle_alloc_dm_memic(context, dm,
2471 					    attr,
2472 					    attrs);
2473 		break;
2474 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2475 		err = handle_alloc_dm_sw_icm(context, dm,
2476 					     attr, attrs,
2477 					     MLX5_SW_ICM_TYPE_STEERING);
2478 		break;
2479 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2480 		err = handle_alloc_dm_sw_icm(context, dm,
2481 					     attr, attrs,
2482 					     MLX5_SW_ICM_TYPE_HEADER_MODIFY);
2483 		break;
2484 	default:
2485 		err = -EOPNOTSUPP;
2486 	}
2487 
2488 	if (err)
2489 		goto err_free;
2490 
2491 	return &dm->ibdm;
2492 
2493 err_free:
2494 	kfree(dm);
2495 	return ERR_PTR(err);
2496 }
2497 
2498 int mlx5_ib_dealloc_dm(struct ib_dm *ibdm, struct uverbs_attr_bundle *attrs)
2499 {
2500 	struct mlx5_ib_ucontext *ctx = rdma_udata_to_drv_context(
2501 		&attrs->driver_udata, struct mlx5_ib_ucontext, ibucontext);
2502 	struct mlx5_core_dev *dev = to_mdev(ibdm->device)->mdev;
2503 	struct mlx5_ib_dm *dm = to_mdm(ibdm);
2504 	int ret;
2505 
2506 	switch (dm->type) {
2507 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2508 		rdma_user_mmap_entry_remove(&dm->mentry.rdma_entry);
2509 		return 0;
2510 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2511 		ret = mlx5_dm_sw_icm_dealloc(dev, MLX5_SW_ICM_TYPE_STEERING,
2512 					     dm->size, ctx->devx_uid, dm->dev_addr,
2513 					     dm->icm_dm.obj_id);
2514 		if (ret)
2515 			return ret;
2516 		break;
2517 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2518 		ret = mlx5_dm_sw_icm_dealloc(dev, MLX5_SW_ICM_TYPE_HEADER_MODIFY,
2519 					     dm->size, ctx->devx_uid, dm->dev_addr,
2520 					     dm->icm_dm.obj_id);
2521 		if (ret)
2522 			return ret;
2523 		break;
2524 	default:
2525 		return -EOPNOTSUPP;
2526 	}
2527 
2528 	kfree(dm);
2529 
2530 	return 0;
2531 }
2532 
2533 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2534 {
2535 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2536 	struct ib_device *ibdev = ibpd->device;
2537 	struct mlx5_ib_alloc_pd_resp resp;
2538 	int err;
2539 	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2540 	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {};
2541 	u16 uid = 0;
2542 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2543 		udata, struct mlx5_ib_ucontext, ibucontext);
2544 
2545 	uid = context ? context->devx_uid : 0;
2546 	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2547 	MLX5_SET(alloc_pd_in, in, uid, uid);
2548 	err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out);
2549 	if (err)
2550 		return err;
2551 
2552 	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2553 	pd->uid = uid;
2554 	if (udata) {
2555 		resp.pdn = pd->pdn;
2556 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2557 			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2558 			return -EFAULT;
2559 		}
2560 	}
2561 
2562 	return 0;
2563 }
2564 
2565 static int mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2566 {
2567 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2568 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2569 
2570 	return mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2571 }
2572 
2573 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2574 {
2575 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2576 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
2577 	int err;
2578 	u16 uid;
2579 
2580 	uid = ibqp->pd ?
2581 		to_mpd(ibqp->pd)->uid : 0;
2582 
2583 	if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) {
2584 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
2585 		return -EOPNOTSUPP;
2586 	}
2587 
2588 	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2589 	if (err)
2590 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2591 			     ibqp->qp_num, gid->raw);
2592 
2593 	return err;
2594 }
2595 
2596 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2597 {
2598 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2599 	int err;
2600 	u16 uid;
2601 
2602 	uid = ibqp->pd ?
2603 		to_mpd(ibqp->pd)->uid : 0;
2604 	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2605 	if (err)
2606 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2607 			     ibqp->qp_num, gid->raw);
2608 
2609 	return err;
2610 }
2611 
2612 static int init_node_data(struct mlx5_ib_dev *dev)
2613 {
2614 	int err;
2615 
2616 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2617 	if (err)
2618 		return err;
2619 
2620 	dev->mdev->rev_id = dev->mdev->pdev->revision;
2621 
2622 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2623 }
2624 
2625 static ssize_t fw_pages_show(struct device *device,
2626 			     struct device_attribute *attr, char *buf)
2627 {
2628 	struct mlx5_ib_dev *dev =
2629 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2630 
2631 	return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
2632 }
2633 static DEVICE_ATTR_RO(fw_pages);
2634 
2635 static ssize_t reg_pages_show(struct device *device,
2636 			      struct device_attribute *attr, char *buf)
2637 {
2638 	struct mlx5_ib_dev *dev =
2639 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2640 
2641 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2642 }
2643 static DEVICE_ATTR_RO(reg_pages);
2644 
2645 static ssize_t hca_type_show(struct device *device,
2646 			     struct device_attribute *attr, char *buf)
2647 {
2648 	struct mlx5_ib_dev *dev =
2649 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2650 
2651 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
2652 }
2653 static DEVICE_ATTR_RO(hca_type);
2654 
2655 static ssize_t hw_rev_show(struct device *device,
2656 			   struct device_attribute *attr, char *buf)
2657 {
2658 	struct mlx5_ib_dev *dev =
2659 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2660 
2661 	return sprintf(buf, "%x\n", dev->mdev->rev_id);
2662 }
2663 static DEVICE_ATTR_RO(hw_rev);
2664 
2665 static ssize_t board_id_show(struct device *device,
2666 			     struct device_attribute *attr, char *buf)
2667 {
2668 	struct mlx5_ib_dev *dev =
2669 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2670 
2671 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2672 		       dev->mdev->board_id);
2673 }
2674 static DEVICE_ATTR_RO(board_id);
2675 
2676 static struct attribute *mlx5_class_attributes[] = {
2677 	&dev_attr_hw_rev.attr,
2678 	&dev_attr_hca_type.attr,
2679 	&dev_attr_board_id.attr,
2680 	&dev_attr_fw_pages.attr,
2681 	&dev_attr_reg_pages.attr,
2682 	NULL,
2683 };
2684 
2685 static const struct attribute_group mlx5_attr_group = {
2686 	.attrs = mlx5_class_attributes,
2687 };
2688 
2689 static void pkey_change_handler(struct work_struct *work)
2690 {
2691 	struct mlx5_ib_port_resources *ports =
2692 		container_of(work, struct mlx5_ib_port_resources,
2693 			     pkey_change_work);
2694 
2695 	mlx5_ib_gsi_pkey_change(ports->gsi);
2696 }
2697 
2698 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2699 {
2700 	struct mlx5_ib_qp *mqp;
2701 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2702 	struct mlx5_core_cq *mcq;
2703 	struct list_head cq_armed_list;
2704 	unsigned long flags_qp;
2705 	unsigned long flags_cq;
2706 	unsigned long flags;
2707 
2708 	INIT_LIST_HEAD(&cq_armed_list);
2709 
2710 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2711 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2712 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2713 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2714 		if (mqp->sq.tail != mqp->sq.head) {
2715 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2716 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2717 			if (send_mcq->mcq.comp &&
2718 			    mqp->ibqp.send_cq->comp_handler) {
2719 				if (!send_mcq->mcq.reset_notify_added) {
2720 					send_mcq->mcq.reset_notify_added = 1;
2721 					list_add_tail(&send_mcq->mcq.reset_notify,
2722 						      &cq_armed_list);
2723 				}
2724 			}
2725 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2726 		}
2727 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2728 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2729 		/* no handling is needed for SRQ */
2730 		if (!mqp->ibqp.srq) {
2731 			if (mqp->rq.tail != mqp->rq.head) {
2732 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2733 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2734 				if (recv_mcq->mcq.comp &&
2735 				    mqp->ibqp.recv_cq->comp_handler) {
2736 					if (!recv_mcq->mcq.reset_notify_added) {
2737 						recv_mcq->mcq.reset_notify_added = 1;
2738 						list_add_tail(&recv_mcq->mcq.reset_notify,
2739 							      &cq_armed_list);
2740 					}
2741 				}
2742 				spin_unlock_irqrestore(&recv_mcq->lock,
2743 						       flags_cq);
2744 			}
2745 		}
2746 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2747 	}
2748 	/*At that point all inflight post send were put to be executed as of we
2749 	 * lock/unlock above locks Now need to arm all involved CQs.
2750 	 */
2751 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2752 		mcq->comp(mcq, NULL);
2753 	}
2754 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2755 }
2756 
2757 static void delay_drop_handler(struct work_struct *work)
2758 {
2759 	int err;
2760 	struct mlx5_ib_delay_drop *delay_drop =
2761 		container_of(work, struct mlx5_ib_delay_drop,
2762 			     delay_drop_work);
2763 
2764 	atomic_inc(&delay_drop->events_cnt);
2765 
2766 	mutex_lock(&delay_drop->lock);
2767 	err = mlx5_core_set_delay_drop(delay_drop->dev, delay_drop->timeout);
2768 	if (err) {
2769 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
2770 			     delay_drop->timeout);
2771 		delay_drop->activate = false;
2772 	}
2773 	mutex_unlock(&delay_drop->lock);
2774 }
2775 
2776 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2777 				 struct ib_event *ibev)
2778 {
2779 	u8 port = (eqe->data.port.port >> 4) & 0xf;
2780 
2781 	switch (eqe->sub_type) {
2782 	case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
2783 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2784 					    IB_LINK_LAYER_ETHERNET)
2785 			schedule_work(&ibdev->delay_drop.delay_drop_work);
2786 		break;
2787 	default: /* do nothing */
2788 		return;
2789 	}
2790 }
2791 
2792 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2793 			      struct ib_event *ibev)
2794 {
2795 	u8 port = (eqe->data.port.port >> 4) & 0xf;
2796 
2797 	ibev->element.port_num = port;
2798 
2799 	switch (eqe->sub_type) {
2800 	case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
2801 	case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
2802 	case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
2803 		/* In RoCE, port up/down events are handled in
2804 		 * mlx5_netdev_event().
2805 		 */
2806 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2807 					    IB_LINK_LAYER_ETHERNET)
2808 			return -EINVAL;
2809 
2810 		ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
2811 				IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2812 		break;
2813 
2814 	case MLX5_PORT_CHANGE_SUBTYPE_LID:
2815 		ibev->event = IB_EVENT_LID_CHANGE;
2816 		break;
2817 
2818 	case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
2819 		ibev->event = IB_EVENT_PKEY_CHANGE;
2820 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2821 		break;
2822 
2823 	case MLX5_PORT_CHANGE_SUBTYPE_GUID:
2824 		ibev->event = IB_EVENT_GID_CHANGE;
2825 		break;
2826 
2827 	case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
2828 		ibev->event = IB_EVENT_CLIENT_REREGISTER;
2829 		break;
2830 	default:
2831 		return -EINVAL;
2832 	}
2833 
2834 	return 0;
2835 }
2836 
2837 static void mlx5_ib_handle_event(struct work_struct *_work)
2838 {
2839 	struct mlx5_ib_event_work *work =
2840 		container_of(_work, struct mlx5_ib_event_work, work);
2841 	struct mlx5_ib_dev *ibdev;
2842 	struct ib_event ibev;
2843 	bool fatal = false;
2844 
2845 	if (work->is_slave) {
2846 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
2847 		if (!ibdev)
2848 			goto out;
2849 	} else {
2850 		ibdev = work->dev;
2851 	}
2852 
2853 	switch (work->event) {
2854 	case MLX5_DEV_EVENT_SYS_ERROR:
2855 		ibev.event = IB_EVENT_DEVICE_FATAL;
2856 		mlx5_ib_handle_internal_error(ibdev);
2857 		ibev.element.port_num  = (u8)(unsigned long)work->param;
2858 		fatal = true;
2859 		break;
2860 	case MLX5_EVENT_TYPE_PORT_CHANGE:
2861 		if (handle_port_change(ibdev, work->param, &ibev))
2862 			goto out;
2863 		break;
2864 	case MLX5_EVENT_TYPE_GENERAL_EVENT:
2865 		handle_general_event(ibdev, work->param, &ibev);
2866 		fallthrough;
2867 	default:
2868 		goto out;
2869 	}
2870 
2871 	ibev.device = &ibdev->ib_dev;
2872 
2873 	if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
2874 		mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
2875 		goto out;
2876 	}
2877 
2878 	if (ibdev->ib_active)
2879 		ib_dispatch_event(&ibev);
2880 
2881 	if (fatal)
2882 		ibdev->ib_active = false;
2883 out:
2884 	kfree(work);
2885 }
2886 
2887 static int mlx5_ib_event(struct notifier_block *nb,
2888 			 unsigned long event, void *param)
2889 {
2890 	struct mlx5_ib_event_work *work;
2891 
2892 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2893 	if (!work)
2894 		return NOTIFY_DONE;
2895 
2896 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2897 	work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
2898 	work->is_slave = false;
2899 	work->param = param;
2900 	work->event = event;
2901 
2902 	queue_work(mlx5_ib_event_wq, &work->work);
2903 
2904 	return NOTIFY_OK;
2905 }
2906 
2907 static int mlx5_ib_event_slave_port(struct notifier_block *nb,
2908 				    unsigned long event, void *param)
2909 {
2910 	struct mlx5_ib_event_work *work;
2911 
2912 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2913 	if (!work)
2914 		return NOTIFY_DONE;
2915 
2916 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2917 	work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
2918 	work->is_slave = true;
2919 	work->param = param;
2920 	work->event = event;
2921 	queue_work(mlx5_ib_event_wq, &work->work);
2922 
2923 	return NOTIFY_OK;
2924 }
2925 
2926 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2927 {
2928 	struct mlx5_hca_vport_context vport_ctx;
2929 	int err;
2930 	int port;
2931 
2932 	for (port = 1; port <= ARRAY_SIZE(dev->mdev->port_caps); port++) {
2933 		dev->mdev->port_caps[port - 1].has_smi = false;
2934 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
2935 		    MLX5_CAP_PORT_TYPE_IB) {
2936 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2937 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
2938 								   port, 0,
2939 								   &vport_ctx);
2940 				if (err) {
2941 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2942 						    port, err);
2943 					return err;
2944 				}
2945 				dev->mdev->port_caps[port - 1].has_smi =
2946 					vport_ctx.has_smi;
2947 			} else {
2948 				dev->mdev->port_caps[port - 1].has_smi = true;
2949 			}
2950 		}
2951 	}
2952 	return 0;
2953 }
2954 
2955 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2956 {
2957 	int port;
2958 
2959 	for (port = 1; port <= dev->num_ports; port++)
2960 		mlx5_query_ext_port_caps(dev, port);
2961 }
2962 
2963 static int __get_port_caps(struct mlx5_ib_dev *dev, u8 port)
2964 {
2965 	struct ib_device_attr *dprops = NULL;
2966 	struct ib_port_attr *pprops = NULL;
2967 	int err = -ENOMEM;
2968 
2969 	pprops = kzalloc(sizeof(*pprops), GFP_KERNEL);
2970 	if (!pprops)
2971 		goto out;
2972 
2973 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2974 	if (!dprops)
2975 		goto out;
2976 
2977 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, NULL);
2978 	if (err) {
2979 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
2980 		goto out;
2981 	}
2982 
2983 	err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2984 	if (err) {
2985 		mlx5_ib_warn(dev, "query_port %d failed %d\n",
2986 			     port, err);
2987 		goto out;
2988 	}
2989 
2990 	dev->mdev->port_caps[port - 1].pkey_table_len =
2991 					dprops->max_pkeys;
2992 	dev->mdev->port_caps[port - 1].gid_table_len =
2993 					pprops->gid_tbl_len;
2994 	mlx5_ib_dbg(dev, "port %d: pkey_table_len %d, gid_table_len %d\n",
2995 		    port, dprops->max_pkeys, pprops->gid_tbl_len);
2996 
2997 out:
2998 	kfree(pprops);
2999 	kfree(dprops);
3000 
3001 	return err;
3002 }
3003 
3004 static int get_port_caps(struct mlx5_ib_dev *dev, u8 port)
3005 {
3006 	/* For representors use port 1, is this is the only native
3007 	 * port
3008 	 */
3009 	if (dev->is_rep)
3010 		return __get_port_caps(dev, 1);
3011 	return __get_port_caps(dev, port);
3012 }
3013 
3014 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
3015 {
3016 	switch (umr_fence_cap) {
3017 	case MLX5_CAP_UMR_FENCE_NONE:
3018 		return MLX5_FENCE_MODE_NONE;
3019 	case MLX5_CAP_UMR_FENCE_SMALL:
3020 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
3021 	default:
3022 		return MLX5_FENCE_MODE_STRONG_ORDERING;
3023 	}
3024 }
3025 
3026 static int mlx5_ib_dev_res_init(struct mlx5_ib_dev *dev)
3027 {
3028 	struct mlx5_ib_resources *devr = &dev->devr;
3029 	struct ib_srq_init_attr attr;
3030 	struct ib_device *ibdev;
3031 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
3032 	int port;
3033 	int ret = 0;
3034 
3035 	ibdev = &dev->ib_dev;
3036 
3037 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
3038 		return -EOPNOTSUPP;
3039 
3040 	mutex_init(&devr->mutex);
3041 
3042 	devr->p0 = rdma_zalloc_drv_obj(ibdev, ib_pd);
3043 	if (!devr->p0)
3044 		return -ENOMEM;
3045 
3046 	devr->p0->device  = ibdev;
3047 	devr->p0->uobject = NULL;
3048 	atomic_set(&devr->p0->usecnt, 0);
3049 
3050 	ret = mlx5_ib_alloc_pd(devr->p0, NULL);
3051 	if (ret)
3052 		goto error0;
3053 
3054 	devr->c0 = rdma_zalloc_drv_obj(ibdev, ib_cq);
3055 	if (!devr->c0) {
3056 		ret = -ENOMEM;
3057 		goto error1;
3058 	}
3059 
3060 	devr->c0->device = &dev->ib_dev;
3061 	atomic_set(&devr->c0->usecnt, 0);
3062 
3063 	ret = mlx5_ib_create_cq(devr->c0, &cq_attr, NULL);
3064 	if (ret)
3065 		goto err_create_cq;
3066 
3067 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0);
3068 	if (ret)
3069 		goto error2;
3070 
3071 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0);
3072 	if (ret)
3073 		goto error3;
3074 
3075 	memset(&attr, 0, sizeof(attr));
3076 	attr.attr.max_sge = 1;
3077 	attr.attr.max_wr = 1;
3078 	attr.srq_type = IB_SRQT_XRC;
3079 	attr.ext.cq = devr->c0;
3080 
3081 	devr->s0 = rdma_zalloc_drv_obj(ibdev, ib_srq);
3082 	if (!devr->s0) {
3083 		ret = -ENOMEM;
3084 		goto error4;
3085 	}
3086 
3087 	devr->s0->device	= &dev->ib_dev;
3088 	devr->s0->pd		= devr->p0;
3089 	devr->s0->srq_type      = IB_SRQT_XRC;
3090 	devr->s0->ext.cq	= devr->c0;
3091 	ret = mlx5_ib_create_srq(devr->s0, &attr, NULL);
3092 	if (ret)
3093 		goto err_create;
3094 
3095 	atomic_inc(&devr->s0->ext.cq->usecnt);
3096 	atomic_inc(&devr->p0->usecnt);
3097 	atomic_set(&devr->s0->usecnt, 0);
3098 
3099 	memset(&attr, 0, sizeof(attr));
3100 	attr.attr.max_sge = 1;
3101 	attr.attr.max_wr = 1;
3102 	attr.srq_type = IB_SRQT_BASIC;
3103 	devr->s1 = rdma_zalloc_drv_obj(ibdev, ib_srq);
3104 	if (!devr->s1) {
3105 		ret = -ENOMEM;
3106 		goto error5;
3107 	}
3108 
3109 	devr->s1->device	= &dev->ib_dev;
3110 	devr->s1->pd		= devr->p0;
3111 	devr->s1->srq_type      = IB_SRQT_BASIC;
3112 	devr->s1->ext.cq	= devr->c0;
3113 
3114 	ret = mlx5_ib_create_srq(devr->s1, &attr, NULL);
3115 	if (ret)
3116 		goto error6;
3117 
3118 	atomic_inc(&devr->p0->usecnt);
3119 	atomic_set(&devr->s1->usecnt, 0);
3120 
3121 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
3122 		INIT_WORK(&devr->ports[port].pkey_change_work,
3123 			  pkey_change_handler);
3124 
3125 	return 0;
3126 
3127 error6:
3128 	kfree(devr->s1);
3129 error5:
3130 	mlx5_ib_destroy_srq(devr->s0, NULL);
3131 err_create:
3132 	kfree(devr->s0);
3133 error4:
3134 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
3135 error3:
3136 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
3137 error2:
3138 	mlx5_ib_destroy_cq(devr->c0, NULL);
3139 err_create_cq:
3140 	kfree(devr->c0);
3141 error1:
3142 	mlx5_ib_dealloc_pd(devr->p0, NULL);
3143 error0:
3144 	kfree(devr->p0);
3145 	return ret;
3146 }
3147 
3148 static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev)
3149 {
3150 	struct mlx5_ib_resources *devr = &dev->devr;
3151 	int port;
3152 
3153 	mlx5_ib_destroy_srq(devr->s1, NULL);
3154 	kfree(devr->s1);
3155 	mlx5_ib_destroy_srq(devr->s0, NULL);
3156 	kfree(devr->s0);
3157 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
3158 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
3159 	mlx5_ib_destroy_cq(devr->c0, NULL);
3160 	kfree(devr->c0);
3161 	mlx5_ib_dealloc_pd(devr->p0, NULL);
3162 	kfree(devr->p0);
3163 
3164 	/* Make sure no change P_Key work items are still executing */
3165 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
3166 		cancel_work_sync(&devr->ports[port].pkey_change_work);
3167 }
3168 
3169 static u32 get_core_cap_flags(struct ib_device *ibdev,
3170 			      struct mlx5_hca_vport_context *rep)
3171 {
3172 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3173 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
3174 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
3175 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
3176 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
3177 	u32 ret = 0;
3178 
3179 	if (rep->grh_required)
3180 		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
3181 
3182 	if (ll == IB_LINK_LAYER_INFINIBAND)
3183 		return ret | RDMA_CORE_PORT_IBA_IB;
3184 
3185 	if (raw_support)
3186 		ret |= RDMA_CORE_PORT_RAW_PACKET;
3187 
3188 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
3189 		return ret;
3190 
3191 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
3192 		return ret;
3193 
3194 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
3195 		ret |= RDMA_CORE_PORT_IBA_ROCE;
3196 
3197 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
3198 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3199 
3200 	return ret;
3201 }
3202 
3203 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
3204 			       struct ib_port_immutable *immutable)
3205 {
3206 	struct ib_port_attr attr;
3207 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3208 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
3209 	struct mlx5_hca_vport_context rep = {0};
3210 	int err;
3211 
3212 	err = ib_query_port(ibdev, port_num, &attr);
3213 	if (err)
3214 		return err;
3215 
3216 	if (ll == IB_LINK_LAYER_INFINIBAND) {
3217 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
3218 						   &rep);
3219 		if (err)
3220 			return err;
3221 	}
3222 
3223 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
3224 	immutable->gid_tbl_len = attr.gid_tbl_len;
3225 	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
3226 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3227 
3228 	return 0;
3229 }
3230 
3231 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u8 port_num,
3232 				   struct ib_port_immutable *immutable)
3233 {
3234 	struct ib_port_attr attr;
3235 	int err;
3236 
3237 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
3238 
3239 	err = ib_query_port(ibdev, port_num, &attr);
3240 	if (err)
3241 		return err;
3242 
3243 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
3244 	immutable->gid_tbl_len = attr.gid_tbl_len;
3245 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
3246 
3247 	return 0;
3248 }
3249 
3250 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
3251 {
3252 	struct mlx5_ib_dev *dev =
3253 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
3254 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
3255 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
3256 		 fw_rev_sub(dev->mdev));
3257 }
3258 
3259 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
3260 {
3261 	struct mlx5_core_dev *mdev = dev->mdev;
3262 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
3263 								 MLX5_FLOW_NAMESPACE_LAG);
3264 	struct mlx5_flow_table *ft;
3265 	int err;
3266 
3267 	if (!ns || !mlx5_lag_is_roce(mdev))
3268 		return 0;
3269 
3270 	err = mlx5_cmd_create_vport_lag(mdev);
3271 	if (err)
3272 		return err;
3273 
3274 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3275 	if (IS_ERR(ft)) {
3276 		err = PTR_ERR(ft);
3277 		goto err_destroy_vport_lag;
3278 	}
3279 
3280 	dev->flow_db->lag_demux_ft = ft;
3281 	dev->lag_active = true;
3282 	return 0;
3283 
3284 err_destroy_vport_lag:
3285 	mlx5_cmd_destroy_vport_lag(mdev);
3286 	return err;
3287 }
3288 
3289 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3290 {
3291 	struct mlx5_core_dev *mdev = dev->mdev;
3292 
3293 	if (dev->lag_active) {
3294 		dev->lag_active = false;
3295 
3296 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
3297 		dev->flow_db->lag_demux_ft = NULL;
3298 
3299 		mlx5_cmd_destroy_vport_lag(mdev);
3300 	}
3301 }
3302 
3303 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
3304 {
3305 	int err;
3306 
3307 	dev->port[port_num].roce.nb.notifier_call = mlx5_netdev_event;
3308 	err = register_netdevice_notifier_net(mlx5_core_net(dev->mdev),
3309 					      &dev->port[port_num].roce.nb);
3310 	if (err) {
3311 		dev->port[port_num].roce.nb.notifier_call = NULL;
3312 		return err;
3313 	}
3314 
3315 	return 0;
3316 }
3317 
3318 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
3319 {
3320 	if (dev->port[port_num].roce.nb.notifier_call) {
3321 		unregister_netdevice_notifier_net(mlx5_core_net(dev->mdev),
3322 						  &dev->port[port_num].roce.nb);
3323 		dev->port[port_num].roce.nb.notifier_call = NULL;
3324 	}
3325 }
3326 
3327 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3328 {
3329 	int err;
3330 
3331 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3332 	if (err)
3333 		return err;
3334 
3335 	err = mlx5_eth_lag_init(dev);
3336 	if (err)
3337 		goto err_disable_roce;
3338 
3339 	return 0;
3340 
3341 err_disable_roce:
3342 	mlx5_nic_vport_disable_roce(dev->mdev);
3343 
3344 	return err;
3345 }
3346 
3347 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3348 {
3349 	mlx5_eth_lag_cleanup(dev);
3350 	mlx5_nic_vport_disable_roce(dev->mdev);
3351 }
3352 
3353 static int mlx5_ib_rn_get_params(struct ib_device *device, u8 port_num,
3354 				 enum rdma_netdev_t type,
3355 				 struct rdma_netdev_alloc_params *params)
3356 {
3357 	if (type != RDMA_NETDEV_IPOIB)
3358 		return -EOPNOTSUPP;
3359 
3360 	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
3361 }
3362 
3363 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
3364 				       size_t count, loff_t *pos)
3365 {
3366 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3367 	char lbuf[20];
3368 	int len;
3369 
3370 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
3371 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
3372 }
3373 
3374 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
3375 					size_t count, loff_t *pos)
3376 {
3377 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3378 	u32 timeout;
3379 	u32 var;
3380 
3381 	if (kstrtouint_from_user(buf, count, 0, &var))
3382 		return -EFAULT;
3383 
3384 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
3385 			1000);
3386 	if (timeout != var)
3387 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
3388 			    timeout);
3389 
3390 	delay_drop->timeout = timeout;
3391 
3392 	return count;
3393 }
3394 
3395 static const struct file_operations fops_delay_drop_timeout = {
3396 	.owner	= THIS_MODULE,
3397 	.open	= simple_open,
3398 	.write	= delay_drop_timeout_write,
3399 	.read	= delay_drop_timeout_read,
3400 };
3401 
3402 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
3403 				      struct mlx5_ib_multiport_info *mpi)
3404 {
3405 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3406 	struct mlx5_ib_port *port = &ibdev->port[port_num];
3407 	int comps;
3408 	int err;
3409 	int i;
3410 
3411 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3412 
3413 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
3414 
3415 	spin_lock(&port->mp.mpi_lock);
3416 	if (!mpi->ibdev) {
3417 		spin_unlock(&port->mp.mpi_lock);
3418 		return;
3419 	}
3420 
3421 	mpi->ibdev = NULL;
3422 
3423 	spin_unlock(&port->mp.mpi_lock);
3424 	if (mpi->mdev_events.notifier_call)
3425 		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
3426 	mpi->mdev_events.notifier_call = NULL;
3427 	mlx5_remove_netdev_notifier(ibdev, port_num);
3428 	spin_lock(&port->mp.mpi_lock);
3429 
3430 	comps = mpi->mdev_refcnt;
3431 	if (comps) {
3432 		mpi->unaffiliate = true;
3433 		init_completion(&mpi->unref_comp);
3434 		spin_unlock(&port->mp.mpi_lock);
3435 
3436 		for (i = 0; i < comps; i++)
3437 			wait_for_completion(&mpi->unref_comp);
3438 
3439 		spin_lock(&port->mp.mpi_lock);
3440 		mpi->unaffiliate = false;
3441 	}
3442 
3443 	port->mp.mpi = NULL;
3444 
3445 	list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
3446 
3447 	spin_unlock(&port->mp.mpi_lock);
3448 
3449 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
3450 
3451 	mlx5_ib_dbg(ibdev, "unaffiliated port %d\n", port_num + 1);
3452 	/* Log an error, still needed to cleanup the pointers and add
3453 	 * it back to the list.
3454 	 */
3455 	if (err)
3456 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
3457 			    port_num + 1);
3458 
3459 	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
3460 }
3461 
3462 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
3463 				    struct mlx5_ib_multiport_info *mpi)
3464 {
3465 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3466 	int err;
3467 
3468 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3469 
3470 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
3471 	if (ibdev->port[port_num].mp.mpi) {
3472 		mlx5_ib_dbg(ibdev, "port %d already affiliated.\n",
3473 			    port_num + 1);
3474 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3475 		return false;
3476 	}
3477 
3478 	ibdev->port[port_num].mp.mpi = mpi;
3479 	mpi->ibdev = ibdev;
3480 	mpi->mdev_events.notifier_call = NULL;
3481 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3482 
3483 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
3484 	if (err)
3485 		goto unbind;
3486 
3487 	err = get_port_caps(ibdev, mlx5_core_native_port_num(mpi->mdev));
3488 	if (err)
3489 		goto unbind;
3490 
3491 	err = mlx5_add_netdev_notifier(ibdev, port_num);
3492 	if (err) {
3493 		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
3494 			    port_num + 1);
3495 		goto unbind;
3496 	}
3497 
3498 	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
3499 	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
3500 
3501 	mlx5_ib_init_cong_debugfs(ibdev, port_num);
3502 
3503 	return true;
3504 
3505 unbind:
3506 	mlx5_ib_unbind_slave_port(ibdev, mpi);
3507 	return false;
3508 }
3509 
3510 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
3511 {
3512 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3513 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3514 							  port_num + 1);
3515 	struct mlx5_ib_multiport_info *mpi;
3516 	int err;
3517 	int i;
3518 
3519 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3520 		return 0;
3521 
3522 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
3523 						     &dev->sys_image_guid);
3524 	if (err)
3525 		return err;
3526 
3527 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3528 	if (err)
3529 		return err;
3530 
3531 	mutex_lock(&mlx5_ib_multiport_mutex);
3532 	for (i = 0; i < dev->num_ports; i++) {
3533 		bool bound = false;
3534 
3535 		/* build a stub multiport info struct for the native port. */
3536 		if (i == port_num) {
3537 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
3538 			if (!mpi) {
3539 				mutex_unlock(&mlx5_ib_multiport_mutex);
3540 				mlx5_nic_vport_disable_roce(dev->mdev);
3541 				return -ENOMEM;
3542 			}
3543 
3544 			mpi->is_master = true;
3545 			mpi->mdev = dev->mdev;
3546 			mpi->sys_image_guid = dev->sys_image_guid;
3547 			dev->port[i].mp.mpi = mpi;
3548 			mpi->ibdev = dev;
3549 			mpi = NULL;
3550 			continue;
3551 		}
3552 
3553 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
3554 				    list) {
3555 			if (dev->sys_image_guid == mpi->sys_image_guid &&
3556 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
3557 				bound = mlx5_ib_bind_slave_port(dev, mpi);
3558 			}
3559 
3560 			if (bound) {
3561 				dev_dbg(mpi->mdev->device,
3562 					"removing port from unaffiliated list.\n");
3563 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
3564 				list_del(&mpi->list);
3565 				break;
3566 			}
3567 		}
3568 		if (!bound) {
3569 			get_port_caps(dev, i + 1);
3570 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
3571 				    i + 1);
3572 		}
3573 	}
3574 
3575 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
3576 	mutex_unlock(&mlx5_ib_multiport_mutex);
3577 	return err;
3578 }
3579 
3580 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
3581 {
3582 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3583 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3584 							  port_num + 1);
3585 	int i;
3586 
3587 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3588 		return;
3589 
3590 	mutex_lock(&mlx5_ib_multiport_mutex);
3591 	for (i = 0; i < dev->num_ports; i++) {
3592 		if (dev->port[i].mp.mpi) {
3593 			/* Destroy the native port stub */
3594 			if (i == port_num) {
3595 				kfree(dev->port[i].mp.mpi);
3596 				dev->port[i].mp.mpi = NULL;
3597 			} else {
3598 				mlx5_ib_dbg(dev, "unbinding port_num: %d\n", i + 1);
3599 				mlx5_ib_unbind_slave_port(dev, dev->port[i].mp.mpi);
3600 			}
3601 		}
3602 	}
3603 
3604 	mlx5_ib_dbg(dev, "removing from devlist\n");
3605 	list_del(&dev->ib_dev_list);
3606 	mutex_unlock(&mlx5_ib_multiport_mutex);
3607 
3608 	mlx5_nic_vport_disable_roce(dev->mdev);
3609 }
3610 
3611 static int mmap_obj_cleanup(struct ib_uobject *uobject,
3612 			    enum rdma_remove_reason why,
3613 			    struct uverbs_attr_bundle *attrs)
3614 {
3615 	struct mlx5_user_mmap_entry *obj = uobject->object;
3616 
3617 	rdma_user_mmap_entry_remove(&obj->rdma_entry);
3618 	return 0;
3619 }
3620 
3621 static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c,
3622 					    struct mlx5_user_mmap_entry *entry,
3623 					    size_t length)
3624 {
3625 	return rdma_user_mmap_entry_insert_range(
3626 		&c->ibucontext, &entry->rdma_entry, length,
3627 		(MLX5_IB_MMAP_OFFSET_START << 16),
3628 		((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1));
3629 }
3630 
3631 static struct mlx5_user_mmap_entry *
3632 alloc_var_entry(struct mlx5_ib_ucontext *c)
3633 {
3634 	struct mlx5_user_mmap_entry *entry;
3635 	struct mlx5_var_table *var_table;
3636 	u32 page_idx;
3637 	int err;
3638 
3639 	var_table = &to_mdev(c->ibucontext.device)->var_table;
3640 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3641 	if (!entry)
3642 		return ERR_PTR(-ENOMEM);
3643 
3644 	mutex_lock(&var_table->bitmap_lock);
3645 	page_idx = find_first_zero_bit(var_table->bitmap,
3646 				       var_table->num_var_hw_entries);
3647 	if (page_idx >= var_table->num_var_hw_entries) {
3648 		err = -ENOSPC;
3649 		mutex_unlock(&var_table->bitmap_lock);
3650 		goto end;
3651 	}
3652 
3653 	set_bit(page_idx, var_table->bitmap);
3654 	mutex_unlock(&var_table->bitmap_lock);
3655 
3656 	entry->address = var_table->hw_start_addr +
3657 				(page_idx * var_table->stride_size);
3658 	entry->page_idx = page_idx;
3659 	entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
3660 
3661 	err = mlx5_rdma_user_mmap_entry_insert(c, entry,
3662 					       var_table->stride_size);
3663 	if (err)
3664 		goto err_insert;
3665 
3666 	return entry;
3667 
3668 err_insert:
3669 	mutex_lock(&var_table->bitmap_lock);
3670 	clear_bit(page_idx, var_table->bitmap);
3671 	mutex_unlock(&var_table->bitmap_lock);
3672 end:
3673 	kfree(entry);
3674 	return ERR_PTR(err);
3675 }
3676 
3677 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
3678 	struct uverbs_attr_bundle *attrs)
3679 {
3680 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3681 		attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3682 	struct mlx5_ib_ucontext *c;
3683 	struct mlx5_user_mmap_entry *entry;
3684 	u64 mmap_offset;
3685 	u32 length;
3686 	int err;
3687 
3688 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3689 	if (IS_ERR(c))
3690 		return PTR_ERR(c);
3691 
3692 	entry = alloc_var_entry(c);
3693 	if (IS_ERR(entry))
3694 		return PTR_ERR(entry);
3695 
3696 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3697 	length = entry->rdma_entry.npages * PAGE_SIZE;
3698 	uobj->object = entry;
3699 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3700 
3701 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3702 			     &mmap_offset, sizeof(mmap_offset));
3703 	if (err)
3704 		return err;
3705 
3706 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3707 			     &entry->page_idx, sizeof(entry->page_idx));
3708 	if (err)
3709 		return err;
3710 
3711 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3712 			     &length, sizeof(length));
3713 	return err;
3714 }
3715 
3716 DECLARE_UVERBS_NAMED_METHOD(
3717 	MLX5_IB_METHOD_VAR_OBJ_ALLOC,
3718 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
3719 			MLX5_IB_OBJECT_VAR,
3720 			UVERBS_ACCESS_NEW,
3721 			UA_MANDATORY),
3722 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3723 			   UVERBS_ATTR_TYPE(u32),
3724 			   UA_MANDATORY),
3725 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3726 			   UVERBS_ATTR_TYPE(u32),
3727 			   UA_MANDATORY),
3728 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3729 			    UVERBS_ATTR_TYPE(u64),
3730 			    UA_MANDATORY));
3731 
3732 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3733 	MLX5_IB_METHOD_VAR_OBJ_DESTROY,
3734 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
3735 			MLX5_IB_OBJECT_VAR,
3736 			UVERBS_ACCESS_DESTROY,
3737 			UA_MANDATORY));
3738 
3739 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
3740 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3741 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
3742 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
3743 
3744 static bool var_is_supported(struct ib_device *device)
3745 {
3746 	struct mlx5_ib_dev *dev = to_mdev(device);
3747 
3748 	return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3749 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
3750 }
3751 
3752 static struct mlx5_user_mmap_entry *
3753 alloc_uar_entry(struct mlx5_ib_ucontext *c,
3754 		enum mlx5_ib_uapi_uar_alloc_type alloc_type)
3755 {
3756 	struct mlx5_user_mmap_entry *entry;
3757 	struct mlx5_ib_dev *dev;
3758 	u32 uar_index;
3759 	int err;
3760 
3761 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3762 	if (!entry)
3763 		return ERR_PTR(-ENOMEM);
3764 
3765 	dev = to_mdev(c->ibucontext.device);
3766 	err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
3767 	if (err)
3768 		goto end;
3769 
3770 	entry->page_idx = uar_index;
3771 	entry->address = uar_index2paddress(dev, uar_index);
3772 	if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3773 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC;
3774 	else
3775 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC;
3776 
3777 	err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE);
3778 	if (err)
3779 		goto err_insert;
3780 
3781 	return entry;
3782 
3783 err_insert:
3784 	mlx5_cmd_free_uar(dev->mdev, uar_index);
3785 end:
3786 	kfree(entry);
3787 	return ERR_PTR(err);
3788 }
3789 
3790 static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)(
3791 	struct uverbs_attr_bundle *attrs)
3792 {
3793 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3794 		attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3795 	enum mlx5_ib_uapi_uar_alloc_type alloc_type;
3796 	struct mlx5_ib_ucontext *c;
3797 	struct mlx5_user_mmap_entry *entry;
3798 	u64 mmap_offset;
3799 	u32 length;
3800 	int err;
3801 
3802 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3803 	if (IS_ERR(c))
3804 		return PTR_ERR(c);
3805 
3806 	err = uverbs_get_const(&alloc_type, attrs,
3807 			       MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE);
3808 	if (err)
3809 		return err;
3810 
3811 	if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF &&
3812 	    alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC)
3813 		return -EOPNOTSUPP;
3814 
3815 	if (!to_mdev(c->ibucontext.device)->wc_support &&
3816 	    alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3817 		return -EOPNOTSUPP;
3818 
3819 	entry = alloc_uar_entry(c, alloc_type);
3820 	if (IS_ERR(entry))
3821 		return PTR_ERR(entry);
3822 
3823 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3824 	length = entry->rdma_entry.npages * PAGE_SIZE;
3825 	uobj->object = entry;
3826 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3827 
3828 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3829 			     &mmap_offset, sizeof(mmap_offset));
3830 	if (err)
3831 		return err;
3832 
3833 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3834 			     &entry->page_idx, sizeof(entry->page_idx));
3835 	if (err)
3836 		return err;
3837 
3838 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3839 			     &length, sizeof(length));
3840 	return err;
3841 }
3842 
3843 DECLARE_UVERBS_NAMED_METHOD(
3844 	MLX5_IB_METHOD_UAR_OBJ_ALLOC,
3845 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE,
3846 			MLX5_IB_OBJECT_UAR,
3847 			UVERBS_ACCESS_NEW,
3848 			UA_MANDATORY),
3849 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE,
3850 			     enum mlx5_ib_uapi_uar_alloc_type,
3851 			     UA_MANDATORY),
3852 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3853 			   UVERBS_ATTR_TYPE(u32),
3854 			   UA_MANDATORY),
3855 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3856 			   UVERBS_ATTR_TYPE(u32),
3857 			   UA_MANDATORY),
3858 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3859 			    UVERBS_ATTR_TYPE(u64),
3860 			    UA_MANDATORY));
3861 
3862 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3863 	MLX5_IB_METHOD_UAR_OBJ_DESTROY,
3864 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE,
3865 			MLX5_IB_OBJECT_UAR,
3866 			UVERBS_ACCESS_DESTROY,
3867 			UA_MANDATORY));
3868 
3869 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR,
3870 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3871 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC),
3872 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY));
3873 
3874 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3875 	mlx5_ib_dm,
3876 	UVERBS_OBJECT_DM,
3877 	UVERBS_METHOD_DM_ALLOC,
3878 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
3879 			    UVERBS_ATTR_TYPE(u64),
3880 			    UA_MANDATORY),
3881 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
3882 			    UVERBS_ATTR_TYPE(u16),
3883 			    UA_OPTIONAL),
3884 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_ALLOC_DM_REQ_TYPE,
3885 			     enum mlx5_ib_uapi_dm_type,
3886 			     UA_OPTIONAL));
3887 
3888 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3889 	mlx5_ib_flow_action,
3890 	UVERBS_OBJECT_FLOW_ACTION,
3891 	UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
3892 	UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
3893 			     enum mlx5_ib_uapi_flow_action_flags));
3894 
3895 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3896 	mlx5_ib_query_context,
3897 	UVERBS_OBJECT_DEVICE,
3898 	UVERBS_METHOD_QUERY_CONTEXT,
3899 	UVERBS_ATTR_PTR_OUT(
3900 		MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
3901 		UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp,
3902 				   dump_fill_mkey),
3903 		UA_MANDATORY));
3904 
3905 static const struct uapi_definition mlx5_ib_defs[] = {
3906 	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
3907 	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
3908 	UAPI_DEF_CHAIN(mlx5_ib_qos_defs),
3909 	UAPI_DEF_CHAIN(mlx5_ib_std_types_defs),
3910 
3911 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_FLOW_ACTION,
3912 				&mlx5_ib_flow_action),
3913 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DM, &mlx5_ib_dm),
3914 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context),
3915 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
3916 				UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
3917 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR),
3918 	{}
3919 };
3920 
3921 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
3922 {
3923 	mlx5_ib_cleanup_multiport_master(dev);
3924 	WARN_ON(!xa_empty(&dev->odp_mkeys));
3925 	cleanup_srcu_struct(&dev->odp_srcu);
3926 
3927 	WARN_ON(!xa_empty(&dev->sig_mrs));
3928 	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
3929 }
3930 
3931 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
3932 {
3933 	struct mlx5_core_dev *mdev = dev->mdev;
3934 	int err;
3935 	int i;
3936 
3937 	for (i = 0; i < dev->num_ports; i++) {
3938 		spin_lock_init(&dev->port[i].mp.mpi_lock);
3939 		rwlock_init(&dev->port[i].roce.netdev_lock);
3940 		dev->port[i].roce.dev = dev;
3941 		dev->port[i].roce.native_port_num = i + 1;
3942 		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
3943 	}
3944 
3945 	mlx5_ib_internal_fill_odp_caps(dev);
3946 
3947 	err = mlx5_ib_init_multiport_master(dev);
3948 	if (err)
3949 		return err;
3950 
3951 	err = set_has_smi_cap(dev);
3952 	if (err)
3953 		return err;
3954 
3955 	if (!mlx5_core_mp_enabled(mdev)) {
3956 		for (i = 1; i <= dev->num_ports; i++) {
3957 			err = get_port_caps(dev, i);
3958 			if (err)
3959 				break;
3960 		}
3961 	} else {
3962 		err = get_port_caps(dev, mlx5_core_native_port_num(mdev));
3963 	}
3964 	if (err)
3965 		goto err_mp;
3966 
3967 	if (mlx5_use_mad_ifc(dev))
3968 		get_ext_port_caps(dev);
3969 
3970 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
3971 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
3972 	dev->ib_dev.phys_port_cnt	= dev->num_ports;
3973 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_count(mdev);
3974 	dev->ib_dev.dev.parent		= mdev->device;
3975 	dev->ib_dev.lag_flags		= RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
3976 
3977 	mutex_init(&dev->cap_mask_mutex);
3978 	INIT_LIST_HEAD(&dev->qp_list);
3979 	spin_lock_init(&dev->reset_flow_resource_lock);
3980 	xa_init(&dev->odp_mkeys);
3981 	xa_init(&dev->sig_mrs);
3982 	atomic_set(&dev->mkey_var, 0);
3983 
3984 	spin_lock_init(&dev->dm.lock);
3985 	dev->dm.dev = mdev;
3986 
3987 	err = init_srcu_struct(&dev->odp_srcu);
3988 	if (err)
3989 		goto err_mp;
3990 
3991 	return 0;
3992 
3993 err_mp:
3994 	mlx5_ib_cleanup_multiport_master(dev);
3995 
3996 	return -ENOMEM;
3997 }
3998 
3999 static int mlx5_ib_enable_driver(struct ib_device *dev)
4000 {
4001 	struct mlx5_ib_dev *mdev = to_mdev(dev);
4002 	int ret;
4003 
4004 	ret = mlx5_ib_test_wc(mdev);
4005 	mlx5_ib_dbg(mdev, "Write-Combining %s",
4006 		    mdev->wc_support ? "supported" : "not supported");
4007 
4008 	return ret;
4009 }
4010 
4011 static const struct ib_device_ops mlx5_ib_dev_ops = {
4012 	.owner = THIS_MODULE,
4013 	.driver_id = RDMA_DRIVER_MLX5,
4014 	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
4015 
4016 	.add_gid = mlx5_ib_add_gid,
4017 	.alloc_mr = mlx5_ib_alloc_mr,
4018 	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
4019 	.alloc_pd = mlx5_ib_alloc_pd,
4020 	.alloc_ucontext = mlx5_ib_alloc_ucontext,
4021 	.attach_mcast = mlx5_ib_mcg_attach,
4022 	.check_mr_status = mlx5_ib_check_mr_status,
4023 	.create_ah = mlx5_ib_create_ah,
4024 	.create_cq = mlx5_ib_create_cq,
4025 	.create_qp = mlx5_ib_create_qp,
4026 	.create_srq = mlx5_ib_create_srq,
4027 	.dealloc_pd = mlx5_ib_dealloc_pd,
4028 	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
4029 	.del_gid = mlx5_ib_del_gid,
4030 	.dereg_mr = mlx5_ib_dereg_mr,
4031 	.destroy_ah = mlx5_ib_destroy_ah,
4032 	.destroy_cq = mlx5_ib_destroy_cq,
4033 	.destroy_qp = mlx5_ib_destroy_qp,
4034 	.destroy_srq = mlx5_ib_destroy_srq,
4035 	.detach_mcast = mlx5_ib_mcg_detach,
4036 	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
4037 	.drain_rq = mlx5_ib_drain_rq,
4038 	.drain_sq = mlx5_ib_drain_sq,
4039 	.enable_driver = mlx5_ib_enable_driver,
4040 	.get_dev_fw_str = get_dev_fw_str,
4041 	.get_dma_mr = mlx5_ib_get_dma_mr,
4042 	.get_link_layer = mlx5_ib_port_link_layer,
4043 	.map_mr_sg = mlx5_ib_map_mr_sg,
4044 	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
4045 	.mmap = mlx5_ib_mmap,
4046 	.mmap_free = mlx5_ib_mmap_free,
4047 	.modify_cq = mlx5_ib_modify_cq,
4048 	.modify_device = mlx5_ib_modify_device,
4049 	.modify_port = mlx5_ib_modify_port,
4050 	.modify_qp = mlx5_ib_modify_qp,
4051 	.modify_srq = mlx5_ib_modify_srq,
4052 	.poll_cq = mlx5_ib_poll_cq,
4053 	.post_recv = mlx5_ib_post_recv_nodrain,
4054 	.post_send = mlx5_ib_post_send_nodrain,
4055 	.post_srq_recv = mlx5_ib_post_srq_recv,
4056 	.process_mad = mlx5_ib_process_mad,
4057 	.query_ah = mlx5_ib_query_ah,
4058 	.query_device = mlx5_ib_query_device,
4059 	.query_gid = mlx5_ib_query_gid,
4060 	.query_pkey = mlx5_ib_query_pkey,
4061 	.query_qp = mlx5_ib_query_qp,
4062 	.query_srq = mlx5_ib_query_srq,
4063 	.query_ucontext = mlx5_ib_query_ucontext,
4064 	.reg_user_mr = mlx5_ib_reg_user_mr,
4065 	.req_notify_cq = mlx5_ib_arm_cq,
4066 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
4067 	.resize_cq = mlx5_ib_resize_cq,
4068 
4069 	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
4070 	INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs),
4071 	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
4072 	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
4073 	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
4074 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
4075 };
4076 
4077 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
4078 	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
4079 };
4080 
4081 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
4082 	.get_vf_config = mlx5_ib_get_vf_config,
4083 	.get_vf_guid = mlx5_ib_get_vf_guid,
4084 	.get_vf_stats = mlx5_ib_get_vf_stats,
4085 	.set_vf_guid = mlx5_ib_set_vf_guid,
4086 	.set_vf_link_state = mlx5_ib_set_vf_link_state,
4087 };
4088 
4089 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
4090 	.alloc_mw = mlx5_ib_alloc_mw,
4091 	.dealloc_mw = mlx5_ib_dealloc_mw,
4092 
4093 	INIT_RDMA_OBJ_SIZE(ib_mw, mlx5_ib_mw, ibmw),
4094 };
4095 
4096 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
4097 	.alloc_xrcd = mlx5_ib_alloc_xrcd,
4098 	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
4099 
4100 	INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd),
4101 };
4102 
4103 static const struct ib_device_ops mlx5_ib_dev_dm_ops = {
4104 	.alloc_dm = mlx5_ib_alloc_dm,
4105 	.dealloc_dm = mlx5_ib_dealloc_dm,
4106 	.reg_dm_mr = mlx5_ib_reg_dm_mr,
4107 };
4108 
4109 static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
4110 {
4111 	struct mlx5_core_dev *mdev = dev->mdev;
4112 	struct mlx5_var_table *var_table = &dev->var_table;
4113 	u8 log_doorbell_bar_size;
4114 	u8 log_doorbell_stride;
4115 	u64 bar_size;
4116 
4117 	log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
4118 					log_doorbell_bar_size);
4119 	log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
4120 					log_doorbell_stride);
4121 	var_table->hw_start_addr = dev->mdev->bar_addr +
4122 				MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
4123 					doorbell_bar_offset);
4124 	bar_size = (1ULL << log_doorbell_bar_size) * 4096;
4125 	var_table->stride_size = 1ULL << log_doorbell_stride;
4126 	var_table->num_var_hw_entries = div_u64(bar_size,
4127 						var_table->stride_size);
4128 	mutex_init(&var_table->bitmap_lock);
4129 	var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
4130 					  GFP_KERNEL);
4131 	return (var_table->bitmap) ? 0 : -ENOMEM;
4132 }
4133 
4134 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
4135 {
4136 	bitmap_free(dev->var_table.bitmap);
4137 }
4138 
4139 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
4140 {
4141 	struct mlx5_core_dev *mdev = dev->mdev;
4142 	int err;
4143 
4144 	dev->ib_dev.uverbs_cmd_mask	=
4145 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
4146 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
4147 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
4148 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
4149 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
4150 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
4151 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
4152 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
4153 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
4154 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
4155 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
4156 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
4157 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
4158 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
4159 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
4160 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
4161 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
4162 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
4163 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
4164 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
4165 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
4166 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
4167 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
4168 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
4169 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
4170 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
4171 	dev->ib_dev.uverbs_ex_cmd_mask =
4172 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
4173 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
4174 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP)	|
4175 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP)	|
4176 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_CQ)	|
4177 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW)	|
4178 		(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW);
4179 
4180 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
4181 	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
4182 		ib_set_device_ops(&dev->ib_dev,
4183 				  &mlx5_ib_dev_ipoib_enhanced_ops);
4184 
4185 	if (mlx5_core_is_pf(mdev))
4186 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
4187 
4188 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
4189 
4190 	if (MLX5_CAP_GEN(mdev, imaicl)) {
4191 		dev->ib_dev.uverbs_cmd_mask |=
4192 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
4193 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
4194 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
4195 	}
4196 
4197 	if (MLX5_CAP_GEN(mdev, xrc)) {
4198 		dev->ib_dev.uverbs_cmd_mask |=
4199 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
4200 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
4201 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
4202 	}
4203 
4204 	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
4205 	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
4206 	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
4207 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
4208 
4209 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
4210 
4211 	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
4212 		dev->ib_dev.driver_def = mlx5_ib_defs;
4213 
4214 	err = init_node_data(dev);
4215 	if (err)
4216 		return err;
4217 
4218 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
4219 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
4220 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
4221 		mutex_init(&dev->lb.mutex);
4222 
4223 	if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
4224 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
4225 		err = mlx5_ib_init_var_table(dev);
4226 		if (err)
4227 			return err;
4228 	}
4229 
4230 	dev->ib_dev.use_cq_dim = true;
4231 
4232 	return 0;
4233 }
4234 
4235 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
4236 	.get_port_immutable = mlx5_port_immutable,
4237 	.query_port = mlx5_ib_query_port,
4238 };
4239 
4240 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
4241 {
4242 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
4243 	return 0;
4244 }
4245 
4246 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
4247 	.get_port_immutable = mlx5_port_rep_immutable,
4248 	.query_port = mlx5_ib_rep_query_port,
4249 };
4250 
4251 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
4252 {
4253 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
4254 	return 0;
4255 }
4256 
4257 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
4258 	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
4259 	.create_wq = mlx5_ib_create_wq,
4260 	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
4261 	.destroy_wq = mlx5_ib_destroy_wq,
4262 	.get_netdev = mlx5_ib_get_netdev,
4263 	.modify_wq = mlx5_ib_modify_wq,
4264 
4265 	INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table,
4266 			   ib_rwq_ind_tbl),
4267 };
4268 
4269 static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev)
4270 {
4271 	struct mlx5_core_dev *mdev = dev->mdev;
4272 	enum rdma_link_layer ll;
4273 	int port_type_cap;
4274 	u8 port_num = 0;
4275 	int err;
4276 
4277 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4278 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4279 
4280 	if (ll == IB_LINK_LAYER_ETHERNET) {
4281 		dev->ib_dev.uverbs_ex_cmd_mask |=
4282 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
4283 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
4284 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
4285 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
4286 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
4287 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
4288 
4289 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4290 
4291 		/* Register only for native ports */
4292 		err = mlx5_add_netdev_notifier(dev, port_num);
4293 		if (err || dev->is_rep || !mlx5_is_roce_enabled(mdev))
4294 			/*
4295 			 * We don't enable ETH interface for
4296 			 * 1. IB representors
4297 			 * 2. User disabled ROCE through devlink interface
4298 			 */
4299 			return err;
4300 
4301 		err = mlx5_enable_eth(dev);
4302 		if (err)
4303 			goto cleanup;
4304 	}
4305 
4306 	return 0;
4307 cleanup:
4308 	mlx5_remove_netdev_notifier(dev, port_num);
4309 	return err;
4310 }
4311 
4312 static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev)
4313 {
4314 	struct mlx5_core_dev *mdev = dev->mdev;
4315 	enum rdma_link_layer ll;
4316 	int port_type_cap;
4317 	u8 port_num;
4318 
4319 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4320 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4321 
4322 	if (ll == IB_LINK_LAYER_ETHERNET) {
4323 		if (!dev->is_rep)
4324 			mlx5_disable_eth(dev);
4325 
4326 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4327 		mlx5_remove_netdev_notifier(dev, port_num);
4328 	}
4329 }
4330 
4331 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
4332 {
4333 	mlx5_ib_init_cong_debugfs(dev,
4334 				  mlx5_core_native_port_num(dev->mdev) - 1);
4335 	return 0;
4336 }
4337 
4338 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
4339 {
4340 	mlx5_ib_cleanup_cong_debugfs(dev,
4341 				     mlx5_core_native_port_num(dev->mdev) - 1);
4342 }
4343 
4344 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
4345 {
4346 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
4347 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
4348 }
4349 
4350 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
4351 {
4352 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
4353 }
4354 
4355 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
4356 {
4357 	int err;
4358 
4359 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
4360 	if (err)
4361 		return err;
4362 
4363 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
4364 	if (err)
4365 		mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4366 
4367 	return err;
4368 }
4369 
4370 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
4371 {
4372 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4373 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4374 }
4375 
4376 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
4377 {
4378 	const char *name;
4379 
4380 	rdma_set_device_sysfs_group(&dev->ib_dev, &mlx5_attr_group);
4381 	if (!mlx5_lag_is_roce(dev->mdev))
4382 		name = "mlx5_%d";
4383 	else
4384 		name = "mlx5_bond_%d";
4385 	return ib_register_device(&dev->ib_dev, name, &dev->mdev->pdev->dev);
4386 }
4387 
4388 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
4389 {
4390 	int err;
4391 
4392 	err = mlx5_mr_cache_cleanup(dev);
4393 	if (err)
4394 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
4395 
4396 	if (dev->umrc.qp)
4397 		mlx5_ib_destroy_qp(dev->umrc.qp, NULL);
4398 	if (dev->umrc.cq)
4399 		ib_free_cq(dev->umrc.cq);
4400 	if (dev->umrc.pd)
4401 		ib_dealloc_pd(dev->umrc.pd);
4402 }
4403 
4404 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
4405 {
4406 	ib_unregister_device(&dev->ib_dev);
4407 }
4408 
4409 enum {
4410 	MAX_UMR_WR = 128,
4411 };
4412 
4413 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
4414 {
4415 	struct ib_qp_init_attr *init_attr = NULL;
4416 	struct ib_qp_attr *attr = NULL;
4417 	struct ib_pd *pd;
4418 	struct ib_cq *cq;
4419 	struct ib_qp *qp;
4420 	int ret;
4421 
4422 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
4423 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
4424 	if (!attr || !init_attr) {
4425 		ret = -ENOMEM;
4426 		goto error_0;
4427 	}
4428 
4429 	pd = ib_alloc_pd(&dev->ib_dev, 0);
4430 	if (IS_ERR(pd)) {
4431 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
4432 		ret = PTR_ERR(pd);
4433 		goto error_0;
4434 	}
4435 
4436 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
4437 	if (IS_ERR(cq)) {
4438 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
4439 		ret = PTR_ERR(cq);
4440 		goto error_2;
4441 	}
4442 
4443 	init_attr->send_cq = cq;
4444 	init_attr->recv_cq = cq;
4445 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
4446 	init_attr->cap.max_send_wr = MAX_UMR_WR;
4447 	init_attr->cap.max_send_sge = 1;
4448 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
4449 	init_attr->port_num = 1;
4450 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
4451 	if (IS_ERR(qp)) {
4452 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
4453 		ret = PTR_ERR(qp);
4454 		goto error_3;
4455 	}
4456 	qp->device     = &dev->ib_dev;
4457 	qp->real_qp    = qp;
4458 	qp->uobject    = NULL;
4459 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
4460 	qp->send_cq    = init_attr->send_cq;
4461 	qp->recv_cq    = init_attr->recv_cq;
4462 
4463 	attr->qp_state = IB_QPS_INIT;
4464 	attr->port_num = 1;
4465 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
4466 				IB_QP_PORT, NULL);
4467 	if (ret) {
4468 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
4469 		goto error_4;
4470 	}
4471 
4472 	memset(attr, 0, sizeof(*attr));
4473 	attr->qp_state = IB_QPS_RTR;
4474 	attr->path_mtu = IB_MTU_256;
4475 
4476 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4477 	if (ret) {
4478 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
4479 		goto error_4;
4480 	}
4481 
4482 	memset(attr, 0, sizeof(*attr));
4483 	attr->qp_state = IB_QPS_RTS;
4484 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4485 	if (ret) {
4486 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
4487 		goto error_4;
4488 	}
4489 
4490 	dev->umrc.qp = qp;
4491 	dev->umrc.cq = cq;
4492 	dev->umrc.pd = pd;
4493 
4494 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
4495 	ret = mlx5_mr_cache_init(dev);
4496 	if (ret) {
4497 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4498 		goto error_4;
4499 	}
4500 
4501 	kfree(attr);
4502 	kfree(init_attr);
4503 
4504 	return 0;
4505 
4506 error_4:
4507 	mlx5_ib_destroy_qp(qp, NULL);
4508 	dev->umrc.qp = NULL;
4509 
4510 error_3:
4511 	ib_free_cq(cq);
4512 	dev->umrc.cq = NULL;
4513 
4514 error_2:
4515 	ib_dealloc_pd(pd);
4516 	dev->umrc.pd = NULL;
4517 
4518 error_0:
4519 	kfree(attr);
4520 	kfree(init_attr);
4521 	return ret;
4522 }
4523 
4524 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
4525 {
4526 	struct dentry *root;
4527 
4528 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4529 		return 0;
4530 
4531 	mutex_init(&dev->delay_drop.lock);
4532 	dev->delay_drop.dev = dev;
4533 	dev->delay_drop.activate = false;
4534 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4535 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4536 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4537 	atomic_set(&dev->delay_drop.events_cnt, 0);
4538 
4539 	if (!mlx5_debugfs_root)
4540 		return 0;
4541 
4542 	root = debugfs_create_dir("delay_drop", dev->mdev->priv.dbg_root);
4543 	dev->delay_drop.dir_debugfs = root;
4544 
4545 	debugfs_create_atomic_t("num_timeout_events", 0400, root,
4546 				&dev->delay_drop.events_cnt);
4547 	debugfs_create_atomic_t("num_rqs", 0400, root,
4548 				&dev->delay_drop.rqs_cnt);
4549 	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
4550 			    &fops_delay_drop_timeout);
4551 	return 0;
4552 }
4553 
4554 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
4555 {
4556 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4557 		return;
4558 
4559 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4560 	if (!dev->delay_drop.dir_debugfs)
4561 		return;
4562 
4563 	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
4564 	dev->delay_drop.dir_debugfs = NULL;
4565 }
4566 
4567 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
4568 {
4569 	dev->mdev_events.notifier_call = mlx5_ib_event;
4570 	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
4571 	return 0;
4572 }
4573 
4574 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
4575 {
4576 	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
4577 }
4578 
4579 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
4580 		      const struct mlx5_ib_profile *profile,
4581 		      int stage)
4582 {
4583 	dev->ib_active = false;
4584 
4585 	/* Number of stages to cleanup */
4586 	while (stage) {
4587 		stage--;
4588 		if (profile->stage[stage].cleanup)
4589 			profile->stage[stage].cleanup(dev);
4590 	}
4591 
4592 	kfree(dev->port);
4593 	ib_dealloc_device(&dev->ib_dev);
4594 }
4595 
4596 void *__mlx5_ib_add(struct mlx5_ib_dev *dev,
4597 		    const struct mlx5_ib_profile *profile)
4598 {
4599 	int err;
4600 	int i;
4601 
4602 	dev->profile = profile;
4603 
4604 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
4605 		if (profile->stage[i].init) {
4606 			err = profile->stage[i].init(dev);
4607 			if (err)
4608 				goto err_out;
4609 		}
4610 	}
4611 
4612 	dev->ib_active = true;
4613 
4614 	return dev;
4615 
4616 err_out:
4617 	__mlx5_ib_remove(dev, profile, i);
4618 
4619 	return NULL;
4620 }
4621 
4622 static const struct mlx5_ib_profile pf_profile = {
4623 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4624 		     mlx5_ib_stage_init_init,
4625 		     mlx5_ib_stage_init_cleanup),
4626 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4627 		     mlx5_ib_fs_init,
4628 		     mlx5_ib_fs_cleanup),
4629 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4630 		     mlx5_ib_stage_caps_init,
4631 		     mlx5_ib_stage_caps_cleanup),
4632 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4633 		     mlx5_ib_stage_non_default_cb,
4634 		     NULL),
4635 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4636 		     mlx5_ib_roce_init,
4637 		     mlx5_ib_roce_cleanup),
4638 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4639 		     mlx5_init_qp_table,
4640 		     mlx5_cleanup_qp_table),
4641 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4642 		     mlx5_init_srq_table,
4643 		     mlx5_cleanup_srq_table),
4644 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4645 		     mlx5_ib_dev_res_init,
4646 		     mlx5_ib_dev_res_cleanup),
4647 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4648 		     mlx5_ib_stage_dev_notifier_init,
4649 		     mlx5_ib_stage_dev_notifier_cleanup),
4650 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
4651 		     mlx5_ib_odp_init_one,
4652 		     mlx5_ib_odp_cleanup_one),
4653 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4654 		     mlx5_ib_counters_init,
4655 		     mlx5_ib_counters_cleanup),
4656 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4657 		     mlx5_ib_stage_cong_debugfs_init,
4658 		     mlx5_ib_stage_cong_debugfs_cleanup),
4659 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4660 		     mlx5_ib_stage_uar_init,
4661 		     mlx5_ib_stage_uar_cleanup),
4662 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4663 		     mlx5_ib_stage_bfrag_init,
4664 		     mlx5_ib_stage_bfrag_cleanup),
4665 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4666 		     NULL,
4667 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4668 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4669 		     mlx5_ib_devx_init,
4670 		     mlx5_ib_devx_cleanup),
4671 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4672 		     mlx5_ib_stage_ib_reg_init,
4673 		     mlx5_ib_stage_ib_reg_cleanup),
4674 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4675 		     mlx5_ib_stage_post_ib_reg_umr_init,
4676 		     NULL),
4677 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4678 		     mlx5_ib_stage_delay_drop_init,
4679 		     mlx5_ib_stage_delay_drop_cleanup),
4680 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4681 		     mlx5_ib_restrack_init,
4682 		     NULL),
4683 };
4684 
4685 const struct mlx5_ib_profile raw_eth_profile = {
4686 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4687 		     mlx5_ib_stage_init_init,
4688 		     mlx5_ib_stage_init_cleanup),
4689 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4690 		     mlx5_ib_fs_init,
4691 		     mlx5_ib_fs_cleanup),
4692 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4693 		     mlx5_ib_stage_caps_init,
4694 		     mlx5_ib_stage_caps_cleanup),
4695 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4696 		     mlx5_ib_stage_raw_eth_non_default_cb,
4697 		     NULL),
4698 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4699 		     mlx5_ib_roce_init,
4700 		     mlx5_ib_roce_cleanup),
4701 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4702 		     mlx5_init_qp_table,
4703 		     mlx5_cleanup_qp_table),
4704 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4705 		     mlx5_init_srq_table,
4706 		     mlx5_cleanup_srq_table),
4707 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4708 		     mlx5_ib_dev_res_init,
4709 		     mlx5_ib_dev_res_cleanup),
4710 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4711 		     mlx5_ib_stage_dev_notifier_init,
4712 		     mlx5_ib_stage_dev_notifier_cleanup),
4713 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4714 		     mlx5_ib_counters_init,
4715 		     mlx5_ib_counters_cleanup),
4716 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4717 		     mlx5_ib_stage_cong_debugfs_init,
4718 		     mlx5_ib_stage_cong_debugfs_cleanup),
4719 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4720 		     mlx5_ib_stage_uar_init,
4721 		     mlx5_ib_stage_uar_cleanup),
4722 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4723 		     mlx5_ib_stage_bfrag_init,
4724 		     mlx5_ib_stage_bfrag_cleanup),
4725 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4726 		     NULL,
4727 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4728 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4729 		     mlx5_ib_devx_init,
4730 		     mlx5_ib_devx_cleanup),
4731 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4732 		     mlx5_ib_stage_ib_reg_init,
4733 		     mlx5_ib_stage_ib_reg_cleanup),
4734 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4735 		     mlx5_ib_stage_post_ib_reg_umr_init,
4736 		     NULL),
4737 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4738 		     mlx5_ib_restrack_init,
4739 		     NULL),
4740 };
4741 
4742 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev)
4743 {
4744 	struct mlx5_ib_multiport_info *mpi;
4745 	struct mlx5_ib_dev *dev;
4746 	bool bound = false;
4747 	int err;
4748 
4749 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4750 	if (!mpi)
4751 		return NULL;
4752 
4753 	mpi->mdev = mdev;
4754 
4755 	err = mlx5_query_nic_vport_system_image_guid(mdev,
4756 						     &mpi->sys_image_guid);
4757 	if (err) {
4758 		kfree(mpi);
4759 		return NULL;
4760 	}
4761 
4762 	mutex_lock(&mlx5_ib_multiport_mutex);
4763 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
4764 		if (dev->sys_image_guid == mpi->sys_image_guid)
4765 			bound = mlx5_ib_bind_slave_port(dev, mpi);
4766 
4767 		if (bound) {
4768 			rdma_roce_rescan_device(&dev->ib_dev);
4769 			break;
4770 		}
4771 	}
4772 
4773 	if (!bound) {
4774 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4775 		dev_dbg(mdev->device,
4776 			"no suitable IB device found to bind to, added to unaffiliated list.\n");
4777 	}
4778 	mutex_unlock(&mlx5_ib_multiport_mutex);
4779 
4780 	return mpi;
4781 }
4782 
4783 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
4784 {
4785 	const struct mlx5_ib_profile *profile;
4786 	enum rdma_link_layer ll;
4787 	struct mlx5_ib_dev *dev;
4788 	int port_type_cap;
4789 	int num_ports;
4790 
4791 	if (MLX5_ESWITCH_MANAGER(mdev) &&
4792 	    mlx5_ib_eswitch_mode(mdev->priv.eswitch) == MLX5_ESWITCH_OFFLOADS) {
4793 		if (!mlx5_core_mp_enabled(mdev))
4794 			mlx5_ib_register_vport_reps(mdev);
4795 		return mdev;
4796 	}
4797 
4798 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4799 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4800 
4801 	if (mlx5_core_is_mp_slave(mdev) && ll == IB_LINK_LAYER_ETHERNET)
4802 		return mlx5_ib_add_slave_port(mdev);
4803 
4804 	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
4805 			MLX5_CAP_GEN(mdev, num_vhca_ports));
4806 	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
4807 	if (!dev)
4808 		return NULL;
4809 	dev->port = kcalloc(num_ports, sizeof(*dev->port),
4810 			     GFP_KERNEL);
4811 	if (!dev->port) {
4812 		ib_dealloc_device(&dev->ib_dev);
4813 		return NULL;
4814 	}
4815 
4816 	dev->mdev = mdev;
4817 	dev->num_ports = num_ports;
4818 
4819 	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_is_roce_enabled(mdev))
4820 		profile = &raw_eth_profile;
4821 	else
4822 		profile = &pf_profile;
4823 
4824 	return __mlx5_ib_add(dev, profile);
4825 }
4826 
4827 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
4828 {
4829 	struct mlx5_ib_multiport_info *mpi;
4830 	struct mlx5_ib_dev *dev;
4831 
4832 	if (MLX5_ESWITCH_MANAGER(mdev) && context == mdev) {
4833 		mlx5_ib_unregister_vport_reps(mdev);
4834 		return;
4835 	}
4836 
4837 	if (mlx5_core_is_mp_slave(mdev)) {
4838 		mpi = context;
4839 		mutex_lock(&mlx5_ib_multiport_mutex);
4840 		if (mpi->ibdev)
4841 			mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
4842 		list_del(&mpi->list);
4843 		mutex_unlock(&mlx5_ib_multiport_mutex);
4844 		kfree(mpi);
4845 		return;
4846 	}
4847 
4848 	dev = context;
4849 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
4850 }
4851 
4852 static struct mlx5_interface mlx5_ib_interface = {
4853 	.add            = mlx5_ib_add,
4854 	.remove         = mlx5_ib_remove,
4855 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
4856 };
4857 
4858 unsigned long mlx5_ib_get_xlt_emergency_page(void)
4859 {
4860 	mutex_lock(&xlt_emergency_page_mutex);
4861 	return xlt_emergency_page;
4862 }
4863 
4864 void mlx5_ib_put_xlt_emergency_page(void)
4865 {
4866 	mutex_unlock(&xlt_emergency_page_mutex);
4867 }
4868 
4869 static int __init mlx5_ib_init(void)
4870 {
4871 	int err;
4872 
4873 	xlt_emergency_page = __get_free_page(GFP_KERNEL);
4874 	if (!xlt_emergency_page)
4875 		return -ENOMEM;
4876 
4877 	mutex_init(&xlt_emergency_page_mutex);
4878 
4879 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
4880 	if (!mlx5_ib_event_wq) {
4881 		free_page(xlt_emergency_page);
4882 		return -ENOMEM;
4883 	}
4884 
4885 	mlx5_ib_odp_init();
4886 
4887 	err = mlx5_register_interface(&mlx5_ib_interface);
4888 
4889 	return err;
4890 }
4891 
4892 static void __exit mlx5_ib_cleanup(void)
4893 {
4894 	mlx5_unregister_interface(&mlx5_ib_interface);
4895 	destroy_workqueue(mlx5_ib_event_wq);
4896 	mutex_destroy(&xlt_emergency_page_mutex);
4897 	free_page(xlt_emergency_page);
4898 }
4899 
4900 module_init(mlx5_ib_init);
4901 module_exit(mlx5_ib_cleanup);
4902