• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/highmem.h>
34 #include <linux/module.h>
35 #include <linux/init.h>
36 #include <linux/errno.h>
37 #include <linux/pci.h>
38 #include <linux/dma-mapping.h>
39 #include <linux/slab.h>
40 #if defined(CONFIG_X86)
41 #include <asm/pat.h>
42 #endif
43 #include <linux/sched.h>
44 #include <linux/delay.h>
45 #include <rdma/ib_user_verbs.h>
46 #include <rdma/ib_addr.h>
47 #include <rdma/ib_cache.h>
48 #include <linux/mlx5/port.h>
49 #include <linux/mlx5/vport.h>
50 #include <linux/list.h>
51 #include <rdma/ib_smi.h>
52 #include <rdma/ib_umem.h>
53 #include <linux/in.h>
54 #include <linux/etherdevice.h>
55 #include <linux/mlx5/fs.h>
56 #include "mlx5_ib.h"
57 
58 #define DRIVER_NAME "mlx5_ib"
59 #define DRIVER_VERSION "2.2-1"
60 #define DRIVER_RELDATE	"Feb 2014"
61 
62 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
63 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
64 MODULE_LICENSE("Dual BSD/GPL");
65 MODULE_VERSION(DRIVER_VERSION);
66 
67 static int deprecated_prof_sel = 2;
68 module_param_named(prof_sel, deprecated_prof_sel, int, 0444);
69 MODULE_PARM_DESC(prof_sel, "profile selector. Deprecated here. Moved to module mlx5_core");
70 
71 static char mlx5_version[] =
72 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
73 	DRIVER_VERSION " (" DRIVER_RELDATE ")\n";
74 
75 enum {
76 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
77 };
78 
79 static enum rdma_link_layer
mlx5_port_type_cap_to_rdma_ll(int port_type_cap)80 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
81 {
82 	switch (port_type_cap) {
83 	case MLX5_CAP_PORT_TYPE_IB:
84 		return IB_LINK_LAYER_INFINIBAND;
85 	case MLX5_CAP_PORT_TYPE_ETH:
86 		return IB_LINK_LAYER_ETHERNET;
87 	default:
88 		return IB_LINK_LAYER_UNSPECIFIED;
89 	}
90 }
91 
92 static enum rdma_link_layer
mlx5_ib_port_link_layer(struct ib_device * device,u8 port_num)93 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
94 {
95 	struct mlx5_ib_dev *dev = to_mdev(device);
96 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
97 
98 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
99 }
100 
mlx5_netdev_event(struct notifier_block * this,unsigned long event,void * ptr)101 static int mlx5_netdev_event(struct notifier_block *this,
102 			     unsigned long event, void *ptr)
103 {
104 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
105 	struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
106 						 roce.nb);
107 
108 	switch (event) {
109 	case NETDEV_REGISTER:
110 	case NETDEV_UNREGISTER:
111 		write_lock(&ibdev->roce.netdev_lock);
112 		if (ndev->dev.parent == &ibdev->mdev->pdev->dev)
113 			ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ?
114 					     NULL : ndev;
115 		write_unlock(&ibdev->roce.netdev_lock);
116 		break;
117 
118 	case NETDEV_UP:
119 	case NETDEV_DOWN: {
120 		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(ibdev->mdev);
121 		struct net_device *upper = NULL;
122 
123 		if (lag_ndev) {
124 			upper = netdev_master_upper_dev_get(lag_ndev);
125 			dev_put(lag_ndev);
126 		}
127 
128 		if ((upper == ndev || (!upper && ndev == ibdev->roce.netdev))
129 		    && ibdev->ib_active) {
130 			struct ib_event ibev = {0};
131 
132 			ibev.device = &ibdev->ib_dev;
133 			ibev.event = (event == NETDEV_UP) ?
134 				     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
135 			ibev.element.port_num = 1;
136 			ib_dispatch_event(&ibev);
137 		}
138 		break;
139 	}
140 
141 	default:
142 		break;
143 	}
144 
145 	return NOTIFY_DONE;
146 }
147 
mlx5_ib_get_netdev(struct ib_device * device,u8 port_num)148 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
149 					     u8 port_num)
150 {
151 	struct mlx5_ib_dev *ibdev = to_mdev(device);
152 	struct net_device *ndev;
153 
154 	ndev = mlx5_lag_get_roce_netdev(ibdev->mdev);
155 	if (ndev)
156 		return ndev;
157 
158 	/* Ensure ndev does not disappear before we invoke dev_hold()
159 	 */
160 	read_lock(&ibdev->roce.netdev_lock);
161 	ndev = ibdev->roce.netdev;
162 	if (ndev)
163 		dev_hold(ndev);
164 	read_unlock(&ibdev->roce.netdev_lock);
165 
166 	return ndev;
167 }
168 
mlx5_query_port_roce(struct ib_device * device,u8 port_num,struct ib_port_attr * props)169 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
170 				struct ib_port_attr *props)
171 {
172 	struct mlx5_ib_dev *dev = to_mdev(device);
173 	struct net_device *ndev, *upper;
174 	enum ib_mtu ndev_ib_mtu;
175 	u16 qkey_viol_cntr;
176 
177 	memset(props, 0, sizeof(*props));
178 
179 	props->port_cap_flags  |= IB_PORT_CM_SUP;
180 	props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
181 
182 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
183 						roce_address_table_size);
184 	props->max_mtu          = IB_MTU_4096;
185 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
186 	props->pkey_tbl_len     = 1;
187 	props->state            = IB_PORT_DOWN;
188 	props->phys_state       = 3;
189 
190 	mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr);
191 	props->qkey_viol_cntr = qkey_viol_cntr;
192 
193 	ndev = mlx5_ib_get_netdev(device, port_num);
194 	if (!ndev)
195 		return 0;
196 
197 	if (mlx5_lag_is_active(dev->mdev)) {
198 		rcu_read_lock();
199 		upper = netdev_master_upper_dev_get_rcu(ndev);
200 		if (upper) {
201 			dev_put(ndev);
202 			ndev = upper;
203 			dev_hold(ndev);
204 		}
205 		rcu_read_unlock();
206 	}
207 
208 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
209 		props->state      = IB_PORT_ACTIVE;
210 		props->phys_state = 5;
211 	}
212 
213 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
214 
215 	dev_put(ndev);
216 
217 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
218 
219 	props->active_width	= IB_WIDTH_4X;  /* TODO */
220 	props->active_speed	= IB_SPEED_QDR; /* TODO */
221 
222 	return 0;
223 }
224 
ib_gid_to_mlx5_roce_addr(const union ib_gid * gid,const struct ib_gid_attr * attr,void * mlx5_addr)225 static void ib_gid_to_mlx5_roce_addr(const union ib_gid *gid,
226 				     const struct ib_gid_attr *attr,
227 				     void *mlx5_addr)
228 {
229 #define MLX5_SET_RA(p, f, v) MLX5_SET(roce_addr_layout, p, f, v)
230 	char *mlx5_addr_l3_addr	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
231 					       source_l3_address);
232 	void *mlx5_addr_mac	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
233 					       source_mac_47_32);
234 
235 	if (!gid)
236 		return;
237 
238 	ether_addr_copy(mlx5_addr_mac, attr->ndev->dev_addr);
239 
240 	if (is_vlan_dev(attr->ndev)) {
241 		MLX5_SET_RA(mlx5_addr, vlan_valid, 1);
242 		MLX5_SET_RA(mlx5_addr, vlan_id, vlan_dev_vlan_id(attr->ndev));
243 	}
244 
245 	switch (attr->gid_type) {
246 	case IB_GID_TYPE_IB:
247 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_1);
248 		break;
249 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
250 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_2);
251 		break;
252 
253 	default:
254 		WARN_ON(true);
255 	}
256 
257 	if (attr->gid_type != IB_GID_TYPE_IB) {
258 		if (ipv6_addr_v4mapped((void *)gid))
259 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
260 				    MLX5_ROCE_L3_TYPE_IPV4);
261 		else
262 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
263 				    MLX5_ROCE_L3_TYPE_IPV6);
264 	}
265 
266 	if ((attr->gid_type == IB_GID_TYPE_IB) ||
267 	    !ipv6_addr_v4mapped((void *)gid))
268 		memcpy(mlx5_addr_l3_addr, gid, sizeof(*gid));
269 	else
270 		memcpy(&mlx5_addr_l3_addr[12], &gid->raw[12], 4);
271 }
272 
set_roce_addr(struct ib_device * device,u8 port_num,unsigned int index,const union ib_gid * gid,const struct ib_gid_attr * attr)273 static int set_roce_addr(struct ib_device *device, u8 port_num,
274 			 unsigned int index,
275 			 const union ib_gid *gid,
276 			 const struct ib_gid_attr *attr)
277 {
278 	struct mlx5_ib_dev *dev = to_mdev(device);
279 	u32  in[MLX5_ST_SZ_DW(set_roce_address_in)]  = {0};
280 	u32 out[MLX5_ST_SZ_DW(set_roce_address_out)] = {0};
281 	void *in_addr = MLX5_ADDR_OF(set_roce_address_in, in, roce_address);
282 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(device, port_num);
283 
284 	if (ll != IB_LINK_LAYER_ETHERNET)
285 		return -EINVAL;
286 
287 	ib_gid_to_mlx5_roce_addr(gid, attr, in_addr);
288 
289 	MLX5_SET(set_roce_address_in, in, roce_address_index, index);
290 	MLX5_SET(set_roce_address_in, in, opcode, MLX5_CMD_OP_SET_ROCE_ADDRESS);
291 	return mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
292 }
293 
mlx5_ib_add_gid(struct ib_device * device,u8 port_num,unsigned int index,const union ib_gid * gid,const struct ib_gid_attr * attr,__always_unused void ** context)294 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
295 			   unsigned int index, const union ib_gid *gid,
296 			   const struct ib_gid_attr *attr,
297 			   __always_unused void **context)
298 {
299 	return set_roce_addr(device, port_num, index, gid, attr);
300 }
301 
mlx5_ib_del_gid(struct ib_device * device,u8 port_num,unsigned int index,__always_unused void ** context)302 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
303 			   unsigned int index, __always_unused void **context)
304 {
305 	return set_roce_addr(device, port_num, index, NULL, NULL);
306 }
307 
mlx5_get_roce_udp_sport(struct mlx5_ib_dev * dev,u8 port_num,int index)308 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
309 			       int index)
310 {
311 	struct ib_gid_attr attr;
312 	union ib_gid gid;
313 
314 	if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
315 		return 0;
316 
317 	if (!attr.ndev)
318 		return 0;
319 
320 	dev_put(attr.ndev);
321 
322 	if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
323 		return 0;
324 
325 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
326 }
327 
mlx5_get_roce_gid_type(struct mlx5_ib_dev * dev,u8 port_num,int index,enum ib_gid_type * gid_type)328 int mlx5_get_roce_gid_type(struct mlx5_ib_dev *dev, u8 port_num,
329 			   int index, enum ib_gid_type *gid_type)
330 {
331 	struct ib_gid_attr attr;
332 	union ib_gid gid;
333 	int ret;
334 
335 	ret = ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr);
336 	if (ret)
337 		return ret;
338 
339 	if (!attr.ndev)
340 		return -ENODEV;
341 
342 	dev_put(attr.ndev);
343 
344 	*gid_type = attr.gid_type;
345 
346 	return 0;
347 }
348 
mlx5_use_mad_ifc(struct mlx5_ib_dev * dev)349 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
350 {
351 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
352 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
353 	return 0;
354 }
355 
356 enum {
357 	MLX5_VPORT_ACCESS_METHOD_MAD,
358 	MLX5_VPORT_ACCESS_METHOD_HCA,
359 	MLX5_VPORT_ACCESS_METHOD_NIC,
360 };
361 
mlx5_get_vport_access_method(struct ib_device * ibdev)362 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
363 {
364 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
365 		return MLX5_VPORT_ACCESS_METHOD_MAD;
366 
367 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
368 	    IB_LINK_LAYER_ETHERNET)
369 		return MLX5_VPORT_ACCESS_METHOD_NIC;
370 
371 	return MLX5_VPORT_ACCESS_METHOD_HCA;
372 }
373 
get_atomic_caps(struct mlx5_ib_dev * dev,struct ib_device_attr * props)374 static void get_atomic_caps(struct mlx5_ib_dev *dev,
375 			    struct ib_device_attr *props)
376 {
377 	u8 tmp;
378 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
379 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
380 	u8 atomic_req_8B_endianness_mode =
381 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianess_mode);
382 
383 	/* Check if HW supports 8 bytes standard atomic operations and capable
384 	 * of host endianness respond
385 	 */
386 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
387 	if (((atomic_operations & tmp) == tmp) &&
388 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
389 	    (atomic_req_8B_endianness_mode)) {
390 		props->atomic_cap = IB_ATOMIC_HCA;
391 	} else {
392 		props->atomic_cap = IB_ATOMIC_NONE;
393 	}
394 }
395 
mlx5_query_system_image_guid(struct ib_device * ibdev,__be64 * sys_image_guid)396 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
397 					__be64 *sys_image_guid)
398 {
399 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
400 	struct mlx5_core_dev *mdev = dev->mdev;
401 	u64 tmp;
402 	int err;
403 
404 	switch (mlx5_get_vport_access_method(ibdev)) {
405 	case MLX5_VPORT_ACCESS_METHOD_MAD:
406 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
407 							    sys_image_guid);
408 
409 	case MLX5_VPORT_ACCESS_METHOD_HCA:
410 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
411 		break;
412 
413 	case MLX5_VPORT_ACCESS_METHOD_NIC:
414 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
415 		break;
416 
417 	default:
418 		return -EINVAL;
419 	}
420 
421 	if (!err)
422 		*sys_image_guid = cpu_to_be64(tmp);
423 
424 	return err;
425 
426 }
427 
mlx5_query_max_pkeys(struct ib_device * ibdev,u16 * max_pkeys)428 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
429 				u16 *max_pkeys)
430 {
431 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
432 	struct mlx5_core_dev *mdev = dev->mdev;
433 
434 	switch (mlx5_get_vport_access_method(ibdev)) {
435 	case MLX5_VPORT_ACCESS_METHOD_MAD:
436 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
437 
438 	case MLX5_VPORT_ACCESS_METHOD_HCA:
439 	case MLX5_VPORT_ACCESS_METHOD_NIC:
440 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
441 						pkey_table_size));
442 		return 0;
443 
444 	default:
445 		return -EINVAL;
446 	}
447 }
448 
mlx5_query_vendor_id(struct ib_device * ibdev,u32 * vendor_id)449 static int mlx5_query_vendor_id(struct ib_device *ibdev,
450 				u32 *vendor_id)
451 {
452 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
453 
454 	switch (mlx5_get_vport_access_method(ibdev)) {
455 	case MLX5_VPORT_ACCESS_METHOD_MAD:
456 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
457 
458 	case MLX5_VPORT_ACCESS_METHOD_HCA:
459 	case MLX5_VPORT_ACCESS_METHOD_NIC:
460 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
461 
462 	default:
463 		return -EINVAL;
464 	}
465 }
466 
mlx5_query_node_guid(struct mlx5_ib_dev * dev,__be64 * node_guid)467 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
468 				__be64 *node_guid)
469 {
470 	u64 tmp;
471 	int err;
472 
473 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
474 	case MLX5_VPORT_ACCESS_METHOD_MAD:
475 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
476 
477 	case MLX5_VPORT_ACCESS_METHOD_HCA:
478 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
479 		break;
480 
481 	case MLX5_VPORT_ACCESS_METHOD_NIC:
482 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
483 		break;
484 
485 	default:
486 		return -EINVAL;
487 	}
488 
489 	if (!err)
490 		*node_guid = cpu_to_be64(tmp);
491 
492 	return err;
493 }
494 
495 struct mlx5_reg_node_desc {
496 	u8	desc[IB_DEVICE_NODE_DESC_MAX];
497 };
498 
mlx5_query_node_desc(struct mlx5_ib_dev * dev,char * node_desc)499 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
500 {
501 	struct mlx5_reg_node_desc in;
502 
503 	if (mlx5_use_mad_ifc(dev))
504 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
505 
506 	memset(&in, 0, sizeof(in));
507 
508 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
509 				    sizeof(struct mlx5_reg_node_desc),
510 				    MLX5_REG_NODE_DESC, 0, 0);
511 }
512 
mlx5_ib_query_device(struct ib_device * ibdev,struct ib_device_attr * props,struct ib_udata * uhw)513 static int mlx5_ib_query_device(struct ib_device *ibdev,
514 				struct ib_device_attr *props,
515 				struct ib_udata *uhw)
516 {
517 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
518 	struct mlx5_core_dev *mdev = dev->mdev;
519 	int err = -ENOMEM;
520 	int max_sq_desc;
521 	int max_rq_sg;
522 	int max_sq_sg;
523 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
524 	struct mlx5_ib_query_device_resp resp = {};
525 	size_t resp_len;
526 	u64 max_tso;
527 
528 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
529 	if (uhw->outlen && uhw->outlen < resp_len)
530 		return -EINVAL;
531 	else
532 		resp.response_length = resp_len;
533 
534 	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
535 		return -EINVAL;
536 
537 	memset(props, 0, sizeof(*props));
538 	err = mlx5_query_system_image_guid(ibdev,
539 					   &props->sys_image_guid);
540 	if (err)
541 		return err;
542 
543 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
544 	if (err)
545 		return err;
546 
547 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
548 	if (err)
549 		return err;
550 
551 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
552 		(fw_rev_min(dev->mdev) << 16) |
553 		fw_rev_sub(dev->mdev);
554 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
555 		IB_DEVICE_PORT_ACTIVE_EVENT		|
556 		IB_DEVICE_SYS_IMAGE_GUID		|
557 		IB_DEVICE_RC_RNR_NAK_GEN;
558 
559 	if (MLX5_CAP_GEN(mdev, pkv))
560 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
561 	if (MLX5_CAP_GEN(mdev, qkv))
562 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
563 	if (MLX5_CAP_GEN(mdev, apm))
564 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
565 	if (MLX5_CAP_GEN(mdev, xrc))
566 		props->device_cap_flags |= IB_DEVICE_XRC;
567 	if (MLX5_CAP_GEN(mdev, imaicl)) {
568 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
569 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
570 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
571 		/* We support 'Gappy' memory registration too */
572 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
573 	}
574 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
575 	if (MLX5_CAP_GEN(mdev, sho)) {
576 		props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
577 		/* At this stage no support for signature handover */
578 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
579 				      IB_PROT_T10DIF_TYPE_2 |
580 				      IB_PROT_T10DIF_TYPE_3;
581 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
582 				       IB_GUARD_T10DIF_CSUM;
583 	}
584 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
585 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
586 
587 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads)) {
588 		if (MLX5_CAP_ETH(mdev, csum_cap))
589 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
590 
591 		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
592 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
593 			if (max_tso) {
594 				resp.tso_caps.max_tso = 1 << max_tso;
595 				resp.tso_caps.supported_qpts |=
596 					1 << IB_QPT_RAW_PACKET;
597 				resp.response_length += sizeof(resp.tso_caps);
598 			}
599 		}
600 
601 		if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
602 			resp.rss_caps.rx_hash_function =
603 						MLX5_RX_HASH_FUNC_TOEPLITZ;
604 			resp.rss_caps.rx_hash_fields_mask =
605 						MLX5_RX_HASH_SRC_IPV4 |
606 						MLX5_RX_HASH_DST_IPV4 |
607 						MLX5_RX_HASH_SRC_IPV6 |
608 						MLX5_RX_HASH_DST_IPV6 |
609 						MLX5_RX_HASH_SRC_PORT_TCP |
610 						MLX5_RX_HASH_DST_PORT_TCP |
611 						MLX5_RX_HASH_SRC_PORT_UDP |
612 						MLX5_RX_HASH_DST_PORT_UDP;
613 			resp.response_length += sizeof(resp.rss_caps);
614 		}
615 	} else {
616 		if (field_avail(typeof(resp), tso_caps, uhw->outlen))
617 			resp.response_length += sizeof(resp.tso_caps);
618 		if (field_avail(typeof(resp), rss_caps, uhw->outlen))
619 			resp.response_length += sizeof(resp.rss_caps);
620 	}
621 
622 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
623 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
624 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
625 	}
626 
627 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
628 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs))
629 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
630 
631 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
632 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
633 
634 	props->vendor_part_id	   = mdev->pdev->device;
635 	props->hw_ver		   = mdev->pdev->revision;
636 
637 	props->max_mr_size	   = ~0ull;
638 	props->page_size_cap	   = ~(min_page_size - 1);
639 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
640 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
641 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
642 		     sizeof(struct mlx5_wqe_data_seg);
643 	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
644 	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
645 		     sizeof(struct mlx5_wqe_raddr_seg)) /
646 		sizeof(struct mlx5_wqe_data_seg);
647 	props->max_sge = min(max_rq_sg, max_sq_sg);
648 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
649 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
650 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
651 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
652 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
653 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
654 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
655 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
656 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
657 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
658 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
659 	props->max_srq_sge	   = max_rq_sg - 1;
660 	props->max_fast_reg_page_list_len =
661 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
662 	get_atomic_caps(dev, props);
663 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
664 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
665 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
666 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
667 					   props->max_mcast_grp;
668 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
669 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
670 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
671 
672 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
673 	if (MLX5_CAP_GEN(mdev, pg))
674 		props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
675 	props->odp_caps = dev->odp_caps;
676 #endif
677 
678 	if (MLX5_CAP_GEN(mdev, cd))
679 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
680 
681 	if (!mlx5_core_is_pf(mdev))
682 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
683 
684 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
685 	    IB_LINK_LAYER_ETHERNET) {
686 		props->rss_caps.max_rwq_indirection_tables =
687 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
688 		props->rss_caps.max_rwq_indirection_table_size =
689 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
690 		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
691 		props->max_wq_type_rq =
692 			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
693 	}
694 
695 	if (uhw->outlen) {
696 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
697 
698 		if (err)
699 			return err;
700 	}
701 
702 	return 0;
703 }
704 
705 enum mlx5_ib_width {
706 	MLX5_IB_WIDTH_1X	= 1 << 0,
707 	MLX5_IB_WIDTH_2X	= 1 << 1,
708 	MLX5_IB_WIDTH_4X	= 1 << 2,
709 	MLX5_IB_WIDTH_8X	= 1 << 3,
710 	MLX5_IB_WIDTH_12X	= 1 << 4
711 };
712 
translate_active_width(struct ib_device * ibdev,u8 active_width,u8 * ib_width)713 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
714 				  u8 *ib_width)
715 {
716 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
717 	int err = 0;
718 
719 	if (active_width & MLX5_IB_WIDTH_1X) {
720 		*ib_width = IB_WIDTH_1X;
721 	} else if (active_width & MLX5_IB_WIDTH_2X) {
722 		mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
723 			    (int)active_width);
724 		err = -EINVAL;
725 	} else if (active_width & MLX5_IB_WIDTH_4X) {
726 		*ib_width = IB_WIDTH_4X;
727 	} else if (active_width & MLX5_IB_WIDTH_8X) {
728 		*ib_width = IB_WIDTH_8X;
729 	} else if (active_width & MLX5_IB_WIDTH_12X) {
730 		*ib_width = IB_WIDTH_12X;
731 	} else {
732 		mlx5_ib_dbg(dev, "Invalid active_width %d\n",
733 			    (int)active_width);
734 		err = -EINVAL;
735 	}
736 
737 	return err;
738 }
739 
mlx5_mtu_to_ib_mtu(int mtu)740 static int mlx5_mtu_to_ib_mtu(int mtu)
741 {
742 	switch (mtu) {
743 	case 256: return 1;
744 	case 512: return 2;
745 	case 1024: return 3;
746 	case 2048: return 4;
747 	case 4096: return 5;
748 	default:
749 		pr_warn("invalid mtu\n");
750 		return -1;
751 	}
752 }
753 
754 enum ib_max_vl_num {
755 	__IB_MAX_VL_0		= 1,
756 	__IB_MAX_VL_0_1		= 2,
757 	__IB_MAX_VL_0_3		= 3,
758 	__IB_MAX_VL_0_7		= 4,
759 	__IB_MAX_VL_0_14	= 5,
760 };
761 
762 enum mlx5_vl_hw_cap {
763 	MLX5_VL_HW_0	= 1,
764 	MLX5_VL_HW_0_1	= 2,
765 	MLX5_VL_HW_0_2	= 3,
766 	MLX5_VL_HW_0_3	= 4,
767 	MLX5_VL_HW_0_4	= 5,
768 	MLX5_VL_HW_0_5	= 6,
769 	MLX5_VL_HW_0_6	= 7,
770 	MLX5_VL_HW_0_7	= 8,
771 	MLX5_VL_HW_0_14	= 15
772 };
773 
translate_max_vl_num(struct ib_device * ibdev,u8 vl_hw_cap,u8 * max_vl_num)774 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
775 				u8 *max_vl_num)
776 {
777 	switch (vl_hw_cap) {
778 	case MLX5_VL_HW_0:
779 		*max_vl_num = __IB_MAX_VL_0;
780 		break;
781 	case MLX5_VL_HW_0_1:
782 		*max_vl_num = __IB_MAX_VL_0_1;
783 		break;
784 	case MLX5_VL_HW_0_3:
785 		*max_vl_num = __IB_MAX_VL_0_3;
786 		break;
787 	case MLX5_VL_HW_0_7:
788 		*max_vl_num = __IB_MAX_VL_0_7;
789 		break;
790 	case MLX5_VL_HW_0_14:
791 		*max_vl_num = __IB_MAX_VL_0_14;
792 		break;
793 
794 	default:
795 		return -EINVAL;
796 	}
797 
798 	return 0;
799 }
800 
mlx5_query_hca_port(struct ib_device * ibdev,u8 port,struct ib_port_attr * props)801 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
802 			       struct ib_port_attr *props)
803 {
804 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
805 	struct mlx5_core_dev *mdev = dev->mdev;
806 	struct mlx5_hca_vport_context *rep;
807 	u16 max_mtu;
808 	u16 oper_mtu;
809 	int err;
810 	u8 ib_link_width_oper;
811 	u8 vl_hw_cap;
812 
813 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
814 	if (!rep) {
815 		err = -ENOMEM;
816 		goto out;
817 	}
818 
819 	memset(props, 0, sizeof(*props));
820 
821 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
822 	if (err)
823 		goto out;
824 
825 	props->lid		= rep->lid;
826 	props->lmc		= rep->lmc;
827 	props->sm_lid		= rep->sm_lid;
828 	props->sm_sl		= rep->sm_sl;
829 	props->state		= rep->vport_state;
830 	props->phys_state	= rep->port_physical_state;
831 	props->port_cap_flags	= rep->cap_mask1;
832 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
833 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
834 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
835 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
836 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
837 	props->subnet_timeout	= rep->subnet_timeout;
838 	props->init_type_reply	= rep->init_type_reply;
839 	props->grh_required	= rep->grh_required;
840 
841 	err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
842 	if (err)
843 		goto out;
844 
845 	err = translate_active_width(ibdev, ib_link_width_oper,
846 				     &props->active_width);
847 	if (err)
848 		goto out;
849 	err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
850 	if (err)
851 		goto out;
852 
853 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
854 
855 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
856 
857 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
858 
859 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
860 
861 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
862 	if (err)
863 		goto out;
864 
865 	err = translate_max_vl_num(ibdev, vl_hw_cap,
866 				   &props->max_vl_num);
867 out:
868 	kfree(rep);
869 	return err;
870 }
871 
mlx5_ib_query_port(struct ib_device * ibdev,u8 port,struct ib_port_attr * props)872 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
873 		       struct ib_port_attr *props)
874 {
875 	switch (mlx5_get_vport_access_method(ibdev)) {
876 	case MLX5_VPORT_ACCESS_METHOD_MAD:
877 		return mlx5_query_mad_ifc_port(ibdev, port, props);
878 
879 	case MLX5_VPORT_ACCESS_METHOD_HCA:
880 		return mlx5_query_hca_port(ibdev, port, props);
881 
882 	case MLX5_VPORT_ACCESS_METHOD_NIC:
883 		return mlx5_query_port_roce(ibdev, port, props);
884 
885 	default:
886 		return -EINVAL;
887 	}
888 }
889 
mlx5_ib_query_gid(struct ib_device * ibdev,u8 port,int index,union ib_gid * gid)890 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
891 			     union ib_gid *gid)
892 {
893 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
894 	struct mlx5_core_dev *mdev = dev->mdev;
895 
896 	switch (mlx5_get_vport_access_method(ibdev)) {
897 	case MLX5_VPORT_ACCESS_METHOD_MAD:
898 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
899 
900 	case MLX5_VPORT_ACCESS_METHOD_HCA:
901 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
902 
903 	default:
904 		return -EINVAL;
905 	}
906 
907 }
908 
mlx5_ib_query_pkey(struct ib_device * ibdev,u8 port,u16 index,u16 * pkey)909 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
910 			      u16 *pkey)
911 {
912 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
913 	struct mlx5_core_dev *mdev = dev->mdev;
914 
915 	switch (mlx5_get_vport_access_method(ibdev)) {
916 	case MLX5_VPORT_ACCESS_METHOD_MAD:
917 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
918 
919 	case MLX5_VPORT_ACCESS_METHOD_HCA:
920 	case MLX5_VPORT_ACCESS_METHOD_NIC:
921 		return mlx5_query_hca_vport_pkey(mdev, 0, port,  0, index,
922 						 pkey);
923 	default:
924 		return -EINVAL;
925 	}
926 }
927 
mlx5_ib_modify_device(struct ib_device * ibdev,int mask,struct ib_device_modify * props)928 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
929 				 struct ib_device_modify *props)
930 {
931 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
932 	struct mlx5_reg_node_desc in;
933 	struct mlx5_reg_node_desc out;
934 	int err;
935 
936 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
937 		return -EOPNOTSUPP;
938 
939 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
940 		return 0;
941 
942 	/*
943 	 * If possible, pass node desc to FW, so it can generate
944 	 * a 144 trap.  If cmd fails, just ignore.
945 	 */
946 	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
947 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
948 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
949 	if (err)
950 		return err;
951 
952 	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
953 
954 	return err;
955 }
956 
mlx5_ib_modify_port(struct ib_device * ibdev,u8 port,int mask,struct ib_port_modify * props)957 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
958 			       struct ib_port_modify *props)
959 {
960 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
961 	struct ib_port_attr attr;
962 	u32 tmp;
963 	int err;
964 
965 	mutex_lock(&dev->cap_mask_mutex);
966 
967 	err = mlx5_ib_query_port(ibdev, port, &attr);
968 	if (err)
969 		goto out;
970 
971 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
972 		~props->clr_port_cap_mask;
973 
974 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
975 
976 out:
977 	mutex_unlock(&dev->cap_mask_mutex);
978 	return err;
979 }
980 
mlx5_ib_alloc_ucontext(struct ib_device * ibdev,struct ib_udata * udata)981 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
982 						  struct ib_udata *udata)
983 {
984 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
985 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
986 	struct mlx5_ib_alloc_ucontext_resp resp = {};
987 	struct mlx5_ib_ucontext *context;
988 	struct mlx5_uuar_info *uuari;
989 	struct mlx5_uar *uars;
990 	int gross_uuars;
991 	int num_uars;
992 	int ver;
993 	int uuarn;
994 	int err;
995 	int i;
996 	size_t reqlen;
997 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
998 				     max_cqe_version);
999 
1000 	if (!dev->ib_active)
1001 		return ERR_PTR(-EAGAIN);
1002 
1003 	if (udata->inlen < sizeof(struct ib_uverbs_cmd_hdr))
1004 		return ERR_PTR(-EINVAL);
1005 
1006 	reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
1007 	if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1008 		ver = 0;
1009 	else if (reqlen >= min_req_v2)
1010 		ver = 2;
1011 	else
1012 		return ERR_PTR(-EINVAL);
1013 
1014 	err = ib_copy_from_udata(&req, udata, min(reqlen, sizeof(req)));
1015 	if (err)
1016 		return ERR_PTR(err);
1017 
1018 	if (req.flags)
1019 		return ERR_PTR(-EINVAL);
1020 
1021 	if (req.total_num_uuars > MLX5_MAX_UUARS)
1022 		return ERR_PTR(-ENOMEM);
1023 
1024 	if (req.total_num_uuars == 0)
1025 		return ERR_PTR(-EINVAL);
1026 
1027 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1028 		return ERR_PTR(-EOPNOTSUPP);
1029 
1030 	if (reqlen > sizeof(req) &&
1031 	    !ib_is_udata_cleared(udata, sizeof(req),
1032 				 reqlen - sizeof(req)))
1033 		return ERR_PTR(-EOPNOTSUPP);
1034 
1035 	req.total_num_uuars = ALIGN(req.total_num_uuars,
1036 				    MLX5_NON_FP_BF_REGS_PER_PAGE);
1037 	if (req.num_low_latency_uuars > req.total_num_uuars - 1)
1038 		return ERR_PTR(-EINVAL);
1039 
1040 	num_uars = req.total_num_uuars / MLX5_NON_FP_BF_REGS_PER_PAGE;
1041 	gross_uuars = num_uars * MLX5_BF_REGS_PER_PAGE;
1042 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1043 	if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1044 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1045 	resp.cache_line_size = cache_line_size();
1046 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1047 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1048 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1049 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1050 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1051 	resp.cqe_version = min_t(__u8,
1052 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1053 				 req.max_cqe_version);
1054 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1055 				   sizeof(resp.response_length), udata->outlen);
1056 
1057 	context = kzalloc(sizeof(*context), GFP_KERNEL);
1058 	if (!context)
1059 		return ERR_PTR(-ENOMEM);
1060 
1061 	uuari = &context->uuari;
1062 	mutex_init(&uuari->lock);
1063 	uars = kcalloc(num_uars, sizeof(*uars), GFP_KERNEL);
1064 	if (!uars) {
1065 		err = -ENOMEM;
1066 		goto out_ctx;
1067 	}
1068 
1069 	uuari->bitmap = kcalloc(BITS_TO_LONGS(gross_uuars),
1070 				sizeof(*uuari->bitmap),
1071 				GFP_KERNEL);
1072 	if (!uuari->bitmap) {
1073 		err = -ENOMEM;
1074 		goto out_uar_ctx;
1075 	}
1076 	/*
1077 	 * clear all fast path uuars
1078 	 */
1079 	for (i = 0; i < gross_uuars; i++) {
1080 		uuarn = i & 3;
1081 		if (uuarn == 2 || uuarn == 3)
1082 			set_bit(i, uuari->bitmap);
1083 	}
1084 
1085 	uuari->count = kcalloc(gross_uuars, sizeof(*uuari->count), GFP_KERNEL);
1086 	if (!uuari->count) {
1087 		err = -ENOMEM;
1088 		goto out_bitmap;
1089 	}
1090 
1091 	for (i = 0; i < num_uars; i++) {
1092 		err = mlx5_cmd_alloc_uar(dev->mdev, &uars[i].index);
1093 		if (err)
1094 			goto out_count;
1095 	}
1096 
1097 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1098 	context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1099 #endif
1100 
1101 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1102 		err = mlx5_core_alloc_transport_domain(dev->mdev,
1103 						       &context->tdn);
1104 		if (err)
1105 			goto out_uars;
1106 	}
1107 
1108 	INIT_LIST_HEAD(&context->vma_private_list);
1109 	INIT_LIST_HEAD(&context->db_page_list);
1110 	mutex_init(&context->db_page_mutex);
1111 
1112 	resp.tot_uuars = req.total_num_uuars;
1113 	resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
1114 
1115 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1116 		resp.response_length += sizeof(resp.cqe_version);
1117 
1118 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1119 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE;
1120 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1121 	}
1122 
1123 	/*
1124 	 * We don't want to expose information from the PCI bar that is located
1125 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1126 	 * pretend we don't support reading the HCA's core clock. This is also
1127 	 * forced by mmap function.
1128 	 */
1129 	if (field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1130 		if (PAGE_SIZE <= 4096) {
1131 			resp.comp_mask |=
1132 				MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1133 			resp.hca_core_clock_offset =
1134 				offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1135 		}
1136 		resp.response_length += sizeof(resp.hca_core_clock_offset) +
1137 					sizeof(resp.reserved2);
1138 	}
1139 
1140 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1141 	if (err)
1142 		goto out_td;
1143 
1144 	uuari->ver = ver;
1145 	uuari->num_low_latency_uuars = req.num_low_latency_uuars;
1146 	uuari->uars = uars;
1147 	uuari->num_uars = num_uars;
1148 	context->cqe_version = resp.cqe_version;
1149 
1150 	return &context->ibucontext;
1151 
1152 out_td:
1153 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1154 		mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1155 
1156 out_uars:
1157 	for (i--; i >= 0; i--)
1158 		mlx5_cmd_free_uar(dev->mdev, uars[i].index);
1159 out_count:
1160 	kfree(uuari->count);
1161 
1162 out_bitmap:
1163 	kfree(uuari->bitmap);
1164 
1165 out_uar_ctx:
1166 	kfree(uars);
1167 
1168 out_ctx:
1169 	kfree(context);
1170 	return ERR_PTR(err);
1171 }
1172 
mlx5_ib_dealloc_ucontext(struct ib_ucontext * ibcontext)1173 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1174 {
1175 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1176 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1177 	struct mlx5_uuar_info *uuari = &context->uuari;
1178 	int i;
1179 
1180 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1181 		mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1182 
1183 	for (i = 0; i < uuari->num_uars; i++) {
1184 		if (mlx5_cmd_free_uar(dev->mdev, uuari->uars[i].index))
1185 			mlx5_ib_warn(dev, "failed to free UAR 0x%x\n", uuari->uars[i].index);
1186 	}
1187 
1188 	kfree(uuari->count);
1189 	kfree(uuari->bitmap);
1190 	kfree(uuari->uars);
1191 	kfree(context);
1192 
1193 	return 0;
1194 }
1195 
uar_index2pfn(struct mlx5_ib_dev * dev,int index)1196 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, int index)
1197 {
1198 	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + index;
1199 }
1200 
get_command(unsigned long offset)1201 static int get_command(unsigned long offset)
1202 {
1203 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1204 }
1205 
get_arg(unsigned long offset)1206 static int get_arg(unsigned long offset)
1207 {
1208 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1209 }
1210 
get_index(unsigned long offset)1211 static int get_index(unsigned long offset)
1212 {
1213 	return get_arg(offset);
1214 }
1215 
mlx5_ib_vma_open(struct vm_area_struct * area)1216 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1217 {
1218 	/* vma_open is called when a new VMA is created on top of our VMA.  This
1219 	 * is done through either mremap flow or split_vma (usually due to
1220 	 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1221 	 * as this VMA is strongly hardware related.  Therefore we set the
1222 	 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1223 	 * calling us again and trying to do incorrect actions.  We assume that
1224 	 * the original VMA size is exactly a single page, and therefore all
1225 	 * "splitting" operation will not happen to it.
1226 	 */
1227 	area->vm_ops = NULL;
1228 }
1229 
mlx5_ib_vma_close(struct vm_area_struct * area)1230 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1231 {
1232 	struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1233 
1234 	/* It's guaranteed that all VMAs opened on a FD are closed before the
1235 	 * file itself is closed, therefore no sync is needed with the regular
1236 	 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1237 	 * However need a sync with accessing the vma as part of
1238 	 * mlx5_ib_disassociate_ucontext.
1239 	 * The close operation is usually called under mm->mmap_sem except when
1240 	 * process is exiting.
1241 	 * The exiting case is handled explicitly as part of
1242 	 * mlx5_ib_disassociate_ucontext.
1243 	 */
1244 	mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1245 
1246 	/* setting the vma context pointer to null in the mlx5_ib driver's
1247 	 * private data, to protect a race condition in
1248 	 * mlx5_ib_disassociate_ucontext().
1249 	 */
1250 	mlx5_ib_vma_priv_data->vma = NULL;
1251 	list_del(&mlx5_ib_vma_priv_data->list);
1252 	kfree(mlx5_ib_vma_priv_data);
1253 }
1254 
1255 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1256 	.open = mlx5_ib_vma_open,
1257 	.close = mlx5_ib_vma_close
1258 };
1259 
mlx5_ib_set_vma_data(struct vm_area_struct * vma,struct mlx5_ib_ucontext * ctx)1260 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1261 				struct mlx5_ib_ucontext *ctx)
1262 {
1263 	struct mlx5_ib_vma_private_data *vma_prv;
1264 	struct list_head *vma_head = &ctx->vma_private_list;
1265 
1266 	vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1267 	if (!vma_prv)
1268 		return -ENOMEM;
1269 
1270 	vma_prv->vma = vma;
1271 	vma->vm_private_data = vma_prv;
1272 	vma->vm_ops =  &mlx5_ib_vm_ops;
1273 
1274 	list_add(&vma_prv->list, vma_head);
1275 
1276 	return 0;
1277 }
1278 
mlx5_ib_disassociate_ucontext(struct ib_ucontext * ibcontext)1279 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1280 {
1281 	int ret;
1282 	struct vm_area_struct *vma;
1283 	struct mlx5_ib_vma_private_data *vma_private, *n;
1284 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1285 	struct task_struct *owning_process  = NULL;
1286 	struct mm_struct   *owning_mm       = NULL;
1287 
1288 	owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1289 	if (!owning_process)
1290 		return;
1291 
1292 	owning_mm = get_task_mm(owning_process);
1293 	if (!owning_mm) {
1294 		pr_info("no mm, disassociate ucontext is pending task termination\n");
1295 		while (1) {
1296 			put_task_struct(owning_process);
1297 			usleep_range(1000, 2000);
1298 			owning_process = get_pid_task(ibcontext->tgid,
1299 						      PIDTYPE_PID);
1300 			if (!owning_process ||
1301 			    owning_process->state == TASK_DEAD) {
1302 				pr_info("disassociate ucontext done, task was terminated\n");
1303 				/* in case task was dead need to release the
1304 				 * task struct.
1305 				 */
1306 				if (owning_process)
1307 					put_task_struct(owning_process);
1308 				return;
1309 			}
1310 		}
1311 	}
1312 
1313 	/* need to protect from a race on closing the vma as part of
1314 	 * mlx5_ib_vma_close.
1315 	 */
1316 	down_write(&owning_mm->mmap_sem);
1317 	list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1318 				 list) {
1319 		vma = vma_private->vma;
1320 		ret = zap_vma_ptes(vma, vma->vm_start,
1321 				   PAGE_SIZE);
1322 		WARN_ONCE(ret, "%s: zap_vma_ptes failed", __func__);
1323 		/* context going to be destroyed, should
1324 		 * not access ops any more.
1325 		 */
1326 		vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
1327 		vma->vm_ops = NULL;
1328 		list_del(&vma_private->list);
1329 		kfree(vma_private);
1330 	}
1331 	up_write(&owning_mm->mmap_sem);
1332 	mmput(owning_mm);
1333 	put_task_struct(owning_process);
1334 }
1335 
mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)1336 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1337 {
1338 	switch (cmd) {
1339 	case MLX5_IB_MMAP_WC_PAGE:
1340 		return "WC";
1341 	case MLX5_IB_MMAP_REGULAR_PAGE:
1342 		return "best effort WC";
1343 	case MLX5_IB_MMAP_NC_PAGE:
1344 		return "NC";
1345 	default:
1346 		return NULL;
1347 	}
1348 }
1349 
uar_mmap(struct mlx5_ib_dev * dev,enum mlx5_ib_mmap_cmd cmd,struct vm_area_struct * vma,struct mlx5_ib_ucontext * context)1350 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
1351 		    struct vm_area_struct *vma,
1352 		    struct mlx5_ib_ucontext *context)
1353 {
1354 	struct mlx5_uuar_info *uuari = &context->uuari;
1355 	int err;
1356 	unsigned long idx;
1357 	phys_addr_t pfn, pa;
1358 	pgprot_t prot;
1359 
1360 	switch (cmd) {
1361 	case MLX5_IB_MMAP_WC_PAGE:
1362 /* Some architectures don't support WC memory */
1363 #if defined(CONFIG_X86)
1364 		if (!pat_enabled())
1365 			return -EPERM;
1366 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
1367 			return -EPERM;
1368 #endif
1369 	/* fall through */
1370 	case MLX5_IB_MMAP_REGULAR_PAGE:
1371 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
1372 		prot = pgprot_writecombine(vma->vm_page_prot);
1373 		break;
1374 	case MLX5_IB_MMAP_NC_PAGE:
1375 		prot = pgprot_noncached(vma->vm_page_prot);
1376 		break;
1377 	default:
1378 		return -EINVAL;
1379 	}
1380 
1381 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1382 		return -EINVAL;
1383 
1384 	idx = get_index(vma->vm_pgoff);
1385 	if (idx >= uuari->num_uars)
1386 		return -EINVAL;
1387 
1388 	pfn = uar_index2pfn(dev, uuari->uars[idx].index);
1389 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
1390 
1391 	vma->vm_page_prot = prot;
1392 	err = io_remap_pfn_range(vma, vma->vm_start, pfn,
1393 				 PAGE_SIZE, vma->vm_page_prot);
1394 	if (err) {
1395 		mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%lx, pfn=%pa, mmap_cmd=%s\n",
1396 			    err, vma->vm_start, &pfn, mmap_cmd2str(cmd));
1397 		return -EAGAIN;
1398 	}
1399 
1400 	pa = pfn << PAGE_SHIFT;
1401 	mlx5_ib_dbg(dev, "mapped %s at 0x%lx, PA %pa\n", mmap_cmd2str(cmd),
1402 		    vma->vm_start, &pa);
1403 
1404 	return mlx5_ib_set_vma_data(vma, context);
1405 }
1406 
mlx5_ib_mmap(struct ib_ucontext * ibcontext,struct vm_area_struct * vma)1407 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1408 {
1409 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1410 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1411 	unsigned long command;
1412 	phys_addr_t pfn;
1413 
1414 	command = get_command(vma->vm_pgoff);
1415 	switch (command) {
1416 	case MLX5_IB_MMAP_WC_PAGE:
1417 	case MLX5_IB_MMAP_NC_PAGE:
1418 	case MLX5_IB_MMAP_REGULAR_PAGE:
1419 		return uar_mmap(dev, command, vma, context);
1420 
1421 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1422 		return -ENOSYS;
1423 
1424 	case MLX5_IB_MMAP_CORE_CLOCK:
1425 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1426 			return -EINVAL;
1427 
1428 		if (vma->vm_flags & VM_WRITE)
1429 			return -EPERM;
1430 
1431 		/* Don't expose to user-space information it shouldn't have */
1432 		if (PAGE_SIZE > 4096)
1433 			return -EOPNOTSUPP;
1434 
1435 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1436 		pfn = (dev->mdev->iseg_base +
1437 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1438 			PAGE_SHIFT;
1439 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1440 				       PAGE_SIZE, vma->vm_page_prot))
1441 			return -EAGAIN;
1442 
1443 		mlx5_ib_dbg(dev, "mapped internal timer at 0x%lx, PA 0x%llx\n",
1444 			    vma->vm_start,
1445 			    (unsigned long long)pfn << PAGE_SHIFT);
1446 		break;
1447 
1448 	default:
1449 		return -EINVAL;
1450 	}
1451 
1452 	return 0;
1453 }
1454 
mlx5_ib_alloc_pd(struct ib_device * ibdev,struct ib_ucontext * context,struct ib_udata * udata)1455 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
1456 				      struct ib_ucontext *context,
1457 				      struct ib_udata *udata)
1458 {
1459 	struct mlx5_ib_alloc_pd_resp resp;
1460 	struct mlx5_ib_pd *pd;
1461 	int err;
1462 
1463 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1464 	if (!pd)
1465 		return ERR_PTR(-ENOMEM);
1466 
1467 	err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
1468 	if (err) {
1469 		kfree(pd);
1470 		return ERR_PTR(err);
1471 	}
1472 
1473 	if (context) {
1474 		resp.pdn = pd->pdn;
1475 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1476 			mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
1477 			kfree(pd);
1478 			return ERR_PTR(-EFAULT);
1479 		}
1480 	}
1481 
1482 	return &pd->ibpd;
1483 }
1484 
mlx5_ib_dealloc_pd(struct ib_pd * pd)1485 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
1486 {
1487 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1488 	struct mlx5_ib_pd *mpd = to_mpd(pd);
1489 
1490 	mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
1491 	kfree(mpd);
1492 
1493 	return 0;
1494 }
1495 
1496 enum {
1497 	MATCH_CRITERIA_ENABLE_OUTER_BIT,
1498 	MATCH_CRITERIA_ENABLE_MISC_BIT,
1499 	MATCH_CRITERIA_ENABLE_INNER_BIT
1500 };
1501 
1502 #define HEADER_IS_ZERO(match_criteria, headers)			           \
1503 	!(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
1504 		    0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
1505 
get_match_criteria_enable(u32 * match_criteria)1506 static u8 get_match_criteria_enable(u32 *match_criteria)
1507 {
1508 	u8 match_criteria_enable;
1509 
1510 	match_criteria_enable =
1511 		(!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
1512 		MATCH_CRITERIA_ENABLE_OUTER_BIT;
1513 	match_criteria_enable |=
1514 		(!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
1515 		MATCH_CRITERIA_ENABLE_MISC_BIT;
1516 	match_criteria_enable |=
1517 		(!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
1518 		MATCH_CRITERIA_ENABLE_INNER_BIT;
1519 
1520 	return match_criteria_enable;
1521 }
1522 
set_proto(void * outer_c,void * outer_v,u8 mask,u8 val)1523 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
1524 {
1525 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
1526 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
1527 }
1528 
set_tos(void * outer_c,void * outer_v,u8 mask,u8 val)1529 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
1530 {
1531 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
1532 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
1533 	MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
1534 	MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
1535 }
1536 
1537 #define LAST_ETH_FIELD vlan_tag
1538 #define LAST_IB_FIELD sl
1539 #define LAST_IPV4_FIELD tos
1540 #define LAST_IPV6_FIELD traffic_class
1541 #define LAST_TCP_UDP_FIELD src_port
1542 
1543 /* Field is the last supported field */
1544 #define FIELDS_NOT_SUPPORTED(filter, field)\
1545 	memchr_inv((void *)&filter.field  +\
1546 		   sizeof(filter.field), 0,\
1547 		   sizeof(filter) -\
1548 		   offsetof(typeof(filter), field) -\
1549 		   sizeof(filter.field))
1550 
parse_flow_attr(u32 * match_c,u32 * match_v,const union ib_flow_spec * ib_spec)1551 static int parse_flow_attr(u32 *match_c, u32 *match_v,
1552 			   const union ib_flow_spec *ib_spec)
1553 {
1554 	void *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1555 					     outer_headers);
1556 	void *outer_headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1557 					     outer_headers);
1558 	void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
1559 					   misc_parameters);
1560 	void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
1561 					   misc_parameters);
1562 
1563 	switch (ib_spec->type) {
1564 	case IB_FLOW_SPEC_ETH:
1565 		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
1566 			return -ENOTSUPP;
1567 
1568 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1569 					     dmac_47_16),
1570 				ib_spec->eth.mask.dst_mac);
1571 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1572 					     dmac_47_16),
1573 				ib_spec->eth.val.dst_mac);
1574 
1575 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1576 					     smac_47_16),
1577 				ib_spec->eth.mask.src_mac);
1578 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1579 					     smac_47_16),
1580 				ib_spec->eth.val.src_mac);
1581 
1582 		if (ib_spec->eth.mask.vlan_tag) {
1583 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1584 				 vlan_tag, 1);
1585 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1586 				 vlan_tag, 1);
1587 
1588 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1589 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
1590 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1591 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
1592 
1593 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1594 				 first_cfi,
1595 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
1596 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1597 				 first_cfi,
1598 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
1599 
1600 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1601 				 first_prio,
1602 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
1603 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1604 				 first_prio,
1605 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
1606 		}
1607 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1608 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
1609 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1610 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
1611 		break;
1612 	case IB_FLOW_SPEC_IPV4:
1613 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
1614 			return -ENOTSUPP;
1615 
1616 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1617 			 ethertype, 0xffff);
1618 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1619 			 ethertype, ETH_P_IP);
1620 
1621 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1622 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1623 		       &ib_spec->ipv4.mask.src_ip,
1624 		       sizeof(ib_spec->ipv4.mask.src_ip));
1625 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1626 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1627 		       &ib_spec->ipv4.val.src_ip,
1628 		       sizeof(ib_spec->ipv4.val.src_ip));
1629 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1630 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1631 		       &ib_spec->ipv4.mask.dst_ip,
1632 		       sizeof(ib_spec->ipv4.mask.dst_ip));
1633 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1634 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1635 		       &ib_spec->ipv4.val.dst_ip,
1636 		       sizeof(ib_spec->ipv4.val.dst_ip));
1637 
1638 		set_tos(outer_headers_c, outer_headers_v,
1639 			ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
1640 
1641 		set_proto(outer_headers_c, outer_headers_v,
1642 			  ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
1643 		break;
1644 	case IB_FLOW_SPEC_IPV6:
1645 		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
1646 			return -ENOTSUPP;
1647 
1648 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1649 			 ethertype, 0xffff);
1650 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1651 			 ethertype, ETH_P_IPV6);
1652 
1653 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1654 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1655 		       &ib_spec->ipv6.mask.src_ip,
1656 		       sizeof(ib_spec->ipv6.mask.src_ip));
1657 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1658 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1659 		       &ib_spec->ipv6.val.src_ip,
1660 		       sizeof(ib_spec->ipv6.val.src_ip));
1661 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1662 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1663 		       &ib_spec->ipv6.mask.dst_ip,
1664 		       sizeof(ib_spec->ipv6.mask.dst_ip));
1665 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1666 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1667 		       &ib_spec->ipv6.val.dst_ip,
1668 		       sizeof(ib_spec->ipv6.val.dst_ip));
1669 
1670 		set_tos(outer_headers_c, outer_headers_v,
1671 			ib_spec->ipv6.mask.traffic_class,
1672 			ib_spec->ipv6.val.traffic_class);
1673 
1674 		set_proto(outer_headers_c, outer_headers_v,
1675 			  ib_spec->ipv6.mask.next_hdr,
1676 			  ib_spec->ipv6.val.next_hdr);
1677 
1678 		MLX5_SET(fte_match_set_misc, misc_params_c,
1679 			 outer_ipv6_flow_label,
1680 			 ntohl(ib_spec->ipv6.mask.flow_label));
1681 		MLX5_SET(fte_match_set_misc, misc_params_v,
1682 			 outer_ipv6_flow_label,
1683 			 ntohl(ib_spec->ipv6.val.flow_label));
1684 		break;
1685 	case IB_FLOW_SPEC_TCP:
1686 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1687 					 LAST_TCP_UDP_FIELD))
1688 			return -ENOTSUPP;
1689 
1690 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1691 			 0xff);
1692 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1693 			 IPPROTO_TCP);
1694 
1695 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_sport,
1696 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1697 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_sport,
1698 			 ntohs(ib_spec->tcp_udp.val.src_port));
1699 
1700 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_dport,
1701 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1702 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_dport,
1703 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1704 		break;
1705 	case IB_FLOW_SPEC_UDP:
1706 		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1707 					 LAST_TCP_UDP_FIELD))
1708 			return -ENOTSUPP;
1709 
1710 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1711 			 0xff);
1712 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1713 			 IPPROTO_UDP);
1714 
1715 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_sport,
1716 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1717 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_sport,
1718 			 ntohs(ib_spec->tcp_udp.val.src_port));
1719 
1720 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_dport,
1721 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1722 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_dport,
1723 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1724 		break;
1725 	default:
1726 		return -EINVAL;
1727 	}
1728 
1729 	return 0;
1730 }
1731 
1732 /* If a flow could catch both multicast and unicast packets,
1733  * it won't fall into the multicast flow steering table and this rule
1734  * could steal other multicast packets.
1735  */
flow_is_multicast_only(struct ib_flow_attr * ib_attr)1736 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr)
1737 {
1738 	struct ib_flow_spec_eth *eth_spec;
1739 
1740 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
1741 	    ib_attr->size < sizeof(struct ib_flow_attr) +
1742 	    sizeof(struct ib_flow_spec_eth) ||
1743 	    ib_attr->num_of_specs < 1)
1744 		return false;
1745 
1746 	eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1);
1747 	if (eth_spec->type != IB_FLOW_SPEC_ETH ||
1748 	    eth_spec->size != sizeof(*eth_spec))
1749 		return false;
1750 
1751 	return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
1752 	       is_multicast_ether_addr(eth_spec->val.dst_mac);
1753 }
1754 
is_valid_attr(const struct ib_flow_attr * flow_attr)1755 static bool is_valid_attr(const struct ib_flow_attr *flow_attr)
1756 {
1757 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
1758 	bool has_ipv4_spec = false;
1759 	bool eth_type_ipv4 = true;
1760 	unsigned int spec_index;
1761 
1762 	/* Validate that ethertype is correct */
1763 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1764 		if (ib_spec->type == IB_FLOW_SPEC_ETH &&
1765 		    ib_spec->eth.mask.ether_type) {
1766 			if (!((ib_spec->eth.mask.ether_type == htons(0xffff)) &&
1767 			      ib_spec->eth.val.ether_type == htons(ETH_P_IP)))
1768 				eth_type_ipv4 = false;
1769 		} else if (ib_spec->type == IB_FLOW_SPEC_IPV4) {
1770 			has_ipv4_spec = true;
1771 		}
1772 		ib_spec = (void *)ib_spec + ib_spec->size;
1773 	}
1774 	return !has_ipv4_spec || eth_type_ipv4;
1775 }
1776 
put_flow_table(struct mlx5_ib_dev * dev,struct mlx5_ib_flow_prio * prio,bool ft_added)1777 static void put_flow_table(struct mlx5_ib_dev *dev,
1778 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
1779 {
1780 	prio->refcount -= !!ft_added;
1781 	if (!prio->refcount) {
1782 		mlx5_destroy_flow_table(prio->flow_table);
1783 		prio->flow_table = NULL;
1784 	}
1785 }
1786 
mlx5_ib_destroy_flow(struct ib_flow * flow_id)1787 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
1788 {
1789 	struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
1790 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
1791 							  struct mlx5_ib_flow_handler,
1792 							  ibflow);
1793 	struct mlx5_ib_flow_handler *iter, *tmp;
1794 
1795 	mutex_lock(&dev->flow_db.lock);
1796 
1797 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
1798 		mlx5_del_flow_rule(iter->rule);
1799 		put_flow_table(dev, iter->prio, true);
1800 		list_del(&iter->list);
1801 		kfree(iter);
1802 	}
1803 
1804 	mlx5_del_flow_rule(handler->rule);
1805 	put_flow_table(dev, handler->prio, true);
1806 	mutex_unlock(&dev->flow_db.lock);
1807 
1808 	kfree(handler);
1809 
1810 	return 0;
1811 }
1812 
ib_prio_to_core_prio(unsigned int priority,bool dont_trap)1813 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
1814 {
1815 	priority *= 2;
1816 	if (!dont_trap)
1817 		priority++;
1818 	return priority;
1819 }
1820 
1821 enum flow_table_type {
1822 	MLX5_IB_FT_RX,
1823 	MLX5_IB_FT_TX
1824 };
1825 
1826 #define MLX5_FS_MAX_TYPES	 10
1827 #define MLX5_FS_MAX_ENTRIES	 32000UL
get_flow_table(struct mlx5_ib_dev * dev,struct ib_flow_attr * flow_attr,enum flow_table_type ft_type)1828 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
1829 						struct ib_flow_attr *flow_attr,
1830 						enum flow_table_type ft_type)
1831 {
1832 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
1833 	struct mlx5_flow_namespace *ns = NULL;
1834 	struct mlx5_ib_flow_prio *prio;
1835 	struct mlx5_flow_table *ft;
1836 	int num_entries;
1837 	int num_groups;
1838 	int priority;
1839 	int err = 0;
1840 
1841 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1842 		if (flow_is_multicast_only(flow_attr) &&
1843 		    !dont_trap)
1844 			priority = MLX5_IB_FLOW_MCAST_PRIO;
1845 		else
1846 			priority = ib_prio_to_core_prio(flow_attr->priority,
1847 							dont_trap);
1848 		ns = mlx5_get_flow_namespace(dev->mdev,
1849 					     MLX5_FLOW_NAMESPACE_BYPASS);
1850 		num_entries = MLX5_FS_MAX_ENTRIES;
1851 		num_groups = MLX5_FS_MAX_TYPES;
1852 		prio = &dev->flow_db.prios[priority];
1853 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
1854 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
1855 		ns = mlx5_get_flow_namespace(dev->mdev,
1856 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
1857 		build_leftovers_ft_param(&priority,
1858 					 &num_entries,
1859 					 &num_groups);
1860 		prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
1861 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
1862 		if (!MLX5_CAP_FLOWTABLE(dev->mdev,
1863 					allow_sniffer_and_nic_rx_shared_tir))
1864 			return ERR_PTR(-ENOTSUPP);
1865 
1866 		ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
1867 					     MLX5_FLOW_NAMESPACE_SNIFFER_RX :
1868 					     MLX5_FLOW_NAMESPACE_SNIFFER_TX);
1869 
1870 		prio = &dev->flow_db.sniffer[ft_type];
1871 		priority = 0;
1872 		num_entries = 1;
1873 		num_groups = 1;
1874 	}
1875 
1876 	if (!ns)
1877 		return ERR_PTR(-ENOTSUPP);
1878 
1879 	ft = prio->flow_table;
1880 	if (!ft) {
1881 		ft = mlx5_create_auto_grouped_flow_table(ns, priority,
1882 							 num_entries,
1883 							 num_groups,
1884 							 0);
1885 
1886 		if (!IS_ERR(ft)) {
1887 			prio->refcount = 0;
1888 			prio->flow_table = ft;
1889 		} else {
1890 			err = PTR_ERR(ft);
1891 		}
1892 	}
1893 
1894 	return err ? ERR_PTR(err) : prio;
1895 }
1896 
create_flow_rule(struct mlx5_ib_dev * dev,struct mlx5_ib_flow_prio * ft_prio,const struct ib_flow_attr * flow_attr,struct mlx5_flow_destination * dst)1897 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
1898 						     struct mlx5_ib_flow_prio *ft_prio,
1899 						     const struct ib_flow_attr *flow_attr,
1900 						     struct mlx5_flow_destination *dst)
1901 {
1902 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
1903 	struct mlx5_ib_flow_handler *handler;
1904 	struct mlx5_flow_spec *spec;
1905 	const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
1906 	unsigned int spec_index;
1907 	u32 action;
1908 	int err = 0;
1909 
1910 	if (!is_valid_attr(flow_attr))
1911 		return ERR_PTR(-EINVAL);
1912 
1913 	spec = mlx5_vzalloc(sizeof(*spec));
1914 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
1915 	if (!handler || !spec) {
1916 		err = -ENOMEM;
1917 		goto free;
1918 	}
1919 
1920 	INIT_LIST_HEAD(&handler->list);
1921 
1922 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1923 		err = parse_flow_attr(spec->match_criteria,
1924 				      spec->match_value, ib_flow);
1925 		if (err < 0)
1926 			goto free;
1927 
1928 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
1929 	}
1930 
1931 	spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
1932 	action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
1933 		MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
1934 	handler->rule = mlx5_add_flow_rule(ft, spec,
1935 					   action,
1936 					   MLX5_FS_DEFAULT_FLOW_TAG,
1937 					   dst);
1938 
1939 	if (IS_ERR(handler->rule)) {
1940 		err = PTR_ERR(handler->rule);
1941 		goto free;
1942 	}
1943 
1944 	ft_prio->refcount++;
1945 	handler->prio = ft_prio;
1946 
1947 	ft_prio->flow_table = ft;
1948 free:
1949 	if (err)
1950 		kfree(handler);
1951 	kvfree(spec);
1952 	return err ? ERR_PTR(err) : handler;
1953 }
1954 
create_dont_trap_rule(struct mlx5_ib_dev * dev,struct mlx5_ib_flow_prio * ft_prio,struct ib_flow_attr * flow_attr,struct mlx5_flow_destination * dst)1955 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
1956 							  struct mlx5_ib_flow_prio *ft_prio,
1957 							  struct ib_flow_attr *flow_attr,
1958 							  struct mlx5_flow_destination *dst)
1959 {
1960 	struct mlx5_ib_flow_handler *handler_dst = NULL;
1961 	struct mlx5_ib_flow_handler *handler = NULL;
1962 
1963 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
1964 	if (!IS_ERR(handler)) {
1965 		handler_dst = create_flow_rule(dev, ft_prio,
1966 					       flow_attr, dst);
1967 		if (IS_ERR(handler_dst)) {
1968 			mlx5_del_flow_rule(handler->rule);
1969 			ft_prio->refcount--;
1970 			kfree(handler);
1971 			handler = handler_dst;
1972 		} else {
1973 			list_add(&handler_dst->list, &handler->list);
1974 		}
1975 	}
1976 
1977 	return handler;
1978 }
1979 enum {
1980 	LEFTOVERS_MC,
1981 	LEFTOVERS_UC,
1982 };
1983 
create_leftovers_rule(struct mlx5_ib_dev * dev,struct mlx5_ib_flow_prio * ft_prio,struct ib_flow_attr * flow_attr,struct mlx5_flow_destination * dst)1984 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
1985 							  struct mlx5_ib_flow_prio *ft_prio,
1986 							  struct ib_flow_attr *flow_attr,
1987 							  struct mlx5_flow_destination *dst)
1988 {
1989 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
1990 	struct mlx5_ib_flow_handler *handler = NULL;
1991 
1992 	static struct {
1993 		struct ib_flow_attr	flow_attr;
1994 		struct ib_flow_spec_eth eth_flow;
1995 	} leftovers_specs[] = {
1996 		[LEFTOVERS_MC] = {
1997 			.flow_attr = {
1998 				.num_of_specs = 1,
1999 				.size = sizeof(leftovers_specs[0])
2000 			},
2001 			.eth_flow = {
2002 				.type = IB_FLOW_SPEC_ETH,
2003 				.size = sizeof(struct ib_flow_spec_eth),
2004 				.mask = {.dst_mac = {0x1} },
2005 				.val =  {.dst_mac = {0x1} }
2006 			}
2007 		},
2008 		[LEFTOVERS_UC] = {
2009 			.flow_attr = {
2010 				.num_of_specs = 1,
2011 				.size = sizeof(leftovers_specs[0])
2012 			},
2013 			.eth_flow = {
2014 				.type = IB_FLOW_SPEC_ETH,
2015 				.size = sizeof(struct ib_flow_spec_eth),
2016 				.mask = {.dst_mac = {0x1} },
2017 				.val = {.dst_mac = {} }
2018 			}
2019 		}
2020 	};
2021 
2022 	handler = create_flow_rule(dev, ft_prio,
2023 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
2024 				   dst);
2025 	if (!IS_ERR(handler) &&
2026 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
2027 		handler_ucast = create_flow_rule(dev, ft_prio,
2028 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
2029 						 dst);
2030 		if (IS_ERR(handler_ucast)) {
2031 			mlx5_del_flow_rule(handler->rule);
2032 			ft_prio->refcount--;
2033 			kfree(handler);
2034 			handler = handler_ucast;
2035 		} else {
2036 			list_add(&handler_ucast->list, &handler->list);
2037 		}
2038 	}
2039 
2040 	return handler;
2041 }
2042 
create_sniffer_rule(struct mlx5_ib_dev * dev,struct mlx5_ib_flow_prio * ft_rx,struct mlx5_ib_flow_prio * ft_tx,struct mlx5_flow_destination * dst)2043 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
2044 							struct mlx5_ib_flow_prio *ft_rx,
2045 							struct mlx5_ib_flow_prio *ft_tx,
2046 							struct mlx5_flow_destination *dst)
2047 {
2048 	struct mlx5_ib_flow_handler *handler_rx;
2049 	struct mlx5_ib_flow_handler *handler_tx;
2050 	int err;
2051 	static const struct ib_flow_attr flow_attr  = {
2052 		.num_of_specs = 0,
2053 		.size = sizeof(flow_attr)
2054 	};
2055 
2056 	handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
2057 	if (IS_ERR(handler_rx)) {
2058 		err = PTR_ERR(handler_rx);
2059 		goto err;
2060 	}
2061 
2062 	handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
2063 	if (IS_ERR(handler_tx)) {
2064 		err = PTR_ERR(handler_tx);
2065 		goto err_tx;
2066 	}
2067 
2068 	list_add(&handler_tx->list, &handler_rx->list);
2069 
2070 	return handler_rx;
2071 
2072 err_tx:
2073 	mlx5_del_flow_rule(handler_rx->rule);
2074 	ft_rx->refcount--;
2075 	kfree(handler_rx);
2076 err:
2077 	return ERR_PTR(err);
2078 }
2079 
mlx5_ib_create_flow(struct ib_qp * qp,struct ib_flow_attr * flow_attr,int domain)2080 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
2081 					   struct ib_flow_attr *flow_attr,
2082 					   int domain)
2083 {
2084 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2085 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2086 	struct mlx5_ib_flow_handler *handler = NULL;
2087 	struct mlx5_flow_destination *dst = NULL;
2088 	struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
2089 	struct mlx5_ib_flow_prio *ft_prio;
2090 	int err;
2091 
2092 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
2093 		return ERR_PTR(-ENOSPC);
2094 
2095 	if (domain != IB_FLOW_DOMAIN_USER ||
2096 	    flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
2097 	    (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
2098 		return ERR_PTR(-EINVAL);
2099 
2100 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
2101 	if (!dst)
2102 		return ERR_PTR(-ENOMEM);
2103 
2104 	mutex_lock(&dev->flow_db.lock);
2105 
2106 	ft_prio = get_flow_table(dev, flow_attr, MLX5_IB_FT_RX);
2107 	if (IS_ERR(ft_prio)) {
2108 		err = PTR_ERR(ft_prio);
2109 		goto unlock;
2110 	}
2111 	if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2112 		ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
2113 		if (IS_ERR(ft_prio_tx)) {
2114 			err = PTR_ERR(ft_prio_tx);
2115 			ft_prio_tx = NULL;
2116 			goto destroy_ft;
2117 		}
2118 	}
2119 
2120 	dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
2121 	if (mqp->flags & MLX5_IB_QP_RSS)
2122 		dst->tir_num = mqp->rss_qp.tirn;
2123 	else
2124 		dst->tir_num = mqp->raw_packet_qp.rq.tirn;
2125 
2126 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2127 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
2128 			handler = create_dont_trap_rule(dev, ft_prio,
2129 							flow_attr, dst);
2130 		} else {
2131 			handler = create_flow_rule(dev, ft_prio, flow_attr,
2132 						   dst);
2133 		}
2134 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2135 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2136 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
2137 						dst);
2138 	} else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2139 		handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
2140 	} else {
2141 		err = -EINVAL;
2142 		goto destroy_ft;
2143 	}
2144 
2145 	if (IS_ERR(handler)) {
2146 		err = PTR_ERR(handler);
2147 		handler = NULL;
2148 		goto destroy_ft;
2149 	}
2150 
2151 	mutex_unlock(&dev->flow_db.lock);
2152 	kfree(dst);
2153 
2154 	return &handler->ibflow;
2155 
2156 destroy_ft:
2157 	put_flow_table(dev, ft_prio, false);
2158 	if (ft_prio_tx)
2159 		put_flow_table(dev, ft_prio_tx, false);
2160 unlock:
2161 	mutex_unlock(&dev->flow_db.lock);
2162 	kfree(dst);
2163 	kfree(handler);
2164 	return ERR_PTR(err);
2165 }
2166 
mlx5_ib_mcg_attach(struct ib_qp * ibqp,union ib_gid * gid,u16 lid)2167 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2168 {
2169 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2170 	int err;
2171 
2172 	err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
2173 	if (err)
2174 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2175 			     ibqp->qp_num, gid->raw);
2176 
2177 	return err;
2178 }
2179 
mlx5_ib_mcg_detach(struct ib_qp * ibqp,union ib_gid * gid,u16 lid)2180 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2181 {
2182 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2183 	int err;
2184 
2185 	err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
2186 	if (err)
2187 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2188 			     ibqp->qp_num, gid->raw);
2189 
2190 	return err;
2191 }
2192 
init_node_data(struct mlx5_ib_dev * dev)2193 static int init_node_data(struct mlx5_ib_dev *dev)
2194 {
2195 	int err;
2196 
2197 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2198 	if (err)
2199 		return err;
2200 
2201 	dev->mdev->rev_id = dev->mdev->pdev->revision;
2202 
2203 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2204 }
2205 
show_fw_pages(struct device * device,struct device_attribute * attr,char * buf)2206 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
2207 			     char *buf)
2208 {
2209 	struct mlx5_ib_dev *dev =
2210 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2211 
2212 	return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
2213 }
2214 
show_reg_pages(struct device * device,struct device_attribute * attr,char * buf)2215 static ssize_t show_reg_pages(struct device *device,
2216 			      struct device_attribute *attr, char *buf)
2217 {
2218 	struct mlx5_ib_dev *dev =
2219 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2220 
2221 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2222 }
2223 
show_hca(struct device * device,struct device_attribute * attr,char * buf)2224 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2225 			char *buf)
2226 {
2227 	struct mlx5_ib_dev *dev =
2228 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2229 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
2230 }
2231 
show_rev(struct device * device,struct device_attribute * attr,char * buf)2232 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2233 			char *buf)
2234 {
2235 	struct mlx5_ib_dev *dev =
2236 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2237 	return sprintf(buf, "%x\n", dev->mdev->rev_id);
2238 }
2239 
show_board(struct device * device,struct device_attribute * attr,char * buf)2240 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2241 			  char *buf)
2242 {
2243 	struct mlx5_ib_dev *dev =
2244 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2245 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2246 		       dev->mdev->board_id);
2247 }
2248 
2249 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
2250 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
2251 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
2252 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
2253 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
2254 
2255 static struct device_attribute *mlx5_class_attributes[] = {
2256 	&dev_attr_hw_rev,
2257 	&dev_attr_hca_type,
2258 	&dev_attr_board_id,
2259 	&dev_attr_fw_pages,
2260 	&dev_attr_reg_pages,
2261 };
2262 
pkey_change_handler(struct work_struct * work)2263 static void pkey_change_handler(struct work_struct *work)
2264 {
2265 	struct mlx5_ib_port_resources *ports =
2266 		container_of(work, struct mlx5_ib_port_resources,
2267 			     pkey_change_work);
2268 
2269 	mutex_lock(&ports->devr->mutex);
2270 	mlx5_ib_gsi_pkey_change(ports->gsi);
2271 	mutex_unlock(&ports->devr->mutex);
2272 }
2273 
mlx5_ib_handle_internal_error(struct mlx5_ib_dev * ibdev)2274 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2275 {
2276 	struct mlx5_ib_qp *mqp;
2277 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2278 	struct mlx5_core_cq *mcq;
2279 	struct list_head cq_armed_list;
2280 	unsigned long flags_qp;
2281 	unsigned long flags_cq;
2282 	unsigned long flags;
2283 
2284 	INIT_LIST_HEAD(&cq_armed_list);
2285 
2286 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2287 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2288 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2289 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2290 		if (mqp->sq.tail != mqp->sq.head) {
2291 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2292 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2293 			if (send_mcq->mcq.comp &&
2294 			    mqp->ibqp.send_cq->comp_handler) {
2295 				if (!send_mcq->mcq.reset_notify_added) {
2296 					send_mcq->mcq.reset_notify_added = 1;
2297 					list_add_tail(&send_mcq->mcq.reset_notify,
2298 						      &cq_armed_list);
2299 				}
2300 			}
2301 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2302 		}
2303 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2304 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2305 		/* no handling is needed for SRQ */
2306 		if (!mqp->ibqp.srq) {
2307 			if (mqp->rq.tail != mqp->rq.head) {
2308 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2309 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2310 				if (recv_mcq->mcq.comp &&
2311 				    mqp->ibqp.recv_cq->comp_handler) {
2312 					if (!recv_mcq->mcq.reset_notify_added) {
2313 						recv_mcq->mcq.reset_notify_added = 1;
2314 						list_add_tail(&recv_mcq->mcq.reset_notify,
2315 							      &cq_armed_list);
2316 					}
2317 				}
2318 				spin_unlock_irqrestore(&recv_mcq->lock,
2319 						       flags_cq);
2320 			}
2321 		}
2322 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2323 	}
2324 	/*At that point all inflight post send were put to be executed as of we
2325 	 * lock/unlock above locks Now need to arm all involved CQs.
2326 	 */
2327 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2328 		mcq->comp(mcq);
2329 	}
2330 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2331 }
2332 
mlx5_ib_event(struct mlx5_core_dev * dev,void * context,enum mlx5_dev_event event,unsigned long param)2333 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
2334 			  enum mlx5_dev_event event, unsigned long param)
2335 {
2336 	struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
2337 	struct ib_event ibev;
2338 	bool fatal = false;
2339 	u8 port = 0;
2340 
2341 	switch (event) {
2342 	case MLX5_DEV_EVENT_SYS_ERROR:
2343 		ibev.event = IB_EVENT_DEVICE_FATAL;
2344 		mlx5_ib_handle_internal_error(ibdev);
2345 		fatal = true;
2346 		break;
2347 
2348 	case MLX5_DEV_EVENT_PORT_UP:
2349 	case MLX5_DEV_EVENT_PORT_DOWN:
2350 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
2351 		port = (u8)param;
2352 
2353 		/* In RoCE, port up/down events are handled in
2354 		 * mlx5_netdev_event().
2355 		 */
2356 		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2357 			IB_LINK_LAYER_ETHERNET)
2358 			return;
2359 
2360 		ibev.event = (event == MLX5_DEV_EVENT_PORT_UP) ?
2361 			     IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2362 		break;
2363 
2364 	case MLX5_DEV_EVENT_LID_CHANGE:
2365 		ibev.event = IB_EVENT_LID_CHANGE;
2366 		port = (u8)param;
2367 		break;
2368 
2369 	case MLX5_DEV_EVENT_PKEY_CHANGE:
2370 		ibev.event = IB_EVENT_PKEY_CHANGE;
2371 		port = (u8)param;
2372 
2373 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2374 		break;
2375 
2376 	case MLX5_DEV_EVENT_GUID_CHANGE:
2377 		ibev.event = IB_EVENT_GID_CHANGE;
2378 		port = (u8)param;
2379 		break;
2380 
2381 	case MLX5_DEV_EVENT_CLIENT_REREG:
2382 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
2383 		port = (u8)param;
2384 		break;
2385 	}
2386 
2387 	ibev.device	      = &ibdev->ib_dev;
2388 	ibev.element.port_num = port;
2389 
2390 	if (port < 1 || port > ibdev->num_ports) {
2391 		mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
2392 		return;
2393 	}
2394 
2395 	if (ibdev->ib_active)
2396 		ib_dispatch_event(&ibev);
2397 
2398 	if (fatal)
2399 		ibdev->ib_active = false;
2400 }
2401 
get_ext_port_caps(struct mlx5_ib_dev * dev)2402 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2403 {
2404 	int port;
2405 
2406 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
2407 		mlx5_query_ext_port_caps(dev, port);
2408 }
2409 
get_port_caps(struct mlx5_ib_dev * dev)2410 static int get_port_caps(struct mlx5_ib_dev *dev)
2411 {
2412 	struct ib_device_attr *dprops = NULL;
2413 	struct ib_port_attr *pprops = NULL;
2414 	int err = -ENOMEM;
2415 	int port;
2416 	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
2417 
2418 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
2419 	if (!pprops)
2420 		goto out;
2421 
2422 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2423 	if (!dprops)
2424 		goto out;
2425 
2426 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
2427 	if (err) {
2428 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
2429 		goto out;
2430 	}
2431 
2432 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2433 		err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2434 		if (err) {
2435 			mlx5_ib_warn(dev, "query_port %d failed %d\n",
2436 				     port, err);
2437 			break;
2438 		}
2439 		dev->mdev->port_caps[port - 1].pkey_table_len =
2440 						dprops->max_pkeys;
2441 		dev->mdev->port_caps[port - 1].gid_table_len =
2442 						pprops->gid_tbl_len;
2443 		mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
2444 			    dprops->max_pkeys, pprops->gid_tbl_len);
2445 	}
2446 
2447 out:
2448 	kfree(pprops);
2449 	kfree(dprops);
2450 
2451 	return err;
2452 }
2453 
destroy_umrc_res(struct mlx5_ib_dev * dev)2454 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
2455 {
2456 	int err;
2457 
2458 	err = mlx5_mr_cache_cleanup(dev);
2459 	if (err)
2460 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
2461 
2462 	mlx5_ib_destroy_qp(dev->umrc.qp);
2463 	ib_free_cq(dev->umrc.cq);
2464 	ib_dealloc_pd(dev->umrc.pd);
2465 }
2466 
2467 enum {
2468 	MAX_UMR_WR = 128,
2469 };
2470 
create_umr_res(struct mlx5_ib_dev * dev)2471 static int create_umr_res(struct mlx5_ib_dev *dev)
2472 {
2473 	struct ib_qp_init_attr *init_attr = NULL;
2474 	struct ib_qp_attr *attr = NULL;
2475 	struct ib_pd *pd;
2476 	struct ib_cq *cq;
2477 	struct ib_qp *qp;
2478 	int ret;
2479 
2480 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
2481 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
2482 	if (!attr || !init_attr) {
2483 		ret = -ENOMEM;
2484 		goto error_0;
2485 	}
2486 
2487 	pd = ib_alloc_pd(&dev->ib_dev, 0);
2488 	if (IS_ERR(pd)) {
2489 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
2490 		ret = PTR_ERR(pd);
2491 		goto error_0;
2492 	}
2493 
2494 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
2495 	if (IS_ERR(cq)) {
2496 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
2497 		ret = PTR_ERR(cq);
2498 		goto error_2;
2499 	}
2500 
2501 	init_attr->send_cq = cq;
2502 	init_attr->recv_cq = cq;
2503 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
2504 	init_attr->cap.max_send_wr = MAX_UMR_WR;
2505 	init_attr->cap.max_send_sge = 1;
2506 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
2507 	init_attr->port_num = 1;
2508 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
2509 	if (IS_ERR(qp)) {
2510 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
2511 		ret = PTR_ERR(qp);
2512 		goto error_3;
2513 	}
2514 	qp->device     = &dev->ib_dev;
2515 	qp->real_qp    = qp;
2516 	qp->uobject    = NULL;
2517 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
2518 	qp->send_cq    = init_attr->send_cq;
2519 	qp->recv_cq    = init_attr->recv_cq;
2520 
2521 	attr->qp_state = IB_QPS_INIT;
2522 	attr->port_num = 1;
2523 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
2524 				IB_QP_PORT, NULL);
2525 	if (ret) {
2526 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
2527 		goto error_4;
2528 	}
2529 
2530 	memset(attr, 0, sizeof(*attr));
2531 	attr->qp_state = IB_QPS_RTR;
2532 	attr->path_mtu = IB_MTU_256;
2533 
2534 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2535 	if (ret) {
2536 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
2537 		goto error_4;
2538 	}
2539 
2540 	memset(attr, 0, sizeof(*attr));
2541 	attr->qp_state = IB_QPS_RTS;
2542 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2543 	if (ret) {
2544 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
2545 		goto error_4;
2546 	}
2547 
2548 	dev->umrc.qp = qp;
2549 	dev->umrc.cq = cq;
2550 	dev->umrc.pd = pd;
2551 
2552 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
2553 	ret = mlx5_mr_cache_init(dev);
2554 	if (ret) {
2555 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
2556 		goto error_4;
2557 	}
2558 
2559 	kfree(attr);
2560 	kfree(init_attr);
2561 
2562 	return 0;
2563 
2564 error_4:
2565 	mlx5_ib_destroy_qp(qp);
2566 
2567 error_3:
2568 	ib_free_cq(cq);
2569 
2570 error_2:
2571 	ib_dealloc_pd(pd);
2572 
2573 error_0:
2574 	kfree(attr);
2575 	kfree(init_attr);
2576 	return ret;
2577 }
2578 
mlx5_get_umr_fence(u8 umr_fence_cap)2579 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
2580 {
2581 	switch (umr_fence_cap) {
2582 	case MLX5_CAP_UMR_FENCE_NONE:
2583 		return MLX5_FENCE_MODE_NONE;
2584 	case MLX5_CAP_UMR_FENCE_SMALL:
2585 		return MLX5_FENCE_MODE_INITIATOR_SMALL;
2586 	default:
2587 		return MLX5_FENCE_MODE_STRONG_ORDERING;
2588 	}
2589 }
2590 
create_dev_resources(struct mlx5_ib_resources * devr)2591 static int create_dev_resources(struct mlx5_ib_resources *devr)
2592 {
2593 	struct ib_srq_init_attr attr;
2594 	struct mlx5_ib_dev *dev;
2595 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2596 	int port;
2597 	int ret = 0;
2598 
2599 	dev = container_of(devr, struct mlx5_ib_dev, devr);
2600 
2601 	mutex_init(&devr->mutex);
2602 
2603 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
2604 	if (IS_ERR(devr->p0)) {
2605 		ret = PTR_ERR(devr->p0);
2606 		goto error0;
2607 	}
2608 	devr->p0->device  = &dev->ib_dev;
2609 	devr->p0->uobject = NULL;
2610 	atomic_set(&devr->p0->usecnt, 0);
2611 
2612 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
2613 	if (IS_ERR(devr->c0)) {
2614 		ret = PTR_ERR(devr->c0);
2615 		goto error1;
2616 	}
2617 	devr->c0->device        = &dev->ib_dev;
2618 	devr->c0->uobject       = NULL;
2619 	devr->c0->comp_handler  = NULL;
2620 	devr->c0->event_handler = NULL;
2621 	devr->c0->cq_context    = NULL;
2622 	atomic_set(&devr->c0->usecnt, 0);
2623 
2624 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2625 	if (IS_ERR(devr->x0)) {
2626 		ret = PTR_ERR(devr->x0);
2627 		goto error2;
2628 	}
2629 	devr->x0->device = &dev->ib_dev;
2630 	devr->x0->inode = NULL;
2631 	atomic_set(&devr->x0->usecnt, 0);
2632 	mutex_init(&devr->x0->tgt_qp_mutex);
2633 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
2634 
2635 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2636 	if (IS_ERR(devr->x1)) {
2637 		ret = PTR_ERR(devr->x1);
2638 		goto error3;
2639 	}
2640 	devr->x1->device = &dev->ib_dev;
2641 	devr->x1->inode = NULL;
2642 	atomic_set(&devr->x1->usecnt, 0);
2643 	mutex_init(&devr->x1->tgt_qp_mutex);
2644 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
2645 
2646 	memset(&attr, 0, sizeof(attr));
2647 	attr.attr.max_sge = 1;
2648 	attr.attr.max_wr = 1;
2649 	attr.srq_type = IB_SRQT_XRC;
2650 	attr.ext.xrc.cq = devr->c0;
2651 	attr.ext.xrc.xrcd = devr->x0;
2652 
2653 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2654 	if (IS_ERR(devr->s0)) {
2655 		ret = PTR_ERR(devr->s0);
2656 		goto error4;
2657 	}
2658 	devr->s0->device	= &dev->ib_dev;
2659 	devr->s0->pd		= devr->p0;
2660 	devr->s0->uobject       = NULL;
2661 	devr->s0->event_handler = NULL;
2662 	devr->s0->srq_context   = NULL;
2663 	devr->s0->srq_type      = IB_SRQT_XRC;
2664 	devr->s0->ext.xrc.xrcd	= devr->x0;
2665 	devr->s0->ext.xrc.cq	= devr->c0;
2666 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
2667 	atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
2668 	atomic_inc(&devr->p0->usecnt);
2669 	atomic_set(&devr->s0->usecnt, 0);
2670 
2671 	memset(&attr, 0, sizeof(attr));
2672 	attr.attr.max_sge = 1;
2673 	attr.attr.max_wr = 1;
2674 	attr.srq_type = IB_SRQT_BASIC;
2675 	devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2676 	if (IS_ERR(devr->s1)) {
2677 		ret = PTR_ERR(devr->s1);
2678 		goto error5;
2679 	}
2680 	devr->s1->device	= &dev->ib_dev;
2681 	devr->s1->pd		= devr->p0;
2682 	devr->s1->uobject       = NULL;
2683 	devr->s1->event_handler = NULL;
2684 	devr->s1->srq_context   = NULL;
2685 	devr->s1->srq_type      = IB_SRQT_BASIC;
2686 	devr->s1->ext.xrc.cq	= devr->c0;
2687 	atomic_inc(&devr->p0->usecnt);
2688 	atomic_set(&devr->s0->usecnt, 0);
2689 
2690 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
2691 		INIT_WORK(&devr->ports[port].pkey_change_work,
2692 			  pkey_change_handler);
2693 		devr->ports[port].devr = devr;
2694 	}
2695 
2696 	return 0;
2697 
2698 error5:
2699 	mlx5_ib_destroy_srq(devr->s0);
2700 error4:
2701 	mlx5_ib_dealloc_xrcd(devr->x1);
2702 error3:
2703 	mlx5_ib_dealloc_xrcd(devr->x0);
2704 error2:
2705 	mlx5_ib_destroy_cq(devr->c0);
2706 error1:
2707 	mlx5_ib_dealloc_pd(devr->p0);
2708 error0:
2709 	return ret;
2710 }
2711 
destroy_dev_resources(struct mlx5_ib_resources * devr)2712 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
2713 {
2714 	struct mlx5_ib_dev *dev =
2715 		container_of(devr, struct mlx5_ib_dev, devr);
2716 	int port;
2717 
2718 	mlx5_ib_destroy_srq(devr->s1);
2719 	mlx5_ib_destroy_srq(devr->s0);
2720 	mlx5_ib_dealloc_xrcd(devr->x0);
2721 	mlx5_ib_dealloc_xrcd(devr->x1);
2722 	mlx5_ib_destroy_cq(devr->c0);
2723 	mlx5_ib_dealloc_pd(devr->p0);
2724 
2725 	/* Make sure no change P_Key work items are still executing */
2726 	for (port = 0; port < dev->num_ports; ++port)
2727 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2728 }
2729 
get_core_cap_flags(struct ib_device * ibdev)2730 static u32 get_core_cap_flags(struct ib_device *ibdev)
2731 {
2732 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2733 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2734 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2735 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2736 	u32 ret = 0;
2737 
2738 	if (ll == IB_LINK_LAYER_INFINIBAND)
2739 		return RDMA_CORE_PORT_IBA_IB;
2740 
2741 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2742 		return 0;
2743 
2744 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2745 		return 0;
2746 
2747 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2748 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2749 
2750 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2751 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2752 
2753 	return ret;
2754 }
2755 
mlx5_port_immutable(struct ib_device * ibdev,u8 port_num,struct ib_port_immutable * immutable)2756 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
2757 			       struct ib_port_immutable *immutable)
2758 {
2759 	struct ib_port_attr attr;
2760 	int err;
2761 
2762 	err = mlx5_ib_query_port(ibdev, port_num, &attr);
2763 	if (err)
2764 		return err;
2765 
2766 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2767 	immutable->gid_tbl_len = attr.gid_tbl_len;
2768 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
2769 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2770 
2771 	return 0;
2772 }
2773 
get_dev_fw_str(struct ib_device * ibdev,char * str,size_t str_len)2774 static void get_dev_fw_str(struct ib_device *ibdev, char *str,
2775 			   size_t str_len)
2776 {
2777 	struct mlx5_ib_dev *dev =
2778 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2779 	snprintf(str, str_len, "%d.%d.%04d", fw_rev_maj(dev->mdev),
2780 		       fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev));
2781 }
2782 
mlx5_roce_lag_init(struct mlx5_ib_dev * dev)2783 static int mlx5_roce_lag_init(struct mlx5_ib_dev *dev)
2784 {
2785 	struct mlx5_core_dev *mdev = dev->mdev;
2786 	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
2787 								 MLX5_FLOW_NAMESPACE_LAG);
2788 	struct mlx5_flow_table *ft;
2789 	int err;
2790 
2791 	if (!ns || !mlx5_lag_is_active(mdev))
2792 		return 0;
2793 
2794 	err = mlx5_cmd_create_vport_lag(mdev);
2795 	if (err)
2796 		return err;
2797 
2798 	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
2799 	if (IS_ERR(ft)) {
2800 		err = PTR_ERR(ft);
2801 		goto err_destroy_vport_lag;
2802 	}
2803 
2804 	dev->flow_db.lag_demux_ft = ft;
2805 	return 0;
2806 
2807 err_destroy_vport_lag:
2808 	mlx5_cmd_destroy_vport_lag(mdev);
2809 	return err;
2810 }
2811 
mlx5_roce_lag_cleanup(struct mlx5_ib_dev * dev)2812 static void mlx5_roce_lag_cleanup(struct mlx5_ib_dev *dev)
2813 {
2814 	struct mlx5_core_dev *mdev = dev->mdev;
2815 
2816 	if (dev->flow_db.lag_demux_ft) {
2817 		mlx5_destroy_flow_table(dev->flow_db.lag_demux_ft);
2818 		dev->flow_db.lag_demux_ft = NULL;
2819 
2820 		mlx5_cmd_destroy_vport_lag(mdev);
2821 	}
2822 }
2823 
mlx5_remove_roce_notifier(struct mlx5_ib_dev * dev)2824 static void mlx5_remove_roce_notifier(struct mlx5_ib_dev *dev)
2825 {
2826 	if (dev->roce.nb.notifier_call) {
2827 		unregister_netdevice_notifier(&dev->roce.nb);
2828 		dev->roce.nb.notifier_call = NULL;
2829 	}
2830 }
2831 
mlx5_enable_roce(struct mlx5_ib_dev * dev)2832 static int mlx5_enable_roce(struct mlx5_ib_dev *dev)
2833 {
2834 	int err;
2835 
2836 	dev->roce.nb.notifier_call = mlx5_netdev_event;
2837 	err = register_netdevice_notifier(&dev->roce.nb);
2838 	if (err) {
2839 		dev->roce.nb.notifier_call = NULL;
2840 		return err;
2841 	}
2842 
2843 	err = mlx5_nic_vport_enable_roce(dev->mdev);
2844 	if (err)
2845 		goto err_unregister_netdevice_notifier;
2846 
2847 	err = mlx5_roce_lag_init(dev);
2848 	if (err)
2849 		goto err_disable_roce;
2850 
2851 	return 0;
2852 
2853 err_disable_roce:
2854 	mlx5_nic_vport_disable_roce(dev->mdev);
2855 
2856 err_unregister_netdevice_notifier:
2857 	mlx5_remove_roce_notifier(dev);
2858 	return err;
2859 }
2860 
mlx5_disable_roce(struct mlx5_ib_dev * dev)2861 static void mlx5_disable_roce(struct mlx5_ib_dev *dev)
2862 {
2863 	mlx5_roce_lag_cleanup(dev);
2864 	mlx5_nic_vport_disable_roce(dev->mdev);
2865 }
2866 
mlx5_ib_dealloc_q_counters(struct mlx5_ib_dev * dev)2867 static void mlx5_ib_dealloc_q_counters(struct mlx5_ib_dev *dev)
2868 {
2869 	unsigned int i;
2870 
2871 	for (i = 0; i < dev->num_ports; i++)
2872 		mlx5_core_dealloc_q_counter(dev->mdev,
2873 					    dev->port[i].q_cnt_id);
2874 }
2875 
mlx5_ib_alloc_q_counters(struct mlx5_ib_dev * dev)2876 static int mlx5_ib_alloc_q_counters(struct mlx5_ib_dev *dev)
2877 {
2878 	int i;
2879 	int ret;
2880 
2881 	for (i = 0; i < dev->num_ports; i++) {
2882 		ret = mlx5_core_alloc_q_counter(dev->mdev,
2883 						&dev->port[i].q_cnt_id);
2884 		if (ret) {
2885 			mlx5_ib_warn(dev,
2886 				     "couldn't allocate queue counter for port %d, err %d\n",
2887 				     i + 1, ret);
2888 			goto dealloc_counters;
2889 		}
2890 	}
2891 
2892 	return 0;
2893 
2894 dealloc_counters:
2895 	while (--i >= 0)
2896 		mlx5_core_dealloc_q_counter(dev->mdev,
2897 					    dev->port[i].q_cnt_id);
2898 
2899 	return ret;
2900 }
2901 
2902 static const char * const names[] = {
2903 	"rx_write_requests",
2904 	"rx_read_requests",
2905 	"rx_atomic_requests",
2906 	"out_of_buffer",
2907 	"out_of_sequence",
2908 	"duplicate_request",
2909 	"rnr_nak_retry_err",
2910 	"packet_seq_err",
2911 	"implied_nak_seq_err",
2912 	"local_ack_timeout_err",
2913 };
2914 
2915 static const size_t stats_offsets[] = {
2916 	MLX5_BYTE_OFF(query_q_counter_out, rx_write_requests),
2917 	MLX5_BYTE_OFF(query_q_counter_out, rx_read_requests),
2918 	MLX5_BYTE_OFF(query_q_counter_out, rx_atomic_requests),
2919 	MLX5_BYTE_OFF(query_q_counter_out, out_of_buffer),
2920 	MLX5_BYTE_OFF(query_q_counter_out, out_of_sequence),
2921 	MLX5_BYTE_OFF(query_q_counter_out, duplicate_request),
2922 	MLX5_BYTE_OFF(query_q_counter_out, rnr_nak_retry_err),
2923 	MLX5_BYTE_OFF(query_q_counter_out, packet_seq_err),
2924 	MLX5_BYTE_OFF(query_q_counter_out, implied_nak_seq_err),
2925 	MLX5_BYTE_OFF(query_q_counter_out, local_ack_timeout_err),
2926 };
2927 
mlx5_ib_alloc_hw_stats(struct ib_device * ibdev,u8 port_num)2928 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
2929 						    u8 port_num)
2930 {
2931 	BUILD_BUG_ON(ARRAY_SIZE(names) != ARRAY_SIZE(stats_offsets));
2932 
2933 	/* We support only per port stats */
2934 	if (port_num == 0)
2935 		return NULL;
2936 
2937 	return rdma_alloc_hw_stats_struct(names, ARRAY_SIZE(names),
2938 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
2939 }
2940 
mlx5_ib_get_hw_stats(struct ib_device * ibdev,struct rdma_hw_stats * stats,u8 port,int index)2941 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
2942 				struct rdma_hw_stats *stats,
2943 				u8 port, int index)
2944 {
2945 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2946 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
2947 	void *out;
2948 	__be32 val;
2949 	int ret;
2950 	int i;
2951 
2952 	if (!port || !stats)
2953 		return -ENOSYS;
2954 
2955 	out = mlx5_vzalloc(outlen);
2956 	if (!out)
2957 		return -ENOMEM;
2958 
2959 	ret = mlx5_core_query_q_counter(dev->mdev,
2960 					dev->port[port - 1].q_cnt_id, 0,
2961 					out, outlen);
2962 	if (ret)
2963 		goto free;
2964 
2965 	for (i = 0; i < ARRAY_SIZE(names); i++) {
2966 		val = *(__be32 *)(out + stats_offsets[i]);
2967 		stats->value[i] = (u64)be32_to_cpu(val);
2968 	}
2969 free:
2970 	kvfree(out);
2971 	return ARRAY_SIZE(names);
2972 }
2973 
mlx5_ib_add(struct mlx5_core_dev * mdev)2974 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
2975 {
2976 	struct mlx5_ib_dev *dev;
2977 	enum rdma_link_layer ll;
2978 	int port_type_cap;
2979 	const char *name;
2980 	int err;
2981 	int i;
2982 
2983 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
2984 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
2985 
2986 	if ((ll == IB_LINK_LAYER_ETHERNET) && !MLX5_CAP_GEN(mdev, roce))
2987 		return NULL;
2988 
2989 	printk_once(KERN_INFO "%s", mlx5_version);
2990 
2991 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
2992 	if (!dev)
2993 		return NULL;
2994 
2995 	dev->mdev = mdev;
2996 
2997 	dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port),
2998 			    GFP_KERNEL);
2999 	if (!dev->port)
3000 		goto err_dealloc;
3001 
3002 	rwlock_init(&dev->roce.netdev_lock);
3003 	err = get_port_caps(dev);
3004 	if (err)
3005 		goto err_free_port;
3006 
3007 	if (mlx5_use_mad_ifc(dev))
3008 		get_ext_port_caps(dev);
3009 
3010 	MLX5_INIT_DOORBELL_LOCK(&dev->uar_lock);
3011 
3012 	if (!mlx5_lag_is_active(mdev))
3013 		name = "mlx5_%d";
3014 	else
3015 		name = "mlx5_bond_%d";
3016 
3017 	strlcpy(dev->ib_dev.name, name, IB_DEVICE_NAME_MAX);
3018 	dev->ib_dev.owner		= THIS_MODULE;
3019 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
3020 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
3021 	dev->num_ports		= MLX5_CAP_GEN(mdev, num_ports);
3022 	dev->ib_dev.phys_port_cnt     = dev->num_ports;
3023 	dev->ib_dev.num_comp_vectors    =
3024 		dev->mdev->priv.eq_table.num_comp_vectors;
3025 	dev->ib_dev.dma_device	= &mdev->pdev->dev;
3026 
3027 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
3028 	dev->ib_dev.uverbs_cmd_mask	=
3029 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
3030 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
3031 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
3032 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
3033 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
3034 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
3035 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
3036 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
3037 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
3038 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
3039 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
3040 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
3041 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
3042 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
3043 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
3044 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
3045 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
3046 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
3047 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
3048 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
3049 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
3050 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
3051 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
3052 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
3053 	dev->ib_dev.uverbs_ex_cmd_mask =
3054 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
3055 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
3056 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP);
3057 
3058 	dev->ib_dev.query_device	= mlx5_ib_query_device;
3059 	dev->ib_dev.query_port		= mlx5_ib_query_port;
3060 	dev->ib_dev.get_link_layer	= mlx5_ib_port_link_layer;
3061 	if (ll == IB_LINK_LAYER_ETHERNET)
3062 		dev->ib_dev.get_netdev	= mlx5_ib_get_netdev;
3063 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
3064 	dev->ib_dev.add_gid		= mlx5_ib_add_gid;
3065 	dev->ib_dev.del_gid		= mlx5_ib_del_gid;
3066 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
3067 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
3068 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
3069 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
3070 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
3071 	dev->ib_dev.mmap		= mlx5_ib_mmap;
3072 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
3073 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
3074 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
3075 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
3076 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
3077 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
3078 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
3079 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
3080 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
3081 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
3082 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
3083 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
3084 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
3085 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
3086 	dev->ib_dev.post_send		= mlx5_ib_post_send;
3087 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
3088 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
3089 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
3090 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
3091 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
3092 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
3093 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
3094 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
3095 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
3096 	dev->ib_dev.rereg_user_mr	= mlx5_ib_rereg_user_mr;
3097 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
3098 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
3099 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
3100 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
3101 	dev->ib_dev.alloc_mr		= mlx5_ib_alloc_mr;
3102 	dev->ib_dev.map_mr_sg		= mlx5_ib_map_mr_sg;
3103 	dev->ib_dev.check_mr_status	= mlx5_ib_check_mr_status;
3104 	dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
3105 	dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
3106 	if (mlx5_core_is_pf(mdev)) {
3107 		dev->ib_dev.get_vf_config	= mlx5_ib_get_vf_config;
3108 		dev->ib_dev.set_vf_link_state	= mlx5_ib_set_vf_link_state;
3109 		dev->ib_dev.get_vf_stats	= mlx5_ib_get_vf_stats;
3110 		dev->ib_dev.set_vf_guid		= mlx5_ib_set_vf_guid;
3111 	}
3112 
3113 	dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
3114 
3115 	mlx5_ib_internal_fill_odp_caps(dev);
3116 
3117 	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
3118 
3119 	if (MLX5_CAP_GEN(mdev, imaicl)) {
3120 		dev->ib_dev.alloc_mw		= mlx5_ib_alloc_mw;
3121 		dev->ib_dev.dealloc_mw		= mlx5_ib_dealloc_mw;
3122 		dev->ib_dev.uverbs_cmd_mask |=
3123 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
3124 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
3125 	}
3126 
3127 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt) &&
3128 	    MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
3129 		dev->ib_dev.get_hw_stats	= mlx5_ib_get_hw_stats;
3130 		dev->ib_dev.alloc_hw_stats	= mlx5_ib_alloc_hw_stats;
3131 	}
3132 
3133 	if (MLX5_CAP_GEN(mdev, xrc)) {
3134 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
3135 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
3136 		dev->ib_dev.uverbs_cmd_mask |=
3137 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
3138 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
3139 	}
3140 
3141 	if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
3142 	    IB_LINK_LAYER_ETHERNET) {
3143 		dev->ib_dev.create_flow	= mlx5_ib_create_flow;
3144 		dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
3145 		dev->ib_dev.create_wq	 = mlx5_ib_create_wq;
3146 		dev->ib_dev.modify_wq	 = mlx5_ib_modify_wq;
3147 		dev->ib_dev.destroy_wq	 = mlx5_ib_destroy_wq;
3148 		dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
3149 		dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
3150 		dev->ib_dev.uverbs_ex_cmd_mask |=
3151 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
3152 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW) |
3153 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
3154 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
3155 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
3156 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
3157 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
3158 	}
3159 	err = init_node_data(dev);
3160 	if (err)
3161 		goto err_free_port;
3162 
3163 	mutex_init(&dev->flow_db.lock);
3164 	mutex_init(&dev->cap_mask_mutex);
3165 	INIT_LIST_HEAD(&dev->qp_list);
3166 	spin_lock_init(&dev->reset_flow_resource_lock);
3167 
3168 	if (ll == IB_LINK_LAYER_ETHERNET) {
3169 		err = mlx5_enable_roce(dev);
3170 		if (err)
3171 			goto err_free_port;
3172 	}
3173 
3174 	err = create_dev_resources(&dev->devr);
3175 	if (err)
3176 		goto err_disable_roce;
3177 
3178 	err = mlx5_ib_odp_init_one(dev);
3179 	if (err)
3180 		goto err_rsrc;
3181 
3182 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
3183 		err = mlx5_ib_alloc_q_counters(dev);
3184 		if (err)
3185 			goto err_odp;
3186 	}
3187 
3188 	err = ib_register_device(&dev->ib_dev, NULL);
3189 	if (err)
3190 		goto err_q_cnt;
3191 
3192 	err = create_umr_res(dev);
3193 	if (err)
3194 		goto err_dev;
3195 
3196 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
3197 		err = device_create_file(&dev->ib_dev.dev,
3198 					 mlx5_class_attributes[i]);
3199 		if (err)
3200 			goto err_umrc;
3201 	}
3202 
3203 	dev->ib_active = true;
3204 
3205 	return dev;
3206 
3207 err_umrc:
3208 	destroy_umrc_res(dev);
3209 
3210 err_dev:
3211 	ib_unregister_device(&dev->ib_dev);
3212 
3213 err_q_cnt:
3214 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
3215 		mlx5_ib_dealloc_q_counters(dev);
3216 
3217 err_odp:
3218 	mlx5_ib_odp_remove_one(dev);
3219 
3220 err_rsrc:
3221 	destroy_dev_resources(&dev->devr);
3222 
3223 err_disable_roce:
3224 	if (ll == IB_LINK_LAYER_ETHERNET) {
3225 		mlx5_disable_roce(dev);
3226 		mlx5_remove_roce_notifier(dev);
3227 	}
3228 
3229 err_free_port:
3230 	kfree(dev->port);
3231 
3232 err_dealloc:
3233 	ib_dealloc_device((struct ib_device *)dev);
3234 
3235 	return NULL;
3236 }
3237 
mlx5_ib_remove(struct mlx5_core_dev * mdev,void * context)3238 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
3239 {
3240 	struct mlx5_ib_dev *dev = context;
3241 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
3242 
3243 	mlx5_remove_roce_notifier(dev);
3244 	ib_unregister_device(&dev->ib_dev);
3245 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
3246 		mlx5_ib_dealloc_q_counters(dev);
3247 	destroy_umrc_res(dev);
3248 	mlx5_ib_odp_remove_one(dev);
3249 	destroy_dev_resources(&dev->devr);
3250 	if (ll == IB_LINK_LAYER_ETHERNET)
3251 		mlx5_disable_roce(dev);
3252 	kfree(dev->port);
3253 	ib_dealloc_device(&dev->ib_dev);
3254 }
3255 
3256 static struct mlx5_interface mlx5_ib_interface = {
3257 	.add            = mlx5_ib_add,
3258 	.remove         = mlx5_ib_remove,
3259 	.event          = mlx5_ib_event,
3260 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
3261 };
3262 
mlx5_ib_init(void)3263 static int __init mlx5_ib_init(void)
3264 {
3265 	int err;
3266 
3267 	if (deprecated_prof_sel != 2)
3268 		pr_warn("prof_sel is deprecated for mlx5_ib, set it for mlx5_core\n");
3269 
3270 	err = mlx5_ib_odp_init();
3271 	if (err)
3272 		return err;
3273 
3274 	err = mlx5_register_interface(&mlx5_ib_interface);
3275 	if (err)
3276 		goto clean_odp;
3277 
3278 	return err;
3279 
3280 clean_odp:
3281 	mlx5_ib_odp_cleanup();
3282 	return err;
3283 }
3284 
mlx5_ib_cleanup(void)3285 static void __exit mlx5_ib_cleanup(void)
3286 {
3287 	mlx5_unregister_interface(&mlx5_ib_interface);
3288 	mlx5_ib_odp_cleanup();
3289 }
3290 
3291 module_init(mlx5_ib_init);
3292 module_exit(mlx5_ib_cleanup);
3293