xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision 52cdded0)
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2013-2020, Mellanox Technologies inc. All rights reserved.
4  */
5 
6 #include <linux/debugfs.h>
7 #include <linux/highmem.h>
8 #include <linux/module.h>
9 #include <linux/init.h>
10 #include <linux/errno.h>
11 #include <linux/pci.h>
12 #include <linux/dma-mapping.h>
13 #include <linux/slab.h>
14 #include <linux/bitmap.h>
15 #include <linux/sched.h>
16 #include <linux/sched/mm.h>
17 #include <linux/sched/task.h>
18 #include <linux/delay.h>
19 #include <rdma/ib_user_verbs.h>
20 #include <rdma/ib_addr.h>
21 #include <rdma/ib_cache.h>
22 #include <linux/mlx5/port.h>
23 #include <linux/mlx5/vport.h>
24 #include <linux/mlx5/fs.h>
25 #include <linux/mlx5/eswitch.h>
26 #include <linux/list.h>
27 #include <rdma/ib_smi.h>
28 #include <rdma/ib_umem.h>
29 #include <rdma/lag.h>
30 #include <linux/in.h>
31 #include <linux/etherdevice.h>
32 #include "mlx5_ib.h"
33 #include "ib_rep.h"
34 #include "cmd.h"
35 #include "devx.h"
36 #include "fs.h"
37 #include "srq.h"
38 #include "qp.h"
39 #include "wr.h"
40 #include "restrack.h"
41 #include "counters.h"
42 #include <linux/mlx5/accel.h>
43 #include <rdma/uverbs_std_types.h>
44 #include <rdma/mlx5_user_ioctl_verbs.h>
45 #include <rdma/mlx5_user_ioctl_cmds.h>
46 #include <rdma/ib_umem_odp.h>
47 
48 #define UVERBS_MODULE_NAME mlx5_ib
49 #include <rdma/uverbs_named_ioctl.h>
50 
51 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
52 MODULE_DESCRIPTION("Mellanox 5th generation network adapters (ConnectX series) IB driver");
53 MODULE_LICENSE("Dual BSD/GPL");
54 
55 struct mlx5_ib_event_work {
56 	struct work_struct	work;
57 	union {
58 		struct mlx5_ib_dev	      *dev;
59 		struct mlx5_ib_multiport_info *mpi;
60 	};
61 	bool			is_slave;
62 	unsigned int		event;
63 	void			*param;
64 };
65 
66 enum {
67 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
68 };
69 
70 static struct workqueue_struct *mlx5_ib_event_wq;
71 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
72 static LIST_HEAD(mlx5_ib_dev_list);
73 /*
74  * This mutex should be held when accessing either of the above lists
75  */
76 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
77 
78 /* We can't use an array for xlt_emergency_page because dma_map_single
79  * doesn't work on kernel modules memory
80  */
81 static unsigned long xlt_emergency_page;
82 static struct mutex xlt_emergency_page_mutex;
83 
84 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
85 {
86 	struct mlx5_ib_dev *dev;
87 
88 	mutex_lock(&mlx5_ib_multiport_mutex);
89 	dev = mpi->ibdev;
90 	mutex_unlock(&mlx5_ib_multiport_mutex);
91 	return dev;
92 }
93 
94 static enum rdma_link_layer
95 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
96 {
97 	switch (port_type_cap) {
98 	case MLX5_CAP_PORT_TYPE_IB:
99 		return IB_LINK_LAYER_INFINIBAND;
100 	case MLX5_CAP_PORT_TYPE_ETH:
101 		return IB_LINK_LAYER_ETHERNET;
102 	default:
103 		return IB_LINK_LAYER_UNSPECIFIED;
104 	}
105 }
106 
107 static enum rdma_link_layer
108 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
109 {
110 	struct mlx5_ib_dev *dev = to_mdev(device);
111 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
112 
113 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
114 }
115 
116 static int get_port_state(struct ib_device *ibdev,
117 			  u8 port_num,
118 			  enum ib_port_state *state)
119 {
120 	struct ib_port_attr attr;
121 	int ret;
122 
123 	memset(&attr, 0, sizeof(attr));
124 	ret = ibdev->ops.query_port(ibdev, port_num, &attr);
125 	if (!ret)
126 		*state = attr.state;
127 	return ret;
128 }
129 
130 static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev,
131 					   struct net_device *ndev,
132 					   u8 *port_num)
133 {
134 	struct mlx5_eswitch *esw = dev->mdev->priv.eswitch;
135 	struct net_device *rep_ndev;
136 	struct mlx5_ib_port *port;
137 	int i;
138 
139 	for (i = 0; i < dev->num_ports; i++) {
140 		port  = &dev->port[i];
141 		if (!port->rep)
142 			continue;
143 
144 		read_lock(&port->roce.netdev_lock);
145 		rep_ndev = mlx5_ib_get_rep_netdev(esw,
146 						  port->rep->vport);
147 		if (rep_ndev == ndev) {
148 			read_unlock(&port->roce.netdev_lock);
149 			*port_num = i + 1;
150 			return &port->roce;
151 		}
152 		read_unlock(&port->roce.netdev_lock);
153 	}
154 
155 	return NULL;
156 }
157 
158 static int mlx5_netdev_event(struct notifier_block *this,
159 			     unsigned long event, void *ptr)
160 {
161 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
162 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
163 	u8 port_num = roce->native_port_num;
164 	struct mlx5_core_dev *mdev;
165 	struct mlx5_ib_dev *ibdev;
166 
167 	ibdev = roce->dev;
168 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
169 	if (!mdev)
170 		return NOTIFY_DONE;
171 
172 	switch (event) {
173 	case NETDEV_REGISTER:
174 		/* Should already be registered during the load */
175 		if (ibdev->is_rep)
176 			break;
177 		write_lock(&roce->netdev_lock);
178 		if (ndev->dev.parent == mdev->device)
179 			roce->netdev = ndev;
180 		write_unlock(&roce->netdev_lock);
181 		break;
182 
183 	case NETDEV_UNREGISTER:
184 		/* In case of reps, ib device goes away before the netdevs */
185 		write_lock(&roce->netdev_lock);
186 		if (roce->netdev == ndev)
187 			roce->netdev = NULL;
188 		write_unlock(&roce->netdev_lock);
189 		break;
190 
191 	case NETDEV_CHANGE:
192 	case NETDEV_UP:
193 	case NETDEV_DOWN: {
194 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
195 		struct net_device *upper = NULL;
196 
197 		if (lag_ndev) {
198 			upper = netdev_master_upper_dev_get(lag_ndev);
199 			dev_put(lag_ndev);
200 		}
201 
202 		if (ibdev->is_rep)
203 			roce = mlx5_get_rep_roce(ibdev, ndev, &port_num);
204 		if (!roce)
205 			return NOTIFY_DONE;
206 		if ((upper == ndev || (!upper && ndev == roce->netdev))
207 		    && ibdev->ib_active) {
208 			struct ib_event ibev = { };
209 			enum ib_port_state port_state;
210 
211 			if (get_port_state(&ibdev->ib_dev, port_num,
212 					   &port_state))
213 				goto done;
214 
215 			if (roce->last_port_state == port_state)
216 				goto done;
217 
218 			roce->last_port_state = port_state;
219 			ibev.device = &ibdev->ib_dev;
220 			if (port_state == IB_PORT_DOWN)
221 				ibev.event = IB_EVENT_PORT_ERR;
222 			else if (port_state == IB_PORT_ACTIVE)
223 				ibev.event = IB_EVENT_PORT_ACTIVE;
224 			else
225 				goto done;
226 
227 			ibev.element.port_num = port_num;
228 			ib_dispatch_event(&ibev);
229 		}
230 		break;
231 	}
232 
233 	default:
234 		break;
235 	}
236 done:
237 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
238 	return NOTIFY_DONE;
239 }
240 
241 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
242 					     u8 port_num)
243 {
244 	struct mlx5_ib_dev *ibdev = to_mdev(device);
245 	struct net_device *ndev;
246 	struct mlx5_core_dev *mdev;
247 
248 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
249 	if (!mdev)
250 		return NULL;
251 
252 	ndev = mlx5_lag_get_roce_netdev(mdev);
253 	if (ndev)
254 		goto out;
255 
256 	/* Ensure ndev does not disappear before we invoke dev_hold()
257 	 */
258 	read_lock(&ibdev->port[port_num - 1].roce.netdev_lock);
259 	ndev = ibdev->port[port_num - 1].roce.netdev;
260 	if (ndev)
261 		dev_hold(ndev);
262 	read_unlock(&ibdev->port[port_num - 1].roce.netdev_lock);
263 
264 out:
265 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
266 	return ndev;
267 }
268 
269 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
270 						   u8 ib_port_num,
271 						   u8 *native_port_num)
272 {
273 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
274 							  ib_port_num);
275 	struct mlx5_core_dev *mdev = NULL;
276 	struct mlx5_ib_multiport_info *mpi;
277 	struct mlx5_ib_port *port;
278 
279 	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
280 	    ll != IB_LINK_LAYER_ETHERNET) {
281 		if (native_port_num)
282 			*native_port_num = ib_port_num;
283 		return ibdev->mdev;
284 	}
285 
286 	if (native_port_num)
287 		*native_port_num = 1;
288 
289 	port = &ibdev->port[ib_port_num - 1];
290 	spin_lock(&port->mp.mpi_lock);
291 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
292 	if (mpi && !mpi->unaffiliate) {
293 		mdev = mpi->mdev;
294 		/* If it's the master no need to refcount, it'll exist
295 		 * as long as the ib_dev exists.
296 		 */
297 		if (!mpi->is_master)
298 			mpi->mdev_refcnt++;
299 	}
300 	spin_unlock(&port->mp.mpi_lock);
301 
302 	return mdev;
303 }
304 
305 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u8 port_num)
306 {
307 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
308 							  port_num);
309 	struct mlx5_ib_multiport_info *mpi;
310 	struct mlx5_ib_port *port;
311 
312 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
313 		return;
314 
315 	port = &ibdev->port[port_num - 1];
316 
317 	spin_lock(&port->mp.mpi_lock);
318 	mpi = ibdev->port[port_num - 1].mp.mpi;
319 	if (mpi->is_master)
320 		goto out;
321 
322 	mpi->mdev_refcnt--;
323 	if (mpi->unaffiliate)
324 		complete(&mpi->unref_comp);
325 out:
326 	spin_unlock(&port->mp.mpi_lock);
327 }
328 
329 static int translate_eth_legacy_proto_oper(u32 eth_proto_oper, u8 *active_speed,
330 					   u8 *active_width)
331 {
332 	switch (eth_proto_oper) {
333 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
334 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
335 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
336 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
337 		*active_width = IB_WIDTH_1X;
338 		*active_speed = IB_SPEED_SDR;
339 		break;
340 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
341 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
342 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
343 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
344 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
345 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
346 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
347 		*active_width = IB_WIDTH_1X;
348 		*active_speed = IB_SPEED_QDR;
349 		break;
350 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
351 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
352 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
353 		*active_width = IB_WIDTH_1X;
354 		*active_speed = IB_SPEED_EDR;
355 		break;
356 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
357 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
358 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
359 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
360 		*active_width = IB_WIDTH_4X;
361 		*active_speed = IB_SPEED_QDR;
362 		break;
363 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
364 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
365 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
366 		*active_width = IB_WIDTH_1X;
367 		*active_speed = IB_SPEED_HDR;
368 		break;
369 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
370 		*active_width = IB_WIDTH_4X;
371 		*active_speed = IB_SPEED_FDR;
372 		break;
373 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
374 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
375 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
376 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
377 		*active_width = IB_WIDTH_4X;
378 		*active_speed = IB_SPEED_EDR;
379 		break;
380 	default:
381 		return -EINVAL;
382 	}
383 
384 	return 0;
385 }
386 
387 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u8 *active_speed,
388 					u8 *active_width)
389 {
390 	switch (eth_proto_oper) {
391 	case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
392 	case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
393 		*active_width = IB_WIDTH_1X;
394 		*active_speed = IB_SPEED_SDR;
395 		break;
396 	case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
397 		*active_width = IB_WIDTH_1X;
398 		*active_speed = IB_SPEED_DDR;
399 		break;
400 	case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
401 		*active_width = IB_WIDTH_1X;
402 		*active_speed = IB_SPEED_QDR;
403 		break;
404 	case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
405 		*active_width = IB_WIDTH_4X;
406 		*active_speed = IB_SPEED_QDR;
407 		break;
408 	case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
409 		*active_width = IB_WIDTH_1X;
410 		*active_speed = IB_SPEED_EDR;
411 		break;
412 	case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
413 		*active_width = IB_WIDTH_2X;
414 		*active_speed = IB_SPEED_EDR;
415 		break;
416 	case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
417 		*active_width = IB_WIDTH_1X;
418 		*active_speed = IB_SPEED_HDR;
419 		break;
420 	case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
421 		*active_width = IB_WIDTH_4X;
422 		*active_speed = IB_SPEED_EDR;
423 		break;
424 	case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
425 		*active_width = IB_WIDTH_2X;
426 		*active_speed = IB_SPEED_HDR;
427 		break;
428 	case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4):
429 		*active_width = IB_WIDTH_4X;
430 		*active_speed = IB_SPEED_HDR;
431 		break;
432 	default:
433 		return -EINVAL;
434 	}
435 
436 	return 0;
437 }
438 
439 static int translate_eth_proto_oper(u32 eth_proto_oper, u8 *active_speed,
440 				    u8 *active_width, bool ext)
441 {
442 	return ext ?
443 		translate_eth_ext_proto_oper(eth_proto_oper, active_speed,
444 					     active_width) :
445 		translate_eth_legacy_proto_oper(eth_proto_oper, active_speed,
446 						active_width);
447 }
448 
449 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
450 				struct ib_port_attr *props)
451 {
452 	struct mlx5_ib_dev *dev = to_mdev(device);
453 	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0};
454 	struct mlx5_core_dev *mdev;
455 	struct net_device *ndev, *upper;
456 	enum ib_mtu ndev_ib_mtu;
457 	bool put_mdev = true;
458 	u16 qkey_viol_cntr;
459 	u32 eth_prot_oper;
460 	u8 mdev_port_num;
461 	bool ext;
462 	int err;
463 
464 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
465 	if (!mdev) {
466 		/* This means the port isn't affiliated yet. Get the
467 		 * info for the master port instead.
468 		 */
469 		put_mdev = false;
470 		mdev = dev->mdev;
471 		mdev_port_num = 1;
472 		port_num = 1;
473 	}
474 
475 	/* Possible bad flows are checked before filling out props so in case
476 	 * of an error it will still be zeroed out.
477 	 * Use native port in case of reps
478 	 */
479 	if (dev->is_rep)
480 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
481 					   1);
482 	else
483 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
484 					   mdev_port_num);
485 	if (err)
486 		goto out;
487 	ext = !!MLX5_GET_ETH_PROTO(ptys_reg, out, true, eth_proto_capability);
488 	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
489 
490 	props->active_width     = IB_WIDTH_4X;
491 	props->active_speed     = IB_SPEED_QDR;
492 
493 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
494 				 &props->active_width, ext);
495 
496 	props->port_cap_flags |= IB_PORT_CM_SUP;
497 	props->ip_gids = true;
498 
499 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
500 						roce_address_table_size);
501 	props->max_mtu          = IB_MTU_4096;
502 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
503 	props->pkey_tbl_len     = 1;
504 	props->state            = IB_PORT_DOWN;
505 	props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
506 
507 	mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
508 	props->qkey_viol_cntr = qkey_viol_cntr;
509 
510 	/* If this is a stub query for an unaffiliated port stop here */
511 	if (!put_mdev)
512 		goto out;
513 
514 	ndev = mlx5_ib_get_netdev(device, port_num);
515 	if (!ndev)
516 		goto out;
517 
518 	if (dev->lag_active) {
519 		rcu_read_lock();
520 		upper = netdev_master_upper_dev_get_rcu(ndev);
521 		if (upper) {
522 			dev_put(ndev);
523 			ndev = upper;
524 			dev_hold(ndev);
525 		}
526 		rcu_read_unlock();
527 	}
528 
529 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
530 		props->state      = IB_PORT_ACTIVE;
531 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
532 	}
533 
534 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
535 
536 	dev_put(ndev);
537 
538 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
539 out:
540 	if (put_mdev)
541 		mlx5_ib_put_native_port_mdev(dev, port_num);
542 	return err;
543 }
544 
545 static int set_roce_addr(struct mlx5_ib_dev *dev, u8 port_num,
546 			 unsigned int index, const union ib_gid *gid,
547 			 const struct ib_gid_attr *attr)
548 {
549 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
550 	u16 vlan_id = 0xffff;
551 	u8 roce_version = 0;
552 	u8 roce_l3_type = 0;
553 	u8 mac[ETH_ALEN];
554 	int ret;
555 
556 	if (gid) {
557 		gid_type = attr->gid_type;
558 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
559 		if (ret)
560 			return ret;
561 	}
562 
563 	switch (gid_type) {
564 	case IB_GID_TYPE_IB:
565 		roce_version = MLX5_ROCE_VERSION_1;
566 		break;
567 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
568 		roce_version = MLX5_ROCE_VERSION_2;
569 		if (ipv6_addr_v4mapped((void *)gid))
570 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
571 		else
572 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
573 		break;
574 
575 	default:
576 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
577 	}
578 
579 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
580 				      roce_l3_type, gid->raw, mac,
581 				      vlan_id < VLAN_CFI_MASK, vlan_id,
582 				      port_num);
583 }
584 
585 static int mlx5_ib_add_gid(const struct ib_gid_attr *attr,
586 			   __always_unused void **context)
587 {
588 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
589 			     attr->index, &attr->gid, attr);
590 }
591 
592 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
593 			   __always_unused void **context)
594 {
595 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
596 			     attr->index, NULL, NULL);
597 }
598 
599 __be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev,
600 				   const struct ib_gid_attr *attr)
601 {
602 	if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
603 		return 0;
604 
605 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
606 }
607 
608 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
609 {
610 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
611 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
612 	return 0;
613 }
614 
615 enum {
616 	MLX5_VPORT_ACCESS_METHOD_MAD,
617 	MLX5_VPORT_ACCESS_METHOD_HCA,
618 	MLX5_VPORT_ACCESS_METHOD_NIC,
619 };
620 
621 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
622 {
623 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
624 		return MLX5_VPORT_ACCESS_METHOD_MAD;
625 
626 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
627 	    IB_LINK_LAYER_ETHERNET)
628 		return MLX5_VPORT_ACCESS_METHOD_NIC;
629 
630 	return MLX5_VPORT_ACCESS_METHOD_HCA;
631 }
632 
633 static void get_atomic_caps(struct mlx5_ib_dev *dev,
634 			    u8 atomic_size_qp,
635 			    struct ib_device_attr *props)
636 {
637 	u8 tmp;
638 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
639 	u8 atomic_req_8B_endianness_mode =
640 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
641 
642 	/* Check if HW supports 8 bytes standard atomic operations and capable
643 	 * of host endianness respond
644 	 */
645 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
646 	if (((atomic_operations & tmp) == tmp) &&
647 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
648 	    (atomic_req_8B_endianness_mode)) {
649 		props->atomic_cap = IB_ATOMIC_HCA;
650 	} else {
651 		props->atomic_cap = IB_ATOMIC_NONE;
652 	}
653 }
654 
655 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
656 			       struct ib_device_attr *props)
657 {
658 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
659 
660 	get_atomic_caps(dev, atomic_size_qp, props);
661 }
662 
663 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
664 					__be64 *sys_image_guid)
665 {
666 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
667 	struct mlx5_core_dev *mdev = dev->mdev;
668 	u64 tmp;
669 	int err;
670 
671 	switch (mlx5_get_vport_access_method(ibdev)) {
672 	case MLX5_VPORT_ACCESS_METHOD_MAD:
673 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
674 							    sys_image_guid);
675 
676 	case MLX5_VPORT_ACCESS_METHOD_HCA:
677 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
678 		break;
679 
680 	case MLX5_VPORT_ACCESS_METHOD_NIC:
681 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
682 		break;
683 
684 	default:
685 		return -EINVAL;
686 	}
687 
688 	if (!err)
689 		*sys_image_guid = cpu_to_be64(tmp);
690 
691 	return err;
692 
693 }
694 
695 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
696 				u16 *max_pkeys)
697 {
698 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
699 	struct mlx5_core_dev *mdev = dev->mdev;
700 
701 	switch (mlx5_get_vport_access_method(ibdev)) {
702 	case MLX5_VPORT_ACCESS_METHOD_MAD:
703 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
704 
705 	case MLX5_VPORT_ACCESS_METHOD_HCA:
706 	case MLX5_VPORT_ACCESS_METHOD_NIC:
707 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
708 						pkey_table_size));
709 		return 0;
710 
711 	default:
712 		return -EINVAL;
713 	}
714 }
715 
716 static int mlx5_query_vendor_id(struct ib_device *ibdev,
717 				u32 *vendor_id)
718 {
719 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
720 
721 	switch (mlx5_get_vport_access_method(ibdev)) {
722 	case MLX5_VPORT_ACCESS_METHOD_MAD:
723 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
724 
725 	case MLX5_VPORT_ACCESS_METHOD_HCA:
726 	case MLX5_VPORT_ACCESS_METHOD_NIC:
727 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
728 
729 	default:
730 		return -EINVAL;
731 	}
732 }
733 
734 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
735 				__be64 *node_guid)
736 {
737 	u64 tmp;
738 	int err;
739 
740 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
741 	case MLX5_VPORT_ACCESS_METHOD_MAD:
742 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
743 
744 	case MLX5_VPORT_ACCESS_METHOD_HCA:
745 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
746 		break;
747 
748 	case MLX5_VPORT_ACCESS_METHOD_NIC:
749 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
750 		break;
751 
752 	default:
753 		return -EINVAL;
754 	}
755 
756 	if (!err)
757 		*node_guid = cpu_to_be64(tmp);
758 
759 	return err;
760 }
761 
762 struct mlx5_reg_node_desc {
763 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
764 };
765 
766 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
767 {
768 	struct mlx5_reg_node_desc in;
769 
770 	if (mlx5_use_mad_ifc(dev))
771 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
772 
773 	memset(&in, 0, sizeof(in));
774 
775 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
776 				    sizeof(struct mlx5_reg_node_desc),
777 				    MLX5_REG_NODE_DESC, 0, 0);
778 }
779 
780 static int mlx5_ib_query_device(struct ib_device *ibdev,
781 				struct ib_device_attr *props,
782 				struct ib_udata *uhw)
783 {
784 	size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
785 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
786 	struct mlx5_core_dev *mdev = dev->mdev;
787 	int err = -ENOMEM;
788 	int max_sq_desc;
789 	int max_rq_sg;
790 	int max_sq_sg;
791 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
792 	bool raw_support = !mlx5_core_mp_enabled(mdev);
793 	struct mlx5_ib_query_device_resp resp = {};
794 	size_t resp_len;
795 	u64 max_tso;
796 
797 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
798 	if (uhw_outlen && uhw_outlen < resp_len)
799 		return -EINVAL;
800 
801 	resp.response_length = resp_len;
802 
803 	if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
804 		return -EINVAL;
805 
806 	memset(props, 0, sizeof(*props));
807 	err = mlx5_query_system_image_guid(ibdev,
808 					   &props->sys_image_guid);
809 	if (err)
810 		return err;
811 
812 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
813 	if (err)
814 		return err;
815 
816 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
817 	if (err)
818 		return err;
819 
820 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
821 		(fw_rev_min(dev->mdev) << 16) |
822 		fw_rev_sub(dev->mdev);
823 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
824 		IB_DEVICE_PORT_ACTIVE_EVENT		|
825 		IB_DEVICE_SYS_IMAGE_GUID		|
826 		IB_DEVICE_RC_RNR_NAK_GEN;
827 
828 	if (MLX5_CAP_GEN(mdev, pkv))
829 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
830 	if (MLX5_CAP_GEN(mdev, qkv))
831 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
832 	if (MLX5_CAP_GEN(mdev, apm))
833 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
834 	if (MLX5_CAP_GEN(mdev, xrc))
835 		props->device_cap_flags |= IB_DEVICE_XRC;
836 	if (MLX5_CAP_GEN(mdev, imaicl)) {
837 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
838 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
839 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
840 		/* We support 'Gappy' memory registration too */
841 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
842 	}
843 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
844 	if (MLX5_CAP_GEN(mdev, sho)) {
845 		props->device_cap_flags |= IB_DEVICE_INTEGRITY_HANDOVER;
846 		/* At this stage no support for signature handover */
847 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
848 				      IB_PROT_T10DIF_TYPE_2 |
849 				      IB_PROT_T10DIF_TYPE_3;
850 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
851 				       IB_GUARD_T10DIF_CSUM;
852 	}
853 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
854 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
855 
856 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
857 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
858 			/* Legacy bit to support old userspace libraries */
859 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
860 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
861 		}
862 
863 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
864 			props->raw_packet_caps |=
865 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
866 
867 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) {
868 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
869 			if (max_tso) {
870 				resp.tso_caps.max_tso = 1 << max_tso;
871 				resp.tso_caps.supported_qpts |=
872 					1 << IB_QPT_RAW_PACKET;
873 				resp.response_length += sizeof(resp.tso_caps);
874 			}
875 		}
876 
877 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) {
878 			resp.rss_caps.rx_hash_function =
879 						MLX5_RX_HASH_FUNC_TOEPLITZ;
880 			resp.rss_caps.rx_hash_fields_mask =
881 						MLX5_RX_HASH_SRC_IPV4 |
882 						MLX5_RX_HASH_DST_IPV4 |
883 						MLX5_RX_HASH_SRC_IPV6 |
884 						MLX5_RX_HASH_DST_IPV6 |
885 						MLX5_RX_HASH_SRC_PORT_TCP |
886 						MLX5_RX_HASH_DST_PORT_TCP |
887 						MLX5_RX_HASH_SRC_PORT_UDP |
888 						MLX5_RX_HASH_DST_PORT_UDP |
889 						MLX5_RX_HASH_INNER;
890 			if (mlx5_accel_ipsec_device_caps(dev->mdev) &
891 			    MLX5_ACCEL_IPSEC_CAP_DEVICE)
892 				resp.rss_caps.rx_hash_fields_mask |=
893 					MLX5_RX_HASH_IPSEC_SPI;
894 			resp.response_length += sizeof(resp.rss_caps);
895 		}
896 	} else {
897 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen)
898 			resp.response_length += sizeof(resp.tso_caps);
899 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen)
900 			resp.response_length += sizeof(resp.rss_caps);
901 	}
902 
903 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
904 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
905 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
906 	}
907 
908 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
909 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
910 	    raw_support)
911 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
912 
913 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
914 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
915 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
916 
917 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
918 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
919 	    raw_support) {
920 		/* Legacy bit to support old userspace libraries */
921 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
922 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
923 	}
924 
925 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
926 		props->max_dm_size =
927 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
928 	}
929 
930 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
931 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
932 
933 	if (MLX5_CAP_GEN(mdev, end_pad))
934 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
935 
936 	props->vendor_part_id	   = mdev->pdev->device;
937 	props->hw_ver		   = mdev->pdev->revision;
938 
939 	props->max_mr_size	   = ~0ull;
940 	props->page_size_cap	   = ~(min_page_size - 1);
941 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
942 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
943 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
944 		     sizeof(struct mlx5_wqe_data_seg);
945 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
946 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
947 		     sizeof(struct mlx5_wqe_raddr_seg)) /
948 		sizeof(struct mlx5_wqe_data_seg);
949 	props->max_send_sge = max_sq_sg;
950 	props->max_recv_sge = max_rq_sg;
951 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
952 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
953 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
954 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
955 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
956 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
957 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
958 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
959 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
960 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
961 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
962 	props->max_srq_sge	   = max_rq_sg - 1;
963 	props->max_fast_reg_page_list_len =
964 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
965 	props->max_pi_fast_reg_page_list_len =
966 		props->max_fast_reg_page_list_len / 2;
967 	props->max_sgl_rd =
968 		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
969 	get_atomic_caps_qp(dev, props);
970 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
971 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
972 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
973 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
974 					   props->max_mcast_grp;
975 	props->max_ah = INT_MAX;
976 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
977 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
978 
979 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
980 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
981 			props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
982 		props->odp_caps = dev->odp_caps;
983 		if (!uhw) {
984 			/* ODP for kernel QPs is not implemented for receive
985 			 * WQEs and SRQ WQEs
986 			 */
987 			props->odp_caps.per_transport_caps.rc_odp_caps &=
988 				~(IB_ODP_SUPPORT_READ |
989 				  IB_ODP_SUPPORT_SRQ_RECV);
990 			props->odp_caps.per_transport_caps.uc_odp_caps &=
991 				~(IB_ODP_SUPPORT_READ |
992 				  IB_ODP_SUPPORT_SRQ_RECV);
993 			props->odp_caps.per_transport_caps.ud_odp_caps &=
994 				~(IB_ODP_SUPPORT_READ |
995 				  IB_ODP_SUPPORT_SRQ_RECV);
996 			props->odp_caps.per_transport_caps.xrc_odp_caps &=
997 				~(IB_ODP_SUPPORT_READ |
998 				  IB_ODP_SUPPORT_SRQ_RECV);
999 		}
1000 	}
1001 
1002 	if (MLX5_CAP_GEN(mdev, cd))
1003 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
1004 
1005 	if (mlx5_core_is_vf(mdev))
1006 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
1007 
1008 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1009 	    IB_LINK_LAYER_ETHERNET && raw_support) {
1010 		props->rss_caps.max_rwq_indirection_tables =
1011 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1012 		props->rss_caps.max_rwq_indirection_table_size =
1013 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1014 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1015 		props->max_wq_type_rq =
1016 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1017 	}
1018 
1019 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1020 		props->tm_caps.max_num_tags =
1021 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1022 		props->tm_caps.max_ops =
1023 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1024 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1025 	}
1026 
1027 	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1028 	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1029 		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1030 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1031 	}
1032 
1033 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1034 		props->cq_caps.max_cq_moderation_count =
1035 						MLX5_MAX_CQ_COUNT;
1036 		props->cq_caps.max_cq_moderation_period =
1037 						MLX5_MAX_CQ_PERIOD;
1038 	}
1039 
1040 	if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) {
1041 		resp.response_length += sizeof(resp.cqe_comp_caps);
1042 
1043 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1044 			resp.cqe_comp_caps.max_num =
1045 				MLX5_CAP_GEN(dev->mdev,
1046 					     cqe_compression_max_num);
1047 
1048 			resp.cqe_comp_caps.supported_format =
1049 				MLX5_IB_CQE_RES_FORMAT_HASH |
1050 				MLX5_IB_CQE_RES_FORMAT_CSUM;
1051 
1052 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1053 				resp.cqe_comp_caps.supported_format |=
1054 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1055 		}
1056 	}
1057 
1058 	if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen &&
1059 	    raw_support) {
1060 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1061 		    MLX5_CAP_GEN(mdev, qos)) {
1062 			resp.packet_pacing_caps.qp_rate_limit_max =
1063 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1064 			resp.packet_pacing_caps.qp_rate_limit_min =
1065 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1066 			resp.packet_pacing_caps.supported_qpts |=
1067 				1 << IB_QPT_RAW_PACKET;
1068 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1069 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1070 				resp.packet_pacing_caps.cap_flags |=
1071 					MLX5_IB_PP_SUPPORT_BURST;
1072 		}
1073 		resp.response_length += sizeof(resp.packet_pacing_caps);
1074 	}
1075 
1076 	if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <=
1077 	    uhw_outlen) {
1078 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1079 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1080 				MLX5_IB_ALLOW_MPW;
1081 
1082 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1083 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1084 				MLX5_IB_SUPPORT_EMPW;
1085 
1086 		resp.response_length +=
1087 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1088 	}
1089 
1090 	if (offsetofend(typeof(resp), flags) <= uhw_outlen) {
1091 		resp.response_length += sizeof(resp.flags);
1092 
1093 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1094 			resp.flags |=
1095 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1096 
1097 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1098 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1099 		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1100 			resp.flags |=
1101 				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1102 
1103 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1104 	}
1105 
1106 	if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) {
1107 		resp.response_length += sizeof(resp.sw_parsing_caps);
1108 		if (MLX5_CAP_ETH(mdev, swp)) {
1109 			resp.sw_parsing_caps.sw_parsing_offloads |=
1110 				MLX5_IB_SW_PARSING;
1111 
1112 			if (MLX5_CAP_ETH(mdev, swp_csum))
1113 				resp.sw_parsing_caps.sw_parsing_offloads |=
1114 					MLX5_IB_SW_PARSING_CSUM;
1115 
1116 			if (MLX5_CAP_ETH(mdev, swp_lso))
1117 				resp.sw_parsing_caps.sw_parsing_offloads |=
1118 					MLX5_IB_SW_PARSING_LSO;
1119 
1120 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1121 				resp.sw_parsing_caps.supported_qpts =
1122 					BIT(IB_QPT_RAW_PACKET);
1123 		}
1124 	}
1125 
1126 	if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen &&
1127 	    raw_support) {
1128 		resp.response_length += sizeof(resp.striding_rq_caps);
1129 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1130 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1131 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1132 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1133 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1134 			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1135 				resp.striding_rq_caps
1136 					.min_single_wqe_log_num_of_strides =
1137 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1138 			else
1139 				resp.striding_rq_caps
1140 					.min_single_wqe_log_num_of_strides =
1141 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1142 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1143 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1144 			resp.striding_rq_caps.supported_qpts =
1145 				BIT(IB_QPT_RAW_PACKET);
1146 		}
1147 	}
1148 
1149 	if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) {
1150 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1151 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1152 			resp.tunnel_offloads_caps |=
1153 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1154 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1155 			resp.tunnel_offloads_caps |=
1156 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1157 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1158 			resp.tunnel_offloads_caps |=
1159 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1160 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre))
1161 			resp.tunnel_offloads_caps |=
1162 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1163 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp))
1164 			resp.tunnel_offloads_caps |=
1165 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1166 	}
1167 
1168 	if (uhw_outlen) {
1169 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1170 
1171 		if (err)
1172 			return err;
1173 	}
1174 
1175 	return 0;
1176 }
1177 
1178 enum mlx5_ib_width {
1179 	MLX5_IB_WIDTH_1X	= 1 << 0,
1180 	MLX5_IB_WIDTH_2X	= 1 << 1,
1181 	MLX5_IB_WIDTH_4X	= 1 << 2,
1182 	MLX5_IB_WIDTH_8X	= 1 << 3,
1183 	MLX5_IB_WIDTH_12X	= 1 << 4
1184 };
1185 
1186 static void translate_active_width(struct ib_device *ibdev, u8 active_width,
1187 				  u8 *ib_width)
1188 {
1189 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1190 
1191 	if (active_width & MLX5_IB_WIDTH_1X)
1192 		*ib_width = IB_WIDTH_1X;
1193 	else if (active_width & MLX5_IB_WIDTH_2X)
1194 		*ib_width = IB_WIDTH_2X;
1195 	else if (active_width & MLX5_IB_WIDTH_4X)
1196 		*ib_width = IB_WIDTH_4X;
1197 	else if (active_width & MLX5_IB_WIDTH_8X)
1198 		*ib_width = IB_WIDTH_8X;
1199 	else if (active_width & MLX5_IB_WIDTH_12X)
1200 		*ib_width = IB_WIDTH_12X;
1201 	else {
1202 		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1203 			    (int)active_width);
1204 		*ib_width = IB_WIDTH_4X;
1205 	}
1206 
1207 	return;
1208 }
1209 
1210 static int mlx5_mtu_to_ib_mtu(int mtu)
1211 {
1212 	switch (mtu) {
1213 	case 256: return 1;
1214 	case 512: return 2;
1215 	case 1024: return 3;
1216 	case 2048: return 4;
1217 	case 4096: return 5;
1218 	default:
1219 		pr_warn("invalid mtu\n");
1220 		return -1;
1221 	}
1222 }
1223 
1224 enum ib_max_vl_num {
1225 	__IB_MAX_VL_0		= 1,
1226 	__IB_MAX_VL_0_1		= 2,
1227 	__IB_MAX_VL_0_3		= 3,
1228 	__IB_MAX_VL_0_7		= 4,
1229 	__IB_MAX_VL_0_14	= 5,
1230 };
1231 
1232 enum mlx5_vl_hw_cap {
1233 	MLX5_VL_HW_0	= 1,
1234 	MLX5_VL_HW_0_1	= 2,
1235 	MLX5_VL_HW_0_2	= 3,
1236 	MLX5_VL_HW_0_3	= 4,
1237 	MLX5_VL_HW_0_4	= 5,
1238 	MLX5_VL_HW_0_5	= 6,
1239 	MLX5_VL_HW_0_6	= 7,
1240 	MLX5_VL_HW_0_7	= 8,
1241 	MLX5_VL_HW_0_14	= 15
1242 };
1243 
1244 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1245 				u8 *max_vl_num)
1246 {
1247 	switch (vl_hw_cap) {
1248 	case MLX5_VL_HW_0:
1249 		*max_vl_num = __IB_MAX_VL_0;
1250 		break;
1251 	case MLX5_VL_HW_0_1:
1252 		*max_vl_num = __IB_MAX_VL_0_1;
1253 		break;
1254 	case MLX5_VL_HW_0_3:
1255 		*max_vl_num = __IB_MAX_VL_0_3;
1256 		break;
1257 	case MLX5_VL_HW_0_7:
1258 		*max_vl_num = __IB_MAX_VL_0_7;
1259 		break;
1260 	case MLX5_VL_HW_0_14:
1261 		*max_vl_num = __IB_MAX_VL_0_14;
1262 		break;
1263 
1264 	default:
1265 		return -EINVAL;
1266 	}
1267 
1268 	return 0;
1269 }
1270 
1271 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
1272 			       struct ib_port_attr *props)
1273 {
1274 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1275 	struct mlx5_core_dev *mdev = dev->mdev;
1276 	struct mlx5_hca_vport_context *rep;
1277 	u16 max_mtu;
1278 	u16 oper_mtu;
1279 	int err;
1280 	u8 ib_link_width_oper;
1281 	u8 vl_hw_cap;
1282 
1283 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1284 	if (!rep) {
1285 		err = -ENOMEM;
1286 		goto out;
1287 	}
1288 
1289 	/* props being zeroed by the caller, avoid zeroing it here */
1290 
1291 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1292 	if (err)
1293 		goto out;
1294 
1295 	props->lid		= rep->lid;
1296 	props->lmc		= rep->lmc;
1297 	props->sm_lid		= rep->sm_lid;
1298 	props->sm_sl		= rep->sm_sl;
1299 	props->state		= rep->vport_state;
1300 	props->phys_state	= rep->port_physical_state;
1301 	props->port_cap_flags	= rep->cap_mask1;
1302 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1303 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1304 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1305 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1306 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1307 	props->subnet_timeout	= rep->subnet_timeout;
1308 	props->init_type_reply	= rep->init_type_reply;
1309 
1310 	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1311 		props->port_cap_flags2 = rep->cap_mask2;
1312 
1313 	err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
1314 	if (err)
1315 		goto out;
1316 
1317 	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1318 
1319 	err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
1320 	if (err)
1321 		goto out;
1322 
1323 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1324 
1325 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1326 
1327 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1328 
1329 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1330 
1331 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1332 	if (err)
1333 		goto out;
1334 
1335 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1336 				   &props->max_vl_num);
1337 out:
1338 	kfree(rep);
1339 	return err;
1340 }
1341 
1342 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
1343 		       struct ib_port_attr *props)
1344 {
1345 	unsigned int count;
1346 	int ret;
1347 
1348 	switch (mlx5_get_vport_access_method(ibdev)) {
1349 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1350 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1351 		break;
1352 
1353 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1354 		ret = mlx5_query_hca_port(ibdev, port, props);
1355 		break;
1356 
1357 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1358 		ret = mlx5_query_port_roce(ibdev, port, props);
1359 		break;
1360 
1361 	default:
1362 		ret = -EINVAL;
1363 	}
1364 
1365 	if (!ret && props) {
1366 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1367 		struct mlx5_core_dev *mdev;
1368 		bool put_mdev = true;
1369 
1370 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1371 		if (!mdev) {
1372 			/* If the port isn't affiliated yet query the master.
1373 			 * The master and slave will have the same values.
1374 			 */
1375 			mdev = dev->mdev;
1376 			port = 1;
1377 			put_mdev = false;
1378 		}
1379 		count = mlx5_core_reserved_gids_count(mdev);
1380 		if (put_mdev)
1381 			mlx5_ib_put_native_port_mdev(dev, port);
1382 		props->gid_tbl_len -= count;
1383 	}
1384 	return ret;
1385 }
1386 
1387 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u8 port,
1388 				  struct ib_port_attr *props)
1389 {
1390 	int ret;
1391 
1392 	/* Only link layer == ethernet is valid for representors
1393 	 * and we always use port 1
1394 	 */
1395 	ret = mlx5_query_port_roce(ibdev, port, props);
1396 	if (ret || !props)
1397 		return ret;
1398 
1399 	/* We don't support GIDS */
1400 	props->gid_tbl_len = 0;
1401 
1402 	return ret;
1403 }
1404 
1405 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1406 			     union ib_gid *gid)
1407 {
1408 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1409 	struct mlx5_core_dev *mdev = dev->mdev;
1410 
1411 	switch (mlx5_get_vport_access_method(ibdev)) {
1412 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1413 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1414 
1415 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1416 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1417 
1418 	default:
1419 		return -EINVAL;
1420 	}
1421 
1422 }
1423 
1424 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u8 port,
1425 				   u16 index, u16 *pkey)
1426 {
1427 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1428 	struct mlx5_core_dev *mdev;
1429 	bool put_mdev = true;
1430 	u8 mdev_port_num;
1431 	int err;
1432 
1433 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1434 	if (!mdev) {
1435 		/* The port isn't affiliated yet, get the PKey from the master
1436 		 * port. For RoCE the PKey tables will be the same.
1437 		 */
1438 		put_mdev = false;
1439 		mdev = dev->mdev;
1440 		mdev_port_num = 1;
1441 	}
1442 
1443 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1444 					index, pkey);
1445 	if (put_mdev)
1446 		mlx5_ib_put_native_port_mdev(dev, port);
1447 
1448 	return err;
1449 }
1450 
1451 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1452 			      u16 *pkey)
1453 {
1454 	switch (mlx5_get_vport_access_method(ibdev)) {
1455 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1456 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1457 
1458 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1459 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1460 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1461 	default:
1462 		return -EINVAL;
1463 	}
1464 }
1465 
1466 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1467 				 struct ib_device_modify *props)
1468 {
1469 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1470 	struct mlx5_reg_node_desc in;
1471 	struct mlx5_reg_node_desc out;
1472 	int err;
1473 
1474 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1475 		return -EOPNOTSUPP;
1476 
1477 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1478 		return 0;
1479 
1480 	/*
1481 	 * If possible, pass node desc to FW, so it can generate
1482 	 * a 144 trap.  If cmd fails, just ignore.
1483 	 */
1484 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1485 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1486 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1487 	if (err)
1488 		return err;
1489 
1490 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1491 
1492 	return err;
1493 }
1494 
1495 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1496 				u32 value)
1497 {
1498 	struct mlx5_hca_vport_context ctx = {};
1499 	struct mlx5_core_dev *mdev;
1500 	u8 mdev_port_num;
1501 	int err;
1502 
1503 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1504 	if (!mdev)
1505 		return -ENODEV;
1506 
1507 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1508 	if (err)
1509 		goto out;
1510 
1511 	if (~ctx.cap_mask1_perm & mask) {
1512 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1513 			     mask, ctx.cap_mask1_perm);
1514 		err = -EINVAL;
1515 		goto out;
1516 	}
1517 
1518 	ctx.cap_mask1 = value;
1519 	ctx.cap_mask1_perm = mask;
1520 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1521 						 0, &ctx);
1522 
1523 out:
1524 	mlx5_ib_put_native_port_mdev(dev, port_num);
1525 
1526 	return err;
1527 }
1528 
1529 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1530 			       struct ib_port_modify *props)
1531 {
1532 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1533 	struct ib_port_attr attr;
1534 	u32 tmp;
1535 	int err;
1536 	u32 change_mask;
1537 	u32 value;
1538 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1539 		      IB_LINK_LAYER_INFINIBAND);
1540 
1541 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1542 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1543 	 */
1544 	if (!is_ib)
1545 		return 0;
1546 
1547 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1548 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1549 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1550 		return set_port_caps_atomic(dev, port, change_mask, value);
1551 	}
1552 
1553 	mutex_lock(&dev->cap_mask_mutex);
1554 
1555 	err = ib_query_port(ibdev, port, &attr);
1556 	if (err)
1557 		goto out;
1558 
1559 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1560 		~props->clr_port_cap_mask;
1561 
1562 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1563 
1564 out:
1565 	mutex_unlock(&dev->cap_mask_mutex);
1566 	return err;
1567 }
1568 
1569 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1570 {
1571 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1572 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1573 }
1574 
1575 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1576 {
1577 	/* Large page with non 4k uar support might limit the dynamic size */
1578 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1579 		return MLX5_MIN_DYN_BFREGS;
1580 
1581 	return MLX5_MAX_DYN_BFREGS;
1582 }
1583 
1584 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1585 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1586 			     struct mlx5_bfreg_info *bfregi)
1587 {
1588 	int uars_per_sys_page;
1589 	int bfregs_per_sys_page;
1590 	int ref_bfregs = req->total_num_bfregs;
1591 
1592 	if (req->total_num_bfregs == 0)
1593 		return -EINVAL;
1594 
1595 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1596 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1597 
1598 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1599 		return -ENOMEM;
1600 
1601 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1602 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1603 	/* This holds the required static allocation asked by the user */
1604 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1605 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1606 		return -EINVAL;
1607 
1608 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1609 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1610 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1611 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1612 
1613 	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",
1614 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1615 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1616 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1617 		    bfregi->num_sys_pages);
1618 
1619 	return 0;
1620 }
1621 
1622 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1623 {
1624 	struct mlx5_bfreg_info *bfregi;
1625 	int err;
1626 	int i;
1627 
1628 	bfregi = &context->bfregi;
1629 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1630 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1631 		if (err)
1632 			goto error;
1633 
1634 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1635 	}
1636 
1637 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1638 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1639 
1640 	return 0;
1641 
1642 error:
1643 	for (--i; i >= 0; i--)
1644 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1645 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1646 
1647 	return err;
1648 }
1649 
1650 static void deallocate_uars(struct mlx5_ib_dev *dev,
1651 			    struct mlx5_ib_ucontext *context)
1652 {
1653 	struct mlx5_bfreg_info *bfregi;
1654 	int i;
1655 
1656 	bfregi = &context->bfregi;
1657 	for (i = 0; i < bfregi->num_sys_pages; i++)
1658 		if (i < bfregi->num_static_sys_pages ||
1659 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1660 			mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1661 }
1662 
1663 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1664 {
1665 	int err = 0;
1666 
1667 	mutex_lock(&dev->lb.mutex);
1668 	if (td)
1669 		dev->lb.user_td++;
1670 	if (qp)
1671 		dev->lb.qps++;
1672 
1673 	if (dev->lb.user_td == 2 ||
1674 	    dev->lb.qps == 1) {
1675 		if (!dev->lb.enabled) {
1676 			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1677 			dev->lb.enabled = true;
1678 		}
1679 	}
1680 
1681 	mutex_unlock(&dev->lb.mutex);
1682 
1683 	return err;
1684 }
1685 
1686 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1687 {
1688 	mutex_lock(&dev->lb.mutex);
1689 	if (td)
1690 		dev->lb.user_td--;
1691 	if (qp)
1692 		dev->lb.qps--;
1693 
1694 	if (dev->lb.user_td == 1 &&
1695 	    dev->lb.qps == 0) {
1696 		if (dev->lb.enabled) {
1697 			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1698 			dev->lb.enabled = false;
1699 		}
1700 	}
1701 
1702 	mutex_unlock(&dev->lb.mutex);
1703 }
1704 
1705 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1706 					  u16 uid)
1707 {
1708 	int err;
1709 
1710 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1711 		return 0;
1712 
1713 	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1714 	if (err)
1715 		return err;
1716 
1717 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1718 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1719 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1720 		return err;
1721 
1722 	return mlx5_ib_enable_lb(dev, true, false);
1723 }
1724 
1725 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1726 					     u16 uid)
1727 {
1728 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1729 		return;
1730 
1731 	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1732 
1733 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1734 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1735 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1736 		return;
1737 
1738 	mlx5_ib_disable_lb(dev, true, false);
1739 }
1740 
1741 static int set_ucontext_resp(struct ib_ucontext *uctx,
1742 			     struct mlx5_ib_alloc_ucontext_resp *resp)
1743 {
1744 	struct ib_device *ibdev = uctx->device;
1745 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1746 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1747 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1748 	int err;
1749 
1750 	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1751 		err = mlx5_cmd_dump_fill_mkey(dev->mdev,
1752 					      &resp->dump_fill_mkey);
1753 		if (err)
1754 			return err;
1755 		resp->comp_mask |=
1756 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1757 	}
1758 
1759 	resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1760 	if (dev->wc_support)
1761 		resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev,
1762 						      log_bf_reg_size);
1763 	resp->cache_line_size = cache_line_size();
1764 	resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1765 	resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1766 	resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1767 	resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1768 	resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1769 	resp->cqe_version = context->cqe_version;
1770 	resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1771 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1772 	resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1773 					MLX5_CAP_GEN(dev->mdev,
1774 						     num_of_uars_per_page) : 1;
1775 
1776 	if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1777 				MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1778 		if (mlx5_get_flow_namespace(dev->mdev,
1779 				MLX5_FLOW_NAMESPACE_EGRESS))
1780 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1781 		if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1782 				MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1783 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1784 		if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1785 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1786 		if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1787 				MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1788 			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1789 		/* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1790 	}
1791 
1792 	resp->tot_bfregs = bfregi->lib_uar_dyn ? 0 :
1793 			bfregi->total_num_bfregs - bfregi->num_dyn_bfregs;
1794 	resp->num_ports = dev->num_ports;
1795 	resp->cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1796 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1797 
1798 	if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1799 		mlx5_query_min_inline(dev->mdev, &resp->eth_min_inline);
1800 		resp->eth_min_inline++;
1801 	}
1802 
1803 	if (dev->mdev->clock_info)
1804 		resp->clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1805 
1806 	/*
1807 	 * We don't want to expose information from the PCI bar that is located
1808 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1809 	 * pretend we don't support reading the HCA's core clock. This is also
1810 	 * forced by mmap function.
1811 	 */
1812 	if (PAGE_SIZE <= 4096) {
1813 		resp->comp_mask |=
1814 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1815 		resp->hca_core_clock_offset =
1816 			offsetof(struct mlx5_init_seg,
1817 				 internal_timer_h) % PAGE_SIZE;
1818 	}
1819 
1820 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
1821 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE;
1822 
1823 	resp->num_dyn_bfregs = bfregi->num_dyn_bfregs;
1824 	return 0;
1825 }
1826 
1827 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1828 				  struct ib_udata *udata)
1829 {
1830 	struct ib_device *ibdev = uctx->device;
1831 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1832 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1833 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1834 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1835 	struct mlx5_bfreg_info *bfregi;
1836 	int ver;
1837 	int err;
1838 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1839 				     max_cqe_version);
1840 	bool lib_uar_4k;
1841 	bool lib_uar_dyn;
1842 
1843 	if (!dev->ib_active)
1844 		return -EAGAIN;
1845 
1846 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1847 		ver = 0;
1848 	else if (udata->inlen >= min_req_v2)
1849 		ver = 2;
1850 	else
1851 		return -EINVAL;
1852 
1853 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1854 	if (err)
1855 		return err;
1856 
1857 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1858 		return -EOPNOTSUPP;
1859 
1860 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1861 		return -EOPNOTSUPP;
1862 
1863 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1864 				    MLX5_NON_FP_BFREGS_PER_UAR);
1865 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1866 		return -EINVAL;
1867 
1868 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1869 	lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
1870 	bfregi = &context->bfregi;
1871 
1872 	if (lib_uar_dyn) {
1873 		bfregi->lib_uar_dyn = lib_uar_dyn;
1874 		goto uar_done;
1875 	}
1876 
1877 	/* updates req->total_num_bfregs */
1878 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1879 	if (err)
1880 		goto out_ctx;
1881 
1882 	mutex_init(&bfregi->lock);
1883 	bfregi->lib_uar_4k = lib_uar_4k;
1884 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1885 				GFP_KERNEL);
1886 	if (!bfregi->count) {
1887 		err = -ENOMEM;
1888 		goto out_ctx;
1889 	}
1890 
1891 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1892 				    sizeof(*bfregi->sys_pages),
1893 				    GFP_KERNEL);
1894 	if (!bfregi->sys_pages) {
1895 		err = -ENOMEM;
1896 		goto out_count;
1897 	}
1898 
1899 	err = allocate_uars(dev, context);
1900 	if (err)
1901 		goto out_sys_pages;
1902 
1903 uar_done:
1904 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1905 		err = mlx5_ib_devx_create(dev, true);
1906 		if (err < 0)
1907 			goto out_uars;
1908 		context->devx_uid = err;
1909 	}
1910 
1911 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1912 					     context->devx_uid);
1913 	if (err)
1914 		goto out_devx;
1915 
1916 	INIT_LIST_HEAD(&context->db_page_list);
1917 	mutex_init(&context->db_page_mutex);
1918 
1919 	context->cqe_version = min_t(__u8,
1920 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1921 				 req.max_cqe_version);
1922 
1923 	err = set_ucontext_resp(uctx, &resp);
1924 	if (err)
1925 		goto out_mdev;
1926 
1927 	resp.response_length = min(udata->outlen, sizeof(resp));
1928 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1929 	if (err)
1930 		goto out_mdev;
1931 
1932 	bfregi->ver = ver;
1933 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1934 	context->lib_caps = req.lib_caps;
1935 	print_lib_caps(dev, context->lib_caps);
1936 
1937 	if (mlx5_ib_lag_should_assign_affinity(dev)) {
1938 		u8 port = mlx5_core_native_port_num(dev->mdev) - 1;
1939 
1940 		atomic_set(&context->tx_port_affinity,
1941 			   atomic_add_return(
1942 				   1, &dev->port[port].roce.tx_port_affinity));
1943 	}
1944 
1945 	return 0;
1946 
1947 out_mdev:
1948 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1949 out_devx:
1950 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1951 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1952 
1953 out_uars:
1954 	deallocate_uars(dev, context);
1955 
1956 out_sys_pages:
1957 	kfree(bfregi->sys_pages);
1958 
1959 out_count:
1960 	kfree(bfregi->count);
1961 
1962 out_ctx:
1963 	return err;
1964 }
1965 
1966 static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext,
1967 				  struct uverbs_attr_bundle *attrs)
1968 {
1969 	struct mlx5_ib_alloc_ucontext_resp uctx_resp = {};
1970 	int ret;
1971 
1972 	ret = set_ucontext_resp(ibcontext, &uctx_resp);
1973 	if (ret)
1974 		return ret;
1975 
1976 	uctx_resp.response_length =
1977 		min_t(size_t,
1978 		      uverbs_attr_get_len(attrs,
1979 				MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX),
1980 		      sizeof(uctx_resp));
1981 
1982 	ret = uverbs_copy_to_struct_or_zero(attrs,
1983 					MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
1984 					&uctx_resp,
1985 					sizeof(uctx_resp));
1986 	return ret;
1987 }
1988 
1989 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1990 {
1991 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1992 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1993 	struct mlx5_bfreg_info *bfregi;
1994 
1995 	bfregi = &context->bfregi;
1996 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1997 
1998 	if (context->devx_uid)
1999 		mlx5_ib_devx_destroy(dev, context->devx_uid);
2000 
2001 	deallocate_uars(dev, context);
2002 	kfree(bfregi->sys_pages);
2003 	kfree(bfregi->count);
2004 }
2005 
2006 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2007 				 int uar_idx)
2008 {
2009 	int fw_uars_per_page;
2010 
2011 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2012 
2013 	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2014 }
2015 
2016 static u64 uar_index2paddress(struct mlx5_ib_dev *dev,
2017 				 int uar_idx)
2018 {
2019 	unsigned int fw_uars_per_page;
2020 
2021 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
2022 				MLX5_UARS_IN_PAGE : 1;
2023 
2024 	return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE);
2025 }
2026 
2027 static int get_command(unsigned long offset)
2028 {
2029 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2030 }
2031 
2032 static int get_arg(unsigned long offset)
2033 {
2034 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2035 }
2036 
2037 static int get_index(unsigned long offset)
2038 {
2039 	return get_arg(offset);
2040 }
2041 
2042 /* Index resides in an extra byte to enable larger values than 255 */
2043 static int get_extended_index(unsigned long offset)
2044 {
2045 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2046 }
2047 
2048 
2049 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2050 {
2051 }
2052 
2053 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2054 {
2055 	switch (cmd) {
2056 	case MLX5_IB_MMAP_WC_PAGE:
2057 		return "WC";
2058 	case MLX5_IB_MMAP_REGULAR_PAGE:
2059 		return "best effort WC";
2060 	case MLX5_IB_MMAP_NC_PAGE:
2061 		return "NC";
2062 	case MLX5_IB_MMAP_DEVICE_MEM:
2063 		return "Device Memory";
2064 	default:
2065 		return NULL;
2066 	}
2067 }
2068 
2069 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2070 					struct vm_area_struct *vma,
2071 					struct mlx5_ib_ucontext *context)
2072 {
2073 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2074 	    !(vma->vm_flags & VM_SHARED))
2075 		return -EINVAL;
2076 
2077 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2078 		return -EOPNOTSUPP;
2079 
2080 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2081 		return -EPERM;
2082 	vma->vm_flags &= ~VM_MAYWRITE;
2083 
2084 	if (!dev->mdev->clock_info)
2085 		return -EOPNOTSUPP;
2086 
2087 	return vm_insert_page(vma, vma->vm_start,
2088 			      virt_to_page(dev->mdev->clock_info));
2089 }
2090 
2091 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2092 {
2093 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2094 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2095 	struct mlx5_var_table *var_table = &dev->var_table;
2096 	struct mlx5_ib_dm *mdm;
2097 
2098 	switch (mentry->mmap_flag) {
2099 	case MLX5_IB_MMAP_TYPE_MEMIC:
2100 		mdm = container_of(mentry, struct mlx5_ib_dm, mentry);
2101 		mlx5_cmd_dealloc_memic(&dev->dm, mdm->dev_addr,
2102 				       mdm->size);
2103 		kfree(mdm);
2104 		break;
2105 	case MLX5_IB_MMAP_TYPE_VAR:
2106 		mutex_lock(&var_table->bitmap_lock);
2107 		clear_bit(mentry->page_idx, var_table->bitmap);
2108 		mutex_unlock(&var_table->bitmap_lock);
2109 		kfree(mentry);
2110 		break;
2111 	case MLX5_IB_MMAP_TYPE_UAR_WC:
2112 	case MLX5_IB_MMAP_TYPE_UAR_NC:
2113 		mlx5_cmd_free_uar(dev->mdev, mentry->page_idx);
2114 		kfree(mentry);
2115 		break;
2116 	default:
2117 		WARN_ON(true);
2118 	}
2119 }
2120 
2121 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2122 		    struct vm_area_struct *vma,
2123 		    struct mlx5_ib_ucontext *context)
2124 {
2125 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2126 	int err;
2127 	unsigned long idx;
2128 	phys_addr_t pfn;
2129 	pgprot_t prot;
2130 	u32 bfreg_dyn_idx = 0;
2131 	u32 uar_index;
2132 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2133 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2134 				bfregi->num_static_sys_pages;
2135 
2136 	if (bfregi->lib_uar_dyn)
2137 		return -EINVAL;
2138 
2139 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2140 		return -EINVAL;
2141 
2142 	if (dyn_uar)
2143 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2144 	else
2145 		idx = get_index(vma->vm_pgoff);
2146 
2147 	if (idx >= max_valid_idx) {
2148 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2149 			     idx, max_valid_idx);
2150 		return -EINVAL;
2151 	}
2152 
2153 	switch (cmd) {
2154 	case MLX5_IB_MMAP_WC_PAGE:
2155 	case MLX5_IB_MMAP_ALLOC_WC:
2156 	case MLX5_IB_MMAP_REGULAR_PAGE:
2157 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2158 		prot = pgprot_writecombine(vma->vm_page_prot);
2159 		break;
2160 	case MLX5_IB_MMAP_NC_PAGE:
2161 		prot = pgprot_noncached(vma->vm_page_prot);
2162 		break;
2163 	default:
2164 		return -EINVAL;
2165 	}
2166 
2167 	if (dyn_uar) {
2168 		int uars_per_page;
2169 
2170 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2171 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2172 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2173 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2174 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2175 			return -EINVAL;
2176 		}
2177 
2178 		mutex_lock(&bfregi->lock);
2179 		/* Fail if uar already allocated, first bfreg index of each
2180 		 * page holds its count.
2181 		 */
2182 		if (bfregi->count[bfreg_dyn_idx]) {
2183 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2184 			mutex_unlock(&bfregi->lock);
2185 			return -EINVAL;
2186 		}
2187 
2188 		bfregi->count[bfreg_dyn_idx]++;
2189 		mutex_unlock(&bfregi->lock);
2190 
2191 		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2192 		if (err) {
2193 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2194 			goto free_bfreg;
2195 		}
2196 	} else {
2197 		uar_index = bfregi->sys_pages[idx];
2198 	}
2199 
2200 	pfn = uar_index2pfn(dev, uar_index);
2201 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2202 
2203 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2204 				prot, NULL);
2205 	if (err) {
2206 		mlx5_ib_err(dev,
2207 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2208 			    err, mmap_cmd2str(cmd));
2209 		goto err;
2210 	}
2211 
2212 	if (dyn_uar)
2213 		bfregi->sys_pages[idx] = uar_index;
2214 	return 0;
2215 
2216 err:
2217 	if (!dyn_uar)
2218 		return err;
2219 
2220 	mlx5_cmd_free_uar(dev->mdev, idx);
2221 
2222 free_bfreg:
2223 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2224 
2225 	return err;
2226 }
2227 
2228 static int add_dm_mmap_entry(struct ib_ucontext *context,
2229 			     struct mlx5_ib_dm *mdm,
2230 			     u64 address)
2231 {
2232 	mdm->mentry.mmap_flag = MLX5_IB_MMAP_TYPE_MEMIC;
2233 	mdm->mentry.address = address;
2234 	return rdma_user_mmap_entry_insert_range(
2235 			context, &mdm->mentry.rdma_entry,
2236 			mdm->size,
2237 			MLX5_IB_MMAP_DEVICE_MEM << 16,
2238 			(MLX5_IB_MMAP_DEVICE_MEM << 16) + (1UL << 16) - 1);
2239 }
2240 
2241 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2242 {
2243 	unsigned long idx;
2244 	u8 command;
2245 
2246 	command = get_command(vma->vm_pgoff);
2247 	idx = get_extended_index(vma->vm_pgoff);
2248 
2249 	return (command << 16 | idx);
2250 }
2251 
2252 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2253 			       struct vm_area_struct *vma,
2254 			       struct ib_ucontext *ucontext)
2255 {
2256 	struct mlx5_user_mmap_entry *mentry;
2257 	struct rdma_user_mmap_entry *entry;
2258 	unsigned long pgoff;
2259 	pgprot_t prot;
2260 	phys_addr_t pfn;
2261 	int ret;
2262 
2263 	pgoff = mlx5_vma_to_pgoff(vma);
2264 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2265 	if (!entry)
2266 		return -EINVAL;
2267 
2268 	mentry = to_mmmap(entry);
2269 	pfn = (mentry->address >> PAGE_SHIFT);
2270 	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
2271 	    mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
2272 		prot = pgprot_noncached(vma->vm_page_prot);
2273 	else
2274 		prot = pgprot_writecombine(vma->vm_page_prot);
2275 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2276 				entry->npages * PAGE_SIZE,
2277 				prot,
2278 				entry);
2279 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2280 	return ret;
2281 }
2282 
2283 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2284 {
2285 	u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2286 	u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2287 
2288 	return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2289 		(index & 0xFF)) << PAGE_SHIFT;
2290 }
2291 
2292 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2293 {
2294 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2295 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2296 	unsigned long command;
2297 	phys_addr_t pfn;
2298 
2299 	command = get_command(vma->vm_pgoff);
2300 	switch (command) {
2301 	case MLX5_IB_MMAP_WC_PAGE:
2302 	case MLX5_IB_MMAP_ALLOC_WC:
2303 		if (!dev->wc_support)
2304 			return -EPERM;
2305 		fallthrough;
2306 	case MLX5_IB_MMAP_NC_PAGE:
2307 	case MLX5_IB_MMAP_REGULAR_PAGE:
2308 		return uar_mmap(dev, command, vma, context);
2309 
2310 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2311 		return -ENOSYS;
2312 
2313 	case MLX5_IB_MMAP_CORE_CLOCK:
2314 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2315 			return -EINVAL;
2316 
2317 		if (vma->vm_flags & VM_WRITE)
2318 			return -EPERM;
2319 		vma->vm_flags &= ~VM_MAYWRITE;
2320 
2321 		/* Don't expose to user-space information it shouldn't have */
2322 		if (PAGE_SIZE > 4096)
2323 			return -EOPNOTSUPP;
2324 
2325 		pfn = (dev->mdev->iseg_base +
2326 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2327 			PAGE_SHIFT;
2328 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2329 					 PAGE_SIZE,
2330 					 pgprot_noncached(vma->vm_page_prot),
2331 					 NULL);
2332 	case MLX5_IB_MMAP_CLOCK_INFO:
2333 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2334 
2335 	default:
2336 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2337 	}
2338 
2339 	return 0;
2340 }
2341 
2342 static inline int check_dm_type_support(struct mlx5_ib_dev *dev,
2343 					u32 type)
2344 {
2345 	switch (type) {
2346 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2347 		if (!MLX5_CAP_DEV_MEM(dev->mdev, memic))
2348 			return -EOPNOTSUPP;
2349 		break;
2350 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2351 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2352 		if (!capable(CAP_SYS_RAWIO) ||
2353 		    !capable(CAP_NET_RAW))
2354 			return -EPERM;
2355 
2356 		if (!(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
2357 		      MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, sw_owner)))
2358 			return -EOPNOTSUPP;
2359 		break;
2360 	}
2361 
2362 	return 0;
2363 }
2364 
2365 static int handle_alloc_dm_memic(struct ib_ucontext *ctx,
2366 				 struct mlx5_ib_dm *dm,
2367 				 struct ib_dm_alloc_attr *attr,
2368 				 struct uverbs_attr_bundle *attrs)
2369 {
2370 	struct mlx5_dm *dm_db = &to_mdev(ctx->device)->dm;
2371 	u64 start_offset;
2372 	u16 page_idx;
2373 	int err;
2374 	u64 address;
2375 
2376 	dm->size = roundup(attr->length, MLX5_MEMIC_BASE_SIZE);
2377 
2378 	err = mlx5_cmd_alloc_memic(dm_db, &dm->dev_addr,
2379 				   dm->size, attr->alignment);
2380 	if (err)
2381 		return err;
2382 
2383 	address = dm->dev_addr & PAGE_MASK;
2384 	err = add_dm_mmap_entry(ctx, dm, address);
2385 	if (err)
2386 		goto err_dealloc;
2387 
2388 	page_idx = dm->mentry.rdma_entry.start_pgoff & 0xFFFF;
2389 	err = uverbs_copy_to(attrs,
2390 			     MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
2391 			     &page_idx,
2392 			     sizeof(page_idx));
2393 	if (err)
2394 		goto err_copy;
2395 
2396 	start_offset = dm->dev_addr & ~PAGE_MASK;
2397 	err = uverbs_copy_to(attrs,
2398 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2399 			     &start_offset, sizeof(start_offset));
2400 	if (err)
2401 		goto err_copy;
2402 
2403 	return 0;
2404 
2405 err_copy:
2406 	rdma_user_mmap_entry_remove(&dm->mentry.rdma_entry);
2407 err_dealloc:
2408 	mlx5_cmd_dealloc_memic(dm_db, dm->dev_addr, dm->size);
2409 
2410 	return err;
2411 }
2412 
2413 static int handle_alloc_dm_sw_icm(struct ib_ucontext *ctx,
2414 				  struct mlx5_ib_dm *dm,
2415 				  struct ib_dm_alloc_attr *attr,
2416 				  struct uverbs_attr_bundle *attrs,
2417 				  int type)
2418 {
2419 	struct mlx5_core_dev *dev = to_mdev(ctx->device)->mdev;
2420 	u64 act_size;
2421 	int err;
2422 
2423 	/* Allocation size must a multiple of the basic block size
2424 	 * and a power of 2.
2425 	 */
2426 	act_size = round_up(attr->length, MLX5_SW_ICM_BLOCK_SIZE(dev));
2427 	act_size = roundup_pow_of_two(act_size);
2428 
2429 	dm->size = act_size;
2430 	err = mlx5_dm_sw_icm_alloc(dev, type, act_size, attr->alignment,
2431 				   to_mucontext(ctx)->devx_uid, &dm->dev_addr,
2432 				   &dm->icm_dm.obj_id);
2433 	if (err)
2434 		return err;
2435 
2436 	err = uverbs_copy_to(attrs,
2437 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2438 			     &dm->dev_addr, sizeof(dm->dev_addr));
2439 	if (err)
2440 		mlx5_dm_sw_icm_dealloc(dev, type, dm->size,
2441 				       to_mucontext(ctx)->devx_uid, dm->dev_addr,
2442 				       dm->icm_dm.obj_id);
2443 
2444 	return err;
2445 }
2446 
2447 struct ib_dm *mlx5_ib_alloc_dm(struct ib_device *ibdev,
2448 			       struct ib_ucontext *context,
2449 			       struct ib_dm_alloc_attr *attr,
2450 			       struct uverbs_attr_bundle *attrs)
2451 {
2452 	struct mlx5_ib_dm *dm;
2453 	enum mlx5_ib_uapi_dm_type type;
2454 	int err;
2455 
2456 	err = uverbs_get_const_default(&type, attrs,
2457 				       MLX5_IB_ATTR_ALLOC_DM_REQ_TYPE,
2458 				       MLX5_IB_UAPI_DM_TYPE_MEMIC);
2459 	if (err)
2460 		return ERR_PTR(err);
2461 
2462 	mlx5_ib_dbg(to_mdev(ibdev), "alloc_dm req: dm_type=%d user_length=0x%llx log_alignment=%d\n",
2463 		    type, attr->length, attr->alignment);
2464 
2465 	err = check_dm_type_support(to_mdev(ibdev), type);
2466 	if (err)
2467 		return ERR_PTR(err);
2468 
2469 	dm = kzalloc(sizeof(*dm), GFP_KERNEL);
2470 	if (!dm)
2471 		return ERR_PTR(-ENOMEM);
2472 
2473 	dm->type = type;
2474 
2475 	switch (type) {
2476 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2477 		err = handle_alloc_dm_memic(context, dm,
2478 					    attr,
2479 					    attrs);
2480 		break;
2481 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2482 		err = handle_alloc_dm_sw_icm(context, dm,
2483 					     attr, attrs,
2484 					     MLX5_SW_ICM_TYPE_STEERING);
2485 		break;
2486 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2487 		err = handle_alloc_dm_sw_icm(context, dm,
2488 					     attr, attrs,
2489 					     MLX5_SW_ICM_TYPE_HEADER_MODIFY);
2490 		break;
2491 	default:
2492 		err = -EOPNOTSUPP;
2493 	}
2494 
2495 	if (err)
2496 		goto err_free;
2497 
2498 	return &dm->ibdm;
2499 
2500 err_free:
2501 	kfree(dm);
2502 	return ERR_PTR(err);
2503 }
2504 
2505 int mlx5_ib_dealloc_dm(struct ib_dm *ibdm, struct uverbs_attr_bundle *attrs)
2506 {
2507 	struct mlx5_ib_ucontext *ctx = rdma_udata_to_drv_context(
2508 		&attrs->driver_udata, struct mlx5_ib_ucontext, ibucontext);
2509 	struct mlx5_core_dev *dev = to_mdev(ibdm->device)->mdev;
2510 	struct mlx5_ib_dm *dm = to_mdm(ibdm);
2511 	int ret;
2512 
2513 	switch (dm->type) {
2514 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2515 		rdma_user_mmap_entry_remove(&dm->mentry.rdma_entry);
2516 		return 0;
2517 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2518 		ret = mlx5_dm_sw_icm_dealloc(dev, MLX5_SW_ICM_TYPE_STEERING,
2519 					     dm->size, ctx->devx_uid, dm->dev_addr,
2520 					     dm->icm_dm.obj_id);
2521 		if (ret)
2522 			return ret;
2523 		break;
2524 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2525 		ret = mlx5_dm_sw_icm_dealloc(dev, MLX5_SW_ICM_TYPE_HEADER_MODIFY,
2526 					     dm->size, ctx->devx_uid, dm->dev_addr,
2527 					     dm->icm_dm.obj_id);
2528 		if (ret)
2529 			return ret;
2530 		break;
2531 	default:
2532 		return -EOPNOTSUPP;
2533 	}
2534 
2535 	kfree(dm);
2536 
2537 	return 0;
2538 }
2539 
2540 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2541 {
2542 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2543 	struct ib_device *ibdev = ibpd->device;
2544 	struct mlx5_ib_alloc_pd_resp resp;
2545 	int err;
2546 	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2547 	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {};
2548 	u16 uid = 0;
2549 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2550 		udata, struct mlx5_ib_ucontext, ibucontext);
2551 
2552 	uid = context ? context->devx_uid : 0;
2553 	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2554 	MLX5_SET(alloc_pd_in, in, uid, uid);
2555 	err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out);
2556 	if (err)
2557 		return err;
2558 
2559 	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2560 	pd->uid = uid;
2561 	if (udata) {
2562 		resp.pdn = pd->pdn;
2563 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2564 			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2565 			return -EFAULT;
2566 		}
2567 	}
2568 
2569 	return 0;
2570 }
2571 
2572 static void mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2573 {
2574 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2575 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2576 
2577 	mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2578 }
2579 
2580 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2581 {
2582 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2583 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
2584 	int err;
2585 	u16 uid;
2586 
2587 	uid = ibqp->pd ?
2588 		to_mpd(ibqp->pd)->uid : 0;
2589 
2590 	if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) {
2591 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
2592 		return -EOPNOTSUPP;
2593 	}
2594 
2595 	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2596 	if (err)
2597 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2598 			     ibqp->qp_num, gid->raw);
2599 
2600 	return err;
2601 }
2602 
2603 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2604 {
2605 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2606 	int err;
2607 	u16 uid;
2608 
2609 	uid = ibqp->pd ?
2610 		to_mpd(ibqp->pd)->uid : 0;
2611 	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2612 	if (err)
2613 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2614 			     ibqp->qp_num, gid->raw);
2615 
2616 	return err;
2617 }
2618 
2619 static int init_node_data(struct mlx5_ib_dev *dev)
2620 {
2621 	int err;
2622 
2623 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2624 	if (err)
2625 		return err;
2626 
2627 	dev->mdev->rev_id = dev->mdev->pdev->revision;
2628 
2629 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2630 }
2631 
2632 static ssize_t fw_pages_show(struct device *device,
2633 			     struct device_attribute *attr, char *buf)
2634 {
2635 	struct mlx5_ib_dev *dev =
2636 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2637 
2638 	return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
2639 }
2640 static DEVICE_ATTR_RO(fw_pages);
2641 
2642 static ssize_t reg_pages_show(struct device *device,
2643 			      struct device_attribute *attr, char *buf)
2644 {
2645 	struct mlx5_ib_dev *dev =
2646 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2647 
2648 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2649 }
2650 static DEVICE_ATTR_RO(reg_pages);
2651 
2652 static ssize_t hca_type_show(struct device *device,
2653 			     struct device_attribute *attr, char *buf)
2654 {
2655 	struct mlx5_ib_dev *dev =
2656 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2657 
2658 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
2659 }
2660 static DEVICE_ATTR_RO(hca_type);
2661 
2662 static ssize_t hw_rev_show(struct device *device,
2663 			   struct device_attribute *attr, char *buf)
2664 {
2665 	struct mlx5_ib_dev *dev =
2666 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2667 
2668 	return sprintf(buf, "%x\n", dev->mdev->rev_id);
2669 }
2670 static DEVICE_ATTR_RO(hw_rev);
2671 
2672 static ssize_t board_id_show(struct device *device,
2673 			     struct device_attribute *attr, char *buf)
2674 {
2675 	struct mlx5_ib_dev *dev =
2676 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2677 
2678 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2679 		       dev->mdev->board_id);
2680 }
2681 static DEVICE_ATTR_RO(board_id);
2682 
2683 static struct attribute *mlx5_class_attributes[] = {
2684 	&dev_attr_hw_rev.attr,
2685 	&dev_attr_hca_type.attr,
2686 	&dev_attr_board_id.attr,
2687 	&dev_attr_fw_pages.attr,
2688 	&dev_attr_reg_pages.attr,
2689 	NULL,
2690 };
2691 
2692 static const struct attribute_group mlx5_attr_group = {
2693 	.attrs = mlx5_class_attributes,
2694 };
2695 
2696 static void pkey_change_handler(struct work_struct *work)
2697 {
2698 	struct mlx5_ib_port_resources *ports =
2699 		container_of(work, struct mlx5_ib_port_resources,
2700 			     pkey_change_work);
2701 
2702 	mutex_lock(&ports->devr->mutex);
2703 	mlx5_ib_gsi_pkey_change(ports->gsi);
2704 	mutex_unlock(&ports->devr->mutex);
2705 }
2706 
2707 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2708 {
2709 	struct mlx5_ib_qp *mqp;
2710 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2711 	struct mlx5_core_cq *mcq;
2712 	struct list_head cq_armed_list;
2713 	unsigned long flags_qp;
2714 	unsigned long flags_cq;
2715 	unsigned long flags;
2716 
2717 	INIT_LIST_HEAD(&cq_armed_list);
2718 
2719 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2720 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2721 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2722 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2723 		if (mqp->sq.tail != mqp->sq.head) {
2724 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2725 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2726 			if (send_mcq->mcq.comp &&
2727 			    mqp->ibqp.send_cq->comp_handler) {
2728 				if (!send_mcq->mcq.reset_notify_added) {
2729 					send_mcq->mcq.reset_notify_added = 1;
2730 					list_add_tail(&send_mcq->mcq.reset_notify,
2731 						      &cq_armed_list);
2732 				}
2733 			}
2734 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2735 		}
2736 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2737 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2738 		/* no handling is needed for SRQ */
2739 		if (!mqp->ibqp.srq) {
2740 			if (mqp->rq.tail != mqp->rq.head) {
2741 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2742 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2743 				if (recv_mcq->mcq.comp &&
2744 				    mqp->ibqp.recv_cq->comp_handler) {
2745 					if (!recv_mcq->mcq.reset_notify_added) {
2746 						recv_mcq->mcq.reset_notify_added = 1;
2747 						list_add_tail(&recv_mcq->mcq.reset_notify,
2748 							      &cq_armed_list);
2749 					}
2750 				}
2751 				spin_unlock_irqrestore(&recv_mcq->lock,
2752 						       flags_cq);
2753 			}
2754 		}
2755 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2756 	}
2757 	/*At that point all inflight post send were put to be executed as of we
2758 	 * lock/unlock above locks Now need to arm all involved CQs.
2759 	 */
2760 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2761 		mcq->comp(mcq, NULL);
2762 	}
2763 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2764 }
2765 
2766 static void delay_drop_handler(struct work_struct *work)
2767 {
2768 	int err;
2769 	struct mlx5_ib_delay_drop *delay_drop =
2770 		container_of(work, struct mlx5_ib_delay_drop,
2771 			     delay_drop_work);
2772 
2773 	atomic_inc(&delay_drop->events_cnt);
2774 
2775 	mutex_lock(&delay_drop->lock);
2776 	err = mlx5_core_set_delay_drop(delay_drop->dev, delay_drop->timeout);
2777 	if (err) {
2778 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
2779 			     delay_drop->timeout);
2780 		delay_drop->activate = false;
2781 	}
2782 	mutex_unlock(&delay_drop->lock);
2783 }
2784 
2785 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2786 				 struct ib_event *ibev)
2787 {
2788 	u8 port = (eqe->data.port.port >> 4) & 0xf;
2789 
2790 	switch (eqe->sub_type) {
2791 	case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
2792 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2793 					    IB_LINK_LAYER_ETHERNET)
2794 			schedule_work(&ibdev->delay_drop.delay_drop_work);
2795 		break;
2796 	default: /* do nothing */
2797 		return;
2798 	}
2799 }
2800 
2801 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2802 			      struct ib_event *ibev)
2803 {
2804 	u8 port = (eqe->data.port.port >> 4) & 0xf;
2805 
2806 	ibev->element.port_num = port;
2807 
2808 	switch (eqe->sub_type) {
2809 	case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
2810 	case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
2811 	case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
2812 		/* In RoCE, port up/down events are handled in
2813 		 * mlx5_netdev_event().
2814 		 */
2815 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2816 					    IB_LINK_LAYER_ETHERNET)
2817 			return -EINVAL;
2818 
2819 		ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
2820 				IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2821 		break;
2822 
2823 	case MLX5_PORT_CHANGE_SUBTYPE_LID:
2824 		ibev->event = IB_EVENT_LID_CHANGE;
2825 		break;
2826 
2827 	case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
2828 		ibev->event = IB_EVENT_PKEY_CHANGE;
2829 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2830 		break;
2831 
2832 	case MLX5_PORT_CHANGE_SUBTYPE_GUID:
2833 		ibev->event = IB_EVENT_GID_CHANGE;
2834 		break;
2835 
2836 	case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
2837 		ibev->event = IB_EVENT_CLIENT_REREGISTER;
2838 		break;
2839 	default:
2840 		return -EINVAL;
2841 	}
2842 
2843 	return 0;
2844 }
2845 
2846 static void mlx5_ib_handle_event(struct work_struct *_work)
2847 {
2848 	struct mlx5_ib_event_work *work =
2849 		container_of(_work, struct mlx5_ib_event_work, work);
2850 	struct mlx5_ib_dev *ibdev;
2851 	struct ib_event ibev;
2852 	bool fatal = false;
2853 
2854 	if (work->is_slave) {
2855 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
2856 		if (!ibdev)
2857 			goto out;
2858 	} else {
2859 		ibdev = work->dev;
2860 	}
2861 
2862 	switch (work->event) {
2863 	case MLX5_DEV_EVENT_SYS_ERROR:
2864 		ibev.event = IB_EVENT_DEVICE_FATAL;
2865 		mlx5_ib_handle_internal_error(ibdev);
2866 		ibev.element.port_num  = (u8)(unsigned long)work->param;
2867 		fatal = true;
2868 		break;
2869 	case MLX5_EVENT_TYPE_PORT_CHANGE:
2870 		if (handle_port_change(ibdev, work->param, &ibev))
2871 			goto out;
2872 		break;
2873 	case MLX5_EVENT_TYPE_GENERAL_EVENT:
2874 		handle_general_event(ibdev, work->param, &ibev);
2875 		/* fall through */
2876 	default:
2877 		goto out;
2878 	}
2879 
2880 	ibev.device = &ibdev->ib_dev;
2881 
2882 	if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
2883 		mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
2884 		goto out;
2885 	}
2886 
2887 	if (ibdev->ib_active)
2888 		ib_dispatch_event(&ibev);
2889 
2890 	if (fatal)
2891 		ibdev->ib_active = false;
2892 out:
2893 	kfree(work);
2894 }
2895 
2896 static int mlx5_ib_event(struct notifier_block *nb,
2897 			 unsigned long event, void *param)
2898 {
2899 	struct mlx5_ib_event_work *work;
2900 
2901 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2902 	if (!work)
2903 		return NOTIFY_DONE;
2904 
2905 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2906 	work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
2907 	work->is_slave = false;
2908 	work->param = param;
2909 	work->event = event;
2910 
2911 	queue_work(mlx5_ib_event_wq, &work->work);
2912 
2913 	return NOTIFY_OK;
2914 }
2915 
2916 static int mlx5_ib_event_slave_port(struct notifier_block *nb,
2917 				    unsigned long event, void *param)
2918 {
2919 	struct mlx5_ib_event_work *work;
2920 
2921 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2922 	if (!work)
2923 		return NOTIFY_DONE;
2924 
2925 	INIT_WORK(&work->work, mlx5_ib_handle_event);
2926 	work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
2927 	work->is_slave = true;
2928 	work->param = param;
2929 	work->event = event;
2930 	queue_work(mlx5_ib_event_wq, &work->work);
2931 
2932 	return NOTIFY_OK;
2933 }
2934 
2935 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2936 {
2937 	struct mlx5_hca_vport_context vport_ctx;
2938 	int err;
2939 	int port;
2940 
2941 	for (port = 1; port <= ARRAY_SIZE(dev->mdev->port_caps); port++) {
2942 		dev->mdev->port_caps[port - 1].has_smi = false;
2943 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
2944 		    MLX5_CAP_PORT_TYPE_IB) {
2945 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2946 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
2947 								   port, 0,
2948 								   &vport_ctx);
2949 				if (err) {
2950 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2951 						    port, err);
2952 					return err;
2953 				}
2954 				dev->mdev->port_caps[port - 1].has_smi =
2955 					vport_ctx.has_smi;
2956 			} else {
2957 				dev->mdev->port_caps[port - 1].has_smi = true;
2958 			}
2959 		}
2960 	}
2961 	return 0;
2962 }
2963 
2964 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2965 {
2966 	int port;
2967 
2968 	for (port = 1; port <= dev->num_ports; port++)
2969 		mlx5_query_ext_port_caps(dev, port);
2970 }
2971 
2972 static int __get_port_caps(struct mlx5_ib_dev *dev, u8 port)
2973 {
2974 	struct ib_device_attr *dprops = NULL;
2975 	struct ib_port_attr *pprops = NULL;
2976 	int err = -ENOMEM;
2977 
2978 	pprops = kzalloc(sizeof(*pprops), GFP_KERNEL);
2979 	if (!pprops)
2980 		goto out;
2981 
2982 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2983 	if (!dprops)
2984 		goto out;
2985 
2986 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, NULL);
2987 	if (err) {
2988 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
2989 		goto out;
2990 	}
2991 
2992 	err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2993 	if (err) {
2994 		mlx5_ib_warn(dev, "query_port %d failed %d\n",
2995 			     port, err);
2996 		goto out;
2997 	}
2998 
2999 	dev->mdev->port_caps[port - 1].pkey_table_len =
3000 					dprops->max_pkeys;
3001 	dev->mdev->port_caps[port - 1].gid_table_len =
3002 					pprops->gid_tbl_len;
3003 	mlx5_ib_dbg(dev, "port %d: pkey_table_len %d, gid_table_len %d\n",
3004 		    port, dprops->max_pkeys, pprops->gid_tbl_len);
3005 
3006 out:
3007 	kfree(pprops);
3008 	kfree(dprops);
3009 
3010 	return err;
3011 }
3012 
3013 static int get_port_caps(struct mlx5_ib_dev *dev, u8 port)
3014 {
3015 	/* For representors use port 1, is this is the only native
3016 	 * port
3017 	 */
3018 	if (dev->is_rep)
3019 		return __get_port_caps(dev, 1);
3020 	return __get_port_caps(dev, port);
3021 }
3022 
3023 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
3024 {
3025 	switch (umr_fence_cap) {
3026 	case MLX5_CAP_UMR_FENCE_NONE:
3027 		return MLX5_FENCE_MODE_NONE;
3028 	case MLX5_CAP_UMR_FENCE_SMALL:
3029 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
3030 	default:
3031 		return MLX5_FENCE_MODE_STRONG_ORDERING;
3032 	}
3033 }
3034 
3035 static int mlx5_ib_dev_res_init(struct mlx5_ib_dev *dev)
3036 {
3037 	struct mlx5_ib_resources *devr = &dev->devr;
3038 	struct ib_srq_init_attr attr;
3039 	struct ib_device *ibdev;
3040 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
3041 	int port;
3042 	int ret = 0;
3043 
3044 	ibdev = &dev->ib_dev;
3045 
3046 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
3047 		return -EOPNOTSUPP;
3048 
3049 	mutex_init(&devr->mutex);
3050 
3051 	devr->p0 = rdma_zalloc_drv_obj(ibdev, ib_pd);
3052 	if (!devr->p0)
3053 		return -ENOMEM;
3054 
3055 	devr->p0->device  = ibdev;
3056 	devr->p0->uobject = NULL;
3057 	atomic_set(&devr->p0->usecnt, 0);
3058 
3059 	ret = mlx5_ib_alloc_pd(devr->p0, NULL);
3060 	if (ret)
3061 		goto error0;
3062 
3063 	devr->c0 = rdma_zalloc_drv_obj(ibdev, ib_cq);
3064 	if (!devr->c0) {
3065 		ret = -ENOMEM;
3066 		goto error1;
3067 	}
3068 
3069 	devr->c0->device = &dev->ib_dev;
3070 	atomic_set(&devr->c0->usecnt, 0);
3071 
3072 	ret = mlx5_ib_create_cq(devr->c0, &cq_attr, NULL);
3073 	if (ret)
3074 		goto err_create_cq;
3075 
3076 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0);
3077 	if (ret)
3078 		goto error2;
3079 
3080 	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0);
3081 	if (ret)
3082 		goto error3;
3083 
3084 	memset(&attr, 0, sizeof(attr));
3085 	attr.attr.max_sge = 1;
3086 	attr.attr.max_wr = 1;
3087 	attr.srq_type = IB_SRQT_XRC;
3088 	attr.ext.cq = devr->c0;
3089 
3090 	devr->s0 = rdma_zalloc_drv_obj(ibdev, ib_srq);
3091 	if (!devr->s0) {
3092 		ret = -ENOMEM;
3093 		goto error4;
3094 	}
3095 
3096 	devr->s0->device	= &dev->ib_dev;
3097 	devr->s0->pd		= devr->p0;
3098 	devr->s0->srq_type      = IB_SRQT_XRC;
3099 	devr->s0->ext.cq	= devr->c0;
3100 	ret = mlx5_ib_create_srq(devr->s0, &attr, NULL);
3101 	if (ret)
3102 		goto err_create;
3103 
3104 	atomic_inc(&devr->s0->ext.cq->usecnt);
3105 	atomic_inc(&devr->p0->usecnt);
3106 	atomic_set(&devr->s0->usecnt, 0);
3107 
3108 	memset(&attr, 0, sizeof(attr));
3109 	attr.attr.max_sge = 1;
3110 	attr.attr.max_wr = 1;
3111 	attr.srq_type = IB_SRQT_BASIC;
3112 	devr->s1 = rdma_zalloc_drv_obj(ibdev, ib_srq);
3113 	if (!devr->s1) {
3114 		ret = -ENOMEM;
3115 		goto error5;
3116 	}
3117 
3118 	devr->s1->device	= &dev->ib_dev;
3119 	devr->s1->pd		= devr->p0;
3120 	devr->s1->srq_type      = IB_SRQT_BASIC;
3121 	devr->s1->ext.cq	= devr->c0;
3122 
3123 	ret = mlx5_ib_create_srq(devr->s1, &attr, NULL);
3124 	if (ret)
3125 		goto error6;
3126 
3127 	atomic_inc(&devr->p0->usecnt);
3128 	atomic_set(&devr->s1->usecnt, 0);
3129 
3130 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
3131 		INIT_WORK(&devr->ports[port].pkey_change_work,
3132 			  pkey_change_handler);
3133 		devr->ports[port].devr = devr;
3134 	}
3135 
3136 	return 0;
3137 
3138 error6:
3139 	kfree(devr->s1);
3140 error5:
3141 	mlx5_ib_destroy_srq(devr->s0, NULL);
3142 err_create:
3143 	kfree(devr->s0);
3144 error4:
3145 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
3146 error3:
3147 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
3148 error2:
3149 	mlx5_ib_destroy_cq(devr->c0, NULL);
3150 err_create_cq:
3151 	kfree(devr->c0);
3152 error1:
3153 	mlx5_ib_dealloc_pd(devr->p0, NULL);
3154 error0:
3155 	kfree(devr->p0);
3156 	return ret;
3157 }
3158 
3159 static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev)
3160 {
3161 	struct mlx5_ib_resources *devr = &dev->devr;
3162 	int port;
3163 
3164 	mlx5_ib_destroy_srq(devr->s1, NULL);
3165 	kfree(devr->s1);
3166 	mlx5_ib_destroy_srq(devr->s0, NULL);
3167 	kfree(devr->s0);
3168 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
3169 	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
3170 	mlx5_ib_destroy_cq(devr->c0, NULL);
3171 	kfree(devr->c0);
3172 	mlx5_ib_dealloc_pd(devr->p0, NULL);
3173 	kfree(devr->p0);
3174 
3175 	/* Make sure no change P_Key work items are still executing */
3176 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
3177 		cancel_work_sync(&devr->ports[port].pkey_change_work);
3178 }
3179 
3180 static u32 get_core_cap_flags(struct ib_device *ibdev,
3181 			      struct mlx5_hca_vport_context *rep)
3182 {
3183 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3184 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
3185 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
3186 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
3187 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
3188 	u32 ret = 0;
3189 
3190 	if (rep->grh_required)
3191 		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
3192 
3193 	if (ll == IB_LINK_LAYER_INFINIBAND)
3194 		return ret | RDMA_CORE_PORT_IBA_IB;
3195 
3196 	if (raw_support)
3197 		ret |= RDMA_CORE_PORT_RAW_PACKET;
3198 
3199 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
3200 		return ret;
3201 
3202 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
3203 		return ret;
3204 
3205 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
3206 		ret |= RDMA_CORE_PORT_IBA_ROCE;
3207 
3208 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
3209 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3210 
3211 	return ret;
3212 }
3213 
3214 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
3215 			       struct ib_port_immutable *immutable)
3216 {
3217 	struct ib_port_attr attr;
3218 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3219 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
3220 	struct mlx5_hca_vport_context rep = {0};
3221 	int err;
3222 
3223 	err = ib_query_port(ibdev, port_num, &attr);
3224 	if (err)
3225 		return err;
3226 
3227 	if (ll == IB_LINK_LAYER_INFINIBAND) {
3228 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
3229 						   &rep);
3230 		if (err)
3231 			return err;
3232 	}
3233 
3234 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
3235 	immutable->gid_tbl_len = attr.gid_tbl_len;
3236 	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
3237 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3238 
3239 	return 0;
3240 }
3241 
3242 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u8 port_num,
3243 				   struct ib_port_immutable *immutable)
3244 {
3245 	struct ib_port_attr attr;
3246 	int err;
3247 
3248 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
3249 
3250 	err = ib_query_port(ibdev, port_num, &attr);
3251 	if (err)
3252 		return err;
3253 
3254 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
3255 	immutable->gid_tbl_len = attr.gid_tbl_len;
3256 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
3257 
3258 	return 0;
3259 }
3260 
3261 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
3262 {
3263 	struct mlx5_ib_dev *dev =
3264 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
3265 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
3266 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
3267 		 fw_rev_sub(dev->mdev));
3268 }
3269 
3270 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
3271 {
3272 	struct mlx5_core_dev *mdev = dev->mdev;
3273 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
3274 								 MLX5_FLOW_NAMESPACE_LAG);
3275 	struct mlx5_flow_table *ft;
3276 	int err;
3277 
3278 	if (!ns || !mlx5_lag_is_roce(mdev))
3279 		return 0;
3280 
3281 	err = mlx5_cmd_create_vport_lag(mdev);
3282 	if (err)
3283 		return err;
3284 
3285 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3286 	if (IS_ERR(ft)) {
3287 		err = PTR_ERR(ft);
3288 		goto err_destroy_vport_lag;
3289 	}
3290 
3291 	dev->flow_db->lag_demux_ft = ft;
3292 	dev->lag_active = true;
3293 	return 0;
3294 
3295 err_destroy_vport_lag:
3296 	mlx5_cmd_destroy_vport_lag(mdev);
3297 	return err;
3298 }
3299 
3300 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3301 {
3302 	struct mlx5_core_dev *mdev = dev->mdev;
3303 
3304 	if (dev->lag_active) {
3305 		dev->lag_active = false;
3306 
3307 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
3308 		dev->flow_db->lag_demux_ft = NULL;
3309 
3310 		mlx5_cmd_destroy_vport_lag(mdev);
3311 	}
3312 }
3313 
3314 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
3315 {
3316 	int err;
3317 
3318 	dev->port[port_num].roce.nb.notifier_call = mlx5_netdev_event;
3319 	err = register_netdevice_notifier(&dev->port[port_num].roce.nb);
3320 	if (err) {
3321 		dev->port[port_num].roce.nb.notifier_call = NULL;
3322 		return err;
3323 	}
3324 
3325 	return 0;
3326 }
3327 
3328 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
3329 {
3330 	if (dev->port[port_num].roce.nb.notifier_call) {
3331 		unregister_netdevice_notifier(&dev->port[port_num].roce.nb);
3332 		dev->port[port_num].roce.nb.notifier_call = NULL;
3333 	}
3334 }
3335 
3336 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3337 {
3338 	int err;
3339 
3340 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3341 	if (err)
3342 		return err;
3343 
3344 	err = mlx5_eth_lag_init(dev);
3345 	if (err)
3346 		goto err_disable_roce;
3347 
3348 	return 0;
3349 
3350 err_disable_roce:
3351 	mlx5_nic_vport_disable_roce(dev->mdev);
3352 
3353 	return err;
3354 }
3355 
3356 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3357 {
3358 	mlx5_eth_lag_cleanup(dev);
3359 	mlx5_nic_vport_disable_roce(dev->mdev);
3360 }
3361 
3362 static int mlx5_ib_rn_get_params(struct ib_device *device, u8 port_num,
3363 				 enum rdma_netdev_t type,
3364 				 struct rdma_netdev_alloc_params *params)
3365 {
3366 	if (type != RDMA_NETDEV_IPOIB)
3367 		return -EOPNOTSUPP;
3368 
3369 	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
3370 }
3371 
3372 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
3373 				       size_t count, loff_t *pos)
3374 {
3375 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3376 	char lbuf[20];
3377 	int len;
3378 
3379 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
3380 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
3381 }
3382 
3383 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
3384 					size_t count, loff_t *pos)
3385 {
3386 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3387 	u32 timeout;
3388 	u32 var;
3389 
3390 	if (kstrtouint_from_user(buf, count, 0, &var))
3391 		return -EFAULT;
3392 
3393 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
3394 			1000);
3395 	if (timeout != var)
3396 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
3397 			    timeout);
3398 
3399 	delay_drop->timeout = timeout;
3400 
3401 	return count;
3402 }
3403 
3404 static const struct file_operations fops_delay_drop_timeout = {
3405 	.owner	= THIS_MODULE,
3406 	.open	= simple_open,
3407 	.write	= delay_drop_timeout_write,
3408 	.read	= delay_drop_timeout_read,
3409 };
3410 
3411 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
3412 				      struct mlx5_ib_multiport_info *mpi)
3413 {
3414 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3415 	struct mlx5_ib_port *port = &ibdev->port[port_num];
3416 	int comps;
3417 	int err;
3418 	int i;
3419 
3420 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3421 
3422 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
3423 
3424 	spin_lock(&port->mp.mpi_lock);
3425 	if (!mpi->ibdev) {
3426 		spin_unlock(&port->mp.mpi_lock);
3427 		return;
3428 	}
3429 
3430 	mpi->ibdev = NULL;
3431 
3432 	spin_unlock(&port->mp.mpi_lock);
3433 	if (mpi->mdev_events.notifier_call)
3434 		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
3435 	mpi->mdev_events.notifier_call = NULL;
3436 	mlx5_remove_netdev_notifier(ibdev, port_num);
3437 	spin_lock(&port->mp.mpi_lock);
3438 
3439 	comps = mpi->mdev_refcnt;
3440 	if (comps) {
3441 		mpi->unaffiliate = true;
3442 		init_completion(&mpi->unref_comp);
3443 		spin_unlock(&port->mp.mpi_lock);
3444 
3445 		for (i = 0; i < comps; i++)
3446 			wait_for_completion(&mpi->unref_comp);
3447 
3448 		spin_lock(&port->mp.mpi_lock);
3449 		mpi->unaffiliate = false;
3450 	}
3451 
3452 	port->mp.mpi = NULL;
3453 
3454 	list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
3455 
3456 	spin_unlock(&port->mp.mpi_lock);
3457 
3458 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
3459 
3460 	mlx5_ib_dbg(ibdev, "unaffiliated port %d\n", port_num + 1);
3461 	/* Log an error, still needed to cleanup the pointers and add
3462 	 * it back to the list.
3463 	 */
3464 	if (err)
3465 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
3466 			    port_num + 1);
3467 
3468 	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
3469 }
3470 
3471 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
3472 				    struct mlx5_ib_multiport_info *mpi)
3473 {
3474 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3475 	int err;
3476 
3477 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3478 
3479 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
3480 	if (ibdev->port[port_num].mp.mpi) {
3481 		mlx5_ib_dbg(ibdev, "port %d already affiliated.\n",
3482 			    port_num + 1);
3483 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3484 		return false;
3485 	}
3486 
3487 	ibdev->port[port_num].mp.mpi = mpi;
3488 	mpi->ibdev = ibdev;
3489 	mpi->mdev_events.notifier_call = NULL;
3490 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3491 
3492 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
3493 	if (err)
3494 		goto unbind;
3495 
3496 	err = get_port_caps(ibdev, mlx5_core_native_port_num(mpi->mdev));
3497 	if (err)
3498 		goto unbind;
3499 
3500 	err = mlx5_add_netdev_notifier(ibdev, port_num);
3501 	if (err) {
3502 		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
3503 			    port_num + 1);
3504 		goto unbind;
3505 	}
3506 
3507 	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
3508 	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
3509 
3510 	mlx5_ib_init_cong_debugfs(ibdev, port_num);
3511 
3512 	return true;
3513 
3514 unbind:
3515 	mlx5_ib_unbind_slave_port(ibdev, mpi);
3516 	return false;
3517 }
3518 
3519 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
3520 {
3521 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3522 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3523 							  port_num + 1);
3524 	struct mlx5_ib_multiport_info *mpi;
3525 	int err;
3526 	int i;
3527 
3528 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3529 		return 0;
3530 
3531 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
3532 						     &dev->sys_image_guid);
3533 	if (err)
3534 		return err;
3535 
3536 	err = mlx5_nic_vport_enable_roce(dev->mdev);
3537 	if (err)
3538 		return err;
3539 
3540 	mutex_lock(&mlx5_ib_multiport_mutex);
3541 	for (i = 0; i < dev->num_ports; i++) {
3542 		bool bound = false;
3543 
3544 		/* build a stub multiport info struct for the native port. */
3545 		if (i == port_num) {
3546 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
3547 			if (!mpi) {
3548 				mutex_unlock(&mlx5_ib_multiport_mutex);
3549 				mlx5_nic_vport_disable_roce(dev->mdev);
3550 				return -ENOMEM;
3551 			}
3552 
3553 			mpi->is_master = true;
3554 			mpi->mdev = dev->mdev;
3555 			mpi->sys_image_guid = dev->sys_image_guid;
3556 			dev->port[i].mp.mpi = mpi;
3557 			mpi->ibdev = dev;
3558 			mpi = NULL;
3559 			continue;
3560 		}
3561 
3562 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
3563 				    list) {
3564 			if (dev->sys_image_guid == mpi->sys_image_guid &&
3565 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
3566 				bound = mlx5_ib_bind_slave_port(dev, mpi);
3567 			}
3568 
3569 			if (bound) {
3570 				dev_dbg(mpi->mdev->device,
3571 					"removing port from unaffiliated list.\n");
3572 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
3573 				list_del(&mpi->list);
3574 				break;
3575 			}
3576 		}
3577 		if (!bound) {
3578 			get_port_caps(dev, i + 1);
3579 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
3580 				    i + 1);
3581 		}
3582 	}
3583 
3584 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
3585 	mutex_unlock(&mlx5_ib_multiport_mutex);
3586 	return err;
3587 }
3588 
3589 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
3590 {
3591 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3592 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3593 							  port_num + 1);
3594 	int i;
3595 
3596 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3597 		return;
3598 
3599 	mutex_lock(&mlx5_ib_multiport_mutex);
3600 	for (i = 0; i < dev->num_ports; i++) {
3601 		if (dev->port[i].mp.mpi) {
3602 			/* Destroy the native port stub */
3603 			if (i == port_num) {
3604 				kfree(dev->port[i].mp.mpi);
3605 				dev->port[i].mp.mpi = NULL;
3606 			} else {
3607 				mlx5_ib_dbg(dev, "unbinding port_num: %d\n", i + 1);
3608 				mlx5_ib_unbind_slave_port(dev, dev->port[i].mp.mpi);
3609 			}
3610 		}
3611 	}
3612 
3613 	mlx5_ib_dbg(dev, "removing from devlist\n");
3614 	list_del(&dev->ib_dev_list);
3615 	mutex_unlock(&mlx5_ib_multiport_mutex);
3616 
3617 	mlx5_nic_vport_disable_roce(dev->mdev);
3618 }
3619 
3620 static int mmap_obj_cleanup(struct ib_uobject *uobject,
3621 			    enum rdma_remove_reason why,
3622 			    struct uverbs_attr_bundle *attrs)
3623 {
3624 	struct mlx5_user_mmap_entry *obj = uobject->object;
3625 
3626 	rdma_user_mmap_entry_remove(&obj->rdma_entry);
3627 	return 0;
3628 }
3629 
3630 static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c,
3631 					    struct mlx5_user_mmap_entry *entry,
3632 					    size_t length)
3633 {
3634 	return rdma_user_mmap_entry_insert_range(
3635 		&c->ibucontext, &entry->rdma_entry, length,
3636 		(MLX5_IB_MMAP_OFFSET_START << 16),
3637 		((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1));
3638 }
3639 
3640 static struct mlx5_user_mmap_entry *
3641 alloc_var_entry(struct mlx5_ib_ucontext *c)
3642 {
3643 	struct mlx5_user_mmap_entry *entry;
3644 	struct mlx5_var_table *var_table;
3645 	u32 page_idx;
3646 	int err;
3647 
3648 	var_table = &to_mdev(c->ibucontext.device)->var_table;
3649 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3650 	if (!entry)
3651 		return ERR_PTR(-ENOMEM);
3652 
3653 	mutex_lock(&var_table->bitmap_lock);
3654 	page_idx = find_first_zero_bit(var_table->bitmap,
3655 				       var_table->num_var_hw_entries);
3656 	if (page_idx >= var_table->num_var_hw_entries) {
3657 		err = -ENOSPC;
3658 		mutex_unlock(&var_table->bitmap_lock);
3659 		goto end;
3660 	}
3661 
3662 	set_bit(page_idx, var_table->bitmap);
3663 	mutex_unlock(&var_table->bitmap_lock);
3664 
3665 	entry->address = var_table->hw_start_addr +
3666 				(page_idx * var_table->stride_size);
3667 	entry->page_idx = page_idx;
3668 	entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
3669 
3670 	err = mlx5_rdma_user_mmap_entry_insert(c, entry,
3671 					       var_table->stride_size);
3672 	if (err)
3673 		goto err_insert;
3674 
3675 	return entry;
3676 
3677 err_insert:
3678 	mutex_lock(&var_table->bitmap_lock);
3679 	clear_bit(page_idx, var_table->bitmap);
3680 	mutex_unlock(&var_table->bitmap_lock);
3681 end:
3682 	kfree(entry);
3683 	return ERR_PTR(err);
3684 }
3685 
3686 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
3687 	struct uverbs_attr_bundle *attrs)
3688 {
3689 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3690 		attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3691 	struct mlx5_ib_ucontext *c;
3692 	struct mlx5_user_mmap_entry *entry;
3693 	u64 mmap_offset;
3694 	u32 length;
3695 	int err;
3696 
3697 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3698 	if (IS_ERR(c))
3699 		return PTR_ERR(c);
3700 
3701 	entry = alloc_var_entry(c);
3702 	if (IS_ERR(entry))
3703 		return PTR_ERR(entry);
3704 
3705 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3706 	length = entry->rdma_entry.npages * PAGE_SIZE;
3707 	uobj->object = entry;
3708 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3709 
3710 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3711 			     &mmap_offset, sizeof(mmap_offset));
3712 	if (err)
3713 		return err;
3714 
3715 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3716 			     &entry->page_idx, sizeof(entry->page_idx));
3717 	if (err)
3718 		return err;
3719 
3720 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3721 			     &length, sizeof(length));
3722 	return err;
3723 }
3724 
3725 DECLARE_UVERBS_NAMED_METHOD(
3726 	MLX5_IB_METHOD_VAR_OBJ_ALLOC,
3727 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
3728 			MLX5_IB_OBJECT_VAR,
3729 			UVERBS_ACCESS_NEW,
3730 			UA_MANDATORY),
3731 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3732 			   UVERBS_ATTR_TYPE(u32),
3733 			   UA_MANDATORY),
3734 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3735 			   UVERBS_ATTR_TYPE(u32),
3736 			   UA_MANDATORY),
3737 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3738 			    UVERBS_ATTR_TYPE(u64),
3739 			    UA_MANDATORY));
3740 
3741 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3742 	MLX5_IB_METHOD_VAR_OBJ_DESTROY,
3743 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
3744 			MLX5_IB_OBJECT_VAR,
3745 			UVERBS_ACCESS_DESTROY,
3746 			UA_MANDATORY));
3747 
3748 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
3749 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3750 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
3751 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
3752 
3753 static bool var_is_supported(struct ib_device *device)
3754 {
3755 	struct mlx5_ib_dev *dev = to_mdev(device);
3756 
3757 	return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3758 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
3759 }
3760 
3761 static struct mlx5_user_mmap_entry *
3762 alloc_uar_entry(struct mlx5_ib_ucontext *c,
3763 		enum mlx5_ib_uapi_uar_alloc_type alloc_type)
3764 {
3765 	struct mlx5_user_mmap_entry *entry;
3766 	struct mlx5_ib_dev *dev;
3767 	u32 uar_index;
3768 	int err;
3769 
3770 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3771 	if (!entry)
3772 		return ERR_PTR(-ENOMEM);
3773 
3774 	dev = to_mdev(c->ibucontext.device);
3775 	err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
3776 	if (err)
3777 		goto end;
3778 
3779 	entry->page_idx = uar_index;
3780 	entry->address = uar_index2paddress(dev, uar_index);
3781 	if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3782 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC;
3783 	else
3784 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC;
3785 
3786 	err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE);
3787 	if (err)
3788 		goto err_insert;
3789 
3790 	return entry;
3791 
3792 err_insert:
3793 	mlx5_cmd_free_uar(dev->mdev, uar_index);
3794 end:
3795 	kfree(entry);
3796 	return ERR_PTR(err);
3797 }
3798 
3799 static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)(
3800 	struct uverbs_attr_bundle *attrs)
3801 {
3802 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3803 		attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3804 	enum mlx5_ib_uapi_uar_alloc_type alloc_type;
3805 	struct mlx5_ib_ucontext *c;
3806 	struct mlx5_user_mmap_entry *entry;
3807 	u64 mmap_offset;
3808 	u32 length;
3809 	int err;
3810 
3811 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3812 	if (IS_ERR(c))
3813 		return PTR_ERR(c);
3814 
3815 	err = uverbs_get_const(&alloc_type, attrs,
3816 			       MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE);
3817 	if (err)
3818 		return err;
3819 
3820 	if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF &&
3821 	    alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC)
3822 		return -EOPNOTSUPP;
3823 
3824 	if (!to_mdev(c->ibucontext.device)->wc_support &&
3825 	    alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3826 		return -EOPNOTSUPP;
3827 
3828 	entry = alloc_uar_entry(c, alloc_type);
3829 	if (IS_ERR(entry))
3830 		return PTR_ERR(entry);
3831 
3832 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3833 	length = entry->rdma_entry.npages * PAGE_SIZE;
3834 	uobj->object = entry;
3835 	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3836 
3837 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3838 			     &mmap_offset, sizeof(mmap_offset));
3839 	if (err)
3840 		return err;
3841 
3842 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3843 			     &entry->page_idx, sizeof(entry->page_idx));
3844 	if (err)
3845 		return err;
3846 
3847 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3848 			     &length, sizeof(length));
3849 	return err;
3850 }
3851 
3852 DECLARE_UVERBS_NAMED_METHOD(
3853 	MLX5_IB_METHOD_UAR_OBJ_ALLOC,
3854 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE,
3855 			MLX5_IB_OBJECT_UAR,
3856 			UVERBS_ACCESS_NEW,
3857 			UA_MANDATORY),
3858 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE,
3859 			     enum mlx5_ib_uapi_uar_alloc_type,
3860 			     UA_MANDATORY),
3861 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3862 			   UVERBS_ATTR_TYPE(u32),
3863 			   UA_MANDATORY),
3864 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3865 			   UVERBS_ATTR_TYPE(u32),
3866 			   UA_MANDATORY),
3867 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3868 			    UVERBS_ATTR_TYPE(u64),
3869 			    UA_MANDATORY));
3870 
3871 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3872 	MLX5_IB_METHOD_UAR_OBJ_DESTROY,
3873 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE,
3874 			MLX5_IB_OBJECT_UAR,
3875 			UVERBS_ACCESS_DESTROY,
3876 			UA_MANDATORY));
3877 
3878 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR,
3879 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3880 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC),
3881 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY));
3882 
3883 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3884 	mlx5_ib_dm,
3885 	UVERBS_OBJECT_DM,
3886 	UVERBS_METHOD_DM_ALLOC,
3887 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
3888 			    UVERBS_ATTR_TYPE(u64),
3889 			    UA_MANDATORY),
3890 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
3891 			    UVERBS_ATTR_TYPE(u16),
3892 			    UA_OPTIONAL),
3893 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_ALLOC_DM_REQ_TYPE,
3894 			     enum mlx5_ib_uapi_dm_type,
3895 			     UA_OPTIONAL));
3896 
3897 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3898 	mlx5_ib_flow_action,
3899 	UVERBS_OBJECT_FLOW_ACTION,
3900 	UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
3901 	UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
3902 			     enum mlx5_ib_uapi_flow_action_flags));
3903 
3904 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3905 	mlx5_ib_query_context,
3906 	UVERBS_OBJECT_DEVICE,
3907 	UVERBS_METHOD_QUERY_CONTEXT,
3908 	UVERBS_ATTR_PTR_OUT(
3909 		MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
3910 		UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp,
3911 				   dump_fill_mkey),
3912 		UA_MANDATORY));
3913 
3914 static const struct uapi_definition mlx5_ib_defs[] = {
3915 	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
3916 	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
3917 	UAPI_DEF_CHAIN(mlx5_ib_qos_defs),
3918 	UAPI_DEF_CHAIN(mlx5_ib_std_types_defs),
3919 
3920 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_FLOW_ACTION,
3921 				&mlx5_ib_flow_action),
3922 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DM, &mlx5_ib_dm),
3923 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context),
3924 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
3925 				UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
3926 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR),
3927 	{}
3928 };
3929 
3930 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
3931 {
3932 	mlx5_ib_cleanup_multiport_master(dev);
3933 	WARN_ON(!xa_empty(&dev->odp_mkeys));
3934 	cleanup_srcu_struct(&dev->odp_srcu);
3935 
3936 	WARN_ON(!xa_empty(&dev->sig_mrs));
3937 	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
3938 }
3939 
3940 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
3941 {
3942 	struct mlx5_core_dev *mdev = dev->mdev;
3943 	int err;
3944 	int i;
3945 
3946 	for (i = 0; i < dev->num_ports; i++) {
3947 		spin_lock_init(&dev->port[i].mp.mpi_lock);
3948 		rwlock_init(&dev->port[i].roce.netdev_lock);
3949 		dev->port[i].roce.dev = dev;
3950 		dev->port[i].roce.native_port_num = i + 1;
3951 		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
3952 	}
3953 
3954 	mlx5_ib_internal_fill_odp_caps(dev);
3955 
3956 	err = mlx5_ib_init_multiport_master(dev);
3957 	if (err)
3958 		return err;
3959 
3960 	err = set_has_smi_cap(dev);
3961 	if (err)
3962 		return err;
3963 
3964 	if (!mlx5_core_mp_enabled(mdev)) {
3965 		for (i = 1; i <= dev->num_ports; i++) {
3966 			err = get_port_caps(dev, i);
3967 			if (err)
3968 				break;
3969 		}
3970 	} else {
3971 		err = get_port_caps(dev, mlx5_core_native_port_num(mdev));
3972 	}
3973 	if (err)
3974 		goto err_mp;
3975 
3976 	if (mlx5_use_mad_ifc(dev))
3977 		get_ext_port_caps(dev);
3978 
3979 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
3980 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
3981 	dev->ib_dev.phys_port_cnt	= dev->num_ports;
3982 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_count(mdev);
3983 	dev->ib_dev.dev.parent		= mdev->device;
3984 	dev->ib_dev.lag_flags		= RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
3985 
3986 	mutex_init(&dev->cap_mask_mutex);
3987 	INIT_LIST_HEAD(&dev->qp_list);
3988 	spin_lock_init(&dev->reset_flow_resource_lock);
3989 	xa_init(&dev->odp_mkeys);
3990 	xa_init(&dev->sig_mrs);
3991 	atomic_set(&dev->mkey_var, 0);
3992 
3993 	spin_lock_init(&dev->dm.lock);
3994 	dev->dm.dev = mdev;
3995 
3996 	err = init_srcu_struct(&dev->odp_srcu);
3997 	if (err)
3998 		goto err_mp;
3999 
4000 	return 0;
4001 
4002 err_mp:
4003 	mlx5_ib_cleanup_multiport_master(dev);
4004 
4005 	return -ENOMEM;
4006 }
4007 
4008 static int mlx5_ib_enable_driver(struct ib_device *dev)
4009 {
4010 	struct mlx5_ib_dev *mdev = to_mdev(dev);
4011 	int ret;
4012 
4013 	ret = mlx5_ib_test_wc(mdev);
4014 	mlx5_ib_dbg(mdev, "Write-Combining %s",
4015 		    mdev->wc_support ? "supported" : "not supported");
4016 
4017 	return ret;
4018 }
4019 
4020 static const struct ib_device_ops mlx5_ib_dev_ops = {
4021 	.owner = THIS_MODULE,
4022 	.driver_id = RDMA_DRIVER_MLX5,
4023 	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
4024 
4025 	.add_gid = mlx5_ib_add_gid,
4026 	.alloc_mr = mlx5_ib_alloc_mr,
4027 	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
4028 	.alloc_pd = mlx5_ib_alloc_pd,
4029 	.alloc_ucontext = mlx5_ib_alloc_ucontext,
4030 	.attach_mcast = mlx5_ib_mcg_attach,
4031 	.check_mr_status = mlx5_ib_check_mr_status,
4032 	.create_ah = mlx5_ib_create_ah,
4033 	.create_cq = mlx5_ib_create_cq,
4034 	.create_qp = mlx5_ib_create_qp,
4035 	.create_srq = mlx5_ib_create_srq,
4036 	.dealloc_pd = mlx5_ib_dealloc_pd,
4037 	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
4038 	.del_gid = mlx5_ib_del_gid,
4039 	.dereg_mr = mlx5_ib_dereg_mr,
4040 	.destroy_ah = mlx5_ib_destroy_ah,
4041 	.destroy_cq = mlx5_ib_destroy_cq,
4042 	.destroy_qp = mlx5_ib_destroy_qp,
4043 	.destroy_srq = mlx5_ib_destroy_srq,
4044 	.detach_mcast = mlx5_ib_mcg_detach,
4045 	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
4046 	.drain_rq = mlx5_ib_drain_rq,
4047 	.drain_sq = mlx5_ib_drain_sq,
4048 	.enable_driver = mlx5_ib_enable_driver,
4049 	.get_dev_fw_str = get_dev_fw_str,
4050 	.get_dma_mr = mlx5_ib_get_dma_mr,
4051 	.get_link_layer = mlx5_ib_port_link_layer,
4052 	.map_mr_sg = mlx5_ib_map_mr_sg,
4053 	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
4054 	.mmap = mlx5_ib_mmap,
4055 	.mmap_free = mlx5_ib_mmap_free,
4056 	.modify_cq = mlx5_ib_modify_cq,
4057 	.modify_device = mlx5_ib_modify_device,
4058 	.modify_port = mlx5_ib_modify_port,
4059 	.modify_qp = mlx5_ib_modify_qp,
4060 	.modify_srq = mlx5_ib_modify_srq,
4061 	.poll_cq = mlx5_ib_poll_cq,
4062 	.post_recv = mlx5_ib_post_recv_nodrain,
4063 	.post_send = mlx5_ib_post_send_nodrain,
4064 	.post_srq_recv = mlx5_ib_post_srq_recv,
4065 	.process_mad = mlx5_ib_process_mad,
4066 	.query_ah = mlx5_ib_query_ah,
4067 	.query_device = mlx5_ib_query_device,
4068 	.query_gid = mlx5_ib_query_gid,
4069 	.query_pkey = mlx5_ib_query_pkey,
4070 	.query_qp = mlx5_ib_query_qp,
4071 	.query_srq = mlx5_ib_query_srq,
4072 	.query_ucontext = mlx5_ib_query_ucontext,
4073 	.reg_user_mr = mlx5_ib_reg_user_mr,
4074 	.req_notify_cq = mlx5_ib_arm_cq,
4075 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
4076 	.resize_cq = mlx5_ib_resize_cq,
4077 
4078 	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
4079 	INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs),
4080 	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
4081 	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
4082 	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
4083 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
4084 };
4085 
4086 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
4087 	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
4088 };
4089 
4090 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
4091 	.get_vf_config = mlx5_ib_get_vf_config,
4092 	.get_vf_guid = mlx5_ib_get_vf_guid,
4093 	.get_vf_stats = mlx5_ib_get_vf_stats,
4094 	.set_vf_guid = mlx5_ib_set_vf_guid,
4095 	.set_vf_link_state = mlx5_ib_set_vf_link_state,
4096 };
4097 
4098 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
4099 	.alloc_mw = mlx5_ib_alloc_mw,
4100 	.dealloc_mw = mlx5_ib_dealloc_mw,
4101 };
4102 
4103 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
4104 	.alloc_xrcd = mlx5_ib_alloc_xrcd,
4105 	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
4106 
4107 	INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd),
4108 };
4109 
4110 static const struct ib_device_ops mlx5_ib_dev_dm_ops = {
4111 	.alloc_dm = mlx5_ib_alloc_dm,
4112 	.dealloc_dm = mlx5_ib_dealloc_dm,
4113 	.reg_dm_mr = mlx5_ib_reg_dm_mr,
4114 };
4115 
4116 static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
4117 {
4118 	struct mlx5_core_dev *mdev = dev->mdev;
4119 	struct mlx5_var_table *var_table = &dev->var_table;
4120 	u8 log_doorbell_bar_size;
4121 	u8 log_doorbell_stride;
4122 	u64 bar_size;
4123 
4124 	log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
4125 					log_doorbell_bar_size);
4126 	log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
4127 					log_doorbell_stride);
4128 	var_table->hw_start_addr = dev->mdev->bar_addr +
4129 				MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
4130 					doorbell_bar_offset);
4131 	bar_size = (1ULL << log_doorbell_bar_size) * 4096;
4132 	var_table->stride_size = 1ULL << log_doorbell_stride;
4133 	var_table->num_var_hw_entries = div_u64(bar_size,
4134 						var_table->stride_size);
4135 	mutex_init(&var_table->bitmap_lock);
4136 	var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
4137 					  GFP_KERNEL);
4138 	return (var_table->bitmap) ? 0 : -ENOMEM;
4139 }
4140 
4141 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
4142 {
4143 	bitmap_free(dev->var_table.bitmap);
4144 }
4145 
4146 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
4147 {
4148 	struct mlx5_core_dev *mdev = dev->mdev;
4149 	int err;
4150 
4151 	dev->ib_dev.uverbs_cmd_mask	=
4152 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
4153 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
4154 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
4155 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
4156 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
4157 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
4158 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
4159 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
4160 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
4161 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
4162 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
4163 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
4164 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
4165 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
4166 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
4167 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
4168 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
4169 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
4170 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
4171 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
4172 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
4173 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
4174 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
4175 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
4176 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
4177 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
4178 	dev->ib_dev.uverbs_ex_cmd_mask =
4179 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
4180 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
4181 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP)	|
4182 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP)	|
4183 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_CQ)	|
4184 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW)	|
4185 		(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW);
4186 
4187 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
4188 	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
4189 		ib_set_device_ops(&dev->ib_dev,
4190 				  &mlx5_ib_dev_ipoib_enhanced_ops);
4191 
4192 	if (mlx5_core_is_pf(mdev))
4193 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
4194 
4195 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
4196 
4197 	if (MLX5_CAP_GEN(mdev, imaicl)) {
4198 		dev->ib_dev.uverbs_cmd_mask |=
4199 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
4200 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
4201 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
4202 	}
4203 
4204 	if (MLX5_CAP_GEN(mdev, xrc)) {
4205 		dev->ib_dev.uverbs_cmd_mask |=
4206 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
4207 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
4208 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
4209 	}
4210 
4211 	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
4212 	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
4213 	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
4214 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
4215 
4216 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
4217 
4218 	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
4219 		dev->ib_dev.driver_def = mlx5_ib_defs;
4220 
4221 	err = init_node_data(dev);
4222 	if (err)
4223 		return err;
4224 
4225 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
4226 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
4227 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
4228 		mutex_init(&dev->lb.mutex);
4229 
4230 	if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
4231 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
4232 		err = mlx5_ib_init_var_table(dev);
4233 		if (err)
4234 			return err;
4235 	}
4236 
4237 	dev->ib_dev.use_cq_dim = true;
4238 
4239 	return 0;
4240 }
4241 
4242 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
4243 	.get_port_immutable = mlx5_port_immutable,
4244 	.query_port = mlx5_ib_query_port,
4245 };
4246 
4247 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
4248 {
4249 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
4250 	return 0;
4251 }
4252 
4253 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
4254 	.get_port_immutable = mlx5_port_rep_immutable,
4255 	.query_port = mlx5_ib_rep_query_port,
4256 };
4257 
4258 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
4259 {
4260 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
4261 	return 0;
4262 }
4263 
4264 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
4265 	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
4266 	.create_wq = mlx5_ib_create_wq,
4267 	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
4268 	.destroy_wq = mlx5_ib_destroy_wq,
4269 	.get_netdev = mlx5_ib_get_netdev,
4270 	.modify_wq = mlx5_ib_modify_wq,
4271 };
4272 
4273 static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev)
4274 {
4275 	struct mlx5_core_dev *mdev = dev->mdev;
4276 	enum rdma_link_layer ll;
4277 	int port_type_cap;
4278 	u8 port_num = 0;
4279 	int err;
4280 
4281 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4282 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4283 
4284 	if (ll == IB_LINK_LAYER_ETHERNET) {
4285 		dev->ib_dev.uverbs_ex_cmd_mask |=
4286 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
4287 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
4288 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
4289 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
4290 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
4291 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
4292 
4293 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4294 
4295 		/* Register only for native ports */
4296 		err = mlx5_add_netdev_notifier(dev, port_num);
4297 		if (err || dev->is_rep || !mlx5_is_roce_enabled(mdev))
4298 			/*
4299 			 * We don't enable ETH interface for
4300 			 * 1. IB representors
4301 			 * 2. User disabled ROCE through devlink interface
4302 			 */
4303 			return err;
4304 
4305 		err = mlx5_enable_eth(dev);
4306 		if (err)
4307 			goto cleanup;
4308 	}
4309 
4310 	return 0;
4311 cleanup:
4312 	mlx5_remove_netdev_notifier(dev, port_num);
4313 	return err;
4314 }
4315 
4316 static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev)
4317 {
4318 	struct mlx5_core_dev *mdev = dev->mdev;
4319 	enum rdma_link_layer ll;
4320 	int port_type_cap;
4321 	u8 port_num;
4322 
4323 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4324 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4325 
4326 	if (ll == IB_LINK_LAYER_ETHERNET) {
4327 		if (!dev->is_rep)
4328 			mlx5_disable_eth(dev);
4329 
4330 		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4331 		mlx5_remove_netdev_notifier(dev, port_num);
4332 	}
4333 }
4334 
4335 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
4336 {
4337 	mlx5_ib_init_cong_debugfs(dev,
4338 				  mlx5_core_native_port_num(dev->mdev) - 1);
4339 	return 0;
4340 }
4341 
4342 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
4343 {
4344 	mlx5_ib_cleanup_cong_debugfs(dev,
4345 				     mlx5_core_native_port_num(dev->mdev) - 1);
4346 }
4347 
4348 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
4349 {
4350 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
4351 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
4352 }
4353 
4354 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
4355 {
4356 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
4357 }
4358 
4359 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
4360 {
4361 	int err;
4362 
4363 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
4364 	if (err)
4365 		return err;
4366 
4367 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
4368 	if (err)
4369 		mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4370 
4371 	return err;
4372 }
4373 
4374 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
4375 {
4376 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4377 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4378 }
4379 
4380 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
4381 {
4382 	const char *name;
4383 
4384 	rdma_set_device_sysfs_group(&dev->ib_dev, &mlx5_attr_group);
4385 	if (!mlx5_lag_is_roce(dev->mdev))
4386 		name = "mlx5_%d";
4387 	else
4388 		name = "mlx5_bond_%d";
4389 	return ib_register_device(&dev->ib_dev, name);
4390 }
4391 
4392 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
4393 {
4394 	int err;
4395 
4396 	err = mlx5_mr_cache_cleanup(dev);
4397 	if (err)
4398 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
4399 
4400 	if (dev->umrc.qp)
4401 		mlx5_ib_destroy_qp(dev->umrc.qp, NULL);
4402 	if (dev->umrc.cq)
4403 		ib_free_cq(dev->umrc.cq);
4404 	if (dev->umrc.pd)
4405 		ib_dealloc_pd(dev->umrc.pd);
4406 }
4407 
4408 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
4409 {
4410 	ib_unregister_device(&dev->ib_dev);
4411 }
4412 
4413 enum {
4414 	MAX_UMR_WR = 128,
4415 };
4416 
4417 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
4418 {
4419 	struct ib_qp_init_attr *init_attr = NULL;
4420 	struct ib_qp_attr *attr = NULL;
4421 	struct ib_pd *pd;
4422 	struct ib_cq *cq;
4423 	struct ib_qp *qp;
4424 	int ret;
4425 
4426 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
4427 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
4428 	if (!attr || !init_attr) {
4429 		ret = -ENOMEM;
4430 		goto error_0;
4431 	}
4432 
4433 	pd = ib_alloc_pd(&dev->ib_dev, 0);
4434 	if (IS_ERR(pd)) {
4435 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
4436 		ret = PTR_ERR(pd);
4437 		goto error_0;
4438 	}
4439 
4440 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
4441 	if (IS_ERR(cq)) {
4442 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
4443 		ret = PTR_ERR(cq);
4444 		goto error_2;
4445 	}
4446 
4447 	init_attr->send_cq = cq;
4448 	init_attr->recv_cq = cq;
4449 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
4450 	init_attr->cap.max_send_wr = MAX_UMR_WR;
4451 	init_attr->cap.max_send_sge = 1;
4452 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
4453 	init_attr->port_num = 1;
4454 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
4455 	if (IS_ERR(qp)) {
4456 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
4457 		ret = PTR_ERR(qp);
4458 		goto error_3;
4459 	}
4460 	qp->device     = &dev->ib_dev;
4461 	qp->real_qp    = qp;
4462 	qp->uobject    = NULL;
4463 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
4464 	qp->send_cq    = init_attr->send_cq;
4465 	qp->recv_cq    = init_attr->recv_cq;
4466 
4467 	attr->qp_state = IB_QPS_INIT;
4468 	attr->port_num = 1;
4469 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
4470 				IB_QP_PORT, NULL);
4471 	if (ret) {
4472 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
4473 		goto error_4;
4474 	}
4475 
4476 	memset(attr, 0, sizeof(*attr));
4477 	attr->qp_state = IB_QPS_RTR;
4478 	attr->path_mtu = IB_MTU_256;
4479 
4480 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4481 	if (ret) {
4482 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
4483 		goto error_4;
4484 	}
4485 
4486 	memset(attr, 0, sizeof(*attr));
4487 	attr->qp_state = IB_QPS_RTS;
4488 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4489 	if (ret) {
4490 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
4491 		goto error_4;
4492 	}
4493 
4494 	dev->umrc.qp = qp;
4495 	dev->umrc.cq = cq;
4496 	dev->umrc.pd = pd;
4497 
4498 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
4499 	ret = mlx5_mr_cache_init(dev);
4500 	if (ret) {
4501 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4502 		goto error_4;
4503 	}
4504 
4505 	kfree(attr);
4506 	kfree(init_attr);
4507 
4508 	return 0;
4509 
4510 error_4:
4511 	mlx5_ib_destroy_qp(qp, NULL);
4512 	dev->umrc.qp = NULL;
4513 
4514 error_3:
4515 	ib_free_cq(cq);
4516 	dev->umrc.cq = NULL;
4517 
4518 error_2:
4519 	ib_dealloc_pd(pd);
4520 	dev->umrc.pd = NULL;
4521 
4522 error_0:
4523 	kfree(attr);
4524 	kfree(init_attr);
4525 	return ret;
4526 }
4527 
4528 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
4529 {
4530 	struct dentry *root;
4531 
4532 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4533 		return 0;
4534 
4535 	mutex_init(&dev->delay_drop.lock);
4536 	dev->delay_drop.dev = dev;
4537 	dev->delay_drop.activate = false;
4538 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4539 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4540 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4541 	atomic_set(&dev->delay_drop.events_cnt, 0);
4542 
4543 	if (!mlx5_debugfs_root)
4544 		return 0;
4545 
4546 	root = debugfs_create_dir("delay_drop", dev->mdev->priv.dbg_root);
4547 	dev->delay_drop.dir_debugfs = root;
4548 
4549 	debugfs_create_atomic_t("num_timeout_events", 0400, root,
4550 				&dev->delay_drop.events_cnt);
4551 	debugfs_create_atomic_t("num_rqs", 0400, root,
4552 				&dev->delay_drop.rqs_cnt);
4553 	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
4554 			    &fops_delay_drop_timeout);
4555 	return 0;
4556 }
4557 
4558 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
4559 {
4560 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4561 		return;
4562 
4563 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4564 	if (!dev->delay_drop.dir_debugfs)
4565 		return;
4566 
4567 	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
4568 	dev->delay_drop.dir_debugfs = NULL;
4569 }
4570 
4571 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
4572 {
4573 	dev->mdev_events.notifier_call = mlx5_ib_event;
4574 	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
4575 	return 0;
4576 }
4577 
4578 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
4579 {
4580 	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
4581 }
4582 
4583 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
4584 		      const struct mlx5_ib_profile *profile,
4585 		      int stage)
4586 {
4587 	dev->ib_active = false;
4588 
4589 	/* Number of stages to cleanup */
4590 	while (stage) {
4591 		stage--;
4592 		if (profile->stage[stage].cleanup)
4593 			profile->stage[stage].cleanup(dev);
4594 	}
4595 
4596 	kfree(dev->port);
4597 	ib_dealloc_device(&dev->ib_dev);
4598 }
4599 
4600 void *__mlx5_ib_add(struct mlx5_ib_dev *dev,
4601 		    const struct mlx5_ib_profile *profile)
4602 {
4603 	int err;
4604 	int i;
4605 
4606 	dev->profile = profile;
4607 
4608 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
4609 		if (profile->stage[i].init) {
4610 			err = profile->stage[i].init(dev);
4611 			if (err)
4612 				goto err_out;
4613 		}
4614 	}
4615 
4616 	dev->ib_active = true;
4617 
4618 	return dev;
4619 
4620 err_out:
4621 	__mlx5_ib_remove(dev, profile, i);
4622 
4623 	return NULL;
4624 }
4625 
4626 static const struct mlx5_ib_profile pf_profile = {
4627 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4628 		     mlx5_ib_stage_init_init,
4629 		     mlx5_ib_stage_init_cleanup),
4630 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4631 		     mlx5_ib_fs_init,
4632 		     mlx5_ib_fs_cleanup),
4633 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4634 		     mlx5_ib_stage_caps_init,
4635 		     mlx5_ib_stage_caps_cleanup),
4636 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4637 		     mlx5_ib_stage_non_default_cb,
4638 		     NULL),
4639 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4640 		     mlx5_ib_roce_init,
4641 		     mlx5_ib_roce_cleanup),
4642 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4643 		     mlx5_init_qp_table,
4644 		     mlx5_cleanup_qp_table),
4645 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4646 		     mlx5_init_srq_table,
4647 		     mlx5_cleanup_srq_table),
4648 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4649 		     mlx5_ib_dev_res_init,
4650 		     mlx5_ib_dev_res_cleanup),
4651 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4652 		     mlx5_ib_stage_dev_notifier_init,
4653 		     mlx5_ib_stage_dev_notifier_cleanup),
4654 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
4655 		     mlx5_ib_odp_init_one,
4656 		     mlx5_ib_odp_cleanup_one),
4657 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4658 		     mlx5_ib_counters_init,
4659 		     mlx5_ib_counters_cleanup),
4660 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4661 		     mlx5_ib_stage_cong_debugfs_init,
4662 		     mlx5_ib_stage_cong_debugfs_cleanup),
4663 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4664 		     mlx5_ib_stage_uar_init,
4665 		     mlx5_ib_stage_uar_cleanup),
4666 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4667 		     mlx5_ib_stage_bfrag_init,
4668 		     mlx5_ib_stage_bfrag_cleanup),
4669 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4670 		     NULL,
4671 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4672 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4673 		     mlx5_ib_devx_init,
4674 		     mlx5_ib_devx_cleanup),
4675 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4676 		     mlx5_ib_stage_ib_reg_init,
4677 		     mlx5_ib_stage_ib_reg_cleanup),
4678 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4679 		     mlx5_ib_stage_post_ib_reg_umr_init,
4680 		     NULL),
4681 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4682 		     mlx5_ib_stage_delay_drop_init,
4683 		     mlx5_ib_stage_delay_drop_cleanup),
4684 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4685 		     mlx5_ib_restrack_init,
4686 		     NULL),
4687 };
4688 
4689 const struct mlx5_ib_profile raw_eth_profile = {
4690 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4691 		     mlx5_ib_stage_init_init,
4692 		     mlx5_ib_stage_init_cleanup),
4693 	STAGE_CREATE(MLX5_IB_STAGE_FS,
4694 		     mlx5_ib_fs_init,
4695 		     mlx5_ib_fs_cleanup),
4696 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4697 		     mlx5_ib_stage_caps_init,
4698 		     mlx5_ib_stage_caps_cleanup),
4699 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4700 		     mlx5_ib_stage_raw_eth_non_default_cb,
4701 		     NULL),
4702 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4703 		     mlx5_ib_roce_init,
4704 		     mlx5_ib_roce_cleanup),
4705 	STAGE_CREATE(MLX5_IB_STAGE_QP,
4706 		     mlx5_init_qp_table,
4707 		     mlx5_cleanup_qp_table),
4708 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4709 		     mlx5_init_srq_table,
4710 		     mlx5_cleanup_srq_table),
4711 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4712 		     mlx5_ib_dev_res_init,
4713 		     mlx5_ib_dev_res_cleanup),
4714 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4715 		     mlx5_ib_stage_dev_notifier_init,
4716 		     mlx5_ib_stage_dev_notifier_cleanup),
4717 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4718 		     mlx5_ib_counters_init,
4719 		     mlx5_ib_counters_cleanup),
4720 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4721 		     mlx5_ib_stage_cong_debugfs_init,
4722 		     mlx5_ib_stage_cong_debugfs_cleanup),
4723 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4724 		     mlx5_ib_stage_uar_init,
4725 		     mlx5_ib_stage_uar_cleanup),
4726 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4727 		     mlx5_ib_stage_bfrag_init,
4728 		     mlx5_ib_stage_bfrag_cleanup),
4729 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4730 		     NULL,
4731 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4732 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4733 		     mlx5_ib_devx_init,
4734 		     mlx5_ib_devx_cleanup),
4735 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4736 		     mlx5_ib_stage_ib_reg_init,
4737 		     mlx5_ib_stage_ib_reg_cleanup),
4738 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4739 		     mlx5_ib_stage_post_ib_reg_umr_init,
4740 		     NULL),
4741 	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4742 		     mlx5_ib_restrack_init,
4743 		     NULL),
4744 };
4745 
4746 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev)
4747 {
4748 	struct mlx5_ib_multiport_info *mpi;
4749 	struct mlx5_ib_dev *dev;
4750 	bool bound = false;
4751 	int err;
4752 
4753 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4754 	if (!mpi)
4755 		return NULL;
4756 
4757 	mpi->mdev = mdev;
4758 
4759 	err = mlx5_query_nic_vport_system_image_guid(mdev,
4760 						     &mpi->sys_image_guid);
4761 	if (err) {
4762 		kfree(mpi);
4763 		return NULL;
4764 	}
4765 
4766 	mutex_lock(&mlx5_ib_multiport_mutex);
4767 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
4768 		if (dev->sys_image_guid == mpi->sys_image_guid)
4769 			bound = mlx5_ib_bind_slave_port(dev, mpi);
4770 
4771 		if (bound) {
4772 			rdma_roce_rescan_device(&dev->ib_dev);
4773 			break;
4774 		}
4775 	}
4776 
4777 	if (!bound) {
4778 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4779 		dev_dbg(mdev->device,
4780 			"no suitable IB device found to bind to, added to unaffiliated list.\n");
4781 	}
4782 	mutex_unlock(&mlx5_ib_multiport_mutex);
4783 
4784 	return mpi;
4785 }
4786 
4787 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
4788 {
4789 	const struct mlx5_ib_profile *profile;
4790 	enum rdma_link_layer ll;
4791 	struct mlx5_ib_dev *dev;
4792 	int port_type_cap;
4793 	int num_ports;
4794 
4795 	if (MLX5_ESWITCH_MANAGER(mdev) &&
4796 	    mlx5_ib_eswitch_mode(mdev->priv.eswitch) == MLX5_ESWITCH_OFFLOADS) {
4797 		if (!mlx5_core_mp_enabled(mdev))
4798 			mlx5_ib_register_vport_reps(mdev);
4799 		return mdev;
4800 	}
4801 
4802 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4803 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4804 
4805 	if (mlx5_core_is_mp_slave(mdev) && ll == IB_LINK_LAYER_ETHERNET)
4806 		return mlx5_ib_add_slave_port(mdev);
4807 
4808 	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
4809 			MLX5_CAP_GEN(mdev, num_vhca_ports));
4810 	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
4811 	if (!dev)
4812 		return NULL;
4813 	dev->port = kcalloc(num_ports, sizeof(*dev->port),
4814 			     GFP_KERNEL);
4815 	if (!dev->port) {
4816 		ib_dealloc_device(&dev->ib_dev);
4817 		return NULL;
4818 	}
4819 
4820 	dev->mdev = mdev;
4821 	dev->num_ports = num_ports;
4822 
4823 	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_is_roce_enabled(mdev))
4824 		profile = &raw_eth_profile;
4825 	else
4826 		profile = &pf_profile;
4827 
4828 	return __mlx5_ib_add(dev, profile);
4829 }
4830 
4831 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
4832 {
4833 	struct mlx5_ib_multiport_info *mpi;
4834 	struct mlx5_ib_dev *dev;
4835 
4836 	if (MLX5_ESWITCH_MANAGER(mdev) && context == mdev) {
4837 		mlx5_ib_unregister_vport_reps(mdev);
4838 		return;
4839 	}
4840 
4841 	if (mlx5_core_is_mp_slave(mdev)) {
4842 		mpi = context;
4843 		mutex_lock(&mlx5_ib_multiport_mutex);
4844 		if (mpi->ibdev)
4845 			mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
4846 		list_del(&mpi->list);
4847 		mutex_unlock(&mlx5_ib_multiport_mutex);
4848 		kfree(mpi);
4849 		return;
4850 	}
4851 
4852 	dev = context;
4853 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
4854 }
4855 
4856 static struct mlx5_interface mlx5_ib_interface = {
4857 	.add            = mlx5_ib_add,
4858 	.remove         = mlx5_ib_remove,
4859 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
4860 };
4861 
4862 unsigned long mlx5_ib_get_xlt_emergency_page(void)
4863 {
4864 	mutex_lock(&xlt_emergency_page_mutex);
4865 	return xlt_emergency_page;
4866 }
4867 
4868 void mlx5_ib_put_xlt_emergency_page(void)
4869 {
4870 	mutex_unlock(&xlt_emergency_page_mutex);
4871 }
4872 
4873 static int __init mlx5_ib_init(void)
4874 {
4875 	int err;
4876 
4877 	xlt_emergency_page = __get_free_page(GFP_KERNEL);
4878 	if (!xlt_emergency_page)
4879 		return -ENOMEM;
4880 
4881 	mutex_init(&xlt_emergency_page_mutex);
4882 
4883 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
4884 	if (!mlx5_ib_event_wq) {
4885 		free_page(xlt_emergency_page);
4886 		return -ENOMEM;
4887 	}
4888 
4889 	mlx5_ib_odp_init();
4890 
4891 	err = mlx5_register_interface(&mlx5_ib_interface);
4892 
4893 	return err;
4894 }
4895 
4896 static void __exit mlx5_ib_cleanup(void)
4897 {
4898 	mlx5_unregister_interface(&mlx5_ib_interface);
4899 	destroy_workqueue(mlx5_ib_event_wq);
4900 	mutex_destroy(&xlt_emergency_page_mutex);
4901 	free_page(xlt_emergency_page);
4902 }
4903 
4904 module_init(mlx5_ib_init);
4905 module_exit(mlx5_ib_cleanup);
4906