xref: /linux/drivers/infiniband/hw/mlx5/main.c (revision 165f2d2858013253042809df082b8df7e34e86d7)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/debugfs.h>
34 #include <linux/highmem.h>
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/errno.h>
38 #include <linux/pci.h>
39 #include <linux/dma-mapping.h>
40 #include <linux/slab.h>
41 #include <linux/bitmap.h>
42 #include <linux/sched.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/task.h>
45 #include <linux/delay.h>
46 #include <rdma/ib_user_verbs.h>
47 #include <rdma/ib_addr.h>
48 #include <rdma/ib_cache.h>
49 #include <linux/mlx5/port.h>
50 #include <linux/mlx5/vport.h>
51 #include <linux/mlx5/fs.h>
52 #include <linux/mlx5/eswitch.h>
53 #include <linux/list.h>
54 #include <rdma/ib_smi.h>
55 #include <rdma/ib_umem.h>
56 #include <linux/in.h>
57 #include <linux/etherdevice.h>
58 #include "mlx5_ib.h"
59 #include "ib_rep.h"
60 #include "cmd.h"
61 #include "srq.h"
62 #include <linux/mlx5/fs_helpers.h>
63 #include <linux/mlx5/accel.h>
64 #include <rdma/uverbs_std_types.h>
65 #include <rdma/mlx5_user_ioctl_verbs.h>
66 #include <rdma/mlx5_user_ioctl_cmds.h>
67 #include <rdma/ib_umem_odp.h>
68 
69 #define UVERBS_MODULE_NAME mlx5_ib
70 #include <rdma/uverbs_named_ioctl.h>
71 
72 #define DRIVER_NAME "mlx5_ib"
73 #define DRIVER_VERSION "5.0-0"
74 
75 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
76 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
77 MODULE_LICENSE("Dual BSD/GPL");
78 
79 static char mlx5_version[] =
80 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
81 	DRIVER_VERSION "\n";
82 
83 struct mlx5_ib_event_work {
84 	struct work_struct	work;
85 	union {
86 		struct mlx5_ib_dev	      *dev;
87 		struct mlx5_ib_multiport_info *mpi;
88 	};
89 	bool			is_slave;
90 	unsigned int		event;
91 	void			*param;
92 };
93 
94 enum {
95 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
96 };
97 
98 static struct workqueue_struct *mlx5_ib_event_wq;
99 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
100 static LIST_HEAD(mlx5_ib_dev_list);
101 /*
102  * This mutex should be held when accessing either of the above lists
103  */
104 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
105 
106 /* We can't use an array for xlt_emergency_page because dma_map_single
107  * doesn't work on kernel modules memory
108  */
109 static unsigned long xlt_emergency_page;
110 static struct mutex xlt_emergency_page_mutex;
111 
112 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
113 {
114 	struct mlx5_ib_dev *dev;
115 
116 	mutex_lock(&mlx5_ib_multiport_mutex);
117 	dev = mpi->ibdev;
118 	mutex_unlock(&mlx5_ib_multiport_mutex);
119 	return dev;
120 }
121 
122 static enum rdma_link_layer
123 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
124 {
125 	switch (port_type_cap) {
126 	case MLX5_CAP_PORT_TYPE_IB:
127 		return IB_LINK_LAYER_INFINIBAND;
128 	case MLX5_CAP_PORT_TYPE_ETH:
129 		return IB_LINK_LAYER_ETHERNET;
130 	default:
131 		return IB_LINK_LAYER_UNSPECIFIED;
132 	}
133 }
134 
135 static enum rdma_link_layer
136 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
137 {
138 	struct mlx5_ib_dev *dev = to_mdev(device);
139 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
140 
141 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
142 }
143 
144 static int get_port_state(struct ib_device *ibdev,
145 			  u8 port_num,
146 			  enum ib_port_state *state)
147 {
148 	struct ib_port_attr attr;
149 	int ret;
150 
151 	memset(&attr, 0, sizeof(attr));
152 	ret = ibdev->ops.query_port(ibdev, port_num, &attr);
153 	if (!ret)
154 		*state = attr.state;
155 	return ret;
156 }
157 
158 static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev,
159 					   struct net_device *ndev,
160 					   u8 *port_num)
161 {
162 	struct mlx5_eswitch *esw = dev->mdev->priv.eswitch;
163 	struct net_device *rep_ndev;
164 	struct mlx5_ib_port *port;
165 	int i;
166 
167 	for (i = 0; i < dev->num_ports; i++) {
168 		port  = &dev->port[i];
169 		if (!port->rep)
170 			continue;
171 
172 		read_lock(&port->roce.netdev_lock);
173 		rep_ndev = mlx5_ib_get_rep_netdev(esw,
174 						  port->rep->vport);
175 		if (rep_ndev == ndev) {
176 			read_unlock(&port->roce.netdev_lock);
177 			*port_num = i + 1;
178 			return &port->roce;
179 		}
180 		read_unlock(&port->roce.netdev_lock);
181 	}
182 
183 	return NULL;
184 }
185 
186 static int mlx5_netdev_event(struct notifier_block *this,
187 			     unsigned long event, void *ptr)
188 {
189 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
190 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
191 	u8 port_num = roce->native_port_num;
192 	struct mlx5_core_dev *mdev;
193 	struct mlx5_ib_dev *ibdev;
194 
195 	ibdev = roce->dev;
196 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
197 	if (!mdev)
198 		return NOTIFY_DONE;
199 
200 	switch (event) {
201 	case NETDEV_REGISTER:
202 		/* Should already be registered during the load */
203 		if (ibdev->is_rep)
204 			break;
205 		write_lock(&roce->netdev_lock);
206 		if (ndev->dev.parent == mdev->device)
207 			roce->netdev = ndev;
208 		write_unlock(&roce->netdev_lock);
209 		break;
210 
211 	case NETDEV_UNREGISTER:
212 		/* In case of reps, ib device goes away before the netdevs */
213 		write_lock(&roce->netdev_lock);
214 		if (roce->netdev == ndev)
215 			roce->netdev = NULL;
216 		write_unlock(&roce->netdev_lock);
217 		break;
218 
219 	case NETDEV_CHANGE:
220 	case NETDEV_UP:
221 	case NETDEV_DOWN: {
222 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
223 		struct net_device *upper = NULL;
224 
225 		if (lag_ndev) {
226 			upper = netdev_master_upper_dev_get(lag_ndev);
227 			dev_put(lag_ndev);
228 		}
229 
230 		if (ibdev->is_rep)
231 			roce = mlx5_get_rep_roce(ibdev, ndev, &port_num);
232 		if (!roce)
233 			return NOTIFY_DONE;
234 		if ((upper == ndev || (!upper && ndev == roce->netdev))
235 		    && ibdev->ib_active) {
236 			struct ib_event ibev = { };
237 			enum ib_port_state port_state;
238 
239 			if (get_port_state(&ibdev->ib_dev, port_num,
240 					   &port_state))
241 				goto done;
242 
243 			if (roce->last_port_state == port_state)
244 				goto done;
245 
246 			roce->last_port_state = port_state;
247 			ibev.device = &ibdev->ib_dev;
248 			if (port_state == IB_PORT_DOWN)
249 				ibev.event = IB_EVENT_PORT_ERR;
250 			else if (port_state == IB_PORT_ACTIVE)
251 				ibev.event = IB_EVENT_PORT_ACTIVE;
252 			else
253 				goto done;
254 
255 			ibev.element.port_num = port_num;
256 			ib_dispatch_event(&ibev);
257 		}
258 		break;
259 	}
260 
261 	default:
262 		break;
263 	}
264 done:
265 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
266 	return NOTIFY_DONE;
267 }
268 
269 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
270 					     u8 port_num)
271 {
272 	struct mlx5_ib_dev *ibdev = to_mdev(device);
273 	struct net_device *ndev;
274 	struct mlx5_core_dev *mdev;
275 
276 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
277 	if (!mdev)
278 		return NULL;
279 
280 	ndev = mlx5_lag_get_roce_netdev(mdev);
281 	if (ndev)
282 		goto out;
283 
284 	/* Ensure ndev does not disappear before we invoke dev_hold()
285 	 */
286 	read_lock(&ibdev->port[port_num - 1].roce.netdev_lock);
287 	ndev = ibdev->port[port_num - 1].roce.netdev;
288 	if (ndev)
289 		dev_hold(ndev);
290 	read_unlock(&ibdev->port[port_num - 1].roce.netdev_lock);
291 
292 out:
293 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
294 	return ndev;
295 }
296 
297 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
298 						   u8 ib_port_num,
299 						   u8 *native_port_num)
300 {
301 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
302 							  ib_port_num);
303 	struct mlx5_core_dev *mdev = NULL;
304 	struct mlx5_ib_multiport_info *mpi;
305 	struct mlx5_ib_port *port;
306 
307 	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
308 	    ll != IB_LINK_LAYER_ETHERNET) {
309 		if (native_port_num)
310 			*native_port_num = ib_port_num;
311 		return ibdev->mdev;
312 	}
313 
314 	if (native_port_num)
315 		*native_port_num = 1;
316 
317 	port = &ibdev->port[ib_port_num - 1];
318 	if (!port)
319 		return NULL;
320 
321 	spin_lock(&port->mp.mpi_lock);
322 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
323 	if (mpi && !mpi->unaffiliate) {
324 		mdev = mpi->mdev;
325 		/* If it's the master no need to refcount, it'll exist
326 		 * as long as the ib_dev exists.
327 		 */
328 		if (!mpi->is_master)
329 			mpi->mdev_refcnt++;
330 	}
331 	spin_unlock(&port->mp.mpi_lock);
332 
333 	return mdev;
334 }
335 
336 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u8 port_num)
337 {
338 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
339 							  port_num);
340 	struct mlx5_ib_multiport_info *mpi;
341 	struct mlx5_ib_port *port;
342 
343 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
344 		return;
345 
346 	port = &ibdev->port[port_num - 1];
347 
348 	spin_lock(&port->mp.mpi_lock);
349 	mpi = ibdev->port[port_num - 1].mp.mpi;
350 	if (mpi->is_master)
351 		goto out;
352 
353 	mpi->mdev_refcnt--;
354 	if (mpi->unaffiliate)
355 		complete(&mpi->unref_comp);
356 out:
357 	spin_unlock(&port->mp.mpi_lock);
358 }
359 
360 static int translate_eth_legacy_proto_oper(u32 eth_proto_oper, u8 *active_speed,
361 					   u8 *active_width)
362 {
363 	switch (eth_proto_oper) {
364 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
365 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
366 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
367 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
368 		*active_width = IB_WIDTH_1X;
369 		*active_speed = IB_SPEED_SDR;
370 		break;
371 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
372 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
373 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
374 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
375 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
376 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
377 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
378 		*active_width = IB_WIDTH_1X;
379 		*active_speed = IB_SPEED_QDR;
380 		break;
381 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
382 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
383 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
384 		*active_width = IB_WIDTH_1X;
385 		*active_speed = IB_SPEED_EDR;
386 		break;
387 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
388 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
389 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
390 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
391 		*active_width = IB_WIDTH_4X;
392 		*active_speed = IB_SPEED_QDR;
393 		break;
394 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
395 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
396 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
397 		*active_width = IB_WIDTH_1X;
398 		*active_speed = IB_SPEED_HDR;
399 		break;
400 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
401 		*active_width = IB_WIDTH_4X;
402 		*active_speed = IB_SPEED_FDR;
403 		break;
404 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
405 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
406 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
407 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
408 		*active_width = IB_WIDTH_4X;
409 		*active_speed = IB_SPEED_EDR;
410 		break;
411 	default:
412 		return -EINVAL;
413 	}
414 
415 	return 0;
416 }
417 
418 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u8 *active_speed,
419 					u8 *active_width)
420 {
421 	switch (eth_proto_oper) {
422 	case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
423 	case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
424 		*active_width = IB_WIDTH_1X;
425 		*active_speed = IB_SPEED_SDR;
426 		break;
427 	case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
428 		*active_width = IB_WIDTH_1X;
429 		*active_speed = IB_SPEED_DDR;
430 		break;
431 	case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
432 		*active_width = IB_WIDTH_1X;
433 		*active_speed = IB_SPEED_QDR;
434 		break;
435 	case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
436 		*active_width = IB_WIDTH_4X;
437 		*active_speed = IB_SPEED_QDR;
438 		break;
439 	case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
440 		*active_width = IB_WIDTH_1X;
441 		*active_speed = IB_SPEED_EDR;
442 		break;
443 	case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
444 		*active_width = IB_WIDTH_2X;
445 		*active_speed = IB_SPEED_EDR;
446 		break;
447 	case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
448 		*active_width = IB_WIDTH_1X;
449 		*active_speed = IB_SPEED_HDR;
450 		break;
451 	case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
452 		*active_width = IB_WIDTH_4X;
453 		*active_speed = IB_SPEED_EDR;
454 		break;
455 	case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
456 		*active_width = IB_WIDTH_2X;
457 		*active_speed = IB_SPEED_HDR;
458 		break;
459 	case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4):
460 		*active_width = IB_WIDTH_4X;
461 		*active_speed = IB_SPEED_HDR;
462 		break;
463 	default:
464 		return -EINVAL;
465 	}
466 
467 	return 0;
468 }
469 
470 static int translate_eth_proto_oper(u32 eth_proto_oper, u8 *active_speed,
471 				    u8 *active_width, bool ext)
472 {
473 	return ext ?
474 		translate_eth_ext_proto_oper(eth_proto_oper, active_speed,
475 					     active_width) :
476 		translate_eth_legacy_proto_oper(eth_proto_oper, active_speed,
477 						active_width);
478 }
479 
480 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
481 				struct ib_port_attr *props)
482 {
483 	struct mlx5_ib_dev *dev = to_mdev(device);
484 	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0};
485 	struct mlx5_core_dev *mdev;
486 	struct net_device *ndev, *upper;
487 	enum ib_mtu ndev_ib_mtu;
488 	bool put_mdev = true;
489 	u16 qkey_viol_cntr;
490 	u32 eth_prot_oper;
491 	u8 mdev_port_num;
492 	bool ext;
493 	int err;
494 
495 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
496 	if (!mdev) {
497 		/* This means the port isn't affiliated yet. Get the
498 		 * info for the master port instead.
499 		 */
500 		put_mdev = false;
501 		mdev = dev->mdev;
502 		mdev_port_num = 1;
503 		port_num = 1;
504 	}
505 
506 	/* Possible bad flows are checked before filling out props so in case
507 	 * of an error it will still be zeroed out.
508 	 * Use native port in case of reps
509 	 */
510 	if (dev->is_rep)
511 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
512 					   1);
513 	else
514 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
515 					   mdev_port_num);
516 	if (err)
517 		goto out;
518 	ext = MLX5_CAP_PCAM_FEATURE(dev->mdev, ptys_extended_ethernet);
519 	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
520 
521 	props->active_width     = IB_WIDTH_4X;
522 	props->active_speed     = IB_SPEED_QDR;
523 
524 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
525 				 &props->active_width, ext);
526 
527 	props->port_cap_flags |= IB_PORT_CM_SUP;
528 	props->ip_gids = true;
529 
530 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
531 						roce_address_table_size);
532 	props->max_mtu          = IB_MTU_4096;
533 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
534 	props->pkey_tbl_len     = 1;
535 	props->state            = IB_PORT_DOWN;
536 	props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
537 
538 	mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
539 	props->qkey_viol_cntr = qkey_viol_cntr;
540 
541 	/* If this is a stub query for an unaffiliated port stop here */
542 	if (!put_mdev)
543 		goto out;
544 
545 	ndev = mlx5_ib_get_netdev(device, port_num);
546 	if (!ndev)
547 		goto out;
548 
549 	if (dev->lag_active) {
550 		rcu_read_lock();
551 		upper = netdev_master_upper_dev_get_rcu(ndev);
552 		if (upper) {
553 			dev_put(ndev);
554 			ndev = upper;
555 			dev_hold(ndev);
556 		}
557 		rcu_read_unlock();
558 	}
559 
560 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
561 		props->state      = IB_PORT_ACTIVE;
562 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
563 	}
564 
565 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
566 
567 	dev_put(ndev);
568 
569 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
570 out:
571 	if (put_mdev)
572 		mlx5_ib_put_native_port_mdev(dev, port_num);
573 	return err;
574 }
575 
576 static int set_roce_addr(struct mlx5_ib_dev *dev, u8 port_num,
577 			 unsigned int index, const union ib_gid *gid,
578 			 const struct ib_gid_attr *attr)
579 {
580 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
581 	u16 vlan_id = 0xffff;
582 	u8 roce_version = 0;
583 	u8 roce_l3_type = 0;
584 	u8 mac[ETH_ALEN];
585 	int ret;
586 
587 	if (gid) {
588 		gid_type = attr->gid_type;
589 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
590 		if (ret)
591 			return ret;
592 	}
593 
594 	switch (gid_type) {
595 	case IB_GID_TYPE_IB:
596 		roce_version = MLX5_ROCE_VERSION_1;
597 		break;
598 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
599 		roce_version = MLX5_ROCE_VERSION_2;
600 		if (ipv6_addr_v4mapped((void *)gid))
601 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
602 		else
603 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
604 		break;
605 
606 	default:
607 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
608 	}
609 
610 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
611 				      roce_l3_type, gid->raw, mac,
612 				      vlan_id < VLAN_CFI_MASK, vlan_id,
613 				      port_num);
614 }
615 
616 static int mlx5_ib_add_gid(const struct ib_gid_attr *attr,
617 			   __always_unused void **context)
618 {
619 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
620 			     attr->index, &attr->gid, attr);
621 }
622 
623 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
624 			   __always_unused void **context)
625 {
626 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
627 			     attr->index, NULL, NULL);
628 }
629 
630 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev,
631 			       const struct ib_gid_attr *attr)
632 {
633 	if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
634 		return 0;
635 
636 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
637 }
638 
639 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
640 {
641 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
642 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
643 	return 0;
644 }
645 
646 enum {
647 	MLX5_VPORT_ACCESS_METHOD_MAD,
648 	MLX5_VPORT_ACCESS_METHOD_HCA,
649 	MLX5_VPORT_ACCESS_METHOD_NIC,
650 };
651 
652 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
653 {
654 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
655 		return MLX5_VPORT_ACCESS_METHOD_MAD;
656 
657 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
658 	    IB_LINK_LAYER_ETHERNET)
659 		return MLX5_VPORT_ACCESS_METHOD_NIC;
660 
661 	return MLX5_VPORT_ACCESS_METHOD_HCA;
662 }
663 
664 static void get_atomic_caps(struct mlx5_ib_dev *dev,
665 			    u8 atomic_size_qp,
666 			    struct ib_device_attr *props)
667 {
668 	u8 tmp;
669 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
670 	u8 atomic_req_8B_endianness_mode =
671 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
672 
673 	/* Check if HW supports 8 bytes standard atomic operations and capable
674 	 * of host endianness respond
675 	 */
676 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
677 	if (((atomic_operations & tmp) == tmp) &&
678 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
679 	    (atomic_req_8B_endianness_mode)) {
680 		props->atomic_cap = IB_ATOMIC_HCA;
681 	} else {
682 		props->atomic_cap = IB_ATOMIC_NONE;
683 	}
684 }
685 
686 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
687 			       struct ib_device_attr *props)
688 {
689 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
690 
691 	get_atomic_caps(dev, atomic_size_qp, props);
692 }
693 
694 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
695 					__be64 *sys_image_guid)
696 {
697 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
698 	struct mlx5_core_dev *mdev = dev->mdev;
699 	u64 tmp;
700 	int err;
701 
702 	switch (mlx5_get_vport_access_method(ibdev)) {
703 	case MLX5_VPORT_ACCESS_METHOD_MAD:
704 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
705 							    sys_image_guid);
706 
707 	case MLX5_VPORT_ACCESS_METHOD_HCA:
708 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
709 		break;
710 
711 	case MLX5_VPORT_ACCESS_METHOD_NIC:
712 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
713 		break;
714 
715 	default:
716 		return -EINVAL;
717 	}
718 
719 	if (!err)
720 		*sys_image_guid = cpu_to_be64(tmp);
721 
722 	return err;
723 
724 }
725 
726 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
727 				u16 *max_pkeys)
728 {
729 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
730 	struct mlx5_core_dev *mdev = dev->mdev;
731 
732 	switch (mlx5_get_vport_access_method(ibdev)) {
733 	case MLX5_VPORT_ACCESS_METHOD_MAD:
734 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
735 
736 	case MLX5_VPORT_ACCESS_METHOD_HCA:
737 	case MLX5_VPORT_ACCESS_METHOD_NIC:
738 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
739 						pkey_table_size));
740 		return 0;
741 
742 	default:
743 		return -EINVAL;
744 	}
745 }
746 
747 static int mlx5_query_vendor_id(struct ib_device *ibdev,
748 				u32 *vendor_id)
749 {
750 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
751 
752 	switch (mlx5_get_vport_access_method(ibdev)) {
753 	case MLX5_VPORT_ACCESS_METHOD_MAD:
754 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
755 
756 	case MLX5_VPORT_ACCESS_METHOD_HCA:
757 	case MLX5_VPORT_ACCESS_METHOD_NIC:
758 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
759 
760 	default:
761 		return -EINVAL;
762 	}
763 }
764 
765 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
766 				__be64 *node_guid)
767 {
768 	u64 tmp;
769 	int err;
770 
771 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
772 	case MLX5_VPORT_ACCESS_METHOD_MAD:
773 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
774 
775 	case MLX5_VPORT_ACCESS_METHOD_HCA:
776 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
777 		break;
778 
779 	case MLX5_VPORT_ACCESS_METHOD_NIC:
780 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
781 		break;
782 
783 	default:
784 		return -EINVAL;
785 	}
786 
787 	if (!err)
788 		*node_guid = cpu_to_be64(tmp);
789 
790 	return err;
791 }
792 
793 struct mlx5_reg_node_desc {
794 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
795 };
796 
797 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
798 {
799 	struct mlx5_reg_node_desc in;
800 
801 	if (mlx5_use_mad_ifc(dev))
802 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
803 
804 	memset(&in, 0, sizeof(in));
805 
806 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
807 				    sizeof(struct mlx5_reg_node_desc),
808 				    MLX5_REG_NODE_DESC, 0, 0);
809 }
810 
811 static int mlx5_ib_query_device(struct ib_device *ibdev,
812 				struct ib_device_attr *props,
813 				struct ib_udata *uhw)
814 {
815 	size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
816 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
817 	struct mlx5_core_dev *mdev = dev->mdev;
818 	int err = -ENOMEM;
819 	int max_sq_desc;
820 	int max_rq_sg;
821 	int max_sq_sg;
822 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
823 	bool raw_support = !mlx5_core_mp_enabled(mdev);
824 	struct mlx5_ib_query_device_resp resp = {};
825 	size_t resp_len;
826 	u64 max_tso;
827 
828 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
829 	if (uhw_outlen && uhw_outlen < resp_len)
830 		return -EINVAL;
831 
832 	resp.response_length = resp_len;
833 
834 	if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
835 		return -EINVAL;
836 
837 	memset(props, 0, sizeof(*props));
838 	err = mlx5_query_system_image_guid(ibdev,
839 					   &props->sys_image_guid);
840 	if (err)
841 		return err;
842 
843 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
844 	if (err)
845 		return err;
846 
847 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
848 	if (err)
849 		return err;
850 
851 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
852 		(fw_rev_min(dev->mdev) << 16) |
853 		fw_rev_sub(dev->mdev);
854 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
855 		IB_DEVICE_PORT_ACTIVE_EVENT		|
856 		IB_DEVICE_SYS_IMAGE_GUID		|
857 		IB_DEVICE_RC_RNR_NAK_GEN;
858 
859 	if (MLX5_CAP_GEN(mdev, pkv))
860 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
861 	if (MLX5_CAP_GEN(mdev, qkv))
862 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
863 	if (MLX5_CAP_GEN(mdev, apm))
864 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
865 	if (MLX5_CAP_GEN(mdev, xrc))
866 		props->device_cap_flags |= IB_DEVICE_XRC;
867 	if (MLX5_CAP_GEN(mdev, imaicl)) {
868 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
869 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
870 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
871 		/* We support 'Gappy' memory registration too */
872 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
873 	}
874 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
875 	if (MLX5_CAP_GEN(mdev, sho)) {
876 		props->device_cap_flags |= IB_DEVICE_INTEGRITY_HANDOVER;
877 		/* At this stage no support for signature handover */
878 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
879 				      IB_PROT_T10DIF_TYPE_2 |
880 				      IB_PROT_T10DIF_TYPE_3;
881 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
882 				       IB_GUARD_T10DIF_CSUM;
883 	}
884 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
885 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
886 
887 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
888 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
889 			/* Legacy bit to support old userspace libraries */
890 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
891 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
892 		}
893 
894 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
895 			props->raw_packet_caps |=
896 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
897 
898 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) {
899 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
900 			if (max_tso) {
901 				resp.tso_caps.max_tso = 1 << max_tso;
902 				resp.tso_caps.supported_qpts |=
903 					1 << IB_QPT_RAW_PACKET;
904 				resp.response_length += sizeof(resp.tso_caps);
905 			}
906 		}
907 
908 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) {
909 			resp.rss_caps.rx_hash_function =
910 						MLX5_RX_HASH_FUNC_TOEPLITZ;
911 			resp.rss_caps.rx_hash_fields_mask =
912 						MLX5_RX_HASH_SRC_IPV4 |
913 						MLX5_RX_HASH_DST_IPV4 |
914 						MLX5_RX_HASH_SRC_IPV6 |
915 						MLX5_RX_HASH_DST_IPV6 |
916 						MLX5_RX_HASH_SRC_PORT_TCP |
917 						MLX5_RX_HASH_DST_PORT_TCP |
918 						MLX5_RX_HASH_SRC_PORT_UDP |
919 						MLX5_RX_HASH_DST_PORT_UDP |
920 						MLX5_RX_HASH_INNER;
921 			if (mlx5_accel_ipsec_device_caps(dev->mdev) &
922 			    MLX5_ACCEL_IPSEC_CAP_DEVICE)
923 				resp.rss_caps.rx_hash_fields_mask |=
924 					MLX5_RX_HASH_IPSEC_SPI;
925 			resp.response_length += sizeof(resp.rss_caps);
926 		}
927 	} else {
928 		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen)
929 			resp.response_length += sizeof(resp.tso_caps);
930 		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen)
931 			resp.response_length += sizeof(resp.rss_caps);
932 	}
933 
934 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
935 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
936 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
937 	}
938 
939 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
940 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
941 	    raw_support)
942 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
943 
944 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
945 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
946 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
947 
948 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
949 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
950 	    raw_support) {
951 		/* Legacy bit to support old userspace libraries */
952 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
953 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
954 	}
955 
956 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
957 		props->max_dm_size =
958 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
959 	}
960 
961 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
962 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
963 
964 	if (MLX5_CAP_GEN(mdev, end_pad))
965 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
966 
967 	props->vendor_part_id	   = mdev->pdev->device;
968 	props->hw_ver		   = mdev->pdev->revision;
969 
970 	props->max_mr_size	   = ~0ull;
971 	props->page_size_cap	   = ~(min_page_size - 1);
972 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
973 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
974 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
975 		     sizeof(struct mlx5_wqe_data_seg);
976 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
977 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
978 		     sizeof(struct mlx5_wqe_raddr_seg)) /
979 		sizeof(struct mlx5_wqe_data_seg);
980 	props->max_send_sge = max_sq_sg;
981 	props->max_recv_sge = max_rq_sg;
982 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
983 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
984 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
985 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
986 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
987 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
988 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
989 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
990 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
991 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
992 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
993 	props->max_srq_sge	   = max_rq_sg - 1;
994 	props->max_fast_reg_page_list_len =
995 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
996 	props->max_pi_fast_reg_page_list_len =
997 		props->max_fast_reg_page_list_len / 2;
998 	props->max_sgl_rd =
999 		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
1000 	get_atomic_caps_qp(dev, props);
1001 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
1002 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
1003 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
1004 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1005 					   props->max_mcast_grp;
1006 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
1007 	props->max_ah = INT_MAX;
1008 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
1009 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
1010 
1011 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
1012 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
1013 			props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
1014 		props->odp_caps = dev->odp_caps;
1015 		if (!uhw) {
1016 			/* ODP for kernel QPs is not implemented for receive
1017 			 * WQEs and SRQ WQEs
1018 			 */
1019 			props->odp_caps.per_transport_caps.rc_odp_caps &=
1020 				~(IB_ODP_SUPPORT_READ |
1021 				  IB_ODP_SUPPORT_SRQ_RECV);
1022 			props->odp_caps.per_transport_caps.uc_odp_caps &=
1023 				~(IB_ODP_SUPPORT_READ |
1024 				  IB_ODP_SUPPORT_SRQ_RECV);
1025 			props->odp_caps.per_transport_caps.ud_odp_caps &=
1026 				~(IB_ODP_SUPPORT_READ |
1027 				  IB_ODP_SUPPORT_SRQ_RECV);
1028 			props->odp_caps.per_transport_caps.xrc_odp_caps &=
1029 				~(IB_ODP_SUPPORT_READ |
1030 				  IB_ODP_SUPPORT_SRQ_RECV);
1031 		}
1032 	}
1033 
1034 	if (MLX5_CAP_GEN(mdev, cd))
1035 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
1036 
1037 	if (mlx5_core_is_vf(mdev))
1038 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
1039 
1040 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1041 	    IB_LINK_LAYER_ETHERNET && raw_support) {
1042 		props->rss_caps.max_rwq_indirection_tables =
1043 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1044 		props->rss_caps.max_rwq_indirection_table_size =
1045 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1046 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1047 		props->max_wq_type_rq =
1048 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1049 	}
1050 
1051 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1052 		props->tm_caps.max_num_tags =
1053 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1054 		props->tm_caps.max_ops =
1055 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1056 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1057 	}
1058 
1059 	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1060 	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1061 		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1062 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1063 	}
1064 
1065 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1066 		props->cq_caps.max_cq_moderation_count =
1067 						MLX5_MAX_CQ_COUNT;
1068 		props->cq_caps.max_cq_moderation_period =
1069 						MLX5_MAX_CQ_PERIOD;
1070 	}
1071 
1072 	if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) {
1073 		resp.response_length += sizeof(resp.cqe_comp_caps);
1074 
1075 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1076 			resp.cqe_comp_caps.max_num =
1077 				MLX5_CAP_GEN(dev->mdev,
1078 					     cqe_compression_max_num);
1079 
1080 			resp.cqe_comp_caps.supported_format =
1081 				MLX5_IB_CQE_RES_FORMAT_HASH |
1082 				MLX5_IB_CQE_RES_FORMAT_CSUM;
1083 
1084 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1085 				resp.cqe_comp_caps.supported_format |=
1086 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1087 		}
1088 	}
1089 
1090 	if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen &&
1091 	    raw_support) {
1092 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1093 		    MLX5_CAP_GEN(mdev, qos)) {
1094 			resp.packet_pacing_caps.qp_rate_limit_max =
1095 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1096 			resp.packet_pacing_caps.qp_rate_limit_min =
1097 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1098 			resp.packet_pacing_caps.supported_qpts |=
1099 				1 << IB_QPT_RAW_PACKET;
1100 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1101 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1102 				resp.packet_pacing_caps.cap_flags |=
1103 					MLX5_IB_PP_SUPPORT_BURST;
1104 		}
1105 		resp.response_length += sizeof(resp.packet_pacing_caps);
1106 	}
1107 
1108 	if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <=
1109 	    uhw_outlen) {
1110 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1111 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1112 				MLX5_IB_ALLOW_MPW;
1113 
1114 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1115 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1116 				MLX5_IB_SUPPORT_EMPW;
1117 
1118 		resp.response_length +=
1119 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1120 	}
1121 
1122 	if (offsetofend(typeof(resp), flags) <= uhw_outlen) {
1123 		resp.response_length += sizeof(resp.flags);
1124 
1125 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1126 			resp.flags |=
1127 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1128 
1129 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1130 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1131 		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1132 			resp.flags |=
1133 				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1134 
1135 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1136 	}
1137 
1138 	if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) {
1139 		resp.response_length += sizeof(resp.sw_parsing_caps);
1140 		if (MLX5_CAP_ETH(mdev, swp)) {
1141 			resp.sw_parsing_caps.sw_parsing_offloads |=
1142 				MLX5_IB_SW_PARSING;
1143 
1144 			if (MLX5_CAP_ETH(mdev, swp_csum))
1145 				resp.sw_parsing_caps.sw_parsing_offloads |=
1146 					MLX5_IB_SW_PARSING_CSUM;
1147 
1148 			if (MLX5_CAP_ETH(mdev, swp_lso))
1149 				resp.sw_parsing_caps.sw_parsing_offloads |=
1150 					MLX5_IB_SW_PARSING_LSO;
1151 
1152 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1153 				resp.sw_parsing_caps.supported_qpts =
1154 					BIT(IB_QPT_RAW_PACKET);
1155 		}
1156 	}
1157 
1158 	if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen &&
1159 	    raw_support) {
1160 		resp.response_length += sizeof(resp.striding_rq_caps);
1161 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1162 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1163 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1164 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1165 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1166 			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1167 				resp.striding_rq_caps
1168 					.min_single_wqe_log_num_of_strides =
1169 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1170 			else
1171 				resp.striding_rq_caps
1172 					.min_single_wqe_log_num_of_strides =
1173 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1174 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1175 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1176 			resp.striding_rq_caps.supported_qpts =
1177 				BIT(IB_QPT_RAW_PACKET);
1178 		}
1179 	}
1180 
1181 	if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) {
1182 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1183 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1184 			resp.tunnel_offloads_caps |=
1185 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1186 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1187 			resp.tunnel_offloads_caps |=
1188 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1189 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1190 			resp.tunnel_offloads_caps |=
1191 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1192 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre))
1193 			resp.tunnel_offloads_caps |=
1194 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1195 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp))
1196 			resp.tunnel_offloads_caps |=
1197 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1198 	}
1199 
1200 	if (uhw_outlen) {
1201 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1202 
1203 		if (err)
1204 			return err;
1205 	}
1206 
1207 	return 0;
1208 }
1209 
1210 enum mlx5_ib_width {
1211 	MLX5_IB_WIDTH_1X	= 1 << 0,
1212 	MLX5_IB_WIDTH_2X	= 1 << 1,
1213 	MLX5_IB_WIDTH_4X	= 1 << 2,
1214 	MLX5_IB_WIDTH_8X	= 1 << 3,
1215 	MLX5_IB_WIDTH_12X	= 1 << 4
1216 };
1217 
1218 static void translate_active_width(struct ib_device *ibdev, u8 active_width,
1219 				  u8 *ib_width)
1220 {
1221 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1222 
1223 	if (active_width & MLX5_IB_WIDTH_1X)
1224 		*ib_width = IB_WIDTH_1X;
1225 	else if (active_width & MLX5_IB_WIDTH_2X)
1226 		*ib_width = IB_WIDTH_2X;
1227 	else if (active_width & MLX5_IB_WIDTH_4X)
1228 		*ib_width = IB_WIDTH_4X;
1229 	else if (active_width & MLX5_IB_WIDTH_8X)
1230 		*ib_width = IB_WIDTH_8X;
1231 	else if (active_width & MLX5_IB_WIDTH_12X)
1232 		*ib_width = IB_WIDTH_12X;
1233 	else {
1234 		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1235 			    (int)active_width);
1236 		*ib_width = IB_WIDTH_4X;
1237 	}
1238 
1239 	return;
1240 }
1241 
1242 static int mlx5_mtu_to_ib_mtu(int mtu)
1243 {
1244 	switch (mtu) {
1245 	case 256: return 1;
1246 	case 512: return 2;
1247 	case 1024: return 3;
1248 	case 2048: return 4;
1249 	case 4096: return 5;
1250 	default:
1251 		pr_warn("invalid mtu\n");
1252 		return -1;
1253 	}
1254 }
1255 
1256 enum ib_max_vl_num {
1257 	__IB_MAX_VL_0		= 1,
1258 	__IB_MAX_VL_0_1		= 2,
1259 	__IB_MAX_VL_0_3		= 3,
1260 	__IB_MAX_VL_0_7		= 4,
1261 	__IB_MAX_VL_0_14	= 5,
1262 };
1263 
1264 enum mlx5_vl_hw_cap {
1265 	MLX5_VL_HW_0	= 1,
1266 	MLX5_VL_HW_0_1	= 2,
1267 	MLX5_VL_HW_0_2	= 3,
1268 	MLX5_VL_HW_0_3	= 4,
1269 	MLX5_VL_HW_0_4	= 5,
1270 	MLX5_VL_HW_0_5	= 6,
1271 	MLX5_VL_HW_0_6	= 7,
1272 	MLX5_VL_HW_0_7	= 8,
1273 	MLX5_VL_HW_0_14	= 15
1274 };
1275 
1276 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1277 				u8 *max_vl_num)
1278 {
1279 	switch (vl_hw_cap) {
1280 	case MLX5_VL_HW_0:
1281 		*max_vl_num = __IB_MAX_VL_0;
1282 		break;
1283 	case MLX5_VL_HW_0_1:
1284 		*max_vl_num = __IB_MAX_VL_0_1;
1285 		break;
1286 	case MLX5_VL_HW_0_3:
1287 		*max_vl_num = __IB_MAX_VL_0_3;
1288 		break;
1289 	case MLX5_VL_HW_0_7:
1290 		*max_vl_num = __IB_MAX_VL_0_7;
1291 		break;
1292 	case MLX5_VL_HW_0_14:
1293 		*max_vl_num = __IB_MAX_VL_0_14;
1294 		break;
1295 
1296 	default:
1297 		return -EINVAL;
1298 	}
1299 
1300 	return 0;
1301 }
1302 
1303 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
1304 			       struct ib_port_attr *props)
1305 {
1306 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1307 	struct mlx5_core_dev *mdev = dev->mdev;
1308 	struct mlx5_hca_vport_context *rep;
1309 	u16 max_mtu;
1310 	u16 oper_mtu;
1311 	int err;
1312 	u8 ib_link_width_oper;
1313 	u8 vl_hw_cap;
1314 
1315 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1316 	if (!rep) {
1317 		err = -ENOMEM;
1318 		goto out;
1319 	}
1320 
1321 	/* props being zeroed by the caller, avoid zeroing it here */
1322 
1323 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1324 	if (err)
1325 		goto out;
1326 
1327 	props->lid		= rep->lid;
1328 	props->lmc		= rep->lmc;
1329 	props->sm_lid		= rep->sm_lid;
1330 	props->sm_sl		= rep->sm_sl;
1331 	props->state		= rep->vport_state;
1332 	props->phys_state	= rep->port_physical_state;
1333 	props->port_cap_flags	= rep->cap_mask1;
1334 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1335 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1336 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1337 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1338 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1339 	props->subnet_timeout	= rep->subnet_timeout;
1340 	props->init_type_reply	= rep->init_type_reply;
1341 
1342 	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1343 		props->port_cap_flags2 = rep->cap_mask2;
1344 
1345 	err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
1346 	if (err)
1347 		goto out;
1348 
1349 	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1350 
1351 	err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
1352 	if (err)
1353 		goto out;
1354 
1355 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1356 
1357 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1358 
1359 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1360 
1361 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1362 
1363 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1364 	if (err)
1365 		goto out;
1366 
1367 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1368 				   &props->max_vl_num);
1369 out:
1370 	kfree(rep);
1371 	return err;
1372 }
1373 
1374 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
1375 		       struct ib_port_attr *props)
1376 {
1377 	unsigned int count;
1378 	int ret;
1379 
1380 	switch (mlx5_get_vport_access_method(ibdev)) {
1381 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1382 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1383 		break;
1384 
1385 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1386 		ret = mlx5_query_hca_port(ibdev, port, props);
1387 		break;
1388 
1389 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1390 		ret = mlx5_query_port_roce(ibdev, port, props);
1391 		break;
1392 
1393 	default:
1394 		ret = -EINVAL;
1395 	}
1396 
1397 	if (!ret && props) {
1398 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1399 		struct mlx5_core_dev *mdev;
1400 		bool put_mdev = true;
1401 
1402 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1403 		if (!mdev) {
1404 			/* If the port isn't affiliated yet query the master.
1405 			 * The master and slave will have the same values.
1406 			 */
1407 			mdev = dev->mdev;
1408 			port = 1;
1409 			put_mdev = false;
1410 		}
1411 		count = mlx5_core_reserved_gids_count(mdev);
1412 		if (put_mdev)
1413 			mlx5_ib_put_native_port_mdev(dev, port);
1414 		props->gid_tbl_len -= count;
1415 	}
1416 	return ret;
1417 }
1418 
1419 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u8 port,
1420 				  struct ib_port_attr *props)
1421 {
1422 	int ret;
1423 
1424 	/* Only link layer == ethernet is valid for representors
1425 	 * and we always use port 1
1426 	 */
1427 	ret = mlx5_query_port_roce(ibdev, port, props);
1428 	if (ret || !props)
1429 		return ret;
1430 
1431 	/* We don't support GIDS */
1432 	props->gid_tbl_len = 0;
1433 
1434 	return ret;
1435 }
1436 
1437 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1438 			     union ib_gid *gid)
1439 {
1440 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1441 	struct mlx5_core_dev *mdev = dev->mdev;
1442 
1443 	switch (mlx5_get_vport_access_method(ibdev)) {
1444 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1445 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1446 
1447 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1448 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1449 
1450 	default:
1451 		return -EINVAL;
1452 	}
1453 
1454 }
1455 
1456 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u8 port,
1457 				   u16 index, u16 *pkey)
1458 {
1459 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1460 	struct mlx5_core_dev *mdev;
1461 	bool put_mdev = true;
1462 	u8 mdev_port_num;
1463 	int err;
1464 
1465 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1466 	if (!mdev) {
1467 		/* The port isn't affiliated yet, get the PKey from the master
1468 		 * port. For RoCE the PKey tables will be the same.
1469 		 */
1470 		put_mdev = false;
1471 		mdev = dev->mdev;
1472 		mdev_port_num = 1;
1473 	}
1474 
1475 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1476 					index, pkey);
1477 	if (put_mdev)
1478 		mlx5_ib_put_native_port_mdev(dev, port);
1479 
1480 	return err;
1481 }
1482 
1483 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1484 			      u16 *pkey)
1485 {
1486 	switch (mlx5_get_vport_access_method(ibdev)) {
1487 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1488 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1489 
1490 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1491 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1492 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1493 	default:
1494 		return -EINVAL;
1495 	}
1496 }
1497 
1498 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1499 				 struct ib_device_modify *props)
1500 {
1501 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1502 	struct mlx5_reg_node_desc in;
1503 	struct mlx5_reg_node_desc out;
1504 	int err;
1505 
1506 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1507 		return -EOPNOTSUPP;
1508 
1509 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1510 		return 0;
1511 
1512 	/*
1513 	 * If possible, pass node desc to FW, so it can generate
1514 	 * a 144 trap.  If cmd fails, just ignore.
1515 	 */
1516 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1517 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1518 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1519 	if (err)
1520 		return err;
1521 
1522 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1523 
1524 	return err;
1525 }
1526 
1527 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1528 				u32 value)
1529 {
1530 	struct mlx5_hca_vport_context ctx = {};
1531 	struct mlx5_core_dev *mdev;
1532 	u8 mdev_port_num;
1533 	int err;
1534 
1535 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1536 	if (!mdev)
1537 		return -ENODEV;
1538 
1539 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1540 	if (err)
1541 		goto out;
1542 
1543 	if (~ctx.cap_mask1_perm & mask) {
1544 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1545 			     mask, ctx.cap_mask1_perm);
1546 		err = -EINVAL;
1547 		goto out;
1548 	}
1549 
1550 	ctx.cap_mask1 = value;
1551 	ctx.cap_mask1_perm = mask;
1552 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1553 						 0, &ctx);
1554 
1555 out:
1556 	mlx5_ib_put_native_port_mdev(dev, port_num);
1557 
1558 	return err;
1559 }
1560 
1561 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1562 			       struct ib_port_modify *props)
1563 {
1564 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1565 	struct ib_port_attr attr;
1566 	u32 tmp;
1567 	int err;
1568 	u32 change_mask;
1569 	u32 value;
1570 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1571 		      IB_LINK_LAYER_INFINIBAND);
1572 
1573 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1574 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1575 	 */
1576 	if (!is_ib)
1577 		return 0;
1578 
1579 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1580 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1581 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1582 		return set_port_caps_atomic(dev, port, change_mask, value);
1583 	}
1584 
1585 	mutex_lock(&dev->cap_mask_mutex);
1586 
1587 	err = ib_query_port(ibdev, port, &attr);
1588 	if (err)
1589 		goto out;
1590 
1591 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1592 		~props->clr_port_cap_mask;
1593 
1594 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1595 
1596 out:
1597 	mutex_unlock(&dev->cap_mask_mutex);
1598 	return err;
1599 }
1600 
1601 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1602 {
1603 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1604 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1605 }
1606 
1607 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1608 {
1609 	/* Large page with non 4k uar support might limit the dynamic size */
1610 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1611 		return MLX5_MIN_DYN_BFREGS;
1612 
1613 	return MLX5_MAX_DYN_BFREGS;
1614 }
1615 
1616 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1617 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1618 			     struct mlx5_bfreg_info *bfregi)
1619 {
1620 	int uars_per_sys_page;
1621 	int bfregs_per_sys_page;
1622 	int ref_bfregs = req->total_num_bfregs;
1623 
1624 	if (req->total_num_bfregs == 0)
1625 		return -EINVAL;
1626 
1627 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1628 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1629 
1630 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1631 		return -ENOMEM;
1632 
1633 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1634 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1635 	/* This holds the required static allocation asked by the user */
1636 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1637 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1638 		return -EINVAL;
1639 
1640 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1641 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1642 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1643 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1644 
1645 	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",
1646 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1647 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1648 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1649 		    bfregi->num_sys_pages);
1650 
1651 	return 0;
1652 }
1653 
1654 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1655 {
1656 	struct mlx5_bfreg_info *bfregi;
1657 	int err;
1658 	int i;
1659 
1660 	bfregi = &context->bfregi;
1661 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1662 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1663 		if (err)
1664 			goto error;
1665 
1666 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1667 	}
1668 
1669 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1670 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1671 
1672 	return 0;
1673 
1674 error:
1675 	for (--i; i >= 0; i--)
1676 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1677 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1678 
1679 	return err;
1680 }
1681 
1682 static void deallocate_uars(struct mlx5_ib_dev *dev,
1683 			    struct mlx5_ib_ucontext *context)
1684 {
1685 	struct mlx5_bfreg_info *bfregi;
1686 	int i;
1687 
1688 	bfregi = &context->bfregi;
1689 	for (i = 0; i < bfregi->num_sys_pages; i++)
1690 		if (i < bfregi->num_static_sys_pages ||
1691 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1692 			mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1693 }
1694 
1695 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1696 {
1697 	int err = 0;
1698 
1699 	mutex_lock(&dev->lb.mutex);
1700 	if (td)
1701 		dev->lb.user_td++;
1702 	if (qp)
1703 		dev->lb.qps++;
1704 
1705 	if (dev->lb.user_td == 2 ||
1706 	    dev->lb.qps == 1) {
1707 		if (!dev->lb.enabled) {
1708 			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1709 			dev->lb.enabled = true;
1710 		}
1711 	}
1712 
1713 	mutex_unlock(&dev->lb.mutex);
1714 
1715 	return err;
1716 }
1717 
1718 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1719 {
1720 	mutex_lock(&dev->lb.mutex);
1721 	if (td)
1722 		dev->lb.user_td--;
1723 	if (qp)
1724 		dev->lb.qps--;
1725 
1726 	if (dev->lb.user_td == 1 &&
1727 	    dev->lb.qps == 0) {
1728 		if (dev->lb.enabled) {
1729 			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1730 			dev->lb.enabled = false;
1731 		}
1732 	}
1733 
1734 	mutex_unlock(&dev->lb.mutex);
1735 }
1736 
1737 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1738 					  u16 uid)
1739 {
1740 	int err;
1741 
1742 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1743 		return 0;
1744 
1745 	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1746 	if (err)
1747 		return err;
1748 
1749 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1750 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1751 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1752 		return err;
1753 
1754 	return mlx5_ib_enable_lb(dev, true, false);
1755 }
1756 
1757 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1758 					     u16 uid)
1759 {
1760 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1761 		return;
1762 
1763 	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1764 
1765 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1766 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1767 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1768 		return;
1769 
1770 	mlx5_ib_disable_lb(dev, true, false);
1771 }
1772 
1773 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1774 				  struct ib_udata *udata)
1775 {
1776 	struct ib_device *ibdev = uctx->device;
1777 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1778 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1779 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1780 	struct mlx5_core_dev *mdev = dev->mdev;
1781 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1782 	struct mlx5_bfreg_info *bfregi;
1783 	int ver;
1784 	int err;
1785 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1786 				     max_cqe_version);
1787 	u32 dump_fill_mkey;
1788 	bool lib_uar_4k;
1789 	bool lib_uar_dyn;
1790 
1791 	if (!dev->ib_active)
1792 		return -EAGAIN;
1793 
1794 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1795 		ver = 0;
1796 	else if (udata->inlen >= min_req_v2)
1797 		ver = 2;
1798 	else
1799 		return -EINVAL;
1800 
1801 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1802 	if (err)
1803 		return err;
1804 
1805 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1806 		return -EOPNOTSUPP;
1807 
1808 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1809 		return -EOPNOTSUPP;
1810 
1811 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1812 				    MLX5_NON_FP_BFREGS_PER_UAR);
1813 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1814 		return -EINVAL;
1815 
1816 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1817 	if (dev->wc_support)
1818 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1819 	resp.cache_line_size = cache_line_size();
1820 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1821 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1822 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1823 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1824 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1825 	resp.cqe_version = min_t(__u8,
1826 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1827 				 req.max_cqe_version);
1828 	resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1829 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1830 	resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1831 					MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1;
1832 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1833 				   sizeof(resp.response_length), udata->outlen);
1834 
1835 	if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1836 		if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_EGRESS))
1837 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1838 		if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1839 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1840 		if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1841 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1842 		if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1843 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1844 		/* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1845 	}
1846 
1847 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1848 	lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
1849 	bfregi = &context->bfregi;
1850 
1851 	if (lib_uar_dyn) {
1852 		bfregi->lib_uar_dyn = lib_uar_dyn;
1853 		goto uar_done;
1854 	}
1855 
1856 	/* updates req->total_num_bfregs */
1857 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1858 	if (err)
1859 		goto out_ctx;
1860 
1861 	mutex_init(&bfregi->lock);
1862 	bfregi->lib_uar_4k = lib_uar_4k;
1863 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1864 				GFP_KERNEL);
1865 	if (!bfregi->count) {
1866 		err = -ENOMEM;
1867 		goto out_ctx;
1868 	}
1869 
1870 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1871 				    sizeof(*bfregi->sys_pages),
1872 				    GFP_KERNEL);
1873 	if (!bfregi->sys_pages) {
1874 		err = -ENOMEM;
1875 		goto out_count;
1876 	}
1877 
1878 	err = allocate_uars(dev, context);
1879 	if (err)
1880 		goto out_sys_pages;
1881 
1882 uar_done:
1883 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1884 		err = mlx5_ib_devx_create(dev, true);
1885 		if (err < 0)
1886 			goto out_uars;
1887 		context->devx_uid = err;
1888 	}
1889 
1890 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1891 					     context->devx_uid);
1892 	if (err)
1893 		goto out_devx;
1894 
1895 	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1896 		err = mlx5_cmd_dump_fill_mkey(dev->mdev, &dump_fill_mkey);
1897 		if (err)
1898 			goto out_mdev;
1899 	}
1900 
1901 	INIT_LIST_HEAD(&context->db_page_list);
1902 	mutex_init(&context->db_page_mutex);
1903 
1904 	resp.tot_bfregs = lib_uar_dyn ? 0 : req.total_num_bfregs;
1905 	resp.num_ports = dev->num_ports;
1906 
1907 	if (offsetofend(typeof(resp), cqe_version) <= udata->outlen)
1908 		resp.response_length += sizeof(resp.cqe_version);
1909 
1910 	if (offsetofend(typeof(resp), cmds_supp_uhw) <= udata->outlen) {
1911 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1912 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1913 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1914 	}
1915 
1916 	if (offsetofend(typeof(resp), eth_min_inline) <= udata->outlen) {
1917 		if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1918 			mlx5_query_min_inline(dev->mdev, &resp.eth_min_inline);
1919 			resp.eth_min_inline++;
1920 		}
1921 		resp.response_length += sizeof(resp.eth_min_inline);
1922 	}
1923 
1924 	if (offsetofend(typeof(resp), clock_info_versions) <= udata->outlen) {
1925 		if (mdev->clock_info)
1926 			resp.clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1927 		resp.response_length += sizeof(resp.clock_info_versions);
1928 	}
1929 
1930 	/*
1931 	 * We don't want to expose information from the PCI bar that is located
1932 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1933 	 * pretend we don't support reading the HCA's core clock. This is also
1934 	 * forced by mmap function.
1935 	 */
1936 	if (offsetofend(typeof(resp), hca_core_clock_offset) <= udata->outlen) {
1937 		if (PAGE_SIZE <= 4096) {
1938 			resp.comp_mask |=
1939 				MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1940 			resp.hca_core_clock_offset =
1941 				offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1942 		}
1943 		resp.response_length += sizeof(resp.hca_core_clock_offset);
1944 	}
1945 
1946 	if (offsetofend(typeof(resp), log_uar_size) <= udata->outlen)
1947 		resp.response_length += sizeof(resp.log_uar_size);
1948 
1949 	if (offsetofend(typeof(resp), num_uars_per_page) <= udata->outlen)
1950 		resp.response_length += sizeof(resp.num_uars_per_page);
1951 
1952 	if (offsetofend(typeof(resp), num_dyn_bfregs) <= udata->outlen) {
1953 		resp.num_dyn_bfregs = bfregi->num_dyn_bfregs;
1954 		resp.response_length += sizeof(resp.num_dyn_bfregs);
1955 	}
1956 
1957 	if (offsetofend(typeof(resp), dump_fill_mkey) <= udata->outlen) {
1958 		if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1959 			resp.dump_fill_mkey = dump_fill_mkey;
1960 			resp.comp_mask |=
1961 				MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1962 		}
1963 		resp.response_length += sizeof(resp.dump_fill_mkey);
1964 	}
1965 
1966 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1967 	if (err)
1968 		goto out_mdev;
1969 
1970 	bfregi->ver = ver;
1971 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1972 	context->cqe_version = resp.cqe_version;
1973 	context->lib_caps = req.lib_caps;
1974 	print_lib_caps(dev, context->lib_caps);
1975 
1976 	if (dev->lag_active) {
1977 		u8 port = mlx5_core_native_port_num(dev->mdev) - 1;
1978 
1979 		atomic_set(&context->tx_port_affinity,
1980 			   atomic_add_return(
1981 				   1, &dev->port[port].roce.tx_port_affinity));
1982 	}
1983 
1984 	return 0;
1985 
1986 out_mdev:
1987 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1988 out_devx:
1989 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1990 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1991 
1992 out_uars:
1993 	deallocate_uars(dev, context);
1994 
1995 out_sys_pages:
1996 	kfree(bfregi->sys_pages);
1997 
1998 out_count:
1999 	kfree(bfregi->count);
2000 
2001 out_ctx:
2002 	return err;
2003 }
2004 
2005 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
2006 {
2007 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2008 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2009 	struct mlx5_bfreg_info *bfregi;
2010 
2011 	bfregi = &context->bfregi;
2012 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
2013 
2014 	if (context->devx_uid)
2015 		mlx5_ib_devx_destroy(dev, context->devx_uid);
2016 
2017 	deallocate_uars(dev, context);
2018 	kfree(bfregi->sys_pages);
2019 	kfree(bfregi->count);
2020 }
2021 
2022 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2023 				 int uar_idx)
2024 {
2025 	int fw_uars_per_page;
2026 
2027 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2028 
2029 	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2030 }
2031 
2032 static u64 uar_index2paddress(struct mlx5_ib_dev *dev,
2033 				 int uar_idx)
2034 {
2035 	unsigned int fw_uars_per_page;
2036 
2037 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
2038 				MLX5_UARS_IN_PAGE : 1;
2039 
2040 	return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE);
2041 }
2042 
2043 static int get_command(unsigned long offset)
2044 {
2045 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2046 }
2047 
2048 static int get_arg(unsigned long offset)
2049 {
2050 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2051 }
2052 
2053 static int get_index(unsigned long offset)
2054 {
2055 	return get_arg(offset);
2056 }
2057 
2058 /* Index resides in an extra byte to enable larger values than 255 */
2059 static int get_extended_index(unsigned long offset)
2060 {
2061 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2062 }
2063 
2064 
2065 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2066 {
2067 }
2068 
2069 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2070 {
2071 	switch (cmd) {
2072 	case MLX5_IB_MMAP_WC_PAGE:
2073 		return "WC";
2074 	case MLX5_IB_MMAP_REGULAR_PAGE:
2075 		return "best effort WC";
2076 	case MLX5_IB_MMAP_NC_PAGE:
2077 		return "NC";
2078 	case MLX5_IB_MMAP_DEVICE_MEM:
2079 		return "Device Memory";
2080 	default:
2081 		return NULL;
2082 	}
2083 }
2084 
2085 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2086 					struct vm_area_struct *vma,
2087 					struct mlx5_ib_ucontext *context)
2088 {
2089 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2090 	    !(vma->vm_flags & VM_SHARED))
2091 		return -EINVAL;
2092 
2093 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2094 		return -EOPNOTSUPP;
2095 
2096 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2097 		return -EPERM;
2098 	vma->vm_flags &= ~VM_MAYWRITE;
2099 
2100 	if (!dev->mdev->clock_info)
2101 		return -EOPNOTSUPP;
2102 
2103 	return vm_insert_page(vma, vma->vm_start,
2104 			      virt_to_page(dev->mdev->clock_info));
2105 }
2106 
2107 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2108 {
2109 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2110 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2111 	struct mlx5_var_table *var_table = &dev->var_table;
2112 	struct mlx5_ib_dm *mdm;
2113 
2114 	switch (mentry->mmap_flag) {
2115 	case MLX5_IB_MMAP_TYPE_MEMIC:
2116 		mdm = container_of(mentry, struct mlx5_ib_dm, mentry);
2117 		mlx5_cmd_dealloc_memic(&dev->dm, mdm->dev_addr,
2118 				       mdm->size);
2119 		kfree(mdm);
2120 		break;
2121 	case MLX5_IB_MMAP_TYPE_VAR:
2122 		mutex_lock(&var_table->bitmap_lock);
2123 		clear_bit(mentry->page_idx, var_table->bitmap);
2124 		mutex_unlock(&var_table->bitmap_lock);
2125 		kfree(mentry);
2126 		break;
2127 	case MLX5_IB_MMAP_TYPE_UAR_WC:
2128 	case MLX5_IB_MMAP_TYPE_UAR_NC:
2129 		mlx5_cmd_free_uar(dev->mdev, mentry->page_idx);
2130 		kfree(mentry);
2131 		break;
2132 	default:
2133 		WARN_ON(true);
2134 	}
2135 }
2136 
2137 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2138 		    struct vm_area_struct *vma,
2139 		    struct mlx5_ib_ucontext *context)
2140 {
2141 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2142 	int err;
2143 	unsigned long idx;
2144 	phys_addr_t pfn;
2145 	pgprot_t prot;
2146 	u32 bfreg_dyn_idx = 0;
2147 	u32 uar_index;
2148 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2149 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2150 				bfregi->num_static_sys_pages;
2151 
2152 	if (bfregi->lib_uar_dyn)
2153 		return -EINVAL;
2154 
2155 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2156 		return -EINVAL;
2157 
2158 	if (dyn_uar)
2159 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2160 	else
2161 		idx = get_index(vma->vm_pgoff);
2162 
2163 	if (idx >= max_valid_idx) {
2164 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2165 			     idx, max_valid_idx);
2166 		return -EINVAL;
2167 	}
2168 
2169 	switch (cmd) {
2170 	case MLX5_IB_MMAP_WC_PAGE:
2171 	case MLX5_IB_MMAP_ALLOC_WC:
2172 	case MLX5_IB_MMAP_REGULAR_PAGE:
2173 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2174 		prot = pgprot_writecombine(vma->vm_page_prot);
2175 		break;
2176 	case MLX5_IB_MMAP_NC_PAGE:
2177 		prot = pgprot_noncached(vma->vm_page_prot);
2178 		break;
2179 	default:
2180 		return -EINVAL;
2181 	}
2182 
2183 	if (dyn_uar) {
2184 		int uars_per_page;
2185 
2186 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2187 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2188 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2189 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2190 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2191 			return -EINVAL;
2192 		}
2193 
2194 		mutex_lock(&bfregi->lock);
2195 		/* Fail if uar already allocated, first bfreg index of each
2196 		 * page holds its count.
2197 		 */
2198 		if (bfregi->count[bfreg_dyn_idx]) {
2199 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2200 			mutex_unlock(&bfregi->lock);
2201 			return -EINVAL;
2202 		}
2203 
2204 		bfregi->count[bfreg_dyn_idx]++;
2205 		mutex_unlock(&bfregi->lock);
2206 
2207 		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2208 		if (err) {
2209 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2210 			goto free_bfreg;
2211 		}
2212 	} else {
2213 		uar_index = bfregi->sys_pages[idx];
2214 	}
2215 
2216 	pfn = uar_index2pfn(dev, uar_index);
2217 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2218 
2219 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2220 				prot, NULL);
2221 	if (err) {
2222 		mlx5_ib_err(dev,
2223 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2224 			    err, mmap_cmd2str(cmd));
2225 		goto err;
2226 	}
2227 
2228 	if (dyn_uar)
2229 		bfregi->sys_pages[idx] = uar_index;
2230 	return 0;
2231 
2232 err:
2233 	if (!dyn_uar)
2234 		return err;
2235 
2236 	mlx5_cmd_free_uar(dev->mdev, idx);
2237 
2238 free_bfreg:
2239 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2240 
2241 	return err;
2242 }
2243 
2244 static int add_dm_mmap_entry(struct ib_ucontext *context,
2245 			     struct mlx5_ib_dm *mdm,
2246 			     u64 address)
2247 {
2248 	mdm->mentry.mmap_flag = MLX5_IB_MMAP_TYPE_MEMIC;
2249 	mdm->mentry.address = address;
2250 	return rdma_user_mmap_entry_insert_range(
2251 			context, &mdm->mentry.rdma_entry,
2252 			mdm->size,
2253 			MLX5_IB_MMAP_DEVICE_MEM << 16,
2254 			(MLX5_IB_MMAP_DEVICE_MEM << 16) + (1UL << 16) - 1);
2255 }
2256 
2257 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2258 {
2259 	unsigned long idx;
2260 	u8 command;
2261 
2262 	command = get_command(vma->vm_pgoff);
2263 	idx = get_extended_index(vma->vm_pgoff);
2264 
2265 	return (command << 16 | idx);
2266 }
2267 
2268 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2269 			       struct vm_area_struct *vma,
2270 			       struct ib_ucontext *ucontext)
2271 {
2272 	struct mlx5_user_mmap_entry *mentry;
2273 	struct rdma_user_mmap_entry *entry;
2274 	unsigned long pgoff;
2275 	pgprot_t prot;
2276 	phys_addr_t pfn;
2277 	int ret;
2278 
2279 	pgoff = mlx5_vma_to_pgoff(vma);
2280 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2281 	if (!entry)
2282 		return -EINVAL;
2283 
2284 	mentry = to_mmmap(entry);
2285 	pfn = (mentry->address >> PAGE_SHIFT);
2286 	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
2287 	    mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
2288 		prot = pgprot_noncached(vma->vm_page_prot);
2289 	else
2290 		prot = pgprot_writecombine(vma->vm_page_prot);
2291 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2292 				entry->npages * PAGE_SIZE,
2293 				prot,
2294 				entry);
2295 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2296 	return ret;
2297 }
2298 
2299 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2300 {
2301 	u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2302 	u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2303 
2304 	return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2305 		(index & 0xFF)) << PAGE_SHIFT;
2306 }
2307 
2308 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2309 {
2310 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2311 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2312 	unsigned long command;
2313 	phys_addr_t pfn;
2314 
2315 	command = get_command(vma->vm_pgoff);
2316 	switch (command) {
2317 	case MLX5_IB_MMAP_WC_PAGE:
2318 	case MLX5_IB_MMAP_ALLOC_WC:
2319 		if (!dev->wc_support)
2320 			return -EPERM;
2321 		fallthrough;
2322 	case MLX5_IB_MMAP_NC_PAGE:
2323 	case MLX5_IB_MMAP_REGULAR_PAGE:
2324 		return uar_mmap(dev, command, vma, context);
2325 
2326 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2327 		return -ENOSYS;
2328 
2329 	case MLX5_IB_MMAP_CORE_CLOCK:
2330 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2331 			return -EINVAL;
2332 
2333 		if (vma->vm_flags & VM_WRITE)
2334 			return -EPERM;
2335 		vma->vm_flags &= ~VM_MAYWRITE;
2336 
2337 		/* Don't expose to user-space information it shouldn't have */
2338 		if (PAGE_SIZE > 4096)
2339 			return -EOPNOTSUPP;
2340 
2341 		pfn = (dev->mdev->iseg_base +
2342 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2343 			PAGE_SHIFT;
2344 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2345 					 PAGE_SIZE,
2346 					 pgprot_noncached(vma->vm_page_prot),
2347 					 NULL);
2348 	case MLX5_IB_MMAP_CLOCK_INFO:
2349 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2350 
2351 	default:
2352 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2353 	}
2354 
2355 	return 0;
2356 }
2357 
2358 static inline int check_dm_type_support(struct mlx5_ib_dev *dev,
2359 					u32 type)
2360 {
2361 	switch (type) {
2362 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2363 		if (!MLX5_CAP_DEV_MEM(dev->mdev, memic))
2364 			return -EOPNOTSUPP;
2365 		break;
2366 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2367 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2368 		if (!capable(CAP_SYS_RAWIO) ||
2369 		    !capable(CAP_NET_RAW))
2370 			return -EPERM;
2371 
2372 		if (!(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
2373 		      MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, sw_owner)))
2374 			return -EOPNOTSUPP;
2375 		break;
2376 	}
2377 
2378 	return 0;
2379 }
2380 
2381 static int handle_alloc_dm_memic(struct ib_ucontext *ctx,
2382 				 struct mlx5_ib_dm *dm,
2383 				 struct ib_dm_alloc_attr *attr,
2384 				 struct uverbs_attr_bundle *attrs)
2385 {
2386 	struct mlx5_dm *dm_db = &to_mdev(ctx->device)->dm;
2387 	u64 start_offset;
2388 	u16 page_idx;
2389 	int err;
2390 	u64 address;
2391 
2392 	dm->size = roundup(attr->length, MLX5_MEMIC_BASE_SIZE);
2393 
2394 	err = mlx5_cmd_alloc_memic(dm_db, &dm->dev_addr,
2395 				   dm->size, attr->alignment);
2396 	if (err)
2397 		return err;
2398 
2399 	address = dm->dev_addr & PAGE_MASK;
2400 	err = add_dm_mmap_entry(ctx, dm, address);
2401 	if (err)
2402 		goto err_dealloc;
2403 
2404 	page_idx = dm->mentry.rdma_entry.start_pgoff & 0xFFFF;
2405 	err = uverbs_copy_to(attrs,
2406 			     MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
2407 			     &page_idx,
2408 			     sizeof(page_idx));
2409 	if (err)
2410 		goto err_copy;
2411 
2412 	start_offset = dm->dev_addr & ~PAGE_MASK;
2413 	err = uverbs_copy_to(attrs,
2414 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2415 			     &start_offset, sizeof(start_offset));
2416 	if (err)
2417 		goto err_copy;
2418 
2419 	return 0;
2420 
2421 err_copy:
2422 	rdma_user_mmap_entry_remove(&dm->mentry.rdma_entry);
2423 err_dealloc:
2424 	mlx5_cmd_dealloc_memic(dm_db, dm->dev_addr, dm->size);
2425 
2426 	return err;
2427 }
2428 
2429 static int handle_alloc_dm_sw_icm(struct ib_ucontext *ctx,
2430 				  struct mlx5_ib_dm *dm,
2431 				  struct ib_dm_alloc_attr *attr,
2432 				  struct uverbs_attr_bundle *attrs,
2433 				  int type)
2434 {
2435 	struct mlx5_core_dev *dev = to_mdev(ctx->device)->mdev;
2436 	u64 act_size;
2437 	int err;
2438 
2439 	/* Allocation size must a multiple of the basic block size
2440 	 * and a power of 2.
2441 	 */
2442 	act_size = round_up(attr->length, MLX5_SW_ICM_BLOCK_SIZE(dev));
2443 	act_size = roundup_pow_of_two(act_size);
2444 
2445 	dm->size = act_size;
2446 	err = mlx5_dm_sw_icm_alloc(dev, type, act_size,
2447 				   to_mucontext(ctx)->devx_uid, &dm->dev_addr,
2448 				   &dm->icm_dm.obj_id);
2449 	if (err)
2450 		return err;
2451 
2452 	err = uverbs_copy_to(attrs,
2453 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2454 			     &dm->dev_addr, sizeof(dm->dev_addr));
2455 	if (err)
2456 		mlx5_dm_sw_icm_dealloc(dev, type, dm->size,
2457 				       to_mucontext(ctx)->devx_uid, dm->dev_addr,
2458 				       dm->icm_dm.obj_id);
2459 
2460 	return err;
2461 }
2462 
2463 struct ib_dm *mlx5_ib_alloc_dm(struct ib_device *ibdev,
2464 			       struct ib_ucontext *context,
2465 			       struct ib_dm_alloc_attr *attr,
2466 			       struct uverbs_attr_bundle *attrs)
2467 {
2468 	struct mlx5_ib_dm *dm;
2469 	enum mlx5_ib_uapi_dm_type type;
2470 	int err;
2471 
2472 	err = uverbs_get_const_default(&type, attrs,
2473 				       MLX5_IB_ATTR_ALLOC_DM_REQ_TYPE,
2474 				       MLX5_IB_UAPI_DM_TYPE_MEMIC);
2475 	if (err)
2476 		return ERR_PTR(err);
2477 
2478 	mlx5_ib_dbg(to_mdev(ibdev), "alloc_dm req: dm_type=%d user_length=0x%llx log_alignment=%d\n",
2479 		    type, attr->length, attr->alignment);
2480 
2481 	err = check_dm_type_support(to_mdev(ibdev), type);
2482 	if (err)
2483 		return ERR_PTR(err);
2484 
2485 	dm = kzalloc(sizeof(*dm), GFP_KERNEL);
2486 	if (!dm)
2487 		return ERR_PTR(-ENOMEM);
2488 
2489 	dm->type = type;
2490 
2491 	switch (type) {
2492 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2493 		err = handle_alloc_dm_memic(context, dm,
2494 					    attr,
2495 					    attrs);
2496 		break;
2497 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2498 		err = handle_alloc_dm_sw_icm(context, dm,
2499 					     attr, attrs,
2500 					     MLX5_SW_ICM_TYPE_STEERING);
2501 		break;
2502 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2503 		err = handle_alloc_dm_sw_icm(context, dm,
2504 					     attr, attrs,
2505 					     MLX5_SW_ICM_TYPE_HEADER_MODIFY);
2506 		break;
2507 	default:
2508 		err = -EOPNOTSUPP;
2509 	}
2510 
2511 	if (err)
2512 		goto err_free;
2513 
2514 	return &dm->ibdm;
2515 
2516 err_free:
2517 	kfree(dm);
2518 	return ERR_PTR(err);
2519 }
2520 
2521 int mlx5_ib_dealloc_dm(struct ib_dm *ibdm, struct uverbs_attr_bundle *attrs)
2522 {
2523 	struct mlx5_ib_ucontext *ctx = rdma_udata_to_drv_context(
2524 		&attrs->driver_udata, struct mlx5_ib_ucontext, ibucontext);
2525 	struct mlx5_core_dev *dev = to_mdev(ibdm->device)->mdev;
2526 	struct mlx5_ib_dm *dm = to_mdm(ibdm);
2527 	int ret;
2528 
2529 	switch (dm->type) {
2530 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2531 		rdma_user_mmap_entry_remove(&dm->mentry.rdma_entry);
2532 		return 0;
2533 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2534 		ret = mlx5_dm_sw_icm_dealloc(dev, MLX5_SW_ICM_TYPE_STEERING,
2535 					     dm->size, ctx->devx_uid, dm->dev_addr,
2536 					     dm->icm_dm.obj_id);
2537 		if (ret)
2538 			return ret;
2539 		break;
2540 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2541 		ret = mlx5_dm_sw_icm_dealloc(dev, MLX5_SW_ICM_TYPE_HEADER_MODIFY,
2542 					     dm->size, ctx->devx_uid, dm->dev_addr,
2543 					     dm->icm_dm.obj_id);
2544 		if (ret)
2545 			return ret;
2546 		break;
2547 	default:
2548 		return -EOPNOTSUPP;
2549 	}
2550 
2551 	kfree(dm);
2552 
2553 	return 0;
2554 }
2555 
2556 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2557 {
2558 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2559 	struct ib_device *ibdev = ibpd->device;
2560 	struct mlx5_ib_alloc_pd_resp resp;
2561 	int err;
2562 	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2563 	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)]   = {};
2564 	u16 uid = 0;
2565 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2566 		udata, struct mlx5_ib_ucontext, ibucontext);
2567 
2568 	uid = context ? context->devx_uid : 0;
2569 	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2570 	MLX5_SET(alloc_pd_in, in, uid, uid);
2571 	err = mlx5_cmd_exec(to_mdev(ibdev)->mdev, in, sizeof(in),
2572 			    out, sizeof(out));
2573 	if (err)
2574 		return err;
2575 
2576 	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2577 	pd->uid = uid;
2578 	if (udata) {
2579 		resp.pdn = pd->pdn;
2580 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2581 			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2582 			return -EFAULT;
2583 		}
2584 	}
2585 
2586 	return 0;
2587 }
2588 
2589 static void mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2590 {
2591 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2592 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2593 
2594 	mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2595 }
2596 
2597 enum {
2598 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
2599 	MATCH_CRITERIA_ENABLE_MISC_BIT,
2600 	MATCH_CRITERIA_ENABLE_INNER_BIT,
2601 	MATCH_CRITERIA_ENABLE_MISC2_BIT
2602 };
2603 
2604 #define HEADER_IS_ZERO(match_criteria, headers)			           \
2605 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
2606 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
2607 
2608 static u8 get_match_criteria_enable(u32 *match_criteria)
2609 {
2610 	u8 match_criteria_enable;
2611 
2612 	match_criteria_enable =
2613 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
2614 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
2615 	match_criteria_enable |=
2616 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
2617 		MATCH_CRITERIA_ENABLE_MISC_BIT;
2618 	match_criteria_enable |=
2619 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
2620 		MATCH_CRITERIA_ENABLE_INNER_BIT;
2621 	match_criteria_enable |=
2622 		(!HEADER_IS_ZERO(match_criteria, misc_parameters_2)) <<
2623 		MATCH_CRITERIA_ENABLE_MISC2_BIT;
2624 
2625 	return match_criteria_enable;
2626 }
2627 
2628 static int set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
2629 {
2630 	u8 entry_mask;
2631 	u8 entry_val;
2632 	int err = 0;
2633 
2634 	if (!mask)
2635 		goto out;
2636 
2637 	entry_mask = MLX5_GET(fte_match_set_lyr_2_4, outer_c,
2638 			      ip_protocol);
2639 	entry_val = MLX5_GET(fte_match_set_lyr_2_4, outer_v,
2640 			     ip_protocol);
2641 	if (!entry_mask) {
2642 		MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
2643 		MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
2644 		goto out;
2645 	}
2646 	/* Don't override existing ip protocol */
2647 	if (mask != entry_mask || val != entry_val)
2648 		err = -EINVAL;
2649 out:
2650 	return err;
2651 }
2652 
2653 static void set_flow_label(void *misc_c, void *misc_v, u32 mask, u32 val,
2654 			   bool inner)
2655 {
2656 	if (inner) {
2657 		MLX5_SET(fte_match_set_misc,
2658 			 misc_c, inner_ipv6_flow_label, mask);
2659 		MLX5_SET(fte_match_set_misc,
2660 			 misc_v, inner_ipv6_flow_label, val);
2661 	} else {
2662 		MLX5_SET(fte_match_set_misc,
2663 			 misc_c, outer_ipv6_flow_label, mask);
2664 		MLX5_SET(fte_match_set_misc,
2665 			 misc_v, outer_ipv6_flow_label, val);
2666 	}
2667 }
2668 
2669 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
2670 {
2671 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
2672 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
2673 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
2674 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
2675 }
2676 
2677 static int check_mpls_supp_fields(u32 field_support, const __be32 *set_mask)
2678 {
2679 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_label) &&
2680 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_LABEL))
2681 		return -EOPNOTSUPP;
2682 
2683 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_exp) &&
2684 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_EXP))
2685 		return -EOPNOTSUPP;
2686 
2687 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_s_bos) &&
2688 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_S_BOS))
2689 		return -EOPNOTSUPP;
2690 
2691 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_ttl) &&
2692 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_TTL))
2693 		return -EOPNOTSUPP;
2694 
2695 	return 0;
2696 }
2697 
2698 #define LAST_ETH_FIELD vlan_tag
2699 #define LAST_IB_FIELD sl
2700 #define LAST_IPV4_FIELD tos
2701 #define LAST_IPV6_FIELD traffic_class
2702 #define LAST_TCP_UDP_FIELD src_port
2703 #define LAST_TUNNEL_FIELD tunnel_id
2704 #define LAST_FLOW_TAG_FIELD tag_id
2705 #define LAST_DROP_FIELD size
2706 #define LAST_COUNTERS_FIELD counters
2707 
2708 /* Field is the last supported field */
2709 #define FIELDS_NOT_SUPPORTED(filter, field)\
2710 	memchr_inv((void *)&filter.field  +\
2711 		   sizeof(filter.field), 0,\
2712 		   sizeof(filter) -\
2713 		   offsetof(typeof(filter), field) -\
2714 		   sizeof(filter.field))
2715 
2716 int parse_flow_flow_action(struct mlx5_ib_flow_action *maction,
2717 			   bool is_egress,
2718 			   struct mlx5_flow_act *action)
2719 {
2720 
2721 	switch (maction->ib_action.type) {
2722 	case IB_FLOW_ACTION_ESP:
2723 		if (action->action & (MLX5_FLOW_CONTEXT_ACTION_ENCRYPT |
2724 				      MLX5_FLOW_CONTEXT_ACTION_DECRYPT))
2725 			return -EINVAL;
2726 		/* Currently only AES_GCM keymat is supported by the driver */
2727 		action->esp_id = (uintptr_t)maction->esp_aes_gcm.ctx;
2728 		action->action |= is_egress ?
2729 			MLX5_FLOW_CONTEXT_ACTION_ENCRYPT :
2730 			MLX5_FLOW_CONTEXT_ACTION_DECRYPT;
2731 		return 0;
2732 	case IB_FLOW_ACTION_UNSPECIFIED:
2733 		if (maction->flow_action_raw.sub_type ==
2734 		    MLX5_IB_FLOW_ACTION_MODIFY_HEADER) {
2735 			if (action->action & MLX5_FLOW_CONTEXT_ACTION_MOD_HDR)
2736 				return -EINVAL;
2737 			action->action |= MLX5_FLOW_CONTEXT_ACTION_MOD_HDR;
2738 			action->modify_hdr =
2739 				maction->flow_action_raw.modify_hdr;
2740 			return 0;
2741 		}
2742 		if (maction->flow_action_raw.sub_type ==
2743 		    MLX5_IB_FLOW_ACTION_DECAP) {
2744 			if (action->action & MLX5_FLOW_CONTEXT_ACTION_DECAP)
2745 				return -EINVAL;
2746 			action->action |= MLX5_FLOW_CONTEXT_ACTION_DECAP;
2747 			return 0;
2748 		}
2749 		if (maction->flow_action_raw.sub_type ==
2750 		    MLX5_IB_FLOW_ACTION_PACKET_REFORMAT) {
2751 			if (action->action &
2752 			    MLX5_FLOW_CONTEXT_ACTION_PACKET_REFORMAT)
2753 				return -EINVAL;
2754 			action->action |=
2755 				MLX5_FLOW_CONTEXT_ACTION_PACKET_REFORMAT;
2756 			action->pkt_reformat =
2757 				maction->flow_action_raw.pkt_reformat;
2758 			return 0;
2759 		}
2760 		/* fall through */
2761 	default:
2762 		return -EOPNOTSUPP;
2763 	}
2764 }
2765 
2766 static int parse_flow_attr(struct mlx5_core_dev *mdev,
2767 			   struct mlx5_flow_spec *spec,
2768 			   const union ib_flow_spec *ib_spec,
2769 			   const struct ib_flow_attr *flow_attr,
2770 			   struct mlx5_flow_act *action, u32 prev_type)
2771 {
2772 	struct mlx5_flow_context *flow_context = &spec->flow_context;
2773 	u32 *match_c = spec->match_criteria;
2774 	u32 *match_v = spec->match_value;
2775 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
2776 					   misc_parameters);
2777 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
2778 					   misc_parameters);
2779 	void *misc_params2_c = MLX5_ADDR_OF(fte_match_param, match_c,
2780 					    misc_parameters_2);
2781 	void *misc_params2_v = MLX5_ADDR_OF(fte_match_param, match_v,
2782 					    misc_parameters_2);
2783 	void *headers_c;
2784 	void *headers_v;
2785 	int match_ipv;
2786 	int ret;
2787 
2788 	if (ib_spec->type & IB_FLOW_SPEC_INNER) {
2789 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2790 					 inner_headers);
2791 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2792 					 inner_headers);
2793 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2794 					ft_field_support.inner_ip_version);
2795 	} else {
2796 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2797 					 outer_headers);
2798 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2799 					 outer_headers);
2800 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2801 					ft_field_support.outer_ip_version);
2802 	}
2803 
2804 	switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
2805 	case IB_FLOW_SPEC_ETH:
2806 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
2807 			return -EOPNOTSUPP;
2808 
2809 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2810 					     dmac_47_16),
2811 				ib_spec->eth.mask.dst_mac);
2812 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2813 					     dmac_47_16),
2814 				ib_spec->eth.val.dst_mac);
2815 
2816 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2817 					     smac_47_16),
2818 				ib_spec->eth.mask.src_mac);
2819 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2820 					     smac_47_16),
2821 				ib_spec->eth.val.src_mac);
2822 
2823 		if (ib_spec->eth.mask.vlan_tag) {
2824 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2825 				 cvlan_tag, 1);
2826 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2827 				 cvlan_tag, 1);
2828 
2829 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2830 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
2831 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2832 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
2833 
2834 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2835 				 first_cfi,
2836 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
2837 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2838 				 first_cfi,
2839 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
2840 
2841 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2842 				 first_prio,
2843 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
2844 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2845 				 first_prio,
2846 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
2847 		}
2848 		MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2849 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
2850 		MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2851 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
2852 		break;
2853 	case IB_FLOW_SPEC_IPV4:
2854 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
2855 			return -EOPNOTSUPP;
2856 
2857 		if (match_ipv) {
2858 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2859 				 ip_version, 0xf);
2860 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2861 				 ip_version, MLX5_FS_IPV4_VERSION);
2862 		} else {
2863 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2864 				 ethertype, 0xffff);
2865 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2866 				 ethertype, ETH_P_IP);
2867 		}
2868 
2869 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2870 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
2871 		       &ib_spec->ipv4.mask.src_ip,
2872 		       sizeof(ib_spec->ipv4.mask.src_ip));
2873 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2874 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
2875 		       &ib_spec->ipv4.val.src_ip,
2876 		       sizeof(ib_spec->ipv4.val.src_ip));
2877 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2878 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2879 		       &ib_spec->ipv4.mask.dst_ip,
2880 		       sizeof(ib_spec->ipv4.mask.dst_ip));
2881 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2882 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2883 		       &ib_spec->ipv4.val.dst_ip,
2884 		       sizeof(ib_spec->ipv4.val.dst_ip));
2885 
2886 		set_tos(headers_c, headers_v,
2887 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
2888 
2889 		if (set_proto(headers_c, headers_v,
2890 			      ib_spec->ipv4.mask.proto,
2891 			      ib_spec->ipv4.val.proto))
2892 			return -EINVAL;
2893 		break;
2894 	case IB_FLOW_SPEC_IPV6:
2895 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
2896 			return -EOPNOTSUPP;
2897 
2898 		if (match_ipv) {
2899 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2900 				 ip_version, 0xf);
2901 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2902 				 ip_version, MLX5_FS_IPV6_VERSION);
2903 		} else {
2904 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2905 				 ethertype, 0xffff);
2906 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2907 				 ethertype, ETH_P_IPV6);
2908 		}
2909 
2910 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2911 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
2912 		       &ib_spec->ipv6.mask.src_ip,
2913 		       sizeof(ib_spec->ipv6.mask.src_ip));
2914 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2915 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
2916 		       &ib_spec->ipv6.val.src_ip,
2917 		       sizeof(ib_spec->ipv6.val.src_ip));
2918 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2919 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2920 		       &ib_spec->ipv6.mask.dst_ip,
2921 		       sizeof(ib_spec->ipv6.mask.dst_ip));
2922 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2923 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2924 		       &ib_spec->ipv6.val.dst_ip,
2925 		       sizeof(ib_spec->ipv6.val.dst_ip));
2926 
2927 		set_tos(headers_c, headers_v,
2928 			ib_spec->ipv6.mask.traffic_class,
2929 			ib_spec->ipv6.val.traffic_class);
2930 
2931 		if (set_proto(headers_c, headers_v,
2932 			      ib_spec->ipv6.mask.next_hdr,
2933 			      ib_spec->ipv6.val.next_hdr))
2934 			return -EINVAL;
2935 
2936 		set_flow_label(misc_params_c, misc_params_v,
2937 			       ntohl(ib_spec->ipv6.mask.flow_label),
2938 			       ntohl(ib_spec->ipv6.val.flow_label),
2939 			       ib_spec->type & IB_FLOW_SPEC_INNER);
2940 		break;
2941 	case IB_FLOW_SPEC_ESP:
2942 		if (ib_spec->esp.mask.seq)
2943 			return -EOPNOTSUPP;
2944 
2945 		MLX5_SET(fte_match_set_misc, misc_params_c, outer_esp_spi,
2946 			 ntohl(ib_spec->esp.mask.spi));
2947 		MLX5_SET(fte_match_set_misc, misc_params_v, outer_esp_spi,
2948 			 ntohl(ib_spec->esp.val.spi));
2949 		break;
2950 	case IB_FLOW_SPEC_TCP:
2951 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2952 					 LAST_TCP_UDP_FIELD))
2953 			return -EOPNOTSUPP;
2954 
2955 		if (set_proto(headers_c, headers_v, 0xff, IPPROTO_TCP))
2956 			return -EINVAL;
2957 
2958 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_sport,
2959 			 ntohs(ib_spec->tcp_udp.mask.src_port));
2960 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_sport,
2961 			 ntohs(ib_spec->tcp_udp.val.src_port));
2962 
2963 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_dport,
2964 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
2965 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_dport,
2966 			 ntohs(ib_spec->tcp_udp.val.dst_port));
2967 		break;
2968 	case IB_FLOW_SPEC_UDP:
2969 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2970 					 LAST_TCP_UDP_FIELD))
2971 			return -EOPNOTSUPP;
2972 
2973 		if (set_proto(headers_c, headers_v, 0xff, IPPROTO_UDP))
2974 			return -EINVAL;
2975 
2976 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_sport,
2977 			 ntohs(ib_spec->tcp_udp.mask.src_port));
2978 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_sport,
2979 			 ntohs(ib_spec->tcp_udp.val.src_port));
2980 
2981 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_dport,
2982 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
2983 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_dport,
2984 			 ntohs(ib_spec->tcp_udp.val.dst_port));
2985 		break;
2986 	case IB_FLOW_SPEC_GRE:
2987 		if (ib_spec->gre.mask.c_ks_res0_ver)
2988 			return -EOPNOTSUPP;
2989 
2990 		if (set_proto(headers_c, headers_v, 0xff, IPPROTO_GRE))
2991 			return -EINVAL;
2992 
2993 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2994 			 0xff);
2995 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2996 			 IPPROTO_GRE);
2997 
2998 		MLX5_SET(fte_match_set_misc, misc_params_c, gre_protocol,
2999 			 ntohs(ib_spec->gre.mask.protocol));
3000 		MLX5_SET(fte_match_set_misc, misc_params_v, gre_protocol,
3001 			 ntohs(ib_spec->gre.val.protocol));
3002 
3003 		memcpy(MLX5_ADDR_OF(fte_match_set_misc, misc_params_c,
3004 				    gre_key.nvgre.hi),
3005 		       &ib_spec->gre.mask.key,
3006 		       sizeof(ib_spec->gre.mask.key));
3007 		memcpy(MLX5_ADDR_OF(fte_match_set_misc, misc_params_v,
3008 				    gre_key.nvgre.hi),
3009 		       &ib_spec->gre.val.key,
3010 		       sizeof(ib_spec->gre.val.key));
3011 		break;
3012 	case IB_FLOW_SPEC_MPLS:
3013 		switch (prev_type) {
3014 		case IB_FLOW_SPEC_UDP:
3015 			if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
3016 						   ft_field_support.outer_first_mpls_over_udp),
3017 						   &ib_spec->mpls.mask.tag))
3018 				return -EOPNOTSUPP;
3019 
3020 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
3021 					    outer_first_mpls_over_udp),
3022 			       &ib_spec->mpls.val.tag,
3023 			       sizeof(ib_spec->mpls.val.tag));
3024 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
3025 					    outer_first_mpls_over_udp),
3026 			       &ib_spec->mpls.mask.tag,
3027 			       sizeof(ib_spec->mpls.mask.tag));
3028 			break;
3029 		case IB_FLOW_SPEC_GRE:
3030 			if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
3031 						   ft_field_support.outer_first_mpls_over_gre),
3032 						   &ib_spec->mpls.mask.tag))
3033 				return -EOPNOTSUPP;
3034 
3035 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
3036 					    outer_first_mpls_over_gre),
3037 			       &ib_spec->mpls.val.tag,
3038 			       sizeof(ib_spec->mpls.val.tag));
3039 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
3040 					    outer_first_mpls_over_gre),
3041 			       &ib_spec->mpls.mask.tag,
3042 			       sizeof(ib_spec->mpls.mask.tag));
3043 			break;
3044 		default:
3045 			if (ib_spec->type & IB_FLOW_SPEC_INNER) {
3046 				if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
3047 							   ft_field_support.inner_first_mpls),
3048 							   &ib_spec->mpls.mask.tag))
3049 					return -EOPNOTSUPP;
3050 
3051 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
3052 						    inner_first_mpls),
3053 				       &ib_spec->mpls.val.tag,
3054 				       sizeof(ib_spec->mpls.val.tag));
3055 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
3056 						    inner_first_mpls),
3057 				       &ib_spec->mpls.mask.tag,
3058 				       sizeof(ib_spec->mpls.mask.tag));
3059 			} else {
3060 				if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
3061 							   ft_field_support.outer_first_mpls),
3062 							   &ib_spec->mpls.mask.tag))
3063 					return -EOPNOTSUPP;
3064 
3065 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
3066 						    outer_first_mpls),
3067 				       &ib_spec->mpls.val.tag,
3068 				       sizeof(ib_spec->mpls.val.tag));
3069 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
3070 						    outer_first_mpls),
3071 				       &ib_spec->mpls.mask.tag,
3072 				       sizeof(ib_spec->mpls.mask.tag));
3073 			}
3074 		}
3075 		break;
3076 	case IB_FLOW_SPEC_VXLAN_TUNNEL:
3077 		if (FIELDS_NOT_SUPPORTED(ib_spec->tunnel.mask,
3078 					 LAST_TUNNEL_FIELD))
3079 			return -EOPNOTSUPP;
3080 
3081 		MLX5_SET(fte_match_set_misc, misc_params_c, vxlan_vni,
3082 			 ntohl(ib_spec->tunnel.mask.tunnel_id));
3083 		MLX5_SET(fte_match_set_misc, misc_params_v, vxlan_vni,
3084 			 ntohl(ib_spec->tunnel.val.tunnel_id));
3085 		break;
3086 	case IB_FLOW_SPEC_ACTION_TAG:
3087 		if (FIELDS_NOT_SUPPORTED(ib_spec->flow_tag,
3088 					 LAST_FLOW_TAG_FIELD))
3089 			return -EOPNOTSUPP;
3090 		if (ib_spec->flow_tag.tag_id >= BIT(24))
3091 			return -EINVAL;
3092 
3093 		flow_context->flow_tag = ib_spec->flow_tag.tag_id;
3094 		flow_context->flags |= FLOW_CONTEXT_HAS_TAG;
3095 		break;
3096 	case IB_FLOW_SPEC_ACTION_DROP:
3097 		if (FIELDS_NOT_SUPPORTED(ib_spec->drop,
3098 					 LAST_DROP_FIELD))
3099 			return -EOPNOTSUPP;
3100 		action->action |= MLX5_FLOW_CONTEXT_ACTION_DROP;
3101 		break;
3102 	case IB_FLOW_SPEC_ACTION_HANDLE:
3103 		ret = parse_flow_flow_action(to_mflow_act(ib_spec->action.act),
3104 			flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS, action);
3105 		if (ret)
3106 			return ret;
3107 		break;
3108 	case IB_FLOW_SPEC_ACTION_COUNT:
3109 		if (FIELDS_NOT_SUPPORTED(ib_spec->flow_count,
3110 					 LAST_COUNTERS_FIELD))
3111 			return -EOPNOTSUPP;
3112 
3113 		/* for now support only one counters spec per flow */
3114 		if (action->action & MLX5_FLOW_CONTEXT_ACTION_COUNT)
3115 			return -EINVAL;
3116 
3117 		action->counters = ib_spec->flow_count.counters;
3118 		action->action |= MLX5_FLOW_CONTEXT_ACTION_COUNT;
3119 		break;
3120 	default:
3121 		return -EINVAL;
3122 	}
3123 
3124 	return 0;
3125 }
3126 
3127 /* If a flow could catch both multicast and unicast packets,
3128  * it won't fall into the multicast flow steering table and this rule
3129  * could steal other multicast packets.
3130  */
3131 static bool flow_is_multicast_only(const struct ib_flow_attr *ib_attr)
3132 {
3133 	union ib_flow_spec *flow_spec;
3134 
3135 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
3136 	    ib_attr->num_of_specs < 1)
3137 		return false;
3138 
3139 	flow_spec = (union ib_flow_spec *)(ib_attr + 1);
3140 	if (flow_spec->type == IB_FLOW_SPEC_IPV4) {
3141 		struct ib_flow_spec_ipv4 *ipv4_spec;
3142 
3143 		ipv4_spec = (struct ib_flow_spec_ipv4 *)flow_spec;
3144 		if (ipv4_is_multicast(ipv4_spec->val.dst_ip))
3145 			return true;
3146 
3147 		return false;
3148 	}
3149 
3150 	if (flow_spec->type == IB_FLOW_SPEC_ETH) {
3151 		struct ib_flow_spec_eth *eth_spec;
3152 
3153 		eth_spec = (struct ib_flow_spec_eth *)flow_spec;
3154 		return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
3155 		       is_multicast_ether_addr(eth_spec->val.dst_mac);
3156 	}
3157 
3158 	return false;
3159 }
3160 
3161 enum valid_spec {
3162 	VALID_SPEC_INVALID,
3163 	VALID_SPEC_VALID,
3164 	VALID_SPEC_NA,
3165 };
3166 
3167 static enum valid_spec
3168 is_valid_esp_aes_gcm(struct mlx5_core_dev *mdev,
3169 		     const struct mlx5_flow_spec *spec,
3170 		     const struct mlx5_flow_act *flow_act,
3171 		     bool egress)
3172 {
3173 	const u32 *match_c = spec->match_criteria;
3174 	bool is_crypto =
3175 		(flow_act->action & (MLX5_FLOW_CONTEXT_ACTION_ENCRYPT |
3176 				     MLX5_FLOW_CONTEXT_ACTION_DECRYPT));
3177 	bool is_ipsec = mlx5_fs_is_ipsec_flow(match_c);
3178 	bool is_drop = flow_act->action & MLX5_FLOW_CONTEXT_ACTION_DROP;
3179 
3180 	/*
3181 	 * Currently only crypto is supported in egress, when regular egress
3182 	 * rules would be supported, always return VALID_SPEC_NA.
3183 	 */
3184 	if (!is_crypto)
3185 		return VALID_SPEC_NA;
3186 
3187 	return is_crypto && is_ipsec &&
3188 		(!egress || (!is_drop &&
3189 			     !(spec->flow_context.flags & FLOW_CONTEXT_HAS_TAG))) ?
3190 		VALID_SPEC_VALID : VALID_SPEC_INVALID;
3191 }
3192 
3193 static bool is_valid_spec(struct mlx5_core_dev *mdev,
3194 			  const struct mlx5_flow_spec *spec,
3195 			  const struct mlx5_flow_act *flow_act,
3196 			  bool egress)
3197 {
3198 	/* We curretly only support ipsec egress flow */
3199 	return is_valid_esp_aes_gcm(mdev, spec, flow_act, egress) != VALID_SPEC_INVALID;
3200 }
3201 
3202 static bool is_valid_ethertype(struct mlx5_core_dev *mdev,
3203 			       const struct ib_flow_attr *flow_attr,
3204 			       bool check_inner)
3205 {
3206 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
3207 	int match_ipv = check_inner ?
3208 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
3209 					ft_field_support.inner_ip_version) :
3210 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
3211 					ft_field_support.outer_ip_version);
3212 	int inner_bit = check_inner ? IB_FLOW_SPEC_INNER : 0;
3213 	bool ipv4_spec_valid, ipv6_spec_valid;
3214 	unsigned int ip_spec_type = 0;
3215 	bool has_ethertype = false;
3216 	unsigned int spec_index;
3217 	bool mask_valid = true;
3218 	u16 eth_type = 0;
3219 	bool type_valid;
3220 
3221 	/* Validate that ethertype is correct */
3222 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
3223 		if ((ib_spec->type == (IB_FLOW_SPEC_ETH | inner_bit)) &&
3224 		    ib_spec->eth.mask.ether_type) {
3225 			mask_valid = (ib_spec->eth.mask.ether_type ==
3226 				      htons(0xffff));
3227 			has_ethertype = true;
3228 			eth_type = ntohs(ib_spec->eth.val.ether_type);
3229 		} else if ((ib_spec->type == (IB_FLOW_SPEC_IPV4 | inner_bit)) ||
3230 			   (ib_spec->type == (IB_FLOW_SPEC_IPV6 | inner_bit))) {
3231 			ip_spec_type = ib_spec->type;
3232 		}
3233 		ib_spec = (void *)ib_spec + ib_spec->size;
3234 	}
3235 
3236 	type_valid = (!has_ethertype) || (!ip_spec_type);
3237 	if (!type_valid && mask_valid) {
3238 		ipv4_spec_valid = (eth_type == ETH_P_IP) &&
3239 			(ip_spec_type == (IB_FLOW_SPEC_IPV4 | inner_bit));
3240 		ipv6_spec_valid = (eth_type == ETH_P_IPV6) &&
3241 			(ip_spec_type == (IB_FLOW_SPEC_IPV6 | inner_bit));
3242 
3243 		type_valid = (ipv4_spec_valid) || (ipv6_spec_valid) ||
3244 			     (((eth_type == ETH_P_MPLS_UC) ||
3245 			       (eth_type == ETH_P_MPLS_MC)) && match_ipv);
3246 	}
3247 
3248 	return type_valid;
3249 }
3250 
3251 static bool is_valid_attr(struct mlx5_core_dev *mdev,
3252 			  const struct ib_flow_attr *flow_attr)
3253 {
3254 	return is_valid_ethertype(mdev, flow_attr, false) &&
3255 	       is_valid_ethertype(mdev, flow_attr, true);
3256 }
3257 
3258 static void put_flow_table(struct mlx5_ib_dev *dev,
3259 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
3260 {
3261 	prio->refcount -= !!ft_added;
3262 	if (!prio->refcount) {
3263 		mlx5_destroy_flow_table(prio->flow_table);
3264 		prio->flow_table = NULL;
3265 	}
3266 }
3267 
3268 static void counters_clear_description(struct ib_counters *counters)
3269 {
3270 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
3271 
3272 	mutex_lock(&mcounters->mcntrs_mutex);
3273 	kfree(mcounters->counters_data);
3274 	mcounters->counters_data = NULL;
3275 	mcounters->cntrs_max_index = 0;
3276 	mutex_unlock(&mcounters->mcntrs_mutex);
3277 }
3278 
3279 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
3280 {
3281 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
3282 							  struct mlx5_ib_flow_handler,
3283 							  ibflow);
3284 	struct mlx5_ib_flow_handler *iter, *tmp;
3285 	struct mlx5_ib_dev *dev = handler->dev;
3286 
3287 	mutex_lock(&dev->flow_db->lock);
3288 
3289 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
3290 		mlx5_del_flow_rules(iter->rule);
3291 		put_flow_table(dev, iter->prio, true);
3292 		list_del(&iter->list);
3293 		kfree(iter);
3294 	}
3295 
3296 	mlx5_del_flow_rules(handler->rule);
3297 	put_flow_table(dev, handler->prio, true);
3298 	if (handler->ibcounters &&
3299 	    atomic_read(&handler->ibcounters->usecnt) == 1)
3300 		counters_clear_description(handler->ibcounters);
3301 
3302 	mutex_unlock(&dev->flow_db->lock);
3303 	if (handler->flow_matcher)
3304 		atomic_dec(&handler->flow_matcher->usecnt);
3305 	kfree(handler);
3306 
3307 	return 0;
3308 }
3309 
3310 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
3311 {
3312 	priority *= 2;
3313 	if (!dont_trap)
3314 		priority++;
3315 	return priority;
3316 }
3317 
3318 enum flow_table_type {
3319 	MLX5_IB_FT_RX,
3320 	MLX5_IB_FT_TX
3321 };
3322 
3323 #define MLX5_FS_MAX_TYPES	 6
3324 #define MLX5_FS_MAX_ENTRIES	 BIT(16)
3325 
3326 static struct mlx5_ib_flow_prio *_get_prio(struct mlx5_flow_namespace *ns,
3327 					   struct mlx5_ib_flow_prio *prio,
3328 					   int priority,
3329 					   int num_entries, int num_groups,
3330 					   u32 flags)
3331 {
3332 	struct mlx5_flow_table_attr ft_attr = {};
3333 	struct mlx5_flow_table *ft;
3334 
3335 	ft_attr.prio = priority;
3336 	ft_attr.max_fte = num_entries;
3337 	ft_attr.flags = flags;
3338 	ft_attr.autogroup.max_num_groups = num_groups;
3339 	ft = mlx5_create_auto_grouped_flow_table(ns, &ft_attr);
3340 	if (IS_ERR(ft))
3341 		return ERR_CAST(ft);
3342 
3343 	prio->flow_table = ft;
3344 	prio->refcount = 0;
3345 	return prio;
3346 }
3347 
3348 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
3349 						struct ib_flow_attr *flow_attr,
3350 						enum flow_table_type ft_type)
3351 {
3352 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
3353 	struct mlx5_flow_namespace *ns = NULL;
3354 	struct mlx5_ib_flow_prio *prio;
3355 	struct mlx5_flow_table *ft;
3356 	int max_table_size;
3357 	int num_entries;
3358 	int num_groups;
3359 	bool esw_encap;
3360 	u32 flags = 0;
3361 	int priority;
3362 
3363 	max_table_size = BIT(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3364 						       log_max_ft_size));
3365 	esw_encap = mlx5_eswitch_get_encap_mode(dev->mdev) !=
3366 		DEVLINK_ESWITCH_ENCAP_MODE_NONE;
3367 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
3368 		enum mlx5_flow_namespace_type fn_type;
3369 
3370 		if (flow_is_multicast_only(flow_attr) &&
3371 		    !dont_trap)
3372 			priority = MLX5_IB_FLOW_MCAST_PRIO;
3373 		else
3374 			priority = ib_prio_to_core_prio(flow_attr->priority,
3375 							dont_trap);
3376 		if (ft_type == MLX5_IB_FT_RX) {
3377 			fn_type = MLX5_FLOW_NAMESPACE_BYPASS;
3378 			prio = &dev->flow_db->prios[priority];
3379 			if (!dev->is_rep && !esw_encap &&
3380 			    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, decap))
3381 				flags |= MLX5_FLOW_TABLE_TUNNEL_EN_DECAP;
3382 			if (!dev->is_rep && !esw_encap &&
3383 			    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3384 					reformat_l3_tunnel_to_l2))
3385 				flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
3386 		} else {
3387 			max_table_size =
3388 				BIT(MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev,
3389 							      log_max_ft_size));
3390 			fn_type = MLX5_FLOW_NAMESPACE_EGRESS;
3391 			prio = &dev->flow_db->egress_prios[priority];
3392 			if (!dev->is_rep && !esw_encap &&
3393 			    MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, reformat))
3394 				flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
3395 		}
3396 		ns = mlx5_get_flow_namespace(dev->mdev, fn_type);
3397 		num_entries = MLX5_FS_MAX_ENTRIES;
3398 		num_groups = MLX5_FS_MAX_TYPES;
3399 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3400 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
3401 		ns = mlx5_get_flow_namespace(dev->mdev,
3402 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
3403 		build_leftovers_ft_param(&priority,
3404 					 &num_entries,
3405 					 &num_groups);
3406 		prio = &dev->flow_db->prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
3407 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3408 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
3409 					allow_sniffer_and_nic_rx_shared_tir))
3410 			return ERR_PTR(-ENOTSUPP);
3411 
3412 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
3413 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
3414 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
3415 
3416 		prio = &dev->flow_db->sniffer[ft_type];
3417 		priority = 0;
3418 		num_entries = 1;
3419 		num_groups = 1;
3420 	}
3421 
3422 	if (!ns)
3423 		return ERR_PTR(-ENOTSUPP);
3424 
3425 	max_table_size = min_t(int, num_entries, max_table_size);
3426 
3427 	ft = prio->flow_table;
3428 	if (!ft)
3429 		return _get_prio(ns, prio, priority, max_table_size, num_groups,
3430 				 flags);
3431 
3432 	return prio;
3433 }
3434 
3435 static void set_underlay_qp(struct mlx5_ib_dev *dev,
3436 			    struct mlx5_flow_spec *spec,
3437 			    u32 underlay_qpn)
3438 {
3439 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param,
3440 					   spec->match_criteria,
3441 					   misc_parameters);
3442 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3443 					   misc_parameters);
3444 
3445 	if (underlay_qpn &&
3446 	    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3447 				      ft_field_support.bth_dst_qp)) {
3448 		MLX5_SET(fte_match_set_misc,
3449 			 misc_params_v, bth_dst_qp, underlay_qpn);
3450 		MLX5_SET(fte_match_set_misc,
3451 			 misc_params_c, bth_dst_qp, 0xffffff);
3452 	}
3453 }
3454 
3455 static int read_flow_counters(struct ib_device *ibdev,
3456 			      struct mlx5_read_counters_attr *read_attr)
3457 {
3458 	struct mlx5_fc *fc = read_attr->hw_cntrs_hndl;
3459 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3460 
3461 	return mlx5_fc_query(dev->mdev, fc,
3462 			     &read_attr->out[IB_COUNTER_PACKETS],
3463 			     &read_attr->out[IB_COUNTER_BYTES]);
3464 }
3465 
3466 /* flow counters currently expose two counters packets and bytes */
3467 #define FLOW_COUNTERS_NUM 2
3468 static int counters_set_description(struct ib_counters *counters,
3469 				    enum mlx5_ib_counters_type counters_type,
3470 				    struct mlx5_ib_flow_counters_desc *desc_data,
3471 				    u32 ncounters)
3472 {
3473 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
3474 	u32 cntrs_max_index = 0;
3475 	int i;
3476 
3477 	if (counters_type != MLX5_IB_COUNTERS_FLOW)
3478 		return -EINVAL;
3479 
3480 	/* init the fields for the object */
3481 	mcounters->type = counters_type;
3482 	mcounters->read_counters = read_flow_counters;
3483 	mcounters->counters_num = FLOW_COUNTERS_NUM;
3484 	mcounters->ncounters = ncounters;
3485 	/* each counter entry have both description and index pair */
3486 	for (i = 0; i < ncounters; i++) {
3487 		if (desc_data[i].description > IB_COUNTER_BYTES)
3488 			return -EINVAL;
3489 
3490 		if (cntrs_max_index <= desc_data[i].index)
3491 			cntrs_max_index = desc_data[i].index + 1;
3492 	}
3493 
3494 	mutex_lock(&mcounters->mcntrs_mutex);
3495 	mcounters->counters_data = desc_data;
3496 	mcounters->cntrs_max_index = cntrs_max_index;
3497 	mutex_unlock(&mcounters->mcntrs_mutex);
3498 
3499 	return 0;
3500 }
3501 
3502 #define MAX_COUNTERS_NUM (USHRT_MAX / (sizeof(u32) * 2))
3503 static int flow_counters_set_data(struct ib_counters *ibcounters,
3504 				  struct mlx5_ib_create_flow *ucmd)
3505 {
3506 	struct mlx5_ib_mcounters *mcounters = to_mcounters(ibcounters);
3507 	struct mlx5_ib_flow_counters_data *cntrs_data = NULL;
3508 	struct mlx5_ib_flow_counters_desc *desc_data = NULL;
3509 	bool hw_hndl = false;
3510 	int ret = 0;
3511 
3512 	if (ucmd && ucmd->ncounters_data != 0) {
3513 		cntrs_data = ucmd->data;
3514 		if (cntrs_data->ncounters > MAX_COUNTERS_NUM)
3515 			return -EINVAL;
3516 
3517 		desc_data = kcalloc(cntrs_data->ncounters,
3518 				    sizeof(*desc_data),
3519 				    GFP_KERNEL);
3520 		if (!desc_data)
3521 			return  -ENOMEM;
3522 
3523 		if (copy_from_user(desc_data,
3524 				   u64_to_user_ptr(cntrs_data->counters_data),
3525 				   sizeof(*desc_data) * cntrs_data->ncounters)) {
3526 			ret = -EFAULT;
3527 			goto free;
3528 		}
3529 	}
3530 
3531 	if (!mcounters->hw_cntrs_hndl) {
3532 		mcounters->hw_cntrs_hndl = mlx5_fc_create(
3533 			to_mdev(ibcounters->device)->mdev, false);
3534 		if (IS_ERR(mcounters->hw_cntrs_hndl)) {
3535 			ret = PTR_ERR(mcounters->hw_cntrs_hndl);
3536 			goto free;
3537 		}
3538 		hw_hndl = true;
3539 	}
3540 
3541 	if (desc_data) {
3542 		/* counters already bound to at least one flow */
3543 		if (mcounters->cntrs_max_index) {
3544 			ret = -EINVAL;
3545 			goto free_hndl;
3546 		}
3547 
3548 		ret = counters_set_description(ibcounters,
3549 					       MLX5_IB_COUNTERS_FLOW,
3550 					       desc_data,
3551 					       cntrs_data->ncounters);
3552 		if (ret)
3553 			goto free_hndl;
3554 
3555 	} else if (!mcounters->cntrs_max_index) {
3556 		/* counters not bound yet, must have udata passed */
3557 		ret = -EINVAL;
3558 		goto free_hndl;
3559 	}
3560 
3561 	return 0;
3562 
3563 free_hndl:
3564 	if (hw_hndl) {
3565 		mlx5_fc_destroy(to_mdev(ibcounters->device)->mdev,
3566 				mcounters->hw_cntrs_hndl);
3567 		mcounters->hw_cntrs_hndl = NULL;
3568 	}
3569 free:
3570 	kfree(desc_data);
3571 	return ret;
3572 }
3573 
3574 static void mlx5_ib_set_rule_source_port(struct mlx5_ib_dev *dev,
3575 					 struct mlx5_flow_spec *spec,
3576 					 struct mlx5_eswitch_rep *rep)
3577 {
3578 	struct mlx5_eswitch *esw = dev->mdev->priv.eswitch;
3579 	void *misc;
3580 
3581 	if (mlx5_eswitch_vport_match_metadata_enabled(esw)) {
3582 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3583 				    misc_parameters_2);
3584 
3585 		MLX5_SET(fte_match_set_misc2, misc, metadata_reg_c_0,
3586 			 mlx5_eswitch_get_vport_metadata_for_match(esw,
3587 								   rep->vport));
3588 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_criteria,
3589 				    misc_parameters_2);
3590 
3591 		MLX5_SET(fte_match_set_misc2, misc, metadata_reg_c_0,
3592 			 mlx5_eswitch_get_vport_metadata_mask());
3593 	} else {
3594 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3595 				    misc_parameters);
3596 
3597 		MLX5_SET(fte_match_set_misc, misc, source_port, rep->vport);
3598 
3599 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_criteria,
3600 				    misc_parameters);
3601 
3602 		MLX5_SET_TO_ONES(fte_match_set_misc, misc, source_port);
3603 	}
3604 }
3605 
3606 static struct mlx5_ib_flow_handler *_create_flow_rule(struct mlx5_ib_dev *dev,
3607 						      struct mlx5_ib_flow_prio *ft_prio,
3608 						      const struct ib_flow_attr *flow_attr,
3609 						      struct mlx5_flow_destination *dst,
3610 						      u32 underlay_qpn,
3611 						      struct mlx5_ib_create_flow *ucmd)
3612 {
3613 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
3614 	struct mlx5_ib_flow_handler *handler;
3615 	struct mlx5_flow_act flow_act = {};
3616 	struct mlx5_flow_spec *spec;
3617 	struct mlx5_flow_destination dest_arr[2] = {};
3618 	struct mlx5_flow_destination *rule_dst = dest_arr;
3619 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
3620 	unsigned int spec_index;
3621 	u32 prev_type = 0;
3622 	int err = 0;
3623 	int dest_num = 0;
3624 	bool is_egress = flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS;
3625 
3626 	if (!is_valid_attr(dev->mdev, flow_attr))
3627 		return ERR_PTR(-EINVAL);
3628 
3629 	if (dev->is_rep && is_egress)
3630 		return ERR_PTR(-EINVAL);
3631 
3632 	spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
3633 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
3634 	if (!handler || !spec) {
3635 		err = -ENOMEM;
3636 		goto free;
3637 	}
3638 
3639 	INIT_LIST_HEAD(&handler->list);
3640 
3641 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
3642 		err = parse_flow_attr(dev->mdev, spec,
3643 				      ib_flow, flow_attr, &flow_act,
3644 				      prev_type);
3645 		if (err < 0)
3646 			goto free;
3647 
3648 		prev_type = ((union ib_flow_spec *)ib_flow)->type;
3649 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
3650 	}
3651 
3652 	if (dst && !(flow_act.action & MLX5_FLOW_CONTEXT_ACTION_DROP)) {
3653 		memcpy(&dest_arr[0], dst, sizeof(*dst));
3654 		dest_num++;
3655 	}
3656 
3657 	if (!flow_is_multicast_only(flow_attr))
3658 		set_underlay_qp(dev, spec, underlay_qpn);
3659 
3660 	if (dev->is_rep) {
3661 		struct mlx5_eswitch_rep *rep;
3662 
3663 		rep = dev->port[flow_attr->port - 1].rep;
3664 		if (!rep) {
3665 			err = -EINVAL;
3666 			goto free;
3667 		}
3668 
3669 		mlx5_ib_set_rule_source_port(dev, spec, rep);
3670 	}
3671 
3672 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
3673 
3674 	if (is_egress &&
3675 	    !is_valid_spec(dev->mdev, spec, &flow_act, is_egress)) {
3676 		err = -EINVAL;
3677 		goto free;
3678 	}
3679 
3680 	if (flow_act.action & MLX5_FLOW_CONTEXT_ACTION_COUNT) {
3681 		struct mlx5_ib_mcounters *mcounters;
3682 
3683 		err = flow_counters_set_data(flow_act.counters, ucmd);
3684 		if (err)
3685 			goto free;
3686 
3687 		mcounters = to_mcounters(flow_act.counters);
3688 		handler->ibcounters = flow_act.counters;
3689 		dest_arr[dest_num].type =
3690 			MLX5_FLOW_DESTINATION_TYPE_COUNTER;
3691 		dest_arr[dest_num].counter_id =
3692 			mlx5_fc_id(mcounters->hw_cntrs_hndl);
3693 		dest_num++;
3694 	}
3695 
3696 	if (flow_act.action & MLX5_FLOW_CONTEXT_ACTION_DROP) {
3697 		if (!dest_num)
3698 			rule_dst = NULL;
3699 	} else {
3700 		if (is_egress)
3701 			flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_ALLOW;
3702 		else
3703 			flow_act.action |=
3704 				dest_num ?  MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
3705 					MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
3706 	}
3707 
3708 	if ((spec->flow_context.flags & FLOW_CONTEXT_HAS_TAG)  &&
3709 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3710 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
3711 		mlx5_ib_warn(dev, "Flow tag %u and attribute type %x isn't allowed in leftovers\n",
3712 			     spec->flow_context.flow_tag, flow_attr->type);
3713 		err = -EINVAL;
3714 		goto free;
3715 	}
3716 	handler->rule = mlx5_add_flow_rules(ft, spec,
3717 					    &flow_act,
3718 					    rule_dst, dest_num);
3719 
3720 	if (IS_ERR(handler->rule)) {
3721 		err = PTR_ERR(handler->rule);
3722 		goto free;
3723 	}
3724 
3725 	ft_prio->refcount++;
3726 	handler->prio = ft_prio;
3727 	handler->dev = dev;
3728 
3729 	ft_prio->flow_table = ft;
3730 free:
3731 	if (err && handler) {
3732 		if (handler->ibcounters &&
3733 		    atomic_read(&handler->ibcounters->usecnt) == 1)
3734 			counters_clear_description(handler->ibcounters);
3735 		kfree(handler);
3736 	}
3737 	kvfree(spec);
3738 	return err ? ERR_PTR(err) : handler;
3739 }
3740 
3741 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
3742 						     struct mlx5_ib_flow_prio *ft_prio,
3743 						     const struct ib_flow_attr *flow_attr,
3744 						     struct mlx5_flow_destination *dst)
3745 {
3746 	return _create_flow_rule(dev, ft_prio, flow_attr, dst, 0, NULL);
3747 }
3748 
3749 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
3750 							  struct mlx5_ib_flow_prio *ft_prio,
3751 							  struct ib_flow_attr *flow_attr,
3752 							  struct mlx5_flow_destination *dst)
3753 {
3754 	struct mlx5_ib_flow_handler *handler_dst = NULL;
3755 	struct mlx5_ib_flow_handler *handler = NULL;
3756 
3757 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
3758 	if (!IS_ERR(handler)) {
3759 		handler_dst = create_flow_rule(dev, ft_prio,
3760 					       flow_attr, dst);
3761 		if (IS_ERR(handler_dst)) {
3762 			mlx5_del_flow_rules(handler->rule);
3763 			ft_prio->refcount--;
3764 			kfree(handler);
3765 			handler = handler_dst;
3766 		} else {
3767 			list_add(&handler_dst->list, &handler->list);
3768 		}
3769 	}
3770 
3771 	return handler;
3772 }
3773 enum {
3774 	LEFTOVERS_MC,
3775 	LEFTOVERS_UC,
3776 };
3777 
3778 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
3779 							  struct mlx5_ib_flow_prio *ft_prio,
3780 							  struct ib_flow_attr *flow_attr,
3781 							  struct mlx5_flow_destination *dst)
3782 {
3783 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
3784 	struct mlx5_ib_flow_handler *handler = NULL;
3785 
3786 	static struct {
3787 		struct ib_flow_attr	flow_attr;
3788 		struct ib_flow_spec_eth eth_flow;
3789 	} leftovers_specs[] = {
3790 		[LEFTOVERS_MC] = {
3791 			.flow_attr = {
3792 				.num_of_specs = 1,
3793 				.size = sizeof(leftovers_specs[0])
3794 			},
3795 			.eth_flow = {
3796 				.type = IB_FLOW_SPEC_ETH,
3797 				.size = sizeof(struct ib_flow_spec_eth),
3798 				.mask = {.dst_mac = {0x1} },
3799 				.val =  {.dst_mac = {0x1} }
3800 			}
3801 		},
3802 		[LEFTOVERS_UC] = {
3803 			.flow_attr = {
3804 				.num_of_specs = 1,
3805 				.size = sizeof(leftovers_specs[0])
3806 			},
3807 			.eth_flow = {
3808 				.type = IB_FLOW_SPEC_ETH,
3809 				.size = sizeof(struct ib_flow_spec_eth),
3810 				.mask = {.dst_mac = {0x1} },
3811 				.val = {.dst_mac = {} }
3812 			}
3813 		}
3814 	};
3815 
3816 	handler = create_flow_rule(dev, ft_prio,
3817 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
3818 				   dst);
3819 	if (!IS_ERR(handler) &&
3820 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
3821 		handler_ucast = create_flow_rule(dev, ft_prio,
3822 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
3823 						 dst);
3824 		if (IS_ERR(handler_ucast)) {
3825 			mlx5_del_flow_rules(handler->rule);
3826 			ft_prio->refcount--;
3827 			kfree(handler);
3828 			handler = handler_ucast;
3829 		} else {
3830 			list_add(&handler_ucast->list, &handler->list);
3831 		}
3832 	}
3833 
3834 	return handler;
3835 }
3836 
3837 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
3838 							struct mlx5_ib_flow_prio *ft_rx,
3839 							struct mlx5_ib_flow_prio *ft_tx,
3840 							struct mlx5_flow_destination *dst)
3841 {
3842 	struct mlx5_ib_flow_handler *handler_rx;
3843 	struct mlx5_ib_flow_handler *handler_tx;
3844 	int err;
3845 	static const struct ib_flow_attr flow_attr  = {
3846 		.num_of_specs = 0,
3847 		.size = sizeof(flow_attr)
3848 	};
3849 
3850 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
3851 	if (IS_ERR(handler_rx)) {
3852 		err = PTR_ERR(handler_rx);
3853 		goto err;
3854 	}
3855 
3856 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
3857 	if (IS_ERR(handler_tx)) {
3858 		err = PTR_ERR(handler_tx);
3859 		goto err_tx;
3860 	}
3861 
3862 	list_add(&handler_tx->list, &handler_rx->list);
3863 
3864 	return handler_rx;
3865 
3866 err_tx:
3867 	mlx5_del_flow_rules(handler_rx->rule);
3868 	ft_rx->refcount--;
3869 	kfree(handler_rx);
3870 err:
3871 	return ERR_PTR(err);
3872 }
3873 
3874 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
3875 					   struct ib_flow_attr *flow_attr,
3876 					   int domain,
3877 					   struct ib_udata *udata)
3878 {
3879 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3880 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3881 	struct mlx5_ib_flow_handler *handler = NULL;
3882 	struct mlx5_flow_destination *dst = NULL;
3883 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
3884 	struct mlx5_ib_flow_prio *ft_prio;
3885 	bool is_egress = flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS;
3886 	struct mlx5_ib_create_flow *ucmd = NULL, ucmd_hdr;
3887 	size_t min_ucmd_sz, required_ucmd_sz;
3888 	int err;
3889 	int underlay_qpn;
3890 
3891 	if (udata && udata->inlen) {
3892 		min_ucmd_sz = offsetof(typeof(ucmd_hdr), reserved) +
3893 				sizeof(ucmd_hdr.reserved);
3894 		if (udata->inlen < min_ucmd_sz)
3895 			return ERR_PTR(-EOPNOTSUPP);
3896 
3897 		err = ib_copy_from_udata(&ucmd_hdr, udata, min_ucmd_sz);
3898 		if (err)
3899 			return ERR_PTR(err);
3900 
3901 		/* currently supports only one counters data */
3902 		if (ucmd_hdr.ncounters_data > 1)
3903 			return ERR_PTR(-EINVAL);
3904 
3905 		required_ucmd_sz = min_ucmd_sz +
3906 			sizeof(struct mlx5_ib_flow_counters_data) *
3907 			ucmd_hdr.ncounters_data;
3908 		if (udata->inlen > required_ucmd_sz &&
3909 		    !ib_is_udata_cleared(udata, required_ucmd_sz,
3910 					 udata->inlen - required_ucmd_sz))
3911 			return ERR_PTR(-EOPNOTSUPP);
3912 
3913 		ucmd = kzalloc(required_ucmd_sz, GFP_KERNEL);
3914 		if (!ucmd)
3915 			return ERR_PTR(-ENOMEM);
3916 
3917 		err = ib_copy_from_udata(ucmd, udata, required_ucmd_sz);
3918 		if (err)
3919 			goto free_ucmd;
3920 	}
3921 
3922 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO) {
3923 		err = -ENOMEM;
3924 		goto free_ucmd;
3925 	}
3926 
3927 	if (domain != IB_FLOW_DOMAIN_USER ||
3928 	    flow_attr->port > dev->num_ports ||
3929 	    (flow_attr->flags & ~(IB_FLOW_ATTR_FLAGS_DONT_TRAP |
3930 				  IB_FLOW_ATTR_FLAGS_EGRESS))) {
3931 		err = -EINVAL;
3932 		goto free_ucmd;
3933 	}
3934 
3935 	if (is_egress &&
3936 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3937 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
3938 		err = -EINVAL;
3939 		goto free_ucmd;
3940 	}
3941 
3942 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
3943 	if (!dst) {
3944 		err = -ENOMEM;
3945 		goto free_ucmd;
3946 	}
3947 
3948 	mutex_lock(&dev->flow_db->lock);
3949 
3950 	ft_prio = get_flow_table(dev, flow_attr,
3951 				 is_egress ? MLX5_IB_FT_TX : MLX5_IB_FT_RX);
3952 	if (IS_ERR(ft_prio)) {
3953 		err = PTR_ERR(ft_prio);
3954 		goto unlock;
3955 	}
3956 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3957 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
3958 		if (IS_ERR(ft_prio_tx)) {
3959 			err = PTR_ERR(ft_prio_tx);
3960 			ft_prio_tx = NULL;
3961 			goto destroy_ft;
3962 		}
3963 	}
3964 
3965 	if (is_egress) {
3966 		dst->type = MLX5_FLOW_DESTINATION_TYPE_PORT;
3967 	} else {
3968 		dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
3969 		if (mqp->flags & MLX5_IB_QP_RSS)
3970 			dst->tir_num = mqp->rss_qp.tirn;
3971 		else
3972 			dst->tir_num = mqp->raw_packet_qp.rq.tirn;
3973 	}
3974 
3975 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
3976 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
3977 			handler = create_dont_trap_rule(dev, ft_prio,
3978 							flow_attr, dst);
3979 		} else {
3980 			underlay_qpn = (mqp->flags & MLX5_IB_QP_UNDERLAY) ?
3981 					mqp->underlay_qpn : 0;
3982 			handler = _create_flow_rule(dev, ft_prio, flow_attr,
3983 						    dst, underlay_qpn, ucmd);
3984 		}
3985 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3986 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
3987 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
3988 						dst);
3989 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3990 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
3991 	} else {
3992 		err = -EINVAL;
3993 		goto destroy_ft;
3994 	}
3995 
3996 	if (IS_ERR(handler)) {
3997 		err = PTR_ERR(handler);
3998 		handler = NULL;
3999 		goto destroy_ft;
4000 	}
4001 
4002 	mutex_unlock(&dev->flow_db->lock);
4003 	kfree(dst);
4004 	kfree(ucmd);
4005 
4006 	return &handler->ibflow;
4007 
4008 destroy_ft:
4009 	put_flow_table(dev, ft_prio, false);
4010 	if (ft_prio_tx)
4011 		put_flow_table(dev, ft_prio_tx, false);
4012 unlock:
4013 	mutex_unlock(&dev->flow_db->lock);
4014 	kfree(dst);
4015 free_ucmd:
4016 	kfree(ucmd);
4017 	return ERR_PTR(err);
4018 }
4019 
4020 static struct mlx5_ib_flow_prio *
4021 _get_flow_table(struct mlx5_ib_dev *dev,
4022 		struct mlx5_ib_flow_matcher *fs_matcher,
4023 		bool mcast)
4024 {
4025 	struct mlx5_flow_namespace *ns = NULL;
4026 	struct mlx5_ib_flow_prio *prio = NULL;
4027 	int max_table_size = 0;
4028 	bool esw_encap;
4029 	u32 flags = 0;
4030 	int priority;
4031 
4032 	if (mcast)
4033 		priority = MLX5_IB_FLOW_MCAST_PRIO;
4034 	else
4035 		priority = ib_prio_to_core_prio(fs_matcher->priority, false);
4036 
4037 	esw_encap = mlx5_eswitch_get_encap_mode(dev->mdev) !=
4038 		DEVLINK_ESWITCH_ENCAP_MODE_NONE;
4039 	if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_BYPASS) {
4040 		max_table_size = BIT(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
4041 					log_max_ft_size));
4042 		if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, decap) && !esw_encap)
4043 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_DECAP;
4044 		if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
4045 					      reformat_l3_tunnel_to_l2) &&
4046 		    !esw_encap)
4047 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
4048 	} else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_EGRESS) {
4049 		max_table_size = BIT(
4050 			MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, log_max_ft_size));
4051 		if (MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, reformat) && !esw_encap)
4052 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
4053 	} else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_FDB) {
4054 		max_table_size = BIT(
4055 			MLX5_CAP_ESW_FLOWTABLE_FDB(dev->mdev, log_max_ft_size));
4056 		if (MLX5_CAP_ESW_FLOWTABLE_FDB(dev->mdev, decap) && esw_encap)
4057 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_DECAP;
4058 		if (MLX5_CAP_ESW_FLOWTABLE_FDB(dev->mdev, reformat_l3_tunnel_to_l2) &&
4059 		    esw_encap)
4060 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
4061 		priority = FDB_BYPASS_PATH;
4062 	} else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_RDMA_RX) {
4063 		max_table_size =
4064 			BIT(MLX5_CAP_FLOWTABLE_RDMA_RX(dev->mdev,
4065 						       log_max_ft_size));
4066 		priority = fs_matcher->priority;
4067 	} else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_RDMA_TX) {
4068 		max_table_size =
4069 			BIT(MLX5_CAP_FLOWTABLE_RDMA_TX(dev->mdev,
4070 						       log_max_ft_size));
4071 		priority = fs_matcher->priority;
4072 	}
4073 
4074 	max_table_size = min_t(int, max_table_size, MLX5_FS_MAX_ENTRIES);
4075 
4076 	ns = mlx5_get_flow_namespace(dev->mdev, fs_matcher->ns_type);
4077 	if (!ns)
4078 		return ERR_PTR(-ENOTSUPP);
4079 
4080 	if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_BYPASS)
4081 		prio = &dev->flow_db->prios[priority];
4082 	else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_EGRESS)
4083 		prio = &dev->flow_db->egress_prios[priority];
4084 	else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_FDB)
4085 		prio = &dev->flow_db->fdb;
4086 	else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_RDMA_RX)
4087 		prio = &dev->flow_db->rdma_rx[priority];
4088 	else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_RDMA_TX)
4089 		prio = &dev->flow_db->rdma_tx[priority];
4090 
4091 	if (!prio)
4092 		return ERR_PTR(-EINVAL);
4093 
4094 	if (prio->flow_table)
4095 		return prio;
4096 
4097 	return _get_prio(ns, prio, priority, max_table_size,
4098 			 MLX5_FS_MAX_TYPES, flags);
4099 }
4100 
4101 static struct mlx5_ib_flow_handler *
4102 _create_raw_flow_rule(struct mlx5_ib_dev *dev,
4103 		      struct mlx5_ib_flow_prio *ft_prio,
4104 		      struct mlx5_flow_destination *dst,
4105 		      struct mlx5_ib_flow_matcher  *fs_matcher,
4106 		      struct mlx5_flow_context *flow_context,
4107 		      struct mlx5_flow_act *flow_act,
4108 		      void *cmd_in, int inlen,
4109 		      int dst_num)
4110 {
4111 	struct mlx5_ib_flow_handler *handler;
4112 	struct mlx5_flow_spec *spec;
4113 	struct mlx5_flow_table *ft = ft_prio->flow_table;
4114 	int err = 0;
4115 
4116 	spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
4117 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
4118 	if (!handler || !spec) {
4119 		err = -ENOMEM;
4120 		goto free;
4121 	}
4122 
4123 	INIT_LIST_HEAD(&handler->list);
4124 
4125 	memcpy(spec->match_value, cmd_in, inlen);
4126 	memcpy(spec->match_criteria, fs_matcher->matcher_mask.match_params,
4127 	       fs_matcher->mask_len);
4128 	spec->match_criteria_enable = fs_matcher->match_criteria_enable;
4129 	spec->flow_context = *flow_context;
4130 
4131 	handler->rule = mlx5_add_flow_rules(ft, spec,
4132 					    flow_act, dst, dst_num);
4133 
4134 	if (IS_ERR(handler->rule)) {
4135 		err = PTR_ERR(handler->rule);
4136 		goto free;
4137 	}
4138 
4139 	ft_prio->refcount++;
4140 	handler->prio = ft_prio;
4141 	handler->dev = dev;
4142 	ft_prio->flow_table = ft;
4143 
4144 free:
4145 	if (err)
4146 		kfree(handler);
4147 	kvfree(spec);
4148 	return err ? ERR_PTR(err) : handler;
4149 }
4150 
4151 static bool raw_fs_is_multicast(struct mlx5_ib_flow_matcher *fs_matcher,
4152 				void *match_v)
4153 {
4154 	void *match_c;
4155 	void *match_v_set_lyr_2_4, *match_c_set_lyr_2_4;
4156 	void *dmac, *dmac_mask;
4157 	void *ipv4, *ipv4_mask;
4158 
4159 	if (!(fs_matcher->match_criteria_enable &
4160 	      (1 << MATCH_CRITERIA_ENABLE_OUTER_BIT)))
4161 		return false;
4162 
4163 	match_c = fs_matcher->matcher_mask.match_params;
4164 	match_v_set_lyr_2_4 = MLX5_ADDR_OF(fte_match_param, match_v,
4165 					   outer_headers);
4166 	match_c_set_lyr_2_4 = MLX5_ADDR_OF(fte_match_param, match_c,
4167 					   outer_headers);
4168 
4169 	dmac = MLX5_ADDR_OF(fte_match_set_lyr_2_4, match_v_set_lyr_2_4,
4170 			    dmac_47_16);
4171 	dmac_mask = MLX5_ADDR_OF(fte_match_set_lyr_2_4, match_c_set_lyr_2_4,
4172 				 dmac_47_16);
4173 
4174 	if (is_multicast_ether_addr(dmac) &&
4175 	    is_multicast_ether_addr(dmac_mask))
4176 		return true;
4177 
4178 	ipv4 = MLX5_ADDR_OF(fte_match_set_lyr_2_4, match_v_set_lyr_2_4,
4179 			    dst_ipv4_dst_ipv6.ipv4_layout.ipv4);
4180 
4181 	ipv4_mask = MLX5_ADDR_OF(fte_match_set_lyr_2_4, match_c_set_lyr_2_4,
4182 				 dst_ipv4_dst_ipv6.ipv4_layout.ipv4);
4183 
4184 	if (ipv4_is_multicast(*(__be32 *)(ipv4)) &&
4185 	    ipv4_is_multicast(*(__be32 *)(ipv4_mask)))
4186 		return true;
4187 
4188 	return false;
4189 }
4190 
4191 struct mlx5_ib_flow_handler *
4192 mlx5_ib_raw_fs_rule_add(struct mlx5_ib_dev *dev,
4193 			struct mlx5_ib_flow_matcher *fs_matcher,
4194 			struct mlx5_flow_context *flow_context,
4195 			struct mlx5_flow_act *flow_act,
4196 			u32 counter_id,
4197 			void *cmd_in, int inlen, int dest_id,
4198 			int dest_type)
4199 {
4200 	struct mlx5_flow_destination *dst;
4201 	struct mlx5_ib_flow_prio *ft_prio;
4202 	struct mlx5_ib_flow_handler *handler;
4203 	int dst_num = 0;
4204 	bool mcast;
4205 	int err;
4206 
4207 	if (fs_matcher->flow_type != MLX5_IB_FLOW_TYPE_NORMAL)
4208 		return ERR_PTR(-EOPNOTSUPP);
4209 
4210 	if (fs_matcher->priority > MLX5_IB_FLOW_LAST_PRIO)
4211 		return ERR_PTR(-ENOMEM);
4212 
4213 	dst = kcalloc(2, sizeof(*dst), GFP_KERNEL);
4214 	if (!dst)
4215 		return ERR_PTR(-ENOMEM);
4216 
4217 	mcast = raw_fs_is_multicast(fs_matcher, cmd_in);
4218 	mutex_lock(&dev->flow_db->lock);
4219 
4220 	ft_prio = _get_flow_table(dev, fs_matcher, mcast);
4221 	if (IS_ERR(ft_prio)) {
4222 		err = PTR_ERR(ft_prio);
4223 		goto unlock;
4224 	}
4225 
4226 	if (dest_type == MLX5_FLOW_DESTINATION_TYPE_TIR) {
4227 		dst[dst_num].type = dest_type;
4228 		dst[dst_num].tir_num = dest_id;
4229 		flow_act->action |= MLX5_FLOW_CONTEXT_ACTION_FWD_DEST;
4230 	} else if (dest_type == MLX5_FLOW_DESTINATION_TYPE_FLOW_TABLE) {
4231 		dst[dst_num].type = MLX5_FLOW_DESTINATION_TYPE_FLOW_TABLE_NUM;
4232 		dst[dst_num].ft_num = dest_id;
4233 		flow_act->action |= MLX5_FLOW_CONTEXT_ACTION_FWD_DEST;
4234 	} else {
4235 		dst[dst_num].type = MLX5_FLOW_DESTINATION_TYPE_PORT;
4236 		flow_act->action |= MLX5_FLOW_CONTEXT_ACTION_ALLOW;
4237 	}
4238 
4239 	dst_num++;
4240 
4241 	if (flow_act->action & MLX5_FLOW_CONTEXT_ACTION_COUNT) {
4242 		dst[dst_num].type = MLX5_FLOW_DESTINATION_TYPE_COUNTER;
4243 		dst[dst_num].counter_id = counter_id;
4244 		dst_num++;
4245 	}
4246 
4247 	handler = _create_raw_flow_rule(dev, ft_prio, dst, fs_matcher,
4248 					flow_context, flow_act,
4249 					cmd_in, inlen, dst_num);
4250 
4251 	if (IS_ERR(handler)) {
4252 		err = PTR_ERR(handler);
4253 		goto destroy_ft;
4254 	}
4255 
4256 	mutex_unlock(&dev->flow_db->lock);
4257 	atomic_inc(&fs_matcher->usecnt);
4258 	handler->flow_matcher = fs_matcher;
4259 
4260 	kfree(dst);
4261 
4262 	return handler;
4263 
4264 destroy_ft:
4265 	put_flow_table(dev, ft_prio, false);
4266 unlock:
4267 	mutex_unlock(&dev->flow_db->lock);
4268 	kfree(dst);
4269 
4270 	return ERR_PTR(err);
4271 }
4272 
4273 static u32 mlx5_ib_flow_action_flags_to_accel_xfrm_flags(u32 mlx5_flags)
4274 {
4275 	u32 flags = 0;
4276 
4277 	if (mlx5_flags & MLX5_IB_UAPI_FLOW_ACTION_FLAGS_REQUIRE_METADATA)
4278 		flags |= MLX5_ACCEL_XFRM_FLAG_REQUIRE_METADATA;
4279 
4280 	return flags;
4281 }
4282 
4283 #define MLX5_FLOW_ACTION_ESP_CREATE_LAST_SUPPORTED	MLX5_IB_UAPI_FLOW_ACTION_FLAGS_REQUIRE_METADATA
4284 static struct ib_flow_action *
4285 mlx5_ib_create_flow_action_esp(struct ib_device *device,
4286 			       const struct ib_flow_action_attrs_esp *attr,
4287 			       struct uverbs_attr_bundle *attrs)
4288 {
4289 	struct mlx5_ib_dev *mdev = to_mdev(device);
4290 	struct ib_uverbs_flow_action_esp_keymat_aes_gcm *aes_gcm;
4291 	struct mlx5_accel_esp_xfrm_attrs accel_attrs = {};
4292 	struct mlx5_ib_flow_action *action;
4293 	u64 action_flags;
4294 	u64 flags;
4295 	int err = 0;
4296 
4297 	err = uverbs_get_flags64(
4298 		&action_flags, attrs, MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
4299 		((MLX5_FLOW_ACTION_ESP_CREATE_LAST_SUPPORTED << 1) - 1));
4300 	if (err)
4301 		return ERR_PTR(err);
4302 
4303 	flags = mlx5_ib_flow_action_flags_to_accel_xfrm_flags(action_flags);
4304 
4305 	/* We current only support a subset of the standard features. Only a
4306 	 * keymat of type AES_GCM, with icv_len == 16, iv_algo == SEQ and esn
4307 	 * (with overlap). Full offload mode isn't supported.
4308 	 */
4309 	if (!attr->keymat || attr->replay || attr->encap ||
4310 	    attr->spi || attr->seq || attr->tfc_pad ||
4311 	    attr->hard_limit_pkts ||
4312 	    (attr->flags & ~(IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
4313 			     IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ENCRYPT)))
4314 		return ERR_PTR(-EOPNOTSUPP);
4315 
4316 	if (attr->keymat->protocol !=
4317 	    IB_UVERBS_FLOW_ACTION_ESP_KEYMAT_AES_GCM)
4318 		return ERR_PTR(-EOPNOTSUPP);
4319 
4320 	aes_gcm = &attr->keymat->keymat.aes_gcm;
4321 
4322 	if (aes_gcm->icv_len != 16 ||
4323 	    aes_gcm->iv_algo != IB_UVERBS_FLOW_ACTION_IV_ALGO_SEQ)
4324 		return ERR_PTR(-EOPNOTSUPP);
4325 
4326 	action = kmalloc(sizeof(*action), GFP_KERNEL);
4327 	if (!action)
4328 		return ERR_PTR(-ENOMEM);
4329 
4330 	action->esp_aes_gcm.ib_flags = attr->flags;
4331 	memcpy(&accel_attrs.keymat.aes_gcm.aes_key, &aes_gcm->aes_key,
4332 	       sizeof(accel_attrs.keymat.aes_gcm.aes_key));
4333 	accel_attrs.keymat.aes_gcm.key_len = aes_gcm->key_len * 8;
4334 	memcpy(&accel_attrs.keymat.aes_gcm.salt, &aes_gcm->salt,
4335 	       sizeof(accel_attrs.keymat.aes_gcm.salt));
4336 	memcpy(&accel_attrs.keymat.aes_gcm.seq_iv, &aes_gcm->iv,
4337 	       sizeof(accel_attrs.keymat.aes_gcm.seq_iv));
4338 	accel_attrs.keymat.aes_gcm.icv_len = aes_gcm->icv_len * 8;
4339 	accel_attrs.keymat.aes_gcm.iv_algo = MLX5_ACCEL_ESP_AES_GCM_IV_ALGO_SEQ;
4340 	accel_attrs.keymat_type = MLX5_ACCEL_ESP_KEYMAT_AES_GCM;
4341 
4342 	accel_attrs.esn = attr->esn;
4343 	if (attr->flags & IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED)
4344 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_TRIGGERED;
4345 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)
4346 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
4347 
4348 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ENCRYPT)
4349 		accel_attrs.action |= MLX5_ACCEL_ESP_ACTION_ENCRYPT;
4350 
4351 	action->esp_aes_gcm.ctx =
4352 		mlx5_accel_esp_create_xfrm(mdev->mdev, &accel_attrs, flags);
4353 	if (IS_ERR(action->esp_aes_gcm.ctx)) {
4354 		err = PTR_ERR(action->esp_aes_gcm.ctx);
4355 		goto err_parse;
4356 	}
4357 
4358 	action->esp_aes_gcm.ib_flags = attr->flags;
4359 
4360 	return &action->ib_action;
4361 
4362 err_parse:
4363 	kfree(action);
4364 	return ERR_PTR(err);
4365 }
4366 
4367 static int
4368 mlx5_ib_modify_flow_action_esp(struct ib_flow_action *action,
4369 			       const struct ib_flow_action_attrs_esp *attr,
4370 			       struct uverbs_attr_bundle *attrs)
4371 {
4372 	struct mlx5_ib_flow_action *maction = to_mflow_act(action);
4373 	struct mlx5_accel_esp_xfrm_attrs accel_attrs;
4374 	int err = 0;
4375 
4376 	if (attr->keymat || attr->replay || attr->encap ||
4377 	    attr->spi || attr->seq || attr->tfc_pad ||
4378 	    attr->hard_limit_pkts ||
4379 	    (attr->flags & ~(IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
4380 			     IB_FLOW_ACTION_ESP_FLAGS_MOD_ESP_ATTRS |
4381 			     IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)))
4382 		return -EOPNOTSUPP;
4383 
4384 	/* Only the ESN value or the MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP can
4385 	 * be modified.
4386 	 */
4387 	if (!(maction->esp_aes_gcm.ib_flags &
4388 	      IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED) &&
4389 	    attr->flags & (IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
4390 			   IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW))
4391 		return -EINVAL;
4392 
4393 	memcpy(&accel_attrs, &maction->esp_aes_gcm.ctx->attrs,
4394 	       sizeof(accel_attrs));
4395 
4396 	accel_attrs.esn = attr->esn;
4397 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)
4398 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
4399 	else
4400 		accel_attrs.flags &= ~MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
4401 
4402 	err = mlx5_accel_esp_modify_xfrm(maction->esp_aes_gcm.ctx,
4403 					 &accel_attrs);
4404 	if (err)
4405 		return err;
4406 
4407 	maction->esp_aes_gcm.ib_flags &=
4408 		~IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW;
4409 	maction->esp_aes_gcm.ib_flags |=
4410 		attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW;
4411 
4412 	return 0;
4413 }
4414 
4415 static int mlx5_ib_destroy_flow_action(struct ib_flow_action *action)
4416 {
4417 	struct mlx5_ib_flow_action *maction = to_mflow_act(action);
4418 
4419 	switch (action->type) {
4420 	case IB_FLOW_ACTION_ESP:
4421 		/*
4422 		 * We only support aes_gcm by now, so we implicitly know this is
4423 		 * the underline crypto.
4424 		 */
4425 		mlx5_accel_esp_destroy_xfrm(maction->esp_aes_gcm.ctx);
4426 		break;
4427 	case IB_FLOW_ACTION_UNSPECIFIED:
4428 		mlx5_ib_destroy_flow_action_raw(maction);
4429 		break;
4430 	default:
4431 		WARN_ON(true);
4432 		break;
4433 	}
4434 
4435 	kfree(maction);
4436 	return 0;
4437 }
4438 
4439 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
4440 {
4441 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4442 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
4443 	int err;
4444 	u16 uid;
4445 
4446 	uid = ibqp->pd ?
4447 		to_mpd(ibqp->pd)->uid : 0;
4448 
4449 	if (mqp->flags & MLX5_IB_QP_UNDERLAY) {
4450 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
4451 		return -EOPNOTSUPP;
4452 	}
4453 
4454 	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
4455 	if (err)
4456 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
4457 			     ibqp->qp_num, gid->raw);
4458 
4459 	return err;
4460 }
4461 
4462 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
4463 {
4464 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4465 	int err;
4466 	u16 uid;
4467 
4468 	uid = ibqp->pd ?
4469 		to_mpd(ibqp->pd)->uid : 0;
4470 	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
4471 	if (err)
4472 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
4473 			     ibqp->qp_num, gid->raw);
4474 
4475 	return err;
4476 }
4477 
4478 static int init_node_data(struct mlx5_ib_dev *dev)
4479 {
4480 	int err;
4481 
4482 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
4483 	if (err)
4484 		return err;
4485 
4486 	dev->mdev->rev_id = dev->mdev->pdev->revision;
4487 
4488 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
4489 }
4490 
4491 static ssize_t fw_pages_show(struct device *device,
4492 			     struct device_attribute *attr, char *buf)
4493 {
4494 	struct mlx5_ib_dev *dev =
4495 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4496 
4497 	return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
4498 }
4499 static DEVICE_ATTR_RO(fw_pages);
4500 
4501 static ssize_t reg_pages_show(struct device *device,
4502 			      struct device_attribute *attr, char *buf)
4503 {
4504 	struct mlx5_ib_dev *dev =
4505 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4506 
4507 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
4508 }
4509 static DEVICE_ATTR_RO(reg_pages);
4510 
4511 static ssize_t hca_type_show(struct device *device,
4512 			     struct device_attribute *attr, char *buf)
4513 {
4514 	struct mlx5_ib_dev *dev =
4515 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4516 
4517 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
4518 }
4519 static DEVICE_ATTR_RO(hca_type);
4520 
4521 static ssize_t hw_rev_show(struct device *device,
4522 			   struct device_attribute *attr, char *buf)
4523 {
4524 	struct mlx5_ib_dev *dev =
4525 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4526 
4527 	return sprintf(buf, "%x\n", dev->mdev->rev_id);
4528 }
4529 static DEVICE_ATTR_RO(hw_rev);
4530 
4531 static ssize_t board_id_show(struct device *device,
4532 			     struct device_attribute *attr, char *buf)
4533 {
4534 	struct mlx5_ib_dev *dev =
4535 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4536 
4537 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
4538 		       dev->mdev->board_id);
4539 }
4540 static DEVICE_ATTR_RO(board_id);
4541 
4542 static struct attribute *mlx5_class_attributes[] = {
4543 	&dev_attr_hw_rev.attr,
4544 	&dev_attr_hca_type.attr,
4545 	&dev_attr_board_id.attr,
4546 	&dev_attr_fw_pages.attr,
4547 	&dev_attr_reg_pages.attr,
4548 	NULL,
4549 };
4550 
4551 static const struct attribute_group mlx5_attr_group = {
4552 	.attrs = mlx5_class_attributes,
4553 };
4554 
4555 static void pkey_change_handler(struct work_struct *work)
4556 {
4557 	struct mlx5_ib_port_resources *ports =
4558 		container_of(work, struct mlx5_ib_port_resources,
4559 			     pkey_change_work);
4560 
4561 	mutex_lock(&ports->devr->mutex);
4562 	mlx5_ib_gsi_pkey_change(ports->gsi);
4563 	mutex_unlock(&ports->devr->mutex);
4564 }
4565 
4566 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
4567 {
4568 	struct mlx5_ib_qp *mqp;
4569 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
4570 	struct mlx5_core_cq *mcq;
4571 	struct list_head cq_armed_list;
4572 	unsigned long flags_qp;
4573 	unsigned long flags_cq;
4574 	unsigned long flags;
4575 
4576 	INIT_LIST_HEAD(&cq_armed_list);
4577 
4578 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
4579 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
4580 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
4581 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
4582 		if (mqp->sq.tail != mqp->sq.head) {
4583 			send_mcq = to_mcq(mqp->ibqp.send_cq);
4584 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
4585 			if (send_mcq->mcq.comp &&
4586 			    mqp->ibqp.send_cq->comp_handler) {
4587 				if (!send_mcq->mcq.reset_notify_added) {
4588 					send_mcq->mcq.reset_notify_added = 1;
4589 					list_add_tail(&send_mcq->mcq.reset_notify,
4590 						      &cq_armed_list);
4591 				}
4592 			}
4593 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
4594 		}
4595 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
4596 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
4597 		/* no handling is needed for SRQ */
4598 		if (!mqp->ibqp.srq) {
4599 			if (mqp->rq.tail != mqp->rq.head) {
4600 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
4601 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
4602 				if (recv_mcq->mcq.comp &&
4603 				    mqp->ibqp.recv_cq->comp_handler) {
4604 					if (!recv_mcq->mcq.reset_notify_added) {
4605 						recv_mcq->mcq.reset_notify_added = 1;
4606 						list_add_tail(&recv_mcq->mcq.reset_notify,
4607 							      &cq_armed_list);
4608 					}
4609 				}
4610 				spin_unlock_irqrestore(&recv_mcq->lock,
4611 						       flags_cq);
4612 			}
4613 		}
4614 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
4615 	}
4616 	/*At that point all inflight post send were put to be executed as of we
4617 	 * lock/unlock above locks Now need to arm all involved CQs.
4618 	 */
4619 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
4620 		mcq->comp(mcq, NULL);
4621 	}
4622 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
4623 }
4624 
4625 static void delay_drop_handler(struct work_struct *work)
4626 {
4627 	int err;
4628 	struct mlx5_ib_delay_drop *delay_drop =
4629 		container_of(work, struct mlx5_ib_delay_drop,
4630 			     delay_drop_work);
4631 
4632 	atomic_inc(&delay_drop->events_cnt);
4633 
4634 	mutex_lock(&delay_drop->lock);
4635 	err = mlx5_core_set_delay_drop(delay_drop->dev->mdev,
4636 				       delay_drop->timeout);
4637 	if (err) {
4638 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
4639 			     delay_drop->timeout);
4640 		delay_drop->activate = false;
4641 	}
4642 	mutex_unlock(&delay_drop->lock);
4643 }
4644 
4645 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
4646 				 struct ib_event *ibev)
4647 {
4648 	u8 port = (eqe->data.port.port >> 4) & 0xf;
4649 
4650 	switch (eqe->sub_type) {
4651 	case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
4652 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
4653 					    IB_LINK_LAYER_ETHERNET)
4654 			schedule_work(&ibdev->delay_drop.delay_drop_work);
4655 		break;
4656 	default: /* do nothing */
4657 		return;
4658 	}
4659 }
4660 
4661 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
4662 			      struct ib_event *ibev)
4663 {
4664 	u8 port = (eqe->data.port.port >> 4) & 0xf;
4665 
4666 	ibev->element.port_num = port;
4667 
4668 	switch (eqe->sub_type) {
4669 	case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
4670 	case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
4671 	case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
4672 		/* In RoCE, port up/down events are handled in
4673 		 * mlx5_netdev_event().
4674 		 */
4675 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
4676 					    IB_LINK_LAYER_ETHERNET)
4677 			return -EINVAL;
4678 
4679 		ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
4680 				IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
4681 		break;
4682 
4683 	case MLX5_PORT_CHANGE_SUBTYPE_LID:
4684 		ibev->event = IB_EVENT_LID_CHANGE;
4685 		break;
4686 
4687 	case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
4688 		ibev->event = IB_EVENT_PKEY_CHANGE;
4689 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
4690 		break;
4691 
4692 	case MLX5_PORT_CHANGE_SUBTYPE_GUID:
4693 		ibev->event = IB_EVENT_GID_CHANGE;
4694 		break;
4695 
4696 	case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
4697 		ibev->event = IB_EVENT_CLIENT_REREGISTER;
4698 		break;
4699 	default:
4700 		return -EINVAL;
4701 	}
4702 
4703 	return 0;
4704 }
4705 
4706 static void mlx5_ib_handle_event(struct work_struct *_work)
4707 {
4708 	struct mlx5_ib_event_work *work =
4709 		container_of(_work, struct mlx5_ib_event_work, work);
4710 	struct mlx5_ib_dev *ibdev;
4711 	struct ib_event ibev;
4712 	bool fatal = false;
4713 
4714 	if (work->is_slave) {
4715 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
4716 		if (!ibdev)
4717 			goto out;
4718 	} else {
4719 		ibdev = work->dev;
4720 	}
4721 
4722 	switch (work->event) {
4723 	case MLX5_DEV_EVENT_SYS_ERROR:
4724 		ibev.event = IB_EVENT_DEVICE_FATAL;
4725 		mlx5_ib_handle_internal_error(ibdev);
4726 		ibev.element.port_num  = (u8)(unsigned long)work->param;
4727 		fatal = true;
4728 		break;
4729 	case MLX5_EVENT_TYPE_PORT_CHANGE:
4730 		if (handle_port_change(ibdev, work->param, &ibev))
4731 			goto out;
4732 		break;
4733 	case MLX5_EVENT_TYPE_GENERAL_EVENT:
4734 		handle_general_event(ibdev, work->param, &ibev);
4735 		/* fall through */
4736 	default:
4737 		goto out;
4738 	}
4739 
4740 	ibev.device = &ibdev->ib_dev;
4741 
4742 	if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
4743 		mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
4744 		goto out;
4745 	}
4746 
4747 	if (ibdev->ib_active)
4748 		ib_dispatch_event(&ibev);
4749 
4750 	if (fatal)
4751 		ibdev->ib_active = false;
4752 out:
4753 	kfree(work);
4754 }
4755 
4756 static int mlx5_ib_event(struct notifier_block *nb,
4757 			 unsigned long event, void *param)
4758 {
4759 	struct mlx5_ib_event_work *work;
4760 
4761 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
4762 	if (!work)
4763 		return NOTIFY_DONE;
4764 
4765 	INIT_WORK(&work->work, mlx5_ib_handle_event);
4766 	work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
4767 	work->is_slave = false;
4768 	work->param = param;
4769 	work->event = event;
4770 
4771 	queue_work(mlx5_ib_event_wq, &work->work);
4772 
4773 	return NOTIFY_OK;
4774 }
4775 
4776 static int mlx5_ib_event_slave_port(struct notifier_block *nb,
4777 				    unsigned long event, void *param)
4778 {
4779 	struct mlx5_ib_event_work *work;
4780 
4781 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
4782 	if (!work)
4783 		return NOTIFY_DONE;
4784 
4785 	INIT_WORK(&work->work, mlx5_ib_handle_event);
4786 	work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
4787 	work->is_slave = true;
4788 	work->param = param;
4789 	work->event = event;
4790 	queue_work(mlx5_ib_event_wq, &work->work);
4791 
4792 	return NOTIFY_OK;
4793 }
4794 
4795 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
4796 {
4797 	struct mlx5_hca_vport_context vport_ctx;
4798 	int err;
4799 	int port;
4800 
4801 	for (port = 1; port <= ARRAY_SIZE(dev->mdev->port_caps); port++) {
4802 		dev->mdev->port_caps[port - 1].has_smi = false;
4803 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
4804 		    MLX5_CAP_PORT_TYPE_IB) {
4805 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
4806 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
4807 								   port, 0,
4808 								   &vport_ctx);
4809 				if (err) {
4810 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
4811 						    port, err);
4812 					return err;
4813 				}
4814 				dev->mdev->port_caps[port - 1].has_smi =
4815 					vport_ctx.has_smi;
4816 			} else {
4817 				dev->mdev->port_caps[port - 1].has_smi = true;
4818 			}
4819 		}
4820 	}
4821 	return 0;
4822 }
4823 
4824 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
4825 {
4826 	int port;
4827 
4828 	for (port = 1; port <= dev->num_ports; port++)
4829 		mlx5_query_ext_port_caps(dev, port);
4830 }
4831 
4832 static int __get_port_caps(struct mlx5_ib_dev *dev, u8 port)
4833 {
4834 	struct ib_device_attr *dprops = NULL;
4835 	struct ib_port_attr *pprops = NULL;
4836 	int err = -ENOMEM;
4837 
4838 	pprops = kzalloc(sizeof(*pprops), GFP_KERNEL);
4839 	if (!pprops)
4840 		goto out;
4841 
4842 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
4843 	if (!dprops)
4844 		goto out;
4845 
4846 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, NULL);
4847 	if (err) {
4848 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
4849 		goto out;
4850 	}
4851 
4852 	err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
4853 	if (err) {
4854 		mlx5_ib_warn(dev, "query_port %d failed %d\n",
4855 			     port, err);
4856 		goto out;
4857 	}
4858 
4859 	dev->mdev->port_caps[port - 1].pkey_table_len =
4860 					dprops->max_pkeys;
4861 	dev->mdev->port_caps[port - 1].gid_table_len =
4862 					pprops->gid_tbl_len;
4863 	mlx5_ib_dbg(dev, "port %d: pkey_table_len %d, gid_table_len %d\n",
4864 		    port, dprops->max_pkeys, pprops->gid_tbl_len);
4865 
4866 out:
4867 	kfree(pprops);
4868 	kfree(dprops);
4869 
4870 	return err;
4871 }
4872 
4873 static int get_port_caps(struct mlx5_ib_dev *dev, u8 port)
4874 {
4875 	/* For representors use port 1, is this is the only native
4876 	 * port
4877 	 */
4878 	if (dev->is_rep)
4879 		return __get_port_caps(dev, 1);
4880 	return __get_port_caps(dev, port);
4881 }
4882 
4883 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
4884 {
4885 	int err;
4886 
4887 	err = mlx5_mr_cache_cleanup(dev);
4888 	if (err)
4889 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
4890 
4891 	if (dev->umrc.qp)
4892 		mlx5_ib_destroy_qp(dev->umrc.qp, NULL);
4893 	if (dev->umrc.cq)
4894 		ib_free_cq(dev->umrc.cq);
4895 	if (dev->umrc.pd)
4896 		ib_dealloc_pd(dev->umrc.pd);
4897 }
4898 
4899 enum {
4900 	MAX_UMR_WR = 128,
4901 };
4902 
4903 static int create_umr_res(struct mlx5_ib_dev *dev)
4904 {
4905 	struct ib_qp_init_attr *init_attr = NULL;
4906 	struct ib_qp_attr *attr = NULL;
4907 	struct ib_pd *pd;
4908 	struct ib_cq *cq;
4909 	struct ib_qp *qp;
4910 	int ret;
4911 
4912 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
4913 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
4914 	if (!attr || !init_attr) {
4915 		ret = -ENOMEM;
4916 		goto error_0;
4917 	}
4918 
4919 	pd = ib_alloc_pd(&dev->ib_dev, 0);
4920 	if (IS_ERR(pd)) {
4921 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
4922 		ret = PTR_ERR(pd);
4923 		goto error_0;
4924 	}
4925 
4926 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
4927 	if (IS_ERR(cq)) {
4928 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
4929 		ret = PTR_ERR(cq);
4930 		goto error_2;
4931 	}
4932 
4933 	init_attr->send_cq = cq;
4934 	init_attr->recv_cq = cq;
4935 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
4936 	init_attr->cap.max_send_wr = MAX_UMR_WR;
4937 	init_attr->cap.max_send_sge = 1;
4938 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
4939 	init_attr->port_num = 1;
4940 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
4941 	if (IS_ERR(qp)) {
4942 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
4943 		ret = PTR_ERR(qp);
4944 		goto error_3;
4945 	}
4946 	qp->device     = &dev->ib_dev;
4947 	qp->real_qp    = qp;
4948 	qp->uobject    = NULL;
4949 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
4950 	qp->send_cq    = init_attr->send_cq;
4951 	qp->recv_cq    = init_attr->recv_cq;
4952 
4953 	attr->qp_state = IB_QPS_INIT;
4954 	attr->port_num = 1;
4955 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
4956 				IB_QP_PORT, NULL);
4957 	if (ret) {
4958 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
4959 		goto error_4;
4960 	}
4961 
4962 	memset(attr, 0, sizeof(*attr));
4963 	attr->qp_state = IB_QPS_RTR;
4964 	attr->path_mtu = IB_MTU_256;
4965 
4966 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4967 	if (ret) {
4968 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
4969 		goto error_4;
4970 	}
4971 
4972 	memset(attr, 0, sizeof(*attr));
4973 	attr->qp_state = IB_QPS_RTS;
4974 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4975 	if (ret) {
4976 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
4977 		goto error_4;
4978 	}
4979 
4980 	dev->umrc.qp = qp;
4981 	dev->umrc.cq = cq;
4982 	dev->umrc.pd = pd;
4983 
4984 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
4985 	ret = mlx5_mr_cache_init(dev);
4986 	if (ret) {
4987 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4988 		goto error_4;
4989 	}
4990 
4991 	kfree(attr);
4992 	kfree(init_attr);
4993 
4994 	return 0;
4995 
4996 error_4:
4997 	mlx5_ib_destroy_qp(qp, NULL);
4998 	dev->umrc.qp = NULL;
4999 
5000 error_3:
5001 	ib_free_cq(cq);
5002 	dev->umrc.cq = NULL;
5003 
5004 error_2:
5005 	ib_dealloc_pd(pd);
5006 	dev->umrc.pd = NULL;
5007 
5008 error_0:
5009 	kfree(attr);
5010 	kfree(init_attr);
5011 	return ret;
5012 }
5013 
5014 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
5015 {
5016 	switch (umr_fence_cap) {
5017 	case MLX5_CAP_UMR_FENCE_NONE:
5018 		return MLX5_FENCE_MODE_NONE;
5019 	case MLX5_CAP_UMR_FENCE_SMALL:
5020 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
5021 	default:
5022 		return MLX5_FENCE_MODE_STRONG_ORDERING;
5023 	}
5024 }
5025 
5026 static int create_dev_resources(struct mlx5_ib_resources *devr)
5027 {
5028 	struct ib_srq_init_attr attr;
5029 	struct mlx5_ib_dev *dev;
5030 	struct ib_device *ibdev;
5031 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
5032 	int port;
5033 	int ret = 0;
5034 
5035 	dev = container_of(devr, struct mlx5_ib_dev, devr);
5036 	ibdev = &dev->ib_dev;
5037 
5038 	mutex_init(&devr->mutex);
5039 
5040 	devr->p0 = rdma_zalloc_drv_obj(ibdev, ib_pd);
5041 	if (!devr->p0)
5042 		return -ENOMEM;
5043 
5044 	devr->p0->device  = ibdev;
5045 	devr->p0->uobject = NULL;
5046 	atomic_set(&devr->p0->usecnt, 0);
5047 
5048 	ret = mlx5_ib_alloc_pd(devr->p0, NULL);
5049 	if (ret)
5050 		goto error0;
5051 
5052 	devr->c0 = rdma_zalloc_drv_obj(ibdev, ib_cq);
5053 	if (!devr->c0) {
5054 		ret = -ENOMEM;
5055 		goto error1;
5056 	}
5057 
5058 	devr->c0->device = &dev->ib_dev;
5059 	atomic_set(&devr->c0->usecnt, 0);
5060 
5061 	ret = mlx5_ib_create_cq(devr->c0, &cq_attr, NULL);
5062 	if (ret)
5063 		goto err_create_cq;
5064 
5065 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL);
5066 	if (IS_ERR(devr->x0)) {
5067 		ret = PTR_ERR(devr->x0);
5068 		goto error2;
5069 	}
5070 	devr->x0->device = &dev->ib_dev;
5071 	devr->x0->inode = NULL;
5072 	atomic_set(&devr->x0->usecnt, 0);
5073 	mutex_init(&devr->x0->tgt_qp_mutex);
5074 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
5075 
5076 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL);
5077 	if (IS_ERR(devr->x1)) {
5078 		ret = PTR_ERR(devr->x1);
5079 		goto error3;
5080 	}
5081 	devr->x1->device = &dev->ib_dev;
5082 	devr->x1->inode = NULL;
5083 	atomic_set(&devr->x1->usecnt, 0);
5084 	mutex_init(&devr->x1->tgt_qp_mutex);
5085 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
5086 
5087 	memset(&attr, 0, sizeof(attr));
5088 	attr.attr.max_sge = 1;
5089 	attr.attr.max_wr = 1;
5090 	attr.srq_type = IB_SRQT_XRC;
5091 	attr.ext.cq = devr->c0;
5092 	attr.ext.xrc.xrcd = devr->x0;
5093 
5094 	devr->s0 = rdma_zalloc_drv_obj(ibdev, ib_srq);
5095 	if (!devr->s0) {
5096 		ret = -ENOMEM;
5097 		goto error4;
5098 	}
5099 
5100 	devr->s0->device	= &dev->ib_dev;
5101 	devr->s0->pd		= devr->p0;
5102 	devr->s0->srq_type      = IB_SRQT_XRC;
5103 	devr->s0->ext.xrc.xrcd	= devr->x0;
5104 	devr->s0->ext.cq	= devr->c0;
5105 	ret = mlx5_ib_create_srq(devr->s0, &attr, NULL);
5106 	if (ret)
5107 		goto err_create;
5108 
5109 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
5110 	atomic_inc(&devr->s0->ext.cq->usecnt);
5111 	atomic_inc(&devr->p0->usecnt);
5112 	atomic_set(&devr->s0->usecnt, 0);
5113 
5114 	memset(&attr, 0, sizeof(attr));
5115 	attr.attr.max_sge = 1;
5116 	attr.attr.max_wr = 1;
5117 	attr.srq_type = IB_SRQT_BASIC;
5118 	devr->s1 = rdma_zalloc_drv_obj(ibdev, ib_srq);
5119 	if (!devr->s1) {
5120 		ret = -ENOMEM;
5121 		goto error5;
5122 	}
5123 
5124 	devr->s1->device	= &dev->ib_dev;
5125 	devr->s1->pd		= devr->p0;
5126 	devr->s1->srq_type      = IB_SRQT_BASIC;
5127 	devr->s1->ext.cq	= devr->c0;
5128 
5129 	ret = mlx5_ib_create_srq(devr->s1, &attr, NULL);
5130 	if (ret)
5131 		goto error6;
5132 
5133 	atomic_inc(&devr->p0->usecnt);
5134 	atomic_set(&devr->s1->usecnt, 0);
5135 
5136 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
5137 		INIT_WORK(&devr->ports[port].pkey_change_work,
5138 			  pkey_change_handler);
5139 		devr->ports[port].devr = devr;
5140 	}
5141 
5142 	return 0;
5143 
5144 error6:
5145 	kfree(devr->s1);
5146 error5:
5147 	mlx5_ib_destroy_srq(devr->s0, NULL);
5148 err_create:
5149 	kfree(devr->s0);
5150 error4:
5151 	mlx5_ib_dealloc_xrcd(devr->x1, NULL);
5152 error3:
5153 	mlx5_ib_dealloc_xrcd(devr->x0, NULL);
5154 error2:
5155 	mlx5_ib_destroy_cq(devr->c0, NULL);
5156 err_create_cq:
5157 	kfree(devr->c0);
5158 error1:
5159 	mlx5_ib_dealloc_pd(devr->p0, NULL);
5160 error0:
5161 	kfree(devr->p0);
5162 	return ret;
5163 }
5164 
5165 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
5166 {
5167 	int port;
5168 
5169 	mlx5_ib_destroy_srq(devr->s1, NULL);
5170 	kfree(devr->s1);
5171 	mlx5_ib_destroy_srq(devr->s0, NULL);
5172 	kfree(devr->s0);
5173 	mlx5_ib_dealloc_xrcd(devr->x0, NULL);
5174 	mlx5_ib_dealloc_xrcd(devr->x1, NULL);
5175 	mlx5_ib_destroy_cq(devr->c0, NULL);
5176 	kfree(devr->c0);
5177 	mlx5_ib_dealloc_pd(devr->p0, NULL);
5178 	kfree(devr->p0);
5179 
5180 	/* Make sure no change P_Key work items are still executing */
5181 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
5182 		cancel_work_sync(&devr->ports[port].pkey_change_work);
5183 }
5184 
5185 static u32 get_core_cap_flags(struct ib_device *ibdev,
5186 			      struct mlx5_hca_vport_context *rep)
5187 {
5188 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5189 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
5190 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
5191 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
5192 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
5193 	u32 ret = 0;
5194 
5195 	if (rep->grh_required)
5196 		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
5197 
5198 	if (ll == IB_LINK_LAYER_INFINIBAND)
5199 		return ret | RDMA_CORE_PORT_IBA_IB;
5200 
5201 	if (raw_support)
5202 		ret |= RDMA_CORE_PORT_RAW_PACKET;
5203 
5204 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
5205 		return ret;
5206 
5207 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
5208 		return ret;
5209 
5210 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
5211 		ret |= RDMA_CORE_PORT_IBA_ROCE;
5212 
5213 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
5214 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
5215 
5216 	return ret;
5217 }
5218 
5219 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
5220 			       struct ib_port_immutable *immutable)
5221 {
5222 	struct ib_port_attr attr;
5223 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5224 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
5225 	struct mlx5_hca_vport_context rep = {0};
5226 	int err;
5227 
5228 	err = ib_query_port(ibdev, port_num, &attr);
5229 	if (err)
5230 		return err;
5231 
5232 	if (ll == IB_LINK_LAYER_INFINIBAND) {
5233 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
5234 						   &rep);
5235 		if (err)
5236 			return err;
5237 	}
5238 
5239 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
5240 	immutable->gid_tbl_len = attr.gid_tbl_len;
5241 	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
5242 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
5243 
5244 	return 0;
5245 }
5246 
5247 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u8 port_num,
5248 				   struct ib_port_immutable *immutable)
5249 {
5250 	struct ib_port_attr attr;
5251 	int err;
5252 
5253 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
5254 
5255 	err = ib_query_port(ibdev, port_num, &attr);
5256 	if (err)
5257 		return err;
5258 
5259 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
5260 	immutable->gid_tbl_len = attr.gid_tbl_len;
5261 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
5262 
5263 	return 0;
5264 }
5265 
5266 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
5267 {
5268 	struct mlx5_ib_dev *dev =
5269 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
5270 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
5271 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
5272 		 fw_rev_sub(dev->mdev));
5273 }
5274 
5275 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
5276 {
5277 	struct mlx5_core_dev *mdev = dev->mdev;
5278 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
5279 								 MLX5_FLOW_NAMESPACE_LAG);
5280 	struct mlx5_flow_table *ft;
5281 	int err;
5282 
5283 	if (!ns || !mlx5_lag_is_roce(mdev))
5284 		return 0;
5285 
5286 	err = mlx5_cmd_create_vport_lag(mdev);
5287 	if (err)
5288 		return err;
5289 
5290 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
5291 	if (IS_ERR(ft)) {
5292 		err = PTR_ERR(ft);
5293 		goto err_destroy_vport_lag;
5294 	}
5295 
5296 	dev->flow_db->lag_demux_ft = ft;
5297 	dev->lag_active = true;
5298 	return 0;
5299 
5300 err_destroy_vport_lag:
5301 	mlx5_cmd_destroy_vport_lag(mdev);
5302 	return err;
5303 }
5304 
5305 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
5306 {
5307 	struct mlx5_core_dev *mdev = dev->mdev;
5308 
5309 	if (dev->lag_active) {
5310 		dev->lag_active = false;
5311 
5312 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
5313 		dev->flow_db->lag_demux_ft = NULL;
5314 
5315 		mlx5_cmd_destroy_vport_lag(mdev);
5316 	}
5317 }
5318 
5319 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
5320 {
5321 	int err;
5322 
5323 	dev->port[port_num].roce.nb.notifier_call = mlx5_netdev_event;
5324 	err = register_netdevice_notifier(&dev->port[port_num].roce.nb);
5325 	if (err) {
5326 		dev->port[port_num].roce.nb.notifier_call = NULL;
5327 		return err;
5328 	}
5329 
5330 	return 0;
5331 }
5332 
5333 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
5334 {
5335 	if (dev->port[port_num].roce.nb.notifier_call) {
5336 		unregister_netdevice_notifier(&dev->port[port_num].roce.nb);
5337 		dev->port[port_num].roce.nb.notifier_call = NULL;
5338 	}
5339 }
5340 
5341 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
5342 {
5343 	int err;
5344 
5345 	err = mlx5_nic_vport_enable_roce(dev->mdev);
5346 	if (err)
5347 		return err;
5348 
5349 	err = mlx5_eth_lag_init(dev);
5350 	if (err)
5351 		goto err_disable_roce;
5352 
5353 	return 0;
5354 
5355 err_disable_roce:
5356 	mlx5_nic_vport_disable_roce(dev->mdev);
5357 
5358 	return err;
5359 }
5360 
5361 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
5362 {
5363 	mlx5_eth_lag_cleanup(dev);
5364 	mlx5_nic_vport_disable_roce(dev->mdev);
5365 }
5366 
5367 struct mlx5_ib_counter {
5368 	const char *name;
5369 	size_t offset;
5370 };
5371 
5372 #define INIT_Q_COUNTER(_name)		\
5373 	{ .name = #_name, .offset = MLX5_BYTE_OFF(query_q_counter_out, _name)}
5374 
5375 static const struct mlx5_ib_counter basic_q_cnts[] = {
5376 	INIT_Q_COUNTER(rx_write_requests),
5377 	INIT_Q_COUNTER(rx_read_requests),
5378 	INIT_Q_COUNTER(rx_atomic_requests),
5379 	INIT_Q_COUNTER(out_of_buffer),
5380 };
5381 
5382 static const struct mlx5_ib_counter out_of_seq_q_cnts[] = {
5383 	INIT_Q_COUNTER(out_of_sequence),
5384 };
5385 
5386 static const struct mlx5_ib_counter retrans_q_cnts[] = {
5387 	INIT_Q_COUNTER(duplicate_request),
5388 	INIT_Q_COUNTER(rnr_nak_retry_err),
5389 	INIT_Q_COUNTER(packet_seq_err),
5390 	INIT_Q_COUNTER(implied_nak_seq_err),
5391 	INIT_Q_COUNTER(local_ack_timeout_err),
5392 };
5393 
5394 #define INIT_CONG_COUNTER(_name)		\
5395 	{ .name = #_name, .offset =	\
5396 		MLX5_BYTE_OFF(query_cong_statistics_out, _name ## _high)}
5397 
5398 static const struct mlx5_ib_counter cong_cnts[] = {
5399 	INIT_CONG_COUNTER(rp_cnp_ignored),
5400 	INIT_CONG_COUNTER(rp_cnp_handled),
5401 	INIT_CONG_COUNTER(np_ecn_marked_roce_packets),
5402 	INIT_CONG_COUNTER(np_cnp_sent),
5403 };
5404 
5405 static const struct mlx5_ib_counter extended_err_cnts[] = {
5406 	INIT_Q_COUNTER(resp_local_length_error),
5407 	INIT_Q_COUNTER(resp_cqe_error),
5408 	INIT_Q_COUNTER(req_cqe_error),
5409 	INIT_Q_COUNTER(req_remote_invalid_request),
5410 	INIT_Q_COUNTER(req_remote_access_errors),
5411 	INIT_Q_COUNTER(resp_remote_access_errors),
5412 	INIT_Q_COUNTER(resp_cqe_flush_error),
5413 	INIT_Q_COUNTER(req_cqe_flush_error),
5414 };
5415 
5416 static const struct mlx5_ib_counter roce_accl_cnts[] = {
5417 	INIT_Q_COUNTER(roce_adp_retrans),
5418 	INIT_Q_COUNTER(roce_adp_retrans_to),
5419 	INIT_Q_COUNTER(roce_slow_restart),
5420 	INIT_Q_COUNTER(roce_slow_restart_cnps),
5421 	INIT_Q_COUNTER(roce_slow_restart_trans),
5422 };
5423 
5424 #define INIT_EXT_PPCNT_COUNTER(_name)		\
5425 	{ .name = #_name, .offset =	\
5426 	MLX5_BYTE_OFF(ppcnt_reg, \
5427 		      counter_set.eth_extended_cntrs_grp_data_layout._name##_high)}
5428 
5429 static const struct mlx5_ib_counter ext_ppcnt_cnts[] = {
5430 	INIT_EXT_PPCNT_COUNTER(rx_icrc_encapsulated),
5431 };
5432 
5433 static bool is_mdev_switchdev_mode(const struct mlx5_core_dev *mdev)
5434 {
5435 	return MLX5_ESWITCH_MANAGER(mdev) &&
5436 	       mlx5_ib_eswitch_mode(mdev->priv.eswitch) ==
5437 		       MLX5_ESWITCH_OFFLOADS;
5438 }
5439 
5440 static void mlx5_ib_dealloc_counters(struct mlx5_ib_dev *dev)
5441 {
5442 	int num_cnt_ports;
5443 	int i;
5444 
5445 	num_cnt_ports = is_mdev_switchdev_mode(dev->mdev) ? 1 : dev->num_ports;
5446 
5447 	for (i = 0; i < num_cnt_ports; i++) {
5448 		if (dev->port[i].cnts.set_id_valid)
5449 			mlx5_core_dealloc_q_counter(dev->mdev,
5450 						    dev->port[i].cnts.set_id);
5451 		kfree(dev->port[i].cnts.names);
5452 		kfree(dev->port[i].cnts.offsets);
5453 	}
5454 }
5455 
5456 static int __mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev,
5457 				    struct mlx5_ib_counters *cnts)
5458 {
5459 	u32 num_counters;
5460 
5461 	num_counters = ARRAY_SIZE(basic_q_cnts);
5462 
5463 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt))
5464 		num_counters += ARRAY_SIZE(out_of_seq_q_cnts);
5465 
5466 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters))
5467 		num_counters += ARRAY_SIZE(retrans_q_cnts);
5468 
5469 	if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters))
5470 		num_counters += ARRAY_SIZE(extended_err_cnts);
5471 
5472 	if (MLX5_CAP_GEN(dev->mdev, roce_accl))
5473 		num_counters += ARRAY_SIZE(roce_accl_cnts);
5474 
5475 	cnts->num_q_counters = num_counters;
5476 
5477 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
5478 		cnts->num_cong_counters = ARRAY_SIZE(cong_cnts);
5479 		num_counters += ARRAY_SIZE(cong_cnts);
5480 	}
5481 	if (MLX5_CAP_PCAM_FEATURE(dev->mdev, rx_icrc_encapsulated_counter)) {
5482 		cnts->num_ext_ppcnt_counters = ARRAY_SIZE(ext_ppcnt_cnts);
5483 		num_counters += ARRAY_SIZE(ext_ppcnt_cnts);
5484 	}
5485 	cnts->names = kcalloc(num_counters, sizeof(cnts->names), GFP_KERNEL);
5486 	if (!cnts->names)
5487 		return -ENOMEM;
5488 
5489 	cnts->offsets = kcalloc(num_counters,
5490 				sizeof(cnts->offsets), GFP_KERNEL);
5491 	if (!cnts->offsets)
5492 		goto err_names;
5493 
5494 	return 0;
5495 
5496 err_names:
5497 	kfree(cnts->names);
5498 	cnts->names = NULL;
5499 	return -ENOMEM;
5500 }
5501 
5502 static void mlx5_ib_fill_counters(struct mlx5_ib_dev *dev,
5503 				  const char **names,
5504 				  size_t *offsets)
5505 {
5506 	int i;
5507 	int j = 0;
5508 
5509 	for (i = 0; i < ARRAY_SIZE(basic_q_cnts); i++, j++) {
5510 		names[j] = basic_q_cnts[i].name;
5511 		offsets[j] = basic_q_cnts[i].offset;
5512 	}
5513 
5514 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt)) {
5515 		for (i = 0; i < ARRAY_SIZE(out_of_seq_q_cnts); i++, j++) {
5516 			names[j] = out_of_seq_q_cnts[i].name;
5517 			offsets[j] = out_of_seq_q_cnts[i].offset;
5518 		}
5519 	}
5520 
5521 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
5522 		for (i = 0; i < ARRAY_SIZE(retrans_q_cnts); i++, j++) {
5523 			names[j] = retrans_q_cnts[i].name;
5524 			offsets[j] = retrans_q_cnts[i].offset;
5525 		}
5526 	}
5527 
5528 	if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters)) {
5529 		for (i = 0; i < ARRAY_SIZE(extended_err_cnts); i++, j++) {
5530 			names[j] = extended_err_cnts[i].name;
5531 			offsets[j] = extended_err_cnts[i].offset;
5532 		}
5533 	}
5534 
5535 	if (MLX5_CAP_GEN(dev->mdev, roce_accl)) {
5536 		for (i = 0; i < ARRAY_SIZE(roce_accl_cnts); i++, j++) {
5537 			names[j] = roce_accl_cnts[i].name;
5538 			offsets[j] = roce_accl_cnts[i].offset;
5539 		}
5540 	}
5541 
5542 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
5543 		for (i = 0; i < ARRAY_SIZE(cong_cnts); i++, j++) {
5544 			names[j] = cong_cnts[i].name;
5545 			offsets[j] = cong_cnts[i].offset;
5546 		}
5547 	}
5548 
5549 	if (MLX5_CAP_PCAM_FEATURE(dev->mdev, rx_icrc_encapsulated_counter)) {
5550 		for (i = 0; i < ARRAY_SIZE(ext_ppcnt_cnts); i++, j++) {
5551 			names[j] = ext_ppcnt_cnts[i].name;
5552 			offsets[j] = ext_ppcnt_cnts[i].offset;
5553 		}
5554 	}
5555 }
5556 
5557 static int mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev)
5558 {
5559 	int num_cnt_ports;
5560 	int err = 0;
5561 	int i;
5562 	bool is_shared;
5563 
5564 	is_shared = MLX5_CAP_GEN(dev->mdev, log_max_uctx) != 0;
5565 	num_cnt_ports = is_mdev_switchdev_mode(dev->mdev) ? 1 : dev->num_ports;
5566 
5567 	for (i = 0; i < num_cnt_ports; i++) {
5568 		err = __mlx5_ib_alloc_counters(dev, &dev->port[i].cnts);
5569 		if (err)
5570 			goto err_alloc;
5571 
5572 		mlx5_ib_fill_counters(dev, dev->port[i].cnts.names,
5573 				      dev->port[i].cnts.offsets);
5574 
5575 		err = mlx5_cmd_alloc_q_counter(dev->mdev,
5576 					       &dev->port[i].cnts.set_id,
5577 					       is_shared ?
5578 					       MLX5_SHARED_RESOURCE_UID : 0);
5579 		if (err) {
5580 			mlx5_ib_warn(dev,
5581 				     "couldn't allocate queue counter for port %d, err %d\n",
5582 				     i + 1, err);
5583 			goto err_alloc;
5584 		}
5585 		dev->port[i].cnts.set_id_valid = true;
5586 	}
5587 	return 0;
5588 
5589 err_alloc:
5590 	mlx5_ib_dealloc_counters(dev);
5591 	return err;
5592 }
5593 
5594 static const struct mlx5_ib_counters *get_counters(struct mlx5_ib_dev *dev,
5595 						   u8 port_num)
5596 {
5597 	return is_mdev_switchdev_mode(dev->mdev) ? &dev->port[0].cnts :
5598 						   &dev->port[port_num].cnts;
5599 }
5600 
5601 /**
5602  * mlx5_ib_get_counters_id - Returns counters id to use for device+port
5603  * @dev:	Pointer to mlx5 IB device
5604  * @port_num:	Zero based port number
5605  *
5606  * mlx5_ib_get_counters_id() Returns counters set id to use for given
5607  * device port combination in switchdev and non switchdev mode of the
5608  * parent device.
5609  */
5610 u16 mlx5_ib_get_counters_id(struct mlx5_ib_dev *dev, u8 port_num)
5611 {
5612 	const struct mlx5_ib_counters *cnts = get_counters(dev, port_num);
5613 
5614 	return cnts->set_id;
5615 }
5616 
5617 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
5618 						    u8 port_num)
5619 {
5620 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5621 	const struct mlx5_ib_counters *cnts;
5622 	bool is_switchdev = is_mdev_switchdev_mode(dev->mdev);
5623 
5624 	if ((is_switchdev && port_num) || (!is_switchdev && !port_num))
5625 		return NULL;
5626 
5627 	cnts = get_counters(dev, port_num - 1);
5628 
5629 	return rdma_alloc_hw_stats_struct(cnts->names,
5630 					  cnts->num_q_counters +
5631 					  cnts->num_cong_counters +
5632 					  cnts->num_ext_ppcnt_counters,
5633 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
5634 }
5635 
5636 static int mlx5_ib_query_q_counters(struct mlx5_core_dev *mdev,
5637 				    const struct mlx5_ib_counters *cnts,
5638 				    struct rdma_hw_stats *stats,
5639 				    u16 set_id)
5640 {
5641 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
5642 	void *out;
5643 	__be32 val;
5644 	int ret, i;
5645 
5646 	out = kvzalloc(outlen, GFP_KERNEL);
5647 	if (!out)
5648 		return -ENOMEM;
5649 
5650 	ret = mlx5_core_query_q_counter(mdev, set_id, 0, out, outlen);
5651 	if (ret)
5652 		goto free;
5653 
5654 	for (i = 0; i < cnts->num_q_counters; i++) {
5655 		val = *(__be32 *)(out + cnts->offsets[i]);
5656 		stats->value[i] = (u64)be32_to_cpu(val);
5657 	}
5658 
5659 free:
5660 	kvfree(out);
5661 	return ret;
5662 }
5663 
5664 static int mlx5_ib_query_ext_ppcnt_counters(struct mlx5_ib_dev *dev,
5665 					    const struct mlx5_ib_counters *cnts,
5666 					    struct rdma_hw_stats *stats)
5667 {
5668 	int offset = cnts->num_q_counters + cnts->num_cong_counters;
5669 	int sz = MLX5_ST_SZ_BYTES(ppcnt_reg);
5670 	int ret, i;
5671 	void *out;
5672 
5673 	out = kvzalloc(sz, GFP_KERNEL);
5674 	if (!out)
5675 		return -ENOMEM;
5676 
5677 	ret = mlx5_cmd_query_ext_ppcnt_counters(dev->mdev, out);
5678 	if (ret)
5679 		goto free;
5680 
5681 	for (i = 0; i < cnts->num_ext_ppcnt_counters; i++)
5682 		stats->value[i + offset] =
5683 			be64_to_cpup((__be64 *)(out +
5684 				    cnts->offsets[i + offset]));
5685 free:
5686 	kvfree(out);
5687 	return ret;
5688 }
5689 
5690 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
5691 				struct rdma_hw_stats *stats,
5692 				u8 port_num, int index)
5693 {
5694 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5695 	const struct mlx5_ib_counters *cnts = get_counters(dev, port_num - 1);
5696 	struct mlx5_core_dev *mdev;
5697 	int ret, num_counters;
5698 	u8 mdev_port_num;
5699 
5700 	if (!stats)
5701 		return -EINVAL;
5702 
5703 	num_counters = cnts->num_q_counters +
5704 		       cnts->num_cong_counters +
5705 		       cnts->num_ext_ppcnt_counters;
5706 
5707 	/* q_counters are per IB device, query the master mdev */
5708 	ret = mlx5_ib_query_q_counters(dev->mdev, cnts, stats, cnts->set_id);
5709 	if (ret)
5710 		return ret;
5711 
5712 	if (MLX5_CAP_PCAM_FEATURE(dev->mdev, rx_icrc_encapsulated_counter)) {
5713 		ret =  mlx5_ib_query_ext_ppcnt_counters(dev, cnts, stats);
5714 		if (ret)
5715 			return ret;
5716 	}
5717 
5718 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
5719 		mdev = mlx5_ib_get_native_port_mdev(dev, port_num,
5720 						    &mdev_port_num);
5721 		if (!mdev) {
5722 			/* If port is not affiliated yet, its in down state
5723 			 * which doesn't have any counters yet, so it would be
5724 			 * zero. So no need to read from the HCA.
5725 			 */
5726 			goto done;
5727 		}
5728 		ret = mlx5_lag_query_cong_counters(dev->mdev,
5729 						   stats->value +
5730 						   cnts->num_q_counters,
5731 						   cnts->num_cong_counters,
5732 						   cnts->offsets +
5733 						   cnts->num_q_counters);
5734 
5735 		mlx5_ib_put_native_port_mdev(dev, port_num);
5736 		if (ret)
5737 			return ret;
5738 	}
5739 
5740 done:
5741 	return num_counters;
5742 }
5743 
5744 static struct rdma_hw_stats *
5745 mlx5_ib_counter_alloc_stats(struct rdma_counter *counter)
5746 {
5747 	struct mlx5_ib_dev *dev = to_mdev(counter->device);
5748 	const struct mlx5_ib_counters *cnts =
5749 		get_counters(dev, counter->port - 1);
5750 
5751 	return rdma_alloc_hw_stats_struct(cnts->names,
5752 					  cnts->num_q_counters +
5753 					  cnts->num_cong_counters +
5754 					  cnts->num_ext_ppcnt_counters,
5755 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
5756 }
5757 
5758 static int mlx5_ib_counter_update_stats(struct rdma_counter *counter)
5759 {
5760 	struct mlx5_ib_dev *dev = to_mdev(counter->device);
5761 	const struct mlx5_ib_counters *cnts =
5762 		get_counters(dev, counter->port - 1);
5763 
5764 	return mlx5_ib_query_q_counters(dev->mdev, cnts,
5765 					counter->stats, counter->id);
5766 }
5767 
5768 static int mlx5_ib_counter_bind_qp(struct rdma_counter *counter,
5769 				   struct ib_qp *qp)
5770 {
5771 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5772 	u16 cnt_set_id = 0;
5773 	int err;
5774 
5775 	if (!counter->id) {
5776 		err = mlx5_cmd_alloc_q_counter(dev->mdev,
5777 					       &cnt_set_id,
5778 					       MLX5_SHARED_RESOURCE_UID);
5779 		if (err)
5780 			return err;
5781 		counter->id = cnt_set_id;
5782 	}
5783 
5784 	err = mlx5_ib_qp_set_counter(qp, counter);
5785 	if (err)
5786 		goto fail_set_counter;
5787 
5788 	return 0;
5789 
5790 fail_set_counter:
5791 	mlx5_core_dealloc_q_counter(dev->mdev, cnt_set_id);
5792 	counter->id = 0;
5793 
5794 	return err;
5795 }
5796 
5797 static int mlx5_ib_counter_unbind_qp(struct ib_qp *qp)
5798 {
5799 	return mlx5_ib_qp_set_counter(qp, NULL);
5800 }
5801 
5802 static int mlx5_ib_counter_dealloc(struct rdma_counter *counter)
5803 {
5804 	struct mlx5_ib_dev *dev = to_mdev(counter->device);
5805 
5806 	return mlx5_core_dealloc_q_counter(dev->mdev, counter->id);
5807 }
5808 
5809 static int mlx5_ib_rn_get_params(struct ib_device *device, u8 port_num,
5810 				 enum rdma_netdev_t type,
5811 				 struct rdma_netdev_alloc_params *params)
5812 {
5813 	if (type != RDMA_NETDEV_IPOIB)
5814 		return -EOPNOTSUPP;
5815 
5816 	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
5817 }
5818 
5819 static void delay_drop_debugfs_cleanup(struct mlx5_ib_dev *dev)
5820 {
5821 	if (!dev->delay_drop.dir_debugfs)
5822 		return;
5823 	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
5824 	dev->delay_drop.dir_debugfs = NULL;
5825 }
5826 
5827 static void cancel_delay_drop(struct mlx5_ib_dev *dev)
5828 {
5829 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
5830 		return;
5831 
5832 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
5833 	delay_drop_debugfs_cleanup(dev);
5834 }
5835 
5836 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
5837 				       size_t count, loff_t *pos)
5838 {
5839 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
5840 	char lbuf[20];
5841 	int len;
5842 
5843 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
5844 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
5845 }
5846 
5847 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
5848 					size_t count, loff_t *pos)
5849 {
5850 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
5851 	u32 timeout;
5852 	u32 var;
5853 
5854 	if (kstrtouint_from_user(buf, count, 0, &var))
5855 		return -EFAULT;
5856 
5857 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
5858 			1000);
5859 	if (timeout != var)
5860 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
5861 			    timeout);
5862 
5863 	delay_drop->timeout = timeout;
5864 
5865 	return count;
5866 }
5867 
5868 static const struct file_operations fops_delay_drop_timeout = {
5869 	.owner	= THIS_MODULE,
5870 	.open	= simple_open,
5871 	.write	= delay_drop_timeout_write,
5872 	.read	= delay_drop_timeout_read,
5873 };
5874 
5875 static void delay_drop_debugfs_init(struct mlx5_ib_dev *dev)
5876 {
5877 	struct dentry *root;
5878 
5879 	if (!mlx5_debugfs_root)
5880 		return;
5881 
5882 	root = debugfs_create_dir("delay_drop", dev->mdev->priv.dbg_root);
5883 	dev->delay_drop.dir_debugfs = root;
5884 
5885 	debugfs_create_atomic_t("num_timeout_events", 0400, root,
5886 				&dev->delay_drop.events_cnt);
5887 	debugfs_create_atomic_t("num_rqs", 0400, root,
5888 				&dev->delay_drop.rqs_cnt);
5889 	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
5890 			    &fops_delay_drop_timeout);
5891 }
5892 
5893 static void init_delay_drop(struct mlx5_ib_dev *dev)
5894 {
5895 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
5896 		return;
5897 
5898 	mutex_init(&dev->delay_drop.lock);
5899 	dev->delay_drop.dev = dev;
5900 	dev->delay_drop.activate = false;
5901 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
5902 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
5903 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
5904 	atomic_set(&dev->delay_drop.events_cnt, 0);
5905 
5906 	delay_drop_debugfs_init(dev);
5907 }
5908 
5909 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
5910 				      struct mlx5_ib_multiport_info *mpi)
5911 {
5912 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
5913 	struct mlx5_ib_port *port = &ibdev->port[port_num];
5914 	int comps;
5915 	int err;
5916 	int i;
5917 
5918 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
5919 
5920 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
5921 
5922 	spin_lock(&port->mp.mpi_lock);
5923 	if (!mpi->ibdev) {
5924 		spin_unlock(&port->mp.mpi_lock);
5925 		return;
5926 	}
5927 
5928 	mpi->ibdev = NULL;
5929 
5930 	spin_unlock(&port->mp.mpi_lock);
5931 	if (mpi->mdev_events.notifier_call)
5932 		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
5933 	mpi->mdev_events.notifier_call = NULL;
5934 	mlx5_remove_netdev_notifier(ibdev, port_num);
5935 	spin_lock(&port->mp.mpi_lock);
5936 
5937 	comps = mpi->mdev_refcnt;
5938 	if (comps) {
5939 		mpi->unaffiliate = true;
5940 		init_completion(&mpi->unref_comp);
5941 		spin_unlock(&port->mp.mpi_lock);
5942 
5943 		for (i = 0; i < comps; i++)
5944 			wait_for_completion(&mpi->unref_comp);
5945 
5946 		spin_lock(&port->mp.mpi_lock);
5947 		mpi->unaffiliate = false;
5948 	}
5949 
5950 	port->mp.mpi = NULL;
5951 
5952 	list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
5953 
5954 	spin_unlock(&port->mp.mpi_lock);
5955 
5956 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
5957 
5958 	mlx5_ib_dbg(ibdev, "unaffiliated port %d\n", port_num + 1);
5959 	/* Log an error, still needed to cleanup the pointers and add
5960 	 * it back to the list.
5961 	 */
5962 	if (err)
5963 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
5964 			    port_num + 1);
5965 
5966 	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
5967 }
5968 
5969 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
5970 				    struct mlx5_ib_multiport_info *mpi)
5971 {
5972 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
5973 	int err;
5974 
5975 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
5976 
5977 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
5978 	if (ibdev->port[port_num].mp.mpi) {
5979 		mlx5_ib_dbg(ibdev, "port %d already affiliated.\n",
5980 			    port_num + 1);
5981 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
5982 		return false;
5983 	}
5984 
5985 	ibdev->port[port_num].mp.mpi = mpi;
5986 	mpi->ibdev = ibdev;
5987 	mpi->mdev_events.notifier_call = NULL;
5988 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
5989 
5990 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
5991 	if (err)
5992 		goto unbind;
5993 
5994 	err = get_port_caps(ibdev, mlx5_core_native_port_num(mpi->mdev));
5995 	if (err)
5996 		goto unbind;
5997 
5998 	err = mlx5_add_netdev_notifier(ibdev, port_num);
5999 	if (err) {
6000 		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
6001 			    port_num + 1);
6002 		goto unbind;
6003 	}
6004 
6005 	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
6006 	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
6007 
6008 	mlx5_ib_init_cong_debugfs(ibdev, port_num);
6009 
6010 	return true;
6011 
6012 unbind:
6013 	mlx5_ib_unbind_slave_port(ibdev, mpi);
6014 	return false;
6015 }
6016 
6017 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
6018 {
6019 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
6020 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
6021 							  port_num + 1);
6022 	struct mlx5_ib_multiport_info *mpi;
6023 	int err;
6024 	int i;
6025 
6026 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
6027 		return 0;
6028 
6029 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
6030 						     &dev->sys_image_guid);
6031 	if (err)
6032 		return err;
6033 
6034 	err = mlx5_nic_vport_enable_roce(dev->mdev);
6035 	if (err)
6036 		return err;
6037 
6038 	mutex_lock(&mlx5_ib_multiport_mutex);
6039 	for (i = 0; i < dev->num_ports; i++) {
6040 		bool bound = false;
6041 
6042 		/* build a stub multiport info struct for the native port. */
6043 		if (i == port_num) {
6044 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
6045 			if (!mpi) {
6046 				mutex_unlock(&mlx5_ib_multiport_mutex);
6047 				mlx5_nic_vport_disable_roce(dev->mdev);
6048 				return -ENOMEM;
6049 			}
6050 
6051 			mpi->is_master = true;
6052 			mpi->mdev = dev->mdev;
6053 			mpi->sys_image_guid = dev->sys_image_guid;
6054 			dev->port[i].mp.mpi = mpi;
6055 			mpi->ibdev = dev;
6056 			mpi = NULL;
6057 			continue;
6058 		}
6059 
6060 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
6061 				    list) {
6062 			if (dev->sys_image_guid == mpi->sys_image_guid &&
6063 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
6064 				bound = mlx5_ib_bind_slave_port(dev, mpi);
6065 			}
6066 
6067 			if (bound) {
6068 				dev_dbg(mpi->mdev->device,
6069 					"removing port from unaffiliated list.\n");
6070 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
6071 				list_del(&mpi->list);
6072 				break;
6073 			}
6074 		}
6075 		if (!bound) {
6076 			get_port_caps(dev, i + 1);
6077 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
6078 				    i + 1);
6079 		}
6080 	}
6081 
6082 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
6083 	mutex_unlock(&mlx5_ib_multiport_mutex);
6084 	return err;
6085 }
6086 
6087 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
6088 {
6089 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
6090 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
6091 							  port_num + 1);
6092 	int i;
6093 
6094 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
6095 		return;
6096 
6097 	mutex_lock(&mlx5_ib_multiport_mutex);
6098 	for (i = 0; i < dev->num_ports; i++) {
6099 		if (dev->port[i].mp.mpi) {
6100 			/* Destroy the native port stub */
6101 			if (i == port_num) {
6102 				kfree(dev->port[i].mp.mpi);
6103 				dev->port[i].mp.mpi = NULL;
6104 			} else {
6105 				mlx5_ib_dbg(dev, "unbinding port_num: %d\n", i + 1);
6106 				mlx5_ib_unbind_slave_port(dev, dev->port[i].mp.mpi);
6107 			}
6108 		}
6109 	}
6110 
6111 	mlx5_ib_dbg(dev, "removing from devlist\n");
6112 	list_del(&dev->ib_dev_list);
6113 	mutex_unlock(&mlx5_ib_multiport_mutex);
6114 
6115 	mlx5_nic_vport_disable_roce(dev->mdev);
6116 }
6117 
6118 static int mmap_obj_cleanup(struct ib_uobject *uobject,
6119 			    enum rdma_remove_reason why,
6120 			    struct uverbs_attr_bundle *attrs)
6121 {
6122 	struct mlx5_user_mmap_entry *obj = uobject->object;
6123 
6124 	rdma_user_mmap_entry_remove(&obj->rdma_entry);
6125 	return 0;
6126 }
6127 
6128 static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c,
6129 					    struct mlx5_user_mmap_entry *entry,
6130 					    size_t length)
6131 {
6132 	return rdma_user_mmap_entry_insert_range(
6133 		&c->ibucontext, &entry->rdma_entry, length,
6134 		(MLX5_IB_MMAP_OFFSET_START << 16),
6135 		((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1));
6136 }
6137 
6138 static struct mlx5_user_mmap_entry *
6139 alloc_var_entry(struct mlx5_ib_ucontext *c)
6140 {
6141 	struct mlx5_user_mmap_entry *entry;
6142 	struct mlx5_var_table *var_table;
6143 	u32 page_idx;
6144 	int err;
6145 
6146 	var_table = &to_mdev(c->ibucontext.device)->var_table;
6147 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
6148 	if (!entry)
6149 		return ERR_PTR(-ENOMEM);
6150 
6151 	mutex_lock(&var_table->bitmap_lock);
6152 	page_idx = find_first_zero_bit(var_table->bitmap,
6153 				       var_table->num_var_hw_entries);
6154 	if (page_idx >= var_table->num_var_hw_entries) {
6155 		err = -ENOSPC;
6156 		mutex_unlock(&var_table->bitmap_lock);
6157 		goto end;
6158 	}
6159 
6160 	set_bit(page_idx, var_table->bitmap);
6161 	mutex_unlock(&var_table->bitmap_lock);
6162 
6163 	entry->address = var_table->hw_start_addr +
6164 				(page_idx * var_table->stride_size);
6165 	entry->page_idx = page_idx;
6166 	entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
6167 
6168 	err = mlx5_rdma_user_mmap_entry_insert(c, entry,
6169 					       var_table->stride_size);
6170 	if (err)
6171 		goto err_insert;
6172 
6173 	return entry;
6174 
6175 err_insert:
6176 	mutex_lock(&var_table->bitmap_lock);
6177 	clear_bit(page_idx, var_table->bitmap);
6178 	mutex_unlock(&var_table->bitmap_lock);
6179 end:
6180 	kfree(entry);
6181 	return ERR_PTR(err);
6182 }
6183 
6184 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
6185 	struct uverbs_attr_bundle *attrs)
6186 {
6187 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
6188 		attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
6189 	struct mlx5_ib_ucontext *c;
6190 	struct mlx5_user_mmap_entry *entry;
6191 	u64 mmap_offset;
6192 	u32 length;
6193 	int err;
6194 
6195 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
6196 	if (IS_ERR(c))
6197 		return PTR_ERR(c);
6198 
6199 	entry = alloc_var_entry(c);
6200 	if (IS_ERR(entry))
6201 		return PTR_ERR(entry);
6202 
6203 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
6204 	length = entry->rdma_entry.npages * PAGE_SIZE;
6205 	uobj->object = entry;
6206 
6207 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
6208 			     &mmap_offset, sizeof(mmap_offset));
6209 	if (err)
6210 		goto err;
6211 
6212 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
6213 			     &entry->page_idx, sizeof(entry->page_idx));
6214 	if (err)
6215 		goto err;
6216 
6217 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
6218 			     &length, sizeof(length));
6219 	if (err)
6220 		goto err;
6221 
6222 	return 0;
6223 
6224 err:
6225 	rdma_user_mmap_entry_remove(&entry->rdma_entry);
6226 	return err;
6227 }
6228 
6229 DECLARE_UVERBS_NAMED_METHOD(
6230 	MLX5_IB_METHOD_VAR_OBJ_ALLOC,
6231 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
6232 			MLX5_IB_OBJECT_VAR,
6233 			UVERBS_ACCESS_NEW,
6234 			UA_MANDATORY),
6235 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
6236 			   UVERBS_ATTR_TYPE(u32),
6237 			   UA_MANDATORY),
6238 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
6239 			   UVERBS_ATTR_TYPE(u32),
6240 			   UA_MANDATORY),
6241 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
6242 			    UVERBS_ATTR_TYPE(u64),
6243 			    UA_MANDATORY));
6244 
6245 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
6246 	MLX5_IB_METHOD_VAR_OBJ_DESTROY,
6247 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
6248 			MLX5_IB_OBJECT_VAR,
6249 			UVERBS_ACCESS_DESTROY,
6250 			UA_MANDATORY));
6251 
6252 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
6253 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
6254 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
6255 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
6256 
6257 static bool var_is_supported(struct ib_device *device)
6258 {
6259 	struct mlx5_ib_dev *dev = to_mdev(device);
6260 
6261 	return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
6262 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
6263 }
6264 
6265 static struct mlx5_user_mmap_entry *
6266 alloc_uar_entry(struct mlx5_ib_ucontext *c,
6267 		enum mlx5_ib_uapi_uar_alloc_type alloc_type)
6268 {
6269 	struct mlx5_user_mmap_entry *entry;
6270 	struct mlx5_ib_dev *dev;
6271 	u32 uar_index;
6272 	int err;
6273 
6274 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
6275 	if (!entry)
6276 		return ERR_PTR(-ENOMEM);
6277 
6278 	dev = to_mdev(c->ibucontext.device);
6279 	err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
6280 	if (err)
6281 		goto end;
6282 
6283 	entry->page_idx = uar_index;
6284 	entry->address = uar_index2paddress(dev, uar_index);
6285 	if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
6286 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC;
6287 	else
6288 		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC;
6289 
6290 	err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE);
6291 	if (err)
6292 		goto err_insert;
6293 
6294 	return entry;
6295 
6296 err_insert:
6297 	mlx5_cmd_free_uar(dev->mdev, uar_index);
6298 end:
6299 	kfree(entry);
6300 	return ERR_PTR(err);
6301 }
6302 
6303 static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)(
6304 	struct uverbs_attr_bundle *attrs)
6305 {
6306 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
6307 		attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
6308 	enum mlx5_ib_uapi_uar_alloc_type alloc_type;
6309 	struct mlx5_ib_ucontext *c;
6310 	struct mlx5_user_mmap_entry *entry;
6311 	u64 mmap_offset;
6312 	u32 length;
6313 	int err;
6314 
6315 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
6316 	if (IS_ERR(c))
6317 		return PTR_ERR(c);
6318 
6319 	err = uverbs_get_const(&alloc_type, attrs,
6320 			       MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE);
6321 	if (err)
6322 		return err;
6323 
6324 	if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF &&
6325 	    alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC)
6326 		return -EOPNOTSUPP;
6327 
6328 	if (!to_mdev(c->ibucontext.device)->wc_support &&
6329 	    alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
6330 		return -EOPNOTSUPP;
6331 
6332 	entry = alloc_uar_entry(c, alloc_type);
6333 	if (IS_ERR(entry))
6334 		return PTR_ERR(entry);
6335 
6336 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
6337 	length = entry->rdma_entry.npages * PAGE_SIZE;
6338 	uobj->object = entry;
6339 
6340 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
6341 			     &mmap_offset, sizeof(mmap_offset));
6342 	if (err)
6343 		goto err;
6344 
6345 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
6346 			     &entry->page_idx, sizeof(entry->page_idx));
6347 	if (err)
6348 		goto err;
6349 
6350 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
6351 			     &length, sizeof(length));
6352 	if (err)
6353 		goto err;
6354 
6355 	return 0;
6356 
6357 err:
6358 	rdma_user_mmap_entry_remove(&entry->rdma_entry);
6359 	return err;
6360 }
6361 
6362 DECLARE_UVERBS_NAMED_METHOD(
6363 	MLX5_IB_METHOD_UAR_OBJ_ALLOC,
6364 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE,
6365 			MLX5_IB_OBJECT_UAR,
6366 			UVERBS_ACCESS_NEW,
6367 			UA_MANDATORY),
6368 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE,
6369 			     enum mlx5_ib_uapi_uar_alloc_type,
6370 			     UA_MANDATORY),
6371 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
6372 			   UVERBS_ATTR_TYPE(u32),
6373 			   UA_MANDATORY),
6374 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
6375 			   UVERBS_ATTR_TYPE(u32),
6376 			   UA_MANDATORY),
6377 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
6378 			    UVERBS_ATTR_TYPE(u64),
6379 			    UA_MANDATORY));
6380 
6381 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
6382 	MLX5_IB_METHOD_UAR_OBJ_DESTROY,
6383 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE,
6384 			MLX5_IB_OBJECT_UAR,
6385 			UVERBS_ACCESS_DESTROY,
6386 			UA_MANDATORY));
6387 
6388 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR,
6389 			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
6390 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC),
6391 			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY));
6392 
6393 ADD_UVERBS_ATTRIBUTES_SIMPLE(
6394 	mlx5_ib_dm,
6395 	UVERBS_OBJECT_DM,
6396 	UVERBS_METHOD_DM_ALLOC,
6397 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
6398 			    UVERBS_ATTR_TYPE(u64),
6399 			    UA_MANDATORY),
6400 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
6401 			    UVERBS_ATTR_TYPE(u16),
6402 			    UA_OPTIONAL),
6403 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_ALLOC_DM_REQ_TYPE,
6404 			     enum mlx5_ib_uapi_dm_type,
6405 			     UA_OPTIONAL));
6406 
6407 ADD_UVERBS_ATTRIBUTES_SIMPLE(
6408 	mlx5_ib_flow_action,
6409 	UVERBS_OBJECT_FLOW_ACTION,
6410 	UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
6411 	UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
6412 			     enum mlx5_ib_uapi_flow_action_flags));
6413 
6414 static const struct uapi_definition mlx5_ib_defs[] = {
6415 	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
6416 	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
6417 	UAPI_DEF_CHAIN(mlx5_ib_qos_defs),
6418 
6419 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_FLOW_ACTION,
6420 				&mlx5_ib_flow_action),
6421 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DM, &mlx5_ib_dm),
6422 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
6423 				UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
6424 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR),
6425 	{}
6426 };
6427 
6428 static int mlx5_ib_read_counters(struct ib_counters *counters,
6429 				 struct ib_counters_read_attr *read_attr,
6430 				 struct uverbs_attr_bundle *attrs)
6431 {
6432 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
6433 	struct mlx5_read_counters_attr mread_attr = {};
6434 	struct mlx5_ib_flow_counters_desc *desc;
6435 	int ret, i;
6436 
6437 	mutex_lock(&mcounters->mcntrs_mutex);
6438 	if (mcounters->cntrs_max_index > read_attr->ncounters) {
6439 		ret = -EINVAL;
6440 		goto err_bound;
6441 	}
6442 
6443 	mread_attr.out = kcalloc(mcounters->counters_num, sizeof(u64),
6444 				 GFP_KERNEL);
6445 	if (!mread_attr.out) {
6446 		ret = -ENOMEM;
6447 		goto err_bound;
6448 	}
6449 
6450 	mread_attr.hw_cntrs_hndl = mcounters->hw_cntrs_hndl;
6451 	mread_attr.flags = read_attr->flags;
6452 	ret = mcounters->read_counters(counters->device, &mread_attr);
6453 	if (ret)
6454 		goto err_read;
6455 
6456 	/* do the pass over the counters data array to assign according to the
6457 	 * descriptions and indexing pairs
6458 	 */
6459 	desc = mcounters->counters_data;
6460 	for (i = 0; i < mcounters->ncounters; i++)
6461 		read_attr->counters_buff[desc[i].index] += mread_attr.out[desc[i].description];
6462 
6463 err_read:
6464 	kfree(mread_attr.out);
6465 err_bound:
6466 	mutex_unlock(&mcounters->mcntrs_mutex);
6467 	return ret;
6468 }
6469 
6470 static int mlx5_ib_destroy_counters(struct ib_counters *counters)
6471 {
6472 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
6473 
6474 	counters_clear_description(counters);
6475 	if (mcounters->hw_cntrs_hndl)
6476 		mlx5_fc_destroy(to_mdev(counters->device)->mdev,
6477 				mcounters->hw_cntrs_hndl);
6478 
6479 	kfree(mcounters);
6480 
6481 	return 0;
6482 }
6483 
6484 static struct ib_counters *mlx5_ib_create_counters(struct ib_device *device,
6485 						   struct uverbs_attr_bundle *attrs)
6486 {
6487 	struct mlx5_ib_mcounters *mcounters;
6488 
6489 	mcounters = kzalloc(sizeof(*mcounters), GFP_KERNEL);
6490 	if (!mcounters)
6491 		return ERR_PTR(-ENOMEM);
6492 
6493 	mutex_init(&mcounters->mcntrs_mutex);
6494 
6495 	return &mcounters->ibcntrs;
6496 }
6497 
6498 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
6499 {
6500 	mlx5_ib_cleanup_multiport_master(dev);
6501 	WARN_ON(!xa_empty(&dev->odp_mkeys));
6502 	cleanup_srcu_struct(&dev->odp_srcu);
6503 
6504 	WARN_ON(!xa_empty(&dev->sig_mrs));
6505 	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
6506 }
6507 
6508 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
6509 {
6510 	struct mlx5_core_dev *mdev = dev->mdev;
6511 	int err;
6512 	int i;
6513 
6514 	for (i = 0; i < dev->num_ports; i++) {
6515 		spin_lock_init(&dev->port[i].mp.mpi_lock);
6516 		rwlock_init(&dev->port[i].roce.netdev_lock);
6517 		dev->port[i].roce.dev = dev;
6518 		dev->port[i].roce.native_port_num = i + 1;
6519 		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
6520 	}
6521 
6522 	mlx5_ib_internal_fill_odp_caps(dev);
6523 
6524 	err = mlx5_ib_init_multiport_master(dev);
6525 	if (err)
6526 		return err;
6527 
6528 	err = set_has_smi_cap(dev);
6529 	if (err)
6530 		return err;
6531 
6532 	if (!mlx5_core_mp_enabled(mdev)) {
6533 		for (i = 1; i <= dev->num_ports; i++) {
6534 			err = get_port_caps(dev, i);
6535 			if (err)
6536 				break;
6537 		}
6538 	} else {
6539 		err = get_port_caps(dev, mlx5_core_native_port_num(mdev));
6540 	}
6541 	if (err)
6542 		goto err_mp;
6543 
6544 	if (mlx5_use_mad_ifc(dev))
6545 		get_ext_port_caps(dev);
6546 
6547 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
6548 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
6549 	dev->ib_dev.phys_port_cnt	= dev->num_ports;
6550 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_count(mdev);
6551 	dev->ib_dev.dev.parent		= mdev->device;
6552 
6553 	mutex_init(&dev->cap_mask_mutex);
6554 	INIT_LIST_HEAD(&dev->qp_list);
6555 	spin_lock_init(&dev->reset_flow_resource_lock);
6556 	xa_init(&dev->odp_mkeys);
6557 	xa_init(&dev->sig_mrs);
6558 	atomic_set(&dev->mkey_var, 0);
6559 
6560 	spin_lock_init(&dev->dm.lock);
6561 	dev->dm.dev = mdev;
6562 
6563 	err = init_srcu_struct(&dev->odp_srcu);
6564 	if (err)
6565 		goto err_mp;
6566 
6567 	return 0;
6568 
6569 err_mp:
6570 	mlx5_ib_cleanup_multiport_master(dev);
6571 
6572 	return -ENOMEM;
6573 }
6574 
6575 static int mlx5_ib_stage_flow_db_init(struct mlx5_ib_dev *dev)
6576 {
6577 	dev->flow_db = kzalloc(sizeof(*dev->flow_db), GFP_KERNEL);
6578 
6579 	if (!dev->flow_db)
6580 		return -ENOMEM;
6581 
6582 	mutex_init(&dev->flow_db->lock);
6583 
6584 	return 0;
6585 }
6586 
6587 static void mlx5_ib_stage_flow_db_cleanup(struct mlx5_ib_dev *dev)
6588 {
6589 	kfree(dev->flow_db);
6590 }
6591 
6592 static const struct ib_device_ops mlx5_ib_dev_ops = {
6593 	.owner = THIS_MODULE,
6594 	.driver_id = RDMA_DRIVER_MLX5,
6595 	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
6596 
6597 	.add_gid = mlx5_ib_add_gid,
6598 	.alloc_mr = mlx5_ib_alloc_mr,
6599 	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
6600 	.alloc_pd = mlx5_ib_alloc_pd,
6601 	.alloc_ucontext = mlx5_ib_alloc_ucontext,
6602 	.attach_mcast = mlx5_ib_mcg_attach,
6603 	.check_mr_status = mlx5_ib_check_mr_status,
6604 	.create_ah = mlx5_ib_create_ah,
6605 	.create_counters = mlx5_ib_create_counters,
6606 	.create_cq = mlx5_ib_create_cq,
6607 	.create_flow = mlx5_ib_create_flow,
6608 	.create_qp = mlx5_ib_create_qp,
6609 	.create_srq = mlx5_ib_create_srq,
6610 	.dealloc_pd = mlx5_ib_dealloc_pd,
6611 	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
6612 	.del_gid = mlx5_ib_del_gid,
6613 	.dereg_mr = mlx5_ib_dereg_mr,
6614 	.destroy_ah = mlx5_ib_destroy_ah,
6615 	.destroy_counters = mlx5_ib_destroy_counters,
6616 	.destroy_cq = mlx5_ib_destroy_cq,
6617 	.destroy_flow = mlx5_ib_destroy_flow,
6618 	.destroy_flow_action = mlx5_ib_destroy_flow_action,
6619 	.destroy_qp = mlx5_ib_destroy_qp,
6620 	.destroy_srq = mlx5_ib_destroy_srq,
6621 	.detach_mcast = mlx5_ib_mcg_detach,
6622 	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
6623 	.drain_rq = mlx5_ib_drain_rq,
6624 	.drain_sq = mlx5_ib_drain_sq,
6625 	.enable_driver = mlx5_ib_enable_driver,
6626 	.fill_res_entry = mlx5_ib_fill_res_entry,
6627 	.fill_stat_entry = mlx5_ib_fill_stat_entry,
6628 	.get_dev_fw_str = get_dev_fw_str,
6629 	.get_dma_mr = mlx5_ib_get_dma_mr,
6630 	.get_link_layer = mlx5_ib_port_link_layer,
6631 	.map_mr_sg = mlx5_ib_map_mr_sg,
6632 	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
6633 	.mmap = mlx5_ib_mmap,
6634 	.mmap_free = mlx5_ib_mmap_free,
6635 	.modify_cq = mlx5_ib_modify_cq,
6636 	.modify_device = mlx5_ib_modify_device,
6637 	.modify_port = mlx5_ib_modify_port,
6638 	.modify_qp = mlx5_ib_modify_qp,
6639 	.modify_srq = mlx5_ib_modify_srq,
6640 	.poll_cq = mlx5_ib_poll_cq,
6641 	.post_recv = mlx5_ib_post_recv,
6642 	.post_send = mlx5_ib_post_send,
6643 	.post_srq_recv = mlx5_ib_post_srq_recv,
6644 	.process_mad = mlx5_ib_process_mad,
6645 	.query_ah = mlx5_ib_query_ah,
6646 	.query_device = mlx5_ib_query_device,
6647 	.query_gid = mlx5_ib_query_gid,
6648 	.query_pkey = mlx5_ib_query_pkey,
6649 	.query_qp = mlx5_ib_query_qp,
6650 	.query_srq = mlx5_ib_query_srq,
6651 	.read_counters = mlx5_ib_read_counters,
6652 	.reg_user_mr = mlx5_ib_reg_user_mr,
6653 	.req_notify_cq = mlx5_ib_arm_cq,
6654 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
6655 	.resize_cq = mlx5_ib_resize_cq,
6656 
6657 	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
6658 	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
6659 	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
6660 	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
6661 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
6662 };
6663 
6664 static const struct ib_device_ops mlx5_ib_dev_flow_ipsec_ops = {
6665 	.create_flow_action_esp = mlx5_ib_create_flow_action_esp,
6666 	.modify_flow_action_esp = mlx5_ib_modify_flow_action_esp,
6667 };
6668 
6669 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
6670 	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
6671 };
6672 
6673 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
6674 	.get_vf_config = mlx5_ib_get_vf_config,
6675 	.get_vf_guid = mlx5_ib_get_vf_guid,
6676 	.get_vf_stats = mlx5_ib_get_vf_stats,
6677 	.set_vf_guid = mlx5_ib_set_vf_guid,
6678 	.set_vf_link_state = mlx5_ib_set_vf_link_state,
6679 };
6680 
6681 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
6682 	.alloc_mw = mlx5_ib_alloc_mw,
6683 	.dealloc_mw = mlx5_ib_dealloc_mw,
6684 };
6685 
6686 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
6687 	.alloc_xrcd = mlx5_ib_alloc_xrcd,
6688 	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
6689 };
6690 
6691 static const struct ib_device_ops mlx5_ib_dev_dm_ops = {
6692 	.alloc_dm = mlx5_ib_alloc_dm,
6693 	.dealloc_dm = mlx5_ib_dealloc_dm,
6694 	.reg_dm_mr = mlx5_ib_reg_dm_mr,
6695 };
6696 
6697 static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
6698 {
6699 	struct mlx5_core_dev *mdev = dev->mdev;
6700 	struct mlx5_var_table *var_table = &dev->var_table;
6701 	u8 log_doorbell_bar_size;
6702 	u8 log_doorbell_stride;
6703 	u64 bar_size;
6704 
6705 	log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
6706 					log_doorbell_bar_size);
6707 	log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
6708 					log_doorbell_stride);
6709 	var_table->hw_start_addr = dev->mdev->bar_addr +
6710 				MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
6711 					doorbell_bar_offset);
6712 	bar_size = (1ULL << log_doorbell_bar_size) * 4096;
6713 	var_table->stride_size = 1ULL << log_doorbell_stride;
6714 	var_table->num_var_hw_entries = div_u64(bar_size,
6715 						var_table->stride_size);
6716 	mutex_init(&var_table->bitmap_lock);
6717 	var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
6718 					  GFP_KERNEL);
6719 	return (var_table->bitmap) ? 0 : -ENOMEM;
6720 }
6721 
6722 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
6723 {
6724 	bitmap_free(dev->var_table.bitmap);
6725 }
6726 
6727 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
6728 {
6729 	struct mlx5_core_dev *mdev = dev->mdev;
6730 	int err;
6731 
6732 	dev->ib_dev.uverbs_cmd_mask	=
6733 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
6734 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
6735 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
6736 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
6737 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
6738 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
6739 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
6740 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
6741 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
6742 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
6743 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
6744 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
6745 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
6746 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
6747 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
6748 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
6749 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
6750 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
6751 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
6752 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
6753 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
6754 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
6755 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
6756 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
6757 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
6758 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
6759 	dev->ib_dev.uverbs_ex_cmd_mask =
6760 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
6761 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
6762 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP)	|
6763 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP)	|
6764 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_CQ)	|
6765 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW)	|
6766 		(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW);
6767 
6768 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
6769 	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
6770 		ib_set_device_ops(&dev->ib_dev,
6771 				  &mlx5_ib_dev_ipoib_enhanced_ops);
6772 
6773 	if (mlx5_core_is_pf(mdev))
6774 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
6775 
6776 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
6777 
6778 	if (MLX5_CAP_GEN(mdev, imaicl)) {
6779 		dev->ib_dev.uverbs_cmd_mask |=
6780 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
6781 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
6782 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
6783 	}
6784 
6785 	if (MLX5_CAP_GEN(mdev, xrc)) {
6786 		dev->ib_dev.uverbs_cmd_mask |=
6787 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
6788 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
6789 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
6790 	}
6791 
6792 	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
6793 	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
6794 	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
6795 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
6796 
6797 	if (mlx5_accel_ipsec_device_caps(dev->mdev) &
6798 	    MLX5_ACCEL_IPSEC_CAP_DEVICE)
6799 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_flow_ipsec_ops);
6800 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
6801 
6802 	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
6803 		dev->ib_dev.driver_def = mlx5_ib_defs;
6804 
6805 	err = init_node_data(dev);
6806 	if (err)
6807 		return err;
6808 
6809 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
6810 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
6811 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
6812 		mutex_init(&dev->lb.mutex);
6813 
6814 	if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
6815 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
6816 		err = mlx5_ib_init_var_table(dev);
6817 		if (err)
6818 			return err;
6819 	}
6820 
6821 	dev->ib_dev.use_cq_dim = true;
6822 
6823 	return 0;
6824 }
6825 
6826 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
6827 	.get_port_immutable = mlx5_port_immutable,
6828 	.query_port = mlx5_ib_query_port,
6829 };
6830 
6831 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
6832 {
6833 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
6834 	return 0;
6835 }
6836 
6837 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
6838 	.get_port_immutable = mlx5_port_rep_immutable,
6839 	.query_port = mlx5_ib_rep_query_port,
6840 };
6841 
6842 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
6843 {
6844 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
6845 	return 0;
6846 }
6847 
6848 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
6849 	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
6850 	.create_wq = mlx5_ib_create_wq,
6851 	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
6852 	.destroy_wq = mlx5_ib_destroy_wq,
6853 	.get_netdev = mlx5_ib_get_netdev,
6854 	.modify_wq = mlx5_ib_modify_wq,
6855 };
6856 
6857 static int mlx5_ib_stage_common_roce_init(struct mlx5_ib_dev *dev)
6858 {
6859 	u8 port_num;
6860 
6861 	dev->ib_dev.uverbs_ex_cmd_mask |=
6862 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
6863 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
6864 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
6865 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
6866 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
6867 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
6868 
6869 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
6870 
6871 	/* Register only for native ports */
6872 	return mlx5_add_netdev_notifier(dev, port_num);
6873 }
6874 
6875 static void mlx5_ib_stage_common_roce_cleanup(struct mlx5_ib_dev *dev)
6876 {
6877 	u8 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
6878 
6879 	mlx5_remove_netdev_notifier(dev, port_num);
6880 }
6881 
6882 static int mlx5_ib_stage_raw_eth_roce_init(struct mlx5_ib_dev *dev)
6883 {
6884 	struct mlx5_core_dev *mdev = dev->mdev;
6885 	enum rdma_link_layer ll;
6886 	int port_type_cap;
6887 	int err = 0;
6888 
6889 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
6890 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
6891 
6892 	if (ll == IB_LINK_LAYER_ETHERNET)
6893 		err = mlx5_ib_stage_common_roce_init(dev);
6894 
6895 	return err;
6896 }
6897 
6898 static void mlx5_ib_stage_raw_eth_roce_cleanup(struct mlx5_ib_dev *dev)
6899 {
6900 	mlx5_ib_stage_common_roce_cleanup(dev);
6901 }
6902 
6903 static int mlx5_ib_stage_roce_init(struct mlx5_ib_dev *dev)
6904 {
6905 	struct mlx5_core_dev *mdev = dev->mdev;
6906 	enum rdma_link_layer ll;
6907 	int port_type_cap;
6908 	int err;
6909 
6910 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
6911 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
6912 
6913 	if (ll == IB_LINK_LAYER_ETHERNET) {
6914 		err = mlx5_ib_stage_common_roce_init(dev);
6915 		if (err)
6916 			return err;
6917 
6918 		err = mlx5_enable_eth(dev);
6919 		if (err)
6920 			goto cleanup;
6921 	}
6922 
6923 	return 0;
6924 cleanup:
6925 	mlx5_ib_stage_common_roce_cleanup(dev);
6926 
6927 	return err;
6928 }
6929 
6930 static void mlx5_ib_stage_roce_cleanup(struct mlx5_ib_dev *dev)
6931 {
6932 	struct mlx5_core_dev *mdev = dev->mdev;
6933 	enum rdma_link_layer ll;
6934 	int port_type_cap;
6935 
6936 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
6937 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
6938 
6939 	if (ll == IB_LINK_LAYER_ETHERNET) {
6940 		mlx5_disable_eth(dev);
6941 		mlx5_ib_stage_common_roce_cleanup(dev);
6942 	}
6943 }
6944 
6945 static int mlx5_ib_stage_dev_res_init(struct mlx5_ib_dev *dev)
6946 {
6947 	return create_dev_resources(&dev->devr);
6948 }
6949 
6950 static void mlx5_ib_stage_dev_res_cleanup(struct mlx5_ib_dev *dev)
6951 {
6952 	destroy_dev_resources(&dev->devr);
6953 }
6954 
6955 static int mlx5_ib_stage_odp_init(struct mlx5_ib_dev *dev)
6956 {
6957 	return mlx5_ib_odp_init_one(dev);
6958 }
6959 
6960 static void mlx5_ib_stage_odp_cleanup(struct mlx5_ib_dev *dev)
6961 {
6962 	mlx5_ib_odp_cleanup_one(dev);
6963 }
6964 
6965 static const struct ib_device_ops mlx5_ib_dev_hw_stats_ops = {
6966 	.alloc_hw_stats = mlx5_ib_alloc_hw_stats,
6967 	.get_hw_stats = mlx5_ib_get_hw_stats,
6968 	.counter_bind_qp = mlx5_ib_counter_bind_qp,
6969 	.counter_unbind_qp = mlx5_ib_counter_unbind_qp,
6970 	.counter_dealloc = mlx5_ib_counter_dealloc,
6971 	.counter_alloc_stats = mlx5_ib_counter_alloc_stats,
6972 	.counter_update_stats = mlx5_ib_counter_update_stats,
6973 };
6974 
6975 static int mlx5_ib_stage_counters_init(struct mlx5_ib_dev *dev)
6976 {
6977 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
6978 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_hw_stats_ops);
6979 
6980 		return mlx5_ib_alloc_counters(dev);
6981 	}
6982 
6983 	return 0;
6984 }
6985 
6986 static void mlx5_ib_stage_counters_cleanup(struct mlx5_ib_dev *dev)
6987 {
6988 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
6989 		mlx5_ib_dealloc_counters(dev);
6990 }
6991 
6992 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
6993 {
6994 	mlx5_ib_init_cong_debugfs(dev,
6995 				  mlx5_core_native_port_num(dev->mdev) - 1);
6996 	return 0;
6997 }
6998 
6999 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
7000 {
7001 	mlx5_ib_cleanup_cong_debugfs(dev,
7002 				     mlx5_core_native_port_num(dev->mdev) - 1);
7003 }
7004 
7005 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
7006 {
7007 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
7008 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
7009 }
7010 
7011 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
7012 {
7013 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
7014 }
7015 
7016 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
7017 {
7018 	int err;
7019 
7020 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
7021 	if (err)
7022 		return err;
7023 
7024 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
7025 	if (err)
7026 		mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
7027 
7028 	return err;
7029 }
7030 
7031 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
7032 {
7033 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
7034 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
7035 }
7036 
7037 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
7038 {
7039 	const char *name;
7040 
7041 	rdma_set_device_sysfs_group(&dev->ib_dev, &mlx5_attr_group);
7042 	if (!mlx5_lag_is_roce(dev->mdev))
7043 		name = "mlx5_%d";
7044 	else
7045 		name = "mlx5_bond_%d";
7046 	return ib_register_device(&dev->ib_dev, name);
7047 }
7048 
7049 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
7050 {
7051 	destroy_umrc_res(dev);
7052 }
7053 
7054 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
7055 {
7056 	ib_unregister_device(&dev->ib_dev);
7057 }
7058 
7059 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
7060 {
7061 	return create_umr_res(dev);
7062 }
7063 
7064 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
7065 {
7066 	init_delay_drop(dev);
7067 
7068 	return 0;
7069 }
7070 
7071 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
7072 {
7073 	cancel_delay_drop(dev);
7074 }
7075 
7076 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
7077 {
7078 	dev->mdev_events.notifier_call = mlx5_ib_event;
7079 	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
7080 	return 0;
7081 }
7082 
7083 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
7084 {
7085 	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
7086 }
7087 
7088 static int mlx5_ib_stage_devx_init(struct mlx5_ib_dev *dev)
7089 {
7090 	int uid;
7091 
7092 	uid = mlx5_ib_devx_create(dev, false);
7093 	if (uid > 0) {
7094 		dev->devx_whitelist_uid = uid;
7095 		mlx5_ib_devx_init_event_table(dev);
7096 	}
7097 
7098 	return 0;
7099 }
7100 static void mlx5_ib_stage_devx_cleanup(struct mlx5_ib_dev *dev)
7101 {
7102 	if (dev->devx_whitelist_uid) {
7103 		mlx5_ib_devx_cleanup_event_table(dev);
7104 		mlx5_ib_devx_destroy(dev, dev->devx_whitelist_uid);
7105 	}
7106 }
7107 
7108 int mlx5_ib_enable_driver(struct ib_device *dev)
7109 {
7110 	struct mlx5_ib_dev *mdev = to_mdev(dev);
7111 	int ret;
7112 
7113 	ret = mlx5_ib_test_wc(mdev);
7114 	mlx5_ib_dbg(mdev, "Write-Combining %s",
7115 		    mdev->wc_support ? "supported" : "not supported");
7116 
7117 	return ret;
7118 }
7119 
7120 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
7121 		      const struct mlx5_ib_profile *profile,
7122 		      int stage)
7123 {
7124 	dev->ib_active = false;
7125 
7126 	/* Number of stages to cleanup */
7127 	while (stage) {
7128 		stage--;
7129 		if (profile->stage[stage].cleanup)
7130 			profile->stage[stage].cleanup(dev);
7131 	}
7132 
7133 	kfree(dev->port);
7134 	ib_dealloc_device(&dev->ib_dev);
7135 }
7136 
7137 void *__mlx5_ib_add(struct mlx5_ib_dev *dev,
7138 		    const struct mlx5_ib_profile *profile)
7139 {
7140 	int err;
7141 	int i;
7142 
7143 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
7144 		if (profile->stage[i].init) {
7145 			err = profile->stage[i].init(dev);
7146 			if (err)
7147 				goto err_out;
7148 		}
7149 	}
7150 
7151 	dev->profile = profile;
7152 	dev->ib_active = true;
7153 
7154 	return dev;
7155 
7156 err_out:
7157 	__mlx5_ib_remove(dev, profile, i);
7158 
7159 	return NULL;
7160 }
7161 
7162 static const struct mlx5_ib_profile pf_profile = {
7163 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
7164 		     mlx5_ib_stage_init_init,
7165 		     mlx5_ib_stage_init_cleanup),
7166 	STAGE_CREATE(MLX5_IB_STAGE_FLOW_DB,
7167 		     mlx5_ib_stage_flow_db_init,
7168 		     mlx5_ib_stage_flow_db_cleanup),
7169 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
7170 		     mlx5_ib_stage_caps_init,
7171 		     mlx5_ib_stage_caps_cleanup),
7172 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
7173 		     mlx5_ib_stage_non_default_cb,
7174 		     NULL),
7175 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
7176 		     mlx5_ib_stage_roce_init,
7177 		     mlx5_ib_stage_roce_cleanup),
7178 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
7179 		     mlx5_init_srq_table,
7180 		     mlx5_cleanup_srq_table),
7181 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
7182 		     mlx5_ib_stage_dev_res_init,
7183 		     mlx5_ib_stage_dev_res_cleanup),
7184 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
7185 		     mlx5_ib_stage_dev_notifier_init,
7186 		     mlx5_ib_stage_dev_notifier_cleanup),
7187 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
7188 		     mlx5_ib_stage_odp_init,
7189 		     mlx5_ib_stage_odp_cleanup),
7190 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
7191 		     mlx5_ib_stage_counters_init,
7192 		     mlx5_ib_stage_counters_cleanup),
7193 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
7194 		     mlx5_ib_stage_cong_debugfs_init,
7195 		     mlx5_ib_stage_cong_debugfs_cleanup),
7196 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
7197 		     mlx5_ib_stage_uar_init,
7198 		     mlx5_ib_stage_uar_cleanup),
7199 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
7200 		     mlx5_ib_stage_bfrag_init,
7201 		     mlx5_ib_stage_bfrag_cleanup),
7202 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
7203 		     NULL,
7204 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
7205 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
7206 		     mlx5_ib_stage_devx_init,
7207 		     mlx5_ib_stage_devx_cleanup),
7208 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
7209 		     mlx5_ib_stage_ib_reg_init,
7210 		     mlx5_ib_stage_ib_reg_cleanup),
7211 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
7212 		     mlx5_ib_stage_post_ib_reg_umr_init,
7213 		     NULL),
7214 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
7215 		     mlx5_ib_stage_delay_drop_init,
7216 		     mlx5_ib_stage_delay_drop_cleanup),
7217 };
7218 
7219 const struct mlx5_ib_profile raw_eth_profile = {
7220 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
7221 		     mlx5_ib_stage_init_init,
7222 		     mlx5_ib_stage_init_cleanup),
7223 	STAGE_CREATE(MLX5_IB_STAGE_FLOW_DB,
7224 		     mlx5_ib_stage_flow_db_init,
7225 		     mlx5_ib_stage_flow_db_cleanup),
7226 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
7227 		     mlx5_ib_stage_caps_init,
7228 		     mlx5_ib_stage_caps_cleanup),
7229 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
7230 		     mlx5_ib_stage_raw_eth_non_default_cb,
7231 		     NULL),
7232 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
7233 		     mlx5_ib_stage_raw_eth_roce_init,
7234 		     mlx5_ib_stage_raw_eth_roce_cleanup),
7235 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
7236 		     mlx5_init_srq_table,
7237 		     mlx5_cleanup_srq_table),
7238 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
7239 		     mlx5_ib_stage_dev_res_init,
7240 		     mlx5_ib_stage_dev_res_cleanup),
7241 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
7242 		     mlx5_ib_stage_dev_notifier_init,
7243 		     mlx5_ib_stage_dev_notifier_cleanup),
7244 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
7245 		     mlx5_ib_stage_counters_init,
7246 		     mlx5_ib_stage_counters_cleanup),
7247 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
7248 		     mlx5_ib_stage_cong_debugfs_init,
7249 		     mlx5_ib_stage_cong_debugfs_cleanup),
7250 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
7251 		     mlx5_ib_stage_uar_init,
7252 		     mlx5_ib_stage_uar_cleanup),
7253 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
7254 		     mlx5_ib_stage_bfrag_init,
7255 		     mlx5_ib_stage_bfrag_cleanup),
7256 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
7257 		     NULL,
7258 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
7259 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
7260 		     mlx5_ib_stage_devx_init,
7261 		     mlx5_ib_stage_devx_cleanup),
7262 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
7263 		     mlx5_ib_stage_ib_reg_init,
7264 		     mlx5_ib_stage_ib_reg_cleanup),
7265 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
7266 		     mlx5_ib_stage_post_ib_reg_umr_init,
7267 		     NULL),
7268 };
7269 
7270 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev)
7271 {
7272 	struct mlx5_ib_multiport_info *mpi;
7273 	struct mlx5_ib_dev *dev;
7274 	bool bound = false;
7275 	int err;
7276 
7277 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
7278 	if (!mpi)
7279 		return NULL;
7280 
7281 	mpi->mdev = mdev;
7282 
7283 	err = mlx5_query_nic_vport_system_image_guid(mdev,
7284 						     &mpi->sys_image_guid);
7285 	if (err) {
7286 		kfree(mpi);
7287 		return NULL;
7288 	}
7289 
7290 	mutex_lock(&mlx5_ib_multiport_mutex);
7291 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
7292 		if (dev->sys_image_guid == mpi->sys_image_guid)
7293 			bound = mlx5_ib_bind_slave_port(dev, mpi);
7294 
7295 		if (bound) {
7296 			rdma_roce_rescan_device(&dev->ib_dev);
7297 			break;
7298 		}
7299 	}
7300 
7301 	if (!bound) {
7302 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
7303 		dev_dbg(mdev->device,
7304 			"no suitable IB device found to bind to, added to unaffiliated list.\n");
7305 	}
7306 	mutex_unlock(&mlx5_ib_multiport_mutex);
7307 
7308 	return mpi;
7309 }
7310 
7311 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
7312 {
7313 	const struct mlx5_ib_profile *profile;
7314 	enum rdma_link_layer ll;
7315 	struct mlx5_ib_dev *dev;
7316 	int port_type_cap;
7317 	int num_ports;
7318 
7319 	printk_once(KERN_INFO "%s", mlx5_version);
7320 
7321 	if (MLX5_ESWITCH_MANAGER(mdev) &&
7322 	    mlx5_ib_eswitch_mode(mdev->priv.eswitch) == MLX5_ESWITCH_OFFLOADS) {
7323 		if (!mlx5_core_mp_enabled(mdev))
7324 			mlx5_ib_register_vport_reps(mdev);
7325 		return mdev;
7326 	}
7327 
7328 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
7329 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
7330 
7331 	if (mlx5_core_is_mp_slave(mdev) && ll == IB_LINK_LAYER_ETHERNET)
7332 		return mlx5_ib_add_slave_port(mdev);
7333 
7334 	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
7335 			MLX5_CAP_GEN(mdev, num_vhca_ports));
7336 	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
7337 	if (!dev)
7338 		return NULL;
7339 	dev->port = kcalloc(num_ports, sizeof(*dev->port),
7340 			     GFP_KERNEL);
7341 	if (!dev->port) {
7342 		ib_dealloc_device(&dev->ib_dev);
7343 		return NULL;
7344 	}
7345 
7346 	dev->mdev = mdev;
7347 	dev->num_ports = num_ports;
7348 
7349 	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_is_roce_enabled(mdev))
7350 		profile = &raw_eth_profile;
7351 	else
7352 		profile = &pf_profile;
7353 
7354 	return __mlx5_ib_add(dev, profile);
7355 }
7356 
7357 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
7358 {
7359 	struct mlx5_ib_multiport_info *mpi;
7360 	struct mlx5_ib_dev *dev;
7361 
7362 	if (MLX5_ESWITCH_MANAGER(mdev) && context == mdev) {
7363 		mlx5_ib_unregister_vport_reps(mdev);
7364 		return;
7365 	}
7366 
7367 	if (mlx5_core_is_mp_slave(mdev)) {
7368 		mpi = context;
7369 		mutex_lock(&mlx5_ib_multiport_mutex);
7370 		if (mpi->ibdev)
7371 			mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
7372 		list_del(&mpi->list);
7373 		mutex_unlock(&mlx5_ib_multiport_mutex);
7374 		kfree(mpi);
7375 		return;
7376 	}
7377 
7378 	dev = context;
7379 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
7380 }
7381 
7382 static struct mlx5_interface mlx5_ib_interface = {
7383 	.add            = mlx5_ib_add,
7384 	.remove         = mlx5_ib_remove,
7385 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
7386 };
7387 
7388 unsigned long mlx5_ib_get_xlt_emergency_page(void)
7389 {
7390 	mutex_lock(&xlt_emergency_page_mutex);
7391 	return xlt_emergency_page;
7392 }
7393 
7394 void mlx5_ib_put_xlt_emergency_page(void)
7395 {
7396 	mutex_unlock(&xlt_emergency_page_mutex);
7397 }
7398 
7399 static int __init mlx5_ib_init(void)
7400 {
7401 	int err;
7402 
7403 	xlt_emergency_page = __get_free_page(GFP_KERNEL);
7404 	if (!xlt_emergency_page)
7405 		return -ENOMEM;
7406 
7407 	mutex_init(&xlt_emergency_page_mutex);
7408 
7409 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
7410 	if (!mlx5_ib_event_wq) {
7411 		free_page(xlt_emergency_page);
7412 		return -ENOMEM;
7413 	}
7414 
7415 	mlx5_ib_odp_init();
7416 
7417 	err = mlx5_register_interface(&mlx5_ib_interface);
7418 
7419 	return err;
7420 }
7421 
7422 static void __exit mlx5_ib_cleanup(void)
7423 {
7424 	mlx5_unregister_interface(&mlx5_ib_interface);
7425 	destroy_workqueue(mlx5_ib_event_wq);
7426 	mutex_destroy(&xlt_emergency_page_mutex);
7427 	free_page(xlt_emergency_page);
7428 }
7429 
7430 module_init(mlx5_ib_init);
7431 module_exit(mlx5_ib_cleanup);
7432