xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision 6396bb221514d2876fd6dc0aa2a1f240d99b37bb)
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/list.h>
56 #include <rdma/ib_smi.h>
57 #include <rdma/ib_umem.h>
58 #include <linux/in.h>
59 #include <linux/etherdevice.h>
60 #include "mlx5_ib.h"
61 #include "ib_rep.h"
62 #include "cmd.h"
63 #include <linux/mlx5/fs_helpers.h>
64 #include <linux/mlx5/accel.h>
65 #include <rdma/uverbs_std_types.h>
66 #include <rdma/mlx5_user_ioctl_verbs.h>
67 #include <rdma/mlx5_user_ioctl_cmds.h>
68 
69 #define UVERBS_MODULE_NAME mlx5_ib
70 #include <rdma/uverbs_named_ioctl.h>
71 
72 #define DRIVER_NAME "mlx5_ib"
73 #define DRIVER_VERSION "5.0-0"
74 
75 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
76 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
77 MODULE_LICENSE("Dual BSD/GPL");
78 
79 static char mlx5_version[] =
80 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
81 	DRIVER_VERSION "\n";
82 
83 struct mlx5_ib_event_work {
84 	struct work_struct	work;
85 	struct mlx5_core_dev	*dev;
86 	void			*context;
87 	enum mlx5_dev_event	event;
88 	unsigned long		param;
89 };
90 
91 enum {
92 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
93 };
94 
95 static struct workqueue_struct *mlx5_ib_event_wq;
96 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
97 static LIST_HEAD(mlx5_ib_dev_list);
98 /*
99  * This mutex should be held when accessing either of the above lists
100  */
101 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
102 
103 /* We can't use an array for xlt_emergency_page because dma_map_single
104  * doesn't work on kernel modules memory
105  */
106 static unsigned long xlt_emergency_page;
107 static struct mutex xlt_emergency_page_mutex;
108 
109 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
110 {
111 	struct mlx5_ib_dev *dev;
112 
113 	mutex_lock(&mlx5_ib_multiport_mutex);
114 	dev = mpi->ibdev;
115 	mutex_unlock(&mlx5_ib_multiport_mutex);
116 	return dev;
117 }
118 
119 static enum rdma_link_layer
120 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
121 {
122 	switch (port_type_cap) {
123 	case MLX5_CAP_PORT_TYPE_IB:
124 		return IB_LINK_LAYER_INFINIBAND;
125 	case MLX5_CAP_PORT_TYPE_ETH:
126 		return IB_LINK_LAYER_ETHERNET;
127 	default:
128 		return IB_LINK_LAYER_UNSPECIFIED;
129 	}
130 }
131 
132 static enum rdma_link_layer
133 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
134 {
135 	struct mlx5_ib_dev *dev = to_mdev(device);
136 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
137 
138 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
139 }
140 
141 static int get_port_state(struct ib_device *ibdev,
142 			  u8 port_num,
143 			  enum ib_port_state *state)
144 {
145 	struct ib_port_attr attr;
146 	int ret;
147 
148 	memset(&attr, 0, sizeof(attr));
149 	ret = ibdev->query_port(ibdev, port_num, &attr);
150 	if (!ret)
151 		*state = attr.state;
152 	return ret;
153 }
154 
155 static int mlx5_netdev_event(struct notifier_block *this,
156 			     unsigned long event, void *ptr)
157 {
158 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
159 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
160 	u8 port_num = roce->native_port_num;
161 	struct mlx5_core_dev *mdev;
162 	struct mlx5_ib_dev *ibdev;
163 
164 	ibdev = roce->dev;
165 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
166 	if (!mdev)
167 		return NOTIFY_DONE;
168 
169 	switch (event) {
170 	case NETDEV_REGISTER:
171 	case NETDEV_UNREGISTER:
172 		write_lock(&roce->netdev_lock);
173 		if (ibdev->rep) {
174 			struct mlx5_eswitch *esw = ibdev->mdev->priv.eswitch;
175 			struct net_device *rep_ndev;
176 
177 			rep_ndev = mlx5_ib_get_rep_netdev(esw,
178 							  ibdev->rep->vport);
179 			if (rep_ndev == ndev)
180 				roce->netdev = (event == NETDEV_UNREGISTER) ?
181 					NULL : ndev;
182 		} else if (ndev->dev.parent == &mdev->pdev->dev) {
183 			roce->netdev = (event == NETDEV_UNREGISTER) ?
184 				NULL : ndev;
185 		}
186 		write_unlock(&roce->netdev_lock);
187 		break;
188 
189 	case NETDEV_CHANGE:
190 	case NETDEV_UP:
191 	case NETDEV_DOWN: {
192 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
193 		struct net_device *upper = NULL;
194 
195 		if (lag_ndev) {
196 			upper = netdev_master_upper_dev_get(lag_ndev);
197 			dev_put(lag_ndev);
198 		}
199 
200 		if ((upper == ndev || (!upper && ndev == roce->netdev))
201 		    && ibdev->ib_active) {
202 			struct ib_event ibev = { };
203 			enum ib_port_state port_state;
204 
205 			if (get_port_state(&ibdev->ib_dev, port_num,
206 					   &port_state))
207 				goto done;
208 
209 			if (roce->last_port_state == port_state)
210 				goto done;
211 
212 			roce->last_port_state = port_state;
213 			ibev.device = &ibdev->ib_dev;
214 			if (port_state == IB_PORT_DOWN)
215 				ibev.event = IB_EVENT_PORT_ERR;
216 			else if (port_state == IB_PORT_ACTIVE)
217 				ibev.event = IB_EVENT_PORT_ACTIVE;
218 			else
219 				goto done;
220 
221 			ibev.element.port_num = port_num;
222 			ib_dispatch_event(&ibev);
223 		}
224 		break;
225 	}
226 
227 	default:
228 		break;
229 	}
230 done:
231 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
232 	return NOTIFY_DONE;
233 }
234 
235 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
236 					     u8 port_num)
237 {
238 	struct mlx5_ib_dev *ibdev = to_mdev(device);
239 	struct net_device *ndev;
240 	struct mlx5_core_dev *mdev;
241 
242 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
243 	if (!mdev)
244 		return NULL;
245 
246 	ndev = mlx5_lag_get_roce_netdev(mdev);
247 	if (ndev)
248 		goto out;
249 
250 	/* Ensure ndev does not disappear before we invoke dev_hold()
251 	 */
252 	read_lock(&ibdev->roce[port_num - 1].netdev_lock);
253 	ndev = ibdev->roce[port_num - 1].netdev;
254 	if (ndev)
255 		dev_hold(ndev);
256 	read_unlock(&ibdev->roce[port_num - 1].netdev_lock);
257 
258 out:
259 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
260 	return ndev;
261 }
262 
263 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
264 						   u8 ib_port_num,
265 						   u8 *native_port_num)
266 {
267 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
268 							  ib_port_num);
269 	struct mlx5_core_dev *mdev = NULL;
270 	struct mlx5_ib_multiport_info *mpi;
271 	struct mlx5_ib_port *port;
272 
273 	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
274 	    ll != IB_LINK_LAYER_ETHERNET) {
275 		if (native_port_num)
276 			*native_port_num = ib_port_num;
277 		return ibdev->mdev;
278 	}
279 
280 	if (native_port_num)
281 		*native_port_num = 1;
282 
283 	port = &ibdev->port[ib_port_num - 1];
284 	if (!port)
285 		return NULL;
286 
287 	spin_lock(&port->mp.mpi_lock);
288 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
289 	if (mpi && !mpi->unaffiliate) {
290 		mdev = mpi->mdev;
291 		/* If it's the master no need to refcount, it'll exist
292 		 * as long as the ib_dev exists.
293 		 */
294 		if (!mpi->is_master)
295 			mpi->mdev_refcnt++;
296 	}
297 	spin_unlock(&port->mp.mpi_lock);
298 
299 	return mdev;
300 }
301 
302 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u8 port_num)
303 {
304 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
305 							  port_num);
306 	struct mlx5_ib_multiport_info *mpi;
307 	struct mlx5_ib_port *port;
308 
309 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
310 		return;
311 
312 	port = &ibdev->port[port_num - 1];
313 
314 	spin_lock(&port->mp.mpi_lock);
315 	mpi = ibdev->port[port_num - 1].mp.mpi;
316 	if (mpi->is_master)
317 		goto out;
318 
319 	mpi->mdev_refcnt--;
320 	if (mpi->unaffiliate)
321 		complete(&mpi->unref_comp);
322 out:
323 	spin_unlock(&port->mp.mpi_lock);
324 }
325 
326 static int translate_eth_proto_oper(u32 eth_proto_oper, u8 *active_speed,
327 				    u8 *active_width)
328 {
329 	switch (eth_proto_oper) {
330 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
331 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
332 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
333 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
334 		*active_width = IB_WIDTH_1X;
335 		*active_speed = IB_SPEED_SDR;
336 		break;
337 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
338 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
339 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
340 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
341 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
342 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
343 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
344 		*active_width = IB_WIDTH_1X;
345 		*active_speed = IB_SPEED_QDR;
346 		break;
347 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
348 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
349 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
350 		*active_width = IB_WIDTH_1X;
351 		*active_speed = IB_SPEED_EDR;
352 		break;
353 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
354 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
355 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
356 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
357 		*active_width = IB_WIDTH_4X;
358 		*active_speed = IB_SPEED_QDR;
359 		break;
360 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
361 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
362 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
363 		*active_width = IB_WIDTH_1X;
364 		*active_speed = IB_SPEED_HDR;
365 		break;
366 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
367 		*active_width = IB_WIDTH_4X;
368 		*active_speed = IB_SPEED_FDR;
369 		break;
370 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
371 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
372 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
373 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
374 		*active_width = IB_WIDTH_4X;
375 		*active_speed = IB_SPEED_EDR;
376 		break;
377 	default:
378 		return -EINVAL;
379 	}
380 
381 	return 0;
382 }
383 
384 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
385 				struct ib_port_attr *props)
386 {
387 	struct mlx5_ib_dev *dev = to_mdev(device);
388 	struct mlx5_core_dev *mdev;
389 	struct net_device *ndev, *upper;
390 	enum ib_mtu ndev_ib_mtu;
391 	bool put_mdev = true;
392 	u16 qkey_viol_cntr;
393 	u32 eth_prot_oper;
394 	u8 mdev_port_num;
395 	int err;
396 
397 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
398 	if (!mdev) {
399 		/* This means the port isn't affiliated yet. Get the
400 		 * info for the master port instead.
401 		 */
402 		put_mdev = false;
403 		mdev = dev->mdev;
404 		mdev_port_num = 1;
405 		port_num = 1;
406 	}
407 
408 	/* Possible bad flows are checked before filling out props so in case
409 	 * of an error it will still be zeroed out.
410 	 */
411 	err = mlx5_query_port_eth_proto_oper(mdev, &eth_prot_oper,
412 					     mdev_port_num);
413 	if (err)
414 		goto out;
415 
416 	props->active_width     = IB_WIDTH_4X;
417 	props->active_speed     = IB_SPEED_QDR;
418 
419 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
420 				 &props->active_width);
421 
422 	props->port_cap_flags  |= IB_PORT_CM_SUP;
423 	props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
424 
425 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
426 						roce_address_table_size);
427 	props->max_mtu          = IB_MTU_4096;
428 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
429 	props->pkey_tbl_len     = 1;
430 	props->state            = IB_PORT_DOWN;
431 	props->phys_state       = 3;
432 
433 	mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
434 	props->qkey_viol_cntr = qkey_viol_cntr;
435 
436 	/* If this is a stub query for an unaffiliated port stop here */
437 	if (!put_mdev)
438 		goto out;
439 
440 	ndev = mlx5_ib_get_netdev(device, port_num);
441 	if (!ndev)
442 		goto out;
443 
444 	if (mlx5_lag_is_active(dev->mdev)) {
445 		rcu_read_lock();
446 		upper = netdev_master_upper_dev_get_rcu(ndev);
447 		if (upper) {
448 			dev_put(ndev);
449 			ndev = upper;
450 			dev_hold(ndev);
451 		}
452 		rcu_read_unlock();
453 	}
454 
455 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
456 		props->state      = IB_PORT_ACTIVE;
457 		props->phys_state = 5;
458 	}
459 
460 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
461 
462 	dev_put(ndev);
463 
464 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
465 out:
466 	if (put_mdev)
467 		mlx5_ib_put_native_port_mdev(dev, port_num);
468 	return err;
469 }
470 
471 static int set_roce_addr(struct mlx5_ib_dev *dev, u8 port_num,
472 			 unsigned int index, const union ib_gid *gid,
473 			 const struct ib_gid_attr *attr)
474 {
475 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
476 	u8 roce_version = 0;
477 	u8 roce_l3_type = 0;
478 	bool vlan = false;
479 	u8 mac[ETH_ALEN];
480 	u16 vlan_id = 0;
481 
482 	if (gid) {
483 		gid_type = attr->gid_type;
484 		ether_addr_copy(mac, attr->ndev->dev_addr);
485 
486 		if (is_vlan_dev(attr->ndev)) {
487 			vlan = true;
488 			vlan_id = vlan_dev_vlan_id(attr->ndev);
489 		}
490 	}
491 
492 	switch (gid_type) {
493 	case IB_GID_TYPE_IB:
494 		roce_version = MLX5_ROCE_VERSION_1;
495 		break;
496 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
497 		roce_version = MLX5_ROCE_VERSION_2;
498 		if (ipv6_addr_v4mapped((void *)gid))
499 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
500 		else
501 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
502 		break;
503 
504 	default:
505 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
506 	}
507 
508 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
509 				      roce_l3_type, gid->raw, mac, vlan,
510 				      vlan_id, port_num);
511 }
512 
513 static int mlx5_ib_add_gid(const union ib_gid *gid,
514 			   const struct ib_gid_attr *attr,
515 			   __always_unused void **context)
516 {
517 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
518 			     attr->index, gid, attr);
519 }
520 
521 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
522 			   __always_unused void **context)
523 {
524 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
525 			     attr->index, NULL, NULL);
526 }
527 
528 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
529 			       int index)
530 {
531 	struct ib_gid_attr attr;
532 	union ib_gid gid;
533 
534 	if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
535 		return 0;
536 
537 	dev_put(attr.ndev);
538 
539 	if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
540 		return 0;
541 
542 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
543 }
544 
545 int mlx5_get_roce_gid_type(struct mlx5_ib_dev *dev, u8 port_num,
546 			   int index, enum ib_gid_type *gid_type)
547 {
548 	struct ib_gid_attr attr;
549 	union ib_gid gid;
550 	int ret;
551 
552 	ret = ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr);
553 	if (ret)
554 		return ret;
555 
556 	dev_put(attr.ndev);
557 
558 	*gid_type = attr.gid_type;
559 
560 	return 0;
561 }
562 
563 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
564 {
565 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
566 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
567 	return 0;
568 }
569 
570 enum {
571 	MLX5_VPORT_ACCESS_METHOD_MAD,
572 	MLX5_VPORT_ACCESS_METHOD_HCA,
573 	MLX5_VPORT_ACCESS_METHOD_NIC,
574 };
575 
576 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
577 {
578 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
579 		return MLX5_VPORT_ACCESS_METHOD_MAD;
580 
581 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
582 	    IB_LINK_LAYER_ETHERNET)
583 		return MLX5_VPORT_ACCESS_METHOD_NIC;
584 
585 	return MLX5_VPORT_ACCESS_METHOD_HCA;
586 }
587 
588 static void get_atomic_caps(struct mlx5_ib_dev *dev,
589 			    u8 atomic_size_qp,
590 			    struct ib_device_attr *props)
591 {
592 	u8 tmp;
593 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
594 	u8 atomic_req_8B_endianness_mode =
595 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
596 
597 	/* Check if HW supports 8 bytes standard atomic operations and capable
598 	 * of host endianness respond
599 	 */
600 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
601 	if (((atomic_operations & tmp) == tmp) &&
602 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
603 	    (atomic_req_8B_endianness_mode)) {
604 		props->atomic_cap = IB_ATOMIC_HCA;
605 	} else {
606 		props->atomic_cap = IB_ATOMIC_NONE;
607 	}
608 }
609 
610 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
611 			       struct ib_device_attr *props)
612 {
613 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
614 
615 	get_atomic_caps(dev, atomic_size_qp, props);
616 }
617 
618 static void get_atomic_caps_dc(struct mlx5_ib_dev *dev,
619 			       struct ib_device_attr *props)
620 {
621 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
622 
623 	get_atomic_caps(dev, atomic_size_qp, props);
624 }
625 
626 bool mlx5_ib_dc_atomic_is_supported(struct mlx5_ib_dev *dev)
627 {
628 	struct ib_device_attr props = {};
629 
630 	get_atomic_caps_dc(dev, &props);
631 	return (props.atomic_cap == IB_ATOMIC_HCA) ? true : false;
632 }
633 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
634 					__be64 *sys_image_guid)
635 {
636 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
637 	struct mlx5_core_dev *mdev = dev->mdev;
638 	u64 tmp;
639 	int err;
640 
641 	switch (mlx5_get_vport_access_method(ibdev)) {
642 	case MLX5_VPORT_ACCESS_METHOD_MAD:
643 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
644 							    sys_image_guid);
645 
646 	case MLX5_VPORT_ACCESS_METHOD_HCA:
647 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
648 		break;
649 
650 	case MLX5_VPORT_ACCESS_METHOD_NIC:
651 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
652 		break;
653 
654 	default:
655 		return -EINVAL;
656 	}
657 
658 	if (!err)
659 		*sys_image_guid = cpu_to_be64(tmp);
660 
661 	return err;
662 
663 }
664 
665 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
666 				u16 *max_pkeys)
667 {
668 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
669 	struct mlx5_core_dev *mdev = dev->mdev;
670 
671 	switch (mlx5_get_vport_access_method(ibdev)) {
672 	case MLX5_VPORT_ACCESS_METHOD_MAD:
673 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
674 
675 	case MLX5_VPORT_ACCESS_METHOD_HCA:
676 	case MLX5_VPORT_ACCESS_METHOD_NIC:
677 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
678 						pkey_table_size));
679 		return 0;
680 
681 	default:
682 		return -EINVAL;
683 	}
684 }
685 
686 static int mlx5_query_vendor_id(struct ib_device *ibdev,
687 				u32 *vendor_id)
688 {
689 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
690 
691 	switch (mlx5_get_vport_access_method(ibdev)) {
692 	case MLX5_VPORT_ACCESS_METHOD_MAD:
693 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
694 
695 	case MLX5_VPORT_ACCESS_METHOD_HCA:
696 	case MLX5_VPORT_ACCESS_METHOD_NIC:
697 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
698 
699 	default:
700 		return -EINVAL;
701 	}
702 }
703 
704 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
705 				__be64 *node_guid)
706 {
707 	u64 tmp;
708 	int err;
709 
710 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
711 	case MLX5_VPORT_ACCESS_METHOD_MAD:
712 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
713 
714 	case MLX5_VPORT_ACCESS_METHOD_HCA:
715 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
716 		break;
717 
718 	case MLX5_VPORT_ACCESS_METHOD_NIC:
719 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
720 		break;
721 
722 	default:
723 		return -EINVAL;
724 	}
725 
726 	if (!err)
727 		*node_guid = cpu_to_be64(tmp);
728 
729 	return err;
730 }
731 
732 struct mlx5_reg_node_desc {
733 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
734 };
735 
736 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
737 {
738 	struct mlx5_reg_node_desc in;
739 
740 	if (mlx5_use_mad_ifc(dev))
741 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
742 
743 	memset(&in, 0, sizeof(in));
744 
745 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
746 				    sizeof(struct mlx5_reg_node_desc),
747 				    MLX5_REG_NODE_DESC, 0, 0);
748 }
749 
750 static int mlx5_ib_query_device(struct ib_device *ibdev,
751 				struct ib_device_attr *props,
752 				struct ib_udata *uhw)
753 {
754 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
755 	struct mlx5_core_dev *mdev = dev->mdev;
756 	int err = -ENOMEM;
757 	int max_sq_desc;
758 	int max_rq_sg;
759 	int max_sq_sg;
760 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
761 	bool raw_support = !mlx5_core_mp_enabled(mdev);
762 	struct mlx5_ib_query_device_resp resp = {};
763 	size_t resp_len;
764 	u64 max_tso;
765 
766 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
767 	if (uhw->outlen && uhw->outlen < resp_len)
768 		return -EINVAL;
769 	else
770 		resp.response_length = resp_len;
771 
772 	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
773 		return -EINVAL;
774 
775 	memset(props, 0, sizeof(*props));
776 	err = mlx5_query_system_image_guid(ibdev,
777 					   &props->sys_image_guid);
778 	if (err)
779 		return err;
780 
781 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
782 	if (err)
783 		return err;
784 
785 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
786 	if (err)
787 		return err;
788 
789 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
790 		(fw_rev_min(dev->mdev) << 16) |
791 		fw_rev_sub(dev->mdev);
792 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
793 		IB_DEVICE_PORT_ACTIVE_EVENT		|
794 		IB_DEVICE_SYS_IMAGE_GUID		|
795 		IB_DEVICE_RC_RNR_NAK_GEN;
796 
797 	if (MLX5_CAP_GEN(mdev, pkv))
798 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
799 	if (MLX5_CAP_GEN(mdev, qkv))
800 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
801 	if (MLX5_CAP_GEN(mdev, apm))
802 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
803 	if (MLX5_CAP_GEN(mdev, xrc))
804 		props->device_cap_flags |= IB_DEVICE_XRC;
805 	if (MLX5_CAP_GEN(mdev, imaicl)) {
806 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
807 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
808 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
809 		/* We support 'Gappy' memory registration too */
810 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
811 	}
812 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
813 	if (MLX5_CAP_GEN(mdev, sho)) {
814 		props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
815 		/* At this stage no support for signature handover */
816 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
817 				      IB_PROT_T10DIF_TYPE_2 |
818 				      IB_PROT_T10DIF_TYPE_3;
819 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
820 				       IB_GUARD_T10DIF_CSUM;
821 	}
822 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
823 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
824 
825 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
826 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
827 			/* Legacy bit to support old userspace libraries */
828 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
829 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
830 		}
831 
832 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
833 			props->raw_packet_caps |=
834 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
835 
836 		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
837 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
838 			if (max_tso) {
839 				resp.tso_caps.max_tso = 1 << max_tso;
840 				resp.tso_caps.supported_qpts |=
841 					1 << IB_QPT_RAW_PACKET;
842 				resp.response_length += sizeof(resp.tso_caps);
843 			}
844 		}
845 
846 		if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
847 			resp.rss_caps.rx_hash_function =
848 						MLX5_RX_HASH_FUNC_TOEPLITZ;
849 			resp.rss_caps.rx_hash_fields_mask =
850 						MLX5_RX_HASH_SRC_IPV4 |
851 						MLX5_RX_HASH_DST_IPV4 |
852 						MLX5_RX_HASH_SRC_IPV6 |
853 						MLX5_RX_HASH_DST_IPV6 |
854 						MLX5_RX_HASH_SRC_PORT_TCP |
855 						MLX5_RX_HASH_DST_PORT_TCP |
856 						MLX5_RX_HASH_SRC_PORT_UDP |
857 						MLX5_RX_HASH_DST_PORT_UDP |
858 						MLX5_RX_HASH_INNER;
859 			if (mlx5_accel_ipsec_device_caps(dev->mdev) &
860 			    MLX5_ACCEL_IPSEC_CAP_DEVICE)
861 				resp.rss_caps.rx_hash_fields_mask |=
862 					MLX5_RX_HASH_IPSEC_SPI;
863 			resp.response_length += sizeof(resp.rss_caps);
864 		}
865 	} else {
866 		if (field_avail(typeof(resp), tso_caps, uhw->outlen))
867 			resp.response_length += sizeof(resp.tso_caps);
868 		if (field_avail(typeof(resp), rss_caps, uhw->outlen))
869 			resp.response_length += sizeof(resp.rss_caps);
870 	}
871 
872 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
873 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
874 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
875 	}
876 
877 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
878 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
879 	    raw_support)
880 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
881 
882 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
883 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
884 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
885 
886 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
887 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
888 	    raw_support) {
889 		/* Legacy bit to support old userspace libraries */
890 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
891 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
892 	}
893 
894 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
895 		props->max_dm_size =
896 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
897 	}
898 
899 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
900 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
901 
902 	if (MLX5_CAP_GEN(mdev, end_pad))
903 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
904 
905 	props->vendor_part_id	   = mdev->pdev->device;
906 	props->hw_ver		   = mdev->pdev->revision;
907 
908 	props->max_mr_size	   = ~0ull;
909 	props->page_size_cap	   = ~(min_page_size - 1);
910 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
911 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
912 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
913 		     sizeof(struct mlx5_wqe_data_seg);
914 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
915 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
916 		     sizeof(struct mlx5_wqe_raddr_seg)) /
917 		sizeof(struct mlx5_wqe_data_seg);
918 	props->max_sge = min(max_rq_sg, max_sq_sg);
919 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
920 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
921 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
922 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
923 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
924 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
925 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
926 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
927 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
928 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
929 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
930 	props->max_srq_sge	   = max_rq_sg - 1;
931 	props->max_fast_reg_page_list_len =
932 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
933 	get_atomic_caps_qp(dev, props);
934 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
935 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
936 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
937 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
938 					   props->max_mcast_grp;
939 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
940 	props->max_ah = INT_MAX;
941 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
942 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
943 
944 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
945 	if (MLX5_CAP_GEN(mdev, pg))
946 		props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
947 	props->odp_caps = dev->odp_caps;
948 #endif
949 
950 	if (MLX5_CAP_GEN(mdev, cd))
951 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
952 
953 	if (!mlx5_core_is_pf(mdev))
954 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
955 
956 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
957 	    IB_LINK_LAYER_ETHERNET && raw_support) {
958 		props->rss_caps.max_rwq_indirection_tables =
959 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
960 		props->rss_caps.max_rwq_indirection_table_size =
961 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
962 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
963 		props->max_wq_type_rq =
964 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
965 	}
966 
967 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
968 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
969 		props->tm_caps.max_num_tags =
970 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
971 		props->tm_caps.flags = IB_TM_CAP_RC;
972 		props->tm_caps.max_ops =
973 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
974 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
975 	}
976 
977 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
978 		props->cq_caps.max_cq_moderation_count =
979 						MLX5_MAX_CQ_COUNT;
980 		props->cq_caps.max_cq_moderation_period =
981 						MLX5_MAX_CQ_PERIOD;
982 	}
983 
984 	if (field_avail(typeof(resp), cqe_comp_caps, uhw->outlen)) {
985 		resp.response_length += sizeof(resp.cqe_comp_caps);
986 
987 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
988 			resp.cqe_comp_caps.max_num =
989 				MLX5_CAP_GEN(dev->mdev,
990 					     cqe_compression_max_num);
991 
992 			resp.cqe_comp_caps.supported_format =
993 				MLX5_IB_CQE_RES_FORMAT_HASH |
994 				MLX5_IB_CQE_RES_FORMAT_CSUM;
995 
996 			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
997 				resp.cqe_comp_caps.supported_format |=
998 					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
999 		}
1000 	}
1001 
1002 	if (field_avail(typeof(resp), packet_pacing_caps, uhw->outlen) &&
1003 	    raw_support) {
1004 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1005 		    MLX5_CAP_GEN(mdev, qos)) {
1006 			resp.packet_pacing_caps.qp_rate_limit_max =
1007 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1008 			resp.packet_pacing_caps.qp_rate_limit_min =
1009 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1010 			resp.packet_pacing_caps.supported_qpts |=
1011 				1 << IB_QPT_RAW_PACKET;
1012 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1013 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1014 				resp.packet_pacing_caps.cap_flags |=
1015 					MLX5_IB_PP_SUPPORT_BURST;
1016 		}
1017 		resp.response_length += sizeof(resp.packet_pacing_caps);
1018 	}
1019 
1020 	if (field_avail(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes,
1021 			uhw->outlen)) {
1022 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1023 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1024 				MLX5_IB_ALLOW_MPW;
1025 
1026 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1027 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1028 				MLX5_IB_SUPPORT_EMPW;
1029 
1030 		resp.response_length +=
1031 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1032 	}
1033 
1034 	if (field_avail(typeof(resp), flags, uhw->outlen)) {
1035 		resp.response_length += sizeof(resp.flags);
1036 
1037 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1038 			resp.flags |=
1039 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1040 
1041 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1042 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1043 	}
1044 
1045 	if (field_avail(typeof(resp), sw_parsing_caps,
1046 			uhw->outlen)) {
1047 		resp.response_length += sizeof(resp.sw_parsing_caps);
1048 		if (MLX5_CAP_ETH(mdev, swp)) {
1049 			resp.sw_parsing_caps.sw_parsing_offloads |=
1050 				MLX5_IB_SW_PARSING;
1051 
1052 			if (MLX5_CAP_ETH(mdev, swp_csum))
1053 				resp.sw_parsing_caps.sw_parsing_offloads |=
1054 					MLX5_IB_SW_PARSING_CSUM;
1055 
1056 			if (MLX5_CAP_ETH(mdev, swp_lso))
1057 				resp.sw_parsing_caps.sw_parsing_offloads |=
1058 					MLX5_IB_SW_PARSING_LSO;
1059 
1060 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1061 				resp.sw_parsing_caps.supported_qpts =
1062 					BIT(IB_QPT_RAW_PACKET);
1063 		}
1064 	}
1065 
1066 	if (field_avail(typeof(resp), striding_rq_caps, uhw->outlen) &&
1067 	    raw_support) {
1068 		resp.response_length += sizeof(resp.striding_rq_caps);
1069 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1070 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1071 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1072 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1073 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1074 			resp.striding_rq_caps.min_single_wqe_log_num_of_strides =
1075 				MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1076 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1077 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1078 			resp.striding_rq_caps.supported_qpts =
1079 				BIT(IB_QPT_RAW_PACKET);
1080 		}
1081 	}
1082 
1083 	if (field_avail(typeof(resp), tunnel_offloads_caps,
1084 			uhw->outlen)) {
1085 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1086 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1087 			resp.tunnel_offloads_caps |=
1088 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1089 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1090 			resp.tunnel_offloads_caps |=
1091 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1092 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1093 			resp.tunnel_offloads_caps |=
1094 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1095 		if (MLX5_CAP_GEN(mdev, flex_parser_protocols) &
1096 		    MLX5_FLEX_PROTO_CW_MPLS_GRE)
1097 			resp.tunnel_offloads_caps |=
1098 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1099 		if (MLX5_CAP_GEN(mdev, flex_parser_protocols) &
1100 		    MLX5_FLEX_PROTO_CW_MPLS_UDP)
1101 			resp.tunnel_offloads_caps |=
1102 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1103 	}
1104 
1105 	if (uhw->outlen) {
1106 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1107 
1108 		if (err)
1109 			return err;
1110 	}
1111 
1112 	return 0;
1113 }
1114 
1115 enum mlx5_ib_width {
1116 	MLX5_IB_WIDTH_1X	= 1 << 0,
1117 	MLX5_IB_WIDTH_2X	= 1 << 1,
1118 	MLX5_IB_WIDTH_4X	= 1 << 2,
1119 	MLX5_IB_WIDTH_8X	= 1 << 3,
1120 	MLX5_IB_WIDTH_12X	= 1 << 4
1121 };
1122 
1123 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
1124 				  u8 *ib_width)
1125 {
1126 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1127 	int err = 0;
1128 
1129 	if (active_width & MLX5_IB_WIDTH_1X) {
1130 		*ib_width = IB_WIDTH_1X;
1131 	} else if (active_width & MLX5_IB_WIDTH_2X) {
1132 		mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
1133 			    (int)active_width);
1134 		err = -EINVAL;
1135 	} else if (active_width & MLX5_IB_WIDTH_4X) {
1136 		*ib_width = IB_WIDTH_4X;
1137 	} else if (active_width & MLX5_IB_WIDTH_8X) {
1138 		*ib_width = IB_WIDTH_8X;
1139 	} else if (active_width & MLX5_IB_WIDTH_12X) {
1140 		*ib_width = IB_WIDTH_12X;
1141 	} else {
1142 		mlx5_ib_dbg(dev, "Invalid active_width %d\n",
1143 			    (int)active_width);
1144 		err = -EINVAL;
1145 	}
1146 
1147 	return err;
1148 }
1149 
1150 static int mlx5_mtu_to_ib_mtu(int mtu)
1151 {
1152 	switch (mtu) {
1153 	case 256: return 1;
1154 	case 512: return 2;
1155 	case 1024: return 3;
1156 	case 2048: return 4;
1157 	case 4096: return 5;
1158 	default:
1159 		pr_warn("invalid mtu\n");
1160 		return -1;
1161 	}
1162 }
1163 
1164 enum ib_max_vl_num {
1165 	__IB_MAX_VL_0		= 1,
1166 	__IB_MAX_VL_0_1		= 2,
1167 	__IB_MAX_VL_0_3		= 3,
1168 	__IB_MAX_VL_0_7		= 4,
1169 	__IB_MAX_VL_0_14	= 5,
1170 };
1171 
1172 enum mlx5_vl_hw_cap {
1173 	MLX5_VL_HW_0	= 1,
1174 	MLX5_VL_HW_0_1	= 2,
1175 	MLX5_VL_HW_0_2	= 3,
1176 	MLX5_VL_HW_0_3	= 4,
1177 	MLX5_VL_HW_0_4	= 5,
1178 	MLX5_VL_HW_0_5	= 6,
1179 	MLX5_VL_HW_0_6	= 7,
1180 	MLX5_VL_HW_0_7	= 8,
1181 	MLX5_VL_HW_0_14	= 15
1182 };
1183 
1184 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1185 				u8 *max_vl_num)
1186 {
1187 	switch (vl_hw_cap) {
1188 	case MLX5_VL_HW_0:
1189 		*max_vl_num = __IB_MAX_VL_0;
1190 		break;
1191 	case MLX5_VL_HW_0_1:
1192 		*max_vl_num = __IB_MAX_VL_0_1;
1193 		break;
1194 	case MLX5_VL_HW_0_3:
1195 		*max_vl_num = __IB_MAX_VL_0_3;
1196 		break;
1197 	case MLX5_VL_HW_0_7:
1198 		*max_vl_num = __IB_MAX_VL_0_7;
1199 		break;
1200 	case MLX5_VL_HW_0_14:
1201 		*max_vl_num = __IB_MAX_VL_0_14;
1202 		break;
1203 
1204 	default:
1205 		return -EINVAL;
1206 	}
1207 
1208 	return 0;
1209 }
1210 
1211 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
1212 			       struct ib_port_attr *props)
1213 {
1214 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1215 	struct mlx5_core_dev *mdev = dev->mdev;
1216 	struct mlx5_hca_vport_context *rep;
1217 	u16 max_mtu;
1218 	u16 oper_mtu;
1219 	int err;
1220 	u8 ib_link_width_oper;
1221 	u8 vl_hw_cap;
1222 
1223 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1224 	if (!rep) {
1225 		err = -ENOMEM;
1226 		goto out;
1227 	}
1228 
1229 	/* props being zeroed by the caller, avoid zeroing it here */
1230 
1231 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1232 	if (err)
1233 		goto out;
1234 
1235 	props->lid		= rep->lid;
1236 	props->lmc		= rep->lmc;
1237 	props->sm_lid		= rep->sm_lid;
1238 	props->sm_sl		= rep->sm_sl;
1239 	props->state		= rep->vport_state;
1240 	props->phys_state	= rep->port_physical_state;
1241 	props->port_cap_flags	= rep->cap_mask1;
1242 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1243 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1244 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1245 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1246 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1247 	props->subnet_timeout	= rep->subnet_timeout;
1248 	props->init_type_reply	= rep->init_type_reply;
1249 	props->grh_required	= rep->grh_required;
1250 
1251 	err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
1252 	if (err)
1253 		goto out;
1254 
1255 	err = translate_active_width(ibdev, ib_link_width_oper,
1256 				     &props->active_width);
1257 	if (err)
1258 		goto out;
1259 	err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
1260 	if (err)
1261 		goto out;
1262 
1263 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1264 
1265 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1266 
1267 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1268 
1269 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1270 
1271 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1272 	if (err)
1273 		goto out;
1274 
1275 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1276 				   &props->max_vl_num);
1277 out:
1278 	kfree(rep);
1279 	return err;
1280 }
1281 
1282 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
1283 		       struct ib_port_attr *props)
1284 {
1285 	unsigned int count;
1286 	int ret;
1287 
1288 	switch (mlx5_get_vport_access_method(ibdev)) {
1289 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1290 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1291 		break;
1292 
1293 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1294 		ret = mlx5_query_hca_port(ibdev, port, props);
1295 		break;
1296 
1297 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1298 		ret = mlx5_query_port_roce(ibdev, port, props);
1299 		break;
1300 
1301 	default:
1302 		ret = -EINVAL;
1303 	}
1304 
1305 	if (!ret && props) {
1306 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1307 		struct mlx5_core_dev *mdev;
1308 		bool put_mdev = true;
1309 
1310 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1311 		if (!mdev) {
1312 			/* If the port isn't affiliated yet query the master.
1313 			 * The master and slave will have the same values.
1314 			 */
1315 			mdev = dev->mdev;
1316 			port = 1;
1317 			put_mdev = false;
1318 		}
1319 		count = mlx5_core_reserved_gids_count(mdev);
1320 		if (put_mdev)
1321 			mlx5_ib_put_native_port_mdev(dev, port);
1322 		props->gid_tbl_len -= count;
1323 	}
1324 	return ret;
1325 }
1326 
1327 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u8 port,
1328 				  struct ib_port_attr *props)
1329 {
1330 	int ret;
1331 
1332 	/* Only link layer == ethernet is valid for representors */
1333 	ret = mlx5_query_port_roce(ibdev, port, props);
1334 	if (ret || !props)
1335 		return ret;
1336 
1337 	/* We don't support GIDS */
1338 	props->gid_tbl_len = 0;
1339 
1340 	return ret;
1341 }
1342 
1343 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1344 			     union ib_gid *gid)
1345 {
1346 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1347 	struct mlx5_core_dev *mdev = dev->mdev;
1348 
1349 	switch (mlx5_get_vport_access_method(ibdev)) {
1350 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1351 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1352 
1353 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1354 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1355 
1356 	default:
1357 		return -EINVAL;
1358 	}
1359 
1360 }
1361 
1362 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u8 port,
1363 				   u16 index, u16 *pkey)
1364 {
1365 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1366 	struct mlx5_core_dev *mdev;
1367 	bool put_mdev = true;
1368 	u8 mdev_port_num;
1369 	int err;
1370 
1371 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1372 	if (!mdev) {
1373 		/* The port isn't affiliated yet, get the PKey from the master
1374 		 * port. For RoCE the PKey tables will be the same.
1375 		 */
1376 		put_mdev = false;
1377 		mdev = dev->mdev;
1378 		mdev_port_num = 1;
1379 	}
1380 
1381 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1382 					index, pkey);
1383 	if (put_mdev)
1384 		mlx5_ib_put_native_port_mdev(dev, port);
1385 
1386 	return err;
1387 }
1388 
1389 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1390 			      u16 *pkey)
1391 {
1392 	switch (mlx5_get_vport_access_method(ibdev)) {
1393 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1394 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1395 
1396 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1397 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1398 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1399 	default:
1400 		return -EINVAL;
1401 	}
1402 }
1403 
1404 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1405 				 struct ib_device_modify *props)
1406 {
1407 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1408 	struct mlx5_reg_node_desc in;
1409 	struct mlx5_reg_node_desc out;
1410 	int err;
1411 
1412 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1413 		return -EOPNOTSUPP;
1414 
1415 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1416 		return 0;
1417 
1418 	/*
1419 	 * If possible, pass node desc to FW, so it can generate
1420 	 * a 144 trap.  If cmd fails, just ignore.
1421 	 */
1422 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1423 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1424 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1425 	if (err)
1426 		return err;
1427 
1428 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1429 
1430 	return err;
1431 }
1432 
1433 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1434 				u32 value)
1435 {
1436 	struct mlx5_hca_vport_context ctx = {};
1437 	struct mlx5_core_dev *mdev;
1438 	u8 mdev_port_num;
1439 	int err;
1440 
1441 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1442 	if (!mdev)
1443 		return -ENODEV;
1444 
1445 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1446 	if (err)
1447 		goto out;
1448 
1449 	if (~ctx.cap_mask1_perm & mask) {
1450 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1451 			     mask, ctx.cap_mask1_perm);
1452 		err = -EINVAL;
1453 		goto out;
1454 	}
1455 
1456 	ctx.cap_mask1 = value;
1457 	ctx.cap_mask1_perm = mask;
1458 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1459 						 0, &ctx);
1460 
1461 out:
1462 	mlx5_ib_put_native_port_mdev(dev, port_num);
1463 
1464 	return err;
1465 }
1466 
1467 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1468 			       struct ib_port_modify *props)
1469 {
1470 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1471 	struct ib_port_attr attr;
1472 	u32 tmp;
1473 	int err;
1474 	u32 change_mask;
1475 	u32 value;
1476 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1477 		      IB_LINK_LAYER_INFINIBAND);
1478 
1479 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1480 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1481 	 */
1482 	if (!is_ib)
1483 		return 0;
1484 
1485 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1486 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1487 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1488 		return set_port_caps_atomic(dev, port, change_mask, value);
1489 	}
1490 
1491 	mutex_lock(&dev->cap_mask_mutex);
1492 
1493 	err = ib_query_port(ibdev, port, &attr);
1494 	if (err)
1495 		goto out;
1496 
1497 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1498 		~props->clr_port_cap_mask;
1499 
1500 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1501 
1502 out:
1503 	mutex_unlock(&dev->cap_mask_mutex);
1504 	return err;
1505 }
1506 
1507 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1508 {
1509 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1510 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1511 }
1512 
1513 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1514 {
1515 	/* Large page with non 4k uar support might limit the dynamic size */
1516 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1517 		return MLX5_MIN_DYN_BFREGS;
1518 
1519 	return MLX5_MAX_DYN_BFREGS;
1520 }
1521 
1522 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1523 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1524 			     struct mlx5_bfreg_info *bfregi)
1525 {
1526 	int uars_per_sys_page;
1527 	int bfregs_per_sys_page;
1528 	int ref_bfregs = req->total_num_bfregs;
1529 
1530 	if (req->total_num_bfregs == 0)
1531 		return -EINVAL;
1532 
1533 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1534 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1535 
1536 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1537 		return -ENOMEM;
1538 
1539 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1540 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1541 	/* This holds the required static allocation asked by the user */
1542 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1543 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1544 		return -EINVAL;
1545 
1546 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1547 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1548 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1549 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1550 
1551 	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",
1552 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1553 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1554 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1555 		    bfregi->num_sys_pages);
1556 
1557 	return 0;
1558 }
1559 
1560 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1561 {
1562 	struct mlx5_bfreg_info *bfregi;
1563 	int err;
1564 	int i;
1565 
1566 	bfregi = &context->bfregi;
1567 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1568 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1569 		if (err)
1570 			goto error;
1571 
1572 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1573 	}
1574 
1575 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1576 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1577 
1578 	return 0;
1579 
1580 error:
1581 	for (--i; i >= 0; i--)
1582 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1583 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1584 
1585 	return err;
1586 }
1587 
1588 static int deallocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1589 {
1590 	struct mlx5_bfreg_info *bfregi;
1591 	int err;
1592 	int i;
1593 
1594 	bfregi = &context->bfregi;
1595 	for (i = 0; i < bfregi->num_sys_pages; i++) {
1596 		if (i < bfregi->num_static_sys_pages ||
1597 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX) {
1598 			err = mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1599 			if (err) {
1600 				mlx5_ib_warn(dev, "failed to free uar %d, err=%d\n", i, err);
1601 				return err;
1602 			}
1603 		}
1604 	}
1605 
1606 	return 0;
1607 }
1608 
1609 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn)
1610 {
1611 	int err;
1612 
1613 	err = mlx5_core_alloc_transport_domain(dev->mdev, tdn);
1614 	if (err)
1615 		return err;
1616 
1617 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1618 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1619 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1620 		return err;
1621 
1622 	mutex_lock(&dev->lb_mutex);
1623 	dev->user_td++;
1624 
1625 	if (dev->user_td == 2)
1626 		err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1627 
1628 	mutex_unlock(&dev->lb_mutex);
1629 	return err;
1630 }
1631 
1632 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn)
1633 {
1634 	mlx5_core_dealloc_transport_domain(dev->mdev, tdn);
1635 
1636 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1637 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1638 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1639 		return;
1640 
1641 	mutex_lock(&dev->lb_mutex);
1642 	dev->user_td--;
1643 
1644 	if (dev->user_td < 2)
1645 		mlx5_nic_vport_update_local_lb(dev->mdev, false);
1646 
1647 	mutex_unlock(&dev->lb_mutex);
1648 }
1649 
1650 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
1651 						  struct ib_udata *udata)
1652 {
1653 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1654 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1655 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1656 	struct mlx5_core_dev *mdev = dev->mdev;
1657 	struct mlx5_ib_ucontext *context;
1658 	struct mlx5_bfreg_info *bfregi;
1659 	int ver;
1660 	int err;
1661 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1662 				     max_cqe_version);
1663 	bool lib_uar_4k;
1664 
1665 	if (!dev->ib_active)
1666 		return ERR_PTR(-EAGAIN);
1667 
1668 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1669 		ver = 0;
1670 	else if (udata->inlen >= min_req_v2)
1671 		ver = 2;
1672 	else
1673 		return ERR_PTR(-EINVAL);
1674 
1675 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1676 	if (err)
1677 		return ERR_PTR(err);
1678 
1679 	if (req.flags)
1680 		return ERR_PTR(-EINVAL);
1681 
1682 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1683 		return ERR_PTR(-EOPNOTSUPP);
1684 
1685 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1686 				    MLX5_NON_FP_BFREGS_PER_UAR);
1687 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1688 		return ERR_PTR(-EINVAL);
1689 
1690 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1691 	if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1692 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1693 	resp.cache_line_size = cache_line_size();
1694 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1695 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1696 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1697 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1698 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1699 	resp.cqe_version = min_t(__u8,
1700 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1701 				 req.max_cqe_version);
1702 	resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1703 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1704 	resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1705 					MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1;
1706 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1707 				   sizeof(resp.response_length), udata->outlen);
1708 
1709 	if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1710 		if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_EGRESS))
1711 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1712 		if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1713 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1714 		if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1715 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1716 		if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1717 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1718 		/* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1719 	}
1720 
1721 	context = kzalloc(sizeof(*context), GFP_KERNEL);
1722 	if (!context)
1723 		return ERR_PTR(-ENOMEM);
1724 
1725 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1726 	bfregi = &context->bfregi;
1727 
1728 	/* updates req->total_num_bfregs */
1729 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1730 	if (err)
1731 		goto out_ctx;
1732 
1733 	mutex_init(&bfregi->lock);
1734 	bfregi->lib_uar_4k = lib_uar_4k;
1735 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1736 				GFP_KERNEL);
1737 	if (!bfregi->count) {
1738 		err = -ENOMEM;
1739 		goto out_ctx;
1740 	}
1741 
1742 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1743 				    sizeof(*bfregi->sys_pages),
1744 				    GFP_KERNEL);
1745 	if (!bfregi->sys_pages) {
1746 		err = -ENOMEM;
1747 		goto out_count;
1748 	}
1749 
1750 	err = allocate_uars(dev, context);
1751 	if (err)
1752 		goto out_sys_pages;
1753 
1754 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1755 	context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1756 #endif
1757 
1758 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1759 		err = mlx5_ib_alloc_transport_domain(dev, &context->tdn);
1760 		if (err)
1761 			goto out_uars;
1762 	}
1763 
1764 	INIT_LIST_HEAD(&context->vma_private_list);
1765 	mutex_init(&context->vma_private_list_mutex);
1766 	INIT_LIST_HEAD(&context->db_page_list);
1767 	mutex_init(&context->db_page_mutex);
1768 
1769 	resp.tot_bfregs = req.total_num_bfregs;
1770 	resp.num_ports = dev->num_ports;
1771 
1772 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1773 		resp.response_length += sizeof(resp.cqe_version);
1774 
1775 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1776 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1777 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1778 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1779 	}
1780 
1781 	if (field_avail(typeof(resp), eth_min_inline, udata->outlen)) {
1782 		if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1783 			mlx5_query_min_inline(dev->mdev, &resp.eth_min_inline);
1784 			resp.eth_min_inline++;
1785 		}
1786 		resp.response_length += sizeof(resp.eth_min_inline);
1787 	}
1788 
1789 	if (field_avail(typeof(resp), clock_info_versions, udata->outlen)) {
1790 		if (mdev->clock_info)
1791 			resp.clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1792 		resp.response_length += sizeof(resp.clock_info_versions);
1793 	}
1794 
1795 	/*
1796 	 * We don't want to expose information from the PCI bar that is located
1797 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1798 	 * pretend we don't support reading the HCA's core clock. This is also
1799 	 * forced by mmap function.
1800 	 */
1801 	if (field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1802 		if (PAGE_SIZE <= 4096) {
1803 			resp.comp_mask |=
1804 				MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1805 			resp.hca_core_clock_offset =
1806 				offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1807 		}
1808 		resp.response_length += sizeof(resp.hca_core_clock_offset);
1809 	}
1810 
1811 	if (field_avail(typeof(resp), log_uar_size, udata->outlen))
1812 		resp.response_length += sizeof(resp.log_uar_size);
1813 
1814 	if (field_avail(typeof(resp), num_uars_per_page, udata->outlen))
1815 		resp.response_length += sizeof(resp.num_uars_per_page);
1816 
1817 	if (field_avail(typeof(resp), num_dyn_bfregs, udata->outlen)) {
1818 		resp.num_dyn_bfregs = bfregi->num_dyn_bfregs;
1819 		resp.response_length += sizeof(resp.num_dyn_bfregs);
1820 	}
1821 
1822 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1823 	if (err)
1824 		goto out_td;
1825 
1826 	bfregi->ver = ver;
1827 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1828 	context->cqe_version = resp.cqe_version;
1829 	context->lib_caps = req.lib_caps;
1830 	print_lib_caps(dev, context->lib_caps);
1831 
1832 	return &context->ibucontext;
1833 
1834 out_td:
1835 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1836 		mlx5_ib_dealloc_transport_domain(dev, context->tdn);
1837 
1838 out_uars:
1839 	deallocate_uars(dev, context);
1840 
1841 out_sys_pages:
1842 	kfree(bfregi->sys_pages);
1843 
1844 out_count:
1845 	kfree(bfregi->count);
1846 
1847 out_ctx:
1848 	kfree(context);
1849 
1850 	return ERR_PTR(err);
1851 }
1852 
1853 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1854 {
1855 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1856 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1857 	struct mlx5_bfreg_info *bfregi;
1858 
1859 	bfregi = &context->bfregi;
1860 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1861 		mlx5_ib_dealloc_transport_domain(dev, context->tdn);
1862 
1863 	deallocate_uars(dev, context);
1864 	kfree(bfregi->sys_pages);
1865 	kfree(bfregi->count);
1866 	kfree(context);
1867 
1868 	return 0;
1869 }
1870 
1871 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
1872 				 int uar_idx)
1873 {
1874 	int fw_uars_per_page;
1875 
1876 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
1877 
1878 	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
1879 }
1880 
1881 static int get_command(unsigned long offset)
1882 {
1883 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1884 }
1885 
1886 static int get_arg(unsigned long offset)
1887 {
1888 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1889 }
1890 
1891 static int get_index(unsigned long offset)
1892 {
1893 	return get_arg(offset);
1894 }
1895 
1896 /* Index resides in an extra byte to enable larger values than 255 */
1897 static int get_extended_index(unsigned long offset)
1898 {
1899 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
1900 }
1901 
1902 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1903 {
1904 	/* vma_open is called when a new VMA is created on top of our VMA.  This
1905 	 * is done through either mremap flow or split_vma (usually due to
1906 	 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1907 	 * as this VMA is strongly hardware related.  Therefore we set the
1908 	 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1909 	 * calling us again and trying to do incorrect actions.  We assume that
1910 	 * the original VMA size is exactly a single page, and therefore all
1911 	 * "splitting" operation will not happen to it.
1912 	 */
1913 	area->vm_ops = NULL;
1914 }
1915 
1916 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1917 {
1918 	struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1919 
1920 	/* It's guaranteed that all VMAs opened on a FD are closed before the
1921 	 * file itself is closed, therefore no sync is needed with the regular
1922 	 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1923 	 * However need a sync with accessing the vma as part of
1924 	 * mlx5_ib_disassociate_ucontext.
1925 	 * The close operation is usually called under mm->mmap_sem except when
1926 	 * process is exiting.
1927 	 * The exiting case is handled explicitly as part of
1928 	 * mlx5_ib_disassociate_ucontext.
1929 	 */
1930 	mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1931 
1932 	/* setting the vma context pointer to null in the mlx5_ib driver's
1933 	 * private data, to protect a race condition in
1934 	 * mlx5_ib_disassociate_ucontext().
1935 	 */
1936 	mlx5_ib_vma_priv_data->vma = NULL;
1937 	mutex_lock(mlx5_ib_vma_priv_data->vma_private_list_mutex);
1938 	list_del(&mlx5_ib_vma_priv_data->list);
1939 	mutex_unlock(mlx5_ib_vma_priv_data->vma_private_list_mutex);
1940 	kfree(mlx5_ib_vma_priv_data);
1941 }
1942 
1943 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1944 	.open = mlx5_ib_vma_open,
1945 	.close = mlx5_ib_vma_close
1946 };
1947 
1948 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1949 				struct mlx5_ib_ucontext *ctx)
1950 {
1951 	struct mlx5_ib_vma_private_data *vma_prv;
1952 	struct list_head *vma_head = &ctx->vma_private_list;
1953 
1954 	vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1955 	if (!vma_prv)
1956 		return -ENOMEM;
1957 
1958 	vma_prv->vma = vma;
1959 	vma_prv->vma_private_list_mutex = &ctx->vma_private_list_mutex;
1960 	vma->vm_private_data = vma_prv;
1961 	vma->vm_ops =  &mlx5_ib_vm_ops;
1962 
1963 	mutex_lock(&ctx->vma_private_list_mutex);
1964 	list_add(&vma_prv->list, vma_head);
1965 	mutex_unlock(&ctx->vma_private_list_mutex);
1966 
1967 	return 0;
1968 }
1969 
1970 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1971 {
1972 	struct vm_area_struct *vma;
1973 	struct mlx5_ib_vma_private_data *vma_private, *n;
1974 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1975 
1976 	mutex_lock(&context->vma_private_list_mutex);
1977 	list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1978 				 list) {
1979 		vma = vma_private->vma;
1980 		zap_vma_ptes(vma, vma->vm_start, PAGE_SIZE);
1981 		/* context going to be destroyed, should
1982 		 * not access ops any more.
1983 		 */
1984 		vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
1985 		vma->vm_ops = NULL;
1986 		list_del(&vma_private->list);
1987 		kfree(vma_private);
1988 	}
1989 	mutex_unlock(&context->vma_private_list_mutex);
1990 }
1991 
1992 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1993 {
1994 	switch (cmd) {
1995 	case MLX5_IB_MMAP_WC_PAGE:
1996 		return "WC";
1997 	case MLX5_IB_MMAP_REGULAR_PAGE:
1998 		return "best effort WC";
1999 	case MLX5_IB_MMAP_NC_PAGE:
2000 		return "NC";
2001 	case MLX5_IB_MMAP_DEVICE_MEM:
2002 		return "Device Memory";
2003 	default:
2004 		return NULL;
2005 	}
2006 }
2007 
2008 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2009 					struct vm_area_struct *vma,
2010 					struct mlx5_ib_ucontext *context)
2011 {
2012 	phys_addr_t pfn;
2013 	int err;
2014 
2015 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2016 		return -EINVAL;
2017 
2018 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2019 		return -EOPNOTSUPP;
2020 
2021 	if (vma->vm_flags & VM_WRITE)
2022 		return -EPERM;
2023 
2024 	if (!dev->mdev->clock_info_page)
2025 		return -EOPNOTSUPP;
2026 
2027 	pfn = page_to_pfn(dev->mdev->clock_info_page);
2028 	err = remap_pfn_range(vma, vma->vm_start, pfn, PAGE_SIZE,
2029 			      vma->vm_page_prot);
2030 	if (err)
2031 		return err;
2032 
2033 	return mlx5_ib_set_vma_data(vma, context);
2034 }
2035 
2036 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2037 		    struct vm_area_struct *vma,
2038 		    struct mlx5_ib_ucontext *context)
2039 {
2040 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2041 	int err;
2042 	unsigned long idx;
2043 	phys_addr_t pfn, pa;
2044 	pgprot_t prot;
2045 	u32 bfreg_dyn_idx = 0;
2046 	u32 uar_index;
2047 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2048 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2049 				bfregi->num_static_sys_pages;
2050 
2051 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2052 		return -EINVAL;
2053 
2054 	if (dyn_uar)
2055 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2056 	else
2057 		idx = get_index(vma->vm_pgoff);
2058 
2059 	if (idx >= max_valid_idx) {
2060 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2061 			     idx, max_valid_idx);
2062 		return -EINVAL;
2063 	}
2064 
2065 	switch (cmd) {
2066 	case MLX5_IB_MMAP_WC_PAGE:
2067 	case MLX5_IB_MMAP_ALLOC_WC:
2068 /* Some architectures don't support WC memory */
2069 #if defined(CONFIG_X86)
2070 		if (!pat_enabled())
2071 			return -EPERM;
2072 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
2073 			return -EPERM;
2074 #endif
2075 	/* fall through */
2076 	case MLX5_IB_MMAP_REGULAR_PAGE:
2077 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2078 		prot = pgprot_writecombine(vma->vm_page_prot);
2079 		break;
2080 	case MLX5_IB_MMAP_NC_PAGE:
2081 		prot = pgprot_noncached(vma->vm_page_prot);
2082 		break;
2083 	default:
2084 		return -EINVAL;
2085 	}
2086 
2087 	if (dyn_uar) {
2088 		int uars_per_page;
2089 
2090 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2091 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2092 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2093 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2094 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2095 			return -EINVAL;
2096 		}
2097 
2098 		mutex_lock(&bfregi->lock);
2099 		/* Fail if uar already allocated, first bfreg index of each
2100 		 * page holds its count.
2101 		 */
2102 		if (bfregi->count[bfreg_dyn_idx]) {
2103 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2104 			mutex_unlock(&bfregi->lock);
2105 			return -EINVAL;
2106 		}
2107 
2108 		bfregi->count[bfreg_dyn_idx]++;
2109 		mutex_unlock(&bfregi->lock);
2110 
2111 		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2112 		if (err) {
2113 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2114 			goto free_bfreg;
2115 		}
2116 	} else {
2117 		uar_index = bfregi->sys_pages[idx];
2118 	}
2119 
2120 	pfn = uar_index2pfn(dev, uar_index);
2121 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2122 
2123 	vma->vm_page_prot = prot;
2124 	err = io_remap_pfn_range(vma, vma->vm_start, pfn,
2125 				 PAGE_SIZE, vma->vm_page_prot);
2126 	if (err) {
2127 		mlx5_ib_err(dev,
2128 			    "io_remap_pfn_range failed with error=%d, mmap_cmd=%s\n",
2129 			    err, mmap_cmd2str(cmd));
2130 		err = -EAGAIN;
2131 		goto err;
2132 	}
2133 
2134 	pa = pfn << PAGE_SHIFT;
2135 
2136 	err = mlx5_ib_set_vma_data(vma, context);
2137 	if (err)
2138 		goto err;
2139 
2140 	if (dyn_uar)
2141 		bfregi->sys_pages[idx] = uar_index;
2142 	return 0;
2143 
2144 err:
2145 	if (!dyn_uar)
2146 		return err;
2147 
2148 	mlx5_cmd_free_uar(dev->mdev, idx);
2149 
2150 free_bfreg:
2151 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2152 
2153 	return err;
2154 }
2155 
2156 static int dm_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
2157 {
2158 	struct mlx5_ib_ucontext *mctx = to_mucontext(context);
2159 	struct mlx5_ib_dev *dev = to_mdev(context->device);
2160 	u16 page_idx = get_extended_index(vma->vm_pgoff);
2161 	size_t map_size = vma->vm_end - vma->vm_start;
2162 	u32 npages = map_size >> PAGE_SHIFT;
2163 	phys_addr_t pfn;
2164 	pgprot_t prot;
2165 
2166 	if (find_next_zero_bit(mctx->dm_pages, page_idx + npages, page_idx) !=
2167 	    page_idx + npages)
2168 		return -EINVAL;
2169 
2170 	pfn = ((pci_resource_start(dev->mdev->pdev, 0) +
2171 	      MLX5_CAP64_DEV_MEM(dev->mdev, memic_bar_start_addr)) >>
2172 	      PAGE_SHIFT) +
2173 	      page_idx;
2174 	prot = pgprot_writecombine(vma->vm_page_prot);
2175 	vma->vm_page_prot = prot;
2176 
2177 	if (io_remap_pfn_range(vma, vma->vm_start, pfn, map_size,
2178 			       vma->vm_page_prot))
2179 		return -EAGAIN;
2180 
2181 	return mlx5_ib_set_vma_data(vma, mctx);
2182 }
2183 
2184 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2185 {
2186 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2187 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2188 	unsigned long command;
2189 	phys_addr_t pfn;
2190 
2191 	command = get_command(vma->vm_pgoff);
2192 	switch (command) {
2193 	case MLX5_IB_MMAP_WC_PAGE:
2194 	case MLX5_IB_MMAP_NC_PAGE:
2195 	case MLX5_IB_MMAP_REGULAR_PAGE:
2196 	case MLX5_IB_MMAP_ALLOC_WC:
2197 		return uar_mmap(dev, command, vma, context);
2198 
2199 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2200 		return -ENOSYS;
2201 
2202 	case MLX5_IB_MMAP_CORE_CLOCK:
2203 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2204 			return -EINVAL;
2205 
2206 		if (vma->vm_flags & VM_WRITE)
2207 			return -EPERM;
2208 
2209 		/* Don't expose to user-space information it shouldn't have */
2210 		if (PAGE_SIZE > 4096)
2211 			return -EOPNOTSUPP;
2212 
2213 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
2214 		pfn = (dev->mdev->iseg_base +
2215 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2216 			PAGE_SHIFT;
2217 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
2218 				       PAGE_SIZE, vma->vm_page_prot))
2219 			return -EAGAIN;
2220 		break;
2221 	case MLX5_IB_MMAP_CLOCK_INFO:
2222 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2223 
2224 	case MLX5_IB_MMAP_DEVICE_MEM:
2225 		return dm_mmap(ibcontext, vma);
2226 
2227 	default:
2228 		return -EINVAL;
2229 	}
2230 
2231 	return 0;
2232 }
2233 
2234 struct ib_dm *mlx5_ib_alloc_dm(struct ib_device *ibdev,
2235 			       struct ib_ucontext *context,
2236 			       struct ib_dm_alloc_attr *attr,
2237 			       struct uverbs_attr_bundle *attrs)
2238 {
2239 	u64 act_size = roundup(attr->length, MLX5_MEMIC_BASE_SIZE);
2240 	struct mlx5_memic *memic = &to_mdev(ibdev)->memic;
2241 	phys_addr_t memic_addr;
2242 	struct mlx5_ib_dm *dm;
2243 	u64 start_offset;
2244 	u32 page_idx;
2245 	int err;
2246 
2247 	dm = kzalloc(sizeof(*dm), GFP_KERNEL);
2248 	if (!dm)
2249 		return ERR_PTR(-ENOMEM);
2250 
2251 	mlx5_ib_dbg(to_mdev(ibdev), "alloc_memic req: user_length=0x%llx act_length=0x%llx log_alignment=%d\n",
2252 		    attr->length, act_size, attr->alignment);
2253 
2254 	err = mlx5_cmd_alloc_memic(memic, &memic_addr,
2255 				   act_size, attr->alignment);
2256 	if (err)
2257 		goto err_free;
2258 
2259 	start_offset = memic_addr & ~PAGE_MASK;
2260 	page_idx = (memic_addr - pci_resource_start(memic->dev->pdev, 0) -
2261 		    MLX5_CAP64_DEV_MEM(memic->dev, memic_bar_start_addr)) >>
2262 		    PAGE_SHIFT;
2263 
2264 	err = uverbs_copy_to(attrs,
2265 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2266 			     &start_offset, sizeof(start_offset));
2267 	if (err)
2268 		goto err_dealloc;
2269 
2270 	err = uverbs_copy_to(attrs,
2271 			     MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
2272 			     &page_idx, sizeof(page_idx));
2273 	if (err)
2274 		goto err_dealloc;
2275 
2276 	bitmap_set(to_mucontext(context)->dm_pages, page_idx,
2277 		   DIV_ROUND_UP(act_size, PAGE_SIZE));
2278 
2279 	dm->dev_addr = memic_addr;
2280 
2281 	return &dm->ibdm;
2282 
2283 err_dealloc:
2284 	mlx5_cmd_dealloc_memic(memic, memic_addr,
2285 			       act_size);
2286 err_free:
2287 	kfree(dm);
2288 	return ERR_PTR(err);
2289 }
2290 
2291 int mlx5_ib_dealloc_dm(struct ib_dm *ibdm)
2292 {
2293 	struct mlx5_memic *memic = &to_mdev(ibdm->device)->memic;
2294 	struct mlx5_ib_dm *dm = to_mdm(ibdm);
2295 	u64 act_size = roundup(dm->ibdm.length, MLX5_MEMIC_BASE_SIZE);
2296 	u32 page_idx;
2297 	int ret;
2298 
2299 	ret = mlx5_cmd_dealloc_memic(memic, dm->dev_addr, act_size);
2300 	if (ret)
2301 		return ret;
2302 
2303 	page_idx = (dm->dev_addr - pci_resource_start(memic->dev->pdev, 0) -
2304 		    MLX5_CAP64_DEV_MEM(memic->dev, memic_bar_start_addr)) >>
2305 		    PAGE_SHIFT;
2306 	bitmap_clear(to_mucontext(ibdm->uobject->context)->dm_pages,
2307 		     page_idx,
2308 		     DIV_ROUND_UP(act_size, PAGE_SIZE));
2309 
2310 	kfree(dm);
2311 
2312 	return 0;
2313 }
2314 
2315 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
2316 				      struct ib_ucontext *context,
2317 				      struct ib_udata *udata)
2318 {
2319 	struct mlx5_ib_alloc_pd_resp resp;
2320 	struct mlx5_ib_pd *pd;
2321 	int err;
2322 
2323 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
2324 	if (!pd)
2325 		return ERR_PTR(-ENOMEM);
2326 
2327 	err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
2328 	if (err) {
2329 		kfree(pd);
2330 		return ERR_PTR(err);
2331 	}
2332 
2333 	if (context) {
2334 		resp.pdn = pd->pdn;
2335 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2336 			mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
2337 			kfree(pd);
2338 			return ERR_PTR(-EFAULT);
2339 		}
2340 	}
2341 
2342 	return &pd->ibpd;
2343 }
2344 
2345 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
2346 {
2347 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2348 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2349 
2350 	mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
2351 	kfree(mpd);
2352 
2353 	return 0;
2354 }
2355 
2356 enum {
2357 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
2358 	MATCH_CRITERIA_ENABLE_MISC_BIT,
2359 	MATCH_CRITERIA_ENABLE_INNER_BIT,
2360 	MATCH_CRITERIA_ENABLE_MISC2_BIT
2361 };
2362 
2363 #define HEADER_IS_ZERO(match_criteria, headers)			           \
2364 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
2365 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
2366 
2367 static u8 get_match_criteria_enable(u32 *match_criteria)
2368 {
2369 	u8 match_criteria_enable;
2370 
2371 	match_criteria_enable =
2372 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
2373 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
2374 	match_criteria_enable |=
2375 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
2376 		MATCH_CRITERIA_ENABLE_MISC_BIT;
2377 	match_criteria_enable |=
2378 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
2379 		MATCH_CRITERIA_ENABLE_INNER_BIT;
2380 	match_criteria_enable |=
2381 		(!HEADER_IS_ZERO(match_criteria, misc_parameters_2)) <<
2382 		MATCH_CRITERIA_ENABLE_MISC2_BIT;
2383 
2384 	return match_criteria_enable;
2385 }
2386 
2387 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
2388 {
2389 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
2390 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
2391 }
2392 
2393 static void set_flow_label(void *misc_c, void *misc_v, u32 mask, u32 val,
2394 			   bool inner)
2395 {
2396 	if (inner) {
2397 		MLX5_SET(fte_match_set_misc,
2398 			 misc_c, inner_ipv6_flow_label, mask);
2399 		MLX5_SET(fte_match_set_misc,
2400 			 misc_v, inner_ipv6_flow_label, val);
2401 	} else {
2402 		MLX5_SET(fte_match_set_misc,
2403 			 misc_c, outer_ipv6_flow_label, mask);
2404 		MLX5_SET(fte_match_set_misc,
2405 			 misc_v, outer_ipv6_flow_label, val);
2406 	}
2407 }
2408 
2409 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
2410 {
2411 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
2412 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
2413 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
2414 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
2415 }
2416 
2417 static int check_mpls_supp_fields(u32 field_support, const __be32 *set_mask)
2418 {
2419 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_label) &&
2420 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_LABEL))
2421 		return -EOPNOTSUPP;
2422 
2423 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_exp) &&
2424 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_EXP))
2425 		return -EOPNOTSUPP;
2426 
2427 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_s_bos) &&
2428 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_S_BOS))
2429 		return -EOPNOTSUPP;
2430 
2431 	if (MLX5_GET(fte_match_mpls, set_mask, mpls_ttl) &&
2432 	    !(field_support & MLX5_FIELD_SUPPORT_MPLS_TTL))
2433 		return -EOPNOTSUPP;
2434 
2435 	return 0;
2436 }
2437 
2438 #define LAST_ETH_FIELD vlan_tag
2439 #define LAST_IB_FIELD sl
2440 #define LAST_IPV4_FIELD tos
2441 #define LAST_IPV6_FIELD traffic_class
2442 #define LAST_TCP_UDP_FIELD src_port
2443 #define LAST_TUNNEL_FIELD tunnel_id
2444 #define LAST_FLOW_TAG_FIELD tag_id
2445 #define LAST_DROP_FIELD size
2446 #define LAST_COUNTERS_FIELD counters
2447 
2448 /* Field is the last supported field */
2449 #define FIELDS_NOT_SUPPORTED(filter, field)\
2450 	memchr_inv((void *)&filter.field  +\
2451 		   sizeof(filter.field), 0,\
2452 		   sizeof(filter) -\
2453 		   offsetof(typeof(filter), field) -\
2454 		   sizeof(filter.field))
2455 
2456 static int parse_flow_flow_action(const union ib_flow_spec *ib_spec,
2457 				  const struct ib_flow_attr *flow_attr,
2458 				  struct mlx5_flow_act *action)
2459 {
2460 	struct mlx5_ib_flow_action *maction = to_mflow_act(ib_spec->action.act);
2461 
2462 	switch (maction->ib_action.type) {
2463 	case IB_FLOW_ACTION_ESP:
2464 		/* Currently only AES_GCM keymat is supported by the driver */
2465 		action->esp_id = (uintptr_t)maction->esp_aes_gcm.ctx;
2466 		action->action |= flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS ?
2467 			MLX5_FLOW_CONTEXT_ACTION_ENCRYPT :
2468 			MLX5_FLOW_CONTEXT_ACTION_DECRYPT;
2469 		return 0;
2470 	default:
2471 		return -EOPNOTSUPP;
2472 	}
2473 }
2474 
2475 static int parse_flow_attr(struct mlx5_core_dev *mdev, u32 *match_c,
2476 			   u32 *match_v, const union ib_flow_spec *ib_spec,
2477 			   const struct ib_flow_attr *flow_attr,
2478 			   struct mlx5_flow_act *action, u32 prev_type)
2479 {
2480 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
2481 					   misc_parameters);
2482 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
2483 					   misc_parameters);
2484 	void *misc_params2_c = MLX5_ADDR_OF(fte_match_param, match_c,
2485 					    misc_parameters_2);
2486 	void *misc_params2_v = MLX5_ADDR_OF(fte_match_param, match_v,
2487 					    misc_parameters_2);
2488 	void *headers_c;
2489 	void *headers_v;
2490 	int match_ipv;
2491 	int ret;
2492 
2493 	if (ib_spec->type & IB_FLOW_SPEC_INNER) {
2494 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2495 					 inner_headers);
2496 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2497 					 inner_headers);
2498 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2499 					ft_field_support.inner_ip_version);
2500 	} else {
2501 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2502 					 outer_headers);
2503 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2504 					 outer_headers);
2505 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2506 					ft_field_support.outer_ip_version);
2507 	}
2508 
2509 	switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
2510 	case IB_FLOW_SPEC_ETH:
2511 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
2512 			return -EOPNOTSUPP;
2513 
2514 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2515 					     dmac_47_16),
2516 				ib_spec->eth.mask.dst_mac);
2517 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2518 					     dmac_47_16),
2519 				ib_spec->eth.val.dst_mac);
2520 
2521 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2522 					     smac_47_16),
2523 				ib_spec->eth.mask.src_mac);
2524 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2525 					     smac_47_16),
2526 				ib_spec->eth.val.src_mac);
2527 
2528 		if (ib_spec->eth.mask.vlan_tag) {
2529 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2530 				 cvlan_tag, 1);
2531 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2532 				 cvlan_tag, 1);
2533 
2534 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2535 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
2536 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2537 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
2538 
2539 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2540 				 first_cfi,
2541 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
2542 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2543 				 first_cfi,
2544 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
2545 
2546 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2547 				 first_prio,
2548 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
2549 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2550 				 first_prio,
2551 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
2552 		}
2553 		MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2554 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
2555 		MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2556 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
2557 		break;
2558 	case IB_FLOW_SPEC_IPV4:
2559 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
2560 			return -EOPNOTSUPP;
2561 
2562 		if (match_ipv) {
2563 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2564 				 ip_version, 0xf);
2565 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2566 				 ip_version, MLX5_FS_IPV4_VERSION);
2567 		} else {
2568 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2569 				 ethertype, 0xffff);
2570 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2571 				 ethertype, ETH_P_IP);
2572 		}
2573 
2574 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2575 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
2576 		       &ib_spec->ipv4.mask.src_ip,
2577 		       sizeof(ib_spec->ipv4.mask.src_ip));
2578 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2579 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
2580 		       &ib_spec->ipv4.val.src_ip,
2581 		       sizeof(ib_spec->ipv4.val.src_ip));
2582 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2583 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2584 		       &ib_spec->ipv4.mask.dst_ip,
2585 		       sizeof(ib_spec->ipv4.mask.dst_ip));
2586 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2587 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2588 		       &ib_spec->ipv4.val.dst_ip,
2589 		       sizeof(ib_spec->ipv4.val.dst_ip));
2590 
2591 		set_tos(headers_c, headers_v,
2592 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
2593 
2594 		set_proto(headers_c, headers_v,
2595 			  ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
2596 		break;
2597 	case IB_FLOW_SPEC_IPV6:
2598 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
2599 			return -EOPNOTSUPP;
2600 
2601 		if (match_ipv) {
2602 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2603 				 ip_version, 0xf);
2604 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2605 				 ip_version, MLX5_FS_IPV6_VERSION);
2606 		} else {
2607 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2608 				 ethertype, 0xffff);
2609 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2610 				 ethertype, ETH_P_IPV6);
2611 		}
2612 
2613 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2614 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
2615 		       &ib_spec->ipv6.mask.src_ip,
2616 		       sizeof(ib_spec->ipv6.mask.src_ip));
2617 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2618 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
2619 		       &ib_spec->ipv6.val.src_ip,
2620 		       sizeof(ib_spec->ipv6.val.src_ip));
2621 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2622 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2623 		       &ib_spec->ipv6.mask.dst_ip,
2624 		       sizeof(ib_spec->ipv6.mask.dst_ip));
2625 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2626 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2627 		       &ib_spec->ipv6.val.dst_ip,
2628 		       sizeof(ib_spec->ipv6.val.dst_ip));
2629 
2630 		set_tos(headers_c, headers_v,
2631 			ib_spec->ipv6.mask.traffic_class,
2632 			ib_spec->ipv6.val.traffic_class);
2633 
2634 		set_proto(headers_c, headers_v,
2635 			  ib_spec->ipv6.mask.next_hdr,
2636 			  ib_spec->ipv6.val.next_hdr);
2637 
2638 		set_flow_label(misc_params_c, misc_params_v,
2639 			       ntohl(ib_spec->ipv6.mask.flow_label),
2640 			       ntohl(ib_spec->ipv6.val.flow_label),
2641 			       ib_spec->type & IB_FLOW_SPEC_INNER);
2642 		break;
2643 	case IB_FLOW_SPEC_ESP:
2644 		if (ib_spec->esp.mask.seq)
2645 			return -EOPNOTSUPP;
2646 
2647 		MLX5_SET(fte_match_set_misc, misc_params_c, outer_esp_spi,
2648 			 ntohl(ib_spec->esp.mask.spi));
2649 		MLX5_SET(fte_match_set_misc, misc_params_v, outer_esp_spi,
2650 			 ntohl(ib_spec->esp.val.spi));
2651 		break;
2652 	case IB_FLOW_SPEC_TCP:
2653 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2654 					 LAST_TCP_UDP_FIELD))
2655 			return -EOPNOTSUPP;
2656 
2657 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2658 			 0xff);
2659 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2660 			 IPPROTO_TCP);
2661 
2662 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_sport,
2663 			 ntohs(ib_spec->tcp_udp.mask.src_port));
2664 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_sport,
2665 			 ntohs(ib_spec->tcp_udp.val.src_port));
2666 
2667 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_dport,
2668 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
2669 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_dport,
2670 			 ntohs(ib_spec->tcp_udp.val.dst_port));
2671 		break;
2672 	case IB_FLOW_SPEC_UDP:
2673 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2674 					 LAST_TCP_UDP_FIELD))
2675 			return -EOPNOTSUPP;
2676 
2677 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2678 			 0xff);
2679 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2680 			 IPPROTO_UDP);
2681 
2682 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_sport,
2683 			 ntohs(ib_spec->tcp_udp.mask.src_port));
2684 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_sport,
2685 			 ntohs(ib_spec->tcp_udp.val.src_port));
2686 
2687 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_dport,
2688 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
2689 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_dport,
2690 			 ntohs(ib_spec->tcp_udp.val.dst_port));
2691 		break;
2692 	case IB_FLOW_SPEC_GRE:
2693 		if (ib_spec->gre.mask.c_ks_res0_ver)
2694 			return -EOPNOTSUPP;
2695 
2696 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2697 			 0xff);
2698 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2699 			 IPPROTO_GRE);
2700 
2701 		MLX5_SET(fte_match_set_misc, misc_params_c, gre_protocol,
2702 			 0xffff);
2703 		MLX5_SET(fte_match_set_misc, misc_params_v, gre_protocol,
2704 			 ntohs(ib_spec->gre.val.protocol));
2705 
2706 		memcpy(MLX5_ADDR_OF(fte_match_set_misc, misc_params_c,
2707 				    gre_key_h),
2708 		       &ib_spec->gre.mask.key,
2709 		       sizeof(ib_spec->gre.mask.key));
2710 		memcpy(MLX5_ADDR_OF(fte_match_set_misc, misc_params_v,
2711 				    gre_key_h),
2712 		       &ib_spec->gre.val.key,
2713 		       sizeof(ib_spec->gre.val.key));
2714 		break;
2715 	case IB_FLOW_SPEC_MPLS:
2716 		switch (prev_type) {
2717 		case IB_FLOW_SPEC_UDP:
2718 			if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2719 						   ft_field_support.outer_first_mpls_over_udp),
2720 						   &ib_spec->mpls.mask.tag))
2721 				return -EOPNOTSUPP;
2722 
2723 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
2724 					    outer_first_mpls_over_udp),
2725 			       &ib_spec->mpls.val.tag,
2726 			       sizeof(ib_spec->mpls.val.tag));
2727 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
2728 					    outer_first_mpls_over_udp),
2729 			       &ib_spec->mpls.mask.tag,
2730 			       sizeof(ib_spec->mpls.mask.tag));
2731 			break;
2732 		case IB_FLOW_SPEC_GRE:
2733 			if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2734 						   ft_field_support.outer_first_mpls_over_gre),
2735 						   &ib_spec->mpls.mask.tag))
2736 				return -EOPNOTSUPP;
2737 
2738 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
2739 					    outer_first_mpls_over_gre),
2740 			       &ib_spec->mpls.val.tag,
2741 			       sizeof(ib_spec->mpls.val.tag));
2742 			memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
2743 					    outer_first_mpls_over_gre),
2744 			       &ib_spec->mpls.mask.tag,
2745 			       sizeof(ib_spec->mpls.mask.tag));
2746 			break;
2747 		default:
2748 			if (ib_spec->type & IB_FLOW_SPEC_INNER) {
2749 				if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2750 							   ft_field_support.inner_first_mpls),
2751 							   &ib_spec->mpls.mask.tag))
2752 					return -EOPNOTSUPP;
2753 
2754 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
2755 						    inner_first_mpls),
2756 				       &ib_spec->mpls.val.tag,
2757 				       sizeof(ib_spec->mpls.val.tag));
2758 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
2759 						    inner_first_mpls),
2760 				       &ib_spec->mpls.mask.tag,
2761 				       sizeof(ib_spec->mpls.mask.tag));
2762 			} else {
2763 				if (check_mpls_supp_fields(MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2764 							   ft_field_support.outer_first_mpls),
2765 							   &ib_spec->mpls.mask.tag))
2766 					return -EOPNOTSUPP;
2767 
2768 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_v,
2769 						    outer_first_mpls),
2770 				       &ib_spec->mpls.val.tag,
2771 				       sizeof(ib_spec->mpls.val.tag));
2772 				memcpy(MLX5_ADDR_OF(fte_match_set_misc2, misc_params2_c,
2773 						    outer_first_mpls),
2774 				       &ib_spec->mpls.mask.tag,
2775 				       sizeof(ib_spec->mpls.mask.tag));
2776 			}
2777 		}
2778 		break;
2779 	case IB_FLOW_SPEC_VXLAN_TUNNEL:
2780 		if (FIELDS_NOT_SUPPORTED(ib_spec->tunnel.mask,
2781 					 LAST_TUNNEL_FIELD))
2782 			return -EOPNOTSUPP;
2783 
2784 		MLX5_SET(fte_match_set_misc, misc_params_c, vxlan_vni,
2785 			 ntohl(ib_spec->tunnel.mask.tunnel_id));
2786 		MLX5_SET(fte_match_set_misc, misc_params_v, vxlan_vni,
2787 			 ntohl(ib_spec->tunnel.val.tunnel_id));
2788 		break;
2789 	case IB_FLOW_SPEC_ACTION_TAG:
2790 		if (FIELDS_NOT_SUPPORTED(ib_spec->flow_tag,
2791 					 LAST_FLOW_TAG_FIELD))
2792 			return -EOPNOTSUPP;
2793 		if (ib_spec->flow_tag.tag_id >= BIT(24))
2794 			return -EINVAL;
2795 
2796 		action->flow_tag = ib_spec->flow_tag.tag_id;
2797 		action->has_flow_tag = true;
2798 		break;
2799 	case IB_FLOW_SPEC_ACTION_DROP:
2800 		if (FIELDS_NOT_SUPPORTED(ib_spec->drop,
2801 					 LAST_DROP_FIELD))
2802 			return -EOPNOTSUPP;
2803 		action->action |= MLX5_FLOW_CONTEXT_ACTION_DROP;
2804 		break;
2805 	case IB_FLOW_SPEC_ACTION_HANDLE:
2806 		ret = parse_flow_flow_action(ib_spec, flow_attr, action);
2807 		if (ret)
2808 			return ret;
2809 		break;
2810 	case IB_FLOW_SPEC_ACTION_COUNT:
2811 		if (FIELDS_NOT_SUPPORTED(ib_spec->flow_count,
2812 					 LAST_COUNTERS_FIELD))
2813 			return -EOPNOTSUPP;
2814 
2815 		/* for now support only one counters spec per flow */
2816 		if (action->action & MLX5_FLOW_CONTEXT_ACTION_COUNT)
2817 			return -EINVAL;
2818 
2819 		action->counters = ib_spec->flow_count.counters;
2820 		action->action |= MLX5_FLOW_CONTEXT_ACTION_COUNT;
2821 		break;
2822 	default:
2823 		return -EINVAL;
2824 	}
2825 
2826 	return 0;
2827 }
2828 
2829 /* If a flow could catch both multicast and unicast packets,
2830  * it won't fall into the multicast flow steering table and this rule
2831  * could steal other multicast packets.
2832  */
2833 static bool flow_is_multicast_only(const struct ib_flow_attr *ib_attr)
2834 {
2835 	union ib_flow_spec *flow_spec;
2836 
2837 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
2838 	    ib_attr->num_of_specs < 1)
2839 		return false;
2840 
2841 	flow_spec = (union ib_flow_spec *)(ib_attr + 1);
2842 	if (flow_spec->type == IB_FLOW_SPEC_IPV4) {
2843 		struct ib_flow_spec_ipv4 *ipv4_spec;
2844 
2845 		ipv4_spec = (struct ib_flow_spec_ipv4 *)flow_spec;
2846 		if (ipv4_is_multicast(ipv4_spec->val.dst_ip))
2847 			return true;
2848 
2849 		return false;
2850 	}
2851 
2852 	if (flow_spec->type == IB_FLOW_SPEC_ETH) {
2853 		struct ib_flow_spec_eth *eth_spec;
2854 
2855 		eth_spec = (struct ib_flow_spec_eth *)flow_spec;
2856 		return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
2857 		       is_multicast_ether_addr(eth_spec->val.dst_mac);
2858 	}
2859 
2860 	return false;
2861 }
2862 
2863 enum valid_spec {
2864 	VALID_SPEC_INVALID,
2865 	VALID_SPEC_VALID,
2866 	VALID_SPEC_NA,
2867 };
2868 
2869 static enum valid_spec
2870 is_valid_esp_aes_gcm(struct mlx5_core_dev *mdev,
2871 		     const struct mlx5_flow_spec *spec,
2872 		     const struct mlx5_flow_act *flow_act,
2873 		     bool egress)
2874 {
2875 	const u32 *match_c = spec->match_criteria;
2876 	bool is_crypto =
2877 		(flow_act->action & (MLX5_FLOW_CONTEXT_ACTION_ENCRYPT |
2878 				     MLX5_FLOW_CONTEXT_ACTION_DECRYPT));
2879 	bool is_ipsec = mlx5_fs_is_ipsec_flow(match_c);
2880 	bool is_drop = flow_act->action & MLX5_FLOW_CONTEXT_ACTION_DROP;
2881 
2882 	/*
2883 	 * Currently only crypto is supported in egress, when regular egress
2884 	 * rules would be supported, always return VALID_SPEC_NA.
2885 	 */
2886 	if (!is_crypto)
2887 		return egress ? VALID_SPEC_INVALID : VALID_SPEC_NA;
2888 
2889 	return is_crypto && is_ipsec &&
2890 		(!egress || (!is_drop && !flow_act->has_flow_tag)) ?
2891 		VALID_SPEC_VALID : VALID_SPEC_INVALID;
2892 }
2893 
2894 static bool is_valid_spec(struct mlx5_core_dev *mdev,
2895 			  const struct mlx5_flow_spec *spec,
2896 			  const struct mlx5_flow_act *flow_act,
2897 			  bool egress)
2898 {
2899 	/* We curretly only support ipsec egress flow */
2900 	return is_valid_esp_aes_gcm(mdev, spec, flow_act, egress) != VALID_SPEC_INVALID;
2901 }
2902 
2903 static bool is_valid_ethertype(struct mlx5_core_dev *mdev,
2904 			       const struct ib_flow_attr *flow_attr,
2905 			       bool check_inner)
2906 {
2907 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
2908 	int match_ipv = check_inner ?
2909 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2910 					ft_field_support.inner_ip_version) :
2911 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2912 					ft_field_support.outer_ip_version);
2913 	int inner_bit = check_inner ? IB_FLOW_SPEC_INNER : 0;
2914 	bool ipv4_spec_valid, ipv6_spec_valid;
2915 	unsigned int ip_spec_type = 0;
2916 	bool has_ethertype = false;
2917 	unsigned int spec_index;
2918 	bool mask_valid = true;
2919 	u16 eth_type = 0;
2920 	bool type_valid;
2921 
2922 	/* Validate that ethertype is correct */
2923 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2924 		if ((ib_spec->type == (IB_FLOW_SPEC_ETH | inner_bit)) &&
2925 		    ib_spec->eth.mask.ether_type) {
2926 			mask_valid = (ib_spec->eth.mask.ether_type ==
2927 				      htons(0xffff));
2928 			has_ethertype = true;
2929 			eth_type = ntohs(ib_spec->eth.val.ether_type);
2930 		} else if ((ib_spec->type == (IB_FLOW_SPEC_IPV4 | inner_bit)) ||
2931 			   (ib_spec->type == (IB_FLOW_SPEC_IPV6 | inner_bit))) {
2932 			ip_spec_type = ib_spec->type;
2933 		}
2934 		ib_spec = (void *)ib_spec + ib_spec->size;
2935 	}
2936 
2937 	type_valid = (!has_ethertype) || (!ip_spec_type);
2938 	if (!type_valid && mask_valid) {
2939 		ipv4_spec_valid = (eth_type == ETH_P_IP) &&
2940 			(ip_spec_type == (IB_FLOW_SPEC_IPV4 | inner_bit));
2941 		ipv6_spec_valid = (eth_type == ETH_P_IPV6) &&
2942 			(ip_spec_type == (IB_FLOW_SPEC_IPV6 | inner_bit));
2943 
2944 		type_valid = (ipv4_spec_valid) || (ipv6_spec_valid) ||
2945 			     (((eth_type == ETH_P_MPLS_UC) ||
2946 			       (eth_type == ETH_P_MPLS_MC)) && match_ipv);
2947 	}
2948 
2949 	return type_valid;
2950 }
2951 
2952 static bool is_valid_attr(struct mlx5_core_dev *mdev,
2953 			  const struct ib_flow_attr *flow_attr)
2954 {
2955 	return is_valid_ethertype(mdev, flow_attr, false) &&
2956 	       is_valid_ethertype(mdev, flow_attr, true);
2957 }
2958 
2959 static void put_flow_table(struct mlx5_ib_dev *dev,
2960 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
2961 {
2962 	prio->refcount -= !!ft_added;
2963 	if (!prio->refcount) {
2964 		mlx5_destroy_flow_table(prio->flow_table);
2965 		prio->flow_table = NULL;
2966 	}
2967 }
2968 
2969 static void counters_clear_description(struct ib_counters *counters)
2970 {
2971 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
2972 
2973 	mutex_lock(&mcounters->mcntrs_mutex);
2974 	kfree(mcounters->counters_data);
2975 	mcounters->counters_data = NULL;
2976 	mcounters->cntrs_max_index = 0;
2977 	mutex_unlock(&mcounters->mcntrs_mutex);
2978 }
2979 
2980 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
2981 {
2982 	struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
2983 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
2984 							  struct mlx5_ib_flow_handler,
2985 							  ibflow);
2986 	struct mlx5_ib_flow_handler *iter, *tmp;
2987 
2988 	mutex_lock(&dev->flow_db->lock);
2989 
2990 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
2991 		mlx5_del_flow_rules(iter->rule);
2992 		put_flow_table(dev, iter->prio, true);
2993 		list_del(&iter->list);
2994 		kfree(iter);
2995 	}
2996 
2997 	mlx5_del_flow_rules(handler->rule);
2998 	put_flow_table(dev, handler->prio, true);
2999 	if (handler->ibcounters &&
3000 	    atomic_read(&handler->ibcounters->usecnt) == 1)
3001 		counters_clear_description(handler->ibcounters);
3002 
3003 	mutex_unlock(&dev->flow_db->lock);
3004 	kfree(handler);
3005 
3006 	return 0;
3007 }
3008 
3009 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
3010 {
3011 	priority *= 2;
3012 	if (!dont_trap)
3013 		priority++;
3014 	return priority;
3015 }
3016 
3017 enum flow_table_type {
3018 	MLX5_IB_FT_RX,
3019 	MLX5_IB_FT_TX
3020 };
3021 
3022 #define MLX5_FS_MAX_TYPES	 6
3023 #define MLX5_FS_MAX_ENTRIES	 BIT(16)
3024 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
3025 						struct ib_flow_attr *flow_attr,
3026 						enum flow_table_type ft_type)
3027 {
3028 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
3029 	struct mlx5_flow_namespace *ns = NULL;
3030 	struct mlx5_ib_flow_prio *prio;
3031 	struct mlx5_flow_table *ft;
3032 	int max_table_size;
3033 	int num_entries;
3034 	int num_groups;
3035 	int priority;
3036 	int err = 0;
3037 
3038 	max_table_size = BIT(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3039 						       log_max_ft_size));
3040 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
3041 		if (ft_type == MLX5_IB_FT_TX)
3042 			priority = 0;
3043 		else if (flow_is_multicast_only(flow_attr) &&
3044 			 !dont_trap)
3045 			priority = MLX5_IB_FLOW_MCAST_PRIO;
3046 		else
3047 			priority = ib_prio_to_core_prio(flow_attr->priority,
3048 							dont_trap);
3049 		ns = mlx5_get_flow_namespace(dev->mdev,
3050 					     ft_type == MLX5_IB_FT_TX ?
3051 					     MLX5_FLOW_NAMESPACE_EGRESS :
3052 					     MLX5_FLOW_NAMESPACE_BYPASS);
3053 		num_entries = MLX5_FS_MAX_ENTRIES;
3054 		num_groups = MLX5_FS_MAX_TYPES;
3055 		prio = &dev->flow_db->prios[priority];
3056 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3057 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
3058 		ns = mlx5_get_flow_namespace(dev->mdev,
3059 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
3060 		build_leftovers_ft_param(&priority,
3061 					 &num_entries,
3062 					 &num_groups);
3063 		prio = &dev->flow_db->prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
3064 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3065 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
3066 					allow_sniffer_and_nic_rx_shared_tir))
3067 			return ERR_PTR(-ENOTSUPP);
3068 
3069 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
3070 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
3071 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
3072 
3073 		prio = &dev->flow_db->sniffer[ft_type];
3074 		priority = 0;
3075 		num_entries = 1;
3076 		num_groups = 1;
3077 	}
3078 
3079 	if (!ns)
3080 		return ERR_PTR(-ENOTSUPP);
3081 
3082 	if (num_entries > max_table_size)
3083 		return ERR_PTR(-ENOMEM);
3084 
3085 	ft = prio->flow_table;
3086 	if (!ft) {
3087 		ft = mlx5_create_auto_grouped_flow_table(ns, priority,
3088 							 num_entries,
3089 							 num_groups,
3090 							 0, 0);
3091 
3092 		if (!IS_ERR(ft)) {
3093 			prio->refcount = 0;
3094 			prio->flow_table = ft;
3095 		} else {
3096 			err = PTR_ERR(ft);
3097 		}
3098 	}
3099 
3100 	return err ? ERR_PTR(err) : prio;
3101 }
3102 
3103 static void set_underlay_qp(struct mlx5_ib_dev *dev,
3104 			    struct mlx5_flow_spec *spec,
3105 			    u32 underlay_qpn)
3106 {
3107 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param,
3108 					   spec->match_criteria,
3109 					   misc_parameters);
3110 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3111 					   misc_parameters);
3112 
3113 	if (underlay_qpn &&
3114 	    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3115 				      ft_field_support.bth_dst_qp)) {
3116 		MLX5_SET(fte_match_set_misc,
3117 			 misc_params_v, bth_dst_qp, underlay_qpn);
3118 		MLX5_SET(fte_match_set_misc,
3119 			 misc_params_c, bth_dst_qp, 0xffffff);
3120 	}
3121 }
3122 
3123 static int read_flow_counters(struct ib_device *ibdev,
3124 			      struct mlx5_read_counters_attr *read_attr)
3125 {
3126 	struct mlx5_fc *fc = read_attr->hw_cntrs_hndl;
3127 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3128 
3129 	return mlx5_fc_query(dev->mdev, fc,
3130 			     &read_attr->out[IB_COUNTER_PACKETS],
3131 			     &read_attr->out[IB_COUNTER_BYTES]);
3132 }
3133 
3134 /* flow counters currently expose two counters packets and bytes */
3135 #define FLOW_COUNTERS_NUM 2
3136 static int counters_set_description(struct ib_counters *counters,
3137 				    enum mlx5_ib_counters_type counters_type,
3138 				    struct mlx5_ib_flow_counters_desc *desc_data,
3139 				    u32 ncounters)
3140 {
3141 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
3142 	u32 cntrs_max_index = 0;
3143 	int i;
3144 
3145 	if (counters_type != MLX5_IB_COUNTERS_FLOW)
3146 		return -EINVAL;
3147 
3148 	/* init the fields for the object */
3149 	mcounters->type = counters_type;
3150 	mcounters->read_counters = read_flow_counters;
3151 	mcounters->counters_num = FLOW_COUNTERS_NUM;
3152 	mcounters->ncounters = ncounters;
3153 	/* each counter entry have both description and index pair */
3154 	for (i = 0; i < ncounters; i++) {
3155 		if (desc_data[i].description > IB_COUNTER_BYTES)
3156 			return -EINVAL;
3157 
3158 		if (cntrs_max_index <= desc_data[i].index)
3159 			cntrs_max_index = desc_data[i].index + 1;
3160 	}
3161 
3162 	mutex_lock(&mcounters->mcntrs_mutex);
3163 	mcounters->counters_data = desc_data;
3164 	mcounters->cntrs_max_index = cntrs_max_index;
3165 	mutex_unlock(&mcounters->mcntrs_mutex);
3166 
3167 	return 0;
3168 }
3169 
3170 #define MAX_COUNTERS_NUM (USHRT_MAX / (sizeof(u32) * 2))
3171 static int flow_counters_set_data(struct ib_counters *ibcounters,
3172 				  struct mlx5_ib_create_flow *ucmd)
3173 {
3174 	struct mlx5_ib_mcounters *mcounters = to_mcounters(ibcounters);
3175 	struct mlx5_ib_flow_counters_data *cntrs_data = NULL;
3176 	struct mlx5_ib_flow_counters_desc *desc_data = NULL;
3177 	bool hw_hndl = false;
3178 	int ret = 0;
3179 
3180 	if (ucmd && ucmd->ncounters_data != 0) {
3181 		cntrs_data = ucmd->data;
3182 		if (cntrs_data->ncounters > MAX_COUNTERS_NUM)
3183 			return -EINVAL;
3184 
3185 		desc_data = kcalloc(cntrs_data->ncounters,
3186 				    sizeof(*desc_data),
3187 				    GFP_KERNEL);
3188 		if (!desc_data)
3189 			return  -ENOMEM;
3190 
3191 		if (copy_from_user(desc_data,
3192 				   u64_to_user_ptr(cntrs_data->counters_data),
3193 				   sizeof(*desc_data) * cntrs_data->ncounters)) {
3194 			ret = -EFAULT;
3195 			goto free;
3196 		}
3197 	}
3198 
3199 	if (!mcounters->hw_cntrs_hndl) {
3200 		mcounters->hw_cntrs_hndl = mlx5_fc_create(
3201 			to_mdev(ibcounters->device)->mdev, false);
3202 		if (!mcounters->hw_cntrs_hndl) {
3203 			ret = -ENOMEM;
3204 			goto free;
3205 		}
3206 		hw_hndl = true;
3207 	}
3208 
3209 	if (desc_data) {
3210 		/* counters already bound to at least one flow */
3211 		if (mcounters->cntrs_max_index) {
3212 			ret = -EINVAL;
3213 			goto free_hndl;
3214 		}
3215 
3216 		ret = counters_set_description(ibcounters,
3217 					       MLX5_IB_COUNTERS_FLOW,
3218 					       desc_data,
3219 					       cntrs_data->ncounters);
3220 		if (ret)
3221 			goto free_hndl;
3222 
3223 	} else if (!mcounters->cntrs_max_index) {
3224 		/* counters not bound yet, must have udata passed */
3225 		ret = -EINVAL;
3226 		goto free_hndl;
3227 	}
3228 
3229 	return 0;
3230 
3231 free_hndl:
3232 	if (hw_hndl) {
3233 		mlx5_fc_destroy(to_mdev(ibcounters->device)->mdev,
3234 				mcounters->hw_cntrs_hndl);
3235 		mcounters->hw_cntrs_hndl = NULL;
3236 	}
3237 free:
3238 	kfree(desc_data);
3239 	return ret;
3240 }
3241 
3242 static struct mlx5_ib_flow_handler *_create_flow_rule(struct mlx5_ib_dev *dev,
3243 						      struct mlx5_ib_flow_prio *ft_prio,
3244 						      const struct ib_flow_attr *flow_attr,
3245 						      struct mlx5_flow_destination *dst,
3246 						      u32 underlay_qpn,
3247 						      struct mlx5_ib_create_flow *ucmd)
3248 {
3249 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
3250 	struct mlx5_ib_flow_handler *handler;
3251 	struct mlx5_flow_act flow_act = {.flow_tag = MLX5_FS_DEFAULT_FLOW_TAG};
3252 	struct mlx5_flow_spec *spec;
3253 	struct mlx5_flow_destination dest_arr[2] = {};
3254 	struct mlx5_flow_destination *rule_dst = dest_arr;
3255 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
3256 	unsigned int spec_index;
3257 	u32 prev_type = 0;
3258 	int err = 0;
3259 	int dest_num = 0;
3260 	bool is_egress = flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS;
3261 
3262 	if (!is_valid_attr(dev->mdev, flow_attr))
3263 		return ERR_PTR(-EINVAL);
3264 
3265 	spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
3266 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
3267 	if (!handler || !spec) {
3268 		err = -ENOMEM;
3269 		goto free;
3270 	}
3271 
3272 	INIT_LIST_HEAD(&handler->list);
3273 	if (dst) {
3274 		memcpy(&dest_arr[0], dst, sizeof(*dst));
3275 		dest_num++;
3276 	}
3277 
3278 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
3279 		err = parse_flow_attr(dev->mdev, spec->match_criteria,
3280 				      spec->match_value,
3281 				      ib_flow, flow_attr, &flow_act,
3282 				      prev_type);
3283 		if (err < 0)
3284 			goto free;
3285 
3286 		prev_type = ((union ib_flow_spec *)ib_flow)->type;
3287 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
3288 	}
3289 
3290 	if (!flow_is_multicast_only(flow_attr))
3291 		set_underlay_qp(dev, spec, underlay_qpn);
3292 
3293 	if (dev->rep) {
3294 		void *misc;
3295 
3296 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3297 				    misc_parameters);
3298 		MLX5_SET(fte_match_set_misc, misc, source_port,
3299 			 dev->rep->vport);
3300 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_criteria,
3301 				    misc_parameters);
3302 		MLX5_SET_TO_ONES(fte_match_set_misc, misc, source_port);
3303 	}
3304 
3305 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
3306 
3307 	if (is_egress &&
3308 	    !is_valid_spec(dev->mdev, spec, &flow_act, is_egress)) {
3309 		err = -EINVAL;
3310 		goto free;
3311 	}
3312 
3313 	if (flow_act.action & MLX5_FLOW_CONTEXT_ACTION_COUNT) {
3314 		err = flow_counters_set_data(flow_act.counters, ucmd);
3315 		if (err)
3316 			goto free;
3317 
3318 		handler->ibcounters = flow_act.counters;
3319 		dest_arr[dest_num].type =
3320 			MLX5_FLOW_DESTINATION_TYPE_COUNTER;
3321 		dest_arr[dest_num].counter =
3322 			to_mcounters(flow_act.counters)->hw_cntrs_hndl;
3323 		dest_num++;
3324 	}
3325 
3326 	if (flow_act.action & MLX5_FLOW_CONTEXT_ACTION_DROP) {
3327 		if (!(flow_act.action & MLX5_FLOW_CONTEXT_ACTION_COUNT)) {
3328 			rule_dst = NULL;
3329 			dest_num = 0;
3330 		}
3331 	} else {
3332 		if (is_egress)
3333 			flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_ALLOW;
3334 		else
3335 			flow_act.action |=
3336 				dest_num ?  MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
3337 					MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
3338 	}
3339 
3340 	if (flow_act.has_flow_tag &&
3341 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3342 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
3343 		mlx5_ib_warn(dev, "Flow tag %u and attribute type %x isn't allowed in leftovers\n",
3344 			     flow_act.flow_tag, flow_attr->type);
3345 		err = -EINVAL;
3346 		goto free;
3347 	}
3348 	handler->rule = mlx5_add_flow_rules(ft, spec,
3349 					    &flow_act,
3350 					    rule_dst, dest_num);
3351 
3352 	if (IS_ERR(handler->rule)) {
3353 		err = PTR_ERR(handler->rule);
3354 		goto free;
3355 	}
3356 
3357 	ft_prio->refcount++;
3358 	handler->prio = ft_prio;
3359 
3360 	ft_prio->flow_table = ft;
3361 free:
3362 	if (err && handler) {
3363 		if (handler->ibcounters &&
3364 		    atomic_read(&handler->ibcounters->usecnt) == 1)
3365 			counters_clear_description(handler->ibcounters);
3366 		kfree(handler);
3367 	}
3368 	kvfree(spec);
3369 	return err ? ERR_PTR(err) : handler;
3370 }
3371 
3372 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
3373 						     struct mlx5_ib_flow_prio *ft_prio,
3374 						     const struct ib_flow_attr *flow_attr,
3375 						     struct mlx5_flow_destination *dst)
3376 {
3377 	return _create_flow_rule(dev, ft_prio, flow_attr, dst, 0, NULL);
3378 }
3379 
3380 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
3381 							  struct mlx5_ib_flow_prio *ft_prio,
3382 							  struct ib_flow_attr *flow_attr,
3383 							  struct mlx5_flow_destination *dst)
3384 {
3385 	struct mlx5_ib_flow_handler *handler_dst = NULL;
3386 	struct mlx5_ib_flow_handler *handler = NULL;
3387 
3388 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
3389 	if (!IS_ERR(handler)) {
3390 		handler_dst = create_flow_rule(dev, ft_prio,
3391 					       flow_attr, dst);
3392 		if (IS_ERR(handler_dst)) {
3393 			mlx5_del_flow_rules(handler->rule);
3394 			ft_prio->refcount--;
3395 			kfree(handler);
3396 			handler = handler_dst;
3397 		} else {
3398 			list_add(&handler_dst->list, &handler->list);
3399 		}
3400 	}
3401 
3402 	return handler;
3403 }
3404 enum {
3405 	LEFTOVERS_MC,
3406 	LEFTOVERS_UC,
3407 };
3408 
3409 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
3410 							  struct mlx5_ib_flow_prio *ft_prio,
3411 							  struct ib_flow_attr *flow_attr,
3412 							  struct mlx5_flow_destination *dst)
3413 {
3414 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
3415 	struct mlx5_ib_flow_handler *handler = NULL;
3416 
3417 	static struct {
3418 		struct ib_flow_attr	flow_attr;
3419 		struct ib_flow_spec_eth eth_flow;
3420 	} leftovers_specs[] = {
3421 		[LEFTOVERS_MC] = {
3422 			.flow_attr = {
3423 				.num_of_specs = 1,
3424 				.size = sizeof(leftovers_specs[0])
3425 			},
3426 			.eth_flow = {
3427 				.type = IB_FLOW_SPEC_ETH,
3428 				.size = sizeof(struct ib_flow_spec_eth),
3429 				.mask = {.dst_mac = {0x1} },
3430 				.val =  {.dst_mac = {0x1} }
3431 			}
3432 		},
3433 		[LEFTOVERS_UC] = {
3434 			.flow_attr = {
3435 				.num_of_specs = 1,
3436 				.size = sizeof(leftovers_specs[0])
3437 			},
3438 			.eth_flow = {
3439 				.type = IB_FLOW_SPEC_ETH,
3440 				.size = sizeof(struct ib_flow_spec_eth),
3441 				.mask = {.dst_mac = {0x1} },
3442 				.val = {.dst_mac = {} }
3443 			}
3444 		}
3445 	};
3446 
3447 	handler = create_flow_rule(dev, ft_prio,
3448 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
3449 				   dst);
3450 	if (!IS_ERR(handler) &&
3451 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
3452 		handler_ucast = create_flow_rule(dev, ft_prio,
3453 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
3454 						 dst);
3455 		if (IS_ERR(handler_ucast)) {
3456 			mlx5_del_flow_rules(handler->rule);
3457 			ft_prio->refcount--;
3458 			kfree(handler);
3459 			handler = handler_ucast;
3460 		} else {
3461 			list_add(&handler_ucast->list, &handler->list);
3462 		}
3463 	}
3464 
3465 	return handler;
3466 }
3467 
3468 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
3469 							struct mlx5_ib_flow_prio *ft_rx,
3470 							struct mlx5_ib_flow_prio *ft_tx,
3471 							struct mlx5_flow_destination *dst)
3472 {
3473 	struct mlx5_ib_flow_handler *handler_rx;
3474 	struct mlx5_ib_flow_handler *handler_tx;
3475 	int err;
3476 	static const struct ib_flow_attr flow_attr  = {
3477 		.num_of_specs = 0,
3478 		.size = sizeof(flow_attr)
3479 	};
3480 
3481 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
3482 	if (IS_ERR(handler_rx)) {
3483 		err = PTR_ERR(handler_rx);
3484 		goto err;
3485 	}
3486 
3487 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
3488 	if (IS_ERR(handler_tx)) {
3489 		err = PTR_ERR(handler_tx);
3490 		goto err_tx;
3491 	}
3492 
3493 	list_add(&handler_tx->list, &handler_rx->list);
3494 
3495 	return handler_rx;
3496 
3497 err_tx:
3498 	mlx5_del_flow_rules(handler_rx->rule);
3499 	ft_rx->refcount--;
3500 	kfree(handler_rx);
3501 err:
3502 	return ERR_PTR(err);
3503 }
3504 
3505 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
3506 					   struct ib_flow_attr *flow_attr,
3507 					   int domain,
3508 					   struct ib_udata *udata)
3509 {
3510 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3511 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3512 	struct mlx5_ib_flow_handler *handler = NULL;
3513 	struct mlx5_flow_destination *dst = NULL;
3514 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
3515 	struct mlx5_ib_flow_prio *ft_prio;
3516 	bool is_egress = flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS;
3517 	struct mlx5_ib_create_flow *ucmd = NULL, ucmd_hdr;
3518 	size_t min_ucmd_sz, required_ucmd_sz;
3519 	int err;
3520 	int underlay_qpn;
3521 
3522 	if (udata && udata->inlen) {
3523 		min_ucmd_sz = offsetof(typeof(ucmd_hdr), reserved) +
3524 				sizeof(ucmd_hdr.reserved);
3525 		if (udata->inlen < min_ucmd_sz)
3526 			return ERR_PTR(-EOPNOTSUPP);
3527 
3528 		err = ib_copy_from_udata(&ucmd_hdr, udata, min_ucmd_sz);
3529 		if (err)
3530 			return ERR_PTR(err);
3531 
3532 		/* currently supports only one counters data */
3533 		if (ucmd_hdr.ncounters_data > 1)
3534 			return ERR_PTR(-EINVAL);
3535 
3536 		required_ucmd_sz = min_ucmd_sz +
3537 			sizeof(struct mlx5_ib_flow_counters_data) *
3538 			ucmd_hdr.ncounters_data;
3539 		if (udata->inlen > required_ucmd_sz &&
3540 		    !ib_is_udata_cleared(udata, required_ucmd_sz,
3541 					 udata->inlen - required_ucmd_sz))
3542 			return ERR_PTR(-EOPNOTSUPP);
3543 
3544 		ucmd = kzalloc(required_ucmd_sz, GFP_KERNEL);
3545 		if (!ucmd)
3546 			return ERR_PTR(-ENOMEM);
3547 
3548 		err = ib_copy_from_udata(ucmd, udata, required_ucmd_sz);
3549 		if (err) {
3550 			kfree(ucmd);
3551 			return ERR_PTR(err);
3552 		}
3553 	}
3554 
3555 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
3556 		return ERR_PTR(-ENOMEM);
3557 
3558 	if (domain != IB_FLOW_DOMAIN_USER ||
3559 	    flow_attr->port > dev->num_ports ||
3560 	    (flow_attr->flags & ~(IB_FLOW_ATTR_FLAGS_DONT_TRAP |
3561 				  IB_FLOW_ATTR_FLAGS_EGRESS)))
3562 		return ERR_PTR(-EINVAL);
3563 
3564 	if (is_egress &&
3565 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3566 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT))
3567 		return ERR_PTR(-EINVAL);
3568 
3569 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
3570 	if (!dst)
3571 		return ERR_PTR(-ENOMEM);
3572 
3573 	mutex_lock(&dev->flow_db->lock);
3574 
3575 	ft_prio = get_flow_table(dev, flow_attr,
3576 				 is_egress ? MLX5_IB_FT_TX : MLX5_IB_FT_RX);
3577 	if (IS_ERR(ft_prio)) {
3578 		err = PTR_ERR(ft_prio);
3579 		goto unlock;
3580 	}
3581 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3582 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
3583 		if (IS_ERR(ft_prio_tx)) {
3584 			err = PTR_ERR(ft_prio_tx);
3585 			ft_prio_tx = NULL;
3586 			goto destroy_ft;
3587 		}
3588 	}
3589 
3590 	if (is_egress) {
3591 		dst->type = MLX5_FLOW_DESTINATION_TYPE_PORT;
3592 	} else {
3593 		dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
3594 		if (mqp->flags & MLX5_IB_QP_RSS)
3595 			dst->tir_num = mqp->rss_qp.tirn;
3596 		else
3597 			dst->tir_num = mqp->raw_packet_qp.rq.tirn;
3598 	}
3599 
3600 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
3601 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
3602 			handler = create_dont_trap_rule(dev, ft_prio,
3603 							flow_attr, dst);
3604 		} else {
3605 			underlay_qpn = (mqp->flags & MLX5_IB_QP_UNDERLAY) ?
3606 					mqp->underlay_qpn : 0;
3607 			handler = _create_flow_rule(dev, ft_prio, flow_attr,
3608 						    dst, underlay_qpn, ucmd);
3609 		}
3610 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3611 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
3612 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
3613 						dst);
3614 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3615 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
3616 	} else {
3617 		err = -EINVAL;
3618 		goto destroy_ft;
3619 	}
3620 
3621 	if (IS_ERR(handler)) {
3622 		err = PTR_ERR(handler);
3623 		handler = NULL;
3624 		goto destroy_ft;
3625 	}
3626 
3627 	mutex_unlock(&dev->flow_db->lock);
3628 	kfree(dst);
3629 	kfree(ucmd);
3630 
3631 	return &handler->ibflow;
3632 
3633 destroy_ft:
3634 	put_flow_table(dev, ft_prio, false);
3635 	if (ft_prio_tx)
3636 		put_flow_table(dev, ft_prio_tx, false);
3637 unlock:
3638 	mutex_unlock(&dev->flow_db->lock);
3639 	kfree(dst);
3640 	kfree(ucmd);
3641 	kfree(handler);
3642 	return ERR_PTR(err);
3643 }
3644 
3645 static u32 mlx5_ib_flow_action_flags_to_accel_xfrm_flags(u32 mlx5_flags)
3646 {
3647 	u32 flags = 0;
3648 
3649 	if (mlx5_flags & MLX5_IB_UAPI_FLOW_ACTION_FLAGS_REQUIRE_METADATA)
3650 		flags |= MLX5_ACCEL_XFRM_FLAG_REQUIRE_METADATA;
3651 
3652 	return flags;
3653 }
3654 
3655 #define MLX5_FLOW_ACTION_ESP_CREATE_LAST_SUPPORTED	MLX5_IB_UAPI_FLOW_ACTION_FLAGS_REQUIRE_METADATA
3656 static struct ib_flow_action *
3657 mlx5_ib_create_flow_action_esp(struct ib_device *device,
3658 			       const struct ib_flow_action_attrs_esp *attr,
3659 			       struct uverbs_attr_bundle *attrs)
3660 {
3661 	struct mlx5_ib_dev *mdev = to_mdev(device);
3662 	struct ib_uverbs_flow_action_esp_keymat_aes_gcm *aes_gcm;
3663 	struct mlx5_accel_esp_xfrm_attrs accel_attrs = {};
3664 	struct mlx5_ib_flow_action *action;
3665 	u64 action_flags;
3666 	u64 flags;
3667 	int err = 0;
3668 
3669 	if (IS_UVERBS_COPY_ERR(uverbs_copy_from(&action_flags, attrs,
3670 						MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS)))
3671 		return ERR_PTR(-EFAULT);
3672 
3673 	if (action_flags >= (MLX5_FLOW_ACTION_ESP_CREATE_LAST_SUPPORTED << 1))
3674 		return ERR_PTR(-EOPNOTSUPP);
3675 
3676 	flags = mlx5_ib_flow_action_flags_to_accel_xfrm_flags(action_flags);
3677 
3678 	/* We current only support a subset of the standard features. Only a
3679 	 * keymat of type AES_GCM, with icv_len == 16, iv_algo == SEQ and esn
3680 	 * (with overlap). Full offload mode isn't supported.
3681 	 */
3682 	if (!attr->keymat || attr->replay || attr->encap ||
3683 	    attr->spi || attr->seq || attr->tfc_pad ||
3684 	    attr->hard_limit_pkts ||
3685 	    (attr->flags & ~(IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
3686 			     IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ENCRYPT)))
3687 		return ERR_PTR(-EOPNOTSUPP);
3688 
3689 	if (attr->keymat->protocol !=
3690 	    IB_UVERBS_FLOW_ACTION_ESP_KEYMAT_AES_GCM)
3691 		return ERR_PTR(-EOPNOTSUPP);
3692 
3693 	aes_gcm = &attr->keymat->keymat.aes_gcm;
3694 
3695 	if (aes_gcm->icv_len != 16 ||
3696 	    aes_gcm->iv_algo != IB_UVERBS_FLOW_ACTION_IV_ALGO_SEQ)
3697 		return ERR_PTR(-EOPNOTSUPP);
3698 
3699 	action = kmalloc(sizeof(*action), GFP_KERNEL);
3700 	if (!action)
3701 		return ERR_PTR(-ENOMEM);
3702 
3703 	action->esp_aes_gcm.ib_flags = attr->flags;
3704 	memcpy(&accel_attrs.keymat.aes_gcm.aes_key, &aes_gcm->aes_key,
3705 	       sizeof(accel_attrs.keymat.aes_gcm.aes_key));
3706 	accel_attrs.keymat.aes_gcm.key_len = aes_gcm->key_len * 8;
3707 	memcpy(&accel_attrs.keymat.aes_gcm.salt, &aes_gcm->salt,
3708 	       sizeof(accel_attrs.keymat.aes_gcm.salt));
3709 	memcpy(&accel_attrs.keymat.aes_gcm.seq_iv, &aes_gcm->iv,
3710 	       sizeof(accel_attrs.keymat.aes_gcm.seq_iv));
3711 	accel_attrs.keymat.aes_gcm.icv_len = aes_gcm->icv_len * 8;
3712 	accel_attrs.keymat.aes_gcm.iv_algo = MLX5_ACCEL_ESP_AES_GCM_IV_ALGO_SEQ;
3713 	accel_attrs.keymat_type = MLX5_ACCEL_ESP_KEYMAT_AES_GCM;
3714 
3715 	accel_attrs.esn = attr->esn;
3716 	if (attr->flags & IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED)
3717 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_TRIGGERED;
3718 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)
3719 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
3720 
3721 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ENCRYPT)
3722 		accel_attrs.action |= MLX5_ACCEL_ESP_ACTION_ENCRYPT;
3723 
3724 	action->esp_aes_gcm.ctx =
3725 		mlx5_accel_esp_create_xfrm(mdev->mdev, &accel_attrs, flags);
3726 	if (IS_ERR(action->esp_aes_gcm.ctx)) {
3727 		err = PTR_ERR(action->esp_aes_gcm.ctx);
3728 		goto err_parse;
3729 	}
3730 
3731 	action->esp_aes_gcm.ib_flags = attr->flags;
3732 
3733 	return &action->ib_action;
3734 
3735 err_parse:
3736 	kfree(action);
3737 	return ERR_PTR(err);
3738 }
3739 
3740 static int
3741 mlx5_ib_modify_flow_action_esp(struct ib_flow_action *action,
3742 			       const struct ib_flow_action_attrs_esp *attr,
3743 			       struct uverbs_attr_bundle *attrs)
3744 {
3745 	struct mlx5_ib_flow_action *maction = to_mflow_act(action);
3746 	struct mlx5_accel_esp_xfrm_attrs accel_attrs;
3747 	int err = 0;
3748 
3749 	if (attr->keymat || attr->replay || attr->encap ||
3750 	    attr->spi || attr->seq || attr->tfc_pad ||
3751 	    attr->hard_limit_pkts ||
3752 	    (attr->flags & ~(IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
3753 			     IB_FLOW_ACTION_ESP_FLAGS_MOD_ESP_ATTRS |
3754 			     IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)))
3755 		return -EOPNOTSUPP;
3756 
3757 	/* Only the ESN value or the MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP can
3758 	 * be modified.
3759 	 */
3760 	if (!(maction->esp_aes_gcm.ib_flags &
3761 	      IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED) &&
3762 	    attr->flags & (IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
3763 			   IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW))
3764 		return -EINVAL;
3765 
3766 	memcpy(&accel_attrs, &maction->esp_aes_gcm.ctx->attrs,
3767 	       sizeof(accel_attrs));
3768 
3769 	accel_attrs.esn = attr->esn;
3770 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)
3771 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
3772 	else
3773 		accel_attrs.flags &= ~MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
3774 
3775 	err = mlx5_accel_esp_modify_xfrm(maction->esp_aes_gcm.ctx,
3776 					 &accel_attrs);
3777 	if (err)
3778 		return err;
3779 
3780 	maction->esp_aes_gcm.ib_flags &=
3781 		~IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW;
3782 	maction->esp_aes_gcm.ib_flags |=
3783 		attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW;
3784 
3785 	return 0;
3786 }
3787 
3788 static int mlx5_ib_destroy_flow_action(struct ib_flow_action *action)
3789 {
3790 	struct mlx5_ib_flow_action *maction = to_mflow_act(action);
3791 
3792 	switch (action->type) {
3793 	case IB_FLOW_ACTION_ESP:
3794 		/*
3795 		 * We only support aes_gcm by now, so we implicitly know this is
3796 		 * the underline crypto.
3797 		 */
3798 		mlx5_accel_esp_destroy_xfrm(maction->esp_aes_gcm.ctx);
3799 		break;
3800 	default:
3801 		WARN_ON(true);
3802 		break;
3803 	}
3804 
3805 	kfree(maction);
3806 	return 0;
3807 }
3808 
3809 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
3810 {
3811 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3812 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
3813 	int err;
3814 
3815 	if (mqp->flags & MLX5_IB_QP_UNDERLAY) {
3816 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
3817 		return -EOPNOTSUPP;
3818 	}
3819 
3820 	err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
3821 	if (err)
3822 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
3823 			     ibqp->qp_num, gid->raw);
3824 
3825 	return err;
3826 }
3827 
3828 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
3829 {
3830 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3831 	int err;
3832 
3833 	err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
3834 	if (err)
3835 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
3836 			     ibqp->qp_num, gid->raw);
3837 
3838 	return err;
3839 }
3840 
3841 static int init_node_data(struct mlx5_ib_dev *dev)
3842 {
3843 	int err;
3844 
3845 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
3846 	if (err)
3847 		return err;
3848 
3849 	dev->mdev->rev_id = dev->mdev->pdev->revision;
3850 
3851 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
3852 }
3853 
3854 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
3855 			     char *buf)
3856 {
3857 	struct mlx5_ib_dev *dev =
3858 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3859 
3860 	return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
3861 }
3862 
3863 static ssize_t show_reg_pages(struct device *device,
3864 			      struct device_attribute *attr, char *buf)
3865 {
3866 	struct mlx5_ib_dev *dev =
3867 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3868 
3869 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
3870 }
3871 
3872 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
3873 			char *buf)
3874 {
3875 	struct mlx5_ib_dev *dev =
3876 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3877 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
3878 }
3879 
3880 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
3881 			char *buf)
3882 {
3883 	struct mlx5_ib_dev *dev =
3884 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3885 	return sprintf(buf, "%x\n", dev->mdev->rev_id);
3886 }
3887 
3888 static ssize_t show_board(struct device *device, struct device_attribute *attr,
3889 			  char *buf)
3890 {
3891 	struct mlx5_ib_dev *dev =
3892 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3893 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
3894 		       dev->mdev->board_id);
3895 }
3896 
3897 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
3898 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
3899 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
3900 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
3901 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
3902 
3903 static struct device_attribute *mlx5_class_attributes[] = {
3904 	&dev_attr_hw_rev,
3905 	&dev_attr_hca_type,
3906 	&dev_attr_board_id,
3907 	&dev_attr_fw_pages,
3908 	&dev_attr_reg_pages,
3909 };
3910 
3911 static void pkey_change_handler(struct work_struct *work)
3912 {
3913 	struct mlx5_ib_port_resources *ports =
3914 		container_of(work, struct mlx5_ib_port_resources,
3915 			     pkey_change_work);
3916 
3917 	mutex_lock(&ports->devr->mutex);
3918 	mlx5_ib_gsi_pkey_change(ports->gsi);
3919 	mutex_unlock(&ports->devr->mutex);
3920 }
3921 
3922 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
3923 {
3924 	struct mlx5_ib_qp *mqp;
3925 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
3926 	struct mlx5_core_cq *mcq;
3927 	struct list_head cq_armed_list;
3928 	unsigned long flags_qp;
3929 	unsigned long flags_cq;
3930 	unsigned long flags;
3931 
3932 	INIT_LIST_HEAD(&cq_armed_list);
3933 
3934 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
3935 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
3936 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
3937 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
3938 		if (mqp->sq.tail != mqp->sq.head) {
3939 			send_mcq = to_mcq(mqp->ibqp.send_cq);
3940 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
3941 			if (send_mcq->mcq.comp &&
3942 			    mqp->ibqp.send_cq->comp_handler) {
3943 				if (!send_mcq->mcq.reset_notify_added) {
3944 					send_mcq->mcq.reset_notify_added = 1;
3945 					list_add_tail(&send_mcq->mcq.reset_notify,
3946 						      &cq_armed_list);
3947 				}
3948 			}
3949 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
3950 		}
3951 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
3952 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
3953 		/* no handling is needed for SRQ */
3954 		if (!mqp->ibqp.srq) {
3955 			if (mqp->rq.tail != mqp->rq.head) {
3956 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
3957 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
3958 				if (recv_mcq->mcq.comp &&
3959 				    mqp->ibqp.recv_cq->comp_handler) {
3960 					if (!recv_mcq->mcq.reset_notify_added) {
3961 						recv_mcq->mcq.reset_notify_added = 1;
3962 						list_add_tail(&recv_mcq->mcq.reset_notify,
3963 							      &cq_armed_list);
3964 					}
3965 				}
3966 				spin_unlock_irqrestore(&recv_mcq->lock,
3967 						       flags_cq);
3968 			}
3969 		}
3970 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
3971 	}
3972 	/*At that point all inflight post send were put to be executed as of we
3973 	 * lock/unlock above locks Now need to arm all involved CQs.
3974 	 */
3975 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
3976 		mcq->comp(mcq);
3977 	}
3978 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
3979 }
3980 
3981 static void delay_drop_handler(struct work_struct *work)
3982 {
3983 	int err;
3984 	struct mlx5_ib_delay_drop *delay_drop =
3985 		container_of(work, struct mlx5_ib_delay_drop,
3986 			     delay_drop_work);
3987 
3988 	atomic_inc(&delay_drop->events_cnt);
3989 
3990 	mutex_lock(&delay_drop->lock);
3991 	err = mlx5_core_set_delay_drop(delay_drop->dev->mdev,
3992 				       delay_drop->timeout);
3993 	if (err) {
3994 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
3995 			     delay_drop->timeout);
3996 		delay_drop->activate = false;
3997 	}
3998 	mutex_unlock(&delay_drop->lock);
3999 }
4000 
4001 static void mlx5_ib_handle_event(struct work_struct *_work)
4002 {
4003 	struct mlx5_ib_event_work *work =
4004 		container_of(_work, struct mlx5_ib_event_work, work);
4005 	struct mlx5_ib_dev *ibdev;
4006 	struct ib_event ibev;
4007 	bool fatal = false;
4008 	u8 port = (u8)work->param;
4009 
4010 	if (mlx5_core_is_mp_slave(work->dev)) {
4011 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->context);
4012 		if (!ibdev)
4013 			goto out;
4014 	} else {
4015 		ibdev = work->context;
4016 	}
4017 
4018 	switch (work->event) {
4019 	case MLX5_DEV_EVENT_SYS_ERROR:
4020 		ibev.event = IB_EVENT_DEVICE_FATAL;
4021 		mlx5_ib_handle_internal_error(ibdev);
4022 		fatal = true;
4023 		break;
4024 
4025 	case MLX5_DEV_EVENT_PORT_UP:
4026 	case MLX5_DEV_EVENT_PORT_DOWN:
4027 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
4028 		/* In RoCE, port up/down events are handled in
4029 		 * mlx5_netdev_event().
4030 		 */
4031 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
4032 			IB_LINK_LAYER_ETHERNET)
4033 			goto out;
4034 
4035 		ibev.event = (work->event == MLX5_DEV_EVENT_PORT_UP) ?
4036 			     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
4037 		break;
4038 
4039 	case MLX5_DEV_EVENT_LID_CHANGE:
4040 		ibev.event = IB_EVENT_LID_CHANGE;
4041 		break;
4042 
4043 	case MLX5_DEV_EVENT_PKEY_CHANGE:
4044 		ibev.event = IB_EVENT_PKEY_CHANGE;
4045 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
4046 		break;
4047 
4048 	case MLX5_DEV_EVENT_GUID_CHANGE:
4049 		ibev.event = IB_EVENT_GID_CHANGE;
4050 		break;
4051 
4052 	case MLX5_DEV_EVENT_CLIENT_REREG:
4053 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
4054 		break;
4055 	case MLX5_DEV_EVENT_DELAY_DROP_TIMEOUT:
4056 		schedule_work(&ibdev->delay_drop.delay_drop_work);
4057 		goto out;
4058 	default:
4059 		goto out;
4060 	}
4061 
4062 	ibev.device	      = &ibdev->ib_dev;
4063 	ibev.element.port_num = port;
4064 
4065 	if (!rdma_is_port_valid(&ibdev->ib_dev, port)) {
4066 		mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
4067 		goto out;
4068 	}
4069 
4070 	if (ibdev->ib_active)
4071 		ib_dispatch_event(&ibev);
4072 
4073 	if (fatal)
4074 		ibdev->ib_active = false;
4075 out:
4076 	kfree(work);
4077 }
4078 
4079 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
4080 			  enum mlx5_dev_event event, unsigned long param)
4081 {
4082 	struct mlx5_ib_event_work *work;
4083 
4084 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
4085 	if (!work)
4086 		return;
4087 
4088 	INIT_WORK(&work->work, mlx5_ib_handle_event);
4089 	work->dev = dev;
4090 	work->param = param;
4091 	work->context = context;
4092 	work->event = event;
4093 
4094 	queue_work(mlx5_ib_event_wq, &work->work);
4095 }
4096 
4097 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
4098 {
4099 	struct mlx5_hca_vport_context vport_ctx;
4100 	int err;
4101 	int port;
4102 
4103 	for (port = 1; port <= dev->num_ports; port++) {
4104 		dev->mdev->port_caps[port - 1].has_smi = false;
4105 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
4106 		    MLX5_CAP_PORT_TYPE_IB) {
4107 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
4108 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
4109 								   port, 0,
4110 								   &vport_ctx);
4111 				if (err) {
4112 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
4113 						    port, err);
4114 					return err;
4115 				}
4116 				dev->mdev->port_caps[port - 1].has_smi =
4117 					vport_ctx.has_smi;
4118 			} else {
4119 				dev->mdev->port_caps[port - 1].has_smi = true;
4120 			}
4121 		}
4122 	}
4123 	return 0;
4124 }
4125 
4126 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
4127 {
4128 	int port;
4129 
4130 	for (port = 1; port <= dev->num_ports; port++)
4131 		mlx5_query_ext_port_caps(dev, port);
4132 }
4133 
4134 static int get_port_caps(struct mlx5_ib_dev *dev, u8 port)
4135 {
4136 	struct ib_device_attr *dprops = NULL;
4137 	struct ib_port_attr *pprops = NULL;
4138 	int err = -ENOMEM;
4139 	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
4140 
4141 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
4142 	if (!pprops)
4143 		goto out;
4144 
4145 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
4146 	if (!dprops)
4147 		goto out;
4148 
4149 	err = set_has_smi_cap(dev);
4150 	if (err)
4151 		goto out;
4152 
4153 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
4154 	if (err) {
4155 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
4156 		goto out;
4157 	}
4158 
4159 	memset(pprops, 0, sizeof(*pprops));
4160 	err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
4161 	if (err) {
4162 		mlx5_ib_warn(dev, "query_port %d failed %d\n",
4163 			     port, err);
4164 		goto out;
4165 	}
4166 
4167 	dev->mdev->port_caps[port - 1].pkey_table_len =
4168 					dprops->max_pkeys;
4169 	dev->mdev->port_caps[port - 1].gid_table_len =
4170 					pprops->gid_tbl_len;
4171 	mlx5_ib_dbg(dev, "port %d: pkey_table_len %d, gid_table_len %d\n",
4172 		    port, dprops->max_pkeys, pprops->gid_tbl_len);
4173 
4174 out:
4175 	kfree(pprops);
4176 	kfree(dprops);
4177 
4178 	return err;
4179 }
4180 
4181 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
4182 {
4183 	int err;
4184 
4185 	err = mlx5_mr_cache_cleanup(dev);
4186 	if (err)
4187 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
4188 
4189 	if (dev->umrc.qp)
4190 		mlx5_ib_destroy_qp(dev->umrc.qp);
4191 	if (dev->umrc.cq)
4192 		ib_free_cq(dev->umrc.cq);
4193 	if (dev->umrc.pd)
4194 		ib_dealloc_pd(dev->umrc.pd);
4195 }
4196 
4197 enum {
4198 	MAX_UMR_WR = 128,
4199 };
4200 
4201 static int create_umr_res(struct mlx5_ib_dev *dev)
4202 {
4203 	struct ib_qp_init_attr *init_attr = NULL;
4204 	struct ib_qp_attr *attr = NULL;
4205 	struct ib_pd *pd;
4206 	struct ib_cq *cq;
4207 	struct ib_qp *qp;
4208 	int ret;
4209 
4210 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
4211 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
4212 	if (!attr || !init_attr) {
4213 		ret = -ENOMEM;
4214 		goto error_0;
4215 	}
4216 
4217 	pd = ib_alloc_pd(&dev->ib_dev, 0);
4218 	if (IS_ERR(pd)) {
4219 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
4220 		ret = PTR_ERR(pd);
4221 		goto error_0;
4222 	}
4223 
4224 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
4225 	if (IS_ERR(cq)) {
4226 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
4227 		ret = PTR_ERR(cq);
4228 		goto error_2;
4229 	}
4230 
4231 	init_attr->send_cq = cq;
4232 	init_attr->recv_cq = cq;
4233 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
4234 	init_attr->cap.max_send_wr = MAX_UMR_WR;
4235 	init_attr->cap.max_send_sge = 1;
4236 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
4237 	init_attr->port_num = 1;
4238 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
4239 	if (IS_ERR(qp)) {
4240 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
4241 		ret = PTR_ERR(qp);
4242 		goto error_3;
4243 	}
4244 	qp->device     = &dev->ib_dev;
4245 	qp->real_qp    = qp;
4246 	qp->uobject    = NULL;
4247 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
4248 	qp->send_cq    = init_attr->send_cq;
4249 	qp->recv_cq    = init_attr->recv_cq;
4250 
4251 	attr->qp_state = IB_QPS_INIT;
4252 	attr->port_num = 1;
4253 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
4254 				IB_QP_PORT, NULL);
4255 	if (ret) {
4256 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
4257 		goto error_4;
4258 	}
4259 
4260 	memset(attr, 0, sizeof(*attr));
4261 	attr->qp_state = IB_QPS_RTR;
4262 	attr->path_mtu = IB_MTU_256;
4263 
4264 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4265 	if (ret) {
4266 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
4267 		goto error_4;
4268 	}
4269 
4270 	memset(attr, 0, sizeof(*attr));
4271 	attr->qp_state = IB_QPS_RTS;
4272 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4273 	if (ret) {
4274 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
4275 		goto error_4;
4276 	}
4277 
4278 	dev->umrc.qp = qp;
4279 	dev->umrc.cq = cq;
4280 	dev->umrc.pd = pd;
4281 
4282 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
4283 	ret = mlx5_mr_cache_init(dev);
4284 	if (ret) {
4285 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4286 		goto error_4;
4287 	}
4288 
4289 	kfree(attr);
4290 	kfree(init_attr);
4291 
4292 	return 0;
4293 
4294 error_4:
4295 	mlx5_ib_destroy_qp(qp);
4296 	dev->umrc.qp = NULL;
4297 
4298 error_3:
4299 	ib_free_cq(cq);
4300 	dev->umrc.cq = NULL;
4301 
4302 error_2:
4303 	ib_dealloc_pd(pd);
4304 	dev->umrc.pd = NULL;
4305 
4306 error_0:
4307 	kfree(attr);
4308 	kfree(init_attr);
4309 	return ret;
4310 }
4311 
4312 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
4313 {
4314 	switch (umr_fence_cap) {
4315 	case MLX5_CAP_UMR_FENCE_NONE:
4316 		return MLX5_FENCE_MODE_NONE;
4317 	case MLX5_CAP_UMR_FENCE_SMALL:
4318 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
4319 	default:
4320 		return MLX5_FENCE_MODE_STRONG_ORDERING;
4321 	}
4322 }
4323 
4324 static int create_dev_resources(struct mlx5_ib_resources *devr)
4325 {
4326 	struct ib_srq_init_attr attr;
4327 	struct mlx5_ib_dev *dev;
4328 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
4329 	int port;
4330 	int ret = 0;
4331 
4332 	dev = container_of(devr, struct mlx5_ib_dev, devr);
4333 
4334 	mutex_init(&devr->mutex);
4335 
4336 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
4337 	if (IS_ERR(devr->p0)) {
4338 		ret = PTR_ERR(devr->p0);
4339 		goto error0;
4340 	}
4341 	devr->p0->device  = &dev->ib_dev;
4342 	devr->p0->uobject = NULL;
4343 	atomic_set(&devr->p0->usecnt, 0);
4344 
4345 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
4346 	if (IS_ERR(devr->c0)) {
4347 		ret = PTR_ERR(devr->c0);
4348 		goto error1;
4349 	}
4350 	devr->c0->device        = &dev->ib_dev;
4351 	devr->c0->uobject       = NULL;
4352 	devr->c0->comp_handler  = NULL;
4353 	devr->c0->event_handler = NULL;
4354 	devr->c0->cq_context    = NULL;
4355 	atomic_set(&devr->c0->usecnt, 0);
4356 
4357 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
4358 	if (IS_ERR(devr->x0)) {
4359 		ret = PTR_ERR(devr->x0);
4360 		goto error2;
4361 	}
4362 	devr->x0->device = &dev->ib_dev;
4363 	devr->x0->inode = NULL;
4364 	atomic_set(&devr->x0->usecnt, 0);
4365 	mutex_init(&devr->x0->tgt_qp_mutex);
4366 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
4367 
4368 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
4369 	if (IS_ERR(devr->x1)) {
4370 		ret = PTR_ERR(devr->x1);
4371 		goto error3;
4372 	}
4373 	devr->x1->device = &dev->ib_dev;
4374 	devr->x1->inode = NULL;
4375 	atomic_set(&devr->x1->usecnt, 0);
4376 	mutex_init(&devr->x1->tgt_qp_mutex);
4377 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
4378 
4379 	memset(&attr, 0, sizeof(attr));
4380 	attr.attr.max_sge = 1;
4381 	attr.attr.max_wr = 1;
4382 	attr.srq_type = IB_SRQT_XRC;
4383 	attr.ext.cq = devr->c0;
4384 	attr.ext.xrc.xrcd = devr->x0;
4385 
4386 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
4387 	if (IS_ERR(devr->s0)) {
4388 		ret = PTR_ERR(devr->s0);
4389 		goto error4;
4390 	}
4391 	devr->s0->device	= &dev->ib_dev;
4392 	devr->s0->pd		= devr->p0;
4393 	devr->s0->uobject       = NULL;
4394 	devr->s0->event_handler = NULL;
4395 	devr->s0->srq_context   = NULL;
4396 	devr->s0->srq_type      = IB_SRQT_XRC;
4397 	devr->s0->ext.xrc.xrcd	= devr->x0;
4398 	devr->s0->ext.cq	= devr->c0;
4399 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
4400 	atomic_inc(&devr->s0->ext.cq->usecnt);
4401 	atomic_inc(&devr->p0->usecnt);
4402 	atomic_set(&devr->s0->usecnt, 0);
4403 
4404 	memset(&attr, 0, sizeof(attr));
4405 	attr.attr.max_sge = 1;
4406 	attr.attr.max_wr = 1;
4407 	attr.srq_type = IB_SRQT_BASIC;
4408 	devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
4409 	if (IS_ERR(devr->s1)) {
4410 		ret = PTR_ERR(devr->s1);
4411 		goto error5;
4412 	}
4413 	devr->s1->device	= &dev->ib_dev;
4414 	devr->s1->pd		= devr->p0;
4415 	devr->s1->uobject       = NULL;
4416 	devr->s1->event_handler = NULL;
4417 	devr->s1->srq_context   = NULL;
4418 	devr->s1->srq_type      = IB_SRQT_BASIC;
4419 	devr->s1->ext.cq	= devr->c0;
4420 	atomic_inc(&devr->p0->usecnt);
4421 	atomic_set(&devr->s1->usecnt, 0);
4422 
4423 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
4424 		INIT_WORK(&devr->ports[port].pkey_change_work,
4425 			  pkey_change_handler);
4426 		devr->ports[port].devr = devr;
4427 	}
4428 
4429 	return 0;
4430 
4431 error5:
4432 	mlx5_ib_destroy_srq(devr->s0);
4433 error4:
4434 	mlx5_ib_dealloc_xrcd(devr->x1);
4435 error3:
4436 	mlx5_ib_dealloc_xrcd(devr->x0);
4437 error2:
4438 	mlx5_ib_destroy_cq(devr->c0);
4439 error1:
4440 	mlx5_ib_dealloc_pd(devr->p0);
4441 error0:
4442 	return ret;
4443 }
4444 
4445 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
4446 {
4447 	struct mlx5_ib_dev *dev =
4448 		container_of(devr, struct mlx5_ib_dev, devr);
4449 	int port;
4450 
4451 	mlx5_ib_destroy_srq(devr->s1);
4452 	mlx5_ib_destroy_srq(devr->s0);
4453 	mlx5_ib_dealloc_xrcd(devr->x0);
4454 	mlx5_ib_dealloc_xrcd(devr->x1);
4455 	mlx5_ib_destroy_cq(devr->c0);
4456 	mlx5_ib_dealloc_pd(devr->p0);
4457 
4458 	/* Make sure no change P_Key work items are still executing */
4459 	for (port = 0; port < dev->num_ports; ++port)
4460 		cancel_work_sync(&devr->ports[port].pkey_change_work);
4461 }
4462 
4463 static u32 get_core_cap_flags(struct ib_device *ibdev)
4464 {
4465 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4466 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
4467 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
4468 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
4469 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
4470 	u32 ret = 0;
4471 
4472 	if (ll == IB_LINK_LAYER_INFINIBAND)
4473 		return RDMA_CORE_PORT_IBA_IB;
4474 
4475 	if (raw_support)
4476 		ret = RDMA_CORE_PORT_RAW_PACKET;
4477 
4478 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
4479 		return ret;
4480 
4481 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
4482 		return ret;
4483 
4484 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
4485 		ret |= RDMA_CORE_PORT_IBA_ROCE;
4486 
4487 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
4488 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
4489 
4490 	return ret;
4491 }
4492 
4493 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
4494 			       struct ib_port_immutable *immutable)
4495 {
4496 	struct ib_port_attr attr;
4497 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4498 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
4499 	int err;
4500 
4501 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
4502 
4503 	err = ib_query_port(ibdev, port_num, &attr);
4504 	if (err)
4505 		return err;
4506 
4507 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
4508 	immutable->gid_tbl_len = attr.gid_tbl_len;
4509 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
4510 	if ((ll == IB_LINK_LAYER_INFINIBAND) || MLX5_CAP_GEN(dev->mdev, roce))
4511 		immutable->max_mad_size = IB_MGMT_MAD_SIZE;
4512 
4513 	return 0;
4514 }
4515 
4516 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u8 port_num,
4517 				   struct ib_port_immutable *immutable)
4518 {
4519 	struct ib_port_attr attr;
4520 	int err;
4521 
4522 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
4523 
4524 	err = ib_query_port(ibdev, port_num, &attr);
4525 	if (err)
4526 		return err;
4527 
4528 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
4529 	immutable->gid_tbl_len = attr.gid_tbl_len;
4530 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
4531 
4532 	return 0;
4533 }
4534 
4535 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
4536 {
4537 	struct mlx5_ib_dev *dev =
4538 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
4539 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
4540 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
4541 		 fw_rev_sub(dev->mdev));
4542 }
4543 
4544 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
4545 {
4546 	struct mlx5_core_dev *mdev = dev->mdev;
4547 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
4548 								 MLX5_FLOW_NAMESPACE_LAG);
4549 	struct mlx5_flow_table *ft;
4550 	int err;
4551 
4552 	if (!ns || !mlx5_lag_is_active(mdev))
4553 		return 0;
4554 
4555 	err = mlx5_cmd_create_vport_lag(mdev);
4556 	if (err)
4557 		return err;
4558 
4559 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
4560 	if (IS_ERR(ft)) {
4561 		err = PTR_ERR(ft);
4562 		goto err_destroy_vport_lag;
4563 	}
4564 
4565 	dev->flow_db->lag_demux_ft = ft;
4566 	return 0;
4567 
4568 err_destroy_vport_lag:
4569 	mlx5_cmd_destroy_vport_lag(mdev);
4570 	return err;
4571 }
4572 
4573 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
4574 {
4575 	struct mlx5_core_dev *mdev = dev->mdev;
4576 
4577 	if (dev->flow_db->lag_demux_ft) {
4578 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
4579 		dev->flow_db->lag_demux_ft = NULL;
4580 
4581 		mlx5_cmd_destroy_vport_lag(mdev);
4582 	}
4583 }
4584 
4585 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
4586 {
4587 	int err;
4588 
4589 	dev->roce[port_num].nb.notifier_call = mlx5_netdev_event;
4590 	err = register_netdevice_notifier(&dev->roce[port_num].nb);
4591 	if (err) {
4592 		dev->roce[port_num].nb.notifier_call = NULL;
4593 		return err;
4594 	}
4595 
4596 	return 0;
4597 }
4598 
4599 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
4600 {
4601 	if (dev->roce[port_num].nb.notifier_call) {
4602 		unregister_netdevice_notifier(&dev->roce[port_num].nb);
4603 		dev->roce[port_num].nb.notifier_call = NULL;
4604 	}
4605 }
4606 
4607 static int mlx5_enable_eth(struct mlx5_ib_dev *dev, u8 port_num)
4608 {
4609 	int err;
4610 
4611 	if (MLX5_CAP_GEN(dev->mdev, roce)) {
4612 		err = mlx5_nic_vport_enable_roce(dev->mdev);
4613 		if (err)
4614 			return err;
4615 	}
4616 
4617 	err = mlx5_eth_lag_init(dev);
4618 	if (err)
4619 		goto err_disable_roce;
4620 
4621 	return 0;
4622 
4623 err_disable_roce:
4624 	if (MLX5_CAP_GEN(dev->mdev, roce))
4625 		mlx5_nic_vport_disable_roce(dev->mdev);
4626 
4627 	return err;
4628 }
4629 
4630 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
4631 {
4632 	mlx5_eth_lag_cleanup(dev);
4633 	if (MLX5_CAP_GEN(dev->mdev, roce))
4634 		mlx5_nic_vport_disable_roce(dev->mdev);
4635 }
4636 
4637 struct mlx5_ib_counter {
4638 	const char *name;
4639 	size_t offset;
4640 };
4641 
4642 #define INIT_Q_COUNTER(_name)		\
4643 	{ .name = #_name, .offset = MLX5_BYTE_OFF(query_q_counter_out, _name)}
4644 
4645 static const struct mlx5_ib_counter basic_q_cnts[] = {
4646 	INIT_Q_COUNTER(rx_write_requests),
4647 	INIT_Q_COUNTER(rx_read_requests),
4648 	INIT_Q_COUNTER(rx_atomic_requests),
4649 	INIT_Q_COUNTER(out_of_buffer),
4650 };
4651 
4652 static const struct mlx5_ib_counter out_of_seq_q_cnts[] = {
4653 	INIT_Q_COUNTER(out_of_sequence),
4654 };
4655 
4656 static const struct mlx5_ib_counter retrans_q_cnts[] = {
4657 	INIT_Q_COUNTER(duplicate_request),
4658 	INIT_Q_COUNTER(rnr_nak_retry_err),
4659 	INIT_Q_COUNTER(packet_seq_err),
4660 	INIT_Q_COUNTER(implied_nak_seq_err),
4661 	INIT_Q_COUNTER(local_ack_timeout_err),
4662 };
4663 
4664 #define INIT_CONG_COUNTER(_name)		\
4665 	{ .name = #_name, .offset =	\
4666 		MLX5_BYTE_OFF(query_cong_statistics_out, _name ## _high)}
4667 
4668 static const struct mlx5_ib_counter cong_cnts[] = {
4669 	INIT_CONG_COUNTER(rp_cnp_ignored),
4670 	INIT_CONG_COUNTER(rp_cnp_handled),
4671 	INIT_CONG_COUNTER(np_ecn_marked_roce_packets),
4672 	INIT_CONG_COUNTER(np_cnp_sent),
4673 };
4674 
4675 static const struct mlx5_ib_counter extended_err_cnts[] = {
4676 	INIT_Q_COUNTER(resp_local_length_error),
4677 	INIT_Q_COUNTER(resp_cqe_error),
4678 	INIT_Q_COUNTER(req_cqe_error),
4679 	INIT_Q_COUNTER(req_remote_invalid_request),
4680 	INIT_Q_COUNTER(req_remote_access_errors),
4681 	INIT_Q_COUNTER(resp_remote_access_errors),
4682 	INIT_Q_COUNTER(resp_cqe_flush_error),
4683 	INIT_Q_COUNTER(req_cqe_flush_error),
4684 };
4685 
4686 static void mlx5_ib_dealloc_counters(struct mlx5_ib_dev *dev)
4687 {
4688 	int i;
4689 
4690 	for (i = 0; i < dev->num_ports; i++) {
4691 		if (dev->port[i].cnts.set_id)
4692 			mlx5_core_dealloc_q_counter(dev->mdev,
4693 						    dev->port[i].cnts.set_id);
4694 		kfree(dev->port[i].cnts.names);
4695 		kfree(dev->port[i].cnts.offsets);
4696 	}
4697 }
4698 
4699 static int __mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev,
4700 				    struct mlx5_ib_counters *cnts)
4701 {
4702 	u32 num_counters;
4703 
4704 	num_counters = ARRAY_SIZE(basic_q_cnts);
4705 
4706 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt))
4707 		num_counters += ARRAY_SIZE(out_of_seq_q_cnts);
4708 
4709 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters))
4710 		num_counters += ARRAY_SIZE(retrans_q_cnts);
4711 
4712 	if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters))
4713 		num_counters += ARRAY_SIZE(extended_err_cnts);
4714 
4715 	cnts->num_q_counters = num_counters;
4716 
4717 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
4718 		cnts->num_cong_counters = ARRAY_SIZE(cong_cnts);
4719 		num_counters += ARRAY_SIZE(cong_cnts);
4720 	}
4721 
4722 	cnts->names = kcalloc(num_counters, sizeof(cnts->names), GFP_KERNEL);
4723 	if (!cnts->names)
4724 		return -ENOMEM;
4725 
4726 	cnts->offsets = kcalloc(num_counters,
4727 				sizeof(cnts->offsets), GFP_KERNEL);
4728 	if (!cnts->offsets)
4729 		goto err_names;
4730 
4731 	return 0;
4732 
4733 err_names:
4734 	kfree(cnts->names);
4735 	cnts->names = NULL;
4736 	return -ENOMEM;
4737 }
4738 
4739 static void mlx5_ib_fill_counters(struct mlx5_ib_dev *dev,
4740 				  const char **names,
4741 				  size_t *offsets)
4742 {
4743 	int i;
4744 	int j = 0;
4745 
4746 	for (i = 0; i < ARRAY_SIZE(basic_q_cnts); i++, j++) {
4747 		names[j] = basic_q_cnts[i].name;
4748 		offsets[j] = basic_q_cnts[i].offset;
4749 	}
4750 
4751 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt)) {
4752 		for (i = 0; i < ARRAY_SIZE(out_of_seq_q_cnts); i++, j++) {
4753 			names[j] = out_of_seq_q_cnts[i].name;
4754 			offsets[j] = out_of_seq_q_cnts[i].offset;
4755 		}
4756 	}
4757 
4758 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
4759 		for (i = 0; i < ARRAY_SIZE(retrans_q_cnts); i++, j++) {
4760 			names[j] = retrans_q_cnts[i].name;
4761 			offsets[j] = retrans_q_cnts[i].offset;
4762 		}
4763 	}
4764 
4765 	if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters)) {
4766 		for (i = 0; i < ARRAY_SIZE(extended_err_cnts); i++, j++) {
4767 			names[j] = extended_err_cnts[i].name;
4768 			offsets[j] = extended_err_cnts[i].offset;
4769 		}
4770 	}
4771 
4772 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
4773 		for (i = 0; i < ARRAY_SIZE(cong_cnts); i++, j++) {
4774 			names[j] = cong_cnts[i].name;
4775 			offsets[j] = cong_cnts[i].offset;
4776 		}
4777 	}
4778 }
4779 
4780 static int mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev)
4781 {
4782 	int err = 0;
4783 	int i;
4784 
4785 	for (i = 0; i < dev->num_ports; i++) {
4786 		err = __mlx5_ib_alloc_counters(dev, &dev->port[i].cnts);
4787 		if (err)
4788 			goto err_alloc;
4789 
4790 		mlx5_ib_fill_counters(dev, dev->port[i].cnts.names,
4791 				      dev->port[i].cnts.offsets);
4792 
4793 		err = mlx5_core_alloc_q_counter(dev->mdev,
4794 						&dev->port[i].cnts.set_id);
4795 		if (err) {
4796 			mlx5_ib_warn(dev,
4797 				     "couldn't allocate queue counter for port %d, err %d\n",
4798 				     i + 1, err);
4799 			goto err_alloc;
4800 		}
4801 		dev->port[i].cnts.set_id_valid = true;
4802 	}
4803 
4804 	return 0;
4805 
4806 err_alloc:
4807 	mlx5_ib_dealloc_counters(dev);
4808 	return err;
4809 }
4810 
4811 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
4812 						    u8 port_num)
4813 {
4814 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4815 	struct mlx5_ib_port *port = &dev->port[port_num - 1];
4816 
4817 	/* We support only per port stats */
4818 	if (port_num == 0)
4819 		return NULL;
4820 
4821 	return rdma_alloc_hw_stats_struct(port->cnts.names,
4822 					  port->cnts.num_q_counters +
4823 					  port->cnts.num_cong_counters,
4824 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
4825 }
4826 
4827 static int mlx5_ib_query_q_counters(struct mlx5_core_dev *mdev,
4828 				    struct mlx5_ib_port *port,
4829 				    struct rdma_hw_stats *stats)
4830 {
4831 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
4832 	void *out;
4833 	__be32 val;
4834 	int ret, i;
4835 
4836 	out = kvzalloc(outlen, GFP_KERNEL);
4837 	if (!out)
4838 		return -ENOMEM;
4839 
4840 	ret = mlx5_core_query_q_counter(mdev,
4841 					port->cnts.set_id, 0,
4842 					out, outlen);
4843 	if (ret)
4844 		goto free;
4845 
4846 	for (i = 0; i < port->cnts.num_q_counters; i++) {
4847 		val = *(__be32 *)(out + port->cnts.offsets[i]);
4848 		stats->value[i] = (u64)be32_to_cpu(val);
4849 	}
4850 
4851 free:
4852 	kvfree(out);
4853 	return ret;
4854 }
4855 
4856 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
4857 				struct rdma_hw_stats *stats,
4858 				u8 port_num, int index)
4859 {
4860 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4861 	struct mlx5_ib_port *port = &dev->port[port_num - 1];
4862 	struct mlx5_core_dev *mdev;
4863 	int ret, num_counters;
4864 	u8 mdev_port_num;
4865 
4866 	if (!stats)
4867 		return -EINVAL;
4868 
4869 	num_counters = port->cnts.num_q_counters + port->cnts.num_cong_counters;
4870 
4871 	/* q_counters are per IB device, query the master mdev */
4872 	ret = mlx5_ib_query_q_counters(dev->mdev, port, stats);
4873 	if (ret)
4874 		return ret;
4875 
4876 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
4877 		mdev = mlx5_ib_get_native_port_mdev(dev, port_num,
4878 						    &mdev_port_num);
4879 		if (!mdev) {
4880 			/* If port is not affiliated yet, its in down state
4881 			 * which doesn't have any counters yet, so it would be
4882 			 * zero. So no need to read from the HCA.
4883 			 */
4884 			goto done;
4885 		}
4886 		ret = mlx5_lag_query_cong_counters(dev->mdev,
4887 						   stats->value +
4888 						   port->cnts.num_q_counters,
4889 						   port->cnts.num_cong_counters,
4890 						   port->cnts.offsets +
4891 						   port->cnts.num_q_counters);
4892 
4893 		mlx5_ib_put_native_port_mdev(dev, port_num);
4894 		if (ret)
4895 			return ret;
4896 	}
4897 
4898 done:
4899 	return num_counters;
4900 }
4901 
4902 static void mlx5_ib_free_rdma_netdev(struct net_device *netdev)
4903 {
4904 	return mlx5_rdma_netdev_free(netdev);
4905 }
4906 
4907 static struct net_device*
4908 mlx5_ib_alloc_rdma_netdev(struct ib_device *hca,
4909 			  u8 port_num,
4910 			  enum rdma_netdev_t type,
4911 			  const char *name,
4912 			  unsigned char name_assign_type,
4913 			  void (*setup)(struct net_device *))
4914 {
4915 	struct net_device *netdev;
4916 	struct rdma_netdev *rn;
4917 
4918 	if (type != RDMA_NETDEV_IPOIB)
4919 		return ERR_PTR(-EOPNOTSUPP);
4920 
4921 	netdev = mlx5_rdma_netdev_alloc(to_mdev(hca)->mdev, hca,
4922 					name, setup);
4923 	if (likely(!IS_ERR_OR_NULL(netdev))) {
4924 		rn = netdev_priv(netdev);
4925 		rn->free_rdma_netdev = mlx5_ib_free_rdma_netdev;
4926 	}
4927 	return netdev;
4928 }
4929 
4930 static void delay_drop_debugfs_cleanup(struct mlx5_ib_dev *dev)
4931 {
4932 	if (!dev->delay_drop.dbg)
4933 		return;
4934 	debugfs_remove_recursive(dev->delay_drop.dbg->dir_debugfs);
4935 	kfree(dev->delay_drop.dbg);
4936 	dev->delay_drop.dbg = NULL;
4937 }
4938 
4939 static void cancel_delay_drop(struct mlx5_ib_dev *dev)
4940 {
4941 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4942 		return;
4943 
4944 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4945 	delay_drop_debugfs_cleanup(dev);
4946 }
4947 
4948 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
4949 				       size_t count, loff_t *pos)
4950 {
4951 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
4952 	char lbuf[20];
4953 	int len;
4954 
4955 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
4956 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
4957 }
4958 
4959 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
4960 					size_t count, loff_t *pos)
4961 {
4962 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
4963 	u32 timeout;
4964 	u32 var;
4965 
4966 	if (kstrtouint_from_user(buf, count, 0, &var))
4967 		return -EFAULT;
4968 
4969 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
4970 			1000);
4971 	if (timeout != var)
4972 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
4973 			    timeout);
4974 
4975 	delay_drop->timeout = timeout;
4976 
4977 	return count;
4978 }
4979 
4980 static const struct file_operations fops_delay_drop_timeout = {
4981 	.owner	= THIS_MODULE,
4982 	.open	= simple_open,
4983 	.write	= delay_drop_timeout_write,
4984 	.read	= delay_drop_timeout_read,
4985 };
4986 
4987 static int delay_drop_debugfs_init(struct mlx5_ib_dev *dev)
4988 {
4989 	struct mlx5_ib_dbg_delay_drop *dbg;
4990 
4991 	if (!mlx5_debugfs_root)
4992 		return 0;
4993 
4994 	dbg = kzalloc(sizeof(*dbg), GFP_KERNEL);
4995 	if (!dbg)
4996 		return -ENOMEM;
4997 
4998 	dev->delay_drop.dbg = dbg;
4999 
5000 	dbg->dir_debugfs =
5001 		debugfs_create_dir("delay_drop",
5002 				   dev->mdev->priv.dbg_root);
5003 	if (!dbg->dir_debugfs)
5004 		goto out_debugfs;
5005 
5006 	dbg->events_cnt_debugfs =
5007 		debugfs_create_atomic_t("num_timeout_events", 0400,
5008 					dbg->dir_debugfs,
5009 					&dev->delay_drop.events_cnt);
5010 	if (!dbg->events_cnt_debugfs)
5011 		goto out_debugfs;
5012 
5013 	dbg->rqs_cnt_debugfs =
5014 		debugfs_create_atomic_t("num_rqs", 0400,
5015 					dbg->dir_debugfs,
5016 					&dev->delay_drop.rqs_cnt);
5017 	if (!dbg->rqs_cnt_debugfs)
5018 		goto out_debugfs;
5019 
5020 	dbg->timeout_debugfs =
5021 		debugfs_create_file("timeout", 0600,
5022 				    dbg->dir_debugfs,
5023 				    &dev->delay_drop,
5024 				    &fops_delay_drop_timeout);
5025 	if (!dbg->timeout_debugfs)
5026 		goto out_debugfs;
5027 
5028 	return 0;
5029 
5030 out_debugfs:
5031 	delay_drop_debugfs_cleanup(dev);
5032 	return -ENOMEM;
5033 }
5034 
5035 static void init_delay_drop(struct mlx5_ib_dev *dev)
5036 {
5037 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
5038 		return;
5039 
5040 	mutex_init(&dev->delay_drop.lock);
5041 	dev->delay_drop.dev = dev;
5042 	dev->delay_drop.activate = false;
5043 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
5044 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
5045 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
5046 	atomic_set(&dev->delay_drop.events_cnt, 0);
5047 
5048 	if (delay_drop_debugfs_init(dev))
5049 		mlx5_ib_warn(dev, "Failed to init delay drop debugfs\n");
5050 }
5051 
5052 static const struct cpumask *
5053 mlx5_ib_get_vector_affinity(struct ib_device *ibdev, int comp_vector)
5054 {
5055 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5056 
5057 	return mlx5_get_vector_affinity_hint(dev->mdev, comp_vector);
5058 }
5059 
5060 /* The mlx5_ib_multiport_mutex should be held when calling this function */
5061 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
5062 				      struct mlx5_ib_multiport_info *mpi)
5063 {
5064 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
5065 	struct mlx5_ib_port *port = &ibdev->port[port_num];
5066 	int comps;
5067 	int err;
5068 	int i;
5069 
5070 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
5071 
5072 	spin_lock(&port->mp.mpi_lock);
5073 	if (!mpi->ibdev) {
5074 		spin_unlock(&port->mp.mpi_lock);
5075 		return;
5076 	}
5077 	mpi->ibdev = NULL;
5078 
5079 	spin_unlock(&port->mp.mpi_lock);
5080 	mlx5_remove_netdev_notifier(ibdev, port_num);
5081 	spin_lock(&port->mp.mpi_lock);
5082 
5083 	comps = mpi->mdev_refcnt;
5084 	if (comps) {
5085 		mpi->unaffiliate = true;
5086 		init_completion(&mpi->unref_comp);
5087 		spin_unlock(&port->mp.mpi_lock);
5088 
5089 		for (i = 0; i < comps; i++)
5090 			wait_for_completion(&mpi->unref_comp);
5091 
5092 		spin_lock(&port->mp.mpi_lock);
5093 		mpi->unaffiliate = false;
5094 	}
5095 
5096 	port->mp.mpi = NULL;
5097 
5098 	list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
5099 
5100 	spin_unlock(&port->mp.mpi_lock);
5101 
5102 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
5103 
5104 	mlx5_ib_dbg(ibdev, "unaffiliated port %d\n", port_num + 1);
5105 	/* Log an error, still needed to cleanup the pointers and add
5106 	 * it back to the list.
5107 	 */
5108 	if (err)
5109 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
5110 			    port_num + 1);
5111 
5112 	ibdev->roce[port_num].last_port_state = IB_PORT_DOWN;
5113 }
5114 
5115 /* The mlx5_ib_multiport_mutex should be held when calling this function */
5116 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
5117 				    struct mlx5_ib_multiport_info *mpi)
5118 {
5119 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
5120 	int err;
5121 
5122 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
5123 	if (ibdev->port[port_num].mp.mpi) {
5124 		mlx5_ib_warn(ibdev, "port %d already affiliated.\n",
5125 			     port_num + 1);
5126 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
5127 		return false;
5128 	}
5129 
5130 	ibdev->port[port_num].mp.mpi = mpi;
5131 	mpi->ibdev = ibdev;
5132 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
5133 
5134 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
5135 	if (err)
5136 		goto unbind;
5137 
5138 	err = get_port_caps(ibdev, mlx5_core_native_port_num(mpi->mdev));
5139 	if (err)
5140 		goto unbind;
5141 
5142 	err = mlx5_add_netdev_notifier(ibdev, port_num);
5143 	if (err) {
5144 		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
5145 			    port_num + 1);
5146 		goto unbind;
5147 	}
5148 
5149 	err = mlx5_ib_init_cong_debugfs(ibdev, port_num);
5150 	if (err)
5151 		goto unbind;
5152 
5153 	return true;
5154 
5155 unbind:
5156 	mlx5_ib_unbind_slave_port(ibdev, mpi);
5157 	return false;
5158 }
5159 
5160 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
5161 {
5162 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5163 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
5164 							  port_num + 1);
5165 	struct mlx5_ib_multiport_info *mpi;
5166 	int err;
5167 	int i;
5168 
5169 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
5170 		return 0;
5171 
5172 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
5173 						     &dev->sys_image_guid);
5174 	if (err)
5175 		return err;
5176 
5177 	err = mlx5_nic_vport_enable_roce(dev->mdev);
5178 	if (err)
5179 		return err;
5180 
5181 	mutex_lock(&mlx5_ib_multiport_mutex);
5182 	for (i = 0; i < dev->num_ports; i++) {
5183 		bool bound = false;
5184 
5185 		/* build a stub multiport info struct for the native port. */
5186 		if (i == port_num) {
5187 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
5188 			if (!mpi) {
5189 				mutex_unlock(&mlx5_ib_multiport_mutex);
5190 				mlx5_nic_vport_disable_roce(dev->mdev);
5191 				return -ENOMEM;
5192 			}
5193 
5194 			mpi->is_master = true;
5195 			mpi->mdev = dev->mdev;
5196 			mpi->sys_image_guid = dev->sys_image_guid;
5197 			dev->port[i].mp.mpi = mpi;
5198 			mpi->ibdev = dev;
5199 			mpi = NULL;
5200 			continue;
5201 		}
5202 
5203 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
5204 				    list) {
5205 			if (dev->sys_image_guid == mpi->sys_image_guid &&
5206 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
5207 				bound = mlx5_ib_bind_slave_port(dev, mpi);
5208 			}
5209 
5210 			if (bound) {
5211 				dev_dbg(&mpi->mdev->pdev->dev, "removing port from unaffiliated list.\n");
5212 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
5213 				list_del(&mpi->list);
5214 				break;
5215 			}
5216 		}
5217 		if (!bound) {
5218 			get_port_caps(dev, i + 1);
5219 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
5220 				    i + 1);
5221 		}
5222 	}
5223 
5224 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
5225 	mutex_unlock(&mlx5_ib_multiport_mutex);
5226 	return err;
5227 }
5228 
5229 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
5230 {
5231 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5232 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
5233 							  port_num + 1);
5234 	int i;
5235 
5236 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
5237 		return;
5238 
5239 	mutex_lock(&mlx5_ib_multiport_mutex);
5240 	for (i = 0; i < dev->num_ports; i++) {
5241 		if (dev->port[i].mp.mpi) {
5242 			/* Destroy the native port stub */
5243 			if (i == port_num) {
5244 				kfree(dev->port[i].mp.mpi);
5245 				dev->port[i].mp.mpi = NULL;
5246 			} else {
5247 				mlx5_ib_dbg(dev, "unbinding port_num: %d\n", i + 1);
5248 				mlx5_ib_unbind_slave_port(dev, dev->port[i].mp.mpi);
5249 			}
5250 		}
5251 	}
5252 
5253 	mlx5_ib_dbg(dev, "removing from devlist\n");
5254 	list_del(&dev->ib_dev_list);
5255 	mutex_unlock(&mlx5_ib_multiport_mutex);
5256 
5257 	mlx5_nic_vport_disable_roce(dev->mdev);
5258 }
5259 
5260 ADD_UVERBS_ATTRIBUTES_SIMPLE(mlx5_ib_dm, UVERBS_OBJECT_DM,
5261 			     UVERBS_METHOD_DM_ALLOC,
5262 			     &UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
5263 						  UVERBS_ATTR_TYPE(u64),
5264 						  UA_FLAGS(UVERBS_ATTR_SPEC_F_MANDATORY)),
5265 			     &UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
5266 						  UVERBS_ATTR_TYPE(u16),
5267 						  UA_FLAGS(UVERBS_ATTR_SPEC_F_MANDATORY)));
5268 
5269 ADD_UVERBS_ATTRIBUTES_SIMPLE(mlx5_ib_flow_action, UVERBS_OBJECT_FLOW_ACTION,
5270 			     UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
5271 			     &UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
5272 						 UVERBS_ATTR_TYPE(u64),
5273 						 UA_FLAGS(UVERBS_ATTR_SPEC_F_MANDATORY)));
5274 
5275 #define NUM_TREES	2
5276 static int populate_specs_root(struct mlx5_ib_dev *dev)
5277 {
5278 	const struct uverbs_object_tree_def *default_root[NUM_TREES + 1] = {
5279 		uverbs_default_get_objects()};
5280 	size_t num_trees = 1;
5281 
5282 	if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_DEVICE &&
5283 	    !WARN_ON(num_trees >= ARRAY_SIZE(default_root)))
5284 		default_root[num_trees++] = &mlx5_ib_flow_action;
5285 
5286 	if (MLX5_CAP_DEV_MEM(dev->mdev, memic) &&
5287 	    !WARN_ON(num_trees >= ARRAY_SIZE(default_root)))
5288 		default_root[num_trees++] = &mlx5_ib_dm;
5289 
5290 	dev->ib_dev.specs_root =
5291 		uverbs_alloc_spec_tree(num_trees, default_root);
5292 
5293 	return PTR_ERR_OR_ZERO(dev->ib_dev.specs_root);
5294 }
5295 
5296 static void depopulate_specs_root(struct mlx5_ib_dev *dev)
5297 {
5298 	uverbs_free_spec_tree(dev->ib_dev.specs_root);
5299 }
5300 
5301 static int mlx5_ib_read_counters(struct ib_counters *counters,
5302 				 struct ib_counters_read_attr *read_attr,
5303 				 struct uverbs_attr_bundle *attrs)
5304 {
5305 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
5306 	struct mlx5_read_counters_attr mread_attr = {};
5307 	struct mlx5_ib_flow_counters_desc *desc;
5308 	int ret, i;
5309 
5310 	mutex_lock(&mcounters->mcntrs_mutex);
5311 	if (mcounters->cntrs_max_index > read_attr->ncounters) {
5312 		ret = -EINVAL;
5313 		goto err_bound;
5314 	}
5315 
5316 	mread_attr.out = kcalloc(mcounters->counters_num, sizeof(u64),
5317 				 GFP_KERNEL);
5318 	if (!mread_attr.out) {
5319 		ret = -ENOMEM;
5320 		goto err_bound;
5321 	}
5322 
5323 	mread_attr.hw_cntrs_hndl = mcounters->hw_cntrs_hndl;
5324 	mread_attr.flags = read_attr->flags;
5325 	ret = mcounters->read_counters(counters->device, &mread_attr);
5326 	if (ret)
5327 		goto err_read;
5328 
5329 	/* do the pass over the counters data array to assign according to the
5330 	 * descriptions and indexing pairs
5331 	 */
5332 	desc = mcounters->counters_data;
5333 	for (i = 0; i < mcounters->ncounters; i++)
5334 		read_attr->counters_buff[desc[i].index] += mread_attr.out[desc[i].description];
5335 
5336 err_read:
5337 	kfree(mread_attr.out);
5338 err_bound:
5339 	mutex_unlock(&mcounters->mcntrs_mutex);
5340 	return ret;
5341 }
5342 
5343 static int mlx5_ib_destroy_counters(struct ib_counters *counters)
5344 {
5345 	struct mlx5_ib_mcounters *mcounters = to_mcounters(counters);
5346 
5347 	counters_clear_description(counters);
5348 	if (mcounters->hw_cntrs_hndl)
5349 		mlx5_fc_destroy(to_mdev(counters->device)->mdev,
5350 				mcounters->hw_cntrs_hndl);
5351 
5352 	kfree(mcounters);
5353 
5354 	return 0;
5355 }
5356 
5357 static struct ib_counters *mlx5_ib_create_counters(struct ib_device *device,
5358 						   struct uverbs_attr_bundle *attrs)
5359 {
5360 	struct mlx5_ib_mcounters *mcounters;
5361 
5362 	mcounters = kzalloc(sizeof(*mcounters), GFP_KERNEL);
5363 	if (!mcounters)
5364 		return ERR_PTR(-ENOMEM);
5365 
5366 	mutex_init(&mcounters->mcntrs_mutex);
5367 
5368 	return &mcounters->ibcntrs;
5369 }
5370 
5371 void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
5372 {
5373 	mlx5_ib_cleanup_multiport_master(dev);
5374 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
5375 	cleanup_srcu_struct(&dev->mr_srcu);
5376 #endif
5377 	kfree(dev->port);
5378 }
5379 
5380 int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
5381 {
5382 	struct mlx5_core_dev *mdev = dev->mdev;
5383 	const char *name;
5384 	int err;
5385 	int i;
5386 
5387 	dev->port = kcalloc(dev->num_ports, sizeof(*dev->port),
5388 			    GFP_KERNEL);
5389 	if (!dev->port)
5390 		return -ENOMEM;
5391 
5392 	for (i = 0; i < dev->num_ports; i++) {
5393 		spin_lock_init(&dev->port[i].mp.mpi_lock);
5394 		rwlock_init(&dev->roce[i].netdev_lock);
5395 	}
5396 
5397 	err = mlx5_ib_init_multiport_master(dev);
5398 	if (err)
5399 		goto err_free_port;
5400 
5401 	if (!mlx5_core_mp_enabled(mdev)) {
5402 		for (i = 1; i <= dev->num_ports; i++) {
5403 			err = get_port_caps(dev, i);
5404 			if (err)
5405 				break;
5406 		}
5407 	} else {
5408 		err = get_port_caps(dev, mlx5_core_native_port_num(mdev));
5409 	}
5410 	if (err)
5411 		goto err_mp;
5412 
5413 	if (mlx5_use_mad_ifc(dev))
5414 		get_ext_port_caps(dev);
5415 
5416 	if (!mlx5_lag_is_active(mdev))
5417 		name = "mlx5_%d";
5418 	else
5419 		name = "mlx5_bond_%d";
5420 
5421 	strlcpy(dev->ib_dev.name, name, IB_DEVICE_NAME_MAX);
5422 	dev->ib_dev.owner		= THIS_MODULE;
5423 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
5424 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
5425 	dev->ib_dev.phys_port_cnt	= dev->num_ports;
5426 	dev->ib_dev.num_comp_vectors    =
5427 		dev->mdev->priv.eq_table.num_comp_vectors;
5428 	dev->ib_dev.dev.parent		= &mdev->pdev->dev;
5429 
5430 	mutex_init(&dev->cap_mask_mutex);
5431 	INIT_LIST_HEAD(&dev->qp_list);
5432 	spin_lock_init(&dev->reset_flow_resource_lock);
5433 
5434 	spin_lock_init(&dev->memic.memic_lock);
5435 	dev->memic.dev = mdev;
5436 
5437 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
5438 	err = init_srcu_struct(&dev->mr_srcu);
5439 	if (err)
5440 		goto err_free_port;
5441 #endif
5442 
5443 	return 0;
5444 err_mp:
5445 	mlx5_ib_cleanup_multiport_master(dev);
5446 
5447 err_free_port:
5448 	kfree(dev->port);
5449 
5450 	return -ENOMEM;
5451 }
5452 
5453 static int mlx5_ib_stage_flow_db_init(struct mlx5_ib_dev *dev)
5454 {
5455 	dev->flow_db = kzalloc(sizeof(*dev->flow_db), GFP_KERNEL);
5456 
5457 	if (!dev->flow_db)
5458 		return -ENOMEM;
5459 
5460 	mutex_init(&dev->flow_db->lock);
5461 
5462 	return 0;
5463 }
5464 
5465 int mlx5_ib_stage_rep_flow_db_init(struct mlx5_ib_dev *dev)
5466 {
5467 	struct mlx5_ib_dev *nic_dev;
5468 
5469 	nic_dev = mlx5_ib_get_uplink_ibdev(dev->mdev->priv.eswitch);
5470 
5471 	if (!nic_dev)
5472 		return -EINVAL;
5473 
5474 	dev->flow_db = nic_dev->flow_db;
5475 
5476 	return 0;
5477 }
5478 
5479 static void mlx5_ib_stage_flow_db_cleanup(struct mlx5_ib_dev *dev)
5480 {
5481 	kfree(dev->flow_db);
5482 }
5483 
5484 int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
5485 {
5486 	struct mlx5_core_dev *mdev = dev->mdev;
5487 	int err;
5488 
5489 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
5490 	dev->ib_dev.uverbs_cmd_mask	=
5491 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
5492 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
5493 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
5494 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
5495 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
5496 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
5497 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
5498 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
5499 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
5500 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
5501 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
5502 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
5503 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
5504 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
5505 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
5506 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
5507 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
5508 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
5509 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
5510 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
5511 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
5512 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
5513 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
5514 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
5515 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
5516 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
5517 	dev->ib_dev.uverbs_ex_cmd_mask =
5518 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
5519 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
5520 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP)	|
5521 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP)	|
5522 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_CQ);
5523 
5524 	dev->ib_dev.query_device	= mlx5_ib_query_device;
5525 	dev->ib_dev.get_link_layer	= mlx5_ib_port_link_layer;
5526 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
5527 	dev->ib_dev.add_gid		= mlx5_ib_add_gid;
5528 	dev->ib_dev.del_gid		= mlx5_ib_del_gid;
5529 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
5530 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
5531 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
5532 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
5533 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
5534 	dev->ib_dev.mmap		= mlx5_ib_mmap;
5535 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
5536 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
5537 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
5538 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
5539 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
5540 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
5541 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
5542 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
5543 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
5544 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
5545 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
5546 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
5547 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
5548 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
5549 	dev->ib_dev.post_send		= mlx5_ib_post_send;
5550 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
5551 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
5552 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
5553 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
5554 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
5555 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
5556 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
5557 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
5558 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
5559 	dev->ib_dev.rereg_user_mr	= mlx5_ib_rereg_user_mr;
5560 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
5561 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
5562 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
5563 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
5564 	dev->ib_dev.alloc_mr		= mlx5_ib_alloc_mr;
5565 	dev->ib_dev.map_mr_sg		= mlx5_ib_map_mr_sg;
5566 	dev->ib_dev.check_mr_status	= mlx5_ib_check_mr_status;
5567 	dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
5568 	dev->ib_dev.get_vector_affinity	= mlx5_ib_get_vector_affinity;
5569 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads))
5570 		dev->ib_dev.alloc_rdma_netdev	= mlx5_ib_alloc_rdma_netdev;
5571 
5572 	if (mlx5_core_is_pf(mdev)) {
5573 		dev->ib_dev.get_vf_config	= mlx5_ib_get_vf_config;
5574 		dev->ib_dev.set_vf_link_state	= mlx5_ib_set_vf_link_state;
5575 		dev->ib_dev.get_vf_stats	= mlx5_ib_get_vf_stats;
5576 		dev->ib_dev.set_vf_guid		= mlx5_ib_set_vf_guid;
5577 	}
5578 
5579 	dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
5580 
5581 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
5582 
5583 	if (MLX5_CAP_GEN(mdev, imaicl)) {
5584 		dev->ib_dev.alloc_mw		= mlx5_ib_alloc_mw;
5585 		dev->ib_dev.dealloc_mw		= mlx5_ib_dealloc_mw;
5586 		dev->ib_dev.uverbs_cmd_mask |=
5587 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
5588 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
5589 	}
5590 
5591 	if (MLX5_CAP_GEN(mdev, xrc)) {
5592 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
5593 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
5594 		dev->ib_dev.uverbs_cmd_mask |=
5595 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
5596 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
5597 	}
5598 
5599 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
5600 		dev->ib_dev.alloc_dm = mlx5_ib_alloc_dm;
5601 		dev->ib_dev.dealloc_dm = mlx5_ib_dealloc_dm;
5602 		dev->ib_dev.reg_dm_mr = mlx5_ib_reg_dm_mr;
5603 	}
5604 
5605 	dev->ib_dev.create_flow	= mlx5_ib_create_flow;
5606 	dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
5607 	dev->ib_dev.uverbs_ex_cmd_mask |=
5608 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
5609 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW);
5610 	dev->ib_dev.create_flow_action_esp = mlx5_ib_create_flow_action_esp;
5611 	dev->ib_dev.destroy_flow_action = mlx5_ib_destroy_flow_action;
5612 	dev->ib_dev.modify_flow_action_esp = mlx5_ib_modify_flow_action_esp;
5613 	dev->ib_dev.driver_id = RDMA_DRIVER_MLX5;
5614 	dev->ib_dev.create_counters = mlx5_ib_create_counters;
5615 	dev->ib_dev.destroy_counters = mlx5_ib_destroy_counters;
5616 	dev->ib_dev.read_counters = mlx5_ib_read_counters;
5617 
5618 	err = init_node_data(dev);
5619 	if (err)
5620 		return err;
5621 
5622 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
5623 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
5624 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
5625 		mutex_init(&dev->lb_mutex);
5626 
5627 	return 0;
5628 }
5629 
5630 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
5631 {
5632 	dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
5633 	dev->ib_dev.query_port		= mlx5_ib_query_port;
5634 
5635 	return 0;
5636 }
5637 
5638 int mlx5_ib_stage_rep_non_default_cb(struct mlx5_ib_dev *dev)
5639 {
5640 	dev->ib_dev.get_port_immutable  = mlx5_port_rep_immutable;
5641 	dev->ib_dev.query_port		= mlx5_ib_rep_query_port;
5642 
5643 	return 0;
5644 }
5645 
5646 static int mlx5_ib_stage_common_roce_init(struct mlx5_ib_dev *dev,
5647 					  u8 port_num)
5648 {
5649 	int i;
5650 
5651 	for (i = 0; i < dev->num_ports; i++) {
5652 		dev->roce[i].dev = dev;
5653 		dev->roce[i].native_port_num = i + 1;
5654 		dev->roce[i].last_port_state = IB_PORT_DOWN;
5655 	}
5656 
5657 	dev->ib_dev.get_netdev	= mlx5_ib_get_netdev;
5658 	dev->ib_dev.create_wq	 = mlx5_ib_create_wq;
5659 	dev->ib_dev.modify_wq	 = mlx5_ib_modify_wq;
5660 	dev->ib_dev.destroy_wq	 = mlx5_ib_destroy_wq;
5661 	dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
5662 	dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
5663 
5664 	dev->ib_dev.uverbs_ex_cmd_mask |=
5665 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
5666 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
5667 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
5668 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
5669 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
5670 
5671 	return mlx5_add_netdev_notifier(dev, port_num);
5672 }
5673 
5674 static void mlx5_ib_stage_common_roce_cleanup(struct mlx5_ib_dev *dev)
5675 {
5676 	u8 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5677 
5678 	mlx5_remove_netdev_notifier(dev, port_num);
5679 }
5680 
5681 int mlx5_ib_stage_rep_roce_init(struct mlx5_ib_dev *dev)
5682 {
5683 	struct mlx5_core_dev *mdev = dev->mdev;
5684 	enum rdma_link_layer ll;
5685 	int port_type_cap;
5686 	int err = 0;
5687 	u8 port_num;
5688 
5689 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5690 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5691 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5692 
5693 	if (ll == IB_LINK_LAYER_ETHERNET)
5694 		err = mlx5_ib_stage_common_roce_init(dev, port_num);
5695 
5696 	return err;
5697 }
5698 
5699 void mlx5_ib_stage_rep_roce_cleanup(struct mlx5_ib_dev *dev)
5700 {
5701 	mlx5_ib_stage_common_roce_cleanup(dev);
5702 }
5703 
5704 static int mlx5_ib_stage_roce_init(struct mlx5_ib_dev *dev)
5705 {
5706 	struct mlx5_core_dev *mdev = dev->mdev;
5707 	enum rdma_link_layer ll;
5708 	int port_type_cap;
5709 	u8 port_num;
5710 	int err;
5711 
5712 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5713 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5714 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5715 
5716 	if (ll == IB_LINK_LAYER_ETHERNET) {
5717 		err = mlx5_ib_stage_common_roce_init(dev, port_num);
5718 		if (err)
5719 			return err;
5720 
5721 		err = mlx5_enable_eth(dev, port_num);
5722 		if (err)
5723 			goto cleanup;
5724 	}
5725 
5726 	return 0;
5727 cleanup:
5728 	mlx5_ib_stage_common_roce_cleanup(dev);
5729 
5730 	return err;
5731 }
5732 
5733 static void mlx5_ib_stage_roce_cleanup(struct mlx5_ib_dev *dev)
5734 {
5735 	struct mlx5_core_dev *mdev = dev->mdev;
5736 	enum rdma_link_layer ll;
5737 	int port_type_cap;
5738 	u8 port_num;
5739 
5740 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5741 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5742 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5743 
5744 	if (ll == IB_LINK_LAYER_ETHERNET) {
5745 		mlx5_disable_eth(dev);
5746 		mlx5_ib_stage_common_roce_cleanup(dev);
5747 	}
5748 }
5749 
5750 int mlx5_ib_stage_dev_res_init(struct mlx5_ib_dev *dev)
5751 {
5752 	return create_dev_resources(&dev->devr);
5753 }
5754 
5755 void mlx5_ib_stage_dev_res_cleanup(struct mlx5_ib_dev *dev)
5756 {
5757 	destroy_dev_resources(&dev->devr);
5758 }
5759 
5760 static int mlx5_ib_stage_odp_init(struct mlx5_ib_dev *dev)
5761 {
5762 	mlx5_ib_internal_fill_odp_caps(dev);
5763 
5764 	return mlx5_ib_odp_init_one(dev);
5765 }
5766 
5767 int mlx5_ib_stage_counters_init(struct mlx5_ib_dev *dev)
5768 {
5769 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
5770 		dev->ib_dev.get_hw_stats	= mlx5_ib_get_hw_stats;
5771 		dev->ib_dev.alloc_hw_stats	= mlx5_ib_alloc_hw_stats;
5772 
5773 		return mlx5_ib_alloc_counters(dev);
5774 	}
5775 
5776 	return 0;
5777 }
5778 
5779 void mlx5_ib_stage_counters_cleanup(struct mlx5_ib_dev *dev)
5780 {
5781 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
5782 		mlx5_ib_dealloc_counters(dev);
5783 }
5784 
5785 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
5786 {
5787 	return mlx5_ib_init_cong_debugfs(dev,
5788 					 mlx5_core_native_port_num(dev->mdev) - 1);
5789 }
5790 
5791 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
5792 {
5793 	mlx5_ib_cleanup_cong_debugfs(dev,
5794 				     mlx5_core_native_port_num(dev->mdev) - 1);
5795 }
5796 
5797 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
5798 {
5799 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
5800 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
5801 }
5802 
5803 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
5804 {
5805 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
5806 }
5807 
5808 int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
5809 {
5810 	int err;
5811 
5812 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
5813 	if (err)
5814 		return err;
5815 
5816 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
5817 	if (err)
5818 		mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
5819 
5820 	return err;
5821 }
5822 
5823 void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
5824 {
5825 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
5826 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
5827 }
5828 
5829 static int mlx5_ib_stage_populate_specs(struct mlx5_ib_dev *dev)
5830 {
5831 	return populate_specs_root(dev);
5832 }
5833 
5834 int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
5835 {
5836 	return ib_register_device(&dev->ib_dev, NULL);
5837 }
5838 
5839 static void mlx5_ib_stage_depopulate_specs(struct mlx5_ib_dev *dev)
5840 {
5841 	depopulate_specs_root(dev);
5842 }
5843 
5844 void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
5845 {
5846 	destroy_umrc_res(dev);
5847 }
5848 
5849 void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
5850 {
5851 	ib_unregister_device(&dev->ib_dev);
5852 }
5853 
5854 int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
5855 {
5856 	return create_umr_res(dev);
5857 }
5858 
5859 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
5860 {
5861 	init_delay_drop(dev);
5862 
5863 	return 0;
5864 }
5865 
5866 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
5867 {
5868 	cancel_delay_drop(dev);
5869 }
5870 
5871 int mlx5_ib_stage_class_attr_init(struct mlx5_ib_dev *dev)
5872 {
5873 	int err;
5874 	int i;
5875 
5876 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
5877 		err = device_create_file(&dev->ib_dev.dev,
5878 					 mlx5_class_attributes[i]);
5879 		if (err)
5880 			return err;
5881 	}
5882 
5883 	return 0;
5884 }
5885 
5886 static int mlx5_ib_stage_rep_reg_init(struct mlx5_ib_dev *dev)
5887 {
5888 	mlx5_ib_register_vport_reps(dev);
5889 
5890 	return 0;
5891 }
5892 
5893 static void mlx5_ib_stage_rep_reg_cleanup(struct mlx5_ib_dev *dev)
5894 {
5895 	mlx5_ib_unregister_vport_reps(dev);
5896 }
5897 
5898 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
5899 		      const struct mlx5_ib_profile *profile,
5900 		      int stage)
5901 {
5902 	/* Number of stages to cleanup */
5903 	while (stage) {
5904 		stage--;
5905 		if (profile->stage[stage].cleanup)
5906 			profile->stage[stage].cleanup(dev);
5907 	}
5908 
5909 	ib_dealloc_device((struct ib_device *)dev);
5910 }
5911 
5912 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev, u8 port_num);
5913 
5914 void *__mlx5_ib_add(struct mlx5_ib_dev *dev,
5915 		    const struct mlx5_ib_profile *profile)
5916 {
5917 	int err;
5918 	int i;
5919 
5920 	printk_once(KERN_INFO "%s", mlx5_version);
5921 
5922 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
5923 		if (profile->stage[i].init) {
5924 			err = profile->stage[i].init(dev);
5925 			if (err)
5926 				goto err_out;
5927 		}
5928 	}
5929 
5930 	dev->profile = profile;
5931 	dev->ib_active = true;
5932 
5933 	return dev;
5934 
5935 err_out:
5936 	__mlx5_ib_remove(dev, profile, i);
5937 
5938 	return NULL;
5939 }
5940 
5941 static const struct mlx5_ib_profile pf_profile = {
5942 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
5943 		     mlx5_ib_stage_init_init,
5944 		     mlx5_ib_stage_init_cleanup),
5945 	STAGE_CREATE(MLX5_IB_STAGE_FLOW_DB,
5946 		     mlx5_ib_stage_flow_db_init,
5947 		     mlx5_ib_stage_flow_db_cleanup),
5948 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
5949 		     mlx5_ib_stage_caps_init,
5950 		     NULL),
5951 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
5952 		     mlx5_ib_stage_non_default_cb,
5953 		     NULL),
5954 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
5955 		     mlx5_ib_stage_roce_init,
5956 		     mlx5_ib_stage_roce_cleanup),
5957 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
5958 		     mlx5_ib_stage_dev_res_init,
5959 		     mlx5_ib_stage_dev_res_cleanup),
5960 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
5961 		     mlx5_ib_stage_odp_init,
5962 		     NULL),
5963 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
5964 		     mlx5_ib_stage_counters_init,
5965 		     mlx5_ib_stage_counters_cleanup),
5966 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
5967 		     mlx5_ib_stage_cong_debugfs_init,
5968 		     mlx5_ib_stage_cong_debugfs_cleanup),
5969 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
5970 		     mlx5_ib_stage_uar_init,
5971 		     mlx5_ib_stage_uar_cleanup),
5972 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
5973 		     mlx5_ib_stage_bfrag_init,
5974 		     mlx5_ib_stage_bfrag_cleanup),
5975 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
5976 		     NULL,
5977 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
5978 	STAGE_CREATE(MLX5_IB_STAGE_SPECS,
5979 		     mlx5_ib_stage_populate_specs,
5980 		     mlx5_ib_stage_depopulate_specs),
5981 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
5982 		     mlx5_ib_stage_ib_reg_init,
5983 		     mlx5_ib_stage_ib_reg_cleanup),
5984 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
5985 		     mlx5_ib_stage_post_ib_reg_umr_init,
5986 		     NULL),
5987 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
5988 		     mlx5_ib_stage_delay_drop_init,
5989 		     mlx5_ib_stage_delay_drop_cleanup),
5990 	STAGE_CREATE(MLX5_IB_STAGE_CLASS_ATTR,
5991 		     mlx5_ib_stage_class_attr_init,
5992 		     NULL),
5993 };
5994 
5995 static const struct mlx5_ib_profile nic_rep_profile = {
5996 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
5997 		     mlx5_ib_stage_init_init,
5998 		     mlx5_ib_stage_init_cleanup),
5999 	STAGE_CREATE(MLX5_IB_STAGE_FLOW_DB,
6000 		     mlx5_ib_stage_flow_db_init,
6001 		     mlx5_ib_stage_flow_db_cleanup),
6002 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
6003 		     mlx5_ib_stage_caps_init,
6004 		     NULL),
6005 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
6006 		     mlx5_ib_stage_rep_non_default_cb,
6007 		     NULL),
6008 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
6009 		     mlx5_ib_stage_rep_roce_init,
6010 		     mlx5_ib_stage_rep_roce_cleanup),
6011 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
6012 		     mlx5_ib_stage_dev_res_init,
6013 		     mlx5_ib_stage_dev_res_cleanup),
6014 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
6015 		     mlx5_ib_stage_counters_init,
6016 		     mlx5_ib_stage_counters_cleanup),
6017 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
6018 		     mlx5_ib_stage_uar_init,
6019 		     mlx5_ib_stage_uar_cleanup),
6020 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
6021 		     mlx5_ib_stage_bfrag_init,
6022 		     mlx5_ib_stage_bfrag_cleanup),
6023 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
6024 		     NULL,
6025 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
6026 	STAGE_CREATE(MLX5_IB_STAGE_SPECS,
6027 		     mlx5_ib_stage_populate_specs,
6028 		     mlx5_ib_stage_depopulate_specs),
6029 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
6030 		     mlx5_ib_stage_ib_reg_init,
6031 		     mlx5_ib_stage_ib_reg_cleanup),
6032 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
6033 		     mlx5_ib_stage_post_ib_reg_umr_init,
6034 		     NULL),
6035 	STAGE_CREATE(MLX5_IB_STAGE_CLASS_ATTR,
6036 		     mlx5_ib_stage_class_attr_init,
6037 		     NULL),
6038 	STAGE_CREATE(MLX5_IB_STAGE_REP_REG,
6039 		     mlx5_ib_stage_rep_reg_init,
6040 		     mlx5_ib_stage_rep_reg_cleanup),
6041 };
6042 
6043 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev, u8 port_num)
6044 {
6045 	struct mlx5_ib_multiport_info *mpi;
6046 	struct mlx5_ib_dev *dev;
6047 	bool bound = false;
6048 	int err;
6049 
6050 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
6051 	if (!mpi)
6052 		return NULL;
6053 
6054 	mpi->mdev = mdev;
6055 
6056 	err = mlx5_query_nic_vport_system_image_guid(mdev,
6057 						     &mpi->sys_image_guid);
6058 	if (err) {
6059 		kfree(mpi);
6060 		return NULL;
6061 	}
6062 
6063 	mutex_lock(&mlx5_ib_multiport_mutex);
6064 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
6065 		if (dev->sys_image_guid == mpi->sys_image_guid)
6066 			bound = mlx5_ib_bind_slave_port(dev, mpi);
6067 
6068 		if (bound) {
6069 			rdma_roce_rescan_device(&dev->ib_dev);
6070 			break;
6071 		}
6072 	}
6073 
6074 	if (!bound) {
6075 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
6076 		dev_dbg(&mdev->pdev->dev, "no suitable IB device found to bind to, added to unaffiliated list.\n");
6077 	} else {
6078 		mlx5_ib_dbg(dev, "bound port %u\n", port_num + 1);
6079 	}
6080 	mutex_unlock(&mlx5_ib_multiport_mutex);
6081 
6082 	return mpi;
6083 }
6084 
6085 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
6086 {
6087 	enum rdma_link_layer ll;
6088 	struct mlx5_ib_dev *dev;
6089 	int port_type_cap;
6090 
6091 	printk_once(KERN_INFO "%s", mlx5_version);
6092 
6093 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
6094 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
6095 
6096 	if (mlx5_core_is_mp_slave(mdev) && ll == IB_LINK_LAYER_ETHERNET) {
6097 		u8 port_num = mlx5_core_native_port_num(mdev) - 1;
6098 
6099 		return mlx5_ib_add_slave_port(mdev, port_num);
6100 	}
6101 
6102 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
6103 	if (!dev)
6104 		return NULL;
6105 
6106 	dev->mdev = mdev;
6107 	dev->num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
6108 			     MLX5_CAP_GEN(mdev, num_vhca_ports));
6109 
6110 	if (MLX5_VPORT_MANAGER(mdev) &&
6111 	    mlx5_ib_eswitch_mode(mdev->priv.eswitch) == SRIOV_OFFLOADS) {
6112 		dev->rep = mlx5_ib_vport_rep(mdev->priv.eswitch, 0);
6113 
6114 		return __mlx5_ib_add(dev, &nic_rep_profile);
6115 	}
6116 
6117 	return __mlx5_ib_add(dev, &pf_profile);
6118 }
6119 
6120 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
6121 {
6122 	struct mlx5_ib_multiport_info *mpi;
6123 	struct mlx5_ib_dev *dev;
6124 
6125 	if (mlx5_core_is_mp_slave(mdev)) {
6126 		mpi = context;
6127 		mutex_lock(&mlx5_ib_multiport_mutex);
6128 		if (mpi->ibdev)
6129 			mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
6130 		list_del(&mpi->list);
6131 		mutex_unlock(&mlx5_ib_multiport_mutex);
6132 		return;
6133 	}
6134 
6135 	dev = context;
6136 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
6137 }
6138 
6139 static struct mlx5_interface mlx5_ib_interface = {
6140 	.add            = mlx5_ib_add,
6141 	.remove         = mlx5_ib_remove,
6142 	.event          = mlx5_ib_event,
6143 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
6144 	.pfault		= mlx5_ib_pfault,
6145 #endif
6146 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
6147 };
6148 
6149 unsigned long mlx5_ib_get_xlt_emergency_page(void)
6150 {
6151 	mutex_lock(&xlt_emergency_page_mutex);
6152 	return xlt_emergency_page;
6153 }
6154 
6155 void mlx5_ib_put_xlt_emergency_page(void)
6156 {
6157 	mutex_unlock(&xlt_emergency_page_mutex);
6158 }
6159 
6160 static int __init mlx5_ib_init(void)
6161 {
6162 	int err;
6163 
6164 	xlt_emergency_page = __get_free_page(GFP_KERNEL);
6165 	if (!xlt_emergency_page)
6166 		return -ENOMEM;
6167 
6168 	mutex_init(&xlt_emergency_page_mutex);
6169 
6170 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
6171 	if (!mlx5_ib_event_wq) {
6172 		free_page(xlt_emergency_page);
6173 		return -ENOMEM;
6174 	}
6175 
6176 	mlx5_ib_odp_init();
6177 
6178 	err = mlx5_register_interface(&mlx5_ib_interface);
6179 
6180 	return err;
6181 }
6182 
6183 static void __exit mlx5_ib_cleanup(void)
6184 {
6185 	mlx5_unregister_interface(&mlx5_ib_interface);
6186 	destroy_workqueue(mlx5_ib_event_wq);
6187 	mutex_destroy(&xlt_emergency_page_mutex);
6188 	free_page(xlt_emergency_page);
6189 }
6190 
6191 module_init(mlx5_ib_init);
6192 module_exit(mlx5_ib_cleanup);
6193