xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision e33bbe69149b802c0c77bfb822685772f85388ca)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/debugfs.h>
34 #include <linux/highmem.h>
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/errno.h>
38 #include <linux/pci.h>
39 #include <linux/dma-mapping.h>
40 #include <linux/slab.h>
41 #include <linux/bitmap.h>
42 #if defined(CONFIG_X86)
43 #include <asm/pat.h>
44 #endif
45 #include <linux/sched.h>
46 #include <linux/sched/mm.h>
47 #include <linux/sched/task.h>
48 #include <linux/delay.h>
49 #include <rdma/ib_user_verbs.h>
50 #include <rdma/ib_addr.h>
51 #include <rdma/ib_cache.h>
52 #include <linux/mlx5/port.h>
53 #include <linux/mlx5/vport.h>
54 #include <linux/mlx5/fs.h>
55 #include <linux/mlx5/fs_helpers.h>
56 #include <linux/list.h>
57 #include <rdma/ib_smi.h>
58 #include <rdma/ib_umem.h>
59 #include <linux/in.h>
60 #include <linux/etherdevice.h>
61 #include "mlx5_ib.h"
62 #include "ib_rep.h"
63 #include "cmd.h"
64 #include <linux/mlx5/fs_helpers.h>
65 #include <linux/mlx5/accel.h>
66 #include <rdma/uverbs_std_types.h>
67 #include <rdma/mlx5_user_ioctl_verbs.h>
68 #include <rdma/mlx5_user_ioctl_cmds.h>
69 
70 #define UVERBS_MODULE_NAME mlx5_ib
71 #include <rdma/uverbs_named_ioctl.h>
72 
73 #define DRIVER_NAME "mlx5_ib"
74 #define DRIVER_VERSION "5.0-0"
75 
76 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
77 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
78 MODULE_LICENSE("Dual BSD/GPL");
79 
80 static char mlx5_version[] =
81 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
82 	DRIVER_VERSION "\n";
83 
84 struct mlx5_ib_event_work {
85 	struct work_struct	work;
86 	struct mlx5_core_dev	*dev;
87 	void			*context;
88 	enum mlx5_dev_event	event;
89 	unsigned long		param;
90 };
91 
92 enum {
93 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
94 };
95 
96 static struct workqueue_struct *mlx5_ib_event_wq;
97 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
98 static LIST_HEAD(mlx5_ib_dev_list);
99 /*
100  * This mutex should be held when accessing either of the above lists
101  */
102 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
103 
104 /* We can't use an array for xlt_emergency_page because dma_map_single
105  * doesn't work on kernel modules memory
106  */
107 static unsigned long xlt_emergency_page;
108 static struct mutex xlt_emergency_page_mutex;
109 
110 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
111 {
112 	struct mlx5_ib_dev *dev;
113 
114 	mutex_lock(&mlx5_ib_multiport_mutex);
115 	dev = mpi->ibdev;
116 	mutex_unlock(&mlx5_ib_multiport_mutex);
117 	return dev;
118 }
119 
120 static enum rdma_link_layer
121 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
122 {
123 	switch (port_type_cap) {
124 	case MLX5_CAP_PORT_TYPE_IB:
125 		return IB_LINK_LAYER_INFINIBAND;
126 	case MLX5_CAP_PORT_TYPE_ETH:
127 		return IB_LINK_LAYER_ETHERNET;
128 	default:
129 		return IB_LINK_LAYER_UNSPECIFIED;
130 	}
131 }
132 
133 static enum rdma_link_layer
134 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
135 {
136 	struct mlx5_ib_dev *dev = to_mdev(device);
137 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
138 
139 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
140 }
141 
142 static int get_port_state(struct ib_device *ibdev,
143 			  u8 port_num,
144 			  enum ib_port_state *state)
145 {
146 	struct ib_port_attr attr;
147 	int ret;
148 
149 	memset(&attr, 0, sizeof(attr));
150 	ret = ibdev->query_port(ibdev, port_num, &attr);
151 	if (!ret)
152 		*state = attr.state;
153 	return ret;
154 }
155 
156 static int mlx5_netdev_event(struct notifier_block *this,
157 			     unsigned long event, void *ptr)
158 {
159 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
160 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
161 	u8 port_num = roce->native_port_num;
162 	struct mlx5_core_dev *mdev;
163 	struct mlx5_ib_dev *ibdev;
164 
165 	ibdev = roce->dev;
166 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
167 	if (!mdev)
168 		return NOTIFY_DONE;
169 
170 	switch (event) {
171 	case NETDEV_REGISTER:
172 	case NETDEV_UNREGISTER:
173 		write_lock(&roce->netdev_lock);
174 		if (ibdev->rep) {
175 			struct mlx5_eswitch *esw = ibdev->mdev->priv.eswitch;
176 			struct net_device *rep_ndev;
177 
178 			rep_ndev = mlx5_ib_get_rep_netdev(esw,
179 							  ibdev->rep->vport);
180 			if (rep_ndev == ndev)
181 				roce->netdev = (event == NETDEV_UNREGISTER) ?
182 					NULL : ndev;
183 		} else if (ndev->dev.parent == &ibdev->mdev->pdev->dev) {
184 			roce->netdev = (event == NETDEV_UNREGISTER) ?
185 				NULL : ndev;
186 		}
187 		write_unlock(&roce->netdev_lock);
188 		break;
189 
190 	case NETDEV_CHANGE:
191 	case NETDEV_UP:
192 	case NETDEV_DOWN: {
193 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
194 		struct net_device *upper = NULL;
195 
196 		if (lag_ndev) {
197 			upper = netdev_master_upper_dev_get(lag_ndev);
198 			dev_put(lag_ndev);
199 		}
200 
201 		if ((upper == ndev || (!upper && ndev == roce->netdev))
202 		    && ibdev->ib_active) {
203 			struct ib_event ibev = { };
204 			enum ib_port_state port_state;
205 
206 			if (get_port_state(&ibdev->ib_dev, port_num,
207 					   &port_state))
208 				goto done;
209 
210 			if (roce->last_port_state == port_state)
211 				goto done;
212 
213 			roce->last_port_state = port_state;
214 			ibev.device = &ibdev->ib_dev;
215 			if (port_state == IB_PORT_DOWN)
216 				ibev.event = IB_EVENT_PORT_ERR;
217 			else if (port_state == IB_PORT_ACTIVE)
218 				ibev.event = IB_EVENT_PORT_ACTIVE;
219 			else
220 				goto done;
221 
222 			ibev.element.port_num = port_num;
223 			ib_dispatch_event(&ibev);
224 		}
225 		break;
226 	}
227 
228 	default:
229 		break;
230 	}
231 done:
232 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
233 	return NOTIFY_DONE;
234 }
235 
236 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
237 					     u8 port_num)
238 {
239 	struct mlx5_ib_dev *ibdev = to_mdev(device);
240 	struct net_device *ndev;
241 	struct mlx5_core_dev *mdev;
242 
243 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
244 	if (!mdev)
245 		return NULL;
246 
247 	ndev = mlx5_lag_get_roce_netdev(mdev);
248 	if (ndev)
249 		goto out;
250 
251 	/* Ensure ndev does not disappear before we invoke dev_hold()
252 	 */
253 	read_lock(&ibdev->roce[port_num - 1].netdev_lock);
254 	ndev = ibdev->roce[port_num - 1].netdev;
255 	if (ndev)
256 		dev_hold(ndev);
257 	read_unlock(&ibdev->roce[port_num - 1].netdev_lock);
258 
259 out:
260 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
261 	return ndev;
262 }
263 
264 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
265 						   u8 ib_port_num,
266 						   u8 *native_port_num)
267 {
268 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
269 							  ib_port_num);
270 	struct mlx5_core_dev *mdev = NULL;
271 	struct mlx5_ib_multiport_info *mpi;
272 	struct mlx5_ib_port *port;
273 
274 	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
275 	    ll != IB_LINK_LAYER_ETHERNET) {
276 		if (native_port_num)
277 			*native_port_num = ib_port_num;
278 		return ibdev->mdev;
279 	}
280 
281 	if (native_port_num)
282 		*native_port_num = 1;
283 
284 	port = &ibdev->port[ib_port_num - 1];
285 	if (!port)
286 		return NULL;
287 
288 	spin_lock(&port->mp.mpi_lock);
289 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
290 	if (mpi && !mpi->unaffiliate) {
291 		mdev = mpi->mdev;
292 		/* If it's the master no need to refcount, it'll exist
293 		 * as long as the ib_dev exists.
294 		 */
295 		if (!mpi->is_master)
296 			mpi->mdev_refcnt++;
297 	}
298 	spin_unlock(&port->mp.mpi_lock);
299 
300 	return mdev;
301 }
302 
303 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u8 port_num)
304 {
305 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
306 							  port_num);
307 	struct mlx5_ib_multiport_info *mpi;
308 	struct mlx5_ib_port *port;
309 
310 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
311 		return;
312 
313 	port = &ibdev->port[port_num - 1];
314 
315 	spin_lock(&port->mp.mpi_lock);
316 	mpi = ibdev->port[port_num - 1].mp.mpi;
317 	if (mpi->is_master)
318 		goto out;
319 
320 	mpi->mdev_refcnt--;
321 	if (mpi->unaffiliate)
322 		complete(&mpi->unref_comp);
323 out:
324 	spin_unlock(&port->mp.mpi_lock);
325 }
326 
327 static int translate_eth_proto_oper(u32 eth_proto_oper, u8 *active_speed,
328 				    u8 *active_width)
329 {
330 	switch (eth_proto_oper) {
331 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
332 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
333 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
334 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
335 		*active_width = IB_WIDTH_1X;
336 		*active_speed = IB_SPEED_SDR;
337 		break;
338 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
339 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
340 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
341 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
342 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
343 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
344 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
345 		*active_width = IB_WIDTH_1X;
346 		*active_speed = IB_SPEED_QDR;
347 		break;
348 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
349 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
350 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
351 		*active_width = IB_WIDTH_1X;
352 		*active_speed = IB_SPEED_EDR;
353 		break;
354 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
355 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
356 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
357 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
358 		*active_width = IB_WIDTH_4X;
359 		*active_speed = IB_SPEED_QDR;
360 		break;
361 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
362 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
363 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
364 		*active_width = IB_WIDTH_1X;
365 		*active_speed = IB_SPEED_HDR;
366 		break;
367 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
368 		*active_width = IB_WIDTH_4X;
369 		*active_speed = IB_SPEED_FDR;
370 		break;
371 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
372 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
373 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
374 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
375 		*active_width = IB_WIDTH_4X;
376 		*active_speed = IB_SPEED_EDR;
377 		break;
378 	default:
379 		return -EINVAL;
380 	}
381 
382 	return 0;
383 }
384 
385 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
386 				struct ib_port_attr *props)
387 {
388 	struct mlx5_ib_dev *dev = to_mdev(device);
389 	struct mlx5_core_dev *mdev;
390 	struct net_device *ndev, *upper;
391 	enum ib_mtu ndev_ib_mtu;
392 	bool put_mdev = true;
393 	u16 qkey_viol_cntr;
394 	u32 eth_prot_oper;
395 	u8 mdev_port_num;
396 	int err;
397 
398 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
399 	if (!mdev) {
400 		/* This means the port isn't affiliated yet. Get the
401 		 * info for the master port instead.
402 		 */
403 		put_mdev = false;
404 		mdev = dev->mdev;
405 		mdev_port_num = 1;
406 		port_num = 1;
407 	}
408 
409 	/* Possible bad flows are checked before filling out props so in case
410 	 * of an error it will still be zeroed out.
411 	 */
412 	err = mlx5_query_port_eth_proto_oper(mdev, &eth_prot_oper,
413 					     mdev_port_num);
414 	if (err)
415 		goto out;
416 
417 	props->active_width     = IB_WIDTH_4X;
418 	props->active_speed     = IB_SPEED_QDR;
419 
420 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
421 				 &props->active_width);
422 
423 	props->port_cap_flags  |= IB_PORT_CM_SUP;
424 	props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
425 
426 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
427 						roce_address_table_size);
428 	props->max_mtu          = IB_MTU_4096;
429 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
430 	props->pkey_tbl_len     = 1;
431 	props->state            = IB_PORT_DOWN;
432 	props->phys_state       = 3;
433 
434 	mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
435 	props->qkey_viol_cntr = qkey_viol_cntr;
436 
437 	/* If this is a stub query for an unaffiliated port stop here */
438 	if (!put_mdev)
439 		goto out;
440 
441 	ndev = mlx5_ib_get_netdev(device, port_num);
442 	if (!ndev)
443 		goto out;
444 
445 	if (mlx5_lag_is_active(dev->mdev)) {
446 		rcu_read_lock();
447 		upper = netdev_master_upper_dev_get_rcu(ndev);
448 		if (upper) {
449 			dev_put(ndev);
450 			ndev = upper;
451 			dev_hold(ndev);
452 		}
453 		rcu_read_unlock();
454 	}
455 
456 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
457 		props->state      = IB_PORT_ACTIVE;
458 		props->phys_state = 5;
459 	}
460 
461 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
462 
463 	dev_put(ndev);
464 
465 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
466 out:
467 	if (put_mdev)
468 		mlx5_ib_put_native_port_mdev(dev, port_num);
469 	return err;
470 }
471 
472 static int set_roce_addr(struct mlx5_ib_dev *dev, u8 port_num,
473 			 unsigned int index, const union ib_gid *gid,
474 			 const struct ib_gid_attr *attr)
475 {
476 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
477 	u8 roce_version = 0;
478 	u8 roce_l3_type = 0;
479 	bool vlan = false;
480 	u8 mac[ETH_ALEN];
481 	u16 vlan_id = 0;
482 
483 	if (gid) {
484 		gid_type = attr->gid_type;
485 		ether_addr_copy(mac, attr->ndev->dev_addr);
486 
487 		if (is_vlan_dev(attr->ndev)) {
488 			vlan = true;
489 			vlan_id = vlan_dev_vlan_id(attr->ndev);
490 		}
491 	}
492 
493 	switch (gid_type) {
494 	case IB_GID_TYPE_IB:
495 		roce_version = MLX5_ROCE_VERSION_1;
496 		break;
497 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
498 		roce_version = MLX5_ROCE_VERSION_2;
499 		if (ipv6_addr_v4mapped((void *)gid))
500 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
501 		else
502 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
503 		break;
504 
505 	default:
506 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
507 	}
508 
509 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
510 				      roce_l3_type, gid->raw, mac, vlan,
511 				      vlan_id, port_num);
512 }
513 
514 static int mlx5_ib_add_gid(const union ib_gid *gid,
515 			   const struct ib_gid_attr *attr,
516 			   __always_unused void **context)
517 {
518 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
519 			     attr->index, gid, attr);
520 }
521 
522 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
523 			   __always_unused void **context)
524 {
525 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
526 			     attr->index, NULL, NULL);
527 }
528 
529 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
530 			       int index)
531 {
532 	struct ib_gid_attr attr;
533 	union ib_gid gid;
534 
535 	if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
536 		return 0;
537 
538 	dev_put(attr.ndev);
539 
540 	if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
541 		return 0;
542 
543 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
544 }
545 
546 int mlx5_get_roce_gid_type(struct mlx5_ib_dev *dev, u8 port_num,
547 			   int index, enum ib_gid_type *gid_type)
548 {
549 	struct ib_gid_attr attr;
550 	union ib_gid gid;
551 	int ret;
552 
553 	ret = ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr);
554 	if (ret)
555 		return ret;
556 
557 	dev_put(attr.ndev);
558 
559 	*gid_type = attr.gid_type;
560 
561 	return 0;
562 }
563 
564 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
565 {
566 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
567 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
568 	return 0;
569 }
570 
571 enum {
572 	MLX5_VPORT_ACCESS_METHOD_MAD,
573 	MLX5_VPORT_ACCESS_METHOD_HCA,
574 	MLX5_VPORT_ACCESS_METHOD_NIC,
575 };
576 
577 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
578 {
579 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
580 		return MLX5_VPORT_ACCESS_METHOD_MAD;
581 
582 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
583 	    IB_LINK_LAYER_ETHERNET)
584 		return MLX5_VPORT_ACCESS_METHOD_NIC;
585 
586 	return MLX5_VPORT_ACCESS_METHOD_HCA;
587 }
588 
589 static void get_atomic_caps(struct mlx5_ib_dev *dev,
590 			    u8 atomic_size_qp,
591 			    struct ib_device_attr *props)
592 {
593 	u8 tmp;
594 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
595 	u8 atomic_req_8B_endianness_mode =
596 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
597 
598 	/* Check if HW supports 8 bytes standard atomic operations and capable
599 	 * of host endianness respond
600 	 */
601 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
602 	if (((atomic_operations & tmp) == tmp) &&
603 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
604 	    (atomic_req_8B_endianness_mode)) {
605 		props->atomic_cap = IB_ATOMIC_HCA;
606 	} else {
607 		props->atomic_cap = IB_ATOMIC_NONE;
608 	}
609 }
610 
611 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
612 			       struct ib_device_attr *props)
613 {
614 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
615 
616 	get_atomic_caps(dev, atomic_size_qp, props);
617 }
618 
619 static void get_atomic_caps_dc(struct mlx5_ib_dev *dev,
620 			       struct ib_device_attr *props)
621 {
622 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
623 
624 	get_atomic_caps(dev, atomic_size_qp, props);
625 }
626 
627 bool mlx5_ib_dc_atomic_is_supported(struct mlx5_ib_dev *dev)
628 {
629 	struct ib_device_attr props = {};
630 
631 	get_atomic_caps_dc(dev, &props);
632 	return (props.atomic_cap == IB_ATOMIC_HCA) ? true : false;
633 }
634 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
635 					__be64 *sys_image_guid)
636 {
637 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
638 	struct mlx5_core_dev *mdev = dev->mdev;
639 	u64 tmp;
640 	int err;
641 
642 	switch (mlx5_get_vport_access_method(ibdev)) {
643 	case MLX5_VPORT_ACCESS_METHOD_MAD:
644 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
645 							    sys_image_guid);
646 
647 	case MLX5_VPORT_ACCESS_METHOD_HCA:
648 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
649 		break;
650 
651 	case MLX5_VPORT_ACCESS_METHOD_NIC:
652 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
653 		break;
654 
655 	default:
656 		return -EINVAL;
657 	}
658 
659 	if (!err)
660 		*sys_image_guid = cpu_to_be64(tmp);
661 
662 	return err;
663 
664 }
665 
666 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
667 				u16 *max_pkeys)
668 {
669 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
670 	struct mlx5_core_dev *mdev = dev->mdev;
671 
672 	switch (mlx5_get_vport_access_method(ibdev)) {
673 	case MLX5_VPORT_ACCESS_METHOD_MAD:
674 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
675 
676 	case MLX5_VPORT_ACCESS_METHOD_HCA:
677 	case MLX5_VPORT_ACCESS_METHOD_NIC:
678 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
679 						pkey_table_size));
680 		return 0;
681 
682 	default:
683 		return -EINVAL;
684 	}
685 }
686 
687 static int mlx5_query_vendor_id(struct ib_device *ibdev,
688 				u32 *vendor_id)
689 {
690 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
691 
692 	switch (mlx5_get_vport_access_method(ibdev)) {
693 	case MLX5_VPORT_ACCESS_METHOD_MAD:
694 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
695 
696 	case MLX5_VPORT_ACCESS_METHOD_HCA:
697 	case MLX5_VPORT_ACCESS_METHOD_NIC:
698 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
699 
700 	default:
701 		return -EINVAL;
702 	}
703 }
704 
705 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
706 				__be64 *node_guid)
707 {
708 	u64 tmp;
709 	int err;
710 
711 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
712 	case MLX5_VPORT_ACCESS_METHOD_MAD:
713 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
714 
715 	case MLX5_VPORT_ACCESS_METHOD_HCA:
716 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
717 		break;
718 
719 	case MLX5_VPORT_ACCESS_METHOD_NIC:
720 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
721 		break;
722 
723 	default:
724 		return -EINVAL;
725 	}
726 
727 	if (!err)
728 		*node_guid = cpu_to_be64(tmp);
729 
730 	return err;
731 }
732 
733 struct mlx5_reg_node_desc {
734 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
735 };
736 
737 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
738 {
739 	struct mlx5_reg_node_desc in;
740 
741 	if (mlx5_use_mad_ifc(dev))
742 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
743 
744 	memset(&in, 0, sizeof(in));
745 
746 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
747 				    sizeof(struct mlx5_reg_node_desc),
748 				    MLX5_REG_NODE_DESC, 0, 0);
749 }
750 
751 static int mlx5_ib_query_device(struct ib_device *ibdev,
752 				struct ib_device_attr *props,
753 				struct ib_udata *uhw)
754 {
755 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
756 	struct mlx5_core_dev *mdev = dev->mdev;
757 	int err = -ENOMEM;
758 	int max_sq_desc;
759 	int max_rq_sg;
760 	int max_sq_sg;
761 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
762 	bool raw_support = !mlx5_core_mp_enabled(mdev);
763 	struct mlx5_ib_query_device_resp resp = {};
764 	size_t resp_len;
765 	u64 max_tso;
766 
767 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
768 	if (uhw->outlen && uhw->outlen < resp_len)
769 		return -EINVAL;
770 	else
771 		resp.response_length = resp_len;
772 
773 	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
774 		return -EINVAL;
775 
776 	memset(props, 0, sizeof(*props));
777 	err = mlx5_query_system_image_guid(ibdev,
778 					   &props->sys_image_guid);
779 	if (err)
780 		return err;
781 
782 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
783 	if (err)
784 		return err;
785 
786 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
787 	if (err)
788 		return err;
789 
790 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
791 		(fw_rev_min(dev->mdev) << 16) |
792 		fw_rev_sub(dev->mdev);
793 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
794 		IB_DEVICE_PORT_ACTIVE_EVENT		|
795 		IB_DEVICE_SYS_IMAGE_GUID		|
796 		IB_DEVICE_RC_RNR_NAK_GEN;
797 
798 	if (MLX5_CAP_GEN(mdev, pkv))
799 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
800 	if (MLX5_CAP_GEN(mdev, qkv))
801 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
802 	if (MLX5_CAP_GEN(mdev, apm))
803 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
804 	if (MLX5_CAP_GEN(mdev, xrc))
805 		props->device_cap_flags |= IB_DEVICE_XRC;
806 	if (MLX5_CAP_GEN(mdev, imaicl)) {
807 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
808 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
809 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
810 		/* We support 'Gappy' memory registration too */
811 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
812 	}
813 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
814 	if (MLX5_CAP_GEN(mdev, sho)) {
815 		props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
816 		/* At this stage no support for signature handover */
817 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
818 				      IB_PROT_T10DIF_TYPE_2 |
819 				      IB_PROT_T10DIF_TYPE_3;
820 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
821 				       IB_GUARD_T10DIF_CSUM;
822 	}
823 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
824 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
825 
826 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
827 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
828 			/* Legacy bit to support old userspace libraries */
829 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
830 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
831 		}
832 
833 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
834 			props->raw_packet_caps |=
835 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
836 
837 		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
838 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
839 			if (max_tso) {
840 				resp.tso_caps.max_tso = 1 << max_tso;
841 				resp.tso_caps.supported_qpts |=
842 					1 << IB_QPT_RAW_PACKET;
843 				resp.response_length += sizeof(resp.tso_caps);
844 			}
845 		}
846 
847 		if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
848 			resp.rss_caps.rx_hash_function =
849 						MLX5_RX_HASH_FUNC_TOEPLITZ;
850 			resp.rss_caps.rx_hash_fields_mask =
851 						MLX5_RX_HASH_SRC_IPV4 |
852 						MLX5_RX_HASH_DST_IPV4 |
853 						MLX5_RX_HASH_SRC_IPV6 |
854 						MLX5_RX_HASH_DST_IPV6 |
855 						MLX5_RX_HASH_SRC_PORT_TCP |
856 						MLX5_RX_HASH_DST_PORT_TCP |
857 						MLX5_RX_HASH_SRC_PORT_UDP |
858 						MLX5_RX_HASH_DST_PORT_UDP |
859 						MLX5_RX_HASH_INNER;
860 			if (mlx5_accel_ipsec_device_caps(dev->mdev) &
861 			    MLX5_ACCEL_IPSEC_CAP_DEVICE)
862 				resp.rss_caps.rx_hash_fields_mask |=
863 					MLX5_RX_HASH_IPSEC_SPI;
864 			resp.response_length += sizeof(resp.rss_caps);
865 		}
866 	} else {
867 		if (field_avail(typeof(resp), tso_caps, uhw->outlen))
868 			resp.response_length += sizeof(resp.tso_caps);
869 		if (field_avail(typeof(resp), rss_caps, uhw->outlen))
870 			resp.response_length += sizeof(resp.rss_caps);
871 	}
872 
873 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
874 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
875 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
876 	}
877 
878 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
879 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
880 	    raw_support)
881 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
882 
883 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
884 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
885 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
886 
887 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
888 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
889 	    raw_support) {
890 		/* Legacy bit to support old userspace libraries */
891 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
892 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
893 	}
894 
895 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
896 		props->max_dm_size =
897 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
898 	}
899 
900 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
901 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
902 
903 	if (MLX5_CAP_GEN(mdev, end_pad))
904 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
905 
906 	props->vendor_part_id	   = mdev->pdev->device;
907 	props->hw_ver		   = mdev->pdev->revision;
908 
909 	props->max_mr_size	   = ~0ull;
910 	props->page_size_cap	   = ~(min_page_size - 1);
911 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
912 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
913 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
914 		     sizeof(struct mlx5_wqe_data_seg);
915 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
916 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
917 		     sizeof(struct mlx5_wqe_raddr_seg)) /
918 		sizeof(struct mlx5_wqe_data_seg);
919 	props->max_sge = min(max_rq_sg, max_sq_sg);
920 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
921 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
922 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
923 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
924 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
925 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
926 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
927 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
928 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
929 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
930 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
931 	props->max_srq_sge	   = max_rq_sg - 1;
932 	props->max_fast_reg_page_list_len =
933 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
934 	get_atomic_caps_qp(dev, props);
935 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
936 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
937 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
938 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
939 					   props->max_mcast_grp;
940 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
941 	props->max_ah = INT_MAX;
942 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
943 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
944 
945 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
946 	if (MLX5_CAP_GEN(mdev, pg))
947 		props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
948 	props->odp_caps = dev->odp_caps;
949 #endif
950 
951 	if (MLX5_CAP_GEN(mdev, cd))
952 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
953 
954 	if (!mlx5_core_is_pf(mdev))
955 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
956 
957 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
958 	    IB_LINK_LAYER_ETHERNET && raw_support) {
959 		props->rss_caps.max_rwq_indirection_tables =
960 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
961 		props->rss_caps.max_rwq_indirection_table_size =
962 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
963 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
964 		props->max_wq_type_rq =
965 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
966 	}
967 
968 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
969 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
970 		props->tm_caps.max_num_tags =
971 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
972 		props->tm_caps.flags = IB_TM_CAP_RC;
973 		props->tm_caps.max_ops =
974 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
975 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
976 	}
977 
978 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
979 		props->cq_caps.max_cq_moderation_count =
980 						MLX5_MAX_CQ_COUNT;
981 		props->cq_caps.max_cq_moderation_period =
982 						MLX5_MAX_CQ_PERIOD;
983 	}
984 
985 	if (field_avail(typeof(resp), cqe_comp_caps, uhw->outlen)) {
986 		resp.cqe_comp_caps.max_num =
987 			MLX5_CAP_GEN(dev->mdev, cqe_compression) ?
988 			MLX5_CAP_GEN(dev->mdev, cqe_compression_max_num) : 0;
989 		resp.cqe_comp_caps.supported_format =
990 			MLX5_IB_CQE_RES_FORMAT_HASH |
991 			MLX5_IB_CQE_RES_FORMAT_CSUM;
992 		resp.response_length += sizeof(resp.cqe_comp_caps);
993 	}
994 
995 	if (field_avail(typeof(resp), packet_pacing_caps, uhw->outlen) &&
996 	    raw_support) {
997 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
998 		    MLX5_CAP_GEN(mdev, qos)) {
999 			resp.packet_pacing_caps.qp_rate_limit_max =
1000 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1001 			resp.packet_pacing_caps.qp_rate_limit_min =
1002 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1003 			resp.packet_pacing_caps.supported_qpts |=
1004 				1 << IB_QPT_RAW_PACKET;
1005 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1006 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1007 				resp.packet_pacing_caps.cap_flags |=
1008 					MLX5_IB_PP_SUPPORT_BURST;
1009 		}
1010 		resp.response_length += sizeof(resp.packet_pacing_caps);
1011 	}
1012 
1013 	if (field_avail(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes,
1014 			uhw->outlen)) {
1015 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1016 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1017 				MLX5_IB_ALLOW_MPW;
1018 
1019 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1020 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1021 				MLX5_IB_SUPPORT_EMPW;
1022 
1023 		resp.response_length +=
1024 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1025 	}
1026 
1027 	if (field_avail(typeof(resp), flags, uhw->outlen)) {
1028 		resp.response_length += sizeof(resp.flags);
1029 
1030 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1031 			resp.flags |=
1032 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1033 
1034 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1035 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1036 	}
1037 
1038 	if (field_avail(typeof(resp), sw_parsing_caps,
1039 			uhw->outlen)) {
1040 		resp.response_length += sizeof(resp.sw_parsing_caps);
1041 		if (MLX5_CAP_ETH(mdev, swp)) {
1042 			resp.sw_parsing_caps.sw_parsing_offloads |=
1043 				MLX5_IB_SW_PARSING;
1044 
1045 			if (MLX5_CAP_ETH(mdev, swp_csum))
1046 				resp.sw_parsing_caps.sw_parsing_offloads |=
1047 					MLX5_IB_SW_PARSING_CSUM;
1048 
1049 			if (MLX5_CAP_ETH(mdev, swp_lso))
1050 				resp.sw_parsing_caps.sw_parsing_offloads |=
1051 					MLX5_IB_SW_PARSING_LSO;
1052 
1053 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1054 				resp.sw_parsing_caps.supported_qpts =
1055 					BIT(IB_QPT_RAW_PACKET);
1056 		}
1057 	}
1058 
1059 	if (field_avail(typeof(resp), striding_rq_caps, uhw->outlen) &&
1060 	    raw_support) {
1061 		resp.response_length += sizeof(resp.striding_rq_caps);
1062 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1063 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1064 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1065 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1066 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1067 			resp.striding_rq_caps.min_single_wqe_log_num_of_strides =
1068 				MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1069 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1070 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1071 			resp.striding_rq_caps.supported_qpts =
1072 				BIT(IB_QPT_RAW_PACKET);
1073 		}
1074 	}
1075 
1076 	if (field_avail(typeof(resp), tunnel_offloads_caps,
1077 			uhw->outlen)) {
1078 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1079 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1080 			resp.tunnel_offloads_caps |=
1081 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1082 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1083 			resp.tunnel_offloads_caps |=
1084 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1085 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1086 			resp.tunnel_offloads_caps |=
1087 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1088 	}
1089 
1090 	if (uhw->outlen) {
1091 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1092 
1093 		if (err)
1094 			return err;
1095 	}
1096 
1097 	return 0;
1098 }
1099 
1100 enum mlx5_ib_width {
1101 	MLX5_IB_WIDTH_1X	= 1 << 0,
1102 	MLX5_IB_WIDTH_2X	= 1 << 1,
1103 	MLX5_IB_WIDTH_4X	= 1 << 2,
1104 	MLX5_IB_WIDTH_8X	= 1 << 3,
1105 	MLX5_IB_WIDTH_12X	= 1 << 4
1106 };
1107 
1108 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
1109 				  u8 *ib_width)
1110 {
1111 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1112 	int err = 0;
1113 
1114 	if (active_width & MLX5_IB_WIDTH_1X) {
1115 		*ib_width = IB_WIDTH_1X;
1116 	} else if (active_width & MLX5_IB_WIDTH_2X) {
1117 		mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
1118 			    (int)active_width);
1119 		err = -EINVAL;
1120 	} else if (active_width & MLX5_IB_WIDTH_4X) {
1121 		*ib_width = IB_WIDTH_4X;
1122 	} else if (active_width & MLX5_IB_WIDTH_8X) {
1123 		*ib_width = IB_WIDTH_8X;
1124 	} else if (active_width & MLX5_IB_WIDTH_12X) {
1125 		*ib_width = IB_WIDTH_12X;
1126 	} else {
1127 		mlx5_ib_dbg(dev, "Invalid active_width %d\n",
1128 			    (int)active_width);
1129 		err = -EINVAL;
1130 	}
1131 
1132 	return err;
1133 }
1134 
1135 static int mlx5_mtu_to_ib_mtu(int mtu)
1136 {
1137 	switch (mtu) {
1138 	case 256: return 1;
1139 	case 512: return 2;
1140 	case 1024: return 3;
1141 	case 2048: return 4;
1142 	case 4096: return 5;
1143 	default:
1144 		pr_warn("invalid mtu\n");
1145 		return -1;
1146 	}
1147 }
1148 
1149 enum ib_max_vl_num {
1150 	__IB_MAX_VL_0		= 1,
1151 	__IB_MAX_VL_0_1		= 2,
1152 	__IB_MAX_VL_0_3		= 3,
1153 	__IB_MAX_VL_0_7		= 4,
1154 	__IB_MAX_VL_0_14	= 5,
1155 };
1156 
1157 enum mlx5_vl_hw_cap {
1158 	MLX5_VL_HW_0	= 1,
1159 	MLX5_VL_HW_0_1	= 2,
1160 	MLX5_VL_HW_0_2	= 3,
1161 	MLX5_VL_HW_0_3	= 4,
1162 	MLX5_VL_HW_0_4	= 5,
1163 	MLX5_VL_HW_0_5	= 6,
1164 	MLX5_VL_HW_0_6	= 7,
1165 	MLX5_VL_HW_0_7	= 8,
1166 	MLX5_VL_HW_0_14	= 15
1167 };
1168 
1169 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1170 				u8 *max_vl_num)
1171 {
1172 	switch (vl_hw_cap) {
1173 	case MLX5_VL_HW_0:
1174 		*max_vl_num = __IB_MAX_VL_0;
1175 		break;
1176 	case MLX5_VL_HW_0_1:
1177 		*max_vl_num = __IB_MAX_VL_0_1;
1178 		break;
1179 	case MLX5_VL_HW_0_3:
1180 		*max_vl_num = __IB_MAX_VL_0_3;
1181 		break;
1182 	case MLX5_VL_HW_0_7:
1183 		*max_vl_num = __IB_MAX_VL_0_7;
1184 		break;
1185 	case MLX5_VL_HW_0_14:
1186 		*max_vl_num = __IB_MAX_VL_0_14;
1187 		break;
1188 
1189 	default:
1190 		return -EINVAL;
1191 	}
1192 
1193 	return 0;
1194 }
1195 
1196 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
1197 			       struct ib_port_attr *props)
1198 {
1199 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1200 	struct mlx5_core_dev *mdev = dev->mdev;
1201 	struct mlx5_hca_vport_context *rep;
1202 	u16 max_mtu;
1203 	u16 oper_mtu;
1204 	int err;
1205 	u8 ib_link_width_oper;
1206 	u8 vl_hw_cap;
1207 
1208 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1209 	if (!rep) {
1210 		err = -ENOMEM;
1211 		goto out;
1212 	}
1213 
1214 	/* props being zeroed by the caller, avoid zeroing it here */
1215 
1216 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1217 	if (err)
1218 		goto out;
1219 
1220 	props->lid		= rep->lid;
1221 	props->lmc		= rep->lmc;
1222 	props->sm_lid		= rep->sm_lid;
1223 	props->sm_sl		= rep->sm_sl;
1224 	props->state		= rep->vport_state;
1225 	props->phys_state	= rep->port_physical_state;
1226 	props->port_cap_flags	= rep->cap_mask1;
1227 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1228 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1229 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1230 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1231 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1232 	props->subnet_timeout	= rep->subnet_timeout;
1233 	props->init_type_reply	= rep->init_type_reply;
1234 	props->grh_required	= rep->grh_required;
1235 
1236 	err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
1237 	if (err)
1238 		goto out;
1239 
1240 	err = translate_active_width(ibdev, ib_link_width_oper,
1241 				     &props->active_width);
1242 	if (err)
1243 		goto out;
1244 	err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
1245 	if (err)
1246 		goto out;
1247 
1248 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1249 
1250 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1251 
1252 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1253 
1254 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1255 
1256 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1257 	if (err)
1258 		goto out;
1259 
1260 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1261 				   &props->max_vl_num);
1262 out:
1263 	kfree(rep);
1264 	return err;
1265 }
1266 
1267 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
1268 		       struct ib_port_attr *props)
1269 {
1270 	unsigned int count;
1271 	int ret;
1272 
1273 	switch (mlx5_get_vport_access_method(ibdev)) {
1274 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1275 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1276 		break;
1277 
1278 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1279 		ret = mlx5_query_hca_port(ibdev, port, props);
1280 		break;
1281 
1282 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1283 		ret = mlx5_query_port_roce(ibdev, port, props);
1284 		break;
1285 
1286 	default:
1287 		ret = -EINVAL;
1288 	}
1289 
1290 	if (!ret && props) {
1291 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1292 		struct mlx5_core_dev *mdev;
1293 		bool put_mdev = true;
1294 
1295 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1296 		if (!mdev) {
1297 			/* If the port isn't affiliated yet query the master.
1298 			 * The master and slave will have the same values.
1299 			 */
1300 			mdev = dev->mdev;
1301 			port = 1;
1302 			put_mdev = false;
1303 		}
1304 		count = mlx5_core_reserved_gids_count(mdev);
1305 		if (put_mdev)
1306 			mlx5_ib_put_native_port_mdev(dev, port);
1307 		props->gid_tbl_len -= count;
1308 	}
1309 	return ret;
1310 }
1311 
1312 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u8 port,
1313 				  struct ib_port_attr *props)
1314 {
1315 	int ret;
1316 
1317 	/* Only link layer == ethernet is valid for representors */
1318 	ret = mlx5_query_port_roce(ibdev, port, props);
1319 	if (ret || !props)
1320 		return ret;
1321 
1322 	/* We don't support GIDS */
1323 	props->gid_tbl_len = 0;
1324 
1325 	return ret;
1326 }
1327 
1328 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1329 			     union ib_gid *gid)
1330 {
1331 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1332 	struct mlx5_core_dev *mdev = dev->mdev;
1333 
1334 	switch (mlx5_get_vport_access_method(ibdev)) {
1335 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1336 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1337 
1338 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1339 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1340 
1341 	default:
1342 		return -EINVAL;
1343 	}
1344 
1345 }
1346 
1347 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u8 port,
1348 				   u16 index, u16 *pkey)
1349 {
1350 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1351 	struct mlx5_core_dev *mdev;
1352 	bool put_mdev = true;
1353 	u8 mdev_port_num;
1354 	int err;
1355 
1356 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1357 	if (!mdev) {
1358 		/* The port isn't affiliated yet, get the PKey from the master
1359 		 * port. For RoCE the PKey tables will be the same.
1360 		 */
1361 		put_mdev = false;
1362 		mdev = dev->mdev;
1363 		mdev_port_num = 1;
1364 	}
1365 
1366 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1367 					index, pkey);
1368 	if (put_mdev)
1369 		mlx5_ib_put_native_port_mdev(dev, port);
1370 
1371 	return err;
1372 }
1373 
1374 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1375 			      u16 *pkey)
1376 {
1377 	switch (mlx5_get_vport_access_method(ibdev)) {
1378 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1379 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1380 
1381 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1382 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1383 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1384 	default:
1385 		return -EINVAL;
1386 	}
1387 }
1388 
1389 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1390 				 struct ib_device_modify *props)
1391 {
1392 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1393 	struct mlx5_reg_node_desc in;
1394 	struct mlx5_reg_node_desc out;
1395 	int err;
1396 
1397 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1398 		return -EOPNOTSUPP;
1399 
1400 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1401 		return 0;
1402 
1403 	/*
1404 	 * If possible, pass node desc to FW, so it can generate
1405 	 * a 144 trap.  If cmd fails, just ignore.
1406 	 */
1407 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1408 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1409 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1410 	if (err)
1411 		return err;
1412 
1413 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1414 
1415 	return err;
1416 }
1417 
1418 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1419 				u32 value)
1420 {
1421 	struct mlx5_hca_vport_context ctx = {};
1422 	struct mlx5_core_dev *mdev;
1423 	u8 mdev_port_num;
1424 	int err;
1425 
1426 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1427 	if (!mdev)
1428 		return -ENODEV;
1429 
1430 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1431 	if (err)
1432 		goto out;
1433 
1434 	if (~ctx.cap_mask1_perm & mask) {
1435 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1436 			     mask, ctx.cap_mask1_perm);
1437 		err = -EINVAL;
1438 		goto out;
1439 	}
1440 
1441 	ctx.cap_mask1 = value;
1442 	ctx.cap_mask1_perm = mask;
1443 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1444 						 0, &ctx);
1445 
1446 out:
1447 	mlx5_ib_put_native_port_mdev(dev, port_num);
1448 
1449 	return err;
1450 }
1451 
1452 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1453 			       struct ib_port_modify *props)
1454 {
1455 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1456 	struct ib_port_attr attr;
1457 	u32 tmp;
1458 	int err;
1459 	u32 change_mask;
1460 	u32 value;
1461 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1462 		      IB_LINK_LAYER_INFINIBAND);
1463 
1464 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1465 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1466 	 */
1467 	if (!is_ib)
1468 		return 0;
1469 
1470 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1471 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1472 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1473 		return set_port_caps_atomic(dev, port, change_mask, value);
1474 	}
1475 
1476 	mutex_lock(&dev->cap_mask_mutex);
1477 
1478 	err = ib_query_port(ibdev, port, &attr);
1479 	if (err)
1480 		goto out;
1481 
1482 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1483 		~props->clr_port_cap_mask;
1484 
1485 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1486 
1487 out:
1488 	mutex_unlock(&dev->cap_mask_mutex);
1489 	return err;
1490 }
1491 
1492 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1493 {
1494 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1495 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1496 }
1497 
1498 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1499 {
1500 	/* Large page with non 4k uar support might limit the dynamic size */
1501 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1502 		return MLX5_MIN_DYN_BFREGS;
1503 
1504 	return MLX5_MAX_DYN_BFREGS;
1505 }
1506 
1507 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1508 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1509 			     struct mlx5_bfreg_info *bfregi)
1510 {
1511 	int uars_per_sys_page;
1512 	int bfregs_per_sys_page;
1513 	int ref_bfregs = req->total_num_bfregs;
1514 
1515 	if (req->total_num_bfregs == 0)
1516 		return -EINVAL;
1517 
1518 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1519 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1520 
1521 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1522 		return -ENOMEM;
1523 
1524 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1525 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1526 	/* This holds the required static allocation asked by the user */
1527 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1528 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1529 		return -EINVAL;
1530 
1531 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1532 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1533 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1534 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1535 
1536 	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",
1537 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1538 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1539 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1540 		    bfregi->num_sys_pages);
1541 
1542 	return 0;
1543 }
1544 
1545 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1546 {
1547 	struct mlx5_bfreg_info *bfregi;
1548 	int err;
1549 	int i;
1550 
1551 	bfregi = &context->bfregi;
1552 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1553 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1554 		if (err)
1555 			goto error;
1556 
1557 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1558 	}
1559 
1560 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1561 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1562 
1563 	return 0;
1564 
1565 error:
1566 	for (--i; i >= 0; i--)
1567 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1568 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1569 
1570 	return err;
1571 }
1572 
1573 static int deallocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1574 {
1575 	struct mlx5_bfreg_info *bfregi;
1576 	int err;
1577 	int i;
1578 
1579 	bfregi = &context->bfregi;
1580 	for (i = 0; i < bfregi->num_sys_pages; i++) {
1581 		if (i < bfregi->num_static_sys_pages ||
1582 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX) {
1583 			err = mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1584 			if (err) {
1585 				mlx5_ib_warn(dev, "failed to free uar %d, err=%d\n", i, err);
1586 				return err;
1587 			}
1588 		}
1589 	}
1590 
1591 	return 0;
1592 }
1593 
1594 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn)
1595 {
1596 	int err;
1597 
1598 	err = mlx5_core_alloc_transport_domain(dev->mdev, tdn);
1599 	if (err)
1600 		return err;
1601 
1602 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1603 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1604 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1605 		return err;
1606 
1607 	mutex_lock(&dev->lb_mutex);
1608 	dev->user_td++;
1609 
1610 	if (dev->user_td == 2)
1611 		err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1612 
1613 	mutex_unlock(&dev->lb_mutex);
1614 	return err;
1615 }
1616 
1617 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn)
1618 {
1619 	mlx5_core_dealloc_transport_domain(dev->mdev, tdn);
1620 
1621 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1622 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1623 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1624 		return;
1625 
1626 	mutex_lock(&dev->lb_mutex);
1627 	dev->user_td--;
1628 
1629 	if (dev->user_td < 2)
1630 		mlx5_nic_vport_update_local_lb(dev->mdev, false);
1631 
1632 	mutex_unlock(&dev->lb_mutex);
1633 }
1634 
1635 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
1636 						  struct ib_udata *udata)
1637 {
1638 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1639 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1640 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1641 	struct mlx5_core_dev *mdev = dev->mdev;
1642 	struct mlx5_ib_ucontext *context;
1643 	struct mlx5_bfreg_info *bfregi;
1644 	int ver;
1645 	int err;
1646 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1647 				     max_cqe_version);
1648 	bool lib_uar_4k;
1649 
1650 	if (!dev->ib_active)
1651 		return ERR_PTR(-EAGAIN);
1652 
1653 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1654 		ver = 0;
1655 	else if (udata->inlen >= min_req_v2)
1656 		ver = 2;
1657 	else
1658 		return ERR_PTR(-EINVAL);
1659 
1660 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1661 	if (err)
1662 		return ERR_PTR(err);
1663 
1664 	if (req.flags)
1665 		return ERR_PTR(-EINVAL);
1666 
1667 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1668 		return ERR_PTR(-EOPNOTSUPP);
1669 
1670 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1671 				    MLX5_NON_FP_BFREGS_PER_UAR);
1672 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1673 		return ERR_PTR(-EINVAL);
1674 
1675 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1676 	if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1677 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1678 	resp.cache_line_size = cache_line_size();
1679 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1680 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1681 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1682 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1683 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1684 	resp.cqe_version = min_t(__u8,
1685 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1686 				 req.max_cqe_version);
1687 	resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1688 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1689 	resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1690 					MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1;
1691 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1692 				   sizeof(resp.response_length), udata->outlen);
1693 
1694 	if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1695 		if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_EGRESS))
1696 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1697 		if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1698 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1699 		if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1700 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1701 		if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1702 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1703 		/* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1704 	}
1705 
1706 	context = kzalloc(sizeof(*context), GFP_KERNEL);
1707 	if (!context)
1708 		return ERR_PTR(-ENOMEM);
1709 
1710 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1711 	bfregi = &context->bfregi;
1712 
1713 	/* updates req->total_num_bfregs */
1714 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1715 	if (err)
1716 		goto out_ctx;
1717 
1718 	mutex_init(&bfregi->lock);
1719 	bfregi->lib_uar_4k = lib_uar_4k;
1720 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1721 				GFP_KERNEL);
1722 	if (!bfregi->count) {
1723 		err = -ENOMEM;
1724 		goto out_ctx;
1725 	}
1726 
1727 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1728 				    sizeof(*bfregi->sys_pages),
1729 				    GFP_KERNEL);
1730 	if (!bfregi->sys_pages) {
1731 		err = -ENOMEM;
1732 		goto out_count;
1733 	}
1734 
1735 	err = allocate_uars(dev, context);
1736 	if (err)
1737 		goto out_sys_pages;
1738 
1739 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1740 	context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1741 #endif
1742 
1743 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1744 		err = mlx5_ib_alloc_transport_domain(dev, &context->tdn);
1745 		if (err)
1746 			goto out_uars;
1747 	}
1748 
1749 	INIT_LIST_HEAD(&context->vma_private_list);
1750 	mutex_init(&context->vma_private_list_mutex);
1751 	INIT_LIST_HEAD(&context->db_page_list);
1752 	mutex_init(&context->db_page_mutex);
1753 
1754 	resp.tot_bfregs = req.total_num_bfregs;
1755 	resp.num_ports = dev->num_ports;
1756 
1757 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1758 		resp.response_length += sizeof(resp.cqe_version);
1759 
1760 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1761 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1762 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1763 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1764 	}
1765 
1766 	if (field_avail(typeof(resp), eth_min_inline, udata->outlen)) {
1767 		if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1768 			mlx5_query_min_inline(dev->mdev, &resp.eth_min_inline);
1769 			resp.eth_min_inline++;
1770 		}
1771 		resp.response_length += sizeof(resp.eth_min_inline);
1772 	}
1773 
1774 	if (field_avail(typeof(resp), clock_info_versions, udata->outlen)) {
1775 		if (mdev->clock_info)
1776 			resp.clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1777 		resp.response_length += sizeof(resp.clock_info_versions);
1778 	}
1779 
1780 	/*
1781 	 * We don't want to expose information from the PCI bar that is located
1782 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1783 	 * pretend we don't support reading the HCA's core clock. This is also
1784 	 * forced by mmap function.
1785 	 */
1786 	if (field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1787 		if (PAGE_SIZE <= 4096) {
1788 			resp.comp_mask |=
1789 				MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1790 			resp.hca_core_clock_offset =
1791 				offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1792 		}
1793 		resp.response_length += sizeof(resp.hca_core_clock_offset);
1794 	}
1795 
1796 	if (field_avail(typeof(resp), log_uar_size, udata->outlen))
1797 		resp.response_length += sizeof(resp.log_uar_size);
1798 
1799 	if (field_avail(typeof(resp), num_uars_per_page, udata->outlen))
1800 		resp.response_length += sizeof(resp.num_uars_per_page);
1801 
1802 	if (field_avail(typeof(resp), num_dyn_bfregs, udata->outlen)) {
1803 		resp.num_dyn_bfregs = bfregi->num_dyn_bfregs;
1804 		resp.response_length += sizeof(resp.num_dyn_bfregs);
1805 	}
1806 
1807 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1808 	if (err)
1809 		goto out_td;
1810 
1811 	bfregi->ver = ver;
1812 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1813 	context->cqe_version = resp.cqe_version;
1814 	context->lib_caps = req.lib_caps;
1815 	print_lib_caps(dev, context->lib_caps);
1816 
1817 	return &context->ibucontext;
1818 
1819 out_td:
1820 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1821 		mlx5_ib_dealloc_transport_domain(dev, context->tdn);
1822 
1823 out_uars:
1824 	deallocate_uars(dev, context);
1825 
1826 out_sys_pages:
1827 	kfree(bfregi->sys_pages);
1828 
1829 out_count:
1830 	kfree(bfregi->count);
1831 
1832 out_ctx:
1833 	kfree(context);
1834 
1835 	return ERR_PTR(err);
1836 }
1837 
1838 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1839 {
1840 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1841 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1842 	struct mlx5_bfreg_info *bfregi;
1843 
1844 	bfregi = &context->bfregi;
1845 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1846 		mlx5_ib_dealloc_transport_domain(dev, context->tdn);
1847 
1848 	deallocate_uars(dev, context);
1849 	kfree(bfregi->sys_pages);
1850 	kfree(bfregi->count);
1851 	kfree(context);
1852 
1853 	return 0;
1854 }
1855 
1856 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
1857 				 int uar_idx)
1858 {
1859 	int fw_uars_per_page;
1860 
1861 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
1862 
1863 	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
1864 }
1865 
1866 static int get_command(unsigned long offset)
1867 {
1868 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1869 }
1870 
1871 static int get_arg(unsigned long offset)
1872 {
1873 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1874 }
1875 
1876 static int get_index(unsigned long offset)
1877 {
1878 	return get_arg(offset);
1879 }
1880 
1881 /* Index resides in an extra byte to enable larger values than 255 */
1882 static int get_extended_index(unsigned long offset)
1883 {
1884 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
1885 }
1886 
1887 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1888 {
1889 	/* vma_open is called when a new VMA is created on top of our VMA.  This
1890 	 * is done through either mremap flow or split_vma (usually due to
1891 	 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1892 	 * as this VMA is strongly hardware related.  Therefore we set the
1893 	 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1894 	 * calling us again and trying to do incorrect actions.  We assume that
1895 	 * the original VMA size is exactly a single page, and therefore all
1896 	 * "splitting" operation will not happen to it.
1897 	 */
1898 	area->vm_ops = NULL;
1899 }
1900 
1901 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1902 {
1903 	struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1904 
1905 	/* It's guaranteed that all VMAs opened on a FD are closed before the
1906 	 * file itself is closed, therefore no sync is needed with the regular
1907 	 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1908 	 * However need a sync with accessing the vma as part of
1909 	 * mlx5_ib_disassociate_ucontext.
1910 	 * The close operation is usually called under mm->mmap_sem except when
1911 	 * process is exiting.
1912 	 * The exiting case is handled explicitly as part of
1913 	 * mlx5_ib_disassociate_ucontext.
1914 	 */
1915 	mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1916 
1917 	/* setting the vma context pointer to null in the mlx5_ib driver's
1918 	 * private data, to protect a race condition in
1919 	 * mlx5_ib_disassociate_ucontext().
1920 	 */
1921 	mlx5_ib_vma_priv_data->vma = NULL;
1922 	mutex_lock(mlx5_ib_vma_priv_data->vma_private_list_mutex);
1923 	list_del(&mlx5_ib_vma_priv_data->list);
1924 	mutex_unlock(mlx5_ib_vma_priv_data->vma_private_list_mutex);
1925 	kfree(mlx5_ib_vma_priv_data);
1926 }
1927 
1928 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1929 	.open = mlx5_ib_vma_open,
1930 	.close = mlx5_ib_vma_close
1931 };
1932 
1933 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1934 				struct mlx5_ib_ucontext *ctx)
1935 {
1936 	struct mlx5_ib_vma_private_data *vma_prv;
1937 	struct list_head *vma_head = &ctx->vma_private_list;
1938 
1939 	vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1940 	if (!vma_prv)
1941 		return -ENOMEM;
1942 
1943 	vma_prv->vma = vma;
1944 	vma_prv->vma_private_list_mutex = &ctx->vma_private_list_mutex;
1945 	vma->vm_private_data = vma_prv;
1946 	vma->vm_ops =  &mlx5_ib_vm_ops;
1947 
1948 	mutex_lock(&ctx->vma_private_list_mutex);
1949 	list_add(&vma_prv->list, vma_head);
1950 	mutex_unlock(&ctx->vma_private_list_mutex);
1951 
1952 	return 0;
1953 }
1954 
1955 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1956 {
1957 	int ret;
1958 	struct vm_area_struct *vma;
1959 	struct mlx5_ib_vma_private_data *vma_private, *n;
1960 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1961 	struct task_struct *owning_process  = NULL;
1962 	struct mm_struct   *owning_mm       = NULL;
1963 
1964 	owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1965 	if (!owning_process)
1966 		return;
1967 
1968 	owning_mm = get_task_mm(owning_process);
1969 	if (!owning_mm) {
1970 		pr_info("no mm, disassociate ucontext is pending task termination\n");
1971 		while (1) {
1972 			put_task_struct(owning_process);
1973 			usleep_range(1000, 2000);
1974 			owning_process = get_pid_task(ibcontext->tgid,
1975 						      PIDTYPE_PID);
1976 			if (!owning_process ||
1977 			    owning_process->state == TASK_DEAD) {
1978 				pr_info("disassociate ucontext done, task was terminated\n");
1979 				/* in case task was dead need to release the
1980 				 * task struct.
1981 				 */
1982 				if (owning_process)
1983 					put_task_struct(owning_process);
1984 				return;
1985 			}
1986 		}
1987 	}
1988 
1989 	/* need to protect from a race on closing the vma as part of
1990 	 * mlx5_ib_vma_close.
1991 	 */
1992 	down_write(&owning_mm->mmap_sem);
1993 	mutex_lock(&context->vma_private_list_mutex);
1994 	list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1995 				 list) {
1996 		vma = vma_private->vma;
1997 		ret = zap_vma_ptes(vma, vma->vm_start,
1998 				   PAGE_SIZE);
1999 		WARN_ONCE(ret, "%s: zap_vma_ptes failed", __func__);
2000 		/* context going to be destroyed, should
2001 		 * not access ops any more.
2002 		 */
2003 		vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
2004 		vma->vm_ops = NULL;
2005 		list_del(&vma_private->list);
2006 		kfree(vma_private);
2007 	}
2008 	mutex_unlock(&context->vma_private_list_mutex);
2009 	up_write(&owning_mm->mmap_sem);
2010 	mmput(owning_mm);
2011 	put_task_struct(owning_process);
2012 }
2013 
2014 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2015 {
2016 	switch (cmd) {
2017 	case MLX5_IB_MMAP_WC_PAGE:
2018 		return "WC";
2019 	case MLX5_IB_MMAP_REGULAR_PAGE:
2020 		return "best effort WC";
2021 	case MLX5_IB_MMAP_NC_PAGE:
2022 		return "NC";
2023 	case MLX5_IB_MMAP_DEVICE_MEM:
2024 		return "Device Memory";
2025 	default:
2026 		return NULL;
2027 	}
2028 }
2029 
2030 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2031 					struct vm_area_struct *vma,
2032 					struct mlx5_ib_ucontext *context)
2033 {
2034 	phys_addr_t pfn;
2035 	int err;
2036 
2037 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2038 		return -EINVAL;
2039 
2040 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2041 		return -EOPNOTSUPP;
2042 
2043 	if (vma->vm_flags & VM_WRITE)
2044 		return -EPERM;
2045 
2046 	if (!dev->mdev->clock_info_page)
2047 		return -EOPNOTSUPP;
2048 
2049 	pfn = page_to_pfn(dev->mdev->clock_info_page);
2050 	err = remap_pfn_range(vma, vma->vm_start, pfn, PAGE_SIZE,
2051 			      vma->vm_page_prot);
2052 	if (err)
2053 		return err;
2054 
2055 	mlx5_ib_dbg(dev, "mapped clock info at 0x%lx, PA 0x%llx\n",
2056 		    vma->vm_start,
2057 		    (unsigned long long)pfn << PAGE_SHIFT);
2058 
2059 	return mlx5_ib_set_vma_data(vma, context);
2060 }
2061 
2062 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2063 		    struct vm_area_struct *vma,
2064 		    struct mlx5_ib_ucontext *context)
2065 {
2066 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2067 	int err;
2068 	unsigned long idx;
2069 	phys_addr_t pfn, pa;
2070 	pgprot_t prot;
2071 	u32 bfreg_dyn_idx = 0;
2072 	u32 uar_index;
2073 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2074 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2075 				bfregi->num_static_sys_pages;
2076 
2077 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2078 		return -EINVAL;
2079 
2080 	if (dyn_uar)
2081 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2082 	else
2083 		idx = get_index(vma->vm_pgoff);
2084 
2085 	if (idx >= max_valid_idx) {
2086 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2087 			     idx, max_valid_idx);
2088 		return -EINVAL;
2089 	}
2090 
2091 	switch (cmd) {
2092 	case MLX5_IB_MMAP_WC_PAGE:
2093 	case MLX5_IB_MMAP_ALLOC_WC:
2094 /* Some architectures don't support WC memory */
2095 #if defined(CONFIG_X86)
2096 		if (!pat_enabled())
2097 			return -EPERM;
2098 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
2099 			return -EPERM;
2100 #endif
2101 	/* fall through */
2102 	case MLX5_IB_MMAP_REGULAR_PAGE:
2103 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2104 		prot = pgprot_writecombine(vma->vm_page_prot);
2105 		break;
2106 	case MLX5_IB_MMAP_NC_PAGE:
2107 		prot = pgprot_noncached(vma->vm_page_prot);
2108 		break;
2109 	default:
2110 		return -EINVAL;
2111 	}
2112 
2113 	if (dyn_uar) {
2114 		int uars_per_page;
2115 
2116 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2117 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2118 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2119 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2120 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2121 			return -EINVAL;
2122 		}
2123 
2124 		mutex_lock(&bfregi->lock);
2125 		/* Fail if uar already allocated, first bfreg index of each
2126 		 * page holds its count.
2127 		 */
2128 		if (bfregi->count[bfreg_dyn_idx]) {
2129 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2130 			mutex_unlock(&bfregi->lock);
2131 			return -EINVAL;
2132 		}
2133 
2134 		bfregi->count[bfreg_dyn_idx]++;
2135 		mutex_unlock(&bfregi->lock);
2136 
2137 		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2138 		if (err) {
2139 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2140 			goto free_bfreg;
2141 		}
2142 	} else {
2143 		uar_index = bfregi->sys_pages[idx];
2144 	}
2145 
2146 	pfn = uar_index2pfn(dev, uar_index);
2147 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2148 
2149 	vma->vm_page_prot = prot;
2150 	err = io_remap_pfn_range(vma, vma->vm_start, pfn,
2151 				 PAGE_SIZE, vma->vm_page_prot);
2152 	if (err) {
2153 		mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%lx, pfn=%pa, mmap_cmd=%s\n",
2154 			    err, vma->vm_start, &pfn, mmap_cmd2str(cmd));
2155 		err = -EAGAIN;
2156 		goto err;
2157 	}
2158 
2159 	pa = pfn << PAGE_SHIFT;
2160 	mlx5_ib_dbg(dev, "mapped %s at 0x%lx, PA %pa\n", mmap_cmd2str(cmd),
2161 		    vma->vm_start, &pa);
2162 
2163 	err = mlx5_ib_set_vma_data(vma, context);
2164 	if (err)
2165 		goto err;
2166 
2167 	if (dyn_uar)
2168 		bfregi->sys_pages[idx] = uar_index;
2169 	return 0;
2170 
2171 err:
2172 	if (!dyn_uar)
2173 		return err;
2174 
2175 	mlx5_cmd_free_uar(dev->mdev, idx);
2176 
2177 free_bfreg:
2178 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2179 
2180 	return err;
2181 }
2182 
2183 static int dm_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
2184 {
2185 	struct mlx5_ib_ucontext *mctx = to_mucontext(context);
2186 	struct mlx5_ib_dev *dev = to_mdev(context->device);
2187 	u16 page_idx = get_extended_index(vma->vm_pgoff);
2188 	size_t map_size = vma->vm_end - vma->vm_start;
2189 	u32 npages = map_size >> PAGE_SHIFT;
2190 	phys_addr_t pfn;
2191 	pgprot_t prot;
2192 
2193 	if (find_next_zero_bit(mctx->dm_pages, page_idx + npages, page_idx) !=
2194 	    page_idx + npages)
2195 		return -EINVAL;
2196 
2197 	pfn = ((pci_resource_start(dev->mdev->pdev, 0) +
2198 	      MLX5_CAP64_DEV_MEM(dev->mdev, memic_bar_start_addr)) >>
2199 	      PAGE_SHIFT) +
2200 	      page_idx;
2201 	prot = pgprot_writecombine(vma->vm_page_prot);
2202 	vma->vm_page_prot = prot;
2203 
2204 	if (io_remap_pfn_range(vma, vma->vm_start, pfn, map_size,
2205 			       vma->vm_page_prot))
2206 		return -EAGAIN;
2207 
2208 	return mlx5_ib_set_vma_data(vma, mctx);
2209 }
2210 
2211 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2212 {
2213 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2214 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2215 	unsigned long command;
2216 	phys_addr_t pfn;
2217 
2218 	command = get_command(vma->vm_pgoff);
2219 	switch (command) {
2220 	case MLX5_IB_MMAP_WC_PAGE:
2221 	case MLX5_IB_MMAP_NC_PAGE:
2222 	case MLX5_IB_MMAP_REGULAR_PAGE:
2223 	case MLX5_IB_MMAP_ALLOC_WC:
2224 		return uar_mmap(dev, command, vma, context);
2225 
2226 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2227 		return -ENOSYS;
2228 
2229 	case MLX5_IB_MMAP_CORE_CLOCK:
2230 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2231 			return -EINVAL;
2232 
2233 		if (vma->vm_flags & VM_WRITE)
2234 			return -EPERM;
2235 
2236 		/* Don't expose to user-space information it shouldn't have */
2237 		if (PAGE_SIZE > 4096)
2238 			return -EOPNOTSUPP;
2239 
2240 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
2241 		pfn = (dev->mdev->iseg_base +
2242 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2243 			PAGE_SHIFT;
2244 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
2245 				       PAGE_SIZE, vma->vm_page_prot))
2246 			return -EAGAIN;
2247 
2248 		mlx5_ib_dbg(dev, "mapped internal timer at 0x%lx, PA 0x%llx\n",
2249 			    vma->vm_start,
2250 			    (unsigned long long)pfn << PAGE_SHIFT);
2251 		break;
2252 	case MLX5_IB_MMAP_CLOCK_INFO:
2253 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2254 
2255 	case MLX5_IB_MMAP_DEVICE_MEM:
2256 		return dm_mmap(ibcontext, vma);
2257 
2258 	default:
2259 		return -EINVAL;
2260 	}
2261 
2262 	return 0;
2263 }
2264 
2265 struct ib_dm *mlx5_ib_alloc_dm(struct ib_device *ibdev,
2266 			       struct ib_ucontext *context,
2267 			       struct ib_dm_alloc_attr *attr,
2268 			       struct uverbs_attr_bundle *attrs)
2269 {
2270 	u64 act_size = roundup(attr->length, MLX5_MEMIC_BASE_SIZE);
2271 	struct mlx5_memic *memic = &to_mdev(ibdev)->memic;
2272 	phys_addr_t memic_addr;
2273 	struct mlx5_ib_dm *dm;
2274 	u64 start_offset;
2275 	u32 page_idx;
2276 	int err;
2277 
2278 	dm = kzalloc(sizeof(*dm), GFP_KERNEL);
2279 	if (!dm)
2280 		return ERR_PTR(-ENOMEM);
2281 
2282 	mlx5_ib_dbg(to_mdev(ibdev), "alloc_memic req: user_length=0x%llx act_length=0x%llx log_alignment=%d\n",
2283 		    attr->length, act_size, attr->alignment);
2284 
2285 	err = mlx5_cmd_alloc_memic(memic, &memic_addr,
2286 				   act_size, attr->alignment);
2287 	if (err)
2288 		goto err_free;
2289 
2290 	start_offset = memic_addr & ~PAGE_MASK;
2291 	page_idx = (memic_addr - pci_resource_start(memic->dev->pdev, 0) -
2292 		    MLX5_CAP64_DEV_MEM(memic->dev, memic_bar_start_addr)) >>
2293 		    PAGE_SHIFT;
2294 
2295 	err = uverbs_copy_to(attrs,
2296 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2297 			     &start_offset, sizeof(start_offset));
2298 	if (err)
2299 		goto err_dealloc;
2300 
2301 	err = uverbs_copy_to(attrs,
2302 			     MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
2303 			     &page_idx, sizeof(page_idx));
2304 	if (err)
2305 		goto err_dealloc;
2306 
2307 	bitmap_set(to_mucontext(context)->dm_pages, page_idx,
2308 		   DIV_ROUND_UP(act_size, PAGE_SIZE));
2309 
2310 	dm->dev_addr = memic_addr;
2311 
2312 	return &dm->ibdm;
2313 
2314 err_dealloc:
2315 	mlx5_cmd_dealloc_memic(memic, memic_addr,
2316 			       act_size);
2317 err_free:
2318 	kfree(dm);
2319 	return ERR_PTR(err);
2320 }
2321 
2322 int mlx5_ib_dealloc_dm(struct ib_dm *ibdm)
2323 {
2324 	struct mlx5_memic *memic = &to_mdev(ibdm->device)->memic;
2325 	struct mlx5_ib_dm *dm = to_mdm(ibdm);
2326 	u64 act_size = roundup(dm->ibdm.length, MLX5_MEMIC_BASE_SIZE);
2327 	u32 page_idx;
2328 	int ret;
2329 
2330 	ret = mlx5_cmd_dealloc_memic(memic, dm->dev_addr, act_size);
2331 	if (ret)
2332 		return ret;
2333 
2334 	page_idx = (dm->dev_addr - pci_resource_start(memic->dev->pdev, 0) -
2335 		    MLX5_CAP64_DEV_MEM(memic->dev, memic_bar_start_addr)) >>
2336 		    PAGE_SHIFT;
2337 	bitmap_clear(to_mucontext(ibdm->uobject->context)->dm_pages,
2338 		     page_idx,
2339 		     DIV_ROUND_UP(act_size, PAGE_SIZE));
2340 
2341 	kfree(dm);
2342 
2343 	return 0;
2344 }
2345 
2346 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
2347 				      struct ib_ucontext *context,
2348 				      struct ib_udata *udata)
2349 {
2350 	struct mlx5_ib_alloc_pd_resp resp;
2351 	struct mlx5_ib_pd *pd;
2352 	int err;
2353 
2354 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
2355 	if (!pd)
2356 		return ERR_PTR(-ENOMEM);
2357 
2358 	err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
2359 	if (err) {
2360 		kfree(pd);
2361 		return ERR_PTR(err);
2362 	}
2363 
2364 	if (context) {
2365 		resp.pdn = pd->pdn;
2366 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2367 			mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
2368 			kfree(pd);
2369 			return ERR_PTR(-EFAULT);
2370 		}
2371 	}
2372 
2373 	return &pd->ibpd;
2374 }
2375 
2376 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
2377 {
2378 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2379 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2380 
2381 	mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
2382 	kfree(mpd);
2383 
2384 	return 0;
2385 }
2386 
2387 enum {
2388 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
2389 	MATCH_CRITERIA_ENABLE_MISC_BIT,
2390 	MATCH_CRITERIA_ENABLE_INNER_BIT
2391 };
2392 
2393 #define HEADER_IS_ZERO(match_criteria, headers)			           \
2394 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
2395 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
2396 
2397 static u8 get_match_criteria_enable(u32 *match_criteria)
2398 {
2399 	u8 match_criteria_enable;
2400 
2401 	match_criteria_enable =
2402 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
2403 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
2404 	match_criteria_enable |=
2405 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
2406 		MATCH_CRITERIA_ENABLE_MISC_BIT;
2407 	match_criteria_enable |=
2408 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
2409 		MATCH_CRITERIA_ENABLE_INNER_BIT;
2410 
2411 	return match_criteria_enable;
2412 }
2413 
2414 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
2415 {
2416 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
2417 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
2418 }
2419 
2420 static void set_flow_label(void *misc_c, void *misc_v, u8 mask, u8 val,
2421 			   bool inner)
2422 {
2423 	if (inner) {
2424 		MLX5_SET(fte_match_set_misc,
2425 			 misc_c, inner_ipv6_flow_label, mask);
2426 		MLX5_SET(fte_match_set_misc,
2427 			 misc_v, inner_ipv6_flow_label, val);
2428 	} else {
2429 		MLX5_SET(fte_match_set_misc,
2430 			 misc_c, outer_ipv6_flow_label, mask);
2431 		MLX5_SET(fte_match_set_misc,
2432 			 misc_v, outer_ipv6_flow_label, val);
2433 	}
2434 }
2435 
2436 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
2437 {
2438 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
2439 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
2440 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
2441 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
2442 }
2443 
2444 #define LAST_ETH_FIELD vlan_tag
2445 #define LAST_IB_FIELD sl
2446 #define LAST_IPV4_FIELD tos
2447 #define LAST_IPV6_FIELD traffic_class
2448 #define LAST_TCP_UDP_FIELD src_port
2449 #define LAST_TUNNEL_FIELD tunnel_id
2450 #define LAST_FLOW_TAG_FIELD tag_id
2451 #define LAST_DROP_FIELD size
2452 
2453 /* Field is the last supported field */
2454 #define FIELDS_NOT_SUPPORTED(filter, field)\
2455 	memchr_inv((void *)&filter.field  +\
2456 		   sizeof(filter.field), 0,\
2457 		   sizeof(filter) -\
2458 		   offsetof(typeof(filter), field) -\
2459 		   sizeof(filter.field))
2460 
2461 static int parse_flow_flow_action(const union ib_flow_spec *ib_spec,
2462 				  const struct ib_flow_attr *flow_attr,
2463 				  struct mlx5_flow_act *action)
2464 {
2465 	struct mlx5_ib_flow_action *maction = to_mflow_act(ib_spec->action.act);
2466 
2467 	switch (maction->ib_action.type) {
2468 	case IB_FLOW_ACTION_ESP:
2469 		/* Currently only AES_GCM keymat is supported by the driver */
2470 		action->esp_id = (uintptr_t)maction->esp_aes_gcm.ctx;
2471 		action->action |= flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS ?
2472 			MLX5_FLOW_CONTEXT_ACTION_ENCRYPT :
2473 			MLX5_FLOW_CONTEXT_ACTION_DECRYPT;
2474 		return 0;
2475 	default:
2476 		return -EOPNOTSUPP;
2477 	}
2478 }
2479 
2480 static int parse_flow_attr(struct mlx5_core_dev *mdev, u32 *match_c,
2481 			   u32 *match_v, const union ib_flow_spec *ib_spec,
2482 			   const struct ib_flow_attr *flow_attr,
2483 			   struct mlx5_flow_act *action)
2484 {
2485 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
2486 					   misc_parameters);
2487 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
2488 					   misc_parameters);
2489 	void *headers_c;
2490 	void *headers_v;
2491 	int match_ipv;
2492 	int ret;
2493 
2494 	if (ib_spec->type & IB_FLOW_SPEC_INNER) {
2495 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2496 					 inner_headers);
2497 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2498 					 inner_headers);
2499 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2500 					ft_field_support.inner_ip_version);
2501 	} else {
2502 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2503 					 outer_headers);
2504 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2505 					 outer_headers);
2506 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2507 					ft_field_support.outer_ip_version);
2508 	}
2509 
2510 	switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
2511 	case IB_FLOW_SPEC_ETH:
2512 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
2513 			return -EOPNOTSUPP;
2514 
2515 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2516 					     dmac_47_16),
2517 				ib_spec->eth.mask.dst_mac);
2518 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2519 					     dmac_47_16),
2520 				ib_spec->eth.val.dst_mac);
2521 
2522 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2523 					     smac_47_16),
2524 				ib_spec->eth.mask.src_mac);
2525 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2526 					     smac_47_16),
2527 				ib_spec->eth.val.src_mac);
2528 
2529 		if (ib_spec->eth.mask.vlan_tag) {
2530 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2531 				 cvlan_tag, 1);
2532 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2533 				 cvlan_tag, 1);
2534 
2535 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2536 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
2537 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2538 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
2539 
2540 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2541 				 first_cfi,
2542 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
2543 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2544 				 first_cfi,
2545 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
2546 
2547 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2548 				 first_prio,
2549 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
2550 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2551 				 first_prio,
2552 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
2553 		}
2554 		MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2555 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
2556 		MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2557 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
2558 		break;
2559 	case IB_FLOW_SPEC_IPV4:
2560 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
2561 			return -EOPNOTSUPP;
2562 
2563 		if (match_ipv) {
2564 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2565 				 ip_version, 0xf);
2566 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2567 				 ip_version, MLX5_FS_IPV4_VERSION);
2568 		} else {
2569 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2570 				 ethertype, 0xffff);
2571 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2572 				 ethertype, ETH_P_IP);
2573 		}
2574 
2575 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2576 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
2577 		       &ib_spec->ipv4.mask.src_ip,
2578 		       sizeof(ib_spec->ipv4.mask.src_ip));
2579 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2580 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
2581 		       &ib_spec->ipv4.val.src_ip,
2582 		       sizeof(ib_spec->ipv4.val.src_ip));
2583 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2584 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2585 		       &ib_spec->ipv4.mask.dst_ip,
2586 		       sizeof(ib_spec->ipv4.mask.dst_ip));
2587 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2588 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2589 		       &ib_spec->ipv4.val.dst_ip,
2590 		       sizeof(ib_spec->ipv4.val.dst_ip));
2591 
2592 		set_tos(headers_c, headers_v,
2593 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
2594 
2595 		set_proto(headers_c, headers_v,
2596 			  ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
2597 		break;
2598 	case IB_FLOW_SPEC_IPV6:
2599 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
2600 			return -EOPNOTSUPP;
2601 
2602 		if (match_ipv) {
2603 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2604 				 ip_version, 0xf);
2605 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2606 				 ip_version, MLX5_FS_IPV6_VERSION);
2607 		} else {
2608 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2609 				 ethertype, 0xffff);
2610 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2611 				 ethertype, ETH_P_IPV6);
2612 		}
2613 
2614 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2615 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
2616 		       &ib_spec->ipv6.mask.src_ip,
2617 		       sizeof(ib_spec->ipv6.mask.src_ip));
2618 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2619 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
2620 		       &ib_spec->ipv6.val.src_ip,
2621 		       sizeof(ib_spec->ipv6.val.src_ip));
2622 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2623 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2624 		       &ib_spec->ipv6.mask.dst_ip,
2625 		       sizeof(ib_spec->ipv6.mask.dst_ip));
2626 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2627 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2628 		       &ib_spec->ipv6.val.dst_ip,
2629 		       sizeof(ib_spec->ipv6.val.dst_ip));
2630 
2631 		set_tos(headers_c, headers_v,
2632 			ib_spec->ipv6.mask.traffic_class,
2633 			ib_spec->ipv6.val.traffic_class);
2634 
2635 		set_proto(headers_c, headers_v,
2636 			  ib_spec->ipv6.mask.next_hdr,
2637 			  ib_spec->ipv6.val.next_hdr);
2638 
2639 		set_flow_label(misc_params_c, misc_params_v,
2640 			       ntohl(ib_spec->ipv6.mask.flow_label),
2641 			       ntohl(ib_spec->ipv6.val.flow_label),
2642 			       ib_spec->type & IB_FLOW_SPEC_INNER);
2643 		break;
2644 	case IB_FLOW_SPEC_ESP:
2645 		if (ib_spec->esp.mask.seq)
2646 			return -EOPNOTSUPP;
2647 
2648 		MLX5_SET(fte_match_set_misc, misc_params_c, outer_esp_spi,
2649 			 ntohl(ib_spec->esp.mask.spi));
2650 		MLX5_SET(fte_match_set_misc, misc_params_v, outer_esp_spi,
2651 			 ntohl(ib_spec->esp.val.spi));
2652 		break;
2653 	case IB_FLOW_SPEC_TCP:
2654 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2655 					 LAST_TCP_UDP_FIELD))
2656 			return -EOPNOTSUPP;
2657 
2658 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2659 			 0xff);
2660 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2661 			 IPPROTO_TCP);
2662 
2663 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_sport,
2664 			 ntohs(ib_spec->tcp_udp.mask.src_port));
2665 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_sport,
2666 			 ntohs(ib_spec->tcp_udp.val.src_port));
2667 
2668 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_dport,
2669 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
2670 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_dport,
2671 			 ntohs(ib_spec->tcp_udp.val.dst_port));
2672 		break;
2673 	case IB_FLOW_SPEC_UDP:
2674 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2675 					 LAST_TCP_UDP_FIELD))
2676 			return -EOPNOTSUPP;
2677 
2678 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2679 			 0xff);
2680 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2681 			 IPPROTO_UDP);
2682 
2683 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_sport,
2684 			 ntohs(ib_spec->tcp_udp.mask.src_port));
2685 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_sport,
2686 			 ntohs(ib_spec->tcp_udp.val.src_port));
2687 
2688 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_dport,
2689 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
2690 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_dport,
2691 			 ntohs(ib_spec->tcp_udp.val.dst_port));
2692 		break;
2693 	case IB_FLOW_SPEC_VXLAN_TUNNEL:
2694 		if (FIELDS_NOT_SUPPORTED(ib_spec->tunnel.mask,
2695 					 LAST_TUNNEL_FIELD))
2696 			return -EOPNOTSUPP;
2697 
2698 		MLX5_SET(fte_match_set_misc, misc_params_c, vxlan_vni,
2699 			 ntohl(ib_spec->tunnel.mask.tunnel_id));
2700 		MLX5_SET(fte_match_set_misc, misc_params_v, vxlan_vni,
2701 			 ntohl(ib_spec->tunnel.val.tunnel_id));
2702 		break;
2703 	case IB_FLOW_SPEC_ACTION_TAG:
2704 		if (FIELDS_NOT_SUPPORTED(ib_spec->flow_tag,
2705 					 LAST_FLOW_TAG_FIELD))
2706 			return -EOPNOTSUPP;
2707 		if (ib_spec->flow_tag.tag_id >= BIT(24))
2708 			return -EINVAL;
2709 
2710 		action->flow_tag = ib_spec->flow_tag.tag_id;
2711 		action->has_flow_tag = true;
2712 		break;
2713 	case IB_FLOW_SPEC_ACTION_DROP:
2714 		if (FIELDS_NOT_SUPPORTED(ib_spec->drop,
2715 					 LAST_DROP_FIELD))
2716 			return -EOPNOTSUPP;
2717 		action->action |= MLX5_FLOW_CONTEXT_ACTION_DROP;
2718 		break;
2719 	case IB_FLOW_SPEC_ACTION_HANDLE:
2720 		ret = parse_flow_flow_action(ib_spec, flow_attr, action);
2721 		if (ret)
2722 			return ret;
2723 		break;
2724 	default:
2725 		return -EINVAL;
2726 	}
2727 
2728 	return 0;
2729 }
2730 
2731 /* If a flow could catch both multicast and unicast packets,
2732  * it won't fall into the multicast flow steering table and this rule
2733  * could steal other multicast packets.
2734  */
2735 static bool flow_is_multicast_only(const struct ib_flow_attr *ib_attr)
2736 {
2737 	union ib_flow_spec *flow_spec;
2738 
2739 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
2740 	    ib_attr->num_of_specs < 1)
2741 		return false;
2742 
2743 	flow_spec = (union ib_flow_spec *)(ib_attr + 1);
2744 	if (flow_spec->type == IB_FLOW_SPEC_IPV4) {
2745 		struct ib_flow_spec_ipv4 *ipv4_spec;
2746 
2747 		ipv4_spec = (struct ib_flow_spec_ipv4 *)flow_spec;
2748 		if (ipv4_is_multicast(ipv4_spec->val.dst_ip))
2749 			return true;
2750 
2751 		return false;
2752 	}
2753 
2754 	if (flow_spec->type == IB_FLOW_SPEC_ETH) {
2755 		struct ib_flow_spec_eth *eth_spec;
2756 
2757 		eth_spec = (struct ib_flow_spec_eth *)flow_spec;
2758 		return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
2759 		       is_multicast_ether_addr(eth_spec->val.dst_mac);
2760 	}
2761 
2762 	return false;
2763 }
2764 
2765 enum valid_spec {
2766 	VALID_SPEC_INVALID,
2767 	VALID_SPEC_VALID,
2768 	VALID_SPEC_NA,
2769 };
2770 
2771 static enum valid_spec
2772 is_valid_esp_aes_gcm(struct mlx5_core_dev *mdev,
2773 		     const struct mlx5_flow_spec *spec,
2774 		     const struct mlx5_flow_act *flow_act,
2775 		     bool egress)
2776 {
2777 	const u32 *match_c = spec->match_criteria;
2778 	bool is_crypto =
2779 		(flow_act->action & (MLX5_FLOW_CONTEXT_ACTION_ENCRYPT |
2780 				     MLX5_FLOW_CONTEXT_ACTION_DECRYPT));
2781 	bool is_ipsec = mlx5_fs_is_ipsec_flow(match_c);
2782 	bool is_drop = flow_act->action & MLX5_FLOW_CONTEXT_ACTION_DROP;
2783 
2784 	/*
2785 	 * Currently only crypto is supported in egress, when regular egress
2786 	 * rules would be supported, always return VALID_SPEC_NA.
2787 	 */
2788 	if (!is_crypto)
2789 		return egress ? VALID_SPEC_INVALID : VALID_SPEC_NA;
2790 
2791 	return is_crypto && is_ipsec &&
2792 		(!egress || (!is_drop && !flow_act->has_flow_tag)) ?
2793 		VALID_SPEC_VALID : VALID_SPEC_INVALID;
2794 }
2795 
2796 static bool is_valid_spec(struct mlx5_core_dev *mdev,
2797 			  const struct mlx5_flow_spec *spec,
2798 			  const struct mlx5_flow_act *flow_act,
2799 			  bool egress)
2800 {
2801 	/* We curretly only support ipsec egress flow */
2802 	return is_valid_esp_aes_gcm(mdev, spec, flow_act, egress) != VALID_SPEC_INVALID;
2803 }
2804 
2805 static bool is_valid_ethertype(struct mlx5_core_dev *mdev,
2806 			       const struct ib_flow_attr *flow_attr,
2807 			       bool check_inner)
2808 {
2809 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
2810 	int match_ipv = check_inner ?
2811 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2812 					ft_field_support.inner_ip_version) :
2813 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2814 					ft_field_support.outer_ip_version);
2815 	int inner_bit = check_inner ? IB_FLOW_SPEC_INNER : 0;
2816 	bool ipv4_spec_valid, ipv6_spec_valid;
2817 	unsigned int ip_spec_type = 0;
2818 	bool has_ethertype = false;
2819 	unsigned int spec_index;
2820 	bool mask_valid = true;
2821 	u16 eth_type = 0;
2822 	bool type_valid;
2823 
2824 	/* Validate that ethertype is correct */
2825 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2826 		if ((ib_spec->type == (IB_FLOW_SPEC_ETH | inner_bit)) &&
2827 		    ib_spec->eth.mask.ether_type) {
2828 			mask_valid = (ib_spec->eth.mask.ether_type ==
2829 				      htons(0xffff));
2830 			has_ethertype = true;
2831 			eth_type = ntohs(ib_spec->eth.val.ether_type);
2832 		} else if ((ib_spec->type == (IB_FLOW_SPEC_IPV4 | inner_bit)) ||
2833 			   (ib_spec->type == (IB_FLOW_SPEC_IPV6 | inner_bit))) {
2834 			ip_spec_type = ib_spec->type;
2835 		}
2836 		ib_spec = (void *)ib_spec + ib_spec->size;
2837 	}
2838 
2839 	type_valid = (!has_ethertype) || (!ip_spec_type);
2840 	if (!type_valid && mask_valid) {
2841 		ipv4_spec_valid = (eth_type == ETH_P_IP) &&
2842 			(ip_spec_type == (IB_FLOW_SPEC_IPV4 | inner_bit));
2843 		ipv6_spec_valid = (eth_type == ETH_P_IPV6) &&
2844 			(ip_spec_type == (IB_FLOW_SPEC_IPV6 | inner_bit));
2845 
2846 		type_valid = (ipv4_spec_valid) || (ipv6_spec_valid) ||
2847 			     (((eth_type == ETH_P_MPLS_UC) ||
2848 			       (eth_type == ETH_P_MPLS_MC)) && match_ipv);
2849 	}
2850 
2851 	return type_valid;
2852 }
2853 
2854 static bool is_valid_attr(struct mlx5_core_dev *mdev,
2855 			  const struct ib_flow_attr *flow_attr)
2856 {
2857 	return is_valid_ethertype(mdev, flow_attr, false) &&
2858 	       is_valid_ethertype(mdev, flow_attr, true);
2859 }
2860 
2861 static void put_flow_table(struct mlx5_ib_dev *dev,
2862 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
2863 {
2864 	prio->refcount -= !!ft_added;
2865 	if (!prio->refcount) {
2866 		mlx5_destroy_flow_table(prio->flow_table);
2867 		prio->flow_table = NULL;
2868 	}
2869 }
2870 
2871 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
2872 {
2873 	struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
2874 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
2875 							  struct mlx5_ib_flow_handler,
2876 							  ibflow);
2877 	struct mlx5_ib_flow_handler *iter, *tmp;
2878 
2879 	mutex_lock(&dev->flow_db->lock);
2880 
2881 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
2882 		mlx5_del_flow_rules(iter->rule);
2883 		put_flow_table(dev, iter->prio, true);
2884 		list_del(&iter->list);
2885 		kfree(iter);
2886 	}
2887 
2888 	mlx5_del_flow_rules(handler->rule);
2889 	put_flow_table(dev, handler->prio, true);
2890 	mutex_unlock(&dev->flow_db->lock);
2891 
2892 	kfree(handler);
2893 
2894 	return 0;
2895 }
2896 
2897 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
2898 {
2899 	priority *= 2;
2900 	if (!dont_trap)
2901 		priority++;
2902 	return priority;
2903 }
2904 
2905 enum flow_table_type {
2906 	MLX5_IB_FT_RX,
2907 	MLX5_IB_FT_TX
2908 };
2909 
2910 #define MLX5_FS_MAX_TYPES	 6
2911 #define MLX5_FS_MAX_ENTRIES	 BIT(16)
2912 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
2913 						struct ib_flow_attr *flow_attr,
2914 						enum flow_table_type ft_type)
2915 {
2916 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
2917 	struct mlx5_flow_namespace *ns = NULL;
2918 	struct mlx5_ib_flow_prio *prio;
2919 	struct mlx5_flow_table *ft;
2920 	int max_table_size;
2921 	int num_entries;
2922 	int num_groups;
2923 	int priority;
2924 	int err = 0;
2925 
2926 	max_table_size = BIT(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
2927 						       log_max_ft_size));
2928 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2929 		if (ft_type == MLX5_IB_FT_TX)
2930 			priority = 0;
2931 		else if (flow_is_multicast_only(flow_attr) &&
2932 			 !dont_trap)
2933 			priority = MLX5_IB_FLOW_MCAST_PRIO;
2934 		else
2935 			priority = ib_prio_to_core_prio(flow_attr->priority,
2936 							dont_trap);
2937 		ns = mlx5_get_flow_namespace(dev->mdev,
2938 					     ft_type == MLX5_IB_FT_TX ?
2939 					     MLX5_FLOW_NAMESPACE_EGRESS :
2940 					     MLX5_FLOW_NAMESPACE_BYPASS);
2941 		num_entries = MLX5_FS_MAX_ENTRIES;
2942 		num_groups = MLX5_FS_MAX_TYPES;
2943 		prio = &dev->flow_db->prios[priority];
2944 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2945 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2946 		ns = mlx5_get_flow_namespace(dev->mdev,
2947 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
2948 		build_leftovers_ft_param(&priority,
2949 					 &num_entries,
2950 					 &num_groups);
2951 		prio = &dev->flow_db->prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
2952 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2953 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
2954 					allow_sniffer_and_nic_rx_shared_tir))
2955 			return ERR_PTR(-ENOTSUPP);
2956 
2957 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
2958 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
2959 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
2960 
2961 		prio = &dev->flow_db->sniffer[ft_type];
2962 		priority = 0;
2963 		num_entries = 1;
2964 		num_groups = 1;
2965 	}
2966 
2967 	if (!ns)
2968 		return ERR_PTR(-ENOTSUPP);
2969 
2970 	if (num_entries > max_table_size)
2971 		return ERR_PTR(-ENOMEM);
2972 
2973 	ft = prio->flow_table;
2974 	if (!ft) {
2975 		ft = mlx5_create_auto_grouped_flow_table(ns, priority,
2976 							 num_entries,
2977 							 num_groups,
2978 							 0, 0);
2979 
2980 		if (!IS_ERR(ft)) {
2981 			prio->refcount = 0;
2982 			prio->flow_table = ft;
2983 		} else {
2984 			err = PTR_ERR(ft);
2985 		}
2986 	}
2987 
2988 	return err ? ERR_PTR(err) : prio;
2989 }
2990 
2991 static void set_underlay_qp(struct mlx5_ib_dev *dev,
2992 			    struct mlx5_flow_spec *spec,
2993 			    u32 underlay_qpn)
2994 {
2995 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param,
2996 					   spec->match_criteria,
2997 					   misc_parameters);
2998 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, spec->match_value,
2999 					   misc_parameters);
3000 
3001 	if (underlay_qpn &&
3002 	    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3003 				      ft_field_support.bth_dst_qp)) {
3004 		MLX5_SET(fte_match_set_misc,
3005 			 misc_params_v, bth_dst_qp, underlay_qpn);
3006 		MLX5_SET(fte_match_set_misc,
3007 			 misc_params_c, bth_dst_qp, 0xffffff);
3008 	}
3009 }
3010 
3011 static struct mlx5_ib_flow_handler *_create_flow_rule(struct mlx5_ib_dev *dev,
3012 						      struct mlx5_ib_flow_prio *ft_prio,
3013 						      const struct ib_flow_attr *flow_attr,
3014 						      struct mlx5_flow_destination *dst,
3015 						      u32 underlay_qpn)
3016 {
3017 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
3018 	struct mlx5_ib_flow_handler *handler;
3019 	struct mlx5_flow_act flow_act = {.flow_tag = MLX5_FS_DEFAULT_FLOW_TAG};
3020 	struct mlx5_flow_spec *spec;
3021 	struct mlx5_flow_destination *rule_dst = dst;
3022 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
3023 	unsigned int spec_index;
3024 	int err = 0;
3025 	int dest_num = 1;
3026 	bool is_egress = flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS;
3027 
3028 	if (!is_valid_attr(dev->mdev, flow_attr))
3029 		return ERR_PTR(-EINVAL);
3030 
3031 	spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
3032 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
3033 	if (!handler || !spec) {
3034 		err = -ENOMEM;
3035 		goto free;
3036 	}
3037 
3038 	INIT_LIST_HEAD(&handler->list);
3039 
3040 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
3041 		err = parse_flow_attr(dev->mdev, spec->match_criteria,
3042 				      spec->match_value,
3043 				      ib_flow, flow_attr, &flow_act);
3044 		if (err < 0)
3045 			goto free;
3046 
3047 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
3048 	}
3049 
3050 	if (!flow_is_multicast_only(flow_attr))
3051 		set_underlay_qp(dev, spec, underlay_qpn);
3052 
3053 	if (dev->rep) {
3054 		void *misc;
3055 
3056 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3057 				    misc_parameters);
3058 		MLX5_SET(fte_match_set_misc, misc, source_port,
3059 			 dev->rep->vport);
3060 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_criteria,
3061 				    misc_parameters);
3062 		MLX5_SET_TO_ONES(fte_match_set_misc, misc, source_port);
3063 	}
3064 
3065 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
3066 
3067 	if (is_egress &&
3068 	    !is_valid_spec(dev->mdev, spec, &flow_act, is_egress)) {
3069 		err = -EINVAL;
3070 		goto free;
3071 	}
3072 
3073 	if (flow_act.action & MLX5_FLOW_CONTEXT_ACTION_DROP) {
3074 		rule_dst = NULL;
3075 		dest_num = 0;
3076 	} else {
3077 		if (is_egress)
3078 			flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_ALLOW;
3079 		else
3080 			flow_act.action |=
3081 				dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
3082 					MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
3083 	}
3084 
3085 	if (flow_act.has_flow_tag &&
3086 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3087 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
3088 		mlx5_ib_warn(dev, "Flow tag %u and attribute type %x isn't allowed in leftovers\n",
3089 			     flow_act.flow_tag, flow_attr->type);
3090 		err = -EINVAL;
3091 		goto free;
3092 	}
3093 	handler->rule = mlx5_add_flow_rules(ft, spec,
3094 					    &flow_act,
3095 					    rule_dst, dest_num);
3096 
3097 	if (IS_ERR(handler->rule)) {
3098 		err = PTR_ERR(handler->rule);
3099 		goto free;
3100 	}
3101 
3102 	ft_prio->refcount++;
3103 	handler->prio = ft_prio;
3104 
3105 	ft_prio->flow_table = ft;
3106 free:
3107 	if (err)
3108 		kfree(handler);
3109 	kvfree(spec);
3110 	return err ? ERR_PTR(err) : handler;
3111 }
3112 
3113 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
3114 						     struct mlx5_ib_flow_prio *ft_prio,
3115 						     const struct ib_flow_attr *flow_attr,
3116 						     struct mlx5_flow_destination *dst)
3117 {
3118 	return _create_flow_rule(dev, ft_prio, flow_attr, dst, 0);
3119 }
3120 
3121 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
3122 							  struct mlx5_ib_flow_prio *ft_prio,
3123 							  struct ib_flow_attr *flow_attr,
3124 							  struct mlx5_flow_destination *dst)
3125 {
3126 	struct mlx5_ib_flow_handler *handler_dst = NULL;
3127 	struct mlx5_ib_flow_handler *handler = NULL;
3128 
3129 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
3130 	if (!IS_ERR(handler)) {
3131 		handler_dst = create_flow_rule(dev, ft_prio,
3132 					       flow_attr, dst);
3133 		if (IS_ERR(handler_dst)) {
3134 			mlx5_del_flow_rules(handler->rule);
3135 			ft_prio->refcount--;
3136 			kfree(handler);
3137 			handler = handler_dst;
3138 		} else {
3139 			list_add(&handler_dst->list, &handler->list);
3140 		}
3141 	}
3142 
3143 	return handler;
3144 }
3145 enum {
3146 	LEFTOVERS_MC,
3147 	LEFTOVERS_UC,
3148 };
3149 
3150 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
3151 							  struct mlx5_ib_flow_prio *ft_prio,
3152 							  struct ib_flow_attr *flow_attr,
3153 							  struct mlx5_flow_destination *dst)
3154 {
3155 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
3156 	struct mlx5_ib_flow_handler *handler = NULL;
3157 
3158 	static struct {
3159 		struct ib_flow_attr	flow_attr;
3160 		struct ib_flow_spec_eth eth_flow;
3161 	} leftovers_specs[] = {
3162 		[LEFTOVERS_MC] = {
3163 			.flow_attr = {
3164 				.num_of_specs = 1,
3165 				.size = sizeof(leftovers_specs[0])
3166 			},
3167 			.eth_flow = {
3168 				.type = IB_FLOW_SPEC_ETH,
3169 				.size = sizeof(struct ib_flow_spec_eth),
3170 				.mask = {.dst_mac = {0x1} },
3171 				.val =  {.dst_mac = {0x1} }
3172 			}
3173 		},
3174 		[LEFTOVERS_UC] = {
3175 			.flow_attr = {
3176 				.num_of_specs = 1,
3177 				.size = sizeof(leftovers_specs[0])
3178 			},
3179 			.eth_flow = {
3180 				.type = IB_FLOW_SPEC_ETH,
3181 				.size = sizeof(struct ib_flow_spec_eth),
3182 				.mask = {.dst_mac = {0x1} },
3183 				.val = {.dst_mac = {} }
3184 			}
3185 		}
3186 	};
3187 
3188 	handler = create_flow_rule(dev, ft_prio,
3189 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
3190 				   dst);
3191 	if (!IS_ERR(handler) &&
3192 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
3193 		handler_ucast = create_flow_rule(dev, ft_prio,
3194 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
3195 						 dst);
3196 		if (IS_ERR(handler_ucast)) {
3197 			mlx5_del_flow_rules(handler->rule);
3198 			ft_prio->refcount--;
3199 			kfree(handler);
3200 			handler = handler_ucast;
3201 		} else {
3202 			list_add(&handler_ucast->list, &handler->list);
3203 		}
3204 	}
3205 
3206 	return handler;
3207 }
3208 
3209 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
3210 							struct mlx5_ib_flow_prio *ft_rx,
3211 							struct mlx5_ib_flow_prio *ft_tx,
3212 							struct mlx5_flow_destination *dst)
3213 {
3214 	struct mlx5_ib_flow_handler *handler_rx;
3215 	struct mlx5_ib_flow_handler *handler_tx;
3216 	int err;
3217 	static const struct ib_flow_attr flow_attr  = {
3218 		.num_of_specs = 0,
3219 		.size = sizeof(flow_attr)
3220 	};
3221 
3222 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
3223 	if (IS_ERR(handler_rx)) {
3224 		err = PTR_ERR(handler_rx);
3225 		goto err;
3226 	}
3227 
3228 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
3229 	if (IS_ERR(handler_tx)) {
3230 		err = PTR_ERR(handler_tx);
3231 		goto err_tx;
3232 	}
3233 
3234 	list_add(&handler_tx->list, &handler_rx->list);
3235 
3236 	return handler_rx;
3237 
3238 err_tx:
3239 	mlx5_del_flow_rules(handler_rx->rule);
3240 	ft_rx->refcount--;
3241 	kfree(handler_rx);
3242 err:
3243 	return ERR_PTR(err);
3244 }
3245 
3246 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
3247 					   struct ib_flow_attr *flow_attr,
3248 					   int domain)
3249 {
3250 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3251 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3252 	struct mlx5_ib_flow_handler *handler = NULL;
3253 	struct mlx5_flow_destination *dst = NULL;
3254 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
3255 	struct mlx5_ib_flow_prio *ft_prio;
3256 	bool is_egress = flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS;
3257 	int err;
3258 	int underlay_qpn;
3259 
3260 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
3261 		return ERR_PTR(-ENOMEM);
3262 
3263 	if (domain != IB_FLOW_DOMAIN_USER ||
3264 	    flow_attr->port > dev->num_ports ||
3265 	    (flow_attr->flags & ~(IB_FLOW_ATTR_FLAGS_DONT_TRAP |
3266 				  IB_FLOW_ATTR_FLAGS_EGRESS)))
3267 		return ERR_PTR(-EINVAL);
3268 
3269 	if (is_egress &&
3270 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3271 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT))
3272 		return ERR_PTR(-EINVAL);
3273 
3274 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
3275 	if (!dst)
3276 		return ERR_PTR(-ENOMEM);
3277 
3278 	mutex_lock(&dev->flow_db->lock);
3279 
3280 	ft_prio = get_flow_table(dev, flow_attr,
3281 				 is_egress ? MLX5_IB_FT_TX : MLX5_IB_FT_RX);
3282 	if (IS_ERR(ft_prio)) {
3283 		err = PTR_ERR(ft_prio);
3284 		goto unlock;
3285 	}
3286 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3287 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
3288 		if (IS_ERR(ft_prio_tx)) {
3289 			err = PTR_ERR(ft_prio_tx);
3290 			ft_prio_tx = NULL;
3291 			goto destroy_ft;
3292 		}
3293 	}
3294 
3295 	if (is_egress) {
3296 		dst->type = MLX5_FLOW_DESTINATION_TYPE_PORT;
3297 	} else {
3298 		dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
3299 		if (mqp->flags & MLX5_IB_QP_RSS)
3300 			dst->tir_num = mqp->rss_qp.tirn;
3301 		else
3302 			dst->tir_num = mqp->raw_packet_qp.rq.tirn;
3303 	}
3304 
3305 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
3306 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
3307 			handler = create_dont_trap_rule(dev, ft_prio,
3308 							flow_attr, dst);
3309 		} else {
3310 			underlay_qpn = (mqp->flags & MLX5_IB_QP_UNDERLAY) ?
3311 					mqp->underlay_qpn : 0;
3312 			handler = _create_flow_rule(dev, ft_prio, flow_attr,
3313 						    dst, underlay_qpn);
3314 		}
3315 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3316 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
3317 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
3318 						dst);
3319 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3320 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
3321 	} else {
3322 		err = -EINVAL;
3323 		goto destroy_ft;
3324 	}
3325 
3326 	if (IS_ERR(handler)) {
3327 		err = PTR_ERR(handler);
3328 		handler = NULL;
3329 		goto destroy_ft;
3330 	}
3331 
3332 	mutex_unlock(&dev->flow_db->lock);
3333 	kfree(dst);
3334 
3335 	return &handler->ibflow;
3336 
3337 destroy_ft:
3338 	put_flow_table(dev, ft_prio, false);
3339 	if (ft_prio_tx)
3340 		put_flow_table(dev, ft_prio_tx, false);
3341 unlock:
3342 	mutex_unlock(&dev->flow_db->lock);
3343 	kfree(dst);
3344 	kfree(handler);
3345 	return ERR_PTR(err);
3346 }
3347 
3348 static u32 mlx5_ib_flow_action_flags_to_accel_xfrm_flags(u32 mlx5_flags)
3349 {
3350 	u32 flags = 0;
3351 
3352 	if (mlx5_flags & MLX5_IB_UAPI_FLOW_ACTION_FLAGS_REQUIRE_METADATA)
3353 		flags |= MLX5_ACCEL_XFRM_FLAG_REQUIRE_METADATA;
3354 
3355 	return flags;
3356 }
3357 
3358 #define MLX5_FLOW_ACTION_ESP_CREATE_LAST_SUPPORTED	MLX5_IB_UAPI_FLOW_ACTION_FLAGS_REQUIRE_METADATA
3359 static struct ib_flow_action *
3360 mlx5_ib_create_flow_action_esp(struct ib_device *device,
3361 			       const struct ib_flow_action_attrs_esp *attr,
3362 			       struct uverbs_attr_bundle *attrs)
3363 {
3364 	struct mlx5_ib_dev *mdev = to_mdev(device);
3365 	struct ib_uverbs_flow_action_esp_keymat_aes_gcm *aes_gcm;
3366 	struct mlx5_accel_esp_xfrm_attrs accel_attrs = {};
3367 	struct mlx5_ib_flow_action *action;
3368 	u64 action_flags;
3369 	u64 flags;
3370 	int err = 0;
3371 
3372 	if (IS_UVERBS_COPY_ERR(uverbs_copy_from(&action_flags, attrs,
3373 						MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS)))
3374 		return ERR_PTR(-EFAULT);
3375 
3376 	if (action_flags >= (MLX5_FLOW_ACTION_ESP_CREATE_LAST_SUPPORTED << 1))
3377 		return ERR_PTR(-EOPNOTSUPP);
3378 
3379 	flags = mlx5_ib_flow_action_flags_to_accel_xfrm_flags(action_flags);
3380 
3381 	/* We current only support a subset of the standard features. Only a
3382 	 * keymat of type AES_GCM, with icv_len == 16, iv_algo == SEQ and esn
3383 	 * (with overlap). Full offload mode isn't supported.
3384 	 */
3385 	if (!attr->keymat || attr->replay || attr->encap ||
3386 	    attr->spi || attr->seq || attr->tfc_pad ||
3387 	    attr->hard_limit_pkts ||
3388 	    (attr->flags & ~(IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
3389 			     IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ENCRYPT)))
3390 		return ERR_PTR(-EOPNOTSUPP);
3391 
3392 	if (attr->keymat->protocol !=
3393 	    IB_UVERBS_FLOW_ACTION_ESP_KEYMAT_AES_GCM)
3394 		return ERR_PTR(-EOPNOTSUPP);
3395 
3396 	aes_gcm = &attr->keymat->keymat.aes_gcm;
3397 
3398 	if (aes_gcm->icv_len != 16 ||
3399 	    aes_gcm->iv_algo != IB_UVERBS_FLOW_ACTION_IV_ALGO_SEQ)
3400 		return ERR_PTR(-EOPNOTSUPP);
3401 
3402 	action = kmalloc(sizeof(*action), GFP_KERNEL);
3403 	if (!action)
3404 		return ERR_PTR(-ENOMEM);
3405 
3406 	action->esp_aes_gcm.ib_flags = attr->flags;
3407 	memcpy(&accel_attrs.keymat.aes_gcm.aes_key, &aes_gcm->aes_key,
3408 	       sizeof(accel_attrs.keymat.aes_gcm.aes_key));
3409 	accel_attrs.keymat.aes_gcm.key_len = aes_gcm->key_len * 8;
3410 	memcpy(&accel_attrs.keymat.aes_gcm.salt, &aes_gcm->salt,
3411 	       sizeof(accel_attrs.keymat.aes_gcm.salt));
3412 	memcpy(&accel_attrs.keymat.aes_gcm.seq_iv, &aes_gcm->iv,
3413 	       sizeof(accel_attrs.keymat.aes_gcm.seq_iv));
3414 	accel_attrs.keymat.aes_gcm.icv_len = aes_gcm->icv_len * 8;
3415 	accel_attrs.keymat.aes_gcm.iv_algo = MLX5_ACCEL_ESP_AES_GCM_IV_ALGO_SEQ;
3416 	accel_attrs.keymat_type = MLX5_ACCEL_ESP_KEYMAT_AES_GCM;
3417 
3418 	accel_attrs.esn = attr->esn;
3419 	if (attr->flags & IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED)
3420 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_TRIGGERED;
3421 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)
3422 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
3423 
3424 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ENCRYPT)
3425 		accel_attrs.action |= MLX5_ACCEL_ESP_ACTION_ENCRYPT;
3426 
3427 	action->esp_aes_gcm.ctx =
3428 		mlx5_accel_esp_create_xfrm(mdev->mdev, &accel_attrs, flags);
3429 	if (IS_ERR(action->esp_aes_gcm.ctx)) {
3430 		err = PTR_ERR(action->esp_aes_gcm.ctx);
3431 		goto err_parse;
3432 	}
3433 
3434 	action->esp_aes_gcm.ib_flags = attr->flags;
3435 
3436 	return &action->ib_action;
3437 
3438 err_parse:
3439 	kfree(action);
3440 	return ERR_PTR(err);
3441 }
3442 
3443 static int
3444 mlx5_ib_modify_flow_action_esp(struct ib_flow_action *action,
3445 			       const struct ib_flow_action_attrs_esp *attr,
3446 			       struct uverbs_attr_bundle *attrs)
3447 {
3448 	struct mlx5_ib_flow_action *maction = to_mflow_act(action);
3449 	struct mlx5_accel_esp_xfrm_attrs accel_attrs;
3450 	int err = 0;
3451 
3452 	if (attr->keymat || attr->replay || attr->encap ||
3453 	    attr->spi || attr->seq || attr->tfc_pad ||
3454 	    attr->hard_limit_pkts ||
3455 	    (attr->flags & ~(IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
3456 			     IB_FLOW_ACTION_ESP_FLAGS_MOD_ESP_ATTRS |
3457 			     IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)))
3458 		return -EOPNOTSUPP;
3459 
3460 	/* Only the ESN value or the MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP can
3461 	 * be modified.
3462 	 */
3463 	if (!(maction->esp_aes_gcm.ib_flags &
3464 	      IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED) &&
3465 	    attr->flags & (IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
3466 			   IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW))
3467 		return -EINVAL;
3468 
3469 	memcpy(&accel_attrs, &maction->esp_aes_gcm.ctx->attrs,
3470 	       sizeof(accel_attrs));
3471 
3472 	accel_attrs.esn = attr->esn;
3473 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)
3474 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
3475 	else
3476 		accel_attrs.flags &= ~MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
3477 
3478 	err = mlx5_accel_esp_modify_xfrm(maction->esp_aes_gcm.ctx,
3479 					 &accel_attrs);
3480 	if (err)
3481 		return err;
3482 
3483 	maction->esp_aes_gcm.ib_flags &=
3484 		~IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW;
3485 	maction->esp_aes_gcm.ib_flags |=
3486 		attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW;
3487 
3488 	return 0;
3489 }
3490 
3491 static int mlx5_ib_destroy_flow_action(struct ib_flow_action *action)
3492 {
3493 	struct mlx5_ib_flow_action *maction = to_mflow_act(action);
3494 
3495 	switch (action->type) {
3496 	case IB_FLOW_ACTION_ESP:
3497 		/*
3498 		 * We only support aes_gcm by now, so we implicitly know this is
3499 		 * the underline crypto.
3500 		 */
3501 		mlx5_accel_esp_destroy_xfrm(maction->esp_aes_gcm.ctx);
3502 		break;
3503 	default:
3504 		WARN_ON(true);
3505 		break;
3506 	}
3507 
3508 	kfree(maction);
3509 	return 0;
3510 }
3511 
3512 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
3513 {
3514 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3515 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
3516 	int err;
3517 
3518 	if (mqp->flags & MLX5_IB_QP_UNDERLAY) {
3519 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
3520 		return -EOPNOTSUPP;
3521 	}
3522 
3523 	err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
3524 	if (err)
3525 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
3526 			     ibqp->qp_num, gid->raw);
3527 
3528 	return err;
3529 }
3530 
3531 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
3532 {
3533 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3534 	int err;
3535 
3536 	err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
3537 	if (err)
3538 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
3539 			     ibqp->qp_num, gid->raw);
3540 
3541 	return err;
3542 }
3543 
3544 static int init_node_data(struct mlx5_ib_dev *dev)
3545 {
3546 	int err;
3547 
3548 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
3549 	if (err)
3550 		return err;
3551 
3552 	dev->mdev->rev_id = dev->mdev->pdev->revision;
3553 
3554 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
3555 }
3556 
3557 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
3558 			     char *buf)
3559 {
3560 	struct mlx5_ib_dev *dev =
3561 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3562 
3563 	return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
3564 }
3565 
3566 static ssize_t show_reg_pages(struct device *device,
3567 			      struct device_attribute *attr, char *buf)
3568 {
3569 	struct mlx5_ib_dev *dev =
3570 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3571 
3572 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
3573 }
3574 
3575 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
3576 			char *buf)
3577 {
3578 	struct mlx5_ib_dev *dev =
3579 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3580 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
3581 }
3582 
3583 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
3584 			char *buf)
3585 {
3586 	struct mlx5_ib_dev *dev =
3587 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3588 	return sprintf(buf, "%x\n", dev->mdev->rev_id);
3589 }
3590 
3591 static ssize_t show_board(struct device *device, struct device_attribute *attr,
3592 			  char *buf)
3593 {
3594 	struct mlx5_ib_dev *dev =
3595 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3596 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
3597 		       dev->mdev->board_id);
3598 }
3599 
3600 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
3601 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
3602 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
3603 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
3604 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
3605 
3606 static struct device_attribute *mlx5_class_attributes[] = {
3607 	&dev_attr_hw_rev,
3608 	&dev_attr_hca_type,
3609 	&dev_attr_board_id,
3610 	&dev_attr_fw_pages,
3611 	&dev_attr_reg_pages,
3612 };
3613 
3614 static void pkey_change_handler(struct work_struct *work)
3615 {
3616 	struct mlx5_ib_port_resources *ports =
3617 		container_of(work, struct mlx5_ib_port_resources,
3618 			     pkey_change_work);
3619 
3620 	mutex_lock(&ports->devr->mutex);
3621 	mlx5_ib_gsi_pkey_change(ports->gsi);
3622 	mutex_unlock(&ports->devr->mutex);
3623 }
3624 
3625 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
3626 {
3627 	struct mlx5_ib_qp *mqp;
3628 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
3629 	struct mlx5_core_cq *mcq;
3630 	struct list_head cq_armed_list;
3631 	unsigned long flags_qp;
3632 	unsigned long flags_cq;
3633 	unsigned long flags;
3634 
3635 	INIT_LIST_HEAD(&cq_armed_list);
3636 
3637 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
3638 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
3639 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
3640 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
3641 		if (mqp->sq.tail != mqp->sq.head) {
3642 			send_mcq = to_mcq(mqp->ibqp.send_cq);
3643 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
3644 			if (send_mcq->mcq.comp &&
3645 			    mqp->ibqp.send_cq->comp_handler) {
3646 				if (!send_mcq->mcq.reset_notify_added) {
3647 					send_mcq->mcq.reset_notify_added = 1;
3648 					list_add_tail(&send_mcq->mcq.reset_notify,
3649 						      &cq_armed_list);
3650 				}
3651 			}
3652 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
3653 		}
3654 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
3655 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
3656 		/* no handling is needed for SRQ */
3657 		if (!mqp->ibqp.srq) {
3658 			if (mqp->rq.tail != mqp->rq.head) {
3659 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
3660 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
3661 				if (recv_mcq->mcq.comp &&
3662 				    mqp->ibqp.recv_cq->comp_handler) {
3663 					if (!recv_mcq->mcq.reset_notify_added) {
3664 						recv_mcq->mcq.reset_notify_added = 1;
3665 						list_add_tail(&recv_mcq->mcq.reset_notify,
3666 							      &cq_armed_list);
3667 					}
3668 				}
3669 				spin_unlock_irqrestore(&recv_mcq->lock,
3670 						       flags_cq);
3671 			}
3672 		}
3673 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
3674 	}
3675 	/*At that point all inflight post send were put to be executed as of we
3676 	 * lock/unlock above locks Now need to arm all involved CQs.
3677 	 */
3678 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
3679 		mcq->comp(mcq);
3680 	}
3681 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
3682 }
3683 
3684 static void delay_drop_handler(struct work_struct *work)
3685 {
3686 	int err;
3687 	struct mlx5_ib_delay_drop *delay_drop =
3688 		container_of(work, struct mlx5_ib_delay_drop,
3689 			     delay_drop_work);
3690 
3691 	atomic_inc(&delay_drop->events_cnt);
3692 
3693 	mutex_lock(&delay_drop->lock);
3694 	err = mlx5_core_set_delay_drop(delay_drop->dev->mdev,
3695 				       delay_drop->timeout);
3696 	if (err) {
3697 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
3698 			     delay_drop->timeout);
3699 		delay_drop->activate = false;
3700 	}
3701 	mutex_unlock(&delay_drop->lock);
3702 }
3703 
3704 static void mlx5_ib_handle_event(struct work_struct *_work)
3705 {
3706 	struct mlx5_ib_event_work *work =
3707 		container_of(_work, struct mlx5_ib_event_work, work);
3708 	struct mlx5_ib_dev *ibdev;
3709 	struct ib_event ibev;
3710 	bool fatal = false;
3711 	u8 port = (u8)work->param;
3712 
3713 	if (mlx5_core_is_mp_slave(work->dev)) {
3714 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->context);
3715 		if (!ibdev)
3716 			goto out;
3717 	} else {
3718 		ibdev = work->context;
3719 	}
3720 
3721 	switch (work->event) {
3722 	case MLX5_DEV_EVENT_SYS_ERROR:
3723 		ibev.event = IB_EVENT_DEVICE_FATAL;
3724 		mlx5_ib_handle_internal_error(ibdev);
3725 		fatal = true;
3726 		break;
3727 
3728 	case MLX5_DEV_EVENT_PORT_UP:
3729 	case MLX5_DEV_EVENT_PORT_DOWN:
3730 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
3731 		/* In RoCE, port up/down events are handled in
3732 		 * mlx5_netdev_event().
3733 		 */
3734 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
3735 			IB_LINK_LAYER_ETHERNET)
3736 			goto out;
3737 
3738 		ibev.event = (work->event == MLX5_DEV_EVENT_PORT_UP) ?
3739 			     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
3740 		break;
3741 
3742 	case MLX5_DEV_EVENT_LID_CHANGE:
3743 		ibev.event = IB_EVENT_LID_CHANGE;
3744 		break;
3745 
3746 	case MLX5_DEV_EVENT_PKEY_CHANGE:
3747 		ibev.event = IB_EVENT_PKEY_CHANGE;
3748 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
3749 		break;
3750 
3751 	case MLX5_DEV_EVENT_GUID_CHANGE:
3752 		ibev.event = IB_EVENT_GID_CHANGE;
3753 		break;
3754 
3755 	case MLX5_DEV_EVENT_CLIENT_REREG:
3756 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
3757 		break;
3758 	case MLX5_DEV_EVENT_DELAY_DROP_TIMEOUT:
3759 		schedule_work(&ibdev->delay_drop.delay_drop_work);
3760 		goto out;
3761 	default:
3762 		goto out;
3763 	}
3764 
3765 	ibev.device	      = &ibdev->ib_dev;
3766 	ibev.element.port_num = port;
3767 
3768 	if (!rdma_is_port_valid(&ibdev->ib_dev, port)) {
3769 		mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
3770 		goto out;
3771 	}
3772 
3773 	if (ibdev->ib_active)
3774 		ib_dispatch_event(&ibev);
3775 
3776 	if (fatal)
3777 		ibdev->ib_active = false;
3778 out:
3779 	kfree(work);
3780 }
3781 
3782 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
3783 			  enum mlx5_dev_event event, unsigned long param)
3784 {
3785 	struct mlx5_ib_event_work *work;
3786 
3787 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
3788 	if (!work)
3789 		return;
3790 
3791 	INIT_WORK(&work->work, mlx5_ib_handle_event);
3792 	work->dev = dev;
3793 	work->param = param;
3794 	work->context = context;
3795 	work->event = event;
3796 
3797 	queue_work(mlx5_ib_event_wq, &work->work);
3798 }
3799 
3800 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
3801 {
3802 	struct mlx5_hca_vport_context vport_ctx;
3803 	int err;
3804 	int port;
3805 
3806 	for (port = 1; port <= dev->num_ports; port++) {
3807 		dev->mdev->port_caps[port - 1].has_smi = false;
3808 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
3809 		    MLX5_CAP_PORT_TYPE_IB) {
3810 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
3811 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
3812 								   port, 0,
3813 								   &vport_ctx);
3814 				if (err) {
3815 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
3816 						    port, err);
3817 					return err;
3818 				}
3819 				dev->mdev->port_caps[port - 1].has_smi =
3820 					vport_ctx.has_smi;
3821 			} else {
3822 				dev->mdev->port_caps[port - 1].has_smi = true;
3823 			}
3824 		}
3825 	}
3826 	return 0;
3827 }
3828 
3829 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
3830 {
3831 	int port;
3832 
3833 	for (port = 1; port <= dev->num_ports; port++)
3834 		mlx5_query_ext_port_caps(dev, port);
3835 }
3836 
3837 static int get_port_caps(struct mlx5_ib_dev *dev, u8 port)
3838 {
3839 	struct ib_device_attr *dprops = NULL;
3840 	struct ib_port_attr *pprops = NULL;
3841 	int err = -ENOMEM;
3842 	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
3843 
3844 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
3845 	if (!pprops)
3846 		goto out;
3847 
3848 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
3849 	if (!dprops)
3850 		goto out;
3851 
3852 	err = set_has_smi_cap(dev);
3853 	if (err)
3854 		goto out;
3855 
3856 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
3857 	if (err) {
3858 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
3859 		goto out;
3860 	}
3861 
3862 	memset(pprops, 0, sizeof(*pprops));
3863 	err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
3864 	if (err) {
3865 		mlx5_ib_warn(dev, "query_port %d failed %d\n",
3866 			     port, err);
3867 		goto out;
3868 	}
3869 
3870 	dev->mdev->port_caps[port - 1].pkey_table_len =
3871 					dprops->max_pkeys;
3872 	dev->mdev->port_caps[port - 1].gid_table_len =
3873 					pprops->gid_tbl_len;
3874 	mlx5_ib_dbg(dev, "port %d: pkey_table_len %d, gid_table_len %d\n",
3875 		    port, dprops->max_pkeys, pprops->gid_tbl_len);
3876 
3877 out:
3878 	kfree(pprops);
3879 	kfree(dprops);
3880 
3881 	return err;
3882 }
3883 
3884 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
3885 {
3886 	int err;
3887 
3888 	err = mlx5_mr_cache_cleanup(dev);
3889 	if (err)
3890 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
3891 
3892 	if (dev->umrc.qp)
3893 		mlx5_ib_destroy_qp(dev->umrc.qp);
3894 	if (dev->umrc.cq)
3895 		ib_free_cq(dev->umrc.cq);
3896 	if (dev->umrc.pd)
3897 		ib_dealloc_pd(dev->umrc.pd);
3898 }
3899 
3900 enum {
3901 	MAX_UMR_WR = 128,
3902 };
3903 
3904 static int create_umr_res(struct mlx5_ib_dev *dev)
3905 {
3906 	struct ib_qp_init_attr *init_attr = NULL;
3907 	struct ib_qp_attr *attr = NULL;
3908 	struct ib_pd *pd;
3909 	struct ib_cq *cq;
3910 	struct ib_qp *qp;
3911 	int ret;
3912 
3913 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
3914 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
3915 	if (!attr || !init_attr) {
3916 		ret = -ENOMEM;
3917 		goto error_0;
3918 	}
3919 
3920 	pd = ib_alloc_pd(&dev->ib_dev, 0);
3921 	if (IS_ERR(pd)) {
3922 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
3923 		ret = PTR_ERR(pd);
3924 		goto error_0;
3925 	}
3926 
3927 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
3928 	if (IS_ERR(cq)) {
3929 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
3930 		ret = PTR_ERR(cq);
3931 		goto error_2;
3932 	}
3933 
3934 	init_attr->send_cq = cq;
3935 	init_attr->recv_cq = cq;
3936 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
3937 	init_attr->cap.max_send_wr = MAX_UMR_WR;
3938 	init_attr->cap.max_send_sge = 1;
3939 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
3940 	init_attr->port_num = 1;
3941 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
3942 	if (IS_ERR(qp)) {
3943 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
3944 		ret = PTR_ERR(qp);
3945 		goto error_3;
3946 	}
3947 	qp->device     = &dev->ib_dev;
3948 	qp->real_qp    = qp;
3949 	qp->uobject    = NULL;
3950 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
3951 	qp->send_cq    = init_attr->send_cq;
3952 	qp->recv_cq    = init_attr->recv_cq;
3953 
3954 	attr->qp_state = IB_QPS_INIT;
3955 	attr->port_num = 1;
3956 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
3957 				IB_QP_PORT, NULL);
3958 	if (ret) {
3959 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
3960 		goto error_4;
3961 	}
3962 
3963 	memset(attr, 0, sizeof(*attr));
3964 	attr->qp_state = IB_QPS_RTR;
3965 	attr->path_mtu = IB_MTU_256;
3966 
3967 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
3968 	if (ret) {
3969 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
3970 		goto error_4;
3971 	}
3972 
3973 	memset(attr, 0, sizeof(*attr));
3974 	attr->qp_state = IB_QPS_RTS;
3975 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
3976 	if (ret) {
3977 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
3978 		goto error_4;
3979 	}
3980 
3981 	dev->umrc.qp = qp;
3982 	dev->umrc.cq = cq;
3983 	dev->umrc.pd = pd;
3984 
3985 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
3986 	ret = mlx5_mr_cache_init(dev);
3987 	if (ret) {
3988 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
3989 		goto error_4;
3990 	}
3991 
3992 	kfree(attr);
3993 	kfree(init_attr);
3994 
3995 	return 0;
3996 
3997 error_4:
3998 	mlx5_ib_destroy_qp(qp);
3999 	dev->umrc.qp = NULL;
4000 
4001 error_3:
4002 	ib_free_cq(cq);
4003 	dev->umrc.cq = NULL;
4004 
4005 error_2:
4006 	ib_dealloc_pd(pd);
4007 	dev->umrc.pd = NULL;
4008 
4009 error_0:
4010 	kfree(attr);
4011 	kfree(init_attr);
4012 	return ret;
4013 }
4014 
4015 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
4016 {
4017 	switch (umr_fence_cap) {
4018 	case MLX5_CAP_UMR_FENCE_NONE:
4019 		return MLX5_FENCE_MODE_NONE;
4020 	case MLX5_CAP_UMR_FENCE_SMALL:
4021 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
4022 	default:
4023 		return MLX5_FENCE_MODE_STRONG_ORDERING;
4024 	}
4025 }
4026 
4027 static int create_dev_resources(struct mlx5_ib_resources *devr)
4028 {
4029 	struct ib_srq_init_attr attr;
4030 	struct mlx5_ib_dev *dev;
4031 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
4032 	int port;
4033 	int ret = 0;
4034 
4035 	dev = container_of(devr, struct mlx5_ib_dev, devr);
4036 
4037 	mutex_init(&devr->mutex);
4038 
4039 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
4040 	if (IS_ERR(devr->p0)) {
4041 		ret = PTR_ERR(devr->p0);
4042 		goto error0;
4043 	}
4044 	devr->p0->device  = &dev->ib_dev;
4045 	devr->p0->uobject = NULL;
4046 	atomic_set(&devr->p0->usecnt, 0);
4047 
4048 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
4049 	if (IS_ERR(devr->c0)) {
4050 		ret = PTR_ERR(devr->c0);
4051 		goto error1;
4052 	}
4053 	devr->c0->device        = &dev->ib_dev;
4054 	devr->c0->uobject       = NULL;
4055 	devr->c0->comp_handler  = NULL;
4056 	devr->c0->event_handler = NULL;
4057 	devr->c0->cq_context    = NULL;
4058 	atomic_set(&devr->c0->usecnt, 0);
4059 
4060 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
4061 	if (IS_ERR(devr->x0)) {
4062 		ret = PTR_ERR(devr->x0);
4063 		goto error2;
4064 	}
4065 	devr->x0->device = &dev->ib_dev;
4066 	devr->x0->inode = NULL;
4067 	atomic_set(&devr->x0->usecnt, 0);
4068 	mutex_init(&devr->x0->tgt_qp_mutex);
4069 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
4070 
4071 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
4072 	if (IS_ERR(devr->x1)) {
4073 		ret = PTR_ERR(devr->x1);
4074 		goto error3;
4075 	}
4076 	devr->x1->device = &dev->ib_dev;
4077 	devr->x1->inode = NULL;
4078 	atomic_set(&devr->x1->usecnt, 0);
4079 	mutex_init(&devr->x1->tgt_qp_mutex);
4080 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
4081 
4082 	memset(&attr, 0, sizeof(attr));
4083 	attr.attr.max_sge = 1;
4084 	attr.attr.max_wr = 1;
4085 	attr.srq_type = IB_SRQT_XRC;
4086 	attr.ext.cq = devr->c0;
4087 	attr.ext.xrc.xrcd = devr->x0;
4088 
4089 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
4090 	if (IS_ERR(devr->s0)) {
4091 		ret = PTR_ERR(devr->s0);
4092 		goto error4;
4093 	}
4094 	devr->s0->device	= &dev->ib_dev;
4095 	devr->s0->pd		= devr->p0;
4096 	devr->s0->uobject       = NULL;
4097 	devr->s0->event_handler = NULL;
4098 	devr->s0->srq_context   = NULL;
4099 	devr->s0->srq_type      = IB_SRQT_XRC;
4100 	devr->s0->ext.xrc.xrcd	= devr->x0;
4101 	devr->s0->ext.cq	= devr->c0;
4102 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
4103 	atomic_inc(&devr->s0->ext.cq->usecnt);
4104 	atomic_inc(&devr->p0->usecnt);
4105 	atomic_set(&devr->s0->usecnt, 0);
4106 
4107 	memset(&attr, 0, sizeof(attr));
4108 	attr.attr.max_sge = 1;
4109 	attr.attr.max_wr = 1;
4110 	attr.srq_type = IB_SRQT_BASIC;
4111 	devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
4112 	if (IS_ERR(devr->s1)) {
4113 		ret = PTR_ERR(devr->s1);
4114 		goto error5;
4115 	}
4116 	devr->s1->device	= &dev->ib_dev;
4117 	devr->s1->pd		= devr->p0;
4118 	devr->s1->uobject       = NULL;
4119 	devr->s1->event_handler = NULL;
4120 	devr->s1->srq_context   = NULL;
4121 	devr->s1->srq_type      = IB_SRQT_BASIC;
4122 	devr->s1->ext.cq	= devr->c0;
4123 	atomic_inc(&devr->p0->usecnt);
4124 	atomic_set(&devr->s1->usecnt, 0);
4125 
4126 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
4127 		INIT_WORK(&devr->ports[port].pkey_change_work,
4128 			  pkey_change_handler);
4129 		devr->ports[port].devr = devr;
4130 	}
4131 
4132 	return 0;
4133 
4134 error5:
4135 	mlx5_ib_destroy_srq(devr->s0);
4136 error4:
4137 	mlx5_ib_dealloc_xrcd(devr->x1);
4138 error3:
4139 	mlx5_ib_dealloc_xrcd(devr->x0);
4140 error2:
4141 	mlx5_ib_destroy_cq(devr->c0);
4142 error1:
4143 	mlx5_ib_dealloc_pd(devr->p0);
4144 error0:
4145 	return ret;
4146 }
4147 
4148 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
4149 {
4150 	struct mlx5_ib_dev *dev =
4151 		container_of(devr, struct mlx5_ib_dev, devr);
4152 	int port;
4153 
4154 	mlx5_ib_destroy_srq(devr->s1);
4155 	mlx5_ib_destroy_srq(devr->s0);
4156 	mlx5_ib_dealloc_xrcd(devr->x0);
4157 	mlx5_ib_dealloc_xrcd(devr->x1);
4158 	mlx5_ib_destroy_cq(devr->c0);
4159 	mlx5_ib_dealloc_pd(devr->p0);
4160 
4161 	/* Make sure no change P_Key work items are still executing */
4162 	for (port = 0; port < dev->num_ports; ++port)
4163 		cancel_work_sync(&devr->ports[port].pkey_change_work);
4164 }
4165 
4166 static u32 get_core_cap_flags(struct ib_device *ibdev)
4167 {
4168 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4169 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
4170 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
4171 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
4172 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
4173 	u32 ret = 0;
4174 
4175 	if (ll == IB_LINK_LAYER_INFINIBAND)
4176 		return RDMA_CORE_PORT_IBA_IB;
4177 
4178 	if (raw_support)
4179 		ret = RDMA_CORE_PORT_RAW_PACKET;
4180 
4181 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
4182 		return ret;
4183 
4184 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
4185 		return ret;
4186 
4187 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
4188 		ret |= RDMA_CORE_PORT_IBA_ROCE;
4189 
4190 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
4191 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
4192 
4193 	return ret;
4194 }
4195 
4196 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
4197 			       struct ib_port_immutable *immutable)
4198 {
4199 	struct ib_port_attr attr;
4200 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4201 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
4202 	int err;
4203 
4204 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
4205 
4206 	err = ib_query_port(ibdev, port_num, &attr);
4207 	if (err)
4208 		return err;
4209 
4210 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
4211 	immutable->gid_tbl_len = attr.gid_tbl_len;
4212 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
4213 	if ((ll == IB_LINK_LAYER_INFINIBAND) || MLX5_CAP_GEN(dev->mdev, roce))
4214 		immutable->max_mad_size = IB_MGMT_MAD_SIZE;
4215 
4216 	return 0;
4217 }
4218 
4219 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u8 port_num,
4220 				   struct ib_port_immutable *immutable)
4221 {
4222 	struct ib_port_attr attr;
4223 	int err;
4224 
4225 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
4226 
4227 	err = ib_query_port(ibdev, port_num, &attr);
4228 	if (err)
4229 		return err;
4230 
4231 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
4232 	immutable->gid_tbl_len = attr.gid_tbl_len;
4233 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
4234 
4235 	return 0;
4236 }
4237 
4238 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
4239 {
4240 	struct mlx5_ib_dev *dev =
4241 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
4242 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
4243 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
4244 		 fw_rev_sub(dev->mdev));
4245 }
4246 
4247 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
4248 {
4249 	struct mlx5_core_dev *mdev = dev->mdev;
4250 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
4251 								 MLX5_FLOW_NAMESPACE_LAG);
4252 	struct mlx5_flow_table *ft;
4253 	int err;
4254 
4255 	if (!ns || !mlx5_lag_is_active(mdev))
4256 		return 0;
4257 
4258 	err = mlx5_cmd_create_vport_lag(mdev);
4259 	if (err)
4260 		return err;
4261 
4262 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
4263 	if (IS_ERR(ft)) {
4264 		err = PTR_ERR(ft);
4265 		goto err_destroy_vport_lag;
4266 	}
4267 
4268 	dev->flow_db->lag_demux_ft = ft;
4269 	return 0;
4270 
4271 err_destroy_vport_lag:
4272 	mlx5_cmd_destroy_vport_lag(mdev);
4273 	return err;
4274 }
4275 
4276 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
4277 {
4278 	struct mlx5_core_dev *mdev = dev->mdev;
4279 
4280 	if (dev->flow_db->lag_demux_ft) {
4281 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
4282 		dev->flow_db->lag_demux_ft = NULL;
4283 
4284 		mlx5_cmd_destroy_vport_lag(mdev);
4285 	}
4286 }
4287 
4288 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
4289 {
4290 	int err;
4291 
4292 	dev->roce[port_num].nb.notifier_call = mlx5_netdev_event;
4293 	err = register_netdevice_notifier(&dev->roce[port_num].nb);
4294 	if (err) {
4295 		dev->roce[port_num].nb.notifier_call = NULL;
4296 		return err;
4297 	}
4298 
4299 	return 0;
4300 }
4301 
4302 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
4303 {
4304 	if (dev->roce[port_num].nb.notifier_call) {
4305 		unregister_netdevice_notifier(&dev->roce[port_num].nb);
4306 		dev->roce[port_num].nb.notifier_call = NULL;
4307 	}
4308 }
4309 
4310 static int mlx5_enable_eth(struct mlx5_ib_dev *dev, u8 port_num)
4311 {
4312 	int err;
4313 
4314 	if (MLX5_CAP_GEN(dev->mdev, roce)) {
4315 		err = mlx5_nic_vport_enable_roce(dev->mdev);
4316 		if (err)
4317 			return err;
4318 	}
4319 
4320 	err = mlx5_eth_lag_init(dev);
4321 	if (err)
4322 		goto err_disable_roce;
4323 
4324 	return 0;
4325 
4326 err_disable_roce:
4327 	if (MLX5_CAP_GEN(dev->mdev, roce))
4328 		mlx5_nic_vport_disable_roce(dev->mdev);
4329 
4330 	return err;
4331 }
4332 
4333 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
4334 {
4335 	mlx5_eth_lag_cleanup(dev);
4336 	if (MLX5_CAP_GEN(dev->mdev, roce))
4337 		mlx5_nic_vport_disable_roce(dev->mdev);
4338 }
4339 
4340 struct mlx5_ib_counter {
4341 	const char *name;
4342 	size_t offset;
4343 };
4344 
4345 #define INIT_Q_COUNTER(_name)		\
4346 	{ .name = #_name, .offset = MLX5_BYTE_OFF(query_q_counter_out, _name)}
4347 
4348 static const struct mlx5_ib_counter basic_q_cnts[] = {
4349 	INIT_Q_COUNTER(rx_write_requests),
4350 	INIT_Q_COUNTER(rx_read_requests),
4351 	INIT_Q_COUNTER(rx_atomic_requests),
4352 	INIT_Q_COUNTER(out_of_buffer),
4353 };
4354 
4355 static const struct mlx5_ib_counter out_of_seq_q_cnts[] = {
4356 	INIT_Q_COUNTER(out_of_sequence),
4357 };
4358 
4359 static const struct mlx5_ib_counter retrans_q_cnts[] = {
4360 	INIT_Q_COUNTER(duplicate_request),
4361 	INIT_Q_COUNTER(rnr_nak_retry_err),
4362 	INIT_Q_COUNTER(packet_seq_err),
4363 	INIT_Q_COUNTER(implied_nak_seq_err),
4364 	INIT_Q_COUNTER(local_ack_timeout_err),
4365 };
4366 
4367 #define INIT_CONG_COUNTER(_name)		\
4368 	{ .name = #_name, .offset =	\
4369 		MLX5_BYTE_OFF(query_cong_statistics_out, _name ## _high)}
4370 
4371 static const struct mlx5_ib_counter cong_cnts[] = {
4372 	INIT_CONG_COUNTER(rp_cnp_ignored),
4373 	INIT_CONG_COUNTER(rp_cnp_handled),
4374 	INIT_CONG_COUNTER(np_ecn_marked_roce_packets),
4375 	INIT_CONG_COUNTER(np_cnp_sent),
4376 };
4377 
4378 static const struct mlx5_ib_counter extended_err_cnts[] = {
4379 	INIT_Q_COUNTER(resp_local_length_error),
4380 	INIT_Q_COUNTER(resp_cqe_error),
4381 	INIT_Q_COUNTER(req_cqe_error),
4382 	INIT_Q_COUNTER(req_remote_invalid_request),
4383 	INIT_Q_COUNTER(req_remote_access_errors),
4384 	INIT_Q_COUNTER(resp_remote_access_errors),
4385 	INIT_Q_COUNTER(resp_cqe_flush_error),
4386 	INIT_Q_COUNTER(req_cqe_flush_error),
4387 };
4388 
4389 static void mlx5_ib_dealloc_counters(struct mlx5_ib_dev *dev)
4390 {
4391 	int i;
4392 
4393 	for (i = 0; i < dev->num_ports; i++) {
4394 		if (dev->port[i].cnts.set_id)
4395 			mlx5_core_dealloc_q_counter(dev->mdev,
4396 						    dev->port[i].cnts.set_id);
4397 		kfree(dev->port[i].cnts.names);
4398 		kfree(dev->port[i].cnts.offsets);
4399 	}
4400 }
4401 
4402 static int __mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev,
4403 				    struct mlx5_ib_counters *cnts)
4404 {
4405 	u32 num_counters;
4406 
4407 	num_counters = ARRAY_SIZE(basic_q_cnts);
4408 
4409 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt))
4410 		num_counters += ARRAY_SIZE(out_of_seq_q_cnts);
4411 
4412 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters))
4413 		num_counters += ARRAY_SIZE(retrans_q_cnts);
4414 
4415 	if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters))
4416 		num_counters += ARRAY_SIZE(extended_err_cnts);
4417 
4418 	cnts->num_q_counters = num_counters;
4419 
4420 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
4421 		cnts->num_cong_counters = ARRAY_SIZE(cong_cnts);
4422 		num_counters += ARRAY_SIZE(cong_cnts);
4423 	}
4424 
4425 	cnts->names = kcalloc(num_counters, sizeof(cnts->names), GFP_KERNEL);
4426 	if (!cnts->names)
4427 		return -ENOMEM;
4428 
4429 	cnts->offsets = kcalloc(num_counters,
4430 				sizeof(cnts->offsets), GFP_KERNEL);
4431 	if (!cnts->offsets)
4432 		goto err_names;
4433 
4434 	return 0;
4435 
4436 err_names:
4437 	kfree(cnts->names);
4438 	cnts->names = NULL;
4439 	return -ENOMEM;
4440 }
4441 
4442 static void mlx5_ib_fill_counters(struct mlx5_ib_dev *dev,
4443 				  const char **names,
4444 				  size_t *offsets)
4445 {
4446 	int i;
4447 	int j = 0;
4448 
4449 	for (i = 0; i < ARRAY_SIZE(basic_q_cnts); i++, j++) {
4450 		names[j] = basic_q_cnts[i].name;
4451 		offsets[j] = basic_q_cnts[i].offset;
4452 	}
4453 
4454 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt)) {
4455 		for (i = 0; i < ARRAY_SIZE(out_of_seq_q_cnts); i++, j++) {
4456 			names[j] = out_of_seq_q_cnts[i].name;
4457 			offsets[j] = out_of_seq_q_cnts[i].offset;
4458 		}
4459 	}
4460 
4461 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
4462 		for (i = 0; i < ARRAY_SIZE(retrans_q_cnts); i++, j++) {
4463 			names[j] = retrans_q_cnts[i].name;
4464 			offsets[j] = retrans_q_cnts[i].offset;
4465 		}
4466 	}
4467 
4468 	if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters)) {
4469 		for (i = 0; i < ARRAY_SIZE(extended_err_cnts); i++, j++) {
4470 			names[j] = extended_err_cnts[i].name;
4471 			offsets[j] = extended_err_cnts[i].offset;
4472 		}
4473 	}
4474 
4475 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
4476 		for (i = 0; i < ARRAY_SIZE(cong_cnts); i++, j++) {
4477 			names[j] = cong_cnts[i].name;
4478 			offsets[j] = cong_cnts[i].offset;
4479 		}
4480 	}
4481 }
4482 
4483 static int mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev)
4484 {
4485 	int err = 0;
4486 	int i;
4487 
4488 	for (i = 0; i < dev->num_ports; i++) {
4489 		err = __mlx5_ib_alloc_counters(dev, &dev->port[i].cnts);
4490 		if (err)
4491 			goto err_alloc;
4492 
4493 		mlx5_ib_fill_counters(dev, dev->port[i].cnts.names,
4494 				      dev->port[i].cnts.offsets);
4495 
4496 		err = mlx5_core_alloc_q_counter(dev->mdev,
4497 						&dev->port[i].cnts.set_id);
4498 		if (err) {
4499 			mlx5_ib_warn(dev,
4500 				     "couldn't allocate queue counter for port %d, err %d\n",
4501 				     i + 1, err);
4502 			goto err_alloc;
4503 		}
4504 		dev->port[i].cnts.set_id_valid = true;
4505 	}
4506 
4507 	return 0;
4508 
4509 err_alloc:
4510 	mlx5_ib_dealloc_counters(dev);
4511 	return err;
4512 }
4513 
4514 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
4515 						    u8 port_num)
4516 {
4517 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4518 	struct mlx5_ib_port *port = &dev->port[port_num - 1];
4519 
4520 	/* We support only per port stats */
4521 	if (port_num == 0)
4522 		return NULL;
4523 
4524 	return rdma_alloc_hw_stats_struct(port->cnts.names,
4525 					  port->cnts.num_q_counters +
4526 					  port->cnts.num_cong_counters,
4527 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
4528 }
4529 
4530 static int mlx5_ib_query_q_counters(struct mlx5_core_dev *mdev,
4531 				    struct mlx5_ib_port *port,
4532 				    struct rdma_hw_stats *stats)
4533 {
4534 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
4535 	void *out;
4536 	__be32 val;
4537 	int ret, i;
4538 
4539 	out = kvzalloc(outlen, GFP_KERNEL);
4540 	if (!out)
4541 		return -ENOMEM;
4542 
4543 	ret = mlx5_core_query_q_counter(mdev,
4544 					port->cnts.set_id, 0,
4545 					out, outlen);
4546 	if (ret)
4547 		goto free;
4548 
4549 	for (i = 0; i < port->cnts.num_q_counters; i++) {
4550 		val = *(__be32 *)(out + port->cnts.offsets[i]);
4551 		stats->value[i] = (u64)be32_to_cpu(val);
4552 	}
4553 
4554 free:
4555 	kvfree(out);
4556 	return ret;
4557 }
4558 
4559 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
4560 				struct rdma_hw_stats *stats,
4561 				u8 port_num, int index)
4562 {
4563 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4564 	struct mlx5_ib_port *port = &dev->port[port_num - 1];
4565 	struct mlx5_core_dev *mdev;
4566 	int ret, num_counters;
4567 	u8 mdev_port_num;
4568 
4569 	if (!stats)
4570 		return -EINVAL;
4571 
4572 	num_counters = port->cnts.num_q_counters + port->cnts.num_cong_counters;
4573 
4574 	/* q_counters are per IB device, query the master mdev */
4575 	ret = mlx5_ib_query_q_counters(dev->mdev, port, stats);
4576 	if (ret)
4577 		return ret;
4578 
4579 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
4580 		mdev = mlx5_ib_get_native_port_mdev(dev, port_num,
4581 						    &mdev_port_num);
4582 		if (!mdev) {
4583 			/* If port is not affiliated yet, its in down state
4584 			 * which doesn't have any counters yet, so it would be
4585 			 * zero. So no need to read from the HCA.
4586 			 */
4587 			goto done;
4588 		}
4589 		ret = mlx5_lag_query_cong_counters(dev->mdev,
4590 						   stats->value +
4591 						   port->cnts.num_q_counters,
4592 						   port->cnts.num_cong_counters,
4593 						   port->cnts.offsets +
4594 						   port->cnts.num_q_counters);
4595 
4596 		mlx5_ib_put_native_port_mdev(dev, port_num);
4597 		if (ret)
4598 			return ret;
4599 	}
4600 
4601 done:
4602 	return num_counters;
4603 }
4604 
4605 static void mlx5_ib_free_rdma_netdev(struct net_device *netdev)
4606 {
4607 	return mlx5_rdma_netdev_free(netdev);
4608 }
4609 
4610 static struct net_device*
4611 mlx5_ib_alloc_rdma_netdev(struct ib_device *hca,
4612 			  u8 port_num,
4613 			  enum rdma_netdev_t type,
4614 			  const char *name,
4615 			  unsigned char name_assign_type,
4616 			  void (*setup)(struct net_device *))
4617 {
4618 	struct net_device *netdev;
4619 	struct rdma_netdev *rn;
4620 
4621 	if (type != RDMA_NETDEV_IPOIB)
4622 		return ERR_PTR(-EOPNOTSUPP);
4623 
4624 	netdev = mlx5_rdma_netdev_alloc(to_mdev(hca)->mdev, hca,
4625 					name, setup);
4626 	if (likely(!IS_ERR_OR_NULL(netdev))) {
4627 		rn = netdev_priv(netdev);
4628 		rn->free_rdma_netdev = mlx5_ib_free_rdma_netdev;
4629 	}
4630 	return netdev;
4631 }
4632 
4633 static void delay_drop_debugfs_cleanup(struct mlx5_ib_dev *dev)
4634 {
4635 	if (!dev->delay_drop.dbg)
4636 		return;
4637 	debugfs_remove_recursive(dev->delay_drop.dbg->dir_debugfs);
4638 	kfree(dev->delay_drop.dbg);
4639 	dev->delay_drop.dbg = NULL;
4640 }
4641 
4642 static void cancel_delay_drop(struct mlx5_ib_dev *dev)
4643 {
4644 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4645 		return;
4646 
4647 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4648 	delay_drop_debugfs_cleanup(dev);
4649 }
4650 
4651 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
4652 				       size_t count, loff_t *pos)
4653 {
4654 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
4655 	char lbuf[20];
4656 	int len;
4657 
4658 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
4659 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
4660 }
4661 
4662 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
4663 					size_t count, loff_t *pos)
4664 {
4665 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
4666 	u32 timeout;
4667 	u32 var;
4668 
4669 	if (kstrtouint_from_user(buf, count, 0, &var))
4670 		return -EFAULT;
4671 
4672 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
4673 			1000);
4674 	if (timeout != var)
4675 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
4676 			    timeout);
4677 
4678 	delay_drop->timeout = timeout;
4679 
4680 	return count;
4681 }
4682 
4683 static const struct file_operations fops_delay_drop_timeout = {
4684 	.owner	= THIS_MODULE,
4685 	.open	= simple_open,
4686 	.write	= delay_drop_timeout_write,
4687 	.read	= delay_drop_timeout_read,
4688 };
4689 
4690 static int delay_drop_debugfs_init(struct mlx5_ib_dev *dev)
4691 {
4692 	struct mlx5_ib_dbg_delay_drop *dbg;
4693 
4694 	if (!mlx5_debugfs_root)
4695 		return 0;
4696 
4697 	dbg = kzalloc(sizeof(*dbg), GFP_KERNEL);
4698 	if (!dbg)
4699 		return -ENOMEM;
4700 
4701 	dev->delay_drop.dbg = dbg;
4702 
4703 	dbg->dir_debugfs =
4704 		debugfs_create_dir("delay_drop",
4705 				   dev->mdev->priv.dbg_root);
4706 	if (!dbg->dir_debugfs)
4707 		goto out_debugfs;
4708 
4709 	dbg->events_cnt_debugfs =
4710 		debugfs_create_atomic_t("num_timeout_events", 0400,
4711 					dbg->dir_debugfs,
4712 					&dev->delay_drop.events_cnt);
4713 	if (!dbg->events_cnt_debugfs)
4714 		goto out_debugfs;
4715 
4716 	dbg->rqs_cnt_debugfs =
4717 		debugfs_create_atomic_t("num_rqs", 0400,
4718 					dbg->dir_debugfs,
4719 					&dev->delay_drop.rqs_cnt);
4720 	if (!dbg->rqs_cnt_debugfs)
4721 		goto out_debugfs;
4722 
4723 	dbg->timeout_debugfs =
4724 		debugfs_create_file("timeout", 0600,
4725 				    dbg->dir_debugfs,
4726 				    &dev->delay_drop,
4727 				    &fops_delay_drop_timeout);
4728 	if (!dbg->timeout_debugfs)
4729 		goto out_debugfs;
4730 
4731 	return 0;
4732 
4733 out_debugfs:
4734 	delay_drop_debugfs_cleanup(dev);
4735 	return -ENOMEM;
4736 }
4737 
4738 static void init_delay_drop(struct mlx5_ib_dev *dev)
4739 {
4740 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4741 		return;
4742 
4743 	mutex_init(&dev->delay_drop.lock);
4744 	dev->delay_drop.dev = dev;
4745 	dev->delay_drop.activate = false;
4746 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4747 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4748 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4749 	atomic_set(&dev->delay_drop.events_cnt, 0);
4750 
4751 	if (delay_drop_debugfs_init(dev))
4752 		mlx5_ib_warn(dev, "Failed to init delay drop debugfs\n");
4753 }
4754 
4755 static const struct cpumask *
4756 mlx5_ib_get_vector_affinity(struct ib_device *ibdev, int comp_vector)
4757 {
4758 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4759 
4760 	return mlx5_get_vector_affinity(dev->mdev, comp_vector);
4761 }
4762 
4763 /* The mlx5_ib_multiport_mutex should be held when calling this function */
4764 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
4765 				      struct mlx5_ib_multiport_info *mpi)
4766 {
4767 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
4768 	struct mlx5_ib_port *port = &ibdev->port[port_num];
4769 	int comps;
4770 	int err;
4771 	int i;
4772 
4773 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
4774 
4775 	spin_lock(&port->mp.mpi_lock);
4776 	if (!mpi->ibdev) {
4777 		spin_unlock(&port->mp.mpi_lock);
4778 		return;
4779 	}
4780 	mpi->ibdev = NULL;
4781 
4782 	spin_unlock(&port->mp.mpi_lock);
4783 	mlx5_remove_netdev_notifier(ibdev, port_num);
4784 	spin_lock(&port->mp.mpi_lock);
4785 
4786 	comps = mpi->mdev_refcnt;
4787 	if (comps) {
4788 		mpi->unaffiliate = true;
4789 		init_completion(&mpi->unref_comp);
4790 		spin_unlock(&port->mp.mpi_lock);
4791 
4792 		for (i = 0; i < comps; i++)
4793 			wait_for_completion(&mpi->unref_comp);
4794 
4795 		spin_lock(&port->mp.mpi_lock);
4796 		mpi->unaffiliate = false;
4797 	}
4798 
4799 	port->mp.mpi = NULL;
4800 
4801 	list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4802 
4803 	spin_unlock(&port->mp.mpi_lock);
4804 
4805 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
4806 
4807 	mlx5_ib_dbg(ibdev, "unaffiliated port %d\n", port_num + 1);
4808 	/* Log an error, still needed to cleanup the pointers and add
4809 	 * it back to the list.
4810 	 */
4811 	if (err)
4812 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
4813 			    port_num + 1);
4814 
4815 	ibdev->roce[port_num].last_port_state = IB_PORT_DOWN;
4816 }
4817 
4818 /* The mlx5_ib_multiport_mutex should be held when calling this function */
4819 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
4820 				    struct mlx5_ib_multiport_info *mpi)
4821 {
4822 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
4823 	int err;
4824 
4825 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
4826 	if (ibdev->port[port_num].mp.mpi) {
4827 		mlx5_ib_warn(ibdev, "port %d already affiliated.\n",
4828 			     port_num + 1);
4829 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
4830 		return false;
4831 	}
4832 
4833 	ibdev->port[port_num].mp.mpi = mpi;
4834 	mpi->ibdev = ibdev;
4835 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
4836 
4837 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
4838 	if (err)
4839 		goto unbind;
4840 
4841 	err = get_port_caps(ibdev, mlx5_core_native_port_num(mpi->mdev));
4842 	if (err)
4843 		goto unbind;
4844 
4845 	err = mlx5_add_netdev_notifier(ibdev, port_num);
4846 	if (err) {
4847 		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
4848 			    port_num + 1);
4849 		goto unbind;
4850 	}
4851 
4852 	err = mlx5_ib_init_cong_debugfs(ibdev, port_num);
4853 	if (err)
4854 		goto unbind;
4855 
4856 	return true;
4857 
4858 unbind:
4859 	mlx5_ib_unbind_slave_port(ibdev, mpi);
4860 	return false;
4861 }
4862 
4863 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
4864 {
4865 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4866 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
4867 							  port_num + 1);
4868 	struct mlx5_ib_multiport_info *mpi;
4869 	int err;
4870 	int i;
4871 
4872 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
4873 		return 0;
4874 
4875 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
4876 						     &dev->sys_image_guid);
4877 	if (err)
4878 		return err;
4879 
4880 	err = mlx5_nic_vport_enable_roce(dev->mdev);
4881 	if (err)
4882 		return err;
4883 
4884 	mutex_lock(&mlx5_ib_multiport_mutex);
4885 	for (i = 0; i < dev->num_ports; i++) {
4886 		bool bound = false;
4887 
4888 		/* build a stub multiport info struct for the native port. */
4889 		if (i == port_num) {
4890 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4891 			if (!mpi) {
4892 				mutex_unlock(&mlx5_ib_multiport_mutex);
4893 				mlx5_nic_vport_disable_roce(dev->mdev);
4894 				return -ENOMEM;
4895 			}
4896 
4897 			mpi->is_master = true;
4898 			mpi->mdev = dev->mdev;
4899 			mpi->sys_image_guid = dev->sys_image_guid;
4900 			dev->port[i].mp.mpi = mpi;
4901 			mpi->ibdev = dev;
4902 			mpi = NULL;
4903 			continue;
4904 		}
4905 
4906 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
4907 				    list) {
4908 			if (dev->sys_image_guid == mpi->sys_image_guid &&
4909 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
4910 				bound = mlx5_ib_bind_slave_port(dev, mpi);
4911 			}
4912 
4913 			if (bound) {
4914 				dev_dbg(&mpi->mdev->pdev->dev, "removing port from unaffiliated list.\n");
4915 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
4916 				list_del(&mpi->list);
4917 				break;
4918 			}
4919 		}
4920 		if (!bound) {
4921 			get_port_caps(dev, i + 1);
4922 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
4923 				    i + 1);
4924 		}
4925 	}
4926 
4927 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
4928 	mutex_unlock(&mlx5_ib_multiport_mutex);
4929 	return err;
4930 }
4931 
4932 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
4933 {
4934 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4935 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
4936 							  port_num + 1);
4937 	int i;
4938 
4939 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
4940 		return;
4941 
4942 	mutex_lock(&mlx5_ib_multiport_mutex);
4943 	for (i = 0; i < dev->num_ports; i++) {
4944 		if (dev->port[i].mp.mpi) {
4945 			/* Destroy the native port stub */
4946 			if (i == port_num) {
4947 				kfree(dev->port[i].mp.mpi);
4948 				dev->port[i].mp.mpi = NULL;
4949 			} else {
4950 				mlx5_ib_dbg(dev, "unbinding port_num: %d\n", i + 1);
4951 				mlx5_ib_unbind_slave_port(dev, dev->port[i].mp.mpi);
4952 			}
4953 		}
4954 	}
4955 
4956 	mlx5_ib_dbg(dev, "removing from devlist\n");
4957 	list_del(&dev->ib_dev_list);
4958 	mutex_unlock(&mlx5_ib_multiport_mutex);
4959 
4960 	mlx5_nic_vport_disable_roce(dev->mdev);
4961 }
4962 
4963 ADD_UVERBS_ATTRIBUTES_SIMPLE(mlx5_ib_dm, UVERBS_OBJECT_DM,
4964 			     UVERBS_METHOD_DM_ALLOC,
4965 			     &UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
4966 						  UVERBS_ATTR_TYPE(u64),
4967 						  UA_FLAGS(UVERBS_ATTR_SPEC_F_MANDATORY)),
4968 			     &UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
4969 						  UVERBS_ATTR_TYPE(u16),
4970 						  UA_FLAGS(UVERBS_ATTR_SPEC_F_MANDATORY)));
4971 
4972 ADD_UVERBS_ATTRIBUTES_SIMPLE(mlx5_ib_flow_action, UVERBS_OBJECT_FLOW_ACTION,
4973 			     UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
4974 			     &UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
4975 						 UVERBS_ATTR_TYPE(u64),
4976 						 UA_FLAGS(UVERBS_ATTR_SPEC_F_MANDATORY)));
4977 
4978 #define NUM_TREES	2
4979 static int populate_specs_root(struct mlx5_ib_dev *dev)
4980 {
4981 	const struct uverbs_object_tree_def *default_root[NUM_TREES + 1] = {
4982 		uverbs_default_get_objects()};
4983 	size_t num_trees = 1;
4984 
4985 	if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_DEVICE &&
4986 	    !WARN_ON(num_trees >= ARRAY_SIZE(default_root)))
4987 		default_root[num_trees++] = &mlx5_ib_flow_action;
4988 
4989 	if (MLX5_CAP_DEV_MEM(dev->mdev, memic) &&
4990 	    !WARN_ON(num_trees >= ARRAY_SIZE(default_root)))
4991 		default_root[num_trees++] = &mlx5_ib_dm;
4992 
4993 	dev->ib_dev.specs_root =
4994 		uverbs_alloc_spec_tree(num_trees, default_root);
4995 
4996 	return PTR_ERR_OR_ZERO(dev->ib_dev.specs_root);
4997 }
4998 
4999 static void depopulate_specs_root(struct mlx5_ib_dev *dev)
5000 {
5001 	uverbs_free_spec_tree(dev->ib_dev.specs_root);
5002 }
5003 
5004 void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
5005 {
5006 	mlx5_ib_cleanup_multiport_master(dev);
5007 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
5008 	cleanup_srcu_struct(&dev->mr_srcu);
5009 #endif
5010 	kfree(dev->port);
5011 }
5012 
5013 int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
5014 {
5015 	struct mlx5_core_dev *mdev = dev->mdev;
5016 	const char *name;
5017 	int err;
5018 	int i;
5019 
5020 	dev->port = kcalloc(dev->num_ports, sizeof(*dev->port),
5021 			    GFP_KERNEL);
5022 	if (!dev->port)
5023 		return -ENOMEM;
5024 
5025 	for (i = 0; i < dev->num_ports; i++) {
5026 		spin_lock_init(&dev->port[i].mp.mpi_lock);
5027 		rwlock_init(&dev->roce[i].netdev_lock);
5028 	}
5029 
5030 	err = mlx5_ib_init_multiport_master(dev);
5031 	if (err)
5032 		goto err_free_port;
5033 
5034 	if (!mlx5_core_mp_enabled(mdev)) {
5035 		for (i = 1; i <= dev->num_ports; i++) {
5036 			err = get_port_caps(dev, i);
5037 			if (err)
5038 				break;
5039 		}
5040 	} else {
5041 		err = get_port_caps(dev, mlx5_core_native_port_num(mdev));
5042 	}
5043 	if (err)
5044 		goto err_mp;
5045 
5046 	if (mlx5_use_mad_ifc(dev))
5047 		get_ext_port_caps(dev);
5048 
5049 	if (!mlx5_lag_is_active(mdev))
5050 		name = "mlx5_%d";
5051 	else
5052 		name = "mlx5_bond_%d";
5053 
5054 	strlcpy(dev->ib_dev.name, name, IB_DEVICE_NAME_MAX);
5055 	dev->ib_dev.owner		= THIS_MODULE;
5056 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
5057 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
5058 	dev->ib_dev.phys_port_cnt	= dev->num_ports;
5059 	dev->ib_dev.num_comp_vectors    =
5060 		dev->mdev->priv.eq_table.num_comp_vectors;
5061 	dev->ib_dev.dev.parent		= &mdev->pdev->dev;
5062 
5063 	mutex_init(&dev->cap_mask_mutex);
5064 	INIT_LIST_HEAD(&dev->qp_list);
5065 	spin_lock_init(&dev->reset_flow_resource_lock);
5066 
5067 	spin_lock_init(&dev->memic.memic_lock);
5068 	dev->memic.dev = mdev;
5069 
5070 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
5071 	err = init_srcu_struct(&dev->mr_srcu);
5072 	if (err)
5073 		goto err_free_port;
5074 #endif
5075 
5076 	return 0;
5077 err_mp:
5078 	mlx5_ib_cleanup_multiport_master(dev);
5079 
5080 err_free_port:
5081 	kfree(dev->port);
5082 
5083 	return -ENOMEM;
5084 }
5085 
5086 static int mlx5_ib_stage_flow_db_init(struct mlx5_ib_dev *dev)
5087 {
5088 	dev->flow_db = kzalloc(sizeof(*dev->flow_db), GFP_KERNEL);
5089 
5090 	if (!dev->flow_db)
5091 		return -ENOMEM;
5092 
5093 	mutex_init(&dev->flow_db->lock);
5094 
5095 	return 0;
5096 }
5097 
5098 int mlx5_ib_stage_rep_flow_db_init(struct mlx5_ib_dev *dev)
5099 {
5100 	struct mlx5_ib_dev *nic_dev;
5101 
5102 	nic_dev = mlx5_ib_get_uplink_ibdev(dev->mdev->priv.eswitch);
5103 
5104 	if (!nic_dev)
5105 		return -EINVAL;
5106 
5107 	dev->flow_db = nic_dev->flow_db;
5108 
5109 	return 0;
5110 }
5111 
5112 static void mlx5_ib_stage_flow_db_cleanup(struct mlx5_ib_dev *dev)
5113 {
5114 	kfree(dev->flow_db);
5115 }
5116 
5117 int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
5118 {
5119 	struct mlx5_core_dev *mdev = dev->mdev;
5120 	int err;
5121 
5122 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
5123 	dev->ib_dev.uverbs_cmd_mask	=
5124 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
5125 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
5126 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
5127 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
5128 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
5129 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
5130 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
5131 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
5132 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
5133 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
5134 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
5135 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
5136 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
5137 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
5138 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
5139 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
5140 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
5141 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
5142 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
5143 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
5144 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
5145 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
5146 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
5147 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
5148 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
5149 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
5150 	dev->ib_dev.uverbs_ex_cmd_mask =
5151 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
5152 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
5153 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP)	|
5154 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP)	|
5155 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_CQ);
5156 
5157 	dev->ib_dev.query_device	= mlx5_ib_query_device;
5158 	dev->ib_dev.get_link_layer	= mlx5_ib_port_link_layer;
5159 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
5160 	dev->ib_dev.add_gid		= mlx5_ib_add_gid;
5161 	dev->ib_dev.del_gid		= mlx5_ib_del_gid;
5162 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
5163 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
5164 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
5165 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
5166 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
5167 	dev->ib_dev.mmap		= mlx5_ib_mmap;
5168 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
5169 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
5170 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
5171 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
5172 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
5173 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
5174 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
5175 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
5176 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
5177 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
5178 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
5179 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
5180 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
5181 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
5182 	dev->ib_dev.post_send		= mlx5_ib_post_send;
5183 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
5184 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
5185 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
5186 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
5187 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
5188 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
5189 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
5190 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
5191 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
5192 	dev->ib_dev.rereg_user_mr	= mlx5_ib_rereg_user_mr;
5193 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
5194 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
5195 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
5196 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
5197 	dev->ib_dev.alloc_mr		= mlx5_ib_alloc_mr;
5198 	dev->ib_dev.map_mr_sg		= mlx5_ib_map_mr_sg;
5199 	dev->ib_dev.check_mr_status	= mlx5_ib_check_mr_status;
5200 	dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
5201 	dev->ib_dev.get_vector_affinity	= mlx5_ib_get_vector_affinity;
5202 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads))
5203 		dev->ib_dev.alloc_rdma_netdev	= mlx5_ib_alloc_rdma_netdev;
5204 
5205 	if (mlx5_core_is_pf(mdev)) {
5206 		dev->ib_dev.get_vf_config	= mlx5_ib_get_vf_config;
5207 		dev->ib_dev.set_vf_link_state	= mlx5_ib_set_vf_link_state;
5208 		dev->ib_dev.get_vf_stats	= mlx5_ib_get_vf_stats;
5209 		dev->ib_dev.set_vf_guid		= mlx5_ib_set_vf_guid;
5210 	}
5211 
5212 	dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
5213 
5214 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
5215 
5216 	if (MLX5_CAP_GEN(mdev, imaicl)) {
5217 		dev->ib_dev.alloc_mw		= mlx5_ib_alloc_mw;
5218 		dev->ib_dev.dealloc_mw		= mlx5_ib_dealloc_mw;
5219 		dev->ib_dev.uverbs_cmd_mask |=
5220 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
5221 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
5222 	}
5223 
5224 	if (MLX5_CAP_GEN(mdev, xrc)) {
5225 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
5226 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
5227 		dev->ib_dev.uverbs_cmd_mask |=
5228 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
5229 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
5230 	}
5231 
5232 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
5233 		dev->ib_dev.alloc_dm = mlx5_ib_alloc_dm;
5234 		dev->ib_dev.dealloc_dm = mlx5_ib_dealloc_dm;
5235 		dev->ib_dev.reg_dm_mr = mlx5_ib_reg_dm_mr;
5236 	}
5237 
5238 	dev->ib_dev.create_flow	= mlx5_ib_create_flow;
5239 	dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
5240 	dev->ib_dev.uverbs_ex_cmd_mask |=
5241 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
5242 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW);
5243 	dev->ib_dev.create_flow_action_esp = mlx5_ib_create_flow_action_esp;
5244 	dev->ib_dev.destroy_flow_action = mlx5_ib_destroy_flow_action;
5245 	dev->ib_dev.modify_flow_action_esp = mlx5_ib_modify_flow_action_esp;
5246 	dev->ib_dev.driver_id = RDMA_DRIVER_MLX5;
5247 
5248 	err = init_node_data(dev);
5249 	if (err)
5250 		return err;
5251 
5252 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
5253 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
5254 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
5255 		mutex_init(&dev->lb_mutex);
5256 
5257 	return 0;
5258 }
5259 
5260 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
5261 {
5262 	dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
5263 	dev->ib_dev.query_port		= mlx5_ib_query_port;
5264 
5265 	return 0;
5266 }
5267 
5268 int mlx5_ib_stage_rep_non_default_cb(struct mlx5_ib_dev *dev)
5269 {
5270 	dev->ib_dev.get_port_immutable  = mlx5_port_rep_immutable;
5271 	dev->ib_dev.query_port		= mlx5_ib_rep_query_port;
5272 
5273 	return 0;
5274 }
5275 
5276 static int mlx5_ib_stage_common_roce_init(struct mlx5_ib_dev *dev,
5277 					  u8 port_num)
5278 {
5279 	int i;
5280 
5281 	for (i = 0; i < dev->num_ports; i++) {
5282 		dev->roce[i].dev = dev;
5283 		dev->roce[i].native_port_num = i + 1;
5284 		dev->roce[i].last_port_state = IB_PORT_DOWN;
5285 	}
5286 
5287 	dev->ib_dev.get_netdev	= mlx5_ib_get_netdev;
5288 	dev->ib_dev.create_wq	 = mlx5_ib_create_wq;
5289 	dev->ib_dev.modify_wq	 = mlx5_ib_modify_wq;
5290 	dev->ib_dev.destroy_wq	 = mlx5_ib_destroy_wq;
5291 	dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
5292 	dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
5293 
5294 	dev->ib_dev.uverbs_ex_cmd_mask |=
5295 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
5296 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
5297 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
5298 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
5299 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
5300 
5301 	return mlx5_add_netdev_notifier(dev, port_num);
5302 }
5303 
5304 static void mlx5_ib_stage_common_roce_cleanup(struct mlx5_ib_dev *dev)
5305 {
5306 	u8 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5307 
5308 	mlx5_remove_netdev_notifier(dev, port_num);
5309 }
5310 
5311 int mlx5_ib_stage_rep_roce_init(struct mlx5_ib_dev *dev)
5312 {
5313 	struct mlx5_core_dev *mdev = dev->mdev;
5314 	enum rdma_link_layer ll;
5315 	int port_type_cap;
5316 	int err = 0;
5317 	u8 port_num;
5318 
5319 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5320 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5321 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5322 
5323 	if (ll == IB_LINK_LAYER_ETHERNET)
5324 		err = mlx5_ib_stage_common_roce_init(dev, port_num);
5325 
5326 	return err;
5327 }
5328 
5329 void mlx5_ib_stage_rep_roce_cleanup(struct mlx5_ib_dev *dev)
5330 {
5331 	mlx5_ib_stage_common_roce_cleanup(dev);
5332 }
5333 
5334 static int mlx5_ib_stage_roce_init(struct mlx5_ib_dev *dev)
5335 {
5336 	struct mlx5_core_dev *mdev = dev->mdev;
5337 	enum rdma_link_layer ll;
5338 	int port_type_cap;
5339 	u8 port_num;
5340 	int err;
5341 
5342 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5343 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5344 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5345 
5346 	if (ll == IB_LINK_LAYER_ETHERNET) {
5347 		err = mlx5_ib_stage_common_roce_init(dev, port_num);
5348 		if (err)
5349 			return err;
5350 
5351 		err = mlx5_enable_eth(dev, port_num);
5352 		if (err)
5353 			goto cleanup;
5354 	}
5355 
5356 	return 0;
5357 cleanup:
5358 	mlx5_ib_stage_common_roce_cleanup(dev);
5359 
5360 	return err;
5361 }
5362 
5363 static void mlx5_ib_stage_roce_cleanup(struct mlx5_ib_dev *dev)
5364 {
5365 	struct mlx5_core_dev *mdev = dev->mdev;
5366 	enum rdma_link_layer ll;
5367 	int port_type_cap;
5368 	u8 port_num;
5369 
5370 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5371 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5372 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5373 
5374 	if (ll == IB_LINK_LAYER_ETHERNET) {
5375 		mlx5_disable_eth(dev);
5376 		mlx5_ib_stage_common_roce_cleanup(dev);
5377 	}
5378 }
5379 
5380 int mlx5_ib_stage_dev_res_init(struct mlx5_ib_dev *dev)
5381 {
5382 	return create_dev_resources(&dev->devr);
5383 }
5384 
5385 void mlx5_ib_stage_dev_res_cleanup(struct mlx5_ib_dev *dev)
5386 {
5387 	destroy_dev_resources(&dev->devr);
5388 }
5389 
5390 static int mlx5_ib_stage_odp_init(struct mlx5_ib_dev *dev)
5391 {
5392 	mlx5_ib_internal_fill_odp_caps(dev);
5393 
5394 	return mlx5_ib_odp_init_one(dev);
5395 }
5396 
5397 int mlx5_ib_stage_counters_init(struct mlx5_ib_dev *dev)
5398 {
5399 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
5400 		dev->ib_dev.get_hw_stats	= mlx5_ib_get_hw_stats;
5401 		dev->ib_dev.alloc_hw_stats	= mlx5_ib_alloc_hw_stats;
5402 
5403 		return mlx5_ib_alloc_counters(dev);
5404 	}
5405 
5406 	return 0;
5407 }
5408 
5409 void mlx5_ib_stage_counters_cleanup(struct mlx5_ib_dev *dev)
5410 {
5411 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
5412 		mlx5_ib_dealloc_counters(dev);
5413 }
5414 
5415 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
5416 {
5417 	return mlx5_ib_init_cong_debugfs(dev,
5418 					 mlx5_core_native_port_num(dev->mdev) - 1);
5419 }
5420 
5421 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
5422 {
5423 	mlx5_ib_cleanup_cong_debugfs(dev,
5424 				     mlx5_core_native_port_num(dev->mdev) - 1);
5425 }
5426 
5427 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
5428 {
5429 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
5430 	if (!dev->mdev->priv.uar)
5431 		return -ENOMEM;
5432 	return 0;
5433 }
5434 
5435 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
5436 {
5437 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
5438 }
5439 
5440 int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
5441 {
5442 	int err;
5443 
5444 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
5445 	if (err)
5446 		return err;
5447 
5448 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
5449 	if (err)
5450 		mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
5451 
5452 	return err;
5453 }
5454 
5455 void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
5456 {
5457 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
5458 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
5459 }
5460 
5461 static int mlx5_ib_stage_populate_specs(struct mlx5_ib_dev *dev)
5462 {
5463 	return populate_specs_root(dev);
5464 }
5465 
5466 int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
5467 {
5468 	return ib_register_device(&dev->ib_dev, NULL);
5469 }
5470 
5471 static void mlx5_ib_stage_depopulate_specs(struct mlx5_ib_dev *dev)
5472 {
5473 	depopulate_specs_root(dev);
5474 }
5475 
5476 void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
5477 {
5478 	destroy_umrc_res(dev);
5479 }
5480 
5481 void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
5482 {
5483 	ib_unregister_device(&dev->ib_dev);
5484 }
5485 
5486 int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
5487 {
5488 	return create_umr_res(dev);
5489 }
5490 
5491 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
5492 {
5493 	init_delay_drop(dev);
5494 
5495 	return 0;
5496 }
5497 
5498 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
5499 {
5500 	cancel_delay_drop(dev);
5501 }
5502 
5503 int mlx5_ib_stage_class_attr_init(struct mlx5_ib_dev *dev)
5504 {
5505 	int err;
5506 	int i;
5507 
5508 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
5509 		err = device_create_file(&dev->ib_dev.dev,
5510 					 mlx5_class_attributes[i]);
5511 		if (err)
5512 			return err;
5513 	}
5514 
5515 	return 0;
5516 }
5517 
5518 static int mlx5_ib_stage_rep_reg_init(struct mlx5_ib_dev *dev)
5519 {
5520 	mlx5_ib_register_vport_reps(dev);
5521 
5522 	return 0;
5523 }
5524 
5525 static void mlx5_ib_stage_rep_reg_cleanup(struct mlx5_ib_dev *dev)
5526 {
5527 	mlx5_ib_unregister_vport_reps(dev);
5528 }
5529 
5530 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
5531 		      const struct mlx5_ib_profile *profile,
5532 		      int stage)
5533 {
5534 	/* Number of stages to cleanup */
5535 	while (stage) {
5536 		stage--;
5537 		if (profile->stage[stage].cleanup)
5538 			profile->stage[stage].cleanup(dev);
5539 	}
5540 
5541 	ib_dealloc_device((struct ib_device *)dev);
5542 }
5543 
5544 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev, u8 port_num);
5545 
5546 void *__mlx5_ib_add(struct mlx5_ib_dev *dev,
5547 		    const struct mlx5_ib_profile *profile)
5548 {
5549 	int err;
5550 	int i;
5551 
5552 	printk_once(KERN_INFO "%s", mlx5_version);
5553 
5554 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
5555 		if (profile->stage[i].init) {
5556 			err = profile->stage[i].init(dev);
5557 			if (err)
5558 				goto err_out;
5559 		}
5560 	}
5561 
5562 	dev->profile = profile;
5563 	dev->ib_active = true;
5564 
5565 	return dev;
5566 
5567 err_out:
5568 	__mlx5_ib_remove(dev, profile, i);
5569 
5570 	return NULL;
5571 }
5572 
5573 static const struct mlx5_ib_profile pf_profile = {
5574 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
5575 		     mlx5_ib_stage_init_init,
5576 		     mlx5_ib_stage_init_cleanup),
5577 	STAGE_CREATE(MLX5_IB_STAGE_FLOW_DB,
5578 		     mlx5_ib_stage_flow_db_init,
5579 		     mlx5_ib_stage_flow_db_cleanup),
5580 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
5581 		     mlx5_ib_stage_caps_init,
5582 		     NULL),
5583 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
5584 		     mlx5_ib_stage_non_default_cb,
5585 		     NULL),
5586 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
5587 		     mlx5_ib_stage_roce_init,
5588 		     mlx5_ib_stage_roce_cleanup),
5589 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
5590 		     mlx5_ib_stage_dev_res_init,
5591 		     mlx5_ib_stage_dev_res_cleanup),
5592 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
5593 		     mlx5_ib_stage_odp_init,
5594 		     NULL),
5595 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
5596 		     mlx5_ib_stage_counters_init,
5597 		     mlx5_ib_stage_counters_cleanup),
5598 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
5599 		     mlx5_ib_stage_cong_debugfs_init,
5600 		     mlx5_ib_stage_cong_debugfs_cleanup),
5601 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
5602 		     mlx5_ib_stage_uar_init,
5603 		     mlx5_ib_stage_uar_cleanup),
5604 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
5605 		     mlx5_ib_stage_bfrag_init,
5606 		     mlx5_ib_stage_bfrag_cleanup),
5607 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
5608 		     NULL,
5609 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
5610 	STAGE_CREATE(MLX5_IB_STAGE_SPECS,
5611 		     mlx5_ib_stage_populate_specs,
5612 		     mlx5_ib_stage_depopulate_specs),
5613 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
5614 		     mlx5_ib_stage_ib_reg_init,
5615 		     mlx5_ib_stage_ib_reg_cleanup),
5616 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
5617 		     mlx5_ib_stage_post_ib_reg_umr_init,
5618 		     NULL),
5619 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
5620 		     mlx5_ib_stage_delay_drop_init,
5621 		     mlx5_ib_stage_delay_drop_cleanup),
5622 	STAGE_CREATE(MLX5_IB_STAGE_CLASS_ATTR,
5623 		     mlx5_ib_stage_class_attr_init,
5624 		     NULL),
5625 };
5626 
5627 static const struct mlx5_ib_profile nic_rep_profile = {
5628 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
5629 		     mlx5_ib_stage_init_init,
5630 		     mlx5_ib_stage_init_cleanup),
5631 	STAGE_CREATE(MLX5_IB_STAGE_FLOW_DB,
5632 		     mlx5_ib_stage_flow_db_init,
5633 		     mlx5_ib_stage_flow_db_cleanup),
5634 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
5635 		     mlx5_ib_stage_caps_init,
5636 		     NULL),
5637 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
5638 		     mlx5_ib_stage_rep_non_default_cb,
5639 		     NULL),
5640 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
5641 		     mlx5_ib_stage_rep_roce_init,
5642 		     mlx5_ib_stage_rep_roce_cleanup),
5643 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
5644 		     mlx5_ib_stage_dev_res_init,
5645 		     mlx5_ib_stage_dev_res_cleanup),
5646 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
5647 		     mlx5_ib_stage_counters_init,
5648 		     mlx5_ib_stage_counters_cleanup),
5649 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
5650 		     mlx5_ib_stage_uar_init,
5651 		     mlx5_ib_stage_uar_cleanup),
5652 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
5653 		     mlx5_ib_stage_bfrag_init,
5654 		     mlx5_ib_stage_bfrag_cleanup),
5655 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
5656 		     NULL,
5657 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
5658 	STAGE_CREATE(MLX5_IB_STAGE_SPECS,
5659 		     mlx5_ib_stage_populate_specs,
5660 		     mlx5_ib_stage_depopulate_specs),
5661 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
5662 		     mlx5_ib_stage_ib_reg_init,
5663 		     mlx5_ib_stage_ib_reg_cleanup),
5664 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
5665 		     mlx5_ib_stage_post_ib_reg_umr_init,
5666 		     NULL),
5667 	STAGE_CREATE(MLX5_IB_STAGE_CLASS_ATTR,
5668 		     mlx5_ib_stage_class_attr_init,
5669 		     NULL),
5670 	STAGE_CREATE(MLX5_IB_STAGE_REP_REG,
5671 		     mlx5_ib_stage_rep_reg_init,
5672 		     mlx5_ib_stage_rep_reg_cleanup),
5673 };
5674 
5675 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev, u8 port_num)
5676 {
5677 	struct mlx5_ib_multiport_info *mpi;
5678 	struct mlx5_ib_dev *dev;
5679 	bool bound = false;
5680 	int err;
5681 
5682 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
5683 	if (!mpi)
5684 		return NULL;
5685 
5686 	mpi->mdev = mdev;
5687 
5688 	err = mlx5_query_nic_vport_system_image_guid(mdev,
5689 						     &mpi->sys_image_guid);
5690 	if (err) {
5691 		kfree(mpi);
5692 		return NULL;
5693 	}
5694 
5695 	mutex_lock(&mlx5_ib_multiport_mutex);
5696 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
5697 		if (dev->sys_image_guid == mpi->sys_image_guid)
5698 			bound = mlx5_ib_bind_slave_port(dev, mpi);
5699 
5700 		if (bound) {
5701 			rdma_roce_rescan_device(&dev->ib_dev);
5702 			break;
5703 		}
5704 	}
5705 
5706 	if (!bound) {
5707 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
5708 		dev_dbg(&mdev->pdev->dev, "no suitable IB device found to bind to, added to unaffiliated list.\n");
5709 	} else {
5710 		mlx5_ib_dbg(dev, "bound port %u\n", port_num + 1);
5711 	}
5712 	mutex_unlock(&mlx5_ib_multiport_mutex);
5713 
5714 	return mpi;
5715 }
5716 
5717 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
5718 {
5719 	enum rdma_link_layer ll;
5720 	struct mlx5_ib_dev *dev;
5721 	int port_type_cap;
5722 
5723 	printk_once(KERN_INFO "%s", mlx5_version);
5724 
5725 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5726 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5727 
5728 	if (mlx5_core_is_mp_slave(mdev) && ll == IB_LINK_LAYER_ETHERNET) {
5729 		u8 port_num = mlx5_core_native_port_num(mdev) - 1;
5730 
5731 		return mlx5_ib_add_slave_port(mdev, port_num);
5732 	}
5733 
5734 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
5735 	if (!dev)
5736 		return NULL;
5737 
5738 	dev->mdev = mdev;
5739 	dev->num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
5740 			     MLX5_CAP_GEN(mdev, num_vhca_ports));
5741 
5742 	if (MLX5_VPORT_MANAGER(mdev) &&
5743 	    mlx5_ib_eswitch_mode(mdev->priv.eswitch) == SRIOV_OFFLOADS) {
5744 		dev->rep = mlx5_ib_vport_rep(mdev->priv.eswitch, 0);
5745 
5746 		return __mlx5_ib_add(dev, &nic_rep_profile);
5747 	}
5748 
5749 	return __mlx5_ib_add(dev, &pf_profile);
5750 }
5751 
5752 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
5753 {
5754 	struct mlx5_ib_multiport_info *mpi;
5755 	struct mlx5_ib_dev *dev;
5756 
5757 	if (mlx5_core_is_mp_slave(mdev)) {
5758 		mpi = context;
5759 		mutex_lock(&mlx5_ib_multiport_mutex);
5760 		if (mpi->ibdev)
5761 			mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
5762 		list_del(&mpi->list);
5763 		mutex_unlock(&mlx5_ib_multiport_mutex);
5764 		return;
5765 	}
5766 
5767 	dev = context;
5768 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
5769 }
5770 
5771 static struct mlx5_interface mlx5_ib_interface = {
5772 	.add            = mlx5_ib_add,
5773 	.remove         = mlx5_ib_remove,
5774 	.event          = mlx5_ib_event,
5775 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
5776 	.pfault		= mlx5_ib_pfault,
5777 #endif
5778 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
5779 };
5780 
5781 unsigned long mlx5_ib_get_xlt_emergency_page(void)
5782 {
5783 	mutex_lock(&xlt_emergency_page_mutex);
5784 	return xlt_emergency_page;
5785 }
5786 
5787 void mlx5_ib_put_xlt_emergency_page(void)
5788 {
5789 	mutex_unlock(&xlt_emergency_page_mutex);
5790 }
5791 
5792 static int __init mlx5_ib_init(void)
5793 {
5794 	int err;
5795 
5796 	xlt_emergency_page = __get_free_page(GFP_KERNEL);
5797 	if (!xlt_emergency_page)
5798 		return -ENOMEM;
5799 
5800 	mutex_init(&xlt_emergency_page_mutex);
5801 
5802 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
5803 	if (!mlx5_ib_event_wq) {
5804 		free_page(xlt_emergency_page);
5805 		return -ENOMEM;
5806 	}
5807 
5808 	mlx5_ib_odp_init();
5809 
5810 	err = mlx5_register_interface(&mlx5_ib_interface);
5811 
5812 	return err;
5813 }
5814 
5815 static void __exit mlx5_ib_cleanup(void)
5816 {
5817 	mlx5_unregister_interface(&mlx5_ib_interface);
5818 	destroy_workqueue(mlx5_ib_event_wq);
5819 	mutex_destroy(&xlt_emergency_page_mutex);
5820 	free_page(xlt_emergency_page);
5821 }
5822 
5823 module_init(mlx5_ib_init);
5824 module_exit(mlx5_ib_cleanup);
5825