xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision f8e17c17)
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 #if defined(CONFIG_X86)
43 #include <asm/memtype.h>
44 #endif
45 #include <linux/sched.h>
46 #include <linux/sched/mm.h>
47 #include <linux/sched/task.h>
48 #include <linux/delay.h>
49 #include <rdma/ib_user_verbs.h>
50 #include <rdma/ib_addr.h>
51 #include <rdma/ib_cache.h>
52 #include <linux/mlx5/port.h>
53 #include <linux/mlx5/vport.h>
54 #include <linux/mlx5/fs.h>
55 #include <linux/mlx5/eswitch.h>
56 #include <linux/list.h>
57 #include <rdma/ib_smi.h>
58 #include <rdma/ib_umem.h>
59 #include <linux/in.h>
60 #include <linux/etherdevice.h>
61 #include "mlx5_ib.h"
62 #include "ib_rep.h"
63 #include "cmd.h"
64 #include "srq.h"
65 #include <linux/mlx5/fs_helpers.h>
66 #include <linux/mlx5/accel.h>
67 #include <rdma/uverbs_std_types.h>
68 #include <rdma/mlx5_user_ioctl_verbs.h>
69 #include <rdma/mlx5_user_ioctl_cmds.h>
70 #include <rdma/ib_umem_odp.h>
71 
72 #define UVERBS_MODULE_NAME mlx5_ib
73 #include <rdma/uverbs_named_ioctl.h>
74 
75 #define DRIVER_NAME "mlx5_ib"
76 #define DRIVER_VERSION "5.0-0"
77 
78 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
79 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
80 MODULE_LICENSE("Dual BSD/GPL");
81 
82 static char mlx5_version[] =
83 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
84 	DRIVER_VERSION "\n";
85 
86 struct mlx5_ib_event_work {
87 	struct work_struct	work;
88 	union {
89 		struct mlx5_ib_dev	      *dev;
90 		struct mlx5_ib_multiport_info *mpi;
91 	};
92 	bool			is_slave;
93 	unsigned int		event;
94 	void			*param;
95 };
96 
97 enum {
98 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
99 };
100 
101 static struct workqueue_struct *mlx5_ib_event_wq;
102 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
103 static LIST_HEAD(mlx5_ib_dev_list);
104 /*
105  * This mutex should be held when accessing either of the above lists
106  */
107 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
108 
109 /* We can't use an array for xlt_emergency_page because dma_map_single
110  * doesn't work on kernel modules memory
111  */
112 static unsigned long xlt_emergency_page;
113 static struct mutex xlt_emergency_page_mutex;
114 
115 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
116 {
117 	struct mlx5_ib_dev *dev;
118 
119 	mutex_lock(&mlx5_ib_multiport_mutex);
120 	dev = mpi->ibdev;
121 	mutex_unlock(&mlx5_ib_multiport_mutex);
122 	return dev;
123 }
124 
125 static enum rdma_link_layer
126 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
127 {
128 	switch (port_type_cap) {
129 	case MLX5_CAP_PORT_TYPE_IB:
130 		return IB_LINK_LAYER_INFINIBAND;
131 	case MLX5_CAP_PORT_TYPE_ETH:
132 		return IB_LINK_LAYER_ETHERNET;
133 	default:
134 		return IB_LINK_LAYER_UNSPECIFIED;
135 	}
136 }
137 
138 static enum rdma_link_layer
139 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
140 {
141 	struct mlx5_ib_dev *dev = to_mdev(device);
142 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
143 
144 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
145 }
146 
147 static int get_port_state(struct ib_device *ibdev,
148 			  u8 port_num,
149 			  enum ib_port_state *state)
150 {
151 	struct ib_port_attr attr;
152 	int ret;
153 
154 	memset(&attr, 0, sizeof(attr));
155 	ret = ibdev->ops.query_port(ibdev, port_num, &attr);
156 	if (!ret)
157 		*state = attr.state;
158 	return ret;
159 }
160 
161 static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev,
162 					   struct net_device *ndev,
163 					   u8 *port_num)
164 {
165 	struct mlx5_eswitch *esw = dev->mdev->priv.eswitch;
166 	struct net_device *rep_ndev;
167 	struct mlx5_ib_port *port;
168 	int i;
169 
170 	for (i = 0; i < dev->num_ports; i++) {
171 		port  = &dev->port[i];
172 		if (!port->rep)
173 			continue;
174 
175 		read_lock(&port->roce.netdev_lock);
176 		rep_ndev = mlx5_ib_get_rep_netdev(esw,
177 						  port->rep->vport);
178 		if (rep_ndev == ndev) {
179 			read_unlock(&port->roce.netdev_lock);
180 			*port_num = i + 1;
181 			return &port->roce;
182 		}
183 		read_unlock(&port->roce.netdev_lock);
184 	}
185 
186 	return NULL;
187 }
188 
189 static int mlx5_netdev_event(struct notifier_block *this,
190 			     unsigned long event, void *ptr)
191 {
192 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
193 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
194 	u8 port_num = roce->native_port_num;
195 	struct mlx5_core_dev *mdev;
196 	struct mlx5_ib_dev *ibdev;
197 
198 	ibdev = roce->dev;
199 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
200 	if (!mdev)
201 		return NOTIFY_DONE;
202 
203 	switch (event) {
204 	case NETDEV_REGISTER:
205 		/* Should already be registered during the load */
206 		if (ibdev->is_rep)
207 			break;
208 		write_lock(&roce->netdev_lock);
209 		if (ndev->dev.parent == mdev->device)
210 			roce->netdev = ndev;
211 		write_unlock(&roce->netdev_lock);
212 		break;
213 
214 	case NETDEV_UNREGISTER:
215 		/* In case of reps, ib device goes away before the netdevs */
216 		write_lock(&roce->netdev_lock);
217 		if (roce->netdev == ndev)
218 			roce->netdev = NULL;
219 		write_unlock(&roce->netdev_lock);
220 		break;
221 
222 	case NETDEV_CHANGE:
223 	case NETDEV_UP:
224 	case NETDEV_DOWN: {
225 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
226 		struct net_device *upper = NULL;
227 
228 		if (lag_ndev) {
229 			upper = netdev_master_upper_dev_get(lag_ndev);
230 			dev_put(lag_ndev);
231 		}
232 
233 		if (ibdev->is_rep)
234 			roce = mlx5_get_rep_roce(ibdev, ndev, &port_num);
235 		if (!roce)
236 			return NOTIFY_DONE;
237 		if ((upper == ndev || (!upper && ndev == roce->netdev))
238 		    && ibdev->ib_active) {
239 			struct ib_event ibev = { };
240 			enum ib_port_state port_state;
241 
242 			if (get_port_state(&ibdev->ib_dev, port_num,
243 					   &port_state))
244 				goto done;
245 
246 			if (roce->last_port_state == port_state)
247 				goto done;
248 
249 			roce->last_port_state = port_state;
250 			ibev.device = &ibdev->ib_dev;
251 			if (port_state == IB_PORT_DOWN)
252 				ibev.event = IB_EVENT_PORT_ERR;
253 			else if (port_state == IB_PORT_ACTIVE)
254 				ibev.event = IB_EVENT_PORT_ACTIVE;
255 			else
256 				goto done;
257 
258 			ibev.element.port_num = port_num;
259 			ib_dispatch_event(&ibev);
260 		}
261 		break;
262 	}
263 
264 	default:
265 		break;
266 	}
267 done:
268 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
269 	return NOTIFY_DONE;
270 }
271 
272 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
273 					     u8 port_num)
274 {
275 	struct mlx5_ib_dev *ibdev = to_mdev(device);
276 	struct net_device *ndev;
277 	struct mlx5_core_dev *mdev;
278 
279 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
280 	if (!mdev)
281 		return NULL;
282 
283 	ndev = mlx5_lag_get_roce_netdev(mdev);
284 	if (ndev)
285 		goto out;
286 
287 	/* Ensure ndev does not disappear before we invoke dev_hold()
288 	 */
289 	read_lock(&ibdev->port[port_num - 1].roce.netdev_lock);
290 	ndev = ibdev->port[port_num - 1].roce.netdev;
291 	if (ndev)
292 		dev_hold(ndev);
293 	read_unlock(&ibdev->port[port_num - 1].roce.netdev_lock);
294 
295 out:
296 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
297 	return ndev;
298 }
299 
300 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
301 						   u8 ib_port_num,
302 						   u8 *native_port_num)
303 {
304 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
305 							  ib_port_num);
306 	struct mlx5_core_dev *mdev = NULL;
307 	struct mlx5_ib_multiport_info *mpi;
308 	struct mlx5_ib_port *port;
309 
310 	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
311 	    ll != IB_LINK_LAYER_ETHERNET) {
312 		if (native_port_num)
313 			*native_port_num = ib_port_num;
314 		return ibdev->mdev;
315 	}
316 
317 	if (native_port_num)
318 		*native_port_num = 1;
319 
320 	port = &ibdev->port[ib_port_num - 1];
321 	if (!port)
322 		return NULL;
323 
324 	spin_lock(&port->mp.mpi_lock);
325 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
326 	if (mpi && !mpi->unaffiliate) {
327 		mdev = mpi->mdev;
328 		/* If it's the master no need to refcount, it'll exist
329 		 * as long as the ib_dev exists.
330 		 */
331 		if (!mpi->is_master)
332 			mpi->mdev_refcnt++;
333 	}
334 	spin_unlock(&port->mp.mpi_lock);
335 
336 	return mdev;
337 }
338 
339 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u8 port_num)
340 {
341 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
342 							  port_num);
343 	struct mlx5_ib_multiport_info *mpi;
344 	struct mlx5_ib_port *port;
345 
346 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
347 		return;
348 
349 	port = &ibdev->port[port_num - 1];
350 
351 	spin_lock(&port->mp.mpi_lock);
352 	mpi = ibdev->port[port_num - 1].mp.mpi;
353 	if (mpi->is_master)
354 		goto out;
355 
356 	mpi->mdev_refcnt--;
357 	if (mpi->unaffiliate)
358 		complete(&mpi->unref_comp);
359 out:
360 	spin_unlock(&port->mp.mpi_lock);
361 }
362 
363 static int translate_eth_legacy_proto_oper(u32 eth_proto_oper, u8 *active_speed,
364 					   u8 *active_width)
365 {
366 	switch (eth_proto_oper) {
367 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
368 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
369 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
370 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
371 		*active_width = IB_WIDTH_1X;
372 		*active_speed = IB_SPEED_SDR;
373 		break;
374 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
375 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
376 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
377 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
378 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
379 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
380 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
381 		*active_width = IB_WIDTH_1X;
382 		*active_speed = IB_SPEED_QDR;
383 		break;
384 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
385 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
386 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
387 		*active_width = IB_WIDTH_1X;
388 		*active_speed = IB_SPEED_EDR;
389 		break;
390 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
391 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
392 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
393 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
394 		*active_width = IB_WIDTH_4X;
395 		*active_speed = IB_SPEED_QDR;
396 		break;
397 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
398 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
399 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
400 		*active_width = IB_WIDTH_1X;
401 		*active_speed = IB_SPEED_HDR;
402 		break;
403 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
404 		*active_width = IB_WIDTH_4X;
405 		*active_speed = IB_SPEED_FDR;
406 		break;
407 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
408 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
409 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
410 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
411 		*active_width = IB_WIDTH_4X;
412 		*active_speed = IB_SPEED_EDR;
413 		break;
414 	default:
415 		return -EINVAL;
416 	}
417 
418 	return 0;
419 }
420 
421 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u8 *active_speed,
422 					u8 *active_width)
423 {
424 	switch (eth_proto_oper) {
425 	case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
426 	case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
427 		*active_width = IB_WIDTH_1X;
428 		*active_speed = IB_SPEED_SDR;
429 		break;
430 	case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
431 		*active_width = IB_WIDTH_1X;
432 		*active_speed = IB_SPEED_DDR;
433 		break;
434 	case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
435 		*active_width = IB_WIDTH_1X;
436 		*active_speed = IB_SPEED_QDR;
437 		break;
438 	case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
439 		*active_width = IB_WIDTH_4X;
440 		*active_speed = IB_SPEED_QDR;
441 		break;
442 	case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
443 		*active_width = IB_WIDTH_1X;
444 		*active_speed = IB_SPEED_EDR;
445 		break;
446 	case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
447 		*active_width = IB_WIDTH_2X;
448 		*active_speed = IB_SPEED_EDR;
449 		break;
450 	case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
451 		*active_width = IB_WIDTH_1X;
452 		*active_speed = IB_SPEED_HDR;
453 		break;
454 	case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
455 		*active_width = IB_WIDTH_4X;
456 		*active_speed = IB_SPEED_EDR;
457 		break;
458 	case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
459 		*active_width = IB_WIDTH_2X;
460 		*active_speed = IB_SPEED_HDR;
461 		break;
462 	case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4):
463 		*active_width = IB_WIDTH_4X;
464 		*active_speed = IB_SPEED_HDR;
465 		break;
466 	default:
467 		return -EINVAL;
468 	}
469 
470 	return 0;
471 }
472 
473 static int translate_eth_proto_oper(u32 eth_proto_oper, u8 *active_speed,
474 				    u8 *active_width, bool ext)
475 {
476 	return ext ?
477 		translate_eth_ext_proto_oper(eth_proto_oper, active_speed,
478 					     active_width) :
479 		translate_eth_legacy_proto_oper(eth_proto_oper, active_speed,
480 						active_width);
481 }
482 
483 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
484 				struct ib_port_attr *props)
485 {
486 	struct mlx5_ib_dev *dev = to_mdev(device);
487 	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0};
488 	struct mlx5_core_dev *mdev;
489 	struct net_device *ndev, *upper;
490 	enum ib_mtu ndev_ib_mtu;
491 	bool put_mdev = true;
492 	u16 qkey_viol_cntr;
493 	u32 eth_prot_oper;
494 	u8 mdev_port_num;
495 	bool ext;
496 	int err;
497 
498 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
499 	if (!mdev) {
500 		/* This means the port isn't affiliated yet. Get the
501 		 * info for the master port instead.
502 		 */
503 		put_mdev = false;
504 		mdev = dev->mdev;
505 		mdev_port_num = 1;
506 		port_num = 1;
507 	}
508 
509 	/* Possible bad flows are checked before filling out props so in case
510 	 * of an error it will still be zeroed out.
511 	 * Use native port in case of reps
512 	 */
513 	if (dev->is_rep)
514 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
515 					   1);
516 	else
517 		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
518 					   mdev_port_num);
519 	if (err)
520 		goto out;
521 	ext = MLX5_CAP_PCAM_FEATURE(dev->mdev, ptys_extended_ethernet);
522 	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
523 
524 	props->active_width     = IB_WIDTH_4X;
525 	props->active_speed     = IB_SPEED_QDR;
526 
527 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
528 				 &props->active_width, ext);
529 
530 	props->port_cap_flags |= IB_PORT_CM_SUP;
531 	props->ip_gids = true;
532 
533 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
534 						roce_address_table_size);
535 	props->max_mtu          = IB_MTU_4096;
536 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
537 	props->pkey_tbl_len     = 1;
538 	props->state            = IB_PORT_DOWN;
539 	props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
540 
541 	mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
542 	props->qkey_viol_cntr = qkey_viol_cntr;
543 
544 	/* If this is a stub query for an unaffiliated port stop here */
545 	if (!put_mdev)
546 		goto out;
547 
548 	ndev = mlx5_ib_get_netdev(device, port_num);
549 	if (!ndev)
550 		goto out;
551 
552 	if (dev->lag_active) {
553 		rcu_read_lock();
554 		upper = netdev_master_upper_dev_get_rcu(ndev);
555 		if (upper) {
556 			dev_put(ndev);
557 			ndev = upper;
558 			dev_hold(ndev);
559 		}
560 		rcu_read_unlock();
561 	}
562 
563 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
564 		props->state      = IB_PORT_ACTIVE;
565 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
566 	}
567 
568 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
569 
570 	dev_put(ndev);
571 
572 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
573 out:
574 	if (put_mdev)
575 		mlx5_ib_put_native_port_mdev(dev, port_num);
576 	return err;
577 }
578 
579 static int set_roce_addr(struct mlx5_ib_dev *dev, u8 port_num,
580 			 unsigned int index, const union ib_gid *gid,
581 			 const struct ib_gid_attr *attr)
582 {
583 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
584 	u16 vlan_id = 0xffff;
585 	u8 roce_version = 0;
586 	u8 roce_l3_type = 0;
587 	u8 mac[ETH_ALEN];
588 	int ret;
589 
590 	if (gid) {
591 		gid_type = attr->gid_type;
592 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
593 		if (ret)
594 			return ret;
595 	}
596 
597 	switch (gid_type) {
598 	case IB_GID_TYPE_IB:
599 		roce_version = MLX5_ROCE_VERSION_1;
600 		break;
601 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
602 		roce_version = MLX5_ROCE_VERSION_2;
603 		if (ipv6_addr_v4mapped((void *)gid))
604 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
605 		else
606 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
607 		break;
608 
609 	default:
610 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
611 	}
612 
613 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
614 				      roce_l3_type, gid->raw, mac,
615 				      vlan_id < VLAN_CFI_MASK, vlan_id,
616 				      port_num);
617 }
618 
619 static int mlx5_ib_add_gid(const struct ib_gid_attr *attr,
620 			   __always_unused void **context)
621 {
622 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
623 			     attr->index, &attr->gid, attr);
624 }
625 
626 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
627 			   __always_unused void **context)
628 {
629 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
630 			     attr->index, NULL, NULL);
631 }
632 
633 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev,
634 			       const struct ib_gid_attr *attr)
635 {
636 	if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
637 		return 0;
638 
639 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
640 }
641 
642 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
643 {
644 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
645 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
646 	return 0;
647 }
648 
649 enum {
650 	MLX5_VPORT_ACCESS_METHOD_MAD,
651 	MLX5_VPORT_ACCESS_METHOD_HCA,
652 	MLX5_VPORT_ACCESS_METHOD_NIC,
653 };
654 
655 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
656 {
657 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
658 		return MLX5_VPORT_ACCESS_METHOD_MAD;
659 
660 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
661 	    IB_LINK_LAYER_ETHERNET)
662 		return MLX5_VPORT_ACCESS_METHOD_NIC;
663 
664 	return MLX5_VPORT_ACCESS_METHOD_HCA;
665 }
666 
667 static void get_atomic_caps(struct mlx5_ib_dev *dev,
668 			    u8 atomic_size_qp,
669 			    struct ib_device_attr *props)
670 {
671 	u8 tmp;
672 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
673 	u8 atomic_req_8B_endianness_mode =
674 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
675 
676 	/* Check if HW supports 8 bytes standard atomic operations and capable
677 	 * of host endianness respond
678 	 */
679 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
680 	if (((atomic_operations & tmp) == tmp) &&
681 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
682 	    (atomic_req_8B_endianness_mode)) {
683 		props->atomic_cap = IB_ATOMIC_HCA;
684 	} else {
685 		props->atomic_cap = IB_ATOMIC_NONE;
686 	}
687 }
688 
689 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
690 			       struct ib_device_attr *props)
691 {
692 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
693 
694 	get_atomic_caps(dev, atomic_size_qp, props);
695 }
696 
697 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
698 					__be64 *sys_image_guid)
699 {
700 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
701 	struct mlx5_core_dev *mdev = dev->mdev;
702 	u64 tmp;
703 	int err;
704 
705 	switch (mlx5_get_vport_access_method(ibdev)) {
706 	case MLX5_VPORT_ACCESS_METHOD_MAD:
707 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
708 							    sys_image_guid);
709 
710 	case MLX5_VPORT_ACCESS_METHOD_HCA:
711 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
712 		break;
713 
714 	case MLX5_VPORT_ACCESS_METHOD_NIC:
715 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
716 		break;
717 
718 	default:
719 		return -EINVAL;
720 	}
721 
722 	if (!err)
723 		*sys_image_guid = cpu_to_be64(tmp);
724 
725 	return err;
726 
727 }
728 
729 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
730 				u16 *max_pkeys)
731 {
732 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
733 	struct mlx5_core_dev *mdev = dev->mdev;
734 
735 	switch (mlx5_get_vport_access_method(ibdev)) {
736 	case MLX5_VPORT_ACCESS_METHOD_MAD:
737 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
738 
739 	case MLX5_VPORT_ACCESS_METHOD_HCA:
740 	case MLX5_VPORT_ACCESS_METHOD_NIC:
741 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
742 						pkey_table_size));
743 		return 0;
744 
745 	default:
746 		return -EINVAL;
747 	}
748 }
749 
750 static int mlx5_query_vendor_id(struct ib_device *ibdev,
751 				u32 *vendor_id)
752 {
753 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
754 
755 	switch (mlx5_get_vport_access_method(ibdev)) {
756 	case MLX5_VPORT_ACCESS_METHOD_MAD:
757 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
758 
759 	case MLX5_VPORT_ACCESS_METHOD_HCA:
760 	case MLX5_VPORT_ACCESS_METHOD_NIC:
761 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
762 
763 	default:
764 		return -EINVAL;
765 	}
766 }
767 
768 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
769 				__be64 *node_guid)
770 {
771 	u64 tmp;
772 	int err;
773 
774 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
775 	case MLX5_VPORT_ACCESS_METHOD_MAD:
776 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
777 
778 	case MLX5_VPORT_ACCESS_METHOD_HCA:
779 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
780 		break;
781 
782 	case MLX5_VPORT_ACCESS_METHOD_NIC:
783 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
784 		break;
785 
786 	default:
787 		return -EINVAL;
788 	}
789 
790 	if (!err)
791 		*node_guid = cpu_to_be64(tmp);
792 
793 	return err;
794 }
795 
796 struct mlx5_reg_node_desc {
797 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
798 };
799 
800 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
801 {
802 	struct mlx5_reg_node_desc in;
803 
804 	if (mlx5_use_mad_ifc(dev))
805 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
806 
807 	memset(&in, 0, sizeof(in));
808 
809 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
810 				    sizeof(struct mlx5_reg_node_desc),
811 				    MLX5_REG_NODE_DESC, 0, 0);
812 }
813 
814 static int mlx5_ib_query_device(struct ib_device *ibdev,
815 				struct ib_device_attr *props,
816 				struct ib_udata *uhw)
817 {
818 	size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
819 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
820 	struct mlx5_core_dev *mdev = dev->mdev;
821 	int err = -ENOMEM;
822 	int max_sq_desc;
823 	int max_rq_sg;
824 	int max_sq_sg;
825 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
826 	bool raw_support = !mlx5_core_mp_enabled(mdev);
827 	struct mlx5_ib_query_device_resp resp = {};
828 	size_t resp_len;
829 	u64 max_tso;
830 
831 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
832 	if (uhw_outlen && uhw_outlen < resp_len)
833 		return -EINVAL;
834 
835 	resp.response_length = resp_len;
836 
837 	if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
838 		return -EINVAL;
839 
840 	memset(props, 0, sizeof(*props));
841 	err = mlx5_query_system_image_guid(ibdev,
842 					   &props->sys_image_guid);
843 	if (err)
844 		return err;
845 
846 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
847 	if (err)
848 		return err;
849 
850 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
851 	if (err)
852 		return err;
853 
854 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
855 		(fw_rev_min(dev->mdev) << 16) |
856 		fw_rev_sub(dev->mdev);
857 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
858 		IB_DEVICE_PORT_ACTIVE_EVENT		|
859 		IB_DEVICE_SYS_IMAGE_GUID		|
860 		IB_DEVICE_RC_RNR_NAK_GEN;
861 
862 	if (MLX5_CAP_GEN(mdev, pkv))
863 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
864 	if (MLX5_CAP_GEN(mdev, qkv))
865 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
866 	if (MLX5_CAP_GEN(mdev, apm))
867 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
868 	if (MLX5_CAP_GEN(mdev, xrc))
869 		props->device_cap_flags |= IB_DEVICE_XRC;
870 	if (MLX5_CAP_GEN(mdev, imaicl)) {
871 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
872 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
873 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
874 		/* We support 'Gappy' memory registration too */
875 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
876 	}
877 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
878 	if (MLX5_CAP_GEN(mdev, sho)) {
879 		props->device_cap_flags |= IB_DEVICE_INTEGRITY_HANDOVER;
880 		/* At this stage no support for signature handover */
881 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
882 				      IB_PROT_T10DIF_TYPE_2 |
883 				      IB_PROT_T10DIF_TYPE_3;
884 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
885 				       IB_GUARD_T10DIF_CSUM;
886 	}
887 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
888 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
889 
890 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
891 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
892 			/* Legacy bit to support old userspace libraries */
893 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
894 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
895 		}
896 
897 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
898 			props->raw_packet_caps |=
899 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
900 
901 		if (field_avail(typeof(resp), tso_caps, uhw_outlen)) {
902 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
903 			if (max_tso) {
904 				resp.tso_caps.max_tso = 1 << max_tso;
905 				resp.tso_caps.supported_qpts |=
906 					1 << IB_QPT_RAW_PACKET;
907 				resp.response_length += sizeof(resp.tso_caps);
908 			}
909 		}
910 
911 		if (field_avail(typeof(resp), rss_caps, uhw_outlen)) {
912 			resp.rss_caps.rx_hash_function =
913 						MLX5_RX_HASH_FUNC_TOEPLITZ;
914 			resp.rss_caps.rx_hash_fields_mask =
915 						MLX5_RX_HASH_SRC_IPV4 |
916 						MLX5_RX_HASH_DST_IPV4 |
917 						MLX5_RX_HASH_SRC_IPV6 |
918 						MLX5_RX_HASH_DST_IPV6 |
919 						MLX5_RX_HASH_SRC_PORT_TCP |
920 						MLX5_RX_HASH_DST_PORT_TCP |
921 						MLX5_RX_HASH_SRC_PORT_UDP |
922 						MLX5_RX_HASH_DST_PORT_UDP |
923 						MLX5_RX_HASH_INNER;
924 			if (mlx5_accel_ipsec_device_caps(dev->mdev) &
925 			    MLX5_ACCEL_IPSEC_CAP_DEVICE)
926 				resp.rss_caps.rx_hash_fields_mask |=
927 					MLX5_RX_HASH_IPSEC_SPI;
928 			resp.response_length += sizeof(resp.rss_caps);
929 		}
930 	} else {
931 		if (field_avail(typeof(resp), tso_caps, uhw_outlen))
932 			resp.response_length += sizeof(resp.tso_caps);
933 		if (field_avail(typeof(resp), rss_caps, uhw_outlen))
934 			resp.response_length += sizeof(resp.rss_caps);
935 	}
936 
937 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
938 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
939 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
940 	}
941 
942 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
943 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
944 	    raw_support)
945 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
946 
947 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
948 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
949 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
950 
951 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
952 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
953 	    raw_support) {
954 		/* Legacy bit to support old userspace libraries */
955 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
956 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
957 	}
958 
959 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
960 		props->max_dm_size =
961 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
962 	}
963 
964 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
965 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
966 
967 	if (MLX5_CAP_GEN(mdev, end_pad))
968 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
969 
970 	props->vendor_part_id	   = mdev->pdev->device;
971 	props->hw_ver		   = mdev->pdev->revision;
972 
973 	props->max_mr_size	   = ~0ull;
974 	props->page_size_cap	   = ~(min_page_size - 1);
975 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
976 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
977 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
978 		     sizeof(struct mlx5_wqe_data_seg);
979 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
980 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
981 		     sizeof(struct mlx5_wqe_raddr_seg)) /
982 		sizeof(struct mlx5_wqe_data_seg);
983 	props->max_send_sge = max_sq_sg;
984 	props->max_recv_sge = max_rq_sg;
985 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
986 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
987 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
988 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
989 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
990 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
991 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
992 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
993 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
994 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
995 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
996 	props->max_srq_sge	   = max_rq_sg - 1;
997 	props->max_fast_reg_page_list_len =
998 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
999 	props->max_pi_fast_reg_page_list_len =
1000 		props->max_fast_reg_page_list_len / 2;
1001 	props->max_sgl_rd =
1002 		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
1003 	get_atomic_caps_qp(dev, props);
1004 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
1005 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
1006 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
1007 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1008 					   props->max_mcast_grp;
1009 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
1010 	props->max_ah = INT_MAX;
1011 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
1012 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
1013 
1014 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
1015 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
1016 			props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
1017 		props->odp_caps = dev->odp_caps;
1018 		if (!uhw) {
1019 			/* ODP for kernel QPs is not implemented for receive
1020 			 * WQEs and SRQ WQEs
1021 			 */
1022 			props->odp_caps.per_transport_caps.rc_odp_caps &=
1023 				~(IB_ODP_SUPPORT_READ |
1024 				  IB_ODP_SUPPORT_SRQ_RECV);
1025 			props->odp_caps.per_transport_caps.uc_odp_caps &=
1026 				~(IB_ODP_SUPPORT_READ |
1027 				  IB_ODP_SUPPORT_SRQ_RECV);
1028 			props->odp_caps.per_transport_caps.ud_odp_caps &=
1029 				~(IB_ODP_SUPPORT_READ |
1030 				  IB_ODP_SUPPORT_SRQ_RECV);
1031 			props->odp_caps.per_transport_caps.xrc_odp_caps &=
1032 				~(IB_ODP_SUPPORT_READ |
1033 				  IB_ODP_SUPPORT_SRQ_RECV);
1034 		}
1035 	}
1036 
1037 	if (MLX5_CAP_GEN(mdev, cd))
1038 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
1039 
1040 	if (mlx5_core_is_vf(mdev))
1041 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
1042 
1043 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1044 	    IB_LINK_LAYER_ETHERNET && raw_support) {
1045 		props->rss_caps.max_rwq_indirection_tables =
1046 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1047 		props->rss_caps.max_rwq_indirection_table_size =
1048 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1049 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1050 		props->max_wq_type_rq =
1051 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1052 	}
1053 
1054 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1055 		props->tm_caps.max_num_tags =
1056 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1057 		props->tm_caps.max_ops =
1058 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1059 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1060 	}
1061 
1062 	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1063 	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1064 		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1065 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1066 	}
1067 
1068 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1069 		props->cq_caps.max_cq_moderation_count =
1070 						MLX5_MAX_CQ_COUNT;
1071 		props->cq_caps.max_cq_moderation_period =
1072 						MLX5_MAX_CQ_PERIOD;
1073 	}
1074 
1075 	if (field_avail(typeof(resp), cqe_comp_caps, uhw_outlen)) {
1076 		resp.response_length += sizeof(resp.cqe_comp_caps);
1077 
1078 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1079 			resp.cqe_comp_caps.max_num =
1080 				MLX5_CAP_GEN(dev->mdev,
1081 					     cqe_compression_max_num);
1082 
1083 			resp.cqe_comp_caps.supported_format =
1084 				MLX5_IB_CQE_RES_FORMAT_HASH |
1085 				MLX5_IB_CQE_RES_FORMAT_CSUM;
1086 
1087 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1088 				resp.cqe_comp_caps.supported_format |=
1089 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1090 		}
1091 	}
1092 
1093 	if (field_avail(typeof(resp), packet_pacing_caps, uhw_outlen) &&
1094 	    raw_support) {
1095 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1096 		    MLX5_CAP_GEN(mdev, qos)) {
1097 			resp.packet_pacing_caps.qp_rate_limit_max =
1098 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1099 			resp.packet_pacing_caps.qp_rate_limit_min =
1100 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1101 			resp.packet_pacing_caps.supported_qpts |=
1102 				1 << IB_QPT_RAW_PACKET;
1103 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1104 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1105 				resp.packet_pacing_caps.cap_flags |=
1106 					MLX5_IB_PP_SUPPORT_BURST;
1107 		}
1108 		resp.response_length += sizeof(resp.packet_pacing_caps);
1109 	}
1110 
1111 	if (field_avail(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes,
1112 			uhw_outlen)) {
1113 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1114 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1115 				MLX5_IB_ALLOW_MPW;
1116 
1117 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1118 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1119 				MLX5_IB_SUPPORT_EMPW;
1120 
1121 		resp.response_length +=
1122 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1123 	}
1124 
1125 	if (field_avail(typeof(resp), flags, uhw_outlen)) {
1126 		resp.response_length += sizeof(resp.flags);
1127 
1128 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1129 			resp.flags |=
1130 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1131 
1132 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1133 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1134 		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1135 			resp.flags |=
1136 				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1137 
1138 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1139 	}
1140 
1141 	if (field_avail(typeof(resp), sw_parsing_caps, uhw_outlen)) {
1142 		resp.response_length += sizeof(resp.sw_parsing_caps);
1143 		if (MLX5_CAP_ETH(mdev, swp)) {
1144 			resp.sw_parsing_caps.sw_parsing_offloads |=
1145 				MLX5_IB_SW_PARSING;
1146 
1147 			if (MLX5_CAP_ETH(mdev, swp_csum))
1148 				resp.sw_parsing_caps.sw_parsing_offloads |=
1149 					MLX5_IB_SW_PARSING_CSUM;
1150 
1151 			if (MLX5_CAP_ETH(mdev, swp_lso))
1152 				resp.sw_parsing_caps.sw_parsing_offloads |=
1153 					MLX5_IB_SW_PARSING_LSO;
1154 
1155 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1156 				resp.sw_parsing_caps.supported_qpts =
1157 					BIT(IB_QPT_RAW_PACKET);
1158 		}
1159 	}
1160 
1161 	if (field_avail(typeof(resp), striding_rq_caps, uhw_outlen) &&
1162 	    raw_support) {
1163 		resp.response_length += sizeof(resp.striding_rq_caps);
1164 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1165 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1166 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1167 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1168 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1169 			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1170 				resp.striding_rq_caps
1171 					.min_single_wqe_log_num_of_strides =
1172 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1173 			else
1174 				resp.striding_rq_caps
1175 					.min_single_wqe_log_num_of_strides =
1176 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1177 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1178 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1179 			resp.striding_rq_caps.supported_qpts =
1180 				BIT(IB_QPT_RAW_PACKET);
1181 		}
1182 	}
1183 
1184 	if (field_avail(typeof(resp), tunnel_offloads_caps, uhw_outlen)) {
1185 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1186 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1187 			resp.tunnel_offloads_caps |=
1188 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1189 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1190 			resp.tunnel_offloads_caps |=
1191 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1192 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1193 			resp.tunnel_offloads_caps |=
1194 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1195 		if (MLX5_CAP_GEN(mdev, flex_parser_protocols) &
1196 		    MLX5_FLEX_PROTO_CW_MPLS_GRE)
1197 			resp.tunnel_offloads_caps |=
1198 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1199 		if (MLX5_CAP_GEN(mdev, flex_parser_protocols) &
1200 		    MLX5_FLEX_PROTO_CW_MPLS_UDP)
1201 			resp.tunnel_offloads_caps |=
1202 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1203 	}
1204 
1205 	if (uhw_outlen) {
1206 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1207 
1208 		if (err)
1209 			return err;
1210 	}
1211 
1212 	return 0;
1213 }
1214 
1215 enum mlx5_ib_width {
1216 	MLX5_IB_WIDTH_1X	= 1 << 0,
1217 	MLX5_IB_WIDTH_2X	= 1 << 1,
1218 	MLX5_IB_WIDTH_4X	= 1 << 2,
1219 	MLX5_IB_WIDTH_8X	= 1 << 3,
1220 	MLX5_IB_WIDTH_12X	= 1 << 4
1221 };
1222 
1223 static void translate_active_width(struct ib_device *ibdev, u8 active_width,
1224 				  u8 *ib_width)
1225 {
1226 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1227 
1228 	if (active_width & MLX5_IB_WIDTH_1X)
1229 		*ib_width = IB_WIDTH_1X;
1230 	else if (active_width & MLX5_IB_WIDTH_2X)
1231 		*ib_width = IB_WIDTH_2X;
1232 	else if (active_width & MLX5_IB_WIDTH_4X)
1233 		*ib_width = IB_WIDTH_4X;
1234 	else if (active_width & MLX5_IB_WIDTH_8X)
1235 		*ib_width = IB_WIDTH_8X;
1236 	else if (active_width & MLX5_IB_WIDTH_12X)
1237 		*ib_width = IB_WIDTH_12X;
1238 	else {
1239 		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1240 			    (int)active_width);
1241 		*ib_width = IB_WIDTH_4X;
1242 	}
1243 
1244 	return;
1245 }
1246 
1247 static int mlx5_mtu_to_ib_mtu(int mtu)
1248 {
1249 	switch (mtu) {
1250 	case 256: return 1;
1251 	case 512: return 2;
1252 	case 1024: return 3;
1253 	case 2048: return 4;
1254 	case 4096: return 5;
1255 	default:
1256 		pr_warn("invalid mtu\n");
1257 		return -1;
1258 	}
1259 }
1260 
1261 enum ib_max_vl_num {
1262 	__IB_MAX_VL_0		= 1,
1263 	__IB_MAX_VL_0_1		= 2,
1264 	__IB_MAX_VL_0_3		= 3,
1265 	__IB_MAX_VL_0_7		= 4,
1266 	__IB_MAX_VL_0_14	= 5,
1267 };
1268 
1269 enum mlx5_vl_hw_cap {
1270 	MLX5_VL_HW_0	= 1,
1271 	MLX5_VL_HW_0_1	= 2,
1272 	MLX5_VL_HW_0_2	= 3,
1273 	MLX5_VL_HW_0_3	= 4,
1274 	MLX5_VL_HW_0_4	= 5,
1275 	MLX5_VL_HW_0_5	= 6,
1276 	MLX5_VL_HW_0_6	= 7,
1277 	MLX5_VL_HW_0_7	= 8,
1278 	MLX5_VL_HW_0_14	= 15
1279 };
1280 
1281 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1282 				u8 *max_vl_num)
1283 {
1284 	switch (vl_hw_cap) {
1285 	case MLX5_VL_HW_0:
1286 		*max_vl_num = __IB_MAX_VL_0;
1287 		break;
1288 	case MLX5_VL_HW_0_1:
1289 		*max_vl_num = __IB_MAX_VL_0_1;
1290 		break;
1291 	case MLX5_VL_HW_0_3:
1292 		*max_vl_num = __IB_MAX_VL_0_3;
1293 		break;
1294 	case MLX5_VL_HW_0_7:
1295 		*max_vl_num = __IB_MAX_VL_0_7;
1296 		break;
1297 	case MLX5_VL_HW_0_14:
1298 		*max_vl_num = __IB_MAX_VL_0_14;
1299 		break;
1300 
1301 	default:
1302 		return -EINVAL;
1303 	}
1304 
1305 	return 0;
1306 }
1307 
1308 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
1309 			       struct ib_port_attr *props)
1310 {
1311 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1312 	struct mlx5_core_dev *mdev = dev->mdev;
1313 	struct mlx5_hca_vport_context *rep;
1314 	u16 max_mtu;
1315 	u16 oper_mtu;
1316 	int err;
1317 	u8 ib_link_width_oper;
1318 	u8 vl_hw_cap;
1319 
1320 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1321 	if (!rep) {
1322 		err = -ENOMEM;
1323 		goto out;
1324 	}
1325 
1326 	/* props being zeroed by the caller, avoid zeroing it here */
1327 
1328 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1329 	if (err)
1330 		goto out;
1331 
1332 	props->lid		= rep->lid;
1333 	props->lmc		= rep->lmc;
1334 	props->sm_lid		= rep->sm_lid;
1335 	props->sm_sl		= rep->sm_sl;
1336 	props->state		= rep->vport_state;
1337 	props->phys_state	= rep->port_physical_state;
1338 	props->port_cap_flags	= rep->cap_mask1;
1339 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1340 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1341 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1342 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1343 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1344 	props->subnet_timeout	= rep->subnet_timeout;
1345 	props->init_type_reply	= rep->init_type_reply;
1346 
1347 	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1348 		props->port_cap_flags2 = rep->cap_mask2;
1349 
1350 	err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
1351 	if (err)
1352 		goto out;
1353 
1354 	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1355 
1356 	err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
1357 	if (err)
1358 		goto out;
1359 
1360 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1361 
1362 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1363 
1364 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1365 
1366 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1367 
1368 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1369 	if (err)
1370 		goto out;
1371 
1372 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1373 				   &props->max_vl_num);
1374 out:
1375 	kfree(rep);
1376 	return err;
1377 }
1378 
1379 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
1380 		       struct ib_port_attr *props)
1381 {
1382 	unsigned int count;
1383 	int ret;
1384 
1385 	switch (mlx5_get_vport_access_method(ibdev)) {
1386 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1387 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1388 		break;
1389 
1390 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1391 		ret = mlx5_query_hca_port(ibdev, port, props);
1392 		break;
1393 
1394 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1395 		ret = mlx5_query_port_roce(ibdev, port, props);
1396 		break;
1397 
1398 	default:
1399 		ret = -EINVAL;
1400 	}
1401 
1402 	if (!ret && props) {
1403 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1404 		struct mlx5_core_dev *mdev;
1405 		bool put_mdev = true;
1406 
1407 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1408 		if (!mdev) {
1409 			/* If the port isn't affiliated yet query the master.
1410 			 * The master and slave will have the same values.
1411 			 */
1412 			mdev = dev->mdev;
1413 			port = 1;
1414 			put_mdev = false;
1415 		}
1416 		count = mlx5_core_reserved_gids_count(mdev);
1417 		if (put_mdev)
1418 			mlx5_ib_put_native_port_mdev(dev, port);
1419 		props->gid_tbl_len -= count;
1420 	}
1421 	return ret;
1422 }
1423 
1424 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u8 port,
1425 				  struct ib_port_attr *props)
1426 {
1427 	int ret;
1428 
1429 	/* Only link layer == ethernet is valid for representors
1430 	 * and we always use port 1
1431 	 */
1432 	ret = mlx5_query_port_roce(ibdev, port, props);
1433 	if (ret || !props)
1434 		return ret;
1435 
1436 	/* We don't support GIDS */
1437 	props->gid_tbl_len = 0;
1438 
1439 	return ret;
1440 }
1441 
1442 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1443 			     union ib_gid *gid)
1444 {
1445 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1446 	struct mlx5_core_dev *mdev = dev->mdev;
1447 
1448 	switch (mlx5_get_vport_access_method(ibdev)) {
1449 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1450 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1451 
1452 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1453 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1454 
1455 	default:
1456 		return -EINVAL;
1457 	}
1458 
1459 }
1460 
1461 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u8 port,
1462 				   u16 index, u16 *pkey)
1463 {
1464 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1465 	struct mlx5_core_dev *mdev;
1466 	bool put_mdev = true;
1467 	u8 mdev_port_num;
1468 	int err;
1469 
1470 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1471 	if (!mdev) {
1472 		/* The port isn't affiliated yet, get the PKey from the master
1473 		 * port. For RoCE the PKey tables will be the same.
1474 		 */
1475 		put_mdev = false;
1476 		mdev = dev->mdev;
1477 		mdev_port_num = 1;
1478 	}
1479 
1480 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1481 					index, pkey);
1482 	if (put_mdev)
1483 		mlx5_ib_put_native_port_mdev(dev, port);
1484 
1485 	return err;
1486 }
1487 
1488 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1489 			      u16 *pkey)
1490 {
1491 	switch (mlx5_get_vport_access_method(ibdev)) {
1492 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1493 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1494 
1495 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1496 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1497 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1498 	default:
1499 		return -EINVAL;
1500 	}
1501 }
1502 
1503 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1504 				 struct ib_device_modify *props)
1505 {
1506 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1507 	struct mlx5_reg_node_desc in;
1508 	struct mlx5_reg_node_desc out;
1509 	int err;
1510 
1511 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1512 		return -EOPNOTSUPP;
1513 
1514 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1515 		return 0;
1516 
1517 	/*
1518 	 * If possible, pass node desc to FW, so it can generate
1519 	 * a 144 trap.  If cmd fails, just ignore.
1520 	 */
1521 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1522 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1523 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1524 	if (err)
1525 		return err;
1526 
1527 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1528 
1529 	return err;
1530 }
1531 
1532 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1533 				u32 value)
1534 {
1535 	struct mlx5_hca_vport_context ctx = {};
1536 	struct mlx5_core_dev *mdev;
1537 	u8 mdev_port_num;
1538 	int err;
1539 
1540 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1541 	if (!mdev)
1542 		return -ENODEV;
1543 
1544 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1545 	if (err)
1546 		goto out;
1547 
1548 	if (~ctx.cap_mask1_perm & mask) {
1549 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1550 			     mask, ctx.cap_mask1_perm);
1551 		err = -EINVAL;
1552 		goto out;
1553 	}
1554 
1555 	ctx.cap_mask1 = value;
1556 	ctx.cap_mask1_perm = mask;
1557 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1558 						 0, &ctx);
1559 
1560 out:
1561 	mlx5_ib_put_native_port_mdev(dev, port_num);
1562 
1563 	return err;
1564 }
1565 
1566 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1567 			       struct ib_port_modify *props)
1568 {
1569 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1570 	struct ib_port_attr attr;
1571 	u32 tmp;
1572 	int err;
1573 	u32 change_mask;
1574 	u32 value;
1575 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1576 		      IB_LINK_LAYER_INFINIBAND);
1577 
1578 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1579 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1580 	 */
1581 	if (!is_ib)
1582 		return 0;
1583 
1584 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1585 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1586 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1587 		return set_port_caps_atomic(dev, port, change_mask, value);
1588 	}
1589 
1590 	mutex_lock(&dev->cap_mask_mutex);
1591 
1592 	err = ib_query_port(ibdev, port, &attr);
1593 	if (err)
1594 		goto out;
1595 
1596 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1597 		~props->clr_port_cap_mask;
1598 
1599 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1600 
1601 out:
1602 	mutex_unlock(&dev->cap_mask_mutex);
1603 	return err;
1604 }
1605 
1606 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1607 {
1608 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1609 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1610 }
1611 
1612 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1613 {
1614 	/* Large page with non 4k uar support might limit the dynamic size */
1615 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1616 		return MLX5_MIN_DYN_BFREGS;
1617 
1618 	return MLX5_MAX_DYN_BFREGS;
1619 }
1620 
1621 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1622 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1623 			     struct mlx5_bfreg_info *bfregi)
1624 {
1625 	int uars_per_sys_page;
1626 	int bfregs_per_sys_page;
1627 	int ref_bfregs = req->total_num_bfregs;
1628 
1629 	if (req->total_num_bfregs == 0)
1630 		return -EINVAL;
1631 
1632 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1633 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1634 
1635 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1636 		return -ENOMEM;
1637 
1638 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1639 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1640 	/* This holds the required static allocation asked by the user */
1641 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1642 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1643 		return -EINVAL;
1644 
1645 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1646 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1647 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1648 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1649 
1650 	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",
1651 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1652 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1653 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1654 		    bfregi->num_sys_pages);
1655 
1656 	return 0;
1657 }
1658 
1659 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1660 {
1661 	struct mlx5_bfreg_info *bfregi;
1662 	int err;
1663 	int i;
1664 
1665 	bfregi = &context->bfregi;
1666 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1667 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1668 		if (err)
1669 			goto error;
1670 
1671 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1672 	}
1673 
1674 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1675 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1676 
1677 	return 0;
1678 
1679 error:
1680 	for (--i; i >= 0; i--)
1681 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1682 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1683 
1684 	return err;
1685 }
1686 
1687 static void deallocate_uars(struct mlx5_ib_dev *dev,
1688 			    struct mlx5_ib_ucontext *context)
1689 {
1690 	struct mlx5_bfreg_info *bfregi;
1691 	int i;
1692 
1693 	bfregi = &context->bfregi;
1694 	for (i = 0; i < bfregi->num_sys_pages; i++)
1695 		if (i < bfregi->num_static_sys_pages ||
1696 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1697 			mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1698 }
1699 
1700 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1701 {
1702 	int err = 0;
1703 
1704 	mutex_lock(&dev->lb.mutex);
1705 	if (td)
1706 		dev->lb.user_td++;
1707 	if (qp)
1708 		dev->lb.qps++;
1709 
1710 	if (dev->lb.user_td == 2 ||
1711 	    dev->lb.qps == 1) {
1712 		if (!dev->lb.enabled) {
1713 			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1714 			dev->lb.enabled = true;
1715 		}
1716 	}
1717 
1718 	mutex_unlock(&dev->lb.mutex);
1719 
1720 	return err;
1721 }
1722 
1723 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1724 {
1725 	mutex_lock(&dev->lb.mutex);
1726 	if (td)
1727 		dev->lb.user_td--;
1728 	if (qp)
1729 		dev->lb.qps--;
1730 
1731 	if (dev->lb.user_td == 1 &&
1732 	    dev->lb.qps == 0) {
1733 		if (dev->lb.enabled) {
1734 			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1735 			dev->lb.enabled = false;
1736 		}
1737 	}
1738 
1739 	mutex_unlock(&dev->lb.mutex);
1740 }
1741 
1742 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1743 					  u16 uid)
1744 {
1745 	int err;
1746 
1747 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1748 		return 0;
1749 
1750 	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1751 	if (err)
1752 		return err;
1753 
1754 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1755 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1756 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1757 		return err;
1758 
1759 	return mlx5_ib_enable_lb(dev, true, false);
1760 }
1761 
1762 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1763 					     u16 uid)
1764 {
1765 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1766 		return;
1767 
1768 	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1769 
1770 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1771 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1772 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1773 		return;
1774 
1775 	mlx5_ib_disable_lb(dev, true, false);
1776 }
1777 
1778 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1779 				  struct ib_udata *udata)
1780 {
1781 	struct ib_device *ibdev = uctx->device;
1782 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1783 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1784 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1785 	struct mlx5_core_dev *mdev = dev->mdev;
1786 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1787 	struct mlx5_bfreg_info *bfregi;
1788 	int ver;
1789 	int err;
1790 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1791 				     max_cqe_version);
1792 	u32 dump_fill_mkey;
1793 	bool lib_uar_4k;
1794 
1795 	if (!dev->ib_active)
1796 		return -EAGAIN;
1797 
1798 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1799 		ver = 0;
1800 	else if (udata->inlen >= min_req_v2)
1801 		ver = 2;
1802 	else
1803 		return -EINVAL;
1804 
1805 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1806 	if (err)
1807 		return err;
1808 
1809 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1810 		return -EOPNOTSUPP;
1811 
1812 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1813 		return -EOPNOTSUPP;
1814 
1815 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1816 				    MLX5_NON_FP_BFREGS_PER_UAR);
1817 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1818 		return -EINVAL;
1819 
1820 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1821 	if (dev->wc_support)
1822 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1823 	resp.cache_line_size = cache_line_size();
1824 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1825 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1826 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1827 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1828 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1829 	resp.cqe_version = min_t(__u8,
1830 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1831 				 req.max_cqe_version);
1832 	resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1833 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1834 	resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1835 					MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1;
1836 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1837 				   sizeof(resp.response_length), udata->outlen);
1838 
1839 	if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1840 		if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_EGRESS))
1841 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1842 		if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1843 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1844 		if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1845 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1846 		if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1847 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1848 		/* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1849 	}
1850 
1851 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1852 	bfregi = &context->bfregi;
1853 
1854 	/* updates req->total_num_bfregs */
1855 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1856 	if (err)
1857 		goto out_ctx;
1858 
1859 	mutex_init(&bfregi->lock);
1860 	bfregi->lib_uar_4k = lib_uar_4k;
1861 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1862 				GFP_KERNEL);
1863 	if (!bfregi->count) {
1864 		err = -ENOMEM;
1865 		goto out_ctx;
1866 	}
1867 
1868 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1869 				    sizeof(*bfregi->sys_pages),
1870 				    GFP_KERNEL);
1871 	if (!bfregi->sys_pages) {
1872 		err = -ENOMEM;
1873 		goto out_count;
1874 	}
1875 
1876 	err = allocate_uars(dev, context);
1877 	if (err)
1878 		goto out_sys_pages;
1879 
1880 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1881 		err = mlx5_ib_devx_create(dev, true);
1882 		if (err < 0)
1883 			goto out_uars;
1884 		context->devx_uid = err;
1885 	}
1886 
1887 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1888 					     context->devx_uid);
1889 	if (err)
1890 		goto out_devx;
1891 
1892 	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1893 		err = mlx5_cmd_dump_fill_mkey(dev->mdev, &dump_fill_mkey);
1894 		if (err)
1895 			goto out_mdev;
1896 	}
1897 
1898 	INIT_LIST_HEAD(&context->db_page_list);
1899 	mutex_init(&context->db_page_mutex);
1900 
1901 	resp.tot_bfregs = req.total_num_bfregs;
1902 	resp.num_ports = dev->num_ports;
1903 
1904 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1905 		resp.response_length += sizeof(resp.cqe_version);
1906 
1907 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1908 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1909 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1910 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1911 	}
1912 
1913 	if (field_avail(typeof(resp), eth_min_inline, udata->outlen)) {
1914 		if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1915 			mlx5_query_min_inline(dev->mdev, &resp.eth_min_inline);
1916 			resp.eth_min_inline++;
1917 		}
1918 		resp.response_length += sizeof(resp.eth_min_inline);
1919 	}
1920 
1921 	if (field_avail(typeof(resp), clock_info_versions, udata->outlen)) {
1922 		if (mdev->clock_info)
1923 			resp.clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1924 		resp.response_length += sizeof(resp.clock_info_versions);
1925 	}
1926 
1927 	/*
1928 	 * We don't want to expose information from the PCI bar that is located
1929 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1930 	 * pretend we don't support reading the HCA's core clock. This is also
1931 	 * forced by mmap function.
1932 	 */
1933 	if (field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1934 		if (PAGE_SIZE <= 4096) {
1935 			resp.comp_mask |=
1936 				MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1937 			resp.hca_core_clock_offset =
1938 				offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1939 		}
1940 		resp.response_length += sizeof(resp.hca_core_clock_offset);
1941 	}
1942 
1943 	if (field_avail(typeof(resp), log_uar_size, udata->outlen))
1944 		resp.response_length += sizeof(resp.log_uar_size);
1945 
1946 	if (field_avail(typeof(resp), num_uars_per_page, udata->outlen))
1947 		resp.response_length += sizeof(resp.num_uars_per_page);
1948 
1949 	if (field_avail(typeof(resp), num_dyn_bfregs, udata->outlen)) {
1950 		resp.num_dyn_bfregs = bfregi->num_dyn_bfregs;
1951 		resp.response_length += sizeof(resp.num_dyn_bfregs);
1952 	}
1953 
1954 	if (field_avail(typeof(resp), dump_fill_mkey, udata->outlen)) {
1955 		if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1956 			resp.dump_fill_mkey = dump_fill_mkey;
1957 			resp.comp_mask |=
1958 				MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1959 		}
1960 		resp.response_length += sizeof(resp.dump_fill_mkey);
1961 	}
1962 
1963 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1964 	if (err)
1965 		goto out_mdev;
1966 
1967 	bfregi->ver = ver;
1968 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1969 	context->cqe_version = resp.cqe_version;
1970 	context->lib_caps = req.lib_caps;
1971 	print_lib_caps(dev, context->lib_caps);
1972 
1973 	if (dev->lag_active) {
1974 		u8 port = mlx5_core_native_port_num(dev->mdev) - 1;
1975 
1976 		atomic_set(&context->tx_port_affinity,
1977 			   atomic_add_return(
1978 				   1, &dev->port[port].roce.tx_port_affinity));
1979 	}
1980 
1981 	return 0;
1982 
1983 out_mdev:
1984 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1985 out_devx:
1986 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1987 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1988 
1989 out_uars:
1990 	deallocate_uars(dev, context);
1991 
1992 out_sys_pages:
1993 	kfree(bfregi->sys_pages);
1994 
1995 out_count:
1996 	kfree(bfregi->count);
1997 
1998 out_ctx:
1999 	return err;
2000 }
2001 
2002 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
2003 {
2004 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2005 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2006 	struct mlx5_bfreg_info *bfregi;
2007 
2008 	bfregi = &context->bfregi;
2009 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
2010 
2011 	if (context->devx_uid)
2012 		mlx5_ib_devx_destroy(dev, context->devx_uid);
2013 
2014 	deallocate_uars(dev, context);
2015 	kfree(bfregi->sys_pages);
2016 	kfree(bfregi->count);
2017 }
2018 
2019 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2020 				 int uar_idx)
2021 {
2022 	int fw_uars_per_page;
2023 
2024 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2025 
2026 	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2027 }
2028 
2029 static int get_command(unsigned long offset)
2030 {
2031 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2032 }
2033 
2034 static int get_arg(unsigned long offset)
2035 {
2036 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2037 }
2038 
2039 static int get_index(unsigned long offset)
2040 {
2041 	return get_arg(offset);
2042 }
2043 
2044 /* Index resides in an extra byte to enable larger values than 255 */
2045 static int get_extended_index(unsigned long offset)
2046 {
2047 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2048 }
2049 
2050 
2051 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2052 {
2053 }
2054 
2055 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2056 {
2057 	switch (cmd) {
2058 	case MLX5_IB_MMAP_WC_PAGE:
2059 		return "WC";
2060 	case MLX5_IB_MMAP_REGULAR_PAGE:
2061 		return "best effort WC";
2062 	case MLX5_IB_MMAP_NC_PAGE:
2063 		return "NC";
2064 	case MLX5_IB_MMAP_DEVICE_MEM:
2065 		return "Device Memory";
2066 	default:
2067 		return NULL;
2068 	}
2069 }
2070 
2071 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2072 					struct vm_area_struct *vma,
2073 					struct mlx5_ib_ucontext *context)
2074 {
2075 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2076 	    !(vma->vm_flags & VM_SHARED))
2077 		return -EINVAL;
2078 
2079 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2080 		return -EOPNOTSUPP;
2081 
2082 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2083 		return -EPERM;
2084 	vma->vm_flags &= ~VM_MAYWRITE;
2085 
2086 	if (!dev->mdev->clock_info)
2087 		return -EOPNOTSUPP;
2088 
2089 	return vm_insert_page(vma, vma->vm_start,
2090 			      virt_to_page(dev->mdev->clock_info));
2091 }
2092 
2093 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2094 {
2095 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2096 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2097 	struct mlx5_var_table *var_table = &dev->var_table;
2098 	struct mlx5_ib_dm *mdm;
2099 
2100 	switch (mentry->mmap_flag) {
2101 	case MLX5_IB_MMAP_TYPE_MEMIC:
2102 		mdm = container_of(mentry, struct mlx5_ib_dm, mentry);
2103 		mlx5_cmd_dealloc_memic(&dev->dm, mdm->dev_addr,
2104 				       mdm->size);
2105 		kfree(mdm);
2106 		break;
2107 	case MLX5_IB_MMAP_TYPE_VAR:
2108 		mutex_lock(&var_table->bitmap_lock);
2109 		clear_bit(mentry->page_idx, var_table->bitmap);
2110 		mutex_unlock(&var_table->bitmap_lock);
2111 		kfree(mentry);
2112 		break;
2113 	default:
2114 		WARN_ON(true);
2115 	}
2116 }
2117 
2118 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2119 		    struct vm_area_struct *vma,
2120 		    struct mlx5_ib_ucontext *context)
2121 {
2122 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2123 	int err;
2124 	unsigned long idx;
2125 	phys_addr_t pfn;
2126 	pgprot_t prot;
2127 	u32 bfreg_dyn_idx = 0;
2128 	u32 uar_index;
2129 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2130 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2131 				bfregi->num_static_sys_pages;
2132 
2133 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2134 		return -EINVAL;
2135 
2136 	if (dyn_uar)
2137 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2138 	else
2139 		idx = get_index(vma->vm_pgoff);
2140 
2141 	if (idx >= max_valid_idx) {
2142 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2143 			     idx, max_valid_idx);
2144 		return -EINVAL;
2145 	}
2146 
2147 	switch (cmd) {
2148 	case MLX5_IB_MMAP_WC_PAGE:
2149 	case MLX5_IB_MMAP_ALLOC_WC:
2150 /* Some architectures don't support WC memory */
2151 #if defined(CONFIG_X86)
2152 		if (!pat_enabled())
2153 			return -EPERM;
2154 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
2155 			return -EPERM;
2156 #endif
2157 	/* fall through */
2158 	case MLX5_IB_MMAP_REGULAR_PAGE:
2159 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2160 		prot = pgprot_writecombine(vma->vm_page_prot);
2161 		break;
2162 	case MLX5_IB_MMAP_NC_PAGE:
2163 		prot = pgprot_noncached(vma->vm_page_prot);
2164 		break;
2165 	default:
2166 		return -EINVAL;
2167 	}
2168 
2169 	if (dyn_uar) {
2170 		int uars_per_page;
2171 
2172 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2173 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2174 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2175 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2176 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2177 			return -EINVAL;
2178 		}
2179 
2180 		mutex_lock(&bfregi->lock);
2181 		/* Fail if uar already allocated, first bfreg index of each
2182 		 * page holds its count.
2183 		 */
2184 		if (bfregi->count[bfreg_dyn_idx]) {
2185 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2186 			mutex_unlock(&bfregi->lock);
2187 			return -EINVAL;
2188 		}
2189 
2190 		bfregi->count[bfreg_dyn_idx]++;
2191 		mutex_unlock(&bfregi->lock);
2192 
2193 		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2194 		if (err) {
2195 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2196 			goto free_bfreg;
2197 		}
2198 	} else {
2199 		uar_index = bfregi->sys_pages[idx];
2200 	}
2201 
2202 	pfn = uar_index2pfn(dev, uar_index);
2203 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2204 
2205 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2206 				prot, NULL);
2207 	if (err) {
2208 		mlx5_ib_err(dev,
2209 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2210 			    err, mmap_cmd2str(cmd));
2211 		goto err;
2212 	}
2213 
2214 	if (dyn_uar)
2215 		bfregi->sys_pages[idx] = uar_index;
2216 	return 0;
2217 
2218 err:
2219 	if (!dyn_uar)
2220 		return err;
2221 
2222 	mlx5_cmd_free_uar(dev->mdev, idx);
2223 
2224 free_bfreg:
2225 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2226 
2227 	return err;
2228 }
2229 
2230 static int add_dm_mmap_entry(struct ib_ucontext *context,
2231 			     struct mlx5_ib_dm *mdm,
2232 			     u64 address)
2233 {
2234 	mdm->mentry.mmap_flag = MLX5_IB_MMAP_TYPE_MEMIC;
2235 	mdm->mentry.address = address;
2236 	return rdma_user_mmap_entry_insert_range(
2237 			context, &mdm->mentry.rdma_entry,
2238 			mdm->size,
2239 			MLX5_IB_MMAP_DEVICE_MEM << 16,
2240 			(MLX5_IB_MMAP_DEVICE_MEM << 16) + (1UL << 16) - 1);
2241 }
2242 
2243 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2244 {
2245 	unsigned long idx;
2246 	u8 command;
2247 
2248 	command = get_command(vma->vm_pgoff);
2249 	idx = get_extended_index(vma->vm_pgoff);
2250 
2251 	return (command << 16 | idx);
2252 }
2253 
2254 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2255 			       struct vm_area_struct *vma,
2256 			       struct ib_ucontext *ucontext)
2257 {
2258 	struct mlx5_user_mmap_entry *mentry;
2259 	struct rdma_user_mmap_entry *entry;
2260 	unsigned long pgoff;
2261 	pgprot_t prot;
2262 	phys_addr_t pfn;
2263 	int ret;
2264 
2265 	pgoff = mlx5_vma_to_pgoff(vma);
2266 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2267 	if (!entry)
2268 		return -EINVAL;
2269 
2270 	mentry = to_mmmap(entry);
2271 	pfn = (mentry->address >> PAGE_SHIFT);
2272 	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR)
2273 		prot = pgprot_noncached(vma->vm_page_prot);
2274 	else
2275 		prot = pgprot_writecombine(vma->vm_page_prot);
2276 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2277 				entry->npages * PAGE_SIZE,
2278 				prot,
2279 				entry);
2280 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2281 	return ret;
2282 }
2283 
2284 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2285 {
2286 	u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2287 	u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2288 
2289 	return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2290 		(index & 0xFF)) << PAGE_SHIFT;
2291 }
2292 
2293 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2294 {
2295 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2296 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2297 	unsigned long command;
2298 	phys_addr_t pfn;
2299 
2300 	command = get_command(vma->vm_pgoff);
2301 	switch (command) {
2302 	case MLX5_IB_MMAP_WC_PAGE:
2303 	case MLX5_IB_MMAP_NC_PAGE:
2304 	case MLX5_IB_MMAP_REGULAR_PAGE:
2305 	case MLX5_IB_MMAP_ALLOC_WC:
2306 		return uar_mmap(dev, command, vma, context);
2307 
2308 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2309 		return -ENOSYS;
2310 
2311 	case MLX5_IB_MMAP_CORE_CLOCK:
2312 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2313 			return -EINVAL;
2314 
2315 		if (vma->vm_flags & VM_WRITE)
2316 			return -EPERM;
2317 		vma->vm_flags &= ~VM_MAYWRITE;
2318 
2319 		/* Don't expose to user-space information it shouldn't have */
2320 		if (PAGE_SIZE > 4096)
2321 			return -EOPNOTSUPP;
2322 
2323 		pfn = (dev->mdev->iseg_base +
2324 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2325 			PAGE_SHIFT;
2326 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2327 					 PAGE_SIZE,
2328 					 pgprot_noncached(vma->vm_page_prot),
2329 					 NULL);
2330 	case MLX5_IB_MMAP_CLOCK_INFO:
2331 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2332 
2333 	default:
2334 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2335 	}
2336 
2337 	return 0;
2338 }
2339 
2340 static inline int check_dm_type_support(struct mlx5_ib_dev *dev,
2341 					u32 type)
2342 {
2343 	switch (type) {
2344 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2345 		if (!MLX5_CAP_DEV_MEM(dev->mdev, memic))
2346 			return -EOPNOTSUPP;
2347 		break;
2348 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2349 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2350 		if (!capable(CAP_SYS_RAWIO) ||
2351 		    !capable(CAP_NET_RAW))
2352 			return -EPERM;
2353 
2354 		if (!(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
2355 		      MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, sw_owner)))
2356 			return -EOPNOTSUPP;
2357 		break;
2358 	}
2359 
2360 	return 0;
2361 }
2362 
2363 static int handle_alloc_dm_memic(struct ib_ucontext *ctx,
2364 				 struct mlx5_ib_dm *dm,
2365 				 struct ib_dm_alloc_attr *attr,
2366 				 struct uverbs_attr_bundle *attrs)
2367 {
2368 	struct mlx5_dm *dm_db = &to_mdev(ctx->device)->dm;
2369 	u64 start_offset;
2370 	u16 page_idx;
2371 	int err;
2372 	u64 address;
2373 
2374 	dm->size = roundup(attr->length, MLX5_MEMIC_BASE_SIZE);
2375 
2376 	err = mlx5_cmd_alloc_memic(dm_db, &dm->dev_addr,
2377 				   dm->size, attr->alignment);
2378 	if (err)
2379 		return err;
2380 
2381 	address = dm->dev_addr & PAGE_MASK;
2382 	err = add_dm_mmap_entry(ctx, dm, address);
2383 	if (err)
2384 		goto err_dealloc;
2385 
2386 	page_idx = dm->mentry.rdma_entry.start_pgoff & 0xFFFF;
2387 	err = uverbs_copy_to(attrs,
2388 			     MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
2389 			     &page_idx,
2390 			     sizeof(page_idx));
2391 	if (err)
2392 		goto err_copy;
2393 
2394 	start_offset = dm->dev_addr & ~PAGE_MASK;
2395 	err = uverbs_copy_to(attrs,
2396 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2397 			     &start_offset, sizeof(start_offset));
2398 	if (err)
2399 		goto err_copy;
2400 
2401 	return 0;
2402 
2403 err_copy:
2404 	rdma_user_mmap_entry_remove(&dm->mentry.rdma_entry);
2405 err_dealloc:
2406 	mlx5_cmd_dealloc_memic(dm_db, dm->dev_addr, dm->size);
2407 
2408 	return err;
2409 }
2410 
2411 static int handle_alloc_dm_sw_icm(struct ib_ucontext *ctx,
2412 				  struct mlx5_ib_dm *dm,
2413 				  struct ib_dm_alloc_attr *attr,
2414 				  struct uverbs_attr_bundle *attrs,
2415 				  int type)
2416 {
2417 	struct mlx5_core_dev *dev = to_mdev(ctx->device)->mdev;
2418 	u64 act_size;
2419 	int err;
2420 
2421 	/* Allocation size must a multiple of the basic block size
2422 	 * and a power of 2.
2423 	 */
2424 	act_size = round_up(attr->length, MLX5_SW_ICM_BLOCK_SIZE(dev));
2425 	act_size = roundup_pow_of_two(act_size);
2426 
2427 	dm->size = act_size;
2428 	err = mlx5_dm_sw_icm_alloc(dev, type, act_size,
2429 				   to_mucontext(ctx)->devx_uid, &dm->dev_addr,
2430 				   &dm->icm_dm.obj_id);
2431 	if (err)
2432 		return err;
2433 
2434 	err = uverbs_copy_to(attrs,
2435 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2436 			     &dm->dev_addr, sizeof(dm->dev_addr));
2437 	if (err)
2438 		mlx5_dm_sw_icm_dealloc(dev, type, dm->size,
2439 				       to_mucontext(ctx)->devx_uid, dm->dev_addr,
2440 				       dm->icm_dm.obj_id);
2441 
2442 	return err;
2443 }
2444 
2445 struct ib_dm *mlx5_ib_alloc_dm(struct ib_device *ibdev,
2446 			       struct ib_ucontext *context,
2447 			       struct ib_dm_alloc_attr *attr,
2448 			       struct uverbs_attr_bundle *attrs)
2449 {
2450 	struct mlx5_ib_dm *dm;
2451 	enum mlx5_ib_uapi_dm_type type;
2452 	int err;
2453 
2454 	err = uverbs_get_const_default(&type, attrs,
2455 				       MLX5_IB_ATTR_ALLOC_DM_REQ_TYPE,
2456 				       MLX5_IB_UAPI_DM_TYPE_MEMIC);
2457 	if (err)
2458 		return ERR_PTR(err);
2459 
2460 	mlx5_ib_dbg(to_mdev(ibdev), "alloc_dm req: dm_type=%d user_length=0x%llx log_alignment=%d\n",
2461 		    type, attr->length, attr->alignment);
2462 
2463 	err = check_dm_type_support(to_mdev(ibdev), type);
2464 	if (err)
2465 		return ERR_PTR(err);
2466 
2467 	dm = kzalloc(sizeof(*dm), GFP_KERNEL);
2468 	if (!dm)
2469 		return ERR_PTR(-ENOMEM);
2470 
2471 	dm->type = type;
2472 
2473 	switch (type) {
2474 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2475 		err = handle_alloc_dm_memic(context, dm,
2476 					    attr,
2477 					    attrs);
2478 		break;
2479 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2480 		err = handle_alloc_dm_sw_icm(context, dm,
2481 					     attr, attrs,
2482 					     MLX5_SW_ICM_TYPE_STEERING);
2483 		break;
2484 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2485 		err = handle_alloc_dm_sw_icm(context, dm,
2486 					     attr, attrs,
2487 					     MLX5_SW_ICM_TYPE_HEADER_MODIFY);
2488 		break;
2489 	default:
2490 		err = -EOPNOTSUPP;
2491 	}
2492 
2493 	if (err)
2494 		goto err_free;
2495 
2496 	return &dm->ibdm;
2497 
2498 err_free:
2499 	kfree(dm);
2500 	return ERR_PTR(err);
2501 }
2502 
2503 int mlx5_ib_dealloc_dm(struct ib_dm *ibdm, struct uverbs_attr_bundle *attrs)
2504 {
2505 	struct mlx5_ib_ucontext *ctx = rdma_udata_to_drv_context(
2506 		&attrs->driver_udata, struct mlx5_ib_ucontext, ibucontext);
2507 	struct mlx5_core_dev *dev = to_mdev(ibdm->device)->mdev;
2508 	struct mlx5_ib_dm *dm = to_mdm(ibdm);
2509 	int ret;
2510 
2511 	switch (dm->type) {
2512 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2513 		rdma_user_mmap_entry_remove(&dm->mentry.rdma_entry);
2514 		return 0;
2515 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2516 		ret = mlx5_dm_sw_icm_dealloc(dev, MLX5_SW_ICM_TYPE_STEERING,
2517 					     dm->size, ctx->devx_uid, dm->dev_addr,
2518 					     dm->icm_dm.obj_id);
2519 		if (ret)
2520 			return ret;
2521 		break;
2522 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2523 		ret = mlx5_dm_sw_icm_dealloc(dev, MLX5_SW_ICM_TYPE_HEADER_MODIFY,
2524 					     dm->size, ctx->devx_uid, dm->dev_addr,
2525 					     dm->icm_dm.obj_id);
2526 		if (ret)
2527 			return ret;
2528 		break;
2529 	default:
2530 		return -EOPNOTSUPP;
2531 	}
2532 
2533 	kfree(dm);
2534 
2535 	return 0;
2536 }
2537 
2538 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2539 {
2540 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2541 	struct ib_device *ibdev = ibpd->device;
2542 	struct mlx5_ib_alloc_pd_resp resp;
2543 	int err;
2544 	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2545 	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)]   = {};
2546 	u16 uid = 0;
2547 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2548 		udata, struct mlx5_ib_ucontext, ibucontext);
2549 
2550 	uid = context ? context->devx_uid : 0;
2551 	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2552 	MLX5_SET(alloc_pd_in, in, uid, uid);
2553 	err = mlx5_cmd_exec(to_mdev(ibdev)->mdev, in, sizeof(in),
2554 			    out, sizeof(out));
2555 	if (err)
2556 		return err;
2557 
2558 	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2559 	pd->uid = uid;
2560 	if (udata) {
2561 		resp.pdn = pd->pdn;
2562 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2563 			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2564 			return -EFAULT;
2565 		}
2566 	}
2567 
2568 	return 0;
2569 }
2570 
2571 static void mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2572 {
2573 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2574 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2575 
2576 	mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2577 }
2578 
2579 enum {
2580 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
2581 	MATCH_CRITERIA_ENABLE_MISC_BIT,
2582 	MATCH_CRITERIA_ENABLE_INNER_BIT,
2583 	MATCH_CRITERIA_ENABLE_MISC2_BIT
2584 };
2585 
2586 #define HEADER_IS_ZERO(match_criteria, headers)			           \
2587 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
2588 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
2589 
2590 static u8 get_match_criteria_enable(u32 *match_criteria)
2591 {
2592 	u8 match_criteria_enable;
2593 
2594 	match_criteria_enable =
2595 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
2596 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
2597 	match_criteria_enable |=
2598 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
2599 		MATCH_CRITERIA_ENABLE_MISC_BIT;
2600 	match_criteria_enable |=
2601 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
2602 		MATCH_CRITERIA_ENABLE_INNER_BIT;
2603 	match_criteria_enable |=
2604 		(!HEADER_IS_ZERO(match_criteria, misc_parameters_2)) <<
2605 		MATCH_CRITERIA_ENABLE_MISC2_BIT;
2606 
2607 	return match_criteria_enable;
2608 }
2609 
2610 static int set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
2611 {
2612 	u8 entry_mask;
2613 	u8 entry_val;
2614 	int err = 0;
2615 
2616 	if (!mask)
2617 		goto out;
2618 
2619 	entry_mask = MLX5_GET(fte_match_set_lyr_2_4, outer_c,
2620 			      ip_protocol);
2621 	entry_val = MLX5_GET(fte_match_set_lyr_2_4, outer_v,
2622 			     ip_protocol);
2623 	if (!entry_mask) {
2624 		MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
2625 		MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
2626 		goto out;
2627 	}
2628 	/* Don't override existing ip protocol */
2629 	if (mask != entry_mask || val != entry_val)
2630 		err = -EINVAL;
2631 out:
2632 	return err;
2633 }
2634 
2635 static void set_flow_label(void *misc_c, void *misc_v, u32 mask, u32 val,
2636 			   bool inner)
2637 {
2638 	if (inner) {
2639 		MLX5_SET(fte_match_set_misc,
2640 			 misc_c, inner_ipv6_flow_label, mask);
2641 		MLX5_SET(fte_match_set_misc,
2642 			 misc_v, inner_ipv6_flow_label, val);
2643 	} else {
2644 		MLX5_SET(fte_match_set_misc,
2645 			 misc_c, outer_ipv6_flow_label, mask);
2646 		MLX5_SET(fte_match_set_misc,
2647 			 misc_v, outer_ipv6_flow_label, val);
2648 	}
2649 }
2650 
2651 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
2652 {
2653 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
2654 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
2655 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
2656 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
2657 }
2658 
2659 static int check_mpls_supp_fields(u32 field_support, const __be32 *set_mask)
2660 {
2661 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_label) &&
2662 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_LABEL))
2663 		return -EOPNOTSUPP;
2664 
2665 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_exp) &&
2666 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_EXP))
2667 		return -EOPNOTSUPP;
2668 
2669 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_s_bos) &&
2670 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_S_BOS))
2671 		return -EOPNOTSUPP;
2672 
2673 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_ttl) &&
2674 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_TTL))
2675 		return -EOPNOTSUPP;
2676 
2677 	return 0;
2678 }
2679 
2680 #define LAST_ETH_FIELD vlan_tag
2681 #define LAST_IB_FIELD sl
2682 #define LAST_IPV4_FIELD tos
2683 #define LAST_IPV6_FIELD traffic_class
2684 #define LAST_TCP_UDP_FIELD src_port
2685 #define LAST_TUNNEL_FIELD tunnel_id
2686 #define LAST_FLOW_TAG_FIELD tag_id
2687 #define LAST_DROP_FIELD size
2688 #define LAST_COUNTERS_FIELD counters
2689 
2690 /* Field is the last supported field */
2691 #define FIELDS_NOT_SUPPORTED(filter, field)\
2692 	memchr_inv((void *)&filter.field  +\
2693 		   sizeof(filter.field), 0,\
2694 		   sizeof(filter) -\
2695 		   offsetof(typeof(filter), field) -\
2696 		   sizeof(filter.field))
2697 
2698 int parse_flow_flow_action(struct mlx5_ib_flow_action *maction,
2699 			   bool is_egress,
2700 			   struct mlx5_flow_act *action)
2701 {
2702 
2703 	switch (maction->ib_action.type) {
2704 	case IB_FLOW_ACTION_ESP:
2705 		if (action->action & (MLX5_FLOW_CONTEXT_ACTION_ENCRYPT |
2706 				      MLX5_FLOW_CONTEXT_ACTION_DECRYPT))
2707 			return -EINVAL;
2708 		/* Currently only AES_GCM keymat is supported by the driver */
2709 		action->esp_id = (uintptr_t)maction->esp_aes_gcm.ctx;
2710 		action->action |= is_egress ?
2711 			MLX5_FLOW_CONTEXT_ACTION_ENCRYPT :
2712 			MLX5_FLOW_CONTEXT_ACTION_DECRYPT;
2713 		return 0;
2714 	case IB_FLOW_ACTION_UNSPECIFIED:
2715 		if (maction->flow_action_raw.sub_type ==
2716 		    MLX5_IB_FLOW_ACTION_MODIFY_HEADER) {
2717 			if (action->action & MLX5_FLOW_CONTEXT_ACTION_MOD_HDR)
2718 				return -EINVAL;
2719 			action->action |= MLX5_FLOW_CONTEXT_ACTION_MOD_HDR;
2720 			action->modify_hdr =
2721 				maction->flow_action_raw.modify_hdr;
2722 			return 0;
2723 		}
2724 		if (maction->flow_action_raw.sub_type ==
2725 		    MLX5_IB_FLOW_ACTION_DECAP) {
2726 			if (action->action & MLX5_FLOW_CONTEXT_ACTION_DECAP)
2727 				return -EINVAL;
2728 			action->action |= MLX5_FLOW_CONTEXT_ACTION_DECAP;
2729 			return 0;
2730 		}
2731 		if (maction->flow_action_raw.sub_type ==
2732 		    MLX5_IB_FLOW_ACTION_PACKET_REFORMAT) {
2733 			if (action->action &
2734 			    MLX5_FLOW_CONTEXT_ACTION_PACKET_REFORMAT)
2735 				return -EINVAL;
2736 			action->action |=
2737 				MLX5_FLOW_CONTEXT_ACTION_PACKET_REFORMAT;
2738 			action->pkt_reformat =
2739 				maction->flow_action_raw.pkt_reformat;
2740 			return 0;
2741 		}
2742 		/* fall through */
2743 	default:
2744 		return -EOPNOTSUPP;
2745 	}
2746 }
2747 
2748 static int parse_flow_attr(struct mlx5_core_dev *mdev,
2749 			   struct mlx5_flow_spec *spec,
2750 			   const union ib_flow_spec *ib_spec,
2751 			   const struct ib_flow_attr *flow_attr,
2752 			   struct mlx5_flow_act *action, u32 prev_type)
2753 {
2754 	struct mlx5_flow_context *flow_context = &spec->flow_context;
2755 	u32 *match_c = spec->match_criteria;
2756 	u32 *match_v = spec->match_value;
2757 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
2758 					   misc_parameters);
2759 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
2760 					   misc_parameters);
2761 	void *misc_params2_c = MLX5_ADDR_OF(fte_match_param, match_c,
2762 					    misc_parameters_2);
2763 	void *misc_params2_v = MLX5_ADDR_OF(fte_match_param, match_v,
2764 					    misc_parameters_2);
2765 	void *headers_c;
2766 	void *headers_v;
2767 	int match_ipv;
2768 	int ret;
2769 
2770 	if (ib_spec->type & IB_FLOW_SPEC_INNER) {
2771 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2772 					 inner_headers);
2773 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2774 					 inner_headers);
2775 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2776 					ft_field_support.inner_ip_version);
2777 	} else {
2778 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2779 					 outer_headers);
2780 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2781 					 outer_headers);
2782 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2783 					ft_field_support.outer_ip_version);
2784 	}
2785 
2786 	switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
2787 	case IB_FLOW_SPEC_ETH:
2788 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
2789 			return -EOPNOTSUPP;
2790 
2791 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2792 					     dmac_47_16),
2793 				ib_spec->eth.mask.dst_mac);
2794 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2795 					     dmac_47_16),
2796 				ib_spec->eth.val.dst_mac);
2797 
2798 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2799 					     smac_47_16),
2800 				ib_spec->eth.mask.src_mac);
2801 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2802 					     smac_47_16),
2803 				ib_spec->eth.val.src_mac);
2804 
2805 		if (ib_spec->eth.mask.vlan_tag) {
2806 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2807 				 cvlan_tag, 1);
2808 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2809 				 cvlan_tag, 1);
2810 
2811 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2812 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
2813 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2814 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
2815 
2816 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2817 				 first_cfi,
2818 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
2819 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2820 				 first_cfi,
2821 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
2822 
2823 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2824 				 first_prio,
2825 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
2826 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2827 				 first_prio,
2828 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
2829 		}
2830 		MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2831 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
2832 		MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2833 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
2834 		break;
2835 	case IB_FLOW_SPEC_IPV4:
2836 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
2837 			return -EOPNOTSUPP;
2838 
2839 		if (match_ipv) {
2840 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2841 				 ip_version, 0xf);
2842 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2843 				 ip_version, MLX5_FS_IPV4_VERSION);
2844 		} else {
2845 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2846 				 ethertype, 0xffff);
2847 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2848 				 ethertype, ETH_P_IP);
2849 		}
2850 
2851 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2852 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
2853 		       &ib_spec->ipv4.mask.src_ip,
2854 		       sizeof(ib_spec->ipv4.mask.src_ip));
2855 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2856 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
2857 		       &ib_spec->ipv4.val.src_ip,
2858 		       sizeof(ib_spec->ipv4.val.src_ip));
2859 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2860 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2861 		       &ib_spec->ipv4.mask.dst_ip,
2862 		       sizeof(ib_spec->ipv4.mask.dst_ip));
2863 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2864 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2865 		       &ib_spec->ipv4.val.dst_ip,
2866 		       sizeof(ib_spec->ipv4.val.dst_ip));
2867 
2868 		set_tos(headers_c, headers_v,
2869 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
2870 
2871 		if (set_proto(headers_c, headers_v,
2872 			      ib_spec->ipv4.mask.proto,
2873 			      ib_spec->ipv4.val.proto))
2874 			return -EINVAL;
2875 		break;
2876 	case IB_FLOW_SPEC_IPV6:
2877 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
2878 			return -EOPNOTSUPP;
2879 
2880 		if (match_ipv) {
2881 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2882 				 ip_version, 0xf);
2883 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2884 				 ip_version, MLX5_FS_IPV6_VERSION);
2885 		} else {
2886 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2887 				 ethertype, 0xffff);
2888 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2889 				 ethertype, ETH_P_IPV6);
2890 		}
2891 
2892 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2893 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
2894 		       &ib_spec->ipv6.mask.src_ip,
2895 		       sizeof(ib_spec->ipv6.mask.src_ip));
2896 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2897 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
2898 		       &ib_spec->ipv6.val.src_ip,
2899 		       sizeof(ib_spec->ipv6.val.src_ip));
2900 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2901 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2902 		       &ib_spec->ipv6.mask.dst_ip,
2903 		       sizeof(ib_spec->ipv6.mask.dst_ip));
2904 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2905 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2906 		       &ib_spec->ipv6.val.dst_ip,
2907 		       sizeof(ib_spec->ipv6.val.dst_ip));
2908 
2909 		set_tos(headers_c, headers_v,
2910 			ib_spec->ipv6.mask.traffic_class,
2911 			ib_spec->ipv6.val.traffic_class);
2912 
2913 		if (set_proto(headers_c, headers_v,
2914 			      ib_spec->ipv6.mask.next_hdr,
2915 			      ib_spec->ipv6.val.next_hdr))
2916 			return -EINVAL;
2917 
2918 		set_flow_label(misc_params_c, misc_params_v,
2919 			       ntohl(ib_spec->ipv6.mask.flow_label),
2920 			       ntohl(ib_spec->ipv6.val.flow_label),
2921 			       ib_spec->type & IB_FLOW_SPEC_INNER);
2922 		break;
2923 	case IB_FLOW_SPEC_ESP:
2924 		if (ib_spec->esp.mask.seq)
2925 			return -EOPNOTSUPP;
2926 
2927 		MLX5_SET(fte_match_set_misc, misc_params_c, outer_esp_spi,
2928 			 ntohl(ib_spec->esp.mask.spi));
2929 		MLX5_SET(fte_match_set_misc, misc_params_v, outer_esp_spi,
2930 			 ntohl(ib_spec->esp.val.spi));
2931 		break;
2932 	case IB_FLOW_SPEC_TCP:
2933 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2934 					 LAST_TCP_UDP_FIELD))
2935 			return -EOPNOTSUPP;
2936 
2937 		if (set_proto(headers_c, headers_v, 0xff, IPPROTO_TCP))
2938 			return -EINVAL;
2939 
2940 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_sport,
2941 			 ntohs(ib_spec->tcp_udp.mask.src_port));
2942 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_sport,
2943 			 ntohs(ib_spec->tcp_udp.val.src_port));
2944 
2945 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_dport,
2946 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
2947 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_dport,
2948 			 ntohs(ib_spec->tcp_udp.val.dst_port));
2949 		break;
2950 	case IB_FLOW_SPEC_UDP:
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_UDP))
2956 			return -EINVAL;
2957 
2958 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_sport,
2959 			 ntohs(ib_spec->tcp_udp.mask.src_port));
2960 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_sport,
2961 			 ntohs(ib_spec->tcp_udp.val.src_port));
2962 
2963 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_dport,
2964 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
2965 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_dport,
2966 			 ntohs(ib_spec->tcp_udp.val.dst_port));
2967 		break;
2968 	case IB_FLOW_SPEC_GRE:
2969 		if (ib_spec->gre.mask.c_ks_res0_ver)
2970 			return -EOPNOTSUPP;
2971 
2972 		if (set_proto(headers_c, headers_v, 0xff, IPPROTO_GRE))
2973 			return -EINVAL;
2974 
2975 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2976 			 0xff);
2977 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2978 			 IPPROTO_GRE);
2979 
2980 		MLX5_SET(fte_match_set_misc, misc_params_c, gre_protocol,
2981 			 ntohs(ib_spec->gre.mask.protocol));
2982 		MLX5_SET(fte_match_set_misc, misc_params_v, gre_protocol,
2983 			 ntohs(ib_spec->gre.val.protocol));
2984 
2985 		memcpy(MLX5_ADDR_OF(fte_match_set_misc, misc_params_c,
2986 				    gre_key.nvgre.hi),
2987 		       &ib_spec->gre.mask.key,
2988 		       sizeof(ib_spec->gre.mask.key));
2989 		memcpy(MLX5_ADDR_OF(fte_match_set_misc, misc_params_v,
2990 				    gre_key.nvgre.hi),
2991 		       &ib_spec->gre.val.key,
2992 		       sizeof(ib_spec->gre.val.key));
2993 		break;
2994 	case IB_FLOW_SPEC_MPLS:
2995 		switch (prev_type) {
2996 		case IB_FLOW_SPEC_UDP:
2997 			if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2998 						   ft_field_support.outer_first_mpls_over_udp),
2999 						   &ib_spec->mpls.mask.tag))
3000 				return -EOPNOTSUPP;
3001 
3002 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
3003 					    outer_first_mpls_over_udp),
3004 			       &ib_spec->mpls.val.tag,
3005 			       sizeof(ib_spec->mpls.val.tag));
3006 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
3007 					    outer_first_mpls_over_udp),
3008 			       &ib_spec->mpls.mask.tag,
3009 			       sizeof(ib_spec->mpls.mask.tag));
3010 			break;
3011 		case IB_FLOW_SPEC_GRE:
3012 			if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
3013 						   ft_field_support.outer_first_mpls_over_gre),
3014 						   &ib_spec->mpls.mask.tag))
3015 				return -EOPNOTSUPP;
3016 
3017 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
3018 					    outer_first_mpls_over_gre),
3019 			       &ib_spec->mpls.val.tag,
3020 			       sizeof(ib_spec->mpls.val.tag));
3021 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
3022 					    outer_first_mpls_over_gre),
3023 			       &ib_spec->mpls.mask.tag,
3024 			       sizeof(ib_spec->mpls.mask.tag));
3025 			break;
3026 		default:
3027 			if (ib_spec->type & IB_FLOW_SPEC_INNER) {
3028 				if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
3029 							   ft_field_support.inner_first_mpls),
3030 							   &ib_spec->mpls.mask.tag))
3031 					return -EOPNOTSUPP;
3032 
3033 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
3034 						    inner_first_mpls),
3035 				       &ib_spec->mpls.val.tag,
3036 				       sizeof(ib_spec->mpls.val.tag));
3037 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
3038 						    inner_first_mpls),
3039 				       &ib_spec->mpls.mask.tag,
3040 				       sizeof(ib_spec->mpls.mask.tag));
3041 			} else {
3042 				if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
3043 							   ft_field_support.outer_first_mpls),
3044 							   &ib_spec->mpls.mask.tag))
3045 					return -EOPNOTSUPP;
3046 
3047 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
3048 						    outer_first_mpls),
3049 				       &ib_spec->mpls.val.tag,
3050 				       sizeof(ib_spec->mpls.val.tag));
3051 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
3052 						    outer_first_mpls),
3053 				       &ib_spec->mpls.mask.tag,
3054 				       sizeof(ib_spec->mpls.mask.tag));
3055 			}
3056 		}
3057 		break;
3058 	case IB_FLOW_SPEC_VXLAN_TUNNEL:
3059 		if (FIELDS_NOT_SUPPORTED(ib_spec->tunnel.mask,
3060 					 LAST_TUNNEL_FIELD))
3061 			return -EOPNOTSUPP;
3062 
3063 		MLX5_SET(fte_match_set_misc, misc_params_c, vxlan_vni,
3064 			 ntohl(ib_spec->tunnel.mask.tunnel_id));
3065 		MLX5_SET(fte_match_set_misc, misc_params_v, vxlan_vni,
3066 			 ntohl(ib_spec->tunnel.val.tunnel_id));
3067 		break;
3068 	case IB_FLOW_SPEC_ACTION_TAG:
3069 		if (FIELDS_NOT_SUPPORTED(ib_spec->flow_tag,
3070 					 LAST_FLOW_TAG_FIELD))
3071 			return -EOPNOTSUPP;
3072 		if (ib_spec->flow_tag.tag_id >= BIT(24))
3073 			return -EINVAL;
3074 
3075 		flow_context->flow_tag = ib_spec->flow_tag.tag_id;
3076 		flow_context->flags |= FLOW_CONTEXT_HAS_TAG;
3077 		break;
3078 	case IB_FLOW_SPEC_ACTION_DROP:
3079 		if (FIELDS_NOT_SUPPORTED(ib_spec->drop,
3080 					 LAST_DROP_FIELD))
3081 			return -EOPNOTSUPP;
3082 		action->action |= MLX5_FLOW_CONTEXT_ACTION_DROP;
3083 		break;
3084 	case IB_FLOW_SPEC_ACTION_HANDLE:
3085 		ret = parse_flow_flow_action(to_mflow_act(ib_spec->action.act),
3086 			flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS, action);
3087 		if (ret)
3088 			return ret;
3089 		break;
3090 	case IB_FLOW_SPEC_ACTION_COUNT:
3091 		if (FIELDS_NOT_SUPPORTED(ib_spec->flow_count,
3092 					 LAST_COUNTERS_FIELD))
3093 			return -EOPNOTSUPP;
3094 
3095 		/* for now support only one counters spec per flow */
3096 		if (action->action & MLX5_FLOW_CONTEXT_ACTION_COUNT)
3097 			return -EINVAL;
3098 
3099 		action->counters = ib_spec->flow_count.counters;
3100 		action->action |= MLX5_FLOW_CONTEXT_ACTION_COUNT;
3101 		break;
3102 	default:
3103 		return -EINVAL;
3104 	}
3105 
3106 	return 0;
3107 }
3108 
3109 /* If a flow could catch both multicast and unicast packets,
3110  * it won't fall into the multicast flow steering table and this rule
3111  * could steal other multicast packets.
3112  */
3113 static bool flow_is_multicast_only(const struct ib_flow_attr *ib_attr)
3114 {
3115 	union ib_flow_spec *flow_spec;
3116 
3117 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
3118 	    ib_attr->num_of_specs < 1)
3119 		return false;
3120 
3121 	flow_spec = (union ib_flow_spec *)(ib_attr + 1);
3122 	if (flow_spec->type == IB_FLOW_SPEC_IPV4) {
3123 		struct ib_flow_spec_ipv4 *ipv4_spec;
3124 
3125 		ipv4_spec = (struct ib_flow_spec_ipv4 *)flow_spec;
3126 		if (ipv4_is_multicast(ipv4_spec->val.dst_ip))
3127 			return true;
3128 
3129 		return false;
3130 	}
3131 
3132 	if (flow_spec->type == IB_FLOW_SPEC_ETH) {
3133 		struct ib_flow_spec_eth *eth_spec;
3134 
3135 		eth_spec = (struct ib_flow_spec_eth *)flow_spec;
3136 		return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
3137 		       is_multicast_ether_addr(eth_spec->val.dst_mac);
3138 	}
3139 
3140 	return false;
3141 }
3142 
3143 enum valid_spec {
3144 	VALID_SPEC_INVALID,
3145 	VALID_SPEC_VALID,
3146 	VALID_SPEC_NA,
3147 };
3148 
3149 static enum valid_spec
3150 is_valid_esp_aes_gcm(struct mlx5_core_dev *mdev,
3151 		     const struct mlx5_flow_spec *spec,
3152 		     const struct mlx5_flow_act *flow_act,
3153 		     bool egress)
3154 {
3155 	const u32 *match_c = spec->match_criteria;
3156 	bool is_crypto =
3157 		(flow_act->action & (MLX5_FLOW_CONTEXT_ACTION_ENCRYPT |
3158 				     MLX5_FLOW_CONTEXT_ACTION_DECRYPT));
3159 	bool is_ipsec = mlx5_fs_is_ipsec_flow(match_c);
3160 	bool is_drop = flow_act->action & MLX5_FLOW_CONTEXT_ACTION_DROP;
3161 
3162 	/*
3163 	 * Currently only crypto is supported in egress, when regular egress
3164 	 * rules would be supported, always return VALID_SPEC_NA.
3165 	 */
3166 	if (!is_crypto)
3167 		return VALID_SPEC_NA;
3168 
3169 	return is_crypto && is_ipsec &&
3170 		(!egress || (!is_drop &&
3171 			     !(spec->flow_context.flags & FLOW_CONTEXT_HAS_TAG))) ?
3172 		VALID_SPEC_VALID : VALID_SPEC_INVALID;
3173 }
3174 
3175 static bool is_valid_spec(struct mlx5_core_dev *mdev,
3176 			  const struct mlx5_flow_spec *spec,
3177 			  const struct mlx5_flow_act *flow_act,
3178 			  bool egress)
3179 {
3180 	/* We curretly only support ipsec egress flow */
3181 	return is_valid_esp_aes_gcm(mdev, spec, flow_act, egress) != VALID_SPEC_INVALID;
3182 }
3183 
3184 static bool is_valid_ethertype(struct mlx5_core_dev *mdev,
3185 			       const struct ib_flow_attr *flow_attr,
3186 			       bool check_inner)
3187 {
3188 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
3189 	int match_ipv = check_inner ?
3190 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
3191 					ft_field_support.inner_ip_version) :
3192 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
3193 					ft_field_support.outer_ip_version);
3194 	int inner_bit = check_inner ? IB_FLOW_SPEC_INNER : 0;
3195 	bool ipv4_spec_valid, ipv6_spec_valid;
3196 	unsigned int ip_spec_type = 0;
3197 	bool has_ethertype = false;
3198 	unsigned int spec_index;
3199 	bool mask_valid = true;
3200 	u16 eth_type = 0;
3201 	bool type_valid;
3202 
3203 	/* Validate that ethertype is correct */
3204 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
3205 		if ((ib_spec->type == (IB_FLOW_SPEC_ETH | inner_bit)) &&
3206 		    ib_spec->eth.mask.ether_type) {
3207 			mask_valid = (ib_spec->eth.mask.ether_type ==
3208 				      htons(0xffff));
3209 			has_ethertype = true;
3210 			eth_type = ntohs(ib_spec->eth.val.ether_type);
3211 		} else if ((ib_spec->type == (IB_FLOW_SPEC_IPV4 | inner_bit)) ||
3212 			   (ib_spec->type == (IB_FLOW_SPEC_IPV6 | inner_bit))) {
3213 			ip_spec_type = ib_spec->type;
3214 		}
3215 		ib_spec = (void *)ib_spec + ib_spec->size;
3216 	}
3217 
3218 	type_valid = (!has_ethertype) || (!ip_spec_type);
3219 	if (!type_valid && mask_valid) {
3220 		ipv4_spec_valid = (eth_type == ETH_P_IP) &&
3221 			(ip_spec_type == (IB_FLOW_SPEC_IPV4 | inner_bit));
3222 		ipv6_spec_valid = (eth_type == ETH_P_IPV6) &&
3223 			(ip_spec_type == (IB_FLOW_SPEC_IPV6 | inner_bit));
3224 
3225 		type_valid = (ipv4_spec_valid) || (ipv6_spec_valid) ||
3226 			     (((eth_type == ETH_P_MPLS_UC) ||
3227 			       (eth_type == ETH_P_MPLS_MC)) && match_ipv);
3228 	}
3229 
3230 	return type_valid;
3231 }
3232 
3233 static bool is_valid_attr(struct mlx5_core_dev *mdev,
3234 			  const struct ib_flow_attr *flow_attr)
3235 {
3236 	return is_valid_ethertype(mdev, flow_attr, false) &&
3237 	       is_valid_ethertype(mdev, flow_attr, true);
3238 }
3239 
3240 static void put_flow_table(struct mlx5_ib_dev *dev,
3241 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
3242 {
3243 	prio->refcount -= !!ft_added;
3244 	if (!prio->refcount) {
3245 		mlx5_destroy_flow_table(prio->flow_table);
3246 		prio->flow_table = NULL;
3247 	}
3248 }
3249 
3250 static void counters_clear_description(struct ib_counters *counters)
3251 {
3252 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
3253 
3254 	mutex_lock(&mcounters->mcntrs_mutex);
3255 	kfree(mcounters->counters_data);
3256 	mcounters->counters_data = NULL;
3257 	mcounters->cntrs_max_index = 0;
3258 	mutex_unlock(&mcounters->mcntrs_mutex);
3259 }
3260 
3261 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
3262 {
3263 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
3264 							  struct mlx5_ib_flow_handler,
3265 							  ibflow);
3266 	struct mlx5_ib_flow_handler *iter, *tmp;
3267 	struct mlx5_ib_dev *dev = handler->dev;
3268 
3269 	mutex_lock(&dev->flow_db->lock);
3270 
3271 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
3272 		mlx5_del_flow_rules(iter->rule);
3273 		put_flow_table(dev, iter->prio, true);
3274 		list_del(&iter->list);
3275 		kfree(iter);
3276 	}
3277 
3278 	mlx5_del_flow_rules(handler->rule);
3279 	put_flow_table(dev, handler->prio, true);
3280 	if (handler->ibcounters &&
3281 	    atomic_read(&handler->ibcounters->usecnt) == 1)
3282 		counters_clear_description(handler->ibcounters);
3283 
3284 	mutex_unlock(&dev->flow_db->lock);
3285 	if (handler->flow_matcher)
3286 		atomic_dec(&handler->flow_matcher->usecnt);
3287 	kfree(handler);
3288 
3289 	return 0;
3290 }
3291 
3292 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
3293 {
3294 	priority *= 2;
3295 	if (!dont_trap)
3296 		priority++;
3297 	return priority;
3298 }
3299 
3300 enum flow_table_type {
3301 	MLX5_IB_FT_RX,
3302 	MLX5_IB_FT_TX
3303 };
3304 
3305 #define MLX5_FS_MAX_TYPES	 6
3306 #define MLX5_FS_MAX_ENTRIES	 BIT(16)
3307 
3308 static struct mlx5_ib_flow_prio *_get_prio(struct mlx5_flow_namespace *ns,
3309 					   struct mlx5_ib_flow_prio *prio,
3310 					   int priority,
3311 					   int num_entries, int num_groups,
3312 					   u32 flags)
3313 {
3314 	struct mlx5_flow_table_attr ft_attr = {};
3315 	struct mlx5_flow_table *ft;
3316 
3317 	ft_attr.prio = priority;
3318 	ft_attr.max_fte = num_entries;
3319 	ft_attr.flags = flags;
3320 	ft_attr.autogroup.max_num_groups = num_groups;
3321 	ft = mlx5_create_auto_grouped_flow_table(ns, &ft_attr);
3322 	if (IS_ERR(ft))
3323 		return ERR_CAST(ft);
3324 
3325 	prio->flow_table = ft;
3326 	prio->refcount = 0;
3327 	return prio;
3328 }
3329 
3330 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
3331 						struct ib_flow_attr *flow_attr,
3332 						enum flow_table_type ft_type)
3333 {
3334 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
3335 	struct mlx5_flow_namespace *ns = NULL;
3336 	struct mlx5_ib_flow_prio *prio;
3337 	struct mlx5_flow_table *ft;
3338 	int max_table_size;
3339 	int num_entries;
3340 	int num_groups;
3341 	bool esw_encap;
3342 	u32 flags = 0;
3343 	int priority;
3344 
3345 	max_table_size = BIT(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3346 						       log_max_ft_size));
3347 	esw_encap = mlx5_eswitch_get_encap_mode(dev->mdev) !=
3348 		DEVLINK_ESWITCH_ENCAP_MODE_NONE;
3349 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
3350 		enum mlx5_flow_namespace_type fn_type;
3351 
3352 		if (flow_is_multicast_only(flow_attr) &&
3353 		    !dont_trap)
3354 			priority = MLX5_IB_FLOW_MCAST_PRIO;
3355 		else
3356 			priority = ib_prio_to_core_prio(flow_attr->priority,
3357 							dont_trap);
3358 		if (ft_type == MLX5_IB_FT_RX) {
3359 			fn_type = MLX5_FLOW_NAMESPACE_BYPASS;
3360 			prio = &dev->flow_db->prios[priority];
3361 			if (!dev->is_rep && !esw_encap &&
3362 			    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, decap))
3363 				flags |= MLX5_FLOW_TABLE_TUNNEL_EN_DECAP;
3364 			if (!dev->is_rep && !esw_encap &&
3365 			    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3366 					reformat_l3_tunnel_to_l2))
3367 				flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
3368 		} else {
3369 			max_table_size =
3370 				BIT(MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev,
3371 							      log_max_ft_size));
3372 			fn_type = MLX5_FLOW_NAMESPACE_EGRESS;
3373 			prio = &dev->flow_db->egress_prios[priority];
3374 			if (!dev->is_rep && !esw_encap &&
3375 			    MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, reformat))
3376 				flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
3377 		}
3378 		ns = mlx5_get_flow_namespace(dev->mdev, fn_type);
3379 		num_entries = MLX5_FS_MAX_ENTRIES;
3380 		num_groups = MLX5_FS_MAX_TYPES;
3381 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3382 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
3383 		ns = mlx5_get_flow_namespace(dev->mdev,
3384 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
3385 		build_leftovers_ft_param(&priority,
3386 					 &num_entries,
3387 					 &num_groups);
3388 		prio = &dev->flow_db->prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
3389 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3390 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
3391 					allow_sniffer_and_nic_rx_shared_tir))
3392 			return ERR_PTR(-ENOTSUPP);
3393 
3394 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
3395 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
3396 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
3397 
3398 		prio = &dev->flow_db->sniffer[ft_type];
3399 		priority = 0;
3400 		num_entries = 1;
3401 		num_groups = 1;
3402 	}
3403 
3404 	if (!ns)
3405 		return ERR_PTR(-ENOTSUPP);
3406 
3407 	max_table_size = min_t(int, num_entries, max_table_size);
3408 
3409 	ft = prio->flow_table;
3410 	if (!ft)
3411 		return _get_prio(ns, prio, priority, max_table_size, num_groups,
3412 				 flags);
3413 
3414 	return prio;
3415 }
3416 
3417 static void set_underlay_qp(struct mlx5_ib_dev *dev,
3418 			    struct mlx5_flow_spec *spec,
3419 			    u32 underlay_qpn)
3420 {
3421 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param,
3422 					   spec->match_criteria,
3423 					   misc_parameters);
3424 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3425 					   misc_parameters);
3426 
3427 	if (underlay_qpn &&
3428 	    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3429 				      ft_field_support.bth_dst_qp)) {
3430 		MLX5_SET(fte_match_set_misc,
3431 			 misc_params_v, bth_dst_qp, underlay_qpn);
3432 		MLX5_SET(fte_match_set_misc,
3433 			 misc_params_c, bth_dst_qp, 0xffffff);
3434 	}
3435 }
3436 
3437 static int read_flow_counters(struct ib_device *ibdev,
3438 			      struct mlx5_read_counters_attr *read_attr)
3439 {
3440 	struct mlx5_fc *fc = read_attr->hw_cntrs_hndl;
3441 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3442 
3443 	return mlx5_fc_query(dev->mdev, fc,
3444 			     &read_attr->out[IB_COUNTER_PACKETS],
3445 			     &read_attr->out[IB_COUNTER_BYTES]);
3446 }
3447 
3448 /* flow counters currently expose two counters packets and bytes */
3449 #define FLOW_COUNTERS_NUM 2
3450 static int counters_set_description(struct ib_counters *counters,
3451 				    enum mlx5_ib_counters_type counters_type,
3452 				    struct mlx5_ib_flow_counters_desc *desc_data,
3453 				    u32 ncounters)
3454 {
3455 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
3456 	u32 cntrs_max_index = 0;
3457 	int i;
3458 
3459 	if (counters_type != MLX5_IB_COUNTERS_FLOW)
3460 		return -EINVAL;
3461 
3462 	/* init the fields for the object */
3463 	mcounters->type = counters_type;
3464 	mcounters->read_counters = read_flow_counters;
3465 	mcounters->counters_num = FLOW_COUNTERS_NUM;
3466 	mcounters->ncounters = ncounters;
3467 	/* each counter entry have both description and index pair */
3468 	for (i = 0; i < ncounters; i++) {
3469 		if (desc_data[i].description > IB_COUNTER_BYTES)
3470 			return -EINVAL;
3471 
3472 		if (cntrs_max_index <= desc_data[i].index)
3473 			cntrs_max_index = desc_data[i].index + 1;
3474 	}
3475 
3476 	mutex_lock(&mcounters->mcntrs_mutex);
3477 	mcounters->counters_data = desc_data;
3478 	mcounters->cntrs_max_index = cntrs_max_index;
3479 	mutex_unlock(&mcounters->mcntrs_mutex);
3480 
3481 	return 0;
3482 }
3483 
3484 #define MAX_COUNTERS_NUM (USHRT_MAX / (sizeof(u32) * 2))
3485 static int flow_counters_set_data(struct ib_counters *ibcounters,
3486 				  struct mlx5_ib_create_flow *ucmd)
3487 {
3488 	struct mlx5_ib_mcounters *mcounters = to_mcounters(ibcounters);
3489 	struct mlx5_ib_flow_counters_data *cntrs_data = NULL;
3490 	struct mlx5_ib_flow_counters_desc *desc_data = NULL;
3491 	bool hw_hndl = false;
3492 	int ret = 0;
3493 
3494 	if (ucmd && ucmd->ncounters_data != 0) {
3495 		cntrs_data = ucmd->data;
3496 		if (cntrs_data->ncounters > MAX_COUNTERS_NUM)
3497 			return -EINVAL;
3498 
3499 		desc_data = kcalloc(cntrs_data->ncounters,
3500 				    sizeof(*desc_data),
3501 				    GFP_KERNEL);
3502 		if (!desc_data)
3503 			return  -ENOMEM;
3504 
3505 		if (copy_from_user(desc_data,
3506 				   u64_to_user_ptr(cntrs_data->counters_data),
3507 				   sizeof(*desc_data) * cntrs_data->ncounters)) {
3508 			ret = -EFAULT;
3509 			goto free;
3510 		}
3511 	}
3512 
3513 	if (!mcounters->hw_cntrs_hndl) {
3514 		mcounters->hw_cntrs_hndl = mlx5_fc_create(
3515 			to_mdev(ibcounters->device)->mdev, false);
3516 		if (IS_ERR(mcounters->hw_cntrs_hndl)) {
3517 			ret = PTR_ERR(mcounters->hw_cntrs_hndl);
3518 			goto free;
3519 		}
3520 		hw_hndl = true;
3521 	}
3522 
3523 	if (desc_data) {
3524 		/* counters already bound to at least one flow */
3525 		if (mcounters->cntrs_max_index) {
3526 			ret = -EINVAL;
3527 			goto free_hndl;
3528 		}
3529 
3530 		ret = counters_set_description(ibcounters,
3531 					       MLX5_IB_COUNTERS_FLOW,
3532 					       desc_data,
3533 					       cntrs_data->ncounters);
3534 		if (ret)
3535 			goto free_hndl;
3536 
3537 	} else if (!mcounters->cntrs_max_index) {
3538 		/* counters not bound yet, must have udata passed */
3539 		ret = -EINVAL;
3540 		goto free_hndl;
3541 	}
3542 
3543 	return 0;
3544 
3545 free_hndl:
3546 	if (hw_hndl) {
3547 		mlx5_fc_destroy(to_mdev(ibcounters->device)->mdev,
3548 				mcounters->hw_cntrs_hndl);
3549 		mcounters->hw_cntrs_hndl = NULL;
3550 	}
3551 free:
3552 	kfree(desc_data);
3553 	return ret;
3554 }
3555 
3556 static void mlx5_ib_set_rule_source_port(struct mlx5_ib_dev *dev,
3557 					 struct mlx5_flow_spec *spec,
3558 					 struct mlx5_eswitch_rep *rep)
3559 {
3560 	struct mlx5_eswitch *esw = dev->mdev->priv.eswitch;
3561 	void *misc;
3562 
3563 	if (mlx5_eswitch_vport_match_metadata_enabled(esw)) {
3564 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3565 				    misc_parameters_2);
3566 
3567 		MLX5_SET(fte_match_set_misc2, misc, metadata_reg_c_0,
3568 			 mlx5_eswitch_get_vport_metadata_for_match(esw,
3569 								   rep->vport));
3570 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_criteria,
3571 				    misc_parameters_2);
3572 
3573 		MLX5_SET_TO_ONES(fte_match_set_misc2, misc, metadata_reg_c_0);
3574 	} else {
3575 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3576 				    misc_parameters);
3577 
3578 		MLX5_SET(fte_match_set_misc, misc, source_port, rep->vport);
3579 
3580 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_criteria,
3581 				    misc_parameters);
3582 
3583 		MLX5_SET_TO_ONES(fte_match_set_misc, misc, source_port);
3584 	}
3585 }
3586 
3587 static struct mlx5_ib_flow_handler *_create_flow_rule(struct mlx5_ib_dev *dev,
3588 						      struct mlx5_ib_flow_prio *ft_prio,
3589 						      const struct ib_flow_attr *flow_attr,
3590 						      struct mlx5_flow_destination *dst,
3591 						      u32 underlay_qpn,
3592 						      struct mlx5_ib_create_flow *ucmd)
3593 {
3594 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
3595 	struct mlx5_ib_flow_handler *handler;
3596 	struct mlx5_flow_act flow_act = {};
3597 	struct mlx5_flow_spec *spec;
3598 	struct mlx5_flow_destination dest_arr[2] = {};
3599 	struct mlx5_flow_destination *rule_dst = dest_arr;
3600 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
3601 	unsigned int spec_index;
3602 	u32 prev_type = 0;
3603 	int err = 0;
3604 	int dest_num = 0;
3605 	bool is_egress = flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS;
3606 
3607 	if (!is_valid_attr(dev->mdev, flow_attr))
3608 		return ERR_PTR(-EINVAL);
3609 
3610 	if (dev->is_rep && is_egress)
3611 		return ERR_PTR(-EINVAL);
3612 
3613 	spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
3614 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
3615 	if (!handler || !spec) {
3616 		err = -ENOMEM;
3617 		goto free;
3618 	}
3619 
3620 	INIT_LIST_HEAD(&handler->list);
3621 
3622 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
3623 		err = parse_flow_attr(dev->mdev, spec,
3624 				      ib_flow, flow_attr, &flow_act,
3625 				      prev_type);
3626 		if (err < 0)
3627 			goto free;
3628 
3629 		prev_type = ((union ib_flow_spec *)ib_flow)->type;
3630 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
3631 	}
3632 
3633 	if (dst && !(flow_act.action & MLX5_FLOW_CONTEXT_ACTION_DROP)) {
3634 		memcpy(&dest_arr[0], dst, sizeof(*dst));
3635 		dest_num++;
3636 	}
3637 
3638 	if (!flow_is_multicast_only(flow_attr))
3639 		set_underlay_qp(dev, spec, underlay_qpn);
3640 
3641 	if (dev->is_rep) {
3642 		struct mlx5_eswitch_rep *rep;
3643 
3644 		rep = dev->port[flow_attr->port - 1].rep;
3645 		if (!rep) {
3646 			err = -EINVAL;
3647 			goto free;
3648 		}
3649 
3650 		mlx5_ib_set_rule_source_port(dev, spec, rep);
3651 	}
3652 
3653 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
3654 
3655 	if (is_egress &&
3656 	    !is_valid_spec(dev->mdev, spec, &flow_act, is_egress)) {
3657 		err = -EINVAL;
3658 		goto free;
3659 	}
3660 
3661 	if (flow_act.action & MLX5_FLOW_CONTEXT_ACTION_COUNT) {
3662 		struct mlx5_ib_mcounters *mcounters;
3663 
3664 		err = flow_counters_set_data(flow_act.counters, ucmd);
3665 		if (err)
3666 			goto free;
3667 
3668 		mcounters = to_mcounters(flow_act.counters);
3669 		handler->ibcounters = flow_act.counters;
3670 		dest_arr[dest_num].type =
3671 			MLX5_FLOW_DESTINATION_TYPE_COUNTER;
3672 		dest_arr[dest_num].counter_id =
3673 			mlx5_fc_id(mcounters->hw_cntrs_hndl);
3674 		dest_num++;
3675 	}
3676 
3677 	if (flow_act.action & MLX5_FLOW_CONTEXT_ACTION_DROP) {
3678 		if (!dest_num)
3679 			rule_dst = NULL;
3680 	} else {
3681 		if (is_egress)
3682 			flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_ALLOW;
3683 		else
3684 			flow_act.action |=
3685 				dest_num ?  MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
3686 					MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
3687 	}
3688 
3689 	if ((spec->flow_context.flags & FLOW_CONTEXT_HAS_TAG)  &&
3690 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3691 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
3692 		mlx5_ib_warn(dev, "Flow tag %u and attribute type %x isn't allowed in leftovers\n",
3693 			     spec->flow_context.flow_tag, flow_attr->type);
3694 		err = -EINVAL;
3695 		goto free;
3696 	}
3697 	handler->rule = mlx5_add_flow_rules(ft, spec,
3698 					    &flow_act,
3699 					    rule_dst, dest_num);
3700 
3701 	if (IS_ERR(handler->rule)) {
3702 		err = PTR_ERR(handler->rule);
3703 		goto free;
3704 	}
3705 
3706 	ft_prio->refcount++;
3707 	handler->prio = ft_prio;
3708 	handler->dev = dev;
3709 
3710 	ft_prio->flow_table = ft;
3711 free:
3712 	if (err && handler) {
3713 		if (handler->ibcounters &&
3714 		    atomic_read(&handler->ibcounters->usecnt) == 1)
3715 			counters_clear_description(handler->ibcounters);
3716 		kfree(handler);
3717 	}
3718 	kvfree(spec);
3719 	return err ? ERR_PTR(err) : handler;
3720 }
3721 
3722 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
3723 						     struct mlx5_ib_flow_prio *ft_prio,
3724 						     const struct ib_flow_attr *flow_attr,
3725 						     struct mlx5_flow_destination *dst)
3726 {
3727 	return _create_flow_rule(dev, ft_prio, flow_attr, dst, 0, NULL);
3728 }
3729 
3730 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
3731 							  struct mlx5_ib_flow_prio *ft_prio,
3732 							  struct ib_flow_attr *flow_attr,
3733 							  struct mlx5_flow_destination *dst)
3734 {
3735 	struct mlx5_ib_flow_handler *handler_dst = NULL;
3736 	struct mlx5_ib_flow_handler *handler = NULL;
3737 
3738 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
3739 	if (!IS_ERR(handler)) {
3740 		handler_dst = create_flow_rule(dev, ft_prio,
3741 					       flow_attr, dst);
3742 		if (IS_ERR(handler_dst)) {
3743 			mlx5_del_flow_rules(handler->rule);
3744 			ft_prio->refcount--;
3745 			kfree(handler);
3746 			handler = handler_dst;
3747 		} else {
3748 			list_add(&handler_dst->list, &handler->list);
3749 		}
3750 	}
3751 
3752 	return handler;
3753 }
3754 enum {
3755 	LEFTOVERS_MC,
3756 	LEFTOVERS_UC,
3757 };
3758 
3759 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
3760 							  struct mlx5_ib_flow_prio *ft_prio,
3761 							  struct ib_flow_attr *flow_attr,
3762 							  struct mlx5_flow_destination *dst)
3763 {
3764 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
3765 	struct mlx5_ib_flow_handler *handler = NULL;
3766 
3767 	static struct {
3768 		struct ib_flow_attr	flow_attr;
3769 		struct ib_flow_spec_eth eth_flow;
3770 	} leftovers_specs[] = {
3771 		[LEFTOVERS_MC] = {
3772 			.flow_attr = {
3773 				.num_of_specs = 1,
3774 				.size = sizeof(leftovers_specs[0])
3775 			},
3776 			.eth_flow = {
3777 				.type = IB_FLOW_SPEC_ETH,
3778 				.size = sizeof(struct ib_flow_spec_eth),
3779 				.mask = {.dst_mac = {0x1} },
3780 				.val =  {.dst_mac = {0x1} }
3781 			}
3782 		},
3783 		[LEFTOVERS_UC] = {
3784 			.flow_attr = {
3785 				.num_of_specs = 1,
3786 				.size = sizeof(leftovers_specs[0])
3787 			},
3788 			.eth_flow = {
3789 				.type = IB_FLOW_SPEC_ETH,
3790 				.size = sizeof(struct ib_flow_spec_eth),
3791 				.mask = {.dst_mac = {0x1} },
3792 				.val = {.dst_mac = {} }
3793 			}
3794 		}
3795 	};
3796 
3797 	handler = create_flow_rule(dev, ft_prio,
3798 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
3799 				   dst);
3800 	if (!IS_ERR(handler) &&
3801 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
3802 		handler_ucast = create_flow_rule(dev, ft_prio,
3803 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
3804 						 dst);
3805 		if (IS_ERR(handler_ucast)) {
3806 			mlx5_del_flow_rules(handler->rule);
3807 			ft_prio->refcount--;
3808 			kfree(handler);
3809 			handler = handler_ucast;
3810 		} else {
3811 			list_add(&handler_ucast->list, &handler->list);
3812 		}
3813 	}
3814 
3815 	return handler;
3816 }
3817 
3818 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
3819 							struct mlx5_ib_flow_prio *ft_rx,
3820 							struct mlx5_ib_flow_prio *ft_tx,
3821 							struct mlx5_flow_destination *dst)
3822 {
3823 	struct mlx5_ib_flow_handler *handler_rx;
3824 	struct mlx5_ib_flow_handler *handler_tx;
3825 	int err;
3826 	static const struct ib_flow_attr flow_attr  = {
3827 		.num_of_specs = 0,
3828 		.size = sizeof(flow_attr)
3829 	};
3830 
3831 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
3832 	if (IS_ERR(handler_rx)) {
3833 		err = PTR_ERR(handler_rx);
3834 		goto err;
3835 	}
3836 
3837 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
3838 	if (IS_ERR(handler_tx)) {
3839 		err = PTR_ERR(handler_tx);
3840 		goto err_tx;
3841 	}
3842 
3843 	list_add(&handler_tx->list, &handler_rx->list);
3844 
3845 	return handler_rx;
3846 
3847 err_tx:
3848 	mlx5_del_flow_rules(handler_rx->rule);
3849 	ft_rx->refcount--;
3850 	kfree(handler_rx);
3851 err:
3852 	return ERR_PTR(err);
3853 }
3854 
3855 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
3856 					   struct ib_flow_attr *flow_attr,
3857 					   int domain,
3858 					   struct ib_udata *udata)
3859 {
3860 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3861 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3862 	struct mlx5_ib_flow_handler *handler = NULL;
3863 	struct mlx5_flow_destination *dst = NULL;
3864 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
3865 	struct mlx5_ib_flow_prio *ft_prio;
3866 	bool is_egress = flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS;
3867 	struct mlx5_ib_create_flow *ucmd = NULL, ucmd_hdr;
3868 	size_t min_ucmd_sz, required_ucmd_sz;
3869 	int err;
3870 	int underlay_qpn;
3871 
3872 	if (udata && udata->inlen) {
3873 		min_ucmd_sz = offsetof(typeof(ucmd_hdr), reserved) +
3874 				sizeof(ucmd_hdr.reserved);
3875 		if (udata->inlen < min_ucmd_sz)
3876 			return ERR_PTR(-EOPNOTSUPP);
3877 
3878 		err = ib_copy_from_udata(&ucmd_hdr, udata, min_ucmd_sz);
3879 		if (err)
3880 			return ERR_PTR(err);
3881 
3882 		/* currently supports only one counters data */
3883 		if (ucmd_hdr.ncounters_data > 1)
3884 			return ERR_PTR(-EINVAL);
3885 
3886 		required_ucmd_sz = min_ucmd_sz +
3887 			sizeof(struct mlx5_ib_flow_counters_data) *
3888 			ucmd_hdr.ncounters_data;
3889 		if (udata->inlen > required_ucmd_sz &&
3890 		    !ib_is_udata_cleared(udata, required_ucmd_sz,
3891 					 udata->inlen - required_ucmd_sz))
3892 			return ERR_PTR(-EOPNOTSUPP);
3893 
3894 		ucmd = kzalloc(required_ucmd_sz, GFP_KERNEL);
3895 		if (!ucmd)
3896 			return ERR_PTR(-ENOMEM);
3897 
3898 		err = ib_copy_from_udata(ucmd, udata, required_ucmd_sz);
3899 		if (err)
3900 			goto free_ucmd;
3901 	}
3902 
3903 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO) {
3904 		err = -ENOMEM;
3905 		goto free_ucmd;
3906 	}
3907 
3908 	if (domain != IB_FLOW_DOMAIN_USER ||
3909 	    flow_attr->port > dev->num_ports ||
3910 	    (flow_attr->flags & ~(IB_FLOW_ATTR_FLAGS_DONT_TRAP |
3911 				  IB_FLOW_ATTR_FLAGS_EGRESS))) {
3912 		err = -EINVAL;
3913 		goto free_ucmd;
3914 	}
3915 
3916 	if (is_egress &&
3917 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3918 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
3919 		err = -EINVAL;
3920 		goto free_ucmd;
3921 	}
3922 
3923 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
3924 	if (!dst) {
3925 		err = -ENOMEM;
3926 		goto free_ucmd;
3927 	}
3928 
3929 	mutex_lock(&dev->flow_db->lock);
3930 
3931 	ft_prio = get_flow_table(dev, flow_attr,
3932 				 is_egress ? MLX5_IB_FT_TX : MLX5_IB_FT_RX);
3933 	if (IS_ERR(ft_prio)) {
3934 		err = PTR_ERR(ft_prio);
3935 		goto unlock;
3936 	}
3937 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3938 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
3939 		if (IS_ERR(ft_prio_tx)) {
3940 			err = PTR_ERR(ft_prio_tx);
3941 			ft_prio_tx = NULL;
3942 			goto destroy_ft;
3943 		}
3944 	}
3945 
3946 	if (is_egress) {
3947 		dst->type = MLX5_FLOW_DESTINATION_TYPE_PORT;
3948 	} else {
3949 		dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
3950 		if (mqp->flags & MLX5_IB_QP_RSS)
3951 			dst->tir_num = mqp->rss_qp.tirn;
3952 		else
3953 			dst->tir_num = mqp->raw_packet_qp.rq.tirn;
3954 	}
3955 
3956 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
3957 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
3958 			handler = create_dont_trap_rule(dev, ft_prio,
3959 							flow_attr, dst);
3960 		} else {
3961 			underlay_qpn = (mqp->flags & MLX5_IB_QP_UNDERLAY) ?
3962 					mqp->underlay_qpn : 0;
3963 			handler = _create_flow_rule(dev, ft_prio, flow_attr,
3964 						    dst, underlay_qpn, ucmd);
3965 		}
3966 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3967 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
3968 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
3969 						dst);
3970 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3971 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
3972 	} else {
3973 		err = -EINVAL;
3974 		goto destroy_ft;
3975 	}
3976 
3977 	if (IS_ERR(handler)) {
3978 		err = PTR_ERR(handler);
3979 		handler = NULL;
3980 		goto destroy_ft;
3981 	}
3982 
3983 	mutex_unlock(&dev->flow_db->lock);
3984 	kfree(dst);
3985 	kfree(ucmd);
3986 
3987 	return &handler->ibflow;
3988 
3989 destroy_ft:
3990 	put_flow_table(dev, ft_prio, false);
3991 	if (ft_prio_tx)
3992 		put_flow_table(dev, ft_prio_tx, false);
3993 unlock:
3994 	mutex_unlock(&dev->flow_db->lock);
3995 	kfree(dst);
3996 free_ucmd:
3997 	kfree(ucmd);
3998 	return ERR_PTR(err);
3999 }
4000 
4001 static struct mlx5_ib_flow_prio *
4002 _get_flow_table(struct mlx5_ib_dev *dev,
4003 		struct mlx5_ib_flow_matcher *fs_matcher,
4004 		bool mcast)
4005 {
4006 	struct mlx5_flow_namespace *ns = NULL;
4007 	struct mlx5_ib_flow_prio *prio = NULL;
4008 	int max_table_size = 0;
4009 	bool esw_encap;
4010 	u32 flags = 0;
4011 	int priority;
4012 
4013 	if (mcast)
4014 		priority = MLX5_IB_FLOW_MCAST_PRIO;
4015 	else
4016 		priority = ib_prio_to_core_prio(fs_matcher->priority, false);
4017 
4018 	esw_encap = mlx5_eswitch_get_encap_mode(dev->mdev) !=
4019 		DEVLINK_ESWITCH_ENCAP_MODE_NONE;
4020 	if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_BYPASS) {
4021 		max_table_size = BIT(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
4022 					log_max_ft_size));
4023 		if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, decap) && !esw_encap)
4024 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_DECAP;
4025 		if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
4026 					      reformat_l3_tunnel_to_l2) &&
4027 		    !esw_encap)
4028 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
4029 	} else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_EGRESS) {
4030 		max_table_size = BIT(
4031 			MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, log_max_ft_size));
4032 		if (MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, reformat) && !esw_encap)
4033 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
4034 	} else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_FDB) {
4035 		max_table_size = BIT(
4036 			MLX5_CAP_ESW_FLOWTABLE_FDB(dev->mdev, log_max_ft_size));
4037 		if (MLX5_CAP_ESW_FLOWTABLE_FDB(dev->mdev, decap) && esw_encap)
4038 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_DECAP;
4039 		if (MLX5_CAP_ESW_FLOWTABLE_FDB(dev->mdev, reformat_l3_tunnel_to_l2) &&
4040 		    esw_encap)
4041 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
4042 		priority = FDB_BYPASS_PATH;
4043 	} else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_RDMA_RX) {
4044 		max_table_size =
4045 			BIT(MLX5_CAP_FLOWTABLE_RDMA_RX(dev->mdev,
4046 						       log_max_ft_size));
4047 		priority = fs_matcher->priority;
4048 	}
4049 
4050 	max_table_size = min_t(int, max_table_size, MLX5_FS_MAX_ENTRIES);
4051 
4052 	ns = mlx5_get_flow_namespace(dev->mdev, fs_matcher->ns_type);
4053 	if (!ns)
4054 		return ERR_PTR(-ENOTSUPP);
4055 
4056 	if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_BYPASS)
4057 		prio = &dev->flow_db->prios[priority];
4058 	else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_EGRESS)
4059 		prio = &dev->flow_db->egress_prios[priority];
4060 	else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_FDB)
4061 		prio = &dev->flow_db->fdb;
4062 	else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_RDMA_RX)
4063 		prio = &dev->flow_db->rdma_rx[priority];
4064 
4065 	if (!prio)
4066 		return ERR_PTR(-EINVAL);
4067 
4068 	if (prio->flow_table)
4069 		return prio;
4070 
4071 	return _get_prio(ns, prio, priority, max_table_size,
4072 			 MLX5_FS_MAX_TYPES, flags);
4073 }
4074 
4075 static struct mlx5_ib_flow_handler *
4076 _create_raw_flow_rule(struct mlx5_ib_dev *dev,
4077 		      struct mlx5_ib_flow_prio *ft_prio,
4078 		      struct mlx5_flow_destination *dst,
4079 		      struct mlx5_ib_flow_matcher  *fs_matcher,
4080 		      struct mlx5_flow_context *flow_context,
4081 		      struct mlx5_flow_act *flow_act,
4082 		      void *cmd_in, int inlen,
4083 		      int dst_num)
4084 {
4085 	struct mlx5_ib_flow_handler *handler;
4086 	struct mlx5_flow_spec *spec;
4087 	struct mlx5_flow_table *ft = ft_prio->flow_table;
4088 	int err = 0;
4089 
4090 	spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
4091 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
4092 	if (!handler || !spec) {
4093 		err = -ENOMEM;
4094 		goto free;
4095 	}
4096 
4097 	INIT_LIST_HEAD(&handler->list);
4098 
4099 	memcpy(spec->match_value, cmd_in, inlen);
4100 	memcpy(spec->match_criteria, fs_matcher->matcher_mask.match_params,
4101 	       fs_matcher->mask_len);
4102 	spec->match_criteria_enable = fs_matcher->match_criteria_enable;
4103 	spec->flow_context = *flow_context;
4104 
4105 	handler->rule = mlx5_add_flow_rules(ft, spec,
4106 					    flow_act, dst, dst_num);
4107 
4108 	if (IS_ERR(handler->rule)) {
4109 		err = PTR_ERR(handler->rule);
4110 		goto free;
4111 	}
4112 
4113 	ft_prio->refcount++;
4114 	handler->prio = ft_prio;
4115 	handler->dev = dev;
4116 	ft_prio->flow_table = ft;
4117 
4118 free:
4119 	if (err)
4120 		kfree(handler);
4121 	kvfree(spec);
4122 	return err ? ERR_PTR(err) : handler;
4123 }
4124 
4125 static bool raw_fs_is_multicast(struct mlx5_ib_flow_matcher *fs_matcher,
4126 				void *match_v)
4127 {
4128 	void *match_c;
4129 	void *match_v_set_lyr_2_4, *match_c_set_lyr_2_4;
4130 	void *dmac, *dmac_mask;
4131 	void *ipv4, *ipv4_mask;
4132 
4133 	if (!(fs_matcher->match_criteria_enable &
4134 	      (1 << MATCH_CRITERIA_ENABLE_OUTER_BIT)))
4135 		return false;
4136 
4137 	match_c = fs_matcher->matcher_mask.match_params;
4138 	match_v_set_lyr_2_4 = MLX5_ADDR_OF(fte_match_param, match_v,
4139 					   outer_headers);
4140 	match_c_set_lyr_2_4 = MLX5_ADDR_OF(fte_match_param, match_c,
4141 					   outer_headers);
4142 
4143 	dmac = MLX5_ADDR_OF(fte_match_set_lyr_2_4, match_v_set_lyr_2_4,
4144 			    dmac_47_16);
4145 	dmac_mask = MLX5_ADDR_OF(fte_match_set_lyr_2_4, match_c_set_lyr_2_4,
4146 				 dmac_47_16);
4147 
4148 	if (is_multicast_ether_addr(dmac) &&
4149 	    is_multicast_ether_addr(dmac_mask))
4150 		return true;
4151 
4152 	ipv4 = MLX5_ADDR_OF(fte_match_set_lyr_2_4, match_v_set_lyr_2_4,
4153 			    dst_ipv4_dst_ipv6.ipv4_layout.ipv4);
4154 
4155 	ipv4_mask = MLX5_ADDR_OF(fte_match_set_lyr_2_4, match_c_set_lyr_2_4,
4156 				 dst_ipv4_dst_ipv6.ipv4_layout.ipv4);
4157 
4158 	if (ipv4_is_multicast(*(__be32 *)(ipv4)) &&
4159 	    ipv4_is_multicast(*(__be32 *)(ipv4_mask)))
4160 		return true;
4161 
4162 	return false;
4163 }
4164 
4165 struct mlx5_ib_flow_handler *
4166 mlx5_ib_raw_fs_rule_add(struct mlx5_ib_dev *dev,
4167 			struct mlx5_ib_flow_matcher *fs_matcher,
4168 			struct mlx5_flow_context *flow_context,
4169 			struct mlx5_flow_act *flow_act,
4170 			u32 counter_id,
4171 			void *cmd_in, int inlen, int dest_id,
4172 			int dest_type)
4173 {
4174 	struct mlx5_flow_destination *dst;
4175 	struct mlx5_ib_flow_prio *ft_prio;
4176 	struct mlx5_ib_flow_handler *handler;
4177 	int dst_num = 0;
4178 	bool mcast;
4179 	int err;
4180 
4181 	if (fs_matcher->flow_type != MLX5_IB_FLOW_TYPE_NORMAL)
4182 		return ERR_PTR(-EOPNOTSUPP);
4183 
4184 	if (fs_matcher->priority > MLX5_IB_FLOW_LAST_PRIO)
4185 		return ERR_PTR(-ENOMEM);
4186 
4187 	dst = kcalloc(2, sizeof(*dst), GFP_KERNEL);
4188 	if (!dst)
4189 		return ERR_PTR(-ENOMEM);
4190 
4191 	mcast = raw_fs_is_multicast(fs_matcher, cmd_in);
4192 	mutex_lock(&dev->flow_db->lock);
4193 
4194 	ft_prio = _get_flow_table(dev, fs_matcher, mcast);
4195 	if (IS_ERR(ft_prio)) {
4196 		err = PTR_ERR(ft_prio);
4197 		goto unlock;
4198 	}
4199 
4200 	if (dest_type == MLX5_FLOW_DESTINATION_TYPE_TIR) {
4201 		dst[dst_num].type = dest_type;
4202 		dst[dst_num].tir_num = dest_id;
4203 		flow_act->action |= MLX5_FLOW_CONTEXT_ACTION_FWD_DEST;
4204 	} else if (dest_type == MLX5_FLOW_DESTINATION_TYPE_FLOW_TABLE) {
4205 		dst[dst_num].type = MLX5_FLOW_DESTINATION_TYPE_FLOW_TABLE_NUM;
4206 		dst[dst_num].ft_num = dest_id;
4207 		flow_act->action |= MLX5_FLOW_CONTEXT_ACTION_FWD_DEST;
4208 	} else {
4209 		dst[dst_num].type = MLX5_FLOW_DESTINATION_TYPE_PORT;
4210 		flow_act->action |= MLX5_FLOW_CONTEXT_ACTION_ALLOW;
4211 	}
4212 
4213 	dst_num++;
4214 
4215 	if (flow_act->action & MLX5_FLOW_CONTEXT_ACTION_COUNT) {
4216 		dst[dst_num].type = MLX5_FLOW_DESTINATION_TYPE_COUNTER;
4217 		dst[dst_num].counter_id = counter_id;
4218 		dst_num++;
4219 	}
4220 
4221 	handler = _create_raw_flow_rule(dev, ft_prio, dst, fs_matcher,
4222 					flow_context, flow_act,
4223 					cmd_in, inlen, dst_num);
4224 
4225 	if (IS_ERR(handler)) {
4226 		err = PTR_ERR(handler);
4227 		goto destroy_ft;
4228 	}
4229 
4230 	mutex_unlock(&dev->flow_db->lock);
4231 	atomic_inc(&fs_matcher->usecnt);
4232 	handler->flow_matcher = fs_matcher;
4233 
4234 	kfree(dst);
4235 
4236 	return handler;
4237 
4238 destroy_ft:
4239 	put_flow_table(dev, ft_prio, false);
4240 unlock:
4241 	mutex_unlock(&dev->flow_db->lock);
4242 	kfree(dst);
4243 
4244 	return ERR_PTR(err);
4245 }
4246 
4247 static u32 mlx5_ib_flow_action_flags_to_accel_xfrm_flags(u32 mlx5_flags)
4248 {
4249 	u32 flags = 0;
4250 
4251 	if (mlx5_flags & MLX5_IB_UAPI_FLOW_ACTION_FLAGS_REQUIRE_METADATA)
4252 		flags |= MLX5_ACCEL_XFRM_FLAG_REQUIRE_METADATA;
4253 
4254 	return flags;
4255 }
4256 
4257 #define MLX5_FLOW_ACTION_ESP_CREATE_LAST_SUPPORTED	MLX5_IB_UAPI_FLOW_ACTION_FLAGS_REQUIRE_METADATA
4258 static struct ib_flow_action *
4259 mlx5_ib_create_flow_action_esp(struct ib_device *device,
4260 			       const struct ib_flow_action_attrs_esp *attr,
4261 			       struct uverbs_attr_bundle *attrs)
4262 {
4263 	struct mlx5_ib_dev *mdev = to_mdev(device);
4264 	struct ib_uverbs_flow_action_esp_keymat_aes_gcm *aes_gcm;
4265 	struct mlx5_accel_esp_xfrm_attrs accel_attrs = {};
4266 	struct mlx5_ib_flow_action *action;
4267 	u64 action_flags;
4268 	u64 flags;
4269 	int err = 0;
4270 
4271 	err = uverbs_get_flags64(
4272 		&action_flags, attrs, MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
4273 		((MLX5_FLOW_ACTION_ESP_CREATE_LAST_SUPPORTED << 1) - 1));
4274 	if (err)
4275 		return ERR_PTR(err);
4276 
4277 	flags = mlx5_ib_flow_action_flags_to_accel_xfrm_flags(action_flags);
4278 
4279 	/* We current only support a subset of the standard features. Only a
4280 	 * keymat of type AES_GCM, with icv_len == 16, iv_algo == SEQ and esn
4281 	 * (with overlap). Full offload mode isn't supported.
4282 	 */
4283 	if (!attr->keymat || attr->replay || attr->encap ||
4284 	    attr->spi || attr->seq || attr->tfc_pad ||
4285 	    attr->hard_limit_pkts ||
4286 	    (attr->flags & ~(IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
4287 			     IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ENCRYPT)))
4288 		return ERR_PTR(-EOPNOTSUPP);
4289 
4290 	if (attr->keymat->protocol !=
4291 	    IB_UVERBS_FLOW_ACTION_ESP_KEYMAT_AES_GCM)
4292 		return ERR_PTR(-EOPNOTSUPP);
4293 
4294 	aes_gcm = &attr->keymat->keymat.aes_gcm;
4295 
4296 	if (aes_gcm->icv_len != 16 ||
4297 	    aes_gcm->iv_algo != IB_UVERBS_FLOW_ACTION_IV_ALGO_SEQ)
4298 		return ERR_PTR(-EOPNOTSUPP);
4299 
4300 	action = kmalloc(sizeof(*action), GFP_KERNEL);
4301 	if (!action)
4302 		return ERR_PTR(-ENOMEM);
4303 
4304 	action->esp_aes_gcm.ib_flags = attr->flags;
4305 	memcpy(&accel_attrs.keymat.aes_gcm.aes_key, &aes_gcm->aes_key,
4306 	       sizeof(accel_attrs.keymat.aes_gcm.aes_key));
4307 	accel_attrs.keymat.aes_gcm.key_len = aes_gcm->key_len * 8;
4308 	memcpy(&accel_attrs.keymat.aes_gcm.salt, &aes_gcm->salt,
4309 	       sizeof(accel_attrs.keymat.aes_gcm.salt));
4310 	memcpy(&accel_attrs.keymat.aes_gcm.seq_iv, &aes_gcm->iv,
4311 	       sizeof(accel_attrs.keymat.aes_gcm.seq_iv));
4312 	accel_attrs.keymat.aes_gcm.icv_len = aes_gcm->icv_len * 8;
4313 	accel_attrs.keymat.aes_gcm.iv_algo = MLX5_ACCEL_ESP_AES_GCM_IV_ALGO_SEQ;
4314 	accel_attrs.keymat_type = MLX5_ACCEL_ESP_KEYMAT_AES_GCM;
4315 
4316 	accel_attrs.esn = attr->esn;
4317 	if (attr->flags & IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED)
4318 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_TRIGGERED;
4319 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)
4320 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
4321 
4322 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ENCRYPT)
4323 		accel_attrs.action |= MLX5_ACCEL_ESP_ACTION_ENCRYPT;
4324 
4325 	action->esp_aes_gcm.ctx =
4326 		mlx5_accel_esp_create_xfrm(mdev->mdev, &accel_attrs, flags);
4327 	if (IS_ERR(action->esp_aes_gcm.ctx)) {
4328 		err = PTR_ERR(action->esp_aes_gcm.ctx);
4329 		goto err_parse;
4330 	}
4331 
4332 	action->esp_aes_gcm.ib_flags = attr->flags;
4333 
4334 	return &action->ib_action;
4335 
4336 err_parse:
4337 	kfree(action);
4338 	return ERR_PTR(err);
4339 }
4340 
4341 static int
4342 mlx5_ib_modify_flow_action_esp(struct ib_flow_action *action,
4343 			       const struct ib_flow_action_attrs_esp *attr,
4344 			       struct uverbs_attr_bundle *attrs)
4345 {
4346 	struct mlx5_ib_flow_action *maction = to_mflow_act(action);
4347 	struct mlx5_accel_esp_xfrm_attrs accel_attrs;
4348 	int err = 0;
4349 
4350 	if (attr->keymat || attr->replay || attr->encap ||
4351 	    attr->spi || attr->seq || attr->tfc_pad ||
4352 	    attr->hard_limit_pkts ||
4353 	    (attr->flags & ~(IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
4354 			     IB_FLOW_ACTION_ESP_FLAGS_MOD_ESP_ATTRS |
4355 			     IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)))
4356 		return -EOPNOTSUPP;
4357 
4358 	/* Only the ESN value or the MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP can
4359 	 * be modified.
4360 	 */
4361 	if (!(maction->esp_aes_gcm.ib_flags &
4362 	      IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED) &&
4363 	    attr->flags & (IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
4364 			   IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW))
4365 		return -EINVAL;
4366 
4367 	memcpy(&accel_attrs, &maction->esp_aes_gcm.ctx->attrs,
4368 	       sizeof(accel_attrs));
4369 
4370 	accel_attrs.esn = attr->esn;
4371 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)
4372 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
4373 	else
4374 		accel_attrs.flags &= ~MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
4375 
4376 	err = mlx5_accel_esp_modify_xfrm(maction->esp_aes_gcm.ctx,
4377 					 &accel_attrs);
4378 	if (err)
4379 		return err;
4380 
4381 	maction->esp_aes_gcm.ib_flags &=
4382 		~IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW;
4383 	maction->esp_aes_gcm.ib_flags |=
4384 		attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW;
4385 
4386 	return 0;
4387 }
4388 
4389 static int mlx5_ib_destroy_flow_action(struct ib_flow_action *action)
4390 {
4391 	struct mlx5_ib_flow_action *maction = to_mflow_act(action);
4392 
4393 	switch (action->type) {
4394 	case IB_FLOW_ACTION_ESP:
4395 		/*
4396 		 * We only support aes_gcm by now, so we implicitly know this is
4397 		 * the underline crypto.
4398 		 */
4399 		mlx5_accel_esp_destroy_xfrm(maction->esp_aes_gcm.ctx);
4400 		break;
4401 	case IB_FLOW_ACTION_UNSPECIFIED:
4402 		mlx5_ib_destroy_flow_action_raw(maction);
4403 		break;
4404 	default:
4405 		WARN_ON(true);
4406 		break;
4407 	}
4408 
4409 	kfree(maction);
4410 	return 0;
4411 }
4412 
4413 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
4414 {
4415 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4416 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
4417 	int err;
4418 	u16 uid;
4419 
4420 	uid = ibqp->pd ?
4421 		to_mpd(ibqp->pd)->uid : 0;
4422 
4423 	if (mqp->flags & MLX5_IB_QP_UNDERLAY) {
4424 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
4425 		return -EOPNOTSUPP;
4426 	}
4427 
4428 	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
4429 	if (err)
4430 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
4431 			     ibqp->qp_num, gid->raw);
4432 
4433 	return err;
4434 }
4435 
4436 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
4437 {
4438 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4439 	int err;
4440 	u16 uid;
4441 
4442 	uid = ibqp->pd ?
4443 		to_mpd(ibqp->pd)->uid : 0;
4444 	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
4445 	if (err)
4446 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
4447 			     ibqp->qp_num, gid->raw);
4448 
4449 	return err;
4450 }
4451 
4452 static int init_node_data(struct mlx5_ib_dev *dev)
4453 {
4454 	int err;
4455 
4456 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
4457 	if (err)
4458 		return err;
4459 
4460 	dev->mdev->rev_id = dev->mdev->pdev->revision;
4461 
4462 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
4463 }
4464 
4465 static ssize_t fw_pages_show(struct device *device,
4466 			     struct device_attribute *attr, char *buf)
4467 {
4468 	struct mlx5_ib_dev *dev =
4469 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4470 
4471 	return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
4472 }
4473 static DEVICE_ATTR_RO(fw_pages);
4474 
4475 static ssize_t reg_pages_show(struct device *device,
4476 			      struct device_attribute *attr, char *buf)
4477 {
4478 	struct mlx5_ib_dev *dev =
4479 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4480 
4481 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
4482 }
4483 static DEVICE_ATTR_RO(reg_pages);
4484 
4485 static ssize_t hca_type_show(struct device *device,
4486 			     struct device_attribute *attr, char *buf)
4487 {
4488 	struct mlx5_ib_dev *dev =
4489 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4490 
4491 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
4492 }
4493 static DEVICE_ATTR_RO(hca_type);
4494 
4495 static ssize_t hw_rev_show(struct device *device,
4496 			   struct device_attribute *attr, char *buf)
4497 {
4498 	struct mlx5_ib_dev *dev =
4499 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4500 
4501 	return sprintf(buf, "%x\n", dev->mdev->rev_id);
4502 }
4503 static DEVICE_ATTR_RO(hw_rev);
4504 
4505 static ssize_t board_id_show(struct device *device,
4506 			     struct device_attribute *attr, char *buf)
4507 {
4508 	struct mlx5_ib_dev *dev =
4509 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4510 
4511 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
4512 		       dev->mdev->board_id);
4513 }
4514 static DEVICE_ATTR_RO(board_id);
4515 
4516 static struct attribute *mlx5_class_attributes[] = {
4517 	&dev_attr_hw_rev.attr,
4518 	&dev_attr_hca_type.attr,
4519 	&dev_attr_board_id.attr,
4520 	&dev_attr_fw_pages.attr,
4521 	&dev_attr_reg_pages.attr,
4522 	NULL,
4523 };
4524 
4525 static const struct attribute_group mlx5_attr_group = {
4526 	.attrs = mlx5_class_attributes,
4527 };
4528 
4529 static void pkey_change_handler(struct work_struct *work)
4530 {
4531 	struct mlx5_ib_port_resources *ports =
4532 		container_of(work, struct mlx5_ib_port_resources,
4533 			     pkey_change_work);
4534 
4535 	mutex_lock(&ports->devr->mutex);
4536 	mlx5_ib_gsi_pkey_change(ports->gsi);
4537 	mutex_unlock(&ports->devr->mutex);
4538 }
4539 
4540 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
4541 {
4542 	struct mlx5_ib_qp *mqp;
4543 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
4544 	struct mlx5_core_cq *mcq;
4545 	struct list_head cq_armed_list;
4546 	unsigned long flags_qp;
4547 	unsigned long flags_cq;
4548 	unsigned long flags;
4549 
4550 	INIT_LIST_HEAD(&cq_armed_list);
4551 
4552 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
4553 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
4554 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
4555 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
4556 		if (mqp->sq.tail != mqp->sq.head) {
4557 			send_mcq = to_mcq(mqp->ibqp.send_cq);
4558 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
4559 			if (send_mcq->mcq.comp &&
4560 			    mqp->ibqp.send_cq->comp_handler) {
4561 				if (!send_mcq->mcq.reset_notify_added) {
4562 					send_mcq->mcq.reset_notify_added = 1;
4563 					list_add_tail(&send_mcq->mcq.reset_notify,
4564 						      &cq_armed_list);
4565 				}
4566 			}
4567 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
4568 		}
4569 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
4570 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
4571 		/* no handling is needed for SRQ */
4572 		if (!mqp->ibqp.srq) {
4573 			if (mqp->rq.tail != mqp->rq.head) {
4574 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
4575 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
4576 				if (recv_mcq->mcq.comp &&
4577 				    mqp->ibqp.recv_cq->comp_handler) {
4578 					if (!recv_mcq->mcq.reset_notify_added) {
4579 						recv_mcq->mcq.reset_notify_added = 1;
4580 						list_add_tail(&recv_mcq->mcq.reset_notify,
4581 							      &cq_armed_list);
4582 					}
4583 				}
4584 				spin_unlock_irqrestore(&recv_mcq->lock,
4585 						       flags_cq);
4586 			}
4587 		}
4588 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
4589 	}
4590 	/*At that point all inflight post send were put to be executed as of we
4591 	 * lock/unlock above locks Now need to arm all involved CQs.
4592 	 */
4593 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
4594 		mcq->comp(mcq, NULL);
4595 	}
4596 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
4597 }
4598 
4599 static void delay_drop_handler(struct work_struct *work)
4600 {
4601 	int err;
4602 	struct mlx5_ib_delay_drop *delay_drop =
4603 		container_of(work, struct mlx5_ib_delay_drop,
4604 			     delay_drop_work);
4605 
4606 	atomic_inc(&delay_drop->events_cnt);
4607 
4608 	mutex_lock(&delay_drop->lock);
4609 	err = mlx5_core_set_delay_drop(delay_drop->dev->mdev,
4610 				       delay_drop->timeout);
4611 	if (err) {
4612 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
4613 			     delay_drop->timeout);
4614 		delay_drop->activate = false;
4615 	}
4616 	mutex_unlock(&delay_drop->lock);
4617 }
4618 
4619 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
4620 				 struct ib_event *ibev)
4621 {
4622 	u8 port = (eqe->data.port.port >> 4) & 0xf;
4623 
4624 	switch (eqe->sub_type) {
4625 	case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
4626 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
4627 					    IB_LINK_LAYER_ETHERNET)
4628 			schedule_work(&ibdev->delay_drop.delay_drop_work);
4629 		break;
4630 	default: /* do nothing */
4631 		return;
4632 	}
4633 }
4634 
4635 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
4636 			      struct ib_event *ibev)
4637 {
4638 	u8 port = (eqe->data.port.port >> 4) & 0xf;
4639 
4640 	ibev->element.port_num = port;
4641 
4642 	switch (eqe->sub_type) {
4643 	case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
4644 	case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
4645 	case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
4646 		/* In RoCE, port up/down events are handled in
4647 		 * mlx5_netdev_event().
4648 		 */
4649 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
4650 					    IB_LINK_LAYER_ETHERNET)
4651 			return -EINVAL;
4652 
4653 		ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
4654 				IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
4655 		break;
4656 
4657 	case MLX5_PORT_CHANGE_SUBTYPE_LID:
4658 		ibev->event = IB_EVENT_LID_CHANGE;
4659 		break;
4660 
4661 	case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
4662 		ibev->event = IB_EVENT_PKEY_CHANGE;
4663 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
4664 		break;
4665 
4666 	case MLX5_PORT_CHANGE_SUBTYPE_GUID:
4667 		ibev->event = IB_EVENT_GID_CHANGE;
4668 		break;
4669 
4670 	case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
4671 		ibev->event = IB_EVENT_CLIENT_REREGISTER;
4672 		break;
4673 	default:
4674 		return -EINVAL;
4675 	}
4676 
4677 	return 0;
4678 }
4679 
4680 static void mlx5_ib_handle_event(struct work_struct *_work)
4681 {
4682 	struct mlx5_ib_event_work *work =
4683 		container_of(_work, struct mlx5_ib_event_work, work);
4684 	struct mlx5_ib_dev *ibdev;
4685 	struct ib_event ibev;
4686 	bool fatal = false;
4687 
4688 	if (work->is_slave) {
4689 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
4690 		if (!ibdev)
4691 			goto out;
4692 	} else {
4693 		ibdev = work->dev;
4694 	}
4695 
4696 	switch (work->event) {
4697 	case MLX5_DEV_EVENT_SYS_ERROR:
4698 		ibev.event = IB_EVENT_DEVICE_FATAL;
4699 		mlx5_ib_handle_internal_error(ibdev);
4700 		ibev.element.port_num  = (u8)(unsigned long)work->param;
4701 		fatal = true;
4702 		break;
4703 	case MLX5_EVENT_TYPE_PORT_CHANGE:
4704 		if (handle_port_change(ibdev, work->param, &ibev))
4705 			goto out;
4706 		break;
4707 	case MLX5_EVENT_TYPE_GENERAL_EVENT:
4708 		handle_general_event(ibdev, work->param, &ibev);
4709 		/* fall through */
4710 	default:
4711 		goto out;
4712 	}
4713 
4714 	ibev.device = &ibdev->ib_dev;
4715 
4716 	if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
4717 		mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
4718 		goto out;
4719 	}
4720 
4721 	if (ibdev->ib_active)
4722 		ib_dispatch_event(&ibev);
4723 
4724 	if (fatal)
4725 		ibdev->ib_active = false;
4726 out:
4727 	kfree(work);
4728 }
4729 
4730 static int mlx5_ib_event(struct notifier_block *nb,
4731 			 unsigned long event, void *param)
4732 {
4733 	struct mlx5_ib_event_work *work;
4734 
4735 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
4736 	if (!work)
4737 		return NOTIFY_DONE;
4738 
4739 	INIT_WORK(&work->work, mlx5_ib_handle_event);
4740 	work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
4741 	work->is_slave = false;
4742 	work->param = param;
4743 	work->event = event;
4744 
4745 	queue_work(mlx5_ib_event_wq, &work->work);
4746 
4747 	return NOTIFY_OK;
4748 }
4749 
4750 static int mlx5_ib_event_slave_port(struct notifier_block *nb,
4751 				    unsigned long event, void *param)
4752 {
4753 	struct mlx5_ib_event_work *work;
4754 
4755 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
4756 	if (!work)
4757 		return NOTIFY_DONE;
4758 
4759 	INIT_WORK(&work->work, mlx5_ib_handle_event);
4760 	work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
4761 	work->is_slave = true;
4762 	work->param = param;
4763 	work->event = event;
4764 	queue_work(mlx5_ib_event_wq, &work->work);
4765 
4766 	return NOTIFY_OK;
4767 }
4768 
4769 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
4770 {
4771 	struct mlx5_hca_vport_context vport_ctx;
4772 	int err;
4773 	int port;
4774 
4775 	for (port = 1; port <= ARRAY_SIZE(dev->mdev->port_caps); port++) {
4776 		dev->mdev->port_caps[port - 1].has_smi = false;
4777 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
4778 		    MLX5_CAP_PORT_TYPE_IB) {
4779 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
4780 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
4781 								   port, 0,
4782 								   &vport_ctx);
4783 				if (err) {
4784 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
4785 						    port, err);
4786 					return err;
4787 				}
4788 				dev->mdev->port_caps[port - 1].has_smi =
4789 					vport_ctx.has_smi;
4790 			} else {
4791 				dev->mdev->port_caps[port - 1].has_smi = true;
4792 			}
4793 		}
4794 	}
4795 	return 0;
4796 }
4797 
4798 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
4799 {
4800 	int port;
4801 
4802 	for (port = 1; port <= dev->num_ports; port++)
4803 		mlx5_query_ext_port_caps(dev, port);
4804 }
4805 
4806 static int __get_port_caps(struct mlx5_ib_dev *dev, u8 port)
4807 {
4808 	struct ib_device_attr *dprops = NULL;
4809 	struct ib_port_attr *pprops = NULL;
4810 	int err = -ENOMEM;
4811 
4812 	pprops = kzalloc(sizeof(*pprops), GFP_KERNEL);
4813 	if (!pprops)
4814 		goto out;
4815 
4816 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
4817 	if (!dprops)
4818 		goto out;
4819 
4820 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, NULL);
4821 	if (err) {
4822 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
4823 		goto out;
4824 	}
4825 
4826 	err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
4827 	if (err) {
4828 		mlx5_ib_warn(dev, "query_port %d failed %d\n",
4829 			     port, err);
4830 		goto out;
4831 	}
4832 
4833 	dev->mdev->port_caps[port - 1].pkey_table_len =
4834 					dprops->max_pkeys;
4835 	dev->mdev->port_caps[port - 1].gid_table_len =
4836 					pprops->gid_tbl_len;
4837 	mlx5_ib_dbg(dev, "port %d: pkey_table_len %d, gid_table_len %d\n",
4838 		    port, dprops->max_pkeys, pprops->gid_tbl_len);
4839 
4840 out:
4841 	kfree(pprops);
4842 	kfree(dprops);
4843 
4844 	return err;
4845 }
4846 
4847 static int get_port_caps(struct mlx5_ib_dev *dev, u8 port)
4848 {
4849 	/* For representors use port 1, is this is the only native
4850 	 * port
4851 	 */
4852 	if (dev->is_rep)
4853 		return __get_port_caps(dev, 1);
4854 	return __get_port_caps(dev, port);
4855 }
4856 
4857 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
4858 {
4859 	int err;
4860 
4861 	err = mlx5_mr_cache_cleanup(dev);
4862 	if (err)
4863 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
4864 
4865 	if (dev->umrc.qp)
4866 		mlx5_ib_destroy_qp(dev->umrc.qp, NULL);
4867 	if (dev->umrc.cq)
4868 		ib_free_cq(dev->umrc.cq);
4869 	if (dev->umrc.pd)
4870 		ib_dealloc_pd(dev->umrc.pd);
4871 }
4872 
4873 enum {
4874 	MAX_UMR_WR = 128,
4875 };
4876 
4877 static int create_umr_res(struct mlx5_ib_dev *dev)
4878 {
4879 	struct ib_qp_init_attr *init_attr = NULL;
4880 	struct ib_qp_attr *attr = NULL;
4881 	struct ib_pd *pd;
4882 	struct ib_cq *cq;
4883 	struct ib_qp *qp;
4884 	int ret;
4885 
4886 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
4887 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
4888 	if (!attr || !init_attr) {
4889 		ret = -ENOMEM;
4890 		goto error_0;
4891 	}
4892 
4893 	pd = ib_alloc_pd(&dev->ib_dev, 0);
4894 	if (IS_ERR(pd)) {
4895 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
4896 		ret = PTR_ERR(pd);
4897 		goto error_0;
4898 	}
4899 
4900 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
4901 	if (IS_ERR(cq)) {
4902 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
4903 		ret = PTR_ERR(cq);
4904 		goto error_2;
4905 	}
4906 
4907 	init_attr->send_cq = cq;
4908 	init_attr->recv_cq = cq;
4909 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
4910 	init_attr->cap.max_send_wr = MAX_UMR_WR;
4911 	init_attr->cap.max_send_sge = 1;
4912 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
4913 	init_attr->port_num = 1;
4914 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
4915 	if (IS_ERR(qp)) {
4916 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
4917 		ret = PTR_ERR(qp);
4918 		goto error_3;
4919 	}
4920 	qp->device     = &dev->ib_dev;
4921 	qp->real_qp    = qp;
4922 	qp->uobject    = NULL;
4923 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
4924 	qp->send_cq    = init_attr->send_cq;
4925 	qp->recv_cq    = init_attr->recv_cq;
4926 
4927 	attr->qp_state = IB_QPS_INIT;
4928 	attr->port_num = 1;
4929 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
4930 				IB_QP_PORT, NULL);
4931 	if (ret) {
4932 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
4933 		goto error_4;
4934 	}
4935 
4936 	memset(attr, 0, sizeof(*attr));
4937 	attr->qp_state = IB_QPS_RTR;
4938 	attr->path_mtu = IB_MTU_256;
4939 
4940 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4941 	if (ret) {
4942 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
4943 		goto error_4;
4944 	}
4945 
4946 	memset(attr, 0, sizeof(*attr));
4947 	attr->qp_state = IB_QPS_RTS;
4948 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4949 	if (ret) {
4950 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
4951 		goto error_4;
4952 	}
4953 
4954 	dev->umrc.qp = qp;
4955 	dev->umrc.cq = cq;
4956 	dev->umrc.pd = pd;
4957 
4958 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
4959 	ret = mlx5_mr_cache_init(dev);
4960 	if (ret) {
4961 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4962 		goto error_4;
4963 	}
4964 
4965 	kfree(attr);
4966 	kfree(init_attr);
4967 
4968 	return 0;
4969 
4970 error_4:
4971 	mlx5_ib_destroy_qp(qp, NULL);
4972 	dev->umrc.qp = NULL;
4973 
4974 error_3:
4975 	ib_free_cq(cq);
4976 	dev->umrc.cq = NULL;
4977 
4978 error_2:
4979 	ib_dealloc_pd(pd);
4980 	dev->umrc.pd = NULL;
4981 
4982 error_0:
4983 	kfree(attr);
4984 	kfree(init_attr);
4985 	return ret;
4986 }
4987 
4988 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
4989 {
4990 	switch (umr_fence_cap) {
4991 	case MLX5_CAP_UMR_FENCE_NONE:
4992 		return MLX5_FENCE_MODE_NONE;
4993 	case MLX5_CAP_UMR_FENCE_SMALL:
4994 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
4995 	default:
4996 		return MLX5_FENCE_MODE_STRONG_ORDERING;
4997 	}
4998 }
4999 
5000 static int create_dev_resources(struct mlx5_ib_resources *devr)
5001 {
5002 	struct ib_srq_init_attr attr;
5003 	struct mlx5_ib_dev *dev;
5004 	struct ib_device *ibdev;
5005 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
5006 	int port;
5007 	int ret = 0;
5008 
5009 	dev = container_of(devr, struct mlx5_ib_dev, devr);
5010 	ibdev = &dev->ib_dev;
5011 
5012 	mutex_init(&devr->mutex);
5013 
5014 	devr->p0 = rdma_zalloc_drv_obj(ibdev, ib_pd);
5015 	if (!devr->p0)
5016 		return -ENOMEM;
5017 
5018 	devr->p0->device  = ibdev;
5019 	devr->p0->uobject = NULL;
5020 	atomic_set(&devr->p0->usecnt, 0);
5021 
5022 	ret = mlx5_ib_alloc_pd(devr->p0, NULL);
5023 	if (ret)
5024 		goto error0;
5025 
5026 	devr->c0 = rdma_zalloc_drv_obj(ibdev, ib_cq);
5027 	if (!devr->c0) {
5028 		ret = -ENOMEM;
5029 		goto error1;
5030 	}
5031 
5032 	devr->c0->device = &dev->ib_dev;
5033 	atomic_set(&devr->c0->usecnt, 0);
5034 
5035 	ret = mlx5_ib_create_cq(devr->c0, &cq_attr, NULL);
5036 	if (ret)
5037 		goto err_create_cq;
5038 
5039 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL);
5040 	if (IS_ERR(devr->x0)) {
5041 		ret = PTR_ERR(devr->x0);
5042 		goto error2;
5043 	}
5044 	devr->x0->device = &dev->ib_dev;
5045 	devr->x0->inode = NULL;
5046 	atomic_set(&devr->x0->usecnt, 0);
5047 	mutex_init(&devr->x0->tgt_qp_mutex);
5048 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
5049 
5050 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL);
5051 	if (IS_ERR(devr->x1)) {
5052 		ret = PTR_ERR(devr->x1);
5053 		goto error3;
5054 	}
5055 	devr->x1->device = &dev->ib_dev;
5056 	devr->x1->inode = NULL;
5057 	atomic_set(&devr->x1->usecnt, 0);
5058 	mutex_init(&devr->x1->tgt_qp_mutex);
5059 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
5060 
5061 	memset(&attr, 0, sizeof(attr));
5062 	attr.attr.max_sge = 1;
5063 	attr.attr.max_wr = 1;
5064 	attr.srq_type = IB_SRQT_XRC;
5065 	attr.ext.cq = devr->c0;
5066 	attr.ext.xrc.xrcd = devr->x0;
5067 
5068 	devr->s0 = rdma_zalloc_drv_obj(ibdev, ib_srq);
5069 	if (!devr->s0) {
5070 		ret = -ENOMEM;
5071 		goto error4;
5072 	}
5073 
5074 	devr->s0->device	= &dev->ib_dev;
5075 	devr->s0->pd		= devr->p0;
5076 	devr->s0->srq_type      = IB_SRQT_XRC;
5077 	devr->s0->ext.xrc.xrcd	= devr->x0;
5078 	devr->s0->ext.cq	= devr->c0;
5079 	ret = mlx5_ib_create_srq(devr->s0, &attr, NULL);
5080 	if (ret)
5081 		goto err_create;
5082 
5083 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
5084 	atomic_inc(&devr->s0->ext.cq->usecnt);
5085 	atomic_inc(&devr->p0->usecnt);
5086 	atomic_set(&devr->s0->usecnt, 0);
5087 
5088 	memset(&attr, 0, sizeof(attr));
5089 	attr.attr.max_sge = 1;
5090 	attr.attr.max_wr = 1;
5091 	attr.srq_type = IB_SRQT_BASIC;
5092 	devr->s1 = rdma_zalloc_drv_obj(ibdev, ib_srq);
5093 	if (!devr->s1) {
5094 		ret = -ENOMEM;
5095 		goto error5;
5096 	}
5097 
5098 	devr->s1->device	= &dev->ib_dev;
5099 	devr->s1->pd		= devr->p0;
5100 	devr->s1->srq_type      = IB_SRQT_BASIC;
5101 	devr->s1->ext.cq	= devr->c0;
5102 
5103 	ret = mlx5_ib_create_srq(devr->s1, &attr, NULL);
5104 	if (ret)
5105 		goto error6;
5106 
5107 	atomic_inc(&devr->p0->usecnt);
5108 	atomic_set(&devr->s1->usecnt, 0);
5109 
5110 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
5111 		INIT_WORK(&devr->ports[port].pkey_change_work,
5112 			  pkey_change_handler);
5113 		devr->ports[port].devr = devr;
5114 	}
5115 
5116 	return 0;
5117 
5118 error6:
5119 	kfree(devr->s1);
5120 error5:
5121 	mlx5_ib_destroy_srq(devr->s0, NULL);
5122 err_create:
5123 	kfree(devr->s0);
5124 error4:
5125 	mlx5_ib_dealloc_xrcd(devr->x1, NULL);
5126 error3:
5127 	mlx5_ib_dealloc_xrcd(devr->x0, NULL);
5128 error2:
5129 	mlx5_ib_destroy_cq(devr->c0, NULL);
5130 err_create_cq:
5131 	kfree(devr->c0);
5132 error1:
5133 	mlx5_ib_dealloc_pd(devr->p0, NULL);
5134 error0:
5135 	kfree(devr->p0);
5136 	return ret;
5137 }
5138 
5139 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
5140 {
5141 	int port;
5142 
5143 	mlx5_ib_destroy_srq(devr->s1, NULL);
5144 	kfree(devr->s1);
5145 	mlx5_ib_destroy_srq(devr->s0, NULL);
5146 	kfree(devr->s0);
5147 	mlx5_ib_dealloc_xrcd(devr->x0, NULL);
5148 	mlx5_ib_dealloc_xrcd(devr->x1, NULL);
5149 	mlx5_ib_destroy_cq(devr->c0, NULL);
5150 	kfree(devr->c0);
5151 	mlx5_ib_dealloc_pd(devr->p0, NULL);
5152 	kfree(devr->p0);
5153 
5154 	/* Make sure no change P_Key work items are still executing */
5155 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
5156 		cancel_work_sync(&devr->ports[port].pkey_change_work);
5157 }
5158 
5159 static u32 get_core_cap_flags(struct ib_device *ibdev,
5160 			      struct mlx5_hca_vport_context *rep)
5161 {
5162 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5163 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
5164 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
5165 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
5166 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
5167 	u32 ret = 0;
5168 
5169 	if (rep->grh_required)
5170 		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
5171 
5172 	if (ll == IB_LINK_LAYER_INFINIBAND)
5173 		return ret | RDMA_CORE_PORT_IBA_IB;
5174 
5175 	if (raw_support)
5176 		ret |= RDMA_CORE_PORT_RAW_PACKET;
5177 
5178 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
5179 		return ret;
5180 
5181 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
5182 		return ret;
5183 
5184 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
5185 		ret |= RDMA_CORE_PORT_IBA_ROCE;
5186 
5187 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
5188 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
5189 
5190 	return ret;
5191 }
5192 
5193 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
5194 			       struct ib_port_immutable *immutable)
5195 {
5196 	struct ib_port_attr attr;
5197 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5198 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
5199 	struct mlx5_hca_vport_context rep = {0};
5200 	int err;
5201 
5202 	err = ib_query_port(ibdev, port_num, &attr);
5203 	if (err)
5204 		return err;
5205 
5206 	if (ll == IB_LINK_LAYER_INFINIBAND) {
5207 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
5208 						   &rep);
5209 		if (err)
5210 			return err;
5211 	}
5212 
5213 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
5214 	immutable->gid_tbl_len = attr.gid_tbl_len;
5215 	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
5216 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
5217 
5218 	return 0;
5219 }
5220 
5221 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u8 port_num,
5222 				   struct ib_port_immutable *immutable)
5223 {
5224 	struct ib_port_attr attr;
5225 	int err;
5226 
5227 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
5228 
5229 	err = ib_query_port(ibdev, port_num, &attr);
5230 	if (err)
5231 		return err;
5232 
5233 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
5234 	immutable->gid_tbl_len = attr.gid_tbl_len;
5235 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
5236 
5237 	return 0;
5238 }
5239 
5240 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
5241 {
5242 	struct mlx5_ib_dev *dev =
5243 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
5244 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
5245 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
5246 		 fw_rev_sub(dev->mdev));
5247 }
5248 
5249 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
5250 {
5251 	struct mlx5_core_dev *mdev = dev->mdev;
5252 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
5253 								 MLX5_FLOW_NAMESPACE_LAG);
5254 	struct mlx5_flow_table *ft;
5255 	int err;
5256 
5257 	if (!ns || !mlx5_lag_is_roce(mdev))
5258 		return 0;
5259 
5260 	err = mlx5_cmd_create_vport_lag(mdev);
5261 	if (err)
5262 		return err;
5263 
5264 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
5265 	if (IS_ERR(ft)) {
5266 		err = PTR_ERR(ft);
5267 		goto err_destroy_vport_lag;
5268 	}
5269 
5270 	dev->flow_db->lag_demux_ft = ft;
5271 	dev->lag_active = true;
5272 	return 0;
5273 
5274 err_destroy_vport_lag:
5275 	mlx5_cmd_destroy_vport_lag(mdev);
5276 	return err;
5277 }
5278 
5279 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
5280 {
5281 	struct mlx5_core_dev *mdev = dev->mdev;
5282 
5283 	if (dev->lag_active) {
5284 		dev->lag_active = false;
5285 
5286 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
5287 		dev->flow_db->lag_demux_ft = NULL;
5288 
5289 		mlx5_cmd_destroy_vport_lag(mdev);
5290 	}
5291 }
5292 
5293 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
5294 {
5295 	int err;
5296 
5297 	dev->port[port_num].roce.nb.notifier_call = mlx5_netdev_event;
5298 	err = register_netdevice_notifier(&dev->port[port_num].roce.nb);
5299 	if (err) {
5300 		dev->port[port_num].roce.nb.notifier_call = NULL;
5301 		return err;
5302 	}
5303 
5304 	return 0;
5305 }
5306 
5307 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
5308 {
5309 	if (dev->port[port_num].roce.nb.notifier_call) {
5310 		unregister_netdevice_notifier(&dev->port[port_num].roce.nb);
5311 		dev->port[port_num].roce.nb.notifier_call = NULL;
5312 	}
5313 }
5314 
5315 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
5316 {
5317 	int err;
5318 
5319 	err = mlx5_nic_vport_enable_roce(dev->mdev);
5320 	if (err)
5321 		return err;
5322 
5323 	err = mlx5_eth_lag_init(dev);
5324 	if (err)
5325 		goto err_disable_roce;
5326 
5327 	return 0;
5328 
5329 err_disable_roce:
5330 	mlx5_nic_vport_disable_roce(dev->mdev);
5331 
5332 	return err;
5333 }
5334 
5335 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
5336 {
5337 	mlx5_eth_lag_cleanup(dev);
5338 	mlx5_nic_vport_disable_roce(dev->mdev);
5339 }
5340 
5341 struct mlx5_ib_counter {
5342 	const char *name;
5343 	size_t offset;
5344 };
5345 
5346 #define INIT_Q_COUNTER(_name)		\
5347 	{ .name = #_name, .offset = MLX5_BYTE_OFF(query_q_counter_out, _name)}
5348 
5349 static const struct mlx5_ib_counter basic_q_cnts[] = {
5350 	INIT_Q_COUNTER(rx_write_requests),
5351 	INIT_Q_COUNTER(rx_read_requests),
5352 	INIT_Q_COUNTER(rx_atomic_requests),
5353 	INIT_Q_COUNTER(out_of_buffer),
5354 };
5355 
5356 static const struct mlx5_ib_counter out_of_seq_q_cnts[] = {
5357 	INIT_Q_COUNTER(out_of_sequence),
5358 };
5359 
5360 static const struct mlx5_ib_counter retrans_q_cnts[] = {
5361 	INIT_Q_COUNTER(duplicate_request),
5362 	INIT_Q_COUNTER(rnr_nak_retry_err),
5363 	INIT_Q_COUNTER(packet_seq_err),
5364 	INIT_Q_COUNTER(implied_nak_seq_err),
5365 	INIT_Q_COUNTER(local_ack_timeout_err),
5366 };
5367 
5368 #define INIT_CONG_COUNTER(_name)		\
5369 	{ .name = #_name, .offset =	\
5370 		MLX5_BYTE_OFF(query_cong_statistics_out, _name ## _high)}
5371 
5372 static const struct mlx5_ib_counter cong_cnts[] = {
5373 	INIT_CONG_COUNTER(rp_cnp_ignored),
5374 	INIT_CONG_COUNTER(rp_cnp_handled),
5375 	INIT_CONG_COUNTER(np_ecn_marked_roce_packets),
5376 	INIT_CONG_COUNTER(np_cnp_sent),
5377 };
5378 
5379 static const struct mlx5_ib_counter extended_err_cnts[] = {
5380 	INIT_Q_COUNTER(resp_local_length_error),
5381 	INIT_Q_COUNTER(resp_cqe_error),
5382 	INIT_Q_COUNTER(req_cqe_error),
5383 	INIT_Q_COUNTER(req_remote_invalid_request),
5384 	INIT_Q_COUNTER(req_remote_access_errors),
5385 	INIT_Q_COUNTER(resp_remote_access_errors),
5386 	INIT_Q_COUNTER(resp_cqe_flush_error),
5387 	INIT_Q_COUNTER(req_cqe_flush_error),
5388 };
5389 
5390 static const struct mlx5_ib_counter roce_accl_cnts[] = {
5391 	INIT_Q_COUNTER(roce_adp_retrans),
5392 	INIT_Q_COUNTER(roce_adp_retrans_to),
5393 	INIT_Q_COUNTER(roce_slow_restart),
5394 	INIT_Q_COUNTER(roce_slow_restart_cnps),
5395 	INIT_Q_COUNTER(roce_slow_restart_trans),
5396 };
5397 
5398 #define INIT_EXT_PPCNT_COUNTER(_name)		\
5399 	{ .name = #_name, .offset =	\
5400 	MLX5_BYTE_OFF(ppcnt_reg, \
5401 		      counter_set.eth_extended_cntrs_grp_data_layout._name##_high)}
5402 
5403 static const struct mlx5_ib_counter ext_ppcnt_cnts[] = {
5404 	INIT_EXT_PPCNT_COUNTER(rx_icrc_encapsulated),
5405 };
5406 
5407 static bool is_mdev_switchdev_mode(const struct mlx5_core_dev *mdev)
5408 {
5409 	return MLX5_ESWITCH_MANAGER(mdev) &&
5410 	       mlx5_ib_eswitch_mode(mdev->priv.eswitch) ==
5411 		       MLX5_ESWITCH_OFFLOADS;
5412 }
5413 
5414 static void mlx5_ib_dealloc_counters(struct mlx5_ib_dev *dev)
5415 {
5416 	int num_cnt_ports;
5417 	int i;
5418 
5419 	num_cnt_ports = is_mdev_switchdev_mode(dev->mdev) ? 1 : dev->num_ports;
5420 
5421 	for (i = 0; i < num_cnt_ports; i++) {
5422 		if (dev->port[i].cnts.set_id_valid)
5423 			mlx5_core_dealloc_q_counter(dev->mdev,
5424 						    dev->port[i].cnts.set_id);
5425 		kfree(dev->port[i].cnts.names);
5426 		kfree(dev->port[i].cnts.offsets);
5427 	}
5428 }
5429 
5430 static int __mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev,
5431 				    struct mlx5_ib_counters *cnts)
5432 {
5433 	u32 num_counters;
5434 
5435 	num_counters = ARRAY_SIZE(basic_q_cnts);
5436 
5437 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt))
5438 		num_counters += ARRAY_SIZE(out_of_seq_q_cnts);
5439 
5440 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters))
5441 		num_counters += ARRAY_SIZE(retrans_q_cnts);
5442 
5443 	if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters))
5444 		num_counters += ARRAY_SIZE(extended_err_cnts);
5445 
5446 	if (MLX5_CAP_GEN(dev->mdev, roce_accl))
5447 		num_counters += ARRAY_SIZE(roce_accl_cnts);
5448 
5449 	cnts->num_q_counters = num_counters;
5450 
5451 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
5452 		cnts->num_cong_counters = ARRAY_SIZE(cong_cnts);
5453 		num_counters += ARRAY_SIZE(cong_cnts);
5454 	}
5455 	if (MLX5_CAP_PCAM_FEATURE(dev->mdev, rx_icrc_encapsulated_counter)) {
5456 		cnts->num_ext_ppcnt_counters = ARRAY_SIZE(ext_ppcnt_cnts);
5457 		num_counters += ARRAY_SIZE(ext_ppcnt_cnts);
5458 	}
5459 	cnts->names = kcalloc(num_counters, sizeof(cnts->names), GFP_KERNEL);
5460 	if (!cnts->names)
5461 		return -ENOMEM;
5462 
5463 	cnts->offsets = kcalloc(num_counters,
5464 				sizeof(cnts->offsets), GFP_KERNEL);
5465 	if (!cnts->offsets)
5466 		goto err_names;
5467 
5468 	return 0;
5469 
5470 err_names:
5471 	kfree(cnts->names);
5472 	cnts->names = NULL;
5473 	return -ENOMEM;
5474 }
5475 
5476 static void mlx5_ib_fill_counters(struct mlx5_ib_dev *dev,
5477 				  const char **names,
5478 				  size_t *offsets)
5479 {
5480 	int i;
5481 	int j = 0;
5482 
5483 	for (i = 0; i < ARRAY_SIZE(basic_q_cnts); i++, j++) {
5484 		names[j] = basic_q_cnts[i].name;
5485 		offsets[j] = basic_q_cnts[i].offset;
5486 	}
5487 
5488 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt)) {
5489 		for (i = 0; i < ARRAY_SIZE(out_of_seq_q_cnts); i++, j++) {
5490 			names[j] = out_of_seq_q_cnts[i].name;
5491 			offsets[j] = out_of_seq_q_cnts[i].offset;
5492 		}
5493 	}
5494 
5495 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
5496 		for (i = 0; i < ARRAY_SIZE(retrans_q_cnts); i++, j++) {
5497 			names[j] = retrans_q_cnts[i].name;
5498 			offsets[j] = retrans_q_cnts[i].offset;
5499 		}
5500 	}
5501 
5502 	if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters)) {
5503 		for (i = 0; i < ARRAY_SIZE(extended_err_cnts); i++, j++) {
5504 			names[j] = extended_err_cnts[i].name;
5505 			offsets[j] = extended_err_cnts[i].offset;
5506 		}
5507 	}
5508 
5509 	if (MLX5_CAP_GEN(dev->mdev, roce_accl)) {
5510 		for (i = 0; i < ARRAY_SIZE(roce_accl_cnts); i++, j++) {
5511 			names[j] = roce_accl_cnts[i].name;
5512 			offsets[j] = roce_accl_cnts[i].offset;
5513 		}
5514 	}
5515 
5516 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
5517 		for (i = 0; i < ARRAY_SIZE(cong_cnts); i++, j++) {
5518 			names[j] = cong_cnts[i].name;
5519 			offsets[j] = cong_cnts[i].offset;
5520 		}
5521 	}
5522 
5523 	if (MLX5_CAP_PCAM_FEATURE(dev->mdev, rx_icrc_encapsulated_counter)) {
5524 		for (i = 0; i < ARRAY_SIZE(ext_ppcnt_cnts); i++, j++) {
5525 			names[j] = ext_ppcnt_cnts[i].name;
5526 			offsets[j] = ext_ppcnt_cnts[i].offset;
5527 		}
5528 	}
5529 }
5530 
5531 static int mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev)
5532 {
5533 	int num_cnt_ports;
5534 	int err = 0;
5535 	int i;
5536 	bool is_shared;
5537 
5538 	is_shared = MLX5_CAP_GEN(dev->mdev, log_max_uctx) != 0;
5539 	num_cnt_ports = is_mdev_switchdev_mode(dev->mdev) ? 1 : dev->num_ports;
5540 
5541 	for (i = 0; i < num_cnt_ports; i++) {
5542 		err = __mlx5_ib_alloc_counters(dev, &dev->port[i].cnts);
5543 		if (err)
5544 			goto err_alloc;
5545 
5546 		mlx5_ib_fill_counters(dev, dev->port[i].cnts.names,
5547 				      dev->port[i].cnts.offsets);
5548 
5549 		err = mlx5_cmd_alloc_q_counter(dev->mdev,
5550 					       &dev->port[i].cnts.set_id,
5551 					       is_shared ?
5552 					       MLX5_SHARED_RESOURCE_UID : 0);
5553 		if (err) {
5554 			mlx5_ib_warn(dev,
5555 				     "couldn't allocate queue counter for port %d, err %d\n",
5556 				     i + 1, err);
5557 			goto err_alloc;
5558 		}
5559 		dev->port[i].cnts.set_id_valid = true;
5560 	}
5561 	return 0;
5562 
5563 err_alloc:
5564 	mlx5_ib_dealloc_counters(dev);
5565 	return err;
5566 }
5567 
5568 static const struct mlx5_ib_counters *get_counters(struct mlx5_ib_dev *dev,
5569 						   u8 port_num)
5570 {
5571 	return is_mdev_switchdev_mode(dev->mdev) ? &dev->port[0].cnts :
5572 						   &dev->port[port_num].cnts;
5573 }
5574 
5575 /**
5576  * mlx5_ib_get_counters_id - Returns counters id to use for device+port
5577  * @dev:	Pointer to mlx5 IB device
5578  * @port_num:	Zero based port number
5579  *
5580  * mlx5_ib_get_counters_id() Returns counters set id to use for given
5581  * device port combination in switchdev and non switchdev mode of the
5582  * parent device.
5583  */
5584 u16 mlx5_ib_get_counters_id(struct mlx5_ib_dev *dev, u8 port_num)
5585 {
5586 	const struct mlx5_ib_counters *cnts = get_counters(dev, port_num);
5587 
5588 	return cnts->set_id;
5589 }
5590 
5591 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
5592 						    u8 port_num)
5593 {
5594 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5595 	const struct mlx5_ib_counters *cnts;
5596 	bool is_switchdev = is_mdev_switchdev_mode(dev->mdev);
5597 
5598 	if ((is_switchdev && port_num) || (!is_switchdev && !port_num))
5599 		return NULL;
5600 
5601 	cnts = get_counters(dev, port_num - 1);
5602 
5603 	return rdma_alloc_hw_stats_struct(cnts->names,
5604 					  cnts->num_q_counters +
5605 					  cnts->num_cong_counters +
5606 					  cnts->num_ext_ppcnt_counters,
5607 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
5608 }
5609 
5610 static int mlx5_ib_query_q_counters(struct mlx5_core_dev *mdev,
5611 				    const struct mlx5_ib_counters *cnts,
5612 				    struct rdma_hw_stats *stats,
5613 				    u16 set_id)
5614 {
5615 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
5616 	void *out;
5617 	__be32 val;
5618 	int ret, i;
5619 
5620 	out = kvzalloc(outlen, GFP_KERNEL);
5621 	if (!out)
5622 		return -ENOMEM;
5623 
5624 	ret = mlx5_core_query_q_counter(mdev, set_id, 0, out, outlen);
5625 	if (ret)
5626 		goto free;
5627 
5628 	for (i = 0; i < cnts->num_q_counters; i++) {
5629 		val = *(__be32 *)(out + cnts->offsets[i]);
5630 		stats->value[i] = (u64)be32_to_cpu(val);
5631 	}
5632 
5633 free:
5634 	kvfree(out);
5635 	return ret;
5636 }
5637 
5638 static int mlx5_ib_query_ext_ppcnt_counters(struct mlx5_ib_dev *dev,
5639 					    const struct mlx5_ib_counters *cnts,
5640 					    struct rdma_hw_stats *stats)
5641 {
5642 	int offset = cnts->num_q_counters + cnts->num_cong_counters;
5643 	int sz = MLX5_ST_SZ_BYTES(ppcnt_reg);
5644 	int ret, i;
5645 	void *out;
5646 
5647 	out = kvzalloc(sz, GFP_KERNEL);
5648 	if (!out)
5649 		return -ENOMEM;
5650 
5651 	ret = mlx5_cmd_query_ext_ppcnt_counters(dev->mdev, out);
5652 	if (ret)
5653 		goto free;
5654 
5655 	for (i = 0; i < cnts->num_ext_ppcnt_counters; i++)
5656 		stats->value[i + offset] =
5657 			be64_to_cpup((__be64 *)(out +
5658 				    cnts->offsets[i + offset]));
5659 free:
5660 	kvfree(out);
5661 	return ret;
5662 }
5663 
5664 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
5665 				struct rdma_hw_stats *stats,
5666 				u8 port_num, int index)
5667 {
5668 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5669 	const struct mlx5_ib_counters *cnts = get_counters(dev, port_num - 1);
5670 	struct mlx5_core_dev *mdev;
5671 	int ret, num_counters;
5672 	u8 mdev_port_num;
5673 
5674 	if (!stats)
5675 		return -EINVAL;
5676 
5677 	num_counters = cnts->num_q_counters +
5678 		       cnts->num_cong_counters +
5679 		       cnts->num_ext_ppcnt_counters;
5680 
5681 	/* q_counters are per IB device, query the master mdev */
5682 	ret = mlx5_ib_query_q_counters(dev->mdev, cnts, stats, cnts->set_id);
5683 	if (ret)
5684 		return ret;
5685 
5686 	if (MLX5_CAP_PCAM_FEATURE(dev->mdev, rx_icrc_encapsulated_counter)) {
5687 		ret =  mlx5_ib_query_ext_ppcnt_counters(dev, cnts, stats);
5688 		if (ret)
5689 			return ret;
5690 	}
5691 
5692 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
5693 		mdev = mlx5_ib_get_native_port_mdev(dev, port_num,
5694 						    &mdev_port_num);
5695 		if (!mdev) {
5696 			/* If port is not affiliated yet, its in down state
5697 			 * which doesn't have any counters yet, so it would be
5698 			 * zero. So no need to read from the HCA.
5699 			 */
5700 			goto done;
5701 		}
5702 		ret = mlx5_lag_query_cong_counters(dev->mdev,
5703 						   stats->value +
5704 						   cnts->num_q_counters,
5705 						   cnts->num_cong_counters,
5706 						   cnts->offsets +
5707 						   cnts->num_q_counters);
5708 
5709 		mlx5_ib_put_native_port_mdev(dev, port_num);
5710 		if (ret)
5711 			return ret;
5712 	}
5713 
5714 done:
5715 	return num_counters;
5716 }
5717 
5718 static struct rdma_hw_stats *
5719 mlx5_ib_counter_alloc_stats(struct rdma_counter *counter)
5720 {
5721 	struct mlx5_ib_dev *dev = to_mdev(counter->device);
5722 	const struct mlx5_ib_counters *cnts =
5723 		get_counters(dev, counter->port - 1);
5724 
5725 	/* Q counters are in the beginning of all counters */
5726 	return rdma_alloc_hw_stats_struct(cnts->names,
5727 					  cnts->num_q_counters,
5728 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
5729 }
5730 
5731 static int mlx5_ib_counter_update_stats(struct rdma_counter *counter)
5732 {
5733 	struct mlx5_ib_dev *dev = to_mdev(counter->device);
5734 	const struct mlx5_ib_counters *cnts =
5735 		get_counters(dev, counter->port - 1);
5736 
5737 	return mlx5_ib_query_q_counters(dev->mdev, cnts,
5738 					counter->stats, counter->id);
5739 }
5740 
5741 static int mlx5_ib_counter_bind_qp(struct rdma_counter *counter,
5742 				   struct ib_qp *qp)
5743 {
5744 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5745 	u16 cnt_set_id = 0;
5746 	int err;
5747 
5748 	if (!counter->id) {
5749 		err = mlx5_cmd_alloc_q_counter(dev->mdev,
5750 					       &cnt_set_id,
5751 					       MLX5_SHARED_RESOURCE_UID);
5752 		if (err)
5753 			return err;
5754 		counter->id = cnt_set_id;
5755 	}
5756 
5757 	err = mlx5_ib_qp_set_counter(qp, counter);
5758 	if (err)
5759 		goto fail_set_counter;
5760 
5761 	return 0;
5762 
5763 fail_set_counter:
5764 	mlx5_core_dealloc_q_counter(dev->mdev, cnt_set_id);
5765 	counter->id = 0;
5766 
5767 	return err;
5768 }
5769 
5770 static int mlx5_ib_counter_unbind_qp(struct ib_qp *qp)
5771 {
5772 	return mlx5_ib_qp_set_counter(qp, NULL);
5773 }
5774 
5775 static int mlx5_ib_counter_dealloc(struct rdma_counter *counter)
5776 {
5777 	struct mlx5_ib_dev *dev = to_mdev(counter->device);
5778 
5779 	return mlx5_core_dealloc_q_counter(dev->mdev, counter->id);
5780 }
5781 
5782 static int mlx5_ib_rn_get_params(struct ib_device *device, u8 port_num,
5783 				 enum rdma_netdev_t type,
5784 				 struct rdma_netdev_alloc_params *params)
5785 {
5786 	if (type != RDMA_NETDEV_IPOIB)
5787 		return -EOPNOTSUPP;
5788 
5789 	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
5790 }
5791 
5792 static void delay_drop_debugfs_cleanup(struct mlx5_ib_dev *dev)
5793 {
5794 	if (!dev->delay_drop.dir_debugfs)
5795 		return;
5796 	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
5797 	dev->delay_drop.dir_debugfs = NULL;
5798 }
5799 
5800 static void cancel_delay_drop(struct mlx5_ib_dev *dev)
5801 {
5802 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
5803 		return;
5804 
5805 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
5806 	delay_drop_debugfs_cleanup(dev);
5807 }
5808 
5809 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
5810 				       size_t count, loff_t *pos)
5811 {
5812 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
5813 	char lbuf[20];
5814 	int len;
5815 
5816 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
5817 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
5818 }
5819 
5820 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
5821 					size_t count, loff_t *pos)
5822 {
5823 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
5824 	u32 timeout;
5825 	u32 var;
5826 
5827 	if (kstrtouint_from_user(buf, count, 0, &var))
5828 		return -EFAULT;
5829 
5830 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
5831 			1000);
5832 	if (timeout != var)
5833 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
5834 			    timeout);
5835 
5836 	delay_drop->timeout = timeout;
5837 
5838 	return count;
5839 }
5840 
5841 static const struct file_operations fops_delay_drop_timeout = {
5842 	.owner	= THIS_MODULE,
5843 	.open	= simple_open,
5844 	.write	= delay_drop_timeout_write,
5845 	.read	= delay_drop_timeout_read,
5846 };
5847 
5848 static void delay_drop_debugfs_init(struct mlx5_ib_dev *dev)
5849 {
5850 	struct dentry *root;
5851 
5852 	if (!mlx5_debugfs_root)
5853 		return;
5854 
5855 	root = debugfs_create_dir("delay_drop", dev->mdev->priv.dbg_root);
5856 	dev->delay_drop.dir_debugfs = root;
5857 
5858 	debugfs_create_atomic_t("num_timeout_events", 0400, root,
5859 				&dev->delay_drop.events_cnt);
5860 	debugfs_create_atomic_t("num_rqs", 0400, root,
5861 				&dev->delay_drop.rqs_cnt);
5862 	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
5863 			    &fops_delay_drop_timeout);
5864 }
5865 
5866 static void init_delay_drop(struct mlx5_ib_dev *dev)
5867 {
5868 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
5869 		return;
5870 
5871 	mutex_init(&dev->delay_drop.lock);
5872 	dev->delay_drop.dev = dev;
5873 	dev->delay_drop.activate = false;
5874 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
5875 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
5876 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
5877 	atomic_set(&dev->delay_drop.events_cnt, 0);
5878 
5879 	delay_drop_debugfs_init(dev);
5880 }
5881 
5882 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
5883 				      struct mlx5_ib_multiport_info *mpi)
5884 {
5885 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
5886 	struct mlx5_ib_port *port = &ibdev->port[port_num];
5887 	int comps;
5888 	int err;
5889 	int i;
5890 
5891 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
5892 
5893 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
5894 
5895 	spin_lock(&port->mp.mpi_lock);
5896 	if (!mpi->ibdev) {
5897 		spin_unlock(&port->mp.mpi_lock);
5898 		return;
5899 	}
5900 
5901 	mpi->ibdev = NULL;
5902 
5903 	spin_unlock(&port->mp.mpi_lock);
5904 	if (mpi->mdev_events.notifier_call)
5905 		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
5906 	mpi->mdev_events.notifier_call = NULL;
5907 	mlx5_remove_netdev_notifier(ibdev, port_num);
5908 	spin_lock(&port->mp.mpi_lock);
5909 
5910 	comps = mpi->mdev_refcnt;
5911 	if (comps) {
5912 		mpi->unaffiliate = true;
5913 		init_completion(&mpi->unref_comp);
5914 		spin_unlock(&port->mp.mpi_lock);
5915 
5916 		for (i = 0; i < comps; i++)
5917 			wait_for_completion(&mpi->unref_comp);
5918 
5919 		spin_lock(&port->mp.mpi_lock);
5920 		mpi->unaffiliate = false;
5921 	}
5922 
5923 	port->mp.mpi = NULL;
5924 
5925 	list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
5926 
5927 	spin_unlock(&port->mp.mpi_lock);
5928 
5929 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
5930 
5931 	mlx5_ib_dbg(ibdev, "unaffiliated port %d\n", port_num + 1);
5932 	/* Log an error, still needed to cleanup the pointers and add
5933 	 * it back to the list.
5934 	 */
5935 	if (err)
5936 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
5937 			    port_num + 1);
5938 
5939 	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
5940 }
5941 
5942 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
5943 				    struct mlx5_ib_multiport_info *mpi)
5944 {
5945 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
5946 	int err;
5947 
5948 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
5949 
5950 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
5951 	if (ibdev->port[port_num].mp.mpi) {
5952 		mlx5_ib_dbg(ibdev, "port %d already affiliated.\n",
5953 			    port_num + 1);
5954 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
5955 		return false;
5956 	}
5957 
5958 	ibdev->port[port_num].mp.mpi = mpi;
5959 	mpi->ibdev = ibdev;
5960 	mpi->mdev_events.notifier_call = NULL;
5961 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
5962 
5963 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
5964 	if (err)
5965 		goto unbind;
5966 
5967 	err = get_port_caps(ibdev, mlx5_core_native_port_num(mpi->mdev));
5968 	if (err)
5969 		goto unbind;
5970 
5971 	err = mlx5_add_netdev_notifier(ibdev, port_num);
5972 	if (err) {
5973 		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
5974 			    port_num + 1);
5975 		goto unbind;
5976 	}
5977 
5978 	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
5979 	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
5980 
5981 	mlx5_ib_init_cong_debugfs(ibdev, port_num);
5982 
5983 	return true;
5984 
5985 unbind:
5986 	mlx5_ib_unbind_slave_port(ibdev, mpi);
5987 	return false;
5988 }
5989 
5990 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
5991 {
5992 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5993 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
5994 							  port_num + 1);
5995 	struct mlx5_ib_multiport_info *mpi;
5996 	int err;
5997 	int i;
5998 
5999 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
6000 		return 0;
6001 
6002 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
6003 						     &dev->sys_image_guid);
6004 	if (err)
6005 		return err;
6006 
6007 	err = mlx5_nic_vport_enable_roce(dev->mdev);
6008 	if (err)
6009 		return err;
6010 
6011 	mutex_lock(&mlx5_ib_multiport_mutex);
6012 	for (i = 0; i < dev->num_ports; i++) {
6013 		bool bound = false;
6014 
6015 		/* build a stub multiport info struct for the native port. */
6016 		if (i == port_num) {
6017 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
6018 			if (!mpi) {
6019 				mutex_unlock(&mlx5_ib_multiport_mutex);
6020 				mlx5_nic_vport_disable_roce(dev->mdev);
6021 				return -ENOMEM;
6022 			}
6023 
6024 			mpi->is_master = true;
6025 			mpi->mdev = dev->mdev;
6026 			mpi->sys_image_guid = dev->sys_image_guid;
6027 			dev->port[i].mp.mpi = mpi;
6028 			mpi->ibdev = dev;
6029 			mpi = NULL;
6030 			continue;
6031 		}
6032 
6033 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
6034 				    list) {
6035 			if (dev->sys_image_guid == mpi->sys_image_guid &&
6036 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
6037 				bound = mlx5_ib_bind_slave_port(dev, mpi);
6038 			}
6039 
6040 			if (bound) {
6041 				dev_dbg(mpi->mdev->device,
6042 					"removing port from unaffiliated list.\n");
6043 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
6044 				list_del(&mpi->list);
6045 				break;
6046 			}
6047 		}
6048 		if (!bound) {
6049 			get_port_caps(dev, i + 1);
6050 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
6051 				    i + 1);
6052 		}
6053 	}
6054 
6055 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
6056 	mutex_unlock(&mlx5_ib_multiport_mutex);
6057 	return err;
6058 }
6059 
6060 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
6061 {
6062 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
6063 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
6064 							  port_num + 1);
6065 	int i;
6066 
6067 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
6068 		return;
6069 
6070 	mutex_lock(&mlx5_ib_multiport_mutex);
6071 	for (i = 0; i < dev->num_ports; i++) {
6072 		if (dev->port[i].mp.mpi) {
6073 			/* Destroy the native port stub */
6074 			if (i == port_num) {
6075 				kfree(dev->port[i].mp.mpi);
6076 				dev->port[i].mp.mpi = NULL;
6077 			} else {
6078 				mlx5_ib_dbg(dev, "unbinding port_num: %d\n", i + 1);
6079 				mlx5_ib_unbind_slave_port(dev, dev->port[i].mp.mpi);
6080 			}
6081 		}
6082 	}
6083 
6084 	mlx5_ib_dbg(dev, "removing from devlist\n");
6085 	list_del(&dev->ib_dev_list);
6086 	mutex_unlock(&mlx5_ib_multiport_mutex);
6087 
6088 	mlx5_nic_vport_disable_roce(dev->mdev);
6089 }
6090 
6091 static int var_obj_cleanup(struct ib_uobject *uobject,
6092 			   enum rdma_remove_reason why,
6093 			   struct uverbs_attr_bundle *attrs)
6094 {
6095 	struct mlx5_user_mmap_entry *obj = uobject->object;
6096 
6097 	rdma_user_mmap_entry_remove(&obj->rdma_entry);
6098 	return 0;
6099 }
6100 
6101 static struct mlx5_user_mmap_entry *
6102 alloc_var_entry(struct mlx5_ib_ucontext *c)
6103 {
6104 	struct mlx5_user_mmap_entry *entry;
6105 	struct mlx5_var_table *var_table;
6106 	u32 page_idx;
6107 	int err;
6108 
6109 	var_table = &to_mdev(c->ibucontext.device)->var_table;
6110 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
6111 	if (!entry)
6112 		return ERR_PTR(-ENOMEM);
6113 
6114 	mutex_lock(&var_table->bitmap_lock);
6115 	page_idx = find_first_zero_bit(var_table->bitmap,
6116 				       var_table->num_var_hw_entries);
6117 	if (page_idx >= var_table->num_var_hw_entries) {
6118 		err = -ENOSPC;
6119 		mutex_unlock(&var_table->bitmap_lock);
6120 		goto end;
6121 	}
6122 
6123 	set_bit(page_idx, var_table->bitmap);
6124 	mutex_unlock(&var_table->bitmap_lock);
6125 
6126 	entry->address = var_table->hw_start_addr +
6127 				(page_idx * var_table->stride_size);
6128 	entry->page_idx = page_idx;
6129 	entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
6130 
6131 	err = rdma_user_mmap_entry_insert_range(
6132 		&c->ibucontext, &entry->rdma_entry, var_table->stride_size,
6133 		MLX5_IB_MMAP_OFFSET_START << 16,
6134 		(MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1);
6135 	if (err)
6136 		goto err_insert;
6137 
6138 	return entry;
6139 
6140 err_insert:
6141 	mutex_lock(&var_table->bitmap_lock);
6142 	clear_bit(page_idx, var_table->bitmap);
6143 	mutex_unlock(&var_table->bitmap_lock);
6144 end:
6145 	kfree(entry);
6146 	return ERR_PTR(err);
6147 }
6148 
6149 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
6150 	struct uverbs_attr_bundle *attrs)
6151 {
6152 	struct ib_uobject *uobj = uverbs_attr_get_uobject(
6153 		attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
6154 	struct mlx5_ib_ucontext *c;
6155 	struct mlx5_user_mmap_entry *entry;
6156 	u64 mmap_offset;
6157 	u32 length;
6158 	int err;
6159 
6160 	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
6161 	if (IS_ERR(c))
6162 		return PTR_ERR(c);
6163 
6164 	entry = alloc_var_entry(c);
6165 	if (IS_ERR(entry))
6166 		return PTR_ERR(entry);
6167 
6168 	mmap_offset = mlx5_entry_to_mmap_offset(entry);
6169 	length = entry->rdma_entry.npages * PAGE_SIZE;
6170 	uobj->object = entry;
6171 
6172 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
6173 			     &mmap_offset, sizeof(mmap_offset));
6174 	if (err)
6175 		goto err;
6176 
6177 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
6178 			     &entry->page_idx, sizeof(entry->page_idx));
6179 	if (err)
6180 		goto err;
6181 
6182 	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
6183 			     &length, sizeof(length));
6184 	if (err)
6185 		goto err;
6186 
6187 	return 0;
6188 
6189 err:
6190 	rdma_user_mmap_entry_remove(&entry->rdma_entry);
6191 	return err;
6192 }
6193 
6194 DECLARE_UVERBS_NAMED_METHOD(
6195 	MLX5_IB_METHOD_VAR_OBJ_ALLOC,
6196 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
6197 			MLX5_IB_OBJECT_VAR,
6198 			UVERBS_ACCESS_NEW,
6199 			UA_MANDATORY),
6200 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
6201 			   UVERBS_ATTR_TYPE(u32),
6202 			   UA_MANDATORY),
6203 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
6204 			   UVERBS_ATTR_TYPE(u32),
6205 			   UA_MANDATORY),
6206 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
6207 			    UVERBS_ATTR_TYPE(u64),
6208 			    UA_MANDATORY));
6209 
6210 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
6211 	MLX5_IB_METHOD_VAR_OBJ_DESTROY,
6212 	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
6213 			MLX5_IB_OBJECT_VAR,
6214 			UVERBS_ACCESS_DESTROY,
6215 			UA_MANDATORY));
6216 
6217 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
6218 			    UVERBS_TYPE_ALLOC_IDR(var_obj_cleanup),
6219 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
6220 			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
6221 
6222 static bool var_is_supported(struct ib_device *device)
6223 {
6224 	struct mlx5_ib_dev *dev = to_mdev(device);
6225 
6226 	return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
6227 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
6228 }
6229 
6230 ADD_UVERBS_ATTRIBUTES_SIMPLE(
6231 	mlx5_ib_dm,
6232 	UVERBS_OBJECT_DM,
6233 	UVERBS_METHOD_DM_ALLOC,
6234 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
6235 			    UVERBS_ATTR_TYPE(u64),
6236 			    UA_MANDATORY),
6237 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
6238 			    UVERBS_ATTR_TYPE(u16),
6239 			    UA_OPTIONAL),
6240 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_ALLOC_DM_REQ_TYPE,
6241 			     enum mlx5_ib_uapi_dm_type,
6242 			     UA_OPTIONAL));
6243 
6244 ADD_UVERBS_ATTRIBUTES_SIMPLE(
6245 	mlx5_ib_flow_action,
6246 	UVERBS_OBJECT_FLOW_ACTION,
6247 	UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
6248 	UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
6249 			     enum mlx5_ib_uapi_flow_action_flags));
6250 
6251 static const struct uapi_definition mlx5_ib_defs[] = {
6252 	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
6253 	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
6254 
6255 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_FLOW_ACTION,
6256 				&mlx5_ib_flow_action),
6257 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DM, &mlx5_ib_dm),
6258 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
6259 				UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
6260 	{}
6261 };
6262 
6263 static int mlx5_ib_read_counters(struct ib_counters *counters,
6264 				 struct ib_counters_read_attr *read_attr,
6265 				 struct uverbs_attr_bundle *attrs)
6266 {
6267 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
6268 	struct mlx5_read_counters_attr mread_attr = {};
6269 	struct mlx5_ib_flow_counters_desc *desc;
6270 	int ret, i;
6271 
6272 	mutex_lock(&mcounters->mcntrs_mutex);
6273 	if (mcounters->cntrs_max_index > read_attr->ncounters) {
6274 		ret = -EINVAL;
6275 		goto err_bound;
6276 	}
6277 
6278 	mread_attr.out = kcalloc(mcounters->counters_num, sizeof(u64),
6279 				 GFP_KERNEL);
6280 	if (!mread_attr.out) {
6281 		ret = -ENOMEM;
6282 		goto err_bound;
6283 	}
6284 
6285 	mread_attr.hw_cntrs_hndl = mcounters->hw_cntrs_hndl;
6286 	mread_attr.flags = read_attr->flags;
6287 	ret = mcounters->read_counters(counters->device, &mread_attr);
6288 	if (ret)
6289 		goto err_read;
6290 
6291 	/* do the pass over the counters data array to assign according to the
6292 	 * descriptions and indexing pairs
6293 	 */
6294 	desc = mcounters->counters_data;
6295 	for (i = 0; i < mcounters->ncounters; i++)
6296 		read_attr->counters_buff[desc[i].index] += mread_attr.out[desc[i].description];
6297 
6298 err_read:
6299 	kfree(mread_attr.out);
6300 err_bound:
6301 	mutex_unlock(&mcounters->mcntrs_mutex);
6302 	return ret;
6303 }
6304 
6305 static int mlx5_ib_destroy_counters(struct ib_counters *counters)
6306 {
6307 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
6308 
6309 	counters_clear_description(counters);
6310 	if (mcounters->hw_cntrs_hndl)
6311 		mlx5_fc_destroy(to_mdev(counters->device)->mdev,
6312 				mcounters->hw_cntrs_hndl);
6313 
6314 	kfree(mcounters);
6315 
6316 	return 0;
6317 }
6318 
6319 static struct ib_counters *mlx5_ib_create_counters(struct ib_device *device,
6320 						   struct uverbs_attr_bundle *attrs)
6321 {
6322 	struct mlx5_ib_mcounters *mcounters;
6323 
6324 	mcounters = kzalloc(sizeof(*mcounters), GFP_KERNEL);
6325 	if (!mcounters)
6326 		return ERR_PTR(-ENOMEM);
6327 
6328 	mutex_init(&mcounters->mcntrs_mutex);
6329 
6330 	return &mcounters->ibcntrs;
6331 }
6332 
6333 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
6334 {
6335 	mlx5_ib_cleanup_multiport_master(dev);
6336 	WARN_ON(!xa_empty(&dev->odp_mkeys));
6337 	cleanup_srcu_struct(&dev->odp_srcu);
6338 
6339 	WARN_ON(!xa_empty(&dev->sig_mrs));
6340 	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
6341 }
6342 
6343 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
6344 {
6345 	struct mlx5_core_dev *mdev = dev->mdev;
6346 	int err;
6347 	int i;
6348 
6349 	for (i = 0; i < dev->num_ports; i++) {
6350 		spin_lock_init(&dev->port[i].mp.mpi_lock);
6351 		rwlock_init(&dev->port[i].roce.netdev_lock);
6352 		dev->port[i].roce.dev = dev;
6353 		dev->port[i].roce.native_port_num = i + 1;
6354 		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
6355 	}
6356 
6357 	mlx5_ib_internal_fill_odp_caps(dev);
6358 
6359 	err = mlx5_ib_init_multiport_master(dev);
6360 	if (err)
6361 		return err;
6362 
6363 	err = set_has_smi_cap(dev);
6364 	if (err)
6365 		return err;
6366 
6367 	if (!mlx5_core_mp_enabled(mdev)) {
6368 		for (i = 1; i <= dev->num_ports; i++) {
6369 			err = get_port_caps(dev, i);
6370 			if (err)
6371 				break;
6372 		}
6373 	} else {
6374 		err = get_port_caps(dev, mlx5_core_native_port_num(mdev));
6375 	}
6376 	if (err)
6377 		goto err_mp;
6378 
6379 	if (mlx5_use_mad_ifc(dev))
6380 		get_ext_port_caps(dev);
6381 
6382 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
6383 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
6384 	dev->ib_dev.phys_port_cnt	= dev->num_ports;
6385 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_count(mdev);
6386 	dev->ib_dev.dev.parent		= mdev->device;
6387 
6388 	mutex_init(&dev->cap_mask_mutex);
6389 	INIT_LIST_HEAD(&dev->qp_list);
6390 	spin_lock_init(&dev->reset_flow_resource_lock);
6391 	xa_init(&dev->odp_mkeys);
6392 	xa_init(&dev->sig_mrs);
6393 
6394 	spin_lock_init(&dev->dm.lock);
6395 	dev->dm.dev = mdev;
6396 
6397 	err = init_srcu_struct(&dev->odp_srcu);
6398 	if (err)
6399 		goto err_mp;
6400 
6401 	return 0;
6402 
6403 err_mp:
6404 	mlx5_ib_cleanup_multiport_master(dev);
6405 
6406 	return -ENOMEM;
6407 }
6408 
6409 static int mlx5_ib_stage_flow_db_init(struct mlx5_ib_dev *dev)
6410 {
6411 	dev->flow_db = kzalloc(sizeof(*dev->flow_db), GFP_KERNEL);
6412 
6413 	if (!dev->flow_db)
6414 		return -ENOMEM;
6415 
6416 	mutex_init(&dev->flow_db->lock);
6417 
6418 	return 0;
6419 }
6420 
6421 static void mlx5_ib_stage_flow_db_cleanup(struct mlx5_ib_dev *dev)
6422 {
6423 	kfree(dev->flow_db);
6424 }
6425 
6426 static const struct ib_device_ops mlx5_ib_dev_ops = {
6427 	.owner = THIS_MODULE,
6428 	.driver_id = RDMA_DRIVER_MLX5,
6429 	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
6430 
6431 	.add_gid = mlx5_ib_add_gid,
6432 	.alloc_mr = mlx5_ib_alloc_mr,
6433 	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
6434 	.alloc_pd = mlx5_ib_alloc_pd,
6435 	.alloc_ucontext = mlx5_ib_alloc_ucontext,
6436 	.attach_mcast = mlx5_ib_mcg_attach,
6437 	.check_mr_status = mlx5_ib_check_mr_status,
6438 	.create_ah = mlx5_ib_create_ah,
6439 	.create_counters = mlx5_ib_create_counters,
6440 	.create_cq = mlx5_ib_create_cq,
6441 	.create_flow = mlx5_ib_create_flow,
6442 	.create_qp = mlx5_ib_create_qp,
6443 	.create_srq = mlx5_ib_create_srq,
6444 	.dealloc_pd = mlx5_ib_dealloc_pd,
6445 	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
6446 	.del_gid = mlx5_ib_del_gid,
6447 	.dereg_mr = mlx5_ib_dereg_mr,
6448 	.destroy_ah = mlx5_ib_destroy_ah,
6449 	.destroy_counters = mlx5_ib_destroy_counters,
6450 	.destroy_cq = mlx5_ib_destroy_cq,
6451 	.destroy_flow = mlx5_ib_destroy_flow,
6452 	.destroy_flow_action = mlx5_ib_destroy_flow_action,
6453 	.destroy_qp = mlx5_ib_destroy_qp,
6454 	.destroy_srq = mlx5_ib_destroy_srq,
6455 	.detach_mcast = mlx5_ib_mcg_detach,
6456 	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
6457 	.drain_rq = mlx5_ib_drain_rq,
6458 	.drain_sq = mlx5_ib_drain_sq,
6459 	.enable_driver = mlx5_ib_enable_driver,
6460 	.fill_res_entry = mlx5_ib_fill_res_entry,
6461 	.fill_stat_entry = mlx5_ib_fill_stat_entry,
6462 	.get_dev_fw_str = get_dev_fw_str,
6463 	.get_dma_mr = mlx5_ib_get_dma_mr,
6464 	.get_link_layer = mlx5_ib_port_link_layer,
6465 	.map_mr_sg = mlx5_ib_map_mr_sg,
6466 	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
6467 	.mmap = mlx5_ib_mmap,
6468 	.mmap_free = mlx5_ib_mmap_free,
6469 	.modify_cq = mlx5_ib_modify_cq,
6470 	.modify_device = mlx5_ib_modify_device,
6471 	.modify_port = mlx5_ib_modify_port,
6472 	.modify_qp = mlx5_ib_modify_qp,
6473 	.modify_srq = mlx5_ib_modify_srq,
6474 	.poll_cq = mlx5_ib_poll_cq,
6475 	.post_recv = mlx5_ib_post_recv,
6476 	.post_send = mlx5_ib_post_send,
6477 	.post_srq_recv = mlx5_ib_post_srq_recv,
6478 	.process_mad = mlx5_ib_process_mad,
6479 	.query_ah = mlx5_ib_query_ah,
6480 	.query_device = mlx5_ib_query_device,
6481 	.query_gid = mlx5_ib_query_gid,
6482 	.query_pkey = mlx5_ib_query_pkey,
6483 	.query_qp = mlx5_ib_query_qp,
6484 	.query_srq = mlx5_ib_query_srq,
6485 	.read_counters = mlx5_ib_read_counters,
6486 	.reg_user_mr = mlx5_ib_reg_user_mr,
6487 	.req_notify_cq = mlx5_ib_arm_cq,
6488 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
6489 	.resize_cq = mlx5_ib_resize_cq,
6490 
6491 	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
6492 	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
6493 	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
6494 	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
6495 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
6496 };
6497 
6498 static const struct ib_device_ops mlx5_ib_dev_flow_ipsec_ops = {
6499 	.create_flow_action_esp = mlx5_ib_create_flow_action_esp,
6500 	.modify_flow_action_esp = mlx5_ib_modify_flow_action_esp,
6501 };
6502 
6503 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
6504 	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
6505 };
6506 
6507 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
6508 	.get_vf_config = mlx5_ib_get_vf_config,
6509 	.get_vf_guid = mlx5_ib_get_vf_guid,
6510 	.get_vf_stats = mlx5_ib_get_vf_stats,
6511 	.set_vf_guid = mlx5_ib_set_vf_guid,
6512 	.set_vf_link_state = mlx5_ib_set_vf_link_state,
6513 };
6514 
6515 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
6516 	.alloc_mw = mlx5_ib_alloc_mw,
6517 	.dealloc_mw = mlx5_ib_dealloc_mw,
6518 };
6519 
6520 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
6521 	.alloc_xrcd = mlx5_ib_alloc_xrcd,
6522 	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
6523 };
6524 
6525 static const struct ib_device_ops mlx5_ib_dev_dm_ops = {
6526 	.alloc_dm = mlx5_ib_alloc_dm,
6527 	.dealloc_dm = mlx5_ib_dealloc_dm,
6528 	.reg_dm_mr = mlx5_ib_reg_dm_mr,
6529 };
6530 
6531 static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
6532 {
6533 	struct mlx5_core_dev *mdev = dev->mdev;
6534 	struct mlx5_var_table *var_table = &dev->var_table;
6535 	u8 log_doorbell_bar_size;
6536 	u8 log_doorbell_stride;
6537 	u64 bar_size;
6538 
6539 	log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
6540 					log_doorbell_bar_size);
6541 	log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
6542 					log_doorbell_stride);
6543 	var_table->hw_start_addr = dev->mdev->bar_addr +
6544 				MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
6545 					doorbell_bar_offset);
6546 	bar_size = (1ULL << log_doorbell_bar_size) * 4096;
6547 	var_table->stride_size = 1ULL << log_doorbell_stride;
6548 	var_table->num_var_hw_entries = div64_u64(bar_size, var_table->stride_size);
6549 	mutex_init(&var_table->bitmap_lock);
6550 	var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
6551 					  GFP_KERNEL);
6552 	return (var_table->bitmap) ? 0 : -ENOMEM;
6553 }
6554 
6555 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
6556 {
6557 	bitmap_free(dev->var_table.bitmap);
6558 }
6559 
6560 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
6561 {
6562 	struct mlx5_core_dev *mdev = dev->mdev;
6563 	int err;
6564 
6565 	dev->ib_dev.uverbs_cmd_mask	=
6566 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
6567 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
6568 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
6569 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
6570 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
6571 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
6572 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
6573 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
6574 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
6575 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
6576 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
6577 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
6578 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
6579 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
6580 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
6581 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
6582 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
6583 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
6584 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
6585 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
6586 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
6587 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
6588 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
6589 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
6590 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
6591 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
6592 	dev->ib_dev.uverbs_ex_cmd_mask =
6593 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
6594 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
6595 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP)	|
6596 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP)	|
6597 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_CQ)	|
6598 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW)	|
6599 		(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW);
6600 
6601 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
6602 	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
6603 		ib_set_device_ops(&dev->ib_dev,
6604 				  &mlx5_ib_dev_ipoib_enhanced_ops);
6605 
6606 	if (mlx5_core_is_pf(mdev))
6607 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
6608 
6609 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
6610 
6611 	if (MLX5_CAP_GEN(mdev, imaicl)) {
6612 		dev->ib_dev.uverbs_cmd_mask |=
6613 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
6614 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
6615 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
6616 	}
6617 
6618 	if (MLX5_CAP_GEN(mdev, xrc)) {
6619 		dev->ib_dev.uverbs_cmd_mask |=
6620 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
6621 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
6622 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
6623 	}
6624 
6625 	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
6626 	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
6627 	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
6628 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
6629 
6630 	if (mlx5_accel_ipsec_device_caps(dev->mdev) &
6631 	    MLX5_ACCEL_IPSEC_CAP_DEVICE)
6632 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_flow_ipsec_ops);
6633 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
6634 
6635 	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
6636 		dev->ib_dev.driver_def = mlx5_ib_defs;
6637 
6638 	err = init_node_data(dev);
6639 	if (err)
6640 		return err;
6641 
6642 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
6643 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
6644 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
6645 		mutex_init(&dev->lb.mutex);
6646 
6647 	if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
6648 			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
6649 		err = mlx5_ib_init_var_table(dev);
6650 		if (err)
6651 			return err;
6652 	}
6653 
6654 	dev->ib_dev.use_cq_dim = true;
6655 
6656 	return 0;
6657 }
6658 
6659 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
6660 	.get_port_immutable = mlx5_port_immutable,
6661 	.query_port = mlx5_ib_query_port,
6662 };
6663 
6664 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
6665 {
6666 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
6667 	return 0;
6668 }
6669 
6670 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
6671 	.get_port_immutable = mlx5_port_rep_immutable,
6672 	.query_port = mlx5_ib_rep_query_port,
6673 };
6674 
6675 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
6676 {
6677 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
6678 	return 0;
6679 }
6680 
6681 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
6682 	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
6683 	.create_wq = mlx5_ib_create_wq,
6684 	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
6685 	.destroy_wq = mlx5_ib_destroy_wq,
6686 	.get_netdev = mlx5_ib_get_netdev,
6687 	.modify_wq = mlx5_ib_modify_wq,
6688 };
6689 
6690 static int mlx5_ib_stage_common_roce_init(struct mlx5_ib_dev *dev)
6691 {
6692 	u8 port_num;
6693 
6694 	dev->ib_dev.uverbs_ex_cmd_mask |=
6695 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
6696 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
6697 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
6698 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
6699 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
6700 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
6701 
6702 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
6703 
6704 	/* Register only for native ports */
6705 	return mlx5_add_netdev_notifier(dev, port_num);
6706 }
6707 
6708 static void mlx5_ib_stage_common_roce_cleanup(struct mlx5_ib_dev *dev)
6709 {
6710 	u8 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
6711 
6712 	mlx5_remove_netdev_notifier(dev, port_num);
6713 }
6714 
6715 static int mlx5_ib_stage_raw_eth_roce_init(struct mlx5_ib_dev *dev)
6716 {
6717 	struct mlx5_core_dev *mdev = dev->mdev;
6718 	enum rdma_link_layer ll;
6719 	int port_type_cap;
6720 	int err = 0;
6721 
6722 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
6723 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
6724 
6725 	if (ll == IB_LINK_LAYER_ETHERNET)
6726 		err = mlx5_ib_stage_common_roce_init(dev);
6727 
6728 	return err;
6729 }
6730 
6731 static void mlx5_ib_stage_raw_eth_roce_cleanup(struct mlx5_ib_dev *dev)
6732 {
6733 	mlx5_ib_stage_common_roce_cleanup(dev);
6734 }
6735 
6736 static int mlx5_ib_stage_roce_init(struct mlx5_ib_dev *dev)
6737 {
6738 	struct mlx5_core_dev *mdev = dev->mdev;
6739 	enum rdma_link_layer ll;
6740 	int port_type_cap;
6741 	int err;
6742 
6743 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
6744 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
6745 
6746 	if (ll == IB_LINK_LAYER_ETHERNET) {
6747 		err = mlx5_ib_stage_common_roce_init(dev);
6748 		if (err)
6749 			return err;
6750 
6751 		err = mlx5_enable_eth(dev);
6752 		if (err)
6753 			goto cleanup;
6754 	}
6755 
6756 	return 0;
6757 cleanup:
6758 	mlx5_ib_stage_common_roce_cleanup(dev);
6759 
6760 	return err;
6761 }
6762 
6763 static void mlx5_ib_stage_roce_cleanup(struct mlx5_ib_dev *dev)
6764 {
6765 	struct mlx5_core_dev *mdev = dev->mdev;
6766 	enum rdma_link_layer ll;
6767 	int port_type_cap;
6768 
6769 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
6770 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
6771 
6772 	if (ll == IB_LINK_LAYER_ETHERNET) {
6773 		mlx5_disable_eth(dev);
6774 		mlx5_ib_stage_common_roce_cleanup(dev);
6775 	}
6776 }
6777 
6778 static int mlx5_ib_stage_dev_res_init(struct mlx5_ib_dev *dev)
6779 {
6780 	return create_dev_resources(&dev->devr);
6781 }
6782 
6783 static void mlx5_ib_stage_dev_res_cleanup(struct mlx5_ib_dev *dev)
6784 {
6785 	destroy_dev_resources(&dev->devr);
6786 }
6787 
6788 static int mlx5_ib_stage_odp_init(struct mlx5_ib_dev *dev)
6789 {
6790 	return mlx5_ib_odp_init_one(dev);
6791 }
6792 
6793 static void mlx5_ib_stage_odp_cleanup(struct mlx5_ib_dev *dev)
6794 {
6795 	mlx5_ib_odp_cleanup_one(dev);
6796 }
6797 
6798 static const struct ib_device_ops mlx5_ib_dev_hw_stats_ops = {
6799 	.alloc_hw_stats = mlx5_ib_alloc_hw_stats,
6800 	.get_hw_stats = mlx5_ib_get_hw_stats,
6801 	.counter_bind_qp = mlx5_ib_counter_bind_qp,
6802 	.counter_unbind_qp = mlx5_ib_counter_unbind_qp,
6803 	.counter_dealloc = mlx5_ib_counter_dealloc,
6804 	.counter_alloc_stats = mlx5_ib_counter_alloc_stats,
6805 	.counter_update_stats = mlx5_ib_counter_update_stats,
6806 };
6807 
6808 static int mlx5_ib_stage_counters_init(struct mlx5_ib_dev *dev)
6809 {
6810 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
6811 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_hw_stats_ops);
6812 
6813 		return mlx5_ib_alloc_counters(dev);
6814 	}
6815 
6816 	return 0;
6817 }
6818 
6819 static void mlx5_ib_stage_counters_cleanup(struct mlx5_ib_dev *dev)
6820 {
6821 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
6822 		mlx5_ib_dealloc_counters(dev);
6823 }
6824 
6825 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
6826 {
6827 	mlx5_ib_init_cong_debugfs(dev,
6828 				  mlx5_core_native_port_num(dev->mdev) - 1);
6829 	return 0;
6830 }
6831 
6832 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
6833 {
6834 	mlx5_ib_cleanup_cong_debugfs(dev,
6835 				     mlx5_core_native_port_num(dev->mdev) - 1);
6836 }
6837 
6838 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
6839 {
6840 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
6841 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
6842 }
6843 
6844 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
6845 {
6846 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
6847 }
6848 
6849 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
6850 {
6851 	int err;
6852 
6853 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
6854 	if (err)
6855 		return err;
6856 
6857 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
6858 	if (err)
6859 		mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
6860 
6861 	return err;
6862 }
6863 
6864 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
6865 {
6866 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
6867 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
6868 }
6869 
6870 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
6871 {
6872 	const char *name;
6873 
6874 	rdma_set_device_sysfs_group(&dev->ib_dev, &mlx5_attr_group);
6875 	if (!mlx5_lag_is_roce(dev->mdev))
6876 		name = "mlx5_%d";
6877 	else
6878 		name = "mlx5_bond_%d";
6879 	return ib_register_device(&dev->ib_dev, name);
6880 }
6881 
6882 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
6883 {
6884 	destroy_umrc_res(dev);
6885 }
6886 
6887 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
6888 {
6889 	ib_unregister_device(&dev->ib_dev);
6890 }
6891 
6892 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
6893 {
6894 	return create_umr_res(dev);
6895 }
6896 
6897 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
6898 {
6899 	init_delay_drop(dev);
6900 
6901 	return 0;
6902 }
6903 
6904 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
6905 {
6906 	cancel_delay_drop(dev);
6907 }
6908 
6909 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
6910 {
6911 	dev->mdev_events.notifier_call = mlx5_ib_event;
6912 	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
6913 	return 0;
6914 }
6915 
6916 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
6917 {
6918 	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
6919 }
6920 
6921 static int mlx5_ib_stage_devx_init(struct mlx5_ib_dev *dev)
6922 {
6923 	int uid;
6924 
6925 	uid = mlx5_ib_devx_create(dev, false);
6926 	if (uid > 0) {
6927 		dev->devx_whitelist_uid = uid;
6928 		mlx5_ib_devx_init_event_table(dev);
6929 	}
6930 
6931 	return 0;
6932 }
6933 static void mlx5_ib_stage_devx_cleanup(struct mlx5_ib_dev *dev)
6934 {
6935 	if (dev->devx_whitelist_uid) {
6936 		mlx5_ib_devx_cleanup_event_table(dev);
6937 		mlx5_ib_devx_destroy(dev, dev->devx_whitelist_uid);
6938 	}
6939 }
6940 
6941 int mlx5_ib_enable_driver(struct ib_device *dev)
6942 {
6943 	struct mlx5_ib_dev *mdev = to_mdev(dev);
6944 	int ret;
6945 
6946 	ret = mlx5_ib_test_wc(mdev);
6947 	mlx5_ib_dbg(mdev, "Write-Combining %s",
6948 		    mdev->wc_support ? "supported" : "not supported");
6949 
6950 	return ret;
6951 }
6952 
6953 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
6954 		      const struct mlx5_ib_profile *profile,
6955 		      int stage)
6956 {
6957 	dev->ib_active = false;
6958 
6959 	/* Number of stages to cleanup */
6960 	while (stage) {
6961 		stage--;
6962 		if (profile->stage[stage].cleanup)
6963 			profile->stage[stage].cleanup(dev);
6964 	}
6965 
6966 	kfree(dev->port);
6967 	ib_dealloc_device(&dev->ib_dev);
6968 }
6969 
6970 void *__mlx5_ib_add(struct mlx5_ib_dev *dev,
6971 		    const struct mlx5_ib_profile *profile)
6972 {
6973 	int err;
6974 	int i;
6975 
6976 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
6977 		if (profile->stage[i].init) {
6978 			err = profile->stage[i].init(dev);
6979 			if (err)
6980 				goto err_out;
6981 		}
6982 	}
6983 
6984 	dev->profile = profile;
6985 	dev->ib_active = true;
6986 
6987 	return dev;
6988 
6989 err_out:
6990 	__mlx5_ib_remove(dev, profile, i);
6991 
6992 	return NULL;
6993 }
6994 
6995 static const struct mlx5_ib_profile pf_profile = {
6996 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
6997 		     mlx5_ib_stage_init_init,
6998 		     mlx5_ib_stage_init_cleanup),
6999 	STAGE_CREATE(MLX5_IB_STAGE_FLOW_DB,
7000 		     mlx5_ib_stage_flow_db_init,
7001 		     mlx5_ib_stage_flow_db_cleanup),
7002 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
7003 		     mlx5_ib_stage_caps_init,
7004 		     mlx5_ib_stage_caps_cleanup),
7005 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
7006 		     mlx5_ib_stage_non_default_cb,
7007 		     NULL),
7008 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
7009 		     mlx5_ib_stage_roce_init,
7010 		     mlx5_ib_stage_roce_cleanup),
7011 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
7012 		     mlx5_init_srq_table,
7013 		     mlx5_cleanup_srq_table),
7014 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
7015 		     mlx5_ib_stage_dev_res_init,
7016 		     mlx5_ib_stage_dev_res_cleanup),
7017 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
7018 		     mlx5_ib_stage_dev_notifier_init,
7019 		     mlx5_ib_stage_dev_notifier_cleanup),
7020 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
7021 		     mlx5_ib_stage_odp_init,
7022 		     mlx5_ib_stage_odp_cleanup),
7023 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
7024 		     mlx5_ib_stage_counters_init,
7025 		     mlx5_ib_stage_counters_cleanup),
7026 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
7027 		     mlx5_ib_stage_cong_debugfs_init,
7028 		     mlx5_ib_stage_cong_debugfs_cleanup),
7029 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
7030 		     mlx5_ib_stage_uar_init,
7031 		     mlx5_ib_stage_uar_cleanup),
7032 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
7033 		     mlx5_ib_stage_bfrag_init,
7034 		     mlx5_ib_stage_bfrag_cleanup),
7035 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
7036 		     NULL,
7037 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
7038 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
7039 		     mlx5_ib_stage_devx_init,
7040 		     mlx5_ib_stage_devx_cleanup),
7041 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
7042 		     mlx5_ib_stage_ib_reg_init,
7043 		     mlx5_ib_stage_ib_reg_cleanup),
7044 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
7045 		     mlx5_ib_stage_post_ib_reg_umr_init,
7046 		     NULL),
7047 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
7048 		     mlx5_ib_stage_delay_drop_init,
7049 		     mlx5_ib_stage_delay_drop_cleanup),
7050 };
7051 
7052 const struct mlx5_ib_profile raw_eth_profile = {
7053 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
7054 		     mlx5_ib_stage_init_init,
7055 		     mlx5_ib_stage_init_cleanup),
7056 	STAGE_CREATE(MLX5_IB_STAGE_FLOW_DB,
7057 		     mlx5_ib_stage_flow_db_init,
7058 		     mlx5_ib_stage_flow_db_cleanup),
7059 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
7060 		     mlx5_ib_stage_caps_init,
7061 		     mlx5_ib_stage_caps_cleanup),
7062 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
7063 		     mlx5_ib_stage_raw_eth_non_default_cb,
7064 		     NULL),
7065 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
7066 		     mlx5_ib_stage_raw_eth_roce_init,
7067 		     mlx5_ib_stage_raw_eth_roce_cleanup),
7068 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
7069 		     mlx5_init_srq_table,
7070 		     mlx5_cleanup_srq_table),
7071 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
7072 		     mlx5_ib_stage_dev_res_init,
7073 		     mlx5_ib_stage_dev_res_cleanup),
7074 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
7075 		     mlx5_ib_stage_dev_notifier_init,
7076 		     mlx5_ib_stage_dev_notifier_cleanup),
7077 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
7078 		     mlx5_ib_stage_counters_init,
7079 		     mlx5_ib_stage_counters_cleanup),
7080 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
7081 		     mlx5_ib_stage_uar_init,
7082 		     mlx5_ib_stage_uar_cleanup),
7083 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
7084 		     mlx5_ib_stage_bfrag_init,
7085 		     mlx5_ib_stage_bfrag_cleanup),
7086 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
7087 		     NULL,
7088 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
7089 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
7090 		     mlx5_ib_stage_devx_init,
7091 		     mlx5_ib_stage_devx_cleanup),
7092 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
7093 		     mlx5_ib_stage_ib_reg_init,
7094 		     mlx5_ib_stage_ib_reg_cleanup),
7095 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
7096 		     mlx5_ib_stage_post_ib_reg_umr_init,
7097 		     NULL),
7098 };
7099 
7100 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev)
7101 {
7102 	struct mlx5_ib_multiport_info *mpi;
7103 	struct mlx5_ib_dev *dev;
7104 	bool bound = false;
7105 	int err;
7106 
7107 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
7108 	if (!mpi)
7109 		return NULL;
7110 
7111 	mpi->mdev = mdev;
7112 
7113 	err = mlx5_query_nic_vport_system_image_guid(mdev,
7114 						     &mpi->sys_image_guid);
7115 	if (err) {
7116 		kfree(mpi);
7117 		return NULL;
7118 	}
7119 
7120 	mutex_lock(&mlx5_ib_multiport_mutex);
7121 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
7122 		if (dev->sys_image_guid == mpi->sys_image_guid)
7123 			bound = mlx5_ib_bind_slave_port(dev, mpi);
7124 
7125 		if (bound) {
7126 			rdma_roce_rescan_device(&dev->ib_dev);
7127 			break;
7128 		}
7129 	}
7130 
7131 	if (!bound) {
7132 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
7133 		dev_dbg(mdev->device,
7134 			"no suitable IB device found to bind to, added to unaffiliated list.\n");
7135 	}
7136 	mutex_unlock(&mlx5_ib_multiport_mutex);
7137 
7138 	return mpi;
7139 }
7140 
7141 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
7142 {
7143 	const struct mlx5_ib_profile *profile;
7144 	enum rdma_link_layer ll;
7145 	struct mlx5_ib_dev *dev;
7146 	int port_type_cap;
7147 	int num_ports;
7148 
7149 	printk_once(KERN_INFO "%s", mlx5_version);
7150 
7151 	if (MLX5_ESWITCH_MANAGER(mdev) &&
7152 	    mlx5_ib_eswitch_mode(mdev->priv.eswitch) == MLX5_ESWITCH_OFFLOADS) {
7153 		if (!mlx5_core_mp_enabled(mdev))
7154 			mlx5_ib_register_vport_reps(mdev);
7155 		return mdev;
7156 	}
7157 
7158 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
7159 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
7160 
7161 	if (mlx5_core_is_mp_slave(mdev) && ll == IB_LINK_LAYER_ETHERNET)
7162 		return mlx5_ib_add_slave_port(mdev);
7163 
7164 	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
7165 			MLX5_CAP_GEN(mdev, num_vhca_ports));
7166 	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
7167 	if (!dev)
7168 		return NULL;
7169 	dev->port = kcalloc(num_ports, sizeof(*dev->port),
7170 			     GFP_KERNEL);
7171 	if (!dev->port) {
7172 		ib_dealloc_device(&dev->ib_dev);
7173 		return NULL;
7174 	}
7175 
7176 	dev->mdev = mdev;
7177 	dev->num_ports = num_ports;
7178 
7179 	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_is_roce_enabled(mdev))
7180 		profile = &raw_eth_profile;
7181 	else
7182 		profile = &pf_profile;
7183 
7184 	return __mlx5_ib_add(dev, profile);
7185 }
7186 
7187 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
7188 {
7189 	struct mlx5_ib_multiport_info *mpi;
7190 	struct mlx5_ib_dev *dev;
7191 
7192 	if (MLX5_ESWITCH_MANAGER(mdev) && context == mdev) {
7193 		mlx5_ib_unregister_vport_reps(mdev);
7194 		return;
7195 	}
7196 
7197 	if (mlx5_core_is_mp_slave(mdev)) {
7198 		mpi = context;
7199 		mutex_lock(&mlx5_ib_multiport_mutex);
7200 		if (mpi->ibdev)
7201 			mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
7202 		list_del(&mpi->list);
7203 		mutex_unlock(&mlx5_ib_multiport_mutex);
7204 		kfree(mpi);
7205 		return;
7206 	}
7207 
7208 	dev = context;
7209 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
7210 }
7211 
7212 static struct mlx5_interface mlx5_ib_interface = {
7213 	.add            = mlx5_ib_add,
7214 	.remove         = mlx5_ib_remove,
7215 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
7216 };
7217 
7218 unsigned long mlx5_ib_get_xlt_emergency_page(void)
7219 {
7220 	mutex_lock(&xlt_emergency_page_mutex);
7221 	return xlt_emergency_page;
7222 }
7223 
7224 void mlx5_ib_put_xlt_emergency_page(void)
7225 {
7226 	mutex_unlock(&xlt_emergency_page_mutex);
7227 }
7228 
7229 static int __init mlx5_ib_init(void)
7230 {
7231 	int err;
7232 
7233 	xlt_emergency_page = __get_free_page(GFP_KERNEL);
7234 	if (!xlt_emergency_page)
7235 		return -ENOMEM;
7236 
7237 	mutex_init(&xlt_emergency_page_mutex);
7238 
7239 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
7240 	if (!mlx5_ib_event_wq) {
7241 		free_page(xlt_emergency_page);
7242 		return -ENOMEM;
7243 	}
7244 
7245 	mlx5_ib_odp_init();
7246 
7247 	err = mlx5_register_interface(&mlx5_ib_interface);
7248 
7249 	return err;
7250 }
7251 
7252 static void __exit mlx5_ib_cleanup(void)
7253 {
7254 	mlx5_unregister_interface(&mlx5_ib_interface);
7255 	destroy_workqueue(mlx5_ib_event_wq);
7256 	mutex_destroy(&xlt_emergency_page_mutex);
7257 	free_page(xlt_emergency_page);
7258 }
7259 
7260 module_init(mlx5_ib_init);
7261 module_exit(mlx5_ib_cleanup);
7262