xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision addee42a)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/debugfs.h>
34 #include <linux/highmem.h>
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/errno.h>
38 #include <linux/pci.h>
39 #include <linux/dma-mapping.h>
40 #include <linux/slab.h>
41 #include <linux/bitmap.h>
42 #if defined(CONFIG_X86)
43 #include <asm/pat.h>
44 #endif
45 #include <linux/sched.h>
46 #include <linux/sched/mm.h>
47 #include <linux/sched/task.h>
48 #include <linux/delay.h>
49 #include <rdma/ib_user_verbs.h>
50 #include <rdma/ib_addr.h>
51 #include <rdma/ib_cache.h>
52 #include <linux/mlx5/port.h>
53 #include <linux/mlx5/vport.h>
54 #include <linux/mlx5/fs.h>
55 #include <linux/list.h>
56 #include <rdma/ib_smi.h>
57 #include <rdma/ib_umem.h>
58 #include <linux/in.h>
59 #include <linux/etherdevice.h>
60 #include "mlx5_ib.h"
61 #include "ib_rep.h"
62 #include "cmd.h"
63 #include <linux/mlx5/fs_helpers.h>
64 #include <linux/mlx5/accel.h>
65 #include <rdma/uverbs_std_types.h>
66 #include <rdma/mlx5_user_ioctl_verbs.h>
67 #include <rdma/mlx5_user_ioctl_cmds.h>
68 
69 #define UVERBS_MODULE_NAME mlx5_ib
70 #include <rdma/uverbs_named_ioctl.h>
71 
72 #define DRIVER_NAME "mlx5_ib"
73 #define DRIVER_VERSION "5.0-0"
74 
75 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
76 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
77 MODULE_LICENSE("Dual BSD/GPL");
78 
79 static char mlx5_version[] =
80 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
81 	DRIVER_VERSION "\n";
82 
83 struct mlx5_ib_event_work {
84 	struct work_struct	work;
85 	struct mlx5_core_dev	*dev;
86 	void			*context;
87 	enum mlx5_dev_event	event;
88 	unsigned long		param;
89 };
90 
91 enum {
92 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
93 };
94 
95 static struct workqueue_struct *mlx5_ib_event_wq;
96 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
97 static LIST_HEAD(mlx5_ib_dev_list);
98 /*
99  * This mutex should be held when accessing either of the above lists
100  */
101 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
102 
103 /* We can't use an array for xlt_emergency_page because dma_map_single
104  * doesn't work on kernel modules memory
105  */
106 static unsigned long xlt_emergency_page;
107 static struct mutex xlt_emergency_page_mutex;
108 
109 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
110 {
111 	struct mlx5_ib_dev *dev;
112 
113 	mutex_lock(&mlx5_ib_multiport_mutex);
114 	dev = mpi->ibdev;
115 	mutex_unlock(&mlx5_ib_multiport_mutex);
116 	return dev;
117 }
118 
119 static enum rdma_link_layer
120 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
121 {
122 	switch (port_type_cap) {
123 	case MLX5_CAP_PORT_TYPE_IB:
124 		return IB_LINK_LAYER_INFINIBAND;
125 	case MLX5_CAP_PORT_TYPE_ETH:
126 		return IB_LINK_LAYER_ETHERNET;
127 	default:
128 		return IB_LINK_LAYER_UNSPECIFIED;
129 	}
130 }
131 
132 static enum rdma_link_layer
133 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
134 {
135 	struct mlx5_ib_dev *dev = to_mdev(device);
136 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
137 
138 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
139 }
140 
141 static int get_port_state(struct ib_device *ibdev,
142 			  u8 port_num,
143 			  enum ib_port_state *state)
144 {
145 	struct ib_port_attr attr;
146 	int ret;
147 
148 	memset(&attr, 0, sizeof(attr));
149 	ret = ibdev->query_port(ibdev, port_num, &attr);
150 	if (!ret)
151 		*state = attr.state;
152 	return ret;
153 }
154 
155 static int mlx5_netdev_event(struct notifier_block *this,
156 			     unsigned long event, void *ptr)
157 {
158 	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
159 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
160 	u8 port_num = roce->native_port_num;
161 	struct mlx5_core_dev *mdev;
162 	struct mlx5_ib_dev *ibdev;
163 
164 	ibdev = roce->dev;
165 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
166 	if (!mdev)
167 		return NOTIFY_DONE;
168 
169 	switch (event) {
170 	case NETDEV_REGISTER:
171 	case NETDEV_UNREGISTER:
172 		write_lock(&roce->netdev_lock);
173 		if (ibdev->rep) {
174 			struct mlx5_eswitch *esw = ibdev->mdev->priv.eswitch;
175 			struct net_device *rep_ndev;
176 
177 			rep_ndev = mlx5_ib_get_rep_netdev(esw,
178 							  ibdev->rep->vport);
179 			if (rep_ndev == ndev)
180 				roce->netdev = (event == NETDEV_UNREGISTER) ?
181 					NULL : ndev;
182 		} else if (ndev->dev.parent == &mdev->pdev->dev) {
183 			roce->netdev = (event == NETDEV_UNREGISTER) ?
184 				NULL : ndev;
185 		}
186 		write_unlock(&roce->netdev_lock);
187 		break;
188 
189 	case NETDEV_CHANGE:
190 	case NETDEV_UP:
191 	case NETDEV_DOWN: {
192 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
193 		struct net_device *upper = NULL;
194 
195 		if (lag_ndev) {
196 			upper = netdev_master_upper_dev_get(lag_ndev);
197 			dev_put(lag_ndev);
198 		}
199 
200 		if ((upper == ndev || (!upper && ndev == roce->netdev))
201 		    && ibdev->ib_active) {
202 			struct ib_event ibev = { };
203 			enum ib_port_state port_state;
204 
205 			if (get_port_state(&ibdev->ib_dev, port_num,
206 					   &port_state))
207 				goto done;
208 
209 			if (roce->last_port_state == port_state)
210 				goto done;
211 
212 			roce->last_port_state = port_state;
213 			ibev.device = &ibdev->ib_dev;
214 			if (port_state == IB_PORT_DOWN)
215 				ibev.event = IB_EVENT_PORT_ERR;
216 			else if (port_state == IB_PORT_ACTIVE)
217 				ibev.event = IB_EVENT_PORT_ACTIVE;
218 			else
219 				goto done;
220 
221 			ibev.element.port_num = port_num;
222 			ib_dispatch_event(&ibev);
223 		}
224 		break;
225 	}
226 
227 	default:
228 		break;
229 	}
230 done:
231 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
232 	return NOTIFY_DONE;
233 }
234 
235 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
236 					     u8 port_num)
237 {
238 	struct mlx5_ib_dev *ibdev = to_mdev(device);
239 	struct net_device *ndev;
240 	struct mlx5_core_dev *mdev;
241 
242 	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
243 	if (!mdev)
244 		return NULL;
245 
246 	ndev = mlx5_lag_get_roce_netdev(mdev);
247 	if (ndev)
248 		goto out;
249 
250 	/* Ensure ndev does not disappear before we invoke dev_hold()
251 	 */
252 	read_lock(&ibdev->roce[port_num - 1].netdev_lock);
253 	ndev = ibdev->roce[port_num - 1].netdev;
254 	if (ndev)
255 		dev_hold(ndev);
256 	read_unlock(&ibdev->roce[port_num - 1].netdev_lock);
257 
258 out:
259 	mlx5_ib_put_native_port_mdev(ibdev, port_num);
260 	return ndev;
261 }
262 
263 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
264 						   u8 ib_port_num,
265 						   u8 *native_port_num)
266 {
267 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
268 							  ib_port_num);
269 	struct mlx5_core_dev *mdev = NULL;
270 	struct mlx5_ib_multiport_info *mpi;
271 	struct mlx5_ib_port *port;
272 
273 	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
274 	    ll != IB_LINK_LAYER_ETHERNET) {
275 		if (native_port_num)
276 			*native_port_num = ib_port_num;
277 		return ibdev->mdev;
278 	}
279 
280 	if (native_port_num)
281 		*native_port_num = 1;
282 
283 	port = &ibdev->port[ib_port_num - 1];
284 	if (!port)
285 		return NULL;
286 
287 	spin_lock(&port->mp.mpi_lock);
288 	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
289 	if (mpi && !mpi->unaffiliate) {
290 		mdev = mpi->mdev;
291 		/* If it's the master no need to refcount, it'll exist
292 		 * as long as the ib_dev exists.
293 		 */
294 		if (!mpi->is_master)
295 			mpi->mdev_refcnt++;
296 	}
297 	spin_unlock(&port->mp.mpi_lock);
298 
299 	return mdev;
300 }
301 
302 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u8 port_num)
303 {
304 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
305 							  port_num);
306 	struct mlx5_ib_multiport_info *mpi;
307 	struct mlx5_ib_port *port;
308 
309 	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
310 		return;
311 
312 	port = &ibdev->port[port_num - 1];
313 
314 	spin_lock(&port->mp.mpi_lock);
315 	mpi = ibdev->port[port_num - 1].mp.mpi;
316 	if (mpi->is_master)
317 		goto out;
318 
319 	mpi->mdev_refcnt--;
320 	if (mpi->unaffiliate)
321 		complete(&mpi->unref_comp);
322 out:
323 	spin_unlock(&port->mp.mpi_lock);
324 }
325 
326 static int translate_eth_proto_oper(u32 eth_proto_oper, u8 *active_speed,
327 				    u8 *active_width)
328 {
329 	switch (eth_proto_oper) {
330 	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
331 	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
332 	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
333 	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
334 		*active_width = IB_WIDTH_1X;
335 		*active_speed = IB_SPEED_SDR;
336 		break;
337 	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
338 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
339 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
340 	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
341 	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
342 	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
343 	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
344 		*active_width = IB_WIDTH_1X;
345 		*active_speed = IB_SPEED_QDR;
346 		break;
347 	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
348 	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
349 	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
350 		*active_width = IB_WIDTH_1X;
351 		*active_speed = IB_SPEED_EDR;
352 		break;
353 	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
354 	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
355 	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
356 	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
357 		*active_width = IB_WIDTH_4X;
358 		*active_speed = IB_SPEED_QDR;
359 		break;
360 	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
361 	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
362 	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
363 		*active_width = IB_WIDTH_1X;
364 		*active_speed = IB_SPEED_HDR;
365 		break;
366 	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
367 		*active_width = IB_WIDTH_4X;
368 		*active_speed = IB_SPEED_FDR;
369 		break;
370 	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
371 	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
372 	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
373 	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
374 		*active_width = IB_WIDTH_4X;
375 		*active_speed = IB_SPEED_EDR;
376 		break;
377 	default:
378 		return -EINVAL;
379 	}
380 
381 	return 0;
382 }
383 
384 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
385 				struct ib_port_attr *props)
386 {
387 	struct mlx5_ib_dev *dev = to_mdev(device);
388 	struct mlx5_core_dev *mdev;
389 	struct net_device *ndev, *upper;
390 	enum ib_mtu ndev_ib_mtu;
391 	bool put_mdev = true;
392 	u16 qkey_viol_cntr;
393 	u32 eth_prot_oper;
394 	u8 mdev_port_num;
395 	int err;
396 
397 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
398 	if (!mdev) {
399 		/* This means the port isn't affiliated yet. Get the
400 		 * info for the master port instead.
401 		 */
402 		put_mdev = false;
403 		mdev = dev->mdev;
404 		mdev_port_num = 1;
405 		port_num = 1;
406 	}
407 
408 	/* Possible bad flows are checked before filling out props so in case
409 	 * of an error it will still be zeroed out.
410 	 */
411 	err = mlx5_query_port_eth_proto_oper(mdev, &eth_prot_oper,
412 					     mdev_port_num);
413 	if (err)
414 		goto out;
415 
416 	props->active_width     = IB_WIDTH_4X;
417 	props->active_speed     = IB_SPEED_QDR;
418 
419 	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
420 				 &props->active_width);
421 
422 	props->port_cap_flags  |= IB_PORT_CM_SUP;
423 	props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
424 
425 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
426 						roce_address_table_size);
427 	props->max_mtu          = IB_MTU_4096;
428 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
429 	props->pkey_tbl_len     = 1;
430 	props->state            = IB_PORT_DOWN;
431 	props->phys_state       = 3;
432 
433 	mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
434 	props->qkey_viol_cntr = qkey_viol_cntr;
435 
436 	/* If this is a stub query for an unaffiliated port stop here */
437 	if (!put_mdev)
438 		goto out;
439 
440 	ndev = mlx5_ib_get_netdev(device, port_num);
441 	if (!ndev)
442 		goto out;
443 
444 	if (mlx5_lag_is_active(dev->mdev)) {
445 		rcu_read_lock();
446 		upper = netdev_master_upper_dev_get_rcu(ndev);
447 		if (upper) {
448 			dev_put(ndev);
449 			ndev = upper;
450 			dev_hold(ndev);
451 		}
452 		rcu_read_unlock();
453 	}
454 
455 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
456 		props->state      = IB_PORT_ACTIVE;
457 		props->phys_state = 5;
458 	}
459 
460 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
461 
462 	dev_put(ndev);
463 
464 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
465 out:
466 	if (put_mdev)
467 		mlx5_ib_put_native_port_mdev(dev, port_num);
468 	return err;
469 }
470 
471 static int set_roce_addr(struct mlx5_ib_dev *dev, u8 port_num,
472 			 unsigned int index, const union ib_gid *gid,
473 			 const struct ib_gid_attr *attr)
474 {
475 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
476 	u8 roce_version = 0;
477 	u8 roce_l3_type = 0;
478 	bool vlan = false;
479 	u8 mac[ETH_ALEN];
480 	u16 vlan_id = 0;
481 
482 	if (gid) {
483 		gid_type = attr->gid_type;
484 		ether_addr_copy(mac, attr->ndev->dev_addr);
485 
486 		if (is_vlan_dev(attr->ndev)) {
487 			vlan = true;
488 			vlan_id = vlan_dev_vlan_id(attr->ndev);
489 		}
490 	}
491 
492 	switch (gid_type) {
493 	case IB_GID_TYPE_IB:
494 		roce_version = MLX5_ROCE_VERSION_1;
495 		break;
496 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
497 		roce_version = MLX5_ROCE_VERSION_2;
498 		if (ipv6_addr_v4mapped((void *)gid))
499 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
500 		else
501 			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
502 		break;
503 
504 	default:
505 		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
506 	}
507 
508 	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
509 				      roce_l3_type, gid->raw, mac, vlan,
510 				      vlan_id, port_num);
511 }
512 
513 static int mlx5_ib_add_gid(const union ib_gid *gid,
514 			   const struct ib_gid_attr *attr,
515 			   __always_unused void **context)
516 {
517 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
518 			     attr->index, gid, attr);
519 }
520 
521 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
522 			   __always_unused void **context)
523 {
524 	return set_roce_addr(to_mdev(attr->device), attr->port_num,
525 			     attr->index, NULL, NULL);
526 }
527 
528 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
529 			       int index)
530 {
531 	struct ib_gid_attr attr;
532 	union ib_gid gid;
533 
534 	if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
535 		return 0;
536 
537 	dev_put(attr.ndev);
538 
539 	if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
540 		return 0;
541 
542 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
543 }
544 
545 int mlx5_get_roce_gid_type(struct mlx5_ib_dev *dev, u8 port_num,
546 			   int index, enum ib_gid_type *gid_type)
547 {
548 	struct ib_gid_attr attr;
549 	union ib_gid gid;
550 	int ret;
551 
552 	ret = ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr);
553 	if (ret)
554 		return ret;
555 
556 	dev_put(attr.ndev);
557 
558 	*gid_type = attr.gid_type;
559 
560 	return 0;
561 }
562 
563 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
564 {
565 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
566 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
567 	return 0;
568 }
569 
570 enum {
571 	MLX5_VPORT_ACCESS_METHOD_MAD,
572 	MLX5_VPORT_ACCESS_METHOD_HCA,
573 	MLX5_VPORT_ACCESS_METHOD_NIC,
574 };
575 
576 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
577 {
578 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
579 		return MLX5_VPORT_ACCESS_METHOD_MAD;
580 
581 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
582 	    IB_LINK_LAYER_ETHERNET)
583 		return MLX5_VPORT_ACCESS_METHOD_NIC;
584 
585 	return MLX5_VPORT_ACCESS_METHOD_HCA;
586 }
587 
588 static void get_atomic_caps(struct mlx5_ib_dev *dev,
589 			    u8 atomic_size_qp,
590 			    struct ib_device_attr *props)
591 {
592 	u8 tmp;
593 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
594 	u8 atomic_req_8B_endianness_mode =
595 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
596 
597 	/* Check if HW supports 8 bytes standard atomic operations and capable
598 	 * of host endianness respond
599 	 */
600 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
601 	if (((atomic_operations & tmp) == tmp) &&
602 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
603 	    (atomic_req_8B_endianness_mode)) {
604 		props->atomic_cap = IB_ATOMIC_HCA;
605 	} else {
606 		props->atomic_cap = IB_ATOMIC_NONE;
607 	}
608 }
609 
610 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
611 			       struct ib_device_attr *props)
612 {
613 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
614 
615 	get_atomic_caps(dev, atomic_size_qp, props);
616 }
617 
618 static void get_atomic_caps_dc(struct mlx5_ib_dev *dev,
619 			       struct ib_device_attr *props)
620 {
621 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
622 
623 	get_atomic_caps(dev, atomic_size_qp, props);
624 }
625 
626 bool mlx5_ib_dc_atomic_is_supported(struct mlx5_ib_dev *dev)
627 {
628 	struct ib_device_attr props = {};
629 
630 	get_atomic_caps_dc(dev, &props);
631 	return (props.atomic_cap == IB_ATOMIC_HCA) ? true : false;
632 }
633 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
634 					__be64 *sys_image_guid)
635 {
636 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
637 	struct mlx5_core_dev *mdev = dev->mdev;
638 	u64 tmp;
639 	int err;
640 
641 	switch (mlx5_get_vport_access_method(ibdev)) {
642 	case MLX5_VPORT_ACCESS_METHOD_MAD:
643 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
644 							    sys_image_guid);
645 
646 	case MLX5_VPORT_ACCESS_METHOD_HCA:
647 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
648 		break;
649 
650 	case MLX5_VPORT_ACCESS_METHOD_NIC:
651 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
652 		break;
653 
654 	default:
655 		return -EINVAL;
656 	}
657 
658 	if (!err)
659 		*sys_image_guid = cpu_to_be64(tmp);
660 
661 	return err;
662 
663 }
664 
665 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
666 				u16 *max_pkeys)
667 {
668 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
669 	struct mlx5_core_dev *mdev = dev->mdev;
670 
671 	switch (mlx5_get_vport_access_method(ibdev)) {
672 	case MLX5_VPORT_ACCESS_METHOD_MAD:
673 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
674 
675 	case MLX5_VPORT_ACCESS_METHOD_HCA:
676 	case MLX5_VPORT_ACCESS_METHOD_NIC:
677 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
678 						pkey_table_size));
679 		return 0;
680 
681 	default:
682 		return -EINVAL;
683 	}
684 }
685 
686 static int mlx5_query_vendor_id(struct ib_device *ibdev,
687 				u32 *vendor_id)
688 {
689 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
690 
691 	switch (mlx5_get_vport_access_method(ibdev)) {
692 	case MLX5_VPORT_ACCESS_METHOD_MAD:
693 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
694 
695 	case MLX5_VPORT_ACCESS_METHOD_HCA:
696 	case MLX5_VPORT_ACCESS_METHOD_NIC:
697 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
698 
699 	default:
700 		return -EINVAL;
701 	}
702 }
703 
704 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
705 				__be64 *node_guid)
706 {
707 	u64 tmp;
708 	int err;
709 
710 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
711 	case MLX5_VPORT_ACCESS_METHOD_MAD:
712 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
713 
714 	case MLX5_VPORT_ACCESS_METHOD_HCA:
715 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
716 		break;
717 
718 	case MLX5_VPORT_ACCESS_METHOD_NIC:
719 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
720 		break;
721 
722 	default:
723 		return -EINVAL;
724 	}
725 
726 	if (!err)
727 		*node_guid = cpu_to_be64(tmp);
728 
729 	return err;
730 }
731 
732 struct mlx5_reg_node_desc {
733 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
734 };
735 
736 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
737 {
738 	struct mlx5_reg_node_desc in;
739 
740 	if (mlx5_use_mad_ifc(dev))
741 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
742 
743 	memset(&in, 0, sizeof(in));
744 
745 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
746 				    sizeof(struct mlx5_reg_node_desc),
747 				    MLX5_REG_NODE_DESC, 0, 0);
748 }
749 
750 static int mlx5_ib_query_device(struct ib_device *ibdev,
751 				struct ib_device_attr *props,
752 				struct ib_udata *uhw)
753 {
754 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
755 	struct mlx5_core_dev *mdev = dev->mdev;
756 	int err = -ENOMEM;
757 	int max_sq_desc;
758 	int max_rq_sg;
759 	int max_sq_sg;
760 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
761 	bool raw_support = !mlx5_core_mp_enabled(mdev);
762 	struct mlx5_ib_query_device_resp resp = {};
763 	size_t resp_len;
764 	u64 max_tso;
765 
766 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
767 	if (uhw->outlen && uhw->outlen < resp_len)
768 		return -EINVAL;
769 	else
770 		resp.response_length = resp_len;
771 
772 	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
773 		return -EINVAL;
774 
775 	memset(props, 0, sizeof(*props));
776 	err = mlx5_query_system_image_guid(ibdev,
777 					   &props->sys_image_guid);
778 	if (err)
779 		return err;
780 
781 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
782 	if (err)
783 		return err;
784 
785 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
786 	if (err)
787 		return err;
788 
789 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
790 		(fw_rev_min(dev->mdev) << 16) |
791 		fw_rev_sub(dev->mdev);
792 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
793 		IB_DEVICE_PORT_ACTIVE_EVENT		|
794 		IB_DEVICE_SYS_IMAGE_GUID		|
795 		IB_DEVICE_RC_RNR_NAK_GEN;
796 
797 	if (MLX5_CAP_GEN(mdev, pkv))
798 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
799 	if (MLX5_CAP_GEN(mdev, qkv))
800 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
801 	if (MLX5_CAP_GEN(mdev, apm))
802 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
803 	if (MLX5_CAP_GEN(mdev, xrc))
804 		props->device_cap_flags |= IB_DEVICE_XRC;
805 	if (MLX5_CAP_GEN(mdev, imaicl)) {
806 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
807 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
808 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
809 		/* We support 'Gappy' memory registration too */
810 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
811 	}
812 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
813 	if (MLX5_CAP_GEN(mdev, sho)) {
814 		props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
815 		/* At this stage no support for signature handover */
816 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
817 				      IB_PROT_T10DIF_TYPE_2 |
818 				      IB_PROT_T10DIF_TYPE_3;
819 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
820 				       IB_GUARD_T10DIF_CSUM;
821 	}
822 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
823 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
824 
825 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
826 		if (MLX5_CAP_ETH(mdev, csum_cap)) {
827 			/* Legacy bit to support old userspace libraries */
828 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
829 			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
830 		}
831 
832 		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
833 			props->raw_packet_caps |=
834 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
835 
836 		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
837 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
838 			if (max_tso) {
839 				resp.tso_caps.max_tso = 1 << max_tso;
840 				resp.tso_caps.supported_qpts |=
841 					1 << IB_QPT_RAW_PACKET;
842 				resp.response_length += sizeof(resp.tso_caps);
843 			}
844 		}
845 
846 		if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
847 			resp.rss_caps.rx_hash_function =
848 						MLX5_RX_HASH_FUNC_TOEPLITZ;
849 			resp.rss_caps.rx_hash_fields_mask =
850 						MLX5_RX_HASH_SRC_IPV4 |
851 						MLX5_RX_HASH_DST_IPV4 |
852 						MLX5_RX_HASH_SRC_IPV6 |
853 						MLX5_RX_HASH_DST_IPV6 |
854 						MLX5_RX_HASH_SRC_PORT_TCP |
855 						MLX5_RX_HASH_DST_PORT_TCP |
856 						MLX5_RX_HASH_SRC_PORT_UDP |
857 						MLX5_RX_HASH_DST_PORT_UDP |
858 						MLX5_RX_HASH_INNER;
859 			if (mlx5_accel_ipsec_device_caps(dev->mdev) &
860 			    MLX5_ACCEL_IPSEC_CAP_DEVICE)
861 				resp.rss_caps.rx_hash_fields_mask |=
862 					MLX5_RX_HASH_IPSEC_SPI;
863 			resp.response_length += sizeof(resp.rss_caps);
864 		}
865 	} else {
866 		if (field_avail(typeof(resp), tso_caps, uhw->outlen))
867 			resp.response_length += sizeof(resp.tso_caps);
868 		if (field_avail(typeof(resp), rss_caps, uhw->outlen))
869 			resp.response_length += sizeof(resp.rss_caps);
870 	}
871 
872 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
873 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
874 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
875 	}
876 
877 	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
878 	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
879 	    raw_support)
880 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
881 
882 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
883 	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
884 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
885 
886 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
887 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
888 	    raw_support) {
889 		/* Legacy bit to support old userspace libraries */
890 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
891 		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
892 	}
893 
894 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
895 		props->max_dm_size =
896 			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
897 	}
898 
899 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
900 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
901 
902 	if (MLX5_CAP_GEN(mdev, end_pad))
903 		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
904 
905 	props->vendor_part_id	   = mdev->pdev->device;
906 	props->hw_ver		   = mdev->pdev->revision;
907 
908 	props->max_mr_size	   = ~0ull;
909 	props->page_size_cap	   = ~(min_page_size - 1);
910 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
911 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
912 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
913 		     sizeof(struct mlx5_wqe_data_seg);
914 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
915 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
916 		     sizeof(struct mlx5_wqe_raddr_seg)) /
917 		sizeof(struct mlx5_wqe_data_seg);
918 	props->max_sge = min(max_rq_sg, max_sq_sg);
919 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
920 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
921 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
922 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
923 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
924 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
925 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
926 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
927 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
928 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
929 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
930 	props->max_srq_sge	   = max_rq_sg - 1;
931 	props->max_fast_reg_page_list_len =
932 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
933 	get_atomic_caps_qp(dev, props);
934 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
935 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
936 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
937 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
938 					   props->max_mcast_grp;
939 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
940 	props->max_ah = INT_MAX;
941 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
942 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
943 
944 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
945 	if (MLX5_CAP_GEN(mdev, pg))
946 		props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
947 	props->odp_caps = dev->odp_caps;
948 #endif
949 
950 	if (MLX5_CAP_GEN(mdev, cd))
951 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
952 
953 	if (!mlx5_core_is_pf(mdev))
954 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
955 
956 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
957 	    IB_LINK_LAYER_ETHERNET && raw_support) {
958 		props->rss_caps.max_rwq_indirection_tables =
959 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
960 		props->rss_caps.max_rwq_indirection_table_size =
961 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
962 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
963 		props->max_wq_type_rq =
964 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
965 	}
966 
967 	if (MLX5_CAP_GEN(mdev, tag_matching)) {
968 		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
969 		props->tm_caps.max_num_tags =
970 			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
971 		props->tm_caps.flags = IB_TM_CAP_RC;
972 		props->tm_caps.max_ops =
973 			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
974 		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
975 	}
976 
977 	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
978 		props->cq_caps.max_cq_moderation_count =
979 						MLX5_MAX_CQ_COUNT;
980 		props->cq_caps.max_cq_moderation_period =
981 						MLX5_MAX_CQ_PERIOD;
982 	}
983 
984 	if (field_avail(typeof(resp), cqe_comp_caps, uhw->outlen)) {
985 		resp.cqe_comp_caps.max_num =
986 			MLX5_CAP_GEN(dev->mdev, cqe_compression) ?
987 			MLX5_CAP_GEN(dev->mdev, cqe_compression_max_num) : 0;
988 		resp.cqe_comp_caps.supported_format =
989 			MLX5_IB_CQE_RES_FORMAT_HASH |
990 			MLX5_IB_CQE_RES_FORMAT_CSUM;
991 		resp.response_length += sizeof(resp.cqe_comp_caps);
992 	}
993 
994 	if (field_avail(typeof(resp), packet_pacing_caps, uhw->outlen) &&
995 	    raw_support) {
996 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
997 		    MLX5_CAP_GEN(mdev, qos)) {
998 			resp.packet_pacing_caps.qp_rate_limit_max =
999 				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1000 			resp.packet_pacing_caps.qp_rate_limit_min =
1001 				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1002 			resp.packet_pacing_caps.supported_qpts |=
1003 				1 << IB_QPT_RAW_PACKET;
1004 			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1005 			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1006 				resp.packet_pacing_caps.cap_flags |=
1007 					MLX5_IB_PP_SUPPORT_BURST;
1008 		}
1009 		resp.response_length += sizeof(resp.packet_pacing_caps);
1010 	}
1011 
1012 	if (field_avail(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes,
1013 			uhw->outlen)) {
1014 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1015 			resp.mlx5_ib_support_multi_pkt_send_wqes =
1016 				MLX5_IB_ALLOW_MPW;
1017 
1018 		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1019 			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1020 				MLX5_IB_SUPPORT_EMPW;
1021 
1022 		resp.response_length +=
1023 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1024 	}
1025 
1026 	if (field_avail(typeof(resp), flags, uhw->outlen)) {
1027 		resp.response_length += sizeof(resp.flags);
1028 
1029 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1030 			resp.flags |=
1031 				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1032 
1033 		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1034 			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1035 	}
1036 
1037 	if (field_avail(typeof(resp), sw_parsing_caps,
1038 			uhw->outlen)) {
1039 		resp.response_length += sizeof(resp.sw_parsing_caps);
1040 		if (MLX5_CAP_ETH(mdev, swp)) {
1041 			resp.sw_parsing_caps.sw_parsing_offloads |=
1042 				MLX5_IB_SW_PARSING;
1043 
1044 			if (MLX5_CAP_ETH(mdev, swp_csum))
1045 				resp.sw_parsing_caps.sw_parsing_offloads |=
1046 					MLX5_IB_SW_PARSING_CSUM;
1047 
1048 			if (MLX5_CAP_ETH(mdev, swp_lso))
1049 				resp.sw_parsing_caps.sw_parsing_offloads |=
1050 					MLX5_IB_SW_PARSING_LSO;
1051 
1052 			if (resp.sw_parsing_caps.sw_parsing_offloads)
1053 				resp.sw_parsing_caps.supported_qpts =
1054 					BIT(IB_QPT_RAW_PACKET);
1055 		}
1056 	}
1057 
1058 	if (field_avail(typeof(resp), striding_rq_caps, uhw->outlen) &&
1059 	    raw_support) {
1060 		resp.response_length += sizeof(resp.striding_rq_caps);
1061 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1062 			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1063 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1064 			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1065 				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1066 			resp.striding_rq_caps.min_single_wqe_log_num_of_strides =
1067 				MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1068 			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1069 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1070 			resp.striding_rq_caps.supported_qpts =
1071 				BIT(IB_QPT_RAW_PACKET);
1072 		}
1073 	}
1074 
1075 	if (field_avail(typeof(resp), tunnel_offloads_caps,
1076 			uhw->outlen)) {
1077 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1078 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1079 			resp.tunnel_offloads_caps |=
1080 				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1081 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1082 			resp.tunnel_offloads_caps |=
1083 				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1084 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1085 			resp.tunnel_offloads_caps |=
1086 				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1087 	}
1088 
1089 	if (uhw->outlen) {
1090 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1091 
1092 		if (err)
1093 			return err;
1094 	}
1095 
1096 	return 0;
1097 }
1098 
1099 enum mlx5_ib_width {
1100 	MLX5_IB_WIDTH_1X	= 1 << 0,
1101 	MLX5_IB_WIDTH_2X	= 1 << 1,
1102 	MLX5_IB_WIDTH_4X	= 1 << 2,
1103 	MLX5_IB_WIDTH_8X	= 1 << 3,
1104 	MLX5_IB_WIDTH_12X	= 1 << 4
1105 };
1106 
1107 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
1108 				  u8 *ib_width)
1109 {
1110 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1111 	int err = 0;
1112 
1113 	if (active_width & MLX5_IB_WIDTH_1X) {
1114 		*ib_width = IB_WIDTH_1X;
1115 	} else if (active_width & MLX5_IB_WIDTH_2X) {
1116 		mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
1117 			    (int)active_width);
1118 		err = -EINVAL;
1119 	} else if (active_width & MLX5_IB_WIDTH_4X) {
1120 		*ib_width = IB_WIDTH_4X;
1121 	} else if (active_width & MLX5_IB_WIDTH_8X) {
1122 		*ib_width = IB_WIDTH_8X;
1123 	} else if (active_width & MLX5_IB_WIDTH_12X) {
1124 		*ib_width = IB_WIDTH_12X;
1125 	} else {
1126 		mlx5_ib_dbg(dev, "Invalid active_width %d\n",
1127 			    (int)active_width);
1128 		err = -EINVAL;
1129 	}
1130 
1131 	return err;
1132 }
1133 
1134 static int mlx5_mtu_to_ib_mtu(int mtu)
1135 {
1136 	switch (mtu) {
1137 	case 256: return 1;
1138 	case 512: return 2;
1139 	case 1024: return 3;
1140 	case 2048: return 4;
1141 	case 4096: return 5;
1142 	default:
1143 		pr_warn("invalid mtu\n");
1144 		return -1;
1145 	}
1146 }
1147 
1148 enum ib_max_vl_num {
1149 	__IB_MAX_VL_0		= 1,
1150 	__IB_MAX_VL_0_1		= 2,
1151 	__IB_MAX_VL_0_3		= 3,
1152 	__IB_MAX_VL_0_7		= 4,
1153 	__IB_MAX_VL_0_14	= 5,
1154 };
1155 
1156 enum mlx5_vl_hw_cap {
1157 	MLX5_VL_HW_0	= 1,
1158 	MLX5_VL_HW_0_1	= 2,
1159 	MLX5_VL_HW_0_2	= 3,
1160 	MLX5_VL_HW_0_3	= 4,
1161 	MLX5_VL_HW_0_4	= 5,
1162 	MLX5_VL_HW_0_5	= 6,
1163 	MLX5_VL_HW_0_6	= 7,
1164 	MLX5_VL_HW_0_7	= 8,
1165 	MLX5_VL_HW_0_14	= 15
1166 };
1167 
1168 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1169 				u8 *max_vl_num)
1170 {
1171 	switch (vl_hw_cap) {
1172 	case MLX5_VL_HW_0:
1173 		*max_vl_num = __IB_MAX_VL_0;
1174 		break;
1175 	case MLX5_VL_HW_0_1:
1176 		*max_vl_num = __IB_MAX_VL_0_1;
1177 		break;
1178 	case MLX5_VL_HW_0_3:
1179 		*max_vl_num = __IB_MAX_VL_0_3;
1180 		break;
1181 	case MLX5_VL_HW_0_7:
1182 		*max_vl_num = __IB_MAX_VL_0_7;
1183 		break;
1184 	case MLX5_VL_HW_0_14:
1185 		*max_vl_num = __IB_MAX_VL_0_14;
1186 		break;
1187 
1188 	default:
1189 		return -EINVAL;
1190 	}
1191 
1192 	return 0;
1193 }
1194 
1195 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
1196 			       struct ib_port_attr *props)
1197 {
1198 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1199 	struct mlx5_core_dev *mdev = dev->mdev;
1200 	struct mlx5_hca_vport_context *rep;
1201 	u16 max_mtu;
1202 	u16 oper_mtu;
1203 	int err;
1204 	u8 ib_link_width_oper;
1205 	u8 vl_hw_cap;
1206 
1207 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1208 	if (!rep) {
1209 		err = -ENOMEM;
1210 		goto out;
1211 	}
1212 
1213 	/* props being zeroed by the caller, avoid zeroing it here */
1214 
1215 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1216 	if (err)
1217 		goto out;
1218 
1219 	props->lid		= rep->lid;
1220 	props->lmc		= rep->lmc;
1221 	props->sm_lid		= rep->sm_lid;
1222 	props->sm_sl		= rep->sm_sl;
1223 	props->state		= rep->vport_state;
1224 	props->phys_state	= rep->port_physical_state;
1225 	props->port_cap_flags	= rep->cap_mask1;
1226 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1227 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1228 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1229 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1230 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1231 	props->subnet_timeout	= rep->subnet_timeout;
1232 	props->init_type_reply	= rep->init_type_reply;
1233 	props->grh_required	= rep->grh_required;
1234 
1235 	err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
1236 	if (err)
1237 		goto out;
1238 
1239 	err = translate_active_width(ibdev, ib_link_width_oper,
1240 				     &props->active_width);
1241 	if (err)
1242 		goto out;
1243 	err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
1244 	if (err)
1245 		goto out;
1246 
1247 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1248 
1249 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1250 
1251 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1252 
1253 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1254 
1255 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1256 	if (err)
1257 		goto out;
1258 
1259 	err = translate_max_vl_num(ibdev, vl_hw_cap,
1260 				   &props->max_vl_num);
1261 out:
1262 	kfree(rep);
1263 	return err;
1264 }
1265 
1266 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
1267 		       struct ib_port_attr *props)
1268 {
1269 	unsigned int count;
1270 	int ret;
1271 
1272 	switch (mlx5_get_vport_access_method(ibdev)) {
1273 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1274 		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1275 		break;
1276 
1277 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1278 		ret = mlx5_query_hca_port(ibdev, port, props);
1279 		break;
1280 
1281 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1282 		ret = mlx5_query_port_roce(ibdev, port, props);
1283 		break;
1284 
1285 	default:
1286 		ret = -EINVAL;
1287 	}
1288 
1289 	if (!ret && props) {
1290 		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1291 		struct mlx5_core_dev *mdev;
1292 		bool put_mdev = true;
1293 
1294 		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1295 		if (!mdev) {
1296 			/* If the port isn't affiliated yet query the master.
1297 			 * The master and slave will have the same values.
1298 			 */
1299 			mdev = dev->mdev;
1300 			port = 1;
1301 			put_mdev = false;
1302 		}
1303 		count = mlx5_core_reserved_gids_count(mdev);
1304 		if (put_mdev)
1305 			mlx5_ib_put_native_port_mdev(dev, port);
1306 		props->gid_tbl_len -= count;
1307 	}
1308 	return ret;
1309 }
1310 
1311 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u8 port,
1312 				  struct ib_port_attr *props)
1313 {
1314 	int ret;
1315 
1316 	/* Only link layer == ethernet is valid for representors */
1317 	ret = mlx5_query_port_roce(ibdev, port, props);
1318 	if (ret || !props)
1319 		return ret;
1320 
1321 	/* We don't support GIDS */
1322 	props->gid_tbl_len = 0;
1323 
1324 	return ret;
1325 }
1326 
1327 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
1328 			     union ib_gid *gid)
1329 {
1330 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1331 	struct mlx5_core_dev *mdev = dev->mdev;
1332 
1333 	switch (mlx5_get_vport_access_method(ibdev)) {
1334 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1335 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1336 
1337 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1338 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1339 
1340 	default:
1341 		return -EINVAL;
1342 	}
1343 
1344 }
1345 
1346 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u8 port,
1347 				   u16 index, u16 *pkey)
1348 {
1349 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1350 	struct mlx5_core_dev *mdev;
1351 	bool put_mdev = true;
1352 	u8 mdev_port_num;
1353 	int err;
1354 
1355 	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1356 	if (!mdev) {
1357 		/* The port isn't affiliated yet, get the PKey from the master
1358 		 * port. For RoCE the PKey tables will be the same.
1359 		 */
1360 		put_mdev = false;
1361 		mdev = dev->mdev;
1362 		mdev_port_num = 1;
1363 	}
1364 
1365 	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1366 					index, pkey);
1367 	if (put_mdev)
1368 		mlx5_ib_put_native_port_mdev(dev, port);
1369 
1370 	return err;
1371 }
1372 
1373 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1374 			      u16 *pkey)
1375 {
1376 	switch (mlx5_get_vport_access_method(ibdev)) {
1377 	case MLX5_VPORT_ACCESS_METHOD_MAD:
1378 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1379 
1380 	case MLX5_VPORT_ACCESS_METHOD_HCA:
1381 	case MLX5_VPORT_ACCESS_METHOD_NIC:
1382 		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1383 	default:
1384 		return -EINVAL;
1385 	}
1386 }
1387 
1388 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1389 				 struct ib_device_modify *props)
1390 {
1391 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1392 	struct mlx5_reg_node_desc in;
1393 	struct mlx5_reg_node_desc out;
1394 	int err;
1395 
1396 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1397 		return -EOPNOTSUPP;
1398 
1399 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1400 		return 0;
1401 
1402 	/*
1403 	 * If possible, pass node desc to FW, so it can generate
1404 	 * a 144 trap.  If cmd fails, just ignore.
1405 	 */
1406 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1407 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1408 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1409 	if (err)
1410 		return err;
1411 
1412 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1413 
1414 	return err;
1415 }
1416 
1417 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u8 port_num, u32 mask,
1418 				u32 value)
1419 {
1420 	struct mlx5_hca_vport_context ctx = {};
1421 	struct mlx5_core_dev *mdev;
1422 	u8 mdev_port_num;
1423 	int err;
1424 
1425 	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1426 	if (!mdev)
1427 		return -ENODEV;
1428 
1429 	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1430 	if (err)
1431 		goto out;
1432 
1433 	if (~ctx.cap_mask1_perm & mask) {
1434 		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1435 			     mask, ctx.cap_mask1_perm);
1436 		err = -EINVAL;
1437 		goto out;
1438 	}
1439 
1440 	ctx.cap_mask1 = value;
1441 	ctx.cap_mask1_perm = mask;
1442 	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1443 						 0, &ctx);
1444 
1445 out:
1446 	mlx5_ib_put_native_port_mdev(dev, port_num);
1447 
1448 	return err;
1449 }
1450 
1451 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1452 			       struct ib_port_modify *props)
1453 {
1454 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1455 	struct ib_port_attr attr;
1456 	u32 tmp;
1457 	int err;
1458 	u32 change_mask;
1459 	u32 value;
1460 	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1461 		      IB_LINK_LAYER_INFINIBAND);
1462 
1463 	/* CM layer calls ib_modify_port() regardless of the link layer. For
1464 	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1465 	 */
1466 	if (!is_ib)
1467 		return 0;
1468 
1469 	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1470 		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1471 		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1472 		return set_port_caps_atomic(dev, port, change_mask, value);
1473 	}
1474 
1475 	mutex_lock(&dev->cap_mask_mutex);
1476 
1477 	err = ib_query_port(ibdev, port, &attr);
1478 	if (err)
1479 		goto out;
1480 
1481 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1482 		~props->clr_port_cap_mask;
1483 
1484 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1485 
1486 out:
1487 	mutex_unlock(&dev->cap_mask_mutex);
1488 	return err;
1489 }
1490 
1491 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1492 {
1493 	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1494 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1495 }
1496 
1497 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1498 {
1499 	/* Large page with non 4k uar support might limit the dynamic size */
1500 	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1501 		return MLX5_MIN_DYN_BFREGS;
1502 
1503 	return MLX5_MAX_DYN_BFREGS;
1504 }
1505 
1506 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1507 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1508 			     struct mlx5_bfreg_info *bfregi)
1509 {
1510 	int uars_per_sys_page;
1511 	int bfregs_per_sys_page;
1512 	int ref_bfregs = req->total_num_bfregs;
1513 
1514 	if (req->total_num_bfregs == 0)
1515 		return -EINVAL;
1516 
1517 	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1518 	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1519 
1520 	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1521 		return -ENOMEM;
1522 
1523 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1524 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1525 	/* This holds the required static allocation asked by the user */
1526 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1527 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1528 		return -EINVAL;
1529 
1530 	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1531 	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1532 	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1533 	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1534 
1535 	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",
1536 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1537 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1538 		    req->total_num_bfregs, bfregi->total_num_bfregs,
1539 		    bfregi->num_sys_pages);
1540 
1541 	return 0;
1542 }
1543 
1544 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1545 {
1546 	struct mlx5_bfreg_info *bfregi;
1547 	int err;
1548 	int i;
1549 
1550 	bfregi = &context->bfregi;
1551 	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1552 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1553 		if (err)
1554 			goto error;
1555 
1556 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1557 	}
1558 
1559 	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1560 		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1561 
1562 	return 0;
1563 
1564 error:
1565 	for (--i; i >= 0; i--)
1566 		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1567 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1568 
1569 	return err;
1570 }
1571 
1572 static int deallocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1573 {
1574 	struct mlx5_bfreg_info *bfregi;
1575 	int err;
1576 	int i;
1577 
1578 	bfregi = &context->bfregi;
1579 	for (i = 0; i < bfregi->num_sys_pages; i++) {
1580 		if (i < bfregi->num_static_sys_pages ||
1581 		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX) {
1582 			err = mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1583 			if (err) {
1584 				mlx5_ib_warn(dev, "failed to free uar %d, err=%d\n", i, err);
1585 				return err;
1586 			}
1587 		}
1588 	}
1589 
1590 	return 0;
1591 }
1592 
1593 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn)
1594 {
1595 	int err;
1596 
1597 	err = mlx5_core_alloc_transport_domain(dev->mdev, tdn);
1598 	if (err)
1599 		return err;
1600 
1601 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1602 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1603 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1604 		return err;
1605 
1606 	mutex_lock(&dev->lb_mutex);
1607 	dev->user_td++;
1608 
1609 	if (dev->user_td == 2)
1610 		err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1611 
1612 	mutex_unlock(&dev->lb_mutex);
1613 	return err;
1614 }
1615 
1616 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn)
1617 {
1618 	mlx5_core_dealloc_transport_domain(dev->mdev, tdn);
1619 
1620 	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1621 	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1622 	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1623 		return;
1624 
1625 	mutex_lock(&dev->lb_mutex);
1626 	dev->user_td--;
1627 
1628 	if (dev->user_td < 2)
1629 		mlx5_nic_vport_update_local_lb(dev->mdev, false);
1630 
1631 	mutex_unlock(&dev->lb_mutex);
1632 }
1633 
1634 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
1635 						  struct ib_udata *udata)
1636 {
1637 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1638 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1639 	struct mlx5_ib_alloc_ucontext_resp resp = {};
1640 	struct mlx5_core_dev *mdev = dev->mdev;
1641 	struct mlx5_ib_ucontext *context;
1642 	struct mlx5_bfreg_info *bfregi;
1643 	int ver;
1644 	int err;
1645 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1646 				     max_cqe_version);
1647 	bool lib_uar_4k;
1648 
1649 	if (!dev->ib_active)
1650 		return ERR_PTR(-EAGAIN);
1651 
1652 	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1653 		ver = 0;
1654 	else if (udata->inlen >= min_req_v2)
1655 		ver = 2;
1656 	else
1657 		return ERR_PTR(-EINVAL);
1658 
1659 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1660 	if (err)
1661 		return ERR_PTR(err);
1662 
1663 	if (req.flags)
1664 		return ERR_PTR(-EINVAL);
1665 
1666 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1667 		return ERR_PTR(-EOPNOTSUPP);
1668 
1669 	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1670 				    MLX5_NON_FP_BFREGS_PER_UAR);
1671 	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1672 		return ERR_PTR(-EINVAL);
1673 
1674 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1675 	if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1676 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1677 	resp.cache_line_size = cache_line_size();
1678 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1679 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1680 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1681 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1682 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1683 	resp.cqe_version = min_t(__u8,
1684 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1685 				 req.max_cqe_version);
1686 	resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1687 				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1688 	resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1689 					MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1;
1690 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1691 				   sizeof(resp.response_length), udata->outlen);
1692 
1693 	if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1694 		if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_EGRESS))
1695 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1696 		if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1697 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1698 		if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1699 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1700 		if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1701 			resp.flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1702 		/* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1703 	}
1704 
1705 	context = kzalloc(sizeof(*context), GFP_KERNEL);
1706 	if (!context)
1707 		return ERR_PTR(-ENOMEM);
1708 
1709 	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1710 	bfregi = &context->bfregi;
1711 
1712 	/* updates req->total_num_bfregs */
1713 	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1714 	if (err)
1715 		goto out_ctx;
1716 
1717 	mutex_init(&bfregi->lock);
1718 	bfregi->lib_uar_4k = lib_uar_4k;
1719 	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1720 				GFP_KERNEL);
1721 	if (!bfregi->count) {
1722 		err = -ENOMEM;
1723 		goto out_ctx;
1724 	}
1725 
1726 	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1727 				    sizeof(*bfregi->sys_pages),
1728 				    GFP_KERNEL);
1729 	if (!bfregi->sys_pages) {
1730 		err = -ENOMEM;
1731 		goto out_count;
1732 	}
1733 
1734 	err = allocate_uars(dev, context);
1735 	if (err)
1736 		goto out_sys_pages;
1737 
1738 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1739 	context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1740 #endif
1741 
1742 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1743 		err = mlx5_ib_alloc_transport_domain(dev, &context->tdn);
1744 		if (err)
1745 			goto out_uars;
1746 	}
1747 
1748 	INIT_LIST_HEAD(&context->vma_private_list);
1749 	mutex_init(&context->vma_private_list_mutex);
1750 	INIT_LIST_HEAD(&context->db_page_list);
1751 	mutex_init(&context->db_page_mutex);
1752 
1753 	resp.tot_bfregs = req.total_num_bfregs;
1754 	resp.num_ports = dev->num_ports;
1755 
1756 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1757 		resp.response_length += sizeof(resp.cqe_version);
1758 
1759 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1760 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1761 				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1762 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1763 	}
1764 
1765 	if (field_avail(typeof(resp), eth_min_inline, udata->outlen)) {
1766 		if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1767 			mlx5_query_min_inline(dev->mdev, &resp.eth_min_inline);
1768 			resp.eth_min_inline++;
1769 		}
1770 		resp.response_length += sizeof(resp.eth_min_inline);
1771 	}
1772 
1773 	if (field_avail(typeof(resp), clock_info_versions, udata->outlen)) {
1774 		if (mdev->clock_info)
1775 			resp.clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1776 		resp.response_length += sizeof(resp.clock_info_versions);
1777 	}
1778 
1779 	/*
1780 	 * We don't want to expose information from the PCI bar that is located
1781 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1782 	 * pretend we don't support reading the HCA's core clock. This is also
1783 	 * forced by mmap function.
1784 	 */
1785 	if (field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1786 		if (PAGE_SIZE <= 4096) {
1787 			resp.comp_mask |=
1788 				MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1789 			resp.hca_core_clock_offset =
1790 				offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1791 		}
1792 		resp.response_length += sizeof(resp.hca_core_clock_offset);
1793 	}
1794 
1795 	if (field_avail(typeof(resp), log_uar_size, udata->outlen))
1796 		resp.response_length += sizeof(resp.log_uar_size);
1797 
1798 	if (field_avail(typeof(resp), num_uars_per_page, udata->outlen))
1799 		resp.response_length += sizeof(resp.num_uars_per_page);
1800 
1801 	if (field_avail(typeof(resp), num_dyn_bfregs, udata->outlen)) {
1802 		resp.num_dyn_bfregs = bfregi->num_dyn_bfregs;
1803 		resp.response_length += sizeof(resp.num_dyn_bfregs);
1804 	}
1805 
1806 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1807 	if (err)
1808 		goto out_td;
1809 
1810 	bfregi->ver = ver;
1811 	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1812 	context->cqe_version = resp.cqe_version;
1813 	context->lib_caps = req.lib_caps;
1814 	print_lib_caps(dev, context->lib_caps);
1815 
1816 	return &context->ibucontext;
1817 
1818 out_td:
1819 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1820 		mlx5_ib_dealloc_transport_domain(dev, context->tdn);
1821 
1822 out_uars:
1823 	deallocate_uars(dev, context);
1824 
1825 out_sys_pages:
1826 	kfree(bfregi->sys_pages);
1827 
1828 out_count:
1829 	kfree(bfregi->count);
1830 
1831 out_ctx:
1832 	kfree(context);
1833 
1834 	return ERR_PTR(err);
1835 }
1836 
1837 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1838 {
1839 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1840 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1841 	struct mlx5_bfreg_info *bfregi;
1842 
1843 	bfregi = &context->bfregi;
1844 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1845 		mlx5_ib_dealloc_transport_domain(dev, context->tdn);
1846 
1847 	deallocate_uars(dev, context);
1848 	kfree(bfregi->sys_pages);
1849 	kfree(bfregi->count);
1850 	kfree(context);
1851 
1852 	return 0;
1853 }
1854 
1855 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
1856 				 int uar_idx)
1857 {
1858 	int fw_uars_per_page;
1859 
1860 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
1861 
1862 	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
1863 }
1864 
1865 static int get_command(unsigned long offset)
1866 {
1867 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1868 }
1869 
1870 static int get_arg(unsigned long offset)
1871 {
1872 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1873 }
1874 
1875 static int get_index(unsigned long offset)
1876 {
1877 	return get_arg(offset);
1878 }
1879 
1880 /* Index resides in an extra byte to enable larger values than 255 */
1881 static int get_extended_index(unsigned long offset)
1882 {
1883 	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
1884 }
1885 
1886 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1887 {
1888 	/* vma_open is called when a new VMA is created on top of our VMA.  This
1889 	 * is done through either mremap flow or split_vma (usually due to
1890 	 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1891 	 * as this VMA is strongly hardware related.  Therefore we set the
1892 	 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1893 	 * calling us again and trying to do incorrect actions.  We assume that
1894 	 * the original VMA size is exactly a single page, and therefore all
1895 	 * "splitting" operation will not happen to it.
1896 	 */
1897 	area->vm_ops = NULL;
1898 }
1899 
1900 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1901 {
1902 	struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1903 
1904 	/* It's guaranteed that all VMAs opened on a FD are closed before the
1905 	 * file itself is closed, therefore no sync is needed with the regular
1906 	 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1907 	 * However need a sync with accessing the vma as part of
1908 	 * mlx5_ib_disassociate_ucontext.
1909 	 * The close operation is usually called under mm->mmap_sem except when
1910 	 * process is exiting.
1911 	 * The exiting case is handled explicitly as part of
1912 	 * mlx5_ib_disassociate_ucontext.
1913 	 */
1914 	mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1915 
1916 	/* setting the vma context pointer to null in the mlx5_ib driver's
1917 	 * private data, to protect a race condition in
1918 	 * mlx5_ib_disassociate_ucontext().
1919 	 */
1920 	mlx5_ib_vma_priv_data->vma = NULL;
1921 	mutex_lock(mlx5_ib_vma_priv_data->vma_private_list_mutex);
1922 	list_del(&mlx5_ib_vma_priv_data->list);
1923 	mutex_unlock(mlx5_ib_vma_priv_data->vma_private_list_mutex);
1924 	kfree(mlx5_ib_vma_priv_data);
1925 }
1926 
1927 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1928 	.open = mlx5_ib_vma_open,
1929 	.close = mlx5_ib_vma_close
1930 };
1931 
1932 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1933 				struct mlx5_ib_ucontext *ctx)
1934 {
1935 	struct mlx5_ib_vma_private_data *vma_prv;
1936 	struct list_head *vma_head = &ctx->vma_private_list;
1937 
1938 	vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1939 	if (!vma_prv)
1940 		return -ENOMEM;
1941 
1942 	vma_prv->vma = vma;
1943 	vma_prv->vma_private_list_mutex = &ctx->vma_private_list_mutex;
1944 	vma->vm_private_data = vma_prv;
1945 	vma->vm_ops =  &mlx5_ib_vm_ops;
1946 
1947 	mutex_lock(&ctx->vma_private_list_mutex);
1948 	list_add(&vma_prv->list, vma_head);
1949 	mutex_unlock(&ctx->vma_private_list_mutex);
1950 
1951 	return 0;
1952 }
1953 
1954 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1955 {
1956 	int ret;
1957 	struct vm_area_struct *vma;
1958 	struct mlx5_ib_vma_private_data *vma_private, *n;
1959 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1960 	struct task_struct *owning_process  = NULL;
1961 	struct mm_struct   *owning_mm       = NULL;
1962 
1963 	owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1964 	if (!owning_process)
1965 		return;
1966 
1967 	owning_mm = get_task_mm(owning_process);
1968 	if (!owning_mm) {
1969 		pr_info("no mm, disassociate ucontext is pending task termination\n");
1970 		while (1) {
1971 			put_task_struct(owning_process);
1972 			usleep_range(1000, 2000);
1973 			owning_process = get_pid_task(ibcontext->tgid,
1974 						      PIDTYPE_PID);
1975 			if (!owning_process ||
1976 			    owning_process->state == TASK_DEAD) {
1977 				pr_info("disassociate ucontext done, task was terminated\n");
1978 				/* in case task was dead need to release the
1979 				 * task struct.
1980 				 */
1981 				if (owning_process)
1982 					put_task_struct(owning_process);
1983 				return;
1984 			}
1985 		}
1986 	}
1987 
1988 	/* need to protect from a race on closing the vma as part of
1989 	 * mlx5_ib_vma_close.
1990 	 */
1991 	down_write(&owning_mm->mmap_sem);
1992 	mutex_lock(&context->vma_private_list_mutex);
1993 	list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1994 				 list) {
1995 		vma = vma_private->vma;
1996 		ret = zap_vma_ptes(vma, vma->vm_start,
1997 				   PAGE_SIZE);
1998 		WARN_ONCE(ret, "%s: zap_vma_ptes failed", __func__);
1999 		/* context going to be destroyed, should
2000 		 * not access ops any more.
2001 		 */
2002 		vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
2003 		vma->vm_ops = NULL;
2004 		list_del(&vma_private->list);
2005 		kfree(vma_private);
2006 	}
2007 	mutex_unlock(&context->vma_private_list_mutex);
2008 	up_write(&owning_mm->mmap_sem);
2009 	mmput(owning_mm);
2010 	put_task_struct(owning_process);
2011 }
2012 
2013 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2014 {
2015 	switch (cmd) {
2016 	case MLX5_IB_MMAP_WC_PAGE:
2017 		return "WC";
2018 	case MLX5_IB_MMAP_REGULAR_PAGE:
2019 		return "best effort WC";
2020 	case MLX5_IB_MMAP_NC_PAGE:
2021 		return "NC";
2022 	case MLX5_IB_MMAP_DEVICE_MEM:
2023 		return "Device Memory";
2024 	default:
2025 		return NULL;
2026 	}
2027 }
2028 
2029 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2030 					struct vm_area_struct *vma,
2031 					struct mlx5_ib_ucontext *context)
2032 {
2033 	phys_addr_t pfn;
2034 	int err;
2035 
2036 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2037 		return -EINVAL;
2038 
2039 	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2040 		return -EOPNOTSUPP;
2041 
2042 	if (vma->vm_flags & VM_WRITE)
2043 		return -EPERM;
2044 
2045 	if (!dev->mdev->clock_info_page)
2046 		return -EOPNOTSUPP;
2047 
2048 	pfn = page_to_pfn(dev->mdev->clock_info_page);
2049 	err = remap_pfn_range(vma, vma->vm_start, pfn, PAGE_SIZE,
2050 			      vma->vm_page_prot);
2051 	if (err)
2052 		return err;
2053 
2054 	mlx5_ib_dbg(dev, "mapped clock info at 0x%lx, PA 0x%llx\n",
2055 		    vma->vm_start,
2056 		    (unsigned long long)pfn << PAGE_SHIFT);
2057 
2058 	return mlx5_ib_set_vma_data(vma, context);
2059 }
2060 
2061 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2062 		    struct vm_area_struct *vma,
2063 		    struct mlx5_ib_ucontext *context)
2064 {
2065 	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2066 	int err;
2067 	unsigned long idx;
2068 	phys_addr_t pfn, pa;
2069 	pgprot_t prot;
2070 	u32 bfreg_dyn_idx = 0;
2071 	u32 uar_index;
2072 	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2073 	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2074 				bfregi->num_static_sys_pages;
2075 
2076 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2077 		return -EINVAL;
2078 
2079 	if (dyn_uar)
2080 		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2081 	else
2082 		idx = get_index(vma->vm_pgoff);
2083 
2084 	if (idx >= max_valid_idx) {
2085 		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2086 			     idx, max_valid_idx);
2087 		return -EINVAL;
2088 	}
2089 
2090 	switch (cmd) {
2091 	case MLX5_IB_MMAP_WC_PAGE:
2092 	case MLX5_IB_MMAP_ALLOC_WC:
2093 /* Some architectures don't support WC memory */
2094 #if defined(CONFIG_X86)
2095 		if (!pat_enabled())
2096 			return -EPERM;
2097 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
2098 			return -EPERM;
2099 #endif
2100 	/* fall through */
2101 	case MLX5_IB_MMAP_REGULAR_PAGE:
2102 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2103 		prot = pgprot_writecombine(vma->vm_page_prot);
2104 		break;
2105 	case MLX5_IB_MMAP_NC_PAGE:
2106 		prot = pgprot_noncached(vma->vm_page_prot);
2107 		break;
2108 	default:
2109 		return -EINVAL;
2110 	}
2111 
2112 	if (dyn_uar) {
2113 		int uars_per_page;
2114 
2115 		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2116 		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2117 		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2118 			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2119 				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2120 			return -EINVAL;
2121 		}
2122 
2123 		mutex_lock(&bfregi->lock);
2124 		/* Fail if uar already allocated, first bfreg index of each
2125 		 * page holds its count.
2126 		 */
2127 		if (bfregi->count[bfreg_dyn_idx]) {
2128 			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2129 			mutex_unlock(&bfregi->lock);
2130 			return -EINVAL;
2131 		}
2132 
2133 		bfregi->count[bfreg_dyn_idx]++;
2134 		mutex_unlock(&bfregi->lock);
2135 
2136 		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2137 		if (err) {
2138 			mlx5_ib_warn(dev, "UAR alloc failed\n");
2139 			goto free_bfreg;
2140 		}
2141 	} else {
2142 		uar_index = bfregi->sys_pages[idx];
2143 	}
2144 
2145 	pfn = uar_index2pfn(dev, uar_index);
2146 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2147 
2148 	vma->vm_page_prot = prot;
2149 	err = io_remap_pfn_range(vma, vma->vm_start, pfn,
2150 				 PAGE_SIZE, vma->vm_page_prot);
2151 	if (err) {
2152 		mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%lx, pfn=%pa, mmap_cmd=%s\n",
2153 			    err, vma->vm_start, &pfn, mmap_cmd2str(cmd));
2154 		err = -EAGAIN;
2155 		goto err;
2156 	}
2157 
2158 	pa = pfn << PAGE_SHIFT;
2159 	mlx5_ib_dbg(dev, "mapped %s at 0x%lx, PA %pa\n", mmap_cmd2str(cmd),
2160 		    vma->vm_start, &pa);
2161 
2162 	err = mlx5_ib_set_vma_data(vma, context);
2163 	if (err)
2164 		goto err;
2165 
2166 	if (dyn_uar)
2167 		bfregi->sys_pages[idx] = uar_index;
2168 	return 0;
2169 
2170 err:
2171 	if (!dyn_uar)
2172 		return err;
2173 
2174 	mlx5_cmd_free_uar(dev->mdev, idx);
2175 
2176 free_bfreg:
2177 	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2178 
2179 	return err;
2180 }
2181 
2182 static int dm_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
2183 {
2184 	struct mlx5_ib_ucontext *mctx = to_mucontext(context);
2185 	struct mlx5_ib_dev *dev = to_mdev(context->device);
2186 	u16 page_idx = get_extended_index(vma->vm_pgoff);
2187 	size_t map_size = vma->vm_end - vma->vm_start;
2188 	u32 npages = map_size >> PAGE_SHIFT;
2189 	phys_addr_t pfn;
2190 	pgprot_t prot;
2191 
2192 	if (find_next_zero_bit(mctx->dm_pages, page_idx + npages, page_idx) !=
2193 	    page_idx + npages)
2194 		return -EINVAL;
2195 
2196 	pfn = ((pci_resource_start(dev->mdev->pdev, 0) +
2197 	      MLX5_CAP64_DEV_MEM(dev->mdev, memic_bar_start_addr)) >>
2198 	      PAGE_SHIFT) +
2199 	      page_idx;
2200 	prot = pgprot_writecombine(vma->vm_page_prot);
2201 	vma->vm_page_prot = prot;
2202 
2203 	if (io_remap_pfn_range(vma, vma->vm_start, pfn, map_size,
2204 			       vma->vm_page_prot))
2205 		return -EAGAIN;
2206 
2207 	return mlx5_ib_set_vma_data(vma, mctx);
2208 }
2209 
2210 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2211 {
2212 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2213 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2214 	unsigned long command;
2215 	phys_addr_t pfn;
2216 
2217 	command = get_command(vma->vm_pgoff);
2218 	switch (command) {
2219 	case MLX5_IB_MMAP_WC_PAGE:
2220 	case MLX5_IB_MMAP_NC_PAGE:
2221 	case MLX5_IB_MMAP_REGULAR_PAGE:
2222 	case MLX5_IB_MMAP_ALLOC_WC:
2223 		return uar_mmap(dev, command, vma, context);
2224 
2225 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2226 		return -ENOSYS;
2227 
2228 	case MLX5_IB_MMAP_CORE_CLOCK:
2229 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2230 			return -EINVAL;
2231 
2232 		if (vma->vm_flags & VM_WRITE)
2233 			return -EPERM;
2234 
2235 		/* Don't expose to user-space information it shouldn't have */
2236 		if (PAGE_SIZE > 4096)
2237 			return -EOPNOTSUPP;
2238 
2239 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
2240 		pfn = (dev->mdev->iseg_base +
2241 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2242 			PAGE_SHIFT;
2243 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
2244 				       PAGE_SIZE, vma->vm_page_prot))
2245 			return -EAGAIN;
2246 
2247 		mlx5_ib_dbg(dev, "mapped internal timer at 0x%lx, PA 0x%llx\n",
2248 			    vma->vm_start,
2249 			    (unsigned long long)pfn << PAGE_SHIFT);
2250 		break;
2251 	case MLX5_IB_MMAP_CLOCK_INFO:
2252 		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2253 
2254 	case MLX5_IB_MMAP_DEVICE_MEM:
2255 		return dm_mmap(ibcontext, vma);
2256 
2257 	default:
2258 		return -EINVAL;
2259 	}
2260 
2261 	return 0;
2262 }
2263 
2264 struct ib_dm *mlx5_ib_alloc_dm(struct ib_device *ibdev,
2265 			       struct ib_ucontext *context,
2266 			       struct ib_dm_alloc_attr *attr,
2267 			       struct uverbs_attr_bundle *attrs)
2268 {
2269 	u64 act_size = roundup(attr->length, MLX5_MEMIC_BASE_SIZE);
2270 	struct mlx5_memic *memic = &to_mdev(ibdev)->memic;
2271 	phys_addr_t memic_addr;
2272 	struct mlx5_ib_dm *dm;
2273 	u64 start_offset;
2274 	u32 page_idx;
2275 	int err;
2276 
2277 	dm = kzalloc(sizeof(*dm), GFP_KERNEL);
2278 	if (!dm)
2279 		return ERR_PTR(-ENOMEM);
2280 
2281 	mlx5_ib_dbg(to_mdev(ibdev), "alloc_memic req: user_length=0x%llx act_length=0x%llx log_alignment=%d\n",
2282 		    attr->length, act_size, attr->alignment);
2283 
2284 	err = mlx5_cmd_alloc_memic(memic, &memic_addr,
2285 				   act_size, attr->alignment);
2286 	if (err)
2287 		goto err_free;
2288 
2289 	start_offset = memic_addr & ~PAGE_MASK;
2290 	page_idx = (memic_addr - pci_resource_start(memic->dev->pdev, 0) -
2291 		    MLX5_CAP64_DEV_MEM(memic->dev, memic_bar_start_addr)) >>
2292 		    PAGE_SHIFT;
2293 
2294 	err = uverbs_copy_to(attrs,
2295 			     MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
2296 			     &start_offset, sizeof(start_offset));
2297 	if (err)
2298 		goto err_dealloc;
2299 
2300 	err = uverbs_copy_to(attrs,
2301 			     MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
2302 			     &page_idx, sizeof(page_idx));
2303 	if (err)
2304 		goto err_dealloc;
2305 
2306 	bitmap_set(to_mucontext(context)->dm_pages, page_idx,
2307 		   DIV_ROUND_UP(act_size, PAGE_SIZE));
2308 
2309 	dm->dev_addr = memic_addr;
2310 
2311 	return &dm->ibdm;
2312 
2313 err_dealloc:
2314 	mlx5_cmd_dealloc_memic(memic, memic_addr,
2315 			       act_size);
2316 err_free:
2317 	kfree(dm);
2318 	return ERR_PTR(err);
2319 }
2320 
2321 int mlx5_ib_dealloc_dm(struct ib_dm *ibdm)
2322 {
2323 	struct mlx5_memic *memic = &to_mdev(ibdm->device)->memic;
2324 	struct mlx5_ib_dm *dm = to_mdm(ibdm);
2325 	u64 act_size = roundup(dm->ibdm.length, MLX5_MEMIC_BASE_SIZE);
2326 	u32 page_idx;
2327 	int ret;
2328 
2329 	ret = mlx5_cmd_dealloc_memic(memic, dm->dev_addr, act_size);
2330 	if (ret)
2331 		return ret;
2332 
2333 	page_idx = (dm->dev_addr - pci_resource_start(memic->dev->pdev, 0) -
2334 		    MLX5_CAP64_DEV_MEM(memic->dev, memic_bar_start_addr)) >>
2335 		    PAGE_SHIFT;
2336 	bitmap_clear(to_mucontext(ibdm->uobject->context)->dm_pages,
2337 		     page_idx,
2338 		     DIV_ROUND_UP(act_size, PAGE_SIZE));
2339 
2340 	kfree(dm);
2341 
2342 	return 0;
2343 }
2344 
2345 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
2346 				      struct ib_ucontext *context,
2347 				      struct ib_udata *udata)
2348 {
2349 	struct mlx5_ib_alloc_pd_resp resp;
2350 	struct mlx5_ib_pd *pd;
2351 	int err;
2352 
2353 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
2354 	if (!pd)
2355 		return ERR_PTR(-ENOMEM);
2356 
2357 	err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
2358 	if (err) {
2359 		kfree(pd);
2360 		return ERR_PTR(err);
2361 	}
2362 
2363 	if (context) {
2364 		resp.pdn = pd->pdn;
2365 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2366 			mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
2367 			kfree(pd);
2368 			return ERR_PTR(-EFAULT);
2369 		}
2370 	}
2371 
2372 	return &pd->ibpd;
2373 }
2374 
2375 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
2376 {
2377 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2378 	struct mlx5_ib_pd *mpd = to_mpd(pd);
2379 
2380 	mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
2381 	kfree(mpd);
2382 
2383 	return 0;
2384 }
2385 
2386 enum {
2387 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
2388 	MATCH_CRITERIA_ENABLE_MISC_BIT,
2389 	MATCH_CRITERIA_ENABLE_INNER_BIT
2390 };
2391 
2392 #define HEADER_IS_ZERO(match_criteria, headers)			           \
2393 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
2394 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
2395 
2396 static u8 get_match_criteria_enable(u32 *match_criteria)
2397 {
2398 	u8 match_criteria_enable;
2399 
2400 	match_criteria_enable =
2401 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
2402 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
2403 	match_criteria_enable |=
2404 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
2405 		MATCH_CRITERIA_ENABLE_MISC_BIT;
2406 	match_criteria_enable |=
2407 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
2408 		MATCH_CRITERIA_ENABLE_INNER_BIT;
2409 
2410 	return match_criteria_enable;
2411 }
2412 
2413 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
2414 {
2415 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
2416 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
2417 }
2418 
2419 static void set_flow_label(void *misc_c, void *misc_v, u32 mask, u32 val,
2420 			   bool inner)
2421 {
2422 	if (inner) {
2423 		MLX5_SET(fte_match_set_misc,
2424 			 misc_c, inner_ipv6_flow_label, mask);
2425 		MLX5_SET(fte_match_set_misc,
2426 			 misc_v, inner_ipv6_flow_label, val);
2427 	} else {
2428 		MLX5_SET(fte_match_set_misc,
2429 			 misc_c, outer_ipv6_flow_label, mask);
2430 		MLX5_SET(fte_match_set_misc,
2431 			 misc_v, outer_ipv6_flow_label, val);
2432 	}
2433 }
2434 
2435 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
2436 {
2437 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
2438 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
2439 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
2440 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
2441 }
2442 
2443 #define LAST_ETH_FIELD vlan_tag
2444 #define LAST_IB_FIELD sl
2445 #define LAST_IPV4_FIELD tos
2446 #define LAST_IPV6_FIELD traffic_class
2447 #define LAST_TCP_UDP_FIELD src_port
2448 #define LAST_TUNNEL_FIELD tunnel_id
2449 #define LAST_FLOW_TAG_FIELD tag_id
2450 #define LAST_DROP_FIELD size
2451 
2452 /* Field is the last supported field */
2453 #define FIELDS_NOT_SUPPORTED(filter, field)\
2454 	memchr_inv((void *)&filter.field  +\
2455 		   sizeof(filter.field), 0,\
2456 		   sizeof(filter) -\
2457 		   offsetof(typeof(filter), field) -\
2458 		   sizeof(filter.field))
2459 
2460 static int parse_flow_flow_action(const union ib_flow_spec *ib_spec,
2461 				  const struct ib_flow_attr *flow_attr,
2462 				  struct mlx5_flow_act *action)
2463 {
2464 	struct mlx5_ib_flow_action *maction = to_mflow_act(ib_spec->action.act);
2465 
2466 	switch (maction->ib_action.type) {
2467 	case IB_FLOW_ACTION_ESP:
2468 		/* Currently only AES_GCM keymat is supported by the driver */
2469 		action->esp_id = (uintptr_t)maction->esp_aes_gcm.ctx;
2470 		action->action |= flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS ?
2471 			MLX5_FLOW_CONTEXT_ACTION_ENCRYPT :
2472 			MLX5_FLOW_CONTEXT_ACTION_DECRYPT;
2473 		return 0;
2474 	default:
2475 		return -EOPNOTSUPP;
2476 	}
2477 }
2478 
2479 static int parse_flow_attr(struct mlx5_core_dev *mdev, u32 *match_c,
2480 			   u32 *match_v, const union ib_flow_spec *ib_spec,
2481 			   const struct ib_flow_attr *flow_attr,
2482 			   struct mlx5_flow_act *action)
2483 {
2484 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
2485 					   misc_parameters);
2486 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
2487 					   misc_parameters);
2488 	void *headers_c;
2489 	void *headers_v;
2490 	int match_ipv;
2491 	int ret;
2492 
2493 	if (ib_spec->type & IB_FLOW_SPEC_INNER) {
2494 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2495 					 inner_headers);
2496 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2497 					 inner_headers);
2498 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2499 					ft_field_support.inner_ip_version);
2500 	} else {
2501 		headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
2502 					 outer_headers);
2503 		headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
2504 					 outer_headers);
2505 		match_ipv = MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2506 					ft_field_support.outer_ip_version);
2507 	}
2508 
2509 	switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
2510 	case IB_FLOW_SPEC_ETH:
2511 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
2512 			return -EOPNOTSUPP;
2513 
2514 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2515 					     dmac_47_16),
2516 				ib_spec->eth.mask.dst_mac);
2517 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2518 					     dmac_47_16),
2519 				ib_spec->eth.val.dst_mac);
2520 
2521 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2522 					     smac_47_16),
2523 				ib_spec->eth.mask.src_mac);
2524 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2525 					     smac_47_16),
2526 				ib_spec->eth.val.src_mac);
2527 
2528 		if (ib_spec->eth.mask.vlan_tag) {
2529 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2530 				 cvlan_tag, 1);
2531 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2532 				 cvlan_tag, 1);
2533 
2534 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2535 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
2536 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2537 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
2538 
2539 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2540 				 first_cfi,
2541 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
2542 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2543 				 first_cfi,
2544 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
2545 
2546 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2547 				 first_prio,
2548 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
2549 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2550 				 first_prio,
2551 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
2552 		}
2553 		MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2554 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
2555 		MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2556 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
2557 		break;
2558 	case IB_FLOW_SPEC_IPV4:
2559 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
2560 			return -EOPNOTSUPP;
2561 
2562 		if (match_ipv) {
2563 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2564 				 ip_version, 0xf);
2565 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2566 				 ip_version, MLX5_FS_IPV4_VERSION);
2567 		} else {
2568 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2569 				 ethertype, 0xffff);
2570 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2571 				 ethertype, ETH_P_IP);
2572 		}
2573 
2574 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2575 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
2576 		       &ib_spec->ipv4.mask.src_ip,
2577 		       sizeof(ib_spec->ipv4.mask.src_ip));
2578 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2579 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
2580 		       &ib_spec->ipv4.val.src_ip,
2581 		       sizeof(ib_spec->ipv4.val.src_ip));
2582 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2583 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2584 		       &ib_spec->ipv4.mask.dst_ip,
2585 		       sizeof(ib_spec->ipv4.mask.dst_ip));
2586 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2587 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
2588 		       &ib_spec->ipv4.val.dst_ip,
2589 		       sizeof(ib_spec->ipv4.val.dst_ip));
2590 
2591 		set_tos(headers_c, headers_v,
2592 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
2593 
2594 		set_proto(headers_c, headers_v,
2595 			  ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
2596 		break;
2597 	case IB_FLOW_SPEC_IPV6:
2598 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
2599 			return -EOPNOTSUPP;
2600 
2601 		if (match_ipv) {
2602 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2603 				 ip_version, 0xf);
2604 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2605 				 ip_version, MLX5_FS_IPV6_VERSION);
2606 		} else {
2607 			MLX5_SET(fte_match_set_lyr_2_4, headers_c,
2608 				 ethertype, 0xffff);
2609 			MLX5_SET(fte_match_set_lyr_2_4, headers_v,
2610 				 ethertype, ETH_P_IPV6);
2611 		}
2612 
2613 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2614 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
2615 		       &ib_spec->ipv6.mask.src_ip,
2616 		       sizeof(ib_spec->ipv6.mask.src_ip));
2617 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2618 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
2619 		       &ib_spec->ipv6.val.src_ip,
2620 		       sizeof(ib_spec->ipv6.val.src_ip));
2621 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
2622 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2623 		       &ib_spec->ipv6.mask.dst_ip,
2624 		       sizeof(ib_spec->ipv6.mask.dst_ip));
2625 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
2626 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
2627 		       &ib_spec->ipv6.val.dst_ip,
2628 		       sizeof(ib_spec->ipv6.val.dst_ip));
2629 
2630 		set_tos(headers_c, headers_v,
2631 			ib_spec->ipv6.mask.traffic_class,
2632 			ib_spec->ipv6.val.traffic_class);
2633 
2634 		set_proto(headers_c, headers_v,
2635 			  ib_spec->ipv6.mask.next_hdr,
2636 			  ib_spec->ipv6.val.next_hdr);
2637 
2638 		set_flow_label(misc_params_c, misc_params_v,
2639 			       ntohl(ib_spec->ipv6.mask.flow_label),
2640 			       ntohl(ib_spec->ipv6.val.flow_label),
2641 			       ib_spec->type & IB_FLOW_SPEC_INNER);
2642 		break;
2643 	case IB_FLOW_SPEC_ESP:
2644 		if (ib_spec->esp.mask.seq)
2645 			return -EOPNOTSUPP;
2646 
2647 		MLX5_SET(fte_match_set_misc, misc_params_c, outer_esp_spi,
2648 			 ntohl(ib_spec->esp.mask.spi));
2649 		MLX5_SET(fte_match_set_misc, misc_params_v, outer_esp_spi,
2650 			 ntohl(ib_spec->esp.val.spi));
2651 		break;
2652 	case IB_FLOW_SPEC_TCP:
2653 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2654 					 LAST_TCP_UDP_FIELD))
2655 			return -EOPNOTSUPP;
2656 
2657 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2658 			 0xff);
2659 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2660 			 IPPROTO_TCP);
2661 
2662 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_sport,
2663 			 ntohs(ib_spec->tcp_udp.mask.src_port));
2664 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_sport,
2665 			 ntohs(ib_spec->tcp_udp.val.src_port));
2666 
2667 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_dport,
2668 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
2669 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_dport,
2670 			 ntohs(ib_spec->tcp_udp.val.dst_port));
2671 		break;
2672 	case IB_FLOW_SPEC_UDP:
2673 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
2674 					 LAST_TCP_UDP_FIELD))
2675 			return -EOPNOTSUPP;
2676 
2677 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
2678 			 0xff);
2679 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
2680 			 IPPROTO_UDP);
2681 
2682 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_sport,
2683 			 ntohs(ib_spec->tcp_udp.mask.src_port));
2684 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_sport,
2685 			 ntohs(ib_spec->tcp_udp.val.src_port));
2686 
2687 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_dport,
2688 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
2689 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_dport,
2690 			 ntohs(ib_spec->tcp_udp.val.dst_port));
2691 		break;
2692 	case IB_FLOW_SPEC_VXLAN_TUNNEL:
2693 		if (FIELDS_NOT_SUPPORTED(ib_spec->tunnel.mask,
2694 					 LAST_TUNNEL_FIELD))
2695 			return -EOPNOTSUPP;
2696 
2697 		MLX5_SET(fte_match_set_misc, misc_params_c, vxlan_vni,
2698 			 ntohl(ib_spec->tunnel.mask.tunnel_id));
2699 		MLX5_SET(fte_match_set_misc, misc_params_v, vxlan_vni,
2700 			 ntohl(ib_spec->tunnel.val.tunnel_id));
2701 		break;
2702 	case IB_FLOW_SPEC_ACTION_TAG:
2703 		if (FIELDS_NOT_SUPPORTED(ib_spec->flow_tag,
2704 					 LAST_FLOW_TAG_FIELD))
2705 			return -EOPNOTSUPP;
2706 		if (ib_spec->flow_tag.tag_id >= BIT(24))
2707 			return -EINVAL;
2708 
2709 		action->flow_tag = ib_spec->flow_tag.tag_id;
2710 		action->has_flow_tag = true;
2711 		break;
2712 	case IB_FLOW_SPEC_ACTION_DROP:
2713 		if (FIELDS_NOT_SUPPORTED(ib_spec->drop,
2714 					 LAST_DROP_FIELD))
2715 			return -EOPNOTSUPP;
2716 		action->action |= MLX5_FLOW_CONTEXT_ACTION_DROP;
2717 		break;
2718 	case IB_FLOW_SPEC_ACTION_HANDLE:
2719 		ret = parse_flow_flow_action(ib_spec, flow_attr, action);
2720 		if (ret)
2721 			return ret;
2722 		break;
2723 	default:
2724 		return -EINVAL;
2725 	}
2726 
2727 	return 0;
2728 }
2729 
2730 /* If a flow could catch both multicast and unicast packets,
2731  * it won't fall into the multicast flow steering table and this rule
2732  * could steal other multicast packets.
2733  */
2734 static bool flow_is_multicast_only(const struct ib_flow_attr *ib_attr)
2735 {
2736 	union ib_flow_spec *flow_spec;
2737 
2738 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
2739 	    ib_attr->num_of_specs < 1)
2740 		return false;
2741 
2742 	flow_spec = (union ib_flow_spec *)(ib_attr + 1);
2743 	if (flow_spec->type == IB_FLOW_SPEC_IPV4) {
2744 		struct ib_flow_spec_ipv4 *ipv4_spec;
2745 
2746 		ipv4_spec = (struct ib_flow_spec_ipv4 *)flow_spec;
2747 		if (ipv4_is_multicast(ipv4_spec->val.dst_ip))
2748 			return true;
2749 
2750 		return false;
2751 	}
2752 
2753 	if (flow_spec->type == IB_FLOW_SPEC_ETH) {
2754 		struct ib_flow_spec_eth *eth_spec;
2755 
2756 		eth_spec = (struct ib_flow_spec_eth *)flow_spec;
2757 		return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
2758 		       is_multicast_ether_addr(eth_spec->val.dst_mac);
2759 	}
2760 
2761 	return false;
2762 }
2763 
2764 enum valid_spec {
2765 	VALID_SPEC_INVALID,
2766 	VALID_SPEC_VALID,
2767 	VALID_SPEC_NA,
2768 };
2769 
2770 static enum valid_spec
2771 is_valid_esp_aes_gcm(struct mlx5_core_dev *mdev,
2772 		     const struct mlx5_flow_spec *spec,
2773 		     const struct mlx5_flow_act *flow_act,
2774 		     bool egress)
2775 {
2776 	const u32 *match_c = spec->match_criteria;
2777 	bool is_crypto =
2778 		(flow_act->action & (MLX5_FLOW_CONTEXT_ACTION_ENCRYPT |
2779 				     MLX5_FLOW_CONTEXT_ACTION_DECRYPT));
2780 	bool is_ipsec = mlx5_fs_is_ipsec_flow(match_c);
2781 	bool is_drop = flow_act->action & MLX5_FLOW_CONTEXT_ACTION_DROP;
2782 
2783 	/*
2784 	 * Currently only crypto is supported in egress, when regular egress
2785 	 * rules would be supported, always return VALID_SPEC_NA.
2786 	 */
2787 	if (!is_crypto)
2788 		return egress ? VALID_SPEC_INVALID : VALID_SPEC_NA;
2789 
2790 	return is_crypto && is_ipsec &&
2791 		(!egress || (!is_drop && !flow_act->has_flow_tag)) ?
2792 		VALID_SPEC_VALID : VALID_SPEC_INVALID;
2793 }
2794 
2795 static bool is_valid_spec(struct mlx5_core_dev *mdev,
2796 			  const struct mlx5_flow_spec *spec,
2797 			  const struct mlx5_flow_act *flow_act,
2798 			  bool egress)
2799 {
2800 	/* We curretly only support ipsec egress flow */
2801 	return is_valid_esp_aes_gcm(mdev, spec, flow_act, egress) != VALID_SPEC_INVALID;
2802 }
2803 
2804 static bool is_valid_ethertype(struct mlx5_core_dev *mdev,
2805 			       const struct ib_flow_attr *flow_attr,
2806 			       bool check_inner)
2807 {
2808 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
2809 	int match_ipv = check_inner ?
2810 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2811 					ft_field_support.inner_ip_version) :
2812 			MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
2813 					ft_field_support.outer_ip_version);
2814 	int inner_bit = check_inner ? IB_FLOW_SPEC_INNER : 0;
2815 	bool ipv4_spec_valid, ipv6_spec_valid;
2816 	unsigned int ip_spec_type = 0;
2817 	bool has_ethertype = false;
2818 	unsigned int spec_index;
2819 	bool mask_valid = true;
2820 	u16 eth_type = 0;
2821 	bool type_valid;
2822 
2823 	/* Validate that ethertype is correct */
2824 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2825 		if ((ib_spec->type == (IB_FLOW_SPEC_ETH | inner_bit)) &&
2826 		    ib_spec->eth.mask.ether_type) {
2827 			mask_valid = (ib_spec->eth.mask.ether_type ==
2828 				      htons(0xffff));
2829 			has_ethertype = true;
2830 			eth_type = ntohs(ib_spec->eth.val.ether_type);
2831 		} else if ((ib_spec->type == (IB_FLOW_SPEC_IPV4 | inner_bit)) ||
2832 			   (ib_spec->type == (IB_FLOW_SPEC_IPV6 | inner_bit))) {
2833 			ip_spec_type = ib_spec->type;
2834 		}
2835 		ib_spec = (void *)ib_spec + ib_spec->size;
2836 	}
2837 
2838 	type_valid = (!has_ethertype) || (!ip_spec_type);
2839 	if (!type_valid && mask_valid) {
2840 		ipv4_spec_valid = (eth_type == ETH_P_IP) &&
2841 			(ip_spec_type == (IB_FLOW_SPEC_IPV4 | inner_bit));
2842 		ipv6_spec_valid = (eth_type == ETH_P_IPV6) &&
2843 			(ip_spec_type == (IB_FLOW_SPEC_IPV6 | inner_bit));
2844 
2845 		type_valid = (ipv4_spec_valid) || (ipv6_spec_valid) ||
2846 			     (((eth_type == ETH_P_MPLS_UC) ||
2847 			       (eth_type == ETH_P_MPLS_MC)) && match_ipv);
2848 	}
2849 
2850 	return type_valid;
2851 }
2852 
2853 static bool is_valid_attr(struct mlx5_core_dev *mdev,
2854 			  const struct ib_flow_attr *flow_attr)
2855 {
2856 	return is_valid_ethertype(mdev, flow_attr, false) &&
2857 	       is_valid_ethertype(mdev, flow_attr, true);
2858 }
2859 
2860 static void put_flow_table(struct mlx5_ib_dev *dev,
2861 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
2862 {
2863 	prio->refcount -= !!ft_added;
2864 	if (!prio->refcount) {
2865 		mlx5_destroy_flow_table(prio->flow_table);
2866 		prio->flow_table = NULL;
2867 	}
2868 }
2869 
2870 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
2871 {
2872 	struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
2873 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
2874 							  struct mlx5_ib_flow_handler,
2875 							  ibflow);
2876 	struct mlx5_ib_flow_handler *iter, *tmp;
2877 
2878 	mutex_lock(&dev->flow_db->lock);
2879 
2880 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
2881 		mlx5_del_flow_rules(iter->rule);
2882 		put_flow_table(dev, iter->prio, true);
2883 		list_del(&iter->list);
2884 		kfree(iter);
2885 	}
2886 
2887 	mlx5_del_flow_rules(handler->rule);
2888 	put_flow_table(dev, handler->prio, true);
2889 	mutex_unlock(&dev->flow_db->lock);
2890 
2891 	kfree(handler);
2892 
2893 	return 0;
2894 }
2895 
2896 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
2897 {
2898 	priority *= 2;
2899 	if (!dont_trap)
2900 		priority++;
2901 	return priority;
2902 }
2903 
2904 enum flow_table_type {
2905 	MLX5_IB_FT_RX,
2906 	MLX5_IB_FT_TX
2907 };
2908 
2909 #define MLX5_FS_MAX_TYPES	 6
2910 #define MLX5_FS_MAX_ENTRIES	 BIT(16)
2911 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
2912 						struct ib_flow_attr *flow_attr,
2913 						enum flow_table_type ft_type)
2914 {
2915 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
2916 	struct mlx5_flow_namespace *ns = NULL;
2917 	struct mlx5_ib_flow_prio *prio;
2918 	struct mlx5_flow_table *ft;
2919 	int max_table_size;
2920 	int num_entries;
2921 	int num_groups;
2922 	int priority;
2923 	int err = 0;
2924 
2925 	max_table_size = BIT(MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
2926 						       log_max_ft_size));
2927 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2928 		if (ft_type == MLX5_IB_FT_TX)
2929 			priority = 0;
2930 		else if (flow_is_multicast_only(flow_attr) &&
2931 			 !dont_trap)
2932 			priority = MLX5_IB_FLOW_MCAST_PRIO;
2933 		else
2934 			priority = ib_prio_to_core_prio(flow_attr->priority,
2935 							dont_trap);
2936 		ns = mlx5_get_flow_namespace(dev->mdev,
2937 					     ft_type == MLX5_IB_FT_TX ?
2938 					     MLX5_FLOW_NAMESPACE_EGRESS :
2939 					     MLX5_FLOW_NAMESPACE_BYPASS);
2940 		num_entries = MLX5_FS_MAX_ENTRIES;
2941 		num_groups = MLX5_FS_MAX_TYPES;
2942 		prio = &dev->flow_db->prios[priority];
2943 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2944 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2945 		ns = mlx5_get_flow_namespace(dev->mdev,
2946 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
2947 		build_leftovers_ft_param(&priority,
2948 					 &num_entries,
2949 					 &num_groups);
2950 		prio = &dev->flow_db->prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
2951 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2952 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
2953 					allow_sniffer_and_nic_rx_shared_tir))
2954 			return ERR_PTR(-ENOTSUPP);
2955 
2956 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
2957 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
2958 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
2959 
2960 		prio = &dev->flow_db->sniffer[ft_type];
2961 		priority = 0;
2962 		num_entries = 1;
2963 		num_groups = 1;
2964 	}
2965 
2966 	if (!ns)
2967 		return ERR_PTR(-ENOTSUPP);
2968 
2969 	if (num_entries > max_table_size)
2970 		return ERR_PTR(-ENOMEM);
2971 
2972 	ft = prio->flow_table;
2973 	if (!ft) {
2974 		ft = mlx5_create_auto_grouped_flow_table(ns, priority,
2975 							 num_entries,
2976 							 num_groups,
2977 							 0, 0);
2978 
2979 		if (!IS_ERR(ft)) {
2980 			prio->refcount = 0;
2981 			prio->flow_table = ft;
2982 		} else {
2983 			err = PTR_ERR(ft);
2984 		}
2985 	}
2986 
2987 	return err ? ERR_PTR(err) : prio;
2988 }
2989 
2990 static void set_underlay_qp(struct mlx5_ib_dev *dev,
2991 			    struct mlx5_flow_spec *spec,
2992 			    u32 underlay_qpn)
2993 {
2994 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param,
2995 					   spec->match_criteria,
2996 					   misc_parameters);
2997 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, spec->match_value,
2998 					   misc_parameters);
2999 
3000 	if (underlay_qpn &&
3001 	    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev,
3002 				      ft_field_support.bth_dst_qp)) {
3003 		MLX5_SET(fte_match_set_misc,
3004 			 misc_params_v, bth_dst_qp, underlay_qpn);
3005 		MLX5_SET(fte_match_set_misc,
3006 			 misc_params_c, bth_dst_qp, 0xffffff);
3007 	}
3008 }
3009 
3010 static struct mlx5_ib_flow_handler *_create_flow_rule(struct mlx5_ib_dev *dev,
3011 						      struct mlx5_ib_flow_prio *ft_prio,
3012 						      const struct ib_flow_attr *flow_attr,
3013 						      struct mlx5_flow_destination *dst,
3014 						      u32 underlay_qpn)
3015 {
3016 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
3017 	struct mlx5_ib_flow_handler *handler;
3018 	struct mlx5_flow_act flow_act = {.flow_tag = MLX5_FS_DEFAULT_FLOW_TAG};
3019 	struct mlx5_flow_spec *spec;
3020 	struct mlx5_flow_destination *rule_dst = dst;
3021 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
3022 	unsigned int spec_index;
3023 	int err = 0;
3024 	int dest_num = 1;
3025 	bool is_egress = flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS;
3026 
3027 	if (!is_valid_attr(dev->mdev, flow_attr))
3028 		return ERR_PTR(-EINVAL);
3029 
3030 	spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
3031 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
3032 	if (!handler || !spec) {
3033 		err = -ENOMEM;
3034 		goto free;
3035 	}
3036 
3037 	INIT_LIST_HEAD(&handler->list);
3038 
3039 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
3040 		err = parse_flow_attr(dev->mdev, spec->match_criteria,
3041 				      spec->match_value,
3042 				      ib_flow, flow_attr, &flow_act);
3043 		if (err < 0)
3044 			goto free;
3045 
3046 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
3047 	}
3048 
3049 	if (!flow_is_multicast_only(flow_attr))
3050 		set_underlay_qp(dev, spec, underlay_qpn);
3051 
3052 	if (dev->rep) {
3053 		void *misc;
3054 
3055 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_value,
3056 				    misc_parameters);
3057 		MLX5_SET(fte_match_set_misc, misc, source_port,
3058 			 dev->rep->vport);
3059 		misc = MLX5_ADDR_OF(fte_match_param, spec->match_criteria,
3060 				    misc_parameters);
3061 		MLX5_SET_TO_ONES(fte_match_set_misc, misc, source_port);
3062 	}
3063 
3064 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
3065 
3066 	if (is_egress &&
3067 	    !is_valid_spec(dev->mdev, spec, &flow_act, is_egress)) {
3068 		err = -EINVAL;
3069 		goto free;
3070 	}
3071 
3072 	if (flow_act.action & MLX5_FLOW_CONTEXT_ACTION_DROP) {
3073 		rule_dst = NULL;
3074 		dest_num = 0;
3075 	} else {
3076 		if (is_egress)
3077 			flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_ALLOW;
3078 		else
3079 			flow_act.action |=
3080 				dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
3081 					MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
3082 	}
3083 
3084 	if (flow_act.has_flow_tag &&
3085 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3086 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT)) {
3087 		mlx5_ib_warn(dev, "Flow tag %u and attribute type %x isn't allowed in leftovers\n",
3088 			     flow_act.flow_tag, flow_attr->type);
3089 		err = -EINVAL;
3090 		goto free;
3091 	}
3092 	handler->rule = mlx5_add_flow_rules(ft, spec,
3093 					    &flow_act,
3094 					    rule_dst, dest_num);
3095 
3096 	if (IS_ERR(handler->rule)) {
3097 		err = PTR_ERR(handler->rule);
3098 		goto free;
3099 	}
3100 
3101 	ft_prio->refcount++;
3102 	handler->prio = ft_prio;
3103 
3104 	ft_prio->flow_table = ft;
3105 free:
3106 	if (err)
3107 		kfree(handler);
3108 	kvfree(spec);
3109 	return err ? ERR_PTR(err) : handler;
3110 }
3111 
3112 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
3113 						     struct mlx5_ib_flow_prio *ft_prio,
3114 						     const struct ib_flow_attr *flow_attr,
3115 						     struct mlx5_flow_destination *dst)
3116 {
3117 	return _create_flow_rule(dev, ft_prio, flow_attr, dst, 0);
3118 }
3119 
3120 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
3121 							  struct mlx5_ib_flow_prio *ft_prio,
3122 							  struct ib_flow_attr *flow_attr,
3123 							  struct mlx5_flow_destination *dst)
3124 {
3125 	struct mlx5_ib_flow_handler *handler_dst = NULL;
3126 	struct mlx5_ib_flow_handler *handler = NULL;
3127 
3128 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
3129 	if (!IS_ERR(handler)) {
3130 		handler_dst = create_flow_rule(dev, ft_prio,
3131 					       flow_attr, dst);
3132 		if (IS_ERR(handler_dst)) {
3133 			mlx5_del_flow_rules(handler->rule);
3134 			ft_prio->refcount--;
3135 			kfree(handler);
3136 			handler = handler_dst;
3137 		} else {
3138 			list_add(&handler_dst->list, &handler->list);
3139 		}
3140 	}
3141 
3142 	return handler;
3143 }
3144 enum {
3145 	LEFTOVERS_MC,
3146 	LEFTOVERS_UC,
3147 };
3148 
3149 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
3150 							  struct mlx5_ib_flow_prio *ft_prio,
3151 							  struct ib_flow_attr *flow_attr,
3152 							  struct mlx5_flow_destination *dst)
3153 {
3154 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
3155 	struct mlx5_ib_flow_handler *handler = NULL;
3156 
3157 	static struct {
3158 		struct ib_flow_attr	flow_attr;
3159 		struct ib_flow_spec_eth eth_flow;
3160 	} leftovers_specs[] = {
3161 		[LEFTOVERS_MC] = {
3162 			.flow_attr = {
3163 				.num_of_specs = 1,
3164 				.size = sizeof(leftovers_specs[0])
3165 			},
3166 			.eth_flow = {
3167 				.type = IB_FLOW_SPEC_ETH,
3168 				.size = sizeof(struct ib_flow_spec_eth),
3169 				.mask = {.dst_mac = {0x1} },
3170 				.val =  {.dst_mac = {0x1} }
3171 			}
3172 		},
3173 		[LEFTOVERS_UC] = {
3174 			.flow_attr = {
3175 				.num_of_specs = 1,
3176 				.size = sizeof(leftovers_specs[0])
3177 			},
3178 			.eth_flow = {
3179 				.type = IB_FLOW_SPEC_ETH,
3180 				.size = sizeof(struct ib_flow_spec_eth),
3181 				.mask = {.dst_mac = {0x1} },
3182 				.val = {.dst_mac = {} }
3183 			}
3184 		}
3185 	};
3186 
3187 	handler = create_flow_rule(dev, ft_prio,
3188 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
3189 				   dst);
3190 	if (!IS_ERR(handler) &&
3191 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
3192 		handler_ucast = create_flow_rule(dev, ft_prio,
3193 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
3194 						 dst);
3195 		if (IS_ERR(handler_ucast)) {
3196 			mlx5_del_flow_rules(handler->rule);
3197 			ft_prio->refcount--;
3198 			kfree(handler);
3199 			handler = handler_ucast;
3200 		} else {
3201 			list_add(&handler_ucast->list, &handler->list);
3202 		}
3203 	}
3204 
3205 	return handler;
3206 }
3207 
3208 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
3209 							struct mlx5_ib_flow_prio *ft_rx,
3210 							struct mlx5_ib_flow_prio *ft_tx,
3211 							struct mlx5_flow_destination *dst)
3212 {
3213 	struct mlx5_ib_flow_handler *handler_rx;
3214 	struct mlx5_ib_flow_handler *handler_tx;
3215 	int err;
3216 	static const struct ib_flow_attr flow_attr  = {
3217 		.num_of_specs = 0,
3218 		.size = sizeof(flow_attr)
3219 	};
3220 
3221 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
3222 	if (IS_ERR(handler_rx)) {
3223 		err = PTR_ERR(handler_rx);
3224 		goto err;
3225 	}
3226 
3227 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
3228 	if (IS_ERR(handler_tx)) {
3229 		err = PTR_ERR(handler_tx);
3230 		goto err_tx;
3231 	}
3232 
3233 	list_add(&handler_tx->list, &handler_rx->list);
3234 
3235 	return handler_rx;
3236 
3237 err_tx:
3238 	mlx5_del_flow_rules(handler_rx->rule);
3239 	ft_rx->refcount--;
3240 	kfree(handler_rx);
3241 err:
3242 	return ERR_PTR(err);
3243 }
3244 
3245 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
3246 					   struct ib_flow_attr *flow_attr,
3247 					   int domain)
3248 {
3249 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3250 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3251 	struct mlx5_ib_flow_handler *handler = NULL;
3252 	struct mlx5_flow_destination *dst = NULL;
3253 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
3254 	struct mlx5_ib_flow_prio *ft_prio;
3255 	bool is_egress = flow_attr->flags & IB_FLOW_ATTR_FLAGS_EGRESS;
3256 	int err;
3257 	int underlay_qpn;
3258 
3259 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
3260 		return ERR_PTR(-ENOMEM);
3261 
3262 	if (domain != IB_FLOW_DOMAIN_USER ||
3263 	    flow_attr->port > dev->num_ports ||
3264 	    (flow_attr->flags & ~(IB_FLOW_ATTR_FLAGS_DONT_TRAP |
3265 				  IB_FLOW_ATTR_FLAGS_EGRESS)))
3266 		return ERR_PTR(-EINVAL);
3267 
3268 	if (is_egress &&
3269 	    (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3270 	     flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT))
3271 		return ERR_PTR(-EINVAL);
3272 
3273 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
3274 	if (!dst)
3275 		return ERR_PTR(-ENOMEM);
3276 
3277 	mutex_lock(&dev->flow_db->lock);
3278 
3279 	ft_prio = get_flow_table(dev, flow_attr,
3280 				 is_egress ? MLX5_IB_FT_TX : MLX5_IB_FT_RX);
3281 	if (IS_ERR(ft_prio)) {
3282 		err = PTR_ERR(ft_prio);
3283 		goto unlock;
3284 	}
3285 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3286 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
3287 		if (IS_ERR(ft_prio_tx)) {
3288 			err = PTR_ERR(ft_prio_tx);
3289 			ft_prio_tx = NULL;
3290 			goto destroy_ft;
3291 		}
3292 	}
3293 
3294 	if (is_egress) {
3295 		dst->type = MLX5_FLOW_DESTINATION_TYPE_PORT;
3296 	} else {
3297 		dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
3298 		if (mqp->flags & MLX5_IB_QP_RSS)
3299 			dst->tir_num = mqp->rss_qp.tirn;
3300 		else
3301 			dst->tir_num = mqp->raw_packet_qp.rq.tirn;
3302 	}
3303 
3304 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
3305 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
3306 			handler = create_dont_trap_rule(dev, ft_prio,
3307 							flow_attr, dst);
3308 		} else {
3309 			underlay_qpn = (mqp->flags & MLX5_IB_QP_UNDERLAY) ?
3310 					mqp->underlay_qpn : 0;
3311 			handler = _create_flow_rule(dev, ft_prio, flow_attr,
3312 						    dst, underlay_qpn);
3313 		}
3314 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
3315 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
3316 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
3317 						dst);
3318 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
3319 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
3320 	} else {
3321 		err = -EINVAL;
3322 		goto destroy_ft;
3323 	}
3324 
3325 	if (IS_ERR(handler)) {
3326 		err = PTR_ERR(handler);
3327 		handler = NULL;
3328 		goto destroy_ft;
3329 	}
3330 
3331 	mutex_unlock(&dev->flow_db->lock);
3332 	kfree(dst);
3333 
3334 	return &handler->ibflow;
3335 
3336 destroy_ft:
3337 	put_flow_table(dev, ft_prio, false);
3338 	if (ft_prio_tx)
3339 		put_flow_table(dev, ft_prio_tx, false);
3340 unlock:
3341 	mutex_unlock(&dev->flow_db->lock);
3342 	kfree(dst);
3343 	kfree(handler);
3344 	return ERR_PTR(err);
3345 }
3346 
3347 static u32 mlx5_ib_flow_action_flags_to_accel_xfrm_flags(u32 mlx5_flags)
3348 {
3349 	u32 flags = 0;
3350 
3351 	if (mlx5_flags & MLX5_IB_UAPI_FLOW_ACTION_FLAGS_REQUIRE_METADATA)
3352 		flags |= MLX5_ACCEL_XFRM_FLAG_REQUIRE_METADATA;
3353 
3354 	return flags;
3355 }
3356 
3357 #define MLX5_FLOW_ACTION_ESP_CREATE_LAST_SUPPORTED	MLX5_IB_UAPI_FLOW_ACTION_FLAGS_REQUIRE_METADATA
3358 static struct ib_flow_action *
3359 mlx5_ib_create_flow_action_esp(struct ib_device *device,
3360 			       const struct ib_flow_action_attrs_esp *attr,
3361 			       struct uverbs_attr_bundle *attrs)
3362 {
3363 	struct mlx5_ib_dev *mdev = to_mdev(device);
3364 	struct ib_uverbs_flow_action_esp_keymat_aes_gcm *aes_gcm;
3365 	struct mlx5_accel_esp_xfrm_attrs accel_attrs = {};
3366 	struct mlx5_ib_flow_action *action;
3367 	u64 action_flags;
3368 	u64 flags;
3369 	int err = 0;
3370 
3371 	if (IS_UVERBS_COPY_ERR(uverbs_copy_from(&action_flags, attrs,
3372 						MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS)))
3373 		return ERR_PTR(-EFAULT);
3374 
3375 	if (action_flags >= (MLX5_FLOW_ACTION_ESP_CREATE_LAST_SUPPORTED << 1))
3376 		return ERR_PTR(-EOPNOTSUPP);
3377 
3378 	flags = mlx5_ib_flow_action_flags_to_accel_xfrm_flags(action_flags);
3379 
3380 	/* We current only support a subset of the standard features. Only a
3381 	 * keymat of type AES_GCM, with icv_len == 16, iv_algo == SEQ and esn
3382 	 * (with overlap). Full offload mode isn't supported.
3383 	 */
3384 	if (!attr->keymat || attr->replay || attr->encap ||
3385 	    attr->spi || attr->seq || attr->tfc_pad ||
3386 	    attr->hard_limit_pkts ||
3387 	    (attr->flags & ~(IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
3388 			     IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ENCRYPT)))
3389 		return ERR_PTR(-EOPNOTSUPP);
3390 
3391 	if (attr->keymat->protocol !=
3392 	    IB_UVERBS_FLOW_ACTION_ESP_KEYMAT_AES_GCM)
3393 		return ERR_PTR(-EOPNOTSUPP);
3394 
3395 	aes_gcm = &attr->keymat->keymat.aes_gcm;
3396 
3397 	if (aes_gcm->icv_len != 16 ||
3398 	    aes_gcm->iv_algo != IB_UVERBS_FLOW_ACTION_IV_ALGO_SEQ)
3399 		return ERR_PTR(-EOPNOTSUPP);
3400 
3401 	action = kmalloc(sizeof(*action), GFP_KERNEL);
3402 	if (!action)
3403 		return ERR_PTR(-ENOMEM);
3404 
3405 	action->esp_aes_gcm.ib_flags = attr->flags;
3406 	memcpy(&accel_attrs.keymat.aes_gcm.aes_key, &aes_gcm->aes_key,
3407 	       sizeof(accel_attrs.keymat.aes_gcm.aes_key));
3408 	accel_attrs.keymat.aes_gcm.key_len = aes_gcm->key_len * 8;
3409 	memcpy(&accel_attrs.keymat.aes_gcm.salt, &aes_gcm->salt,
3410 	       sizeof(accel_attrs.keymat.aes_gcm.salt));
3411 	memcpy(&accel_attrs.keymat.aes_gcm.seq_iv, &aes_gcm->iv,
3412 	       sizeof(accel_attrs.keymat.aes_gcm.seq_iv));
3413 	accel_attrs.keymat.aes_gcm.icv_len = aes_gcm->icv_len * 8;
3414 	accel_attrs.keymat.aes_gcm.iv_algo = MLX5_ACCEL_ESP_AES_GCM_IV_ALGO_SEQ;
3415 	accel_attrs.keymat_type = MLX5_ACCEL_ESP_KEYMAT_AES_GCM;
3416 
3417 	accel_attrs.esn = attr->esn;
3418 	if (attr->flags & IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED)
3419 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_TRIGGERED;
3420 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)
3421 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
3422 
3423 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ENCRYPT)
3424 		accel_attrs.action |= MLX5_ACCEL_ESP_ACTION_ENCRYPT;
3425 
3426 	action->esp_aes_gcm.ctx =
3427 		mlx5_accel_esp_create_xfrm(mdev->mdev, &accel_attrs, flags);
3428 	if (IS_ERR(action->esp_aes_gcm.ctx)) {
3429 		err = PTR_ERR(action->esp_aes_gcm.ctx);
3430 		goto err_parse;
3431 	}
3432 
3433 	action->esp_aes_gcm.ib_flags = attr->flags;
3434 
3435 	return &action->ib_action;
3436 
3437 err_parse:
3438 	kfree(action);
3439 	return ERR_PTR(err);
3440 }
3441 
3442 static int
3443 mlx5_ib_modify_flow_action_esp(struct ib_flow_action *action,
3444 			       const struct ib_flow_action_attrs_esp *attr,
3445 			       struct uverbs_attr_bundle *attrs)
3446 {
3447 	struct mlx5_ib_flow_action *maction = to_mflow_act(action);
3448 	struct mlx5_accel_esp_xfrm_attrs accel_attrs;
3449 	int err = 0;
3450 
3451 	if (attr->keymat || attr->replay || attr->encap ||
3452 	    attr->spi || attr->seq || attr->tfc_pad ||
3453 	    attr->hard_limit_pkts ||
3454 	    (attr->flags & ~(IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
3455 			     IB_FLOW_ACTION_ESP_FLAGS_MOD_ESP_ATTRS |
3456 			     IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)))
3457 		return -EOPNOTSUPP;
3458 
3459 	/* Only the ESN value or the MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP can
3460 	 * be modified.
3461 	 */
3462 	if (!(maction->esp_aes_gcm.ib_flags &
3463 	      IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED) &&
3464 	    attr->flags & (IB_FLOW_ACTION_ESP_FLAGS_ESN_TRIGGERED |
3465 			   IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW))
3466 		return -EINVAL;
3467 
3468 	memcpy(&accel_attrs, &maction->esp_aes_gcm.ctx->attrs,
3469 	       sizeof(accel_attrs));
3470 
3471 	accel_attrs.esn = attr->esn;
3472 	if (attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW)
3473 		accel_attrs.flags |= MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
3474 	else
3475 		accel_attrs.flags &= ~MLX5_ACCEL_ESP_FLAGS_ESN_STATE_OVERLAP;
3476 
3477 	err = mlx5_accel_esp_modify_xfrm(maction->esp_aes_gcm.ctx,
3478 					 &accel_attrs);
3479 	if (err)
3480 		return err;
3481 
3482 	maction->esp_aes_gcm.ib_flags &=
3483 		~IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW;
3484 	maction->esp_aes_gcm.ib_flags |=
3485 		attr->flags & IB_UVERBS_FLOW_ACTION_ESP_FLAGS_ESN_NEW_WINDOW;
3486 
3487 	return 0;
3488 }
3489 
3490 static int mlx5_ib_destroy_flow_action(struct ib_flow_action *action)
3491 {
3492 	struct mlx5_ib_flow_action *maction = to_mflow_act(action);
3493 
3494 	switch (action->type) {
3495 	case IB_FLOW_ACTION_ESP:
3496 		/*
3497 		 * We only support aes_gcm by now, so we implicitly know this is
3498 		 * the underline crypto.
3499 		 */
3500 		mlx5_accel_esp_destroy_xfrm(maction->esp_aes_gcm.ctx);
3501 		break;
3502 	default:
3503 		WARN_ON(true);
3504 		break;
3505 	}
3506 
3507 	kfree(maction);
3508 	return 0;
3509 }
3510 
3511 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
3512 {
3513 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3514 	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
3515 	int err;
3516 
3517 	if (mqp->flags & MLX5_IB_QP_UNDERLAY) {
3518 		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
3519 		return -EOPNOTSUPP;
3520 	}
3521 
3522 	err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
3523 	if (err)
3524 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
3525 			     ibqp->qp_num, gid->raw);
3526 
3527 	return err;
3528 }
3529 
3530 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
3531 {
3532 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3533 	int err;
3534 
3535 	err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
3536 	if (err)
3537 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
3538 			     ibqp->qp_num, gid->raw);
3539 
3540 	return err;
3541 }
3542 
3543 static int init_node_data(struct mlx5_ib_dev *dev)
3544 {
3545 	int err;
3546 
3547 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
3548 	if (err)
3549 		return err;
3550 
3551 	dev->mdev->rev_id = dev->mdev->pdev->revision;
3552 
3553 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
3554 }
3555 
3556 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
3557 			     char *buf)
3558 {
3559 	struct mlx5_ib_dev *dev =
3560 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3561 
3562 	return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
3563 }
3564 
3565 static ssize_t show_reg_pages(struct device *device,
3566 			      struct device_attribute *attr, char *buf)
3567 {
3568 	struct mlx5_ib_dev *dev =
3569 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3570 
3571 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
3572 }
3573 
3574 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
3575 			char *buf)
3576 {
3577 	struct mlx5_ib_dev *dev =
3578 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3579 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
3580 }
3581 
3582 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
3583 			char *buf)
3584 {
3585 	struct mlx5_ib_dev *dev =
3586 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3587 	return sprintf(buf, "%x\n", dev->mdev->rev_id);
3588 }
3589 
3590 static ssize_t show_board(struct device *device, struct device_attribute *attr,
3591 			  char *buf)
3592 {
3593 	struct mlx5_ib_dev *dev =
3594 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
3595 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
3596 		       dev->mdev->board_id);
3597 }
3598 
3599 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
3600 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
3601 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
3602 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
3603 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
3604 
3605 static struct device_attribute *mlx5_class_attributes[] = {
3606 	&dev_attr_hw_rev,
3607 	&dev_attr_hca_type,
3608 	&dev_attr_board_id,
3609 	&dev_attr_fw_pages,
3610 	&dev_attr_reg_pages,
3611 };
3612 
3613 static void pkey_change_handler(struct work_struct *work)
3614 {
3615 	struct mlx5_ib_port_resources *ports =
3616 		container_of(work, struct mlx5_ib_port_resources,
3617 			     pkey_change_work);
3618 
3619 	mutex_lock(&ports->devr->mutex);
3620 	mlx5_ib_gsi_pkey_change(ports->gsi);
3621 	mutex_unlock(&ports->devr->mutex);
3622 }
3623 
3624 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
3625 {
3626 	struct mlx5_ib_qp *mqp;
3627 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
3628 	struct mlx5_core_cq *mcq;
3629 	struct list_head cq_armed_list;
3630 	unsigned long flags_qp;
3631 	unsigned long flags_cq;
3632 	unsigned long flags;
3633 
3634 	INIT_LIST_HEAD(&cq_armed_list);
3635 
3636 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
3637 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
3638 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
3639 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
3640 		if (mqp->sq.tail != mqp->sq.head) {
3641 			send_mcq = to_mcq(mqp->ibqp.send_cq);
3642 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
3643 			if (send_mcq->mcq.comp &&
3644 			    mqp->ibqp.send_cq->comp_handler) {
3645 				if (!send_mcq->mcq.reset_notify_added) {
3646 					send_mcq->mcq.reset_notify_added = 1;
3647 					list_add_tail(&send_mcq->mcq.reset_notify,
3648 						      &cq_armed_list);
3649 				}
3650 			}
3651 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
3652 		}
3653 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
3654 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
3655 		/* no handling is needed for SRQ */
3656 		if (!mqp->ibqp.srq) {
3657 			if (mqp->rq.tail != mqp->rq.head) {
3658 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
3659 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
3660 				if (recv_mcq->mcq.comp &&
3661 				    mqp->ibqp.recv_cq->comp_handler) {
3662 					if (!recv_mcq->mcq.reset_notify_added) {
3663 						recv_mcq->mcq.reset_notify_added = 1;
3664 						list_add_tail(&recv_mcq->mcq.reset_notify,
3665 							      &cq_armed_list);
3666 					}
3667 				}
3668 				spin_unlock_irqrestore(&recv_mcq->lock,
3669 						       flags_cq);
3670 			}
3671 		}
3672 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
3673 	}
3674 	/*At that point all inflight post send were put to be executed as of we
3675 	 * lock/unlock above locks Now need to arm all involved CQs.
3676 	 */
3677 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
3678 		mcq->comp(mcq);
3679 	}
3680 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
3681 }
3682 
3683 static void delay_drop_handler(struct work_struct *work)
3684 {
3685 	int err;
3686 	struct mlx5_ib_delay_drop *delay_drop =
3687 		container_of(work, struct mlx5_ib_delay_drop,
3688 			     delay_drop_work);
3689 
3690 	atomic_inc(&delay_drop->events_cnt);
3691 
3692 	mutex_lock(&delay_drop->lock);
3693 	err = mlx5_core_set_delay_drop(delay_drop->dev->mdev,
3694 				       delay_drop->timeout);
3695 	if (err) {
3696 		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
3697 			     delay_drop->timeout);
3698 		delay_drop->activate = false;
3699 	}
3700 	mutex_unlock(&delay_drop->lock);
3701 }
3702 
3703 static void mlx5_ib_handle_event(struct work_struct *_work)
3704 {
3705 	struct mlx5_ib_event_work *work =
3706 		container_of(_work, struct mlx5_ib_event_work, work);
3707 	struct mlx5_ib_dev *ibdev;
3708 	struct ib_event ibev;
3709 	bool fatal = false;
3710 	u8 port = (u8)work->param;
3711 
3712 	if (mlx5_core_is_mp_slave(work->dev)) {
3713 		ibdev = mlx5_ib_get_ibdev_from_mpi(work->context);
3714 		if (!ibdev)
3715 			goto out;
3716 	} else {
3717 		ibdev = work->context;
3718 	}
3719 
3720 	switch (work->event) {
3721 	case MLX5_DEV_EVENT_SYS_ERROR:
3722 		ibev.event = IB_EVENT_DEVICE_FATAL;
3723 		mlx5_ib_handle_internal_error(ibdev);
3724 		fatal = true;
3725 		break;
3726 
3727 	case MLX5_DEV_EVENT_PORT_UP:
3728 	case MLX5_DEV_EVENT_PORT_DOWN:
3729 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
3730 		/* In RoCE, port up/down events are handled in
3731 		 * mlx5_netdev_event().
3732 		 */
3733 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
3734 			IB_LINK_LAYER_ETHERNET)
3735 			goto out;
3736 
3737 		ibev.event = (work->event == MLX5_DEV_EVENT_PORT_UP) ?
3738 			     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
3739 		break;
3740 
3741 	case MLX5_DEV_EVENT_LID_CHANGE:
3742 		ibev.event = IB_EVENT_LID_CHANGE;
3743 		break;
3744 
3745 	case MLX5_DEV_EVENT_PKEY_CHANGE:
3746 		ibev.event = IB_EVENT_PKEY_CHANGE;
3747 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
3748 		break;
3749 
3750 	case MLX5_DEV_EVENT_GUID_CHANGE:
3751 		ibev.event = IB_EVENT_GID_CHANGE;
3752 		break;
3753 
3754 	case MLX5_DEV_EVENT_CLIENT_REREG:
3755 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
3756 		break;
3757 	case MLX5_DEV_EVENT_DELAY_DROP_TIMEOUT:
3758 		schedule_work(&ibdev->delay_drop.delay_drop_work);
3759 		goto out;
3760 	default:
3761 		goto out;
3762 	}
3763 
3764 	ibev.device	      = &ibdev->ib_dev;
3765 	ibev.element.port_num = port;
3766 
3767 	if (!rdma_is_port_valid(&ibdev->ib_dev, port)) {
3768 		mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
3769 		goto out;
3770 	}
3771 
3772 	if (ibdev->ib_active)
3773 		ib_dispatch_event(&ibev);
3774 
3775 	if (fatal)
3776 		ibdev->ib_active = false;
3777 out:
3778 	kfree(work);
3779 }
3780 
3781 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
3782 			  enum mlx5_dev_event event, unsigned long param)
3783 {
3784 	struct mlx5_ib_event_work *work;
3785 
3786 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
3787 	if (!work)
3788 		return;
3789 
3790 	INIT_WORK(&work->work, mlx5_ib_handle_event);
3791 	work->dev = dev;
3792 	work->param = param;
3793 	work->context = context;
3794 	work->event = event;
3795 
3796 	queue_work(mlx5_ib_event_wq, &work->work);
3797 }
3798 
3799 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
3800 {
3801 	struct mlx5_hca_vport_context vport_ctx;
3802 	int err;
3803 	int port;
3804 
3805 	for (port = 1; port <= dev->num_ports; port++) {
3806 		dev->mdev->port_caps[port - 1].has_smi = false;
3807 		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
3808 		    MLX5_CAP_PORT_TYPE_IB) {
3809 			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
3810 				err = mlx5_query_hca_vport_context(dev->mdev, 0,
3811 								   port, 0,
3812 								   &vport_ctx);
3813 				if (err) {
3814 					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
3815 						    port, err);
3816 					return err;
3817 				}
3818 				dev->mdev->port_caps[port - 1].has_smi =
3819 					vport_ctx.has_smi;
3820 			} else {
3821 				dev->mdev->port_caps[port - 1].has_smi = true;
3822 			}
3823 		}
3824 	}
3825 	return 0;
3826 }
3827 
3828 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
3829 {
3830 	int port;
3831 
3832 	for (port = 1; port <= dev->num_ports; port++)
3833 		mlx5_query_ext_port_caps(dev, port);
3834 }
3835 
3836 static int get_port_caps(struct mlx5_ib_dev *dev, u8 port)
3837 {
3838 	struct ib_device_attr *dprops = NULL;
3839 	struct ib_port_attr *pprops = NULL;
3840 	int err = -ENOMEM;
3841 	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
3842 
3843 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
3844 	if (!pprops)
3845 		goto out;
3846 
3847 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
3848 	if (!dprops)
3849 		goto out;
3850 
3851 	err = set_has_smi_cap(dev);
3852 	if (err)
3853 		goto out;
3854 
3855 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
3856 	if (err) {
3857 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
3858 		goto out;
3859 	}
3860 
3861 	memset(pprops, 0, sizeof(*pprops));
3862 	err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
3863 	if (err) {
3864 		mlx5_ib_warn(dev, "query_port %d failed %d\n",
3865 			     port, err);
3866 		goto out;
3867 	}
3868 
3869 	dev->mdev->port_caps[port - 1].pkey_table_len =
3870 					dprops->max_pkeys;
3871 	dev->mdev->port_caps[port - 1].gid_table_len =
3872 					pprops->gid_tbl_len;
3873 	mlx5_ib_dbg(dev, "port %d: pkey_table_len %d, gid_table_len %d\n",
3874 		    port, dprops->max_pkeys, pprops->gid_tbl_len);
3875 
3876 out:
3877 	kfree(pprops);
3878 	kfree(dprops);
3879 
3880 	return err;
3881 }
3882 
3883 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
3884 {
3885 	int err;
3886 
3887 	err = mlx5_mr_cache_cleanup(dev);
3888 	if (err)
3889 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
3890 
3891 	if (dev->umrc.qp)
3892 		mlx5_ib_destroy_qp(dev->umrc.qp);
3893 	if (dev->umrc.cq)
3894 		ib_free_cq(dev->umrc.cq);
3895 	if (dev->umrc.pd)
3896 		ib_dealloc_pd(dev->umrc.pd);
3897 }
3898 
3899 enum {
3900 	MAX_UMR_WR = 128,
3901 };
3902 
3903 static int create_umr_res(struct mlx5_ib_dev *dev)
3904 {
3905 	struct ib_qp_init_attr *init_attr = NULL;
3906 	struct ib_qp_attr *attr = NULL;
3907 	struct ib_pd *pd;
3908 	struct ib_cq *cq;
3909 	struct ib_qp *qp;
3910 	int ret;
3911 
3912 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
3913 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
3914 	if (!attr || !init_attr) {
3915 		ret = -ENOMEM;
3916 		goto error_0;
3917 	}
3918 
3919 	pd = ib_alloc_pd(&dev->ib_dev, 0);
3920 	if (IS_ERR(pd)) {
3921 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
3922 		ret = PTR_ERR(pd);
3923 		goto error_0;
3924 	}
3925 
3926 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
3927 	if (IS_ERR(cq)) {
3928 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
3929 		ret = PTR_ERR(cq);
3930 		goto error_2;
3931 	}
3932 
3933 	init_attr->send_cq = cq;
3934 	init_attr->recv_cq = cq;
3935 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
3936 	init_attr->cap.max_send_wr = MAX_UMR_WR;
3937 	init_attr->cap.max_send_sge = 1;
3938 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
3939 	init_attr->port_num = 1;
3940 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
3941 	if (IS_ERR(qp)) {
3942 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
3943 		ret = PTR_ERR(qp);
3944 		goto error_3;
3945 	}
3946 	qp->device     = &dev->ib_dev;
3947 	qp->real_qp    = qp;
3948 	qp->uobject    = NULL;
3949 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
3950 	qp->send_cq    = init_attr->send_cq;
3951 	qp->recv_cq    = init_attr->recv_cq;
3952 
3953 	attr->qp_state = IB_QPS_INIT;
3954 	attr->port_num = 1;
3955 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
3956 				IB_QP_PORT, NULL);
3957 	if (ret) {
3958 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
3959 		goto error_4;
3960 	}
3961 
3962 	memset(attr, 0, sizeof(*attr));
3963 	attr->qp_state = IB_QPS_RTR;
3964 	attr->path_mtu = IB_MTU_256;
3965 
3966 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
3967 	if (ret) {
3968 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
3969 		goto error_4;
3970 	}
3971 
3972 	memset(attr, 0, sizeof(*attr));
3973 	attr->qp_state = IB_QPS_RTS;
3974 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
3975 	if (ret) {
3976 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
3977 		goto error_4;
3978 	}
3979 
3980 	dev->umrc.qp = qp;
3981 	dev->umrc.cq = cq;
3982 	dev->umrc.pd = pd;
3983 
3984 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
3985 	ret = mlx5_mr_cache_init(dev);
3986 	if (ret) {
3987 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
3988 		goto error_4;
3989 	}
3990 
3991 	kfree(attr);
3992 	kfree(init_attr);
3993 
3994 	return 0;
3995 
3996 error_4:
3997 	mlx5_ib_destroy_qp(qp);
3998 	dev->umrc.qp = NULL;
3999 
4000 error_3:
4001 	ib_free_cq(cq);
4002 	dev->umrc.cq = NULL;
4003 
4004 error_2:
4005 	ib_dealloc_pd(pd);
4006 	dev->umrc.pd = NULL;
4007 
4008 error_0:
4009 	kfree(attr);
4010 	kfree(init_attr);
4011 	return ret;
4012 }
4013 
4014 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
4015 {
4016 	switch (umr_fence_cap) {
4017 	case MLX5_CAP_UMR_FENCE_NONE:
4018 		return MLX5_FENCE_MODE_NONE;
4019 	case MLX5_CAP_UMR_FENCE_SMALL:
4020 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
4021 	default:
4022 		return MLX5_FENCE_MODE_STRONG_ORDERING;
4023 	}
4024 }
4025 
4026 static int create_dev_resources(struct mlx5_ib_resources *devr)
4027 {
4028 	struct ib_srq_init_attr attr;
4029 	struct mlx5_ib_dev *dev;
4030 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
4031 	int port;
4032 	int ret = 0;
4033 
4034 	dev = container_of(devr, struct mlx5_ib_dev, devr);
4035 
4036 	mutex_init(&devr->mutex);
4037 
4038 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
4039 	if (IS_ERR(devr->p0)) {
4040 		ret = PTR_ERR(devr->p0);
4041 		goto error0;
4042 	}
4043 	devr->p0->device  = &dev->ib_dev;
4044 	devr->p0->uobject = NULL;
4045 	atomic_set(&devr->p0->usecnt, 0);
4046 
4047 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
4048 	if (IS_ERR(devr->c0)) {
4049 		ret = PTR_ERR(devr->c0);
4050 		goto error1;
4051 	}
4052 	devr->c0->device        = &dev->ib_dev;
4053 	devr->c0->uobject       = NULL;
4054 	devr->c0->comp_handler  = NULL;
4055 	devr->c0->event_handler = NULL;
4056 	devr->c0->cq_context    = NULL;
4057 	atomic_set(&devr->c0->usecnt, 0);
4058 
4059 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
4060 	if (IS_ERR(devr->x0)) {
4061 		ret = PTR_ERR(devr->x0);
4062 		goto error2;
4063 	}
4064 	devr->x0->device = &dev->ib_dev;
4065 	devr->x0->inode = NULL;
4066 	atomic_set(&devr->x0->usecnt, 0);
4067 	mutex_init(&devr->x0->tgt_qp_mutex);
4068 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
4069 
4070 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
4071 	if (IS_ERR(devr->x1)) {
4072 		ret = PTR_ERR(devr->x1);
4073 		goto error3;
4074 	}
4075 	devr->x1->device = &dev->ib_dev;
4076 	devr->x1->inode = NULL;
4077 	atomic_set(&devr->x1->usecnt, 0);
4078 	mutex_init(&devr->x1->tgt_qp_mutex);
4079 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
4080 
4081 	memset(&attr, 0, sizeof(attr));
4082 	attr.attr.max_sge = 1;
4083 	attr.attr.max_wr = 1;
4084 	attr.srq_type = IB_SRQT_XRC;
4085 	attr.ext.cq = devr->c0;
4086 	attr.ext.xrc.xrcd = devr->x0;
4087 
4088 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
4089 	if (IS_ERR(devr->s0)) {
4090 		ret = PTR_ERR(devr->s0);
4091 		goto error4;
4092 	}
4093 	devr->s0->device	= &dev->ib_dev;
4094 	devr->s0->pd		= devr->p0;
4095 	devr->s0->uobject       = NULL;
4096 	devr->s0->event_handler = NULL;
4097 	devr->s0->srq_context   = NULL;
4098 	devr->s0->srq_type      = IB_SRQT_XRC;
4099 	devr->s0->ext.xrc.xrcd	= devr->x0;
4100 	devr->s0->ext.cq	= devr->c0;
4101 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
4102 	atomic_inc(&devr->s0->ext.cq->usecnt);
4103 	atomic_inc(&devr->p0->usecnt);
4104 	atomic_set(&devr->s0->usecnt, 0);
4105 
4106 	memset(&attr, 0, sizeof(attr));
4107 	attr.attr.max_sge = 1;
4108 	attr.attr.max_wr = 1;
4109 	attr.srq_type = IB_SRQT_BASIC;
4110 	devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
4111 	if (IS_ERR(devr->s1)) {
4112 		ret = PTR_ERR(devr->s1);
4113 		goto error5;
4114 	}
4115 	devr->s1->device	= &dev->ib_dev;
4116 	devr->s1->pd		= devr->p0;
4117 	devr->s1->uobject       = NULL;
4118 	devr->s1->event_handler = NULL;
4119 	devr->s1->srq_context   = NULL;
4120 	devr->s1->srq_type      = IB_SRQT_BASIC;
4121 	devr->s1->ext.cq	= devr->c0;
4122 	atomic_inc(&devr->p0->usecnt);
4123 	atomic_set(&devr->s1->usecnt, 0);
4124 
4125 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
4126 		INIT_WORK(&devr->ports[port].pkey_change_work,
4127 			  pkey_change_handler);
4128 		devr->ports[port].devr = devr;
4129 	}
4130 
4131 	return 0;
4132 
4133 error5:
4134 	mlx5_ib_destroy_srq(devr->s0);
4135 error4:
4136 	mlx5_ib_dealloc_xrcd(devr->x1);
4137 error3:
4138 	mlx5_ib_dealloc_xrcd(devr->x0);
4139 error2:
4140 	mlx5_ib_destroy_cq(devr->c0);
4141 error1:
4142 	mlx5_ib_dealloc_pd(devr->p0);
4143 error0:
4144 	return ret;
4145 }
4146 
4147 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
4148 {
4149 	struct mlx5_ib_dev *dev =
4150 		container_of(devr, struct mlx5_ib_dev, devr);
4151 	int port;
4152 
4153 	mlx5_ib_destroy_srq(devr->s1);
4154 	mlx5_ib_destroy_srq(devr->s0);
4155 	mlx5_ib_dealloc_xrcd(devr->x0);
4156 	mlx5_ib_dealloc_xrcd(devr->x1);
4157 	mlx5_ib_destroy_cq(devr->c0);
4158 	mlx5_ib_dealloc_pd(devr->p0);
4159 
4160 	/* Make sure no change P_Key work items are still executing */
4161 	for (port = 0; port < dev->num_ports; ++port)
4162 		cancel_work_sync(&devr->ports[port].pkey_change_work);
4163 }
4164 
4165 static u32 get_core_cap_flags(struct ib_device *ibdev)
4166 {
4167 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4168 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
4169 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
4170 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
4171 	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
4172 	u32 ret = 0;
4173 
4174 	if (ll == IB_LINK_LAYER_INFINIBAND)
4175 		return RDMA_CORE_PORT_IBA_IB;
4176 
4177 	if (raw_support)
4178 		ret = RDMA_CORE_PORT_RAW_PACKET;
4179 
4180 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
4181 		return ret;
4182 
4183 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
4184 		return ret;
4185 
4186 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
4187 		ret |= RDMA_CORE_PORT_IBA_ROCE;
4188 
4189 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
4190 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
4191 
4192 	return ret;
4193 }
4194 
4195 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
4196 			       struct ib_port_immutable *immutable)
4197 {
4198 	struct ib_port_attr attr;
4199 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4200 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
4201 	int err;
4202 
4203 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
4204 
4205 	err = ib_query_port(ibdev, port_num, &attr);
4206 	if (err)
4207 		return err;
4208 
4209 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
4210 	immutable->gid_tbl_len = attr.gid_tbl_len;
4211 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
4212 	if ((ll == IB_LINK_LAYER_INFINIBAND) || MLX5_CAP_GEN(dev->mdev, roce))
4213 		immutable->max_mad_size = IB_MGMT_MAD_SIZE;
4214 
4215 	return 0;
4216 }
4217 
4218 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u8 port_num,
4219 				   struct ib_port_immutable *immutable)
4220 {
4221 	struct ib_port_attr attr;
4222 	int err;
4223 
4224 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
4225 
4226 	err = ib_query_port(ibdev, port_num, &attr);
4227 	if (err)
4228 		return err;
4229 
4230 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
4231 	immutable->gid_tbl_len = attr.gid_tbl_len;
4232 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
4233 
4234 	return 0;
4235 }
4236 
4237 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
4238 {
4239 	struct mlx5_ib_dev *dev =
4240 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
4241 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
4242 		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
4243 		 fw_rev_sub(dev->mdev));
4244 }
4245 
4246 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
4247 {
4248 	struct mlx5_core_dev *mdev = dev->mdev;
4249 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
4250 								 MLX5_FLOW_NAMESPACE_LAG);
4251 	struct mlx5_flow_table *ft;
4252 	int err;
4253 
4254 	if (!ns || !mlx5_lag_is_active(mdev))
4255 		return 0;
4256 
4257 	err = mlx5_cmd_create_vport_lag(mdev);
4258 	if (err)
4259 		return err;
4260 
4261 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
4262 	if (IS_ERR(ft)) {
4263 		err = PTR_ERR(ft);
4264 		goto err_destroy_vport_lag;
4265 	}
4266 
4267 	dev->flow_db->lag_demux_ft = ft;
4268 	return 0;
4269 
4270 err_destroy_vport_lag:
4271 	mlx5_cmd_destroy_vport_lag(mdev);
4272 	return err;
4273 }
4274 
4275 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
4276 {
4277 	struct mlx5_core_dev *mdev = dev->mdev;
4278 
4279 	if (dev->flow_db->lag_demux_ft) {
4280 		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
4281 		dev->flow_db->lag_demux_ft = NULL;
4282 
4283 		mlx5_cmd_destroy_vport_lag(mdev);
4284 	}
4285 }
4286 
4287 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
4288 {
4289 	int err;
4290 
4291 	dev->roce[port_num].nb.notifier_call = mlx5_netdev_event;
4292 	err = register_netdevice_notifier(&dev->roce[port_num].nb);
4293 	if (err) {
4294 		dev->roce[port_num].nb.notifier_call = NULL;
4295 		return err;
4296 	}
4297 
4298 	return 0;
4299 }
4300 
4301 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
4302 {
4303 	if (dev->roce[port_num].nb.notifier_call) {
4304 		unregister_netdevice_notifier(&dev->roce[port_num].nb);
4305 		dev->roce[port_num].nb.notifier_call = NULL;
4306 	}
4307 }
4308 
4309 static int mlx5_enable_eth(struct mlx5_ib_dev *dev, u8 port_num)
4310 {
4311 	int err;
4312 
4313 	if (MLX5_CAP_GEN(dev->mdev, roce)) {
4314 		err = mlx5_nic_vport_enable_roce(dev->mdev);
4315 		if (err)
4316 			return err;
4317 	}
4318 
4319 	err = mlx5_eth_lag_init(dev);
4320 	if (err)
4321 		goto err_disable_roce;
4322 
4323 	return 0;
4324 
4325 err_disable_roce:
4326 	if (MLX5_CAP_GEN(dev->mdev, roce))
4327 		mlx5_nic_vport_disable_roce(dev->mdev);
4328 
4329 	return err;
4330 }
4331 
4332 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
4333 {
4334 	mlx5_eth_lag_cleanup(dev);
4335 	if (MLX5_CAP_GEN(dev->mdev, roce))
4336 		mlx5_nic_vport_disable_roce(dev->mdev);
4337 }
4338 
4339 struct mlx5_ib_counter {
4340 	const char *name;
4341 	size_t offset;
4342 };
4343 
4344 #define INIT_Q_COUNTER(_name)		\
4345 	{ .name = #_name, .offset = MLX5_BYTE_OFF(query_q_counter_out, _name)}
4346 
4347 static const struct mlx5_ib_counter basic_q_cnts[] = {
4348 	INIT_Q_COUNTER(rx_write_requests),
4349 	INIT_Q_COUNTER(rx_read_requests),
4350 	INIT_Q_COUNTER(rx_atomic_requests),
4351 	INIT_Q_COUNTER(out_of_buffer),
4352 };
4353 
4354 static const struct mlx5_ib_counter out_of_seq_q_cnts[] = {
4355 	INIT_Q_COUNTER(out_of_sequence),
4356 };
4357 
4358 static const struct mlx5_ib_counter retrans_q_cnts[] = {
4359 	INIT_Q_COUNTER(duplicate_request),
4360 	INIT_Q_COUNTER(rnr_nak_retry_err),
4361 	INIT_Q_COUNTER(packet_seq_err),
4362 	INIT_Q_COUNTER(implied_nak_seq_err),
4363 	INIT_Q_COUNTER(local_ack_timeout_err),
4364 };
4365 
4366 #define INIT_CONG_COUNTER(_name)		\
4367 	{ .name = #_name, .offset =	\
4368 		MLX5_BYTE_OFF(query_cong_statistics_out, _name ## _high)}
4369 
4370 static const struct mlx5_ib_counter cong_cnts[] = {
4371 	INIT_CONG_COUNTER(rp_cnp_ignored),
4372 	INIT_CONG_COUNTER(rp_cnp_handled),
4373 	INIT_CONG_COUNTER(np_ecn_marked_roce_packets),
4374 	INIT_CONG_COUNTER(np_cnp_sent),
4375 };
4376 
4377 static const struct mlx5_ib_counter extended_err_cnts[] = {
4378 	INIT_Q_COUNTER(resp_local_length_error),
4379 	INIT_Q_COUNTER(resp_cqe_error),
4380 	INIT_Q_COUNTER(req_cqe_error),
4381 	INIT_Q_COUNTER(req_remote_invalid_request),
4382 	INIT_Q_COUNTER(req_remote_access_errors),
4383 	INIT_Q_COUNTER(resp_remote_access_errors),
4384 	INIT_Q_COUNTER(resp_cqe_flush_error),
4385 	INIT_Q_COUNTER(req_cqe_flush_error),
4386 };
4387 
4388 static void mlx5_ib_dealloc_counters(struct mlx5_ib_dev *dev)
4389 {
4390 	int i;
4391 
4392 	for (i = 0; i < dev->num_ports; i++) {
4393 		if (dev->port[i].cnts.set_id)
4394 			mlx5_core_dealloc_q_counter(dev->mdev,
4395 						    dev->port[i].cnts.set_id);
4396 		kfree(dev->port[i].cnts.names);
4397 		kfree(dev->port[i].cnts.offsets);
4398 	}
4399 }
4400 
4401 static int __mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev,
4402 				    struct mlx5_ib_counters *cnts)
4403 {
4404 	u32 num_counters;
4405 
4406 	num_counters = ARRAY_SIZE(basic_q_cnts);
4407 
4408 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt))
4409 		num_counters += ARRAY_SIZE(out_of_seq_q_cnts);
4410 
4411 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters))
4412 		num_counters += ARRAY_SIZE(retrans_q_cnts);
4413 
4414 	if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters))
4415 		num_counters += ARRAY_SIZE(extended_err_cnts);
4416 
4417 	cnts->num_q_counters = num_counters;
4418 
4419 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
4420 		cnts->num_cong_counters = ARRAY_SIZE(cong_cnts);
4421 		num_counters += ARRAY_SIZE(cong_cnts);
4422 	}
4423 
4424 	cnts->names = kcalloc(num_counters, sizeof(cnts->names), GFP_KERNEL);
4425 	if (!cnts->names)
4426 		return -ENOMEM;
4427 
4428 	cnts->offsets = kcalloc(num_counters,
4429 				sizeof(cnts->offsets), GFP_KERNEL);
4430 	if (!cnts->offsets)
4431 		goto err_names;
4432 
4433 	return 0;
4434 
4435 err_names:
4436 	kfree(cnts->names);
4437 	cnts->names = NULL;
4438 	return -ENOMEM;
4439 }
4440 
4441 static void mlx5_ib_fill_counters(struct mlx5_ib_dev *dev,
4442 				  const char **names,
4443 				  size_t *offsets)
4444 {
4445 	int i;
4446 	int j = 0;
4447 
4448 	for (i = 0; i < ARRAY_SIZE(basic_q_cnts); i++, j++) {
4449 		names[j] = basic_q_cnts[i].name;
4450 		offsets[j] = basic_q_cnts[i].offset;
4451 	}
4452 
4453 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt)) {
4454 		for (i = 0; i < ARRAY_SIZE(out_of_seq_q_cnts); i++, j++) {
4455 			names[j] = out_of_seq_q_cnts[i].name;
4456 			offsets[j] = out_of_seq_q_cnts[i].offset;
4457 		}
4458 	}
4459 
4460 	if (MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
4461 		for (i = 0; i < ARRAY_SIZE(retrans_q_cnts); i++, j++) {
4462 			names[j] = retrans_q_cnts[i].name;
4463 			offsets[j] = retrans_q_cnts[i].offset;
4464 		}
4465 	}
4466 
4467 	if (MLX5_CAP_GEN(dev->mdev, enhanced_error_q_counters)) {
4468 		for (i = 0; i < ARRAY_SIZE(extended_err_cnts); i++, j++) {
4469 			names[j] = extended_err_cnts[i].name;
4470 			offsets[j] = extended_err_cnts[i].offset;
4471 		}
4472 	}
4473 
4474 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
4475 		for (i = 0; i < ARRAY_SIZE(cong_cnts); i++, j++) {
4476 			names[j] = cong_cnts[i].name;
4477 			offsets[j] = cong_cnts[i].offset;
4478 		}
4479 	}
4480 }
4481 
4482 static int mlx5_ib_alloc_counters(struct mlx5_ib_dev *dev)
4483 {
4484 	int err = 0;
4485 	int i;
4486 
4487 	for (i = 0; i < dev->num_ports; i++) {
4488 		err = __mlx5_ib_alloc_counters(dev, &dev->port[i].cnts);
4489 		if (err)
4490 			goto err_alloc;
4491 
4492 		mlx5_ib_fill_counters(dev, dev->port[i].cnts.names,
4493 				      dev->port[i].cnts.offsets);
4494 
4495 		err = mlx5_core_alloc_q_counter(dev->mdev,
4496 						&dev->port[i].cnts.set_id);
4497 		if (err) {
4498 			mlx5_ib_warn(dev,
4499 				     "couldn't allocate queue counter for port %d, err %d\n",
4500 				     i + 1, err);
4501 			goto err_alloc;
4502 		}
4503 		dev->port[i].cnts.set_id_valid = true;
4504 	}
4505 
4506 	return 0;
4507 
4508 err_alloc:
4509 	mlx5_ib_dealloc_counters(dev);
4510 	return err;
4511 }
4512 
4513 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
4514 						    u8 port_num)
4515 {
4516 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4517 	struct mlx5_ib_port *port = &dev->port[port_num - 1];
4518 
4519 	/* We support only per port stats */
4520 	if (port_num == 0)
4521 		return NULL;
4522 
4523 	return rdma_alloc_hw_stats_struct(port->cnts.names,
4524 					  port->cnts.num_q_counters +
4525 					  port->cnts.num_cong_counters,
4526 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
4527 }
4528 
4529 static int mlx5_ib_query_q_counters(struct mlx5_core_dev *mdev,
4530 				    struct mlx5_ib_port *port,
4531 				    struct rdma_hw_stats *stats)
4532 {
4533 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
4534 	void *out;
4535 	__be32 val;
4536 	int ret, i;
4537 
4538 	out = kvzalloc(outlen, GFP_KERNEL);
4539 	if (!out)
4540 		return -ENOMEM;
4541 
4542 	ret = mlx5_core_query_q_counter(mdev,
4543 					port->cnts.set_id, 0,
4544 					out, outlen);
4545 	if (ret)
4546 		goto free;
4547 
4548 	for (i = 0; i < port->cnts.num_q_counters; i++) {
4549 		val = *(__be32 *)(out + port->cnts.offsets[i]);
4550 		stats->value[i] = (u64)be32_to_cpu(val);
4551 	}
4552 
4553 free:
4554 	kvfree(out);
4555 	return ret;
4556 }
4557 
4558 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
4559 				struct rdma_hw_stats *stats,
4560 				u8 port_num, int index)
4561 {
4562 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4563 	struct mlx5_ib_port *port = &dev->port[port_num - 1];
4564 	struct mlx5_core_dev *mdev;
4565 	int ret, num_counters;
4566 	u8 mdev_port_num;
4567 
4568 	if (!stats)
4569 		return -EINVAL;
4570 
4571 	num_counters = port->cnts.num_q_counters + port->cnts.num_cong_counters;
4572 
4573 	/* q_counters are per IB device, query the master mdev */
4574 	ret = mlx5_ib_query_q_counters(dev->mdev, port, stats);
4575 	if (ret)
4576 		return ret;
4577 
4578 	if (MLX5_CAP_GEN(dev->mdev, cc_query_allowed)) {
4579 		mdev = mlx5_ib_get_native_port_mdev(dev, port_num,
4580 						    &mdev_port_num);
4581 		if (!mdev) {
4582 			/* If port is not affiliated yet, its in down state
4583 			 * which doesn't have any counters yet, so it would be
4584 			 * zero. So no need to read from the HCA.
4585 			 */
4586 			goto done;
4587 		}
4588 		ret = mlx5_lag_query_cong_counters(dev->mdev,
4589 						   stats->value +
4590 						   port->cnts.num_q_counters,
4591 						   port->cnts.num_cong_counters,
4592 						   port->cnts.offsets +
4593 						   port->cnts.num_q_counters);
4594 
4595 		mlx5_ib_put_native_port_mdev(dev, port_num);
4596 		if (ret)
4597 			return ret;
4598 	}
4599 
4600 done:
4601 	return num_counters;
4602 }
4603 
4604 static void mlx5_ib_free_rdma_netdev(struct net_device *netdev)
4605 {
4606 	return mlx5_rdma_netdev_free(netdev);
4607 }
4608 
4609 static struct net_device*
4610 mlx5_ib_alloc_rdma_netdev(struct ib_device *hca,
4611 			  u8 port_num,
4612 			  enum rdma_netdev_t type,
4613 			  const char *name,
4614 			  unsigned char name_assign_type,
4615 			  void (*setup)(struct net_device *))
4616 {
4617 	struct net_device *netdev;
4618 	struct rdma_netdev *rn;
4619 
4620 	if (type != RDMA_NETDEV_IPOIB)
4621 		return ERR_PTR(-EOPNOTSUPP);
4622 
4623 	netdev = mlx5_rdma_netdev_alloc(to_mdev(hca)->mdev, hca,
4624 					name, setup);
4625 	if (likely(!IS_ERR_OR_NULL(netdev))) {
4626 		rn = netdev_priv(netdev);
4627 		rn->free_rdma_netdev = mlx5_ib_free_rdma_netdev;
4628 	}
4629 	return netdev;
4630 }
4631 
4632 static void delay_drop_debugfs_cleanup(struct mlx5_ib_dev *dev)
4633 {
4634 	if (!dev->delay_drop.dbg)
4635 		return;
4636 	debugfs_remove_recursive(dev->delay_drop.dbg->dir_debugfs);
4637 	kfree(dev->delay_drop.dbg);
4638 	dev->delay_drop.dbg = NULL;
4639 }
4640 
4641 static void cancel_delay_drop(struct mlx5_ib_dev *dev)
4642 {
4643 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4644 		return;
4645 
4646 	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4647 	delay_drop_debugfs_cleanup(dev);
4648 }
4649 
4650 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
4651 				       size_t count, loff_t *pos)
4652 {
4653 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
4654 	char lbuf[20];
4655 	int len;
4656 
4657 	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
4658 	return simple_read_from_buffer(buf, count, pos, lbuf, len);
4659 }
4660 
4661 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
4662 					size_t count, loff_t *pos)
4663 {
4664 	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
4665 	u32 timeout;
4666 	u32 var;
4667 
4668 	if (kstrtouint_from_user(buf, count, 0, &var))
4669 		return -EFAULT;
4670 
4671 	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
4672 			1000);
4673 	if (timeout != var)
4674 		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
4675 			    timeout);
4676 
4677 	delay_drop->timeout = timeout;
4678 
4679 	return count;
4680 }
4681 
4682 static const struct file_operations fops_delay_drop_timeout = {
4683 	.owner	= THIS_MODULE,
4684 	.open	= simple_open,
4685 	.write	= delay_drop_timeout_write,
4686 	.read	= delay_drop_timeout_read,
4687 };
4688 
4689 static int delay_drop_debugfs_init(struct mlx5_ib_dev *dev)
4690 {
4691 	struct mlx5_ib_dbg_delay_drop *dbg;
4692 
4693 	if (!mlx5_debugfs_root)
4694 		return 0;
4695 
4696 	dbg = kzalloc(sizeof(*dbg), GFP_KERNEL);
4697 	if (!dbg)
4698 		return -ENOMEM;
4699 
4700 	dev->delay_drop.dbg = dbg;
4701 
4702 	dbg->dir_debugfs =
4703 		debugfs_create_dir("delay_drop",
4704 				   dev->mdev->priv.dbg_root);
4705 	if (!dbg->dir_debugfs)
4706 		goto out_debugfs;
4707 
4708 	dbg->events_cnt_debugfs =
4709 		debugfs_create_atomic_t("num_timeout_events", 0400,
4710 					dbg->dir_debugfs,
4711 					&dev->delay_drop.events_cnt);
4712 	if (!dbg->events_cnt_debugfs)
4713 		goto out_debugfs;
4714 
4715 	dbg->rqs_cnt_debugfs =
4716 		debugfs_create_atomic_t("num_rqs", 0400,
4717 					dbg->dir_debugfs,
4718 					&dev->delay_drop.rqs_cnt);
4719 	if (!dbg->rqs_cnt_debugfs)
4720 		goto out_debugfs;
4721 
4722 	dbg->timeout_debugfs =
4723 		debugfs_create_file("timeout", 0600,
4724 				    dbg->dir_debugfs,
4725 				    &dev->delay_drop,
4726 				    &fops_delay_drop_timeout);
4727 	if (!dbg->timeout_debugfs)
4728 		goto out_debugfs;
4729 
4730 	return 0;
4731 
4732 out_debugfs:
4733 	delay_drop_debugfs_cleanup(dev);
4734 	return -ENOMEM;
4735 }
4736 
4737 static void init_delay_drop(struct mlx5_ib_dev *dev)
4738 {
4739 	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4740 		return;
4741 
4742 	mutex_init(&dev->delay_drop.lock);
4743 	dev->delay_drop.dev = dev;
4744 	dev->delay_drop.activate = false;
4745 	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4746 	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4747 	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4748 	atomic_set(&dev->delay_drop.events_cnt, 0);
4749 
4750 	if (delay_drop_debugfs_init(dev))
4751 		mlx5_ib_warn(dev, "Failed to init delay drop debugfs\n");
4752 }
4753 
4754 static const struct cpumask *
4755 mlx5_ib_get_vector_affinity(struct ib_device *ibdev, int comp_vector)
4756 {
4757 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4758 
4759 	return mlx5_get_vector_affinity_hint(dev->mdev, comp_vector);
4760 }
4761 
4762 /* The mlx5_ib_multiport_mutex should be held when calling this function */
4763 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
4764 				      struct mlx5_ib_multiport_info *mpi)
4765 {
4766 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
4767 	struct mlx5_ib_port *port = &ibdev->port[port_num];
4768 	int comps;
4769 	int err;
4770 	int i;
4771 
4772 	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
4773 
4774 	spin_lock(&port->mp.mpi_lock);
4775 	if (!mpi->ibdev) {
4776 		spin_unlock(&port->mp.mpi_lock);
4777 		return;
4778 	}
4779 	mpi->ibdev = NULL;
4780 
4781 	spin_unlock(&port->mp.mpi_lock);
4782 	mlx5_remove_netdev_notifier(ibdev, port_num);
4783 	spin_lock(&port->mp.mpi_lock);
4784 
4785 	comps = mpi->mdev_refcnt;
4786 	if (comps) {
4787 		mpi->unaffiliate = true;
4788 		init_completion(&mpi->unref_comp);
4789 		spin_unlock(&port->mp.mpi_lock);
4790 
4791 		for (i = 0; i < comps; i++)
4792 			wait_for_completion(&mpi->unref_comp);
4793 
4794 		spin_lock(&port->mp.mpi_lock);
4795 		mpi->unaffiliate = false;
4796 	}
4797 
4798 	port->mp.mpi = NULL;
4799 
4800 	list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4801 
4802 	spin_unlock(&port->mp.mpi_lock);
4803 
4804 	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
4805 
4806 	mlx5_ib_dbg(ibdev, "unaffiliated port %d\n", port_num + 1);
4807 	/* Log an error, still needed to cleanup the pointers and add
4808 	 * it back to the list.
4809 	 */
4810 	if (err)
4811 		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
4812 			    port_num + 1);
4813 
4814 	ibdev->roce[port_num].last_port_state = IB_PORT_DOWN;
4815 }
4816 
4817 /* The mlx5_ib_multiport_mutex should be held when calling this function */
4818 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
4819 				    struct mlx5_ib_multiport_info *mpi)
4820 {
4821 	u8 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
4822 	int err;
4823 
4824 	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
4825 	if (ibdev->port[port_num].mp.mpi) {
4826 		mlx5_ib_warn(ibdev, "port %d already affiliated.\n",
4827 			     port_num + 1);
4828 		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
4829 		return false;
4830 	}
4831 
4832 	ibdev->port[port_num].mp.mpi = mpi;
4833 	mpi->ibdev = ibdev;
4834 	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
4835 
4836 	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
4837 	if (err)
4838 		goto unbind;
4839 
4840 	err = get_port_caps(ibdev, mlx5_core_native_port_num(mpi->mdev));
4841 	if (err)
4842 		goto unbind;
4843 
4844 	err = mlx5_add_netdev_notifier(ibdev, port_num);
4845 	if (err) {
4846 		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
4847 			    port_num + 1);
4848 		goto unbind;
4849 	}
4850 
4851 	err = mlx5_ib_init_cong_debugfs(ibdev, port_num);
4852 	if (err)
4853 		goto unbind;
4854 
4855 	return true;
4856 
4857 unbind:
4858 	mlx5_ib_unbind_slave_port(ibdev, mpi);
4859 	return false;
4860 }
4861 
4862 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
4863 {
4864 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4865 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
4866 							  port_num + 1);
4867 	struct mlx5_ib_multiport_info *mpi;
4868 	int err;
4869 	int i;
4870 
4871 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
4872 		return 0;
4873 
4874 	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
4875 						     &dev->sys_image_guid);
4876 	if (err)
4877 		return err;
4878 
4879 	err = mlx5_nic_vport_enable_roce(dev->mdev);
4880 	if (err)
4881 		return err;
4882 
4883 	mutex_lock(&mlx5_ib_multiport_mutex);
4884 	for (i = 0; i < dev->num_ports; i++) {
4885 		bool bound = false;
4886 
4887 		/* build a stub multiport info struct for the native port. */
4888 		if (i == port_num) {
4889 			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4890 			if (!mpi) {
4891 				mutex_unlock(&mlx5_ib_multiport_mutex);
4892 				mlx5_nic_vport_disable_roce(dev->mdev);
4893 				return -ENOMEM;
4894 			}
4895 
4896 			mpi->is_master = true;
4897 			mpi->mdev = dev->mdev;
4898 			mpi->sys_image_guid = dev->sys_image_guid;
4899 			dev->port[i].mp.mpi = mpi;
4900 			mpi->ibdev = dev;
4901 			mpi = NULL;
4902 			continue;
4903 		}
4904 
4905 		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
4906 				    list) {
4907 			if (dev->sys_image_guid == mpi->sys_image_guid &&
4908 			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
4909 				bound = mlx5_ib_bind_slave_port(dev, mpi);
4910 			}
4911 
4912 			if (bound) {
4913 				dev_dbg(&mpi->mdev->pdev->dev, "removing port from unaffiliated list.\n");
4914 				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
4915 				list_del(&mpi->list);
4916 				break;
4917 			}
4918 		}
4919 		if (!bound) {
4920 			get_port_caps(dev, i + 1);
4921 			mlx5_ib_dbg(dev, "no free port found for port %d\n",
4922 				    i + 1);
4923 		}
4924 	}
4925 
4926 	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
4927 	mutex_unlock(&mlx5_ib_multiport_mutex);
4928 	return err;
4929 }
4930 
4931 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
4932 {
4933 	int port_num = mlx5_core_native_port_num(dev->mdev) - 1;
4934 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
4935 							  port_num + 1);
4936 	int i;
4937 
4938 	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
4939 		return;
4940 
4941 	mutex_lock(&mlx5_ib_multiport_mutex);
4942 	for (i = 0; i < dev->num_ports; i++) {
4943 		if (dev->port[i].mp.mpi) {
4944 			/* Destroy the native port stub */
4945 			if (i == port_num) {
4946 				kfree(dev->port[i].mp.mpi);
4947 				dev->port[i].mp.mpi = NULL;
4948 			} else {
4949 				mlx5_ib_dbg(dev, "unbinding port_num: %d\n", i + 1);
4950 				mlx5_ib_unbind_slave_port(dev, dev->port[i].mp.mpi);
4951 			}
4952 		}
4953 	}
4954 
4955 	mlx5_ib_dbg(dev, "removing from devlist\n");
4956 	list_del(&dev->ib_dev_list);
4957 	mutex_unlock(&mlx5_ib_multiport_mutex);
4958 
4959 	mlx5_nic_vport_disable_roce(dev->mdev);
4960 }
4961 
4962 ADD_UVERBS_ATTRIBUTES_SIMPLE(mlx5_ib_dm, UVERBS_OBJECT_DM,
4963 			     UVERBS_METHOD_DM_ALLOC,
4964 			     &UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_START_OFFSET,
4965 						  UVERBS_ATTR_TYPE(u64),
4966 						  UA_FLAGS(UVERBS_ATTR_SPEC_F_MANDATORY)),
4967 			     &UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_ALLOC_DM_RESP_PAGE_INDEX,
4968 						  UVERBS_ATTR_TYPE(u16),
4969 						  UA_FLAGS(UVERBS_ATTR_SPEC_F_MANDATORY)));
4970 
4971 ADD_UVERBS_ATTRIBUTES_SIMPLE(mlx5_ib_flow_action, UVERBS_OBJECT_FLOW_ACTION,
4972 			     UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
4973 			     &UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
4974 						 UVERBS_ATTR_TYPE(u64),
4975 						 UA_FLAGS(UVERBS_ATTR_SPEC_F_MANDATORY)));
4976 
4977 #define NUM_TREES	2
4978 static int populate_specs_root(struct mlx5_ib_dev *dev)
4979 {
4980 	const struct uverbs_object_tree_def *default_root[NUM_TREES + 1] = {
4981 		uverbs_default_get_objects()};
4982 	size_t num_trees = 1;
4983 
4984 	if (mlx5_accel_ipsec_device_caps(dev->mdev) & MLX5_ACCEL_IPSEC_CAP_DEVICE &&
4985 	    !WARN_ON(num_trees >= ARRAY_SIZE(default_root)))
4986 		default_root[num_trees++] = &mlx5_ib_flow_action;
4987 
4988 	if (MLX5_CAP_DEV_MEM(dev->mdev, memic) &&
4989 	    !WARN_ON(num_trees >= ARRAY_SIZE(default_root)))
4990 		default_root[num_trees++] = &mlx5_ib_dm;
4991 
4992 	dev->ib_dev.specs_root =
4993 		uverbs_alloc_spec_tree(num_trees, default_root);
4994 
4995 	return PTR_ERR_OR_ZERO(dev->ib_dev.specs_root);
4996 }
4997 
4998 static void depopulate_specs_root(struct mlx5_ib_dev *dev)
4999 {
5000 	uverbs_free_spec_tree(dev->ib_dev.specs_root);
5001 }
5002 
5003 void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
5004 {
5005 	mlx5_ib_cleanup_multiport_master(dev);
5006 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
5007 	cleanup_srcu_struct(&dev->mr_srcu);
5008 #endif
5009 	kfree(dev->port);
5010 }
5011 
5012 int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
5013 {
5014 	struct mlx5_core_dev *mdev = dev->mdev;
5015 	const char *name;
5016 	int err;
5017 	int i;
5018 
5019 	dev->port = kcalloc(dev->num_ports, sizeof(*dev->port),
5020 			    GFP_KERNEL);
5021 	if (!dev->port)
5022 		return -ENOMEM;
5023 
5024 	for (i = 0; i < dev->num_ports; i++) {
5025 		spin_lock_init(&dev->port[i].mp.mpi_lock);
5026 		rwlock_init(&dev->roce[i].netdev_lock);
5027 	}
5028 
5029 	err = mlx5_ib_init_multiport_master(dev);
5030 	if (err)
5031 		goto err_free_port;
5032 
5033 	if (!mlx5_core_mp_enabled(mdev)) {
5034 		for (i = 1; i <= dev->num_ports; i++) {
5035 			err = get_port_caps(dev, i);
5036 			if (err)
5037 				break;
5038 		}
5039 	} else {
5040 		err = get_port_caps(dev, mlx5_core_native_port_num(mdev));
5041 	}
5042 	if (err)
5043 		goto err_mp;
5044 
5045 	if (mlx5_use_mad_ifc(dev))
5046 		get_ext_port_caps(dev);
5047 
5048 	if (!mlx5_lag_is_active(mdev))
5049 		name = "mlx5_%d";
5050 	else
5051 		name = "mlx5_bond_%d";
5052 
5053 	strlcpy(dev->ib_dev.name, name, IB_DEVICE_NAME_MAX);
5054 	dev->ib_dev.owner		= THIS_MODULE;
5055 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
5056 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
5057 	dev->ib_dev.phys_port_cnt	= dev->num_ports;
5058 	dev->ib_dev.num_comp_vectors    =
5059 		dev->mdev->priv.eq_table.num_comp_vectors;
5060 	dev->ib_dev.dev.parent		= &mdev->pdev->dev;
5061 
5062 	mutex_init(&dev->cap_mask_mutex);
5063 	INIT_LIST_HEAD(&dev->qp_list);
5064 	spin_lock_init(&dev->reset_flow_resource_lock);
5065 
5066 	spin_lock_init(&dev->memic.memic_lock);
5067 	dev->memic.dev = mdev;
5068 
5069 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
5070 	err = init_srcu_struct(&dev->mr_srcu);
5071 	if (err)
5072 		goto err_free_port;
5073 #endif
5074 
5075 	return 0;
5076 err_mp:
5077 	mlx5_ib_cleanup_multiport_master(dev);
5078 
5079 err_free_port:
5080 	kfree(dev->port);
5081 
5082 	return -ENOMEM;
5083 }
5084 
5085 static int mlx5_ib_stage_flow_db_init(struct mlx5_ib_dev *dev)
5086 {
5087 	dev->flow_db = kzalloc(sizeof(*dev->flow_db), GFP_KERNEL);
5088 
5089 	if (!dev->flow_db)
5090 		return -ENOMEM;
5091 
5092 	mutex_init(&dev->flow_db->lock);
5093 
5094 	return 0;
5095 }
5096 
5097 int mlx5_ib_stage_rep_flow_db_init(struct mlx5_ib_dev *dev)
5098 {
5099 	struct mlx5_ib_dev *nic_dev;
5100 
5101 	nic_dev = mlx5_ib_get_uplink_ibdev(dev->mdev->priv.eswitch);
5102 
5103 	if (!nic_dev)
5104 		return -EINVAL;
5105 
5106 	dev->flow_db = nic_dev->flow_db;
5107 
5108 	return 0;
5109 }
5110 
5111 static void mlx5_ib_stage_flow_db_cleanup(struct mlx5_ib_dev *dev)
5112 {
5113 	kfree(dev->flow_db);
5114 }
5115 
5116 int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
5117 {
5118 	struct mlx5_core_dev *mdev = dev->mdev;
5119 	int err;
5120 
5121 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
5122 	dev->ib_dev.uverbs_cmd_mask	=
5123 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
5124 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
5125 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
5126 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
5127 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
5128 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
5129 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
5130 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
5131 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
5132 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
5133 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
5134 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
5135 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
5136 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
5137 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
5138 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
5139 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
5140 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
5141 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
5142 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
5143 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
5144 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
5145 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
5146 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
5147 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
5148 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
5149 	dev->ib_dev.uverbs_ex_cmd_mask =
5150 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
5151 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
5152 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP)	|
5153 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP)	|
5154 		(1ull << IB_USER_VERBS_EX_CMD_MODIFY_CQ);
5155 
5156 	dev->ib_dev.query_device	= mlx5_ib_query_device;
5157 	dev->ib_dev.get_link_layer	= mlx5_ib_port_link_layer;
5158 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
5159 	dev->ib_dev.add_gid		= mlx5_ib_add_gid;
5160 	dev->ib_dev.del_gid		= mlx5_ib_del_gid;
5161 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
5162 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
5163 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
5164 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
5165 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
5166 	dev->ib_dev.mmap		= mlx5_ib_mmap;
5167 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
5168 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
5169 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
5170 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
5171 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
5172 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
5173 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
5174 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
5175 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
5176 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
5177 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
5178 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
5179 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
5180 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
5181 	dev->ib_dev.post_send		= mlx5_ib_post_send;
5182 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
5183 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
5184 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
5185 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
5186 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
5187 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
5188 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
5189 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
5190 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
5191 	dev->ib_dev.rereg_user_mr	= mlx5_ib_rereg_user_mr;
5192 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
5193 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
5194 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
5195 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
5196 	dev->ib_dev.alloc_mr		= mlx5_ib_alloc_mr;
5197 	dev->ib_dev.map_mr_sg		= mlx5_ib_map_mr_sg;
5198 	dev->ib_dev.check_mr_status	= mlx5_ib_check_mr_status;
5199 	dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
5200 	dev->ib_dev.get_vector_affinity	= mlx5_ib_get_vector_affinity;
5201 	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads))
5202 		dev->ib_dev.alloc_rdma_netdev	= mlx5_ib_alloc_rdma_netdev;
5203 
5204 	if (mlx5_core_is_pf(mdev)) {
5205 		dev->ib_dev.get_vf_config	= mlx5_ib_get_vf_config;
5206 		dev->ib_dev.set_vf_link_state	= mlx5_ib_set_vf_link_state;
5207 		dev->ib_dev.get_vf_stats	= mlx5_ib_get_vf_stats;
5208 		dev->ib_dev.set_vf_guid		= mlx5_ib_set_vf_guid;
5209 	}
5210 
5211 	dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
5212 
5213 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
5214 
5215 	if (MLX5_CAP_GEN(mdev, imaicl)) {
5216 		dev->ib_dev.alloc_mw		= mlx5_ib_alloc_mw;
5217 		dev->ib_dev.dealloc_mw		= mlx5_ib_dealloc_mw;
5218 		dev->ib_dev.uverbs_cmd_mask |=
5219 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
5220 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
5221 	}
5222 
5223 	if (MLX5_CAP_GEN(mdev, xrc)) {
5224 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
5225 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
5226 		dev->ib_dev.uverbs_cmd_mask |=
5227 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
5228 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
5229 	}
5230 
5231 	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
5232 		dev->ib_dev.alloc_dm = mlx5_ib_alloc_dm;
5233 		dev->ib_dev.dealloc_dm = mlx5_ib_dealloc_dm;
5234 		dev->ib_dev.reg_dm_mr = mlx5_ib_reg_dm_mr;
5235 	}
5236 
5237 	dev->ib_dev.create_flow	= mlx5_ib_create_flow;
5238 	dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
5239 	dev->ib_dev.uverbs_ex_cmd_mask |=
5240 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
5241 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW);
5242 	dev->ib_dev.create_flow_action_esp = mlx5_ib_create_flow_action_esp;
5243 	dev->ib_dev.destroy_flow_action = mlx5_ib_destroy_flow_action;
5244 	dev->ib_dev.modify_flow_action_esp = mlx5_ib_modify_flow_action_esp;
5245 	dev->ib_dev.driver_id = RDMA_DRIVER_MLX5;
5246 
5247 	err = init_node_data(dev);
5248 	if (err)
5249 		return err;
5250 
5251 	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
5252 	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
5253 	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
5254 		mutex_init(&dev->lb_mutex);
5255 
5256 	return 0;
5257 }
5258 
5259 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
5260 {
5261 	dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
5262 	dev->ib_dev.query_port		= mlx5_ib_query_port;
5263 
5264 	return 0;
5265 }
5266 
5267 int mlx5_ib_stage_rep_non_default_cb(struct mlx5_ib_dev *dev)
5268 {
5269 	dev->ib_dev.get_port_immutable  = mlx5_port_rep_immutable;
5270 	dev->ib_dev.query_port		= mlx5_ib_rep_query_port;
5271 
5272 	return 0;
5273 }
5274 
5275 static int mlx5_ib_stage_common_roce_init(struct mlx5_ib_dev *dev,
5276 					  u8 port_num)
5277 {
5278 	int i;
5279 
5280 	for (i = 0; i < dev->num_ports; i++) {
5281 		dev->roce[i].dev = dev;
5282 		dev->roce[i].native_port_num = i + 1;
5283 		dev->roce[i].last_port_state = IB_PORT_DOWN;
5284 	}
5285 
5286 	dev->ib_dev.get_netdev	= mlx5_ib_get_netdev;
5287 	dev->ib_dev.create_wq	 = mlx5_ib_create_wq;
5288 	dev->ib_dev.modify_wq	 = mlx5_ib_modify_wq;
5289 	dev->ib_dev.destroy_wq	 = mlx5_ib_destroy_wq;
5290 	dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
5291 	dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
5292 
5293 	dev->ib_dev.uverbs_ex_cmd_mask |=
5294 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
5295 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
5296 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
5297 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
5298 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
5299 
5300 	return mlx5_add_netdev_notifier(dev, port_num);
5301 }
5302 
5303 static void mlx5_ib_stage_common_roce_cleanup(struct mlx5_ib_dev *dev)
5304 {
5305 	u8 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5306 
5307 	mlx5_remove_netdev_notifier(dev, port_num);
5308 }
5309 
5310 int mlx5_ib_stage_rep_roce_init(struct mlx5_ib_dev *dev)
5311 {
5312 	struct mlx5_core_dev *mdev = dev->mdev;
5313 	enum rdma_link_layer ll;
5314 	int port_type_cap;
5315 	int err = 0;
5316 	u8 port_num;
5317 
5318 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5319 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5320 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5321 
5322 	if (ll == IB_LINK_LAYER_ETHERNET)
5323 		err = mlx5_ib_stage_common_roce_init(dev, port_num);
5324 
5325 	return err;
5326 }
5327 
5328 void mlx5_ib_stage_rep_roce_cleanup(struct mlx5_ib_dev *dev)
5329 {
5330 	mlx5_ib_stage_common_roce_cleanup(dev);
5331 }
5332 
5333 static int mlx5_ib_stage_roce_init(struct mlx5_ib_dev *dev)
5334 {
5335 	struct mlx5_core_dev *mdev = dev->mdev;
5336 	enum rdma_link_layer ll;
5337 	int port_type_cap;
5338 	u8 port_num;
5339 	int err;
5340 
5341 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5342 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5343 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5344 
5345 	if (ll == IB_LINK_LAYER_ETHERNET) {
5346 		err = mlx5_ib_stage_common_roce_init(dev, port_num);
5347 		if (err)
5348 			return err;
5349 
5350 		err = mlx5_enable_eth(dev, port_num);
5351 		if (err)
5352 			goto cleanup;
5353 	}
5354 
5355 	return 0;
5356 cleanup:
5357 	mlx5_ib_stage_common_roce_cleanup(dev);
5358 
5359 	return err;
5360 }
5361 
5362 static void mlx5_ib_stage_roce_cleanup(struct mlx5_ib_dev *dev)
5363 {
5364 	struct mlx5_core_dev *mdev = dev->mdev;
5365 	enum rdma_link_layer ll;
5366 	int port_type_cap;
5367 	u8 port_num;
5368 
5369 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
5370 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5371 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5372 
5373 	if (ll == IB_LINK_LAYER_ETHERNET) {
5374 		mlx5_disable_eth(dev);
5375 		mlx5_ib_stage_common_roce_cleanup(dev);
5376 	}
5377 }
5378 
5379 int mlx5_ib_stage_dev_res_init(struct mlx5_ib_dev *dev)
5380 {
5381 	return create_dev_resources(&dev->devr);
5382 }
5383 
5384 void mlx5_ib_stage_dev_res_cleanup(struct mlx5_ib_dev *dev)
5385 {
5386 	destroy_dev_resources(&dev->devr);
5387 }
5388 
5389 static int mlx5_ib_stage_odp_init(struct mlx5_ib_dev *dev)
5390 {
5391 	mlx5_ib_internal_fill_odp_caps(dev);
5392 
5393 	return mlx5_ib_odp_init_one(dev);
5394 }
5395 
5396 int mlx5_ib_stage_counters_init(struct mlx5_ib_dev *dev)
5397 {
5398 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
5399 		dev->ib_dev.get_hw_stats	= mlx5_ib_get_hw_stats;
5400 		dev->ib_dev.alloc_hw_stats	= mlx5_ib_alloc_hw_stats;
5401 
5402 		return mlx5_ib_alloc_counters(dev);
5403 	}
5404 
5405 	return 0;
5406 }
5407 
5408 void mlx5_ib_stage_counters_cleanup(struct mlx5_ib_dev *dev)
5409 {
5410 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
5411 		mlx5_ib_dealloc_counters(dev);
5412 }
5413 
5414 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
5415 {
5416 	return mlx5_ib_init_cong_debugfs(dev,
5417 					 mlx5_core_native_port_num(dev->mdev) - 1);
5418 }
5419 
5420 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
5421 {
5422 	mlx5_ib_cleanup_cong_debugfs(dev,
5423 				     mlx5_core_native_port_num(dev->mdev) - 1);
5424 }
5425 
5426 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
5427 {
5428 	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
5429 	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
5430 }
5431 
5432 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
5433 {
5434 	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
5435 }
5436 
5437 int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
5438 {
5439 	int err;
5440 
5441 	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
5442 	if (err)
5443 		return err;
5444 
5445 	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
5446 	if (err)
5447 		mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
5448 
5449 	return err;
5450 }
5451 
5452 void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
5453 {
5454 	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
5455 	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
5456 }
5457 
5458 static int mlx5_ib_stage_populate_specs(struct mlx5_ib_dev *dev)
5459 {
5460 	return populate_specs_root(dev);
5461 }
5462 
5463 int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
5464 {
5465 	return ib_register_device(&dev->ib_dev, NULL);
5466 }
5467 
5468 static void mlx5_ib_stage_depopulate_specs(struct mlx5_ib_dev *dev)
5469 {
5470 	depopulate_specs_root(dev);
5471 }
5472 
5473 void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
5474 {
5475 	destroy_umrc_res(dev);
5476 }
5477 
5478 void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
5479 {
5480 	ib_unregister_device(&dev->ib_dev);
5481 }
5482 
5483 int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
5484 {
5485 	return create_umr_res(dev);
5486 }
5487 
5488 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
5489 {
5490 	init_delay_drop(dev);
5491 
5492 	return 0;
5493 }
5494 
5495 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
5496 {
5497 	cancel_delay_drop(dev);
5498 }
5499 
5500 int mlx5_ib_stage_class_attr_init(struct mlx5_ib_dev *dev)
5501 {
5502 	int err;
5503 	int i;
5504 
5505 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
5506 		err = device_create_file(&dev->ib_dev.dev,
5507 					 mlx5_class_attributes[i]);
5508 		if (err)
5509 			return err;
5510 	}
5511 
5512 	return 0;
5513 }
5514 
5515 static int mlx5_ib_stage_rep_reg_init(struct mlx5_ib_dev *dev)
5516 {
5517 	mlx5_ib_register_vport_reps(dev);
5518 
5519 	return 0;
5520 }
5521 
5522 static void mlx5_ib_stage_rep_reg_cleanup(struct mlx5_ib_dev *dev)
5523 {
5524 	mlx5_ib_unregister_vport_reps(dev);
5525 }
5526 
5527 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
5528 		      const struct mlx5_ib_profile *profile,
5529 		      int stage)
5530 {
5531 	/* Number of stages to cleanup */
5532 	while (stage) {
5533 		stage--;
5534 		if (profile->stage[stage].cleanup)
5535 			profile->stage[stage].cleanup(dev);
5536 	}
5537 
5538 	ib_dealloc_device((struct ib_device *)dev);
5539 }
5540 
5541 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev, u8 port_num);
5542 
5543 void *__mlx5_ib_add(struct mlx5_ib_dev *dev,
5544 		    const struct mlx5_ib_profile *profile)
5545 {
5546 	int err;
5547 	int i;
5548 
5549 	printk_once(KERN_INFO "%s", mlx5_version);
5550 
5551 	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
5552 		if (profile->stage[i].init) {
5553 			err = profile->stage[i].init(dev);
5554 			if (err)
5555 				goto err_out;
5556 		}
5557 	}
5558 
5559 	dev->profile = profile;
5560 	dev->ib_active = true;
5561 
5562 	return dev;
5563 
5564 err_out:
5565 	__mlx5_ib_remove(dev, profile, i);
5566 
5567 	return NULL;
5568 }
5569 
5570 static const struct mlx5_ib_profile pf_profile = {
5571 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
5572 		     mlx5_ib_stage_init_init,
5573 		     mlx5_ib_stage_init_cleanup),
5574 	STAGE_CREATE(MLX5_IB_STAGE_FLOW_DB,
5575 		     mlx5_ib_stage_flow_db_init,
5576 		     mlx5_ib_stage_flow_db_cleanup),
5577 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
5578 		     mlx5_ib_stage_caps_init,
5579 		     NULL),
5580 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
5581 		     mlx5_ib_stage_non_default_cb,
5582 		     NULL),
5583 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
5584 		     mlx5_ib_stage_roce_init,
5585 		     mlx5_ib_stage_roce_cleanup),
5586 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
5587 		     mlx5_ib_stage_dev_res_init,
5588 		     mlx5_ib_stage_dev_res_cleanup),
5589 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
5590 		     mlx5_ib_stage_odp_init,
5591 		     NULL),
5592 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
5593 		     mlx5_ib_stage_counters_init,
5594 		     mlx5_ib_stage_counters_cleanup),
5595 	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
5596 		     mlx5_ib_stage_cong_debugfs_init,
5597 		     mlx5_ib_stage_cong_debugfs_cleanup),
5598 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
5599 		     mlx5_ib_stage_uar_init,
5600 		     mlx5_ib_stage_uar_cleanup),
5601 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
5602 		     mlx5_ib_stage_bfrag_init,
5603 		     mlx5_ib_stage_bfrag_cleanup),
5604 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
5605 		     NULL,
5606 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
5607 	STAGE_CREATE(MLX5_IB_STAGE_SPECS,
5608 		     mlx5_ib_stage_populate_specs,
5609 		     mlx5_ib_stage_depopulate_specs),
5610 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
5611 		     mlx5_ib_stage_ib_reg_init,
5612 		     mlx5_ib_stage_ib_reg_cleanup),
5613 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
5614 		     mlx5_ib_stage_post_ib_reg_umr_init,
5615 		     NULL),
5616 	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
5617 		     mlx5_ib_stage_delay_drop_init,
5618 		     mlx5_ib_stage_delay_drop_cleanup),
5619 	STAGE_CREATE(MLX5_IB_STAGE_CLASS_ATTR,
5620 		     mlx5_ib_stage_class_attr_init,
5621 		     NULL),
5622 };
5623 
5624 static const struct mlx5_ib_profile nic_rep_profile = {
5625 	STAGE_CREATE(MLX5_IB_STAGE_INIT,
5626 		     mlx5_ib_stage_init_init,
5627 		     mlx5_ib_stage_init_cleanup),
5628 	STAGE_CREATE(MLX5_IB_STAGE_FLOW_DB,
5629 		     mlx5_ib_stage_flow_db_init,
5630 		     mlx5_ib_stage_flow_db_cleanup),
5631 	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
5632 		     mlx5_ib_stage_caps_init,
5633 		     NULL),
5634 	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
5635 		     mlx5_ib_stage_rep_non_default_cb,
5636 		     NULL),
5637 	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
5638 		     mlx5_ib_stage_rep_roce_init,
5639 		     mlx5_ib_stage_rep_roce_cleanup),
5640 	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
5641 		     mlx5_ib_stage_dev_res_init,
5642 		     mlx5_ib_stage_dev_res_cleanup),
5643 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
5644 		     mlx5_ib_stage_counters_init,
5645 		     mlx5_ib_stage_counters_cleanup),
5646 	STAGE_CREATE(MLX5_IB_STAGE_UAR,
5647 		     mlx5_ib_stage_uar_init,
5648 		     mlx5_ib_stage_uar_cleanup),
5649 	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
5650 		     mlx5_ib_stage_bfrag_init,
5651 		     mlx5_ib_stage_bfrag_cleanup),
5652 	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
5653 		     NULL,
5654 		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
5655 	STAGE_CREATE(MLX5_IB_STAGE_SPECS,
5656 		     mlx5_ib_stage_populate_specs,
5657 		     mlx5_ib_stage_depopulate_specs),
5658 	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
5659 		     mlx5_ib_stage_ib_reg_init,
5660 		     mlx5_ib_stage_ib_reg_cleanup),
5661 	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
5662 		     mlx5_ib_stage_post_ib_reg_umr_init,
5663 		     NULL),
5664 	STAGE_CREATE(MLX5_IB_STAGE_CLASS_ATTR,
5665 		     mlx5_ib_stage_class_attr_init,
5666 		     NULL),
5667 	STAGE_CREATE(MLX5_IB_STAGE_REP_REG,
5668 		     mlx5_ib_stage_rep_reg_init,
5669 		     mlx5_ib_stage_rep_reg_cleanup),
5670 };
5671 
5672 static void *mlx5_ib_add_slave_port(struct mlx5_core_dev *mdev, u8 port_num)
5673 {
5674 	struct mlx5_ib_multiport_info *mpi;
5675 	struct mlx5_ib_dev *dev;
5676 	bool bound = false;
5677 	int err;
5678 
5679 	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
5680 	if (!mpi)
5681 		return NULL;
5682 
5683 	mpi->mdev = mdev;
5684 
5685 	err = mlx5_query_nic_vport_system_image_guid(mdev,
5686 						     &mpi->sys_image_guid);
5687 	if (err) {
5688 		kfree(mpi);
5689 		return NULL;
5690 	}
5691 
5692 	mutex_lock(&mlx5_ib_multiport_mutex);
5693 	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
5694 		if (dev->sys_image_guid == mpi->sys_image_guid)
5695 			bound = mlx5_ib_bind_slave_port(dev, mpi);
5696 
5697 		if (bound) {
5698 			rdma_roce_rescan_device(&dev->ib_dev);
5699 			break;
5700 		}
5701 	}
5702 
5703 	if (!bound) {
5704 		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
5705 		dev_dbg(&mdev->pdev->dev, "no suitable IB device found to bind to, added to unaffiliated list.\n");
5706 	} else {
5707 		mlx5_ib_dbg(dev, "bound port %u\n", port_num + 1);
5708 	}
5709 	mutex_unlock(&mlx5_ib_multiport_mutex);
5710 
5711 	return mpi;
5712 }
5713 
5714 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
5715 {
5716 	enum rdma_link_layer ll;
5717 	struct mlx5_ib_dev *dev;
5718 	int port_type_cap;
5719 
5720 	printk_once(KERN_INFO "%s", mlx5_version);
5721 
5722 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
5723 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
5724 
5725 	if (mlx5_core_is_mp_slave(mdev) && ll == IB_LINK_LAYER_ETHERNET) {
5726 		u8 port_num = mlx5_core_native_port_num(mdev) - 1;
5727 
5728 		return mlx5_ib_add_slave_port(mdev, port_num);
5729 	}
5730 
5731 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
5732 	if (!dev)
5733 		return NULL;
5734 
5735 	dev->mdev = mdev;
5736 	dev->num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
5737 			     MLX5_CAP_GEN(mdev, num_vhca_ports));
5738 
5739 	if (MLX5_VPORT_MANAGER(mdev) &&
5740 	    mlx5_ib_eswitch_mode(mdev->priv.eswitch) == SRIOV_OFFLOADS) {
5741 		dev->rep = mlx5_ib_vport_rep(mdev->priv.eswitch, 0);
5742 
5743 		return __mlx5_ib_add(dev, &nic_rep_profile);
5744 	}
5745 
5746 	return __mlx5_ib_add(dev, &pf_profile);
5747 }
5748 
5749 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
5750 {
5751 	struct mlx5_ib_multiport_info *mpi;
5752 	struct mlx5_ib_dev *dev;
5753 
5754 	if (mlx5_core_is_mp_slave(mdev)) {
5755 		mpi = context;
5756 		mutex_lock(&mlx5_ib_multiport_mutex);
5757 		if (mpi->ibdev)
5758 			mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
5759 		list_del(&mpi->list);
5760 		mutex_unlock(&mlx5_ib_multiport_mutex);
5761 		return;
5762 	}
5763 
5764 	dev = context;
5765 	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
5766 }
5767 
5768 static struct mlx5_interface mlx5_ib_interface = {
5769 	.add            = mlx5_ib_add,
5770 	.remove         = mlx5_ib_remove,
5771 	.event          = mlx5_ib_event,
5772 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
5773 	.pfault		= mlx5_ib_pfault,
5774 #endif
5775 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
5776 };
5777 
5778 unsigned long mlx5_ib_get_xlt_emergency_page(void)
5779 {
5780 	mutex_lock(&xlt_emergency_page_mutex);
5781 	return xlt_emergency_page;
5782 }
5783 
5784 void mlx5_ib_put_xlt_emergency_page(void)
5785 {
5786 	mutex_unlock(&xlt_emergency_page_mutex);
5787 }
5788 
5789 static int __init mlx5_ib_init(void)
5790 {
5791 	int err;
5792 
5793 	xlt_emergency_page = __get_free_page(GFP_KERNEL);
5794 	if (!xlt_emergency_page)
5795 		return -ENOMEM;
5796 
5797 	mutex_init(&xlt_emergency_page_mutex);
5798 
5799 	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
5800 	if (!mlx5_ib_event_wq) {
5801 		free_page(xlt_emergency_page);
5802 		return -ENOMEM;
5803 	}
5804 
5805 	mlx5_ib_odp_init();
5806 
5807 	err = mlx5_register_interface(&mlx5_ib_interface);
5808 
5809 	return err;
5810 }
5811 
5812 static void __exit mlx5_ib_cleanup(void)
5813 {
5814 	mlx5_unregister_interface(&mlx5_ib_interface);
5815 	destroy_workqueue(mlx5_ib_event_wq);
5816 	mutex_destroy(&xlt_emergency_page_mutex);
5817 	free_page(xlt_emergency_page);
5818 }
5819 
5820 module_init(mlx5_ib_init);
5821 module_exit(mlx5_ib_cleanup);
5822