xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision 6db6b729)
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_odp.h>
30 #include <rdma/lag.h>
31 #include <linux/in.h>
32 #include <linux/etherdevice.h>
33 #include "mlx5_ib.h"
34 #include "ib_rep.h"
35 #include "cmd.h"
36 #include "devx.h"
37 #include "dm.h"
38 #include "fs.h"
39 #include "srq.h"
40 #include "qp.h"
41 #include "wr.h"
42 #include "restrack.h"
43 #include "counters.h"
44 #include "umr.h"
45 #include <rdma/uverbs_std_types.h>
46 #include <rdma/uverbs_ioctl.h>
47 #include <rdma/mlx5_user_ioctl_verbs.h>
48 #include <rdma/mlx5_user_ioctl_cmds.h>
49 #include "macsec.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_8):
447 		*active_width = IB_WIDTH_8X;
448 		*active_speed = IB_SPEED_HDR;
449 		break;
450 	case MLX5E_PROT_MASK(MLX5E_400GAUI_4_400GBASE_CR4_KR4):
451 		*active_width = IB_WIDTH_4X;
452 		*active_speed = IB_SPEED_NDR;
453 		break;
454 	default:
455 		return -EINVAL;
456 	}
457 
458 	return 0;
459 }
460 
461 static int translate_eth_proto_oper(u32 eth_proto_oper, u16 *active_speed,
462 				    u8 *active_width, bool ext)
463 {
464 	return ext ?
465 		translate_eth_ext_proto_oper(eth_proto_oper, active_speed,
466 					     active_width) :
467 		translate_eth_legacy_proto_oper(eth_proto_oper, active_speed,
468 						active_width);
469 }
470 
471 static int mlx5_query_port_roce(struct ib_device *device, u32 port_num,
472 				struct ib_port_attr *props)
473 {
474 	struct mlx5_ib_dev *dev = to_mdev(device);
475 	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0};
476 	struct mlx5_core_dev *mdev;
477 	struct net_device *ndev, *upper;
478 	enum ib_mtu ndev_ib_mtu;
479 	bool put_mdev = true;
480 	u32 eth_prot_oper;
481 	u32 mdev_port_num;
482 	bool ext;
483 	int err;
484 
485 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
486 	if (!mdev) {
487 		/* This means the port isn't affiliated yet. Get the
488 		 * info for the master port instead.
489 		 */
490 		put_mdev = false;
491 		mdev = dev->mdev;
492 		mdev_port_num = 1;
493 		port_num = 1;
494 	}
495 
496 	/* Possible bad flows are checked before filling out props so in case
497 	 * of an error it will still be zeroed out.
498 	 * Use native port in case of reps
499 	 */
500 	if (dev->is_rep)
501 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
502 					   1);
503 	else
504 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
505 					   mdev_port_num);
506 	if (err)
507 		goto out;
508 	ext = !!MLX5_GET_ETH_PROTO(ptys_reg, out, true, eth_proto_capability);
509 	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
510 
511 	props->active_width     = IB_WIDTH_4X;
512 	props->active_speed     = IB_SPEED_QDR;
513 
514 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
515 				 &props->active_width, ext);
516 
517 	if (!dev->is_rep && dev->mdev->roce.roce_en) {
518 		u16 qkey_viol_cntr;
519 
520 		props->port_cap_flags |= IB_PORT_CM_SUP;
521 		props->ip_gids = true;
522 		props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev,
523 						   roce_address_table_size);
524 		mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
525 		props->qkey_viol_cntr = qkey_viol_cntr;
526 	}
527 	props->max_mtu          = IB_MTU_4096;
528 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
529 	props->pkey_tbl_len     = 1;
530 	props->state            = IB_PORT_DOWN;
531 	props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
532 
533 	/* If this is a stub query for an unaffiliated port stop here */
534 	if (!put_mdev)
535 		goto out;
536 
537 	ndev = mlx5_ib_get_netdev(device, port_num);
538 	if (!ndev)
539 		goto out;
540 
541 	if (dev->lag_active) {
542 		rcu_read_lock();
543 		upper = netdev_master_upper_dev_get_rcu(ndev);
544 		if (upper) {
545 			dev_put(ndev);
546 			ndev = upper;
547 			dev_hold(ndev);
548 		}
549 		rcu_read_unlock();
550 	}
551 
552 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
553 		props->state      = IB_PORT_ACTIVE;
554 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
555 	}
556 
557 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
558 
559 	dev_put(ndev);
560 
561 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
562 out:
563 	if (put_mdev)
564 		mlx5_ib_put_native_port_mdev(dev, port_num);
565 	return err;
566 }
567 
568 int set_roce_addr(struct mlx5_ib_dev *dev, u32 port_num,
569 		  unsigned int index, const union ib_gid *gid,
570 		  const struct ib_gid_attr *attr)
571 {
572 	enum ib_gid_type gid_type;
573 	u16 vlan_id = 0xffff;
574 	u8 roce_version = 0;
575 	u8 roce_l3_type = 0;
576 	u8 mac[ETH_ALEN];
577 	int ret;
578 
579 	gid_type = attr->gid_type;
580 	if (gid) {
581 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
582 		if (ret)
583 			return ret;
584 	}
585 
586 	switch (gid_type) {
587 	case IB_GID_TYPE_ROCE:
588 		roce_version = MLX5_ROCE_VERSION_1;
589 		break;
590 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
591 		roce_version = MLX5_ROCE_VERSION_2;
592 		if (gid && ipv6_addr_v4mapped((void *)gid))
593 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
594 		else
595 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
596 		break;
597 
598 	default:
599 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
600 	}
601 
602 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
603 				      roce_l3_type, gid->raw, mac,
604 				      vlan_id < VLAN_CFI_MASK, vlan_id,
605 				      port_num);
606 }
607 
608 static int mlx5_ib_add_gid(const struct ib_gid_attr *attr,
609 			   __always_unused void **context)
610 {
611 	int ret;
612 
613 	ret = mlx5r_add_gid_macsec_operations(attr);
614 	if (ret)
615 		return ret;
616 
617 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
618 			     attr->index, &attr->gid, attr);
619 }
620 
621 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
622 			   __always_unused void **context)
623 {
624 	int ret;
625 
626 	ret = set_roce_addr(to_mdev(attr->device), attr->port_num,
627 			    attr->index, NULL, attr);
628 	if (ret)
629 		return ret;
630 
631 	mlx5r_del_gid_macsec_operations(attr);
632 	return 0;
633 }
634 
635 __be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev,
636 				   const struct ib_gid_attr *attr)
637 {
638 	if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
639 		return 0;
640 
641 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
642 }
643 
644 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
645 {
646 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
647 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
648 	return 0;
649 }
650 
651 enum {
652 	MLX5_VPORT_ACCESS_METHOD_MAD,
653 	MLX5_VPORT_ACCESS_METHOD_HCA,
654 	MLX5_VPORT_ACCESS_METHOD_NIC,
655 };
656 
657 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
658 {
659 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
660 		return MLX5_VPORT_ACCESS_METHOD_MAD;
661 
662 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
663 	    IB_LINK_LAYER_ETHERNET)
664 		return MLX5_VPORT_ACCESS_METHOD_NIC;
665 
666 	return MLX5_VPORT_ACCESS_METHOD_HCA;
667 }
668 
669 static void get_atomic_caps(struct mlx5_ib_dev *dev,
670 			    u8 atomic_size_qp,
671 			    struct ib_device_attr *props)
672 {
673 	u8 tmp;
674 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
675 	u8 atomic_req_8B_endianness_mode =
676 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
677 
678 	/* Check if HW supports 8 bytes standard atomic operations and capable
679 	 * of host endianness respond
680 	 */
681 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
682 	if (((atomic_operations & tmp) == tmp) &&
683 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
684 	    (atomic_req_8B_endianness_mode)) {
685 		props->atomic_cap = IB_ATOMIC_HCA;
686 	} else {
687 		props->atomic_cap = IB_ATOMIC_NONE;
688 	}
689 }
690 
691 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
692 			       struct ib_device_attr *props)
693 {
694 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
695 
696 	get_atomic_caps(dev, atomic_size_qp, props);
697 }
698 
699 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
700 					__be64 *sys_image_guid)
701 {
702 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
703 	struct mlx5_core_dev *mdev = dev->mdev;
704 	u64 tmp;
705 	int err;
706 
707 	switch (mlx5_get_vport_access_method(ibdev)) {
708 	case MLX5_VPORT_ACCESS_METHOD_MAD:
709 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
710 							    sys_image_guid);
711 
712 	case MLX5_VPORT_ACCESS_METHOD_HCA:
713 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
714 		break;
715 
716 	case MLX5_VPORT_ACCESS_METHOD_NIC:
717 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
718 		break;
719 
720 	default:
721 		return -EINVAL;
722 	}
723 
724 	if (!err)
725 		*sys_image_guid = cpu_to_be64(tmp);
726 
727 	return err;
728 
729 }
730 
731 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
732 				u16 *max_pkeys)
733 {
734 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
735 	struct mlx5_core_dev *mdev = dev->mdev;
736 
737 	switch (mlx5_get_vport_access_method(ibdev)) {
738 	case MLX5_VPORT_ACCESS_METHOD_MAD:
739 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
740 
741 	case MLX5_VPORT_ACCESS_METHOD_HCA:
742 	case MLX5_VPORT_ACCESS_METHOD_NIC:
743 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
744 						pkey_table_size));
745 		return 0;
746 
747 	default:
748 		return -EINVAL;
749 	}
750 }
751 
752 static int mlx5_query_vendor_id(struct ib_device *ibdev,
753 				u32 *vendor_id)
754 {
755 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
756 
757 	switch (mlx5_get_vport_access_method(ibdev)) {
758 	case MLX5_VPORT_ACCESS_METHOD_MAD:
759 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
760 
761 	case MLX5_VPORT_ACCESS_METHOD_HCA:
762 	case MLX5_VPORT_ACCESS_METHOD_NIC:
763 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
764 
765 	default:
766 		return -EINVAL;
767 	}
768 }
769 
770 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
771 				__be64 *node_guid)
772 {
773 	u64 tmp;
774 	int err;
775 
776 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
777 	case MLX5_VPORT_ACCESS_METHOD_MAD:
778 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
779 
780 	case MLX5_VPORT_ACCESS_METHOD_HCA:
781 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
782 		break;
783 
784 	case MLX5_VPORT_ACCESS_METHOD_NIC:
785 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
786 		break;
787 
788 	default:
789 		return -EINVAL;
790 	}
791 
792 	if (!err)
793 		*node_guid = cpu_to_be64(tmp);
794 
795 	return err;
796 }
797 
798 struct mlx5_reg_node_desc {
799 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
800 };
801 
802 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
803 {
804 	struct mlx5_reg_node_desc in;
805 
806 	if (mlx5_use_mad_ifc(dev))
807 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
808 
809 	memset(&in, 0, sizeof(in));
810 
811 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
812 				    sizeof(struct mlx5_reg_node_desc),
813 				    MLX5_REG_NODE_DESC, 0, 0);
814 }
815 
816 static int mlx5_ib_query_device(struct ib_device *ibdev,
817 				struct ib_device_attr *props,
818 				struct ib_udata *uhw)
819 {
820 	size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
821 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
822 	struct mlx5_core_dev *mdev = dev->mdev;
823 	int err = -ENOMEM;
824 	int max_sq_desc;
825 	int max_rq_sg;
826 	int max_sq_sg;
827 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
828 	bool raw_support = !mlx5_core_mp_enabled(mdev);
829 	struct mlx5_ib_query_device_resp resp = {};
830 	size_t resp_len;
831 	u64 max_tso;
832 
833 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
834 	if (uhw_outlen && uhw_outlen < resp_len)
835 		return -EINVAL;
836 
837 	resp.response_length = resp_len;
838 
839 	if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
840 		return -EINVAL;
841 
842 	memset(props, 0, sizeof(*props));
843 	err = mlx5_query_system_image_guid(ibdev,
844 					   &props->sys_image_guid);
845 	if (err)
846 		return err;
847 
848 	props->max_pkeys = dev->pkey_table_len;
849 
850 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
851 	if (err)
852 		return err;
853 
854 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
855 		(fw_rev_min(dev->mdev) << 16) |
856 		fw_rev_sub(dev->mdev);
857 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
858 		IB_DEVICE_PORT_ACTIVE_EVENT		|
859 		IB_DEVICE_SYS_IMAGE_GUID		|
860 		IB_DEVICE_RC_RNR_NAK_GEN;
861 
862 	if (MLX5_CAP_GEN(mdev, pkv))
863 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
864 	if (MLX5_CAP_GEN(mdev, qkv))
865 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
866 	if (MLX5_CAP_GEN(mdev, apm))
867 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
868 	if (MLX5_CAP_GEN(mdev, xrc))
869 		props->device_cap_flags |= IB_DEVICE_XRC;
870 	if (MLX5_CAP_GEN(mdev, imaicl)) {
871 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
872 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
873 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
874 		/* We support 'Gappy' memory registration too */
875 		props->kernel_cap_flags |= IBK_SG_GAPS_REG;
876 	}
877 	/* IB_WR_REG_MR always requires changing the entity size with UMR */
878 	if (!MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled))
879 		props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
880 	if (MLX5_CAP_GEN(mdev, sho)) {
881 		props->kernel_cap_flags |= IBK_INTEGRITY_HANDOVER;
882 		/* At this stage no support for signature handover */
883 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
884 				      IB_PROT_T10DIF_TYPE_2 |
885 				      IB_PROT_T10DIF_TYPE_3;
886 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
887 				       IB_GUARD_T10DIF_CSUM;
888 	}
889 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
890 		props->kernel_cap_flags |= IBK_BLOCK_MULTICAST_LOOPBACK;
891 
892 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
893 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
894 			/* Legacy bit to support old userspace libraries */
895 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
896 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
897 		}
898 
899 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
900 			props->raw_packet_caps |=
901 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
902 
903 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) {
904 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
905 			if (max_tso) {
906 				resp.tso_caps.max_tso = 1 << max_tso;
907 				resp.tso_caps.supported_qpts |=
908 					1 << IB_QPT_RAW_PACKET;
909 				resp.response_length += sizeof(resp.tso_caps);
910 			}
911 		}
912 
913 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) {
914 			resp.rss_caps.rx_hash_function =
915 						MLX5_RX_HASH_FUNC_TOEPLITZ;
916 			resp.rss_caps.rx_hash_fields_mask =
917 						MLX5_RX_HASH_SRC_IPV4 |
918 						MLX5_RX_HASH_DST_IPV4 |
919 						MLX5_RX_HASH_SRC_IPV6 |
920 						MLX5_RX_HASH_DST_IPV6 |
921 						MLX5_RX_HASH_SRC_PORT_TCP |
922 						MLX5_RX_HASH_DST_PORT_TCP |
923 						MLX5_RX_HASH_SRC_PORT_UDP |
924 						MLX5_RX_HASH_DST_PORT_UDP |
925 						MLX5_RX_HASH_INNER;
926 			resp.response_length += sizeof(resp.rss_caps);
927 		}
928 	} else {
929 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen)
930 			resp.response_length += sizeof(resp.tso_caps);
931 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen)
932 			resp.response_length += sizeof(resp.rss_caps);
933 	}
934 
935 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
936 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
937 		props->kernel_cap_flags |= IBK_UD_TSO;
938 	}
939 
940 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
941 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
942 	    raw_support)
943 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
944 
945 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
946 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
947 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
948 
949 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
950 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
951 	    raw_support) {
952 		/* Legacy bit to support old userspace libraries */
953 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
954 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
955 	}
956 
957 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
958 		props->max_dm_size =
959 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
960 	}
961 
962 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
963 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
964 
965 	if (MLX5_CAP_GEN(mdev, end_pad))
966 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
967 
968 	props->vendor_part_id	   = mdev->pdev->device;
969 	props->hw_ver		   = mdev->pdev->revision;
970 
971 	props->max_mr_size	   = ~0ull;
972 	props->page_size_cap	   = ~(min_page_size - 1);
973 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
974 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
975 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
976 		     sizeof(struct mlx5_wqe_data_seg);
977 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
978 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
979 		     sizeof(struct mlx5_wqe_raddr_seg)) /
980 		sizeof(struct mlx5_wqe_data_seg);
981 	props->max_send_sge = max_sq_sg;
982 	props->max_recv_sge = max_rq_sg;
983 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
984 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
985 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
986 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
987 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
988 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
989 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
990 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
991 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
992 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
993 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
994 	props->max_srq_sge	   = max_rq_sg - 1;
995 	props->max_fast_reg_page_list_len =
996 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
997 	props->max_pi_fast_reg_page_list_len =
998 		props->max_fast_reg_page_list_len / 2;
999 	props->max_sgl_rd =
1000 		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
1001 	get_atomic_caps_qp(dev, props);
1002 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
1003 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
1004 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
1005 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1006 					   props->max_mcast_grp;
1007 	props->max_ah = INT_MAX;
1008 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
1009 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
1010 
1011 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
1012 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
1013 			props->kernel_cap_flags |= IBK_ON_DEMAND_PAGING;
1014 		props->odp_caps = dev->odp_caps;
1015 		if (!uhw) {
1016 			/* ODP for kernel QPs is not implemented for receive
1017 			 * WQEs and SRQ WQEs
1018 			 */
1019 			props->odp_caps.per_transport_caps.rc_odp_caps &=
1020 				~(IB_ODP_SUPPORT_READ |
1021 				  IB_ODP_SUPPORT_SRQ_RECV);
1022 			props->odp_caps.per_transport_caps.uc_odp_caps &=
1023 				~(IB_ODP_SUPPORT_READ |
1024 				  IB_ODP_SUPPORT_SRQ_RECV);
1025 			props->odp_caps.per_transport_caps.ud_odp_caps &=
1026 				~(IB_ODP_SUPPORT_READ |
1027 				  IB_ODP_SUPPORT_SRQ_RECV);
1028 			props->odp_caps.per_transport_caps.xrc_odp_caps &=
1029 				~(IB_ODP_SUPPORT_READ |
1030 				  IB_ODP_SUPPORT_SRQ_RECV);
1031 		}
1032 	}
1033 
1034 	if (mlx5_core_is_vf(mdev))
1035 		props->kernel_cap_flags |= IBK_VIRTUAL_FUNCTION;
1036 
1037 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1038 	    IB_LINK_LAYER_ETHERNET && raw_support) {
1039 		props->rss_caps.max_rwq_indirection_tables =
1040 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1041 		props->rss_caps.max_rwq_indirection_table_size =
1042 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1043 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1044 		props->max_wq_type_rq =
1045 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1046 	}
1047 
1048 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1049 		props->tm_caps.max_num_tags =
1050 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1051 		props->tm_caps.max_ops =
1052 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1053 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1054 	}
1055 
1056 	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1057 	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1058 		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1059 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1060 	}
1061 
1062 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1063 		props->cq_caps.max_cq_moderation_count =
1064 						MLX5_MAX_CQ_COUNT;
1065 		props->cq_caps.max_cq_moderation_period =
1066 						MLX5_MAX_CQ_PERIOD;
1067 	}
1068 
1069 	if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) {
1070 		resp.response_length += sizeof(resp.cqe_comp_caps);
1071 
1072 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1073 			resp.cqe_comp_caps.max_num =
1074 				MLX5_CAP_GEN(dev->mdev,
1075 					     cqe_compression_max_num);
1076 
1077 			resp.cqe_comp_caps.supported_format =
1078 				MLX5_IB_CQE_RES_FORMAT_HASH |
1079 				MLX5_IB_CQE_RES_FORMAT_CSUM;
1080 
1081 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1082 				resp.cqe_comp_caps.supported_format |=
1083 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1084 		}
1085 	}
1086 
1087 	if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen &&
1088 	    raw_support) {
1089 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1090 		    MLX5_CAP_GEN(mdev, qos)) {
1091 			resp.packet_pacing_caps.qp_rate_limit_max =
1092 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1093 			resp.packet_pacing_caps.qp_rate_limit_min =
1094 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1095 			resp.packet_pacing_caps.supported_qpts |=
1096 				1 << IB_QPT_RAW_PACKET;
1097 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1098 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1099 				resp.packet_pacing_caps.cap_flags |=
1100 					MLX5_IB_PP_SUPPORT_BURST;
1101 		}
1102 		resp.response_length += sizeof(resp.packet_pacing_caps);
1103 	}
1104 
1105 	if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <=
1106 	    uhw_outlen) {
1107 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1108 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1109 				MLX5_IB_ALLOW_MPW;
1110 
1111 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1112 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1113 				MLX5_IB_SUPPORT_EMPW;
1114 
1115 		resp.response_length +=
1116 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1117 	}
1118 
1119 	if (offsetofend(typeof(resp), flags) <= uhw_outlen) {
1120 		resp.response_length += sizeof(resp.flags);
1121 
1122 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1123 			resp.flags |=
1124 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1125 
1126 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1127 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1128 		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1129 			resp.flags |=
1130 				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1131 
1132 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1133 	}
1134 
1135 	if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) {
1136 		resp.response_length += sizeof(resp.sw_parsing_caps);
1137 		if (MLX5_CAP_ETH(mdev, swp)) {
1138 			resp.sw_parsing_caps.sw_parsing_offloads |=
1139 				MLX5_IB_SW_PARSING;
1140 
1141 			if (MLX5_CAP_ETH(mdev, swp_csum))
1142 				resp.sw_parsing_caps.sw_parsing_offloads |=
1143 					MLX5_IB_SW_PARSING_CSUM;
1144 
1145 			if (MLX5_CAP_ETH(mdev, swp_lso))
1146 				resp.sw_parsing_caps.sw_parsing_offloads |=
1147 					MLX5_IB_SW_PARSING_LSO;
1148 
1149 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1150 				resp.sw_parsing_caps.supported_qpts =
1151 					BIT(IB_QPT_RAW_PACKET);
1152 		}
1153 	}
1154 
1155 	if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen &&
1156 	    raw_support) {
1157 		resp.response_length += sizeof(resp.striding_rq_caps);
1158 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1159 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1160 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1161 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1162 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1163 			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1164 				resp.striding_rq_caps
1165 					.min_single_wqe_log_num_of_strides =
1166 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1167 			else
1168 				resp.striding_rq_caps
1169 					.min_single_wqe_log_num_of_strides =
1170 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1171 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1172 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1173 			resp.striding_rq_caps.supported_qpts =
1174 				BIT(IB_QPT_RAW_PACKET);
1175 		}
1176 	}
1177 
1178 	if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) {
1179 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1180 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1181 			resp.tunnel_offloads_caps |=
1182 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1183 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1184 			resp.tunnel_offloads_caps |=
1185 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1186 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1187 			resp.tunnel_offloads_caps |=
1188 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1189 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre))
1190 			resp.tunnel_offloads_caps |=
1191 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1192 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp))
1193 			resp.tunnel_offloads_caps |=
1194 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1195 	}
1196 
1197 	if (offsetofend(typeof(resp), dci_streams_caps) <= uhw_outlen) {
1198 		resp.response_length += sizeof(resp.dci_streams_caps);
1199 
1200 		resp.dci_streams_caps.max_log_num_concurent =
1201 			MLX5_CAP_GEN(mdev, log_max_dci_stream_channels);
1202 
1203 		resp.dci_streams_caps.max_log_num_errored =
1204 			MLX5_CAP_GEN(mdev, log_max_dci_errored_streams);
1205 	}
1206 
1207 	if (uhw_outlen) {
1208 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1209 
1210 		if (err)
1211 			return err;
1212 	}
1213 
1214 	return 0;
1215 }
1216 
1217 static void translate_active_width(struct ib_device *ibdev, u16 active_width,
1218 				   u8 *ib_width)
1219 {
1220 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1221 
1222 	if (active_width & MLX5_PTYS_WIDTH_1X)
1223 		*ib_width = IB_WIDTH_1X;
1224 	else if (active_width & MLX5_PTYS_WIDTH_2X)
1225 		*ib_width = IB_WIDTH_2X;
1226 	else if (active_width & MLX5_PTYS_WIDTH_4X)
1227 		*ib_width = IB_WIDTH_4X;
1228 	else if (active_width & MLX5_PTYS_WIDTH_8X)
1229 		*ib_width = IB_WIDTH_8X;
1230 	else if (active_width & MLX5_PTYS_WIDTH_12X)
1231 		*ib_width = IB_WIDTH_12X;
1232 	else {
1233 		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1234 			    active_width);
1235 		*ib_width = IB_WIDTH_4X;
1236 	}
1237 
1238 	return;
1239 }
1240 
1241 static int mlx5_mtu_to_ib_mtu(int mtu)
1242 {
1243 	switch (mtu) {
1244 	case 256: return 1;
1245 	case 512: return 2;
1246 	case 1024: return 3;
1247 	case 2048: return 4;
1248 	case 4096: return 5;
1249 	default:
1250 		pr_warn("invalid mtu\n");
1251 		return -1;
1252 	}
1253 }
1254 
1255 enum ib_max_vl_num {
1256 	__IB_MAX_VL_0		= 1,
1257 	__IB_MAX_VL_0_1		= 2,
1258 	__IB_MAX_VL_0_3		= 3,
1259 	__IB_MAX_VL_0_7		= 4,
1260 	__IB_MAX_VL_0_14	= 5,
1261 };
1262 
1263 enum mlx5_vl_hw_cap {
1264 	MLX5_VL_HW_0	= 1,
1265 	MLX5_VL_HW_0_1	= 2,
1266 	MLX5_VL_HW_0_2	= 3,
1267 	MLX5_VL_HW_0_3	= 4,
1268 	MLX5_VL_HW_0_4	= 5,
1269 	MLX5_VL_HW_0_5	= 6,
1270 	MLX5_VL_HW_0_6	= 7,
1271 	MLX5_VL_HW_0_7	= 8,
1272 	MLX5_VL_HW_0_14	= 15
1273 };
1274 
1275 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1276 				u8 *max_vl_num)
1277 {
1278 	switch (vl_hw_cap) {
1279 	case MLX5_VL_HW_0:
1280 		*max_vl_num = __IB_MAX_VL_0;
1281 		break;
1282 	case MLX5_VL_HW_0_1:
1283 		*max_vl_num = __IB_MAX_VL_0_1;
1284 		break;
1285 	case MLX5_VL_HW_0_3:
1286 		*max_vl_num = __IB_MAX_VL_0_3;
1287 		break;
1288 	case MLX5_VL_HW_0_7:
1289 		*max_vl_num = __IB_MAX_VL_0_7;
1290 		break;
1291 	case MLX5_VL_HW_0_14:
1292 		*max_vl_num = __IB_MAX_VL_0_14;
1293 		break;
1294 
1295 	default:
1296 		return -EINVAL;
1297 	}
1298 
1299 	return 0;
1300 }
1301 
1302 static int mlx5_query_hca_port(struct ib_device *ibdev, u32 port,
1303 			       struct ib_port_attr *props)
1304 {
1305 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1306 	struct mlx5_core_dev *mdev = dev->mdev;
1307 	struct mlx5_hca_vport_context *rep;
1308 	u16 max_mtu;
1309 	u16 oper_mtu;
1310 	int err;
1311 	u16 ib_link_width_oper;
1312 	u8 vl_hw_cap;
1313 
1314 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1315 	if (!rep) {
1316 		err = -ENOMEM;
1317 		goto out;
1318 	}
1319 
1320 	/* props being zeroed by the caller, avoid zeroing it here */
1321 
1322 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1323 	if (err)
1324 		goto out;
1325 
1326 	props->lid		= rep->lid;
1327 	props->lmc		= rep->lmc;
1328 	props->sm_lid		= rep->sm_lid;
1329 	props->sm_sl		= rep->sm_sl;
1330 	props->state		= rep->vport_state;
1331 	props->phys_state	= rep->port_physical_state;
1332 	props->port_cap_flags	= rep->cap_mask1;
1333 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1334 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1335 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1336 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1337 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1338 	props->subnet_timeout	= rep->subnet_timeout;
1339 	props->init_type_reply	= rep->init_type_reply;
1340 
1341 	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1342 		props->port_cap_flags2 = rep->cap_mask2;
1343 
1344 	err = mlx5_query_ib_port_oper(mdev, &ib_link_width_oper,
1345 				      &props->active_speed, port);
1346 	if (err)
1347 		goto out;
1348 
1349 	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1350 
1351 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1352 
1353 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1354 
1355 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1356 
1357 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1358 
1359 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1360 	if (err)
1361 		goto out;
1362 
1363 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1364 				   &props->max_vl_num);
1365 out:
1366 	kfree(rep);
1367 	return err;
1368 }
1369 
1370 int mlx5_ib_query_port(struct ib_device *ibdev, u32 port,
1371 		       struct ib_port_attr *props)
1372 {
1373 	unsigned int count;
1374 	int ret;
1375 
1376 	switch (mlx5_get_vport_access_method(ibdev)) {
1377 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1378 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1379 		break;
1380 
1381 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1382 		ret = mlx5_query_hca_port(ibdev, port, props);
1383 		break;
1384 
1385 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1386 		ret = mlx5_query_port_roce(ibdev, port, props);
1387 		break;
1388 
1389 	default:
1390 		ret = -EINVAL;
1391 	}
1392 
1393 	if (!ret && props) {
1394 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1395 		struct mlx5_core_dev *mdev;
1396 		bool put_mdev = true;
1397 
1398 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1399 		if (!mdev) {
1400 			/* If the port isn't affiliated yet query the master.
1401 			 * The master and slave will have the same values.
1402 			 */
1403 			mdev = dev->mdev;
1404 			port = 1;
1405 			put_mdev = false;
1406 		}
1407 		count = mlx5_core_reserved_gids_count(mdev);
1408 		if (put_mdev)
1409 			mlx5_ib_put_native_port_mdev(dev, port);
1410 		props->gid_tbl_len -= count;
1411 	}
1412 	return ret;
1413 }
1414 
1415 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u32 port,
1416 				  struct ib_port_attr *props)
1417 {
1418 	return mlx5_query_port_roce(ibdev, port, props);
1419 }
1420 
1421 static int mlx5_ib_rep_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1422 				  u16 *pkey)
1423 {
1424 	/* Default special Pkey for representor device port as per the
1425 	 * IB specification 1.3 section 10.9.1.2.
1426 	 */
1427 	*pkey = 0xffff;
1428 	return 0;
1429 }
1430 
1431 static int mlx5_ib_query_gid(struct ib_device *ibdev, u32 port, int index,
1432 			     union ib_gid *gid)
1433 {
1434 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1435 	struct mlx5_core_dev *mdev = dev->mdev;
1436 
1437 	switch (mlx5_get_vport_access_method(ibdev)) {
1438 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1439 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1440 
1441 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1442 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1443 
1444 	default:
1445 		return -EINVAL;
1446 	}
1447 
1448 }
1449 
1450 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u32 port,
1451 				   u16 index, u16 *pkey)
1452 {
1453 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1454 	struct mlx5_core_dev *mdev;
1455 	bool put_mdev = true;
1456 	u32 mdev_port_num;
1457 	int err;
1458 
1459 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1460 	if (!mdev) {
1461 		/* The port isn't affiliated yet, get the PKey from the master
1462 		 * port. For RoCE the PKey tables will be the same.
1463 		 */
1464 		put_mdev = false;
1465 		mdev = dev->mdev;
1466 		mdev_port_num = 1;
1467 	}
1468 
1469 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1470 					index, pkey);
1471 	if (put_mdev)
1472 		mlx5_ib_put_native_port_mdev(dev, port);
1473 
1474 	return err;
1475 }
1476 
1477 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1478 			      u16 *pkey)
1479 {
1480 	switch (mlx5_get_vport_access_method(ibdev)) {
1481 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1482 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1483 
1484 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1485 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1486 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1487 	default:
1488 		return -EINVAL;
1489 	}
1490 }
1491 
1492 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1493 				 struct ib_device_modify *props)
1494 {
1495 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1496 	struct mlx5_reg_node_desc in;
1497 	struct mlx5_reg_node_desc out;
1498 	int err;
1499 
1500 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1501 		return -EOPNOTSUPP;
1502 
1503 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1504 		return 0;
1505 
1506 	/*
1507 	 * If possible, pass node desc to FW, so it can generate
1508 	 * a 144 trap.  If cmd fails, just ignore.
1509 	 */
1510 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1511 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1512 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1513 	if (err)
1514 		return err;
1515 
1516 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1517 
1518 	return err;
1519 }
1520 
1521 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u32 port_num, u32 mask,
1522 				u32 value)
1523 {
1524 	struct mlx5_hca_vport_context ctx = {};
1525 	struct mlx5_core_dev *mdev;
1526 	u32 mdev_port_num;
1527 	int err;
1528 
1529 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1530 	if (!mdev)
1531 		return -ENODEV;
1532 
1533 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1534 	if (err)
1535 		goto out;
1536 
1537 	if (~ctx.cap_mask1_perm & mask) {
1538 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1539 			     mask, ctx.cap_mask1_perm);
1540 		err = -EINVAL;
1541 		goto out;
1542 	}
1543 
1544 	ctx.cap_mask1 = value;
1545 	ctx.cap_mask1_perm = mask;
1546 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1547 						 0, &ctx);
1548 
1549 out:
1550 	mlx5_ib_put_native_port_mdev(dev, port_num);
1551 
1552 	return err;
1553 }
1554 
1555 static int mlx5_ib_modify_port(struct ib_device *ibdev, u32 port, int mask,
1556 			       struct ib_port_modify *props)
1557 {
1558 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1559 	struct ib_port_attr attr;
1560 	u32 tmp;
1561 	int err;
1562 	u32 change_mask;
1563 	u32 value;
1564 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1565 		      IB_LINK_LAYER_INFINIBAND);
1566 
1567 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1568 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1569 	 */
1570 	if (!is_ib)
1571 		return 0;
1572 
1573 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1574 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1575 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1576 		return set_port_caps_atomic(dev, port, change_mask, value);
1577 	}
1578 
1579 	mutex_lock(&dev->cap_mask_mutex);
1580 
1581 	err = ib_query_port(ibdev, port, &attr);
1582 	if (err)
1583 		goto out;
1584 
1585 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1586 		~props->clr_port_cap_mask;
1587 
1588 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1589 
1590 out:
1591 	mutex_unlock(&dev->cap_mask_mutex);
1592 	return err;
1593 }
1594 
1595 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1596 {
1597 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1598 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1599 }
1600 
1601 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1602 {
1603 	/* Large page with non 4k uar support might limit the dynamic size */
1604 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1605 		return MLX5_MIN_DYN_BFREGS;
1606 
1607 	return MLX5_MAX_DYN_BFREGS;
1608 }
1609 
1610 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1611 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1612 			     struct mlx5_bfreg_info *bfregi)
1613 {
1614 	int uars_per_sys_page;
1615 	int bfregs_per_sys_page;
1616 	int ref_bfregs = req->total_num_bfregs;
1617 
1618 	if (req->total_num_bfregs == 0)
1619 		return -EINVAL;
1620 
1621 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1622 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1623 
1624 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1625 		return -ENOMEM;
1626 
1627 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1628 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1629 	/* This holds the required static allocation asked by the user */
1630 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1631 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1632 		return -EINVAL;
1633 
1634 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1635 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1636 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1637 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1638 
1639 	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",
1640 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1641 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1642 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1643 		    bfregi->num_sys_pages);
1644 
1645 	return 0;
1646 }
1647 
1648 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1649 {
1650 	struct mlx5_bfreg_info *bfregi;
1651 	int err;
1652 	int i;
1653 
1654 	bfregi = &context->bfregi;
1655 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1656 		err = mlx5_cmd_uar_alloc(dev->mdev, &bfregi->sys_pages[i],
1657 					 context->devx_uid);
1658 		if (err)
1659 			goto error;
1660 
1661 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1662 	}
1663 
1664 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1665 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1666 
1667 	return 0;
1668 
1669 error:
1670 	for (--i; i >= 0; i--)
1671 		if (mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i],
1672 					 context->devx_uid))
1673 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1674 
1675 	return err;
1676 }
1677 
1678 static void deallocate_uars(struct mlx5_ib_dev *dev,
1679 			    struct mlx5_ib_ucontext *context)
1680 {
1681 	struct mlx5_bfreg_info *bfregi;
1682 	int i;
1683 
1684 	bfregi = &context->bfregi;
1685 	for (i = 0; i < bfregi->num_sys_pages; i++)
1686 		if (i < bfregi->num_static_sys_pages ||
1687 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1688 			mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i],
1689 					     context->devx_uid);
1690 }
1691 
1692 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1693 {
1694 	int err = 0;
1695 
1696 	mutex_lock(&dev->lb.mutex);
1697 	if (td)
1698 		dev->lb.user_td++;
1699 	if (qp)
1700 		dev->lb.qps++;
1701 
1702 	if (dev->lb.user_td == 2 ||
1703 	    dev->lb.qps == 1) {
1704 		if (!dev->lb.enabled) {
1705 			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1706 			dev->lb.enabled = true;
1707 		}
1708 	}
1709 
1710 	mutex_unlock(&dev->lb.mutex);
1711 
1712 	return err;
1713 }
1714 
1715 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1716 {
1717 	mutex_lock(&dev->lb.mutex);
1718 	if (td)
1719 		dev->lb.user_td--;
1720 	if (qp)
1721 		dev->lb.qps--;
1722 
1723 	if (dev->lb.user_td == 1 &&
1724 	    dev->lb.qps == 0) {
1725 		if (dev->lb.enabled) {
1726 			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1727 			dev->lb.enabled = false;
1728 		}
1729 	}
1730 
1731 	mutex_unlock(&dev->lb.mutex);
1732 }
1733 
1734 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1735 					  u16 uid)
1736 {
1737 	int err;
1738 
1739 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1740 		return 0;
1741 
1742 	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1743 	if (err)
1744 		return err;
1745 
1746 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1747 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1748 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1749 		return err;
1750 
1751 	return mlx5_ib_enable_lb(dev, true, false);
1752 }
1753 
1754 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1755 					     u16 uid)
1756 {
1757 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1758 		return;
1759 
1760 	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1761 
1762 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1763 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1764 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1765 		return;
1766 
1767 	mlx5_ib_disable_lb(dev, true, false);
1768 }
1769 
1770 static int set_ucontext_resp(struct ib_ucontext *uctx,
1771 			     struct mlx5_ib_alloc_ucontext_resp *resp)
1772 {
1773 	struct ib_device *ibdev = uctx->device;
1774 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1775 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1776 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1777 
1778 	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1779 		resp->dump_fill_mkey = dev->mkeys.dump_fill_mkey;
1780 		resp->comp_mask |=
1781 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1782 	}
1783 
1784 	resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1785 	if (dev->wc_support)
1786 		resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev,
1787 						      log_bf_reg_size);
1788 	resp->cache_line_size = cache_line_size();
1789 	resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1790 	resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1791 	resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1792 	resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1793 	resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1794 	resp->cqe_version = context->cqe_version;
1795 	resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1796 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1797 	resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1798 					MLX5_CAP_GEN(dev->mdev,
1799 						     num_of_uars_per_page) : 1;
1800 	resp->tot_bfregs = bfregi->lib_uar_dyn ? 0 :
1801 			bfregi->total_num_bfregs - bfregi->num_dyn_bfregs;
1802 	resp->num_ports = dev->num_ports;
1803 	resp->cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1804 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1805 
1806 	if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1807 		mlx5_query_min_inline(dev->mdev, &resp->eth_min_inline);
1808 		resp->eth_min_inline++;
1809 	}
1810 
1811 	if (dev->mdev->clock_info)
1812 		resp->clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1813 
1814 	/*
1815 	 * We don't want to expose information from the PCI bar that is located
1816 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1817 	 * pretend we don't support reading the HCA's core clock. This is also
1818 	 * forced by mmap function.
1819 	 */
1820 	if (PAGE_SIZE <= 4096) {
1821 		resp->comp_mask |=
1822 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1823 		resp->hca_core_clock_offset =
1824 			offsetof(struct mlx5_init_seg,
1825 				 internal_timer_h) % PAGE_SIZE;
1826 	}
1827 
1828 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
1829 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE;
1830 
1831 	if (rt_supported(MLX5_CAP_GEN(dev->mdev, sq_ts_format)) &&
1832 	    rt_supported(MLX5_CAP_GEN(dev->mdev, rq_ts_format)) &&
1833 	    rt_supported(MLX5_CAP_ROCE(dev->mdev, qp_ts_format)))
1834 		resp->comp_mask |=
1835 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_REAL_TIME_TS;
1836 
1837 	resp->num_dyn_bfregs = bfregi->num_dyn_bfregs;
1838 
1839 	if (MLX5_CAP_GEN(dev->mdev, drain_sigerr))
1840 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_SQD2RTS;
1841 
1842 	resp->comp_mask |=
1843 		MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_MKEY_UPDATE_TAG;
1844 
1845 	return 0;
1846 }
1847 
1848 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1849 				  struct ib_udata *udata)
1850 {
1851 	struct ib_device *ibdev = uctx->device;
1852 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1853 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1854 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1855 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1856 	struct mlx5_bfreg_info *bfregi;
1857 	int ver;
1858 	int err;
1859 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1860 				     max_cqe_version);
1861 	bool lib_uar_4k;
1862 	bool lib_uar_dyn;
1863 
1864 	if (!dev->ib_active)
1865 		return -EAGAIN;
1866 
1867 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1868 		ver = 0;
1869 	else if (udata->inlen >= min_req_v2)
1870 		ver = 2;
1871 	else
1872 		return -EINVAL;
1873 
1874 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1875 	if (err)
1876 		return err;
1877 
1878 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1879 		return -EOPNOTSUPP;
1880 
1881 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1882 		return -EOPNOTSUPP;
1883 
1884 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1885 				    MLX5_NON_FP_BFREGS_PER_UAR);
1886 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1887 		return -EINVAL;
1888 
1889 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1890 		err = mlx5_ib_devx_create(dev, true);
1891 		if (err < 0)
1892 			goto out_ctx;
1893 		context->devx_uid = err;
1894 	}
1895 
1896 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1897 	lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
1898 	bfregi = &context->bfregi;
1899 
1900 	if (lib_uar_dyn) {
1901 		bfregi->lib_uar_dyn = lib_uar_dyn;
1902 		goto uar_done;
1903 	}
1904 
1905 	/* updates req->total_num_bfregs */
1906 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1907 	if (err)
1908 		goto out_devx;
1909 
1910 	mutex_init(&bfregi->lock);
1911 	bfregi->lib_uar_4k = lib_uar_4k;
1912 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1913 				GFP_KERNEL);
1914 	if (!bfregi->count) {
1915 		err = -ENOMEM;
1916 		goto out_devx;
1917 	}
1918 
1919 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1920 				    sizeof(*bfregi->sys_pages),
1921 				    GFP_KERNEL);
1922 	if (!bfregi->sys_pages) {
1923 		err = -ENOMEM;
1924 		goto out_count;
1925 	}
1926 
1927 	err = allocate_uars(dev, context);
1928 	if (err)
1929 		goto out_sys_pages;
1930 
1931 uar_done:
1932 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1933 					     context->devx_uid);
1934 	if (err)
1935 		goto out_uars;
1936 
1937 	INIT_LIST_HEAD(&context->db_page_list);
1938 	mutex_init(&context->db_page_mutex);
1939 
1940 	context->cqe_version = min_t(__u8,
1941 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1942 				 req.max_cqe_version);
1943 
1944 	err = set_ucontext_resp(uctx, &resp);
1945 	if (err)
1946 		goto out_mdev;
1947 
1948 	resp.response_length = min(udata->outlen, sizeof(resp));
1949 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1950 	if (err)
1951 		goto out_mdev;
1952 
1953 	bfregi->ver = ver;
1954 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1955 	context->lib_caps = req.lib_caps;
1956 	print_lib_caps(dev, context->lib_caps);
1957 
1958 	if (mlx5_ib_lag_should_assign_affinity(dev)) {
1959 		u32 port = mlx5_core_native_port_num(dev->mdev) - 1;
1960 
1961 		atomic_set(&context->tx_port_affinity,
1962 			   atomic_add_return(
1963 				   1, &dev->port[port].roce.tx_port_affinity));
1964 	}
1965 
1966 	return 0;
1967 
1968 out_mdev:
1969 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1970 
1971 out_uars:
1972 	deallocate_uars(dev, context);
1973 
1974 out_sys_pages:
1975 	kfree(bfregi->sys_pages);
1976 
1977 out_count:
1978 	kfree(bfregi->count);
1979 
1980 out_devx:
1981 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1982 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1983 
1984 out_ctx:
1985 	return err;
1986 }
1987 
1988 static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext,
1989 				  struct uverbs_attr_bundle *attrs)
1990 {
1991 	struct mlx5_ib_alloc_ucontext_resp uctx_resp = {};
1992 	int ret;
1993 
1994 	ret = set_ucontext_resp(ibcontext, &uctx_resp);
1995 	if (ret)
1996 		return ret;
1997 
1998 	uctx_resp.response_length =
1999 		min_t(size_t,
2000 		      uverbs_attr_get_len(attrs,
2001 				MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX),
2002 		      sizeof(uctx_resp));
2003 
2004 	ret = uverbs_copy_to_struct_or_zero(attrs,
2005 					MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
2006 					&uctx_resp,
2007 					sizeof(uctx_resp));
2008 	return ret;
2009 }
2010 
2011 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
2012 {
2013 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2014 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2015 	struct mlx5_bfreg_info *bfregi;
2016 
2017 	bfregi = &context->bfregi;
2018 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
2019 
2020 	deallocate_uars(dev, context);
2021 	kfree(bfregi->sys_pages);
2022 	kfree(bfregi->count);
2023 
2024 	if (context->devx_uid)
2025 		mlx5_ib_devx_destroy(dev, context->devx_uid);
2026 }
2027 
2028 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2029 				 int uar_idx)
2030 {
2031 	int fw_uars_per_page;
2032 
2033 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2034 
2035 	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2036 }
2037 
2038 static u64 uar_index2paddress(struct mlx5_ib_dev *dev,
2039 				 int uar_idx)
2040 {
2041 	unsigned int fw_uars_per_page;
2042 
2043 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
2044 				MLX5_UARS_IN_PAGE : 1;
2045 
2046 	return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE);
2047 }
2048 
2049 static int get_command(unsigned long offset)
2050 {
2051 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2052 }
2053 
2054 static int get_arg(unsigned long offset)
2055 {
2056 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2057 }
2058 
2059 static int get_index(unsigned long offset)
2060 {
2061 	return get_arg(offset);
2062 }
2063 
2064 /* Index resides in an extra byte to enable larger values than 255 */
2065 static int get_extended_index(unsigned long offset)
2066 {
2067 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2068 }
2069 
2070 
2071 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2072 {
2073 }
2074 
2075 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2076 {
2077 	switch (cmd) {
2078 	case MLX5_IB_MMAP_WC_PAGE:
2079 		return "WC";
2080 	case MLX5_IB_MMAP_REGULAR_PAGE:
2081 		return "best effort WC";
2082 	case MLX5_IB_MMAP_NC_PAGE:
2083 		return "NC";
2084 	case MLX5_IB_MMAP_DEVICE_MEM:
2085 		return "Device Memory";
2086 	default:
2087 		return "Unknown";
2088 	}
2089 }
2090 
2091 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2092 					struct vm_area_struct *vma,
2093 					struct mlx5_ib_ucontext *context)
2094 {
2095 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2096 	    !(vma->vm_flags & VM_SHARED))
2097 		return -EINVAL;
2098 
2099 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2100 		return -EOPNOTSUPP;
2101 
2102 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2103 		return -EPERM;
2104 	vm_flags_clear(vma, VM_MAYWRITE);
2105 
2106 	if (!dev->mdev->clock_info)
2107 		return -EOPNOTSUPP;
2108 
2109 	return vm_insert_page(vma, vma->vm_start,
2110 			      virt_to_page(dev->mdev->clock_info));
2111 }
2112 
2113 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2114 {
2115 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2116 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2117 	struct mlx5_var_table *var_table = &dev->var_table;
2118 	struct mlx5_ib_ucontext *context = to_mucontext(entry->ucontext);
2119 
2120 	switch (mentry->mmap_flag) {
2121 	case MLX5_IB_MMAP_TYPE_MEMIC:
2122 	case MLX5_IB_MMAP_TYPE_MEMIC_OP:
2123 		mlx5_ib_dm_mmap_free(dev, mentry);
2124 		break;
2125 	case MLX5_IB_MMAP_TYPE_VAR:
2126 		mutex_lock(&var_table->bitmap_lock);
2127 		clear_bit(mentry->page_idx, var_table->bitmap);
2128 		mutex_unlock(&var_table->bitmap_lock);
2129 		kfree(mentry);
2130 		break;
2131 	case MLX5_IB_MMAP_TYPE_UAR_WC:
2132 	case MLX5_IB_MMAP_TYPE_UAR_NC:
2133 		mlx5_cmd_uar_dealloc(dev->mdev, mentry->page_idx,
2134 				     context->devx_uid);
2135 		kfree(mentry);
2136 		break;
2137 	default:
2138 		WARN_ON(true);
2139 	}
2140 }
2141 
2142 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2143 		    struct vm_area_struct *vma,
2144 		    struct mlx5_ib_ucontext *context)
2145 {
2146 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2147 	int err;
2148 	unsigned long idx;
2149 	phys_addr_t pfn;
2150 	pgprot_t prot;
2151 	u32 bfreg_dyn_idx = 0;
2152 	u32 uar_index;
2153 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2154 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2155 				bfregi->num_static_sys_pages;
2156 
2157 	if (bfregi->lib_uar_dyn)
2158 		return -EINVAL;
2159 
2160 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2161 		return -EINVAL;
2162 
2163 	if (dyn_uar)
2164 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2165 	else
2166 		idx = get_index(vma->vm_pgoff);
2167 
2168 	if (idx >= max_valid_idx) {
2169 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2170 			     idx, max_valid_idx);
2171 		return -EINVAL;
2172 	}
2173 
2174 	switch (cmd) {
2175 	case MLX5_IB_MMAP_WC_PAGE:
2176 	case MLX5_IB_MMAP_ALLOC_WC:
2177 	case MLX5_IB_MMAP_REGULAR_PAGE:
2178 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2179 		prot = pgprot_writecombine(vma->vm_page_prot);
2180 		break;
2181 	case MLX5_IB_MMAP_NC_PAGE:
2182 		prot = pgprot_noncached(vma->vm_page_prot);
2183 		break;
2184 	default:
2185 		return -EINVAL;
2186 	}
2187 
2188 	if (dyn_uar) {
2189 		int uars_per_page;
2190 
2191 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2192 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2193 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2194 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2195 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2196 			return -EINVAL;
2197 		}
2198 
2199 		mutex_lock(&bfregi->lock);
2200 		/* Fail if uar already allocated, first bfreg index of each
2201 		 * page holds its count.
2202 		 */
2203 		if (bfregi->count[bfreg_dyn_idx]) {
2204 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2205 			mutex_unlock(&bfregi->lock);
2206 			return -EINVAL;
2207 		}
2208 
2209 		bfregi->count[bfreg_dyn_idx]++;
2210 		mutex_unlock(&bfregi->lock);
2211 
2212 		err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index,
2213 					 context->devx_uid);
2214 		if (err) {
2215 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2216 			goto free_bfreg;
2217 		}
2218 	} else {
2219 		uar_index = bfregi->sys_pages[idx];
2220 	}
2221 
2222 	pfn = uar_index2pfn(dev, uar_index);
2223 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2224 
2225 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2226 				prot, NULL);
2227 	if (err) {
2228 		mlx5_ib_err(dev,
2229 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2230 			    err, mmap_cmd2str(cmd));
2231 		goto err;
2232 	}
2233 
2234 	if (dyn_uar)
2235 		bfregi->sys_pages[idx] = uar_index;
2236 	return 0;
2237 
2238 err:
2239 	if (!dyn_uar)
2240 		return err;
2241 
2242 	mlx5_cmd_uar_dealloc(dev->mdev, idx, context->devx_uid);
2243 
2244 free_bfreg:
2245 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2246 
2247 	return err;
2248 }
2249 
2250 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2251 {
2252 	unsigned long idx;
2253 	u8 command;
2254 
2255 	command = get_command(vma->vm_pgoff);
2256 	idx = get_extended_index(vma->vm_pgoff);
2257 
2258 	return (command << 16 | idx);
2259 }
2260 
2261 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2262 			       struct vm_area_struct *vma,
2263 			       struct ib_ucontext *ucontext)
2264 {
2265 	struct mlx5_user_mmap_entry *mentry;
2266 	struct rdma_user_mmap_entry *entry;
2267 	unsigned long pgoff;
2268 	pgprot_t prot;
2269 	phys_addr_t pfn;
2270 	int ret;
2271 
2272 	pgoff = mlx5_vma_to_pgoff(vma);
2273 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2274 	if (!entry)
2275 		return -EINVAL;
2276 
2277 	mentry = to_mmmap(entry);
2278 	pfn = (mentry->address >> PAGE_SHIFT);
2279 	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
2280 	    mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
2281 		prot = pgprot_noncached(vma->vm_page_prot);
2282 	else
2283 		prot = pgprot_writecombine(vma->vm_page_prot);
2284 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2285 				entry->npages * PAGE_SIZE,
2286 				prot,
2287 				entry);
2288 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2289 	return ret;
2290 }
2291 
2292 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2293 {
2294 	u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2295 	u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2296 
2297 	return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2298 		(index & 0xFF)) << PAGE_SHIFT;
2299 }
2300 
2301 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2302 {
2303 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2304 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2305 	unsigned long command;
2306 	phys_addr_t pfn;
2307 
2308 	command = get_command(vma->vm_pgoff);
2309 	switch (command) {
2310 	case MLX5_IB_MMAP_WC_PAGE:
2311 	case MLX5_IB_MMAP_ALLOC_WC:
2312 		if (!dev->wc_support)
2313 			return -EPERM;
2314 		fallthrough;
2315 	case MLX5_IB_MMAP_NC_PAGE:
2316 	case MLX5_IB_MMAP_REGULAR_PAGE:
2317 		return uar_mmap(dev, command, vma, context);
2318 
2319 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2320 		return -ENOSYS;
2321 
2322 	case MLX5_IB_MMAP_CORE_CLOCK:
2323 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2324 			return -EINVAL;
2325 
2326 		if (vma->vm_flags & VM_WRITE)
2327 			return -EPERM;
2328 		vm_flags_clear(vma, VM_MAYWRITE);
2329 
2330 		/* Don't expose to user-space information it shouldn't have */
2331 		if (PAGE_SIZE > 4096)
2332 			return -EOPNOTSUPP;
2333 
2334 		pfn = (dev->mdev->iseg_base +
2335 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2336 			PAGE_SHIFT;
2337 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2338 					 PAGE_SIZE,
2339 					 pgprot_noncached(vma->vm_page_prot),
2340 					 NULL);
2341 	case MLX5_IB_MMAP_CLOCK_INFO:
2342 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2343 
2344 	default:
2345 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2346 	}
2347 
2348 	return 0;
2349 }
2350 
2351 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2352 {
2353 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2354 	struct ib_device *ibdev = ibpd->device;
2355 	struct mlx5_ib_alloc_pd_resp resp;
2356 	int err;
2357 	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2358 	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {};
2359 	u16 uid = 0;
2360 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2361 		udata, struct mlx5_ib_ucontext, ibucontext);
2362 
2363 	uid = context ? context->devx_uid : 0;
2364 	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2365 	MLX5_SET(alloc_pd_in, in, uid, uid);
2366 	err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out);
2367 	if (err)
2368 		return err;
2369 
2370 	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2371 	pd->uid = uid;
2372 	if (udata) {
2373 		resp.pdn = pd->pdn;
2374 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2375 			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2376 			return -EFAULT;
2377 		}
2378 	}
2379 
2380 	return 0;
2381 }
2382 
2383 static int mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2384 {
2385 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2386 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2387 
2388 	return mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2389 }
2390 
2391 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2392 {
2393 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2394 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
2395 	int err;
2396 	u16 uid;
2397 
2398 	uid = ibqp->pd ?
2399 		to_mpd(ibqp->pd)->uid : 0;
2400 
2401 	if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) {
2402 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
2403 		return -EOPNOTSUPP;
2404 	}
2405 
2406 	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2407 	if (err)
2408 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2409 			     ibqp->qp_num, gid->raw);
2410 
2411 	return err;
2412 }
2413 
2414 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2415 {
2416 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2417 	int err;
2418 	u16 uid;
2419 
2420 	uid = ibqp->pd ?
2421 		to_mpd(ibqp->pd)->uid : 0;
2422 	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2423 	if (err)
2424 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2425 			     ibqp->qp_num, gid->raw);
2426 
2427 	return err;
2428 }
2429 
2430 static int init_node_data(struct mlx5_ib_dev *dev)
2431 {
2432 	int err;
2433 
2434 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2435 	if (err)
2436 		return err;
2437 
2438 	dev->mdev->rev_id = dev->mdev->pdev->revision;
2439 
2440 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2441 }
2442 
2443 static ssize_t fw_pages_show(struct device *device,
2444 			     struct device_attribute *attr, char *buf)
2445 {
2446 	struct mlx5_ib_dev *dev =
2447 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2448 
2449 	return sysfs_emit(buf, "%d\n", dev->mdev->priv.fw_pages);
2450 }
2451 static DEVICE_ATTR_RO(fw_pages);
2452 
2453 static ssize_t reg_pages_show(struct device *device,
2454 			      struct device_attribute *attr, char *buf)
2455 {
2456 	struct mlx5_ib_dev *dev =
2457 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2458 
2459 	return sysfs_emit(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2460 }
2461 static DEVICE_ATTR_RO(reg_pages);
2462 
2463 static ssize_t hca_type_show(struct device *device,
2464 			     struct device_attribute *attr, char *buf)
2465 {
2466 	struct mlx5_ib_dev *dev =
2467 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2468 
2469 	return sysfs_emit(buf, "MT%d\n", dev->mdev->pdev->device);
2470 }
2471 static DEVICE_ATTR_RO(hca_type);
2472 
2473 static ssize_t hw_rev_show(struct device *device,
2474 			   struct device_attribute *attr, char *buf)
2475 {
2476 	struct mlx5_ib_dev *dev =
2477 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2478 
2479 	return sysfs_emit(buf, "%x\n", dev->mdev->rev_id);
2480 }
2481 static DEVICE_ATTR_RO(hw_rev);
2482 
2483 static ssize_t board_id_show(struct device *device,
2484 			     struct device_attribute *attr, char *buf)
2485 {
2486 	struct mlx5_ib_dev *dev =
2487 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2488 
2489 	return sysfs_emit(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2490 			  dev->mdev->board_id);
2491 }
2492 static DEVICE_ATTR_RO(board_id);
2493 
2494 static struct attribute *mlx5_class_attributes[] = {
2495 	&dev_attr_hw_rev.attr,
2496 	&dev_attr_hca_type.attr,
2497 	&dev_attr_board_id.attr,
2498 	&dev_attr_fw_pages.attr,
2499 	&dev_attr_reg_pages.attr,
2500 	NULL,
2501 };
2502 
2503 static const struct attribute_group mlx5_attr_group = {
2504 	.attrs = mlx5_class_attributes,
2505 };
2506 
2507 static void pkey_change_handler(struct work_struct *work)
2508 {
2509 	struct mlx5_ib_port_resources *ports =
2510 		container_of(work, struct mlx5_ib_port_resources,
2511 			     pkey_change_work);
2512 
2513 	if (!ports->gsi)
2514 		/*
2515 		 * We got this event before device was fully configured
2516 		 * and MAD registration code wasn't called/finished yet.
2517 		 */
2518 		return;
2519 
2520 	mlx5_ib_gsi_pkey_change(ports->gsi);
2521 }
2522 
2523 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2524 {
2525 	struct mlx5_ib_qp *mqp;
2526 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2527 	struct mlx5_core_cq *mcq;
2528 	struct list_head cq_armed_list;
2529 	unsigned long flags_qp;
2530 	unsigned long flags_cq;
2531 	unsigned long flags;
2532 
2533 	INIT_LIST_HEAD(&cq_armed_list);
2534 
2535 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2536 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2537 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2538 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2539 		if (mqp->sq.tail != mqp->sq.head) {
2540 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2541 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2542 			if (send_mcq->mcq.comp &&
2543 			    mqp->ibqp.send_cq->comp_handler) {
2544 				if (!send_mcq->mcq.reset_notify_added) {
2545 					send_mcq->mcq.reset_notify_added = 1;
2546 					list_add_tail(&send_mcq->mcq.reset_notify,
2547 						      &cq_armed_list);
2548 				}
2549 			}
2550 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2551 		}
2552 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2553 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2554 		/* no handling is needed for SRQ */
2555 		if (!mqp->ibqp.srq) {
2556 			if (mqp->rq.tail != mqp->rq.head) {
2557 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2558 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2559 				if (recv_mcq->mcq.comp &&
2560 				    mqp->ibqp.recv_cq->comp_handler) {
2561 					if (!recv_mcq->mcq.reset_notify_added) {
2562 						recv_mcq->mcq.reset_notify_added = 1;
2563 						list_add_tail(&recv_mcq->mcq.reset_notify,
2564 							      &cq_armed_list);
2565 					}
2566 				}
2567 				spin_unlock_irqrestore(&recv_mcq->lock,
2568 						       flags_cq);
2569 			}
2570 		}
2571 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2572 	}
2573 	/*At that point all inflight post send were put to be executed as of we
2574 	 * lock/unlock above locks Now need to arm all involved CQs.
2575 	 */
2576 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2577 		mcq->comp(mcq, NULL);
2578 	}
2579 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2580 }
2581 
2582 static void delay_drop_handler(struct work_struct *work)
2583 {
2584 	int err;
2585 	struct mlx5_ib_delay_drop *delay_drop =
2586 		container_of(work, struct mlx5_ib_delay_drop,
2587 			     delay_drop_work);
2588 
2589 	atomic_inc(&delay_drop->events_cnt);
2590 
2591 	mutex_lock(&delay_drop->lock);
2592 	err = mlx5_core_set_delay_drop(delay_drop->dev, delay_drop->timeout);
2593 	if (err) {
2594 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
2595 			     delay_drop->timeout);
2596 		delay_drop->activate = false;
2597 	}
2598 	mutex_unlock(&delay_drop->lock);
2599 }
2600 
2601 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2602 				 struct ib_event *ibev)
2603 {
2604 	u32 port = (eqe->data.port.port >> 4) & 0xf;
2605 
2606 	switch (eqe->sub_type) {
2607 	case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
2608 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2609 					    IB_LINK_LAYER_ETHERNET)
2610 			schedule_work(&ibdev->delay_drop.delay_drop_work);
2611 		break;
2612 	default: /* do nothing */
2613 		return;
2614 	}
2615 }
2616 
2617 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2618 			      struct ib_event *ibev)
2619 {
2620 	u32 port = (eqe->data.port.port >> 4) & 0xf;
2621 
2622 	ibev->element.port_num = port;
2623 
2624 	switch (eqe->sub_type) {
2625 	case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
2626 	case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
2627 	case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
2628 		/* In RoCE, port up/down events are handled in
2629 		 * mlx5_netdev_event().
2630 		 */
2631 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2632 					    IB_LINK_LAYER_ETHERNET)
2633 			return -EINVAL;
2634 
2635 		ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
2636 				IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2637 		break;
2638 
2639 	case MLX5_PORT_CHANGE_SUBTYPE_LID:
2640 		ibev->event = IB_EVENT_LID_CHANGE;
2641 		break;
2642 
2643 	case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
2644 		ibev->event = IB_EVENT_PKEY_CHANGE;
2645 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2646 		break;
2647 
2648 	case MLX5_PORT_CHANGE_SUBTYPE_GUID:
2649 		ibev->event = IB_EVENT_GID_CHANGE;
2650 		break;
2651 
2652 	case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
2653 		ibev->event = IB_EVENT_CLIENT_REREGISTER;
2654 		break;
2655 	default:
2656 		return -EINVAL;
2657 	}
2658 
2659 	return 0;
2660 }
2661 
2662 static void mlx5_ib_handle_event(struct work_struct *_work)
2663 {
2664 	struct mlx5_ib_event_work *work =
2665 		container_of(_work, struct mlx5_ib_event_work, work);
2666 	struct mlx5_ib_dev *ibdev;
2667 	struct ib_event ibev;
2668 	bool fatal = false;
2669 
2670 	if (work->is_slave) {
2671 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
2672 		if (!ibdev)
2673 			goto out;
2674 	} else {
2675 		ibdev = work->dev;
2676 	}
2677 
2678 	switch (work->event) {
2679 	case MLX5_DEV_EVENT_SYS_ERROR:
2680 		ibev.event = IB_EVENT_DEVICE_FATAL;
2681 		mlx5_ib_handle_internal_error(ibdev);
2682 		ibev.element.port_num  = (u8)(unsigned long)work->param;
2683 		fatal = true;
2684 		break;
2685 	case MLX5_EVENT_TYPE_PORT_CHANGE:
2686 		if (handle_port_change(ibdev, work->param, &ibev))
2687 			goto out;
2688 		break;
2689 	case MLX5_EVENT_TYPE_GENERAL_EVENT:
2690 		handle_general_event(ibdev, work->param, &ibev);
2691 		fallthrough;
2692 	default:
2693 		goto out;
2694 	}
2695 
2696 	ibev.device = &ibdev->ib_dev;
2697 
2698 	if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
2699 		mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
2700 		goto out;
2701 	}
2702 
2703 	if (ibdev->ib_active)
2704 		ib_dispatch_event(&ibev);
2705 
2706 	if (fatal)
2707 		ibdev->ib_active = false;
2708 out:
2709 	kfree(work);
2710 }
2711 
2712 static int mlx5_ib_event(struct notifier_block *nb,
2713 			 unsigned long event, void *param)
2714 {
2715 	struct mlx5_ib_event_work *work;
2716 
2717 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2718 	if (!work)
2719 		return NOTIFY_DONE;
2720 
2721 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2722 	work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
2723 	work->is_slave = false;
2724 	work->param = param;
2725 	work->event = event;
2726 
2727 	queue_work(mlx5_ib_event_wq, &work->work);
2728 
2729 	return NOTIFY_OK;
2730 }
2731 
2732 static int mlx5_ib_event_slave_port(struct notifier_block *nb,
2733 				    unsigned long event, void *param)
2734 {
2735 	struct mlx5_ib_event_work *work;
2736 
2737 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2738 	if (!work)
2739 		return NOTIFY_DONE;
2740 
2741 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2742 	work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
2743 	work->is_slave = true;
2744 	work->param = param;
2745 	work->event = event;
2746 	queue_work(mlx5_ib_event_wq, &work->work);
2747 
2748 	return NOTIFY_OK;
2749 }
2750 
2751 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2752 {
2753 	struct mlx5_hca_vport_context vport_ctx;
2754 	int err;
2755 	int port;
2756 
2757 	if (MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_IB)
2758 		return 0;
2759 
2760 	for (port = 1; port <= dev->num_ports; port++) {
2761 		if (!MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2762 			dev->port_caps[port - 1].has_smi = true;
2763 			continue;
2764 		}
2765 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port, 0,
2766 						   &vport_ctx);
2767 		if (err) {
2768 			mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2769 				    port, err);
2770 			return err;
2771 		}
2772 		dev->port_caps[port - 1].has_smi = vport_ctx.has_smi;
2773 	}
2774 
2775 	return 0;
2776 }
2777 
2778 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2779 {
2780 	unsigned int port;
2781 
2782 	rdma_for_each_port (&dev->ib_dev, port)
2783 		mlx5_query_ext_port_caps(dev, port);
2784 }
2785 
2786 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
2787 {
2788 	switch (umr_fence_cap) {
2789 	case MLX5_CAP_UMR_FENCE_NONE:
2790 		return MLX5_FENCE_MODE_NONE;
2791 	case MLX5_CAP_UMR_FENCE_SMALL:
2792 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
2793 	default:
2794 		return MLX5_FENCE_MODE_STRONG_ORDERING;
2795 	}
2796 }
2797 
2798 static int mlx5_ib_dev_res_init(struct mlx5_ib_dev *dev)
2799 {
2800 	struct mlx5_ib_resources *devr = &dev->devr;
2801 	struct ib_srq_init_attr attr;
2802 	struct ib_device *ibdev;
2803 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2804 	int port;
2805 	int ret = 0;
2806 
2807 	ibdev = &dev->ib_dev;
2808 
2809 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
2810 		return -EOPNOTSUPP;
2811 
2812 	devr->p0 = ib_alloc_pd(ibdev, 0);
2813 	if (IS_ERR(devr->p0))
2814 		return PTR_ERR(devr->p0);
2815 
2816 	devr->c0 = ib_create_cq(ibdev, NULL, NULL, NULL, &cq_attr);
2817 	if (IS_ERR(devr->c0)) {
2818 		ret = PTR_ERR(devr->c0);
2819 		goto error1;
2820 	}
2821 
2822 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0);
2823 	if (ret)
2824 		goto error2;
2825 
2826 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0);
2827 	if (ret)
2828 		goto error3;
2829 
2830 	memset(&attr, 0, sizeof(attr));
2831 	attr.attr.max_sge = 1;
2832 	attr.attr.max_wr = 1;
2833 	attr.srq_type = IB_SRQT_XRC;
2834 	attr.ext.cq = devr->c0;
2835 
2836 	devr->s0 = ib_create_srq(devr->p0, &attr);
2837 	if (IS_ERR(devr->s0)) {
2838 		ret = PTR_ERR(devr->s0);
2839 		goto err_create;
2840 	}
2841 
2842 	memset(&attr, 0, sizeof(attr));
2843 	attr.attr.max_sge = 1;
2844 	attr.attr.max_wr = 1;
2845 	attr.srq_type = IB_SRQT_BASIC;
2846 
2847 	devr->s1 = ib_create_srq(devr->p0, &attr);
2848 	if (IS_ERR(devr->s1)) {
2849 		ret = PTR_ERR(devr->s1);
2850 		goto error6;
2851 	}
2852 
2853 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2854 		INIT_WORK(&devr->ports[port].pkey_change_work,
2855 			  pkey_change_handler);
2856 
2857 	return 0;
2858 
2859 error6:
2860 	ib_destroy_srq(devr->s0);
2861 err_create:
2862 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2863 error3:
2864 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2865 error2:
2866 	ib_destroy_cq(devr->c0);
2867 error1:
2868 	ib_dealloc_pd(devr->p0);
2869 	return ret;
2870 }
2871 
2872 static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev)
2873 {
2874 	struct mlx5_ib_resources *devr = &dev->devr;
2875 	int port;
2876 
2877 	/*
2878 	 * Make sure no change P_Key work items are still executing.
2879 	 *
2880 	 * At this stage, the mlx5_ib_event should be unregistered
2881 	 * and it ensures that no new works are added.
2882 	 */
2883 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2884 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2885 
2886 	ib_destroy_srq(devr->s1);
2887 	ib_destroy_srq(devr->s0);
2888 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2889 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2890 	ib_destroy_cq(devr->c0);
2891 	ib_dealloc_pd(devr->p0);
2892 }
2893 
2894 static u32 get_core_cap_flags(struct ib_device *ibdev,
2895 			      struct mlx5_hca_vport_context *rep)
2896 {
2897 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2898 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2899 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2900 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2901 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
2902 	u32 ret = 0;
2903 
2904 	if (rep->grh_required)
2905 		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
2906 
2907 	if (ll == IB_LINK_LAYER_INFINIBAND)
2908 		return ret | RDMA_CORE_PORT_IBA_IB;
2909 
2910 	if (raw_support)
2911 		ret |= RDMA_CORE_PORT_RAW_PACKET;
2912 
2913 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2914 		return ret;
2915 
2916 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2917 		return ret;
2918 
2919 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2920 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2921 
2922 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2923 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2924 
2925 	return ret;
2926 }
2927 
2928 static int mlx5_port_immutable(struct ib_device *ibdev, u32 port_num,
2929 			       struct ib_port_immutable *immutable)
2930 {
2931 	struct ib_port_attr attr;
2932 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2933 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
2934 	struct mlx5_hca_vport_context rep = {0};
2935 	int err;
2936 
2937 	err = ib_query_port(ibdev, port_num, &attr);
2938 	if (err)
2939 		return err;
2940 
2941 	if (ll == IB_LINK_LAYER_INFINIBAND) {
2942 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
2943 						   &rep);
2944 		if (err)
2945 			return err;
2946 	}
2947 
2948 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2949 	immutable->gid_tbl_len = attr.gid_tbl_len;
2950 	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
2951 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2952 
2953 	return 0;
2954 }
2955 
2956 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u32 port_num,
2957 				   struct ib_port_immutable *immutable)
2958 {
2959 	struct ib_port_attr attr;
2960 	int err;
2961 
2962 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2963 
2964 	err = ib_query_port(ibdev, port_num, &attr);
2965 	if (err)
2966 		return err;
2967 
2968 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2969 	immutable->gid_tbl_len = attr.gid_tbl_len;
2970 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2971 
2972 	return 0;
2973 }
2974 
2975 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
2976 {
2977 	struct mlx5_ib_dev *dev =
2978 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2979 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
2980 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
2981 		 fw_rev_sub(dev->mdev));
2982 }
2983 
2984 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
2985 {
2986 	struct mlx5_core_dev *mdev = dev->mdev;
2987 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
2988 								 MLX5_FLOW_NAMESPACE_LAG);
2989 	struct mlx5_flow_table *ft;
2990 	int err;
2991 
2992 	if (!ns || !mlx5_lag_is_active(mdev))
2993 		return 0;
2994 
2995 	err = mlx5_cmd_create_vport_lag(mdev);
2996 	if (err)
2997 		return err;
2998 
2999 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3000 	if (IS_ERR(ft)) {
3001 		err = PTR_ERR(ft);
3002 		goto err_destroy_vport_lag;
3003 	}
3004 
3005 	dev->flow_db->lag_demux_ft = ft;
3006 	dev->lag_ports = mlx5_lag_get_num_ports(mdev);
3007 	dev->lag_active = true;
3008 	return 0;
3009 
3010 err_destroy_vport_lag:
3011 	mlx5_cmd_destroy_vport_lag(mdev);
3012 	return err;
3013 }
3014 
3015 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3016 {
3017 	struct mlx5_core_dev *mdev = dev->mdev;
3018 
3019 	if (dev->lag_active) {
3020 		dev->lag_active = false;
3021 
3022 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
3023 		dev->flow_db->lag_demux_ft = NULL;
3024 
3025 		mlx5_cmd_destroy_vport_lag(mdev);
3026 	}
3027 }
3028 
3029 static void mlx5_netdev_notifier_register(struct mlx5_roce *roce,
3030 					  struct net_device *netdev)
3031 {
3032 	int err;
3033 
3034 	if (roce->tracking_netdev)
3035 		return;
3036 	roce->tracking_netdev = netdev;
3037 	roce->nb.notifier_call = mlx5_netdev_event;
3038 	err = register_netdevice_notifier_dev_net(netdev, &roce->nb, &roce->nn);
3039 	WARN_ON(err);
3040 }
3041 
3042 static void mlx5_netdev_notifier_unregister(struct mlx5_roce *roce)
3043 {
3044 	if (!roce->tracking_netdev)
3045 		return;
3046 	unregister_netdevice_notifier_dev_net(roce->tracking_netdev, &roce->nb,
3047 					      &roce->nn);
3048 	roce->tracking_netdev = NULL;
3049 }
3050 
3051 static int mlx5e_mdev_notifier_event(struct notifier_block *nb,
3052 				     unsigned long event, void *data)
3053 {
3054 	struct mlx5_roce *roce = container_of(nb, struct mlx5_roce, mdev_nb);
3055 	struct net_device *netdev = data;
3056 
3057 	switch (event) {
3058 	case MLX5_DRIVER_EVENT_UPLINK_NETDEV:
3059 		if (netdev)
3060 			mlx5_netdev_notifier_register(roce, netdev);
3061 		else
3062 			mlx5_netdev_notifier_unregister(roce);
3063 		break;
3064 	default:
3065 		return NOTIFY_DONE;
3066 	}
3067 
3068 	return NOTIFY_OK;
3069 }
3070 
3071 static void mlx5_mdev_netdev_track(struct mlx5_ib_dev *dev, u32 port_num)
3072 {
3073 	struct mlx5_roce *roce = &dev->port[port_num].roce;
3074 
3075 	roce->mdev_nb.notifier_call = mlx5e_mdev_notifier_event;
3076 	mlx5_blocking_notifier_register(dev->mdev, &roce->mdev_nb);
3077 	mlx5_core_uplink_netdev_event_replay(dev->mdev);
3078 }
3079 
3080 static void mlx5_mdev_netdev_untrack(struct mlx5_ib_dev *dev, u32 port_num)
3081 {
3082 	struct mlx5_roce *roce = &dev->port[port_num].roce;
3083 
3084 	mlx5_blocking_notifier_unregister(dev->mdev, &roce->mdev_nb);
3085 	mlx5_netdev_notifier_unregister(roce);
3086 }
3087 
3088 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3089 {
3090 	int err;
3091 
3092 	if (!dev->is_rep && dev->profile != &raw_eth_profile) {
3093 		err = mlx5_nic_vport_enable_roce(dev->mdev);
3094 		if (err)
3095 			return err;
3096 	}
3097 
3098 	err = mlx5_eth_lag_init(dev);
3099 	if (err)
3100 		goto err_disable_roce;
3101 
3102 	return 0;
3103 
3104 err_disable_roce:
3105 	if (!dev->is_rep && dev->profile != &raw_eth_profile)
3106 		mlx5_nic_vport_disable_roce(dev->mdev);
3107 
3108 	return err;
3109 }
3110 
3111 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3112 {
3113 	mlx5_eth_lag_cleanup(dev);
3114 	if (!dev->is_rep && dev->profile != &raw_eth_profile)
3115 		mlx5_nic_vport_disable_roce(dev->mdev);
3116 }
3117 
3118 static int mlx5_ib_rn_get_params(struct ib_device *device, u32 port_num,
3119 				 enum rdma_netdev_t type,
3120 				 struct rdma_netdev_alloc_params *params)
3121 {
3122 	if (type != RDMA_NETDEV_IPOIB)
3123 		return -EOPNOTSUPP;
3124 
3125 	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
3126 }
3127 
3128 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
3129 				       size_t count, loff_t *pos)
3130 {
3131 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3132 	char lbuf[20];
3133 	int len;
3134 
3135 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
3136 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
3137 }
3138 
3139 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
3140 					size_t count, loff_t *pos)
3141 {
3142 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3143 	u32 timeout;
3144 	u32 var;
3145 
3146 	if (kstrtouint_from_user(buf, count, 0, &var))
3147 		return -EFAULT;
3148 
3149 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
3150 			1000);
3151 	if (timeout != var)
3152 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
3153 			    timeout);
3154 
3155 	delay_drop->timeout = timeout;
3156 
3157 	return count;
3158 }
3159 
3160 static const struct file_operations fops_delay_drop_timeout = {
3161 	.owner	= THIS_MODULE,
3162 	.open	= simple_open,
3163 	.write	= delay_drop_timeout_write,
3164 	.read	= delay_drop_timeout_read,
3165 };
3166 
3167 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
3168 				      struct mlx5_ib_multiport_info *mpi)
3169 {
3170 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3171 	struct mlx5_ib_port *port = &ibdev->port[port_num];
3172 	int comps;
3173 	int err;
3174 	int i;
3175 
3176 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3177 
3178 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
3179 
3180 	spin_lock(&port->mp.mpi_lock);
3181 	if (!mpi->ibdev) {
3182 		spin_unlock(&port->mp.mpi_lock);
3183 		return;
3184 	}
3185 
3186 	mpi->ibdev = NULL;
3187 
3188 	spin_unlock(&port->mp.mpi_lock);
3189 	if (mpi->mdev_events.notifier_call)
3190 		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
3191 	mpi->mdev_events.notifier_call = NULL;
3192 	mlx5_mdev_netdev_untrack(ibdev, port_num);
3193 	spin_lock(&port->mp.mpi_lock);
3194 
3195 	comps = mpi->mdev_refcnt;
3196 	if (comps) {
3197 		mpi->unaffiliate = true;
3198 		init_completion(&mpi->unref_comp);
3199 		spin_unlock(&port->mp.mpi_lock);
3200 
3201 		for (i = 0; i < comps; i++)
3202 			wait_for_completion(&mpi->unref_comp);
3203 
3204 		spin_lock(&port->mp.mpi_lock);
3205 		mpi->unaffiliate = false;
3206 	}
3207 
3208 	port->mp.mpi = NULL;
3209 
3210 	spin_unlock(&port->mp.mpi_lock);
3211 
3212 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
3213 
3214 	mlx5_ib_dbg(ibdev, "unaffiliated port %u\n", port_num + 1);
3215 	/* Log an error, still needed to cleanup the pointers and add
3216 	 * it back to the list.
3217 	 */
3218 	if (err)
3219 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
3220 			    port_num + 1);
3221 
3222 	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
3223 }
3224 
3225 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
3226 				    struct mlx5_ib_multiport_info *mpi)
3227 {
3228 	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3229 	int err;
3230 
3231 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3232 
3233 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
3234 	if (ibdev->port[port_num].mp.mpi) {
3235 		mlx5_ib_dbg(ibdev, "port %u already affiliated.\n",
3236 			    port_num + 1);
3237 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3238 		return false;
3239 	}
3240 
3241 	ibdev->port[port_num].mp.mpi = mpi;
3242 	mpi->ibdev = ibdev;
3243 	mpi->mdev_events.notifier_call = NULL;
3244 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3245 
3246 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
3247 	if (err)
3248 		goto unbind;
3249 
3250 	mlx5_mdev_netdev_track(ibdev, port_num);
3251 
3252 	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
3253 	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
3254 
3255 	mlx5_ib_init_cong_debugfs(ibdev, port_num);
3256 
3257 	return true;
3258 
3259 unbind:
3260 	mlx5_ib_unbind_slave_port(ibdev, mpi);
3261 	return false;
3262 }
3263 
3264 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
3265 {
3266 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3267 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3268 							  port_num + 1);
3269 	struct mlx5_ib_multiport_info *mpi;
3270 	int err;
3271 	u32 i;
3272 
3273 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3274 		return 0;
3275 
3276 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
3277 						     &dev->sys_image_guid);
3278 	if (err)
3279 		return err;
3280 
3281 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3282 	if (err)
3283 		return err;
3284 
3285 	mutex_lock(&mlx5_ib_multiport_mutex);
3286 	for (i = 0; i < dev->num_ports; i++) {
3287 		bool bound = false;
3288 
3289 		/* build a stub multiport info struct for the native port. */
3290 		if (i == port_num) {
3291 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
3292 			if (!mpi) {
3293 				mutex_unlock(&mlx5_ib_multiport_mutex);
3294 				mlx5_nic_vport_disable_roce(dev->mdev);
3295 				return -ENOMEM;
3296 			}
3297 
3298 			mpi->is_master = true;
3299 			mpi->mdev = dev->mdev;
3300 			mpi->sys_image_guid = dev->sys_image_guid;
3301 			dev->port[i].mp.mpi = mpi;
3302 			mpi->ibdev = dev;
3303 			mpi = NULL;
3304 			continue;
3305 		}
3306 
3307 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
3308 				    list) {
3309 			if (dev->sys_image_guid == mpi->sys_image_guid &&
3310 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
3311 				bound = mlx5_ib_bind_slave_port(dev, mpi);
3312 			}
3313 
3314 			if (bound) {
3315 				dev_dbg(mpi->mdev->device,
3316 					"removing port from unaffiliated list.\n");
3317 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
3318 				list_del(&mpi->list);
3319 				break;
3320 			}
3321 		}
3322 		if (!bound)
3323 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
3324 				    i + 1);
3325 	}
3326 
3327 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
3328 	mutex_unlock(&mlx5_ib_multiport_mutex);
3329 	return err;
3330 }
3331 
3332 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
3333 {
3334 	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3335 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3336 							  port_num + 1);
3337 	u32 i;
3338 
3339 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3340 		return;
3341 
3342 	mutex_lock(&mlx5_ib_multiport_mutex);
3343 	for (i = 0; i < dev->num_ports; i++) {
3344 		if (dev->port[i].mp.mpi) {
3345 			/* Destroy the native port stub */
3346 			if (i == port_num) {
3347 				kfree(dev->port[i].mp.mpi);
3348 				dev->port[i].mp.mpi = NULL;
3349 			} else {
3350 				mlx5_ib_dbg(dev, "unbinding port_num: %u\n",
3351 					    i + 1);
3352 				list_add_tail(&dev->port[i].mp.mpi->list,
3353 					      &mlx5_ib_unaffiliated_port_list);
3354 				mlx5_ib_unbind_slave_port(dev,
3355 							  dev->port[i].mp.mpi);
3356 			}
3357 		}
3358 	}
3359 
3360 	mlx5_ib_dbg(dev, "removing from devlist\n");
3361 	list_del(&dev->ib_dev_list);
3362 	mutex_unlock(&mlx5_ib_multiport_mutex);
3363 
3364 	mlx5_nic_vport_disable_roce(dev->mdev);
3365 }
3366 
3367 static int mmap_obj_cleanup(struct ib_uobject *uobject,
3368 			    enum rdma_remove_reason why,
3369 			    struct uverbs_attr_bundle *attrs)
3370 {
3371 	struct mlx5_user_mmap_entry *obj = uobject->object;
3372 
3373 	rdma_user_mmap_entry_remove(&obj->rdma_entry);
3374 	return 0;
3375 }
3376 
3377 static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c,
3378 					    struct mlx5_user_mmap_entry *entry,
3379 					    size_t length)
3380 {
3381 	return rdma_user_mmap_entry_insert_range(
3382 		&c->ibucontext, &entry->rdma_entry, length,
3383 		(MLX5_IB_MMAP_OFFSET_START << 16),
3384 		((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1));
3385 }
3386 
3387 static struct mlx5_user_mmap_entry *
3388 alloc_var_entry(struct mlx5_ib_ucontext *c)
3389 {
3390 	struct mlx5_user_mmap_entry *entry;
3391 	struct mlx5_var_table *var_table;
3392 	u32 page_idx;
3393 	int err;
3394 
3395 	var_table = &to_mdev(c->ibucontext.device)->var_table;
3396 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3397 	if (!entry)
3398 		return ERR_PTR(-ENOMEM);
3399 
3400 	mutex_lock(&var_table->bitmap_lock);
3401 	page_idx = find_first_zero_bit(var_table->bitmap,
3402 				       var_table->num_var_hw_entries);
3403 	if (page_idx >= var_table->num_var_hw_entries) {
3404 		err = -ENOSPC;
3405 		mutex_unlock(&var_table->bitmap_lock);
3406 		goto end;
3407 	}
3408 
3409 	set_bit(page_idx, var_table->bitmap);
3410 	mutex_unlock(&var_table->bitmap_lock);
3411 
3412 	entry->address = var_table->hw_start_addr +
3413 				(page_idx * var_table->stride_size);
3414 	entry->page_idx = page_idx;
3415 	entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
3416 
3417 	err = mlx5_rdma_user_mmap_entry_insert(c, entry,
3418 					       var_table->stride_size);
3419 	if (err)
3420 		goto err_insert;
3421 
3422 	return entry;
3423 
3424 err_insert:
3425 	mutex_lock(&var_table->bitmap_lock);
3426 	clear_bit(page_idx, var_table->bitmap);
3427 	mutex_unlock(&var_table->bitmap_lock);
3428 end:
3429 	kfree(entry);
3430 	return ERR_PTR(err);
3431 }
3432 
3433 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
3434 	struct uverbs_attr_bundle *attrs)
3435 {
3436 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3437 		attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3438 	struct mlx5_ib_ucontext *c;
3439 	struct mlx5_user_mmap_entry *entry;
3440 	u64 mmap_offset;
3441 	u32 length;
3442 	int err;
3443 
3444 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3445 	if (IS_ERR(c))
3446 		return PTR_ERR(c);
3447 
3448 	entry = alloc_var_entry(c);
3449 	if (IS_ERR(entry))
3450 		return PTR_ERR(entry);
3451 
3452 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3453 	length = entry->rdma_entry.npages * PAGE_SIZE;
3454 	uobj->object = entry;
3455 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3456 
3457 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3458 			     &mmap_offset, sizeof(mmap_offset));
3459 	if (err)
3460 		return err;
3461 
3462 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3463 			     &entry->page_idx, sizeof(entry->page_idx));
3464 	if (err)
3465 		return err;
3466 
3467 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3468 			     &length, sizeof(length));
3469 	return err;
3470 }
3471 
3472 DECLARE_UVERBS_NAMED_METHOD(
3473 	MLX5_IB_METHOD_VAR_OBJ_ALLOC,
3474 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
3475 			MLX5_IB_OBJECT_VAR,
3476 			UVERBS_ACCESS_NEW,
3477 			UA_MANDATORY),
3478 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3479 			   UVERBS_ATTR_TYPE(u32),
3480 			   UA_MANDATORY),
3481 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3482 			   UVERBS_ATTR_TYPE(u32),
3483 			   UA_MANDATORY),
3484 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3485 			    UVERBS_ATTR_TYPE(u64),
3486 			    UA_MANDATORY));
3487 
3488 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3489 	MLX5_IB_METHOD_VAR_OBJ_DESTROY,
3490 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
3491 			MLX5_IB_OBJECT_VAR,
3492 			UVERBS_ACCESS_DESTROY,
3493 			UA_MANDATORY));
3494 
3495 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
3496 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3497 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
3498 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
3499 
3500 static bool var_is_supported(struct ib_device *device)
3501 {
3502 	struct mlx5_ib_dev *dev = to_mdev(device);
3503 
3504 	return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3505 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
3506 }
3507 
3508 static struct mlx5_user_mmap_entry *
3509 alloc_uar_entry(struct mlx5_ib_ucontext *c,
3510 		enum mlx5_ib_uapi_uar_alloc_type alloc_type)
3511 {
3512 	struct mlx5_user_mmap_entry *entry;
3513 	struct mlx5_ib_dev *dev;
3514 	u32 uar_index;
3515 	int err;
3516 
3517 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3518 	if (!entry)
3519 		return ERR_PTR(-ENOMEM);
3520 
3521 	dev = to_mdev(c->ibucontext.device);
3522 	err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index, c->devx_uid);
3523 	if (err)
3524 		goto end;
3525 
3526 	entry->page_idx = uar_index;
3527 	entry->address = uar_index2paddress(dev, uar_index);
3528 	if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3529 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC;
3530 	else
3531 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC;
3532 
3533 	err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE);
3534 	if (err)
3535 		goto err_insert;
3536 
3537 	return entry;
3538 
3539 err_insert:
3540 	mlx5_cmd_uar_dealloc(dev->mdev, uar_index, c->devx_uid);
3541 end:
3542 	kfree(entry);
3543 	return ERR_PTR(err);
3544 }
3545 
3546 static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)(
3547 	struct uverbs_attr_bundle *attrs)
3548 {
3549 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3550 		attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3551 	enum mlx5_ib_uapi_uar_alloc_type alloc_type;
3552 	struct mlx5_ib_ucontext *c;
3553 	struct mlx5_user_mmap_entry *entry;
3554 	u64 mmap_offset;
3555 	u32 length;
3556 	int err;
3557 
3558 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3559 	if (IS_ERR(c))
3560 		return PTR_ERR(c);
3561 
3562 	err = uverbs_get_const(&alloc_type, attrs,
3563 			       MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE);
3564 	if (err)
3565 		return err;
3566 
3567 	if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF &&
3568 	    alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC)
3569 		return -EOPNOTSUPP;
3570 
3571 	if (!to_mdev(c->ibucontext.device)->wc_support &&
3572 	    alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3573 		return -EOPNOTSUPP;
3574 
3575 	entry = alloc_uar_entry(c, alloc_type);
3576 	if (IS_ERR(entry))
3577 		return PTR_ERR(entry);
3578 
3579 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3580 	length = entry->rdma_entry.npages * PAGE_SIZE;
3581 	uobj->object = entry;
3582 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3583 
3584 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3585 			     &mmap_offset, sizeof(mmap_offset));
3586 	if (err)
3587 		return err;
3588 
3589 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3590 			     &entry->page_idx, sizeof(entry->page_idx));
3591 	if (err)
3592 		return err;
3593 
3594 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3595 			     &length, sizeof(length));
3596 	return err;
3597 }
3598 
3599 DECLARE_UVERBS_NAMED_METHOD(
3600 	MLX5_IB_METHOD_UAR_OBJ_ALLOC,
3601 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE,
3602 			MLX5_IB_OBJECT_UAR,
3603 			UVERBS_ACCESS_NEW,
3604 			UA_MANDATORY),
3605 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE,
3606 			     enum mlx5_ib_uapi_uar_alloc_type,
3607 			     UA_MANDATORY),
3608 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3609 			   UVERBS_ATTR_TYPE(u32),
3610 			   UA_MANDATORY),
3611 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3612 			   UVERBS_ATTR_TYPE(u32),
3613 			   UA_MANDATORY),
3614 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3615 			    UVERBS_ATTR_TYPE(u64),
3616 			    UA_MANDATORY));
3617 
3618 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3619 	MLX5_IB_METHOD_UAR_OBJ_DESTROY,
3620 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE,
3621 			MLX5_IB_OBJECT_UAR,
3622 			UVERBS_ACCESS_DESTROY,
3623 			UA_MANDATORY));
3624 
3625 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR,
3626 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3627 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC),
3628 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY));
3629 
3630 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3631 	mlx5_ib_query_context,
3632 	UVERBS_OBJECT_DEVICE,
3633 	UVERBS_METHOD_QUERY_CONTEXT,
3634 	UVERBS_ATTR_PTR_OUT(
3635 		MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
3636 		UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp,
3637 				   dump_fill_mkey),
3638 		UA_MANDATORY));
3639 
3640 static const struct uapi_definition mlx5_ib_defs[] = {
3641 	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
3642 	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
3643 	UAPI_DEF_CHAIN(mlx5_ib_qos_defs),
3644 	UAPI_DEF_CHAIN(mlx5_ib_std_types_defs),
3645 	UAPI_DEF_CHAIN(mlx5_ib_dm_defs),
3646 
3647 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context),
3648 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
3649 				UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
3650 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR),
3651 	{}
3652 };
3653 
3654 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
3655 {
3656 	mlx5_ib_cleanup_multiport_master(dev);
3657 	WARN_ON(!xa_empty(&dev->odp_mkeys));
3658 	mutex_destroy(&dev->cap_mask_mutex);
3659 	WARN_ON(!xa_empty(&dev->sig_mrs));
3660 	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
3661 	mlx5r_macsec_dealloc_gids(dev);
3662 }
3663 
3664 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
3665 {
3666 	struct mlx5_core_dev *mdev = dev->mdev;
3667 	int err, i;
3668 
3669 	dev->ib_dev.node_type = RDMA_NODE_IB_CA;
3670 	dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
3671 	dev->ib_dev.phys_port_cnt = dev->num_ports;
3672 	dev->ib_dev.dev.parent = mdev->device;
3673 	dev->ib_dev.lag_flags = RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
3674 
3675 	for (i = 0; i < dev->num_ports; i++) {
3676 		spin_lock_init(&dev->port[i].mp.mpi_lock);
3677 		rwlock_init(&dev->port[i].roce.netdev_lock);
3678 		dev->port[i].roce.dev = dev;
3679 		dev->port[i].roce.native_port_num = i + 1;
3680 		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
3681 	}
3682 
3683 	err = mlx5r_cmd_query_special_mkeys(dev);
3684 	if (err)
3685 		return err;
3686 
3687 	err = mlx5r_macsec_init_gids_and_devlist(dev);
3688 	if (err)
3689 		return err;
3690 
3691 	err = mlx5_ib_init_multiport_master(dev);
3692 	if (err)
3693 		goto err;
3694 
3695 	err = set_has_smi_cap(dev);
3696 	if (err)
3697 		goto err_mp;
3698 
3699 	err = mlx5_query_max_pkeys(&dev->ib_dev, &dev->pkey_table_len);
3700 	if (err)
3701 		goto err_mp;
3702 
3703 	if (mlx5_use_mad_ifc(dev))
3704 		get_ext_port_caps(dev);
3705 
3706 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_max(mdev);
3707 
3708 	mutex_init(&dev->cap_mask_mutex);
3709 	INIT_LIST_HEAD(&dev->qp_list);
3710 	spin_lock_init(&dev->reset_flow_resource_lock);
3711 	xa_init(&dev->odp_mkeys);
3712 	xa_init(&dev->sig_mrs);
3713 	atomic_set(&dev->mkey_var, 0);
3714 
3715 	spin_lock_init(&dev->dm.lock);
3716 	dev->dm.dev = mdev;
3717 	return 0;
3718 err:
3719 	mlx5r_macsec_dealloc_gids(dev);
3720 err_mp:
3721 	mlx5_ib_cleanup_multiport_master(dev);
3722 	return err;
3723 }
3724 
3725 static int mlx5_ib_enable_driver(struct ib_device *dev)
3726 {
3727 	struct mlx5_ib_dev *mdev = to_mdev(dev);
3728 	int ret;
3729 
3730 	ret = mlx5_ib_test_wc(mdev);
3731 	mlx5_ib_dbg(mdev, "Write-Combining %s",
3732 		    mdev->wc_support ? "supported" : "not supported");
3733 
3734 	return ret;
3735 }
3736 
3737 static const struct ib_device_ops mlx5_ib_dev_ops = {
3738 	.owner = THIS_MODULE,
3739 	.driver_id = RDMA_DRIVER_MLX5,
3740 	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
3741 
3742 	.add_gid = mlx5_ib_add_gid,
3743 	.alloc_mr = mlx5_ib_alloc_mr,
3744 	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
3745 	.alloc_pd = mlx5_ib_alloc_pd,
3746 	.alloc_ucontext = mlx5_ib_alloc_ucontext,
3747 	.attach_mcast = mlx5_ib_mcg_attach,
3748 	.check_mr_status = mlx5_ib_check_mr_status,
3749 	.create_ah = mlx5_ib_create_ah,
3750 	.create_cq = mlx5_ib_create_cq,
3751 	.create_qp = mlx5_ib_create_qp,
3752 	.create_srq = mlx5_ib_create_srq,
3753 	.create_user_ah = mlx5_ib_create_ah,
3754 	.dealloc_pd = mlx5_ib_dealloc_pd,
3755 	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
3756 	.del_gid = mlx5_ib_del_gid,
3757 	.dereg_mr = mlx5_ib_dereg_mr,
3758 	.destroy_ah = mlx5_ib_destroy_ah,
3759 	.destroy_cq = mlx5_ib_destroy_cq,
3760 	.destroy_qp = mlx5_ib_destroy_qp,
3761 	.destroy_srq = mlx5_ib_destroy_srq,
3762 	.detach_mcast = mlx5_ib_mcg_detach,
3763 	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
3764 	.drain_rq = mlx5_ib_drain_rq,
3765 	.drain_sq = mlx5_ib_drain_sq,
3766 	.device_group = &mlx5_attr_group,
3767 	.enable_driver = mlx5_ib_enable_driver,
3768 	.get_dev_fw_str = get_dev_fw_str,
3769 	.get_dma_mr = mlx5_ib_get_dma_mr,
3770 	.get_link_layer = mlx5_ib_port_link_layer,
3771 	.map_mr_sg = mlx5_ib_map_mr_sg,
3772 	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
3773 	.mmap = mlx5_ib_mmap,
3774 	.mmap_free = mlx5_ib_mmap_free,
3775 	.modify_cq = mlx5_ib_modify_cq,
3776 	.modify_device = mlx5_ib_modify_device,
3777 	.modify_port = mlx5_ib_modify_port,
3778 	.modify_qp = mlx5_ib_modify_qp,
3779 	.modify_srq = mlx5_ib_modify_srq,
3780 	.poll_cq = mlx5_ib_poll_cq,
3781 	.post_recv = mlx5_ib_post_recv_nodrain,
3782 	.post_send = mlx5_ib_post_send_nodrain,
3783 	.post_srq_recv = mlx5_ib_post_srq_recv,
3784 	.process_mad = mlx5_ib_process_mad,
3785 	.query_ah = mlx5_ib_query_ah,
3786 	.query_device = mlx5_ib_query_device,
3787 	.query_gid = mlx5_ib_query_gid,
3788 	.query_pkey = mlx5_ib_query_pkey,
3789 	.query_qp = mlx5_ib_query_qp,
3790 	.query_srq = mlx5_ib_query_srq,
3791 	.query_ucontext = mlx5_ib_query_ucontext,
3792 	.reg_user_mr = mlx5_ib_reg_user_mr,
3793 	.reg_user_mr_dmabuf = mlx5_ib_reg_user_mr_dmabuf,
3794 	.req_notify_cq = mlx5_ib_arm_cq,
3795 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
3796 	.resize_cq = mlx5_ib_resize_cq,
3797 
3798 	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
3799 	INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs),
3800 	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
3801 	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
3802 	INIT_RDMA_OBJ_SIZE(ib_qp, mlx5_ib_qp, ibqp),
3803 	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
3804 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
3805 };
3806 
3807 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
3808 	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
3809 };
3810 
3811 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
3812 	.get_vf_config = mlx5_ib_get_vf_config,
3813 	.get_vf_guid = mlx5_ib_get_vf_guid,
3814 	.get_vf_stats = mlx5_ib_get_vf_stats,
3815 	.set_vf_guid = mlx5_ib_set_vf_guid,
3816 	.set_vf_link_state = mlx5_ib_set_vf_link_state,
3817 };
3818 
3819 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
3820 	.alloc_mw = mlx5_ib_alloc_mw,
3821 	.dealloc_mw = mlx5_ib_dealloc_mw,
3822 
3823 	INIT_RDMA_OBJ_SIZE(ib_mw, mlx5_ib_mw, ibmw),
3824 };
3825 
3826 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
3827 	.alloc_xrcd = mlx5_ib_alloc_xrcd,
3828 	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
3829 
3830 	INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd),
3831 };
3832 
3833 static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
3834 {
3835 	struct mlx5_core_dev *mdev = dev->mdev;
3836 	struct mlx5_var_table *var_table = &dev->var_table;
3837 	u8 log_doorbell_bar_size;
3838 	u8 log_doorbell_stride;
3839 	u64 bar_size;
3840 
3841 	log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3842 					log_doorbell_bar_size);
3843 	log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3844 					log_doorbell_stride);
3845 	var_table->hw_start_addr = dev->mdev->bar_addr +
3846 				MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
3847 					doorbell_bar_offset);
3848 	bar_size = (1ULL << log_doorbell_bar_size) * 4096;
3849 	var_table->stride_size = 1ULL << log_doorbell_stride;
3850 	var_table->num_var_hw_entries = div_u64(bar_size,
3851 						var_table->stride_size);
3852 	mutex_init(&var_table->bitmap_lock);
3853 	var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
3854 					  GFP_KERNEL);
3855 	return (var_table->bitmap) ? 0 : -ENOMEM;
3856 }
3857 
3858 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
3859 {
3860 	bitmap_free(dev->var_table.bitmap);
3861 }
3862 
3863 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
3864 {
3865 	struct mlx5_core_dev *mdev = dev->mdev;
3866 	int err;
3867 
3868 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
3869 	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
3870 		ib_set_device_ops(&dev->ib_dev,
3871 				  &mlx5_ib_dev_ipoib_enhanced_ops);
3872 
3873 	if (mlx5_core_is_pf(mdev))
3874 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
3875 
3876 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
3877 
3878 	if (MLX5_CAP_GEN(mdev, imaicl))
3879 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
3880 
3881 	if (MLX5_CAP_GEN(mdev, xrc))
3882 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
3883 
3884 	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
3885 	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3886 	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
3887 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
3888 
3889 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
3890 
3891 	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
3892 		dev->ib_dev.driver_def = mlx5_ib_defs;
3893 
3894 	err = init_node_data(dev);
3895 	if (err)
3896 		return err;
3897 
3898 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
3899 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
3900 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
3901 		mutex_init(&dev->lb.mutex);
3902 
3903 	if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3904 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
3905 		err = mlx5_ib_init_var_table(dev);
3906 		if (err)
3907 			return err;
3908 	}
3909 
3910 	dev->ib_dev.use_cq_dim = true;
3911 
3912 	return 0;
3913 }
3914 
3915 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
3916 	.get_port_immutable = mlx5_port_immutable,
3917 	.query_port = mlx5_ib_query_port,
3918 };
3919 
3920 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
3921 {
3922 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
3923 	return 0;
3924 }
3925 
3926 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
3927 	.get_port_immutable = mlx5_port_rep_immutable,
3928 	.query_port = mlx5_ib_rep_query_port,
3929 	.query_pkey = mlx5_ib_rep_query_pkey,
3930 };
3931 
3932 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
3933 {
3934 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
3935 	return 0;
3936 }
3937 
3938 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
3939 	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
3940 	.create_wq = mlx5_ib_create_wq,
3941 	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
3942 	.destroy_wq = mlx5_ib_destroy_wq,
3943 	.get_netdev = mlx5_ib_get_netdev,
3944 	.modify_wq = mlx5_ib_modify_wq,
3945 
3946 	INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table,
3947 			   ib_rwq_ind_tbl),
3948 };
3949 
3950 static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev)
3951 {
3952 	struct mlx5_core_dev *mdev = dev->mdev;
3953 	enum rdma_link_layer ll;
3954 	int port_type_cap;
3955 	u32 port_num = 0;
3956 	int err;
3957 
3958 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3959 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3960 
3961 	if (ll == IB_LINK_LAYER_ETHERNET) {
3962 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
3963 
3964 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3965 
3966 		/* Register only for native ports */
3967 		mlx5_mdev_netdev_track(dev, port_num);
3968 
3969 		err = mlx5_enable_eth(dev);
3970 		if (err)
3971 			goto cleanup;
3972 	}
3973 
3974 	return 0;
3975 cleanup:
3976 	mlx5_mdev_netdev_untrack(dev, port_num);
3977 	return err;
3978 }
3979 
3980 static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev)
3981 {
3982 	struct mlx5_core_dev *mdev = dev->mdev;
3983 	enum rdma_link_layer ll;
3984 	int port_type_cap;
3985 	u32 port_num;
3986 
3987 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3988 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3989 
3990 	if (ll == IB_LINK_LAYER_ETHERNET) {
3991 		mlx5_disable_eth(dev);
3992 
3993 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3994 		mlx5_mdev_netdev_untrack(dev, port_num);
3995 	}
3996 }
3997 
3998 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
3999 {
4000 	mlx5_ib_init_cong_debugfs(dev,
4001 				  mlx5_core_native_port_num(dev->mdev) - 1);
4002 	return 0;
4003 }
4004 
4005 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
4006 {
4007 	mlx5_ib_cleanup_cong_debugfs(dev,
4008 				     mlx5_core_native_port_num(dev->mdev) - 1);
4009 }
4010 
4011 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
4012 {
4013 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
4014 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
4015 }
4016 
4017 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
4018 {
4019 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
4020 }
4021 
4022 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
4023 {
4024 	int err;
4025 
4026 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
4027 	if (err)
4028 		return err;
4029 
4030 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
4031 	if (err)
4032 		mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4033 
4034 	return err;
4035 }
4036 
4037 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
4038 {
4039 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4040 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4041 }
4042 
4043 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
4044 {
4045 	const char *name;
4046 
4047 	if (!mlx5_lag_is_active(dev->mdev))
4048 		name = "mlx5_%d";
4049 	else
4050 		name = "mlx5_bond_%d";
4051 	return ib_register_device(&dev->ib_dev, name, &dev->mdev->pdev->dev);
4052 }
4053 
4054 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
4055 {
4056 	mlx5_mkey_cache_cleanup(dev);
4057 	mlx5r_umr_resource_cleanup(dev);
4058 }
4059 
4060 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
4061 {
4062 	ib_unregister_device(&dev->ib_dev);
4063 }
4064 
4065 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
4066 {
4067 	int ret;
4068 
4069 	ret = mlx5r_umr_resource_init(dev);
4070 	if (ret)
4071 		return ret;
4072 
4073 	ret = mlx5_mkey_cache_init(dev);
4074 	if (ret) {
4075 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4076 		mlx5r_umr_resource_cleanup(dev);
4077 	}
4078 	return ret;
4079 }
4080 
4081 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
4082 {
4083 	struct dentry *root;
4084 
4085 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4086 		return 0;
4087 
4088 	mutex_init(&dev->delay_drop.lock);
4089 	dev->delay_drop.dev = dev;
4090 	dev->delay_drop.activate = false;
4091 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4092 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4093 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4094 	atomic_set(&dev->delay_drop.events_cnt, 0);
4095 
4096 	if (!mlx5_debugfs_root)
4097 		return 0;
4098 
4099 	root = debugfs_create_dir("delay_drop", mlx5_debugfs_get_dev_root(dev->mdev));
4100 	dev->delay_drop.dir_debugfs = root;
4101 
4102 	debugfs_create_atomic_t("num_timeout_events", 0400, root,
4103 				&dev->delay_drop.events_cnt);
4104 	debugfs_create_atomic_t("num_rqs", 0400, root,
4105 				&dev->delay_drop.rqs_cnt);
4106 	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
4107 			    &fops_delay_drop_timeout);
4108 	return 0;
4109 }
4110 
4111 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
4112 {
4113 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4114 		return;
4115 
4116 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4117 	if (!dev->delay_drop.dir_debugfs)
4118 		return;
4119 
4120 	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
4121 	dev->delay_drop.dir_debugfs = NULL;
4122 }
4123 
4124 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
4125 {
4126 	dev->mdev_events.notifier_call = mlx5_ib_event;
4127 	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
4128 
4129 	mlx5r_macsec_event_register(dev);
4130 
4131 	return 0;
4132 }
4133 
4134 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
4135 {
4136 	mlx5r_macsec_event_unregister(dev);
4137 	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
4138 }
4139 
4140 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
4141 		      const struct mlx5_ib_profile *profile,
4142 		      int stage)
4143 {
4144 	dev->ib_active = false;
4145 
4146 	/* Number of stages to cleanup */
4147 	while (stage) {
4148 		stage--;
4149 		if (profile->stage[stage].cleanup)
4150 			profile->stage[stage].cleanup(dev);
4151 	}
4152 
4153 	kfree(dev->port);
4154 	ib_dealloc_device(&dev->ib_dev);
4155 }
4156 
4157 int __mlx5_ib_add(struct mlx5_ib_dev *dev,
4158 		  const struct mlx5_ib_profile *profile)
4159 {
4160 	int err;
4161 	int i;
4162 
4163 	dev->profile = profile;
4164 
4165 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
4166 		if (profile->stage[i].init) {
4167 			err = profile->stage[i].init(dev);
4168 			if (err)
4169 				goto err_out;
4170 		}
4171 	}
4172 
4173 	dev->ib_active = true;
4174 	return 0;
4175 
4176 err_out:
4177 	/* Clean up stages which were initialized */
4178 	while (i) {
4179 		i--;
4180 		if (profile->stage[i].cleanup)
4181 			profile->stage[i].cleanup(dev);
4182 	}
4183 	return -ENOMEM;
4184 }
4185 
4186 static const struct mlx5_ib_profile pf_profile = {
4187 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4188 		     mlx5_ib_stage_init_init,
4189 		     mlx5_ib_stage_init_cleanup),
4190 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4191 		     mlx5_ib_fs_init,
4192 		     mlx5_ib_fs_cleanup),
4193 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4194 		     mlx5_ib_stage_caps_init,
4195 		     mlx5_ib_stage_caps_cleanup),
4196 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4197 		     mlx5_ib_stage_non_default_cb,
4198 		     NULL),
4199 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4200 		     mlx5_ib_roce_init,
4201 		     mlx5_ib_roce_cleanup),
4202 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4203 		     mlx5_init_qp_table,
4204 		     mlx5_cleanup_qp_table),
4205 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4206 		     mlx5_init_srq_table,
4207 		     mlx5_cleanup_srq_table),
4208 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4209 		     mlx5_ib_dev_res_init,
4210 		     mlx5_ib_dev_res_cleanup),
4211 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4212 		     mlx5_ib_stage_dev_notifier_init,
4213 		     mlx5_ib_stage_dev_notifier_cleanup),
4214 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
4215 		     mlx5_ib_odp_init_one,
4216 		     mlx5_ib_odp_cleanup_one),
4217 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4218 		     mlx5_ib_counters_init,
4219 		     mlx5_ib_counters_cleanup),
4220 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4221 		     mlx5_ib_stage_cong_debugfs_init,
4222 		     mlx5_ib_stage_cong_debugfs_cleanup),
4223 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4224 		     mlx5_ib_stage_uar_init,
4225 		     mlx5_ib_stage_uar_cleanup),
4226 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4227 		     mlx5_ib_stage_bfrag_init,
4228 		     mlx5_ib_stage_bfrag_cleanup),
4229 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4230 		     NULL,
4231 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4232 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4233 		     mlx5_ib_devx_init,
4234 		     mlx5_ib_devx_cleanup),
4235 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4236 		     mlx5_ib_stage_ib_reg_init,
4237 		     mlx5_ib_stage_ib_reg_cleanup),
4238 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4239 		     mlx5_ib_stage_post_ib_reg_umr_init,
4240 		     NULL),
4241 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4242 		     mlx5_ib_stage_delay_drop_init,
4243 		     mlx5_ib_stage_delay_drop_cleanup),
4244 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4245 		     mlx5_ib_restrack_init,
4246 		     NULL),
4247 };
4248 
4249 const struct mlx5_ib_profile raw_eth_profile = {
4250 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4251 		     mlx5_ib_stage_init_init,
4252 		     mlx5_ib_stage_init_cleanup),
4253 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4254 		     mlx5_ib_fs_init,
4255 		     mlx5_ib_fs_cleanup),
4256 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4257 		     mlx5_ib_stage_caps_init,
4258 		     mlx5_ib_stage_caps_cleanup),
4259 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4260 		     mlx5_ib_stage_raw_eth_non_default_cb,
4261 		     NULL),
4262 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4263 		     mlx5_ib_roce_init,
4264 		     mlx5_ib_roce_cleanup),
4265 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4266 		     mlx5_init_qp_table,
4267 		     mlx5_cleanup_qp_table),
4268 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4269 		     mlx5_init_srq_table,
4270 		     mlx5_cleanup_srq_table),
4271 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4272 		     mlx5_ib_dev_res_init,
4273 		     mlx5_ib_dev_res_cleanup),
4274 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4275 		     mlx5_ib_stage_dev_notifier_init,
4276 		     mlx5_ib_stage_dev_notifier_cleanup),
4277 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4278 		     mlx5_ib_counters_init,
4279 		     mlx5_ib_counters_cleanup),
4280 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4281 		     mlx5_ib_stage_cong_debugfs_init,
4282 		     mlx5_ib_stage_cong_debugfs_cleanup),
4283 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4284 		     mlx5_ib_stage_uar_init,
4285 		     mlx5_ib_stage_uar_cleanup),
4286 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4287 		     mlx5_ib_stage_bfrag_init,
4288 		     mlx5_ib_stage_bfrag_cleanup),
4289 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4290 		     NULL,
4291 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4292 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4293 		     mlx5_ib_devx_init,
4294 		     mlx5_ib_devx_cleanup),
4295 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4296 		     mlx5_ib_stage_ib_reg_init,
4297 		     mlx5_ib_stage_ib_reg_cleanup),
4298 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4299 		     mlx5_ib_stage_post_ib_reg_umr_init,
4300 		     NULL),
4301 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4302 		     mlx5_ib_stage_delay_drop_init,
4303 		     mlx5_ib_stage_delay_drop_cleanup),
4304 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4305 		     mlx5_ib_restrack_init,
4306 		     NULL),
4307 };
4308 
4309 static int mlx5r_mp_probe(struct auxiliary_device *adev,
4310 			  const struct auxiliary_device_id *id)
4311 {
4312 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4313 	struct mlx5_core_dev *mdev = idev->mdev;
4314 	struct mlx5_ib_multiport_info *mpi;
4315 	struct mlx5_ib_dev *dev;
4316 	bool bound = false;
4317 	int err;
4318 
4319 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4320 	if (!mpi)
4321 		return -ENOMEM;
4322 
4323 	mpi->mdev = mdev;
4324 	err = mlx5_query_nic_vport_system_image_guid(mdev,
4325 						     &mpi->sys_image_guid);
4326 	if (err) {
4327 		kfree(mpi);
4328 		return err;
4329 	}
4330 
4331 	mutex_lock(&mlx5_ib_multiport_mutex);
4332 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
4333 		if (dev->sys_image_guid == mpi->sys_image_guid)
4334 			bound = mlx5_ib_bind_slave_port(dev, mpi);
4335 
4336 		if (bound) {
4337 			rdma_roce_rescan_device(&dev->ib_dev);
4338 			mpi->ibdev->ib_active = true;
4339 			break;
4340 		}
4341 	}
4342 
4343 	if (!bound) {
4344 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4345 		dev_dbg(mdev->device,
4346 			"no suitable IB device found to bind to, added to unaffiliated list.\n");
4347 	}
4348 	mutex_unlock(&mlx5_ib_multiport_mutex);
4349 
4350 	auxiliary_set_drvdata(adev, mpi);
4351 	return 0;
4352 }
4353 
4354 static void mlx5r_mp_remove(struct auxiliary_device *adev)
4355 {
4356 	struct mlx5_ib_multiport_info *mpi;
4357 
4358 	mpi = auxiliary_get_drvdata(adev);
4359 	mutex_lock(&mlx5_ib_multiport_mutex);
4360 	if (mpi->ibdev)
4361 		mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
4362 	else
4363 		list_del(&mpi->list);
4364 	mutex_unlock(&mlx5_ib_multiport_mutex);
4365 	kfree(mpi);
4366 }
4367 
4368 static int mlx5r_probe(struct auxiliary_device *adev,
4369 		       const struct auxiliary_device_id *id)
4370 {
4371 	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4372 	struct mlx5_core_dev *mdev = idev->mdev;
4373 	const struct mlx5_ib_profile *profile;
4374 	int port_type_cap, num_ports, ret;
4375 	enum rdma_link_layer ll;
4376 	struct mlx5_ib_dev *dev;
4377 
4378 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4379 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4380 
4381 	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
4382 			MLX5_CAP_GEN(mdev, num_vhca_ports));
4383 	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
4384 	if (!dev)
4385 		return -ENOMEM;
4386 	dev->port = kcalloc(num_ports, sizeof(*dev->port),
4387 			     GFP_KERNEL);
4388 	if (!dev->port) {
4389 		ib_dealloc_device(&dev->ib_dev);
4390 		return -ENOMEM;
4391 	}
4392 
4393 	dev->mdev = mdev;
4394 	dev->num_ports = num_ports;
4395 
4396 	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_get_roce_state(mdev))
4397 		profile = &raw_eth_profile;
4398 	else
4399 		profile = &pf_profile;
4400 
4401 	ret = __mlx5_ib_add(dev, profile);
4402 	if (ret) {
4403 		kfree(dev->port);
4404 		ib_dealloc_device(&dev->ib_dev);
4405 		return ret;
4406 	}
4407 
4408 	auxiliary_set_drvdata(adev, dev);
4409 	return 0;
4410 }
4411 
4412 static void mlx5r_remove(struct auxiliary_device *adev)
4413 {
4414 	struct mlx5_ib_dev *dev;
4415 
4416 	dev = auxiliary_get_drvdata(adev);
4417 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
4418 }
4419 
4420 static const struct auxiliary_device_id mlx5r_mp_id_table[] = {
4421 	{ .name = MLX5_ADEV_NAME ".multiport", },
4422 	{},
4423 };
4424 
4425 static const struct auxiliary_device_id mlx5r_id_table[] = {
4426 	{ .name = MLX5_ADEV_NAME ".rdma", },
4427 	{},
4428 };
4429 
4430 MODULE_DEVICE_TABLE(auxiliary, mlx5r_mp_id_table);
4431 MODULE_DEVICE_TABLE(auxiliary, mlx5r_id_table);
4432 
4433 static struct auxiliary_driver mlx5r_mp_driver = {
4434 	.name = "multiport",
4435 	.probe = mlx5r_mp_probe,
4436 	.remove = mlx5r_mp_remove,
4437 	.id_table = mlx5r_mp_id_table,
4438 };
4439 
4440 static struct auxiliary_driver mlx5r_driver = {
4441 	.name = "rdma",
4442 	.probe = mlx5r_probe,
4443 	.remove = mlx5r_remove,
4444 	.id_table = mlx5r_id_table,
4445 };
4446 
4447 static int __init mlx5_ib_init(void)
4448 {
4449 	int ret;
4450 
4451 	xlt_emergency_page = (void *)__get_free_page(GFP_KERNEL);
4452 	if (!xlt_emergency_page)
4453 		return -ENOMEM;
4454 
4455 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
4456 	if (!mlx5_ib_event_wq) {
4457 		free_page((unsigned long)xlt_emergency_page);
4458 		return -ENOMEM;
4459 	}
4460 
4461 	ret = mlx5_ib_qp_event_init();
4462 	if (ret)
4463 		goto qp_event_err;
4464 
4465 	mlx5_ib_odp_init();
4466 	ret = mlx5r_rep_init();
4467 	if (ret)
4468 		goto rep_err;
4469 	ret = auxiliary_driver_register(&mlx5r_mp_driver);
4470 	if (ret)
4471 		goto mp_err;
4472 	ret = auxiliary_driver_register(&mlx5r_driver);
4473 	if (ret)
4474 		goto drv_err;
4475 	return 0;
4476 
4477 drv_err:
4478 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4479 mp_err:
4480 	mlx5r_rep_cleanup();
4481 rep_err:
4482 	mlx5_ib_qp_event_cleanup();
4483 qp_event_err:
4484 	destroy_workqueue(mlx5_ib_event_wq);
4485 	free_page((unsigned long)xlt_emergency_page);
4486 	return ret;
4487 }
4488 
4489 static void __exit mlx5_ib_cleanup(void)
4490 {
4491 	auxiliary_driver_unregister(&mlx5r_driver);
4492 	auxiliary_driver_unregister(&mlx5r_mp_driver);
4493 	mlx5r_rep_cleanup();
4494 
4495 	mlx5_ib_qp_event_cleanup();
4496 	destroy_workqueue(mlx5_ib_event_wq);
4497 	free_page((unsigned long)xlt_emergency_page);
4498 }
4499 
4500 module_init(mlx5_ib_init);
4501 module_exit(mlx5_ib_cleanup);
4502