xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision c4a11bf4)
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2013-2020, Mellanox Technologies inc. All rights reserved.
4  * Copyright (c) 2020, Intel Corporation. All rights reserved.
5  */
6 
7 #include <linux/debugfs.h>
8 #include <linux/highmem.h>
9 #include <linux/module.h>
10 #include <linux/init.h>
11 #include <linux/errno.h>
12 #include <linux/pci.h>
13 #include <linux/dma-mapping.h>
14 #include <linux/slab.h>
15 #include <linux/bitmap.h>
16 #include <linux/sched.h>
17 #include <linux/sched/mm.h>
18 #include <linux/sched/task.h>
19 #include <linux/delay.h>
20 #include <rdma/ib_user_verbs.h>
21 #include <rdma/ib_addr.h>
22 #include <rdma/ib_cache.h>
23 #include <linux/mlx5/port.h>
24 #include <linux/mlx5/vport.h>
25 #include <linux/mlx5/fs.h>
26 #include <linux/mlx5/eswitch.h>
27 #include <linux/list.h>
28 #include <rdma/ib_smi.h>
29 #include <rdma/ib_umem.h>
30 #include <rdma/lag.h>
31 #include <linux/in.h>
32 #include <linux/etherdevice.h>
33 #include "mlx5_ib.h"
34 #include "ib_rep.h"
35 #include "cmd.h"
36 #include "devx.h"
37 #include "dm.h"
38 #include "fs.h"
39 #include "srq.h"
40 #include "qp.h"
41 #include "wr.h"
42 #include "restrack.h"
43 #include "counters.h"
44 #include <linux/mlx5/accel.h>
45 #include <rdma/uverbs_std_types.h>
46 #include <rdma/uverbs_ioctl.h>
47 #include <rdma/mlx5_user_ioctl_verbs.h>
48 #include <rdma/mlx5_user_ioctl_cmds.h>
49 #include <rdma/ib_umem_odp.h>
50 
51 #define UVERBS_MODULE_NAME mlx5_ib
52 #include <rdma/uverbs_named_ioctl.h>
53 
54 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
55 MODULE_DESCRIPTION("Mellanox 5th generation network adapters (ConnectX series) IB driver");
56 MODULE_LICENSE("Dual BSD/GPL");
57 
58 struct mlx5_ib_event_work {
59 	struct work_struct	work;
60 	union {
61 		struct mlx5_ib_dev	      *dev;
62 		struct mlx5_ib_multiport_info *mpi;
63 	};
64 	bool			is_slave;
65 	unsigned int		event;
66 	void			*param;
67 };
68 
69 enum {
70 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
71 };
72 
73 static struct workqueue_struct *mlx5_ib_event_wq;
74 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
75 static LIST_HEAD(mlx5_ib_dev_list);
76 /*
77  * This mutex should be held when accessing either of the above lists
78  */
79 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
80 
81 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
82 {
83 	struct mlx5_ib_dev *dev;
84 
85 	mutex_lock(&mlx5_ib_multiport_mutex);
86 	dev = mpi->ibdev;
87 	mutex_unlock(&mlx5_ib_multiport_mutex);
88 	return dev;
89 }
90 
91 static enum rdma_link_layer
92 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
93 {
94 	switch (port_type_cap) {
95 	case MLX5_CAP_PORT_TYPE_IB:
96 		return IB_LINK_LAYER_INFINIBAND;
97 	case MLX5_CAP_PORT_TYPE_ETH:
98 		return IB_LINK_LAYER_ETHERNET;
99 	default:
100 		return IB_LINK_LAYER_UNSPECIFIED;
101 	}
102 }
103 
104 static enum rdma_link_layer
105 mlx5_ib_port_link_layer(struct ib_device *device, u32 port_num)
106 {
107 	struct mlx5_ib_dev *dev = to_mdev(device);
108 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
109 
110 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
111 }
112 
113 static int get_port_state(struct ib_device *ibdev,
114 			  u32 port_num,
115 			  enum ib_port_state *state)
116 {
117 	struct ib_port_attr attr;
118 	int ret;
119 
120 	memset(&attr, 0, sizeof(attr));
121 	ret = ibdev->ops.query_port(ibdev, port_num, &attr);
122 	if (!ret)
123 		*state = attr.state;
124 	return ret;
125 }
126 
127 static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev,
128 					   struct net_device *ndev,
129 					   struct net_device *upper,
130 					   u32 *port_num)
131 {
132 	struct net_device *rep_ndev;
133 	struct mlx5_ib_port *port;
134 	int i;
135 
136 	for (i = 0; i < dev->num_ports; i++) {
137 		port  = &dev->port[i];
138 		if (!port->rep)
139 			continue;
140 
141 		if (upper == ndev && port->rep->vport == MLX5_VPORT_UPLINK) {
142 			*port_num = i + 1;
143 			return &port->roce;
144 		}
145 
146 		if (upper && port->rep->vport == MLX5_VPORT_UPLINK)
147 			continue;
148 
149 		read_lock(&port->roce.netdev_lock);
150 		rep_ndev = mlx5_ib_get_rep_netdev(port->rep->esw,
151 						  port->rep->vport);
152 		if (rep_ndev == ndev) {
153 			read_unlock(&port->roce.netdev_lock);
154 			*port_num = i + 1;
155 			return &port->roce;
156 		}
157 		read_unlock(&port->roce.netdev_lock);
158 	}
159 
160 	return NULL;
161 }
162 
163 static int mlx5_netdev_event(struct notifier_block *this,
164 			     unsigned long event, void *ptr)
165 {
166 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
167 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
168 	u32 port_num = roce->native_port_num;
169 	struct mlx5_core_dev *mdev;
170 	struct mlx5_ib_dev *ibdev;
171 
172 	ibdev = roce->dev;
173 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
174 	if (!mdev)
175 		return NOTIFY_DONE;
176 
177 	switch (event) {
178 	case NETDEV_REGISTER:
179 		/* Should already be registered during the load */
180 		if (ibdev->is_rep)
181 			break;
182 		write_lock(&roce->netdev_lock);
183 		if (ndev->dev.parent == mdev->device)
184 			roce->netdev = ndev;
185 		write_unlock(&roce->netdev_lock);
186 		break;
187 
188 	case NETDEV_UNREGISTER:
189 		/* In case of reps, ib device goes away before the netdevs */
190 		write_lock(&roce->netdev_lock);
191 		if (roce->netdev == ndev)
192 			roce->netdev = NULL;
193 		write_unlock(&roce->netdev_lock);
194 		break;
195 
196 	case NETDEV_CHANGE:
197 	case NETDEV_UP:
198 	case NETDEV_DOWN: {
199 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
200 		struct net_device *upper = NULL;
201 
202 		if (lag_ndev) {
203 			upper = netdev_master_upper_dev_get(lag_ndev);
204 			dev_put(lag_ndev);
205 		}
206 
207 		if (ibdev->is_rep)
208 			roce = mlx5_get_rep_roce(ibdev, ndev, upper, &port_num);
209 		if (!roce)
210 			return NOTIFY_DONE;
211 		if ((upper == ndev ||
212 		     ((!upper || ibdev->is_rep) && ndev == roce->netdev)) &&
213 		    ibdev->ib_active) {
214 			struct ib_event ibev = { };
215 			enum ib_port_state port_state;
216 
217 			if (get_port_state(&ibdev->ib_dev, port_num,
218 					   &port_state))
219 				goto done;
220 
221 			if (roce->last_port_state == port_state)
222 				goto done;
223 
224 			roce->last_port_state = port_state;
225 			ibev.device = &ibdev->ib_dev;
226 			if (port_state == IB_PORT_DOWN)
227 				ibev.event = IB_EVENT_PORT_ERR;
228 			else if (port_state == IB_PORT_ACTIVE)
229 				ibev.event = IB_EVENT_PORT_ACTIVE;
230 			else
231 				goto done;
232 
233 			ibev.element.port_num = port_num;
234 			ib_dispatch_event(&ibev);
235 		}
236 		break;
237 	}
238 
239 	default:
240 		break;
241 	}
242 done:
243 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
244 	return NOTIFY_DONE;
245 }
246 
247 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
248 					     u32 port_num)
249 {
250 	struct mlx5_ib_dev *ibdev = to_mdev(device);
251 	struct net_device *ndev;
252 	struct mlx5_core_dev *mdev;
253 
254 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
255 	if (!mdev)
256 		return NULL;
257 
258 	ndev = mlx5_lag_get_roce_netdev(mdev);
259 	if (ndev)
260 		goto out;
261 
262 	/* Ensure ndev does not disappear before we invoke dev_hold()
263 	 */
264 	read_lock(&ibdev->port[port_num - 1].roce.netdev_lock);
265 	ndev = ibdev->port[port_num - 1].roce.netdev;
266 	if (ndev)
267 		dev_hold(ndev);
268 	read_unlock(&ibdev->port[port_num - 1].roce.netdev_lock);
269 
270 out:
271 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
272 	return ndev;
273 }
274 
275 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
276 						   u32 ib_port_num,
277 						   u32 *native_port_num)
278 {
279 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
280 							  ib_port_num);
281 	struct mlx5_core_dev *mdev = NULL;
282 	struct mlx5_ib_multiport_info *mpi;
283 	struct mlx5_ib_port *port;
284 
285 	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
286 	    ll != IB_LINK_LAYER_ETHERNET) {
287 		if (native_port_num)
288 			*native_port_num = ib_port_num;
289 		return ibdev->mdev;
290 	}
291 
292 	if (native_port_num)
293 		*native_port_num = 1;
294 
295 	port = &ibdev->port[ib_port_num - 1];
296 	spin_lock(&port->mp.mpi_lock);
297 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
298 	if (mpi && !mpi->unaffiliate) {
299 		mdev = mpi->mdev;
300 		/* If it's the master no need to refcount, it'll exist
301 		 * as long as the ib_dev exists.
302 		 */
303 		if (!mpi->is_master)
304 			mpi->mdev_refcnt++;
305 	}
306 	spin_unlock(&port->mp.mpi_lock);
307 
308 	return mdev;
309 }
310 
311 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u32 port_num)
312 {
313 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
314 							  port_num);
315 	struct mlx5_ib_multiport_info *mpi;
316 	struct mlx5_ib_port *port;
317 
318 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
319 		return;
320 
321 	port = &ibdev->port[port_num - 1];
322 
323 	spin_lock(&port->mp.mpi_lock);
324 	mpi = ibdev->port[port_num - 1].mp.mpi;
325 	if (mpi->is_master)
326 		goto out;
327 
328 	mpi->mdev_refcnt--;
329 	if (mpi->unaffiliate)
330 		complete(&mpi->unref_comp);
331 out:
332 	spin_unlock(&port->mp.mpi_lock);
333 }
334 
335 static int translate_eth_legacy_proto_oper(u32 eth_proto_oper,
336 					   u16 *active_speed, u8 *active_width)
337 {
338 	switch (eth_proto_oper) {
339 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
340 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
341 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
342 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
343 		*active_width = IB_WIDTH_1X;
344 		*active_speed = IB_SPEED_SDR;
345 		break;
346 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
347 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
348 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
349 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
350 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
351 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
352 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
353 		*active_width = IB_WIDTH_1X;
354 		*active_speed = IB_SPEED_QDR;
355 		break;
356 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
357 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
358 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
359 		*active_width = IB_WIDTH_1X;
360 		*active_speed = IB_SPEED_EDR;
361 		break;
362 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
363 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
364 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
365 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
366 		*active_width = IB_WIDTH_4X;
367 		*active_speed = IB_SPEED_QDR;
368 		break;
369 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
370 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
371 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
372 		*active_width = IB_WIDTH_1X;
373 		*active_speed = IB_SPEED_HDR;
374 		break;
375 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
376 		*active_width = IB_WIDTH_4X;
377 		*active_speed = IB_SPEED_FDR;
378 		break;
379 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
380 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
381 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
382 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
383 		*active_width = IB_WIDTH_4X;
384 		*active_speed = IB_SPEED_EDR;
385 		break;
386 	default:
387 		return -EINVAL;
388 	}
389 
390 	return 0;
391 }
392 
393 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u16 *active_speed,
394 					u8 *active_width)
395 {
396 	switch (eth_proto_oper) {
397 	case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
398 	case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
399 		*active_width = IB_WIDTH_1X;
400 		*active_speed = IB_SPEED_SDR;
401 		break;
402 	case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
403 		*active_width = IB_WIDTH_1X;
404 		*active_speed = IB_SPEED_DDR;
405 		break;
406 	case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
407 		*active_width = IB_WIDTH_1X;
408 		*active_speed = IB_SPEED_QDR;
409 		break;
410 	case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
411 		*active_width = IB_WIDTH_4X;
412 		*active_speed = IB_SPEED_QDR;
413 		break;
414 	case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
415 		*active_width = IB_WIDTH_1X;
416 		*active_speed = IB_SPEED_EDR;
417 		break;
418 	case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
419 		*active_width = IB_WIDTH_2X;
420 		*active_speed = IB_SPEED_EDR;
421 		break;
422 	case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
423 		*active_width = IB_WIDTH_1X;
424 		*active_speed = IB_SPEED_HDR;
425 		break;
426 	case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
427 		*active_width = IB_WIDTH_4X;
428 		*active_speed = IB_SPEED_EDR;
429 		break;
430 	case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
431 		*active_width = IB_WIDTH_2X;
432 		*active_speed = IB_SPEED_HDR;
433 		break;
434 	case MLX5E_PROT_MASK(MLX5E_100GAUI_1_100GBASE_CR_KR):
435 		*active_width = IB_WIDTH_1X;
436 		*active_speed = IB_SPEED_NDR;
437 		break;
438 	case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4):
439 		*active_width = IB_WIDTH_4X;
440 		*active_speed = IB_SPEED_HDR;
441 		break;
442 	case MLX5E_PROT_MASK(MLX5E_200GAUI_2_200GBASE_CR2_KR2):
443 		*active_width = IB_WIDTH_2X;
444 		*active_speed = IB_SPEED_NDR;
445 		break;
446 	case MLX5E_PROT_MASK(MLX5E_400GAUI_4_400GBASE_CR4_KR4):
447 		*active_width = IB_WIDTH_4X;
448 		*active_speed = IB_SPEED_NDR;
449 		break;
450 	default:
451 		return -EINVAL;
452 	}
453 
454 	return 0;
455 }
456 
457 static int translate_eth_proto_oper(u32 eth_proto_oper, u16 *active_speed,
458 				    u8 *active_width, bool ext)
459 {
460 	return ext ?
461 		translate_eth_ext_proto_oper(eth_proto_oper, active_speed,
462 					     active_width) :
463 		translate_eth_legacy_proto_oper(eth_proto_oper, active_speed,
464 						active_width);
465 }
466 
467 static int mlx5_query_port_roce(struct ib_device *device, u32 port_num,
468 				struct ib_port_attr *props)
469 {
470 	struct mlx5_ib_dev *dev = to_mdev(device);
471 	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0};
472 	struct mlx5_core_dev *mdev;
473 	struct net_device *ndev, *upper;
474 	enum ib_mtu ndev_ib_mtu;
475 	bool put_mdev = true;
476 	u32 eth_prot_oper;
477 	u32 mdev_port_num;
478 	bool ext;
479 	int err;
480 
481 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
482 	if (!mdev) {
483 		/* This means the port isn't affiliated yet. Get the
484 		 * info for the master port instead.
485 		 */
486 		put_mdev = false;
487 		mdev = dev->mdev;
488 		mdev_port_num = 1;
489 		port_num = 1;
490 	}
491 
492 	/* Possible bad flows are checked before filling out props so in case
493 	 * of an error it will still be zeroed out.
494 	 * Use native port in case of reps
495 	 */
496 	if (dev->is_rep)
497 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
498 					   1);
499 	else
500 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
501 					   mdev_port_num);
502 	if (err)
503 		goto out;
504 	ext = !!MLX5_GET_ETH_PROTO(ptys_reg, out, true, eth_proto_capability);
505 	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
506 
507 	props->active_width     = IB_WIDTH_4X;
508 	props->active_speed     = IB_SPEED_QDR;
509 
510 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
511 				 &props->active_width, ext);
512 
513 	if (!dev->is_rep && dev->mdev->roce.roce_en) {
514 		u16 qkey_viol_cntr;
515 
516 		props->port_cap_flags |= IB_PORT_CM_SUP;
517 		props->ip_gids = true;
518 		props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev,
519 						   roce_address_table_size);
520 		mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
521 		props->qkey_viol_cntr = qkey_viol_cntr;
522 	}
523 	props->max_mtu          = IB_MTU_4096;
524 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
525 	props->pkey_tbl_len     = 1;
526 	props->state            = IB_PORT_DOWN;
527 	props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
528 
529 	/* If this is a stub query for an unaffiliated port stop here */
530 	if (!put_mdev)
531 		goto out;
532 
533 	ndev = mlx5_ib_get_netdev(device, port_num);
534 	if (!ndev)
535 		goto out;
536 
537 	if (dev->lag_active) {
538 		rcu_read_lock();
539 		upper = netdev_master_upper_dev_get_rcu(ndev);
540 		if (upper) {
541 			dev_put(ndev);
542 			ndev = upper;
543 			dev_hold(ndev);
544 		}
545 		rcu_read_unlock();
546 	}
547 
548 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
549 		props->state      = IB_PORT_ACTIVE;
550 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
551 	}
552 
553 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
554 
555 	dev_put(ndev);
556 
557 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
558 out:
559 	if (put_mdev)
560 		mlx5_ib_put_native_port_mdev(dev, port_num);
561 	return err;
562 }
563 
564 static int set_roce_addr(struct mlx5_ib_dev *dev, u32 port_num,
565 			 unsigned int index, const union ib_gid *gid,
566 			 const struct ib_gid_attr *attr)
567 {
568 	enum ib_gid_type gid_type;
569 	u16 vlan_id = 0xffff;
570 	u8 roce_version = 0;
571 	u8 roce_l3_type = 0;
572 	u8 mac[ETH_ALEN];
573 	int ret;
574 
575 	gid_type = attr->gid_type;
576 	if (gid) {
577 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
578 		if (ret)
579 			return ret;
580 	}
581 
582 	switch (gid_type) {
583 	case IB_GID_TYPE_ROCE:
584 		roce_version = MLX5_ROCE_VERSION_1;
585 		break;
586 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
587 		roce_version = MLX5_ROCE_VERSION_2;
588 		if (gid && ipv6_addr_v4mapped((void *)gid))
589 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
590 		else
591 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
592 		break;
593 
594 	default:
595 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
596 	}
597 
598 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
599 				      roce_l3_type, gid->raw, mac,
600 				      vlan_id < VLAN_CFI_MASK, vlan_id,
601 				      port_num);
602 }
603 
604 static int mlx5_ib_add_gid(const struct ib_gid_attr *attr,
605 			   __always_unused void **context)
606 {
607 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
608 			     attr->index, &attr->gid, attr);
609 }
610 
611 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
612 			   __always_unused void **context)
613 {
614 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
615 			     attr->index, NULL, attr);
616 }
617 
618 __be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev,
619 				   const struct ib_gid_attr *attr)
620 {
621 	if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
622 		return 0;
623 
624 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
625 }
626 
627 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
628 {
629 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
630 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
631 	return 0;
632 }
633 
634 enum {
635 	MLX5_VPORT_ACCESS_METHOD_MAD,
636 	MLX5_VPORT_ACCESS_METHOD_HCA,
637 	MLX5_VPORT_ACCESS_METHOD_NIC,
638 };
639 
640 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
641 {
642 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
643 		return MLX5_VPORT_ACCESS_METHOD_MAD;
644 
645 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
646 	    IB_LINK_LAYER_ETHERNET)
647 		return MLX5_VPORT_ACCESS_METHOD_NIC;
648 
649 	return MLX5_VPORT_ACCESS_METHOD_HCA;
650 }
651 
652 static void get_atomic_caps(struct mlx5_ib_dev *dev,
653 			    u8 atomic_size_qp,
654 			    struct ib_device_attr *props)
655 {
656 	u8 tmp;
657 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
658 	u8 atomic_req_8B_endianness_mode =
659 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
660 
661 	/* Check if HW supports 8 bytes standard atomic operations and capable
662 	 * of host endianness respond
663 	 */
664 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
665 	if (((atomic_operations & tmp) == tmp) &&
666 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
667 	    (atomic_req_8B_endianness_mode)) {
668 		props->atomic_cap = IB_ATOMIC_HCA;
669 	} else {
670 		props->atomic_cap = IB_ATOMIC_NONE;
671 	}
672 }
673 
674 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
675 			       struct ib_device_attr *props)
676 {
677 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
678 
679 	get_atomic_caps(dev, atomic_size_qp, props);
680 }
681 
682 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
683 					__be64 *sys_image_guid)
684 {
685 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
686 	struct mlx5_core_dev *mdev = dev->mdev;
687 	u64 tmp;
688 	int err;
689 
690 	switch (mlx5_get_vport_access_method(ibdev)) {
691 	case MLX5_VPORT_ACCESS_METHOD_MAD:
692 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
693 							    sys_image_guid);
694 
695 	case MLX5_VPORT_ACCESS_METHOD_HCA:
696 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
697 		break;
698 
699 	case MLX5_VPORT_ACCESS_METHOD_NIC:
700 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
701 		break;
702 
703 	default:
704 		return -EINVAL;
705 	}
706 
707 	if (!err)
708 		*sys_image_guid = cpu_to_be64(tmp);
709 
710 	return err;
711 
712 }
713 
714 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
715 				u16 *max_pkeys)
716 {
717 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
718 	struct mlx5_core_dev *mdev = dev->mdev;
719 
720 	switch (mlx5_get_vport_access_method(ibdev)) {
721 	case MLX5_VPORT_ACCESS_METHOD_MAD:
722 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
723 
724 	case MLX5_VPORT_ACCESS_METHOD_HCA:
725 	case MLX5_VPORT_ACCESS_METHOD_NIC:
726 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
727 						pkey_table_size));
728 		return 0;
729 
730 	default:
731 		return -EINVAL;
732 	}
733 }
734 
735 static int mlx5_query_vendor_id(struct ib_device *ibdev,
736 				u32 *vendor_id)
737 {
738 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
739 
740 	switch (mlx5_get_vport_access_method(ibdev)) {
741 	case MLX5_VPORT_ACCESS_METHOD_MAD:
742 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
743 
744 	case MLX5_VPORT_ACCESS_METHOD_HCA:
745 	case MLX5_VPORT_ACCESS_METHOD_NIC:
746 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
747 
748 	default:
749 		return -EINVAL;
750 	}
751 }
752 
753 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
754 				__be64 *node_guid)
755 {
756 	u64 tmp;
757 	int err;
758 
759 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
760 	case MLX5_VPORT_ACCESS_METHOD_MAD:
761 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
762 
763 	case MLX5_VPORT_ACCESS_METHOD_HCA:
764 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
765 		break;
766 
767 	case MLX5_VPORT_ACCESS_METHOD_NIC:
768 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
769 		break;
770 
771 	default:
772 		return -EINVAL;
773 	}
774 
775 	if (!err)
776 		*node_guid = cpu_to_be64(tmp);
777 
778 	return err;
779 }
780 
781 struct mlx5_reg_node_desc {
782 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
783 };
784 
785 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
786 {
787 	struct mlx5_reg_node_desc in;
788 
789 	if (mlx5_use_mad_ifc(dev))
790 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
791 
792 	memset(&in, 0, sizeof(in));
793 
794 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
795 				    sizeof(struct mlx5_reg_node_desc),
796 				    MLX5_REG_NODE_DESC, 0, 0);
797 }
798 
799 static int mlx5_ib_query_device(struct ib_device *ibdev,
800 				struct ib_device_attr *props,
801 				struct ib_udata *uhw)
802 {
803 	size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
804 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
805 	struct mlx5_core_dev *mdev = dev->mdev;
806 	int err = -ENOMEM;
807 	int max_sq_desc;
808 	int max_rq_sg;
809 	int max_sq_sg;
810 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
811 	bool raw_support = !mlx5_core_mp_enabled(mdev);
812 	struct mlx5_ib_query_device_resp resp = {};
813 	size_t resp_len;
814 	u64 max_tso;
815 
816 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
817 	if (uhw_outlen && uhw_outlen < resp_len)
818 		return -EINVAL;
819 
820 	resp.response_length = resp_len;
821 
822 	if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
823 		return -EINVAL;
824 
825 	memset(props, 0, sizeof(*props));
826 	err = mlx5_query_system_image_guid(ibdev,
827 					   &props->sys_image_guid);
828 	if (err)
829 		return err;
830 
831 	props->max_pkeys = dev->pkey_table_len;
832 
833 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
834 	if (err)
835 		return err;
836 
837 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
838 		(fw_rev_min(dev->mdev) << 16) |
839 		fw_rev_sub(dev->mdev);
840 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
841 		IB_DEVICE_PORT_ACTIVE_EVENT		|
842 		IB_DEVICE_SYS_IMAGE_GUID		|
843 		IB_DEVICE_RC_RNR_NAK_GEN;
844 
845 	if (MLX5_CAP_GEN(mdev, pkv))
846 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
847 	if (MLX5_CAP_GEN(mdev, qkv))
848 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
849 	if (MLX5_CAP_GEN(mdev, apm))
850 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
851 	if (MLX5_CAP_GEN(mdev, xrc))
852 		props->device_cap_flags |= IB_DEVICE_XRC;
853 	if (MLX5_CAP_GEN(mdev, imaicl)) {
854 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
855 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
856 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
857 		/* We support 'Gappy' memory registration too */
858 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
859 	}
860 	/* IB_WR_REG_MR always requires changing the entity size with UMR */
861 	if (!MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled))
862 		props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
863 	if (MLX5_CAP_GEN(mdev, sho)) {
864 		props->device_cap_flags |= IB_DEVICE_INTEGRITY_HANDOVER;
865 		/* At this stage no support for signature handover */
866 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
867 				      IB_PROT_T10DIF_TYPE_2 |
868 				      IB_PROT_T10DIF_TYPE_3;
869 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
870 				       IB_GUARD_T10DIF_CSUM;
871 	}
872 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
873 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
874 
875 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
876 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
877 			/* Legacy bit to support old userspace libraries */
878 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
879 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
880 		}
881 
882 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
883 			props->raw_packet_caps |=
884 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
885 
886 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) {
887 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
888 			if (max_tso) {
889 				resp.tso_caps.max_tso = 1 << max_tso;
890 				resp.tso_caps.supported_qpts |=
891 					1 << IB_QPT_RAW_PACKET;
892 				resp.response_length += sizeof(resp.tso_caps);
893 			}
894 		}
895 
896 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) {
897 			resp.rss_caps.rx_hash_function =
898 						MLX5_RX_HASH_FUNC_TOEPLITZ;
899 			resp.rss_caps.rx_hash_fields_mask =
900 						MLX5_RX_HASH_SRC_IPV4 |
901 						MLX5_RX_HASH_DST_IPV4 |
902 						MLX5_RX_HASH_SRC_IPV6 |
903 						MLX5_RX_HASH_DST_IPV6 |
904 						MLX5_RX_HASH_SRC_PORT_TCP |
905 						MLX5_RX_HASH_DST_PORT_TCP |
906 						MLX5_RX_HASH_SRC_PORT_UDP |
907 						MLX5_RX_HASH_DST_PORT_UDP |
908 						MLX5_RX_HASH_INNER;
909 			if (mlx5_accel_ipsec_device_caps(dev->mdev) &
910 			    MLX5_ACCEL_IPSEC_CAP_DEVICE)
911 				resp.rss_caps.rx_hash_fields_mask |=
912 					MLX5_RX_HASH_IPSEC_SPI;
913 			resp.response_length += sizeof(resp.rss_caps);
914 		}
915 	} else {
916 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen)
917 			resp.response_length += sizeof(resp.tso_caps);
918 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen)
919 			resp.response_length += sizeof(resp.rss_caps);
920 	}
921 
922 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
923 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
924 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
925 	}
926 
927 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
928 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
929 	    raw_support)
930 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
931 
932 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
933 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
934 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
935 
936 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
937 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
938 	    raw_support) {
939 		/* Legacy bit to support old userspace libraries */
940 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
941 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
942 	}
943 
944 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
945 		props->max_dm_size =
946 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
947 	}
948 
949 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
950 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
951 
952 	if (MLX5_CAP_GEN(mdev, end_pad))
953 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
954 
955 	props->vendor_part_id	   = mdev->pdev->device;
956 	props->hw_ver		   = mdev->pdev->revision;
957 
958 	props->max_mr_size	   = ~0ull;
959 	props->page_size_cap	   = ~(min_page_size - 1);
960 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
961 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
962 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
963 		     sizeof(struct mlx5_wqe_data_seg);
964 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
965 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
966 		     sizeof(struct mlx5_wqe_raddr_seg)) /
967 		sizeof(struct mlx5_wqe_data_seg);
968 	props->max_send_sge = max_sq_sg;
969 	props->max_recv_sge = max_rq_sg;
970 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
971 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
972 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
973 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
974 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
975 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
976 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
977 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
978 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
979 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
980 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
981 	props->max_srq_sge	   = max_rq_sg - 1;
982 	props->max_fast_reg_page_list_len =
983 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
984 	props->max_pi_fast_reg_page_list_len =
985 		props->max_fast_reg_page_list_len / 2;
986 	props->max_sgl_rd =
987 		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
988 	get_atomic_caps_qp(dev, props);
989 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
990 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
991 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
992 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
993 					   props->max_mcast_grp;
994 	props->max_ah = INT_MAX;
995 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
996 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
997 
998 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
999 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
1000 			props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
1001 		props->odp_caps = dev->odp_caps;
1002 		if (!uhw) {
1003 			/* ODP for kernel QPs is not implemented for receive
1004 			 * WQEs and SRQ WQEs
1005 			 */
1006 			props->odp_caps.per_transport_caps.rc_odp_caps &=
1007 				~(IB_ODP_SUPPORT_READ |
1008 				  IB_ODP_SUPPORT_SRQ_RECV);
1009 			props->odp_caps.per_transport_caps.uc_odp_caps &=
1010 				~(IB_ODP_SUPPORT_READ |
1011 				  IB_ODP_SUPPORT_SRQ_RECV);
1012 			props->odp_caps.per_transport_caps.ud_odp_caps &=
1013 				~(IB_ODP_SUPPORT_READ |
1014 				  IB_ODP_SUPPORT_SRQ_RECV);
1015 			props->odp_caps.per_transport_caps.xrc_odp_caps &=
1016 				~(IB_ODP_SUPPORT_READ |
1017 				  IB_ODP_SUPPORT_SRQ_RECV);
1018 		}
1019 	}
1020 
1021 	if (MLX5_CAP_GEN(mdev, cd))
1022 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
1023 
1024 	if (mlx5_core_is_vf(mdev))
1025 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
1026 
1027 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1028 	    IB_LINK_LAYER_ETHERNET && raw_support) {
1029 		props->rss_caps.max_rwq_indirection_tables =
1030 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1031 		props->rss_caps.max_rwq_indirection_table_size =
1032 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1033 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1034 		props->max_wq_type_rq =
1035 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1036 	}
1037 
1038 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1039 		props->tm_caps.max_num_tags =
1040 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1041 		props->tm_caps.max_ops =
1042 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1043 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1044 	}
1045 
1046 	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1047 	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1048 		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1049 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1050 	}
1051 
1052 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1053 		props->cq_caps.max_cq_moderation_count =
1054 						MLX5_MAX_CQ_COUNT;
1055 		props->cq_caps.max_cq_moderation_period =
1056 						MLX5_MAX_CQ_PERIOD;
1057 	}
1058 
1059 	if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) {
1060 		resp.response_length += sizeof(resp.cqe_comp_caps);
1061 
1062 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1063 			resp.cqe_comp_caps.max_num =
1064 				MLX5_CAP_GEN(dev->mdev,
1065 					     cqe_compression_max_num);
1066 
1067 			resp.cqe_comp_caps.supported_format =
1068 				MLX5_IB_CQE_RES_FORMAT_HASH |
1069 				MLX5_IB_CQE_RES_FORMAT_CSUM;
1070 
1071 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1072 				resp.cqe_comp_caps.supported_format |=
1073 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1074 		}
1075 	}
1076 
1077 	if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen &&
1078 	    raw_support) {
1079 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1080 		    MLX5_CAP_GEN(mdev, qos)) {
1081 			resp.packet_pacing_caps.qp_rate_limit_max =
1082 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1083 			resp.packet_pacing_caps.qp_rate_limit_min =
1084 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1085 			resp.packet_pacing_caps.supported_qpts |=
1086 				1 << IB_QPT_RAW_PACKET;
1087 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1088 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1089 				resp.packet_pacing_caps.cap_flags |=
1090 					MLX5_IB_PP_SUPPORT_BURST;
1091 		}
1092 		resp.response_length += sizeof(resp.packet_pacing_caps);
1093 	}
1094 
1095 	if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <=
1096 	    uhw_outlen) {
1097 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1098 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1099 				MLX5_IB_ALLOW_MPW;
1100 
1101 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1102 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1103 				MLX5_IB_SUPPORT_EMPW;
1104 
1105 		resp.response_length +=
1106 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1107 	}
1108 
1109 	if (offsetofend(typeof(resp), flags) <= uhw_outlen) {
1110 		resp.response_length += sizeof(resp.flags);
1111 
1112 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1113 			resp.flags |=
1114 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1115 
1116 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1117 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1118 		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1119 			resp.flags |=
1120 				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1121 
1122 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1123 	}
1124 
1125 	if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) {
1126 		resp.response_length += sizeof(resp.sw_parsing_caps);
1127 		if (MLX5_CAP_ETH(mdev, swp)) {
1128 			resp.sw_parsing_caps.sw_parsing_offloads |=
1129 				MLX5_IB_SW_PARSING;
1130 
1131 			if (MLX5_CAP_ETH(mdev, swp_csum))
1132 				resp.sw_parsing_caps.sw_parsing_offloads |=
1133 					MLX5_IB_SW_PARSING_CSUM;
1134 
1135 			if (MLX5_CAP_ETH(mdev, swp_lso))
1136 				resp.sw_parsing_caps.sw_parsing_offloads |=
1137 					MLX5_IB_SW_PARSING_LSO;
1138 
1139 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1140 				resp.sw_parsing_caps.supported_qpts =
1141 					BIT(IB_QPT_RAW_PACKET);
1142 		}
1143 	}
1144 
1145 	if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen &&
1146 	    raw_support) {
1147 		resp.response_length += sizeof(resp.striding_rq_caps);
1148 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1149 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1150 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1151 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1152 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1153 			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1154 				resp.striding_rq_caps
1155 					.min_single_wqe_log_num_of_strides =
1156 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1157 			else
1158 				resp.striding_rq_caps
1159 					.min_single_wqe_log_num_of_strides =
1160 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1161 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1162 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1163 			resp.striding_rq_caps.supported_qpts =
1164 				BIT(IB_QPT_RAW_PACKET);
1165 		}
1166 	}
1167 
1168 	if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) {
1169 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1170 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1171 			resp.tunnel_offloads_caps |=
1172 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1173 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1174 			resp.tunnel_offloads_caps |=
1175 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1176 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1177 			resp.tunnel_offloads_caps |=
1178 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1179 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre))
1180 			resp.tunnel_offloads_caps |=
1181 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1182 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp))
1183 			resp.tunnel_offloads_caps |=
1184 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1185 	}
1186 
1187 	if (offsetofend(typeof(resp), dci_streams_caps) <= uhw_outlen) {
1188 		resp.response_length += sizeof(resp.dci_streams_caps);
1189 
1190 		resp.dci_streams_caps.max_log_num_concurent =
1191 			MLX5_CAP_GEN(mdev, log_max_dci_stream_channels);
1192 
1193 		resp.dci_streams_caps.max_log_num_errored =
1194 			MLX5_CAP_GEN(mdev, log_max_dci_errored_streams);
1195 	}
1196 
1197 	if (uhw_outlen) {
1198 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1199 
1200 		if (err)
1201 			return err;
1202 	}
1203 
1204 	return 0;
1205 }
1206 
1207 static void translate_active_width(struct ib_device *ibdev, u16 active_width,
1208 				   u8 *ib_width)
1209 {
1210 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1211 
1212 	if (active_width & MLX5_PTYS_WIDTH_1X)
1213 		*ib_width = IB_WIDTH_1X;
1214 	else if (active_width & MLX5_PTYS_WIDTH_2X)
1215 		*ib_width = IB_WIDTH_2X;
1216 	else if (active_width & MLX5_PTYS_WIDTH_4X)
1217 		*ib_width = IB_WIDTH_4X;
1218 	else if (active_width & MLX5_PTYS_WIDTH_8X)
1219 		*ib_width = IB_WIDTH_8X;
1220 	else if (active_width & MLX5_PTYS_WIDTH_12X)
1221 		*ib_width = IB_WIDTH_12X;
1222 	else {
1223 		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1224 			    active_width);
1225 		*ib_width = IB_WIDTH_4X;
1226 	}
1227 
1228 	return;
1229 }
1230 
1231 static int mlx5_mtu_to_ib_mtu(int mtu)
1232 {
1233 	switch (mtu) {
1234 	case 256: return 1;
1235 	case 512: return 2;
1236 	case 1024: return 3;
1237 	case 2048: return 4;
1238 	case 4096: return 5;
1239 	default:
1240 		pr_warn("invalid mtu\n");
1241 		return -1;
1242 	}
1243 }
1244 
1245 enum ib_max_vl_num {
1246 	__IB_MAX_VL_0		= 1,
1247 	__IB_MAX_VL_0_1		= 2,
1248 	__IB_MAX_VL_0_3		= 3,
1249 	__IB_MAX_VL_0_7		= 4,
1250 	__IB_MAX_VL_0_14	= 5,
1251 };
1252 
1253 enum mlx5_vl_hw_cap {
1254 	MLX5_VL_HW_0	= 1,
1255 	MLX5_VL_HW_0_1	= 2,
1256 	MLX5_VL_HW_0_2	= 3,
1257 	MLX5_VL_HW_0_3	= 4,
1258 	MLX5_VL_HW_0_4	= 5,
1259 	MLX5_VL_HW_0_5	= 6,
1260 	MLX5_VL_HW_0_6	= 7,
1261 	MLX5_VL_HW_0_7	= 8,
1262 	MLX5_VL_HW_0_14	= 15
1263 };
1264 
1265 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1266 				u8 *max_vl_num)
1267 {
1268 	switch (vl_hw_cap) {
1269 	case MLX5_VL_HW_0:
1270 		*max_vl_num = __IB_MAX_VL_0;
1271 		break;
1272 	case MLX5_VL_HW_0_1:
1273 		*max_vl_num = __IB_MAX_VL_0_1;
1274 		break;
1275 	case MLX5_VL_HW_0_3:
1276 		*max_vl_num = __IB_MAX_VL_0_3;
1277 		break;
1278 	case MLX5_VL_HW_0_7:
1279 		*max_vl_num = __IB_MAX_VL_0_7;
1280 		break;
1281 	case MLX5_VL_HW_0_14:
1282 		*max_vl_num = __IB_MAX_VL_0_14;
1283 		break;
1284 
1285 	default:
1286 		return -EINVAL;
1287 	}
1288 
1289 	return 0;
1290 }
1291 
1292 static int mlx5_query_hca_port(struct ib_device *ibdev, u32 port,
1293 			       struct ib_port_attr *props)
1294 {
1295 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1296 	struct mlx5_core_dev *mdev = dev->mdev;
1297 	struct mlx5_hca_vport_context *rep;
1298 	u16 max_mtu;
1299 	u16 oper_mtu;
1300 	int err;
1301 	u16 ib_link_width_oper;
1302 	u8 vl_hw_cap;
1303 
1304 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1305 	if (!rep) {
1306 		err = -ENOMEM;
1307 		goto out;
1308 	}
1309 
1310 	/* props being zeroed by the caller, avoid zeroing it here */
1311 
1312 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1313 	if (err)
1314 		goto out;
1315 
1316 	props->lid		= rep->lid;
1317 	props->lmc		= rep->lmc;
1318 	props->sm_lid		= rep->sm_lid;
1319 	props->sm_sl		= rep->sm_sl;
1320 	props->state		= rep->vport_state;
1321 	props->phys_state	= rep->port_physical_state;
1322 	props->port_cap_flags	= rep->cap_mask1;
1323 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1324 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1325 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1326 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1327 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1328 	props->subnet_timeout	= rep->subnet_timeout;
1329 	props->init_type_reply	= rep->init_type_reply;
1330 
1331 	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1332 		props->port_cap_flags2 = rep->cap_mask2;
1333 
1334 	err = mlx5_query_ib_port_oper(mdev, &ib_link_width_oper,
1335 				      &props->active_speed, port);
1336 	if (err)
1337 		goto out;
1338 
1339 	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1340 
1341 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1342 
1343 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1344 
1345 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1346 
1347 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1348 
1349 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1350 	if (err)
1351 		goto out;
1352 
1353 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1354 				   &props->max_vl_num);
1355 out:
1356 	kfree(rep);
1357 	return err;
1358 }
1359 
1360 int mlx5_ib_query_port(struct ib_device *ibdev, u32 port,
1361 		       struct ib_port_attr *props)
1362 {
1363 	unsigned int count;
1364 	int ret;
1365 
1366 	switch (mlx5_get_vport_access_method(ibdev)) {
1367 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1368 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1369 		break;
1370 
1371 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1372 		ret = mlx5_query_hca_port(ibdev, port, props);
1373 		break;
1374 
1375 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1376 		ret = mlx5_query_port_roce(ibdev, port, props);
1377 		break;
1378 
1379 	default:
1380 		ret = -EINVAL;
1381 	}
1382 
1383 	if (!ret && props) {
1384 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1385 		struct mlx5_core_dev *mdev;
1386 		bool put_mdev = true;
1387 
1388 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1389 		if (!mdev) {
1390 			/* If the port isn't affiliated yet query the master.
1391 			 * The master and slave will have the same values.
1392 			 */
1393 			mdev = dev->mdev;
1394 			port = 1;
1395 			put_mdev = false;
1396 		}
1397 		count = mlx5_core_reserved_gids_count(mdev);
1398 		if (put_mdev)
1399 			mlx5_ib_put_native_port_mdev(dev, port);
1400 		props->gid_tbl_len -= count;
1401 	}
1402 	return ret;
1403 }
1404 
1405 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u32 port,
1406 				  struct ib_port_attr *props)
1407 {
1408 	return mlx5_query_port_roce(ibdev, port, props);
1409 }
1410 
1411 static int mlx5_ib_rep_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1412 				  u16 *pkey)
1413 {
1414 	/* Default special Pkey for representor device port as per the
1415 	 * IB specification 1.3 section 10.9.1.2.
1416 	 */
1417 	*pkey = 0xffff;
1418 	return 0;
1419 }
1420 
1421 static int mlx5_ib_query_gid(struct ib_device *ibdev, u32 port, int index,
1422 			     union ib_gid *gid)
1423 {
1424 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1425 	struct mlx5_core_dev *mdev = dev->mdev;
1426 
1427 	switch (mlx5_get_vport_access_method(ibdev)) {
1428 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1429 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1430 
1431 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1432 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1433 
1434 	default:
1435 		return -EINVAL;
1436 	}
1437 
1438 }
1439 
1440 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u32 port,
1441 				   u16 index, u16 *pkey)
1442 {
1443 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1444 	struct mlx5_core_dev *mdev;
1445 	bool put_mdev = true;
1446 	u32 mdev_port_num;
1447 	int err;
1448 
1449 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1450 	if (!mdev) {
1451 		/* The port isn't affiliated yet, get the PKey from the master
1452 		 * port. For RoCE the PKey tables will be the same.
1453 		 */
1454 		put_mdev = false;
1455 		mdev = dev->mdev;
1456 		mdev_port_num = 1;
1457 	}
1458 
1459 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1460 					index, pkey);
1461 	if (put_mdev)
1462 		mlx5_ib_put_native_port_mdev(dev, port);
1463 
1464 	return err;
1465 }
1466 
1467 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1468 			      u16 *pkey)
1469 {
1470 	switch (mlx5_get_vport_access_method(ibdev)) {
1471 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1472 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1473 
1474 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1475 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1476 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1477 	default:
1478 		return -EINVAL;
1479 	}
1480 }
1481 
1482 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1483 				 struct ib_device_modify *props)
1484 {
1485 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1486 	struct mlx5_reg_node_desc in;
1487 	struct mlx5_reg_node_desc out;
1488 	int err;
1489 
1490 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1491 		return -EOPNOTSUPP;
1492 
1493 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1494 		return 0;
1495 
1496 	/*
1497 	 * If possible, pass node desc to FW, so it can generate
1498 	 * a 144 trap.  If cmd fails, just ignore.
1499 	 */
1500 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1501 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1502 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1503 	if (err)
1504 		return err;
1505 
1506 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1507 
1508 	return err;
1509 }
1510 
1511 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u32 port_num, u32 mask,
1512 				u32 value)
1513 {
1514 	struct mlx5_hca_vport_context ctx = {};
1515 	struct mlx5_core_dev *mdev;
1516 	u32 mdev_port_num;
1517 	int err;
1518 
1519 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1520 	if (!mdev)
1521 		return -ENODEV;
1522 
1523 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1524 	if (err)
1525 		goto out;
1526 
1527 	if (~ctx.cap_mask1_perm & mask) {
1528 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1529 			     mask, ctx.cap_mask1_perm);
1530 		err = -EINVAL;
1531 		goto out;
1532 	}
1533 
1534 	ctx.cap_mask1 = value;
1535 	ctx.cap_mask1_perm = mask;
1536 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1537 						 0, &ctx);
1538 
1539 out:
1540 	mlx5_ib_put_native_port_mdev(dev, port_num);
1541 
1542 	return err;
1543 }
1544 
1545 static int mlx5_ib_modify_port(struct ib_device *ibdev, u32 port, int mask,
1546 			       struct ib_port_modify *props)
1547 {
1548 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1549 	struct ib_port_attr attr;
1550 	u32 tmp;
1551 	int err;
1552 	u32 change_mask;
1553 	u32 value;
1554 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1555 		      IB_LINK_LAYER_INFINIBAND);
1556 
1557 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1558 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1559 	 */
1560 	if (!is_ib)
1561 		return 0;
1562 
1563 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1564 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1565 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1566 		return set_port_caps_atomic(dev, port, change_mask, value);
1567 	}
1568 
1569 	mutex_lock(&dev->cap_mask_mutex);
1570 
1571 	err = ib_query_port(ibdev, port, &attr);
1572 	if (err)
1573 		goto out;
1574 
1575 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1576 		~props->clr_port_cap_mask;
1577 
1578 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1579 
1580 out:
1581 	mutex_unlock(&dev->cap_mask_mutex);
1582 	return err;
1583 }
1584 
1585 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1586 {
1587 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1588 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1589 }
1590 
1591 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1592 {
1593 	/* Large page with non 4k uar support might limit the dynamic size */
1594 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1595 		return MLX5_MIN_DYN_BFREGS;
1596 
1597 	return MLX5_MAX_DYN_BFREGS;
1598 }
1599 
1600 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1601 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1602 			     struct mlx5_bfreg_info *bfregi)
1603 {
1604 	int uars_per_sys_page;
1605 	int bfregs_per_sys_page;
1606 	int ref_bfregs = req->total_num_bfregs;
1607 
1608 	if (req->total_num_bfregs == 0)
1609 		return -EINVAL;
1610 
1611 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1612 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1613 
1614 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1615 		return -ENOMEM;
1616 
1617 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1618 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1619 	/* This holds the required static allocation asked by the user */
1620 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1621 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1622 		return -EINVAL;
1623 
1624 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1625 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1626 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1627 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1628 
1629 	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",
1630 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1631 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1632 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1633 		    bfregi->num_sys_pages);
1634 
1635 	return 0;
1636 }
1637 
1638 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1639 {
1640 	struct mlx5_bfreg_info *bfregi;
1641 	int err;
1642 	int i;
1643 
1644 	bfregi = &context->bfregi;
1645 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1646 		err = mlx5_cmd_uar_alloc(dev->mdev, &bfregi->sys_pages[i],
1647 					 context->devx_uid);
1648 		if (err)
1649 			goto error;
1650 
1651 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1652 	}
1653 
1654 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1655 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1656 
1657 	return 0;
1658 
1659 error:
1660 	for (--i; i >= 0; i--)
1661 		if (mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i],
1662 					 context->devx_uid))
1663 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1664 
1665 	return err;
1666 }
1667 
1668 static void deallocate_uars(struct mlx5_ib_dev *dev,
1669 			    struct mlx5_ib_ucontext *context)
1670 {
1671 	struct mlx5_bfreg_info *bfregi;
1672 	int i;
1673 
1674 	bfregi = &context->bfregi;
1675 	for (i = 0; i < bfregi->num_sys_pages; i++)
1676 		if (i < bfregi->num_static_sys_pages ||
1677 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1678 			mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i],
1679 					     context->devx_uid);
1680 }
1681 
1682 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1683 {
1684 	int err = 0;
1685 
1686 	mutex_lock(&dev->lb.mutex);
1687 	if (td)
1688 		dev->lb.user_td++;
1689 	if (qp)
1690 		dev->lb.qps++;
1691 
1692 	if (dev->lb.user_td == 2 ||
1693 	    dev->lb.qps == 1) {
1694 		if (!dev->lb.enabled) {
1695 			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1696 			dev->lb.enabled = true;
1697 		}
1698 	}
1699 
1700 	mutex_unlock(&dev->lb.mutex);
1701 
1702 	return err;
1703 }
1704 
1705 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1706 {
1707 	mutex_lock(&dev->lb.mutex);
1708 	if (td)
1709 		dev->lb.user_td--;
1710 	if (qp)
1711 		dev->lb.qps--;
1712 
1713 	if (dev->lb.user_td == 1 &&
1714 	    dev->lb.qps == 0) {
1715 		if (dev->lb.enabled) {
1716 			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1717 			dev->lb.enabled = false;
1718 		}
1719 	}
1720 
1721 	mutex_unlock(&dev->lb.mutex);
1722 }
1723 
1724 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1725 					  u16 uid)
1726 {
1727 	int err;
1728 
1729 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1730 		return 0;
1731 
1732 	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1733 	if (err)
1734 		return err;
1735 
1736 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1737 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1738 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1739 		return err;
1740 
1741 	return mlx5_ib_enable_lb(dev, true, false);
1742 }
1743 
1744 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1745 					     u16 uid)
1746 {
1747 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1748 		return;
1749 
1750 	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1751 
1752 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1753 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1754 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1755 		return;
1756 
1757 	mlx5_ib_disable_lb(dev, true, false);
1758 }
1759 
1760 static int set_ucontext_resp(struct ib_ucontext *uctx,
1761 			     struct mlx5_ib_alloc_ucontext_resp *resp)
1762 {
1763 	struct ib_device *ibdev = uctx->device;
1764 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1765 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1766 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1767 	int err;
1768 
1769 	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1770 		err = mlx5_cmd_dump_fill_mkey(dev->mdev,
1771 					      &resp->dump_fill_mkey);
1772 		if (err)
1773 			return err;
1774 		resp->comp_mask |=
1775 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1776 	}
1777 
1778 	resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1779 	if (dev->wc_support)
1780 		resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev,
1781 						      log_bf_reg_size);
1782 	resp->cache_line_size = cache_line_size();
1783 	resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1784 	resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1785 	resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1786 	resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1787 	resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1788 	resp->cqe_version = context->cqe_version;
1789 	resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1790 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1791 	resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1792 					MLX5_CAP_GEN(dev->mdev,
1793 						     num_of_uars_per_page) : 1;
1794 
1795 	if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1796 				MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1797 		if (mlx5_get_flow_namespace(dev->mdev,
1798 				MLX5_FLOW_NAMESPACE_EGRESS))
1799 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1800 		if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1801 				MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1802 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1803 		if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1804 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1805 		if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1806 				MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1807 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1808 		/* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1809 	}
1810 
1811 	resp->tot_bfregs = bfregi->lib_uar_dyn ? 0 :
1812 			bfregi->total_num_bfregs - bfregi->num_dyn_bfregs;
1813 	resp->num_ports = dev->num_ports;
1814 	resp->cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1815 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1816 
1817 	if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1818 		mlx5_query_min_inline(dev->mdev, &resp->eth_min_inline);
1819 		resp->eth_min_inline++;
1820 	}
1821 
1822 	if (dev->mdev->clock_info)
1823 		resp->clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1824 
1825 	/*
1826 	 * We don't want to expose information from the PCI bar that is located
1827 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1828 	 * pretend we don't support reading the HCA's core clock. This is also
1829 	 * forced by mmap function.
1830 	 */
1831 	if (PAGE_SIZE <= 4096) {
1832 		resp->comp_mask |=
1833 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1834 		resp->hca_core_clock_offset =
1835 			offsetof(struct mlx5_init_seg,
1836 				 internal_timer_h) % PAGE_SIZE;
1837 	}
1838 
1839 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
1840 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE;
1841 
1842 	if (rt_supported(MLX5_CAP_GEN(dev->mdev, sq_ts_format)) &&
1843 	    rt_supported(MLX5_CAP_GEN(dev->mdev, rq_ts_format)) &&
1844 	    rt_supported(MLX5_CAP_ROCE(dev->mdev, qp_ts_format)))
1845 		resp->comp_mask |=
1846 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_REAL_TIME_TS;
1847 
1848 	resp->num_dyn_bfregs = bfregi->num_dyn_bfregs;
1849 
1850 	if (MLX5_CAP_GEN(dev->mdev, drain_sigerr))
1851 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_SQD2RTS;
1852 
1853 	return 0;
1854 }
1855 
1856 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1857 				  struct ib_udata *udata)
1858 {
1859 	struct ib_device *ibdev = uctx->device;
1860 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1861 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1862 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1863 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1864 	struct mlx5_bfreg_info *bfregi;
1865 	int ver;
1866 	int err;
1867 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1868 				     max_cqe_version);
1869 	bool lib_uar_4k;
1870 	bool lib_uar_dyn;
1871 
1872 	if (!dev->ib_active)
1873 		return -EAGAIN;
1874 
1875 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1876 		ver = 0;
1877 	else if (udata->inlen >= min_req_v2)
1878 		ver = 2;
1879 	else
1880 		return -EINVAL;
1881 
1882 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1883 	if (err)
1884 		return err;
1885 
1886 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1887 		return -EOPNOTSUPP;
1888 
1889 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1890 		return -EOPNOTSUPP;
1891 
1892 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1893 				    MLX5_NON_FP_BFREGS_PER_UAR);
1894 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1895 		return -EINVAL;
1896 
1897 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1898 		err = mlx5_ib_devx_create(dev, true);
1899 		if (err < 0)
1900 			goto out_ctx;
1901 		context->devx_uid = err;
1902 	}
1903 
1904 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1905 	lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
1906 	bfregi = &context->bfregi;
1907 
1908 	if (lib_uar_dyn) {
1909 		bfregi->lib_uar_dyn = lib_uar_dyn;
1910 		goto uar_done;
1911 	}
1912 
1913 	/* updates req->total_num_bfregs */
1914 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1915 	if (err)
1916 		goto out_devx;
1917 
1918 	mutex_init(&bfregi->lock);
1919 	bfregi->lib_uar_4k = lib_uar_4k;
1920 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1921 				GFP_KERNEL);
1922 	if (!bfregi->count) {
1923 		err = -ENOMEM;
1924 		goto out_devx;
1925 	}
1926 
1927 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1928 				    sizeof(*bfregi->sys_pages),
1929 				    GFP_KERNEL);
1930 	if (!bfregi->sys_pages) {
1931 		err = -ENOMEM;
1932 		goto out_count;
1933 	}
1934 
1935 	err = allocate_uars(dev, context);
1936 	if (err)
1937 		goto out_sys_pages;
1938 
1939 uar_done:
1940 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1941 					     context->devx_uid);
1942 	if (err)
1943 		goto out_uars;
1944 
1945 	INIT_LIST_HEAD(&context->db_page_list);
1946 	mutex_init(&context->db_page_mutex);
1947 
1948 	context->cqe_version = min_t(__u8,
1949 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1950 				 req.max_cqe_version);
1951 
1952 	err = set_ucontext_resp(uctx, &resp);
1953 	if (err)
1954 		goto out_mdev;
1955 
1956 	resp.response_length = min(udata->outlen, sizeof(resp));
1957 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1958 	if (err)
1959 		goto out_mdev;
1960 
1961 	bfregi->ver = ver;
1962 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1963 	context->lib_caps = req.lib_caps;
1964 	print_lib_caps(dev, context->lib_caps);
1965 
1966 	if (mlx5_ib_lag_should_assign_affinity(dev)) {
1967 		u32 port = mlx5_core_native_port_num(dev->mdev) - 1;
1968 
1969 		atomic_set(&context->tx_port_affinity,
1970 			   atomic_add_return(
1971 				   1, &dev->port[port].roce.tx_port_affinity));
1972 	}
1973 
1974 	return 0;
1975 
1976 out_mdev:
1977 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1978 
1979 out_uars:
1980 	deallocate_uars(dev, context);
1981 
1982 out_sys_pages:
1983 	kfree(bfregi->sys_pages);
1984 
1985 out_count:
1986 	kfree(bfregi->count);
1987 
1988 out_devx:
1989 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1990 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1991 
1992 out_ctx:
1993 	return err;
1994 }
1995 
1996 static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext,
1997 				  struct uverbs_attr_bundle *attrs)
1998 {
1999 	struct mlx5_ib_alloc_ucontext_resp uctx_resp = {};
2000 	int ret;
2001 
2002 	ret = set_ucontext_resp(ibcontext, &uctx_resp);
2003 	if (ret)
2004 		return ret;
2005 
2006 	uctx_resp.response_length =
2007 		min_t(size_t,
2008 		      uverbs_attr_get_len(attrs,
2009 				MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX),
2010 		      sizeof(uctx_resp));
2011 
2012 	ret = uverbs_copy_to_struct_or_zero(attrs,
2013 					MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
2014 					&uctx_resp,
2015 					sizeof(uctx_resp));
2016 	return ret;
2017 }
2018 
2019 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
2020 {
2021 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2022 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2023 	struct mlx5_bfreg_info *bfregi;
2024 
2025 	bfregi = &context->bfregi;
2026 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
2027 
2028 	deallocate_uars(dev, context);
2029 	kfree(bfregi->sys_pages);
2030 	kfree(bfregi->count);
2031 
2032 	if (context->devx_uid)
2033 		mlx5_ib_devx_destroy(dev, context->devx_uid);
2034 }
2035 
2036 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2037 				 int uar_idx)
2038 {
2039 	int fw_uars_per_page;
2040 
2041 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2042 
2043 	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2044 }
2045 
2046 static u64 uar_index2paddress(struct mlx5_ib_dev *dev,
2047 				 int uar_idx)
2048 {
2049 	unsigned int fw_uars_per_page;
2050 
2051 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
2052 				MLX5_UARS_IN_PAGE : 1;
2053 
2054 	return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE);
2055 }
2056 
2057 static int get_command(unsigned long offset)
2058 {
2059 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2060 }
2061 
2062 static int get_arg(unsigned long offset)
2063 {
2064 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2065 }
2066 
2067 static int get_index(unsigned long offset)
2068 {
2069 	return get_arg(offset);
2070 }
2071 
2072 /* Index resides in an extra byte to enable larger values than 255 */
2073 static int get_extended_index(unsigned long offset)
2074 {
2075 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2076 }
2077 
2078 
2079 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2080 {
2081 }
2082 
2083 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2084 {
2085 	switch (cmd) {
2086 	case MLX5_IB_MMAP_WC_PAGE:
2087 		return "WC";
2088 	case MLX5_IB_MMAP_REGULAR_PAGE:
2089 		return "best effort WC";
2090 	case MLX5_IB_MMAP_NC_PAGE:
2091 		return "NC";
2092 	case MLX5_IB_MMAP_DEVICE_MEM:
2093 		return "Device Memory";
2094 	default:
2095 		return NULL;
2096 	}
2097 }
2098 
2099 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2100 					struct vm_area_struct *vma,
2101 					struct mlx5_ib_ucontext *context)
2102 {
2103 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2104 	    !(vma->vm_flags & VM_SHARED))
2105 		return -EINVAL;
2106 
2107 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2108 		return -EOPNOTSUPP;
2109 
2110 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2111 		return -EPERM;
2112 	vma->vm_flags &= ~VM_MAYWRITE;
2113 
2114 	if (!dev->mdev->clock_info)
2115 		return -EOPNOTSUPP;
2116 
2117 	return vm_insert_page(vma, vma->vm_start,
2118 			      virt_to_page(dev->mdev->clock_info));
2119 }
2120 
2121 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2122 {
2123 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2124 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2125 	struct mlx5_var_table *var_table = &dev->var_table;
2126 	struct mlx5_ib_ucontext *context = to_mucontext(entry->ucontext);
2127 
2128 	switch (mentry->mmap_flag) {
2129 	case MLX5_IB_MMAP_TYPE_MEMIC:
2130 	case MLX5_IB_MMAP_TYPE_MEMIC_OP:
2131 		mlx5_ib_dm_mmap_free(dev, mentry);
2132 		break;
2133 	case MLX5_IB_MMAP_TYPE_VAR:
2134 		mutex_lock(&var_table->bitmap_lock);
2135 		clear_bit(mentry->page_idx, var_table->bitmap);
2136 		mutex_unlock(&var_table->bitmap_lock);
2137 		kfree(mentry);
2138 		break;
2139 	case MLX5_IB_MMAP_TYPE_UAR_WC:
2140 	case MLX5_IB_MMAP_TYPE_UAR_NC:
2141 		mlx5_cmd_uar_dealloc(dev->mdev, mentry->page_idx,
2142 				     context->devx_uid);
2143 		kfree(mentry);
2144 		break;
2145 	default:
2146 		WARN_ON(true);
2147 	}
2148 }
2149 
2150 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2151 		    struct vm_area_struct *vma,
2152 		    struct mlx5_ib_ucontext *context)
2153 {
2154 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2155 	int err;
2156 	unsigned long idx;
2157 	phys_addr_t pfn;
2158 	pgprot_t prot;
2159 	u32 bfreg_dyn_idx = 0;
2160 	u32 uar_index;
2161 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2162 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2163 				bfregi->num_static_sys_pages;
2164 
2165 	if (bfregi->lib_uar_dyn)
2166 		return -EINVAL;
2167 
2168 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2169 		return -EINVAL;
2170 
2171 	if (dyn_uar)
2172 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2173 	else
2174 		idx = get_index(vma->vm_pgoff);
2175 
2176 	if (idx >= max_valid_idx) {
2177 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2178 			     idx, max_valid_idx);
2179 		return -EINVAL;
2180 	}
2181 
2182 	switch (cmd) {
2183 	case MLX5_IB_MMAP_WC_PAGE:
2184 	case MLX5_IB_MMAP_ALLOC_WC:
2185 	case MLX5_IB_MMAP_REGULAR_PAGE:
2186 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2187 		prot = pgprot_writecombine(vma->vm_page_prot);
2188 		break;
2189 	case MLX5_IB_MMAP_NC_PAGE:
2190 		prot = pgprot_noncached(vma->vm_page_prot);
2191 		break;
2192 	default:
2193 		return -EINVAL;
2194 	}
2195 
2196 	if (dyn_uar) {
2197 		int uars_per_page;
2198 
2199 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2200 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2201 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2202 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2203 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2204 			return -EINVAL;
2205 		}
2206 
2207 		mutex_lock(&bfregi->lock);
2208 		/* Fail if uar already allocated, first bfreg index of each
2209 		 * page holds its count.
2210 		 */
2211 		if (bfregi->count[bfreg_dyn_idx]) {
2212 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2213 			mutex_unlock(&bfregi->lock);
2214 			return -EINVAL;
2215 		}
2216 
2217 		bfregi->count[bfreg_dyn_idx]++;
2218 		mutex_unlock(&bfregi->lock);
2219 
2220 		err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index,
2221 					 context->devx_uid);
2222 		if (err) {
2223 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2224 			goto free_bfreg;
2225 		}
2226 	} else {
2227 		uar_index = bfregi->sys_pages[idx];
2228 	}
2229 
2230 	pfn = uar_index2pfn(dev, uar_index);
2231 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2232 
2233 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2234 				prot, NULL);
2235 	if (err) {
2236 		mlx5_ib_err(dev,
2237 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2238 			    err, mmap_cmd2str(cmd));
2239 		goto err;
2240 	}
2241 
2242 	if (dyn_uar)
2243 		bfregi->sys_pages[idx] = uar_index;
2244 	return 0;
2245 
2246 err:
2247 	if (!dyn_uar)
2248 		return err;
2249 
2250 	mlx5_cmd_uar_dealloc(dev->mdev, idx, context->devx_uid);
2251 
2252 free_bfreg:
2253 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2254 
2255 	return err;
2256 }
2257 
2258 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2259 {
2260 	unsigned long idx;
2261 	u8 command;
2262 
2263 	command = get_command(vma->vm_pgoff);
2264 	idx = get_extended_index(vma->vm_pgoff);
2265 
2266 	return (command << 16 | idx);
2267 }
2268 
2269 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2270 			       struct vm_area_struct *vma,
2271 			       struct ib_ucontext *ucontext)
2272 {
2273 	struct mlx5_user_mmap_entry *mentry;
2274 	struct rdma_user_mmap_entry *entry;
2275 	unsigned long pgoff;
2276 	pgprot_t prot;
2277 	phys_addr_t pfn;
2278 	int ret;
2279 
2280 	pgoff = mlx5_vma_to_pgoff(vma);
2281 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2282 	if (!entry)
2283 		return -EINVAL;
2284 
2285 	mentry = to_mmmap(entry);
2286 	pfn = (mentry->address >> PAGE_SHIFT);
2287 	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
2288 	    mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
2289 		prot = pgprot_noncached(vma->vm_page_prot);
2290 	else
2291 		prot = pgprot_writecombine(vma->vm_page_prot);
2292 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2293 				entry->npages * PAGE_SIZE,
2294 				prot,
2295 				entry);
2296 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2297 	return ret;
2298 }
2299 
2300 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2301 {
2302 	u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2303 	u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2304 
2305 	return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2306 		(index & 0xFF)) << PAGE_SHIFT;
2307 }
2308 
2309 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2310 {
2311 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2312 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2313 	unsigned long command;
2314 	phys_addr_t pfn;
2315 
2316 	command = get_command(vma->vm_pgoff);
2317 	switch (command) {
2318 	case MLX5_IB_MMAP_WC_PAGE:
2319 	case MLX5_IB_MMAP_ALLOC_WC:
2320 		if (!dev->wc_support)
2321 			return -EPERM;
2322 		fallthrough;
2323 	case MLX5_IB_MMAP_NC_PAGE:
2324 	case MLX5_IB_MMAP_REGULAR_PAGE:
2325 		return uar_mmap(dev, command, vma, context);
2326 
2327 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2328 		return -ENOSYS;
2329 
2330 	case MLX5_IB_MMAP_CORE_CLOCK:
2331 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2332 			return -EINVAL;
2333 
2334 		if (vma->vm_flags & VM_WRITE)
2335 			return -EPERM;
2336 		vma->vm_flags &= ~VM_MAYWRITE;
2337 
2338 		/* Don't expose to user-space information it shouldn't have */
2339 		if (PAGE_SIZE > 4096)
2340 			return -EOPNOTSUPP;
2341 
2342 		pfn = (dev->mdev->iseg_base +
2343 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2344 			PAGE_SHIFT;
2345 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2346 					 PAGE_SIZE,
2347 					 pgprot_noncached(vma->vm_page_prot),
2348 					 NULL);
2349 	case MLX5_IB_MMAP_CLOCK_INFO:
2350 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2351 
2352 	default:
2353 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2354 	}
2355 
2356 	return 0;
2357 }
2358 
2359 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2360 {
2361 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2362 	struct ib_device *ibdev = ibpd->device;
2363 	struct mlx5_ib_alloc_pd_resp resp;
2364 	int err;
2365 	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2366 	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {};
2367 	u16 uid = 0;
2368 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2369 		udata, struct mlx5_ib_ucontext, ibucontext);
2370 
2371 	uid = context ? context->devx_uid : 0;
2372 	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2373 	MLX5_SET(alloc_pd_in, in, uid, uid);
2374 	err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out);
2375 	if (err)
2376 		return err;
2377 
2378 	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2379 	pd->uid = uid;
2380 	if (udata) {
2381 		resp.pdn = pd->pdn;
2382 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2383 			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2384 			return -EFAULT;
2385 		}
2386 	}
2387 
2388 	return 0;
2389 }
2390 
2391 static int mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2392 {
2393 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2394 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2395 
2396 	return mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2397 }
2398 
2399 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2400 {
2401 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2402 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
2403 	int err;
2404 	u16 uid;
2405 
2406 	uid = ibqp->pd ?
2407 		to_mpd(ibqp->pd)->uid : 0;
2408 
2409 	if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) {
2410 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
2411 		return -EOPNOTSUPP;
2412 	}
2413 
2414 	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2415 	if (err)
2416 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2417 			     ibqp->qp_num, gid->raw);
2418 
2419 	return err;
2420 }
2421 
2422 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2423 {
2424 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2425 	int err;
2426 	u16 uid;
2427 
2428 	uid = ibqp->pd ?
2429 		to_mpd(ibqp->pd)->uid : 0;
2430 	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2431 	if (err)
2432 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2433 			     ibqp->qp_num, gid->raw);
2434 
2435 	return err;
2436 }
2437 
2438 static int init_node_data(struct mlx5_ib_dev *dev)
2439 {
2440 	int err;
2441 
2442 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2443 	if (err)
2444 		return err;
2445 
2446 	dev->mdev->rev_id = dev->mdev->pdev->revision;
2447 
2448 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2449 }
2450 
2451 static ssize_t fw_pages_show(struct device *device,
2452 			     struct device_attribute *attr, char *buf)
2453 {
2454 	struct mlx5_ib_dev *dev =
2455 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2456 
2457 	return sysfs_emit(buf, "%d\n", dev->mdev->priv.fw_pages);
2458 }
2459 static DEVICE_ATTR_RO(fw_pages);
2460 
2461 static ssize_t reg_pages_show(struct device *device,
2462 			      struct device_attribute *attr, char *buf)
2463 {
2464 	struct mlx5_ib_dev *dev =
2465 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2466 
2467 	return sysfs_emit(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2468 }
2469 static DEVICE_ATTR_RO(reg_pages);
2470 
2471 static ssize_t hca_type_show(struct device *device,
2472 			     struct device_attribute *attr, char *buf)
2473 {
2474 	struct mlx5_ib_dev *dev =
2475 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2476 
2477 	return sysfs_emit(buf, "MT%d\n", dev->mdev->pdev->device);
2478 }
2479 static DEVICE_ATTR_RO(hca_type);
2480 
2481 static ssize_t hw_rev_show(struct device *device,
2482 			   struct device_attribute *attr, char *buf)
2483 {
2484 	struct mlx5_ib_dev *dev =
2485 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2486 
2487 	return sysfs_emit(buf, "%x\n", dev->mdev->rev_id);
2488 }
2489 static DEVICE_ATTR_RO(hw_rev);
2490 
2491 static ssize_t board_id_show(struct device *device,
2492 			     struct device_attribute *attr, char *buf)
2493 {
2494 	struct mlx5_ib_dev *dev =
2495 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2496 
2497 	return sysfs_emit(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2498 			  dev->mdev->board_id);
2499 }
2500 static DEVICE_ATTR_RO(board_id);
2501 
2502 static struct attribute *mlx5_class_attributes[] = {
2503 	&dev_attr_hw_rev.attr,
2504 	&dev_attr_hca_type.attr,
2505 	&dev_attr_board_id.attr,
2506 	&dev_attr_fw_pages.attr,
2507 	&dev_attr_reg_pages.attr,
2508 	NULL,
2509 };
2510 
2511 static const struct attribute_group mlx5_attr_group = {
2512 	.attrs = mlx5_class_attributes,
2513 };
2514 
2515 static void pkey_change_handler(struct work_struct *work)
2516 {
2517 	struct mlx5_ib_port_resources *ports =
2518 		container_of(work, struct mlx5_ib_port_resources,
2519 			     pkey_change_work);
2520 
2521 	if (!ports->gsi)
2522 		/*
2523 		 * We got this event before device was fully configured
2524 		 * and MAD registration code wasn't called/finished yet.
2525 		 */
2526 		return;
2527 
2528 	mlx5_ib_gsi_pkey_change(ports->gsi);
2529 }
2530 
2531 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2532 {
2533 	struct mlx5_ib_qp *mqp;
2534 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2535 	struct mlx5_core_cq *mcq;
2536 	struct list_head cq_armed_list;
2537 	unsigned long flags_qp;
2538 	unsigned long flags_cq;
2539 	unsigned long flags;
2540 
2541 	INIT_LIST_HEAD(&cq_armed_list);
2542 
2543 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2544 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2545 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2546 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2547 		if (mqp->sq.tail != mqp->sq.head) {
2548 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2549 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2550 			if (send_mcq->mcq.comp &&
2551 			    mqp->ibqp.send_cq->comp_handler) {
2552 				if (!send_mcq->mcq.reset_notify_added) {
2553 					send_mcq->mcq.reset_notify_added = 1;
2554 					list_add_tail(&send_mcq->mcq.reset_notify,
2555 						      &cq_armed_list);
2556 				}
2557 			}
2558 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2559 		}
2560 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2561 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2562 		/* no handling is needed for SRQ */
2563 		if (!mqp->ibqp.srq) {
2564 			if (mqp->rq.tail != mqp->rq.head) {
2565 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2566 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2567 				if (recv_mcq->mcq.comp &&
2568 				    mqp->ibqp.recv_cq->comp_handler) {
2569 					if (!recv_mcq->mcq.reset_notify_added) {
2570 						recv_mcq->mcq.reset_notify_added = 1;
2571 						list_add_tail(&recv_mcq->mcq.reset_notify,
2572 							      &cq_armed_list);
2573 					}
2574 				}
2575 				spin_unlock_irqrestore(&recv_mcq->lock,
2576 						       flags_cq);
2577 			}
2578 		}
2579 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2580 	}
2581 	/*At that point all inflight post send were put to be executed as of we
2582 	 * lock/unlock above locks Now need to arm all involved CQs.
2583 	 */
2584 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2585 		mcq->comp(mcq, NULL);
2586 	}
2587 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2588 }
2589 
2590 static void delay_drop_handler(struct work_struct *work)
2591 {
2592 	int err;
2593 	struct mlx5_ib_delay_drop *delay_drop =
2594 		container_of(work, struct mlx5_ib_delay_drop,
2595 			     delay_drop_work);
2596 
2597 	atomic_inc(&delay_drop->events_cnt);
2598 
2599 	mutex_lock(&delay_drop->lock);
2600 	err = mlx5_core_set_delay_drop(delay_drop->dev, delay_drop->timeout);
2601 	if (err) {
2602 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
2603 			     delay_drop->timeout);
2604 		delay_drop->activate = false;
2605 	}
2606 	mutex_unlock(&delay_drop->lock);
2607 }
2608 
2609 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2610 				 struct ib_event *ibev)
2611 {
2612 	u32 port = (eqe->data.port.port >> 4) & 0xf;
2613 
2614 	switch (eqe->sub_type) {
2615 	case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
2616 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2617 					    IB_LINK_LAYER_ETHERNET)
2618 			schedule_work(&ibdev->delay_drop.delay_drop_work);
2619 		break;
2620 	default: /* do nothing */
2621 		return;
2622 	}
2623 }
2624 
2625 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2626 			      struct ib_event *ibev)
2627 {
2628 	u32 port = (eqe->data.port.port >> 4) & 0xf;
2629 
2630 	ibev->element.port_num = port;
2631 
2632 	switch (eqe->sub_type) {
2633 	case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
2634 	case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
2635 	case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
2636 		/* In RoCE, port up/down events are handled in
2637 		 * mlx5_netdev_event().
2638 		 */
2639 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2640 					    IB_LINK_LAYER_ETHERNET)
2641 			return -EINVAL;
2642 
2643 		ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
2644 				IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2645 		break;
2646 
2647 	case MLX5_PORT_CHANGE_SUBTYPE_LID:
2648 		ibev->event = IB_EVENT_LID_CHANGE;
2649 		break;
2650 
2651 	case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
2652 		ibev->event = IB_EVENT_PKEY_CHANGE;
2653 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2654 		break;
2655 
2656 	case MLX5_PORT_CHANGE_SUBTYPE_GUID:
2657 		ibev->event = IB_EVENT_GID_CHANGE;
2658 		break;
2659 
2660 	case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
2661 		ibev->event = IB_EVENT_CLIENT_REREGISTER;
2662 		break;
2663 	default:
2664 		return -EINVAL;
2665 	}
2666 
2667 	return 0;
2668 }
2669 
2670 static void mlx5_ib_handle_event(struct work_struct *_work)
2671 {
2672 	struct mlx5_ib_event_work *work =
2673 		container_of(_work, struct mlx5_ib_event_work, work);
2674 	struct mlx5_ib_dev *ibdev;
2675 	struct ib_event ibev;
2676 	bool fatal = false;
2677 
2678 	if (work->is_slave) {
2679 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
2680 		if (!ibdev)
2681 			goto out;
2682 	} else {
2683 		ibdev = work->dev;
2684 	}
2685 
2686 	switch (work->event) {
2687 	case MLX5_DEV_EVENT_SYS_ERROR:
2688 		ibev.event = IB_EVENT_DEVICE_FATAL;
2689 		mlx5_ib_handle_internal_error(ibdev);
2690 		ibev.element.port_num  = (u8)(unsigned long)work->param;
2691 		fatal = true;
2692 		break;
2693 	case MLX5_EVENT_TYPE_PORT_CHANGE:
2694 		if (handle_port_change(ibdev, work->param, &ibev))
2695 			goto out;
2696 		break;
2697 	case MLX5_EVENT_TYPE_GENERAL_EVENT:
2698 		handle_general_event(ibdev, work->param, &ibev);
2699 		fallthrough;
2700 	default:
2701 		goto out;
2702 	}
2703 
2704 	ibev.device = &ibdev->ib_dev;
2705 
2706 	if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
2707 		mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
2708 		goto out;
2709 	}
2710 
2711 	if (ibdev->ib_active)
2712 		ib_dispatch_event(&ibev);
2713 
2714 	if (fatal)
2715 		ibdev->ib_active = false;
2716 out:
2717 	kfree(work);
2718 }
2719 
2720 static int mlx5_ib_event(struct notifier_block *nb,
2721 			 unsigned long event, void *param)
2722 {
2723 	struct mlx5_ib_event_work *work;
2724 
2725 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2726 	if (!work)
2727 		return NOTIFY_DONE;
2728 
2729 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2730 	work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
2731 	work->is_slave = false;
2732 	work->param = param;
2733 	work->event = event;
2734 
2735 	queue_work(mlx5_ib_event_wq, &work->work);
2736 
2737 	return NOTIFY_OK;
2738 }
2739 
2740 static int mlx5_ib_event_slave_port(struct notifier_block *nb,
2741 				    unsigned long event, void *param)
2742 {
2743 	struct mlx5_ib_event_work *work;
2744 
2745 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2746 	if (!work)
2747 		return NOTIFY_DONE;
2748 
2749 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2750 	work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
2751 	work->is_slave = true;
2752 	work->param = param;
2753 	work->event = event;
2754 	queue_work(mlx5_ib_event_wq, &work->work);
2755 
2756 	return NOTIFY_OK;
2757 }
2758 
2759 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2760 {
2761 	struct mlx5_hca_vport_context vport_ctx;
2762 	int err;
2763 	int port;
2764 
2765 	for (port = 1; port <= ARRAY_SIZE(dev->port_caps); port++) {
2766 		dev->port_caps[port - 1].has_smi = false;
2767 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
2768 		    MLX5_CAP_PORT_TYPE_IB) {
2769 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2770 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
2771 								   port, 0,
2772 								   &vport_ctx);
2773 				if (err) {
2774 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2775 						    port, err);
2776 					return err;
2777 				}
2778 				dev->port_caps[port - 1].has_smi =
2779 					vport_ctx.has_smi;
2780 			} else {
2781 				dev->port_caps[port - 1].has_smi = true;
2782 			}
2783 		}
2784 	}
2785 	return 0;
2786 }
2787 
2788 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2789 {
2790 	unsigned int port;
2791 
2792 	rdma_for_each_port (&dev->ib_dev, port)
2793 		mlx5_query_ext_port_caps(dev, port);
2794 }
2795 
2796 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
2797 {
2798 	switch (umr_fence_cap) {
2799 	case MLX5_CAP_UMR_FENCE_NONE:
2800 		return MLX5_FENCE_MODE_NONE;
2801 	case MLX5_CAP_UMR_FENCE_SMALL:
2802 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
2803 	default:
2804 		return MLX5_FENCE_MODE_STRONG_ORDERING;
2805 	}
2806 }
2807 
2808 static int mlx5_ib_dev_res_init(struct mlx5_ib_dev *dev)
2809 {
2810 	struct mlx5_ib_resources *devr = &dev->devr;
2811 	struct ib_srq_init_attr attr;
2812 	struct ib_device *ibdev;
2813 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2814 	int port;
2815 	int ret = 0;
2816 
2817 	ibdev = &dev->ib_dev;
2818 
2819 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
2820 		return -EOPNOTSUPP;
2821 
2822 	devr->p0 = ib_alloc_pd(ibdev, 0);
2823 	if (IS_ERR(devr->p0))
2824 		return PTR_ERR(devr->p0);
2825 
2826 	devr->c0 = ib_create_cq(ibdev, NULL, NULL, NULL, &cq_attr);
2827 	if (IS_ERR(devr->c0)) {
2828 		ret = PTR_ERR(devr->c0);
2829 		goto error1;
2830 	}
2831 
2832 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0);
2833 	if (ret)
2834 		goto error2;
2835 
2836 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0);
2837 	if (ret)
2838 		goto error3;
2839 
2840 	memset(&attr, 0, sizeof(attr));
2841 	attr.attr.max_sge = 1;
2842 	attr.attr.max_wr = 1;
2843 	attr.srq_type = IB_SRQT_XRC;
2844 	attr.ext.cq = devr->c0;
2845 
2846 	devr->s0 = ib_create_srq(devr->p0, &attr);
2847 	if (IS_ERR(devr->s0)) {
2848 		ret = PTR_ERR(devr->s0);
2849 		goto err_create;
2850 	}
2851 
2852 	memset(&attr, 0, sizeof(attr));
2853 	attr.attr.max_sge = 1;
2854 	attr.attr.max_wr = 1;
2855 	attr.srq_type = IB_SRQT_BASIC;
2856 
2857 	devr->s1 = ib_create_srq(devr->p0, &attr);
2858 	if (IS_ERR(devr->s1)) {
2859 		ret = PTR_ERR(devr->s1);
2860 		goto error6;
2861 	}
2862 
2863 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2864 		INIT_WORK(&devr->ports[port].pkey_change_work,
2865 			  pkey_change_handler);
2866 
2867 	return 0;
2868 
2869 error6:
2870 	ib_destroy_srq(devr->s0);
2871 err_create:
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 	ib_destroy_cq(devr->c0);
2877 error1:
2878 	ib_dealloc_pd(devr->p0);
2879 	return ret;
2880 }
2881 
2882 static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev)
2883 {
2884 	struct mlx5_ib_resources *devr = &dev->devr;
2885 	int port;
2886 
2887 	/*
2888 	 * Make sure no change P_Key work items are still executing.
2889 	 *
2890 	 * At this stage, the mlx5_ib_event should be unregistered
2891 	 * and it ensures that no new works are added.
2892 	 */
2893 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2894 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2895 
2896 	ib_destroy_srq(devr->s1);
2897 	ib_destroy_srq(devr->s0);
2898 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2899 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2900 	ib_destroy_cq(devr->c0);
2901 	ib_dealloc_pd(devr->p0);
2902 }
2903 
2904 static u32 get_core_cap_flags(struct ib_device *ibdev,
2905 			      struct mlx5_hca_vport_context *rep)
2906 {
2907 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2908 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2909 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2910 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2911 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
2912 	u32 ret = 0;
2913 
2914 	if (rep->grh_required)
2915 		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
2916 
2917 	if (ll == IB_LINK_LAYER_INFINIBAND)
2918 		return ret | RDMA_CORE_PORT_IBA_IB;
2919 
2920 	if (raw_support)
2921 		ret |= RDMA_CORE_PORT_RAW_PACKET;
2922 
2923 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2924 		return ret;
2925 
2926 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2927 		return ret;
2928 
2929 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2930 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2931 
2932 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2933 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2934 
2935 	return ret;
2936 }
2937 
2938 static int mlx5_port_immutable(struct ib_device *ibdev, u32 port_num,
2939 			       struct ib_port_immutable *immutable)
2940 {
2941 	struct ib_port_attr attr;
2942 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2943 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
2944 	struct mlx5_hca_vport_context rep = {0};
2945 	int err;
2946 
2947 	err = ib_query_port(ibdev, port_num, &attr);
2948 	if (err)
2949 		return err;
2950 
2951 	if (ll == IB_LINK_LAYER_INFINIBAND) {
2952 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
2953 						   &rep);
2954 		if (err)
2955 			return err;
2956 	}
2957 
2958 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2959 	immutable->gid_tbl_len = attr.gid_tbl_len;
2960 	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
2961 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2962 
2963 	return 0;
2964 }
2965 
2966 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u32 port_num,
2967 				   struct ib_port_immutable *immutable)
2968 {
2969 	struct ib_port_attr attr;
2970 	int err;
2971 
2972 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2973 
2974 	err = ib_query_port(ibdev, port_num, &attr);
2975 	if (err)
2976 		return err;
2977 
2978 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2979 	immutable->gid_tbl_len = attr.gid_tbl_len;
2980 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2981 
2982 	return 0;
2983 }
2984 
2985 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
2986 {
2987 	struct mlx5_ib_dev *dev =
2988 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2989 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
2990 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
2991 		 fw_rev_sub(dev->mdev));
2992 }
2993 
2994 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
2995 {
2996 	struct mlx5_core_dev *mdev = dev->mdev;
2997 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
2998 								 MLX5_FLOW_NAMESPACE_LAG);
2999 	struct mlx5_flow_table *ft;
3000 	int err;
3001 
3002 	if (!ns || !mlx5_lag_is_active(mdev))
3003 		return 0;
3004 
3005 	err = mlx5_cmd_create_vport_lag(mdev);
3006 	if (err)
3007 		return err;
3008 
3009 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3010 	if (IS_ERR(ft)) {
3011 		err = PTR_ERR(ft);
3012 		goto err_destroy_vport_lag;
3013 	}
3014 
3015 	dev->flow_db->lag_demux_ft = ft;
3016 	dev->lag_active = true;
3017 	return 0;
3018 
3019 err_destroy_vport_lag:
3020 	mlx5_cmd_destroy_vport_lag(mdev);
3021 	return err;
3022 }
3023 
3024 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3025 {
3026 	struct mlx5_core_dev *mdev = dev->mdev;
3027 
3028 	if (dev->lag_active) {
3029 		dev->lag_active = false;
3030 
3031 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
3032 		dev->flow_db->lag_demux_ft = NULL;
3033 
3034 		mlx5_cmd_destroy_vport_lag(mdev);
3035 	}
3036 }
3037 
3038 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u32 port_num)
3039 {
3040 	int err;
3041 
3042 	dev->port[port_num].roce.nb.notifier_call = mlx5_netdev_event;
3043 	err = register_netdevice_notifier(&dev->port[port_num].roce.nb);
3044 	if (err) {
3045 		dev->port[port_num].roce.nb.notifier_call = NULL;
3046 		return err;
3047 	}
3048 
3049 	return 0;
3050 }
3051 
3052 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u32 port_num)
3053 {
3054 	if (dev->port[port_num].roce.nb.notifier_call) {
3055 		unregister_netdevice_notifier(&dev->port[port_num].roce.nb);
3056 		dev->port[port_num].roce.nb.notifier_call = NULL;
3057 	}
3058 }
3059 
3060 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3061 {
3062 	int err;
3063 
3064 	if (!dev->is_rep && dev->profile != &raw_eth_profile) {
3065 		err = mlx5_nic_vport_enable_roce(dev->mdev);
3066 		if (err)
3067 			return err;
3068 	}
3069 
3070 	err = mlx5_eth_lag_init(dev);
3071 	if (err)
3072 		goto err_disable_roce;
3073 
3074 	return 0;
3075 
3076 err_disable_roce:
3077 	if (!dev->is_rep && dev->profile != &raw_eth_profile)
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 	if (!dev->is_rep && dev->profile != &raw_eth_profile)
3087 		mlx5_nic_vport_disable_roce(dev->mdev);
3088 }
3089 
3090 static int mlx5_ib_rn_get_params(struct ib_device *device, u32 port_num,
3091 				 enum rdma_netdev_t type,
3092 				 struct rdma_netdev_alloc_params *params)
3093 {
3094 	if (type != RDMA_NETDEV_IPOIB)
3095 		return -EOPNOTSUPP;
3096 
3097 	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
3098 }
3099 
3100 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
3101 				       size_t count, loff_t *pos)
3102 {
3103 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3104 	char lbuf[20];
3105 	int len;
3106 
3107 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
3108 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
3109 }
3110 
3111 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
3112 					size_t count, loff_t *pos)
3113 {
3114 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3115 	u32 timeout;
3116 	u32 var;
3117 
3118 	if (kstrtouint_from_user(buf, count, 0, &var))
3119 		return -EFAULT;
3120 
3121 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
3122 			1000);
3123 	if (timeout != var)
3124 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
3125 			    timeout);
3126 
3127 	delay_drop->timeout = timeout;
3128 
3129 	return count;
3130 }
3131 
3132 static const struct file_operations fops_delay_drop_timeout = {
3133 	.owner	= THIS_MODULE,
3134 	.open	= simple_open,
3135 	.write	= delay_drop_timeout_write,
3136 	.read	= delay_drop_timeout_read,
3137 };
3138 
3139 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
3140 				      struct mlx5_ib_multiport_info *mpi)
3141 {
3142 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3143 	struct mlx5_ib_port *port = &ibdev->port[port_num];
3144 	int comps;
3145 	int err;
3146 	int i;
3147 
3148 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3149 
3150 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
3151 
3152 	spin_lock(&port->mp.mpi_lock);
3153 	if (!mpi->ibdev) {
3154 		spin_unlock(&port->mp.mpi_lock);
3155 		return;
3156 	}
3157 
3158 	mpi->ibdev = NULL;
3159 
3160 	spin_unlock(&port->mp.mpi_lock);
3161 	if (mpi->mdev_events.notifier_call)
3162 		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
3163 	mpi->mdev_events.notifier_call = NULL;
3164 	mlx5_remove_netdev_notifier(ibdev, port_num);
3165 	spin_lock(&port->mp.mpi_lock);
3166 
3167 	comps = mpi->mdev_refcnt;
3168 	if (comps) {
3169 		mpi->unaffiliate = true;
3170 		init_completion(&mpi->unref_comp);
3171 		spin_unlock(&port->mp.mpi_lock);
3172 
3173 		for (i = 0; i < comps; i++)
3174 			wait_for_completion(&mpi->unref_comp);
3175 
3176 		spin_lock(&port->mp.mpi_lock);
3177 		mpi->unaffiliate = false;
3178 	}
3179 
3180 	port->mp.mpi = NULL;
3181 
3182 	spin_unlock(&port->mp.mpi_lock);
3183 
3184 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
3185 
3186 	mlx5_ib_dbg(ibdev, "unaffiliated port %u\n", port_num + 1);
3187 	/* Log an error, still needed to cleanup the pointers and add
3188 	 * it back to the list.
3189 	 */
3190 	if (err)
3191 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
3192 			    port_num + 1);
3193 
3194 	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
3195 }
3196 
3197 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
3198 				    struct mlx5_ib_multiport_info *mpi)
3199 {
3200 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3201 	int err;
3202 
3203 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3204 
3205 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
3206 	if (ibdev->port[port_num].mp.mpi) {
3207 		mlx5_ib_dbg(ibdev, "port %u already affiliated.\n",
3208 			    port_num + 1);
3209 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3210 		return false;
3211 	}
3212 
3213 	ibdev->port[port_num].mp.mpi = mpi;
3214 	mpi->ibdev = ibdev;
3215 	mpi->mdev_events.notifier_call = NULL;
3216 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3217 
3218 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
3219 	if (err)
3220 		goto unbind;
3221 
3222 	err = mlx5_add_netdev_notifier(ibdev, port_num);
3223 	if (err) {
3224 		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
3225 			    port_num + 1);
3226 		goto unbind;
3227 	}
3228 
3229 	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
3230 	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
3231 
3232 	mlx5_ib_init_cong_debugfs(ibdev, port_num);
3233 
3234 	return true;
3235 
3236 unbind:
3237 	mlx5_ib_unbind_slave_port(ibdev, mpi);
3238 	return false;
3239 }
3240 
3241 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
3242 {
3243 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3244 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3245 							  port_num + 1);
3246 	struct mlx5_ib_multiport_info *mpi;
3247 	int err;
3248 	u32 i;
3249 
3250 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3251 		return 0;
3252 
3253 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
3254 						     &dev->sys_image_guid);
3255 	if (err)
3256 		return err;
3257 
3258 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3259 	if (err)
3260 		return err;
3261 
3262 	mutex_lock(&mlx5_ib_multiport_mutex);
3263 	for (i = 0; i < dev->num_ports; i++) {
3264 		bool bound = false;
3265 
3266 		/* build a stub multiport info struct for the native port. */
3267 		if (i == port_num) {
3268 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
3269 			if (!mpi) {
3270 				mutex_unlock(&mlx5_ib_multiport_mutex);
3271 				mlx5_nic_vport_disable_roce(dev->mdev);
3272 				return -ENOMEM;
3273 			}
3274 
3275 			mpi->is_master = true;
3276 			mpi->mdev = dev->mdev;
3277 			mpi->sys_image_guid = dev->sys_image_guid;
3278 			dev->port[i].mp.mpi = mpi;
3279 			mpi->ibdev = dev;
3280 			mpi = NULL;
3281 			continue;
3282 		}
3283 
3284 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
3285 				    list) {
3286 			if (dev->sys_image_guid == mpi->sys_image_guid &&
3287 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
3288 				bound = mlx5_ib_bind_slave_port(dev, mpi);
3289 			}
3290 
3291 			if (bound) {
3292 				dev_dbg(mpi->mdev->device,
3293 					"removing port from unaffiliated list.\n");
3294 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
3295 				list_del(&mpi->list);
3296 				break;
3297 			}
3298 		}
3299 		if (!bound)
3300 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
3301 				    i + 1);
3302 	}
3303 
3304 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
3305 	mutex_unlock(&mlx5_ib_multiport_mutex);
3306 	return err;
3307 }
3308 
3309 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
3310 {
3311 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3312 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3313 							  port_num + 1);
3314 	u32 i;
3315 
3316 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3317 		return;
3318 
3319 	mutex_lock(&mlx5_ib_multiport_mutex);
3320 	for (i = 0; i < dev->num_ports; i++) {
3321 		if (dev->port[i].mp.mpi) {
3322 			/* Destroy the native port stub */
3323 			if (i == port_num) {
3324 				kfree(dev->port[i].mp.mpi);
3325 				dev->port[i].mp.mpi = NULL;
3326 			} else {
3327 				mlx5_ib_dbg(dev, "unbinding port_num: %u\n",
3328 					    i + 1);
3329 				list_add_tail(&dev->port[i].mp.mpi->list,
3330 					      &mlx5_ib_unaffiliated_port_list);
3331 				mlx5_ib_unbind_slave_port(dev,
3332 							  dev->port[i].mp.mpi);
3333 			}
3334 		}
3335 	}
3336 
3337 	mlx5_ib_dbg(dev, "removing from devlist\n");
3338 	list_del(&dev->ib_dev_list);
3339 	mutex_unlock(&mlx5_ib_multiport_mutex);
3340 
3341 	mlx5_nic_vport_disable_roce(dev->mdev);
3342 }
3343 
3344 static int mmap_obj_cleanup(struct ib_uobject *uobject,
3345 			    enum rdma_remove_reason why,
3346 			    struct uverbs_attr_bundle *attrs)
3347 {
3348 	struct mlx5_user_mmap_entry *obj = uobject->object;
3349 
3350 	rdma_user_mmap_entry_remove(&obj->rdma_entry);
3351 	return 0;
3352 }
3353 
3354 static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c,
3355 					    struct mlx5_user_mmap_entry *entry,
3356 					    size_t length)
3357 {
3358 	return rdma_user_mmap_entry_insert_range(
3359 		&c->ibucontext, &entry->rdma_entry, length,
3360 		(MLX5_IB_MMAP_OFFSET_START << 16),
3361 		((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1));
3362 }
3363 
3364 static struct mlx5_user_mmap_entry *
3365 alloc_var_entry(struct mlx5_ib_ucontext *c)
3366 {
3367 	struct mlx5_user_mmap_entry *entry;
3368 	struct mlx5_var_table *var_table;
3369 	u32 page_idx;
3370 	int err;
3371 
3372 	var_table = &to_mdev(c->ibucontext.device)->var_table;
3373 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3374 	if (!entry)
3375 		return ERR_PTR(-ENOMEM);
3376 
3377 	mutex_lock(&var_table->bitmap_lock);
3378 	page_idx = find_first_zero_bit(var_table->bitmap,
3379 				       var_table->num_var_hw_entries);
3380 	if (page_idx >= var_table->num_var_hw_entries) {
3381 		err = -ENOSPC;
3382 		mutex_unlock(&var_table->bitmap_lock);
3383 		goto end;
3384 	}
3385 
3386 	set_bit(page_idx, var_table->bitmap);
3387 	mutex_unlock(&var_table->bitmap_lock);
3388 
3389 	entry->address = var_table->hw_start_addr +
3390 				(page_idx * var_table->stride_size);
3391 	entry->page_idx = page_idx;
3392 	entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
3393 
3394 	err = mlx5_rdma_user_mmap_entry_insert(c, entry,
3395 					       var_table->stride_size);
3396 	if (err)
3397 		goto err_insert;
3398 
3399 	return entry;
3400 
3401 err_insert:
3402 	mutex_lock(&var_table->bitmap_lock);
3403 	clear_bit(page_idx, var_table->bitmap);
3404 	mutex_unlock(&var_table->bitmap_lock);
3405 end:
3406 	kfree(entry);
3407 	return ERR_PTR(err);
3408 }
3409 
3410 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
3411 	struct uverbs_attr_bundle *attrs)
3412 {
3413 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3414 		attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3415 	struct mlx5_ib_ucontext *c;
3416 	struct mlx5_user_mmap_entry *entry;
3417 	u64 mmap_offset;
3418 	u32 length;
3419 	int err;
3420 
3421 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3422 	if (IS_ERR(c))
3423 		return PTR_ERR(c);
3424 
3425 	entry = alloc_var_entry(c);
3426 	if (IS_ERR(entry))
3427 		return PTR_ERR(entry);
3428 
3429 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3430 	length = entry->rdma_entry.npages * PAGE_SIZE;
3431 	uobj->object = entry;
3432 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3433 
3434 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3435 			     &mmap_offset, sizeof(mmap_offset));
3436 	if (err)
3437 		return err;
3438 
3439 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3440 			     &entry->page_idx, sizeof(entry->page_idx));
3441 	if (err)
3442 		return err;
3443 
3444 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3445 			     &length, sizeof(length));
3446 	return err;
3447 }
3448 
3449 DECLARE_UVERBS_NAMED_METHOD(
3450 	MLX5_IB_METHOD_VAR_OBJ_ALLOC,
3451 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
3452 			MLX5_IB_OBJECT_VAR,
3453 			UVERBS_ACCESS_NEW,
3454 			UA_MANDATORY),
3455 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3456 			   UVERBS_ATTR_TYPE(u32),
3457 			   UA_MANDATORY),
3458 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3459 			   UVERBS_ATTR_TYPE(u32),
3460 			   UA_MANDATORY),
3461 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3462 			    UVERBS_ATTR_TYPE(u64),
3463 			    UA_MANDATORY));
3464 
3465 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3466 	MLX5_IB_METHOD_VAR_OBJ_DESTROY,
3467 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
3468 			MLX5_IB_OBJECT_VAR,
3469 			UVERBS_ACCESS_DESTROY,
3470 			UA_MANDATORY));
3471 
3472 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
3473 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3474 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
3475 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
3476 
3477 static bool var_is_supported(struct ib_device *device)
3478 {
3479 	struct mlx5_ib_dev *dev = to_mdev(device);
3480 
3481 	return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3482 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
3483 }
3484 
3485 static struct mlx5_user_mmap_entry *
3486 alloc_uar_entry(struct mlx5_ib_ucontext *c,
3487 		enum mlx5_ib_uapi_uar_alloc_type alloc_type)
3488 {
3489 	struct mlx5_user_mmap_entry *entry;
3490 	struct mlx5_ib_dev *dev;
3491 	u32 uar_index;
3492 	int err;
3493 
3494 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3495 	if (!entry)
3496 		return ERR_PTR(-ENOMEM);
3497 
3498 	dev = to_mdev(c->ibucontext.device);
3499 	err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index, c->devx_uid);
3500 	if (err)
3501 		goto end;
3502 
3503 	entry->page_idx = uar_index;
3504 	entry->address = uar_index2paddress(dev, uar_index);
3505 	if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3506 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC;
3507 	else
3508 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC;
3509 
3510 	err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE);
3511 	if (err)
3512 		goto err_insert;
3513 
3514 	return entry;
3515 
3516 err_insert:
3517 	mlx5_cmd_uar_dealloc(dev->mdev, uar_index, c->devx_uid);
3518 end:
3519 	kfree(entry);
3520 	return ERR_PTR(err);
3521 }
3522 
3523 static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)(
3524 	struct uverbs_attr_bundle *attrs)
3525 {
3526 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3527 		attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3528 	enum mlx5_ib_uapi_uar_alloc_type alloc_type;
3529 	struct mlx5_ib_ucontext *c;
3530 	struct mlx5_user_mmap_entry *entry;
3531 	u64 mmap_offset;
3532 	u32 length;
3533 	int err;
3534 
3535 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3536 	if (IS_ERR(c))
3537 		return PTR_ERR(c);
3538 
3539 	err = uverbs_get_const(&alloc_type, attrs,
3540 			       MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE);
3541 	if (err)
3542 		return err;
3543 
3544 	if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF &&
3545 	    alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC)
3546 		return -EOPNOTSUPP;
3547 
3548 	if (!to_mdev(c->ibucontext.device)->wc_support &&
3549 	    alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3550 		return -EOPNOTSUPP;
3551 
3552 	entry = alloc_uar_entry(c, alloc_type);
3553 	if (IS_ERR(entry))
3554 		return PTR_ERR(entry);
3555 
3556 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3557 	length = entry->rdma_entry.npages * PAGE_SIZE;
3558 	uobj->object = entry;
3559 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3560 
3561 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3562 			     &mmap_offset, sizeof(mmap_offset));
3563 	if (err)
3564 		return err;
3565 
3566 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3567 			     &entry->page_idx, sizeof(entry->page_idx));
3568 	if (err)
3569 		return err;
3570 
3571 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3572 			     &length, sizeof(length));
3573 	return err;
3574 }
3575 
3576 DECLARE_UVERBS_NAMED_METHOD(
3577 	MLX5_IB_METHOD_UAR_OBJ_ALLOC,
3578 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE,
3579 			MLX5_IB_OBJECT_UAR,
3580 			UVERBS_ACCESS_NEW,
3581 			UA_MANDATORY),
3582 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE,
3583 			     enum mlx5_ib_uapi_uar_alloc_type,
3584 			     UA_MANDATORY),
3585 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3586 			   UVERBS_ATTR_TYPE(u32),
3587 			   UA_MANDATORY),
3588 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3589 			   UVERBS_ATTR_TYPE(u32),
3590 			   UA_MANDATORY),
3591 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3592 			    UVERBS_ATTR_TYPE(u64),
3593 			    UA_MANDATORY));
3594 
3595 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3596 	MLX5_IB_METHOD_UAR_OBJ_DESTROY,
3597 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE,
3598 			MLX5_IB_OBJECT_UAR,
3599 			UVERBS_ACCESS_DESTROY,
3600 			UA_MANDATORY));
3601 
3602 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR,
3603 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3604 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC),
3605 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY));
3606 
3607 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3608 	mlx5_ib_flow_action,
3609 	UVERBS_OBJECT_FLOW_ACTION,
3610 	UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
3611 	UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
3612 			     enum mlx5_ib_uapi_flow_action_flags));
3613 
3614 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3615 	mlx5_ib_query_context,
3616 	UVERBS_OBJECT_DEVICE,
3617 	UVERBS_METHOD_QUERY_CONTEXT,
3618 	UVERBS_ATTR_PTR_OUT(
3619 		MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
3620 		UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp,
3621 				   dump_fill_mkey),
3622 		UA_MANDATORY));
3623 
3624 static const struct uapi_definition mlx5_ib_defs[] = {
3625 	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
3626 	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
3627 	UAPI_DEF_CHAIN(mlx5_ib_qos_defs),
3628 	UAPI_DEF_CHAIN(mlx5_ib_std_types_defs),
3629 	UAPI_DEF_CHAIN(mlx5_ib_dm_defs),
3630 
3631 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_FLOW_ACTION,
3632 				&mlx5_ib_flow_action),
3633 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context),
3634 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
3635 				UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
3636 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR),
3637 	{}
3638 };
3639 
3640 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
3641 {
3642 	mlx5_ib_cleanup_multiport_master(dev);
3643 	WARN_ON(!xa_empty(&dev->odp_mkeys));
3644 	mutex_destroy(&dev->cap_mask_mutex);
3645 	WARN_ON(!xa_empty(&dev->sig_mrs));
3646 	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
3647 }
3648 
3649 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
3650 {
3651 	struct mlx5_core_dev *mdev = dev->mdev;
3652 	int err;
3653 	int i;
3654 
3655 	dev->ib_dev.node_type = RDMA_NODE_IB_CA;
3656 	dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
3657 	dev->ib_dev.phys_port_cnt = dev->num_ports;
3658 	dev->ib_dev.dev.parent = mdev->device;
3659 	dev->ib_dev.lag_flags = RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
3660 
3661 	for (i = 0; i < dev->num_ports; i++) {
3662 		spin_lock_init(&dev->port[i].mp.mpi_lock);
3663 		rwlock_init(&dev->port[i].roce.netdev_lock);
3664 		dev->port[i].roce.dev = dev;
3665 		dev->port[i].roce.native_port_num = i + 1;
3666 		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
3667 	}
3668 
3669 	err = mlx5_ib_init_multiport_master(dev);
3670 	if (err)
3671 		return err;
3672 
3673 	err = set_has_smi_cap(dev);
3674 	if (err)
3675 		goto err_mp;
3676 
3677 	err = mlx5_query_max_pkeys(&dev->ib_dev, &dev->pkey_table_len);
3678 	if (err)
3679 		goto err_mp;
3680 
3681 	if (mlx5_use_mad_ifc(dev))
3682 		get_ext_port_caps(dev);
3683 
3684 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_count(mdev);
3685 
3686 	mutex_init(&dev->cap_mask_mutex);
3687 	INIT_LIST_HEAD(&dev->qp_list);
3688 	spin_lock_init(&dev->reset_flow_resource_lock);
3689 	xa_init(&dev->odp_mkeys);
3690 	xa_init(&dev->sig_mrs);
3691 	atomic_set(&dev->mkey_var, 0);
3692 
3693 	spin_lock_init(&dev->dm.lock);
3694 	dev->dm.dev = mdev;
3695 	return 0;
3696 
3697 err_mp:
3698 	mlx5_ib_cleanup_multiport_master(dev);
3699 	return err;
3700 }
3701 
3702 static int mlx5_ib_enable_driver(struct ib_device *dev)
3703 {
3704 	struct mlx5_ib_dev *mdev = to_mdev(dev);
3705 	int ret;
3706 
3707 	ret = mlx5_ib_test_wc(mdev);
3708 	mlx5_ib_dbg(mdev, "Write-Combining %s",
3709 		    mdev->wc_support ? "supported" : "not supported");
3710 
3711 	return ret;
3712 }
3713 
3714 static const struct ib_device_ops mlx5_ib_dev_ops = {
3715 	.owner = THIS_MODULE,
3716 	.driver_id = RDMA_DRIVER_MLX5,
3717 	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
3718 
3719 	.add_gid = mlx5_ib_add_gid,
3720 	.alloc_mr = mlx5_ib_alloc_mr,
3721 	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
3722 	.alloc_pd = mlx5_ib_alloc_pd,
3723 	.alloc_ucontext = mlx5_ib_alloc_ucontext,
3724 	.attach_mcast = mlx5_ib_mcg_attach,
3725 	.check_mr_status = mlx5_ib_check_mr_status,
3726 	.create_ah = mlx5_ib_create_ah,
3727 	.create_cq = mlx5_ib_create_cq,
3728 	.create_qp = mlx5_ib_create_qp,
3729 	.create_srq = mlx5_ib_create_srq,
3730 	.create_user_ah = mlx5_ib_create_ah,
3731 	.dealloc_pd = mlx5_ib_dealloc_pd,
3732 	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
3733 	.del_gid = mlx5_ib_del_gid,
3734 	.dereg_mr = mlx5_ib_dereg_mr,
3735 	.destroy_ah = mlx5_ib_destroy_ah,
3736 	.destroy_cq = mlx5_ib_destroy_cq,
3737 	.destroy_qp = mlx5_ib_destroy_qp,
3738 	.destroy_srq = mlx5_ib_destroy_srq,
3739 	.detach_mcast = mlx5_ib_mcg_detach,
3740 	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
3741 	.drain_rq = mlx5_ib_drain_rq,
3742 	.drain_sq = mlx5_ib_drain_sq,
3743 	.device_group = &mlx5_attr_group,
3744 	.enable_driver = mlx5_ib_enable_driver,
3745 	.get_dev_fw_str = get_dev_fw_str,
3746 	.get_dma_mr = mlx5_ib_get_dma_mr,
3747 	.get_link_layer = mlx5_ib_port_link_layer,
3748 	.map_mr_sg = mlx5_ib_map_mr_sg,
3749 	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
3750 	.mmap = mlx5_ib_mmap,
3751 	.mmap_free = mlx5_ib_mmap_free,
3752 	.modify_cq = mlx5_ib_modify_cq,
3753 	.modify_device = mlx5_ib_modify_device,
3754 	.modify_port = mlx5_ib_modify_port,
3755 	.modify_qp = mlx5_ib_modify_qp,
3756 	.modify_srq = mlx5_ib_modify_srq,
3757 	.poll_cq = mlx5_ib_poll_cq,
3758 	.post_recv = mlx5_ib_post_recv_nodrain,
3759 	.post_send = mlx5_ib_post_send_nodrain,
3760 	.post_srq_recv = mlx5_ib_post_srq_recv,
3761 	.process_mad = mlx5_ib_process_mad,
3762 	.query_ah = mlx5_ib_query_ah,
3763 	.query_device = mlx5_ib_query_device,
3764 	.query_gid = mlx5_ib_query_gid,
3765 	.query_pkey = mlx5_ib_query_pkey,
3766 	.query_qp = mlx5_ib_query_qp,
3767 	.query_srq = mlx5_ib_query_srq,
3768 	.query_ucontext = mlx5_ib_query_ucontext,
3769 	.reg_user_mr = mlx5_ib_reg_user_mr,
3770 	.reg_user_mr_dmabuf = mlx5_ib_reg_user_mr_dmabuf,
3771 	.req_notify_cq = mlx5_ib_arm_cq,
3772 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
3773 	.resize_cq = mlx5_ib_resize_cq,
3774 
3775 	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
3776 	INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs),
3777 	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
3778 	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
3779 	INIT_RDMA_OBJ_SIZE(ib_qp, mlx5_ib_qp, ibqp),
3780 	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
3781 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
3782 };
3783 
3784 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
3785 	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
3786 };
3787 
3788 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
3789 	.get_vf_config = mlx5_ib_get_vf_config,
3790 	.get_vf_guid = mlx5_ib_get_vf_guid,
3791 	.get_vf_stats = mlx5_ib_get_vf_stats,
3792 	.set_vf_guid = mlx5_ib_set_vf_guid,
3793 	.set_vf_link_state = mlx5_ib_set_vf_link_state,
3794 };
3795 
3796 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
3797 	.alloc_mw = mlx5_ib_alloc_mw,
3798 	.dealloc_mw = mlx5_ib_dealloc_mw,
3799 
3800 	INIT_RDMA_OBJ_SIZE(ib_mw, mlx5_ib_mw, ibmw),
3801 };
3802 
3803 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
3804 	.alloc_xrcd = mlx5_ib_alloc_xrcd,
3805 	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
3806 
3807 	INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd),
3808 };
3809 
3810 static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
3811 {
3812 	struct mlx5_core_dev *mdev = dev->mdev;
3813 	struct mlx5_var_table *var_table = &dev->var_table;
3814 	u8 log_doorbell_bar_size;
3815 	u8 log_doorbell_stride;
3816 	u64 bar_size;
3817 
3818 	log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3819 					log_doorbell_bar_size);
3820 	log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3821 					log_doorbell_stride);
3822 	var_table->hw_start_addr = dev->mdev->bar_addr +
3823 				MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
3824 					doorbell_bar_offset);
3825 	bar_size = (1ULL << log_doorbell_bar_size) * 4096;
3826 	var_table->stride_size = 1ULL << log_doorbell_stride;
3827 	var_table->num_var_hw_entries = div_u64(bar_size,
3828 						var_table->stride_size);
3829 	mutex_init(&var_table->bitmap_lock);
3830 	var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
3831 					  GFP_KERNEL);
3832 	return (var_table->bitmap) ? 0 : -ENOMEM;
3833 }
3834 
3835 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
3836 {
3837 	bitmap_free(dev->var_table.bitmap);
3838 }
3839 
3840 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
3841 {
3842 	struct mlx5_core_dev *mdev = dev->mdev;
3843 	int err;
3844 
3845 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
3846 	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
3847 		ib_set_device_ops(&dev->ib_dev,
3848 				  &mlx5_ib_dev_ipoib_enhanced_ops);
3849 
3850 	if (mlx5_core_is_pf(mdev))
3851 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
3852 
3853 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
3854 
3855 	if (MLX5_CAP_GEN(mdev, imaicl))
3856 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
3857 
3858 	if (MLX5_CAP_GEN(mdev, xrc))
3859 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
3860 
3861 	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
3862 	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3863 	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
3864 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
3865 
3866 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
3867 
3868 	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
3869 		dev->ib_dev.driver_def = mlx5_ib_defs;
3870 
3871 	err = init_node_data(dev);
3872 	if (err)
3873 		return err;
3874 
3875 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
3876 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
3877 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
3878 		mutex_init(&dev->lb.mutex);
3879 
3880 	if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3881 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
3882 		err = mlx5_ib_init_var_table(dev);
3883 		if (err)
3884 			return err;
3885 	}
3886 
3887 	dev->ib_dev.use_cq_dim = true;
3888 
3889 	return 0;
3890 }
3891 
3892 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
3893 	.get_port_immutable = mlx5_port_immutable,
3894 	.query_port = mlx5_ib_query_port,
3895 };
3896 
3897 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
3898 {
3899 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
3900 	return 0;
3901 }
3902 
3903 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
3904 	.get_port_immutable = mlx5_port_rep_immutable,
3905 	.query_port = mlx5_ib_rep_query_port,
3906 	.query_pkey = mlx5_ib_rep_query_pkey,
3907 };
3908 
3909 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
3910 {
3911 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
3912 	return 0;
3913 }
3914 
3915 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
3916 	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
3917 	.create_wq = mlx5_ib_create_wq,
3918 	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
3919 	.destroy_wq = mlx5_ib_destroy_wq,
3920 	.get_netdev = mlx5_ib_get_netdev,
3921 	.modify_wq = mlx5_ib_modify_wq,
3922 
3923 	INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table,
3924 			   ib_rwq_ind_tbl),
3925 };
3926 
3927 static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev)
3928 {
3929 	struct mlx5_core_dev *mdev = dev->mdev;
3930 	enum rdma_link_layer ll;
3931 	int port_type_cap;
3932 	u32 port_num = 0;
3933 	int err;
3934 
3935 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3936 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3937 
3938 	if (ll == IB_LINK_LAYER_ETHERNET) {
3939 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
3940 
3941 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3942 
3943 		/* Register only for native ports */
3944 		err = mlx5_add_netdev_notifier(dev, port_num);
3945 		if (err)
3946 			return err;
3947 
3948 		err = mlx5_enable_eth(dev);
3949 		if (err)
3950 			goto cleanup;
3951 	}
3952 
3953 	return 0;
3954 cleanup:
3955 	mlx5_remove_netdev_notifier(dev, port_num);
3956 	return err;
3957 }
3958 
3959 static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev)
3960 {
3961 	struct mlx5_core_dev *mdev = dev->mdev;
3962 	enum rdma_link_layer ll;
3963 	int port_type_cap;
3964 	u32 port_num;
3965 
3966 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3967 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3968 
3969 	if (ll == IB_LINK_LAYER_ETHERNET) {
3970 		mlx5_disable_eth(dev);
3971 
3972 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3973 		mlx5_remove_netdev_notifier(dev, port_num);
3974 	}
3975 }
3976 
3977 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
3978 {
3979 	mlx5_ib_init_cong_debugfs(dev,
3980 				  mlx5_core_native_port_num(dev->mdev) - 1);
3981 	return 0;
3982 }
3983 
3984 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
3985 {
3986 	mlx5_ib_cleanup_cong_debugfs(dev,
3987 				     mlx5_core_native_port_num(dev->mdev) - 1);
3988 }
3989 
3990 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
3991 {
3992 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
3993 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
3994 }
3995 
3996 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
3997 {
3998 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
3999 }
4000 
4001 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
4002 {
4003 	int err;
4004 
4005 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
4006 	if (err)
4007 		return err;
4008 
4009 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
4010 	if (err)
4011 		mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4012 
4013 	return err;
4014 }
4015 
4016 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
4017 {
4018 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4019 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4020 }
4021 
4022 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
4023 {
4024 	const char *name;
4025 
4026 	if (!mlx5_lag_is_active(dev->mdev))
4027 		name = "mlx5_%d";
4028 	else
4029 		name = "mlx5_bond_%d";
4030 	return ib_register_device(&dev->ib_dev, name, &dev->mdev->pdev->dev);
4031 }
4032 
4033 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
4034 {
4035 	int err;
4036 
4037 	err = mlx5_mr_cache_cleanup(dev);
4038 	if (err)
4039 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
4040 
4041 	if (dev->umrc.qp)
4042 		ib_destroy_qp(dev->umrc.qp);
4043 	if (dev->umrc.cq)
4044 		ib_free_cq(dev->umrc.cq);
4045 	if (dev->umrc.pd)
4046 		ib_dealloc_pd(dev->umrc.pd);
4047 }
4048 
4049 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
4050 {
4051 	ib_unregister_device(&dev->ib_dev);
4052 }
4053 
4054 enum {
4055 	MAX_UMR_WR = 128,
4056 };
4057 
4058 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
4059 {
4060 	struct ib_qp_init_attr *init_attr = NULL;
4061 	struct ib_qp_attr *attr = NULL;
4062 	struct ib_pd *pd;
4063 	struct ib_cq *cq;
4064 	struct ib_qp *qp;
4065 	int ret;
4066 
4067 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
4068 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
4069 	if (!attr || !init_attr) {
4070 		ret = -ENOMEM;
4071 		goto error_0;
4072 	}
4073 
4074 	pd = ib_alloc_pd(&dev->ib_dev, 0);
4075 	if (IS_ERR(pd)) {
4076 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
4077 		ret = PTR_ERR(pd);
4078 		goto error_0;
4079 	}
4080 
4081 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
4082 	if (IS_ERR(cq)) {
4083 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
4084 		ret = PTR_ERR(cq);
4085 		goto error_2;
4086 	}
4087 
4088 	init_attr->send_cq = cq;
4089 	init_attr->recv_cq = cq;
4090 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
4091 	init_attr->cap.max_send_wr = MAX_UMR_WR;
4092 	init_attr->cap.max_send_sge = 1;
4093 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
4094 	init_attr->port_num = 1;
4095 	qp = ib_create_qp(pd, init_attr);
4096 	if (IS_ERR(qp)) {
4097 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
4098 		ret = PTR_ERR(qp);
4099 		goto error_3;
4100 	}
4101 
4102 	attr->qp_state = IB_QPS_INIT;
4103 	attr->port_num = 1;
4104 	ret = ib_modify_qp(qp, attr,
4105 			   IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT);
4106 	if (ret) {
4107 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
4108 		goto error_4;
4109 	}
4110 
4111 	memset(attr, 0, sizeof(*attr));
4112 	attr->qp_state = IB_QPS_RTR;
4113 	attr->path_mtu = IB_MTU_256;
4114 
4115 	ret = ib_modify_qp(qp, attr, IB_QP_STATE);
4116 	if (ret) {
4117 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
4118 		goto error_4;
4119 	}
4120 
4121 	memset(attr, 0, sizeof(*attr));
4122 	attr->qp_state = IB_QPS_RTS;
4123 	ret = ib_modify_qp(qp, attr, IB_QP_STATE);
4124 	if (ret) {
4125 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
4126 		goto error_4;
4127 	}
4128 
4129 	dev->umrc.qp = qp;
4130 	dev->umrc.cq = cq;
4131 	dev->umrc.pd = pd;
4132 
4133 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
4134 	ret = mlx5_mr_cache_init(dev);
4135 	if (ret) {
4136 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4137 		goto error_4;
4138 	}
4139 
4140 	kfree(attr);
4141 	kfree(init_attr);
4142 
4143 	return 0;
4144 
4145 error_4:
4146 	ib_destroy_qp(qp);
4147 	dev->umrc.qp = NULL;
4148 
4149 error_3:
4150 	ib_free_cq(cq);
4151 	dev->umrc.cq = NULL;
4152 
4153 error_2:
4154 	ib_dealloc_pd(pd);
4155 	dev->umrc.pd = NULL;
4156 
4157 error_0:
4158 	kfree(attr);
4159 	kfree(init_attr);
4160 	return ret;
4161 }
4162 
4163 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
4164 {
4165 	struct dentry *root;
4166 
4167 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4168 		return 0;
4169 
4170 	mutex_init(&dev->delay_drop.lock);
4171 	dev->delay_drop.dev = dev;
4172 	dev->delay_drop.activate = false;
4173 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4174 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4175 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4176 	atomic_set(&dev->delay_drop.events_cnt, 0);
4177 
4178 	if (!mlx5_debugfs_root)
4179 		return 0;
4180 
4181 	root = debugfs_create_dir("delay_drop", dev->mdev->priv.dbg_root);
4182 	dev->delay_drop.dir_debugfs = root;
4183 
4184 	debugfs_create_atomic_t("num_timeout_events", 0400, root,
4185 				&dev->delay_drop.events_cnt);
4186 	debugfs_create_atomic_t("num_rqs", 0400, root,
4187 				&dev->delay_drop.rqs_cnt);
4188 	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
4189 			    &fops_delay_drop_timeout);
4190 	return 0;
4191 }
4192 
4193 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
4194 {
4195 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4196 		return;
4197 
4198 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4199 	if (!dev->delay_drop.dir_debugfs)
4200 		return;
4201 
4202 	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
4203 	dev->delay_drop.dir_debugfs = NULL;
4204 }
4205 
4206 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
4207 {
4208 	dev->mdev_events.notifier_call = mlx5_ib_event;
4209 	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
4210 	return 0;
4211 }
4212 
4213 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
4214 {
4215 	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
4216 }
4217 
4218 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
4219 		      const struct mlx5_ib_profile *profile,
4220 		      int stage)
4221 {
4222 	dev->ib_active = false;
4223 
4224 	/* Number of stages to cleanup */
4225 	while (stage) {
4226 		stage--;
4227 		if (profile->stage[stage].cleanup)
4228 			profile->stage[stage].cleanup(dev);
4229 	}
4230 
4231 	kfree(dev->port);
4232 	ib_dealloc_device(&dev->ib_dev);
4233 }
4234 
4235 int __mlx5_ib_add(struct mlx5_ib_dev *dev,
4236 		  const struct mlx5_ib_profile *profile)
4237 {
4238 	int err;
4239 	int i;
4240 
4241 	dev->profile = profile;
4242 
4243 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
4244 		if (profile->stage[i].init) {
4245 			err = profile->stage[i].init(dev);
4246 			if (err)
4247 				goto err_out;
4248 		}
4249 	}
4250 
4251 	dev->ib_active = true;
4252 	return 0;
4253 
4254 err_out:
4255 	/* Clean up stages which were initialized */
4256 	while (i) {
4257 		i--;
4258 		if (profile->stage[i].cleanup)
4259 			profile->stage[i].cleanup(dev);
4260 	}
4261 	return -ENOMEM;
4262 }
4263 
4264 static const struct mlx5_ib_profile pf_profile = {
4265 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4266 		     mlx5_ib_stage_init_init,
4267 		     mlx5_ib_stage_init_cleanup),
4268 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4269 		     mlx5_ib_fs_init,
4270 		     mlx5_ib_fs_cleanup),
4271 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4272 		     mlx5_ib_stage_caps_init,
4273 		     mlx5_ib_stage_caps_cleanup),
4274 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4275 		     mlx5_ib_stage_non_default_cb,
4276 		     NULL),
4277 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4278 		     mlx5_ib_roce_init,
4279 		     mlx5_ib_roce_cleanup),
4280 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4281 		     mlx5_init_qp_table,
4282 		     mlx5_cleanup_qp_table),
4283 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4284 		     mlx5_init_srq_table,
4285 		     mlx5_cleanup_srq_table),
4286 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4287 		     mlx5_ib_dev_res_init,
4288 		     mlx5_ib_dev_res_cleanup),
4289 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4290 		     mlx5_ib_stage_dev_notifier_init,
4291 		     mlx5_ib_stage_dev_notifier_cleanup),
4292 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
4293 		     mlx5_ib_odp_init_one,
4294 		     mlx5_ib_odp_cleanup_one),
4295 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4296 		     mlx5_ib_counters_init,
4297 		     mlx5_ib_counters_cleanup),
4298 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4299 		     mlx5_ib_stage_cong_debugfs_init,
4300 		     mlx5_ib_stage_cong_debugfs_cleanup),
4301 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4302 		     mlx5_ib_stage_uar_init,
4303 		     mlx5_ib_stage_uar_cleanup),
4304 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4305 		     mlx5_ib_stage_bfrag_init,
4306 		     mlx5_ib_stage_bfrag_cleanup),
4307 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4308 		     NULL,
4309 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4310 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4311 		     mlx5_ib_devx_init,
4312 		     mlx5_ib_devx_cleanup),
4313 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4314 		     mlx5_ib_stage_ib_reg_init,
4315 		     mlx5_ib_stage_ib_reg_cleanup),
4316 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4317 		     mlx5_ib_stage_post_ib_reg_umr_init,
4318 		     NULL),
4319 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4320 		     mlx5_ib_stage_delay_drop_init,
4321 		     mlx5_ib_stage_delay_drop_cleanup),
4322 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4323 		     mlx5_ib_restrack_init,
4324 		     NULL),
4325 };
4326 
4327 const struct mlx5_ib_profile raw_eth_profile = {
4328 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4329 		     mlx5_ib_stage_init_init,
4330 		     mlx5_ib_stage_init_cleanup),
4331 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4332 		     mlx5_ib_fs_init,
4333 		     mlx5_ib_fs_cleanup),
4334 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4335 		     mlx5_ib_stage_caps_init,
4336 		     mlx5_ib_stage_caps_cleanup),
4337 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4338 		     mlx5_ib_stage_raw_eth_non_default_cb,
4339 		     NULL),
4340 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4341 		     mlx5_ib_roce_init,
4342 		     mlx5_ib_roce_cleanup),
4343 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4344 		     mlx5_init_qp_table,
4345 		     mlx5_cleanup_qp_table),
4346 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4347 		     mlx5_init_srq_table,
4348 		     mlx5_cleanup_srq_table),
4349 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4350 		     mlx5_ib_dev_res_init,
4351 		     mlx5_ib_dev_res_cleanup),
4352 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4353 		     mlx5_ib_stage_dev_notifier_init,
4354 		     mlx5_ib_stage_dev_notifier_cleanup),
4355 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4356 		     mlx5_ib_counters_init,
4357 		     mlx5_ib_counters_cleanup),
4358 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4359 		     mlx5_ib_stage_cong_debugfs_init,
4360 		     mlx5_ib_stage_cong_debugfs_cleanup),
4361 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4362 		     mlx5_ib_stage_uar_init,
4363 		     mlx5_ib_stage_uar_cleanup),
4364 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4365 		     mlx5_ib_stage_bfrag_init,
4366 		     mlx5_ib_stage_bfrag_cleanup),
4367 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4368 		     NULL,
4369 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4370 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4371 		     mlx5_ib_devx_init,
4372 		     mlx5_ib_devx_cleanup),
4373 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4374 		     mlx5_ib_stage_ib_reg_init,
4375 		     mlx5_ib_stage_ib_reg_cleanup),
4376 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4377 		     mlx5_ib_stage_post_ib_reg_umr_init,
4378 		     NULL),
4379 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4380 		     mlx5_ib_restrack_init,
4381 		     NULL),
4382 };
4383 
4384 static int mlx5r_mp_probe(struct auxiliary_device *adev,
4385 			  const struct auxiliary_device_id *id)
4386 {
4387 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4388 	struct mlx5_core_dev *mdev = idev->mdev;
4389 	struct mlx5_ib_multiport_info *mpi;
4390 	struct mlx5_ib_dev *dev;
4391 	bool bound = false;
4392 	int err;
4393 
4394 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4395 	if (!mpi)
4396 		return -ENOMEM;
4397 
4398 	mpi->mdev = mdev;
4399 	err = mlx5_query_nic_vport_system_image_guid(mdev,
4400 						     &mpi->sys_image_guid);
4401 	if (err) {
4402 		kfree(mpi);
4403 		return err;
4404 	}
4405 
4406 	mutex_lock(&mlx5_ib_multiport_mutex);
4407 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
4408 		if (dev->sys_image_guid == mpi->sys_image_guid)
4409 			bound = mlx5_ib_bind_slave_port(dev, mpi);
4410 
4411 		if (bound) {
4412 			rdma_roce_rescan_device(&dev->ib_dev);
4413 			mpi->ibdev->ib_active = true;
4414 			break;
4415 		}
4416 	}
4417 
4418 	if (!bound) {
4419 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4420 		dev_dbg(mdev->device,
4421 			"no suitable IB device found to bind to, added to unaffiliated list.\n");
4422 	}
4423 	mutex_unlock(&mlx5_ib_multiport_mutex);
4424 
4425 	dev_set_drvdata(&adev->dev, mpi);
4426 	return 0;
4427 }
4428 
4429 static void mlx5r_mp_remove(struct auxiliary_device *adev)
4430 {
4431 	struct mlx5_ib_multiport_info *mpi;
4432 
4433 	mpi = dev_get_drvdata(&adev->dev);
4434 	mutex_lock(&mlx5_ib_multiport_mutex);
4435 	if (mpi->ibdev)
4436 		mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
4437 	else
4438 		list_del(&mpi->list);
4439 	mutex_unlock(&mlx5_ib_multiport_mutex);
4440 	kfree(mpi);
4441 }
4442 
4443 static int mlx5r_probe(struct auxiliary_device *adev,
4444 		       const struct auxiliary_device_id *id)
4445 {
4446 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4447 	struct mlx5_core_dev *mdev = idev->mdev;
4448 	const struct mlx5_ib_profile *profile;
4449 	int port_type_cap, num_ports, ret;
4450 	enum rdma_link_layer ll;
4451 	struct mlx5_ib_dev *dev;
4452 
4453 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4454 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4455 
4456 	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
4457 			MLX5_CAP_GEN(mdev, num_vhca_ports));
4458 	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
4459 	if (!dev)
4460 		return -ENOMEM;
4461 	dev->port = kcalloc(num_ports, sizeof(*dev->port),
4462 			     GFP_KERNEL);
4463 	if (!dev->port) {
4464 		ib_dealloc_device(&dev->ib_dev);
4465 		return -ENOMEM;
4466 	}
4467 
4468 	dev->mdev = mdev;
4469 	dev->num_ports = num_ports;
4470 
4471 	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_is_roce_init_enabled(mdev))
4472 		profile = &raw_eth_profile;
4473 	else
4474 		profile = &pf_profile;
4475 
4476 	ret = __mlx5_ib_add(dev, profile);
4477 	if (ret) {
4478 		kfree(dev->port);
4479 		ib_dealloc_device(&dev->ib_dev);
4480 		return ret;
4481 	}
4482 
4483 	dev_set_drvdata(&adev->dev, dev);
4484 	return 0;
4485 }
4486 
4487 static void mlx5r_remove(struct auxiliary_device *adev)
4488 {
4489 	struct mlx5_ib_dev *dev;
4490 
4491 	dev = dev_get_drvdata(&adev->dev);
4492 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
4493 }
4494 
4495 static const struct auxiliary_device_id mlx5r_mp_id_table[] = {
4496 	{ .name = MLX5_ADEV_NAME ".multiport", },
4497 	{},
4498 };
4499 
4500 static const struct auxiliary_device_id mlx5r_id_table[] = {
4501 	{ .name = MLX5_ADEV_NAME ".rdma", },
4502 	{},
4503 };
4504 
4505 MODULE_DEVICE_TABLE(auxiliary, mlx5r_mp_id_table);
4506 MODULE_DEVICE_TABLE(auxiliary, mlx5r_id_table);
4507 
4508 static struct auxiliary_driver mlx5r_mp_driver = {
4509 	.name = "multiport",
4510 	.probe = mlx5r_mp_probe,
4511 	.remove = mlx5r_mp_remove,
4512 	.id_table = mlx5r_mp_id_table,
4513 };
4514 
4515 static struct auxiliary_driver mlx5r_driver = {
4516 	.name = "rdma",
4517 	.probe = mlx5r_probe,
4518 	.remove = mlx5r_remove,
4519 	.id_table = mlx5r_id_table,
4520 };
4521 
4522 static int __init mlx5_ib_init(void)
4523 {
4524 	int ret;
4525 
4526 	xlt_emergency_page = (void *)__get_free_page(GFP_KERNEL);
4527 	if (!xlt_emergency_page)
4528 		return -ENOMEM;
4529 
4530 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
4531 	if (!mlx5_ib_event_wq) {
4532 		free_page((unsigned long)xlt_emergency_page);
4533 		return -ENOMEM;
4534 	}
4535 
4536 	mlx5_ib_odp_init();
4537 	ret = mlx5r_rep_init();
4538 	if (ret)
4539 		goto rep_err;
4540 	ret = auxiliary_driver_register(&mlx5r_mp_driver);
4541 	if (ret)
4542 		goto mp_err;
4543 	ret = auxiliary_driver_register(&mlx5r_driver);
4544 	if (ret)
4545 		goto drv_err;
4546 	return 0;
4547 
4548 drv_err:
4549 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4550 mp_err:
4551 	mlx5r_rep_cleanup();
4552 rep_err:
4553 	destroy_workqueue(mlx5_ib_event_wq);
4554 	free_page((unsigned long)xlt_emergency_page);
4555 	return ret;
4556 }
4557 
4558 static void __exit mlx5_ib_cleanup(void)
4559 {
4560 	auxiliary_driver_unregister(&mlx5r_driver);
4561 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4562 	mlx5r_rep_cleanup();
4563 
4564 	destroy_workqueue(mlx5_ib_event_wq);
4565 	free_page((unsigned long)xlt_emergency_page);
4566 }
4567 
4568 module_init(mlx5_ib_init);
4569 module_exit(mlx5_ib_cleanup);
4570