xref: /linux/drivers/infiniband/hw/mlx5/main.c (revision a4cc96d1f0170b779c32c6b2cc58764f5d2cdef0)
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 "user.h"
57 #include "mlx5_ib.h"
58 
59 #define DRIVER_NAME "mlx5_ib"
60 #define DRIVER_VERSION "2.2-1"
61 #define DRIVER_RELDATE	"Feb 2014"
62 
63 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
64 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
65 MODULE_LICENSE("Dual BSD/GPL");
66 MODULE_VERSION(DRIVER_VERSION);
67 
68 static int deprecated_prof_sel = 2;
69 module_param_named(prof_sel, deprecated_prof_sel, int, 0444);
70 MODULE_PARM_DESC(prof_sel, "profile selector. Deprecated here. Moved to module mlx5_core");
71 
72 static char mlx5_version[] =
73 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
74 	DRIVER_VERSION " (" DRIVER_RELDATE ")\n";
75 
76 enum {
77 	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
78 };
79 
80 static enum rdma_link_layer
81 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
82 {
83 	switch (port_type_cap) {
84 	case MLX5_CAP_PORT_TYPE_IB:
85 		return IB_LINK_LAYER_INFINIBAND;
86 	case MLX5_CAP_PORT_TYPE_ETH:
87 		return IB_LINK_LAYER_ETHERNET;
88 	default:
89 		return IB_LINK_LAYER_UNSPECIFIED;
90 	}
91 }
92 
93 static enum rdma_link_layer
94 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
95 {
96 	struct mlx5_ib_dev *dev = to_mdev(device);
97 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
98 
99 	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
100 }
101 
102 static int mlx5_netdev_event(struct notifier_block *this,
103 			     unsigned long event, void *ptr)
104 {
105 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
106 	struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
107 						 roce.nb);
108 
109 	if ((event != NETDEV_UNREGISTER) && (event != NETDEV_REGISTER))
110 		return NOTIFY_DONE;
111 
112 	write_lock(&ibdev->roce.netdev_lock);
113 	if (ndev->dev.parent == &ibdev->mdev->pdev->dev)
114 		ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ? NULL : ndev;
115 	write_unlock(&ibdev->roce.netdev_lock);
116 
117 	return NOTIFY_DONE;
118 }
119 
120 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
121 					     u8 port_num)
122 {
123 	struct mlx5_ib_dev *ibdev = to_mdev(device);
124 	struct net_device *ndev;
125 
126 	/* Ensure ndev does not disappear before we invoke dev_hold()
127 	 */
128 	read_lock(&ibdev->roce.netdev_lock);
129 	ndev = ibdev->roce.netdev;
130 	if (ndev)
131 		dev_hold(ndev);
132 	read_unlock(&ibdev->roce.netdev_lock);
133 
134 	return ndev;
135 }
136 
137 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
138 				struct ib_port_attr *props)
139 {
140 	struct mlx5_ib_dev *dev = to_mdev(device);
141 	struct net_device *ndev;
142 	enum ib_mtu ndev_ib_mtu;
143 	u16 qkey_viol_cntr;
144 
145 	memset(props, 0, sizeof(*props));
146 
147 	props->port_cap_flags  |= IB_PORT_CM_SUP;
148 	props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
149 
150 	props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
151 						roce_address_table_size);
152 	props->max_mtu          = IB_MTU_4096;
153 	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
154 	props->pkey_tbl_len     = 1;
155 	props->state            = IB_PORT_DOWN;
156 	props->phys_state       = 3;
157 
158 	mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr);
159 	props->qkey_viol_cntr = qkey_viol_cntr;
160 
161 	ndev = mlx5_ib_get_netdev(device, port_num);
162 	if (!ndev)
163 		return 0;
164 
165 	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
166 		props->state      = IB_PORT_ACTIVE;
167 		props->phys_state = 5;
168 	}
169 
170 	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
171 
172 	dev_put(ndev);
173 
174 	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
175 
176 	props->active_width	= IB_WIDTH_4X;  /* TODO */
177 	props->active_speed	= IB_SPEED_QDR; /* TODO */
178 
179 	return 0;
180 }
181 
182 static void ib_gid_to_mlx5_roce_addr(const union ib_gid *gid,
183 				     const struct ib_gid_attr *attr,
184 				     void *mlx5_addr)
185 {
186 #define MLX5_SET_RA(p, f, v) MLX5_SET(roce_addr_layout, p, f, v)
187 	char *mlx5_addr_l3_addr	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
188 					       source_l3_address);
189 	void *mlx5_addr_mac	= MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
190 					       source_mac_47_32);
191 
192 	if (!gid)
193 		return;
194 
195 	ether_addr_copy(mlx5_addr_mac, attr->ndev->dev_addr);
196 
197 	if (is_vlan_dev(attr->ndev)) {
198 		MLX5_SET_RA(mlx5_addr, vlan_valid, 1);
199 		MLX5_SET_RA(mlx5_addr, vlan_id, vlan_dev_vlan_id(attr->ndev));
200 	}
201 
202 	switch (attr->gid_type) {
203 	case IB_GID_TYPE_IB:
204 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_1);
205 		break;
206 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
207 		MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_2);
208 		break;
209 
210 	default:
211 		WARN_ON(true);
212 	}
213 
214 	if (attr->gid_type != IB_GID_TYPE_IB) {
215 		if (ipv6_addr_v4mapped((void *)gid))
216 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
217 				    MLX5_ROCE_L3_TYPE_IPV4);
218 		else
219 			MLX5_SET_RA(mlx5_addr, roce_l3_type,
220 				    MLX5_ROCE_L3_TYPE_IPV6);
221 	}
222 
223 	if ((attr->gid_type == IB_GID_TYPE_IB) ||
224 	    !ipv6_addr_v4mapped((void *)gid))
225 		memcpy(mlx5_addr_l3_addr, gid, sizeof(*gid));
226 	else
227 		memcpy(&mlx5_addr_l3_addr[12], &gid->raw[12], 4);
228 }
229 
230 static int set_roce_addr(struct ib_device *device, u8 port_num,
231 			 unsigned int index,
232 			 const union ib_gid *gid,
233 			 const struct ib_gid_attr *attr)
234 {
235 	struct mlx5_ib_dev *dev = to_mdev(device);
236 	u32  in[MLX5_ST_SZ_DW(set_roce_address_in)]  = {0};
237 	u32 out[MLX5_ST_SZ_DW(set_roce_address_out)] = {0};
238 	void *in_addr = MLX5_ADDR_OF(set_roce_address_in, in, roce_address);
239 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(device, port_num);
240 
241 	if (ll != IB_LINK_LAYER_ETHERNET)
242 		return -EINVAL;
243 
244 	ib_gid_to_mlx5_roce_addr(gid, attr, in_addr);
245 
246 	MLX5_SET(set_roce_address_in, in, roce_address_index, index);
247 	MLX5_SET(set_roce_address_in, in, opcode, MLX5_CMD_OP_SET_ROCE_ADDRESS);
248 	return mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
249 }
250 
251 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
252 			   unsigned int index, const union ib_gid *gid,
253 			   const struct ib_gid_attr *attr,
254 			   __always_unused void **context)
255 {
256 	return set_roce_addr(device, port_num, index, gid, attr);
257 }
258 
259 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
260 			   unsigned int index, __always_unused void **context)
261 {
262 	return set_roce_addr(device, port_num, index, NULL, NULL);
263 }
264 
265 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
266 			       int index)
267 {
268 	struct ib_gid_attr attr;
269 	union ib_gid gid;
270 
271 	if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
272 		return 0;
273 
274 	if (!attr.ndev)
275 		return 0;
276 
277 	dev_put(attr.ndev);
278 
279 	if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
280 		return 0;
281 
282 	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
283 }
284 
285 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
286 {
287 	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
288 		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
289 	return 0;
290 }
291 
292 enum {
293 	MLX5_VPORT_ACCESS_METHOD_MAD,
294 	MLX5_VPORT_ACCESS_METHOD_HCA,
295 	MLX5_VPORT_ACCESS_METHOD_NIC,
296 };
297 
298 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
299 {
300 	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
301 		return MLX5_VPORT_ACCESS_METHOD_MAD;
302 
303 	if (mlx5_ib_port_link_layer(ibdev, 1) ==
304 	    IB_LINK_LAYER_ETHERNET)
305 		return MLX5_VPORT_ACCESS_METHOD_NIC;
306 
307 	return MLX5_VPORT_ACCESS_METHOD_HCA;
308 }
309 
310 static void get_atomic_caps(struct mlx5_ib_dev *dev,
311 			    struct ib_device_attr *props)
312 {
313 	u8 tmp;
314 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
315 	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
316 	u8 atomic_req_8B_endianness_mode =
317 		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianess_mode);
318 
319 	/* Check if HW supports 8 bytes standard atomic operations and capable
320 	 * of host endianness respond
321 	 */
322 	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
323 	if (((atomic_operations & tmp) == tmp) &&
324 	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
325 	    (atomic_req_8B_endianness_mode)) {
326 		props->atomic_cap = IB_ATOMIC_HCA;
327 	} else {
328 		props->atomic_cap = IB_ATOMIC_NONE;
329 	}
330 }
331 
332 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
333 					__be64 *sys_image_guid)
334 {
335 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
336 	struct mlx5_core_dev *mdev = dev->mdev;
337 	u64 tmp;
338 	int err;
339 
340 	switch (mlx5_get_vport_access_method(ibdev)) {
341 	case MLX5_VPORT_ACCESS_METHOD_MAD:
342 		return mlx5_query_mad_ifc_system_image_guid(ibdev,
343 							    sys_image_guid);
344 
345 	case MLX5_VPORT_ACCESS_METHOD_HCA:
346 		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
347 		break;
348 
349 	case MLX5_VPORT_ACCESS_METHOD_NIC:
350 		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
351 		break;
352 
353 	default:
354 		return -EINVAL;
355 	}
356 
357 	if (!err)
358 		*sys_image_guid = cpu_to_be64(tmp);
359 
360 	return err;
361 
362 }
363 
364 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
365 				u16 *max_pkeys)
366 {
367 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
368 	struct mlx5_core_dev *mdev = dev->mdev;
369 
370 	switch (mlx5_get_vport_access_method(ibdev)) {
371 	case MLX5_VPORT_ACCESS_METHOD_MAD:
372 		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
373 
374 	case MLX5_VPORT_ACCESS_METHOD_HCA:
375 	case MLX5_VPORT_ACCESS_METHOD_NIC:
376 		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
377 						pkey_table_size));
378 		return 0;
379 
380 	default:
381 		return -EINVAL;
382 	}
383 }
384 
385 static int mlx5_query_vendor_id(struct ib_device *ibdev,
386 				u32 *vendor_id)
387 {
388 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
389 
390 	switch (mlx5_get_vport_access_method(ibdev)) {
391 	case MLX5_VPORT_ACCESS_METHOD_MAD:
392 		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
393 
394 	case MLX5_VPORT_ACCESS_METHOD_HCA:
395 	case MLX5_VPORT_ACCESS_METHOD_NIC:
396 		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
397 
398 	default:
399 		return -EINVAL;
400 	}
401 }
402 
403 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
404 				__be64 *node_guid)
405 {
406 	u64 tmp;
407 	int err;
408 
409 	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
410 	case MLX5_VPORT_ACCESS_METHOD_MAD:
411 		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
412 
413 	case MLX5_VPORT_ACCESS_METHOD_HCA:
414 		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
415 		break;
416 
417 	case MLX5_VPORT_ACCESS_METHOD_NIC:
418 		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
419 		break;
420 
421 	default:
422 		return -EINVAL;
423 	}
424 
425 	if (!err)
426 		*node_guid = cpu_to_be64(tmp);
427 
428 	return err;
429 }
430 
431 struct mlx5_reg_node_desc {
432 	u8	desc[64];
433 };
434 
435 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
436 {
437 	struct mlx5_reg_node_desc in;
438 
439 	if (mlx5_use_mad_ifc(dev))
440 		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
441 
442 	memset(&in, 0, sizeof(in));
443 
444 	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
445 				    sizeof(struct mlx5_reg_node_desc),
446 				    MLX5_REG_NODE_DESC, 0, 0);
447 }
448 
449 static int mlx5_ib_query_device(struct ib_device *ibdev,
450 				struct ib_device_attr *props,
451 				struct ib_udata *uhw)
452 {
453 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
454 	struct mlx5_core_dev *mdev = dev->mdev;
455 	int err = -ENOMEM;
456 	int max_rq_sg;
457 	int max_sq_sg;
458 	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
459 	struct mlx5_ib_query_device_resp resp = {};
460 	size_t resp_len;
461 	u64 max_tso;
462 
463 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
464 	if (uhw->outlen && uhw->outlen < resp_len)
465 		return -EINVAL;
466 	else
467 		resp.response_length = resp_len;
468 
469 	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
470 		return -EINVAL;
471 
472 	memset(props, 0, sizeof(*props));
473 	err = mlx5_query_system_image_guid(ibdev,
474 					   &props->sys_image_guid);
475 	if (err)
476 		return err;
477 
478 	err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
479 	if (err)
480 		return err;
481 
482 	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
483 	if (err)
484 		return err;
485 
486 	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
487 		(fw_rev_min(dev->mdev) << 16) |
488 		fw_rev_sub(dev->mdev);
489 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
490 		IB_DEVICE_PORT_ACTIVE_EVENT		|
491 		IB_DEVICE_SYS_IMAGE_GUID		|
492 		IB_DEVICE_RC_RNR_NAK_GEN;
493 
494 	if (MLX5_CAP_GEN(mdev, pkv))
495 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
496 	if (MLX5_CAP_GEN(mdev, qkv))
497 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
498 	if (MLX5_CAP_GEN(mdev, apm))
499 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
500 	if (MLX5_CAP_GEN(mdev, xrc))
501 		props->device_cap_flags |= IB_DEVICE_XRC;
502 	if (MLX5_CAP_GEN(mdev, imaicl)) {
503 		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
504 					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
505 		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
506 		/* We support 'Gappy' memory registration too */
507 		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
508 	}
509 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
510 	if (MLX5_CAP_GEN(mdev, sho)) {
511 		props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
512 		/* At this stage no support for signature handover */
513 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
514 				      IB_PROT_T10DIF_TYPE_2 |
515 				      IB_PROT_T10DIF_TYPE_3;
516 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
517 				       IB_GUARD_T10DIF_CSUM;
518 	}
519 	if (MLX5_CAP_GEN(mdev, block_lb_mc))
520 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
521 
522 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads)) {
523 		if (MLX5_CAP_ETH(mdev, csum_cap))
524 			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
525 
526 		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
527 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
528 			if (max_tso) {
529 				resp.tso_caps.max_tso = 1 << max_tso;
530 				resp.tso_caps.supported_qpts |=
531 					1 << IB_QPT_RAW_PACKET;
532 				resp.response_length += sizeof(resp.tso_caps);
533 			}
534 		}
535 	}
536 
537 	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
538 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
539 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
540 	}
541 
542 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
543 	    MLX5_CAP_ETH(dev->mdev, scatter_fcs))
544 		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
545 
546 	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
547 		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
548 
549 	props->vendor_part_id	   = mdev->pdev->device;
550 	props->hw_ver		   = mdev->pdev->revision;
551 
552 	props->max_mr_size	   = ~0ull;
553 	props->page_size_cap	   = ~(min_page_size - 1);
554 	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
555 	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
556 	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
557 		     sizeof(struct mlx5_wqe_data_seg);
558 	max_sq_sg = (MLX5_CAP_GEN(mdev, max_wqe_sz_sq) -
559 		     sizeof(struct mlx5_wqe_ctrl_seg)) /
560 		     sizeof(struct mlx5_wqe_data_seg);
561 	props->max_sge = min(max_rq_sg, max_sq_sg);
562 	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
563 	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
564 	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
565 	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
566 	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
567 	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
568 	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
569 	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
570 	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
571 	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
572 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
573 	props->max_srq_sge	   = max_rq_sg - 1;
574 	props->max_fast_reg_page_list_len =
575 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
576 	get_atomic_caps(dev, props);
577 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
578 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
579 	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
580 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
581 					   props->max_mcast_grp;
582 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
583 	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
584 	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
585 
586 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
587 	if (MLX5_CAP_GEN(mdev, pg))
588 		props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
589 	props->odp_caps = dev->odp_caps;
590 #endif
591 
592 	if (MLX5_CAP_GEN(mdev, cd))
593 		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
594 
595 	if (!mlx5_core_is_pf(mdev))
596 		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
597 
598 	if (uhw->outlen) {
599 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
600 
601 		if (err)
602 			return err;
603 	}
604 
605 	return 0;
606 }
607 
608 enum mlx5_ib_width {
609 	MLX5_IB_WIDTH_1X	= 1 << 0,
610 	MLX5_IB_WIDTH_2X	= 1 << 1,
611 	MLX5_IB_WIDTH_4X	= 1 << 2,
612 	MLX5_IB_WIDTH_8X	= 1 << 3,
613 	MLX5_IB_WIDTH_12X	= 1 << 4
614 };
615 
616 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
617 				  u8 *ib_width)
618 {
619 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
620 	int err = 0;
621 
622 	if (active_width & MLX5_IB_WIDTH_1X) {
623 		*ib_width = IB_WIDTH_1X;
624 	} else if (active_width & MLX5_IB_WIDTH_2X) {
625 		mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
626 			    (int)active_width);
627 		err = -EINVAL;
628 	} else if (active_width & MLX5_IB_WIDTH_4X) {
629 		*ib_width = IB_WIDTH_4X;
630 	} else if (active_width & MLX5_IB_WIDTH_8X) {
631 		*ib_width = IB_WIDTH_8X;
632 	} else if (active_width & MLX5_IB_WIDTH_12X) {
633 		*ib_width = IB_WIDTH_12X;
634 	} else {
635 		mlx5_ib_dbg(dev, "Invalid active_width %d\n",
636 			    (int)active_width);
637 		err = -EINVAL;
638 	}
639 
640 	return err;
641 }
642 
643 static int mlx5_mtu_to_ib_mtu(int mtu)
644 {
645 	switch (mtu) {
646 	case 256: return 1;
647 	case 512: return 2;
648 	case 1024: return 3;
649 	case 2048: return 4;
650 	case 4096: return 5;
651 	default:
652 		pr_warn("invalid mtu\n");
653 		return -1;
654 	}
655 }
656 
657 enum ib_max_vl_num {
658 	__IB_MAX_VL_0		= 1,
659 	__IB_MAX_VL_0_1		= 2,
660 	__IB_MAX_VL_0_3		= 3,
661 	__IB_MAX_VL_0_7		= 4,
662 	__IB_MAX_VL_0_14	= 5,
663 };
664 
665 enum mlx5_vl_hw_cap {
666 	MLX5_VL_HW_0	= 1,
667 	MLX5_VL_HW_0_1	= 2,
668 	MLX5_VL_HW_0_2	= 3,
669 	MLX5_VL_HW_0_3	= 4,
670 	MLX5_VL_HW_0_4	= 5,
671 	MLX5_VL_HW_0_5	= 6,
672 	MLX5_VL_HW_0_6	= 7,
673 	MLX5_VL_HW_0_7	= 8,
674 	MLX5_VL_HW_0_14	= 15
675 };
676 
677 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
678 				u8 *max_vl_num)
679 {
680 	switch (vl_hw_cap) {
681 	case MLX5_VL_HW_0:
682 		*max_vl_num = __IB_MAX_VL_0;
683 		break;
684 	case MLX5_VL_HW_0_1:
685 		*max_vl_num = __IB_MAX_VL_0_1;
686 		break;
687 	case MLX5_VL_HW_0_3:
688 		*max_vl_num = __IB_MAX_VL_0_3;
689 		break;
690 	case MLX5_VL_HW_0_7:
691 		*max_vl_num = __IB_MAX_VL_0_7;
692 		break;
693 	case MLX5_VL_HW_0_14:
694 		*max_vl_num = __IB_MAX_VL_0_14;
695 		break;
696 
697 	default:
698 		return -EINVAL;
699 	}
700 
701 	return 0;
702 }
703 
704 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
705 			       struct ib_port_attr *props)
706 {
707 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
708 	struct mlx5_core_dev *mdev = dev->mdev;
709 	struct mlx5_hca_vport_context *rep;
710 	u16 max_mtu;
711 	u16 oper_mtu;
712 	int err;
713 	u8 ib_link_width_oper;
714 	u8 vl_hw_cap;
715 
716 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
717 	if (!rep) {
718 		err = -ENOMEM;
719 		goto out;
720 	}
721 
722 	memset(props, 0, sizeof(*props));
723 
724 	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
725 	if (err)
726 		goto out;
727 
728 	props->lid		= rep->lid;
729 	props->lmc		= rep->lmc;
730 	props->sm_lid		= rep->sm_lid;
731 	props->sm_sl		= rep->sm_sl;
732 	props->state		= rep->vport_state;
733 	props->phys_state	= rep->port_physical_state;
734 	props->port_cap_flags	= rep->cap_mask1;
735 	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
736 	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
737 	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
738 	props->bad_pkey_cntr	= rep->pkey_violation_counter;
739 	props->qkey_viol_cntr	= rep->qkey_violation_counter;
740 	props->subnet_timeout	= rep->subnet_timeout;
741 	props->init_type_reply	= rep->init_type_reply;
742 	props->grh_required	= rep->grh_required;
743 
744 	err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
745 	if (err)
746 		goto out;
747 
748 	err = translate_active_width(ibdev, ib_link_width_oper,
749 				     &props->active_width);
750 	if (err)
751 		goto out;
752 	err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
753 	if (err)
754 		goto out;
755 
756 	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
757 
758 	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
759 
760 	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
761 
762 	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
763 
764 	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
765 	if (err)
766 		goto out;
767 
768 	err = translate_max_vl_num(ibdev, vl_hw_cap,
769 				   &props->max_vl_num);
770 out:
771 	kfree(rep);
772 	return err;
773 }
774 
775 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
776 		       struct ib_port_attr *props)
777 {
778 	switch (mlx5_get_vport_access_method(ibdev)) {
779 	case MLX5_VPORT_ACCESS_METHOD_MAD:
780 		return mlx5_query_mad_ifc_port(ibdev, port, props);
781 
782 	case MLX5_VPORT_ACCESS_METHOD_HCA:
783 		return mlx5_query_hca_port(ibdev, port, props);
784 
785 	case MLX5_VPORT_ACCESS_METHOD_NIC:
786 		return mlx5_query_port_roce(ibdev, port, props);
787 
788 	default:
789 		return -EINVAL;
790 	}
791 }
792 
793 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
794 			     union ib_gid *gid)
795 {
796 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
797 	struct mlx5_core_dev *mdev = dev->mdev;
798 
799 	switch (mlx5_get_vport_access_method(ibdev)) {
800 	case MLX5_VPORT_ACCESS_METHOD_MAD:
801 		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
802 
803 	case MLX5_VPORT_ACCESS_METHOD_HCA:
804 		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
805 
806 	default:
807 		return -EINVAL;
808 	}
809 
810 }
811 
812 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
813 			      u16 *pkey)
814 {
815 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
816 	struct mlx5_core_dev *mdev = dev->mdev;
817 
818 	switch (mlx5_get_vport_access_method(ibdev)) {
819 	case MLX5_VPORT_ACCESS_METHOD_MAD:
820 		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
821 
822 	case MLX5_VPORT_ACCESS_METHOD_HCA:
823 	case MLX5_VPORT_ACCESS_METHOD_NIC:
824 		return mlx5_query_hca_vport_pkey(mdev, 0, port,  0, index,
825 						 pkey);
826 	default:
827 		return -EINVAL;
828 	}
829 }
830 
831 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
832 				 struct ib_device_modify *props)
833 {
834 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
835 	struct mlx5_reg_node_desc in;
836 	struct mlx5_reg_node_desc out;
837 	int err;
838 
839 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
840 		return -EOPNOTSUPP;
841 
842 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
843 		return 0;
844 
845 	/*
846 	 * If possible, pass node desc to FW, so it can generate
847 	 * a 144 trap.  If cmd fails, just ignore.
848 	 */
849 	memcpy(&in, props->node_desc, 64);
850 	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
851 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
852 	if (err)
853 		return err;
854 
855 	memcpy(ibdev->node_desc, props->node_desc, 64);
856 
857 	return err;
858 }
859 
860 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
861 			       struct ib_port_modify *props)
862 {
863 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
864 	struct ib_port_attr attr;
865 	u32 tmp;
866 	int err;
867 
868 	mutex_lock(&dev->cap_mask_mutex);
869 
870 	err = mlx5_ib_query_port(ibdev, port, &attr);
871 	if (err)
872 		goto out;
873 
874 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
875 		~props->clr_port_cap_mask;
876 
877 	err = mlx5_set_port_caps(dev->mdev, port, tmp);
878 
879 out:
880 	mutex_unlock(&dev->cap_mask_mutex);
881 	return err;
882 }
883 
884 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
885 						  struct ib_udata *udata)
886 {
887 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
888 	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
889 	struct mlx5_ib_alloc_ucontext_resp resp = {};
890 	struct mlx5_ib_ucontext *context;
891 	struct mlx5_uuar_info *uuari;
892 	struct mlx5_uar *uars;
893 	int gross_uuars;
894 	int num_uars;
895 	int ver;
896 	int uuarn;
897 	int err;
898 	int i;
899 	size_t reqlen;
900 	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
901 				     max_cqe_version);
902 
903 	if (!dev->ib_active)
904 		return ERR_PTR(-EAGAIN);
905 
906 	if (udata->inlen < sizeof(struct ib_uverbs_cmd_hdr))
907 		return ERR_PTR(-EINVAL);
908 
909 	reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
910 	if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
911 		ver = 0;
912 	else if (reqlen >= min_req_v2)
913 		ver = 2;
914 	else
915 		return ERR_PTR(-EINVAL);
916 
917 	err = ib_copy_from_udata(&req, udata, min(reqlen, sizeof(req)));
918 	if (err)
919 		return ERR_PTR(err);
920 
921 	if (req.flags)
922 		return ERR_PTR(-EINVAL);
923 
924 	if (req.total_num_uuars > MLX5_MAX_UUARS)
925 		return ERR_PTR(-ENOMEM);
926 
927 	if (req.total_num_uuars == 0)
928 		return ERR_PTR(-EINVAL);
929 
930 	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
931 		return ERR_PTR(-EOPNOTSUPP);
932 
933 	if (reqlen > sizeof(req) &&
934 	    !ib_is_udata_cleared(udata, sizeof(req),
935 				 reqlen - sizeof(req)))
936 		return ERR_PTR(-EOPNOTSUPP);
937 
938 	req.total_num_uuars = ALIGN(req.total_num_uuars,
939 				    MLX5_NON_FP_BF_REGS_PER_PAGE);
940 	if (req.num_low_latency_uuars > req.total_num_uuars - 1)
941 		return ERR_PTR(-EINVAL);
942 
943 	num_uars = req.total_num_uuars / MLX5_NON_FP_BF_REGS_PER_PAGE;
944 	gross_uuars = num_uars * MLX5_BF_REGS_PER_PAGE;
945 	resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
946 	if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
947 		resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
948 	resp.cache_line_size = L1_CACHE_BYTES;
949 	resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
950 	resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
951 	resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
952 	resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
953 	resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
954 	resp.cqe_version = min_t(__u8,
955 				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
956 				 req.max_cqe_version);
957 	resp.response_length = min(offsetof(typeof(resp), response_length) +
958 				   sizeof(resp.response_length), udata->outlen);
959 
960 	context = kzalloc(sizeof(*context), GFP_KERNEL);
961 	if (!context)
962 		return ERR_PTR(-ENOMEM);
963 
964 	uuari = &context->uuari;
965 	mutex_init(&uuari->lock);
966 	uars = kcalloc(num_uars, sizeof(*uars), GFP_KERNEL);
967 	if (!uars) {
968 		err = -ENOMEM;
969 		goto out_ctx;
970 	}
971 
972 	uuari->bitmap = kcalloc(BITS_TO_LONGS(gross_uuars),
973 				sizeof(*uuari->bitmap),
974 				GFP_KERNEL);
975 	if (!uuari->bitmap) {
976 		err = -ENOMEM;
977 		goto out_uar_ctx;
978 	}
979 	/*
980 	 * clear all fast path uuars
981 	 */
982 	for (i = 0; i < gross_uuars; i++) {
983 		uuarn = i & 3;
984 		if (uuarn == 2 || uuarn == 3)
985 			set_bit(i, uuari->bitmap);
986 	}
987 
988 	uuari->count = kcalloc(gross_uuars, sizeof(*uuari->count), GFP_KERNEL);
989 	if (!uuari->count) {
990 		err = -ENOMEM;
991 		goto out_bitmap;
992 	}
993 
994 	for (i = 0; i < num_uars; i++) {
995 		err = mlx5_cmd_alloc_uar(dev->mdev, &uars[i].index);
996 		if (err)
997 			goto out_count;
998 	}
999 
1000 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1001 	context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1002 #endif
1003 
1004 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1005 		err = mlx5_core_alloc_transport_domain(dev->mdev,
1006 						       &context->tdn);
1007 		if (err)
1008 			goto out_uars;
1009 	}
1010 
1011 	INIT_LIST_HEAD(&context->vma_private_list);
1012 	INIT_LIST_HEAD(&context->db_page_list);
1013 	mutex_init(&context->db_page_mutex);
1014 
1015 	resp.tot_uuars = req.total_num_uuars;
1016 	resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
1017 
1018 	if (field_avail(typeof(resp), cqe_version, udata->outlen))
1019 		resp.response_length += sizeof(resp.cqe_version);
1020 
1021 	if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1022 		resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE;
1023 		resp.response_length += sizeof(resp.cmds_supp_uhw);
1024 	}
1025 
1026 	/*
1027 	 * We don't want to expose information from the PCI bar that is located
1028 	 * after 4096 bytes, so if the arch only supports larger pages, let's
1029 	 * pretend we don't support reading the HCA's core clock. This is also
1030 	 * forced by mmap function.
1031 	 */
1032 	if (PAGE_SIZE <= 4096 &&
1033 	    field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1034 		resp.comp_mask |=
1035 			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1036 		resp.hca_core_clock_offset =
1037 			offsetof(struct mlx5_init_seg, internal_timer_h) %
1038 			PAGE_SIZE;
1039 		resp.response_length += sizeof(resp.hca_core_clock_offset) +
1040 					sizeof(resp.reserved2);
1041 	}
1042 
1043 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1044 	if (err)
1045 		goto out_td;
1046 
1047 	uuari->ver = ver;
1048 	uuari->num_low_latency_uuars = req.num_low_latency_uuars;
1049 	uuari->uars = uars;
1050 	uuari->num_uars = num_uars;
1051 	context->cqe_version = resp.cqe_version;
1052 
1053 	return &context->ibucontext;
1054 
1055 out_td:
1056 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1057 		mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1058 
1059 out_uars:
1060 	for (i--; i >= 0; i--)
1061 		mlx5_cmd_free_uar(dev->mdev, uars[i].index);
1062 out_count:
1063 	kfree(uuari->count);
1064 
1065 out_bitmap:
1066 	kfree(uuari->bitmap);
1067 
1068 out_uar_ctx:
1069 	kfree(uars);
1070 
1071 out_ctx:
1072 	kfree(context);
1073 	return ERR_PTR(err);
1074 }
1075 
1076 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1077 {
1078 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1079 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1080 	struct mlx5_uuar_info *uuari = &context->uuari;
1081 	int i;
1082 
1083 	if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1084 		mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1085 
1086 	for (i = 0; i < uuari->num_uars; i++) {
1087 		if (mlx5_cmd_free_uar(dev->mdev, uuari->uars[i].index))
1088 			mlx5_ib_warn(dev, "failed to free UAR 0x%x\n", uuari->uars[i].index);
1089 	}
1090 
1091 	kfree(uuari->count);
1092 	kfree(uuari->bitmap);
1093 	kfree(uuari->uars);
1094 	kfree(context);
1095 
1096 	return 0;
1097 }
1098 
1099 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, int index)
1100 {
1101 	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + index;
1102 }
1103 
1104 static int get_command(unsigned long offset)
1105 {
1106 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1107 }
1108 
1109 static int get_arg(unsigned long offset)
1110 {
1111 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1112 }
1113 
1114 static int get_index(unsigned long offset)
1115 {
1116 	return get_arg(offset);
1117 }
1118 
1119 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1120 {
1121 	/* vma_open is called when a new VMA is created on top of our VMA.  This
1122 	 * is done through either mremap flow or split_vma (usually due to
1123 	 * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1124 	 * as this VMA is strongly hardware related.  Therefore we set the
1125 	 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1126 	 * calling us again and trying to do incorrect actions.  We assume that
1127 	 * the original VMA size is exactly a single page, and therefore all
1128 	 * "splitting" operation will not happen to it.
1129 	 */
1130 	area->vm_ops = NULL;
1131 }
1132 
1133 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1134 {
1135 	struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1136 
1137 	/* It's guaranteed that all VMAs opened on a FD are closed before the
1138 	 * file itself is closed, therefore no sync is needed with the regular
1139 	 * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1140 	 * However need a sync with accessing the vma as part of
1141 	 * mlx5_ib_disassociate_ucontext.
1142 	 * The close operation is usually called under mm->mmap_sem except when
1143 	 * process is exiting.
1144 	 * The exiting case is handled explicitly as part of
1145 	 * mlx5_ib_disassociate_ucontext.
1146 	 */
1147 	mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1148 
1149 	/* setting the vma context pointer to null in the mlx5_ib driver's
1150 	 * private data, to protect a race condition in
1151 	 * mlx5_ib_disassociate_ucontext().
1152 	 */
1153 	mlx5_ib_vma_priv_data->vma = NULL;
1154 	list_del(&mlx5_ib_vma_priv_data->list);
1155 	kfree(mlx5_ib_vma_priv_data);
1156 }
1157 
1158 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1159 	.open = mlx5_ib_vma_open,
1160 	.close = mlx5_ib_vma_close
1161 };
1162 
1163 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1164 				struct mlx5_ib_ucontext *ctx)
1165 {
1166 	struct mlx5_ib_vma_private_data *vma_prv;
1167 	struct list_head *vma_head = &ctx->vma_private_list;
1168 
1169 	vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1170 	if (!vma_prv)
1171 		return -ENOMEM;
1172 
1173 	vma_prv->vma = vma;
1174 	vma->vm_private_data = vma_prv;
1175 	vma->vm_ops =  &mlx5_ib_vm_ops;
1176 
1177 	list_add(&vma_prv->list, vma_head);
1178 
1179 	return 0;
1180 }
1181 
1182 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1183 {
1184 	int ret;
1185 	struct vm_area_struct *vma;
1186 	struct mlx5_ib_vma_private_data *vma_private, *n;
1187 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1188 	struct task_struct *owning_process  = NULL;
1189 	struct mm_struct   *owning_mm       = NULL;
1190 
1191 	owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1192 	if (!owning_process)
1193 		return;
1194 
1195 	owning_mm = get_task_mm(owning_process);
1196 	if (!owning_mm) {
1197 		pr_info("no mm, disassociate ucontext is pending task termination\n");
1198 		while (1) {
1199 			put_task_struct(owning_process);
1200 			usleep_range(1000, 2000);
1201 			owning_process = get_pid_task(ibcontext->tgid,
1202 						      PIDTYPE_PID);
1203 			if (!owning_process ||
1204 			    owning_process->state == TASK_DEAD) {
1205 				pr_info("disassociate ucontext done, task was terminated\n");
1206 				/* in case task was dead need to release the
1207 				 * task struct.
1208 				 */
1209 				if (owning_process)
1210 					put_task_struct(owning_process);
1211 				return;
1212 			}
1213 		}
1214 	}
1215 
1216 	/* need to protect from a race on closing the vma as part of
1217 	 * mlx5_ib_vma_close.
1218 	 */
1219 	down_read(&owning_mm->mmap_sem);
1220 	list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1221 				 list) {
1222 		vma = vma_private->vma;
1223 		ret = zap_vma_ptes(vma, vma->vm_start,
1224 				   PAGE_SIZE);
1225 		WARN_ONCE(ret, "%s: zap_vma_ptes failed", __func__);
1226 		/* context going to be destroyed, should
1227 		 * not access ops any more.
1228 		 */
1229 		vma->vm_ops = NULL;
1230 		list_del(&vma_private->list);
1231 		kfree(vma_private);
1232 	}
1233 	up_read(&owning_mm->mmap_sem);
1234 	mmput(owning_mm);
1235 	put_task_struct(owning_process);
1236 }
1237 
1238 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1239 {
1240 	switch (cmd) {
1241 	case MLX5_IB_MMAP_WC_PAGE:
1242 		return "WC";
1243 	case MLX5_IB_MMAP_REGULAR_PAGE:
1244 		return "best effort WC";
1245 	case MLX5_IB_MMAP_NC_PAGE:
1246 		return "NC";
1247 	default:
1248 		return NULL;
1249 	}
1250 }
1251 
1252 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
1253 		    struct vm_area_struct *vma,
1254 		    struct mlx5_ib_ucontext *context)
1255 {
1256 	struct mlx5_uuar_info *uuari = &context->uuari;
1257 	int err;
1258 	unsigned long idx;
1259 	phys_addr_t pfn, pa;
1260 	pgprot_t prot;
1261 
1262 	switch (cmd) {
1263 	case MLX5_IB_MMAP_WC_PAGE:
1264 /* Some architectures don't support WC memory */
1265 #if defined(CONFIG_X86)
1266 		if (!pat_enabled())
1267 			return -EPERM;
1268 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
1269 			return -EPERM;
1270 #endif
1271 	/* fall through */
1272 	case MLX5_IB_MMAP_REGULAR_PAGE:
1273 		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
1274 		prot = pgprot_writecombine(vma->vm_page_prot);
1275 		break;
1276 	case MLX5_IB_MMAP_NC_PAGE:
1277 		prot = pgprot_noncached(vma->vm_page_prot);
1278 		break;
1279 	default:
1280 		return -EINVAL;
1281 	}
1282 
1283 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1284 		return -EINVAL;
1285 
1286 	idx = get_index(vma->vm_pgoff);
1287 	if (idx >= uuari->num_uars)
1288 		return -EINVAL;
1289 
1290 	pfn = uar_index2pfn(dev, uuari->uars[idx].index);
1291 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
1292 
1293 	vma->vm_page_prot = prot;
1294 	err = io_remap_pfn_range(vma, vma->vm_start, pfn,
1295 				 PAGE_SIZE, vma->vm_page_prot);
1296 	if (err) {
1297 		mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%lx, pfn=%pa, mmap_cmd=%s\n",
1298 			    err, vma->vm_start, &pfn, mmap_cmd2str(cmd));
1299 		return -EAGAIN;
1300 	}
1301 
1302 	pa = pfn << PAGE_SHIFT;
1303 	mlx5_ib_dbg(dev, "mapped %s at 0x%lx, PA %pa\n", mmap_cmd2str(cmd),
1304 		    vma->vm_start, &pa);
1305 
1306 	return mlx5_ib_set_vma_data(vma, context);
1307 }
1308 
1309 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1310 {
1311 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1312 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1313 	unsigned long command;
1314 	phys_addr_t pfn;
1315 
1316 	command = get_command(vma->vm_pgoff);
1317 	switch (command) {
1318 	case MLX5_IB_MMAP_WC_PAGE:
1319 	case MLX5_IB_MMAP_NC_PAGE:
1320 	case MLX5_IB_MMAP_REGULAR_PAGE:
1321 		return uar_mmap(dev, command, vma, context);
1322 
1323 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1324 		return -ENOSYS;
1325 
1326 	case MLX5_IB_MMAP_CORE_CLOCK:
1327 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1328 			return -EINVAL;
1329 
1330 		if (vma->vm_flags & VM_WRITE)
1331 			return -EPERM;
1332 
1333 		/* Don't expose to user-space information it shouldn't have */
1334 		if (PAGE_SIZE > 4096)
1335 			return -EOPNOTSUPP;
1336 
1337 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1338 		pfn = (dev->mdev->iseg_base +
1339 		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1340 			PAGE_SHIFT;
1341 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1342 				       PAGE_SIZE, vma->vm_page_prot))
1343 			return -EAGAIN;
1344 
1345 		mlx5_ib_dbg(dev, "mapped internal timer at 0x%lx, PA 0x%llx\n",
1346 			    vma->vm_start,
1347 			    (unsigned long long)pfn << PAGE_SHIFT);
1348 		break;
1349 
1350 	default:
1351 		return -EINVAL;
1352 	}
1353 
1354 	return 0;
1355 }
1356 
1357 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
1358 				      struct ib_ucontext *context,
1359 				      struct ib_udata *udata)
1360 {
1361 	struct mlx5_ib_alloc_pd_resp resp;
1362 	struct mlx5_ib_pd *pd;
1363 	int err;
1364 
1365 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1366 	if (!pd)
1367 		return ERR_PTR(-ENOMEM);
1368 
1369 	err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
1370 	if (err) {
1371 		kfree(pd);
1372 		return ERR_PTR(err);
1373 	}
1374 
1375 	if (context) {
1376 		resp.pdn = pd->pdn;
1377 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1378 			mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
1379 			kfree(pd);
1380 			return ERR_PTR(-EFAULT);
1381 		}
1382 	}
1383 
1384 	return &pd->ibpd;
1385 }
1386 
1387 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
1388 {
1389 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1390 	struct mlx5_ib_pd *mpd = to_mpd(pd);
1391 
1392 	mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
1393 	kfree(mpd);
1394 
1395 	return 0;
1396 }
1397 
1398 static bool outer_header_zero(u32 *match_criteria)
1399 {
1400 	int size = MLX5_ST_SZ_BYTES(fte_match_param);
1401 	char *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_criteria,
1402 					     outer_headers);
1403 
1404 	return outer_headers_c[0] == 0 && !memcmp(outer_headers_c,
1405 						  outer_headers_c + 1,
1406 						  size - 1);
1407 }
1408 
1409 static int parse_flow_attr(u32 *match_c, u32 *match_v,
1410 			   union ib_flow_spec *ib_spec)
1411 {
1412 	void *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1413 					     outer_headers);
1414 	void *outer_headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1415 					     outer_headers);
1416 	switch (ib_spec->type) {
1417 	case IB_FLOW_SPEC_ETH:
1418 		if (ib_spec->size != sizeof(ib_spec->eth))
1419 			return -EINVAL;
1420 
1421 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1422 					     dmac_47_16),
1423 				ib_spec->eth.mask.dst_mac);
1424 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1425 					     dmac_47_16),
1426 				ib_spec->eth.val.dst_mac);
1427 
1428 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1429 					     smac_47_16),
1430 				ib_spec->eth.mask.src_mac);
1431 		ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1432 					     smac_47_16),
1433 				ib_spec->eth.val.src_mac);
1434 
1435 		if (ib_spec->eth.mask.vlan_tag) {
1436 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1437 				 vlan_tag, 1);
1438 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1439 				 vlan_tag, 1);
1440 
1441 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1442 				 first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
1443 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1444 				 first_vid, ntohs(ib_spec->eth.val.vlan_tag));
1445 
1446 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1447 				 first_cfi,
1448 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
1449 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1450 				 first_cfi,
1451 				 ntohs(ib_spec->eth.val.vlan_tag) >> 12);
1452 
1453 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1454 				 first_prio,
1455 				 ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
1456 			MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1457 				 first_prio,
1458 				 ntohs(ib_spec->eth.val.vlan_tag) >> 13);
1459 		}
1460 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1461 			 ethertype, ntohs(ib_spec->eth.mask.ether_type));
1462 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1463 			 ethertype, ntohs(ib_spec->eth.val.ether_type));
1464 		break;
1465 	case IB_FLOW_SPEC_IPV4:
1466 		if (ib_spec->size != sizeof(ib_spec->ipv4))
1467 			return -EINVAL;
1468 
1469 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1470 			 ethertype, 0xffff);
1471 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1472 			 ethertype, ETH_P_IP);
1473 
1474 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1475 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1476 		       &ib_spec->ipv4.mask.src_ip,
1477 		       sizeof(ib_spec->ipv4.mask.src_ip));
1478 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1479 				    src_ipv4_src_ipv6.ipv4_layout.ipv4),
1480 		       &ib_spec->ipv4.val.src_ip,
1481 		       sizeof(ib_spec->ipv4.val.src_ip));
1482 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1483 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1484 		       &ib_spec->ipv4.mask.dst_ip,
1485 		       sizeof(ib_spec->ipv4.mask.dst_ip));
1486 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1487 				    dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1488 		       &ib_spec->ipv4.val.dst_ip,
1489 		       sizeof(ib_spec->ipv4.val.dst_ip));
1490 		break;
1491 	case IB_FLOW_SPEC_IPV6:
1492 		if (ib_spec->size != sizeof(ib_spec->ipv6))
1493 			return -EINVAL;
1494 
1495 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1496 			 ethertype, 0xffff);
1497 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1498 			 ethertype, ETH_P_IPV6);
1499 
1500 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1501 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1502 		       &ib_spec->ipv6.mask.src_ip,
1503 		       sizeof(ib_spec->ipv6.mask.src_ip));
1504 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1505 				    src_ipv4_src_ipv6.ipv6_layout.ipv6),
1506 		       &ib_spec->ipv6.val.src_ip,
1507 		       sizeof(ib_spec->ipv6.val.src_ip));
1508 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1509 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1510 		       &ib_spec->ipv6.mask.dst_ip,
1511 		       sizeof(ib_spec->ipv6.mask.dst_ip));
1512 		memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1513 				    dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1514 		       &ib_spec->ipv6.val.dst_ip,
1515 		       sizeof(ib_spec->ipv6.val.dst_ip));
1516 		break;
1517 	case IB_FLOW_SPEC_TCP:
1518 		if (ib_spec->size != sizeof(ib_spec->tcp_udp))
1519 			return -EINVAL;
1520 
1521 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1522 			 0xff);
1523 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1524 			 IPPROTO_TCP);
1525 
1526 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_sport,
1527 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1528 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_sport,
1529 			 ntohs(ib_spec->tcp_udp.val.src_port));
1530 
1531 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_dport,
1532 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1533 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_dport,
1534 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1535 		break;
1536 	case IB_FLOW_SPEC_UDP:
1537 		if (ib_spec->size != sizeof(ib_spec->tcp_udp))
1538 			return -EINVAL;
1539 
1540 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1541 			 0xff);
1542 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1543 			 IPPROTO_UDP);
1544 
1545 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_sport,
1546 			 ntohs(ib_spec->tcp_udp.mask.src_port));
1547 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_sport,
1548 			 ntohs(ib_spec->tcp_udp.val.src_port));
1549 
1550 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_dport,
1551 			 ntohs(ib_spec->tcp_udp.mask.dst_port));
1552 		MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_dport,
1553 			 ntohs(ib_spec->tcp_udp.val.dst_port));
1554 		break;
1555 	default:
1556 		return -EINVAL;
1557 	}
1558 
1559 	return 0;
1560 }
1561 
1562 /* If a flow could catch both multicast and unicast packets,
1563  * it won't fall into the multicast flow steering table and this rule
1564  * could steal other multicast packets.
1565  */
1566 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr)
1567 {
1568 	struct ib_flow_spec_eth *eth_spec;
1569 
1570 	if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
1571 	    ib_attr->size < sizeof(struct ib_flow_attr) +
1572 	    sizeof(struct ib_flow_spec_eth) ||
1573 	    ib_attr->num_of_specs < 1)
1574 		return false;
1575 
1576 	eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1);
1577 	if (eth_spec->type != IB_FLOW_SPEC_ETH ||
1578 	    eth_spec->size != sizeof(*eth_spec))
1579 		return false;
1580 
1581 	return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
1582 	       is_multicast_ether_addr(eth_spec->val.dst_mac);
1583 }
1584 
1585 static bool is_valid_attr(struct ib_flow_attr *flow_attr)
1586 {
1587 	union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
1588 	bool has_ipv4_spec = false;
1589 	bool eth_type_ipv4 = true;
1590 	unsigned int spec_index;
1591 
1592 	/* Validate that ethertype is correct */
1593 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1594 		if (ib_spec->type == IB_FLOW_SPEC_ETH &&
1595 		    ib_spec->eth.mask.ether_type) {
1596 			if (!((ib_spec->eth.mask.ether_type == htons(0xffff)) &&
1597 			      ib_spec->eth.val.ether_type == htons(ETH_P_IP)))
1598 				eth_type_ipv4 = false;
1599 		} else if (ib_spec->type == IB_FLOW_SPEC_IPV4) {
1600 			has_ipv4_spec = true;
1601 		}
1602 		ib_spec = (void *)ib_spec + ib_spec->size;
1603 	}
1604 	return !has_ipv4_spec || eth_type_ipv4;
1605 }
1606 
1607 static void put_flow_table(struct mlx5_ib_dev *dev,
1608 			   struct mlx5_ib_flow_prio *prio, bool ft_added)
1609 {
1610 	prio->refcount -= !!ft_added;
1611 	if (!prio->refcount) {
1612 		mlx5_destroy_flow_table(prio->flow_table);
1613 		prio->flow_table = NULL;
1614 	}
1615 }
1616 
1617 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
1618 {
1619 	struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
1620 	struct mlx5_ib_flow_handler *handler = container_of(flow_id,
1621 							  struct mlx5_ib_flow_handler,
1622 							  ibflow);
1623 	struct mlx5_ib_flow_handler *iter, *tmp;
1624 
1625 	mutex_lock(&dev->flow_db.lock);
1626 
1627 	list_for_each_entry_safe(iter, tmp, &handler->list, list) {
1628 		mlx5_del_flow_rule(iter->rule);
1629 		list_del(&iter->list);
1630 		kfree(iter);
1631 	}
1632 
1633 	mlx5_del_flow_rule(handler->rule);
1634 	put_flow_table(dev, &dev->flow_db.prios[handler->prio], true);
1635 	mutex_unlock(&dev->flow_db.lock);
1636 
1637 	kfree(handler);
1638 
1639 	return 0;
1640 }
1641 
1642 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
1643 {
1644 	priority *= 2;
1645 	if (!dont_trap)
1646 		priority++;
1647 	return priority;
1648 }
1649 
1650 #define MLX5_FS_MAX_TYPES	 10
1651 #define MLX5_FS_MAX_ENTRIES	 32000UL
1652 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
1653 						struct ib_flow_attr *flow_attr)
1654 {
1655 	bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
1656 	struct mlx5_flow_namespace *ns = NULL;
1657 	struct mlx5_ib_flow_prio *prio;
1658 	struct mlx5_flow_table *ft;
1659 	int num_entries;
1660 	int num_groups;
1661 	int priority;
1662 	int err = 0;
1663 
1664 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1665 		if (flow_is_multicast_only(flow_attr) &&
1666 		    !dont_trap)
1667 			priority = MLX5_IB_FLOW_MCAST_PRIO;
1668 		else
1669 			priority = ib_prio_to_core_prio(flow_attr->priority,
1670 							dont_trap);
1671 		ns = mlx5_get_flow_namespace(dev->mdev,
1672 					     MLX5_FLOW_NAMESPACE_BYPASS);
1673 		num_entries = MLX5_FS_MAX_ENTRIES;
1674 		num_groups = MLX5_FS_MAX_TYPES;
1675 		prio = &dev->flow_db.prios[priority];
1676 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
1677 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
1678 		ns = mlx5_get_flow_namespace(dev->mdev,
1679 					     MLX5_FLOW_NAMESPACE_LEFTOVERS);
1680 		build_leftovers_ft_param(&priority,
1681 					 &num_entries,
1682 					 &num_groups);
1683 		prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
1684 	}
1685 
1686 	if (!ns)
1687 		return ERR_PTR(-ENOTSUPP);
1688 
1689 	ft = prio->flow_table;
1690 	if (!ft) {
1691 		ft = mlx5_create_auto_grouped_flow_table(ns, priority,
1692 							 num_entries,
1693 							 num_groups,
1694 							 0);
1695 
1696 		if (!IS_ERR(ft)) {
1697 			prio->refcount = 0;
1698 			prio->flow_table = ft;
1699 		} else {
1700 			err = PTR_ERR(ft);
1701 		}
1702 	}
1703 
1704 	return err ? ERR_PTR(err) : prio;
1705 }
1706 
1707 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
1708 						     struct mlx5_ib_flow_prio *ft_prio,
1709 						     struct ib_flow_attr *flow_attr,
1710 						     struct mlx5_flow_destination *dst)
1711 {
1712 	struct mlx5_flow_table	*ft = ft_prio->flow_table;
1713 	struct mlx5_ib_flow_handler *handler;
1714 	struct mlx5_flow_spec *spec;
1715 	void *ib_flow = flow_attr + 1;
1716 	unsigned int spec_index;
1717 	u32 action;
1718 	int err = 0;
1719 
1720 	if (!is_valid_attr(flow_attr))
1721 		return ERR_PTR(-EINVAL);
1722 
1723 	spec = mlx5_vzalloc(sizeof(*spec));
1724 	handler = kzalloc(sizeof(*handler), GFP_KERNEL);
1725 	if (!handler || !spec) {
1726 		err = -ENOMEM;
1727 		goto free;
1728 	}
1729 
1730 	INIT_LIST_HEAD(&handler->list);
1731 
1732 	for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1733 		err = parse_flow_attr(spec->match_criteria,
1734 				      spec->match_value, ib_flow);
1735 		if (err < 0)
1736 			goto free;
1737 
1738 		ib_flow += ((union ib_flow_spec *)ib_flow)->size;
1739 	}
1740 
1741 	/* Outer header support only */
1742 	spec->match_criteria_enable = (!outer_header_zero(spec->match_criteria))
1743 		<< 0;
1744 	action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
1745 		MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
1746 	handler->rule = mlx5_add_flow_rule(ft, spec,
1747 					   action,
1748 					   MLX5_FS_DEFAULT_FLOW_TAG,
1749 					   dst);
1750 
1751 	if (IS_ERR(handler->rule)) {
1752 		err = PTR_ERR(handler->rule);
1753 		goto free;
1754 	}
1755 
1756 	handler->prio = ft_prio - dev->flow_db.prios;
1757 
1758 	ft_prio->flow_table = ft;
1759 free:
1760 	if (err)
1761 		kfree(handler);
1762 	kvfree(spec);
1763 	return err ? ERR_PTR(err) : handler;
1764 }
1765 
1766 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
1767 							  struct mlx5_ib_flow_prio *ft_prio,
1768 							  struct ib_flow_attr *flow_attr,
1769 							  struct mlx5_flow_destination *dst)
1770 {
1771 	struct mlx5_ib_flow_handler *handler_dst = NULL;
1772 	struct mlx5_ib_flow_handler *handler = NULL;
1773 
1774 	handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
1775 	if (!IS_ERR(handler)) {
1776 		handler_dst = create_flow_rule(dev, ft_prio,
1777 					       flow_attr, dst);
1778 		if (IS_ERR(handler_dst)) {
1779 			mlx5_del_flow_rule(handler->rule);
1780 			kfree(handler);
1781 			handler = handler_dst;
1782 		} else {
1783 			list_add(&handler_dst->list, &handler->list);
1784 		}
1785 	}
1786 
1787 	return handler;
1788 }
1789 enum {
1790 	LEFTOVERS_MC,
1791 	LEFTOVERS_UC,
1792 };
1793 
1794 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
1795 							  struct mlx5_ib_flow_prio *ft_prio,
1796 							  struct ib_flow_attr *flow_attr,
1797 							  struct mlx5_flow_destination *dst)
1798 {
1799 	struct mlx5_ib_flow_handler *handler_ucast = NULL;
1800 	struct mlx5_ib_flow_handler *handler = NULL;
1801 
1802 	static struct {
1803 		struct ib_flow_attr	flow_attr;
1804 		struct ib_flow_spec_eth eth_flow;
1805 	} leftovers_specs[] = {
1806 		[LEFTOVERS_MC] = {
1807 			.flow_attr = {
1808 				.num_of_specs = 1,
1809 				.size = sizeof(leftovers_specs[0])
1810 			},
1811 			.eth_flow = {
1812 				.type = IB_FLOW_SPEC_ETH,
1813 				.size = sizeof(struct ib_flow_spec_eth),
1814 				.mask = {.dst_mac = {0x1} },
1815 				.val =  {.dst_mac = {0x1} }
1816 			}
1817 		},
1818 		[LEFTOVERS_UC] = {
1819 			.flow_attr = {
1820 				.num_of_specs = 1,
1821 				.size = sizeof(leftovers_specs[0])
1822 			},
1823 			.eth_flow = {
1824 				.type = IB_FLOW_SPEC_ETH,
1825 				.size = sizeof(struct ib_flow_spec_eth),
1826 				.mask = {.dst_mac = {0x1} },
1827 				.val = {.dst_mac = {} }
1828 			}
1829 		}
1830 	};
1831 
1832 	handler = create_flow_rule(dev, ft_prio,
1833 				   &leftovers_specs[LEFTOVERS_MC].flow_attr,
1834 				   dst);
1835 	if (!IS_ERR(handler) &&
1836 	    flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
1837 		handler_ucast = create_flow_rule(dev, ft_prio,
1838 						 &leftovers_specs[LEFTOVERS_UC].flow_attr,
1839 						 dst);
1840 		if (IS_ERR(handler_ucast)) {
1841 			kfree(handler);
1842 			handler = handler_ucast;
1843 		} else {
1844 			list_add(&handler_ucast->list, &handler->list);
1845 		}
1846 	}
1847 
1848 	return handler;
1849 }
1850 
1851 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
1852 					   struct ib_flow_attr *flow_attr,
1853 					   int domain)
1854 {
1855 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
1856 	struct mlx5_ib_qp *mqp = to_mqp(qp);
1857 	struct mlx5_ib_flow_handler *handler = NULL;
1858 	struct mlx5_flow_destination *dst = NULL;
1859 	struct mlx5_ib_flow_prio *ft_prio;
1860 	int err;
1861 
1862 	if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
1863 		return ERR_PTR(-ENOSPC);
1864 
1865 	if (domain != IB_FLOW_DOMAIN_USER ||
1866 	    flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
1867 	    (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
1868 		return ERR_PTR(-EINVAL);
1869 
1870 	dst = kzalloc(sizeof(*dst), GFP_KERNEL);
1871 	if (!dst)
1872 		return ERR_PTR(-ENOMEM);
1873 
1874 	mutex_lock(&dev->flow_db.lock);
1875 
1876 	ft_prio = get_flow_table(dev, flow_attr);
1877 	if (IS_ERR(ft_prio)) {
1878 		err = PTR_ERR(ft_prio);
1879 		goto unlock;
1880 	}
1881 
1882 	dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
1883 	if (mqp->flags & MLX5_IB_QP_RSS)
1884 		dst->tir_num = mqp->rss_qp.tirn;
1885 	else
1886 		dst->tir_num = mqp->raw_packet_qp.rq.tirn;
1887 
1888 	if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1889 		if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
1890 			handler = create_dont_trap_rule(dev, ft_prio,
1891 							flow_attr, dst);
1892 		} else {
1893 			handler = create_flow_rule(dev, ft_prio, flow_attr,
1894 						   dst);
1895 		}
1896 	} else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
1897 		   flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
1898 		handler = create_leftovers_rule(dev, ft_prio, flow_attr,
1899 						dst);
1900 	} else {
1901 		err = -EINVAL;
1902 		goto destroy_ft;
1903 	}
1904 
1905 	if (IS_ERR(handler)) {
1906 		err = PTR_ERR(handler);
1907 		handler = NULL;
1908 		goto destroy_ft;
1909 	}
1910 
1911 	ft_prio->refcount++;
1912 	mutex_unlock(&dev->flow_db.lock);
1913 	kfree(dst);
1914 
1915 	return &handler->ibflow;
1916 
1917 destroy_ft:
1918 	put_flow_table(dev, ft_prio, false);
1919 unlock:
1920 	mutex_unlock(&dev->flow_db.lock);
1921 	kfree(dst);
1922 	kfree(handler);
1923 	return ERR_PTR(err);
1924 }
1925 
1926 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
1927 {
1928 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1929 	int err;
1930 
1931 	err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
1932 	if (err)
1933 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
1934 			     ibqp->qp_num, gid->raw);
1935 
1936 	return err;
1937 }
1938 
1939 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
1940 {
1941 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1942 	int err;
1943 
1944 	err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
1945 	if (err)
1946 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
1947 			     ibqp->qp_num, gid->raw);
1948 
1949 	return err;
1950 }
1951 
1952 static int init_node_data(struct mlx5_ib_dev *dev)
1953 {
1954 	int err;
1955 
1956 	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
1957 	if (err)
1958 		return err;
1959 
1960 	dev->mdev->rev_id = dev->mdev->pdev->revision;
1961 
1962 	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
1963 }
1964 
1965 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
1966 			     char *buf)
1967 {
1968 	struct mlx5_ib_dev *dev =
1969 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
1970 
1971 	return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
1972 }
1973 
1974 static ssize_t show_reg_pages(struct device *device,
1975 			      struct device_attribute *attr, char *buf)
1976 {
1977 	struct mlx5_ib_dev *dev =
1978 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
1979 
1980 	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
1981 }
1982 
1983 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
1984 			char *buf)
1985 {
1986 	struct mlx5_ib_dev *dev =
1987 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
1988 	return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
1989 }
1990 
1991 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
1992 			char *buf)
1993 {
1994 	struct mlx5_ib_dev *dev =
1995 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
1996 	return sprintf(buf, "%x\n", dev->mdev->rev_id);
1997 }
1998 
1999 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2000 			  char *buf)
2001 {
2002 	struct mlx5_ib_dev *dev =
2003 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2004 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2005 		       dev->mdev->board_id);
2006 }
2007 
2008 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
2009 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
2010 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
2011 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
2012 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
2013 
2014 static struct device_attribute *mlx5_class_attributes[] = {
2015 	&dev_attr_hw_rev,
2016 	&dev_attr_hca_type,
2017 	&dev_attr_board_id,
2018 	&dev_attr_fw_pages,
2019 	&dev_attr_reg_pages,
2020 };
2021 
2022 static void pkey_change_handler(struct work_struct *work)
2023 {
2024 	struct mlx5_ib_port_resources *ports =
2025 		container_of(work, struct mlx5_ib_port_resources,
2026 			     pkey_change_work);
2027 
2028 	mutex_lock(&ports->devr->mutex);
2029 	mlx5_ib_gsi_pkey_change(ports->gsi);
2030 	mutex_unlock(&ports->devr->mutex);
2031 }
2032 
2033 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2034 {
2035 	struct mlx5_ib_qp *mqp;
2036 	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2037 	struct mlx5_core_cq *mcq;
2038 	struct list_head cq_armed_list;
2039 	unsigned long flags_qp;
2040 	unsigned long flags_cq;
2041 	unsigned long flags;
2042 
2043 	INIT_LIST_HEAD(&cq_armed_list);
2044 
2045 	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2046 	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2047 	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2048 		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2049 		if (mqp->sq.tail != mqp->sq.head) {
2050 			send_mcq = to_mcq(mqp->ibqp.send_cq);
2051 			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2052 			if (send_mcq->mcq.comp &&
2053 			    mqp->ibqp.send_cq->comp_handler) {
2054 				if (!send_mcq->mcq.reset_notify_added) {
2055 					send_mcq->mcq.reset_notify_added = 1;
2056 					list_add_tail(&send_mcq->mcq.reset_notify,
2057 						      &cq_armed_list);
2058 				}
2059 			}
2060 			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2061 		}
2062 		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2063 		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2064 		/* no handling is needed for SRQ */
2065 		if (!mqp->ibqp.srq) {
2066 			if (mqp->rq.tail != mqp->rq.head) {
2067 				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2068 				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2069 				if (recv_mcq->mcq.comp &&
2070 				    mqp->ibqp.recv_cq->comp_handler) {
2071 					if (!recv_mcq->mcq.reset_notify_added) {
2072 						recv_mcq->mcq.reset_notify_added = 1;
2073 						list_add_tail(&recv_mcq->mcq.reset_notify,
2074 							      &cq_armed_list);
2075 					}
2076 				}
2077 				spin_unlock_irqrestore(&recv_mcq->lock,
2078 						       flags_cq);
2079 			}
2080 		}
2081 		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2082 	}
2083 	/*At that point all inflight post send were put to be executed as of we
2084 	 * lock/unlock above locks Now need to arm all involved CQs.
2085 	 */
2086 	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2087 		mcq->comp(mcq);
2088 	}
2089 	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2090 }
2091 
2092 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
2093 			  enum mlx5_dev_event event, unsigned long param)
2094 {
2095 	struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
2096 	struct ib_event ibev;
2097 
2098 	u8 port = 0;
2099 
2100 	switch (event) {
2101 	case MLX5_DEV_EVENT_SYS_ERROR:
2102 		ibdev->ib_active = false;
2103 		ibev.event = IB_EVENT_DEVICE_FATAL;
2104 		mlx5_ib_handle_internal_error(ibdev);
2105 		break;
2106 
2107 	case MLX5_DEV_EVENT_PORT_UP:
2108 		ibev.event = IB_EVENT_PORT_ACTIVE;
2109 		port = (u8)param;
2110 		break;
2111 
2112 	case MLX5_DEV_EVENT_PORT_DOWN:
2113 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
2114 		ibev.event = IB_EVENT_PORT_ERR;
2115 		port = (u8)param;
2116 		break;
2117 
2118 	case MLX5_DEV_EVENT_LID_CHANGE:
2119 		ibev.event = IB_EVENT_LID_CHANGE;
2120 		port = (u8)param;
2121 		break;
2122 
2123 	case MLX5_DEV_EVENT_PKEY_CHANGE:
2124 		ibev.event = IB_EVENT_PKEY_CHANGE;
2125 		port = (u8)param;
2126 
2127 		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2128 		break;
2129 
2130 	case MLX5_DEV_EVENT_GUID_CHANGE:
2131 		ibev.event = IB_EVENT_GID_CHANGE;
2132 		port = (u8)param;
2133 		break;
2134 
2135 	case MLX5_DEV_EVENT_CLIENT_REREG:
2136 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
2137 		port = (u8)param;
2138 		break;
2139 	}
2140 
2141 	ibev.device	      = &ibdev->ib_dev;
2142 	ibev.element.port_num = port;
2143 
2144 	if (port < 1 || port > ibdev->num_ports) {
2145 		mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
2146 		return;
2147 	}
2148 
2149 	if (ibdev->ib_active)
2150 		ib_dispatch_event(&ibev);
2151 }
2152 
2153 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2154 {
2155 	int port;
2156 
2157 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
2158 		mlx5_query_ext_port_caps(dev, port);
2159 }
2160 
2161 static int get_port_caps(struct mlx5_ib_dev *dev)
2162 {
2163 	struct ib_device_attr *dprops = NULL;
2164 	struct ib_port_attr *pprops = NULL;
2165 	int err = -ENOMEM;
2166 	int port;
2167 	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
2168 
2169 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
2170 	if (!pprops)
2171 		goto out;
2172 
2173 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2174 	if (!dprops)
2175 		goto out;
2176 
2177 	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
2178 	if (err) {
2179 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
2180 		goto out;
2181 	}
2182 
2183 	for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2184 		err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2185 		if (err) {
2186 			mlx5_ib_warn(dev, "query_port %d failed %d\n",
2187 				     port, err);
2188 			break;
2189 		}
2190 		dev->mdev->port_caps[port - 1].pkey_table_len =
2191 						dprops->max_pkeys;
2192 		dev->mdev->port_caps[port - 1].gid_table_len =
2193 						pprops->gid_tbl_len;
2194 		mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
2195 			    dprops->max_pkeys, pprops->gid_tbl_len);
2196 	}
2197 
2198 out:
2199 	kfree(pprops);
2200 	kfree(dprops);
2201 
2202 	return err;
2203 }
2204 
2205 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
2206 {
2207 	int err;
2208 
2209 	err = mlx5_mr_cache_cleanup(dev);
2210 	if (err)
2211 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
2212 
2213 	mlx5_ib_destroy_qp(dev->umrc.qp);
2214 	ib_free_cq(dev->umrc.cq);
2215 	ib_dealloc_pd(dev->umrc.pd);
2216 }
2217 
2218 enum {
2219 	MAX_UMR_WR = 128,
2220 };
2221 
2222 static int create_umr_res(struct mlx5_ib_dev *dev)
2223 {
2224 	struct ib_qp_init_attr *init_attr = NULL;
2225 	struct ib_qp_attr *attr = NULL;
2226 	struct ib_pd *pd;
2227 	struct ib_cq *cq;
2228 	struct ib_qp *qp;
2229 	int ret;
2230 
2231 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
2232 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
2233 	if (!attr || !init_attr) {
2234 		ret = -ENOMEM;
2235 		goto error_0;
2236 	}
2237 
2238 	pd = ib_alloc_pd(&dev->ib_dev);
2239 	if (IS_ERR(pd)) {
2240 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
2241 		ret = PTR_ERR(pd);
2242 		goto error_0;
2243 	}
2244 
2245 	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
2246 	if (IS_ERR(cq)) {
2247 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
2248 		ret = PTR_ERR(cq);
2249 		goto error_2;
2250 	}
2251 
2252 	init_attr->send_cq = cq;
2253 	init_attr->recv_cq = cq;
2254 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
2255 	init_attr->cap.max_send_wr = MAX_UMR_WR;
2256 	init_attr->cap.max_send_sge = 1;
2257 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
2258 	init_attr->port_num = 1;
2259 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
2260 	if (IS_ERR(qp)) {
2261 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
2262 		ret = PTR_ERR(qp);
2263 		goto error_3;
2264 	}
2265 	qp->device     = &dev->ib_dev;
2266 	qp->real_qp    = qp;
2267 	qp->uobject    = NULL;
2268 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
2269 
2270 	attr->qp_state = IB_QPS_INIT;
2271 	attr->port_num = 1;
2272 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
2273 				IB_QP_PORT, NULL);
2274 	if (ret) {
2275 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
2276 		goto error_4;
2277 	}
2278 
2279 	memset(attr, 0, sizeof(*attr));
2280 	attr->qp_state = IB_QPS_RTR;
2281 	attr->path_mtu = IB_MTU_256;
2282 
2283 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2284 	if (ret) {
2285 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
2286 		goto error_4;
2287 	}
2288 
2289 	memset(attr, 0, sizeof(*attr));
2290 	attr->qp_state = IB_QPS_RTS;
2291 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2292 	if (ret) {
2293 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
2294 		goto error_4;
2295 	}
2296 
2297 	dev->umrc.qp = qp;
2298 	dev->umrc.cq = cq;
2299 	dev->umrc.pd = pd;
2300 
2301 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
2302 	ret = mlx5_mr_cache_init(dev);
2303 	if (ret) {
2304 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
2305 		goto error_4;
2306 	}
2307 
2308 	kfree(attr);
2309 	kfree(init_attr);
2310 
2311 	return 0;
2312 
2313 error_4:
2314 	mlx5_ib_destroy_qp(qp);
2315 
2316 error_3:
2317 	ib_free_cq(cq);
2318 
2319 error_2:
2320 	ib_dealloc_pd(pd);
2321 
2322 error_0:
2323 	kfree(attr);
2324 	kfree(init_attr);
2325 	return ret;
2326 }
2327 
2328 static int create_dev_resources(struct mlx5_ib_resources *devr)
2329 {
2330 	struct ib_srq_init_attr attr;
2331 	struct mlx5_ib_dev *dev;
2332 	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2333 	int port;
2334 	int ret = 0;
2335 
2336 	dev = container_of(devr, struct mlx5_ib_dev, devr);
2337 
2338 	mutex_init(&devr->mutex);
2339 
2340 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
2341 	if (IS_ERR(devr->p0)) {
2342 		ret = PTR_ERR(devr->p0);
2343 		goto error0;
2344 	}
2345 	devr->p0->device  = &dev->ib_dev;
2346 	devr->p0->uobject = NULL;
2347 	atomic_set(&devr->p0->usecnt, 0);
2348 
2349 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
2350 	if (IS_ERR(devr->c0)) {
2351 		ret = PTR_ERR(devr->c0);
2352 		goto error1;
2353 	}
2354 	devr->c0->device        = &dev->ib_dev;
2355 	devr->c0->uobject       = NULL;
2356 	devr->c0->comp_handler  = NULL;
2357 	devr->c0->event_handler = NULL;
2358 	devr->c0->cq_context    = NULL;
2359 	atomic_set(&devr->c0->usecnt, 0);
2360 
2361 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2362 	if (IS_ERR(devr->x0)) {
2363 		ret = PTR_ERR(devr->x0);
2364 		goto error2;
2365 	}
2366 	devr->x0->device = &dev->ib_dev;
2367 	devr->x0->inode = NULL;
2368 	atomic_set(&devr->x0->usecnt, 0);
2369 	mutex_init(&devr->x0->tgt_qp_mutex);
2370 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
2371 
2372 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2373 	if (IS_ERR(devr->x1)) {
2374 		ret = PTR_ERR(devr->x1);
2375 		goto error3;
2376 	}
2377 	devr->x1->device = &dev->ib_dev;
2378 	devr->x1->inode = NULL;
2379 	atomic_set(&devr->x1->usecnt, 0);
2380 	mutex_init(&devr->x1->tgt_qp_mutex);
2381 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
2382 
2383 	memset(&attr, 0, sizeof(attr));
2384 	attr.attr.max_sge = 1;
2385 	attr.attr.max_wr = 1;
2386 	attr.srq_type = IB_SRQT_XRC;
2387 	attr.ext.xrc.cq = devr->c0;
2388 	attr.ext.xrc.xrcd = devr->x0;
2389 
2390 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2391 	if (IS_ERR(devr->s0)) {
2392 		ret = PTR_ERR(devr->s0);
2393 		goto error4;
2394 	}
2395 	devr->s0->device	= &dev->ib_dev;
2396 	devr->s0->pd		= devr->p0;
2397 	devr->s0->uobject       = NULL;
2398 	devr->s0->event_handler = NULL;
2399 	devr->s0->srq_context   = NULL;
2400 	devr->s0->srq_type      = IB_SRQT_XRC;
2401 	devr->s0->ext.xrc.xrcd	= devr->x0;
2402 	devr->s0->ext.xrc.cq	= devr->c0;
2403 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
2404 	atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
2405 	atomic_inc(&devr->p0->usecnt);
2406 	atomic_set(&devr->s0->usecnt, 0);
2407 
2408 	memset(&attr, 0, sizeof(attr));
2409 	attr.attr.max_sge = 1;
2410 	attr.attr.max_wr = 1;
2411 	attr.srq_type = IB_SRQT_BASIC;
2412 	devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2413 	if (IS_ERR(devr->s1)) {
2414 		ret = PTR_ERR(devr->s1);
2415 		goto error5;
2416 	}
2417 	devr->s1->device	= &dev->ib_dev;
2418 	devr->s1->pd		= devr->p0;
2419 	devr->s1->uobject       = NULL;
2420 	devr->s1->event_handler = NULL;
2421 	devr->s1->srq_context   = NULL;
2422 	devr->s1->srq_type      = IB_SRQT_BASIC;
2423 	devr->s1->ext.xrc.cq	= devr->c0;
2424 	atomic_inc(&devr->p0->usecnt);
2425 	atomic_set(&devr->s0->usecnt, 0);
2426 
2427 	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
2428 		INIT_WORK(&devr->ports[port].pkey_change_work,
2429 			  pkey_change_handler);
2430 		devr->ports[port].devr = devr;
2431 	}
2432 
2433 	return 0;
2434 
2435 error5:
2436 	mlx5_ib_destroy_srq(devr->s0);
2437 error4:
2438 	mlx5_ib_dealloc_xrcd(devr->x1);
2439 error3:
2440 	mlx5_ib_dealloc_xrcd(devr->x0);
2441 error2:
2442 	mlx5_ib_destroy_cq(devr->c0);
2443 error1:
2444 	mlx5_ib_dealloc_pd(devr->p0);
2445 error0:
2446 	return ret;
2447 }
2448 
2449 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
2450 {
2451 	struct mlx5_ib_dev *dev =
2452 		container_of(devr, struct mlx5_ib_dev, devr);
2453 	int port;
2454 
2455 	mlx5_ib_destroy_srq(devr->s1);
2456 	mlx5_ib_destroy_srq(devr->s0);
2457 	mlx5_ib_dealloc_xrcd(devr->x0);
2458 	mlx5_ib_dealloc_xrcd(devr->x1);
2459 	mlx5_ib_destroy_cq(devr->c0);
2460 	mlx5_ib_dealloc_pd(devr->p0);
2461 
2462 	/* Make sure no change P_Key work items are still executing */
2463 	for (port = 0; port < dev->num_ports; ++port)
2464 		cancel_work_sync(&devr->ports[port].pkey_change_work);
2465 }
2466 
2467 static u32 get_core_cap_flags(struct ib_device *ibdev)
2468 {
2469 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2470 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2471 	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2472 	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2473 	u32 ret = 0;
2474 
2475 	if (ll == IB_LINK_LAYER_INFINIBAND)
2476 		return RDMA_CORE_PORT_IBA_IB;
2477 
2478 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2479 		return 0;
2480 
2481 	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2482 		return 0;
2483 
2484 	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2485 		ret |= RDMA_CORE_PORT_IBA_ROCE;
2486 
2487 	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2488 		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2489 
2490 	return ret;
2491 }
2492 
2493 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
2494 			       struct ib_port_immutable *immutable)
2495 {
2496 	struct ib_port_attr attr;
2497 	int err;
2498 
2499 	err = mlx5_ib_query_port(ibdev, port_num, &attr);
2500 	if (err)
2501 		return err;
2502 
2503 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2504 	immutable->gid_tbl_len = attr.gid_tbl_len;
2505 	immutable->core_cap_flags = get_core_cap_flags(ibdev);
2506 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2507 
2508 	return 0;
2509 }
2510 
2511 static void get_dev_fw_str(struct ib_device *ibdev, char *str,
2512 			   size_t str_len)
2513 {
2514 	struct mlx5_ib_dev *dev =
2515 		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2516 	snprintf(str, str_len, "%d.%d.%04d", fw_rev_maj(dev->mdev),
2517 		       fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev));
2518 }
2519 
2520 static int mlx5_enable_roce(struct mlx5_ib_dev *dev)
2521 {
2522 	int err;
2523 
2524 	dev->roce.nb.notifier_call = mlx5_netdev_event;
2525 	err = register_netdevice_notifier(&dev->roce.nb);
2526 	if (err)
2527 		return err;
2528 
2529 	err = mlx5_nic_vport_enable_roce(dev->mdev);
2530 	if (err)
2531 		goto err_unregister_netdevice_notifier;
2532 
2533 	return 0;
2534 
2535 err_unregister_netdevice_notifier:
2536 	unregister_netdevice_notifier(&dev->roce.nb);
2537 	return err;
2538 }
2539 
2540 static void mlx5_disable_roce(struct mlx5_ib_dev *dev)
2541 {
2542 	mlx5_nic_vport_disable_roce(dev->mdev);
2543 	unregister_netdevice_notifier(&dev->roce.nb);
2544 }
2545 
2546 static void mlx5_ib_dealloc_q_counters(struct mlx5_ib_dev *dev)
2547 {
2548 	unsigned int i;
2549 
2550 	for (i = 0; i < dev->num_ports; i++)
2551 		mlx5_core_dealloc_q_counter(dev->mdev,
2552 					    dev->port[i].q_cnt_id);
2553 }
2554 
2555 static int mlx5_ib_alloc_q_counters(struct mlx5_ib_dev *dev)
2556 {
2557 	int i;
2558 	int ret;
2559 
2560 	for (i = 0; i < dev->num_ports; i++) {
2561 		ret = mlx5_core_alloc_q_counter(dev->mdev,
2562 						&dev->port[i].q_cnt_id);
2563 		if (ret) {
2564 			mlx5_ib_warn(dev,
2565 				     "couldn't allocate queue counter for port %d, err %d\n",
2566 				     i + 1, ret);
2567 			goto dealloc_counters;
2568 		}
2569 	}
2570 
2571 	return 0;
2572 
2573 dealloc_counters:
2574 	while (--i >= 0)
2575 		mlx5_core_dealloc_q_counter(dev->mdev,
2576 					    dev->port[i].q_cnt_id);
2577 
2578 	return ret;
2579 }
2580 
2581 static const char * const names[] = {
2582 	"rx_write_requests",
2583 	"rx_read_requests",
2584 	"rx_atomic_requests",
2585 	"out_of_buffer",
2586 	"out_of_sequence",
2587 	"duplicate_request",
2588 	"rnr_nak_retry_err",
2589 	"packet_seq_err",
2590 	"implied_nak_seq_err",
2591 	"local_ack_timeout_err",
2592 };
2593 
2594 static const size_t stats_offsets[] = {
2595 	MLX5_BYTE_OFF(query_q_counter_out, rx_write_requests),
2596 	MLX5_BYTE_OFF(query_q_counter_out, rx_read_requests),
2597 	MLX5_BYTE_OFF(query_q_counter_out, rx_atomic_requests),
2598 	MLX5_BYTE_OFF(query_q_counter_out, out_of_buffer),
2599 	MLX5_BYTE_OFF(query_q_counter_out, out_of_sequence),
2600 	MLX5_BYTE_OFF(query_q_counter_out, duplicate_request),
2601 	MLX5_BYTE_OFF(query_q_counter_out, rnr_nak_retry_err),
2602 	MLX5_BYTE_OFF(query_q_counter_out, packet_seq_err),
2603 	MLX5_BYTE_OFF(query_q_counter_out, implied_nak_seq_err),
2604 	MLX5_BYTE_OFF(query_q_counter_out, local_ack_timeout_err),
2605 };
2606 
2607 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
2608 						    u8 port_num)
2609 {
2610 	BUILD_BUG_ON(ARRAY_SIZE(names) != ARRAY_SIZE(stats_offsets));
2611 
2612 	/* We support only per port stats */
2613 	if (port_num == 0)
2614 		return NULL;
2615 
2616 	return rdma_alloc_hw_stats_struct(names, ARRAY_SIZE(names),
2617 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
2618 }
2619 
2620 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
2621 				struct rdma_hw_stats *stats,
2622 				u8 port, int index)
2623 {
2624 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2625 	int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
2626 	void *out;
2627 	__be32 val;
2628 	int ret;
2629 	int i;
2630 
2631 	if (!port || !stats)
2632 		return -ENOSYS;
2633 
2634 	out = mlx5_vzalloc(outlen);
2635 	if (!out)
2636 		return -ENOMEM;
2637 
2638 	ret = mlx5_core_query_q_counter(dev->mdev,
2639 					dev->port[port - 1].q_cnt_id, 0,
2640 					out, outlen);
2641 	if (ret)
2642 		goto free;
2643 
2644 	for (i = 0; i < ARRAY_SIZE(names); i++) {
2645 		val = *(__be32 *)(out + stats_offsets[i]);
2646 		stats->value[i] = (u64)be32_to_cpu(val);
2647 	}
2648 free:
2649 	kvfree(out);
2650 	return ARRAY_SIZE(names);
2651 }
2652 
2653 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
2654 {
2655 	struct mlx5_ib_dev *dev;
2656 	enum rdma_link_layer ll;
2657 	int port_type_cap;
2658 	int err;
2659 	int i;
2660 
2661 	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
2662 	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
2663 
2664 	if ((ll == IB_LINK_LAYER_ETHERNET) && !MLX5_CAP_GEN(mdev, roce))
2665 		return NULL;
2666 
2667 	printk_once(KERN_INFO "%s", mlx5_version);
2668 
2669 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
2670 	if (!dev)
2671 		return NULL;
2672 
2673 	dev->mdev = mdev;
2674 
2675 	dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port),
2676 			    GFP_KERNEL);
2677 	if (!dev->port)
2678 		goto err_dealloc;
2679 
2680 	rwlock_init(&dev->roce.netdev_lock);
2681 	err = get_port_caps(dev);
2682 	if (err)
2683 		goto err_free_port;
2684 
2685 	if (mlx5_use_mad_ifc(dev))
2686 		get_ext_port_caps(dev);
2687 
2688 	MLX5_INIT_DOORBELL_LOCK(&dev->uar_lock);
2689 
2690 	strlcpy(dev->ib_dev.name, "mlx5_%d", IB_DEVICE_NAME_MAX);
2691 	dev->ib_dev.owner		= THIS_MODULE;
2692 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
2693 	dev->ib_dev.local_dma_lkey	= 0 /* not supported for now */;
2694 	dev->num_ports		= MLX5_CAP_GEN(mdev, num_ports);
2695 	dev->ib_dev.phys_port_cnt     = dev->num_ports;
2696 	dev->ib_dev.num_comp_vectors    =
2697 		dev->mdev->priv.eq_table.num_comp_vectors;
2698 	dev->ib_dev.dma_device	= &mdev->pdev->dev;
2699 
2700 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
2701 	dev->ib_dev.uverbs_cmd_mask	=
2702 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
2703 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
2704 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
2705 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
2706 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
2707 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
2708 		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
2709 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
2710 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
2711 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
2712 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
2713 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
2714 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
2715 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
2716 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
2717 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
2718 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
2719 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
2720 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
2721 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
2722 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
2723 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
2724 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
2725 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
2726 	dev->ib_dev.uverbs_ex_cmd_mask =
2727 		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)	|
2728 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)	|
2729 		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP);
2730 
2731 	dev->ib_dev.query_device	= mlx5_ib_query_device;
2732 	dev->ib_dev.query_port		= mlx5_ib_query_port;
2733 	dev->ib_dev.get_link_layer	= mlx5_ib_port_link_layer;
2734 	if (ll == IB_LINK_LAYER_ETHERNET)
2735 		dev->ib_dev.get_netdev	= mlx5_ib_get_netdev;
2736 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
2737 	dev->ib_dev.add_gid		= mlx5_ib_add_gid;
2738 	dev->ib_dev.del_gid		= mlx5_ib_del_gid;
2739 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
2740 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
2741 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
2742 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
2743 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
2744 	dev->ib_dev.mmap		= mlx5_ib_mmap;
2745 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
2746 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
2747 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
2748 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
2749 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
2750 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
2751 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
2752 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
2753 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
2754 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
2755 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
2756 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
2757 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
2758 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
2759 	dev->ib_dev.post_send		= mlx5_ib_post_send;
2760 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
2761 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
2762 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
2763 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
2764 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
2765 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
2766 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
2767 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
2768 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
2769 	dev->ib_dev.rereg_user_mr	= mlx5_ib_rereg_user_mr;
2770 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
2771 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
2772 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
2773 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
2774 	dev->ib_dev.alloc_mr		= mlx5_ib_alloc_mr;
2775 	dev->ib_dev.map_mr_sg		= mlx5_ib_map_mr_sg;
2776 	dev->ib_dev.check_mr_status	= mlx5_ib_check_mr_status;
2777 	dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
2778 	dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
2779 	if (mlx5_core_is_pf(mdev)) {
2780 		dev->ib_dev.get_vf_config	= mlx5_ib_get_vf_config;
2781 		dev->ib_dev.set_vf_link_state	= mlx5_ib_set_vf_link_state;
2782 		dev->ib_dev.get_vf_stats	= mlx5_ib_get_vf_stats;
2783 		dev->ib_dev.set_vf_guid		= mlx5_ib_set_vf_guid;
2784 	}
2785 
2786 	dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
2787 
2788 	mlx5_ib_internal_fill_odp_caps(dev);
2789 
2790 	if (MLX5_CAP_GEN(mdev, imaicl)) {
2791 		dev->ib_dev.alloc_mw		= mlx5_ib_alloc_mw;
2792 		dev->ib_dev.dealloc_mw		= mlx5_ib_dealloc_mw;
2793 		dev->ib_dev.uverbs_cmd_mask |=
2794 			(1ull << IB_USER_VERBS_CMD_ALLOC_MW)	|
2795 			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
2796 	}
2797 
2798 	if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt) &&
2799 	    MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
2800 		dev->ib_dev.get_hw_stats	= mlx5_ib_get_hw_stats;
2801 		dev->ib_dev.alloc_hw_stats	= mlx5_ib_alloc_hw_stats;
2802 	}
2803 
2804 	if (MLX5_CAP_GEN(mdev, xrc)) {
2805 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
2806 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
2807 		dev->ib_dev.uverbs_cmd_mask |=
2808 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
2809 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
2810 	}
2811 
2812 	if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
2813 	    IB_LINK_LAYER_ETHERNET) {
2814 		dev->ib_dev.create_flow	= mlx5_ib_create_flow;
2815 		dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
2816 		dev->ib_dev.create_wq	 = mlx5_ib_create_wq;
2817 		dev->ib_dev.modify_wq	 = mlx5_ib_modify_wq;
2818 		dev->ib_dev.destroy_wq	 = mlx5_ib_destroy_wq;
2819 		dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
2820 		dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
2821 		dev->ib_dev.uverbs_ex_cmd_mask |=
2822 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
2823 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW) |
2824 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
2825 			(1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
2826 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
2827 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
2828 			(1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
2829 	}
2830 	err = init_node_data(dev);
2831 	if (err)
2832 		goto err_dealloc;
2833 
2834 	mutex_init(&dev->flow_db.lock);
2835 	mutex_init(&dev->cap_mask_mutex);
2836 	INIT_LIST_HEAD(&dev->qp_list);
2837 	spin_lock_init(&dev->reset_flow_resource_lock);
2838 
2839 	if (ll == IB_LINK_LAYER_ETHERNET) {
2840 		err = mlx5_enable_roce(dev);
2841 		if (err)
2842 			goto err_dealloc;
2843 	}
2844 
2845 	err = create_dev_resources(&dev->devr);
2846 	if (err)
2847 		goto err_disable_roce;
2848 
2849 	err = mlx5_ib_odp_init_one(dev);
2850 	if (err)
2851 		goto err_rsrc;
2852 
2853 	err = mlx5_ib_alloc_q_counters(dev);
2854 	if (err)
2855 		goto err_odp;
2856 
2857 	err = ib_register_device(&dev->ib_dev, NULL);
2858 	if (err)
2859 		goto err_q_cnt;
2860 
2861 	err = create_umr_res(dev);
2862 	if (err)
2863 		goto err_dev;
2864 
2865 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
2866 		err = device_create_file(&dev->ib_dev.dev,
2867 					 mlx5_class_attributes[i]);
2868 		if (err)
2869 			goto err_umrc;
2870 	}
2871 
2872 	dev->ib_active = true;
2873 
2874 	return dev;
2875 
2876 err_umrc:
2877 	destroy_umrc_res(dev);
2878 
2879 err_dev:
2880 	ib_unregister_device(&dev->ib_dev);
2881 
2882 err_q_cnt:
2883 	mlx5_ib_dealloc_q_counters(dev);
2884 
2885 err_odp:
2886 	mlx5_ib_odp_remove_one(dev);
2887 
2888 err_rsrc:
2889 	destroy_dev_resources(&dev->devr);
2890 
2891 err_disable_roce:
2892 	if (ll == IB_LINK_LAYER_ETHERNET)
2893 		mlx5_disable_roce(dev);
2894 
2895 err_free_port:
2896 	kfree(dev->port);
2897 
2898 err_dealloc:
2899 	ib_dealloc_device((struct ib_device *)dev);
2900 
2901 	return NULL;
2902 }
2903 
2904 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
2905 {
2906 	struct mlx5_ib_dev *dev = context;
2907 	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
2908 
2909 	ib_unregister_device(&dev->ib_dev);
2910 	mlx5_ib_dealloc_q_counters(dev);
2911 	destroy_umrc_res(dev);
2912 	mlx5_ib_odp_remove_one(dev);
2913 	destroy_dev_resources(&dev->devr);
2914 	if (ll == IB_LINK_LAYER_ETHERNET)
2915 		mlx5_disable_roce(dev);
2916 	kfree(dev->port);
2917 	ib_dealloc_device(&dev->ib_dev);
2918 }
2919 
2920 static struct mlx5_interface mlx5_ib_interface = {
2921 	.add            = mlx5_ib_add,
2922 	.remove         = mlx5_ib_remove,
2923 	.event          = mlx5_ib_event,
2924 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
2925 };
2926 
2927 static int __init mlx5_ib_init(void)
2928 {
2929 	int err;
2930 
2931 	if (deprecated_prof_sel != 2)
2932 		pr_warn("prof_sel is deprecated for mlx5_ib, set it for mlx5_core\n");
2933 
2934 	err = mlx5_ib_odp_init();
2935 	if (err)
2936 		return err;
2937 
2938 	err = mlx5_register_interface(&mlx5_ib_interface);
2939 	if (err)
2940 		goto clean_odp;
2941 
2942 	return err;
2943 
2944 clean_odp:
2945 	mlx5_ib_odp_cleanup();
2946 	return err;
2947 }
2948 
2949 static void __exit mlx5_ib_cleanup(void)
2950 {
2951 	mlx5_unregister_interface(&mlx5_ib_interface);
2952 	mlx5_ib_odp_cleanup();
2953 }
2954 
2955 module_init(mlx5_ib_init);
2956 module_exit(mlx5_ib_cleanup);
2957