xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision b923cda9)
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2013-2020, Mellanox Technologies inc. All rights reserved.
4  * Copyright (c) 2020, Intel Corporation. All rights reserved.
5  */
6 
7 #include <linux/debugfs.h>
8 #include <linux/highmem.h>
9 #include <linux/module.h>
10 #include <linux/init.h>
11 #include <linux/errno.h>
12 #include <linux/pci.h>
13 #include <linux/dma-mapping.h>
14 #include <linux/slab.h>
15 #include <linux/bitmap.h>
16 #include <linux/sched.h>
17 #include <linux/sched/mm.h>
18 #include <linux/sched/task.h>
19 #include <linux/delay.h>
20 #include <rdma/ib_user_verbs.h>
21 #include <rdma/ib_addr.h>
22 #include <rdma/ib_cache.h>
23 #include <linux/mlx5/port.h>
24 #include <linux/mlx5/vport.h>
25 #include <linux/mlx5/fs.h>
26 #include <linux/mlx5/eswitch.h>
27 #include <linux/list.h>
28 #include <rdma/ib_smi.h>
29 #include <rdma/ib_umem.h>
30 #include <rdma/lag.h>
31 #include <linux/in.h>
32 #include <linux/etherdevice.h>
33 #include "mlx5_ib.h"
34 #include "ib_rep.h"
35 #include "cmd.h"
36 #include "devx.h"
37 #include "dm.h"
38 #include "fs.h"
39 #include "srq.h"
40 #include "qp.h"
41 #include "wr.h"
42 #include "restrack.h"
43 #include "counters.h"
44 #include <linux/mlx5/accel.h>
45 #include <rdma/uverbs_std_types.h>
46 #include <rdma/uverbs_ioctl.h>
47 #include <rdma/mlx5_user_ioctl_verbs.h>
48 #include <rdma/mlx5_user_ioctl_cmds.h>
49 #include <rdma/ib_umem_odp.h>
50 
51 #define UVERBS_MODULE_NAME mlx5_ib
52 #include <rdma/uverbs_named_ioctl.h>
53 
54 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
55 MODULE_DESCRIPTION("Mellanox 5th generation network adapters (ConnectX series) IB driver");
56 MODULE_LICENSE("Dual BSD/GPL");
57 
58 struct mlx5_ib_event_work {
59 	struct work_struct	work;
60 	union {
61 		struct mlx5_ib_dev	      *dev;
62 		struct mlx5_ib_multiport_info *mpi;
63 	};
64 	bool			is_slave;
65 	unsigned int		event;
66 	void			*param;
67 };
68 
69 enum {
70 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
71 };
72 
73 static struct workqueue_struct *mlx5_ib_event_wq;
74 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
75 static LIST_HEAD(mlx5_ib_dev_list);
76 /*
77  * This mutex should be held when accessing either of the above lists
78  */
79 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
80 
81 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
82 {
83 	struct mlx5_ib_dev *dev;
84 
85 	mutex_lock(&mlx5_ib_multiport_mutex);
86 	dev = mpi->ibdev;
87 	mutex_unlock(&mlx5_ib_multiport_mutex);
88 	return dev;
89 }
90 
91 static enum rdma_link_layer
92 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
93 {
94 	switch (port_type_cap) {
95 	case MLX5_CAP_PORT_TYPE_IB:
96 		return IB_LINK_LAYER_INFINIBAND;
97 	case MLX5_CAP_PORT_TYPE_ETH:
98 		return IB_LINK_LAYER_ETHERNET;
99 	default:
100 		return IB_LINK_LAYER_UNSPECIFIED;
101 	}
102 }
103 
104 static enum rdma_link_layer
105 mlx5_ib_port_link_layer(struct ib_device *device, u32 port_num)
106 {
107 	struct mlx5_ib_dev *dev = to_mdev(device);
108 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
109 
110 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
111 }
112 
113 static int get_port_state(struct ib_device *ibdev,
114 			  u32 port_num,
115 			  enum ib_port_state *state)
116 {
117 	struct ib_port_attr attr;
118 	int ret;
119 
120 	memset(&attr, 0, sizeof(attr));
121 	ret = ibdev->ops.query_port(ibdev, port_num, &attr);
122 	if (!ret)
123 		*state = attr.state;
124 	return ret;
125 }
126 
127 static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev,
128 					   struct net_device *ndev,
129 					   u32 *port_num)
130 {
131 	struct net_device *rep_ndev;
132 	struct mlx5_ib_port *port;
133 	int i;
134 
135 	for (i = 0; i < dev->num_ports; i++) {
136 		port  = &dev->port[i];
137 		if (!port->rep)
138 			continue;
139 
140 		read_lock(&port->roce.netdev_lock);
141 		rep_ndev = mlx5_ib_get_rep_netdev(port->rep->esw,
142 						  port->rep->vport);
143 		if (rep_ndev == ndev) {
144 			read_unlock(&port->roce.netdev_lock);
145 			*port_num = i + 1;
146 			return &port->roce;
147 		}
148 		read_unlock(&port->roce.netdev_lock);
149 	}
150 
151 	return NULL;
152 }
153 
154 static int mlx5_netdev_event(struct notifier_block *this,
155 			     unsigned long event, void *ptr)
156 {
157 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
158 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
159 	u32 port_num = roce->native_port_num;
160 	struct mlx5_core_dev *mdev;
161 	struct mlx5_ib_dev *ibdev;
162 
163 	ibdev = roce->dev;
164 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
165 	if (!mdev)
166 		return NOTIFY_DONE;
167 
168 	switch (event) {
169 	case NETDEV_REGISTER:
170 		/* Should already be registered during the load */
171 		if (ibdev->is_rep)
172 			break;
173 		write_lock(&roce->netdev_lock);
174 		if (ndev->dev.parent == mdev->device)
175 			roce->netdev = ndev;
176 		write_unlock(&roce->netdev_lock);
177 		break;
178 
179 	case NETDEV_UNREGISTER:
180 		/* In case of reps, ib device goes away before the netdevs */
181 		write_lock(&roce->netdev_lock);
182 		if (roce->netdev == ndev)
183 			roce->netdev = NULL;
184 		write_unlock(&roce->netdev_lock);
185 		break;
186 
187 	case NETDEV_CHANGE:
188 	case NETDEV_UP:
189 	case NETDEV_DOWN: {
190 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
191 		struct net_device *upper = NULL;
192 
193 		if (lag_ndev) {
194 			upper = netdev_master_upper_dev_get(lag_ndev);
195 			dev_put(lag_ndev);
196 		}
197 
198 		if (ibdev->is_rep)
199 			roce = mlx5_get_rep_roce(ibdev, ndev, &port_num);
200 		if (!roce)
201 			return NOTIFY_DONE;
202 		if ((upper == ndev || (!upper && ndev == roce->netdev))
203 		    && ibdev->ib_active) {
204 			struct ib_event ibev = { };
205 			enum ib_port_state port_state;
206 
207 			if (get_port_state(&ibdev->ib_dev, port_num,
208 					   &port_state))
209 				goto done;
210 
211 			if (roce->last_port_state == port_state)
212 				goto done;
213 
214 			roce->last_port_state = port_state;
215 			ibev.device = &ibdev->ib_dev;
216 			if (port_state == IB_PORT_DOWN)
217 				ibev.event = IB_EVENT_PORT_ERR;
218 			else if (port_state == IB_PORT_ACTIVE)
219 				ibev.event = IB_EVENT_PORT_ACTIVE;
220 			else
221 				goto done;
222 
223 			ibev.element.port_num = port_num;
224 			ib_dispatch_event(&ibev);
225 		}
226 		break;
227 	}
228 
229 	default:
230 		break;
231 	}
232 done:
233 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
234 	return NOTIFY_DONE;
235 }
236 
237 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
238 					     u32 port_num)
239 {
240 	struct mlx5_ib_dev *ibdev = to_mdev(device);
241 	struct net_device *ndev;
242 	struct mlx5_core_dev *mdev;
243 
244 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
245 	if (!mdev)
246 		return NULL;
247 
248 	ndev = mlx5_lag_get_roce_netdev(mdev);
249 	if (ndev)
250 		goto out;
251 
252 	/* Ensure ndev does not disappear before we invoke dev_hold()
253 	 */
254 	read_lock(&ibdev->port[port_num - 1].roce.netdev_lock);
255 	ndev = ibdev->port[port_num - 1].roce.netdev;
256 	if (ndev)
257 		dev_hold(ndev);
258 	read_unlock(&ibdev->port[port_num - 1].roce.netdev_lock);
259 
260 out:
261 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
262 	return ndev;
263 }
264 
265 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
266 						   u32 ib_port_num,
267 						   u32 *native_port_num)
268 {
269 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
270 							  ib_port_num);
271 	struct mlx5_core_dev *mdev = NULL;
272 	struct mlx5_ib_multiport_info *mpi;
273 	struct mlx5_ib_port *port;
274 
275 	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
276 	    ll != IB_LINK_LAYER_ETHERNET) {
277 		if (native_port_num)
278 			*native_port_num = ib_port_num;
279 		return ibdev->mdev;
280 	}
281 
282 	if (native_port_num)
283 		*native_port_num = 1;
284 
285 	port = &ibdev->port[ib_port_num - 1];
286 	spin_lock(&port->mp.mpi_lock);
287 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
288 	if (mpi && !mpi->unaffiliate) {
289 		mdev = mpi->mdev;
290 		/* If it's the master no need to refcount, it'll exist
291 		 * as long as the ib_dev exists.
292 		 */
293 		if (!mpi->is_master)
294 			mpi->mdev_refcnt++;
295 	}
296 	spin_unlock(&port->mp.mpi_lock);
297 
298 	return mdev;
299 }
300 
301 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u32 port_num)
302 {
303 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
304 							  port_num);
305 	struct mlx5_ib_multiport_info *mpi;
306 	struct mlx5_ib_port *port;
307 
308 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
309 		return;
310 
311 	port = &ibdev->port[port_num - 1];
312 
313 	spin_lock(&port->mp.mpi_lock);
314 	mpi = ibdev->port[port_num - 1].mp.mpi;
315 	if (mpi->is_master)
316 		goto out;
317 
318 	mpi->mdev_refcnt--;
319 	if (mpi->unaffiliate)
320 		complete(&mpi->unref_comp);
321 out:
322 	spin_unlock(&port->mp.mpi_lock);
323 }
324 
325 static int translate_eth_legacy_proto_oper(u32 eth_proto_oper,
326 					   u16 *active_speed, u8 *active_width)
327 {
328 	switch (eth_proto_oper) {
329 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
330 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
331 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
332 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
333 		*active_width = IB_WIDTH_1X;
334 		*active_speed = IB_SPEED_SDR;
335 		break;
336 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
337 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
338 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
339 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
340 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
341 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
342 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
343 		*active_width = IB_WIDTH_1X;
344 		*active_speed = IB_SPEED_QDR;
345 		break;
346 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
347 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
348 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
349 		*active_width = IB_WIDTH_1X;
350 		*active_speed = IB_SPEED_EDR;
351 		break;
352 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
353 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
354 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
355 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
356 		*active_width = IB_WIDTH_4X;
357 		*active_speed = IB_SPEED_QDR;
358 		break;
359 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
360 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
361 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
362 		*active_width = IB_WIDTH_1X;
363 		*active_speed = IB_SPEED_HDR;
364 		break;
365 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
366 		*active_width = IB_WIDTH_4X;
367 		*active_speed = IB_SPEED_FDR;
368 		break;
369 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
370 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
371 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
372 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
373 		*active_width = IB_WIDTH_4X;
374 		*active_speed = IB_SPEED_EDR;
375 		break;
376 	default:
377 		return -EINVAL;
378 	}
379 
380 	return 0;
381 }
382 
383 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u16 *active_speed,
384 					u8 *active_width)
385 {
386 	switch (eth_proto_oper) {
387 	case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
388 	case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
389 		*active_width = IB_WIDTH_1X;
390 		*active_speed = IB_SPEED_SDR;
391 		break;
392 	case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
393 		*active_width = IB_WIDTH_1X;
394 		*active_speed = IB_SPEED_DDR;
395 		break;
396 	case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
397 		*active_width = IB_WIDTH_1X;
398 		*active_speed = IB_SPEED_QDR;
399 		break;
400 	case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
401 		*active_width = IB_WIDTH_4X;
402 		*active_speed = IB_SPEED_QDR;
403 		break;
404 	case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
405 		*active_width = IB_WIDTH_1X;
406 		*active_speed = IB_SPEED_EDR;
407 		break;
408 	case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
409 		*active_width = IB_WIDTH_2X;
410 		*active_speed = IB_SPEED_EDR;
411 		break;
412 	case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
413 		*active_width = IB_WIDTH_1X;
414 		*active_speed = IB_SPEED_HDR;
415 		break;
416 	case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
417 		*active_width = IB_WIDTH_4X;
418 		*active_speed = IB_SPEED_EDR;
419 		break;
420 	case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
421 		*active_width = IB_WIDTH_2X;
422 		*active_speed = IB_SPEED_HDR;
423 		break;
424 	case MLX5E_PROT_MASK(MLX5E_100GAUI_1_100GBASE_CR_KR):
425 		*active_width = IB_WIDTH_1X;
426 		*active_speed = IB_SPEED_NDR;
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 	case MLX5E_PROT_MASK(MLX5E_200GAUI_2_200GBASE_CR2_KR2):
433 		*active_width = IB_WIDTH_2X;
434 		*active_speed = IB_SPEED_NDR;
435 		break;
436 	case MLX5E_PROT_MASK(MLX5E_400GAUI_4_400GBASE_CR4_KR4):
437 		*active_width = IB_WIDTH_4X;
438 		*active_speed = IB_SPEED_NDR;
439 		break;
440 	default:
441 		return -EINVAL;
442 	}
443 
444 	return 0;
445 }
446 
447 static int translate_eth_proto_oper(u32 eth_proto_oper, u16 *active_speed,
448 				    u8 *active_width, bool ext)
449 {
450 	return ext ?
451 		translate_eth_ext_proto_oper(eth_proto_oper, active_speed,
452 					     active_width) :
453 		translate_eth_legacy_proto_oper(eth_proto_oper, active_speed,
454 						active_width);
455 }
456 
457 static int mlx5_query_port_roce(struct ib_device *device, u32 port_num,
458 				struct ib_port_attr *props)
459 {
460 	struct mlx5_ib_dev *dev = to_mdev(device);
461 	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0};
462 	struct mlx5_core_dev *mdev;
463 	struct net_device *ndev, *upper;
464 	enum ib_mtu ndev_ib_mtu;
465 	bool put_mdev = true;
466 	u32 eth_prot_oper;
467 	u32 mdev_port_num;
468 	bool ext;
469 	int err;
470 
471 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
472 	if (!mdev) {
473 		/* This means the port isn't affiliated yet. Get the
474 		 * info for the master port instead.
475 		 */
476 		put_mdev = false;
477 		mdev = dev->mdev;
478 		mdev_port_num = 1;
479 		port_num = 1;
480 	}
481 
482 	/* Possible bad flows are checked before filling out props so in case
483 	 * of an error it will still be zeroed out.
484 	 * Use native port in case of reps
485 	 */
486 	if (dev->is_rep)
487 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
488 					   1);
489 	else
490 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
491 					   mdev_port_num);
492 	if (err)
493 		goto out;
494 	ext = !!MLX5_GET_ETH_PROTO(ptys_reg, out, true, eth_proto_capability);
495 	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
496 
497 	props->active_width     = IB_WIDTH_4X;
498 	props->active_speed     = IB_SPEED_QDR;
499 
500 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
501 				 &props->active_width, ext);
502 
503 	if (!dev->is_rep && dev->mdev->roce.roce_en) {
504 		u16 qkey_viol_cntr;
505 
506 		props->port_cap_flags |= IB_PORT_CM_SUP;
507 		props->ip_gids = true;
508 		props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev,
509 						   roce_address_table_size);
510 		mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
511 		props->qkey_viol_cntr = qkey_viol_cntr;
512 	}
513 	props->max_mtu          = IB_MTU_4096;
514 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
515 	props->pkey_tbl_len     = 1;
516 	props->state            = IB_PORT_DOWN;
517 	props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
518 
519 	/* If this is a stub query for an unaffiliated port stop here */
520 	if (!put_mdev)
521 		goto out;
522 
523 	ndev = mlx5_ib_get_netdev(device, port_num);
524 	if (!ndev)
525 		goto out;
526 
527 	if (dev->lag_active) {
528 		rcu_read_lock();
529 		upper = netdev_master_upper_dev_get_rcu(ndev);
530 		if (upper) {
531 			dev_put(ndev);
532 			ndev = upper;
533 			dev_hold(ndev);
534 		}
535 		rcu_read_unlock();
536 	}
537 
538 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
539 		props->state      = IB_PORT_ACTIVE;
540 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
541 	}
542 
543 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
544 
545 	dev_put(ndev);
546 
547 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
548 out:
549 	if (put_mdev)
550 		mlx5_ib_put_native_port_mdev(dev, port_num);
551 	return err;
552 }
553 
554 static int set_roce_addr(struct mlx5_ib_dev *dev, u32 port_num,
555 			 unsigned int index, const union ib_gid *gid,
556 			 const struct ib_gid_attr *attr)
557 {
558 	enum ib_gid_type gid_type;
559 	u16 vlan_id = 0xffff;
560 	u8 roce_version = 0;
561 	u8 roce_l3_type = 0;
562 	u8 mac[ETH_ALEN];
563 	int ret;
564 
565 	gid_type = attr->gid_type;
566 	if (gid) {
567 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
568 		if (ret)
569 			return ret;
570 	}
571 
572 	switch (gid_type) {
573 	case IB_GID_TYPE_ROCE:
574 		roce_version = MLX5_ROCE_VERSION_1;
575 		break;
576 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
577 		roce_version = MLX5_ROCE_VERSION_2;
578 		if (gid && ipv6_addr_v4mapped((void *)gid))
579 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
580 		else
581 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
582 		break;
583 
584 	default:
585 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
586 	}
587 
588 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
589 				      roce_l3_type, gid->raw, mac,
590 				      vlan_id < VLAN_CFI_MASK, vlan_id,
591 				      port_num);
592 }
593 
594 static int mlx5_ib_add_gid(const struct ib_gid_attr *attr,
595 			   __always_unused void **context)
596 {
597 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
598 			     attr->index, &attr->gid, attr);
599 }
600 
601 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
602 			   __always_unused void **context)
603 {
604 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
605 			     attr->index, NULL, attr);
606 }
607 
608 __be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev,
609 				   const struct ib_gid_attr *attr)
610 {
611 	if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
612 		return 0;
613 
614 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
615 }
616 
617 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
618 {
619 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
620 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
621 	return 0;
622 }
623 
624 enum {
625 	MLX5_VPORT_ACCESS_METHOD_MAD,
626 	MLX5_VPORT_ACCESS_METHOD_HCA,
627 	MLX5_VPORT_ACCESS_METHOD_NIC,
628 };
629 
630 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
631 {
632 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
633 		return MLX5_VPORT_ACCESS_METHOD_MAD;
634 
635 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
636 	    IB_LINK_LAYER_ETHERNET)
637 		return MLX5_VPORT_ACCESS_METHOD_NIC;
638 
639 	return MLX5_VPORT_ACCESS_METHOD_HCA;
640 }
641 
642 static void get_atomic_caps(struct mlx5_ib_dev *dev,
643 			    u8 atomic_size_qp,
644 			    struct ib_device_attr *props)
645 {
646 	u8 tmp;
647 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
648 	u8 atomic_req_8B_endianness_mode =
649 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
650 
651 	/* Check if HW supports 8 bytes standard atomic operations and capable
652 	 * of host endianness respond
653 	 */
654 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
655 	if (((atomic_operations & tmp) == tmp) &&
656 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
657 	    (atomic_req_8B_endianness_mode)) {
658 		props->atomic_cap = IB_ATOMIC_HCA;
659 	} else {
660 		props->atomic_cap = IB_ATOMIC_NONE;
661 	}
662 }
663 
664 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
665 			       struct ib_device_attr *props)
666 {
667 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
668 
669 	get_atomic_caps(dev, atomic_size_qp, props);
670 }
671 
672 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
673 					__be64 *sys_image_guid)
674 {
675 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
676 	struct mlx5_core_dev *mdev = dev->mdev;
677 	u64 tmp;
678 	int err;
679 
680 	switch (mlx5_get_vport_access_method(ibdev)) {
681 	case MLX5_VPORT_ACCESS_METHOD_MAD:
682 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
683 							    sys_image_guid);
684 
685 	case MLX5_VPORT_ACCESS_METHOD_HCA:
686 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
687 		break;
688 
689 	case MLX5_VPORT_ACCESS_METHOD_NIC:
690 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
691 		break;
692 
693 	default:
694 		return -EINVAL;
695 	}
696 
697 	if (!err)
698 		*sys_image_guid = cpu_to_be64(tmp);
699 
700 	return err;
701 
702 }
703 
704 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
705 				u16 *max_pkeys)
706 {
707 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
708 	struct mlx5_core_dev *mdev = dev->mdev;
709 
710 	switch (mlx5_get_vport_access_method(ibdev)) {
711 	case MLX5_VPORT_ACCESS_METHOD_MAD:
712 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
713 
714 	case MLX5_VPORT_ACCESS_METHOD_HCA:
715 	case MLX5_VPORT_ACCESS_METHOD_NIC:
716 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
717 						pkey_table_size));
718 		return 0;
719 
720 	default:
721 		return -EINVAL;
722 	}
723 }
724 
725 static int mlx5_query_vendor_id(struct ib_device *ibdev,
726 				u32 *vendor_id)
727 {
728 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
729 
730 	switch (mlx5_get_vport_access_method(ibdev)) {
731 	case MLX5_VPORT_ACCESS_METHOD_MAD:
732 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
733 
734 	case MLX5_VPORT_ACCESS_METHOD_HCA:
735 	case MLX5_VPORT_ACCESS_METHOD_NIC:
736 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
737 
738 	default:
739 		return -EINVAL;
740 	}
741 }
742 
743 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
744 				__be64 *node_guid)
745 {
746 	u64 tmp;
747 	int err;
748 
749 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
750 	case MLX5_VPORT_ACCESS_METHOD_MAD:
751 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
752 
753 	case MLX5_VPORT_ACCESS_METHOD_HCA:
754 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
755 		break;
756 
757 	case MLX5_VPORT_ACCESS_METHOD_NIC:
758 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
759 		break;
760 
761 	default:
762 		return -EINVAL;
763 	}
764 
765 	if (!err)
766 		*node_guid = cpu_to_be64(tmp);
767 
768 	return err;
769 }
770 
771 struct mlx5_reg_node_desc {
772 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
773 };
774 
775 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
776 {
777 	struct mlx5_reg_node_desc in;
778 
779 	if (mlx5_use_mad_ifc(dev))
780 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
781 
782 	memset(&in, 0, sizeof(in));
783 
784 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
785 				    sizeof(struct mlx5_reg_node_desc),
786 				    MLX5_REG_NODE_DESC, 0, 0);
787 }
788 
789 static int mlx5_ib_query_device(struct ib_device *ibdev,
790 				struct ib_device_attr *props,
791 				struct ib_udata *uhw)
792 {
793 	size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
794 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
795 	struct mlx5_core_dev *mdev = dev->mdev;
796 	int err = -ENOMEM;
797 	int max_sq_desc;
798 	int max_rq_sg;
799 	int max_sq_sg;
800 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
801 	bool raw_support = !mlx5_core_mp_enabled(mdev);
802 	struct mlx5_ib_query_device_resp resp = {};
803 	size_t resp_len;
804 	u64 max_tso;
805 
806 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
807 	if (uhw_outlen && uhw_outlen < resp_len)
808 		return -EINVAL;
809 
810 	resp.response_length = resp_len;
811 
812 	if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
813 		return -EINVAL;
814 
815 	memset(props, 0, sizeof(*props));
816 	err = mlx5_query_system_image_guid(ibdev,
817 					   &props->sys_image_guid);
818 	if (err)
819 		return err;
820 
821 	props->max_pkeys = dev->pkey_table_len;
822 
823 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
824 	if (err)
825 		return err;
826 
827 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
828 		(fw_rev_min(dev->mdev) << 16) |
829 		fw_rev_sub(dev->mdev);
830 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
831 		IB_DEVICE_PORT_ACTIVE_EVENT		|
832 		IB_DEVICE_SYS_IMAGE_GUID		|
833 		IB_DEVICE_RC_RNR_NAK_GEN;
834 
835 	if (MLX5_CAP_GEN(mdev, pkv))
836 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
837 	if (MLX5_CAP_GEN(mdev, qkv))
838 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
839 	if (MLX5_CAP_GEN(mdev, apm))
840 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
841 	if (MLX5_CAP_GEN(mdev, xrc))
842 		props->device_cap_flags |= IB_DEVICE_XRC;
843 	if (MLX5_CAP_GEN(mdev, imaicl)) {
844 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
845 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
846 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
847 		/* We support 'Gappy' memory registration too */
848 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
849 	}
850 	/* IB_WR_REG_MR always requires changing the entity size with UMR */
851 	if (!MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled))
852 		props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
853 	if (MLX5_CAP_GEN(mdev, sho)) {
854 		props->device_cap_flags |= IB_DEVICE_INTEGRITY_HANDOVER;
855 		/* At this stage no support for signature handover */
856 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
857 				      IB_PROT_T10DIF_TYPE_2 |
858 				      IB_PROT_T10DIF_TYPE_3;
859 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
860 				       IB_GUARD_T10DIF_CSUM;
861 	}
862 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
863 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
864 
865 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
866 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
867 			/* Legacy bit to support old userspace libraries */
868 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
869 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
870 		}
871 
872 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
873 			props->raw_packet_caps |=
874 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
875 
876 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) {
877 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
878 			if (max_tso) {
879 				resp.tso_caps.max_tso = 1 << max_tso;
880 				resp.tso_caps.supported_qpts |=
881 					1 << IB_QPT_RAW_PACKET;
882 				resp.response_length += sizeof(resp.tso_caps);
883 			}
884 		}
885 
886 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) {
887 			resp.rss_caps.rx_hash_function =
888 						MLX5_RX_HASH_FUNC_TOEPLITZ;
889 			resp.rss_caps.rx_hash_fields_mask =
890 						MLX5_RX_HASH_SRC_IPV4 |
891 						MLX5_RX_HASH_DST_IPV4 |
892 						MLX5_RX_HASH_SRC_IPV6 |
893 						MLX5_RX_HASH_DST_IPV6 |
894 						MLX5_RX_HASH_SRC_PORT_TCP |
895 						MLX5_RX_HASH_DST_PORT_TCP |
896 						MLX5_RX_HASH_SRC_PORT_UDP |
897 						MLX5_RX_HASH_DST_PORT_UDP |
898 						MLX5_RX_HASH_INNER;
899 			if (mlx5_accel_ipsec_device_caps(dev->mdev) &
900 			    MLX5_ACCEL_IPSEC_CAP_DEVICE)
901 				resp.rss_caps.rx_hash_fields_mask |=
902 					MLX5_RX_HASH_IPSEC_SPI;
903 			resp.response_length += sizeof(resp.rss_caps);
904 		}
905 	} else {
906 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen)
907 			resp.response_length += sizeof(resp.tso_caps);
908 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen)
909 			resp.response_length += sizeof(resp.rss_caps);
910 	}
911 
912 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
913 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
914 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
915 	}
916 
917 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
918 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
919 	    raw_support)
920 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
921 
922 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
923 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
924 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
925 
926 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
927 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
928 	    raw_support) {
929 		/* Legacy bit to support old userspace libraries */
930 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
931 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
932 	}
933 
934 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
935 		props->max_dm_size =
936 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
937 	}
938 
939 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
940 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
941 
942 	if (MLX5_CAP_GEN(mdev, end_pad))
943 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
944 
945 	props->vendor_part_id	   = mdev->pdev->device;
946 	props->hw_ver		   = mdev->pdev->revision;
947 
948 	props->max_mr_size	   = ~0ull;
949 	props->page_size_cap	   = ~(min_page_size - 1);
950 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
951 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
952 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
953 		     sizeof(struct mlx5_wqe_data_seg);
954 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
955 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
956 		     sizeof(struct mlx5_wqe_raddr_seg)) /
957 		sizeof(struct mlx5_wqe_data_seg);
958 	props->max_send_sge = max_sq_sg;
959 	props->max_recv_sge = max_rq_sg;
960 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
961 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
962 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
963 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
964 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
965 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
966 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
967 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
968 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
969 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
970 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
971 	props->max_srq_sge	   = max_rq_sg - 1;
972 	props->max_fast_reg_page_list_len =
973 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
974 	props->max_pi_fast_reg_page_list_len =
975 		props->max_fast_reg_page_list_len / 2;
976 	props->max_sgl_rd =
977 		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
978 	get_atomic_caps_qp(dev, props);
979 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
980 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
981 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
982 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
983 					   props->max_mcast_grp;
984 	props->max_ah = INT_MAX;
985 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
986 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
987 
988 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
989 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
990 			props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
991 		props->odp_caps = dev->odp_caps;
992 		if (!uhw) {
993 			/* ODP for kernel QPs is not implemented for receive
994 			 * WQEs and SRQ WQEs
995 			 */
996 			props->odp_caps.per_transport_caps.rc_odp_caps &=
997 				~(IB_ODP_SUPPORT_READ |
998 				  IB_ODP_SUPPORT_SRQ_RECV);
999 			props->odp_caps.per_transport_caps.uc_odp_caps &=
1000 				~(IB_ODP_SUPPORT_READ |
1001 				  IB_ODP_SUPPORT_SRQ_RECV);
1002 			props->odp_caps.per_transport_caps.ud_odp_caps &=
1003 				~(IB_ODP_SUPPORT_READ |
1004 				  IB_ODP_SUPPORT_SRQ_RECV);
1005 			props->odp_caps.per_transport_caps.xrc_odp_caps &=
1006 				~(IB_ODP_SUPPORT_READ |
1007 				  IB_ODP_SUPPORT_SRQ_RECV);
1008 		}
1009 	}
1010 
1011 	if (MLX5_CAP_GEN(mdev, cd))
1012 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
1013 
1014 	if (mlx5_core_is_vf(mdev))
1015 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
1016 
1017 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1018 	    IB_LINK_LAYER_ETHERNET && raw_support) {
1019 		props->rss_caps.max_rwq_indirection_tables =
1020 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1021 		props->rss_caps.max_rwq_indirection_table_size =
1022 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1023 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1024 		props->max_wq_type_rq =
1025 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1026 	}
1027 
1028 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1029 		props->tm_caps.max_num_tags =
1030 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1031 		props->tm_caps.max_ops =
1032 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1033 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1034 	}
1035 
1036 	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1037 	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1038 		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1039 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1040 	}
1041 
1042 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1043 		props->cq_caps.max_cq_moderation_count =
1044 						MLX5_MAX_CQ_COUNT;
1045 		props->cq_caps.max_cq_moderation_period =
1046 						MLX5_MAX_CQ_PERIOD;
1047 	}
1048 
1049 	if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) {
1050 		resp.response_length += sizeof(resp.cqe_comp_caps);
1051 
1052 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1053 			resp.cqe_comp_caps.max_num =
1054 				MLX5_CAP_GEN(dev->mdev,
1055 					     cqe_compression_max_num);
1056 
1057 			resp.cqe_comp_caps.supported_format =
1058 				MLX5_IB_CQE_RES_FORMAT_HASH |
1059 				MLX5_IB_CQE_RES_FORMAT_CSUM;
1060 
1061 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1062 				resp.cqe_comp_caps.supported_format |=
1063 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1064 		}
1065 	}
1066 
1067 	if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen &&
1068 	    raw_support) {
1069 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1070 		    MLX5_CAP_GEN(mdev, qos)) {
1071 			resp.packet_pacing_caps.qp_rate_limit_max =
1072 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1073 			resp.packet_pacing_caps.qp_rate_limit_min =
1074 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1075 			resp.packet_pacing_caps.supported_qpts |=
1076 				1 << IB_QPT_RAW_PACKET;
1077 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1078 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1079 				resp.packet_pacing_caps.cap_flags |=
1080 					MLX5_IB_PP_SUPPORT_BURST;
1081 		}
1082 		resp.response_length += sizeof(resp.packet_pacing_caps);
1083 	}
1084 
1085 	if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <=
1086 	    uhw_outlen) {
1087 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1088 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1089 				MLX5_IB_ALLOW_MPW;
1090 
1091 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1092 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1093 				MLX5_IB_SUPPORT_EMPW;
1094 
1095 		resp.response_length +=
1096 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1097 	}
1098 
1099 	if (offsetofend(typeof(resp), flags) <= uhw_outlen) {
1100 		resp.response_length += sizeof(resp.flags);
1101 
1102 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1103 			resp.flags |=
1104 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1105 
1106 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1107 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1108 		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1109 			resp.flags |=
1110 				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1111 
1112 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1113 	}
1114 
1115 	if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) {
1116 		resp.response_length += sizeof(resp.sw_parsing_caps);
1117 		if (MLX5_CAP_ETH(mdev, swp)) {
1118 			resp.sw_parsing_caps.sw_parsing_offloads |=
1119 				MLX5_IB_SW_PARSING;
1120 
1121 			if (MLX5_CAP_ETH(mdev, swp_csum))
1122 				resp.sw_parsing_caps.sw_parsing_offloads |=
1123 					MLX5_IB_SW_PARSING_CSUM;
1124 
1125 			if (MLX5_CAP_ETH(mdev, swp_lso))
1126 				resp.sw_parsing_caps.sw_parsing_offloads |=
1127 					MLX5_IB_SW_PARSING_LSO;
1128 
1129 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1130 				resp.sw_parsing_caps.supported_qpts =
1131 					BIT(IB_QPT_RAW_PACKET);
1132 		}
1133 	}
1134 
1135 	if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen &&
1136 	    raw_support) {
1137 		resp.response_length += sizeof(resp.striding_rq_caps);
1138 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1139 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1140 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1141 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1142 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1143 			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1144 				resp.striding_rq_caps
1145 					.min_single_wqe_log_num_of_strides =
1146 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1147 			else
1148 				resp.striding_rq_caps
1149 					.min_single_wqe_log_num_of_strides =
1150 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1151 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1152 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1153 			resp.striding_rq_caps.supported_qpts =
1154 				BIT(IB_QPT_RAW_PACKET);
1155 		}
1156 	}
1157 
1158 	if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) {
1159 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1160 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1161 			resp.tunnel_offloads_caps |=
1162 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1163 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1164 			resp.tunnel_offloads_caps |=
1165 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1166 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1167 			resp.tunnel_offloads_caps |=
1168 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1169 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre))
1170 			resp.tunnel_offloads_caps |=
1171 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1172 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp))
1173 			resp.tunnel_offloads_caps |=
1174 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1175 	}
1176 
1177 	if (uhw_outlen) {
1178 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1179 
1180 		if (err)
1181 			return err;
1182 	}
1183 
1184 	return 0;
1185 }
1186 
1187 static void translate_active_width(struct ib_device *ibdev, u16 active_width,
1188 				   u8 *ib_width)
1189 {
1190 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1191 
1192 	if (active_width & MLX5_PTYS_WIDTH_1X)
1193 		*ib_width = IB_WIDTH_1X;
1194 	else if (active_width & MLX5_PTYS_WIDTH_2X)
1195 		*ib_width = IB_WIDTH_2X;
1196 	else if (active_width & MLX5_PTYS_WIDTH_4X)
1197 		*ib_width = IB_WIDTH_4X;
1198 	else if (active_width & MLX5_PTYS_WIDTH_8X)
1199 		*ib_width = IB_WIDTH_8X;
1200 	else if (active_width & MLX5_PTYS_WIDTH_12X)
1201 		*ib_width = IB_WIDTH_12X;
1202 	else {
1203 		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1204 			    active_width);
1205 		*ib_width = IB_WIDTH_4X;
1206 	}
1207 
1208 	return;
1209 }
1210 
1211 static int mlx5_mtu_to_ib_mtu(int mtu)
1212 {
1213 	switch (mtu) {
1214 	case 256: return 1;
1215 	case 512: return 2;
1216 	case 1024: return 3;
1217 	case 2048: return 4;
1218 	case 4096: return 5;
1219 	default:
1220 		pr_warn("invalid mtu\n");
1221 		return -1;
1222 	}
1223 }
1224 
1225 enum ib_max_vl_num {
1226 	__IB_MAX_VL_0		= 1,
1227 	__IB_MAX_VL_0_1		= 2,
1228 	__IB_MAX_VL_0_3		= 3,
1229 	__IB_MAX_VL_0_7		= 4,
1230 	__IB_MAX_VL_0_14	= 5,
1231 };
1232 
1233 enum mlx5_vl_hw_cap {
1234 	MLX5_VL_HW_0	= 1,
1235 	MLX5_VL_HW_0_1	= 2,
1236 	MLX5_VL_HW_0_2	= 3,
1237 	MLX5_VL_HW_0_3	= 4,
1238 	MLX5_VL_HW_0_4	= 5,
1239 	MLX5_VL_HW_0_5	= 6,
1240 	MLX5_VL_HW_0_6	= 7,
1241 	MLX5_VL_HW_0_7	= 8,
1242 	MLX5_VL_HW_0_14	= 15
1243 };
1244 
1245 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1246 				u8 *max_vl_num)
1247 {
1248 	switch (vl_hw_cap) {
1249 	case MLX5_VL_HW_0:
1250 		*max_vl_num = __IB_MAX_VL_0;
1251 		break;
1252 	case MLX5_VL_HW_0_1:
1253 		*max_vl_num = __IB_MAX_VL_0_1;
1254 		break;
1255 	case MLX5_VL_HW_0_3:
1256 		*max_vl_num = __IB_MAX_VL_0_3;
1257 		break;
1258 	case MLX5_VL_HW_0_7:
1259 		*max_vl_num = __IB_MAX_VL_0_7;
1260 		break;
1261 	case MLX5_VL_HW_0_14:
1262 		*max_vl_num = __IB_MAX_VL_0_14;
1263 		break;
1264 
1265 	default:
1266 		return -EINVAL;
1267 	}
1268 
1269 	return 0;
1270 }
1271 
1272 static int mlx5_query_hca_port(struct ib_device *ibdev, u32 port,
1273 			       struct ib_port_attr *props)
1274 {
1275 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1276 	struct mlx5_core_dev *mdev = dev->mdev;
1277 	struct mlx5_hca_vport_context *rep;
1278 	u16 max_mtu;
1279 	u16 oper_mtu;
1280 	int err;
1281 	u16 ib_link_width_oper;
1282 	u8 vl_hw_cap;
1283 
1284 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1285 	if (!rep) {
1286 		err = -ENOMEM;
1287 		goto out;
1288 	}
1289 
1290 	/* props being zeroed by the caller, avoid zeroing it here */
1291 
1292 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1293 	if (err)
1294 		goto out;
1295 
1296 	props->lid		= rep->lid;
1297 	props->lmc		= rep->lmc;
1298 	props->sm_lid		= rep->sm_lid;
1299 	props->sm_sl		= rep->sm_sl;
1300 	props->state		= rep->vport_state;
1301 	props->phys_state	= rep->port_physical_state;
1302 	props->port_cap_flags	= rep->cap_mask1;
1303 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1304 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1305 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1306 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1307 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1308 	props->subnet_timeout	= rep->subnet_timeout;
1309 	props->init_type_reply	= rep->init_type_reply;
1310 
1311 	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1312 		props->port_cap_flags2 = rep->cap_mask2;
1313 
1314 	err = mlx5_query_ib_port_oper(mdev, &ib_link_width_oper,
1315 				      &props->active_speed, port);
1316 	if (err)
1317 		goto out;
1318 
1319 	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1320 
1321 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1322 
1323 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1324 
1325 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1326 
1327 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1328 
1329 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1330 	if (err)
1331 		goto out;
1332 
1333 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1334 				   &props->max_vl_num);
1335 out:
1336 	kfree(rep);
1337 	return err;
1338 }
1339 
1340 int mlx5_ib_query_port(struct ib_device *ibdev, u32 port,
1341 		       struct ib_port_attr *props)
1342 {
1343 	unsigned int count;
1344 	int ret;
1345 
1346 	switch (mlx5_get_vport_access_method(ibdev)) {
1347 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1348 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1349 		break;
1350 
1351 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1352 		ret = mlx5_query_hca_port(ibdev, port, props);
1353 		break;
1354 
1355 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1356 		ret = mlx5_query_port_roce(ibdev, port, props);
1357 		break;
1358 
1359 	default:
1360 		ret = -EINVAL;
1361 	}
1362 
1363 	if (!ret && props) {
1364 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1365 		struct mlx5_core_dev *mdev;
1366 		bool put_mdev = true;
1367 
1368 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1369 		if (!mdev) {
1370 			/* If the port isn't affiliated yet query the master.
1371 			 * The master and slave will have the same values.
1372 			 */
1373 			mdev = dev->mdev;
1374 			port = 1;
1375 			put_mdev = false;
1376 		}
1377 		count = mlx5_core_reserved_gids_count(mdev);
1378 		if (put_mdev)
1379 			mlx5_ib_put_native_port_mdev(dev, port);
1380 		props->gid_tbl_len -= count;
1381 	}
1382 	return ret;
1383 }
1384 
1385 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u32 port,
1386 				  struct ib_port_attr *props)
1387 {
1388 	return mlx5_query_port_roce(ibdev, port, props);
1389 }
1390 
1391 static int mlx5_ib_rep_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1392 				  u16 *pkey)
1393 {
1394 	/* Default special Pkey for representor device port as per the
1395 	 * IB specification 1.3 section 10.9.1.2.
1396 	 */
1397 	*pkey = 0xffff;
1398 	return 0;
1399 }
1400 
1401 static int mlx5_ib_query_gid(struct ib_device *ibdev, u32 port, int index,
1402 			     union ib_gid *gid)
1403 {
1404 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1405 	struct mlx5_core_dev *mdev = dev->mdev;
1406 
1407 	switch (mlx5_get_vport_access_method(ibdev)) {
1408 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1409 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1410 
1411 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1412 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1413 
1414 	default:
1415 		return -EINVAL;
1416 	}
1417 
1418 }
1419 
1420 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u32 port,
1421 				   u16 index, u16 *pkey)
1422 {
1423 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1424 	struct mlx5_core_dev *mdev;
1425 	bool put_mdev = true;
1426 	u32 mdev_port_num;
1427 	int err;
1428 
1429 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1430 	if (!mdev) {
1431 		/* The port isn't affiliated yet, get the PKey from the master
1432 		 * port. For RoCE the PKey tables will be the same.
1433 		 */
1434 		put_mdev = false;
1435 		mdev = dev->mdev;
1436 		mdev_port_num = 1;
1437 	}
1438 
1439 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1440 					index, pkey);
1441 	if (put_mdev)
1442 		mlx5_ib_put_native_port_mdev(dev, port);
1443 
1444 	return err;
1445 }
1446 
1447 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1448 			      u16 *pkey)
1449 {
1450 	switch (mlx5_get_vport_access_method(ibdev)) {
1451 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1452 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1453 
1454 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1455 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1456 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1457 	default:
1458 		return -EINVAL;
1459 	}
1460 }
1461 
1462 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1463 				 struct ib_device_modify *props)
1464 {
1465 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1466 	struct mlx5_reg_node_desc in;
1467 	struct mlx5_reg_node_desc out;
1468 	int err;
1469 
1470 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1471 		return -EOPNOTSUPP;
1472 
1473 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1474 		return 0;
1475 
1476 	/*
1477 	 * If possible, pass node desc to FW, so it can generate
1478 	 * a 144 trap.  If cmd fails, just ignore.
1479 	 */
1480 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1481 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1482 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1483 	if (err)
1484 		return err;
1485 
1486 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1487 
1488 	return err;
1489 }
1490 
1491 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u32 port_num, u32 mask,
1492 				u32 value)
1493 {
1494 	struct mlx5_hca_vport_context ctx = {};
1495 	struct mlx5_core_dev *mdev;
1496 	u32 mdev_port_num;
1497 	int err;
1498 
1499 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1500 	if (!mdev)
1501 		return -ENODEV;
1502 
1503 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1504 	if (err)
1505 		goto out;
1506 
1507 	if (~ctx.cap_mask1_perm & mask) {
1508 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1509 			     mask, ctx.cap_mask1_perm);
1510 		err = -EINVAL;
1511 		goto out;
1512 	}
1513 
1514 	ctx.cap_mask1 = value;
1515 	ctx.cap_mask1_perm = mask;
1516 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1517 						 0, &ctx);
1518 
1519 out:
1520 	mlx5_ib_put_native_port_mdev(dev, port_num);
1521 
1522 	return err;
1523 }
1524 
1525 static int mlx5_ib_modify_port(struct ib_device *ibdev, u32 port, int mask,
1526 			       struct ib_port_modify *props)
1527 {
1528 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1529 	struct ib_port_attr attr;
1530 	u32 tmp;
1531 	int err;
1532 	u32 change_mask;
1533 	u32 value;
1534 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1535 		      IB_LINK_LAYER_INFINIBAND);
1536 
1537 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1538 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1539 	 */
1540 	if (!is_ib)
1541 		return 0;
1542 
1543 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1544 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1545 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1546 		return set_port_caps_atomic(dev, port, change_mask, value);
1547 	}
1548 
1549 	mutex_lock(&dev->cap_mask_mutex);
1550 
1551 	err = ib_query_port(ibdev, port, &attr);
1552 	if (err)
1553 		goto out;
1554 
1555 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1556 		~props->clr_port_cap_mask;
1557 
1558 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1559 
1560 out:
1561 	mutex_unlock(&dev->cap_mask_mutex);
1562 	return err;
1563 }
1564 
1565 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1566 {
1567 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1568 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1569 }
1570 
1571 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1572 {
1573 	/* Large page with non 4k uar support might limit the dynamic size */
1574 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1575 		return MLX5_MIN_DYN_BFREGS;
1576 
1577 	return MLX5_MAX_DYN_BFREGS;
1578 }
1579 
1580 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1581 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1582 			     struct mlx5_bfreg_info *bfregi)
1583 {
1584 	int uars_per_sys_page;
1585 	int bfregs_per_sys_page;
1586 	int ref_bfregs = req->total_num_bfregs;
1587 
1588 	if (req->total_num_bfregs == 0)
1589 		return -EINVAL;
1590 
1591 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1592 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1593 
1594 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1595 		return -ENOMEM;
1596 
1597 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1598 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1599 	/* This holds the required static allocation asked by the user */
1600 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1601 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1602 		return -EINVAL;
1603 
1604 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1605 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1606 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1607 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1608 
1609 	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",
1610 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1611 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1612 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1613 		    bfregi->num_sys_pages);
1614 
1615 	return 0;
1616 }
1617 
1618 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1619 {
1620 	struct mlx5_bfreg_info *bfregi;
1621 	int err;
1622 	int i;
1623 
1624 	bfregi = &context->bfregi;
1625 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1626 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1627 		if (err)
1628 			goto error;
1629 
1630 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1631 	}
1632 
1633 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1634 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1635 
1636 	return 0;
1637 
1638 error:
1639 	for (--i; i >= 0; i--)
1640 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1641 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1642 
1643 	return err;
1644 }
1645 
1646 static void deallocate_uars(struct mlx5_ib_dev *dev,
1647 			    struct mlx5_ib_ucontext *context)
1648 {
1649 	struct mlx5_bfreg_info *bfregi;
1650 	int i;
1651 
1652 	bfregi = &context->bfregi;
1653 	for (i = 0; i < bfregi->num_sys_pages; i++)
1654 		if (i < bfregi->num_static_sys_pages ||
1655 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1656 			mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1657 }
1658 
1659 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1660 {
1661 	int err = 0;
1662 
1663 	mutex_lock(&dev->lb.mutex);
1664 	if (td)
1665 		dev->lb.user_td++;
1666 	if (qp)
1667 		dev->lb.qps++;
1668 
1669 	if (dev->lb.user_td == 2 ||
1670 	    dev->lb.qps == 1) {
1671 		if (!dev->lb.enabled) {
1672 			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1673 			dev->lb.enabled = true;
1674 		}
1675 	}
1676 
1677 	mutex_unlock(&dev->lb.mutex);
1678 
1679 	return err;
1680 }
1681 
1682 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1683 {
1684 	mutex_lock(&dev->lb.mutex);
1685 	if (td)
1686 		dev->lb.user_td--;
1687 	if (qp)
1688 		dev->lb.qps--;
1689 
1690 	if (dev->lb.user_td == 1 &&
1691 	    dev->lb.qps == 0) {
1692 		if (dev->lb.enabled) {
1693 			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1694 			dev->lb.enabled = false;
1695 		}
1696 	}
1697 
1698 	mutex_unlock(&dev->lb.mutex);
1699 }
1700 
1701 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1702 					  u16 uid)
1703 {
1704 	int err;
1705 
1706 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1707 		return 0;
1708 
1709 	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1710 	if (err)
1711 		return err;
1712 
1713 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1714 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1715 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1716 		return err;
1717 
1718 	return mlx5_ib_enable_lb(dev, true, false);
1719 }
1720 
1721 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1722 					     u16 uid)
1723 {
1724 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1725 		return;
1726 
1727 	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1728 
1729 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1730 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1731 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1732 		return;
1733 
1734 	mlx5_ib_disable_lb(dev, true, false);
1735 }
1736 
1737 static int set_ucontext_resp(struct ib_ucontext *uctx,
1738 			     struct mlx5_ib_alloc_ucontext_resp *resp)
1739 {
1740 	struct ib_device *ibdev = uctx->device;
1741 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1742 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1743 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1744 	int err;
1745 
1746 	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1747 		err = mlx5_cmd_dump_fill_mkey(dev->mdev,
1748 					      &resp->dump_fill_mkey);
1749 		if (err)
1750 			return err;
1751 		resp->comp_mask |=
1752 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1753 	}
1754 
1755 	resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1756 	if (dev->wc_support)
1757 		resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev,
1758 						      log_bf_reg_size);
1759 	resp->cache_line_size = cache_line_size();
1760 	resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1761 	resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1762 	resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1763 	resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1764 	resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1765 	resp->cqe_version = context->cqe_version;
1766 	resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1767 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1768 	resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1769 					MLX5_CAP_GEN(dev->mdev,
1770 						     num_of_uars_per_page) : 1;
1771 
1772 	if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1773 				MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1774 		if (mlx5_get_flow_namespace(dev->mdev,
1775 				MLX5_FLOW_NAMESPACE_EGRESS))
1776 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1777 		if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1778 				MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1779 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1780 		if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1781 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1782 		if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1783 				MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1784 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1785 		/* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1786 	}
1787 
1788 	resp->tot_bfregs = bfregi->lib_uar_dyn ? 0 :
1789 			bfregi->total_num_bfregs - bfregi->num_dyn_bfregs;
1790 	resp->num_ports = dev->num_ports;
1791 	resp->cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1792 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1793 
1794 	if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1795 		mlx5_query_min_inline(dev->mdev, &resp->eth_min_inline);
1796 		resp->eth_min_inline++;
1797 	}
1798 
1799 	if (dev->mdev->clock_info)
1800 		resp->clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1801 
1802 	/*
1803 	 * We don't want to expose information from the PCI bar that is located
1804 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1805 	 * pretend we don't support reading the HCA's core clock. This is also
1806 	 * forced by mmap function.
1807 	 */
1808 	if (PAGE_SIZE <= 4096) {
1809 		resp->comp_mask |=
1810 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1811 		resp->hca_core_clock_offset =
1812 			offsetof(struct mlx5_init_seg,
1813 				 internal_timer_h) % PAGE_SIZE;
1814 	}
1815 
1816 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
1817 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE;
1818 
1819 	resp->num_dyn_bfregs = bfregi->num_dyn_bfregs;
1820 	return 0;
1821 }
1822 
1823 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1824 				  struct ib_udata *udata)
1825 {
1826 	struct ib_device *ibdev = uctx->device;
1827 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1828 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1829 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1830 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1831 	struct mlx5_bfreg_info *bfregi;
1832 	int ver;
1833 	int err;
1834 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1835 				     max_cqe_version);
1836 	bool lib_uar_4k;
1837 	bool lib_uar_dyn;
1838 
1839 	if (!dev->ib_active)
1840 		return -EAGAIN;
1841 
1842 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1843 		ver = 0;
1844 	else if (udata->inlen >= min_req_v2)
1845 		ver = 2;
1846 	else
1847 		return -EINVAL;
1848 
1849 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1850 	if (err)
1851 		return err;
1852 
1853 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1854 		return -EOPNOTSUPP;
1855 
1856 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1857 		return -EOPNOTSUPP;
1858 
1859 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1860 				    MLX5_NON_FP_BFREGS_PER_UAR);
1861 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1862 		return -EINVAL;
1863 
1864 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1865 	lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
1866 	bfregi = &context->bfregi;
1867 
1868 	if (lib_uar_dyn) {
1869 		bfregi->lib_uar_dyn = lib_uar_dyn;
1870 		goto uar_done;
1871 	}
1872 
1873 	/* updates req->total_num_bfregs */
1874 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1875 	if (err)
1876 		goto out_ctx;
1877 
1878 	mutex_init(&bfregi->lock);
1879 	bfregi->lib_uar_4k = lib_uar_4k;
1880 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1881 				GFP_KERNEL);
1882 	if (!bfregi->count) {
1883 		err = -ENOMEM;
1884 		goto out_ctx;
1885 	}
1886 
1887 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1888 				    sizeof(*bfregi->sys_pages),
1889 				    GFP_KERNEL);
1890 	if (!bfregi->sys_pages) {
1891 		err = -ENOMEM;
1892 		goto out_count;
1893 	}
1894 
1895 	err = allocate_uars(dev, context);
1896 	if (err)
1897 		goto out_sys_pages;
1898 
1899 uar_done:
1900 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1901 		err = mlx5_ib_devx_create(dev, true);
1902 		if (err < 0)
1903 			goto out_uars;
1904 		context->devx_uid = err;
1905 	}
1906 
1907 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1908 					     context->devx_uid);
1909 	if (err)
1910 		goto out_devx;
1911 
1912 	INIT_LIST_HEAD(&context->db_page_list);
1913 	mutex_init(&context->db_page_mutex);
1914 
1915 	context->cqe_version = min_t(__u8,
1916 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1917 				 req.max_cqe_version);
1918 
1919 	err = set_ucontext_resp(uctx, &resp);
1920 	if (err)
1921 		goto out_mdev;
1922 
1923 	resp.response_length = min(udata->outlen, sizeof(resp));
1924 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1925 	if (err)
1926 		goto out_mdev;
1927 
1928 	bfregi->ver = ver;
1929 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1930 	context->lib_caps = req.lib_caps;
1931 	print_lib_caps(dev, context->lib_caps);
1932 
1933 	if (mlx5_ib_lag_should_assign_affinity(dev)) {
1934 		u32 port = mlx5_core_native_port_num(dev->mdev) - 1;
1935 
1936 		atomic_set(&context->tx_port_affinity,
1937 			   atomic_add_return(
1938 				   1, &dev->port[port].roce.tx_port_affinity));
1939 	}
1940 
1941 	return 0;
1942 
1943 out_mdev:
1944 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1945 out_devx:
1946 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1947 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1948 
1949 out_uars:
1950 	deallocate_uars(dev, context);
1951 
1952 out_sys_pages:
1953 	kfree(bfregi->sys_pages);
1954 
1955 out_count:
1956 	kfree(bfregi->count);
1957 
1958 out_ctx:
1959 	return err;
1960 }
1961 
1962 static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext,
1963 				  struct uverbs_attr_bundle *attrs)
1964 {
1965 	struct mlx5_ib_alloc_ucontext_resp uctx_resp = {};
1966 	int ret;
1967 
1968 	ret = set_ucontext_resp(ibcontext, &uctx_resp);
1969 	if (ret)
1970 		return ret;
1971 
1972 	uctx_resp.response_length =
1973 		min_t(size_t,
1974 		      uverbs_attr_get_len(attrs,
1975 				MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX),
1976 		      sizeof(uctx_resp));
1977 
1978 	ret = uverbs_copy_to_struct_or_zero(attrs,
1979 					MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
1980 					&uctx_resp,
1981 					sizeof(uctx_resp));
1982 	return ret;
1983 }
1984 
1985 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1986 {
1987 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1988 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1989 	struct mlx5_bfreg_info *bfregi;
1990 
1991 	bfregi = &context->bfregi;
1992 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1993 
1994 	if (context->devx_uid)
1995 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1996 
1997 	deallocate_uars(dev, context);
1998 	kfree(bfregi->sys_pages);
1999 	kfree(bfregi->count);
2000 }
2001 
2002 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2003 				 int uar_idx)
2004 {
2005 	int fw_uars_per_page;
2006 
2007 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2008 
2009 	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2010 }
2011 
2012 static u64 uar_index2paddress(struct mlx5_ib_dev *dev,
2013 				 int uar_idx)
2014 {
2015 	unsigned int fw_uars_per_page;
2016 
2017 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
2018 				MLX5_UARS_IN_PAGE : 1;
2019 
2020 	return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE);
2021 }
2022 
2023 static int get_command(unsigned long offset)
2024 {
2025 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2026 }
2027 
2028 static int get_arg(unsigned long offset)
2029 {
2030 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2031 }
2032 
2033 static int get_index(unsigned long offset)
2034 {
2035 	return get_arg(offset);
2036 }
2037 
2038 /* Index resides in an extra byte to enable larger values than 255 */
2039 static int get_extended_index(unsigned long offset)
2040 {
2041 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2042 }
2043 
2044 
2045 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2046 {
2047 }
2048 
2049 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2050 {
2051 	switch (cmd) {
2052 	case MLX5_IB_MMAP_WC_PAGE:
2053 		return "WC";
2054 	case MLX5_IB_MMAP_REGULAR_PAGE:
2055 		return "best effort WC";
2056 	case MLX5_IB_MMAP_NC_PAGE:
2057 		return "NC";
2058 	case MLX5_IB_MMAP_DEVICE_MEM:
2059 		return "Device Memory";
2060 	default:
2061 		return NULL;
2062 	}
2063 }
2064 
2065 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2066 					struct vm_area_struct *vma,
2067 					struct mlx5_ib_ucontext *context)
2068 {
2069 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2070 	    !(vma->vm_flags & VM_SHARED))
2071 		return -EINVAL;
2072 
2073 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2074 		return -EOPNOTSUPP;
2075 
2076 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2077 		return -EPERM;
2078 	vma->vm_flags &= ~VM_MAYWRITE;
2079 
2080 	if (!dev->mdev->clock_info)
2081 		return -EOPNOTSUPP;
2082 
2083 	return vm_insert_page(vma, vma->vm_start,
2084 			      virt_to_page(dev->mdev->clock_info));
2085 }
2086 
2087 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2088 {
2089 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2090 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2091 	struct mlx5_var_table *var_table = &dev->var_table;
2092 
2093 	switch (mentry->mmap_flag) {
2094 	case MLX5_IB_MMAP_TYPE_MEMIC:
2095 	case MLX5_IB_MMAP_TYPE_MEMIC_OP:
2096 		mlx5_ib_dm_mmap_free(dev, mentry);
2097 		break;
2098 	case MLX5_IB_MMAP_TYPE_VAR:
2099 		mutex_lock(&var_table->bitmap_lock);
2100 		clear_bit(mentry->page_idx, var_table->bitmap);
2101 		mutex_unlock(&var_table->bitmap_lock);
2102 		kfree(mentry);
2103 		break;
2104 	case MLX5_IB_MMAP_TYPE_UAR_WC:
2105 	case MLX5_IB_MMAP_TYPE_UAR_NC:
2106 		mlx5_cmd_free_uar(dev->mdev, mentry->page_idx);
2107 		kfree(mentry);
2108 		break;
2109 	default:
2110 		WARN_ON(true);
2111 	}
2112 }
2113 
2114 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2115 		    struct vm_area_struct *vma,
2116 		    struct mlx5_ib_ucontext *context)
2117 {
2118 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2119 	int err;
2120 	unsigned long idx;
2121 	phys_addr_t pfn;
2122 	pgprot_t prot;
2123 	u32 bfreg_dyn_idx = 0;
2124 	u32 uar_index;
2125 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2126 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2127 				bfregi->num_static_sys_pages;
2128 
2129 	if (bfregi->lib_uar_dyn)
2130 		return -EINVAL;
2131 
2132 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2133 		return -EINVAL;
2134 
2135 	if (dyn_uar)
2136 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2137 	else
2138 		idx = get_index(vma->vm_pgoff);
2139 
2140 	if (idx >= max_valid_idx) {
2141 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2142 			     idx, max_valid_idx);
2143 		return -EINVAL;
2144 	}
2145 
2146 	switch (cmd) {
2147 	case MLX5_IB_MMAP_WC_PAGE:
2148 	case MLX5_IB_MMAP_ALLOC_WC:
2149 	case MLX5_IB_MMAP_REGULAR_PAGE:
2150 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2151 		prot = pgprot_writecombine(vma->vm_page_prot);
2152 		break;
2153 	case MLX5_IB_MMAP_NC_PAGE:
2154 		prot = pgprot_noncached(vma->vm_page_prot);
2155 		break;
2156 	default:
2157 		return -EINVAL;
2158 	}
2159 
2160 	if (dyn_uar) {
2161 		int uars_per_page;
2162 
2163 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2164 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2165 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2166 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2167 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2168 			return -EINVAL;
2169 		}
2170 
2171 		mutex_lock(&bfregi->lock);
2172 		/* Fail if uar already allocated, first bfreg index of each
2173 		 * page holds its count.
2174 		 */
2175 		if (bfregi->count[bfreg_dyn_idx]) {
2176 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2177 			mutex_unlock(&bfregi->lock);
2178 			return -EINVAL;
2179 		}
2180 
2181 		bfregi->count[bfreg_dyn_idx]++;
2182 		mutex_unlock(&bfregi->lock);
2183 
2184 		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2185 		if (err) {
2186 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2187 			goto free_bfreg;
2188 		}
2189 	} else {
2190 		uar_index = bfregi->sys_pages[idx];
2191 	}
2192 
2193 	pfn = uar_index2pfn(dev, uar_index);
2194 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2195 
2196 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2197 				prot, NULL);
2198 	if (err) {
2199 		mlx5_ib_err(dev,
2200 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2201 			    err, mmap_cmd2str(cmd));
2202 		goto err;
2203 	}
2204 
2205 	if (dyn_uar)
2206 		bfregi->sys_pages[idx] = uar_index;
2207 	return 0;
2208 
2209 err:
2210 	if (!dyn_uar)
2211 		return err;
2212 
2213 	mlx5_cmd_free_uar(dev->mdev, idx);
2214 
2215 free_bfreg:
2216 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2217 
2218 	return err;
2219 }
2220 
2221 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2222 {
2223 	unsigned long idx;
2224 	u8 command;
2225 
2226 	command = get_command(vma->vm_pgoff);
2227 	idx = get_extended_index(vma->vm_pgoff);
2228 
2229 	return (command << 16 | idx);
2230 }
2231 
2232 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2233 			       struct vm_area_struct *vma,
2234 			       struct ib_ucontext *ucontext)
2235 {
2236 	struct mlx5_user_mmap_entry *mentry;
2237 	struct rdma_user_mmap_entry *entry;
2238 	unsigned long pgoff;
2239 	pgprot_t prot;
2240 	phys_addr_t pfn;
2241 	int ret;
2242 
2243 	pgoff = mlx5_vma_to_pgoff(vma);
2244 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2245 	if (!entry)
2246 		return -EINVAL;
2247 
2248 	mentry = to_mmmap(entry);
2249 	pfn = (mentry->address >> PAGE_SHIFT);
2250 	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
2251 	    mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
2252 		prot = pgprot_noncached(vma->vm_page_prot);
2253 	else
2254 		prot = pgprot_writecombine(vma->vm_page_prot);
2255 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2256 				entry->npages * PAGE_SIZE,
2257 				prot,
2258 				entry);
2259 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2260 	return ret;
2261 }
2262 
2263 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2264 {
2265 	u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2266 	u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2267 
2268 	return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2269 		(index & 0xFF)) << PAGE_SHIFT;
2270 }
2271 
2272 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2273 {
2274 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2275 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2276 	unsigned long command;
2277 	phys_addr_t pfn;
2278 
2279 	command = get_command(vma->vm_pgoff);
2280 	switch (command) {
2281 	case MLX5_IB_MMAP_WC_PAGE:
2282 	case MLX5_IB_MMAP_ALLOC_WC:
2283 		if (!dev->wc_support)
2284 			return -EPERM;
2285 		fallthrough;
2286 	case MLX5_IB_MMAP_NC_PAGE:
2287 	case MLX5_IB_MMAP_REGULAR_PAGE:
2288 		return uar_mmap(dev, command, vma, context);
2289 
2290 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2291 		return -ENOSYS;
2292 
2293 	case MLX5_IB_MMAP_CORE_CLOCK:
2294 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2295 			return -EINVAL;
2296 
2297 		if (vma->vm_flags & VM_WRITE)
2298 			return -EPERM;
2299 		vma->vm_flags &= ~VM_MAYWRITE;
2300 
2301 		/* Don't expose to user-space information it shouldn't have */
2302 		if (PAGE_SIZE > 4096)
2303 			return -EOPNOTSUPP;
2304 
2305 		pfn = (dev->mdev->iseg_base +
2306 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2307 			PAGE_SHIFT;
2308 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2309 					 PAGE_SIZE,
2310 					 pgprot_noncached(vma->vm_page_prot),
2311 					 NULL);
2312 	case MLX5_IB_MMAP_CLOCK_INFO:
2313 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2314 
2315 	default:
2316 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2317 	}
2318 
2319 	return 0;
2320 }
2321 
2322 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2323 {
2324 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2325 	struct ib_device *ibdev = ibpd->device;
2326 	struct mlx5_ib_alloc_pd_resp resp;
2327 	int err;
2328 	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2329 	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {};
2330 	u16 uid = 0;
2331 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2332 		udata, struct mlx5_ib_ucontext, ibucontext);
2333 
2334 	uid = context ? context->devx_uid : 0;
2335 	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2336 	MLX5_SET(alloc_pd_in, in, uid, uid);
2337 	err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out);
2338 	if (err)
2339 		return err;
2340 
2341 	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2342 	pd->uid = uid;
2343 	if (udata) {
2344 		resp.pdn = pd->pdn;
2345 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2346 			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2347 			return -EFAULT;
2348 		}
2349 	}
2350 
2351 	return 0;
2352 }
2353 
2354 static int mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2355 {
2356 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2357 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2358 
2359 	return mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2360 }
2361 
2362 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2363 {
2364 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2365 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
2366 	int err;
2367 	u16 uid;
2368 
2369 	uid = ibqp->pd ?
2370 		to_mpd(ibqp->pd)->uid : 0;
2371 
2372 	if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) {
2373 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
2374 		return -EOPNOTSUPP;
2375 	}
2376 
2377 	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2378 	if (err)
2379 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2380 			     ibqp->qp_num, gid->raw);
2381 
2382 	return err;
2383 }
2384 
2385 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2386 {
2387 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2388 	int err;
2389 	u16 uid;
2390 
2391 	uid = ibqp->pd ?
2392 		to_mpd(ibqp->pd)->uid : 0;
2393 	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2394 	if (err)
2395 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2396 			     ibqp->qp_num, gid->raw);
2397 
2398 	return err;
2399 }
2400 
2401 static int init_node_data(struct mlx5_ib_dev *dev)
2402 {
2403 	int err;
2404 
2405 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2406 	if (err)
2407 		return err;
2408 
2409 	dev->mdev->rev_id = dev->mdev->pdev->revision;
2410 
2411 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2412 }
2413 
2414 static ssize_t fw_pages_show(struct device *device,
2415 			     struct device_attribute *attr, char *buf)
2416 {
2417 	struct mlx5_ib_dev *dev =
2418 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2419 
2420 	return sysfs_emit(buf, "%d\n", dev->mdev->priv.fw_pages);
2421 }
2422 static DEVICE_ATTR_RO(fw_pages);
2423 
2424 static ssize_t reg_pages_show(struct device *device,
2425 			      struct device_attribute *attr, char *buf)
2426 {
2427 	struct mlx5_ib_dev *dev =
2428 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2429 
2430 	return sysfs_emit(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2431 }
2432 static DEVICE_ATTR_RO(reg_pages);
2433 
2434 static ssize_t hca_type_show(struct device *device,
2435 			     struct device_attribute *attr, char *buf)
2436 {
2437 	struct mlx5_ib_dev *dev =
2438 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2439 
2440 	return sysfs_emit(buf, "MT%d\n", dev->mdev->pdev->device);
2441 }
2442 static DEVICE_ATTR_RO(hca_type);
2443 
2444 static ssize_t hw_rev_show(struct device *device,
2445 			   struct device_attribute *attr, char *buf)
2446 {
2447 	struct mlx5_ib_dev *dev =
2448 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2449 
2450 	return sysfs_emit(buf, "%x\n", dev->mdev->rev_id);
2451 }
2452 static DEVICE_ATTR_RO(hw_rev);
2453 
2454 static ssize_t board_id_show(struct device *device,
2455 			     struct device_attribute *attr, char *buf)
2456 {
2457 	struct mlx5_ib_dev *dev =
2458 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2459 
2460 	return sysfs_emit(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2461 			  dev->mdev->board_id);
2462 }
2463 static DEVICE_ATTR_RO(board_id);
2464 
2465 static struct attribute *mlx5_class_attributes[] = {
2466 	&dev_attr_hw_rev.attr,
2467 	&dev_attr_hca_type.attr,
2468 	&dev_attr_board_id.attr,
2469 	&dev_attr_fw_pages.attr,
2470 	&dev_attr_reg_pages.attr,
2471 	NULL,
2472 };
2473 
2474 static const struct attribute_group mlx5_attr_group = {
2475 	.attrs = mlx5_class_attributes,
2476 };
2477 
2478 static void pkey_change_handler(struct work_struct *work)
2479 {
2480 	struct mlx5_ib_port_resources *ports =
2481 		container_of(work, struct mlx5_ib_port_resources,
2482 			     pkey_change_work);
2483 
2484 	mlx5_ib_gsi_pkey_change(ports->gsi);
2485 }
2486 
2487 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2488 {
2489 	struct mlx5_ib_qp *mqp;
2490 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2491 	struct mlx5_core_cq *mcq;
2492 	struct list_head cq_armed_list;
2493 	unsigned long flags_qp;
2494 	unsigned long flags_cq;
2495 	unsigned long flags;
2496 
2497 	INIT_LIST_HEAD(&cq_armed_list);
2498 
2499 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2500 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2501 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2502 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2503 		if (mqp->sq.tail != mqp->sq.head) {
2504 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2505 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2506 			if (send_mcq->mcq.comp &&
2507 			    mqp->ibqp.send_cq->comp_handler) {
2508 				if (!send_mcq->mcq.reset_notify_added) {
2509 					send_mcq->mcq.reset_notify_added = 1;
2510 					list_add_tail(&send_mcq->mcq.reset_notify,
2511 						      &cq_armed_list);
2512 				}
2513 			}
2514 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2515 		}
2516 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2517 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2518 		/* no handling is needed for SRQ */
2519 		if (!mqp->ibqp.srq) {
2520 			if (mqp->rq.tail != mqp->rq.head) {
2521 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2522 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2523 				if (recv_mcq->mcq.comp &&
2524 				    mqp->ibqp.recv_cq->comp_handler) {
2525 					if (!recv_mcq->mcq.reset_notify_added) {
2526 						recv_mcq->mcq.reset_notify_added = 1;
2527 						list_add_tail(&recv_mcq->mcq.reset_notify,
2528 							      &cq_armed_list);
2529 					}
2530 				}
2531 				spin_unlock_irqrestore(&recv_mcq->lock,
2532 						       flags_cq);
2533 			}
2534 		}
2535 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2536 	}
2537 	/*At that point all inflight post send were put to be executed as of we
2538 	 * lock/unlock above locks Now need to arm all involved CQs.
2539 	 */
2540 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2541 		mcq->comp(mcq, NULL);
2542 	}
2543 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2544 }
2545 
2546 static void delay_drop_handler(struct work_struct *work)
2547 {
2548 	int err;
2549 	struct mlx5_ib_delay_drop *delay_drop =
2550 		container_of(work, struct mlx5_ib_delay_drop,
2551 			     delay_drop_work);
2552 
2553 	atomic_inc(&delay_drop->events_cnt);
2554 
2555 	mutex_lock(&delay_drop->lock);
2556 	err = mlx5_core_set_delay_drop(delay_drop->dev, delay_drop->timeout);
2557 	if (err) {
2558 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
2559 			     delay_drop->timeout);
2560 		delay_drop->activate = false;
2561 	}
2562 	mutex_unlock(&delay_drop->lock);
2563 }
2564 
2565 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2566 				 struct ib_event *ibev)
2567 {
2568 	u32 port = (eqe->data.port.port >> 4) & 0xf;
2569 
2570 	switch (eqe->sub_type) {
2571 	case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
2572 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2573 					    IB_LINK_LAYER_ETHERNET)
2574 			schedule_work(&ibdev->delay_drop.delay_drop_work);
2575 		break;
2576 	default: /* do nothing */
2577 		return;
2578 	}
2579 }
2580 
2581 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2582 			      struct ib_event *ibev)
2583 {
2584 	u32 port = (eqe->data.port.port >> 4) & 0xf;
2585 
2586 	ibev->element.port_num = port;
2587 
2588 	switch (eqe->sub_type) {
2589 	case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
2590 	case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
2591 	case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
2592 		/* In RoCE, port up/down events are handled in
2593 		 * mlx5_netdev_event().
2594 		 */
2595 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2596 					    IB_LINK_LAYER_ETHERNET)
2597 			return -EINVAL;
2598 
2599 		ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
2600 				IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2601 		break;
2602 
2603 	case MLX5_PORT_CHANGE_SUBTYPE_LID:
2604 		ibev->event = IB_EVENT_LID_CHANGE;
2605 		break;
2606 
2607 	case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
2608 		ibev->event = IB_EVENT_PKEY_CHANGE;
2609 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2610 		break;
2611 
2612 	case MLX5_PORT_CHANGE_SUBTYPE_GUID:
2613 		ibev->event = IB_EVENT_GID_CHANGE;
2614 		break;
2615 
2616 	case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
2617 		ibev->event = IB_EVENT_CLIENT_REREGISTER;
2618 		break;
2619 	default:
2620 		return -EINVAL;
2621 	}
2622 
2623 	return 0;
2624 }
2625 
2626 static void mlx5_ib_handle_event(struct work_struct *_work)
2627 {
2628 	struct mlx5_ib_event_work *work =
2629 		container_of(_work, struct mlx5_ib_event_work, work);
2630 	struct mlx5_ib_dev *ibdev;
2631 	struct ib_event ibev;
2632 	bool fatal = false;
2633 
2634 	if (work->is_slave) {
2635 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
2636 		if (!ibdev)
2637 			goto out;
2638 	} else {
2639 		ibdev = work->dev;
2640 	}
2641 
2642 	switch (work->event) {
2643 	case MLX5_DEV_EVENT_SYS_ERROR:
2644 		ibev.event = IB_EVENT_DEVICE_FATAL;
2645 		mlx5_ib_handle_internal_error(ibdev);
2646 		ibev.element.port_num  = (u8)(unsigned long)work->param;
2647 		fatal = true;
2648 		break;
2649 	case MLX5_EVENT_TYPE_PORT_CHANGE:
2650 		if (handle_port_change(ibdev, work->param, &ibev))
2651 			goto out;
2652 		break;
2653 	case MLX5_EVENT_TYPE_GENERAL_EVENT:
2654 		handle_general_event(ibdev, work->param, &ibev);
2655 		fallthrough;
2656 	default:
2657 		goto out;
2658 	}
2659 
2660 	ibev.device = &ibdev->ib_dev;
2661 
2662 	if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
2663 		mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
2664 		goto out;
2665 	}
2666 
2667 	if (ibdev->ib_active)
2668 		ib_dispatch_event(&ibev);
2669 
2670 	if (fatal)
2671 		ibdev->ib_active = false;
2672 out:
2673 	kfree(work);
2674 }
2675 
2676 static int mlx5_ib_event(struct notifier_block *nb,
2677 			 unsigned long event, void *param)
2678 {
2679 	struct mlx5_ib_event_work *work;
2680 
2681 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2682 	if (!work)
2683 		return NOTIFY_DONE;
2684 
2685 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2686 	work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
2687 	work->is_slave = false;
2688 	work->param = param;
2689 	work->event = event;
2690 
2691 	queue_work(mlx5_ib_event_wq, &work->work);
2692 
2693 	return NOTIFY_OK;
2694 }
2695 
2696 static int mlx5_ib_event_slave_port(struct notifier_block *nb,
2697 				    unsigned long event, void *param)
2698 {
2699 	struct mlx5_ib_event_work *work;
2700 
2701 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2702 	if (!work)
2703 		return NOTIFY_DONE;
2704 
2705 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2706 	work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
2707 	work->is_slave = true;
2708 	work->param = param;
2709 	work->event = event;
2710 	queue_work(mlx5_ib_event_wq, &work->work);
2711 
2712 	return NOTIFY_OK;
2713 }
2714 
2715 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2716 {
2717 	struct mlx5_hca_vport_context vport_ctx;
2718 	int err;
2719 	int port;
2720 
2721 	for (port = 1; port <= ARRAY_SIZE(dev->port_caps); port++) {
2722 		dev->port_caps[port - 1].has_smi = false;
2723 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
2724 		    MLX5_CAP_PORT_TYPE_IB) {
2725 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2726 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
2727 								   port, 0,
2728 								   &vport_ctx);
2729 				if (err) {
2730 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2731 						    port, err);
2732 					return err;
2733 				}
2734 				dev->port_caps[port - 1].has_smi =
2735 					vport_ctx.has_smi;
2736 			} else {
2737 				dev->port_caps[port - 1].has_smi = true;
2738 			}
2739 		}
2740 	}
2741 	return 0;
2742 }
2743 
2744 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2745 {
2746 	unsigned int port;
2747 
2748 	rdma_for_each_port (&dev->ib_dev, port)
2749 		mlx5_query_ext_port_caps(dev, port);
2750 }
2751 
2752 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
2753 {
2754 	switch (umr_fence_cap) {
2755 	case MLX5_CAP_UMR_FENCE_NONE:
2756 		return MLX5_FENCE_MODE_NONE;
2757 	case MLX5_CAP_UMR_FENCE_SMALL:
2758 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
2759 	default:
2760 		return MLX5_FENCE_MODE_STRONG_ORDERING;
2761 	}
2762 }
2763 
2764 static int mlx5_ib_dev_res_init(struct mlx5_ib_dev *dev)
2765 {
2766 	struct mlx5_ib_resources *devr = &dev->devr;
2767 	struct ib_srq_init_attr attr;
2768 	struct ib_device *ibdev;
2769 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2770 	int port;
2771 	int ret = 0;
2772 
2773 	ibdev = &dev->ib_dev;
2774 
2775 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
2776 		return -EOPNOTSUPP;
2777 
2778 	mutex_init(&devr->mutex);
2779 
2780 	devr->p0 = rdma_zalloc_drv_obj(ibdev, ib_pd);
2781 	if (!devr->p0)
2782 		return -ENOMEM;
2783 
2784 	devr->p0->device  = ibdev;
2785 	devr->p0->uobject = NULL;
2786 	atomic_set(&devr->p0->usecnt, 0);
2787 
2788 	ret = mlx5_ib_alloc_pd(devr->p0, NULL);
2789 	if (ret)
2790 		goto error0;
2791 
2792 	devr->c0 = rdma_zalloc_drv_obj(ibdev, ib_cq);
2793 	if (!devr->c0) {
2794 		ret = -ENOMEM;
2795 		goto error1;
2796 	}
2797 
2798 	devr->c0->device = &dev->ib_dev;
2799 	atomic_set(&devr->c0->usecnt, 0);
2800 
2801 	ret = mlx5_ib_create_cq(devr->c0, &cq_attr, NULL);
2802 	if (ret)
2803 		goto err_create_cq;
2804 
2805 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0);
2806 	if (ret)
2807 		goto error2;
2808 
2809 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0);
2810 	if (ret)
2811 		goto error3;
2812 
2813 	memset(&attr, 0, sizeof(attr));
2814 	attr.attr.max_sge = 1;
2815 	attr.attr.max_wr = 1;
2816 	attr.srq_type = IB_SRQT_XRC;
2817 	attr.ext.cq = devr->c0;
2818 
2819 	devr->s0 = rdma_zalloc_drv_obj(ibdev, ib_srq);
2820 	if (!devr->s0) {
2821 		ret = -ENOMEM;
2822 		goto error4;
2823 	}
2824 
2825 	devr->s0->device	= &dev->ib_dev;
2826 	devr->s0->pd		= devr->p0;
2827 	devr->s0->srq_type      = IB_SRQT_XRC;
2828 	devr->s0->ext.cq	= devr->c0;
2829 	ret = mlx5_ib_create_srq(devr->s0, &attr, NULL);
2830 	if (ret)
2831 		goto err_create;
2832 
2833 	atomic_inc(&devr->s0->ext.cq->usecnt);
2834 	atomic_inc(&devr->p0->usecnt);
2835 	atomic_set(&devr->s0->usecnt, 0);
2836 
2837 	memset(&attr, 0, sizeof(attr));
2838 	attr.attr.max_sge = 1;
2839 	attr.attr.max_wr = 1;
2840 	attr.srq_type = IB_SRQT_BASIC;
2841 	devr->s1 = rdma_zalloc_drv_obj(ibdev, ib_srq);
2842 	if (!devr->s1) {
2843 		ret = -ENOMEM;
2844 		goto error5;
2845 	}
2846 
2847 	devr->s1->device	= &dev->ib_dev;
2848 	devr->s1->pd		= devr->p0;
2849 	devr->s1->srq_type      = IB_SRQT_BASIC;
2850 	devr->s1->ext.cq	= devr->c0;
2851 
2852 	ret = mlx5_ib_create_srq(devr->s1, &attr, NULL);
2853 	if (ret)
2854 		goto error6;
2855 
2856 	atomic_inc(&devr->p0->usecnt);
2857 	atomic_set(&devr->s1->usecnt, 0);
2858 
2859 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2860 		INIT_WORK(&devr->ports[port].pkey_change_work,
2861 			  pkey_change_handler);
2862 
2863 	return 0;
2864 
2865 error6:
2866 	kfree(devr->s1);
2867 error5:
2868 	mlx5_ib_destroy_srq(devr->s0, NULL);
2869 err_create:
2870 	kfree(devr->s0);
2871 error4:
2872 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2873 error3:
2874 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2875 error2:
2876 	mlx5_ib_destroy_cq(devr->c0, NULL);
2877 err_create_cq:
2878 	kfree(devr->c0);
2879 error1:
2880 	mlx5_ib_dealloc_pd(devr->p0, NULL);
2881 error0:
2882 	kfree(devr->p0);
2883 	return ret;
2884 }
2885 
2886 static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev)
2887 {
2888 	struct mlx5_ib_resources *devr = &dev->devr;
2889 	int port;
2890 
2891 	mlx5_ib_destroy_srq(devr->s1, NULL);
2892 	kfree(devr->s1);
2893 	mlx5_ib_destroy_srq(devr->s0, NULL);
2894 	kfree(devr->s0);
2895 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2896 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2897 	mlx5_ib_destroy_cq(devr->c0, NULL);
2898 	kfree(devr->c0);
2899 	mlx5_ib_dealloc_pd(devr->p0, NULL);
2900 	kfree(devr->p0);
2901 
2902 	/* Make sure no change P_Key work items are still executing */
2903 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2904 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2905 }
2906 
2907 static u32 get_core_cap_flags(struct ib_device *ibdev,
2908 			      struct mlx5_hca_vport_context *rep)
2909 {
2910 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2911 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2912 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2913 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2914 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
2915 	u32 ret = 0;
2916 
2917 	if (rep->grh_required)
2918 		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
2919 
2920 	if (ll == IB_LINK_LAYER_INFINIBAND)
2921 		return ret | RDMA_CORE_PORT_IBA_IB;
2922 
2923 	if (raw_support)
2924 		ret |= RDMA_CORE_PORT_RAW_PACKET;
2925 
2926 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2927 		return ret;
2928 
2929 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2930 		return ret;
2931 
2932 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2933 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2934 
2935 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2936 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2937 
2938 	return ret;
2939 }
2940 
2941 static int mlx5_port_immutable(struct ib_device *ibdev, u32 port_num,
2942 			       struct ib_port_immutable *immutable)
2943 {
2944 	struct ib_port_attr attr;
2945 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2946 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
2947 	struct mlx5_hca_vport_context rep = {0};
2948 	int err;
2949 
2950 	err = ib_query_port(ibdev, port_num, &attr);
2951 	if (err)
2952 		return err;
2953 
2954 	if (ll == IB_LINK_LAYER_INFINIBAND) {
2955 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
2956 						   &rep);
2957 		if (err)
2958 			return err;
2959 	}
2960 
2961 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2962 	immutable->gid_tbl_len = attr.gid_tbl_len;
2963 	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
2964 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2965 
2966 	return 0;
2967 }
2968 
2969 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u32 port_num,
2970 				   struct ib_port_immutable *immutable)
2971 {
2972 	struct ib_port_attr attr;
2973 	int err;
2974 
2975 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2976 
2977 	err = ib_query_port(ibdev, port_num, &attr);
2978 	if (err)
2979 		return err;
2980 
2981 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2982 	immutable->gid_tbl_len = attr.gid_tbl_len;
2983 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2984 
2985 	return 0;
2986 }
2987 
2988 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
2989 {
2990 	struct mlx5_ib_dev *dev =
2991 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2992 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
2993 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
2994 		 fw_rev_sub(dev->mdev));
2995 }
2996 
2997 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
2998 {
2999 	struct mlx5_core_dev *mdev = dev->mdev;
3000 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
3001 								 MLX5_FLOW_NAMESPACE_LAG);
3002 	struct mlx5_flow_table *ft;
3003 	int err;
3004 
3005 	if (!ns || !mlx5_lag_is_roce(mdev))
3006 		return 0;
3007 
3008 	err = mlx5_cmd_create_vport_lag(mdev);
3009 	if (err)
3010 		return err;
3011 
3012 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3013 	if (IS_ERR(ft)) {
3014 		err = PTR_ERR(ft);
3015 		goto err_destroy_vport_lag;
3016 	}
3017 
3018 	dev->flow_db->lag_demux_ft = ft;
3019 	dev->lag_active = true;
3020 	return 0;
3021 
3022 err_destroy_vport_lag:
3023 	mlx5_cmd_destroy_vport_lag(mdev);
3024 	return err;
3025 }
3026 
3027 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3028 {
3029 	struct mlx5_core_dev *mdev = dev->mdev;
3030 
3031 	if (dev->lag_active) {
3032 		dev->lag_active = false;
3033 
3034 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
3035 		dev->flow_db->lag_demux_ft = NULL;
3036 
3037 		mlx5_cmd_destroy_vport_lag(mdev);
3038 	}
3039 }
3040 
3041 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u32 port_num)
3042 {
3043 	int err;
3044 
3045 	dev->port[port_num].roce.nb.notifier_call = mlx5_netdev_event;
3046 	err = register_netdevice_notifier(&dev->port[port_num].roce.nb);
3047 	if (err) {
3048 		dev->port[port_num].roce.nb.notifier_call = NULL;
3049 		return err;
3050 	}
3051 
3052 	return 0;
3053 }
3054 
3055 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u32 port_num)
3056 {
3057 	if (dev->port[port_num].roce.nb.notifier_call) {
3058 		unregister_netdevice_notifier(&dev->port[port_num].roce.nb);
3059 		dev->port[port_num].roce.nb.notifier_call = NULL;
3060 	}
3061 }
3062 
3063 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3064 {
3065 	int err;
3066 
3067 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3068 	if (err)
3069 		return err;
3070 
3071 	err = mlx5_eth_lag_init(dev);
3072 	if (err)
3073 		goto err_disable_roce;
3074 
3075 	return 0;
3076 
3077 err_disable_roce:
3078 	mlx5_nic_vport_disable_roce(dev->mdev);
3079 
3080 	return err;
3081 }
3082 
3083 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3084 {
3085 	mlx5_eth_lag_cleanup(dev);
3086 	mlx5_nic_vport_disable_roce(dev->mdev);
3087 }
3088 
3089 static int mlx5_ib_rn_get_params(struct ib_device *device, u32 port_num,
3090 				 enum rdma_netdev_t type,
3091 				 struct rdma_netdev_alloc_params *params)
3092 {
3093 	if (type != RDMA_NETDEV_IPOIB)
3094 		return -EOPNOTSUPP;
3095 
3096 	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
3097 }
3098 
3099 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
3100 				       size_t count, loff_t *pos)
3101 {
3102 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3103 	char lbuf[20];
3104 	int len;
3105 
3106 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
3107 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
3108 }
3109 
3110 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
3111 					size_t count, loff_t *pos)
3112 {
3113 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3114 	u32 timeout;
3115 	u32 var;
3116 
3117 	if (kstrtouint_from_user(buf, count, 0, &var))
3118 		return -EFAULT;
3119 
3120 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
3121 			1000);
3122 	if (timeout != var)
3123 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
3124 			    timeout);
3125 
3126 	delay_drop->timeout = timeout;
3127 
3128 	return count;
3129 }
3130 
3131 static const struct file_operations fops_delay_drop_timeout = {
3132 	.owner	= THIS_MODULE,
3133 	.open	= simple_open,
3134 	.write	= delay_drop_timeout_write,
3135 	.read	= delay_drop_timeout_read,
3136 };
3137 
3138 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
3139 				      struct mlx5_ib_multiport_info *mpi)
3140 {
3141 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3142 	struct mlx5_ib_port *port = &ibdev->port[port_num];
3143 	int comps;
3144 	int err;
3145 	int i;
3146 
3147 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3148 
3149 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
3150 
3151 	spin_lock(&port->mp.mpi_lock);
3152 	if (!mpi->ibdev) {
3153 		spin_unlock(&port->mp.mpi_lock);
3154 		return;
3155 	}
3156 
3157 	mpi->ibdev = NULL;
3158 
3159 	spin_unlock(&port->mp.mpi_lock);
3160 	if (mpi->mdev_events.notifier_call)
3161 		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
3162 	mpi->mdev_events.notifier_call = NULL;
3163 	mlx5_remove_netdev_notifier(ibdev, port_num);
3164 	spin_lock(&port->mp.mpi_lock);
3165 
3166 	comps = mpi->mdev_refcnt;
3167 	if (comps) {
3168 		mpi->unaffiliate = true;
3169 		init_completion(&mpi->unref_comp);
3170 		spin_unlock(&port->mp.mpi_lock);
3171 
3172 		for (i = 0; i < comps; i++)
3173 			wait_for_completion(&mpi->unref_comp);
3174 
3175 		spin_lock(&port->mp.mpi_lock);
3176 		mpi->unaffiliate = false;
3177 	}
3178 
3179 	port->mp.mpi = NULL;
3180 
3181 	list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
3182 
3183 	spin_unlock(&port->mp.mpi_lock);
3184 
3185 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
3186 
3187 	mlx5_ib_dbg(ibdev, "unaffiliated port %u\n", port_num + 1);
3188 	/* Log an error, still needed to cleanup the pointers and add
3189 	 * it back to the list.
3190 	 */
3191 	if (err)
3192 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
3193 			    port_num + 1);
3194 
3195 	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
3196 }
3197 
3198 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
3199 				    struct mlx5_ib_multiport_info *mpi)
3200 {
3201 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3202 	int err;
3203 
3204 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3205 
3206 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
3207 	if (ibdev->port[port_num].mp.mpi) {
3208 		mlx5_ib_dbg(ibdev, "port %u already affiliated.\n",
3209 			    port_num + 1);
3210 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3211 		return false;
3212 	}
3213 
3214 	ibdev->port[port_num].mp.mpi = mpi;
3215 	mpi->ibdev = ibdev;
3216 	mpi->mdev_events.notifier_call = NULL;
3217 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3218 
3219 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
3220 	if (err)
3221 		goto unbind;
3222 
3223 	err = mlx5_add_netdev_notifier(ibdev, port_num);
3224 	if (err) {
3225 		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
3226 			    port_num + 1);
3227 		goto unbind;
3228 	}
3229 
3230 	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
3231 	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
3232 
3233 	mlx5_ib_init_cong_debugfs(ibdev, port_num);
3234 
3235 	return true;
3236 
3237 unbind:
3238 	mlx5_ib_unbind_slave_port(ibdev, mpi);
3239 	return false;
3240 }
3241 
3242 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
3243 {
3244 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3245 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3246 							  port_num + 1);
3247 	struct mlx5_ib_multiport_info *mpi;
3248 	int err;
3249 	u32 i;
3250 
3251 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3252 		return 0;
3253 
3254 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
3255 						     &dev->sys_image_guid);
3256 	if (err)
3257 		return err;
3258 
3259 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3260 	if (err)
3261 		return err;
3262 
3263 	mutex_lock(&mlx5_ib_multiport_mutex);
3264 	for (i = 0; i < dev->num_ports; i++) {
3265 		bool bound = false;
3266 
3267 		/* build a stub multiport info struct for the native port. */
3268 		if (i == port_num) {
3269 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
3270 			if (!mpi) {
3271 				mutex_unlock(&mlx5_ib_multiport_mutex);
3272 				mlx5_nic_vport_disable_roce(dev->mdev);
3273 				return -ENOMEM;
3274 			}
3275 
3276 			mpi->is_master = true;
3277 			mpi->mdev = dev->mdev;
3278 			mpi->sys_image_guid = dev->sys_image_guid;
3279 			dev->port[i].mp.mpi = mpi;
3280 			mpi->ibdev = dev;
3281 			mpi = NULL;
3282 			continue;
3283 		}
3284 
3285 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
3286 				    list) {
3287 			if (dev->sys_image_guid == mpi->sys_image_guid &&
3288 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
3289 				bound = mlx5_ib_bind_slave_port(dev, mpi);
3290 			}
3291 
3292 			if (bound) {
3293 				dev_dbg(mpi->mdev->device,
3294 					"removing port from unaffiliated list.\n");
3295 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
3296 				list_del(&mpi->list);
3297 				break;
3298 			}
3299 		}
3300 		if (!bound)
3301 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
3302 				    i + 1);
3303 	}
3304 
3305 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
3306 	mutex_unlock(&mlx5_ib_multiport_mutex);
3307 	return err;
3308 }
3309 
3310 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
3311 {
3312 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3313 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3314 							  port_num + 1);
3315 	u32 i;
3316 
3317 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3318 		return;
3319 
3320 	mutex_lock(&mlx5_ib_multiport_mutex);
3321 	for (i = 0; i < dev->num_ports; i++) {
3322 		if (dev->port[i].mp.mpi) {
3323 			/* Destroy the native port stub */
3324 			if (i == port_num) {
3325 				kfree(dev->port[i].mp.mpi);
3326 				dev->port[i].mp.mpi = NULL;
3327 			} else {
3328 				mlx5_ib_dbg(dev, "unbinding port_num: %u\n",
3329 					    i + 1);
3330 				mlx5_ib_unbind_slave_port(dev, dev->port[i].mp.mpi);
3331 			}
3332 		}
3333 	}
3334 
3335 	mlx5_ib_dbg(dev, "removing from devlist\n");
3336 	list_del(&dev->ib_dev_list);
3337 	mutex_unlock(&mlx5_ib_multiport_mutex);
3338 
3339 	mlx5_nic_vport_disable_roce(dev->mdev);
3340 }
3341 
3342 static int mmap_obj_cleanup(struct ib_uobject *uobject,
3343 			    enum rdma_remove_reason why,
3344 			    struct uverbs_attr_bundle *attrs)
3345 {
3346 	struct mlx5_user_mmap_entry *obj = uobject->object;
3347 
3348 	rdma_user_mmap_entry_remove(&obj->rdma_entry);
3349 	return 0;
3350 }
3351 
3352 static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c,
3353 					    struct mlx5_user_mmap_entry *entry,
3354 					    size_t length)
3355 {
3356 	return rdma_user_mmap_entry_insert_range(
3357 		&c->ibucontext, &entry->rdma_entry, length,
3358 		(MLX5_IB_MMAP_OFFSET_START << 16),
3359 		((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1));
3360 }
3361 
3362 static struct mlx5_user_mmap_entry *
3363 alloc_var_entry(struct mlx5_ib_ucontext *c)
3364 {
3365 	struct mlx5_user_mmap_entry *entry;
3366 	struct mlx5_var_table *var_table;
3367 	u32 page_idx;
3368 	int err;
3369 
3370 	var_table = &to_mdev(c->ibucontext.device)->var_table;
3371 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3372 	if (!entry)
3373 		return ERR_PTR(-ENOMEM);
3374 
3375 	mutex_lock(&var_table->bitmap_lock);
3376 	page_idx = find_first_zero_bit(var_table->bitmap,
3377 				       var_table->num_var_hw_entries);
3378 	if (page_idx >= var_table->num_var_hw_entries) {
3379 		err = -ENOSPC;
3380 		mutex_unlock(&var_table->bitmap_lock);
3381 		goto end;
3382 	}
3383 
3384 	set_bit(page_idx, var_table->bitmap);
3385 	mutex_unlock(&var_table->bitmap_lock);
3386 
3387 	entry->address = var_table->hw_start_addr +
3388 				(page_idx * var_table->stride_size);
3389 	entry->page_idx = page_idx;
3390 	entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
3391 
3392 	err = mlx5_rdma_user_mmap_entry_insert(c, entry,
3393 					       var_table->stride_size);
3394 	if (err)
3395 		goto err_insert;
3396 
3397 	return entry;
3398 
3399 err_insert:
3400 	mutex_lock(&var_table->bitmap_lock);
3401 	clear_bit(page_idx, var_table->bitmap);
3402 	mutex_unlock(&var_table->bitmap_lock);
3403 end:
3404 	kfree(entry);
3405 	return ERR_PTR(err);
3406 }
3407 
3408 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
3409 	struct uverbs_attr_bundle *attrs)
3410 {
3411 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3412 		attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3413 	struct mlx5_ib_ucontext *c;
3414 	struct mlx5_user_mmap_entry *entry;
3415 	u64 mmap_offset;
3416 	u32 length;
3417 	int err;
3418 
3419 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3420 	if (IS_ERR(c))
3421 		return PTR_ERR(c);
3422 
3423 	entry = alloc_var_entry(c);
3424 	if (IS_ERR(entry))
3425 		return PTR_ERR(entry);
3426 
3427 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3428 	length = entry->rdma_entry.npages * PAGE_SIZE;
3429 	uobj->object = entry;
3430 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3431 
3432 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3433 			     &mmap_offset, sizeof(mmap_offset));
3434 	if (err)
3435 		return err;
3436 
3437 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3438 			     &entry->page_idx, sizeof(entry->page_idx));
3439 	if (err)
3440 		return err;
3441 
3442 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3443 			     &length, sizeof(length));
3444 	return err;
3445 }
3446 
3447 DECLARE_UVERBS_NAMED_METHOD(
3448 	MLX5_IB_METHOD_VAR_OBJ_ALLOC,
3449 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
3450 			MLX5_IB_OBJECT_VAR,
3451 			UVERBS_ACCESS_NEW,
3452 			UA_MANDATORY),
3453 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3454 			   UVERBS_ATTR_TYPE(u32),
3455 			   UA_MANDATORY),
3456 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3457 			   UVERBS_ATTR_TYPE(u32),
3458 			   UA_MANDATORY),
3459 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3460 			    UVERBS_ATTR_TYPE(u64),
3461 			    UA_MANDATORY));
3462 
3463 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3464 	MLX5_IB_METHOD_VAR_OBJ_DESTROY,
3465 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
3466 			MLX5_IB_OBJECT_VAR,
3467 			UVERBS_ACCESS_DESTROY,
3468 			UA_MANDATORY));
3469 
3470 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
3471 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3472 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
3473 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
3474 
3475 static bool var_is_supported(struct ib_device *device)
3476 {
3477 	struct mlx5_ib_dev *dev = to_mdev(device);
3478 
3479 	return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3480 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
3481 }
3482 
3483 static struct mlx5_user_mmap_entry *
3484 alloc_uar_entry(struct mlx5_ib_ucontext *c,
3485 		enum mlx5_ib_uapi_uar_alloc_type alloc_type)
3486 {
3487 	struct mlx5_user_mmap_entry *entry;
3488 	struct mlx5_ib_dev *dev;
3489 	u32 uar_index;
3490 	int err;
3491 
3492 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3493 	if (!entry)
3494 		return ERR_PTR(-ENOMEM);
3495 
3496 	dev = to_mdev(c->ibucontext.device);
3497 	err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
3498 	if (err)
3499 		goto end;
3500 
3501 	entry->page_idx = uar_index;
3502 	entry->address = uar_index2paddress(dev, uar_index);
3503 	if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3504 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC;
3505 	else
3506 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC;
3507 
3508 	err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE);
3509 	if (err)
3510 		goto err_insert;
3511 
3512 	return entry;
3513 
3514 err_insert:
3515 	mlx5_cmd_free_uar(dev->mdev, uar_index);
3516 end:
3517 	kfree(entry);
3518 	return ERR_PTR(err);
3519 }
3520 
3521 static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)(
3522 	struct uverbs_attr_bundle *attrs)
3523 {
3524 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3525 		attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3526 	enum mlx5_ib_uapi_uar_alloc_type alloc_type;
3527 	struct mlx5_ib_ucontext *c;
3528 	struct mlx5_user_mmap_entry *entry;
3529 	u64 mmap_offset;
3530 	u32 length;
3531 	int err;
3532 
3533 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3534 	if (IS_ERR(c))
3535 		return PTR_ERR(c);
3536 
3537 	err = uverbs_get_const(&alloc_type, attrs,
3538 			       MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE);
3539 	if (err)
3540 		return err;
3541 
3542 	if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF &&
3543 	    alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC)
3544 		return -EOPNOTSUPP;
3545 
3546 	if (!to_mdev(c->ibucontext.device)->wc_support &&
3547 	    alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3548 		return -EOPNOTSUPP;
3549 
3550 	entry = alloc_uar_entry(c, alloc_type);
3551 	if (IS_ERR(entry))
3552 		return PTR_ERR(entry);
3553 
3554 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3555 	length = entry->rdma_entry.npages * PAGE_SIZE;
3556 	uobj->object = entry;
3557 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3558 
3559 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3560 			     &mmap_offset, sizeof(mmap_offset));
3561 	if (err)
3562 		return err;
3563 
3564 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3565 			     &entry->page_idx, sizeof(entry->page_idx));
3566 	if (err)
3567 		return err;
3568 
3569 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3570 			     &length, sizeof(length));
3571 	return err;
3572 }
3573 
3574 DECLARE_UVERBS_NAMED_METHOD(
3575 	MLX5_IB_METHOD_UAR_OBJ_ALLOC,
3576 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE,
3577 			MLX5_IB_OBJECT_UAR,
3578 			UVERBS_ACCESS_NEW,
3579 			UA_MANDATORY),
3580 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE,
3581 			     enum mlx5_ib_uapi_uar_alloc_type,
3582 			     UA_MANDATORY),
3583 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3584 			   UVERBS_ATTR_TYPE(u32),
3585 			   UA_MANDATORY),
3586 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3587 			   UVERBS_ATTR_TYPE(u32),
3588 			   UA_MANDATORY),
3589 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3590 			    UVERBS_ATTR_TYPE(u64),
3591 			    UA_MANDATORY));
3592 
3593 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3594 	MLX5_IB_METHOD_UAR_OBJ_DESTROY,
3595 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE,
3596 			MLX5_IB_OBJECT_UAR,
3597 			UVERBS_ACCESS_DESTROY,
3598 			UA_MANDATORY));
3599 
3600 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR,
3601 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3602 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC),
3603 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY));
3604 
3605 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3606 	mlx5_ib_flow_action,
3607 	UVERBS_OBJECT_FLOW_ACTION,
3608 	UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
3609 	UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
3610 			     enum mlx5_ib_uapi_flow_action_flags));
3611 
3612 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3613 	mlx5_ib_query_context,
3614 	UVERBS_OBJECT_DEVICE,
3615 	UVERBS_METHOD_QUERY_CONTEXT,
3616 	UVERBS_ATTR_PTR_OUT(
3617 		MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
3618 		UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp,
3619 				   dump_fill_mkey),
3620 		UA_MANDATORY));
3621 
3622 static const struct uapi_definition mlx5_ib_defs[] = {
3623 	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
3624 	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
3625 	UAPI_DEF_CHAIN(mlx5_ib_qos_defs),
3626 	UAPI_DEF_CHAIN(mlx5_ib_std_types_defs),
3627 	UAPI_DEF_CHAIN(mlx5_ib_dm_defs),
3628 
3629 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_FLOW_ACTION,
3630 				&mlx5_ib_flow_action),
3631 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context),
3632 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
3633 				UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
3634 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR),
3635 	{}
3636 };
3637 
3638 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
3639 {
3640 	mlx5_ib_cleanup_multiport_master(dev);
3641 	WARN_ON(!xa_empty(&dev->odp_mkeys));
3642 	mutex_destroy(&dev->cap_mask_mutex);
3643 	WARN_ON(!xa_empty(&dev->sig_mrs));
3644 	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
3645 }
3646 
3647 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
3648 {
3649 	struct mlx5_core_dev *mdev = dev->mdev;
3650 	int err;
3651 	int i;
3652 
3653 	dev->ib_dev.node_type = RDMA_NODE_IB_CA;
3654 	dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
3655 	dev->ib_dev.phys_port_cnt = dev->num_ports;
3656 	dev->ib_dev.dev.parent = mdev->device;
3657 	dev->ib_dev.lag_flags = RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
3658 
3659 	for (i = 0; i < dev->num_ports; i++) {
3660 		spin_lock_init(&dev->port[i].mp.mpi_lock);
3661 		rwlock_init(&dev->port[i].roce.netdev_lock);
3662 		dev->port[i].roce.dev = dev;
3663 		dev->port[i].roce.native_port_num = i + 1;
3664 		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
3665 	}
3666 
3667 	err = mlx5_ib_init_multiport_master(dev);
3668 	if (err)
3669 		return err;
3670 
3671 	err = set_has_smi_cap(dev);
3672 	if (err)
3673 		goto err_mp;
3674 
3675 	err = mlx5_query_max_pkeys(&dev->ib_dev, &dev->pkey_table_len);
3676 	if (err)
3677 		goto err_mp;
3678 
3679 	if (mlx5_use_mad_ifc(dev))
3680 		get_ext_port_caps(dev);
3681 
3682 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_count(mdev);
3683 
3684 	mutex_init(&dev->cap_mask_mutex);
3685 	INIT_LIST_HEAD(&dev->qp_list);
3686 	spin_lock_init(&dev->reset_flow_resource_lock);
3687 	xa_init(&dev->odp_mkeys);
3688 	xa_init(&dev->sig_mrs);
3689 	atomic_set(&dev->mkey_var, 0);
3690 
3691 	spin_lock_init(&dev->dm.lock);
3692 	dev->dm.dev = mdev;
3693 	return 0;
3694 
3695 err_mp:
3696 	mlx5_ib_cleanup_multiport_master(dev);
3697 	return err;
3698 }
3699 
3700 static int mlx5_ib_enable_driver(struct ib_device *dev)
3701 {
3702 	struct mlx5_ib_dev *mdev = to_mdev(dev);
3703 	int ret;
3704 
3705 	ret = mlx5_ib_test_wc(mdev);
3706 	mlx5_ib_dbg(mdev, "Write-Combining %s",
3707 		    mdev->wc_support ? "supported" : "not supported");
3708 
3709 	return ret;
3710 }
3711 
3712 static const struct ib_device_ops mlx5_ib_dev_ops = {
3713 	.owner = THIS_MODULE,
3714 	.driver_id = RDMA_DRIVER_MLX5,
3715 	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
3716 
3717 	.add_gid = mlx5_ib_add_gid,
3718 	.alloc_mr = mlx5_ib_alloc_mr,
3719 	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
3720 	.alloc_pd = mlx5_ib_alloc_pd,
3721 	.alloc_ucontext = mlx5_ib_alloc_ucontext,
3722 	.attach_mcast = mlx5_ib_mcg_attach,
3723 	.check_mr_status = mlx5_ib_check_mr_status,
3724 	.create_ah = mlx5_ib_create_ah,
3725 	.create_cq = mlx5_ib_create_cq,
3726 	.create_qp = mlx5_ib_create_qp,
3727 	.create_srq = mlx5_ib_create_srq,
3728 	.create_user_ah = mlx5_ib_create_ah,
3729 	.dealloc_pd = mlx5_ib_dealloc_pd,
3730 	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
3731 	.del_gid = mlx5_ib_del_gid,
3732 	.dereg_mr = mlx5_ib_dereg_mr,
3733 	.destroy_ah = mlx5_ib_destroy_ah,
3734 	.destroy_cq = mlx5_ib_destroy_cq,
3735 	.destroy_qp = mlx5_ib_destroy_qp,
3736 	.destroy_srq = mlx5_ib_destroy_srq,
3737 	.detach_mcast = mlx5_ib_mcg_detach,
3738 	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
3739 	.drain_rq = mlx5_ib_drain_rq,
3740 	.drain_sq = mlx5_ib_drain_sq,
3741 	.enable_driver = mlx5_ib_enable_driver,
3742 	.get_dev_fw_str = get_dev_fw_str,
3743 	.get_dma_mr = mlx5_ib_get_dma_mr,
3744 	.get_link_layer = mlx5_ib_port_link_layer,
3745 	.map_mr_sg = mlx5_ib_map_mr_sg,
3746 	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
3747 	.mmap = mlx5_ib_mmap,
3748 	.mmap_free = mlx5_ib_mmap_free,
3749 	.modify_cq = mlx5_ib_modify_cq,
3750 	.modify_device = mlx5_ib_modify_device,
3751 	.modify_port = mlx5_ib_modify_port,
3752 	.modify_qp = mlx5_ib_modify_qp,
3753 	.modify_srq = mlx5_ib_modify_srq,
3754 	.poll_cq = mlx5_ib_poll_cq,
3755 	.post_recv = mlx5_ib_post_recv_nodrain,
3756 	.post_send = mlx5_ib_post_send_nodrain,
3757 	.post_srq_recv = mlx5_ib_post_srq_recv,
3758 	.process_mad = mlx5_ib_process_mad,
3759 	.query_ah = mlx5_ib_query_ah,
3760 	.query_device = mlx5_ib_query_device,
3761 	.query_gid = mlx5_ib_query_gid,
3762 	.query_pkey = mlx5_ib_query_pkey,
3763 	.query_qp = mlx5_ib_query_qp,
3764 	.query_srq = mlx5_ib_query_srq,
3765 	.query_ucontext = mlx5_ib_query_ucontext,
3766 	.reg_user_mr = mlx5_ib_reg_user_mr,
3767 	.reg_user_mr_dmabuf = mlx5_ib_reg_user_mr_dmabuf,
3768 	.req_notify_cq = mlx5_ib_arm_cq,
3769 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
3770 	.resize_cq = mlx5_ib_resize_cq,
3771 
3772 	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
3773 	INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs),
3774 	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
3775 	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
3776 	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
3777 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
3778 };
3779 
3780 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
3781 	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
3782 };
3783 
3784 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
3785 	.get_vf_config = mlx5_ib_get_vf_config,
3786 	.get_vf_guid = mlx5_ib_get_vf_guid,
3787 	.get_vf_stats = mlx5_ib_get_vf_stats,
3788 	.set_vf_guid = mlx5_ib_set_vf_guid,
3789 	.set_vf_link_state = mlx5_ib_set_vf_link_state,
3790 };
3791 
3792 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
3793 	.alloc_mw = mlx5_ib_alloc_mw,
3794 	.dealloc_mw = mlx5_ib_dealloc_mw,
3795 
3796 	INIT_RDMA_OBJ_SIZE(ib_mw, mlx5_ib_mw, ibmw),
3797 };
3798 
3799 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
3800 	.alloc_xrcd = mlx5_ib_alloc_xrcd,
3801 	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
3802 
3803 	INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd),
3804 };
3805 
3806 static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
3807 {
3808 	struct mlx5_core_dev *mdev = dev->mdev;
3809 	struct mlx5_var_table *var_table = &dev->var_table;
3810 	u8 log_doorbell_bar_size;
3811 	u8 log_doorbell_stride;
3812 	u64 bar_size;
3813 
3814 	log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3815 					log_doorbell_bar_size);
3816 	log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3817 					log_doorbell_stride);
3818 	var_table->hw_start_addr = dev->mdev->bar_addr +
3819 				MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
3820 					doorbell_bar_offset);
3821 	bar_size = (1ULL << log_doorbell_bar_size) * 4096;
3822 	var_table->stride_size = 1ULL << log_doorbell_stride;
3823 	var_table->num_var_hw_entries = div_u64(bar_size,
3824 						var_table->stride_size);
3825 	mutex_init(&var_table->bitmap_lock);
3826 	var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
3827 					  GFP_KERNEL);
3828 	return (var_table->bitmap) ? 0 : -ENOMEM;
3829 }
3830 
3831 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
3832 {
3833 	bitmap_free(dev->var_table.bitmap);
3834 }
3835 
3836 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
3837 {
3838 	struct mlx5_core_dev *mdev = dev->mdev;
3839 	int err;
3840 
3841 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
3842 	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
3843 		ib_set_device_ops(&dev->ib_dev,
3844 				  &mlx5_ib_dev_ipoib_enhanced_ops);
3845 
3846 	if (mlx5_core_is_pf(mdev))
3847 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
3848 
3849 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
3850 
3851 	if (MLX5_CAP_GEN(mdev, imaicl))
3852 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
3853 
3854 	if (MLX5_CAP_GEN(mdev, xrc))
3855 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
3856 
3857 	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
3858 	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3859 	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
3860 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
3861 
3862 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
3863 
3864 	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
3865 		dev->ib_dev.driver_def = mlx5_ib_defs;
3866 
3867 	err = init_node_data(dev);
3868 	if (err)
3869 		return err;
3870 
3871 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
3872 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
3873 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
3874 		mutex_init(&dev->lb.mutex);
3875 
3876 	if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3877 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
3878 		err = mlx5_ib_init_var_table(dev);
3879 		if (err)
3880 			return err;
3881 	}
3882 
3883 	dev->ib_dev.use_cq_dim = true;
3884 
3885 	return 0;
3886 }
3887 
3888 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
3889 	.get_port_immutable = mlx5_port_immutable,
3890 	.query_port = mlx5_ib_query_port,
3891 };
3892 
3893 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
3894 {
3895 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
3896 	return 0;
3897 }
3898 
3899 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
3900 	.get_port_immutable = mlx5_port_rep_immutable,
3901 	.query_port = mlx5_ib_rep_query_port,
3902 	.query_pkey = mlx5_ib_rep_query_pkey,
3903 };
3904 
3905 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
3906 {
3907 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
3908 	return 0;
3909 }
3910 
3911 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
3912 	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
3913 	.create_wq = mlx5_ib_create_wq,
3914 	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
3915 	.destroy_wq = mlx5_ib_destroy_wq,
3916 	.get_netdev = mlx5_ib_get_netdev,
3917 	.modify_wq = mlx5_ib_modify_wq,
3918 
3919 	INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table,
3920 			   ib_rwq_ind_tbl),
3921 };
3922 
3923 static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev)
3924 {
3925 	struct mlx5_core_dev *mdev = dev->mdev;
3926 	enum rdma_link_layer ll;
3927 	int port_type_cap;
3928 	u32 port_num = 0;
3929 	int err;
3930 
3931 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3932 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3933 
3934 	if (ll == IB_LINK_LAYER_ETHERNET) {
3935 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
3936 
3937 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3938 
3939 		/* Register only for native ports */
3940 		err = mlx5_add_netdev_notifier(dev, port_num);
3941 		if (err || dev->is_rep || !mlx5_is_roce_init_enabled(mdev))
3942 			/*
3943 			 * We don't enable ETH interface for
3944 			 * 1. IB representors
3945 			 * 2. User disabled ROCE through devlink interface
3946 			 */
3947 			return err;
3948 
3949 		err = mlx5_enable_eth(dev);
3950 		if (err)
3951 			goto cleanup;
3952 	}
3953 
3954 	return 0;
3955 cleanup:
3956 	mlx5_remove_netdev_notifier(dev, port_num);
3957 	return err;
3958 }
3959 
3960 static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev)
3961 {
3962 	struct mlx5_core_dev *mdev = dev->mdev;
3963 	enum rdma_link_layer ll;
3964 	int port_type_cap;
3965 	u32 port_num;
3966 
3967 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3968 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3969 
3970 	if (ll == IB_LINK_LAYER_ETHERNET) {
3971 		if (!dev->is_rep)
3972 			mlx5_disable_eth(dev);
3973 
3974 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3975 		mlx5_remove_netdev_notifier(dev, port_num);
3976 	}
3977 }
3978 
3979 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
3980 {
3981 	mlx5_ib_init_cong_debugfs(dev,
3982 				  mlx5_core_native_port_num(dev->mdev) - 1);
3983 	return 0;
3984 }
3985 
3986 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
3987 {
3988 	mlx5_ib_cleanup_cong_debugfs(dev,
3989 				     mlx5_core_native_port_num(dev->mdev) - 1);
3990 }
3991 
3992 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
3993 {
3994 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
3995 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
3996 }
3997 
3998 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
3999 {
4000 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
4001 }
4002 
4003 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
4004 {
4005 	int err;
4006 
4007 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
4008 	if (err)
4009 		return err;
4010 
4011 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
4012 	if (err)
4013 		mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4014 
4015 	return err;
4016 }
4017 
4018 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
4019 {
4020 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4021 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4022 }
4023 
4024 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
4025 {
4026 	const char *name;
4027 
4028 	rdma_set_device_sysfs_group(&dev->ib_dev, &mlx5_attr_group);
4029 	if (!mlx5_lag_is_roce(dev->mdev))
4030 		name = "mlx5_%d";
4031 	else
4032 		name = "mlx5_bond_%d";
4033 	return ib_register_device(&dev->ib_dev, name, &dev->mdev->pdev->dev);
4034 }
4035 
4036 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
4037 {
4038 	int err;
4039 
4040 	err = mlx5_mr_cache_cleanup(dev);
4041 	if (err)
4042 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
4043 
4044 	if (dev->umrc.qp)
4045 		mlx5_ib_destroy_qp(dev->umrc.qp, NULL);
4046 	if (dev->umrc.cq)
4047 		ib_free_cq(dev->umrc.cq);
4048 	if (dev->umrc.pd)
4049 		ib_dealloc_pd(dev->umrc.pd);
4050 }
4051 
4052 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
4053 {
4054 	ib_unregister_device(&dev->ib_dev);
4055 }
4056 
4057 enum {
4058 	MAX_UMR_WR = 128,
4059 };
4060 
4061 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
4062 {
4063 	struct ib_qp_init_attr *init_attr = NULL;
4064 	struct ib_qp_attr *attr = NULL;
4065 	struct ib_pd *pd;
4066 	struct ib_cq *cq;
4067 	struct ib_qp *qp;
4068 	int ret;
4069 
4070 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
4071 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
4072 	if (!attr || !init_attr) {
4073 		ret = -ENOMEM;
4074 		goto error_0;
4075 	}
4076 
4077 	pd = ib_alloc_pd(&dev->ib_dev, 0);
4078 	if (IS_ERR(pd)) {
4079 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
4080 		ret = PTR_ERR(pd);
4081 		goto error_0;
4082 	}
4083 
4084 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
4085 	if (IS_ERR(cq)) {
4086 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
4087 		ret = PTR_ERR(cq);
4088 		goto error_2;
4089 	}
4090 
4091 	init_attr->send_cq = cq;
4092 	init_attr->recv_cq = cq;
4093 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
4094 	init_attr->cap.max_send_wr = MAX_UMR_WR;
4095 	init_attr->cap.max_send_sge = 1;
4096 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
4097 	init_attr->port_num = 1;
4098 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
4099 	if (IS_ERR(qp)) {
4100 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
4101 		ret = PTR_ERR(qp);
4102 		goto error_3;
4103 	}
4104 	qp->device     = &dev->ib_dev;
4105 	qp->real_qp    = qp;
4106 	qp->uobject    = NULL;
4107 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
4108 	qp->send_cq    = init_attr->send_cq;
4109 	qp->recv_cq    = init_attr->recv_cq;
4110 
4111 	attr->qp_state = IB_QPS_INIT;
4112 	attr->port_num = 1;
4113 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
4114 				IB_QP_PORT, NULL);
4115 	if (ret) {
4116 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
4117 		goto error_4;
4118 	}
4119 
4120 	memset(attr, 0, sizeof(*attr));
4121 	attr->qp_state = IB_QPS_RTR;
4122 	attr->path_mtu = IB_MTU_256;
4123 
4124 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4125 	if (ret) {
4126 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
4127 		goto error_4;
4128 	}
4129 
4130 	memset(attr, 0, sizeof(*attr));
4131 	attr->qp_state = IB_QPS_RTS;
4132 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4133 	if (ret) {
4134 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
4135 		goto error_4;
4136 	}
4137 
4138 	dev->umrc.qp = qp;
4139 	dev->umrc.cq = cq;
4140 	dev->umrc.pd = pd;
4141 
4142 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
4143 	ret = mlx5_mr_cache_init(dev);
4144 	if (ret) {
4145 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4146 		goto error_4;
4147 	}
4148 
4149 	kfree(attr);
4150 	kfree(init_attr);
4151 
4152 	return 0;
4153 
4154 error_4:
4155 	mlx5_ib_destroy_qp(qp, NULL);
4156 	dev->umrc.qp = NULL;
4157 
4158 error_3:
4159 	ib_free_cq(cq);
4160 	dev->umrc.cq = NULL;
4161 
4162 error_2:
4163 	ib_dealloc_pd(pd);
4164 	dev->umrc.pd = NULL;
4165 
4166 error_0:
4167 	kfree(attr);
4168 	kfree(init_attr);
4169 	return ret;
4170 }
4171 
4172 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
4173 {
4174 	struct dentry *root;
4175 
4176 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4177 		return 0;
4178 
4179 	mutex_init(&dev->delay_drop.lock);
4180 	dev->delay_drop.dev = dev;
4181 	dev->delay_drop.activate = false;
4182 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4183 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4184 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4185 	atomic_set(&dev->delay_drop.events_cnt, 0);
4186 
4187 	if (!mlx5_debugfs_root)
4188 		return 0;
4189 
4190 	root = debugfs_create_dir("delay_drop", dev->mdev->priv.dbg_root);
4191 	dev->delay_drop.dir_debugfs = root;
4192 
4193 	debugfs_create_atomic_t("num_timeout_events", 0400, root,
4194 				&dev->delay_drop.events_cnt);
4195 	debugfs_create_atomic_t("num_rqs", 0400, root,
4196 				&dev->delay_drop.rqs_cnt);
4197 	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
4198 			    &fops_delay_drop_timeout);
4199 	return 0;
4200 }
4201 
4202 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
4203 {
4204 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4205 		return;
4206 
4207 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4208 	if (!dev->delay_drop.dir_debugfs)
4209 		return;
4210 
4211 	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
4212 	dev->delay_drop.dir_debugfs = NULL;
4213 }
4214 
4215 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
4216 {
4217 	dev->mdev_events.notifier_call = mlx5_ib_event;
4218 	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
4219 	return 0;
4220 }
4221 
4222 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
4223 {
4224 	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
4225 }
4226 
4227 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
4228 		      const struct mlx5_ib_profile *profile,
4229 		      int stage)
4230 {
4231 	dev->ib_active = false;
4232 
4233 	/* Number of stages to cleanup */
4234 	while (stage) {
4235 		stage--;
4236 		if (profile->stage[stage].cleanup)
4237 			profile->stage[stage].cleanup(dev);
4238 	}
4239 
4240 	kfree(dev->port);
4241 	ib_dealloc_device(&dev->ib_dev);
4242 }
4243 
4244 int __mlx5_ib_add(struct mlx5_ib_dev *dev,
4245 		  const struct mlx5_ib_profile *profile)
4246 {
4247 	int err;
4248 	int i;
4249 
4250 	dev->profile = profile;
4251 
4252 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
4253 		if (profile->stage[i].init) {
4254 			err = profile->stage[i].init(dev);
4255 			if (err)
4256 				goto err_out;
4257 		}
4258 	}
4259 
4260 	dev->ib_active = true;
4261 	return 0;
4262 
4263 err_out:
4264 	/* Clean up stages which were initialized */
4265 	while (i) {
4266 		i--;
4267 		if (profile->stage[i].cleanup)
4268 			profile->stage[i].cleanup(dev);
4269 	}
4270 	return -ENOMEM;
4271 }
4272 
4273 static const struct mlx5_ib_profile pf_profile = {
4274 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4275 		     mlx5_ib_stage_init_init,
4276 		     mlx5_ib_stage_init_cleanup),
4277 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4278 		     mlx5_ib_fs_init,
4279 		     mlx5_ib_fs_cleanup),
4280 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4281 		     mlx5_ib_stage_caps_init,
4282 		     mlx5_ib_stage_caps_cleanup),
4283 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4284 		     mlx5_ib_stage_non_default_cb,
4285 		     NULL),
4286 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4287 		     mlx5_ib_roce_init,
4288 		     mlx5_ib_roce_cleanup),
4289 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4290 		     mlx5_init_qp_table,
4291 		     mlx5_cleanup_qp_table),
4292 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4293 		     mlx5_init_srq_table,
4294 		     mlx5_cleanup_srq_table),
4295 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4296 		     mlx5_ib_dev_res_init,
4297 		     mlx5_ib_dev_res_cleanup),
4298 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4299 		     mlx5_ib_stage_dev_notifier_init,
4300 		     mlx5_ib_stage_dev_notifier_cleanup),
4301 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
4302 		     mlx5_ib_odp_init_one,
4303 		     mlx5_ib_odp_cleanup_one),
4304 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4305 		     mlx5_ib_counters_init,
4306 		     mlx5_ib_counters_cleanup),
4307 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4308 		     mlx5_ib_stage_cong_debugfs_init,
4309 		     mlx5_ib_stage_cong_debugfs_cleanup),
4310 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4311 		     mlx5_ib_stage_uar_init,
4312 		     mlx5_ib_stage_uar_cleanup),
4313 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4314 		     mlx5_ib_stage_bfrag_init,
4315 		     mlx5_ib_stage_bfrag_cleanup),
4316 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4317 		     NULL,
4318 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4319 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4320 		     mlx5_ib_devx_init,
4321 		     mlx5_ib_devx_cleanup),
4322 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4323 		     mlx5_ib_stage_ib_reg_init,
4324 		     mlx5_ib_stage_ib_reg_cleanup),
4325 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4326 		     mlx5_ib_stage_post_ib_reg_umr_init,
4327 		     NULL),
4328 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4329 		     mlx5_ib_stage_delay_drop_init,
4330 		     mlx5_ib_stage_delay_drop_cleanup),
4331 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4332 		     mlx5_ib_restrack_init,
4333 		     NULL),
4334 };
4335 
4336 const struct mlx5_ib_profile raw_eth_profile = {
4337 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4338 		     mlx5_ib_stage_init_init,
4339 		     mlx5_ib_stage_init_cleanup),
4340 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4341 		     mlx5_ib_fs_init,
4342 		     mlx5_ib_fs_cleanup),
4343 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4344 		     mlx5_ib_stage_caps_init,
4345 		     mlx5_ib_stage_caps_cleanup),
4346 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4347 		     mlx5_ib_stage_raw_eth_non_default_cb,
4348 		     NULL),
4349 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4350 		     mlx5_ib_roce_init,
4351 		     mlx5_ib_roce_cleanup),
4352 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4353 		     mlx5_init_qp_table,
4354 		     mlx5_cleanup_qp_table),
4355 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4356 		     mlx5_init_srq_table,
4357 		     mlx5_cleanup_srq_table),
4358 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4359 		     mlx5_ib_dev_res_init,
4360 		     mlx5_ib_dev_res_cleanup),
4361 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4362 		     mlx5_ib_stage_dev_notifier_init,
4363 		     mlx5_ib_stage_dev_notifier_cleanup),
4364 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4365 		     mlx5_ib_counters_init,
4366 		     mlx5_ib_counters_cleanup),
4367 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4368 		     mlx5_ib_stage_cong_debugfs_init,
4369 		     mlx5_ib_stage_cong_debugfs_cleanup),
4370 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4371 		     mlx5_ib_stage_uar_init,
4372 		     mlx5_ib_stage_uar_cleanup),
4373 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4374 		     mlx5_ib_stage_bfrag_init,
4375 		     mlx5_ib_stage_bfrag_cleanup),
4376 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4377 		     NULL,
4378 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4379 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4380 		     mlx5_ib_devx_init,
4381 		     mlx5_ib_devx_cleanup),
4382 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4383 		     mlx5_ib_stage_ib_reg_init,
4384 		     mlx5_ib_stage_ib_reg_cleanup),
4385 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4386 		     mlx5_ib_stage_post_ib_reg_umr_init,
4387 		     NULL),
4388 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4389 		     mlx5_ib_restrack_init,
4390 		     NULL),
4391 };
4392 
4393 static int mlx5r_mp_probe(struct auxiliary_device *adev,
4394 			  const struct auxiliary_device_id *id)
4395 {
4396 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4397 	struct mlx5_core_dev *mdev = idev->mdev;
4398 	struct mlx5_ib_multiport_info *mpi;
4399 	struct mlx5_ib_dev *dev;
4400 	bool bound = false;
4401 	int err;
4402 
4403 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4404 	if (!mpi)
4405 		return -ENOMEM;
4406 
4407 	mpi->mdev = mdev;
4408 	err = mlx5_query_nic_vport_system_image_guid(mdev,
4409 						     &mpi->sys_image_guid);
4410 	if (err) {
4411 		kfree(mpi);
4412 		return err;
4413 	}
4414 
4415 	mutex_lock(&mlx5_ib_multiport_mutex);
4416 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
4417 		if (dev->sys_image_guid == mpi->sys_image_guid)
4418 			bound = mlx5_ib_bind_slave_port(dev, mpi);
4419 
4420 		if (bound) {
4421 			rdma_roce_rescan_device(&dev->ib_dev);
4422 			mpi->ibdev->ib_active = true;
4423 			break;
4424 		}
4425 	}
4426 
4427 	if (!bound) {
4428 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4429 		dev_dbg(mdev->device,
4430 			"no suitable IB device found to bind to, added to unaffiliated list.\n");
4431 	}
4432 	mutex_unlock(&mlx5_ib_multiport_mutex);
4433 
4434 	dev_set_drvdata(&adev->dev, mpi);
4435 	return 0;
4436 }
4437 
4438 static void mlx5r_mp_remove(struct auxiliary_device *adev)
4439 {
4440 	struct mlx5_ib_multiport_info *mpi;
4441 
4442 	mpi = dev_get_drvdata(&adev->dev);
4443 	mutex_lock(&mlx5_ib_multiport_mutex);
4444 	if (mpi->ibdev)
4445 		mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
4446 	list_del(&mpi->list);
4447 	mutex_unlock(&mlx5_ib_multiport_mutex);
4448 	kfree(mpi);
4449 }
4450 
4451 static int mlx5r_probe(struct auxiliary_device *adev,
4452 		       const struct auxiliary_device_id *id)
4453 {
4454 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4455 	struct mlx5_core_dev *mdev = idev->mdev;
4456 	const struct mlx5_ib_profile *profile;
4457 	int port_type_cap, num_ports, ret;
4458 	enum rdma_link_layer ll;
4459 	struct mlx5_ib_dev *dev;
4460 
4461 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4462 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4463 
4464 	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
4465 			MLX5_CAP_GEN(mdev, num_vhca_ports));
4466 	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
4467 	if (!dev)
4468 		return -ENOMEM;
4469 	dev->port = kcalloc(num_ports, sizeof(*dev->port),
4470 			     GFP_KERNEL);
4471 	if (!dev->port) {
4472 		ib_dealloc_device(&dev->ib_dev);
4473 		return -ENOMEM;
4474 	}
4475 
4476 	dev->mdev = mdev;
4477 	dev->num_ports = num_ports;
4478 
4479 	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_is_roce_init_enabled(mdev))
4480 		profile = &raw_eth_profile;
4481 	else
4482 		profile = &pf_profile;
4483 
4484 	ret = __mlx5_ib_add(dev, profile);
4485 	if (ret) {
4486 		kfree(dev->port);
4487 		ib_dealloc_device(&dev->ib_dev);
4488 		return ret;
4489 	}
4490 
4491 	dev_set_drvdata(&adev->dev, dev);
4492 	return 0;
4493 }
4494 
4495 static void mlx5r_remove(struct auxiliary_device *adev)
4496 {
4497 	struct mlx5_ib_dev *dev;
4498 
4499 	dev = dev_get_drvdata(&adev->dev);
4500 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
4501 }
4502 
4503 static const struct auxiliary_device_id mlx5r_mp_id_table[] = {
4504 	{ .name = MLX5_ADEV_NAME ".multiport", },
4505 	{},
4506 };
4507 
4508 static const struct auxiliary_device_id mlx5r_id_table[] = {
4509 	{ .name = MLX5_ADEV_NAME ".rdma", },
4510 	{},
4511 };
4512 
4513 MODULE_DEVICE_TABLE(auxiliary, mlx5r_mp_id_table);
4514 MODULE_DEVICE_TABLE(auxiliary, mlx5r_id_table);
4515 
4516 static struct auxiliary_driver mlx5r_mp_driver = {
4517 	.name = "multiport",
4518 	.probe = mlx5r_mp_probe,
4519 	.remove = mlx5r_mp_remove,
4520 	.id_table = mlx5r_mp_id_table,
4521 };
4522 
4523 static struct auxiliary_driver mlx5r_driver = {
4524 	.name = "rdma",
4525 	.probe = mlx5r_probe,
4526 	.remove = mlx5r_remove,
4527 	.id_table = mlx5r_id_table,
4528 };
4529 
4530 static int __init mlx5_ib_init(void)
4531 {
4532 	int ret;
4533 
4534 	xlt_emergency_page = (void *)__get_free_page(GFP_KERNEL);
4535 	if (!xlt_emergency_page)
4536 		return -ENOMEM;
4537 
4538 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
4539 	if (!mlx5_ib_event_wq) {
4540 		free_page((unsigned long)xlt_emergency_page);
4541 		return -ENOMEM;
4542 	}
4543 
4544 	mlx5_ib_odp_init();
4545 	ret = mlx5r_rep_init();
4546 	if (ret)
4547 		goto rep_err;
4548 	ret = auxiliary_driver_register(&mlx5r_mp_driver);
4549 	if (ret)
4550 		goto mp_err;
4551 	ret = auxiliary_driver_register(&mlx5r_driver);
4552 	if (ret)
4553 		goto drv_err;
4554 	return 0;
4555 
4556 drv_err:
4557 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4558 mp_err:
4559 	mlx5r_rep_cleanup();
4560 rep_err:
4561 	destroy_workqueue(mlx5_ib_event_wq);
4562 	free_page((unsigned long)xlt_emergency_page);
4563 	return ret;
4564 }
4565 
4566 static void __exit mlx5_ib_cleanup(void)
4567 {
4568 	auxiliary_driver_unregister(&mlx5r_driver);
4569 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4570 	mlx5r_rep_cleanup();
4571 
4572 	destroy_workqueue(mlx5_ib_event_wq);
4573 	free_page((unsigned long)xlt_emergency_page);
4574 }
4575 
4576 module_init(mlx5_ib_init);
4577 module_exit(mlx5_ib_cleanup);
4578