xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision 715f23b6)
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/pat.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 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
819 	struct mlx5_core_dev *mdev = dev->mdev;
820 	int err = -ENOMEM;
821 	int max_sq_desc;
822 	int max_rq_sg;
823 	int max_sq_sg;
824 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
825 	bool raw_support = !mlx5_core_mp_enabled(mdev);
826 	struct mlx5_ib_query_device_resp resp = {};
827 	size_t resp_len;
828 	u64 max_tso;
829 
830 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
831 	if (uhw->outlen && uhw->outlen < resp_len)
832 		return -EINVAL;
833 
834 	resp.response_length = resp_len;
835 
836 	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
837 		return -EINVAL;
838 
839 	memset(props, 0, sizeof(*props));
840 	err = mlx5_query_system_image_guid(ibdev,
841 					   &props->sys_image_guid);
842 	if (err)
843 		return err;
844 
845 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
846 	if (err)
847 		return err;
848 
849 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
850 	if (err)
851 		return err;
852 
853 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
854 		(fw_rev_min(dev->mdev) << 16) |
855 		fw_rev_sub(dev->mdev);
856 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
857 		IB_DEVICE_PORT_ACTIVE_EVENT		|
858 		IB_DEVICE_SYS_IMAGE_GUID		|
859 		IB_DEVICE_RC_RNR_NAK_GEN;
860 
861 	if (MLX5_CAP_GEN(mdev, pkv))
862 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
863 	if (MLX5_CAP_GEN(mdev, qkv))
864 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
865 	if (MLX5_CAP_GEN(mdev, apm))
866 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
867 	if (MLX5_CAP_GEN(mdev, xrc))
868 		props->device_cap_flags |= IB_DEVICE_XRC;
869 	if (MLX5_CAP_GEN(mdev, imaicl)) {
870 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
871 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
872 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
873 		/* We support 'Gappy' memory registration too */
874 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
875 	}
876 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
877 	if (MLX5_CAP_GEN(mdev, sho)) {
878 		props->device_cap_flags |= IB_DEVICE_INTEGRITY_HANDOVER;
879 		/* At this stage no support for signature handover */
880 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
881 				      IB_PROT_T10DIF_TYPE_2 |
882 				      IB_PROT_T10DIF_TYPE_3;
883 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
884 				       IB_GUARD_T10DIF_CSUM;
885 	}
886 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
887 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
888 
889 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
890 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
891 			/* Legacy bit to support old userspace libraries */
892 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
893 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
894 		}
895 
896 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
897 			props->raw_packet_caps |=
898 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
899 
900 		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
901 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
902 			if (max_tso) {
903 				resp.tso_caps.max_tso = 1 << max_tso;
904 				resp.tso_caps.supported_qpts |=
905 					1 << IB_QPT_RAW_PACKET;
906 				resp.response_length += sizeof(resp.tso_caps);
907 			}
908 		}
909 
910 		if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
911 			resp.rss_caps.rx_hash_function =
912 						MLX5_RX_HASH_FUNC_TOEPLITZ;
913 			resp.rss_caps.rx_hash_fields_mask =
914 						MLX5_RX_HASH_SRC_IPV4 |
915 						MLX5_RX_HASH_DST_IPV4 |
916 						MLX5_RX_HASH_SRC_IPV6 |
917 						MLX5_RX_HASH_DST_IPV6 |
918 						MLX5_RX_HASH_SRC_PORT_TCP |
919 						MLX5_RX_HASH_DST_PORT_TCP |
920 						MLX5_RX_HASH_SRC_PORT_UDP |
921 						MLX5_RX_HASH_DST_PORT_UDP |
922 						MLX5_RX_HASH_INNER;
923 			if (mlx5_accel_ipsec_device_caps(dev->mdev) &
924 			    MLX5_ACCEL_IPSEC_CAP_DEVICE)
925 				resp.rss_caps.rx_hash_fields_mask |=
926 					MLX5_RX_HASH_IPSEC_SPI;
927 			resp.response_length += sizeof(resp.rss_caps);
928 		}
929 	} else {
930 		if (field_avail(typeof(resp), tso_caps, uhw->outlen))
931 			resp.response_length += sizeof(resp.tso_caps);
932 		if (field_avail(typeof(resp), rss_caps, uhw->outlen))
933 			resp.response_length += sizeof(resp.rss_caps);
934 	}
935 
936 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
937 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
938 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
939 	}
940 
941 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
942 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
943 	    raw_support)
944 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
945 
946 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
947 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
948 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
949 
950 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
951 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
952 	    raw_support) {
953 		/* Legacy bit to support old userspace libraries */
954 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
955 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
956 	}
957 
958 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
959 		props->max_dm_size =
960 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
961 	}
962 
963 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
964 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
965 
966 	if (MLX5_CAP_GEN(mdev, end_pad))
967 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
968 
969 	props->vendor_part_id	   = mdev->pdev->device;
970 	props->hw_ver		   = mdev->pdev->revision;
971 
972 	props->max_mr_size	   = ~0ull;
973 	props->page_size_cap	   = ~(min_page_size - 1);
974 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
975 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
976 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
977 		     sizeof(struct mlx5_wqe_data_seg);
978 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
979 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
980 		     sizeof(struct mlx5_wqe_raddr_seg)) /
981 		sizeof(struct mlx5_wqe_data_seg);
982 	props->max_send_sge = max_sq_sg;
983 	props->max_recv_sge = max_rq_sg;
984 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
985 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
986 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
987 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
988 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
989 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
990 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
991 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
992 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
993 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
994 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
995 	props->max_srq_sge	   = max_rq_sg - 1;
996 	props->max_fast_reg_page_list_len =
997 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
998 	props->max_pi_fast_reg_page_list_len =
999 		props->max_fast_reg_page_list_len / 2;
1000 	props->max_sgl_rd =
1001 		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
1002 	get_atomic_caps_qp(dev, props);
1003 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
1004 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
1005 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
1006 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1007 					   props->max_mcast_grp;
1008 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
1009 	props->max_ah = INT_MAX;
1010 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
1011 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
1012 
1013 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
1014 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
1015 			props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
1016 		props->odp_caps = dev->odp_caps;
1017 	}
1018 
1019 	if (MLX5_CAP_GEN(mdev, cd))
1020 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
1021 
1022 	if (mlx5_core_is_vf(mdev))
1023 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
1024 
1025 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1026 	    IB_LINK_LAYER_ETHERNET && raw_support) {
1027 		props->rss_caps.max_rwq_indirection_tables =
1028 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1029 		props->rss_caps.max_rwq_indirection_table_size =
1030 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1031 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1032 		props->max_wq_type_rq =
1033 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1034 	}
1035 
1036 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1037 		props->tm_caps.max_num_tags =
1038 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1039 		props->tm_caps.max_ops =
1040 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1041 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1042 	}
1043 
1044 	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1045 	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1046 		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1047 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1048 	}
1049 
1050 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1051 		props->cq_caps.max_cq_moderation_count =
1052 						MLX5_MAX_CQ_COUNT;
1053 		props->cq_caps.max_cq_moderation_period =
1054 						MLX5_MAX_CQ_PERIOD;
1055 	}
1056 
1057 	if (field_avail(typeof(resp), cqe_comp_caps, uhw->outlen)) {
1058 		resp.response_length += sizeof(resp.cqe_comp_caps);
1059 
1060 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1061 			resp.cqe_comp_caps.max_num =
1062 				MLX5_CAP_GEN(dev->mdev,
1063 					     cqe_compression_max_num);
1064 
1065 			resp.cqe_comp_caps.supported_format =
1066 				MLX5_IB_CQE_RES_FORMAT_HASH |
1067 				MLX5_IB_CQE_RES_FORMAT_CSUM;
1068 
1069 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1070 				resp.cqe_comp_caps.supported_format |=
1071 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1072 		}
1073 	}
1074 
1075 	if (field_avail(typeof(resp), packet_pacing_caps, uhw->outlen) &&
1076 	    raw_support) {
1077 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1078 		    MLX5_CAP_GEN(mdev, qos)) {
1079 			resp.packet_pacing_caps.qp_rate_limit_max =
1080 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1081 			resp.packet_pacing_caps.qp_rate_limit_min =
1082 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1083 			resp.packet_pacing_caps.supported_qpts |=
1084 				1 << IB_QPT_RAW_PACKET;
1085 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1086 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1087 				resp.packet_pacing_caps.cap_flags |=
1088 					MLX5_IB_PP_SUPPORT_BURST;
1089 		}
1090 		resp.response_length += sizeof(resp.packet_pacing_caps);
1091 	}
1092 
1093 	if (field_avail(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes,
1094 			uhw->outlen)) {
1095 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1096 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1097 				MLX5_IB_ALLOW_MPW;
1098 
1099 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1100 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1101 				MLX5_IB_SUPPORT_EMPW;
1102 
1103 		resp.response_length +=
1104 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1105 	}
1106 
1107 	if (field_avail(typeof(resp), flags, uhw->outlen)) {
1108 		resp.response_length += sizeof(resp.flags);
1109 
1110 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1111 			resp.flags |=
1112 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1113 
1114 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1115 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1116 		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1117 			resp.flags |=
1118 				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1119 
1120 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1121 	}
1122 
1123 	if (field_avail(typeof(resp), sw_parsing_caps,
1124 			uhw->outlen)) {
1125 		resp.response_length += sizeof(resp.sw_parsing_caps);
1126 		if (MLX5_CAP_ETH(mdev, swp)) {
1127 			resp.sw_parsing_caps.sw_parsing_offloads |=
1128 				MLX5_IB_SW_PARSING;
1129 
1130 			if (MLX5_CAP_ETH(mdev, swp_csum))
1131 				resp.sw_parsing_caps.sw_parsing_offloads |=
1132 					MLX5_IB_SW_PARSING_CSUM;
1133 
1134 			if (MLX5_CAP_ETH(mdev, swp_lso))
1135 				resp.sw_parsing_caps.sw_parsing_offloads |=
1136 					MLX5_IB_SW_PARSING_LSO;
1137 
1138 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1139 				resp.sw_parsing_caps.supported_qpts =
1140 					BIT(IB_QPT_RAW_PACKET);
1141 		}
1142 	}
1143 
1144 	if (field_avail(typeof(resp), striding_rq_caps, uhw->outlen) &&
1145 	    raw_support) {
1146 		resp.response_length += sizeof(resp.striding_rq_caps);
1147 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1148 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1149 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1150 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1151 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1152 			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1153 				resp.striding_rq_caps
1154 					.min_single_wqe_log_num_of_strides =
1155 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1156 			else
1157 				resp.striding_rq_caps
1158 					.min_single_wqe_log_num_of_strides =
1159 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1160 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1161 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1162 			resp.striding_rq_caps.supported_qpts =
1163 				BIT(IB_QPT_RAW_PACKET);
1164 		}
1165 	}
1166 
1167 	if (field_avail(typeof(resp), tunnel_offloads_caps,
1168 			uhw->outlen)) {
1169 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1170 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1171 			resp.tunnel_offloads_caps |=
1172 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1173 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1174 			resp.tunnel_offloads_caps |=
1175 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1176 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1177 			resp.tunnel_offloads_caps |=
1178 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1179 		if (MLX5_CAP_GEN(mdev, flex_parser_protocols) &
1180 		    MLX5_FLEX_PROTO_CW_MPLS_GRE)
1181 			resp.tunnel_offloads_caps |=
1182 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1183 		if (MLX5_CAP_GEN(mdev, flex_parser_protocols) &
1184 		    MLX5_FLEX_PROTO_CW_MPLS_UDP)
1185 			resp.tunnel_offloads_caps |=
1186 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1187 	}
1188 
1189 	if (uhw->outlen) {
1190 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1191 
1192 		if (err)
1193 			return err;
1194 	}
1195 
1196 	return 0;
1197 }
1198 
1199 enum mlx5_ib_width {
1200 	MLX5_IB_WIDTH_1X	= 1 << 0,
1201 	MLX5_IB_WIDTH_2X	= 1 << 1,
1202 	MLX5_IB_WIDTH_4X	= 1 << 2,
1203 	MLX5_IB_WIDTH_8X	= 1 << 3,
1204 	MLX5_IB_WIDTH_12X	= 1 << 4
1205 };
1206 
1207 static void translate_active_width(struct ib_device *ibdev, u8 active_width,
1208 				  u8 *ib_width)
1209 {
1210 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1211 
1212 	if (active_width & MLX5_IB_WIDTH_1X)
1213 		*ib_width = IB_WIDTH_1X;
1214 	else if (active_width & MLX5_IB_WIDTH_2X)
1215 		*ib_width = IB_WIDTH_2X;
1216 	else if (active_width & MLX5_IB_WIDTH_4X)
1217 		*ib_width = IB_WIDTH_4X;
1218 	else if (active_width & MLX5_IB_WIDTH_8X)
1219 		*ib_width = IB_WIDTH_8X;
1220 	else if (active_width & MLX5_IB_WIDTH_12X)
1221 		*ib_width = IB_WIDTH_12X;
1222 	else {
1223 		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1224 			    (int)active_width);
1225 		*ib_width = IB_WIDTH_4X;
1226 	}
1227 
1228 	return;
1229 }
1230 
1231 static int mlx5_mtu_to_ib_mtu(int mtu)
1232 {
1233 	switch (mtu) {
1234 	case 256: return 1;
1235 	case 512: return 2;
1236 	case 1024: return 3;
1237 	case 2048: return 4;
1238 	case 4096: return 5;
1239 	default:
1240 		pr_warn("invalid mtu\n");
1241 		return -1;
1242 	}
1243 }
1244 
1245 enum ib_max_vl_num {
1246 	__IB_MAX_VL_0		= 1,
1247 	__IB_MAX_VL_0_1		= 2,
1248 	__IB_MAX_VL_0_3		= 3,
1249 	__IB_MAX_VL_0_7		= 4,
1250 	__IB_MAX_VL_0_14	= 5,
1251 };
1252 
1253 enum mlx5_vl_hw_cap {
1254 	MLX5_VL_HW_0	= 1,
1255 	MLX5_VL_HW_0_1	= 2,
1256 	MLX5_VL_HW_0_2	= 3,
1257 	MLX5_VL_HW_0_3	= 4,
1258 	MLX5_VL_HW_0_4	= 5,
1259 	MLX5_VL_HW_0_5	= 6,
1260 	MLX5_VL_HW_0_6	= 7,
1261 	MLX5_VL_HW_0_7	= 8,
1262 	MLX5_VL_HW_0_14	= 15
1263 };
1264 
1265 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1266 				u8 *max_vl_num)
1267 {
1268 	switch (vl_hw_cap) {
1269 	case MLX5_VL_HW_0:
1270 		*max_vl_num = __IB_MAX_VL_0;
1271 		break;
1272 	case MLX5_VL_HW_0_1:
1273 		*max_vl_num = __IB_MAX_VL_0_1;
1274 		break;
1275 	case MLX5_VL_HW_0_3:
1276 		*max_vl_num = __IB_MAX_VL_0_3;
1277 		break;
1278 	case MLX5_VL_HW_0_7:
1279 		*max_vl_num = __IB_MAX_VL_0_7;
1280 		break;
1281 	case MLX5_VL_HW_0_14:
1282 		*max_vl_num = __IB_MAX_VL_0_14;
1283 		break;
1284 
1285 	default:
1286 		return -EINVAL;
1287 	}
1288 
1289 	return 0;
1290 }
1291 
1292 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
1293 			       struct ib_port_attr *props)
1294 {
1295 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1296 	struct mlx5_core_dev *mdev = dev->mdev;
1297 	struct mlx5_hca_vport_context *rep;
1298 	u16 max_mtu;
1299 	u16 oper_mtu;
1300 	int err;
1301 	u8 ib_link_width_oper;
1302 	u8 vl_hw_cap;
1303 
1304 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1305 	if (!rep) {
1306 		err = -ENOMEM;
1307 		goto out;
1308 	}
1309 
1310 	/* props being zeroed by the caller, avoid zeroing it here */
1311 
1312 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1313 	if (err)
1314 		goto out;
1315 
1316 	props->lid		= rep->lid;
1317 	props->lmc		= rep->lmc;
1318 	props->sm_lid		= rep->sm_lid;
1319 	props->sm_sl		= rep->sm_sl;
1320 	props->state		= rep->vport_state;
1321 	props->phys_state	= rep->port_physical_state;
1322 	props->port_cap_flags	= rep->cap_mask1;
1323 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1324 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1325 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1326 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1327 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1328 	props->subnet_timeout	= rep->subnet_timeout;
1329 	props->init_type_reply	= rep->init_type_reply;
1330 
1331 	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1332 		props->port_cap_flags2 = rep->cap_mask2;
1333 
1334 	err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
1335 	if (err)
1336 		goto out;
1337 
1338 	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1339 
1340 	err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
1341 	if (err)
1342 		goto out;
1343 
1344 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1345 
1346 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1347 
1348 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1349 
1350 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1351 
1352 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1353 	if (err)
1354 		goto out;
1355 
1356 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1357 				   &props->max_vl_num);
1358 out:
1359 	kfree(rep);
1360 	return err;
1361 }
1362 
1363 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
1364 		       struct ib_port_attr *props)
1365 {
1366 	unsigned int count;
1367 	int ret;
1368 
1369 	switch (mlx5_get_vport_access_method(ibdev)) {
1370 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1371 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1372 		break;
1373 
1374 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1375 		ret = mlx5_query_hca_port(ibdev, port, props);
1376 		break;
1377 
1378 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1379 		ret = mlx5_query_port_roce(ibdev, port, props);
1380 		break;
1381 
1382 	default:
1383 		ret = -EINVAL;
1384 	}
1385 
1386 	if (!ret && props) {
1387 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1388 		struct mlx5_core_dev *mdev;
1389 		bool put_mdev = true;
1390 
1391 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1392 		if (!mdev) {
1393 			/* If the port isn't affiliated yet query the master.
1394 			 * The master and slave will have the same values.
1395 			 */
1396 			mdev = dev->mdev;
1397 			port = 1;
1398 			put_mdev = false;
1399 		}
1400 		count = mlx5_core_reserved_gids_count(mdev);
1401 		if (put_mdev)
1402 			mlx5_ib_put_native_port_mdev(dev, port);
1403 		props->gid_tbl_len -= count;
1404 	}
1405 	return ret;
1406 }
1407 
1408 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u8 port,
1409 				  struct ib_port_attr *props)
1410 {
1411 	int ret;
1412 
1413 	/* Only link layer == ethernet is valid for representors
1414 	 * and we always use port 1
1415 	 */
1416 	ret = mlx5_query_port_roce(ibdev, port, props);
1417 	if (ret || !props)
1418 		return ret;
1419 
1420 	/* We don't support GIDS */
1421 	props->gid_tbl_len = 0;
1422 
1423 	return ret;
1424 }
1425 
1426 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1427 			     union ib_gid *gid)
1428 {
1429 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1430 	struct mlx5_core_dev *mdev = dev->mdev;
1431 
1432 	switch (mlx5_get_vport_access_method(ibdev)) {
1433 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1434 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1435 
1436 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1437 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1438 
1439 	default:
1440 		return -EINVAL;
1441 	}
1442 
1443 }
1444 
1445 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u8 port,
1446 				   u16 index, u16 *pkey)
1447 {
1448 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1449 	struct mlx5_core_dev *mdev;
1450 	bool put_mdev = true;
1451 	u8 mdev_port_num;
1452 	int err;
1453 
1454 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1455 	if (!mdev) {
1456 		/* The port isn't affiliated yet, get the PKey from the master
1457 		 * port. For RoCE the PKey tables will be the same.
1458 		 */
1459 		put_mdev = false;
1460 		mdev = dev->mdev;
1461 		mdev_port_num = 1;
1462 	}
1463 
1464 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1465 					index, pkey);
1466 	if (put_mdev)
1467 		mlx5_ib_put_native_port_mdev(dev, port);
1468 
1469 	return err;
1470 }
1471 
1472 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1473 			      u16 *pkey)
1474 {
1475 	switch (mlx5_get_vport_access_method(ibdev)) {
1476 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1477 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1478 
1479 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1480 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1481 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1482 	default:
1483 		return -EINVAL;
1484 	}
1485 }
1486 
1487 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1488 				 struct ib_device_modify *props)
1489 {
1490 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1491 	struct mlx5_reg_node_desc in;
1492 	struct mlx5_reg_node_desc out;
1493 	int err;
1494 
1495 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1496 		return -EOPNOTSUPP;
1497 
1498 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1499 		return 0;
1500 
1501 	/*
1502 	 * If possible, pass node desc to FW, so it can generate
1503 	 * a 144 trap.  If cmd fails, just ignore.
1504 	 */
1505 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1506 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1507 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1508 	if (err)
1509 		return err;
1510 
1511 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1512 
1513 	return err;
1514 }
1515 
1516 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1517 				u32 value)
1518 {
1519 	struct mlx5_hca_vport_context ctx = {};
1520 	struct mlx5_core_dev *mdev;
1521 	u8 mdev_port_num;
1522 	int err;
1523 
1524 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1525 	if (!mdev)
1526 		return -ENODEV;
1527 
1528 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1529 	if (err)
1530 		goto out;
1531 
1532 	if (~ctx.cap_mask1_perm & mask) {
1533 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1534 			     mask, ctx.cap_mask1_perm);
1535 		err = -EINVAL;
1536 		goto out;
1537 	}
1538 
1539 	ctx.cap_mask1 = value;
1540 	ctx.cap_mask1_perm = mask;
1541 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1542 						 0, &ctx);
1543 
1544 out:
1545 	mlx5_ib_put_native_port_mdev(dev, port_num);
1546 
1547 	return err;
1548 }
1549 
1550 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1551 			       struct ib_port_modify *props)
1552 {
1553 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1554 	struct ib_port_attr attr;
1555 	u32 tmp;
1556 	int err;
1557 	u32 change_mask;
1558 	u32 value;
1559 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1560 		      IB_LINK_LAYER_INFINIBAND);
1561 
1562 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1563 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1564 	 */
1565 	if (!is_ib)
1566 		return 0;
1567 
1568 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1569 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1570 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1571 		return set_port_caps_atomic(dev, port, change_mask, value);
1572 	}
1573 
1574 	mutex_lock(&dev->cap_mask_mutex);
1575 
1576 	err = ib_query_port(ibdev, port, &attr);
1577 	if (err)
1578 		goto out;
1579 
1580 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1581 		~props->clr_port_cap_mask;
1582 
1583 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1584 
1585 out:
1586 	mutex_unlock(&dev->cap_mask_mutex);
1587 	return err;
1588 }
1589 
1590 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1591 {
1592 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1593 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1594 }
1595 
1596 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1597 {
1598 	/* Large page with non 4k uar support might limit the dynamic size */
1599 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1600 		return MLX5_MIN_DYN_BFREGS;
1601 
1602 	return MLX5_MAX_DYN_BFREGS;
1603 }
1604 
1605 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1606 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1607 			     struct mlx5_bfreg_info *bfregi)
1608 {
1609 	int uars_per_sys_page;
1610 	int bfregs_per_sys_page;
1611 	int ref_bfregs = req->total_num_bfregs;
1612 
1613 	if (req->total_num_bfregs == 0)
1614 		return -EINVAL;
1615 
1616 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1617 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1618 
1619 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1620 		return -ENOMEM;
1621 
1622 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1623 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1624 	/* This holds the required static allocation asked by the user */
1625 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1626 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1627 		return -EINVAL;
1628 
1629 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1630 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1631 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1632 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1633 
1634 	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",
1635 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1636 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1637 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1638 		    bfregi->num_sys_pages);
1639 
1640 	return 0;
1641 }
1642 
1643 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1644 {
1645 	struct mlx5_bfreg_info *bfregi;
1646 	int err;
1647 	int i;
1648 
1649 	bfregi = &context->bfregi;
1650 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1651 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1652 		if (err)
1653 			goto error;
1654 
1655 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1656 	}
1657 
1658 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1659 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1660 
1661 	return 0;
1662 
1663 error:
1664 	for (--i; i >= 0; i--)
1665 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1666 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1667 
1668 	return err;
1669 }
1670 
1671 static void deallocate_uars(struct mlx5_ib_dev *dev,
1672 			    struct mlx5_ib_ucontext *context)
1673 {
1674 	struct mlx5_bfreg_info *bfregi;
1675 	int i;
1676 
1677 	bfregi = &context->bfregi;
1678 	for (i = 0; i < bfregi->num_sys_pages; i++)
1679 		if (i < bfregi->num_static_sys_pages ||
1680 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1681 			mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1682 }
1683 
1684 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1685 {
1686 	int err = 0;
1687 
1688 	mutex_lock(&dev->lb.mutex);
1689 	if (td)
1690 		dev->lb.user_td++;
1691 	if (qp)
1692 		dev->lb.qps++;
1693 
1694 	if (dev->lb.user_td == 2 ||
1695 	    dev->lb.qps == 1) {
1696 		if (!dev->lb.enabled) {
1697 			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1698 			dev->lb.enabled = true;
1699 		}
1700 	}
1701 
1702 	mutex_unlock(&dev->lb.mutex);
1703 
1704 	return err;
1705 }
1706 
1707 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1708 {
1709 	mutex_lock(&dev->lb.mutex);
1710 	if (td)
1711 		dev->lb.user_td--;
1712 	if (qp)
1713 		dev->lb.qps--;
1714 
1715 	if (dev->lb.user_td == 1 &&
1716 	    dev->lb.qps == 0) {
1717 		if (dev->lb.enabled) {
1718 			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1719 			dev->lb.enabled = false;
1720 		}
1721 	}
1722 
1723 	mutex_unlock(&dev->lb.mutex);
1724 }
1725 
1726 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1727 					  u16 uid)
1728 {
1729 	int err;
1730 
1731 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1732 		return 0;
1733 
1734 	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1735 	if (err)
1736 		return err;
1737 
1738 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1739 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1740 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1741 		return err;
1742 
1743 	return mlx5_ib_enable_lb(dev, true, false);
1744 }
1745 
1746 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1747 					     u16 uid)
1748 {
1749 	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1750 		return;
1751 
1752 	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
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;
1758 
1759 	mlx5_ib_disable_lb(dev, true, false);
1760 }
1761 
1762 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1763 				  struct ib_udata *udata)
1764 {
1765 	struct ib_device *ibdev = uctx->device;
1766 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1767 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1768 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1769 	struct mlx5_core_dev *mdev = dev->mdev;
1770 	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1771 	struct mlx5_bfreg_info *bfregi;
1772 	int ver;
1773 	int err;
1774 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1775 				     max_cqe_version);
1776 	u32 dump_fill_mkey;
1777 	bool lib_uar_4k;
1778 
1779 	if (!dev->ib_active)
1780 		return -EAGAIN;
1781 
1782 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1783 		ver = 0;
1784 	else if (udata->inlen >= min_req_v2)
1785 		ver = 2;
1786 	else
1787 		return -EINVAL;
1788 
1789 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1790 	if (err)
1791 		return err;
1792 
1793 	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1794 		return -EOPNOTSUPP;
1795 
1796 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1797 		return -EOPNOTSUPP;
1798 
1799 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1800 				    MLX5_NON_FP_BFREGS_PER_UAR);
1801 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1802 		return -EINVAL;
1803 
1804 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1805 	if (dev->wc_support)
1806 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1807 	resp.cache_line_size = cache_line_size();
1808 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1809 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1810 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1811 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1812 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1813 	resp.cqe_version = min_t(__u8,
1814 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1815 				 req.max_cqe_version);
1816 	resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1817 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1818 	resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1819 					MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1;
1820 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1821 				   sizeof(resp.response_length), udata->outlen);
1822 
1823 	if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1824 		if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_EGRESS))
1825 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1826 		if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1827 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1828 		if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1829 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1830 		if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1831 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1832 		/* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1833 	}
1834 
1835 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1836 	bfregi = &context->bfregi;
1837 
1838 	/* updates req->total_num_bfregs */
1839 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1840 	if (err)
1841 		goto out_ctx;
1842 
1843 	mutex_init(&bfregi->lock);
1844 	bfregi->lib_uar_4k = lib_uar_4k;
1845 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1846 				GFP_KERNEL);
1847 	if (!bfregi->count) {
1848 		err = -ENOMEM;
1849 		goto out_ctx;
1850 	}
1851 
1852 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1853 				    sizeof(*bfregi->sys_pages),
1854 				    GFP_KERNEL);
1855 	if (!bfregi->sys_pages) {
1856 		err = -ENOMEM;
1857 		goto out_count;
1858 	}
1859 
1860 	err = allocate_uars(dev, context);
1861 	if (err)
1862 		goto out_sys_pages;
1863 
1864 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1865 		err = mlx5_ib_devx_create(dev, true);
1866 		if (err < 0)
1867 			goto out_uars;
1868 		context->devx_uid = err;
1869 	}
1870 
1871 	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1872 					     context->devx_uid);
1873 	if (err)
1874 		goto out_devx;
1875 
1876 	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1877 		err = mlx5_cmd_dump_fill_mkey(dev->mdev, &dump_fill_mkey);
1878 		if (err)
1879 			goto out_mdev;
1880 	}
1881 
1882 	INIT_LIST_HEAD(&context->db_page_list);
1883 	mutex_init(&context->db_page_mutex);
1884 
1885 	resp.tot_bfregs = req.total_num_bfregs;
1886 	resp.num_ports = dev->num_ports;
1887 
1888 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1889 		resp.response_length += sizeof(resp.cqe_version);
1890 
1891 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1892 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1893 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1894 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1895 	}
1896 
1897 	if (field_avail(typeof(resp), eth_min_inline, udata->outlen)) {
1898 		if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1899 			mlx5_query_min_inline(dev->mdev, &resp.eth_min_inline);
1900 			resp.eth_min_inline++;
1901 		}
1902 		resp.response_length += sizeof(resp.eth_min_inline);
1903 	}
1904 
1905 	if (field_avail(typeof(resp), clock_info_versions, udata->outlen)) {
1906 		if (mdev->clock_info)
1907 			resp.clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1908 		resp.response_length += sizeof(resp.clock_info_versions);
1909 	}
1910 
1911 	/*
1912 	 * We don't want to expose information from the PCI bar that is located
1913 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1914 	 * pretend we don't support reading the HCA's core clock. This is also
1915 	 * forced by mmap function.
1916 	 */
1917 	if (field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1918 		if (PAGE_SIZE <= 4096) {
1919 			resp.comp_mask |=
1920 				MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1921 			resp.hca_core_clock_offset =
1922 				offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1923 		}
1924 		resp.response_length += sizeof(resp.hca_core_clock_offset);
1925 	}
1926 
1927 	if (field_avail(typeof(resp), log_uar_size, udata->outlen))
1928 		resp.response_length += sizeof(resp.log_uar_size);
1929 
1930 	if (field_avail(typeof(resp), num_uars_per_page, udata->outlen))
1931 		resp.response_length += sizeof(resp.num_uars_per_page);
1932 
1933 	if (field_avail(typeof(resp), num_dyn_bfregs, udata->outlen)) {
1934 		resp.num_dyn_bfregs = bfregi->num_dyn_bfregs;
1935 		resp.response_length += sizeof(resp.num_dyn_bfregs);
1936 	}
1937 
1938 	if (field_avail(typeof(resp), dump_fill_mkey, udata->outlen)) {
1939 		if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1940 			resp.dump_fill_mkey = dump_fill_mkey;
1941 			resp.comp_mask |=
1942 				MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1943 		}
1944 		resp.response_length += sizeof(resp.dump_fill_mkey);
1945 	}
1946 
1947 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1948 	if (err)
1949 		goto out_mdev;
1950 
1951 	bfregi->ver = ver;
1952 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1953 	context->cqe_version = resp.cqe_version;
1954 	context->lib_caps = req.lib_caps;
1955 	print_lib_caps(dev, context->lib_caps);
1956 
1957 	if (dev->lag_active) {
1958 		u8 port = mlx5_core_native_port_num(dev->mdev) - 1;
1959 
1960 		atomic_set(&context->tx_port_affinity,
1961 			   atomic_add_return(
1962 				   1, &dev->port[port].roce.tx_port_affinity));
1963 	}
1964 
1965 	return 0;
1966 
1967 out_mdev:
1968 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1969 out_devx:
1970 	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1971 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1972 
1973 out_uars:
1974 	deallocate_uars(dev, context);
1975 
1976 out_sys_pages:
1977 	kfree(bfregi->sys_pages);
1978 
1979 out_count:
1980 	kfree(bfregi->count);
1981 
1982 out_ctx:
1983 	return err;
1984 }
1985 
1986 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1987 {
1988 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1989 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1990 	struct mlx5_bfreg_info *bfregi;
1991 
1992 	bfregi = &context->bfregi;
1993 	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1994 
1995 	if (context->devx_uid)
1996 		mlx5_ib_devx_destroy(dev, context->devx_uid);
1997 
1998 	deallocate_uars(dev, context);
1999 	kfree(bfregi->sys_pages);
2000 	kfree(bfregi->count);
2001 }
2002 
2003 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2004 				 int uar_idx)
2005 {
2006 	int fw_uars_per_page;
2007 
2008 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2009 
2010 	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2011 }
2012 
2013 static int get_command(unsigned long offset)
2014 {
2015 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2016 }
2017 
2018 static int get_arg(unsigned long offset)
2019 {
2020 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2021 }
2022 
2023 static int get_index(unsigned long offset)
2024 {
2025 	return get_arg(offset);
2026 }
2027 
2028 /* Index resides in an extra byte to enable larger values than 255 */
2029 static int get_extended_index(unsigned long offset)
2030 {
2031 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2032 }
2033 
2034 
2035 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2036 {
2037 }
2038 
2039 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2040 {
2041 	switch (cmd) {
2042 	case MLX5_IB_MMAP_WC_PAGE:
2043 		return "WC";
2044 	case MLX5_IB_MMAP_REGULAR_PAGE:
2045 		return "best effort WC";
2046 	case MLX5_IB_MMAP_NC_PAGE:
2047 		return "NC";
2048 	case MLX5_IB_MMAP_DEVICE_MEM:
2049 		return "Device Memory";
2050 	default:
2051 		return NULL;
2052 	}
2053 }
2054 
2055 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2056 					struct vm_area_struct *vma,
2057 					struct mlx5_ib_ucontext *context)
2058 {
2059 	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2060 	    !(vma->vm_flags & VM_SHARED))
2061 		return -EINVAL;
2062 
2063 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2064 		return -EOPNOTSUPP;
2065 
2066 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2067 		return -EPERM;
2068 	vma->vm_flags &= ~VM_MAYWRITE;
2069 
2070 	if (!dev->mdev->clock_info)
2071 		return -EOPNOTSUPP;
2072 
2073 	return vm_insert_page(vma, vma->vm_start,
2074 			      virt_to_page(dev->mdev->clock_info));
2075 }
2076 
2077 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2078 {
2079 	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2080 	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2081 	struct mlx5_ib_dm *mdm;
2082 
2083 	switch (mentry->mmap_flag) {
2084 	case MLX5_IB_MMAP_TYPE_MEMIC:
2085 		mdm = container_of(mentry, struct mlx5_ib_dm, mentry);
2086 		mlx5_cmd_dealloc_memic(&dev->dm, mdm->dev_addr,
2087 				       mdm->size);
2088 		kfree(mdm);
2089 		break;
2090 	default:
2091 		WARN_ON(true);
2092 	}
2093 }
2094 
2095 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2096 		    struct vm_area_struct *vma,
2097 		    struct mlx5_ib_ucontext *context)
2098 {
2099 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2100 	int err;
2101 	unsigned long idx;
2102 	phys_addr_t pfn;
2103 	pgprot_t prot;
2104 	u32 bfreg_dyn_idx = 0;
2105 	u32 uar_index;
2106 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2107 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2108 				bfregi->num_static_sys_pages;
2109 
2110 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2111 		return -EINVAL;
2112 
2113 	if (dyn_uar)
2114 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2115 	else
2116 		idx = get_index(vma->vm_pgoff);
2117 
2118 	if (idx >= max_valid_idx) {
2119 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2120 			     idx, max_valid_idx);
2121 		return -EINVAL;
2122 	}
2123 
2124 	switch (cmd) {
2125 	case MLX5_IB_MMAP_WC_PAGE:
2126 	case MLX5_IB_MMAP_ALLOC_WC:
2127 /* Some architectures don't support WC memory */
2128 #if defined(CONFIG_X86)
2129 		if (!pat_enabled())
2130 			return -EPERM;
2131 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
2132 			return -EPERM;
2133 #endif
2134 	/* fall through */
2135 	case MLX5_IB_MMAP_REGULAR_PAGE:
2136 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2137 		prot = pgprot_writecombine(vma->vm_page_prot);
2138 		break;
2139 	case MLX5_IB_MMAP_NC_PAGE:
2140 		prot = pgprot_noncached(vma->vm_page_prot);
2141 		break;
2142 	default:
2143 		return -EINVAL;
2144 	}
2145 
2146 	if (dyn_uar) {
2147 		int uars_per_page;
2148 
2149 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2150 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2151 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2152 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2153 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2154 			return -EINVAL;
2155 		}
2156 
2157 		mutex_lock(&bfregi->lock);
2158 		/* Fail if uar already allocated, first bfreg index of each
2159 		 * page holds its count.
2160 		 */
2161 		if (bfregi->count[bfreg_dyn_idx]) {
2162 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2163 			mutex_unlock(&bfregi->lock);
2164 			return -EINVAL;
2165 		}
2166 
2167 		bfregi->count[bfreg_dyn_idx]++;
2168 		mutex_unlock(&bfregi->lock);
2169 
2170 		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2171 		if (err) {
2172 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2173 			goto free_bfreg;
2174 		}
2175 	} else {
2176 		uar_index = bfregi->sys_pages[idx];
2177 	}
2178 
2179 	pfn = uar_index2pfn(dev, uar_index);
2180 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2181 
2182 	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2183 				prot, NULL);
2184 	if (err) {
2185 		mlx5_ib_err(dev,
2186 			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2187 			    err, mmap_cmd2str(cmd));
2188 		goto err;
2189 	}
2190 
2191 	if (dyn_uar)
2192 		bfregi->sys_pages[idx] = uar_index;
2193 	return 0;
2194 
2195 err:
2196 	if (!dyn_uar)
2197 		return err;
2198 
2199 	mlx5_cmd_free_uar(dev->mdev, idx);
2200 
2201 free_bfreg:
2202 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2203 
2204 	return err;
2205 }
2206 
2207 static int add_dm_mmap_entry(struct ib_ucontext *context,
2208 			     struct mlx5_ib_dm *mdm,
2209 			     u64 address)
2210 {
2211 	mdm->mentry.mmap_flag = MLX5_IB_MMAP_TYPE_MEMIC;
2212 	mdm->mentry.address = address;
2213 	return rdma_user_mmap_entry_insert_range(
2214 			context, &mdm->mentry.rdma_entry,
2215 			mdm->size,
2216 			MLX5_IB_MMAP_DEVICE_MEM << 16,
2217 			(MLX5_IB_MMAP_DEVICE_MEM << 16) + (1UL << 16) - 1);
2218 }
2219 
2220 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2221 {
2222 	unsigned long idx;
2223 	u8 command;
2224 
2225 	command = get_command(vma->vm_pgoff);
2226 	idx = get_extended_index(vma->vm_pgoff);
2227 
2228 	return (command << 16 | idx);
2229 }
2230 
2231 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2232 			       struct vm_area_struct *vma,
2233 			       struct ib_ucontext *ucontext)
2234 {
2235 	struct mlx5_user_mmap_entry *mentry;
2236 	struct rdma_user_mmap_entry *entry;
2237 	unsigned long pgoff;
2238 	pgprot_t prot;
2239 	phys_addr_t pfn;
2240 	int ret;
2241 
2242 	pgoff = mlx5_vma_to_pgoff(vma);
2243 	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2244 	if (!entry)
2245 		return -EINVAL;
2246 
2247 	mentry = to_mmmap(entry);
2248 	pfn = (mentry->address >> PAGE_SHIFT);
2249 	prot = pgprot_writecombine(vma->vm_page_prot);
2250 	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2251 				entry->npages * PAGE_SIZE,
2252 				prot,
2253 				entry);
2254 	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2255 	return ret;
2256 }
2257 
2258 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2259 {
2260 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2261 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2262 	unsigned long command;
2263 	phys_addr_t pfn;
2264 
2265 	command = get_command(vma->vm_pgoff);
2266 	switch (command) {
2267 	case MLX5_IB_MMAP_WC_PAGE:
2268 	case MLX5_IB_MMAP_NC_PAGE:
2269 	case MLX5_IB_MMAP_REGULAR_PAGE:
2270 	case MLX5_IB_MMAP_ALLOC_WC:
2271 		return uar_mmap(dev, command, vma, context);
2272 
2273 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2274 		return -ENOSYS;
2275 
2276 	case MLX5_IB_MMAP_CORE_CLOCK:
2277 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2278 			return -EINVAL;
2279 
2280 		if (vma->vm_flags & VM_WRITE)
2281 			return -EPERM;
2282 		vma->vm_flags &= ~VM_MAYWRITE;
2283 
2284 		/* Don't expose to user-space information it shouldn't have */
2285 		if (PAGE_SIZE > 4096)
2286 			return -EOPNOTSUPP;
2287 
2288 		pfn = (dev->mdev->iseg_base +
2289 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2290 			PAGE_SHIFT;
2291 		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2292 					 PAGE_SIZE,
2293 					 pgprot_noncached(vma->vm_page_prot),
2294 					 NULL);
2295 	case MLX5_IB_MMAP_CLOCK_INFO:
2296 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2297 
2298 	default:
2299 		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2300 	}
2301 
2302 	return 0;
2303 }
2304 
2305 static inline int check_dm_type_support(struct mlx5_ib_dev *dev,
2306 					u32 type)
2307 {
2308 	switch (type) {
2309 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2310 		if (!MLX5_CAP_DEV_MEM(dev->mdev, memic))
2311 			return -EOPNOTSUPP;
2312 		break;
2313 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2314 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2315 		if (!capable(CAP_SYS_RAWIO) ||
2316 		    !capable(CAP_NET_RAW))
2317 			return -EPERM;
2318 
2319 		if (!(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
2320 		      MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, sw_owner)))
2321 			return -EOPNOTSUPP;
2322 		break;
2323 	}
2324 
2325 	return 0;
2326 }
2327 
2328 static int handle_alloc_dm_memic(struct ib_ucontext *ctx,
2329 				 struct mlx5_ib_dm *dm,
2330 				 struct ib_dm_alloc_attr *attr,
2331 				 struct uverbs_attr_bundle *attrs)
2332 {
2333 	struct mlx5_dm *dm_db = &to_mdev(ctx->device)->dm;
2334 	u64 start_offset;
2335 	u16 page_idx;
2336 	int err;
2337 	u64 address;
2338 
2339 	dm->size = roundup(attr->length, MLX5_MEMIC_BASE_SIZE);
2340 
2341 	err = mlx5_cmd_alloc_memic(dm_db, &dm->dev_addr,
2342 				   dm->size, attr->alignment);
2343 	if (err)
2344 		return err;
2345 
2346 	address = dm->dev_addr & PAGE_MASK;
2347 	err = add_dm_mmap_entry(ctx, dm, address);
2348 	if (err)
2349 		goto err_dealloc;
2350 
2351 	page_idx = dm->mentry.rdma_entry.start_pgoff & 0xFFFF;
2352 	err = uverbs_copy_to(attrs,
2353 			     MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
2354 			     &page_idx,
2355 			     sizeof(page_idx));
2356 	if (err)
2357 		goto err_copy;
2358 
2359 	start_offset = dm->dev_addr & ~PAGE_MASK;
2360 	err = uverbs_copy_to(attrs,
2361 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2362 			     &start_offset, sizeof(start_offset));
2363 	if (err)
2364 		goto err_copy;
2365 
2366 	return 0;
2367 
2368 err_copy:
2369 	rdma_user_mmap_entry_remove(&dm->mentry.rdma_entry);
2370 err_dealloc:
2371 	mlx5_cmd_dealloc_memic(dm_db, dm->dev_addr, dm->size);
2372 
2373 	return err;
2374 }
2375 
2376 static int handle_alloc_dm_sw_icm(struct ib_ucontext *ctx,
2377 				  struct mlx5_ib_dm *dm,
2378 				  struct ib_dm_alloc_attr *attr,
2379 				  struct uverbs_attr_bundle *attrs,
2380 				  int type)
2381 {
2382 	struct mlx5_core_dev *dev = to_mdev(ctx->device)->mdev;
2383 	u64 act_size;
2384 	int err;
2385 
2386 	/* Allocation size must a multiple of the basic block size
2387 	 * and a power of 2.
2388 	 */
2389 	act_size = round_up(attr->length, MLX5_SW_ICM_BLOCK_SIZE(dev));
2390 	act_size = roundup_pow_of_two(act_size);
2391 
2392 	dm->size = act_size;
2393 	err = mlx5_dm_sw_icm_alloc(dev, type, act_size,
2394 				   to_mucontext(ctx)->devx_uid, &dm->dev_addr,
2395 				   &dm->icm_dm.obj_id);
2396 	if (err)
2397 		return err;
2398 
2399 	err = uverbs_copy_to(attrs,
2400 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2401 			     &dm->dev_addr, sizeof(dm->dev_addr));
2402 	if (err)
2403 		mlx5_dm_sw_icm_dealloc(dev, type, dm->size,
2404 				       to_mucontext(ctx)->devx_uid, dm->dev_addr,
2405 				       dm->icm_dm.obj_id);
2406 
2407 	return err;
2408 }
2409 
2410 struct ib_dm *mlx5_ib_alloc_dm(struct ib_device *ibdev,
2411 			       struct ib_ucontext *context,
2412 			       struct ib_dm_alloc_attr *attr,
2413 			       struct uverbs_attr_bundle *attrs)
2414 {
2415 	struct mlx5_ib_dm *dm;
2416 	enum mlx5_ib_uapi_dm_type type;
2417 	int err;
2418 
2419 	err = uverbs_get_const_default(&type, attrs,
2420 				       MLX5_IB_ATTR_ALLOC_DM_REQ_TYPE,
2421 				       MLX5_IB_UAPI_DM_TYPE_MEMIC);
2422 	if (err)
2423 		return ERR_PTR(err);
2424 
2425 	mlx5_ib_dbg(to_mdev(ibdev), "alloc_dm req: dm_type=%d user_length=0x%llx log_alignment=%d\n",
2426 		    type, attr->length, attr->alignment);
2427 
2428 	err = check_dm_type_support(to_mdev(ibdev), type);
2429 	if (err)
2430 		return ERR_PTR(err);
2431 
2432 	dm = kzalloc(sizeof(*dm), GFP_KERNEL);
2433 	if (!dm)
2434 		return ERR_PTR(-ENOMEM);
2435 
2436 	dm->type = type;
2437 
2438 	switch (type) {
2439 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2440 		err = handle_alloc_dm_memic(context, dm,
2441 					    attr,
2442 					    attrs);
2443 		break;
2444 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2445 		err = handle_alloc_dm_sw_icm(context, dm,
2446 					     attr, attrs,
2447 					     MLX5_SW_ICM_TYPE_STEERING);
2448 		break;
2449 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2450 		err = handle_alloc_dm_sw_icm(context, dm,
2451 					     attr, attrs,
2452 					     MLX5_SW_ICM_TYPE_HEADER_MODIFY);
2453 		break;
2454 	default:
2455 		err = -EOPNOTSUPP;
2456 	}
2457 
2458 	if (err)
2459 		goto err_free;
2460 
2461 	return &dm->ibdm;
2462 
2463 err_free:
2464 	kfree(dm);
2465 	return ERR_PTR(err);
2466 }
2467 
2468 int mlx5_ib_dealloc_dm(struct ib_dm *ibdm, struct uverbs_attr_bundle *attrs)
2469 {
2470 	struct mlx5_ib_ucontext *ctx = rdma_udata_to_drv_context(
2471 		&attrs->driver_udata, struct mlx5_ib_ucontext, ibucontext);
2472 	struct mlx5_core_dev *dev = to_mdev(ibdm->device)->mdev;
2473 	struct mlx5_ib_dm *dm = to_mdm(ibdm);
2474 	int ret;
2475 
2476 	switch (dm->type) {
2477 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
2478 		rdma_user_mmap_entry_remove(&dm->mentry.rdma_entry);
2479 		return 0;
2480 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
2481 		ret = mlx5_dm_sw_icm_dealloc(dev, MLX5_SW_ICM_TYPE_STEERING,
2482 					     dm->size, ctx->devx_uid, dm->dev_addr,
2483 					     dm->icm_dm.obj_id);
2484 		if (ret)
2485 			return ret;
2486 		break;
2487 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
2488 		ret = mlx5_dm_sw_icm_dealloc(dev, MLX5_SW_ICM_TYPE_HEADER_MODIFY,
2489 					     dm->size, ctx->devx_uid, dm->dev_addr,
2490 					     dm->icm_dm.obj_id);
2491 		if (ret)
2492 			return ret;
2493 		break;
2494 	default:
2495 		return -EOPNOTSUPP;
2496 	}
2497 
2498 	kfree(dm);
2499 
2500 	return 0;
2501 }
2502 
2503 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2504 {
2505 	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2506 	struct ib_device *ibdev = ibpd->device;
2507 	struct mlx5_ib_alloc_pd_resp resp;
2508 	int err;
2509 	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2510 	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)]   = {};
2511 	u16 uid = 0;
2512 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2513 		udata, struct mlx5_ib_ucontext, ibucontext);
2514 
2515 	uid = context ? context->devx_uid : 0;
2516 	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2517 	MLX5_SET(alloc_pd_in, in, uid, uid);
2518 	err = mlx5_cmd_exec(to_mdev(ibdev)->mdev, in, sizeof(in),
2519 			    out, sizeof(out));
2520 	if (err)
2521 		return err;
2522 
2523 	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2524 	pd->uid = uid;
2525 	if (udata) {
2526 		resp.pdn = pd->pdn;
2527 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2528 			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2529 			return -EFAULT;
2530 		}
2531 	}
2532 
2533 	return 0;
2534 }
2535 
2536 static void mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2537 {
2538 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2539 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2540 
2541 	mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2542 }
2543 
2544 enum {
2545 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
2546 	MATCH_CRITERIA_ENABLE_MISC_BIT,
2547 	MATCH_CRITERIA_ENABLE_INNER_BIT,
2548 	MATCH_CRITERIA_ENABLE_MISC2_BIT
2549 };
2550 
2551 #define HEADER_IS_ZERO(match_criteria, headers)			           \
2552 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
2553 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
2554 
2555 static u8 get_match_criteria_enable(u32 *match_criteria)
2556 {
2557 	u8 match_criteria_enable;
2558 
2559 	match_criteria_enable =
2560 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
2561 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
2562 	match_criteria_enable |=
2563 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
2564 		MATCH_CRITERIA_ENABLE_MISC_BIT;
2565 	match_criteria_enable |=
2566 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
2567 		MATCH_CRITERIA_ENABLE_INNER_BIT;
2568 	match_criteria_enable |=
2569 		(!HEADER_IS_ZERO(match_criteria, misc_parameters_2)) <<
2570 		MATCH_CRITERIA_ENABLE_MISC2_BIT;
2571 
2572 	return match_criteria_enable;
2573 }
2574 
2575 static int set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
2576 {
2577 	u8 entry_mask;
2578 	u8 entry_val;
2579 	int err = 0;
2580 
2581 	if (!mask)
2582 		goto out;
2583 
2584 	entry_mask = MLX5_GET(fte_match_set_lyr_2_4, outer_c,
2585 			      ip_protocol);
2586 	entry_val = MLX5_GET(fte_match_set_lyr_2_4, outer_v,
2587 			     ip_protocol);
2588 	if (!entry_mask) {
2589 		MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
2590 		MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
2591 		goto out;
2592 	}
2593 	/* Don't override existing ip protocol */
2594 	if (mask != entry_mask || val != entry_val)
2595 		err = -EINVAL;
2596 out:
2597 	return err;
2598 }
2599 
2600 static void set_flow_label(void *misc_c, void *misc_v, u32 mask, u32 val,
2601 			   bool inner)
2602 {
2603 	if (inner) {
2604 		MLX5_SET(fte_match_set_misc,
2605 			 misc_c, inner_ipv6_flow_label, mask);
2606 		MLX5_SET(fte_match_set_misc,
2607 			 misc_v, inner_ipv6_flow_label, val);
2608 	} else {
2609 		MLX5_SET(fte_match_set_misc,
2610 			 misc_c, outer_ipv6_flow_label, mask);
2611 		MLX5_SET(fte_match_set_misc,
2612 			 misc_v, outer_ipv6_flow_label, val);
2613 	}
2614 }
2615 
2616 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
2617 {
2618 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
2619 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
2620 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
2621 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
2622 }
2623 
2624 static int check_mpls_supp_fields(u32 field_support, const __be32 *set_mask)
2625 {
2626 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_label) &&
2627 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_LABEL))
2628 		return -EOPNOTSUPP;
2629 
2630 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_exp) &&
2631 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_EXP))
2632 		return -EOPNOTSUPP;
2633 
2634 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_s_bos) &&
2635 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_S_BOS))
2636 		return -EOPNOTSUPP;
2637 
2638 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_ttl) &&
2639 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_TTL))
2640 		return -EOPNOTSUPP;
2641 
2642 	return 0;
2643 }
2644 
2645 #define LAST_ETH_FIELD vlan_tag
2646 #define LAST_IB_FIELD sl
2647 #define LAST_IPV4_FIELD tos
2648 #define LAST_IPV6_FIELD traffic_class
2649 #define LAST_TCP_UDP_FIELD src_port
2650 #define LAST_TUNNEL_FIELD tunnel_id
2651 #define LAST_FLOW_TAG_FIELD tag_id
2652 #define LAST_DROP_FIELD size
2653 #define LAST_COUNTERS_FIELD counters
2654 
2655 /* Field is the last supported field */
2656 #define FIELDS_NOT_SUPPORTED(filter, field)\
2657 	memchr_inv((void *)&filter.field  +\
2658 		   sizeof(filter.field), 0,\
2659 		   sizeof(filter) -\
2660 		   offsetof(typeof(filter), field) -\
2661 		   sizeof(filter.field))
2662 
2663 int parse_flow_flow_action(struct mlx5_ib_flow_action *maction,
2664 			   bool is_egress,
2665 			   struct mlx5_flow_act *action)
2666 {
2667 
2668 	switch (maction->ib_action.type) {
2669 	case IB_FLOW_ACTION_ESP:
2670 		if (action->action & (MLX5_FLOW_CONTEXT_ACTION_ENCRYPT |
2671 				      MLX5_FLOW_CONTEXT_ACTION_DECRYPT))
2672 			return -EINVAL;
2673 		/* Currently only AES_GCM keymat is supported by the driver */
2674 		action->esp_id = (uintptr_t)maction->esp_aes_gcm.ctx;
2675 		action->action |= is_egress ?
2676 			MLX5_FLOW_CONTEXT_ACTION_ENCRYPT :
2677 			MLX5_FLOW_CONTEXT_ACTION_DECRYPT;
2678 		return 0;
2679 	case IB_FLOW_ACTION_UNSPECIFIED:
2680 		if (maction->flow_action_raw.sub_type ==
2681 		    MLX5_IB_FLOW_ACTION_MODIFY_HEADER) {
2682 			if (action->action & MLX5_FLOW_CONTEXT_ACTION_MOD_HDR)
2683 				return -EINVAL;
2684 			action->action |= MLX5_FLOW_CONTEXT_ACTION_MOD_HDR;
2685 			action->modify_hdr =
2686 				maction->flow_action_raw.modify_hdr;
2687 			return 0;
2688 		}
2689 		if (maction->flow_action_raw.sub_type ==
2690 		    MLX5_IB_FLOW_ACTION_DECAP) {
2691 			if (action->action & MLX5_FLOW_CONTEXT_ACTION_DECAP)
2692 				return -EINVAL;
2693 			action->action |= MLX5_FLOW_CONTEXT_ACTION_DECAP;
2694 			return 0;
2695 		}
2696 		if (maction->flow_action_raw.sub_type ==
2697 		    MLX5_IB_FLOW_ACTION_PACKET_REFORMAT) {
2698 			if (action->action &
2699 			    MLX5_FLOW_CONTEXT_ACTION_PACKET_REFORMAT)
2700 				return -EINVAL;
2701 			action->action |=
2702 				MLX5_FLOW_CONTEXT_ACTION_PACKET_REFORMAT;
2703 			action->pkt_reformat =
2704 				maction->flow_action_raw.pkt_reformat;
2705 			return 0;
2706 		}
2707 		/* fall through */
2708 	default:
2709 		return -EOPNOTSUPP;
2710 	}
2711 }
2712 
2713 static int parse_flow_attr(struct mlx5_core_dev *mdev,
2714 			   struct mlx5_flow_spec *spec,
2715 			   const union ib_flow_spec *ib_spec,
2716 			   const struct ib_flow_attr *flow_attr,
2717 			   struct mlx5_flow_act *action, u32 prev_type)
2718 {
2719 	struct mlx5_flow_context *flow_context = &spec->flow_context;
2720 	u32 *match_c = spec->match_criteria;
2721 	u32 *match_v = spec->match_value;
2722 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
2723 					   misc_parameters);
2724 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
2725 					   misc_parameters);
2726 	void *misc_params2_c = MLX5_ADDR_OF(fte_match_param, match_c,
2727 					    misc_parameters_2);
2728 	void *misc_params2_v = MLX5_ADDR_OF(fte_match_param, match_v,
2729 					    misc_parameters_2);
2730 	void *headers_c;
2731 	void *headers_v;
2732 	int match_ipv;
2733 	int ret;
2734 
2735 	if (ib_spec->type & IB_FLOW_SPEC_INNER) {
2736 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2737 					 inner_headers);
2738 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2739 					 inner_headers);
2740 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2741 					ft_field_support.inner_ip_version);
2742 	} else {
2743 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2744 					 outer_headers);
2745 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2746 					 outer_headers);
2747 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2748 					ft_field_support.outer_ip_version);
2749 	}
2750 
2751 	switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
2752 	case IB_FLOW_SPEC_ETH:
2753 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
2754 			return -EOPNOTSUPP;
2755 
2756 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2757 					     dmac_47_16),
2758 				ib_spec->eth.mask.dst_mac);
2759 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2760 					     dmac_47_16),
2761 				ib_spec->eth.val.dst_mac);
2762 
2763 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2764 					     smac_47_16),
2765 				ib_spec->eth.mask.src_mac);
2766 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2767 					     smac_47_16),
2768 				ib_spec->eth.val.src_mac);
2769 
2770 		if (ib_spec->eth.mask.vlan_tag) {
2771 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2772 				 cvlan_tag, 1);
2773 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2774 				 cvlan_tag, 1);
2775 
2776 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2777 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
2778 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2779 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
2780 
2781 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2782 				 first_cfi,
2783 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
2784 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2785 				 first_cfi,
2786 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
2787 
2788 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2789 				 first_prio,
2790 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
2791 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2792 				 first_prio,
2793 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
2794 		}
2795 		MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2796 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
2797 		MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2798 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
2799 		break;
2800 	case IB_FLOW_SPEC_IPV4:
2801 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
2802 			return -EOPNOTSUPP;
2803 
2804 		if (match_ipv) {
2805 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2806 				 ip_version, 0xf);
2807 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2808 				 ip_version, MLX5_FS_IPV4_VERSION);
2809 		} else {
2810 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2811 				 ethertype, 0xffff);
2812 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2813 				 ethertype, ETH_P_IP);
2814 		}
2815 
2816 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2817 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
2818 		       &ib_spec->ipv4.mask.src_ip,
2819 		       sizeof(ib_spec->ipv4.mask.src_ip));
2820 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2821 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
2822 		       &ib_spec->ipv4.val.src_ip,
2823 		       sizeof(ib_spec->ipv4.val.src_ip));
2824 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2825 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2826 		       &ib_spec->ipv4.mask.dst_ip,
2827 		       sizeof(ib_spec->ipv4.mask.dst_ip));
2828 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2829 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2830 		       &ib_spec->ipv4.val.dst_ip,
2831 		       sizeof(ib_spec->ipv4.val.dst_ip));
2832 
2833 		set_tos(headers_c, headers_v,
2834 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
2835 
2836 		if (set_proto(headers_c, headers_v,
2837 			      ib_spec->ipv4.mask.proto,
2838 			      ib_spec->ipv4.val.proto))
2839 			return -EINVAL;
2840 		break;
2841 	case IB_FLOW_SPEC_IPV6:
2842 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
2843 			return -EOPNOTSUPP;
2844 
2845 		if (match_ipv) {
2846 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2847 				 ip_version, 0xf);
2848 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2849 				 ip_version, MLX5_FS_IPV6_VERSION);
2850 		} else {
2851 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2852 				 ethertype, 0xffff);
2853 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2854 				 ethertype, ETH_P_IPV6);
2855 		}
2856 
2857 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2858 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
2859 		       &ib_spec->ipv6.mask.src_ip,
2860 		       sizeof(ib_spec->ipv6.mask.src_ip));
2861 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2862 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
2863 		       &ib_spec->ipv6.val.src_ip,
2864 		       sizeof(ib_spec->ipv6.val.src_ip));
2865 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2866 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2867 		       &ib_spec->ipv6.mask.dst_ip,
2868 		       sizeof(ib_spec->ipv6.mask.dst_ip));
2869 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2870 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2871 		       &ib_spec->ipv6.val.dst_ip,
2872 		       sizeof(ib_spec->ipv6.val.dst_ip));
2873 
2874 		set_tos(headers_c, headers_v,
2875 			ib_spec->ipv6.mask.traffic_class,
2876 			ib_spec->ipv6.val.traffic_class);
2877 
2878 		if (set_proto(headers_c, headers_v,
2879 			      ib_spec->ipv6.mask.next_hdr,
2880 			      ib_spec->ipv6.val.next_hdr))
2881 			return -EINVAL;
2882 
2883 		set_flow_label(misc_params_c, misc_params_v,
2884 			       ntohl(ib_spec->ipv6.mask.flow_label),
2885 			       ntohl(ib_spec->ipv6.val.flow_label),
2886 			       ib_spec->type & IB_FLOW_SPEC_INNER);
2887 		break;
2888 	case IB_FLOW_SPEC_ESP:
2889 		if (ib_spec->esp.mask.seq)
2890 			return -EOPNOTSUPP;
2891 
2892 		MLX5_SET(fte_match_set_misc, misc_params_c, outer_esp_spi,
2893 			 ntohl(ib_spec->esp.mask.spi));
2894 		MLX5_SET(fte_match_set_misc, misc_params_v, outer_esp_spi,
2895 			 ntohl(ib_spec->esp.val.spi));
2896 		break;
2897 	case IB_FLOW_SPEC_TCP:
2898 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2899 					 LAST_TCP_UDP_FIELD))
2900 			return -EOPNOTSUPP;
2901 
2902 		if (set_proto(headers_c, headers_v, 0xff, IPPROTO_TCP))
2903 			return -EINVAL;
2904 
2905 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_sport,
2906 			 ntohs(ib_spec->tcp_udp.mask.src_port));
2907 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_sport,
2908 			 ntohs(ib_spec->tcp_udp.val.src_port));
2909 
2910 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_dport,
2911 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
2912 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_dport,
2913 			 ntohs(ib_spec->tcp_udp.val.dst_port));
2914 		break;
2915 	case IB_FLOW_SPEC_UDP:
2916 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2917 					 LAST_TCP_UDP_FIELD))
2918 			return -EOPNOTSUPP;
2919 
2920 		if (set_proto(headers_c, headers_v, 0xff, IPPROTO_UDP))
2921 			return -EINVAL;
2922 
2923 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_sport,
2924 			 ntohs(ib_spec->tcp_udp.mask.src_port));
2925 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_sport,
2926 			 ntohs(ib_spec->tcp_udp.val.src_port));
2927 
2928 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_dport,
2929 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
2930 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_dport,
2931 			 ntohs(ib_spec->tcp_udp.val.dst_port));
2932 		break;
2933 	case IB_FLOW_SPEC_GRE:
2934 		if (ib_spec->gre.mask.c_ks_res0_ver)
2935 			return -EOPNOTSUPP;
2936 
2937 		if (set_proto(headers_c, headers_v, 0xff, IPPROTO_GRE))
2938 			return -EINVAL;
2939 
2940 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2941 			 0xff);
2942 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2943 			 IPPROTO_GRE);
2944 
2945 		MLX5_SET(fte_match_set_misc, misc_params_c, gre_protocol,
2946 			 ntohs(ib_spec->gre.mask.protocol));
2947 		MLX5_SET(fte_match_set_misc, misc_params_v, gre_protocol,
2948 			 ntohs(ib_spec->gre.val.protocol));
2949 
2950 		memcpy(MLX5_ADDR_OF(fte_match_set_misc, misc_params_c,
2951 				    gre_key.nvgre.hi),
2952 		       &ib_spec->gre.mask.key,
2953 		       sizeof(ib_spec->gre.mask.key));
2954 		memcpy(MLX5_ADDR_OF(fte_match_set_misc, misc_params_v,
2955 				    gre_key.nvgre.hi),
2956 		       &ib_spec->gre.val.key,
2957 		       sizeof(ib_spec->gre.val.key));
2958 		break;
2959 	case IB_FLOW_SPEC_MPLS:
2960 		switch (prev_type) {
2961 		case IB_FLOW_SPEC_UDP:
2962 			if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2963 						   ft_field_support.outer_first_mpls_over_udp),
2964 						   &ib_spec->mpls.mask.tag))
2965 				return -EOPNOTSUPP;
2966 
2967 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
2968 					    outer_first_mpls_over_udp),
2969 			       &ib_spec->mpls.val.tag,
2970 			       sizeof(ib_spec->mpls.val.tag));
2971 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
2972 					    outer_first_mpls_over_udp),
2973 			       &ib_spec->mpls.mask.tag,
2974 			       sizeof(ib_spec->mpls.mask.tag));
2975 			break;
2976 		case IB_FLOW_SPEC_GRE:
2977 			if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2978 						   ft_field_support.outer_first_mpls_over_gre),
2979 						   &ib_spec->mpls.mask.tag))
2980 				return -EOPNOTSUPP;
2981 
2982 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
2983 					    outer_first_mpls_over_gre),
2984 			       &ib_spec->mpls.val.tag,
2985 			       sizeof(ib_spec->mpls.val.tag));
2986 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
2987 					    outer_first_mpls_over_gre),
2988 			       &ib_spec->mpls.mask.tag,
2989 			       sizeof(ib_spec->mpls.mask.tag));
2990 			break;
2991 		default:
2992 			if (ib_spec->type & IB_FLOW_SPEC_INNER) {
2993 				if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2994 							   ft_field_support.inner_first_mpls),
2995 							   &ib_spec->mpls.mask.tag))
2996 					return -EOPNOTSUPP;
2997 
2998 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
2999 						    inner_first_mpls),
3000 				       &ib_spec->mpls.val.tag,
3001 				       sizeof(ib_spec->mpls.val.tag));
3002 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
3003 						    inner_first_mpls),
3004 				       &ib_spec->mpls.mask.tag,
3005 				       sizeof(ib_spec->mpls.mask.tag));
3006 			} else {
3007 				if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
3008 							   ft_field_support.outer_first_mpls),
3009 							   &ib_spec->mpls.mask.tag))
3010 					return -EOPNOTSUPP;
3011 
3012 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
3013 						    outer_first_mpls),
3014 				       &ib_spec->mpls.val.tag,
3015 				       sizeof(ib_spec->mpls.val.tag));
3016 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
3017 						    outer_first_mpls),
3018 				       &ib_spec->mpls.mask.tag,
3019 				       sizeof(ib_spec->mpls.mask.tag));
3020 			}
3021 		}
3022 		break;
3023 	case IB_FLOW_SPEC_VXLAN_TUNNEL:
3024 		if (FIELDS_NOT_SUPPORTED(ib_spec->tunnel.mask,
3025 					 LAST_TUNNEL_FIELD))
3026 			return -EOPNOTSUPP;
3027 
3028 		MLX5_SET(fte_match_set_misc, misc_params_c, vxlan_vni,
3029 			 ntohl(ib_spec->tunnel.mask.tunnel_id));
3030 		MLX5_SET(fte_match_set_misc, misc_params_v, vxlan_vni,
3031 			 ntohl(ib_spec->tunnel.val.tunnel_id));
3032 		break;
3033 	case IB_FLOW_SPEC_ACTION_TAG:
3034 		if (FIELDS_NOT_SUPPORTED(ib_spec->flow_tag,
3035 					 LAST_FLOW_TAG_FIELD))
3036 			return -EOPNOTSUPP;
3037 		if (ib_spec->flow_tag.tag_id >= BIT(24))
3038 			return -EINVAL;
3039 
3040 		flow_context->flow_tag = ib_spec->flow_tag.tag_id;
3041 		flow_context->flags |= FLOW_CONTEXT_HAS_TAG;
3042 		break;
3043 	case IB_FLOW_SPEC_ACTION_DROP:
3044 		if (FIELDS_NOT_SUPPORTED(ib_spec->drop,
3045 					 LAST_DROP_FIELD))
3046 			return -EOPNOTSUPP;
3047 		action->action |= MLX5_FLOW_CONTEXT_ACTION_DROP;
3048 		break;
3049 	case IB_FLOW_SPEC_ACTION_HANDLE:
3050 		ret = parse_flow_flow_action(to_mflow_act(ib_spec->action.act),
3051 			flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS, action);
3052 		if (ret)
3053 			return ret;
3054 		break;
3055 	case IB_FLOW_SPEC_ACTION_COUNT:
3056 		if (FIELDS_NOT_SUPPORTED(ib_spec->flow_count,
3057 					 LAST_COUNTERS_FIELD))
3058 			return -EOPNOTSUPP;
3059 
3060 		/* for now support only one counters spec per flow */
3061 		if (action->action & MLX5_FLOW_CONTEXT_ACTION_COUNT)
3062 			return -EINVAL;
3063 
3064 		action->counters = ib_spec->flow_count.counters;
3065 		action->action |= MLX5_FLOW_CONTEXT_ACTION_COUNT;
3066 		break;
3067 	default:
3068 		return -EINVAL;
3069 	}
3070 
3071 	return 0;
3072 }
3073 
3074 /* If a flow could catch both multicast and unicast packets,
3075  * it won't fall into the multicast flow steering table and this rule
3076  * could steal other multicast packets.
3077  */
3078 static bool flow_is_multicast_only(const struct ib_flow_attr *ib_attr)
3079 {
3080 	union ib_flow_spec *flow_spec;
3081 
3082 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
3083 	    ib_attr->num_of_specs < 1)
3084 		return false;
3085 
3086 	flow_spec = (union ib_flow_spec *)(ib_attr + 1);
3087 	if (flow_spec->type == IB_FLOW_SPEC_IPV4) {
3088 		struct ib_flow_spec_ipv4 *ipv4_spec;
3089 
3090 		ipv4_spec = (struct ib_flow_spec_ipv4 *)flow_spec;
3091 		if (ipv4_is_multicast(ipv4_spec->val.dst_ip))
3092 			return true;
3093 
3094 		return false;
3095 	}
3096 
3097 	if (flow_spec->type == IB_FLOW_SPEC_ETH) {
3098 		struct ib_flow_spec_eth *eth_spec;
3099 
3100 		eth_spec = (struct ib_flow_spec_eth *)flow_spec;
3101 		return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
3102 		       is_multicast_ether_addr(eth_spec->val.dst_mac);
3103 	}
3104 
3105 	return false;
3106 }
3107 
3108 enum valid_spec {
3109 	VALID_SPEC_INVALID,
3110 	VALID_SPEC_VALID,
3111 	VALID_SPEC_NA,
3112 };
3113 
3114 static enum valid_spec
3115 is_valid_esp_aes_gcm(struct mlx5_core_dev *mdev,
3116 		     const struct mlx5_flow_spec *spec,
3117 		     const struct mlx5_flow_act *flow_act,
3118 		     bool egress)
3119 {
3120 	const u32 *match_c = spec->match_criteria;
3121 	bool is_crypto =
3122 		(flow_act->action & (MLX5_FLOW_CONTEXT_ACTION_ENCRYPT |
3123 				     MLX5_FLOW_CONTEXT_ACTION_DECRYPT));
3124 	bool is_ipsec = mlx5_fs_is_ipsec_flow(match_c);
3125 	bool is_drop = flow_act->action & MLX5_FLOW_CONTEXT_ACTION_DROP;
3126 
3127 	/*
3128 	 * Currently only crypto is supported in egress, when regular egress
3129 	 * rules would be supported, always return VALID_SPEC_NA.
3130 	 */
3131 	if (!is_crypto)
3132 		return VALID_SPEC_NA;
3133 
3134 	return is_crypto && is_ipsec &&
3135 		(!egress || (!is_drop &&
3136 			     !(spec->flow_context.flags & FLOW_CONTEXT_HAS_TAG))) ?
3137 		VALID_SPEC_VALID : VALID_SPEC_INVALID;
3138 }
3139 
3140 static bool is_valid_spec(struct mlx5_core_dev *mdev,
3141 			  const struct mlx5_flow_spec *spec,
3142 			  const struct mlx5_flow_act *flow_act,
3143 			  bool egress)
3144 {
3145 	/* We curretly only support ipsec egress flow */
3146 	return is_valid_esp_aes_gcm(mdev, spec, flow_act, egress) != VALID_SPEC_INVALID;
3147 }
3148 
3149 static bool is_valid_ethertype(struct mlx5_core_dev *mdev,
3150 			       const struct ib_flow_attr *flow_attr,
3151 			       bool check_inner)
3152 {
3153 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
3154 	int match_ipv = check_inner ?
3155 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
3156 					ft_field_support.inner_ip_version) :
3157 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
3158 					ft_field_support.outer_ip_version);
3159 	int inner_bit = check_inner ? IB_FLOW_SPEC_INNER : 0;
3160 	bool ipv4_spec_valid, ipv6_spec_valid;
3161 	unsigned int ip_spec_type = 0;
3162 	bool has_ethertype = false;
3163 	unsigned int spec_index;
3164 	bool mask_valid = true;
3165 	u16 eth_type = 0;
3166 	bool type_valid;
3167 
3168 	/* Validate that ethertype is correct */
3169 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
3170 		if ((ib_spec->type == (IB_FLOW_SPEC_ETH | inner_bit)) &&
3171 		    ib_spec->eth.mask.ether_type) {
3172 			mask_valid = (ib_spec->eth.mask.ether_type ==
3173 				      htons(0xffff));
3174 			has_ethertype = true;
3175 			eth_type = ntohs(ib_spec->eth.val.ether_type);
3176 		} else if ((ib_spec->type == (IB_FLOW_SPEC_IPV4 | inner_bit)) ||
3177 			   (ib_spec->type == (IB_FLOW_SPEC_IPV6 | inner_bit))) {
3178 			ip_spec_type = ib_spec->type;
3179 		}
3180 		ib_spec = (void *)ib_spec + ib_spec->size;
3181 	}
3182 
3183 	type_valid = (!has_ethertype) || (!ip_spec_type);
3184 	if (!type_valid && mask_valid) {
3185 		ipv4_spec_valid = (eth_type == ETH_P_IP) &&
3186 			(ip_spec_type == (IB_FLOW_SPEC_IPV4 | inner_bit));
3187 		ipv6_spec_valid = (eth_type == ETH_P_IPV6) &&
3188 			(ip_spec_type == (IB_FLOW_SPEC_IPV6 | inner_bit));
3189 
3190 		type_valid = (ipv4_spec_valid) || (ipv6_spec_valid) ||
3191 			     (((eth_type == ETH_P_MPLS_UC) ||
3192 			       (eth_type == ETH_P_MPLS_MC)) && match_ipv);
3193 	}
3194 
3195 	return type_valid;
3196 }
3197 
3198 static bool is_valid_attr(struct mlx5_core_dev *mdev,
3199 			  const struct ib_flow_attr *flow_attr)
3200 {
3201 	return is_valid_ethertype(mdev, flow_attr, false) &&
3202 	       is_valid_ethertype(mdev, flow_attr, true);
3203 }
3204 
3205 static void put_flow_table(struct mlx5_ib_dev *dev,
3206 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
3207 {
3208 	prio->refcount -= !!ft_added;
3209 	if (!prio->refcount) {
3210 		mlx5_destroy_flow_table(prio->flow_table);
3211 		prio->flow_table = NULL;
3212 	}
3213 }
3214 
3215 static void counters_clear_description(struct ib_counters *counters)
3216 {
3217 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
3218 
3219 	mutex_lock(&mcounters->mcntrs_mutex);
3220 	kfree(mcounters->counters_data);
3221 	mcounters->counters_data = NULL;
3222 	mcounters->cntrs_max_index = 0;
3223 	mutex_unlock(&mcounters->mcntrs_mutex);
3224 }
3225 
3226 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
3227 {
3228 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
3229 							  struct mlx5_ib_flow_handler,
3230 							  ibflow);
3231 	struct mlx5_ib_flow_handler *iter, *tmp;
3232 	struct mlx5_ib_dev *dev = handler->dev;
3233 
3234 	mutex_lock(&dev->flow_db->lock);
3235 
3236 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
3237 		mlx5_del_flow_rules(iter->rule);
3238 		put_flow_table(dev, iter->prio, true);
3239 		list_del(&iter->list);
3240 		kfree(iter);
3241 	}
3242 
3243 	mlx5_del_flow_rules(handler->rule);
3244 	put_flow_table(dev, handler->prio, true);
3245 	if (handler->ibcounters &&
3246 	    atomic_read(&handler->ibcounters->usecnt) == 1)
3247 		counters_clear_description(handler->ibcounters);
3248 
3249 	mutex_unlock(&dev->flow_db->lock);
3250 	if (handler->flow_matcher)
3251 		atomic_dec(&handler->flow_matcher->usecnt);
3252 	kfree(handler);
3253 
3254 	return 0;
3255 }
3256 
3257 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
3258 {
3259 	priority *= 2;
3260 	if (!dont_trap)
3261 		priority++;
3262 	return priority;
3263 }
3264 
3265 enum flow_table_type {
3266 	MLX5_IB_FT_RX,
3267 	MLX5_IB_FT_TX
3268 };
3269 
3270 #define MLX5_FS_MAX_TYPES	 6
3271 #define MLX5_FS_MAX_ENTRIES	 BIT(16)
3272 
3273 static struct mlx5_ib_flow_prio *_get_prio(struct mlx5_flow_namespace *ns,
3274 					   struct mlx5_ib_flow_prio *prio,
3275 					   int priority,
3276 					   int num_entries, int num_groups,
3277 					   u32 flags)
3278 {
3279 	struct mlx5_flow_table *ft;
3280 
3281 	ft = mlx5_create_auto_grouped_flow_table(ns, priority,
3282 						 num_entries,
3283 						 num_groups,
3284 						 0, flags);
3285 	if (IS_ERR(ft))
3286 		return ERR_CAST(ft);
3287 
3288 	prio->flow_table = ft;
3289 	prio->refcount = 0;
3290 	return prio;
3291 }
3292 
3293 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
3294 						struct ib_flow_attr *flow_attr,
3295 						enum flow_table_type ft_type)
3296 {
3297 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
3298 	struct mlx5_flow_namespace *ns = NULL;
3299 	struct mlx5_ib_flow_prio *prio;
3300 	struct mlx5_flow_table *ft;
3301 	int max_table_size;
3302 	int num_entries;
3303 	int num_groups;
3304 	bool esw_encap;
3305 	u32 flags = 0;
3306 	int priority;
3307 
3308 	max_table_size = BIT(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3309 						       log_max_ft_size));
3310 	esw_encap = mlx5_eswitch_get_encap_mode(dev->mdev) !=
3311 		DEVLINK_ESWITCH_ENCAP_MODE_NONE;
3312 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
3313 		enum mlx5_flow_namespace_type fn_type;
3314 
3315 		if (flow_is_multicast_only(flow_attr) &&
3316 		    !dont_trap)
3317 			priority = MLX5_IB_FLOW_MCAST_PRIO;
3318 		else
3319 			priority = ib_prio_to_core_prio(flow_attr->priority,
3320 							dont_trap);
3321 		if (ft_type == MLX5_IB_FT_RX) {
3322 			fn_type = MLX5_FLOW_NAMESPACE_BYPASS;
3323 			prio = &dev->flow_db->prios[priority];
3324 			if (!dev->is_rep && !esw_encap &&
3325 			    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, decap))
3326 				flags |= MLX5_FLOW_TABLE_TUNNEL_EN_DECAP;
3327 			if (!dev->is_rep && !esw_encap &&
3328 			    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3329 					reformat_l3_tunnel_to_l2))
3330 				flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
3331 		} else {
3332 			max_table_size =
3333 				BIT(MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev,
3334 							      log_max_ft_size));
3335 			fn_type = MLX5_FLOW_NAMESPACE_EGRESS;
3336 			prio = &dev->flow_db->egress_prios[priority];
3337 			if (!dev->is_rep && !esw_encap &&
3338 			    MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, reformat))
3339 				flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
3340 		}
3341 		ns = mlx5_get_flow_namespace(dev->mdev, fn_type);
3342 		num_entries = MLX5_FS_MAX_ENTRIES;
3343 		num_groups = MLX5_FS_MAX_TYPES;
3344 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3345 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
3346 		ns = mlx5_get_flow_namespace(dev->mdev,
3347 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
3348 		build_leftovers_ft_param(&priority,
3349 					 &num_entries,
3350 					 &num_groups);
3351 		prio = &dev->flow_db->prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
3352 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3353 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
3354 					allow_sniffer_and_nic_rx_shared_tir))
3355 			return ERR_PTR(-ENOTSUPP);
3356 
3357 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
3358 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
3359 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
3360 
3361 		prio = &dev->flow_db->sniffer[ft_type];
3362 		priority = 0;
3363 		num_entries = 1;
3364 		num_groups = 1;
3365 	}
3366 
3367 	if (!ns)
3368 		return ERR_PTR(-ENOTSUPP);
3369 
3370 	max_table_size = min_t(int, num_entries, max_table_size);
3371 
3372 	ft = prio->flow_table;
3373 	if (!ft)
3374 		return _get_prio(ns, prio, priority, max_table_size, num_groups,
3375 				 flags);
3376 
3377 	return prio;
3378 }
3379 
3380 static void set_underlay_qp(struct mlx5_ib_dev *dev,
3381 			    struct mlx5_flow_spec *spec,
3382 			    u32 underlay_qpn)
3383 {
3384 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param,
3385 					   spec->match_criteria,
3386 					   misc_parameters);
3387 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3388 					   misc_parameters);
3389 
3390 	if (underlay_qpn &&
3391 	    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3392 				      ft_field_support.bth_dst_qp)) {
3393 		MLX5_SET(fte_match_set_misc,
3394 			 misc_params_v, bth_dst_qp, underlay_qpn);
3395 		MLX5_SET(fte_match_set_misc,
3396 			 misc_params_c, bth_dst_qp, 0xffffff);
3397 	}
3398 }
3399 
3400 static int read_flow_counters(struct ib_device *ibdev,
3401 			      struct mlx5_read_counters_attr *read_attr)
3402 {
3403 	struct mlx5_fc *fc = read_attr->hw_cntrs_hndl;
3404 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3405 
3406 	return mlx5_fc_query(dev->mdev, fc,
3407 			     &read_attr->out[IB_COUNTER_PACKETS],
3408 			     &read_attr->out[IB_COUNTER_BYTES]);
3409 }
3410 
3411 /* flow counters currently expose two counters packets and bytes */
3412 #define FLOW_COUNTERS_NUM 2
3413 static int counters_set_description(struct ib_counters *counters,
3414 				    enum mlx5_ib_counters_type counters_type,
3415 				    struct mlx5_ib_flow_counters_desc *desc_data,
3416 				    u32 ncounters)
3417 {
3418 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
3419 	u32 cntrs_max_index = 0;
3420 	int i;
3421 
3422 	if (counters_type != MLX5_IB_COUNTERS_FLOW)
3423 		return -EINVAL;
3424 
3425 	/* init the fields for the object */
3426 	mcounters->type = counters_type;
3427 	mcounters->read_counters = read_flow_counters;
3428 	mcounters->counters_num = FLOW_COUNTERS_NUM;
3429 	mcounters->ncounters = ncounters;
3430 	/* each counter entry have both description and index pair */
3431 	for (i = 0; i < ncounters; i++) {
3432 		if (desc_data[i].description > IB_COUNTER_BYTES)
3433 			return -EINVAL;
3434 
3435 		if (cntrs_max_index <= desc_data[i].index)
3436 			cntrs_max_index = desc_data[i].index + 1;
3437 	}
3438 
3439 	mutex_lock(&mcounters->mcntrs_mutex);
3440 	mcounters->counters_data = desc_data;
3441 	mcounters->cntrs_max_index = cntrs_max_index;
3442 	mutex_unlock(&mcounters->mcntrs_mutex);
3443 
3444 	return 0;
3445 }
3446 
3447 #define MAX_COUNTERS_NUM (USHRT_MAX / (sizeof(u32) * 2))
3448 static int flow_counters_set_data(struct ib_counters *ibcounters,
3449 				  struct mlx5_ib_create_flow *ucmd)
3450 {
3451 	struct mlx5_ib_mcounters *mcounters = to_mcounters(ibcounters);
3452 	struct mlx5_ib_flow_counters_data *cntrs_data = NULL;
3453 	struct mlx5_ib_flow_counters_desc *desc_data = NULL;
3454 	bool hw_hndl = false;
3455 	int ret = 0;
3456 
3457 	if (ucmd && ucmd->ncounters_data != 0) {
3458 		cntrs_data = ucmd->data;
3459 		if (cntrs_data->ncounters > MAX_COUNTERS_NUM)
3460 			return -EINVAL;
3461 
3462 		desc_data = kcalloc(cntrs_data->ncounters,
3463 				    sizeof(*desc_data),
3464 				    GFP_KERNEL);
3465 		if (!desc_data)
3466 			return  -ENOMEM;
3467 
3468 		if (copy_from_user(desc_data,
3469 				   u64_to_user_ptr(cntrs_data->counters_data),
3470 				   sizeof(*desc_data) * cntrs_data->ncounters)) {
3471 			ret = -EFAULT;
3472 			goto free;
3473 		}
3474 	}
3475 
3476 	if (!mcounters->hw_cntrs_hndl) {
3477 		mcounters->hw_cntrs_hndl = mlx5_fc_create(
3478 			to_mdev(ibcounters->device)->mdev, false);
3479 		if (IS_ERR(mcounters->hw_cntrs_hndl)) {
3480 			ret = PTR_ERR(mcounters->hw_cntrs_hndl);
3481 			goto free;
3482 		}
3483 		hw_hndl = true;
3484 	}
3485 
3486 	if (desc_data) {
3487 		/* counters already bound to at least one flow */
3488 		if (mcounters->cntrs_max_index) {
3489 			ret = -EINVAL;
3490 			goto free_hndl;
3491 		}
3492 
3493 		ret = counters_set_description(ibcounters,
3494 					       MLX5_IB_COUNTERS_FLOW,
3495 					       desc_data,
3496 					       cntrs_data->ncounters);
3497 		if (ret)
3498 			goto free_hndl;
3499 
3500 	} else if (!mcounters->cntrs_max_index) {
3501 		/* counters not bound yet, must have udata passed */
3502 		ret = -EINVAL;
3503 		goto free_hndl;
3504 	}
3505 
3506 	return 0;
3507 
3508 free_hndl:
3509 	if (hw_hndl) {
3510 		mlx5_fc_destroy(to_mdev(ibcounters->device)->mdev,
3511 				mcounters->hw_cntrs_hndl);
3512 		mcounters->hw_cntrs_hndl = NULL;
3513 	}
3514 free:
3515 	kfree(desc_data);
3516 	return ret;
3517 }
3518 
3519 static void mlx5_ib_set_rule_source_port(struct mlx5_ib_dev *dev,
3520 					 struct mlx5_flow_spec *spec,
3521 					 struct mlx5_eswitch_rep *rep)
3522 {
3523 	struct mlx5_eswitch *esw = dev->mdev->priv.eswitch;
3524 	void *misc;
3525 
3526 	if (mlx5_eswitch_vport_match_metadata_enabled(esw)) {
3527 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3528 				    misc_parameters_2);
3529 
3530 		MLX5_SET(fte_match_set_misc2, misc, metadata_reg_c_0,
3531 			 mlx5_eswitch_get_vport_metadata_for_match(esw,
3532 								   rep->vport));
3533 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_criteria,
3534 				    misc_parameters_2);
3535 
3536 		MLX5_SET_TO_ONES(fte_match_set_misc2, misc, metadata_reg_c_0);
3537 	} else {
3538 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3539 				    misc_parameters);
3540 
3541 		MLX5_SET(fte_match_set_misc, misc, source_port, rep->vport);
3542 
3543 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_criteria,
3544 				    misc_parameters);
3545 
3546 		MLX5_SET_TO_ONES(fte_match_set_misc, misc, source_port);
3547 	}
3548 }
3549 
3550 static struct mlx5_ib_flow_handler *_create_flow_rule(struct mlx5_ib_dev *dev,
3551 						      struct mlx5_ib_flow_prio *ft_prio,
3552 						      const struct ib_flow_attr *flow_attr,
3553 						      struct mlx5_flow_destination *dst,
3554 						      u32 underlay_qpn,
3555 						      struct mlx5_ib_create_flow *ucmd)
3556 {
3557 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
3558 	struct mlx5_ib_flow_handler *handler;
3559 	struct mlx5_flow_act flow_act = {};
3560 	struct mlx5_flow_spec *spec;
3561 	struct mlx5_flow_destination dest_arr[2] = {};
3562 	struct mlx5_flow_destination *rule_dst = dest_arr;
3563 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
3564 	unsigned int spec_index;
3565 	u32 prev_type = 0;
3566 	int err = 0;
3567 	int dest_num = 0;
3568 	bool is_egress = flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS;
3569 
3570 	if (!is_valid_attr(dev->mdev, flow_attr))
3571 		return ERR_PTR(-EINVAL);
3572 
3573 	if (dev->is_rep && is_egress)
3574 		return ERR_PTR(-EINVAL);
3575 
3576 	spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
3577 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
3578 	if (!handler || !spec) {
3579 		err = -ENOMEM;
3580 		goto free;
3581 	}
3582 
3583 	INIT_LIST_HEAD(&handler->list);
3584 
3585 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
3586 		err = parse_flow_attr(dev->mdev, spec,
3587 				      ib_flow, flow_attr, &flow_act,
3588 				      prev_type);
3589 		if (err < 0)
3590 			goto free;
3591 
3592 		prev_type = ((union ib_flow_spec *)ib_flow)->type;
3593 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
3594 	}
3595 
3596 	if (dst && !(flow_act.action & MLX5_FLOW_CONTEXT_ACTION_DROP)) {
3597 		memcpy(&dest_arr[0], dst, sizeof(*dst));
3598 		dest_num++;
3599 	}
3600 
3601 	if (!flow_is_multicast_only(flow_attr))
3602 		set_underlay_qp(dev, spec, underlay_qpn);
3603 
3604 	if (dev->is_rep) {
3605 		struct mlx5_eswitch_rep *rep;
3606 
3607 		rep = dev->port[flow_attr->port - 1].rep;
3608 		if (!rep) {
3609 			err = -EINVAL;
3610 			goto free;
3611 		}
3612 
3613 		mlx5_ib_set_rule_source_port(dev, spec, rep);
3614 	}
3615 
3616 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
3617 
3618 	if (is_egress &&
3619 	    !is_valid_spec(dev->mdev, spec, &flow_act, is_egress)) {
3620 		err = -EINVAL;
3621 		goto free;
3622 	}
3623 
3624 	if (flow_act.action & MLX5_FLOW_CONTEXT_ACTION_COUNT) {
3625 		struct mlx5_ib_mcounters *mcounters;
3626 
3627 		err = flow_counters_set_data(flow_act.counters, ucmd);
3628 		if (err)
3629 			goto free;
3630 
3631 		mcounters = to_mcounters(flow_act.counters);
3632 		handler->ibcounters = flow_act.counters;
3633 		dest_arr[dest_num].type =
3634 			MLX5_FLOW_DESTINATION_TYPE_COUNTER;
3635 		dest_arr[dest_num].counter_id =
3636 			mlx5_fc_id(mcounters->hw_cntrs_hndl);
3637 		dest_num++;
3638 	}
3639 
3640 	if (flow_act.action & MLX5_FLOW_CONTEXT_ACTION_DROP) {
3641 		if (!dest_num)
3642 			rule_dst = NULL;
3643 	} else {
3644 		if (is_egress)
3645 			flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_ALLOW;
3646 		else
3647 			flow_act.action |=
3648 				dest_num ?  MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
3649 					MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
3650 	}
3651 
3652 	if ((spec->flow_context.flags & FLOW_CONTEXT_HAS_TAG)  &&
3653 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3654 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
3655 		mlx5_ib_warn(dev, "Flow tag %u and attribute type %x isn't allowed in leftovers\n",
3656 			     spec->flow_context.flow_tag, flow_attr->type);
3657 		err = -EINVAL;
3658 		goto free;
3659 	}
3660 	handler->rule = mlx5_add_flow_rules(ft, spec,
3661 					    &flow_act,
3662 					    rule_dst, dest_num);
3663 
3664 	if (IS_ERR(handler->rule)) {
3665 		err = PTR_ERR(handler->rule);
3666 		goto free;
3667 	}
3668 
3669 	ft_prio->refcount++;
3670 	handler->prio = ft_prio;
3671 	handler->dev = dev;
3672 
3673 	ft_prio->flow_table = ft;
3674 free:
3675 	if (err && handler) {
3676 		if (handler->ibcounters &&
3677 		    atomic_read(&handler->ibcounters->usecnt) == 1)
3678 			counters_clear_description(handler->ibcounters);
3679 		kfree(handler);
3680 	}
3681 	kvfree(spec);
3682 	return err ? ERR_PTR(err) : handler;
3683 }
3684 
3685 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
3686 						     struct mlx5_ib_flow_prio *ft_prio,
3687 						     const struct ib_flow_attr *flow_attr,
3688 						     struct mlx5_flow_destination *dst)
3689 {
3690 	return _create_flow_rule(dev, ft_prio, flow_attr, dst, 0, NULL);
3691 }
3692 
3693 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
3694 							  struct mlx5_ib_flow_prio *ft_prio,
3695 							  struct ib_flow_attr *flow_attr,
3696 							  struct mlx5_flow_destination *dst)
3697 {
3698 	struct mlx5_ib_flow_handler *handler_dst = NULL;
3699 	struct mlx5_ib_flow_handler *handler = NULL;
3700 
3701 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
3702 	if (!IS_ERR(handler)) {
3703 		handler_dst = create_flow_rule(dev, ft_prio,
3704 					       flow_attr, dst);
3705 		if (IS_ERR(handler_dst)) {
3706 			mlx5_del_flow_rules(handler->rule);
3707 			ft_prio->refcount--;
3708 			kfree(handler);
3709 			handler = handler_dst;
3710 		} else {
3711 			list_add(&handler_dst->list, &handler->list);
3712 		}
3713 	}
3714 
3715 	return handler;
3716 }
3717 enum {
3718 	LEFTOVERS_MC,
3719 	LEFTOVERS_UC,
3720 };
3721 
3722 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
3723 							  struct mlx5_ib_flow_prio *ft_prio,
3724 							  struct ib_flow_attr *flow_attr,
3725 							  struct mlx5_flow_destination *dst)
3726 {
3727 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
3728 	struct mlx5_ib_flow_handler *handler = NULL;
3729 
3730 	static struct {
3731 		struct ib_flow_attr	flow_attr;
3732 		struct ib_flow_spec_eth eth_flow;
3733 	} leftovers_specs[] = {
3734 		[LEFTOVERS_MC] = {
3735 			.flow_attr = {
3736 				.num_of_specs = 1,
3737 				.size = sizeof(leftovers_specs[0])
3738 			},
3739 			.eth_flow = {
3740 				.type = IB_FLOW_SPEC_ETH,
3741 				.size = sizeof(struct ib_flow_spec_eth),
3742 				.mask = {.dst_mac = {0x1} },
3743 				.val =  {.dst_mac = {0x1} }
3744 			}
3745 		},
3746 		[LEFTOVERS_UC] = {
3747 			.flow_attr = {
3748 				.num_of_specs = 1,
3749 				.size = sizeof(leftovers_specs[0])
3750 			},
3751 			.eth_flow = {
3752 				.type = IB_FLOW_SPEC_ETH,
3753 				.size = sizeof(struct ib_flow_spec_eth),
3754 				.mask = {.dst_mac = {0x1} },
3755 				.val = {.dst_mac = {} }
3756 			}
3757 		}
3758 	};
3759 
3760 	handler = create_flow_rule(dev, ft_prio,
3761 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
3762 				   dst);
3763 	if (!IS_ERR(handler) &&
3764 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
3765 		handler_ucast = create_flow_rule(dev, ft_prio,
3766 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
3767 						 dst);
3768 		if (IS_ERR(handler_ucast)) {
3769 			mlx5_del_flow_rules(handler->rule);
3770 			ft_prio->refcount--;
3771 			kfree(handler);
3772 			handler = handler_ucast;
3773 		} else {
3774 			list_add(&handler_ucast->list, &handler->list);
3775 		}
3776 	}
3777 
3778 	return handler;
3779 }
3780 
3781 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
3782 							struct mlx5_ib_flow_prio *ft_rx,
3783 							struct mlx5_ib_flow_prio *ft_tx,
3784 							struct mlx5_flow_destination *dst)
3785 {
3786 	struct mlx5_ib_flow_handler *handler_rx;
3787 	struct mlx5_ib_flow_handler *handler_tx;
3788 	int err;
3789 	static const struct ib_flow_attr flow_attr  = {
3790 		.num_of_specs = 0,
3791 		.size = sizeof(flow_attr)
3792 	};
3793 
3794 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
3795 	if (IS_ERR(handler_rx)) {
3796 		err = PTR_ERR(handler_rx);
3797 		goto err;
3798 	}
3799 
3800 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
3801 	if (IS_ERR(handler_tx)) {
3802 		err = PTR_ERR(handler_tx);
3803 		goto err_tx;
3804 	}
3805 
3806 	list_add(&handler_tx->list, &handler_rx->list);
3807 
3808 	return handler_rx;
3809 
3810 err_tx:
3811 	mlx5_del_flow_rules(handler_rx->rule);
3812 	ft_rx->refcount--;
3813 	kfree(handler_rx);
3814 err:
3815 	return ERR_PTR(err);
3816 }
3817 
3818 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
3819 					   struct ib_flow_attr *flow_attr,
3820 					   int domain,
3821 					   struct ib_udata *udata)
3822 {
3823 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3824 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3825 	struct mlx5_ib_flow_handler *handler = NULL;
3826 	struct mlx5_flow_destination *dst = NULL;
3827 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
3828 	struct mlx5_ib_flow_prio *ft_prio;
3829 	bool is_egress = flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS;
3830 	struct mlx5_ib_create_flow *ucmd = NULL, ucmd_hdr;
3831 	size_t min_ucmd_sz, required_ucmd_sz;
3832 	int err;
3833 	int underlay_qpn;
3834 
3835 	if (udata && udata->inlen) {
3836 		min_ucmd_sz = offsetof(typeof(ucmd_hdr), reserved) +
3837 				sizeof(ucmd_hdr.reserved);
3838 		if (udata->inlen < min_ucmd_sz)
3839 			return ERR_PTR(-EOPNOTSUPP);
3840 
3841 		err = ib_copy_from_udata(&ucmd_hdr, udata, min_ucmd_sz);
3842 		if (err)
3843 			return ERR_PTR(err);
3844 
3845 		/* currently supports only one counters data */
3846 		if (ucmd_hdr.ncounters_data > 1)
3847 			return ERR_PTR(-EINVAL);
3848 
3849 		required_ucmd_sz = min_ucmd_sz +
3850 			sizeof(struct mlx5_ib_flow_counters_data) *
3851 			ucmd_hdr.ncounters_data;
3852 		if (udata->inlen > required_ucmd_sz &&
3853 		    !ib_is_udata_cleared(udata, required_ucmd_sz,
3854 					 udata->inlen - required_ucmd_sz))
3855 			return ERR_PTR(-EOPNOTSUPP);
3856 
3857 		ucmd = kzalloc(required_ucmd_sz, GFP_KERNEL);
3858 		if (!ucmd)
3859 			return ERR_PTR(-ENOMEM);
3860 
3861 		err = ib_copy_from_udata(ucmd, udata, required_ucmd_sz);
3862 		if (err)
3863 			goto free_ucmd;
3864 	}
3865 
3866 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO) {
3867 		err = -ENOMEM;
3868 		goto free_ucmd;
3869 	}
3870 
3871 	if (domain != IB_FLOW_DOMAIN_USER ||
3872 	    flow_attr->port > dev->num_ports ||
3873 	    (flow_attr->flags & ~(IB_FLOW_ATTR_FLAGS_DONT_TRAP |
3874 				  IB_FLOW_ATTR_FLAGS_EGRESS))) {
3875 		err = -EINVAL;
3876 		goto free_ucmd;
3877 	}
3878 
3879 	if (is_egress &&
3880 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3881 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
3882 		err = -EINVAL;
3883 		goto free_ucmd;
3884 	}
3885 
3886 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
3887 	if (!dst) {
3888 		err = -ENOMEM;
3889 		goto free_ucmd;
3890 	}
3891 
3892 	mutex_lock(&dev->flow_db->lock);
3893 
3894 	ft_prio = get_flow_table(dev, flow_attr,
3895 				 is_egress ? MLX5_IB_FT_TX : MLX5_IB_FT_RX);
3896 	if (IS_ERR(ft_prio)) {
3897 		err = PTR_ERR(ft_prio);
3898 		goto unlock;
3899 	}
3900 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3901 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
3902 		if (IS_ERR(ft_prio_tx)) {
3903 			err = PTR_ERR(ft_prio_tx);
3904 			ft_prio_tx = NULL;
3905 			goto destroy_ft;
3906 		}
3907 	}
3908 
3909 	if (is_egress) {
3910 		dst->type = MLX5_FLOW_DESTINATION_TYPE_PORT;
3911 	} else {
3912 		dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
3913 		if (mqp->flags & MLX5_IB_QP_RSS)
3914 			dst->tir_num = mqp->rss_qp.tirn;
3915 		else
3916 			dst->tir_num = mqp->raw_packet_qp.rq.tirn;
3917 	}
3918 
3919 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
3920 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
3921 			handler = create_dont_trap_rule(dev, ft_prio,
3922 							flow_attr, dst);
3923 		} else {
3924 			underlay_qpn = (mqp->flags & MLX5_IB_QP_UNDERLAY) ?
3925 					mqp->underlay_qpn : 0;
3926 			handler = _create_flow_rule(dev, ft_prio, flow_attr,
3927 						    dst, underlay_qpn, ucmd);
3928 		}
3929 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3930 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
3931 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
3932 						dst);
3933 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3934 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
3935 	} else {
3936 		err = -EINVAL;
3937 		goto destroy_ft;
3938 	}
3939 
3940 	if (IS_ERR(handler)) {
3941 		err = PTR_ERR(handler);
3942 		handler = NULL;
3943 		goto destroy_ft;
3944 	}
3945 
3946 	mutex_unlock(&dev->flow_db->lock);
3947 	kfree(dst);
3948 	kfree(ucmd);
3949 
3950 	return &handler->ibflow;
3951 
3952 destroy_ft:
3953 	put_flow_table(dev, ft_prio, false);
3954 	if (ft_prio_tx)
3955 		put_flow_table(dev, ft_prio_tx, false);
3956 unlock:
3957 	mutex_unlock(&dev->flow_db->lock);
3958 	kfree(dst);
3959 free_ucmd:
3960 	kfree(ucmd);
3961 	return ERR_PTR(err);
3962 }
3963 
3964 static struct mlx5_ib_flow_prio *
3965 _get_flow_table(struct mlx5_ib_dev *dev,
3966 		struct mlx5_ib_flow_matcher *fs_matcher,
3967 		bool mcast)
3968 {
3969 	struct mlx5_flow_namespace *ns = NULL;
3970 	struct mlx5_ib_flow_prio *prio = NULL;
3971 	int max_table_size = 0;
3972 	bool esw_encap;
3973 	u32 flags = 0;
3974 	int priority;
3975 
3976 	if (mcast)
3977 		priority = MLX5_IB_FLOW_MCAST_PRIO;
3978 	else
3979 		priority = ib_prio_to_core_prio(fs_matcher->priority, false);
3980 
3981 	esw_encap = mlx5_eswitch_get_encap_mode(dev->mdev) !=
3982 		DEVLINK_ESWITCH_ENCAP_MODE_NONE;
3983 	if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_BYPASS) {
3984 		max_table_size = BIT(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3985 					log_max_ft_size));
3986 		if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, decap) && !esw_encap)
3987 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_DECAP;
3988 		if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3989 					      reformat_l3_tunnel_to_l2) &&
3990 		    !esw_encap)
3991 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
3992 	} else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_EGRESS) {
3993 		max_table_size = BIT(
3994 			MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, log_max_ft_size));
3995 		if (MLX5_CAP_FLOWTABLE_NIC_TX(dev->mdev, reformat) && !esw_encap)
3996 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
3997 	} else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_FDB) {
3998 		max_table_size = BIT(
3999 			MLX5_CAP_ESW_FLOWTABLE_FDB(dev->mdev, log_max_ft_size));
4000 		if (MLX5_CAP_ESW_FLOWTABLE_FDB(dev->mdev, decap) && esw_encap)
4001 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_DECAP;
4002 		if (MLX5_CAP_ESW_FLOWTABLE_FDB(dev->mdev, reformat_l3_tunnel_to_l2) &&
4003 		    esw_encap)
4004 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
4005 		priority = FDB_BYPASS_PATH;
4006 	} else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_RDMA_RX) {
4007 		max_table_size =
4008 			BIT(MLX5_CAP_FLOWTABLE_RDMA_RX(dev->mdev,
4009 						       log_max_ft_size));
4010 		priority = fs_matcher->priority;
4011 	}
4012 
4013 	max_table_size = min_t(int, max_table_size, MLX5_FS_MAX_ENTRIES);
4014 
4015 	ns = mlx5_get_flow_namespace(dev->mdev, fs_matcher->ns_type);
4016 	if (!ns)
4017 		return ERR_PTR(-ENOTSUPP);
4018 
4019 	if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_BYPASS)
4020 		prio = &dev->flow_db->prios[priority];
4021 	else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_EGRESS)
4022 		prio = &dev->flow_db->egress_prios[priority];
4023 	else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_FDB)
4024 		prio = &dev->flow_db->fdb;
4025 	else if (fs_matcher->ns_type == MLX5_FLOW_NAMESPACE_RDMA_RX)
4026 		prio = &dev->flow_db->rdma_rx[priority];
4027 
4028 	if (!prio)
4029 		return ERR_PTR(-EINVAL);
4030 
4031 	if (prio->flow_table)
4032 		return prio;
4033 
4034 	return _get_prio(ns, prio, priority, max_table_size,
4035 			 MLX5_FS_MAX_TYPES, flags);
4036 }
4037 
4038 static struct mlx5_ib_flow_handler *
4039 _create_raw_flow_rule(struct mlx5_ib_dev *dev,
4040 		      struct mlx5_ib_flow_prio *ft_prio,
4041 		      struct mlx5_flow_destination *dst,
4042 		      struct mlx5_ib_flow_matcher  *fs_matcher,
4043 		      struct mlx5_flow_context *flow_context,
4044 		      struct mlx5_flow_act *flow_act,
4045 		      void *cmd_in, int inlen,
4046 		      int dst_num)
4047 {
4048 	struct mlx5_ib_flow_handler *handler;
4049 	struct mlx5_flow_spec *spec;
4050 	struct mlx5_flow_table *ft = ft_prio->flow_table;
4051 	int err = 0;
4052 
4053 	spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
4054 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
4055 	if (!handler || !spec) {
4056 		err = -ENOMEM;
4057 		goto free;
4058 	}
4059 
4060 	INIT_LIST_HEAD(&handler->list);
4061 
4062 	memcpy(spec->match_value, cmd_in, inlen);
4063 	memcpy(spec->match_criteria, fs_matcher->matcher_mask.match_params,
4064 	       fs_matcher->mask_len);
4065 	spec->match_criteria_enable = fs_matcher->match_criteria_enable;
4066 	spec->flow_context = *flow_context;
4067 
4068 	handler->rule = mlx5_add_flow_rules(ft, spec,
4069 					    flow_act, dst, dst_num);
4070 
4071 	if (IS_ERR(handler->rule)) {
4072 		err = PTR_ERR(handler->rule);
4073 		goto free;
4074 	}
4075 
4076 	ft_prio->refcount++;
4077 	handler->prio = ft_prio;
4078 	handler->dev = dev;
4079 	ft_prio->flow_table = ft;
4080 
4081 free:
4082 	if (err)
4083 		kfree(handler);
4084 	kvfree(spec);
4085 	return err ? ERR_PTR(err) : handler;
4086 }
4087 
4088 static bool raw_fs_is_multicast(struct mlx5_ib_flow_matcher *fs_matcher,
4089 				void *match_v)
4090 {
4091 	void *match_c;
4092 	void *match_v_set_lyr_2_4, *match_c_set_lyr_2_4;
4093 	void *dmac, *dmac_mask;
4094 	void *ipv4, *ipv4_mask;
4095 
4096 	if (!(fs_matcher->match_criteria_enable &
4097 	      (1 << MATCH_CRITERIA_ENABLE_OUTER_BIT)))
4098 		return false;
4099 
4100 	match_c = fs_matcher->matcher_mask.match_params;
4101 	match_v_set_lyr_2_4 = MLX5_ADDR_OF(fte_match_param, match_v,
4102 					   outer_headers);
4103 	match_c_set_lyr_2_4 = MLX5_ADDR_OF(fte_match_param, match_c,
4104 					   outer_headers);
4105 
4106 	dmac = MLX5_ADDR_OF(fte_match_set_lyr_2_4, match_v_set_lyr_2_4,
4107 			    dmac_47_16);
4108 	dmac_mask = MLX5_ADDR_OF(fte_match_set_lyr_2_4, match_c_set_lyr_2_4,
4109 				 dmac_47_16);
4110 
4111 	if (is_multicast_ether_addr(dmac) &&
4112 	    is_multicast_ether_addr(dmac_mask))
4113 		return true;
4114 
4115 	ipv4 = MLX5_ADDR_OF(fte_match_set_lyr_2_4, match_v_set_lyr_2_4,
4116 			    dst_ipv4_dst_ipv6.ipv4_layout.ipv4);
4117 
4118 	ipv4_mask = MLX5_ADDR_OF(fte_match_set_lyr_2_4, match_c_set_lyr_2_4,
4119 				 dst_ipv4_dst_ipv6.ipv4_layout.ipv4);
4120 
4121 	if (ipv4_is_multicast(*(__be32 *)(ipv4)) &&
4122 	    ipv4_is_multicast(*(__be32 *)(ipv4_mask)))
4123 		return true;
4124 
4125 	return false;
4126 }
4127 
4128 struct mlx5_ib_flow_handler *
4129 mlx5_ib_raw_fs_rule_add(struct mlx5_ib_dev *dev,
4130 			struct mlx5_ib_flow_matcher *fs_matcher,
4131 			struct mlx5_flow_context *flow_context,
4132 			struct mlx5_flow_act *flow_act,
4133 			u32 counter_id,
4134 			void *cmd_in, int inlen, int dest_id,
4135 			int dest_type)
4136 {
4137 	struct mlx5_flow_destination *dst;
4138 	struct mlx5_ib_flow_prio *ft_prio;
4139 	struct mlx5_ib_flow_handler *handler;
4140 	int dst_num = 0;
4141 	bool mcast;
4142 	int err;
4143 
4144 	if (fs_matcher->flow_type != MLX5_IB_FLOW_TYPE_NORMAL)
4145 		return ERR_PTR(-EOPNOTSUPP);
4146 
4147 	if (fs_matcher->priority > MLX5_IB_FLOW_LAST_PRIO)
4148 		return ERR_PTR(-ENOMEM);
4149 
4150 	dst = kcalloc(2, sizeof(*dst), GFP_KERNEL);
4151 	if (!dst)
4152 		return ERR_PTR(-ENOMEM);
4153 
4154 	mcast = raw_fs_is_multicast(fs_matcher, cmd_in);
4155 	mutex_lock(&dev->flow_db->lock);
4156 
4157 	ft_prio = _get_flow_table(dev, fs_matcher, mcast);
4158 	if (IS_ERR(ft_prio)) {
4159 		err = PTR_ERR(ft_prio);
4160 		goto unlock;
4161 	}
4162 
4163 	if (dest_type == MLX5_FLOW_DESTINATION_TYPE_TIR) {
4164 		dst[dst_num].type = dest_type;
4165 		dst[dst_num].tir_num = dest_id;
4166 		flow_act->action |= MLX5_FLOW_CONTEXT_ACTION_FWD_DEST;
4167 	} else if (dest_type == MLX5_FLOW_DESTINATION_TYPE_FLOW_TABLE) {
4168 		dst[dst_num].type = MLX5_FLOW_DESTINATION_TYPE_FLOW_TABLE_NUM;
4169 		dst[dst_num].ft_num = dest_id;
4170 		flow_act->action |= MLX5_FLOW_CONTEXT_ACTION_FWD_DEST;
4171 	} else {
4172 		dst[dst_num].type = MLX5_FLOW_DESTINATION_TYPE_PORT;
4173 		flow_act->action |= MLX5_FLOW_CONTEXT_ACTION_ALLOW;
4174 	}
4175 
4176 	dst_num++;
4177 
4178 	if (flow_act->action & MLX5_FLOW_CONTEXT_ACTION_COUNT) {
4179 		dst[dst_num].type = MLX5_FLOW_DESTINATION_TYPE_COUNTER;
4180 		dst[dst_num].counter_id = counter_id;
4181 		dst_num++;
4182 	}
4183 
4184 	handler = _create_raw_flow_rule(dev, ft_prio, dst, fs_matcher,
4185 					flow_context, flow_act,
4186 					cmd_in, inlen, dst_num);
4187 
4188 	if (IS_ERR(handler)) {
4189 		err = PTR_ERR(handler);
4190 		goto destroy_ft;
4191 	}
4192 
4193 	mutex_unlock(&dev->flow_db->lock);
4194 	atomic_inc(&fs_matcher->usecnt);
4195 	handler->flow_matcher = fs_matcher;
4196 
4197 	kfree(dst);
4198 
4199 	return handler;
4200 
4201 destroy_ft:
4202 	put_flow_table(dev, ft_prio, false);
4203 unlock:
4204 	mutex_unlock(&dev->flow_db->lock);
4205 	kfree(dst);
4206 
4207 	return ERR_PTR(err);
4208 }
4209 
4210 static u32 mlx5_ib_flow_action_flags_to_accel_xfrm_flags(u32 mlx5_flags)
4211 {
4212 	u32 flags = 0;
4213 
4214 	if (mlx5_flags & MLX5_IB_UAPI_FLOW_ACTION_FLAGS_REQUIRE_METADATA)
4215 		flags |= MLX5_ACCEL_XFRM_FLAG_REQUIRE_METADATA;
4216 
4217 	return flags;
4218 }
4219 
4220 #define MLX5_FLOW_ACTION_ESP_CREATE_LAST_SUPPORTED	MLX5_IB_UAPI_FLOW_ACTION_FLAGS_REQUIRE_METADATA
4221 static struct ib_flow_action *
4222 mlx5_ib_create_flow_action_esp(struct ib_device *device,
4223 			       const struct ib_flow_action_attrs_esp *attr,
4224 			       struct uverbs_attr_bundle *attrs)
4225 {
4226 	struct mlx5_ib_dev *mdev = to_mdev(device);
4227 	struct ib_uverbs_flow_action_esp_keymat_aes_gcm *aes_gcm;
4228 	struct mlx5_accel_esp_xfrm_attrs accel_attrs = {};
4229 	struct mlx5_ib_flow_action *action;
4230 	u64 action_flags;
4231 	u64 flags;
4232 	int err = 0;
4233 
4234 	err = uverbs_get_flags64(
4235 		&action_flags, attrs, MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
4236 		((MLX5_FLOW_ACTION_ESP_CREATE_LAST_SUPPORTED << 1) - 1));
4237 	if (err)
4238 		return ERR_PTR(err);
4239 
4240 	flags = mlx5_ib_flow_action_flags_to_accel_xfrm_flags(action_flags);
4241 
4242 	/* We current only support a subset of the standard features. Only a
4243 	 * keymat of type AES_GCM, with icv_len == 16, iv_algo == SEQ and esn
4244 	 * (with overlap). Full offload mode isn't supported.
4245 	 */
4246 	if (!attr->keymat || attr->replay || attr->encap ||
4247 	    attr->spi || attr->seq || attr->tfc_pad ||
4248 	    attr->hard_limit_pkts ||
4249 	    (attr->flags & ~(IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
4250 			     IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ENCRYPT)))
4251 		return ERR_PTR(-EOPNOTSUPP);
4252 
4253 	if (attr->keymat->protocol !=
4254 	    IB_UVERBS_FLOW_ACTION_ESP_KEYMAT_AES_GCM)
4255 		return ERR_PTR(-EOPNOTSUPP);
4256 
4257 	aes_gcm = &attr->keymat->keymat.aes_gcm;
4258 
4259 	if (aes_gcm->icv_len != 16 ||
4260 	    aes_gcm->iv_algo != IB_UVERBS_FLOW_ACTION_IV_ALGO_SEQ)
4261 		return ERR_PTR(-EOPNOTSUPP);
4262 
4263 	action = kmalloc(sizeof(*action), GFP_KERNEL);
4264 	if (!action)
4265 		return ERR_PTR(-ENOMEM);
4266 
4267 	action->esp_aes_gcm.ib_flags = attr->flags;
4268 	memcpy(&accel_attrs.keymat.aes_gcm.aes_key, &aes_gcm->aes_key,
4269 	       sizeof(accel_attrs.keymat.aes_gcm.aes_key));
4270 	accel_attrs.keymat.aes_gcm.key_len = aes_gcm->key_len * 8;
4271 	memcpy(&accel_attrs.keymat.aes_gcm.salt, &aes_gcm->salt,
4272 	       sizeof(accel_attrs.keymat.aes_gcm.salt));
4273 	memcpy(&accel_attrs.keymat.aes_gcm.seq_iv, &aes_gcm->iv,
4274 	       sizeof(accel_attrs.keymat.aes_gcm.seq_iv));
4275 	accel_attrs.keymat.aes_gcm.icv_len = aes_gcm->icv_len * 8;
4276 	accel_attrs.keymat.aes_gcm.iv_algo = MLX5_ACCEL_ESP_AES_GCM_IV_ALGO_SEQ;
4277 	accel_attrs.keymat_type = MLX5_ACCEL_ESP_KEYMAT_AES_GCM;
4278 
4279 	accel_attrs.esn = attr->esn;
4280 	if (attr->flags & IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED)
4281 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_TRIGGERED;
4282 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)
4283 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
4284 
4285 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ENCRYPT)
4286 		accel_attrs.action |= MLX5_ACCEL_ESP_ACTION_ENCRYPT;
4287 
4288 	action->esp_aes_gcm.ctx =
4289 		mlx5_accel_esp_create_xfrm(mdev->mdev, &accel_attrs, flags);
4290 	if (IS_ERR(action->esp_aes_gcm.ctx)) {
4291 		err = PTR_ERR(action->esp_aes_gcm.ctx);
4292 		goto err_parse;
4293 	}
4294 
4295 	action->esp_aes_gcm.ib_flags = attr->flags;
4296 
4297 	return &action->ib_action;
4298 
4299 err_parse:
4300 	kfree(action);
4301 	return ERR_PTR(err);
4302 }
4303 
4304 static int
4305 mlx5_ib_modify_flow_action_esp(struct ib_flow_action *action,
4306 			       const struct ib_flow_action_attrs_esp *attr,
4307 			       struct uverbs_attr_bundle *attrs)
4308 {
4309 	struct mlx5_ib_flow_action *maction = to_mflow_act(action);
4310 	struct mlx5_accel_esp_xfrm_attrs accel_attrs;
4311 	int err = 0;
4312 
4313 	if (attr->keymat || attr->replay || attr->encap ||
4314 	    attr->spi || attr->seq || attr->tfc_pad ||
4315 	    attr->hard_limit_pkts ||
4316 	    (attr->flags & ~(IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
4317 			     IB_FLOW_ACTION_ESP_FLAGS_MOD_ESP_ATTRS |
4318 			     IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)))
4319 		return -EOPNOTSUPP;
4320 
4321 	/* Only the ESN value or the MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP can
4322 	 * be modified.
4323 	 */
4324 	if (!(maction->esp_aes_gcm.ib_flags &
4325 	      IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED) &&
4326 	    attr->flags & (IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
4327 			   IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW))
4328 		return -EINVAL;
4329 
4330 	memcpy(&accel_attrs, &maction->esp_aes_gcm.ctx->attrs,
4331 	       sizeof(accel_attrs));
4332 
4333 	accel_attrs.esn = attr->esn;
4334 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)
4335 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
4336 	else
4337 		accel_attrs.flags &= ~MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
4338 
4339 	err = mlx5_accel_esp_modify_xfrm(maction->esp_aes_gcm.ctx,
4340 					 &accel_attrs);
4341 	if (err)
4342 		return err;
4343 
4344 	maction->esp_aes_gcm.ib_flags &=
4345 		~IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW;
4346 	maction->esp_aes_gcm.ib_flags |=
4347 		attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW;
4348 
4349 	return 0;
4350 }
4351 
4352 static int mlx5_ib_destroy_flow_action(struct ib_flow_action *action)
4353 {
4354 	struct mlx5_ib_flow_action *maction = to_mflow_act(action);
4355 
4356 	switch (action->type) {
4357 	case IB_FLOW_ACTION_ESP:
4358 		/*
4359 		 * We only support aes_gcm by now, so we implicitly know this is
4360 		 * the underline crypto.
4361 		 */
4362 		mlx5_accel_esp_destroy_xfrm(maction->esp_aes_gcm.ctx);
4363 		break;
4364 	case IB_FLOW_ACTION_UNSPECIFIED:
4365 		mlx5_ib_destroy_flow_action_raw(maction);
4366 		break;
4367 	default:
4368 		WARN_ON(true);
4369 		break;
4370 	}
4371 
4372 	kfree(maction);
4373 	return 0;
4374 }
4375 
4376 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
4377 {
4378 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4379 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
4380 	int err;
4381 	u16 uid;
4382 
4383 	uid = ibqp->pd ?
4384 		to_mpd(ibqp->pd)->uid : 0;
4385 
4386 	if (mqp->flags & MLX5_IB_QP_UNDERLAY) {
4387 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
4388 		return -EOPNOTSUPP;
4389 	}
4390 
4391 	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
4392 	if (err)
4393 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
4394 			     ibqp->qp_num, gid->raw);
4395 
4396 	return err;
4397 }
4398 
4399 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
4400 {
4401 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4402 	int err;
4403 	u16 uid;
4404 
4405 	uid = ibqp->pd ?
4406 		to_mpd(ibqp->pd)->uid : 0;
4407 	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
4408 	if (err)
4409 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
4410 			     ibqp->qp_num, gid->raw);
4411 
4412 	return err;
4413 }
4414 
4415 static int init_node_data(struct mlx5_ib_dev *dev)
4416 {
4417 	int err;
4418 
4419 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
4420 	if (err)
4421 		return err;
4422 
4423 	dev->mdev->rev_id = dev->mdev->pdev->revision;
4424 
4425 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
4426 }
4427 
4428 static ssize_t fw_pages_show(struct device *device,
4429 			     struct device_attribute *attr, char *buf)
4430 {
4431 	struct mlx5_ib_dev *dev =
4432 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4433 
4434 	return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
4435 }
4436 static DEVICE_ATTR_RO(fw_pages);
4437 
4438 static ssize_t reg_pages_show(struct device *device,
4439 			      struct device_attribute *attr, char *buf)
4440 {
4441 	struct mlx5_ib_dev *dev =
4442 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4443 
4444 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
4445 }
4446 static DEVICE_ATTR_RO(reg_pages);
4447 
4448 static ssize_t hca_type_show(struct device *device,
4449 			     struct device_attribute *attr, char *buf)
4450 {
4451 	struct mlx5_ib_dev *dev =
4452 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4453 
4454 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
4455 }
4456 static DEVICE_ATTR_RO(hca_type);
4457 
4458 static ssize_t hw_rev_show(struct device *device,
4459 			   struct device_attribute *attr, char *buf)
4460 {
4461 	struct mlx5_ib_dev *dev =
4462 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4463 
4464 	return sprintf(buf, "%x\n", dev->mdev->rev_id);
4465 }
4466 static DEVICE_ATTR_RO(hw_rev);
4467 
4468 static ssize_t board_id_show(struct device *device,
4469 			     struct device_attribute *attr, char *buf)
4470 {
4471 	struct mlx5_ib_dev *dev =
4472 		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
4473 
4474 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
4475 		       dev->mdev->board_id);
4476 }
4477 static DEVICE_ATTR_RO(board_id);
4478 
4479 static struct attribute *mlx5_class_attributes[] = {
4480 	&dev_attr_hw_rev.attr,
4481 	&dev_attr_hca_type.attr,
4482 	&dev_attr_board_id.attr,
4483 	&dev_attr_fw_pages.attr,
4484 	&dev_attr_reg_pages.attr,
4485 	NULL,
4486 };
4487 
4488 static const struct attribute_group mlx5_attr_group = {
4489 	.attrs = mlx5_class_attributes,
4490 };
4491 
4492 static void pkey_change_handler(struct work_struct *work)
4493 {
4494 	struct mlx5_ib_port_resources *ports =
4495 		container_of(work, struct mlx5_ib_port_resources,
4496 			     pkey_change_work);
4497 
4498 	mutex_lock(&ports->devr->mutex);
4499 	mlx5_ib_gsi_pkey_change(ports->gsi);
4500 	mutex_unlock(&ports->devr->mutex);
4501 }
4502 
4503 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
4504 {
4505 	struct mlx5_ib_qp *mqp;
4506 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
4507 	struct mlx5_core_cq *mcq;
4508 	struct list_head cq_armed_list;
4509 	unsigned long flags_qp;
4510 	unsigned long flags_cq;
4511 	unsigned long flags;
4512 
4513 	INIT_LIST_HEAD(&cq_armed_list);
4514 
4515 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
4516 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
4517 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
4518 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
4519 		if (mqp->sq.tail != mqp->sq.head) {
4520 			send_mcq = to_mcq(mqp->ibqp.send_cq);
4521 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
4522 			if (send_mcq->mcq.comp &&
4523 			    mqp->ibqp.send_cq->comp_handler) {
4524 				if (!send_mcq->mcq.reset_notify_added) {
4525 					send_mcq->mcq.reset_notify_added = 1;
4526 					list_add_tail(&send_mcq->mcq.reset_notify,
4527 						      &cq_armed_list);
4528 				}
4529 			}
4530 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
4531 		}
4532 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
4533 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
4534 		/* no handling is needed for SRQ */
4535 		if (!mqp->ibqp.srq) {
4536 			if (mqp->rq.tail != mqp->rq.head) {
4537 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
4538 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
4539 				if (recv_mcq->mcq.comp &&
4540 				    mqp->ibqp.recv_cq->comp_handler) {
4541 					if (!recv_mcq->mcq.reset_notify_added) {
4542 						recv_mcq->mcq.reset_notify_added = 1;
4543 						list_add_tail(&recv_mcq->mcq.reset_notify,
4544 							      &cq_armed_list);
4545 					}
4546 				}
4547 				spin_unlock_irqrestore(&recv_mcq->lock,
4548 						       flags_cq);
4549 			}
4550 		}
4551 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
4552 	}
4553 	/*At that point all inflight post send were put to be executed as of we
4554 	 * lock/unlock above locks Now need to arm all involved CQs.
4555 	 */
4556 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
4557 		mcq->comp(mcq, NULL);
4558 	}
4559 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
4560 }
4561 
4562 static void delay_drop_handler(struct work_struct *work)
4563 {
4564 	int err;
4565 	struct mlx5_ib_delay_drop *delay_drop =
4566 		container_of(work, struct mlx5_ib_delay_drop,
4567 			     delay_drop_work);
4568 
4569 	atomic_inc(&delay_drop->events_cnt);
4570 
4571 	mutex_lock(&delay_drop->lock);
4572 	err = mlx5_core_set_delay_drop(delay_drop->dev->mdev,
4573 				       delay_drop->timeout);
4574 	if (err) {
4575 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
4576 			     delay_drop->timeout);
4577 		delay_drop->activate = false;
4578 	}
4579 	mutex_unlock(&delay_drop->lock);
4580 }
4581 
4582 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
4583 				 struct ib_event *ibev)
4584 {
4585 	u8 port = (eqe->data.port.port >> 4) & 0xf;
4586 
4587 	switch (eqe->sub_type) {
4588 	case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
4589 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
4590 					    IB_LINK_LAYER_ETHERNET)
4591 			schedule_work(&ibdev->delay_drop.delay_drop_work);
4592 		break;
4593 	default: /* do nothing */
4594 		return;
4595 	}
4596 }
4597 
4598 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
4599 			      struct ib_event *ibev)
4600 {
4601 	u8 port = (eqe->data.port.port >> 4) & 0xf;
4602 
4603 	ibev->element.port_num = port;
4604 
4605 	switch (eqe->sub_type) {
4606 	case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
4607 	case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
4608 	case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
4609 		/* In RoCE, port up/down events are handled in
4610 		 * mlx5_netdev_event().
4611 		 */
4612 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
4613 					    IB_LINK_LAYER_ETHERNET)
4614 			return -EINVAL;
4615 
4616 		ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
4617 				IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
4618 		break;
4619 
4620 	case MLX5_PORT_CHANGE_SUBTYPE_LID:
4621 		ibev->event = IB_EVENT_LID_CHANGE;
4622 		break;
4623 
4624 	case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
4625 		ibev->event = IB_EVENT_PKEY_CHANGE;
4626 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
4627 		break;
4628 
4629 	case MLX5_PORT_CHANGE_SUBTYPE_GUID:
4630 		ibev->event = IB_EVENT_GID_CHANGE;
4631 		break;
4632 
4633 	case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
4634 		ibev->event = IB_EVENT_CLIENT_REREGISTER;
4635 		break;
4636 	default:
4637 		return -EINVAL;
4638 	}
4639 
4640 	return 0;
4641 }
4642 
4643 static void mlx5_ib_handle_event(struct work_struct *_work)
4644 {
4645 	struct mlx5_ib_event_work *work =
4646 		container_of(_work, struct mlx5_ib_event_work, work);
4647 	struct mlx5_ib_dev *ibdev;
4648 	struct ib_event ibev;
4649 	bool fatal = false;
4650 
4651 	if (work->is_slave) {
4652 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
4653 		if (!ibdev)
4654 			goto out;
4655 	} else {
4656 		ibdev = work->dev;
4657 	}
4658 
4659 	switch (work->event) {
4660 	case MLX5_DEV_EVENT_SYS_ERROR:
4661 		ibev.event = IB_EVENT_DEVICE_FATAL;
4662 		mlx5_ib_handle_internal_error(ibdev);
4663 		ibev.element.port_num  = (u8)(unsigned long)work->param;
4664 		fatal = true;
4665 		break;
4666 	case MLX5_EVENT_TYPE_PORT_CHANGE:
4667 		if (handle_port_change(ibdev, work->param, &ibev))
4668 			goto out;
4669 		break;
4670 	case MLX5_EVENT_TYPE_GENERAL_EVENT:
4671 		handle_general_event(ibdev, work->param, &ibev);
4672 		/* fall through */
4673 	default:
4674 		goto out;
4675 	}
4676 
4677 	ibev.device = &ibdev->ib_dev;
4678 
4679 	if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
4680 		mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
4681 		goto out;
4682 	}
4683 
4684 	if (ibdev->ib_active)
4685 		ib_dispatch_event(&ibev);
4686 
4687 	if (fatal)
4688 		ibdev->ib_active = false;
4689 out:
4690 	kfree(work);
4691 }
4692 
4693 static int mlx5_ib_event(struct notifier_block *nb,
4694 			 unsigned long event, void *param)
4695 {
4696 	struct mlx5_ib_event_work *work;
4697 
4698 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
4699 	if (!work)
4700 		return NOTIFY_DONE;
4701 
4702 	INIT_WORK(&work->work, mlx5_ib_handle_event);
4703 	work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
4704 	work->is_slave = false;
4705 	work->param = param;
4706 	work->event = event;
4707 
4708 	queue_work(mlx5_ib_event_wq, &work->work);
4709 
4710 	return NOTIFY_OK;
4711 }
4712 
4713 static int mlx5_ib_event_slave_port(struct notifier_block *nb,
4714 				    unsigned long event, void *param)
4715 {
4716 	struct mlx5_ib_event_work *work;
4717 
4718 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
4719 	if (!work)
4720 		return NOTIFY_DONE;
4721 
4722 	INIT_WORK(&work->work, mlx5_ib_handle_event);
4723 	work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
4724 	work->is_slave = true;
4725 	work->param = param;
4726 	work->event = event;
4727 	queue_work(mlx5_ib_event_wq, &work->work);
4728 
4729 	return NOTIFY_OK;
4730 }
4731 
4732 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
4733 {
4734 	struct mlx5_hca_vport_context vport_ctx;
4735 	int err;
4736 	int port;
4737 
4738 	for (port = 1; port <= ARRAY_SIZE(dev->mdev->port_caps); port++) {
4739 		dev->mdev->port_caps[port - 1].has_smi = false;
4740 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
4741 		    MLX5_CAP_PORT_TYPE_IB) {
4742 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
4743 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
4744 								   port, 0,
4745 								   &vport_ctx);
4746 				if (err) {
4747 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
4748 						    port, err);
4749 					return err;
4750 				}
4751 				dev->mdev->port_caps[port - 1].has_smi =
4752 					vport_ctx.has_smi;
4753 			} else {
4754 				dev->mdev->port_caps[port - 1].has_smi = true;
4755 			}
4756 		}
4757 	}
4758 	return 0;
4759 }
4760 
4761 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
4762 {
4763 	int port;
4764 
4765 	for (port = 1; port <= dev->num_ports; port++)
4766 		mlx5_query_ext_port_caps(dev, port);
4767 }
4768 
4769 static int __get_port_caps(struct mlx5_ib_dev *dev, u8 port)
4770 {
4771 	struct ib_device_attr *dprops = NULL;
4772 	struct ib_port_attr *pprops = NULL;
4773 	int err = -ENOMEM;
4774 	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
4775 
4776 	pprops = kzalloc(sizeof(*pprops), GFP_KERNEL);
4777 	if (!pprops)
4778 		goto out;
4779 
4780 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
4781 	if (!dprops)
4782 		goto out;
4783 
4784 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
4785 	if (err) {
4786 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
4787 		goto out;
4788 	}
4789 
4790 	err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
4791 	if (err) {
4792 		mlx5_ib_warn(dev, "query_port %d failed %d\n",
4793 			     port, err);
4794 		goto out;
4795 	}
4796 
4797 	dev->mdev->port_caps[port - 1].pkey_table_len =
4798 					dprops->max_pkeys;
4799 	dev->mdev->port_caps[port - 1].gid_table_len =
4800 					pprops->gid_tbl_len;
4801 	mlx5_ib_dbg(dev, "port %d: pkey_table_len %d, gid_table_len %d\n",
4802 		    port, dprops->max_pkeys, pprops->gid_tbl_len);
4803 
4804 out:
4805 	kfree(pprops);
4806 	kfree(dprops);
4807 
4808 	return err;
4809 }
4810 
4811 static int get_port_caps(struct mlx5_ib_dev *dev, u8 port)
4812 {
4813 	/* For representors use port 1, is this is the only native
4814 	 * port
4815 	 */
4816 	if (dev->is_rep)
4817 		return __get_port_caps(dev, 1);
4818 	return __get_port_caps(dev, port);
4819 }
4820 
4821 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
4822 {
4823 	int err;
4824 
4825 	err = mlx5_mr_cache_cleanup(dev);
4826 	if (err)
4827 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
4828 
4829 	if (dev->umrc.qp)
4830 		mlx5_ib_destroy_qp(dev->umrc.qp, NULL);
4831 	if (dev->umrc.cq)
4832 		ib_free_cq(dev->umrc.cq);
4833 	if (dev->umrc.pd)
4834 		ib_dealloc_pd(dev->umrc.pd);
4835 }
4836 
4837 enum {
4838 	MAX_UMR_WR = 128,
4839 };
4840 
4841 static int create_umr_res(struct mlx5_ib_dev *dev)
4842 {
4843 	struct ib_qp_init_attr *init_attr = NULL;
4844 	struct ib_qp_attr *attr = NULL;
4845 	struct ib_pd *pd;
4846 	struct ib_cq *cq;
4847 	struct ib_qp *qp;
4848 	int ret;
4849 
4850 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
4851 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
4852 	if (!attr || !init_attr) {
4853 		ret = -ENOMEM;
4854 		goto error_0;
4855 	}
4856 
4857 	pd = ib_alloc_pd(&dev->ib_dev, 0);
4858 	if (IS_ERR(pd)) {
4859 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
4860 		ret = PTR_ERR(pd);
4861 		goto error_0;
4862 	}
4863 
4864 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
4865 	if (IS_ERR(cq)) {
4866 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
4867 		ret = PTR_ERR(cq);
4868 		goto error_2;
4869 	}
4870 
4871 	init_attr->send_cq = cq;
4872 	init_attr->recv_cq = cq;
4873 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
4874 	init_attr->cap.max_send_wr = MAX_UMR_WR;
4875 	init_attr->cap.max_send_sge = 1;
4876 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
4877 	init_attr->port_num = 1;
4878 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
4879 	if (IS_ERR(qp)) {
4880 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
4881 		ret = PTR_ERR(qp);
4882 		goto error_3;
4883 	}
4884 	qp->device     = &dev->ib_dev;
4885 	qp->real_qp    = qp;
4886 	qp->uobject    = NULL;
4887 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
4888 	qp->send_cq    = init_attr->send_cq;
4889 	qp->recv_cq    = init_attr->recv_cq;
4890 
4891 	attr->qp_state = IB_QPS_INIT;
4892 	attr->port_num = 1;
4893 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
4894 				IB_QP_PORT, NULL);
4895 	if (ret) {
4896 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
4897 		goto error_4;
4898 	}
4899 
4900 	memset(attr, 0, sizeof(*attr));
4901 	attr->qp_state = IB_QPS_RTR;
4902 	attr->path_mtu = IB_MTU_256;
4903 
4904 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4905 	if (ret) {
4906 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
4907 		goto error_4;
4908 	}
4909 
4910 	memset(attr, 0, sizeof(*attr));
4911 	attr->qp_state = IB_QPS_RTS;
4912 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4913 	if (ret) {
4914 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
4915 		goto error_4;
4916 	}
4917 
4918 	dev->umrc.qp = qp;
4919 	dev->umrc.cq = cq;
4920 	dev->umrc.pd = pd;
4921 
4922 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
4923 	ret = mlx5_mr_cache_init(dev);
4924 	if (ret) {
4925 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4926 		goto error_4;
4927 	}
4928 
4929 	kfree(attr);
4930 	kfree(init_attr);
4931 
4932 	return 0;
4933 
4934 error_4:
4935 	mlx5_ib_destroy_qp(qp, NULL);
4936 	dev->umrc.qp = NULL;
4937 
4938 error_3:
4939 	ib_free_cq(cq);
4940 	dev->umrc.cq = NULL;
4941 
4942 error_2:
4943 	ib_dealloc_pd(pd);
4944 	dev->umrc.pd = NULL;
4945 
4946 error_0:
4947 	kfree(attr);
4948 	kfree(init_attr);
4949 	return ret;
4950 }
4951 
4952 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
4953 {
4954 	switch (umr_fence_cap) {
4955 	case MLX5_CAP_UMR_FENCE_NONE:
4956 		return MLX5_FENCE_MODE_NONE;
4957 	case MLX5_CAP_UMR_FENCE_SMALL:
4958 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
4959 	default:
4960 		return MLX5_FENCE_MODE_STRONG_ORDERING;
4961 	}
4962 }
4963 
4964 static int create_dev_resources(struct mlx5_ib_resources *devr)
4965 {
4966 	struct ib_srq_init_attr attr;
4967 	struct mlx5_ib_dev *dev;
4968 	struct ib_device *ibdev;
4969 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
4970 	int port;
4971 	int ret = 0;
4972 
4973 	dev = container_of(devr, struct mlx5_ib_dev, devr);
4974 	ibdev = &dev->ib_dev;
4975 
4976 	mutex_init(&devr->mutex);
4977 
4978 	devr->p0 = rdma_zalloc_drv_obj(ibdev, ib_pd);
4979 	if (!devr->p0)
4980 		return -ENOMEM;
4981 
4982 	devr->p0->device  = ibdev;
4983 	devr->p0->uobject = NULL;
4984 	atomic_set(&devr->p0->usecnt, 0);
4985 
4986 	ret = mlx5_ib_alloc_pd(devr->p0, NULL);
4987 	if (ret)
4988 		goto error0;
4989 
4990 	devr->c0 = rdma_zalloc_drv_obj(ibdev, ib_cq);
4991 	if (!devr->c0) {
4992 		ret = -ENOMEM;
4993 		goto error1;
4994 	}
4995 
4996 	devr->c0->device = &dev->ib_dev;
4997 	atomic_set(&devr->c0->usecnt, 0);
4998 
4999 	ret = mlx5_ib_create_cq(devr->c0, &cq_attr, NULL);
5000 	if (ret)
5001 		goto err_create_cq;
5002 
5003 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL);
5004 	if (IS_ERR(devr->x0)) {
5005 		ret = PTR_ERR(devr->x0);
5006 		goto error2;
5007 	}
5008 	devr->x0->device = &dev->ib_dev;
5009 	devr->x0->inode = NULL;
5010 	atomic_set(&devr->x0->usecnt, 0);
5011 	mutex_init(&devr->x0->tgt_qp_mutex);
5012 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
5013 
5014 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL);
5015 	if (IS_ERR(devr->x1)) {
5016 		ret = PTR_ERR(devr->x1);
5017 		goto error3;
5018 	}
5019 	devr->x1->device = &dev->ib_dev;
5020 	devr->x1->inode = NULL;
5021 	atomic_set(&devr->x1->usecnt, 0);
5022 	mutex_init(&devr->x1->tgt_qp_mutex);
5023 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
5024 
5025 	memset(&attr, 0, sizeof(attr));
5026 	attr.attr.max_sge = 1;
5027 	attr.attr.max_wr = 1;
5028 	attr.srq_type = IB_SRQT_XRC;
5029 	attr.ext.cq = devr->c0;
5030 	attr.ext.xrc.xrcd = devr->x0;
5031 
5032 	devr->s0 = rdma_zalloc_drv_obj(ibdev, ib_srq);
5033 	if (!devr->s0) {
5034 		ret = -ENOMEM;
5035 		goto error4;
5036 	}
5037 
5038 	devr->s0->device	= &dev->ib_dev;
5039 	devr->s0->pd		= devr->p0;
5040 	devr->s0->srq_type      = IB_SRQT_XRC;
5041 	devr->s0->ext.xrc.xrcd	= devr->x0;
5042 	devr->s0->ext.cq	= devr->c0;
5043 	ret = mlx5_ib_create_srq(devr->s0, &attr, NULL);
5044 	if (ret)
5045 		goto err_create;
5046 
5047 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
5048 	atomic_inc(&devr->s0->ext.cq->usecnt);
5049 	atomic_inc(&devr->p0->usecnt);
5050 	atomic_set(&devr->s0->usecnt, 0);
5051 
5052 	memset(&attr, 0, sizeof(attr));
5053 	attr.attr.max_sge = 1;
5054 	attr.attr.max_wr = 1;
5055 	attr.srq_type = IB_SRQT_BASIC;
5056 	devr->s1 = rdma_zalloc_drv_obj(ibdev, ib_srq);
5057 	if (!devr->s1) {
5058 		ret = -ENOMEM;
5059 		goto error5;
5060 	}
5061 
5062 	devr->s1->device	= &dev->ib_dev;
5063 	devr->s1->pd		= devr->p0;
5064 	devr->s1->srq_type      = IB_SRQT_BASIC;
5065 	devr->s1->ext.cq	= devr->c0;
5066 
5067 	ret = mlx5_ib_create_srq(devr->s1, &attr, NULL);
5068 	if (ret)
5069 		goto error6;
5070 
5071 	atomic_inc(&devr->p0->usecnt);
5072 	atomic_set(&devr->s1->usecnt, 0);
5073 
5074 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
5075 		INIT_WORK(&devr->ports[port].pkey_change_work,
5076 			  pkey_change_handler);
5077 		devr->ports[port].devr = devr;
5078 	}
5079 
5080 	return 0;
5081 
5082 error6:
5083 	kfree(devr->s1);
5084 error5:
5085 	mlx5_ib_destroy_srq(devr->s0, NULL);
5086 err_create:
5087 	kfree(devr->s0);
5088 error4:
5089 	mlx5_ib_dealloc_xrcd(devr->x1, NULL);
5090 error3:
5091 	mlx5_ib_dealloc_xrcd(devr->x0, NULL);
5092 error2:
5093 	mlx5_ib_destroy_cq(devr->c0, NULL);
5094 err_create_cq:
5095 	kfree(devr->c0);
5096 error1:
5097 	mlx5_ib_dealloc_pd(devr->p0, NULL);
5098 error0:
5099 	kfree(devr->p0);
5100 	return ret;
5101 }
5102 
5103 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
5104 {
5105 	int port;
5106 
5107 	mlx5_ib_destroy_srq(devr->s1, NULL);
5108 	kfree(devr->s1);
5109 	mlx5_ib_destroy_srq(devr->s0, NULL);
5110 	kfree(devr->s0);
5111 	mlx5_ib_dealloc_xrcd(devr->x0, NULL);
5112 	mlx5_ib_dealloc_xrcd(devr->x1, NULL);
5113 	mlx5_ib_destroy_cq(devr->c0, NULL);
5114 	kfree(devr->c0);
5115 	mlx5_ib_dealloc_pd(devr->p0, NULL);
5116 	kfree(devr->p0);
5117 
5118 	/* Make sure no change P_Key work items are still executing */
5119 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
5120 		cancel_work_sync(&devr->ports[port].pkey_change_work);
5121 }
5122 
5123 static u32 get_core_cap_flags(struct ib_device *ibdev,
5124 			      struct mlx5_hca_vport_context *rep)
5125 {
5126 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5127 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
5128 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
5129 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
5130 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
5131 	u32 ret = 0;
5132 
5133 	if (rep->grh_required)
5134 		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
5135 
5136 	if (ll == IB_LINK_LAYER_INFINIBAND)
5137 		return ret | RDMA_CORE_PORT_IBA_IB;
5138 
5139 	if (raw_support)
5140 		ret |= RDMA_CORE_PORT_RAW_PACKET;
5141 
5142 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
5143 		return ret;
5144 
5145 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
5146 		return ret;
5147 
5148 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
5149 		ret |= RDMA_CORE_PORT_IBA_ROCE;
5150 
5151 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
5152 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
5153 
5154 	return ret;
5155 }
5156 
5157 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
5158 			       struct ib_port_immutable *immutable)
5159 {
5160 	struct ib_port_attr attr;
5161 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5162 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
5163 	struct mlx5_hca_vport_context rep = {0};
5164 	int err;
5165 
5166 	err = ib_query_port(ibdev, port_num, &attr);
5167 	if (err)
5168 		return err;
5169 
5170 	if (ll == IB_LINK_LAYER_INFINIBAND) {
5171 		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
5172 						   &rep);
5173 		if (err)
5174 			return err;
5175 	}
5176 
5177 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
5178 	immutable->gid_tbl_len = attr.gid_tbl_len;
5179 	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
5180 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
5181 
5182 	return 0;
5183 }
5184 
5185 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u8 port_num,
5186 				   struct ib_port_immutable *immutable)
5187 {
5188 	struct ib_port_attr attr;
5189 	int err;
5190 
5191 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
5192 
5193 	err = ib_query_port(ibdev, port_num, &attr);
5194 	if (err)
5195 		return err;
5196 
5197 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
5198 	immutable->gid_tbl_len = attr.gid_tbl_len;
5199 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
5200 
5201 	return 0;
5202 }
5203 
5204 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
5205 {
5206 	struct mlx5_ib_dev *dev =
5207 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
5208 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
5209 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
5210 		 fw_rev_sub(dev->mdev));
5211 }
5212 
5213 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
5214 {
5215 	struct mlx5_core_dev *mdev = dev->mdev;
5216 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
5217 								 MLX5_FLOW_NAMESPACE_LAG);
5218 	struct mlx5_flow_table *ft;
5219 	int err;
5220 
5221 	if (!ns || !mlx5_lag_is_roce(mdev))
5222 		return 0;
5223 
5224 	err = mlx5_cmd_create_vport_lag(mdev);
5225 	if (err)
5226 		return err;
5227 
5228 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
5229 	if (IS_ERR(ft)) {
5230 		err = PTR_ERR(ft);
5231 		goto err_destroy_vport_lag;
5232 	}
5233 
5234 	dev->flow_db->lag_demux_ft = ft;
5235 	dev->lag_active = true;
5236 	return 0;
5237 
5238 err_destroy_vport_lag:
5239 	mlx5_cmd_destroy_vport_lag(mdev);
5240 	return err;
5241 }
5242 
5243 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
5244 {
5245 	struct mlx5_core_dev *mdev = dev->mdev;
5246 
5247 	if (dev->lag_active) {
5248 		dev->lag_active = false;
5249 
5250 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
5251 		dev->flow_db->lag_demux_ft = NULL;
5252 
5253 		mlx5_cmd_destroy_vport_lag(mdev);
5254 	}
5255 }
5256 
5257 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
5258 {
5259 	int err;
5260 
5261 	dev->port[port_num].roce.nb.notifier_call = mlx5_netdev_event;
5262 	err = register_netdevice_notifier(&dev->port[port_num].roce.nb);
5263 	if (err) {
5264 		dev->port[port_num].roce.nb.notifier_call = NULL;
5265 		return err;
5266 	}
5267 
5268 	return 0;
5269 }
5270 
5271 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
5272 {
5273 	if (dev->port[port_num].roce.nb.notifier_call) {
5274 		unregister_netdevice_notifier(&dev->port[port_num].roce.nb);
5275 		dev->port[port_num].roce.nb.notifier_call = NULL;
5276 	}
5277 }
5278 
5279 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
5280 {
5281 	int err;
5282 
5283 	err = mlx5_nic_vport_enable_roce(dev->mdev);
5284 	if (err)
5285 		return err;
5286 
5287 	err = mlx5_eth_lag_init(dev);
5288 	if (err)
5289 		goto err_disable_roce;
5290 
5291 	return 0;
5292 
5293 err_disable_roce:
5294 	mlx5_nic_vport_disable_roce(dev->mdev);
5295 
5296 	return err;
5297 }
5298 
5299 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
5300 {
5301 	mlx5_eth_lag_cleanup(dev);
5302 	mlx5_nic_vport_disable_roce(dev->mdev);
5303 }
5304 
5305 struct mlx5_ib_counter {
5306 	const char *name;
5307 	size_t offset;
5308 };
5309 
5310 #define INIT_Q_COUNTER(_name)		\
5311 	{ .name = #_name, .offset = MLX5_BYTE_OFF(query_q_counter_out, _name)}
5312 
5313 static const struct mlx5_ib_counter basic_q_cnts[] = {
5314 	INIT_Q_COUNTER(rx_write_requests),
5315 	INIT_Q_COUNTER(rx_read_requests),
5316 	INIT_Q_COUNTER(rx_atomic_requests),
5317 	INIT_Q_COUNTER(out_of_buffer),
5318 };
5319 
5320 static const struct mlx5_ib_counter out_of_seq_q_cnts[] = {
5321 	INIT_Q_COUNTER(out_of_sequence),
5322 };
5323 
5324 static const struct mlx5_ib_counter retrans_q_cnts[] = {
5325 	INIT_Q_COUNTER(duplicate_request),
5326 	INIT_Q_COUNTER(rnr_nak_retry_err),
5327 	INIT_Q_COUNTER(packet_seq_err),
5328 	INIT_Q_COUNTER(implied_nak_seq_err),
5329 	INIT_Q_COUNTER(local_ack_timeout_err),
5330 };
5331 
5332 #define INIT_CONG_COUNTER(_name)		\
5333 	{ .name = #_name, .offset =	\
5334 		MLX5_BYTE_OFF(query_cong_statistics_out, _name ## _high)}
5335 
5336 static const struct mlx5_ib_counter cong_cnts[] = {
5337 	INIT_CONG_COUNTER(rp_cnp_ignored),
5338 	INIT_CONG_COUNTER(rp_cnp_handled),
5339 	INIT_CONG_COUNTER(np_ecn_marked_roce_packets),
5340 	INIT_CONG_COUNTER(np_cnp_sent),
5341 };
5342 
5343 static const struct mlx5_ib_counter extended_err_cnts[] = {
5344 	INIT_Q_COUNTER(resp_local_length_error),
5345 	INIT_Q_COUNTER(resp_cqe_error),
5346 	INIT_Q_COUNTER(req_cqe_error),
5347 	INIT_Q_COUNTER(req_remote_invalid_request),
5348 	INIT_Q_COUNTER(req_remote_access_errors),
5349 	INIT_Q_COUNTER(resp_remote_access_errors),
5350 	INIT_Q_COUNTER(resp_cqe_flush_error),
5351 	INIT_Q_COUNTER(req_cqe_flush_error),
5352 };
5353 
5354 #define INIT_EXT_PPCNT_COUNTER(_name)		\
5355 	{ .name = #_name, .offset =	\
5356 	MLX5_BYTE_OFF(ppcnt_reg, \
5357 		      counter_set.eth_extended_cntrs_grp_data_layout._name##_high)}
5358 
5359 static const struct mlx5_ib_counter ext_ppcnt_cnts[] = {
5360 	INIT_EXT_PPCNT_COUNTER(rx_icrc_encapsulated),
5361 };
5362 
5363 static bool is_mdev_switchdev_mode(const struct mlx5_core_dev *mdev)
5364 {
5365 	return MLX5_ESWITCH_MANAGER(mdev) &&
5366 	       mlx5_ib_eswitch_mode(mdev->priv.eswitch) ==
5367 		       MLX5_ESWITCH_OFFLOADS;
5368 }
5369 
5370 static void mlx5_ib_dealloc_counters(struct mlx5_ib_dev *dev)
5371 {
5372 	int num_cnt_ports;
5373 	int i;
5374 
5375 	num_cnt_ports = is_mdev_switchdev_mode(dev->mdev) ? 1 : dev->num_ports;
5376 
5377 	for (i = 0; i < num_cnt_ports; i++) {
5378 		if (dev->port[i].cnts.set_id_valid)
5379 			mlx5_core_dealloc_q_counter(dev->mdev,
5380 						    dev->port[i].cnts.set_id);
5381 		kfree(dev->port[i].cnts.names);
5382 		kfree(dev->port[i].cnts.offsets);
5383 	}
5384 }
5385 
5386 static int __mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev,
5387 				    struct mlx5_ib_counters *cnts)
5388 {
5389 	u32 num_counters;
5390 
5391 	num_counters = ARRAY_SIZE(basic_q_cnts);
5392 
5393 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt))
5394 		num_counters += ARRAY_SIZE(out_of_seq_q_cnts);
5395 
5396 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters))
5397 		num_counters += ARRAY_SIZE(retrans_q_cnts);
5398 
5399 	if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters))
5400 		num_counters += ARRAY_SIZE(extended_err_cnts);
5401 
5402 	cnts->num_q_counters = num_counters;
5403 
5404 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
5405 		cnts->num_cong_counters = ARRAY_SIZE(cong_cnts);
5406 		num_counters += ARRAY_SIZE(cong_cnts);
5407 	}
5408 	if (MLX5_CAP_PCAM_FEATURE(dev->mdev, rx_icrc_encapsulated_counter)) {
5409 		cnts->num_ext_ppcnt_counters = ARRAY_SIZE(ext_ppcnt_cnts);
5410 		num_counters += ARRAY_SIZE(ext_ppcnt_cnts);
5411 	}
5412 	cnts->names = kcalloc(num_counters, sizeof(cnts->names), GFP_KERNEL);
5413 	if (!cnts->names)
5414 		return -ENOMEM;
5415 
5416 	cnts->offsets = kcalloc(num_counters,
5417 				sizeof(cnts->offsets), GFP_KERNEL);
5418 	if (!cnts->offsets)
5419 		goto err_names;
5420 
5421 	return 0;
5422 
5423 err_names:
5424 	kfree(cnts->names);
5425 	cnts->names = NULL;
5426 	return -ENOMEM;
5427 }
5428 
5429 static void mlx5_ib_fill_counters(struct mlx5_ib_dev *dev,
5430 				  const char **names,
5431 				  size_t *offsets)
5432 {
5433 	int i;
5434 	int j = 0;
5435 
5436 	for (i = 0; i < ARRAY_SIZE(basic_q_cnts); i++, j++) {
5437 		names[j] = basic_q_cnts[i].name;
5438 		offsets[j] = basic_q_cnts[i].offset;
5439 	}
5440 
5441 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt)) {
5442 		for (i = 0; i < ARRAY_SIZE(out_of_seq_q_cnts); i++, j++) {
5443 			names[j] = out_of_seq_q_cnts[i].name;
5444 			offsets[j] = out_of_seq_q_cnts[i].offset;
5445 		}
5446 	}
5447 
5448 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
5449 		for (i = 0; i < ARRAY_SIZE(retrans_q_cnts); i++, j++) {
5450 			names[j] = retrans_q_cnts[i].name;
5451 			offsets[j] = retrans_q_cnts[i].offset;
5452 		}
5453 	}
5454 
5455 	if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters)) {
5456 		for (i = 0; i < ARRAY_SIZE(extended_err_cnts); i++, j++) {
5457 			names[j] = extended_err_cnts[i].name;
5458 			offsets[j] = extended_err_cnts[i].offset;
5459 		}
5460 	}
5461 
5462 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
5463 		for (i = 0; i < ARRAY_SIZE(cong_cnts); i++, j++) {
5464 			names[j] = cong_cnts[i].name;
5465 			offsets[j] = cong_cnts[i].offset;
5466 		}
5467 	}
5468 
5469 	if (MLX5_CAP_PCAM_FEATURE(dev->mdev, rx_icrc_encapsulated_counter)) {
5470 		for (i = 0; i < ARRAY_SIZE(ext_ppcnt_cnts); i++, j++) {
5471 			names[j] = ext_ppcnt_cnts[i].name;
5472 			offsets[j] = ext_ppcnt_cnts[i].offset;
5473 		}
5474 	}
5475 }
5476 
5477 static int mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev)
5478 {
5479 	int num_cnt_ports;
5480 	int err = 0;
5481 	int i;
5482 	bool is_shared;
5483 
5484 	is_shared = MLX5_CAP_GEN(dev->mdev, log_max_uctx) != 0;
5485 	num_cnt_ports = is_mdev_switchdev_mode(dev->mdev) ? 1 : dev->num_ports;
5486 
5487 	for (i = 0; i < num_cnt_ports; i++) {
5488 		err = __mlx5_ib_alloc_counters(dev, &dev->port[i].cnts);
5489 		if (err)
5490 			goto err_alloc;
5491 
5492 		mlx5_ib_fill_counters(dev, dev->port[i].cnts.names,
5493 				      dev->port[i].cnts.offsets);
5494 
5495 		err = mlx5_cmd_alloc_q_counter(dev->mdev,
5496 					       &dev->port[i].cnts.set_id,
5497 					       is_shared ?
5498 					       MLX5_SHARED_RESOURCE_UID : 0);
5499 		if (err) {
5500 			mlx5_ib_warn(dev,
5501 				     "couldn't allocate queue counter for port %d, err %d\n",
5502 				     i + 1, err);
5503 			goto err_alloc;
5504 		}
5505 		dev->port[i].cnts.set_id_valid = true;
5506 	}
5507 	return 0;
5508 
5509 err_alloc:
5510 	mlx5_ib_dealloc_counters(dev);
5511 	return err;
5512 }
5513 
5514 static const struct mlx5_ib_counters *get_counters(struct mlx5_ib_dev *dev,
5515 						   u8 port_num)
5516 {
5517 	return is_mdev_switchdev_mode(dev->mdev) ? &dev->port[0].cnts :
5518 						   &dev->port[port_num].cnts;
5519 }
5520 
5521 /**
5522  * mlx5_ib_get_counters_id - Returns counters id to use for device+port
5523  * @dev:	Pointer to mlx5 IB device
5524  * @port_num:	Zero based port number
5525  *
5526  * mlx5_ib_get_counters_id() Returns counters set id to use for given
5527  * device port combination in switchdev and non switchdev mode of the
5528  * parent device.
5529  */
5530 u16 mlx5_ib_get_counters_id(struct mlx5_ib_dev *dev, u8 port_num)
5531 {
5532 	const struct mlx5_ib_counters *cnts = get_counters(dev, port_num);
5533 
5534 	return cnts->set_id;
5535 }
5536 
5537 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
5538 						    u8 port_num)
5539 {
5540 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5541 	const struct mlx5_ib_counters *cnts;
5542 	bool is_switchdev = is_mdev_switchdev_mode(dev->mdev);
5543 
5544 	if ((is_switchdev && port_num) || (!is_switchdev && !port_num))
5545 		return NULL;
5546 
5547 	cnts = get_counters(dev, port_num - 1);
5548 
5549 	return rdma_alloc_hw_stats_struct(cnts->names,
5550 					  cnts->num_q_counters +
5551 					  cnts->num_cong_counters +
5552 					  cnts->num_ext_ppcnt_counters,
5553 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
5554 }
5555 
5556 static int mlx5_ib_query_q_counters(struct mlx5_core_dev *mdev,
5557 				    const struct mlx5_ib_counters *cnts,
5558 				    struct rdma_hw_stats *stats,
5559 				    u16 set_id)
5560 {
5561 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
5562 	void *out;
5563 	__be32 val;
5564 	int ret, i;
5565 
5566 	out = kvzalloc(outlen, GFP_KERNEL);
5567 	if (!out)
5568 		return -ENOMEM;
5569 
5570 	ret = mlx5_core_query_q_counter(mdev, set_id, 0, out, outlen);
5571 	if (ret)
5572 		goto free;
5573 
5574 	for (i = 0; i < cnts->num_q_counters; i++) {
5575 		val = *(__be32 *)(out + cnts->offsets[i]);
5576 		stats->value[i] = (u64)be32_to_cpu(val);
5577 	}
5578 
5579 free:
5580 	kvfree(out);
5581 	return ret;
5582 }
5583 
5584 static int mlx5_ib_query_ext_ppcnt_counters(struct mlx5_ib_dev *dev,
5585 					    const struct mlx5_ib_counters *cnts,
5586 					    struct rdma_hw_stats *stats)
5587 {
5588 	int offset = cnts->num_q_counters + cnts->num_cong_counters;
5589 	int sz = MLX5_ST_SZ_BYTES(ppcnt_reg);
5590 	int ret, i;
5591 	void *out;
5592 
5593 	out = kvzalloc(sz, GFP_KERNEL);
5594 	if (!out)
5595 		return -ENOMEM;
5596 
5597 	ret = mlx5_cmd_query_ext_ppcnt_counters(dev->mdev, out);
5598 	if (ret)
5599 		goto free;
5600 
5601 	for (i = 0; i < cnts->num_ext_ppcnt_counters; i++)
5602 		stats->value[i + offset] =
5603 			be64_to_cpup((__be64 *)(out +
5604 				    cnts->offsets[i + offset]));
5605 free:
5606 	kvfree(out);
5607 	return ret;
5608 }
5609 
5610 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
5611 				struct rdma_hw_stats *stats,
5612 				u8 port_num, int index)
5613 {
5614 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5615 	const struct mlx5_ib_counters *cnts = get_counters(dev, port_num - 1);
5616 	struct mlx5_core_dev *mdev;
5617 	int ret, num_counters;
5618 	u8 mdev_port_num;
5619 
5620 	if (!stats)
5621 		return -EINVAL;
5622 
5623 	num_counters = cnts->num_q_counters +
5624 		       cnts->num_cong_counters +
5625 		       cnts->num_ext_ppcnt_counters;
5626 
5627 	/* q_counters are per IB device, query the master mdev */
5628 	ret = mlx5_ib_query_q_counters(dev->mdev, cnts, stats, cnts->set_id);
5629 	if (ret)
5630 		return ret;
5631 
5632 	if (MLX5_CAP_PCAM_FEATURE(dev->mdev, rx_icrc_encapsulated_counter)) {
5633 		ret =  mlx5_ib_query_ext_ppcnt_counters(dev, cnts, stats);
5634 		if (ret)
5635 			return ret;
5636 	}
5637 
5638 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
5639 		mdev = mlx5_ib_get_native_port_mdev(dev, port_num,
5640 						    &mdev_port_num);
5641 		if (!mdev) {
5642 			/* If port is not affiliated yet, its in down state
5643 			 * which doesn't have any counters yet, so it would be
5644 			 * zero. So no need to read from the HCA.
5645 			 */
5646 			goto done;
5647 		}
5648 		ret = mlx5_lag_query_cong_counters(dev->mdev,
5649 						   stats->value +
5650 						   cnts->num_q_counters,
5651 						   cnts->num_cong_counters,
5652 						   cnts->offsets +
5653 						   cnts->num_q_counters);
5654 
5655 		mlx5_ib_put_native_port_mdev(dev, port_num);
5656 		if (ret)
5657 			return ret;
5658 	}
5659 
5660 done:
5661 	return num_counters;
5662 }
5663 
5664 static struct rdma_hw_stats *
5665 mlx5_ib_counter_alloc_stats(struct rdma_counter *counter)
5666 {
5667 	struct mlx5_ib_dev *dev = to_mdev(counter->device);
5668 	const struct mlx5_ib_counters *cnts =
5669 		get_counters(dev, counter->port - 1);
5670 
5671 	/* Q counters are in the beginning of all counters */
5672 	return rdma_alloc_hw_stats_struct(cnts->names,
5673 					  cnts->num_q_counters,
5674 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
5675 }
5676 
5677 static int mlx5_ib_counter_update_stats(struct rdma_counter *counter)
5678 {
5679 	struct mlx5_ib_dev *dev = to_mdev(counter->device);
5680 	const struct mlx5_ib_counters *cnts =
5681 		get_counters(dev, counter->port - 1);
5682 
5683 	return mlx5_ib_query_q_counters(dev->mdev, cnts,
5684 					counter->stats, counter->id);
5685 }
5686 
5687 static int mlx5_ib_counter_bind_qp(struct rdma_counter *counter,
5688 				   struct ib_qp *qp)
5689 {
5690 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5691 	u16 cnt_set_id = 0;
5692 	int err;
5693 
5694 	if (!counter->id) {
5695 		err = mlx5_cmd_alloc_q_counter(dev->mdev,
5696 					       &cnt_set_id,
5697 					       MLX5_SHARED_RESOURCE_UID);
5698 		if (err)
5699 			return err;
5700 		counter->id = cnt_set_id;
5701 	}
5702 
5703 	err = mlx5_ib_qp_set_counter(qp, counter);
5704 	if (err)
5705 		goto fail_set_counter;
5706 
5707 	return 0;
5708 
5709 fail_set_counter:
5710 	mlx5_core_dealloc_q_counter(dev->mdev, cnt_set_id);
5711 	counter->id = 0;
5712 
5713 	return err;
5714 }
5715 
5716 static int mlx5_ib_counter_unbind_qp(struct ib_qp *qp)
5717 {
5718 	return mlx5_ib_qp_set_counter(qp, NULL);
5719 }
5720 
5721 static int mlx5_ib_counter_dealloc(struct rdma_counter *counter)
5722 {
5723 	struct mlx5_ib_dev *dev = to_mdev(counter->device);
5724 
5725 	return mlx5_core_dealloc_q_counter(dev->mdev, counter->id);
5726 }
5727 
5728 static int mlx5_ib_rn_get_params(struct ib_device *device, u8 port_num,
5729 				 enum rdma_netdev_t type,
5730 				 struct rdma_netdev_alloc_params *params)
5731 {
5732 	if (type != RDMA_NETDEV_IPOIB)
5733 		return -EOPNOTSUPP;
5734 
5735 	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
5736 }
5737 
5738 static void delay_drop_debugfs_cleanup(struct mlx5_ib_dev *dev)
5739 {
5740 	if (!dev->delay_drop.dir_debugfs)
5741 		return;
5742 	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
5743 	dev->delay_drop.dir_debugfs = NULL;
5744 }
5745 
5746 static void cancel_delay_drop(struct mlx5_ib_dev *dev)
5747 {
5748 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
5749 		return;
5750 
5751 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
5752 	delay_drop_debugfs_cleanup(dev);
5753 }
5754 
5755 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
5756 				       size_t count, loff_t *pos)
5757 {
5758 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
5759 	char lbuf[20];
5760 	int len;
5761 
5762 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
5763 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
5764 }
5765 
5766 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
5767 					size_t count, loff_t *pos)
5768 {
5769 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
5770 	u32 timeout;
5771 	u32 var;
5772 
5773 	if (kstrtouint_from_user(buf, count, 0, &var))
5774 		return -EFAULT;
5775 
5776 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
5777 			1000);
5778 	if (timeout != var)
5779 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
5780 			    timeout);
5781 
5782 	delay_drop->timeout = timeout;
5783 
5784 	return count;
5785 }
5786 
5787 static const struct file_operations fops_delay_drop_timeout = {
5788 	.owner	= THIS_MODULE,
5789 	.open	= simple_open,
5790 	.write	= delay_drop_timeout_write,
5791 	.read	= delay_drop_timeout_read,
5792 };
5793 
5794 static void delay_drop_debugfs_init(struct mlx5_ib_dev *dev)
5795 {
5796 	struct dentry *root;
5797 
5798 	if (!mlx5_debugfs_root)
5799 		return;
5800 
5801 	root = debugfs_create_dir("delay_drop", dev->mdev->priv.dbg_root);
5802 	dev->delay_drop.dir_debugfs = root;
5803 
5804 	debugfs_create_atomic_t("num_timeout_events", 0400, root,
5805 				&dev->delay_drop.events_cnt);
5806 	debugfs_create_atomic_t("num_rqs", 0400, root,
5807 				&dev->delay_drop.rqs_cnt);
5808 	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
5809 			    &fops_delay_drop_timeout);
5810 }
5811 
5812 static void init_delay_drop(struct mlx5_ib_dev *dev)
5813 {
5814 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
5815 		return;
5816 
5817 	mutex_init(&dev->delay_drop.lock);
5818 	dev->delay_drop.dev = dev;
5819 	dev->delay_drop.activate = false;
5820 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
5821 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
5822 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
5823 	atomic_set(&dev->delay_drop.events_cnt, 0);
5824 
5825 	delay_drop_debugfs_init(dev);
5826 }
5827 
5828 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
5829 				      struct mlx5_ib_multiport_info *mpi)
5830 {
5831 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
5832 	struct mlx5_ib_port *port = &ibdev->port[port_num];
5833 	int comps;
5834 	int err;
5835 	int i;
5836 
5837 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
5838 
5839 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
5840 
5841 	spin_lock(&port->mp.mpi_lock);
5842 	if (!mpi->ibdev) {
5843 		spin_unlock(&port->mp.mpi_lock);
5844 		return;
5845 	}
5846 
5847 	mpi->ibdev = NULL;
5848 
5849 	spin_unlock(&port->mp.mpi_lock);
5850 	if (mpi->mdev_events.notifier_call)
5851 		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
5852 	mpi->mdev_events.notifier_call = NULL;
5853 	mlx5_remove_netdev_notifier(ibdev, port_num);
5854 	spin_lock(&port->mp.mpi_lock);
5855 
5856 	comps = mpi->mdev_refcnt;
5857 	if (comps) {
5858 		mpi->unaffiliate = true;
5859 		init_completion(&mpi->unref_comp);
5860 		spin_unlock(&port->mp.mpi_lock);
5861 
5862 		for (i = 0; i < comps; i++)
5863 			wait_for_completion(&mpi->unref_comp);
5864 
5865 		spin_lock(&port->mp.mpi_lock);
5866 		mpi->unaffiliate = false;
5867 	}
5868 
5869 	port->mp.mpi = NULL;
5870 
5871 	list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
5872 
5873 	spin_unlock(&port->mp.mpi_lock);
5874 
5875 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
5876 
5877 	mlx5_ib_dbg(ibdev, "unaffiliated port %d\n", port_num + 1);
5878 	/* Log an error, still needed to cleanup the pointers and add
5879 	 * it back to the list.
5880 	 */
5881 	if (err)
5882 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
5883 			    port_num + 1);
5884 
5885 	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
5886 }
5887 
5888 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
5889 				    struct mlx5_ib_multiport_info *mpi)
5890 {
5891 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
5892 	int err;
5893 
5894 	lockdep_assert_held(&mlx5_ib_multiport_mutex);
5895 
5896 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
5897 	if (ibdev->port[port_num].mp.mpi) {
5898 		mlx5_ib_dbg(ibdev, "port %d already affiliated.\n",
5899 			    port_num + 1);
5900 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
5901 		return false;
5902 	}
5903 
5904 	ibdev->port[port_num].mp.mpi = mpi;
5905 	mpi->ibdev = ibdev;
5906 	mpi->mdev_events.notifier_call = NULL;
5907 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
5908 
5909 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
5910 	if (err)
5911 		goto unbind;
5912 
5913 	err = get_port_caps(ibdev, mlx5_core_native_port_num(mpi->mdev));
5914 	if (err)
5915 		goto unbind;
5916 
5917 	err = mlx5_add_netdev_notifier(ibdev, port_num);
5918 	if (err) {
5919 		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
5920 			    port_num + 1);
5921 		goto unbind;
5922 	}
5923 
5924 	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
5925 	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
5926 
5927 	mlx5_ib_init_cong_debugfs(ibdev, port_num);
5928 
5929 	return true;
5930 
5931 unbind:
5932 	mlx5_ib_unbind_slave_port(ibdev, mpi);
5933 	return false;
5934 }
5935 
5936 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
5937 {
5938 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5939 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
5940 							  port_num + 1);
5941 	struct mlx5_ib_multiport_info *mpi;
5942 	int err;
5943 	int i;
5944 
5945 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
5946 		return 0;
5947 
5948 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
5949 						     &dev->sys_image_guid);
5950 	if (err)
5951 		return err;
5952 
5953 	err = mlx5_nic_vport_enable_roce(dev->mdev);
5954 	if (err)
5955 		return err;
5956 
5957 	mutex_lock(&mlx5_ib_multiport_mutex);
5958 	for (i = 0; i < dev->num_ports; i++) {
5959 		bool bound = false;
5960 
5961 		/* build a stub multiport info struct for the native port. */
5962 		if (i == port_num) {
5963 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
5964 			if (!mpi) {
5965 				mutex_unlock(&mlx5_ib_multiport_mutex);
5966 				mlx5_nic_vport_disable_roce(dev->mdev);
5967 				return -ENOMEM;
5968 			}
5969 
5970 			mpi->is_master = true;
5971 			mpi->mdev = dev->mdev;
5972 			mpi->sys_image_guid = dev->sys_image_guid;
5973 			dev->port[i].mp.mpi = mpi;
5974 			mpi->ibdev = dev;
5975 			mpi = NULL;
5976 			continue;
5977 		}
5978 
5979 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
5980 				    list) {
5981 			if (dev->sys_image_guid == mpi->sys_image_guid &&
5982 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
5983 				bound = mlx5_ib_bind_slave_port(dev, mpi);
5984 			}
5985 
5986 			if (bound) {
5987 				dev_dbg(mpi->mdev->device,
5988 					"removing port from unaffiliated list.\n");
5989 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
5990 				list_del(&mpi->list);
5991 				break;
5992 			}
5993 		}
5994 		if (!bound) {
5995 			get_port_caps(dev, i + 1);
5996 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
5997 				    i + 1);
5998 		}
5999 	}
6000 
6001 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
6002 	mutex_unlock(&mlx5_ib_multiport_mutex);
6003 	return err;
6004 }
6005 
6006 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
6007 {
6008 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
6009 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
6010 							  port_num + 1);
6011 	int i;
6012 
6013 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
6014 		return;
6015 
6016 	mutex_lock(&mlx5_ib_multiport_mutex);
6017 	for (i = 0; i < dev->num_ports; i++) {
6018 		if (dev->port[i].mp.mpi) {
6019 			/* Destroy the native port stub */
6020 			if (i == port_num) {
6021 				kfree(dev->port[i].mp.mpi);
6022 				dev->port[i].mp.mpi = NULL;
6023 			} else {
6024 				mlx5_ib_dbg(dev, "unbinding port_num: %d\n", i + 1);
6025 				mlx5_ib_unbind_slave_port(dev, dev->port[i].mp.mpi);
6026 			}
6027 		}
6028 	}
6029 
6030 	mlx5_ib_dbg(dev, "removing from devlist\n");
6031 	list_del(&dev->ib_dev_list);
6032 	mutex_unlock(&mlx5_ib_multiport_mutex);
6033 
6034 	mlx5_nic_vport_disable_roce(dev->mdev);
6035 }
6036 
6037 ADD_UVERBS_ATTRIBUTES_SIMPLE(
6038 	mlx5_ib_dm,
6039 	UVERBS_OBJECT_DM,
6040 	UVERBS_METHOD_DM_ALLOC,
6041 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
6042 			    UVERBS_ATTR_TYPE(u64),
6043 			    UA_MANDATORY),
6044 	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
6045 			    UVERBS_ATTR_TYPE(u16),
6046 			    UA_OPTIONAL),
6047 	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_ALLOC_DM_REQ_TYPE,
6048 			     enum mlx5_ib_uapi_dm_type,
6049 			     UA_OPTIONAL));
6050 
6051 ADD_UVERBS_ATTRIBUTES_SIMPLE(
6052 	mlx5_ib_flow_action,
6053 	UVERBS_OBJECT_FLOW_ACTION,
6054 	UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
6055 	UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
6056 			     enum mlx5_ib_uapi_flow_action_flags));
6057 
6058 static const struct uapi_definition mlx5_ib_defs[] = {
6059 #if IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS)
6060 	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
6061 	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
6062 #endif
6063 
6064 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_FLOW_ACTION,
6065 				&mlx5_ib_flow_action),
6066 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DM, &mlx5_ib_dm),
6067 	{}
6068 };
6069 
6070 static int mlx5_ib_read_counters(struct ib_counters *counters,
6071 				 struct ib_counters_read_attr *read_attr,
6072 				 struct uverbs_attr_bundle *attrs)
6073 {
6074 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
6075 	struct mlx5_read_counters_attr mread_attr = {};
6076 	struct mlx5_ib_flow_counters_desc *desc;
6077 	int ret, i;
6078 
6079 	mutex_lock(&mcounters->mcntrs_mutex);
6080 	if (mcounters->cntrs_max_index > read_attr->ncounters) {
6081 		ret = -EINVAL;
6082 		goto err_bound;
6083 	}
6084 
6085 	mread_attr.out = kcalloc(mcounters->counters_num, sizeof(u64),
6086 				 GFP_KERNEL);
6087 	if (!mread_attr.out) {
6088 		ret = -ENOMEM;
6089 		goto err_bound;
6090 	}
6091 
6092 	mread_attr.hw_cntrs_hndl = mcounters->hw_cntrs_hndl;
6093 	mread_attr.flags = read_attr->flags;
6094 	ret = mcounters->read_counters(counters->device, &mread_attr);
6095 	if (ret)
6096 		goto err_read;
6097 
6098 	/* do the pass over the counters data array to assign according to the
6099 	 * descriptions and indexing pairs
6100 	 */
6101 	desc = mcounters->counters_data;
6102 	for (i = 0; i < mcounters->ncounters; i++)
6103 		read_attr->counters_buff[desc[i].index] += mread_attr.out[desc[i].description];
6104 
6105 err_read:
6106 	kfree(mread_attr.out);
6107 err_bound:
6108 	mutex_unlock(&mcounters->mcntrs_mutex);
6109 	return ret;
6110 }
6111 
6112 static int mlx5_ib_destroy_counters(struct ib_counters *counters)
6113 {
6114 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
6115 
6116 	counters_clear_description(counters);
6117 	if (mcounters->hw_cntrs_hndl)
6118 		mlx5_fc_destroy(to_mdev(counters->device)->mdev,
6119 				mcounters->hw_cntrs_hndl);
6120 
6121 	kfree(mcounters);
6122 
6123 	return 0;
6124 }
6125 
6126 static struct ib_counters *mlx5_ib_create_counters(struct ib_device *device,
6127 						   struct uverbs_attr_bundle *attrs)
6128 {
6129 	struct mlx5_ib_mcounters *mcounters;
6130 
6131 	mcounters = kzalloc(sizeof(*mcounters), GFP_KERNEL);
6132 	if (!mcounters)
6133 		return ERR_PTR(-ENOMEM);
6134 
6135 	mutex_init(&mcounters->mcntrs_mutex);
6136 
6137 	return &mcounters->ibcntrs;
6138 }
6139 
6140 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
6141 {
6142 	mlx5_ib_cleanup_multiport_master(dev);
6143 	WARN_ON(!xa_empty(&dev->odp_mkeys));
6144 	cleanup_srcu_struct(&dev->odp_srcu);
6145 
6146 	WARN_ON(!xa_empty(&dev->sig_mrs));
6147 	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
6148 }
6149 
6150 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
6151 {
6152 	struct mlx5_core_dev *mdev = dev->mdev;
6153 	int err;
6154 	int i;
6155 
6156 	for (i = 0; i < dev->num_ports; i++) {
6157 		spin_lock_init(&dev->port[i].mp.mpi_lock);
6158 		rwlock_init(&dev->port[i].roce.netdev_lock);
6159 		dev->port[i].roce.dev = dev;
6160 		dev->port[i].roce.native_port_num = i + 1;
6161 		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
6162 	}
6163 
6164 	mlx5_ib_internal_fill_odp_caps(dev);
6165 
6166 	err = mlx5_ib_init_multiport_master(dev);
6167 	if (err)
6168 		return err;
6169 
6170 	err = set_has_smi_cap(dev);
6171 	if (err)
6172 		return err;
6173 
6174 	if (!mlx5_core_mp_enabled(mdev)) {
6175 		for (i = 1; i <= dev->num_ports; i++) {
6176 			err = get_port_caps(dev, i);
6177 			if (err)
6178 				break;
6179 		}
6180 	} else {
6181 		err = get_port_caps(dev, mlx5_core_native_port_num(mdev));
6182 	}
6183 	if (err)
6184 		goto err_mp;
6185 
6186 	if (mlx5_use_mad_ifc(dev))
6187 		get_ext_port_caps(dev);
6188 
6189 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
6190 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
6191 	dev->ib_dev.phys_port_cnt	= dev->num_ports;
6192 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_count(mdev);
6193 	dev->ib_dev.dev.parent		= mdev->device;
6194 
6195 	mutex_init(&dev->cap_mask_mutex);
6196 	INIT_LIST_HEAD(&dev->qp_list);
6197 	spin_lock_init(&dev->reset_flow_resource_lock);
6198 	xa_init(&dev->odp_mkeys);
6199 	xa_init(&dev->sig_mrs);
6200 
6201 	spin_lock_init(&dev->dm.lock);
6202 	dev->dm.dev = mdev;
6203 
6204 	err = init_srcu_struct(&dev->odp_srcu);
6205 	if (err)
6206 		goto err_mp;
6207 
6208 	return 0;
6209 
6210 err_mp:
6211 	mlx5_ib_cleanup_multiport_master(dev);
6212 
6213 	return -ENOMEM;
6214 }
6215 
6216 static int mlx5_ib_stage_flow_db_init(struct mlx5_ib_dev *dev)
6217 {
6218 	dev->flow_db = kzalloc(sizeof(*dev->flow_db), GFP_KERNEL);
6219 
6220 	if (!dev->flow_db)
6221 		return -ENOMEM;
6222 
6223 	mutex_init(&dev->flow_db->lock);
6224 
6225 	return 0;
6226 }
6227 
6228 static void mlx5_ib_stage_flow_db_cleanup(struct mlx5_ib_dev *dev)
6229 {
6230 	kfree(dev->flow_db);
6231 }
6232 
6233 static const struct ib_device_ops mlx5_ib_dev_ops = {
6234 	.owner = THIS_MODULE,
6235 	.driver_id = RDMA_DRIVER_MLX5,
6236 	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
6237 
6238 	.add_gid = mlx5_ib_add_gid,
6239 	.alloc_mr = mlx5_ib_alloc_mr,
6240 	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
6241 	.alloc_pd = mlx5_ib_alloc_pd,
6242 	.alloc_ucontext = mlx5_ib_alloc_ucontext,
6243 	.attach_mcast = mlx5_ib_mcg_attach,
6244 	.check_mr_status = mlx5_ib_check_mr_status,
6245 	.create_ah = mlx5_ib_create_ah,
6246 	.create_counters = mlx5_ib_create_counters,
6247 	.create_cq = mlx5_ib_create_cq,
6248 	.create_flow = mlx5_ib_create_flow,
6249 	.create_qp = mlx5_ib_create_qp,
6250 	.create_srq = mlx5_ib_create_srq,
6251 	.dealloc_pd = mlx5_ib_dealloc_pd,
6252 	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
6253 	.del_gid = mlx5_ib_del_gid,
6254 	.dereg_mr = mlx5_ib_dereg_mr,
6255 	.destroy_ah = mlx5_ib_destroy_ah,
6256 	.destroy_counters = mlx5_ib_destroy_counters,
6257 	.destroy_cq = mlx5_ib_destroy_cq,
6258 	.destroy_flow = mlx5_ib_destroy_flow,
6259 	.destroy_flow_action = mlx5_ib_destroy_flow_action,
6260 	.destroy_qp = mlx5_ib_destroy_qp,
6261 	.destroy_srq = mlx5_ib_destroy_srq,
6262 	.detach_mcast = mlx5_ib_mcg_detach,
6263 	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
6264 	.drain_rq = mlx5_ib_drain_rq,
6265 	.drain_sq = mlx5_ib_drain_sq,
6266 	.enable_driver = mlx5_ib_enable_driver,
6267 	.fill_res_entry = mlx5_ib_fill_res_entry,
6268 	.fill_stat_entry = mlx5_ib_fill_stat_entry,
6269 	.get_dev_fw_str = get_dev_fw_str,
6270 	.get_dma_mr = mlx5_ib_get_dma_mr,
6271 	.get_link_layer = mlx5_ib_port_link_layer,
6272 	.map_mr_sg = mlx5_ib_map_mr_sg,
6273 	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
6274 	.mmap = mlx5_ib_mmap,
6275 	.mmap_free = mlx5_ib_mmap_free,
6276 	.modify_cq = mlx5_ib_modify_cq,
6277 	.modify_device = mlx5_ib_modify_device,
6278 	.modify_port = mlx5_ib_modify_port,
6279 	.modify_qp = mlx5_ib_modify_qp,
6280 	.modify_srq = mlx5_ib_modify_srq,
6281 	.poll_cq = mlx5_ib_poll_cq,
6282 	.post_recv = mlx5_ib_post_recv,
6283 	.post_send = mlx5_ib_post_send,
6284 	.post_srq_recv = mlx5_ib_post_srq_recv,
6285 	.process_mad = mlx5_ib_process_mad,
6286 	.query_ah = mlx5_ib_query_ah,
6287 	.query_device = mlx5_ib_query_device,
6288 	.query_gid = mlx5_ib_query_gid,
6289 	.query_pkey = mlx5_ib_query_pkey,
6290 	.query_qp = mlx5_ib_query_qp,
6291 	.query_srq = mlx5_ib_query_srq,
6292 	.read_counters = mlx5_ib_read_counters,
6293 	.reg_user_mr = mlx5_ib_reg_user_mr,
6294 	.req_notify_cq = mlx5_ib_arm_cq,
6295 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
6296 	.resize_cq = mlx5_ib_resize_cq,
6297 
6298 	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
6299 	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
6300 	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
6301 	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
6302 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
6303 };
6304 
6305 static const struct ib_device_ops mlx5_ib_dev_flow_ipsec_ops = {
6306 	.create_flow_action_esp = mlx5_ib_create_flow_action_esp,
6307 	.modify_flow_action_esp = mlx5_ib_modify_flow_action_esp,
6308 };
6309 
6310 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
6311 	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
6312 };
6313 
6314 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
6315 	.get_vf_config = mlx5_ib_get_vf_config,
6316 	.get_vf_guid = mlx5_ib_get_vf_guid,
6317 	.get_vf_stats = mlx5_ib_get_vf_stats,
6318 	.set_vf_guid = mlx5_ib_set_vf_guid,
6319 	.set_vf_link_state = mlx5_ib_set_vf_link_state,
6320 };
6321 
6322 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
6323 	.alloc_mw = mlx5_ib_alloc_mw,
6324 	.dealloc_mw = mlx5_ib_dealloc_mw,
6325 };
6326 
6327 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
6328 	.alloc_xrcd = mlx5_ib_alloc_xrcd,
6329 	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
6330 };
6331 
6332 static const struct ib_device_ops mlx5_ib_dev_dm_ops = {
6333 	.alloc_dm = mlx5_ib_alloc_dm,
6334 	.dealloc_dm = mlx5_ib_dealloc_dm,
6335 	.reg_dm_mr = mlx5_ib_reg_dm_mr,
6336 };
6337 
6338 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
6339 {
6340 	struct mlx5_core_dev *mdev = dev->mdev;
6341 	int err;
6342 
6343 	dev->ib_dev.uverbs_cmd_mask	=
6344 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
6345 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
6346 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
6347 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
6348 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
6349 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
6350 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
6351 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
6352 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
6353 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
6354 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
6355 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
6356 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
6357 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
6358 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
6359 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
6360 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
6361 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
6362 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
6363 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
6364 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
6365 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
6366 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
6367 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
6368 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
6369 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
6370 	dev->ib_dev.uverbs_ex_cmd_mask =
6371 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
6372 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
6373 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP)	|
6374 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP)	|
6375 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_CQ)	|
6376 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW)	|
6377 		(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW);
6378 
6379 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
6380 	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
6381 		ib_set_device_ops(&dev->ib_dev,
6382 				  &mlx5_ib_dev_ipoib_enhanced_ops);
6383 
6384 	if (mlx5_core_is_pf(mdev))
6385 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
6386 
6387 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
6388 
6389 	if (MLX5_CAP_GEN(mdev, imaicl)) {
6390 		dev->ib_dev.uverbs_cmd_mask |=
6391 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
6392 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
6393 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
6394 	}
6395 
6396 	if (MLX5_CAP_GEN(mdev, xrc)) {
6397 		dev->ib_dev.uverbs_cmd_mask |=
6398 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
6399 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
6400 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
6401 	}
6402 
6403 	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
6404 	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
6405 	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
6406 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
6407 
6408 	if (mlx5_accel_ipsec_device_caps(dev->mdev) &
6409 	    MLX5_ACCEL_IPSEC_CAP_DEVICE)
6410 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_flow_ipsec_ops);
6411 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
6412 
6413 	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
6414 		dev->ib_dev.driver_def = mlx5_ib_defs;
6415 
6416 	err = init_node_data(dev);
6417 	if (err)
6418 		return err;
6419 
6420 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
6421 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
6422 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
6423 		mutex_init(&dev->lb.mutex);
6424 
6425 	dev->ib_dev.use_cq_dim = true;
6426 
6427 	return 0;
6428 }
6429 
6430 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
6431 	.get_port_immutable = mlx5_port_immutable,
6432 	.query_port = mlx5_ib_query_port,
6433 };
6434 
6435 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
6436 {
6437 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
6438 	return 0;
6439 }
6440 
6441 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
6442 	.get_port_immutable = mlx5_port_rep_immutable,
6443 	.query_port = mlx5_ib_rep_query_port,
6444 };
6445 
6446 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
6447 {
6448 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
6449 	return 0;
6450 }
6451 
6452 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
6453 	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
6454 	.create_wq = mlx5_ib_create_wq,
6455 	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
6456 	.destroy_wq = mlx5_ib_destroy_wq,
6457 	.get_netdev = mlx5_ib_get_netdev,
6458 	.modify_wq = mlx5_ib_modify_wq,
6459 };
6460 
6461 static int mlx5_ib_stage_common_roce_init(struct mlx5_ib_dev *dev)
6462 {
6463 	u8 port_num;
6464 
6465 	dev->ib_dev.uverbs_ex_cmd_mask |=
6466 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
6467 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
6468 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
6469 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
6470 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
6471 	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
6472 
6473 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
6474 
6475 	/* Register only for native ports */
6476 	return mlx5_add_netdev_notifier(dev, port_num);
6477 }
6478 
6479 static void mlx5_ib_stage_common_roce_cleanup(struct mlx5_ib_dev *dev)
6480 {
6481 	u8 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
6482 
6483 	mlx5_remove_netdev_notifier(dev, port_num);
6484 }
6485 
6486 static int mlx5_ib_stage_raw_eth_roce_init(struct mlx5_ib_dev *dev)
6487 {
6488 	struct mlx5_core_dev *mdev = dev->mdev;
6489 	enum rdma_link_layer ll;
6490 	int port_type_cap;
6491 	int err = 0;
6492 
6493 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
6494 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
6495 
6496 	if (ll == IB_LINK_LAYER_ETHERNET)
6497 		err = mlx5_ib_stage_common_roce_init(dev);
6498 
6499 	return err;
6500 }
6501 
6502 static void mlx5_ib_stage_raw_eth_roce_cleanup(struct mlx5_ib_dev *dev)
6503 {
6504 	mlx5_ib_stage_common_roce_cleanup(dev);
6505 }
6506 
6507 static int mlx5_ib_stage_roce_init(struct mlx5_ib_dev *dev)
6508 {
6509 	struct mlx5_core_dev *mdev = dev->mdev;
6510 	enum rdma_link_layer ll;
6511 	int port_type_cap;
6512 	int err;
6513 
6514 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
6515 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
6516 
6517 	if (ll == IB_LINK_LAYER_ETHERNET) {
6518 		err = mlx5_ib_stage_common_roce_init(dev);
6519 		if (err)
6520 			return err;
6521 
6522 		err = mlx5_enable_eth(dev);
6523 		if (err)
6524 			goto cleanup;
6525 	}
6526 
6527 	return 0;
6528 cleanup:
6529 	mlx5_ib_stage_common_roce_cleanup(dev);
6530 
6531 	return err;
6532 }
6533 
6534 static void mlx5_ib_stage_roce_cleanup(struct mlx5_ib_dev *dev)
6535 {
6536 	struct mlx5_core_dev *mdev = dev->mdev;
6537 	enum rdma_link_layer ll;
6538 	int port_type_cap;
6539 
6540 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
6541 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
6542 
6543 	if (ll == IB_LINK_LAYER_ETHERNET) {
6544 		mlx5_disable_eth(dev);
6545 		mlx5_ib_stage_common_roce_cleanup(dev);
6546 	}
6547 }
6548 
6549 static int mlx5_ib_stage_dev_res_init(struct mlx5_ib_dev *dev)
6550 {
6551 	return create_dev_resources(&dev->devr);
6552 }
6553 
6554 static void mlx5_ib_stage_dev_res_cleanup(struct mlx5_ib_dev *dev)
6555 {
6556 	destroy_dev_resources(&dev->devr);
6557 }
6558 
6559 static int mlx5_ib_stage_odp_init(struct mlx5_ib_dev *dev)
6560 {
6561 	return mlx5_ib_odp_init_one(dev);
6562 }
6563 
6564 static void mlx5_ib_stage_odp_cleanup(struct mlx5_ib_dev *dev)
6565 {
6566 	mlx5_ib_odp_cleanup_one(dev);
6567 }
6568 
6569 static const struct ib_device_ops mlx5_ib_dev_hw_stats_ops = {
6570 	.alloc_hw_stats = mlx5_ib_alloc_hw_stats,
6571 	.get_hw_stats = mlx5_ib_get_hw_stats,
6572 	.counter_bind_qp = mlx5_ib_counter_bind_qp,
6573 	.counter_unbind_qp = mlx5_ib_counter_unbind_qp,
6574 	.counter_dealloc = mlx5_ib_counter_dealloc,
6575 	.counter_alloc_stats = mlx5_ib_counter_alloc_stats,
6576 	.counter_update_stats = mlx5_ib_counter_update_stats,
6577 };
6578 
6579 static int mlx5_ib_stage_counters_init(struct mlx5_ib_dev *dev)
6580 {
6581 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
6582 		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_hw_stats_ops);
6583 
6584 		return mlx5_ib_alloc_counters(dev);
6585 	}
6586 
6587 	return 0;
6588 }
6589 
6590 static void mlx5_ib_stage_counters_cleanup(struct mlx5_ib_dev *dev)
6591 {
6592 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
6593 		mlx5_ib_dealloc_counters(dev);
6594 }
6595 
6596 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
6597 {
6598 	mlx5_ib_init_cong_debugfs(dev,
6599 				  mlx5_core_native_port_num(dev->mdev) - 1);
6600 	return 0;
6601 }
6602 
6603 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
6604 {
6605 	mlx5_ib_cleanup_cong_debugfs(dev,
6606 				     mlx5_core_native_port_num(dev->mdev) - 1);
6607 }
6608 
6609 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
6610 {
6611 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
6612 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
6613 }
6614 
6615 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
6616 {
6617 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
6618 }
6619 
6620 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
6621 {
6622 	int err;
6623 
6624 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
6625 	if (err)
6626 		return err;
6627 
6628 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
6629 	if (err)
6630 		mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
6631 
6632 	return err;
6633 }
6634 
6635 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
6636 {
6637 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
6638 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
6639 }
6640 
6641 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
6642 {
6643 	const char *name;
6644 
6645 	rdma_set_device_sysfs_group(&dev->ib_dev, &mlx5_attr_group);
6646 	if (!mlx5_lag_is_roce(dev->mdev))
6647 		name = "mlx5_%d";
6648 	else
6649 		name = "mlx5_bond_%d";
6650 	return ib_register_device(&dev->ib_dev, name);
6651 }
6652 
6653 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
6654 {
6655 	destroy_umrc_res(dev);
6656 }
6657 
6658 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
6659 {
6660 	ib_unregister_device(&dev->ib_dev);
6661 }
6662 
6663 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
6664 {
6665 	return create_umr_res(dev);
6666 }
6667 
6668 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
6669 {
6670 	init_delay_drop(dev);
6671 
6672 	return 0;
6673 }
6674 
6675 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
6676 {
6677 	cancel_delay_drop(dev);
6678 }
6679 
6680 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
6681 {
6682 	dev->mdev_events.notifier_call = mlx5_ib_event;
6683 	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
6684 	return 0;
6685 }
6686 
6687 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
6688 {
6689 	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
6690 }
6691 
6692 static int mlx5_ib_stage_devx_init(struct mlx5_ib_dev *dev)
6693 {
6694 	int uid;
6695 
6696 	uid = mlx5_ib_devx_create(dev, false);
6697 	if (uid > 0) {
6698 		dev->devx_whitelist_uid = uid;
6699 		mlx5_ib_devx_init_event_table(dev);
6700 	}
6701 
6702 	return 0;
6703 }
6704 static void mlx5_ib_stage_devx_cleanup(struct mlx5_ib_dev *dev)
6705 {
6706 	if (dev->devx_whitelist_uid) {
6707 		mlx5_ib_devx_cleanup_event_table(dev);
6708 		mlx5_ib_devx_destroy(dev, dev->devx_whitelist_uid);
6709 	}
6710 }
6711 
6712 int mlx5_ib_enable_driver(struct ib_device *dev)
6713 {
6714 	struct mlx5_ib_dev *mdev = to_mdev(dev);
6715 	int ret;
6716 
6717 	ret = mlx5_ib_test_wc(mdev);
6718 	mlx5_ib_dbg(mdev, "Write-Combining %s",
6719 		    mdev->wc_support ? "supported" : "not supported");
6720 
6721 	return ret;
6722 }
6723 
6724 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
6725 		      const struct mlx5_ib_profile *profile,
6726 		      int stage)
6727 {
6728 	/* Number of stages to cleanup */
6729 	while (stage) {
6730 		stage--;
6731 		if (profile->stage[stage].cleanup)
6732 			profile->stage[stage].cleanup(dev);
6733 	}
6734 
6735 	kfree(dev->port);
6736 	ib_dealloc_device(&dev->ib_dev);
6737 }
6738 
6739 void *__mlx5_ib_add(struct mlx5_ib_dev *dev,
6740 		    const struct mlx5_ib_profile *profile)
6741 {
6742 	int err;
6743 	int i;
6744 
6745 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
6746 		if (profile->stage[i].init) {
6747 			err = profile->stage[i].init(dev);
6748 			if (err)
6749 				goto err_out;
6750 		}
6751 	}
6752 
6753 	dev->profile = profile;
6754 	dev->ib_active = true;
6755 
6756 	return dev;
6757 
6758 err_out:
6759 	__mlx5_ib_remove(dev, profile, i);
6760 
6761 	return NULL;
6762 }
6763 
6764 static const struct mlx5_ib_profile pf_profile = {
6765 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
6766 		     mlx5_ib_stage_init_init,
6767 		     mlx5_ib_stage_init_cleanup),
6768 	STAGE_CREATE(MLX5_IB_STAGE_FLOW_DB,
6769 		     mlx5_ib_stage_flow_db_init,
6770 		     mlx5_ib_stage_flow_db_cleanup),
6771 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
6772 		     mlx5_ib_stage_caps_init,
6773 		     NULL),
6774 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
6775 		     mlx5_ib_stage_non_default_cb,
6776 		     NULL),
6777 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
6778 		     mlx5_ib_stage_roce_init,
6779 		     mlx5_ib_stage_roce_cleanup),
6780 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
6781 		     mlx5_init_srq_table,
6782 		     mlx5_cleanup_srq_table),
6783 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
6784 		     mlx5_ib_stage_dev_res_init,
6785 		     mlx5_ib_stage_dev_res_cleanup),
6786 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
6787 		     mlx5_ib_stage_dev_notifier_init,
6788 		     mlx5_ib_stage_dev_notifier_cleanup),
6789 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
6790 		     mlx5_ib_stage_odp_init,
6791 		     mlx5_ib_stage_odp_cleanup),
6792 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
6793 		     mlx5_ib_stage_counters_init,
6794 		     mlx5_ib_stage_counters_cleanup),
6795 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
6796 		     mlx5_ib_stage_cong_debugfs_init,
6797 		     mlx5_ib_stage_cong_debugfs_cleanup),
6798 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
6799 		     mlx5_ib_stage_uar_init,
6800 		     mlx5_ib_stage_uar_cleanup),
6801 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
6802 		     mlx5_ib_stage_bfrag_init,
6803 		     mlx5_ib_stage_bfrag_cleanup),
6804 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
6805 		     NULL,
6806 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
6807 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
6808 		     mlx5_ib_stage_devx_init,
6809 		     mlx5_ib_stage_devx_cleanup),
6810 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
6811 		     mlx5_ib_stage_ib_reg_init,
6812 		     mlx5_ib_stage_ib_reg_cleanup),
6813 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
6814 		     mlx5_ib_stage_post_ib_reg_umr_init,
6815 		     NULL),
6816 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
6817 		     mlx5_ib_stage_delay_drop_init,
6818 		     mlx5_ib_stage_delay_drop_cleanup),
6819 };
6820 
6821 const struct mlx5_ib_profile raw_eth_profile = {
6822 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
6823 		     mlx5_ib_stage_init_init,
6824 		     mlx5_ib_stage_init_cleanup),
6825 	STAGE_CREATE(MLX5_IB_STAGE_FLOW_DB,
6826 		     mlx5_ib_stage_flow_db_init,
6827 		     mlx5_ib_stage_flow_db_cleanup),
6828 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
6829 		     mlx5_ib_stage_caps_init,
6830 		     NULL),
6831 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
6832 		     mlx5_ib_stage_raw_eth_non_default_cb,
6833 		     NULL),
6834 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
6835 		     mlx5_ib_stage_raw_eth_roce_init,
6836 		     mlx5_ib_stage_raw_eth_roce_cleanup),
6837 	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
6838 		     mlx5_init_srq_table,
6839 		     mlx5_cleanup_srq_table),
6840 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
6841 		     mlx5_ib_stage_dev_res_init,
6842 		     mlx5_ib_stage_dev_res_cleanup),
6843 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
6844 		     mlx5_ib_stage_dev_notifier_init,
6845 		     mlx5_ib_stage_dev_notifier_cleanup),
6846 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
6847 		     mlx5_ib_stage_counters_init,
6848 		     mlx5_ib_stage_counters_cleanup),
6849 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
6850 		     mlx5_ib_stage_uar_init,
6851 		     mlx5_ib_stage_uar_cleanup),
6852 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
6853 		     mlx5_ib_stage_bfrag_init,
6854 		     mlx5_ib_stage_bfrag_cleanup),
6855 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
6856 		     NULL,
6857 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
6858 	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
6859 		     mlx5_ib_stage_devx_init,
6860 		     mlx5_ib_stage_devx_cleanup),
6861 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
6862 		     mlx5_ib_stage_ib_reg_init,
6863 		     mlx5_ib_stage_ib_reg_cleanup),
6864 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
6865 		     mlx5_ib_stage_post_ib_reg_umr_init,
6866 		     NULL),
6867 };
6868 
6869 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev)
6870 {
6871 	struct mlx5_ib_multiport_info *mpi;
6872 	struct mlx5_ib_dev *dev;
6873 	bool bound = false;
6874 	int err;
6875 
6876 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
6877 	if (!mpi)
6878 		return NULL;
6879 
6880 	mpi->mdev = mdev;
6881 
6882 	err = mlx5_query_nic_vport_system_image_guid(mdev,
6883 						     &mpi->sys_image_guid);
6884 	if (err) {
6885 		kfree(mpi);
6886 		return NULL;
6887 	}
6888 
6889 	mutex_lock(&mlx5_ib_multiport_mutex);
6890 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
6891 		if (dev->sys_image_guid == mpi->sys_image_guid)
6892 			bound = mlx5_ib_bind_slave_port(dev, mpi);
6893 
6894 		if (bound) {
6895 			rdma_roce_rescan_device(&dev->ib_dev);
6896 			break;
6897 		}
6898 	}
6899 
6900 	if (!bound) {
6901 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
6902 		dev_dbg(mdev->device,
6903 			"no suitable IB device found to bind to, added to unaffiliated list.\n");
6904 	}
6905 	mutex_unlock(&mlx5_ib_multiport_mutex);
6906 
6907 	return mpi;
6908 }
6909 
6910 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
6911 {
6912 	const struct mlx5_ib_profile *profile;
6913 	enum rdma_link_layer ll;
6914 	struct mlx5_ib_dev *dev;
6915 	int port_type_cap;
6916 	int num_ports;
6917 
6918 	printk_once(KERN_INFO "%s", mlx5_version);
6919 
6920 	if (MLX5_ESWITCH_MANAGER(mdev) &&
6921 	    mlx5_ib_eswitch_mode(mdev->priv.eswitch) == MLX5_ESWITCH_OFFLOADS) {
6922 		if (!mlx5_core_mp_enabled(mdev))
6923 			mlx5_ib_register_vport_reps(mdev);
6924 		return mdev;
6925 	}
6926 
6927 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
6928 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
6929 
6930 	if (mlx5_core_is_mp_slave(mdev) && ll == IB_LINK_LAYER_ETHERNET)
6931 		return mlx5_ib_add_slave_port(mdev);
6932 
6933 	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
6934 			MLX5_CAP_GEN(mdev, num_vhca_ports));
6935 	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
6936 	if (!dev)
6937 		return NULL;
6938 	dev->port = kcalloc(num_ports, sizeof(*dev->port),
6939 			     GFP_KERNEL);
6940 	if (!dev->port) {
6941 		ib_dealloc_device(&dev->ib_dev);
6942 		return NULL;
6943 	}
6944 
6945 	dev->mdev = mdev;
6946 	dev->num_ports = num_ports;
6947 
6948 	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_is_roce_enabled(mdev))
6949 		profile = &raw_eth_profile;
6950 	else
6951 		profile = &pf_profile;
6952 
6953 	return __mlx5_ib_add(dev, profile);
6954 }
6955 
6956 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
6957 {
6958 	struct mlx5_ib_multiport_info *mpi;
6959 	struct mlx5_ib_dev *dev;
6960 
6961 	if (MLX5_ESWITCH_MANAGER(mdev) && context == mdev) {
6962 		mlx5_ib_unregister_vport_reps(mdev);
6963 		return;
6964 	}
6965 
6966 	if (mlx5_core_is_mp_slave(mdev)) {
6967 		mpi = context;
6968 		mutex_lock(&mlx5_ib_multiport_mutex);
6969 		if (mpi->ibdev)
6970 			mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
6971 		list_del(&mpi->list);
6972 		mutex_unlock(&mlx5_ib_multiport_mutex);
6973 		kfree(mpi);
6974 		return;
6975 	}
6976 
6977 	dev = context;
6978 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
6979 }
6980 
6981 static struct mlx5_interface mlx5_ib_interface = {
6982 	.add            = mlx5_ib_add,
6983 	.remove         = mlx5_ib_remove,
6984 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
6985 };
6986 
6987 unsigned long mlx5_ib_get_xlt_emergency_page(void)
6988 {
6989 	mutex_lock(&xlt_emergency_page_mutex);
6990 	return xlt_emergency_page;
6991 }
6992 
6993 void mlx5_ib_put_xlt_emergency_page(void)
6994 {
6995 	mutex_unlock(&xlt_emergency_page_mutex);
6996 }
6997 
6998 static int __init mlx5_ib_init(void)
6999 {
7000 	int err;
7001 
7002 	xlt_emergency_page = __get_free_page(GFP_KERNEL);
7003 	if (!xlt_emergency_page)
7004 		return -ENOMEM;
7005 
7006 	mutex_init(&xlt_emergency_page_mutex);
7007 
7008 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
7009 	if (!mlx5_ib_event_wq) {
7010 		free_page(xlt_emergency_page);
7011 		return -ENOMEM;
7012 	}
7013 
7014 	mlx5_ib_odp_init();
7015 
7016 	err = mlx5_register_interface(&mlx5_ib_interface);
7017 
7018 	return err;
7019 }
7020 
7021 static void __exit mlx5_ib_cleanup(void)
7022 {
7023 	mlx5_unregister_interface(&mlx5_ib_interface);
7024 	destroy_workqueue(mlx5_ib_event_wq);
7025 	mutex_destroy(&xlt_emergency_page_mutex);
7026 	free_page(xlt_emergency_page);
7027 }
7028 
7029 module_init(mlx5_ib_init);
7030 module_exit(mlx5_ib_cleanup);
7031