xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision 37be287c)
1 /*
2  * Copyright (c) 2013, Mellanox Technologies inc.  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 <asm-generic/kmap_types.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 #include <linux/io-mapping.h>
41 #include <linux/sched.h>
42 #include <rdma/ib_user_verbs.h>
43 #include <rdma/ib_smi.h>
44 #include <rdma/ib_umem.h>
45 #include "user.h"
46 #include "mlx5_ib.h"
47 
48 #define DRIVER_NAME "mlx5_ib"
49 #define DRIVER_VERSION "1.0"
50 #define DRIVER_RELDATE	"June 2013"
51 
52 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
53 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
54 MODULE_LICENSE("Dual BSD/GPL");
55 MODULE_VERSION(DRIVER_VERSION);
56 
57 static int prof_sel = 2;
58 module_param_named(prof_sel, prof_sel, int, 0444);
59 MODULE_PARM_DESC(prof_sel, "profile selector. Valid range 0 - 2");
60 
61 static char mlx5_version[] =
62 	DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
63 	DRIVER_VERSION " (" DRIVER_RELDATE ")\n";
64 
65 static struct mlx5_profile profile[] = {
66 	[0] = {
67 		.mask		= 0,
68 	},
69 	[1] = {
70 		.mask		= MLX5_PROF_MASK_QP_SIZE,
71 		.log_max_qp	= 12,
72 	},
73 	[2] = {
74 		.mask		= MLX5_PROF_MASK_QP_SIZE |
75 				  MLX5_PROF_MASK_MR_CACHE,
76 		.log_max_qp	= 17,
77 		.mr_cache[0]	= {
78 			.size	= 500,
79 			.limit	= 250
80 		},
81 		.mr_cache[1]	= {
82 			.size	= 500,
83 			.limit	= 250
84 		},
85 		.mr_cache[2]	= {
86 			.size	= 500,
87 			.limit	= 250
88 		},
89 		.mr_cache[3]	= {
90 			.size	= 500,
91 			.limit	= 250
92 		},
93 		.mr_cache[4]	= {
94 			.size	= 500,
95 			.limit	= 250
96 		},
97 		.mr_cache[5]	= {
98 			.size	= 500,
99 			.limit	= 250
100 		},
101 		.mr_cache[6]	= {
102 			.size	= 500,
103 			.limit	= 250
104 		},
105 		.mr_cache[7]	= {
106 			.size	= 500,
107 			.limit	= 250
108 		},
109 		.mr_cache[8]	= {
110 			.size	= 500,
111 			.limit	= 250
112 		},
113 		.mr_cache[9]	= {
114 			.size	= 500,
115 			.limit	= 250
116 		},
117 		.mr_cache[10]	= {
118 			.size	= 500,
119 			.limit	= 250
120 		},
121 		.mr_cache[11]	= {
122 			.size	= 500,
123 			.limit	= 250
124 		},
125 		.mr_cache[12]	= {
126 			.size	= 64,
127 			.limit	= 32
128 		},
129 		.mr_cache[13]	= {
130 			.size	= 32,
131 			.limit	= 16
132 		},
133 		.mr_cache[14]	= {
134 			.size	= 16,
135 			.limit	= 8
136 		},
137 		.mr_cache[15]	= {
138 			.size	= 8,
139 			.limit	= 4
140 		},
141 	},
142 };
143 
144 int mlx5_vector2eqn(struct mlx5_ib_dev *dev, int vector, int *eqn, int *irqn)
145 {
146 	struct mlx5_eq_table *table = &dev->mdev.priv.eq_table;
147 	struct mlx5_eq *eq, *n;
148 	int err = -ENOENT;
149 
150 	spin_lock(&table->lock);
151 	list_for_each_entry_safe(eq, n, &dev->eqs_list, list) {
152 		if (eq->index == vector) {
153 			*eqn = eq->eqn;
154 			*irqn = eq->irqn;
155 			err = 0;
156 			break;
157 		}
158 	}
159 	spin_unlock(&table->lock);
160 
161 	return err;
162 }
163 
164 static int alloc_comp_eqs(struct mlx5_ib_dev *dev)
165 {
166 	struct mlx5_eq_table *table = &dev->mdev.priv.eq_table;
167 	char name[MLX5_MAX_EQ_NAME];
168 	struct mlx5_eq *eq, *n;
169 	int ncomp_vec;
170 	int nent;
171 	int err;
172 	int i;
173 
174 	INIT_LIST_HEAD(&dev->eqs_list);
175 	ncomp_vec = table->num_comp_vectors;
176 	nent = MLX5_COMP_EQ_SIZE;
177 	for (i = 0; i < ncomp_vec; i++) {
178 		eq = kzalloc(sizeof(*eq), GFP_KERNEL);
179 		if (!eq) {
180 			err = -ENOMEM;
181 			goto clean;
182 		}
183 
184 		snprintf(name, MLX5_MAX_EQ_NAME, "mlx5_comp%d", i);
185 		err = mlx5_create_map_eq(&dev->mdev, eq,
186 					 i + MLX5_EQ_VEC_COMP_BASE, nent, 0,
187 					 name, &dev->mdev.priv.uuari.uars[0]);
188 		if (err) {
189 			kfree(eq);
190 			goto clean;
191 		}
192 		mlx5_ib_dbg(dev, "allocated completion EQN %d\n", eq->eqn);
193 		eq->index = i;
194 		spin_lock(&table->lock);
195 		list_add_tail(&eq->list, &dev->eqs_list);
196 		spin_unlock(&table->lock);
197 	}
198 
199 	dev->num_comp_vectors = ncomp_vec;
200 	return 0;
201 
202 clean:
203 	spin_lock(&table->lock);
204 	list_for_each_entry_safe(eq, n, &dev->eqs_list, list) {
205 		list_del(&eq->list);
206 		spin_unlock(&table->lock);
207 		if (mlx5_destroy_unmap_eq(&dev->mdev, eq))
208 			mlx5_ib_warn(dev, "failed to destroy EQ 0x%x\n", eq->eqn);
209 		kfree(eq);
210 		spin_lock(&table->lock);
211 	}
212 	spin_unlock(&table->lock);
213 	return err;
214 }
215 
216 static void free_comp_eqs(struct mlx5_ib_dev *dev)
217 {
218 	struct mlx5_eq_table *table = &dev->mdev.priv.eq_table;
219 	struct mlx5_eq *eq, *n;
220 
221 	spin_lock(&table->lock);
222 	list_for_each_entry_safe(eq, n, &dev->eqs_list, list) {
223 		list_del(&eq->list);
224 		spin_unlock(&table->lock);
225 		if (mlx5_destroy_unmap_eq(&dev->mdev, eq))
226 			mlx5_ib_warn(dev, "failed to destroy EQ 0x%x\n", eq->eqn);
227 		kfree(eq);
228 		spin_lock(&table->lock);
229 	}
230 	spin_unlock(&table->lock);
231 }
232 
233 static int mlx5_ib_query_device(struct ib_device *ibdev,
234 				struct ib_device_attr *props)
235 {
236 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
237 	struct ib_smp *in_mad  = NULL;
238 	struct ib_smp *out_mad = NULL;
239 	int err = -ENOMEM;
240 	int max_rq_sg;
241 	int max_sq_sg;
242 	u64 flags;
243 
244 	in_mad  = kzalloc(sizeof(*in_mad), GFP_KERNEL);
245 	out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
246 	if (!in_mad || !out_mad)
247 		goto out;
248 
249 	init_query_mad(in_mad);
250 	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
251 
252 	err = mlx5_MAD_IFC(to_mdev(ibdev), 1, 1, 1, NULL, NULL, in_mad, out_mad);
253 	if (err)
254 		goto out;
255 
256 	memset(props, 0, sizeof(*props));
257 
258 	props->fw_ver = ((u64)fw_rev_maj(&dev->mdev) << 32) |
259 		(fw_rev_min(&dev->mdev) << 16) |
260 		fw_rev_sub(&dev->mdev);
261 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
262 		IB_DEVICE_PORT_ACTIVE_EVENT		|
263 		IB_DEVICE_SYS_IMAGE_GUID		|
264 		IB_DEVICE_RC_RNR_NAK_GEN		|
265 		IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
266 	flags = dev->mdev.caps.flags;
267 	if (flags & MLX5_DEV_CAP_FLAG_BAD_PKEY_CNTR)
268 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
269 	if (flags & MLX5_DEV_CAP_FLAG_BAD_QKEY_CNTR)
270 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
271 	if (flags & MLX5_DEV_CAP_FLAG_APM)
272 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
273 	props->device_cap_flags |= IB_DEVICE_LOCAL_DMA_LKEY;
274 	if (flags & MLX5_DEV_CAP_FLAG_XRC)
275 		props->device_cap_flags |= IB_DEVICE_XRC;
276 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
277 
278 	props->vendor_id	   = be32_to_cpup((__be32 *)(out_mad->data + 36)) &
279 		0xffffff;
280 	props->vendor_part_id	   = be16_to_cpup((__be16 *)(out_mad->data + 30));
281 	props->hw_ver		   = be32_to_cpup((__be32 *)(out_mad->data + 32));
282 	memcpy(&props->sys_image_guid, out_mad->data +	4, 8);
283 
284 	props->max_mr_size	   = ~0ull;
285 	props->page_size_cap	   = dev->mdev.caps.min_page_sz;
286 	props->max_qp		   = 1 << dev->mdev.caps.log_max_qp;
287 	props->max_qp_wr	   = dev->mdev.caps.max_wqes;
288 	max_rq_sg = dev->mdev.caps.max_rq_desc_sz / sizeof(struct mlx5_wqe_data_seg);
289 	max_sq_sg = (dev->mdev.caps.max_sq_desc_sz - sizeof(struct mlx5_wqe_ctrl_seg)) /
290 		sizeof(struct mlx5_wqe_data_seg);
291 	props->max_sge = min(max_rq_sg, max_sq_sg);
292 	props->max_cq		   = 1 << dev->mdev.caps.log_max_cq;
293 	props->max_cqe		   = dev->mdev.caps.max_cqes - 1;
294 	props->max_mr		   = 1 << dev->mdev.caps.log_max_mkey;
295 	props->max_pd		   = 1 << dev->mdev.caps.log_max_pd;
296 	props->max_qp_rd_atom	   = dev->mdev.caps.max_ra_req_qp;
297 	props->max_qp_init_rd_atom = dev->mdev.caps.max_ra_res_qp;
298 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
299 	props->max_srq		   = 1 << dev->mdev.caps.log_max_srq;
300 	props->max_srq_wr	   = dev->mdev.caps.max_srq_wqes - 1;
301 	props->max_srq_sge	   = max_rq_sg - 1;
302 	props->max_fast_reg_page_list_len = (unsigned int)-1;
303 	props->local_ca_ack_delay  = dev->mdev.caps.local_ca_ack_delay;
304 	props->atomic_cap	   = IB_ATOMIC_NONE;
305 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
306 	props->max_pkeys	   = be16_to_cpup((__be16 *)(out_mad->data + 28));
307 	props->max_mcast_grp	   = 1 << dev->mdev.caps.log_max_mcg;
308 	props->max_mcast_qp_attach = dev->mdev.caps.max_qp_mcg;
309 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
310 					   props->max_mcast_grp;
311 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
312 
313 out:
314 	kfree(in_mad);
315 	kfree(out_mad);
316 
317 	return err;
318 }
319 
320 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
321 		       struct ib_port_attr *props)
322 {
323 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
324 	struct ib_smp *in_mad  = NULL;
325 	struct ib_smp *out_mad = NULL;
326 	int ext_active_speed;
327 	int err = -ENOMEM;
328 
329 	if (port < 1 || port > dev->mdev.caps.num_ports) {
330 		mlx5_ib_warn(dev, "invalid port number %d\n", port);
331 		return -EINVAL;
332 	}
333 
334 	in_mad  = kzalloc(sizeof(*in_mad), GFP_KERNEL);
335 	out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
336 	if (!in_mad || !out_mad)
337 		goto out;
338 
339 	memset(props, 0, sizeof(*props));
340 
341 	init_query_mad(in_mad);
342 	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
343 	in_mad->attr_mod = cpu_to_be32(port);
344 
345 	err = mlx5_MAD_IFC(dev, 1, 1, port, NULL, NULL, in_mad, out_mad);
346 	if (err) {
347 		mlx5_ib_warn(dev, "err %d\n", err);
348 		goto out;
349 	}
350 
351 
352 	props->lid		= be16_to_cpup((__be16 *)(out_mad->data + 16));
353 	props->lmc		= out_mad->data[34] & 0x7;
354 	props->sm_lid		= be16_to_cpup((__be16 *)(out_mad->data + 18));
355 	props->sm_sl		= out_mad->data[36] & 0xf;
356 	props->state		= out_mad->data[32] & 0xf;
357 	props->phys_state	= out_mad->data[33] >> 4;
358 	props->port_cap_flags	= be32_to_cpup((__be32 *)(out_mad->data + 20));
359 	props->gid_tbl_len	= out_mad->data[50];
360 	props->max_msg_sz	= 1 << to_mdev(ibdev)->mdev.caps.log_max_msg;
361 	props->pkey_tbl_len	= to_mdev(ibdev)->mdev.caps.port[port - 1].pkey_table_len;
362 	props->bad_pkey_cntr	= be16_to_cpup((__be16 *)(out_mad->data + 46));
363 	props->qkey_viol_cntr	= be16_to_cpup((__be16 *)(out_mad->data + 48));
364 	props->active_width	= out_mad->data[31] & 0xf;
365 	props->active_speed	= out_mad->data[35] >> 4;
366 	props->max_mtu		= out_mad->data[41] & 0xf;
367 	props->active_mtu	= out_mad->data[36] >> 4;
368 	props->subnet_timeout	= out_mad->data[51] & 0x1f;
369 	props->max_vl_num	= out_mad->data[37] >> 4;
370 	props->init_type_reply	= out_mad->data[41] >> 4;
371 
372 	/* Check if extended speeds (EDR/FDR/...) are supported */
373 	if (props->port_cap_flags & IB_PORT_EXTENDED_SPEEDS_SUP) {
374 		ext_active_speed = out_mad->data[62] >> 4;
375 
376 		switch (ext_active_speed) {
377 		case 1:
378 			props->active_speed = 16; /* FDR */
379 			break;
380 		case 2:
381 			props->active_speed = 32; /* EDR */
382 			break;
383 		}
384 	}
385 
386 	/* If reported active speed is QDR, check if is FDR-10 */
387 	if (props->active_speed == 4) {
388 		if (dev->mdev.caps.ext_port_cap[port - 1] &
389 		    MLX_EXT_PORT_CAP_FLAG_EXTENDED_PORT_INFO) {
390 			init_query_mad(in_mad);
391 			in_mad->attr_id = MLX5_ATTR_EXTENDED_PORT_INFO;
392 			in_mad->attr_mod = cpu_to_be32(port);
393 
394 			err = mlx5_MAD_IFC(dev, 1, 1, port,
395 					   NULL, NULL, in_mad, out_mad);
396 			if (err)
397 				goto out;
398 
399 			/* Checking LinkSpeedActive for FDR-10 */
400 			if (out_mad->data[15] & 0x1)
401 				props->active_speed = 8;
402 		}
403 	}
404 
405 out:
406 	kfree(in_mad);
407 	kfree(out_mad);
408 
409 	return err;
410 }
411 
412 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
413 			     union ib_gid *gid)
414 {
415 	struct ib_smp *in_mad  = NULL;
416 	struct ib_smp *out_mad = NULL;
417 	int err = -ENOMEM;
418 
419 	in_mad  = kzalloc(sizeof(*in_mad), GFP_KERNEL);
420 	out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
421 	if (!in_mad || !out_mad)
422 		goto out;
423 
424 	init_query_mad(in_mad);
425 	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
426 	in_mad->attr_mod = cpu_to_be32(port);
427 
428 	err = mlx5_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
429 	if (err)
430 		goto out;
431 
432 	memcpy(gid->raw, out_mad->data + 8, 8);
433 
434 	init_query_mad(in_mad);
435 	in_mad->attr_id  = IB_SMP_ATTR_GUID_INFO;
436 	in_mad->attr_mod = cpu_to_be32(index / 8);
437 
438 	err = mlx5_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
439 	if (err)
440 		goto out;
441 
442 	memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
443 
444 out:
445 	kfree(in_mad);
446 	kfree(out_mad);
447 	return err;
448 }
449 
450 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
451 			      u16 *pkey)
452 {
453 	struct ib_smp *in_mad  = NULL;
454 	struct ib_smp *out_mad = NULL;
455 	int err = -ENOMEM;
456 
457 	in_mad  = kzalloc(sizeof(*in_mad), GFP_KERNEL);
458 	out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
459 	if (!in_mad || !out_mad)
460 		goto out;
461 
462 	init_query_mad(in_mad);
463 	in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
464 	in_mad->attr_mod = cpu_to_be32(index / 32);
465 
466 	err = mlx5_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
467 	if (err)
468 		goto out;
469 
470 	*pkey = be16_to_cpu(((__be16 *)out_mad->data)[index % 32]);
471 
472 out:
473 	kfree(in_mad);
474 	kfree(out_mad);
475 	return err;
476 }
477 
478 struct mlx5_reg_node_desc {
479 	u8	desc[64];
480 };
481 
482 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
483 				 struct ib_device_modify *props)
484 {
485 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
486 	struct mlx5_reg_node_desc in;
487 	struct mlx5_reg_node_desc out;
488 	int err;
489 
490 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
491 		return -EOPNOTSUPP;
492 
493 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
494 		return 0;
495 
496 	/*
497 	 * If possible, pass node desc to FW, so it can generate
498 	 * a 144 trap.  If cmd fails, just ignore.
499 	 */
500 	memcpy(&in, props->node_desc, 64);
501 	err = mlx5_core_access_reg(&dev->mdev, &in, sizeof(in), &out,
502 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
503 	if (err)
504 		return err;
505 
506 	memcpy(ibdev->node_desc, props->node_desc, 64);
507 
508 	return err;
509 }
510 
511 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
512 			       struct ib_port_modify *props)
513 {
514 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
515 	struct ib_port_attr attr;
516 	u32 tmp;
517 	int err;
518 
519 	mutex_lock(&dev->cap_mask_mutex);
520 
521 	err = mlx5_ib_query_port(ibdev, port, &attr);
522 	if (err)
523 		goto out;
524 
525 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
526 		~props->clr_port_cap_mask;
527 
528 	err = mlx5_set_port_caps(&dev->mdev, port, tmp);
529 
530 out:
531 	mutex_unlock(&dev->cap_mask_mutex);
532 	return err;
533 }
534 
535 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
536 						  struct ib_udata *udata)
537 {
538 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
539 	struct mlx5_ib_alloc_ucontext_req req;
540 	struct mlx5_ib_alloc_ucontext_resp resp;
541 	struct mlx5_ib_ucontext *context;
542 	struct mlx5_uuar_info *uuari;
543 	struct mlx5_uar *uars;
544 	int gross_uuars;
545 	int num_uars;
546 	int uuarn;
547 	int err;
548 	int i;
549 
550 	if (!dev->ib_active)
551 		return ERR_PTR(-EAGAIN);
552 
553 	err = ib_copy_from_udata(&req, udata, sizeof(req));
554 	if (err)
555 		return ERR_PTR(err);
556 
557 	if (req.total_num_uuars > MLX5_MAX_UUARS)
558 		return ERR_PTR(-ENOMEM);
559 
560 	if (req.total_num_uuars == 0)
561 		return ERR_PTR(-EINVAL);
562 
563 	req.total_num_uuars = ALIGN(req.total_num_uuars,
564 				    MLX5_NON_FP_BF_REGS_PER_PAGE);
565 	if (req.num_low_latency_uuars > req.total_num_uuars - 1)
566 		return ERR_PTR(-EINVAL);
567 
568 	num_uars = req.total_num_uuars / MLX5_NON_FP_BF_REGS_PER_PAGE;
569 	gross_uuars = num_uars * MLX5_BF_REGS_PER_PAGE;
570 	resp.qp_tab_size      = 1 << dev->mdev.caps.log_max_qp;
571 	resp.bf_reg_size      = dev->mdev.caps.bf_reg_size;
572 	resp.cache_line_size  = L1_CACHE_BYTES;
573 	resp.max_sq_desc_sz = dev->mdev.caps.max_sq_desc_sz;
574 	resp.max_rq_desc_sz = dev->mdev.caps.max_rq_desc_sz;
575 	resp.max_send_wqebb = dev->mdev.caps.max_wqes;
576 	resp.max_recv_wr = dev->mdev.caps.max_wqes;
577 	resp.max_srq_recv_wr = dev->mdev.caps.max_srq_wqes;
578 
579 	context = kzalloc(sizeof(*context), GFP_KERNEL);
580 	if (!context)
581 		return ERR_PTR(-ENOMEM);
582 
583 	uuari = &context->uuari;
584 	mutex_init(&uuari->lock);
585 	uars = kcalloc(num_uars, sizeof(*uars), GFP_KERNEL);
586 	if (!uars) {
587 		err = -ENOMEM;
588 		goto out_ctx;
589 	}
590 
591 	uuari->bitmap = kcalloc(BITS_TO_LONGS(gross_uuars),
592 				sizeof(*uuari->bitmap),
593 				GFP_KERNEL);
594 	if (!uuari->bitmap) {
595 		err = -ENOMEM;
596 		goto out_uar_ctx;
597 	}
598 	/*
599 	 * clear all fast path uuars
600 	 */
601 	for (i = 0; i < gross_uuars; i++) {
602 		uuarn = i & 3;
603 		if (uuarn == 2 || uuarn == 3)
604 			set_bit(i, uuari->bitmap);
605 	}
606 
607 	uuari->count = kcalloc(gross_uuars, sizeof(*uuari->count), GFP_KERNEL);
608 	if (!uuari->count) {
609 		err = -ENOMEM;
610 		goto out_bitmap;
611 	}
612 
613 	for (i = 0; i < num_uars; i++) {
614 		err = mlx5_cmd_alloc_uar(&dev->mdev, &uars[i].index);
615 		if (err)
616 			goto out_count;
617 	}
618 
619 	INIT_LIST_HEAD(&context->db_page_list);
620 	mutex_init(&context->db_page_mutex);
621 
622 	resp.tot_uuars = req.total_num_uuars;
623 	resp.num_ports = dev->mdev.caps.num_ports;
624 	err = ib_copy_to_udata(udata, &resp,
625 			       sizeof(resp) - sizeof(resp.reserved));
626 	if (err)
627 		goto out_uars;
628 
629 	uuari->num_low_latency_uuars = req.num_low_latency_uuars;
630 	uuari->uars = uars;
631 	uuari->num_uars = num_uars;
632 	return &context->ibucontext;
633 
634 out_uars:
635 	for (i--; i >= 0; i--)
636 		mlx5_cmd_free_uar(&dev->mdev, uars[i].index);
637 out_count:
638 	kfree(uuari->count);
639 
640 out_bitmap:
641 	kfree(uuari->bitmap);
642 
643 out_uar_ctx:
644 	kfree(uars);
645 
646 out_ctx:
647 	kfree(context);
648 	return ERR_PTR(err);
649 }
650 
651 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
652 {
653 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
654 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
655 	struct mlx5_uuar_info *uuari = &context->uuari;
656 	int i;
657 
658 	for (i = 0; i < uuari->num_uars; i++) {
659 		if (mlx5_cmd_free_uar(&dev->mdev, uuari->uars[i].index))
660 			mlx5_ib_warn(dev, "failed to free UAR 0x%x\n", uuari->uars[i].index);
661 	}
662 
663 	kfree(uuari->count);
664 	kfree(uuari->bitmap);
665 	kfree(uuari->uars);
666 	kfree(context);
667 
668 	return 0;
669 }
670 
671 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, int index)
672 {
673 	return (pci_resource_start(dev->mdev.pdev, 0) >> PAGE_SHIFT) + index;
674 }
675 
676 static int get_command(unsigned long offset)
677 {
678 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
679 }
680 
681 static int get_arg(unsigned long offset)
682 {
683 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
684 }
685 
686 static int get_index(unsigned long offset)
687 {
688 	return get_arg(offset);
689 }
690 
691 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
692 {
693 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
694 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
695 	struct mlx5_uuar_info *uuari = &context->uuari;
696 	unsigned long command;
697 	unsigned long idx;
698 	phys_addr_t pfn;
699 
700 	command = get_command(vma->vm_pgoff);
701 	switch (command) {
702 	case MLX5_IB_MMAP_REGULAR_PAGE:
703 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
704 			return -EINVAL;
705 
706 		idx = get_index(vma->vm_pgoff);
707 		pfn = uar_index2pfn(dev, uuari->uars[idx].index);
708 		mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn 0x%llx\n", idx,
709 			    (unsigned long long)pfn);
710 
711 		if (idx >= uuari->num_uars)
712 			return -EINVAL;
713 
714 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
715 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
716 				       PAGE_SIZE, vma->vm_page_prot))
717 			return -EAGAIN;
718 
719 		mlx5_ib_dbg(dev, "mapped WC at 0x%lx, PA 0x%llx\n",
720 			    vma->vm_start,
721 			    (unsigned long long)pfn << PAGE_SHIFT);
722 		break;
723 
724 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
725 		return -ENOSYS;
726 
727 	default:
728 		return -EINVAL;
729 	}
730 
731 	return 0;
732 }
733 
734 static int alloc_pa_mkey(struct mlx5_ib_dev *dev, u32 *key, u32 pdn)
735 {
736 	struct mlx5_create_mkey_mbox_in *in;
737 	struct mlx5_mkey_seg *seg;
738 	struct mlx5_core_mr mr;
739 	int err;
740 
741 	in = kzalloc(sizeof(*in), GFP_KERNEL);
742 	if (!in)
743 		return -ENOMEM;
744 
745 	seg = &in->seg;
746 	seg->flags = MLX5_PERM_LOCAL_READ | MLX5_ACCESS_MODE_PA;
747 	seg->flags_pd = cpu_to_be32(pdn | MLX5_MKEY_LEN64);
748 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff << 8);
749 	seg->start_addr = 0;
750 
751 	err = mlx5_core_create_mkey(&dev->mdev, &mr, in, sizeof(*in),
752 				    NULL, NULL, NULL);
753 	if (err) {
754 		mlx5_ib_warn(dev, "failed to create mkey, %d\n", err);
755 		goto err_in;
756 	}
757 
758 	kfree(in);
759 	*key = mr.key;
760 
761 	return 0;
762 
763 err_in:
764 	kfree(in);
765 
766 	return err;
767 }
768 
769 static void free_pa_mkey(struct mlx5_ib_dev *dev, u32 key)
770 {
771 	struct mlx5_core_mr mr;
772 	int err;
773 
774 	memset(&mr, 0, sizeof(mr));
775 	mr.key = key;
776 	err = mlx5_core_destroy_mkey(&dev->mdev, &mr);
777 	if (err)
778 		mlx5_ib_warn(dev, "failed to destroy mkey 0x%x\n", key);
779 }
780 
781 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
782 				      struct ib_ucontext *context,
783 				      struct ib_udata *udata)
784 {
785 	struct mlx5_ib_alloc_pd_resp resp;
786 	struct mlx5_ib_pd *pd;
787 	int err;
788 
789 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
790 	if (!pd)
791 		return ERR_PTR(-ENOMEM);
792 
793 	err = mlx5_core_alloc_pd(&to_mdev(ibdev)->mdev, &pd->pdn);
794 	if (err) {
795 		kfree(pd);
796 		return ERR_PTR(err);
797 	}
798 
799 	if (context) {
800 		resp.pdn = pd->pdn;
801 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
802 			mlx5_core_dealloc_pd(&to_mdev(ibdev)->mdev, pd->pdn);
803 			kfree(pd);
804 			return ERR_PTR(-EFAULT);
805 		}
806 	} else {
807 		err = alloc_pa_mkey(to_mdev(ibdev), &pd->pa_lkey, pd->pdn);
808 		if (err) {
809 			mlx5_core_dealloc_pd(&to_mdev(ibdev)->mdev, pd->pdn);
810 			kfree(pd);
811 			return ERR_PTR(err);
812 		}
813 	}
814 
815 	return &pd->ibpd;
816 }
817 
818 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
819 {
820 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
821 	struct mlx5_ib_pd *mpd = to_mpd(pd);
822 
823 	if (!pd->uobject)
824 		free_pa_mkey(mdev, mpd->pa_lkey);
825 
826 	mlx5_core_dealloc_pd(&mdev->mdev, mpd->pdn);
827 	kfree(mpd);
828 
829 	return 0;
830 }
831 
832 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
833 {
834 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
835 	int err;
836 
837 	err = mlx5_core_attach_mcg(&dev->mdev, gid, ibqp->qp_num);
838 	if (err)
839 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
840 			     ibqp->qp_num, gid->raw);
841 
842 	return err;
843 }
844 
845 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
846 {
847 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
848 	int err;
849 
850 	err = mlx5_core_detach_mcg(&dev->mdev, gid, ibqp->qp_num);
851 	if (err)
852 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
853 			     ibqp->qp_num, gid->raw);
854 
855 	return err;
856 }
857 
858 static int init_node_data(struct mlx5_ib_dev *dev)
859 {
860 	struct ib_smp *in_mad  = NULL;
861 	struct ib_smp *out_mad = NULL;
862 	int err = -ENOMEM;
863 
864 	in_mad  = kzalloc(sizeof(*in_mad), GFP_KERNEL);
865 	out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
866 	if (!in_mad || !out_mad)
867 		goto out;
868 
869 	init_query_mad(in_mad);
870 	in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
871 
872 	err = mlx5_MAD_IFC(dev, 1, 1, 1, NULL, NULL, in_mad, out_mad);
873 	if (err)
874 		goto out;
875 
876 	memcpy(dev->ib_dev.node_desc, out_mad->data, 64);
877 
878 	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
879 
880 	err = mlx5_MAD_IFC(dev, 1, 1, 1, NULL, NULL, in_mad, out_mad);
881 	if (err)
882 		goto out;
883 
884 	dev->mdev.rev_id = be32_to_cpup((__be32 *)(out_mad->data + 32));
885 	memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
886 
887 out:
888 	kfree(in_mad);
889 	kfree(out_mad);
890 	return err;
891 }
892 
893 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
894 			     char *buf)
895 {
896 	struct mlx5_ib_dev *dev =
897 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
898 
899 	return sprintf(buf, "%d\n", dev->mdev.priv.fw_pages);
900 }
901 
902 static ssize_t show_reg_pages(struct device *device,
903 			      struct device_attribute *attr, char *buf)
904 {
905 	struct mlx5_ib_dev *dev =
906 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
907 
908 	return sprintf(buf, "%d\n", dev->mdev.priv.reg_pages);
909 }
910 
911 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
912 			char *buf)
913 {
914 	struct mlx5_ib_dev *dev =
915 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
916 	return sprintf(buf, "MT%d\n", dev->mdev.pdev->device);
917 }
918 
919 static ssize_t show_fw_ver(struct device *device, struct device_attribute *attr,
920 			   char *buf)
921 {
922 	struct mlx5_ib_dev *dev =
923 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
924 	return sprintf(buf, "%d.%d.%d\n", fw_rev_maj(&dev->mdev),
925 		       fw_rev_min(&dev->mdev), fw_rev_sub(&dev->mdev));
926 }
927 
928 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
929 			char *buf)
930 {
931 	struct mlx5_ib_dev *dev =
932 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
933 	return sprintf(buf, "%x\n", dev->mdev.rev_id);
934 }
935 
936 static ssize_t show_board(struct device *device, struct device_attribute *attr,
937 			  char *buf)
938 {
939 	struct mlx5_ib_dev *dev =
940 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
941 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
942 		       dev->mdev.board_id);
943 }
944 
945 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
946 static DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
947 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
948 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
949 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
950 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
951 
952 static struct device_attribute *mlx5_class_attributes[] = {
953 	&dev_attr_hw_rev,
954 	&dev_attr_fw_ver,
955 	&dev_attr_hca_type,
956 	&dev_attr_board_id,
957 	&dev_attr_fw_pages,
958 	&dev_attr_reg_pages,
959 };
960 
961 static void mlx5_ib_event(struct mlx5_core_dev *dev, enum mlx5_dev_event event,
962 			  void *data)
963 {
964 	struct mlx5_ib_dev *ibdev = container_of(dev, struct mlx5_ib_dev, mdev);
965 	struct ib_event ibev;
966 	u8 port = 0;
967 
968 	switch (event) {
969 	case MLX5_DEV_EVENT_SYS_ERROR:
970 		ibdev->ib_active = false;
971 		ibev.event = IB_EVENT_DEVICE_FATAL;
972 		break;
973 
974 	case MLX5_DEV_EVENT_PORT_UP:
975 		ibev.event = IB_EVENT_PORT_ACTIVE;
976 		port = *(u8 *)data;
977 		break;
978 
979 	case MLX5_DEV_EVENT_PORT_DOWN:
980 		ibev.event = IB_EVENT_PORT_ERR;
981 		port = *(u8 *)data;
982 		break;
983 
984 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
985 		/* not used by ULPs */
986 		return;
987 
988 	case MLX5_DEV_EVENT_LID_CHANGE:
989 		ibev.event = IB_EVENT_LID_CHANGE;
990 		port = *(u8 *)data;
991 		break;
992 
993 	case MLX5_DEV_EVENT_PKEY_CHANGE:
994 		ibev.event = IB_EVENT_PKEY_CHANGE;
995 		port = *(u8 *)data;
996 		break;
997 
998 	case MLX5_DEV_EVENT_GUID_CHANGE:
999 		ibev.event = IB_EVENT_GID_CHANGE;
1000 		port = *(u8 *)data;
1001 		break;
1002 
1003 	case MLX5_DEV_EVENT_CLIENT_REREG:
1004 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
1005 		port = *(u8 *)data;
1006 		break;
1007 	}
1008 
1009 	ibev.device	      = &ibdev->ib_dev;
1010 	ibev.element.port_num = port;
1011 
1012 	if (port < 1 || port > ibdev->num_ports) {
1013 		mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
1014 		return;
1015 	}
1016 
1017 	if (ibdev->ib_active)
1018 		ib_dispatch_event(&ibev);
1019 }
1020 
1021 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
1022 {
1023 	int port;
1024 
1025 	for (port = 1; port <= dev->mdev.caps.num_ports; port++)
1026 		mlx5_query_ext_port_caps(dev, port);
1027 }
1028 
1029 static int get_port_caps(struct mlx5_ib_dev *dev)
1030 {
1031 	struct ib_device_attr *dprops = NULL;
1032 	struct ib_port_attr *pprops = NULL;
1033 	int err = 0;
1034 	int port;
1035 
1036 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
1037 	if (!pprops)
1038 		goto out;
1039 
1040 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
1041 	if (!dprops)
1042 		goto out;
1043 
1044 	err = mlx5_ib_query_device(&dev->ib_dev, dprops);
1045 	if (err) {
1046 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
1047 		goto out;
1048 	}
1049 
1050 	for (port = 1; port <= dev->mdev.caps.num_ports; port++) {
1051 		err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
1052 		if (err) {
1053 			mlx5_ib_warn(dev, "query_port %d failed %d\n", port, err);
1054 			break;
1055 		}
1056 		dev->mdev.caps.port[port - 1].pkey_table_len = dprops->max_pkeys;
1057 		dev->mdev.caps.port[port - 1].gid_table_len = pprops->gid_tbl_len;
1058 		mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
1059 			    dprops->max_pkeys, pprops->gid_tbl_len);
1060 	}
1061 
1062 out:
1063 	kfree(pprops);
1064 	kfree(dprops);
1065 
1066 	return err;
1067 }
1068 
1069 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
1070 {
1071 	int err;
1072 
1073 	err = mlx5_mr_cache_cleanup(dev);
1074 	if (err)
1075 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
1076 
1077 	mlx5_ib_destroy_qp(dev->umrc.qp);
1078 	ib_destroy_cq(dev->umrc.cq);
1079 	ib_dereg_mr(dev->umrc.mr);
1080 	ib_dealloc_pd(dev->umrc.pd);
1081 }
1082 
1083 enum {
1084 	MAX_UMR_WR = 128,
1085 };
1086 
1087 static int create_umr_res(struct mlx5_ib_dev *dev)
1088 {
1089 	struct ib_qp_init_attr *init_attr = NULL;
1090 	struct ib_qp_attr *attr = NULL;
1091 	struct ib_pd *pd;
1092 	struct ib_cq *cq;
1093 	struct ib_qp *qp;
1094 	struct ib_mr *mr;
1095 	int ret;
1096 
1097 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
1098 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
1099 	if (!attr || !init_attr) {
1100 		ret = -ENOMEM;
1101 		goto error_0;
1102 	}
1103 
1104 	pd = ib_alloc_pd(&dev->ib_dev);
1105 	if (IS_ERR(pd)) {
1106 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
1107 		ret = PTR_ERR(pd);
1108 		goto error_0;
1109 	}
1110 
1111 	mr = ib_get_dma_mr(pd,  IB_ACCESS_LOCAL_WRITE);
1112 	if (IS_ERR(mr)) {
1113 		mlx5_ib_dbg(dev, "Couldn't create DMA MR for sync UMR QP\n");
1114 		ret = PTR_ERR(mr);
1115 		goto error_1;
1116 	}
1117 
1118 	cq = ib_create_cq(&dev->ib_dev, mlx5_umr_cq_handler, NULL, NULL, 128,
1119 			  0);
1120 	if (IS_ERR(cq)) {
1121 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
1122 		ret = PTR_ERR(cq);
1123 		goto error_2;
1124 	}
1125 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1126 
1127 	init_attr->send_cq = cq;
1128 	init_attr->recv_cq = cq;
1129 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
1130 	init_attr->cap.max_send_wr = MAX_UMR_WR;
1131 	init_attr->cap.max_send_sge = 1;
1132 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
1133 	init_attr->port_num = 1;
1134 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
1135 	if (IS_ERR(qp)) {
1136 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
1137 		ret = PTR_ERR(qp);
1138 		goto error_3;
1139 	}
1140 	qp->device     = &dev->ib_dev;
1141 	qp->real_qp    = qp;
1142 	qp->uobject    = NULL;
1143 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
1144 
1145 	attr->qp_state = IB_QPS_INIT;
1146 	attr->port_num = 1;
1147 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
1148 				IB_QP_PORT, NULL);
1149 	if (ret) {
1150 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
1151 		goto error_4;
1152 	}
1153 
1154 	memset(attr, 0, sizeof(*attr));
1155 	attr->qp_state = IB_QPS_RTR;
1156 	attr->path_mtu = IB_MTU_256;
1157 
1158 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
1159 	if (ret) {
1160 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
1161 		goto error_4;
1162 	}
1163 
1164 	memset(attr, 0, sizeof(*attr));
1165 	attr->qp_state = IB_QPS_RTS;
1166 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
1167 	if (ret) {
1168 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
1169 		goto error_4;
1170 	}
1171 
1172 	dev->umrc.qp = qp;
1173 	dev->umrc.cq = cq;
1174 	dev->umrc.mr = mr;
1175 	dev->umrc.pd = pd;
1176 
1177 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
1178 	ret = mlx5_mr_cache_init(dev);
1179 	if (ret) {
1180 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
1181 		goto error_4;
1182 	}
1183 
1184 	kfree(attr);
1185 	kfree(init_attr);
1186 
1187 	return 0;
1188 
1189 error_4:
1190 	mlx5_ib_destroy_qp(qp);
1191 
1192 error_3:
1193 	ib_destroy_cq(cq);
1194 
1195 error_2:
1196 	ib_dereg_mr(mr);
1197 
1198 error_1:
1199 	ib_dealloc_pd(pd);
1200 
1201 error_0:
1202 	kfree(attr);
1203 	kfree(init_attr);
1204 	return ret;
1205 }
1206 
1207 static int create_dev_resources(struct mlx5_ib_resources *devr)
1208 {
1209 	struct ib_srq_init_attr attr;
1210 	struct mlx5_ib_dev *dev;
1211 	int ret = 0;
1212 
1213 	dev = container_of(devr, struct mlx5_ib_dev, devr);
1214 
1215 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
1216 	if (IS_ERR(devr->p0)) {
1217 		ret = PTR_ERR(devr->p0);
1218 		goto error0;
1219 	}
1220 	devr->p0->device  = &dev->ib_dev;
1221 	devr->p0->uobject = NULL;
1222 	atomic_set(&devr->p0->usecnt, 0);
1223 
1224 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, 1, 0, NULL, NULL);
1225 	if (IS_ERR(devr->c0)) {
1226 		ret = PTR_ERR(devr->c0);
1227 		goto error1;
1228 	}
1229 	devr->c0->device        = &dev->ib_dev;
1230 	devr->c0->uobject       = NULL;
1231 	devr->c0->comp_handler  = NULL;
1232 	devr->c0->event_handler = NULL;
1233 	devr->c0->cq_context    = NULL;
1234 	atomic_set(&devr->c0->usecnt, 0);
1235 
1236 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
1237 	if (IS_ERR(devr->x0)) {
1238 		ret = PTR_ERR(devr->x0);
1239 		goto error2;
1240 	}
1241 	devr->x0->device = &dev->ib_dev;
1242 	devr->x0->inode = NULL;
1243 	atomic_set(&devr->x0->usecnt, 0);
1244 	mutex_init(&devr->x0->tgt_qp_mutex);
1245 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
1246 
1247 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
1248 	if (IS_ERR(devr->x1)) {
1249 		ret = PTR_ERR(devr->x1);
1250 		goto error3;
1251 	}
1252 	devr->x1->device = &dev->ib_dev;
1253 	devr->x1->inode = NULL;
1254 	atomic_set(&devr->x1->usecnt, 0);
1255 	mutex_init(&devr->x1->tgt_qp_mutex);
1256 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
1257 
1258 	memset(&attr, 0, sizeof(attr));
1259 	attr.attr.max_sge = 1;
1260 	attr.attr.max_wr = 1;
1261 	attr.srq_type = IB_SRQT_XRC;
1262 	attr.ext.xrc.cq = devr->c0;
1263 	attr.ext.xrc.xrcd = devr->x0;
1264 
1265 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
1266 	if (IS_ERR(devr->s0)) {
1267 		ret = PTR_ERR(devr->s0);
1268 		goto error4;
1269 	}
1270 	devr->s0->device	= &dev->ib_dev;
1271 	devr->s0->pd		= devr->p0;
1272 	devr->s0->uobject       = NULL;
1273 	devr->s0->event_handler = NULL;
1274 	devr->s0->srq_context   = NULL;
1275 	devr->s0->srq_type      = IB_SRQT_XRC;
1276 	devr->s0->ext.xrc.xrcd	= devr->x0;
1277 	devr->s0->ext.xrc.cq	= devr->c0;
1278 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
1279 	atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
1280 	atomic_inc(&devr->p0->usecnt);
1281 	atomic_set(&devr->s0->usecnt, 0);
1282 
1283 	return 0;
1284 
1285 error4:
1286 	mlx5_ib_dealloc_xrcd(devr->x1);
1287 error3:
1288 	mlx5_ib_dealloc_xrcd(devr->x0);
1289 error2:
1290 	mlx5_ib_destroy_cq(devr->c0);
1291 error1:
1292 	mlx5_ib_dealloc_pd(devr->p0);
1293 error0:
1294 	return ret;
1295 }
1296 
1297 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
1298 {
1299 	mlx5_ib_destroy_srq(devr->s0);
1300 	mlx5_ib_dealloc_xrcd(devr->x0);
1301 	mlx5_ib_dealloc_xrcd(devr->x1);
1302 	mlx5_ib_destroy_cq(devr->c0);
1303 	mlx5_ib_dealloc_pd(devr->p0);
1304 }
1305 
1306 static int init_one(struct pci_dev *pdev,
1307 		    const struct pci_device_id *id)
1308 {
1309 	struct mlx5_core_dev *mdev;
1310 	struct mlx5_ib_dev *dev;
1311 	int err;
1312 	int i;
1313 
1314 	printk_once(KERN_INFO "%s", mlx5_version);
1315 
1316 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
1317 	if (!dev)
1318 		return -ENOMEM;
1319 
1320 	mdev = &dev->mdev;
1321 	mdev->event = mlx5_ib_event;
1322 	if (prof_sel >= ARRAY_SIZE(profile)) {
1323 		pr_warn("selected pofile out of range, selceting default\n");
1324 		prof_sel = 0;
1325 	}
1326 	mdev->profile = &profile[prof_sel];
1327 	err = mlx5_dev_init(mdev, pdev);
1328 	if (err)
1329 		goto err_free;
1330 
1331 	err = get_port_caps(dev);
1332 	if (err)
1333 		goto err_cleanup;
1334 
1335 	get_ext_port_caps(dev);
1336 
1337 	err = alloc_comp_eqs(dev);
1338 	if (err)
1339 		goto err_cleanup;
1340 
1341 	MLX5_INIT_DOORBELL_LOCK(&dev->uar_lock);
1342 
1343 	strlcpy(dev->ib_dev.name, "mlx5_%d", IB_DEVICE_NAME_MAX);
1344 	dev->ib_dev.owner		= THIS_MODULE;
1345 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
1346 	dev->ib_dev.local_dma_lkey	= mdev->caps.reserved_lkey;
1347 	dev->num_ports		= mdev->caps.num_ports;
1348 	dev->ib_dev.phys_port_cnt     = dev->num_ports;
1349 	dev->ib_dev.num_comp_vectors	= dev->num_comp_vectors;
1350 	dev->ib_dev.dma_device	= &mdev->pdev->dev;
1351 
1352 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
1353 	dev->ib_dev.uverbs_cmd_mask	=
1354 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
1355 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
1356 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
1357 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
1358 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
1359 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
1360 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
1361 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
1362 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
1363 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
1364 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
1365 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
1366 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
1367 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
1368 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
1369 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
1370 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
1371 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
1372 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
1373 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
1374 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
1375 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
1376 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
1377 
1378 	dev->ib_dev.query_device	= mlx5_ib_query_device;
1379 	dev->ib_dev.query_port		= mlx5_ib_query_port;
1380 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
1381 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
1382 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
1383 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
1384 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
1385 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
1386 	dev->ib_dev.mmap		= mlx5_ib_mmap;
1387 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
1388 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
1389 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
1390 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
1391 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
1392 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
1393 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
1394 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
1395 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
1396 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
1397 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
1398 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
1399 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
1400 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
1401 	dev->ib_dev.post_send		= mlx5_ib_post_send;
1402 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
1403 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
1404 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
1405 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
1406 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
1407 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
1408 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
1409 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
1410 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
1411 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
1412 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
1413 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
1414 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
1415 	dev->ib_dev.alloc_fast_reg_mr	= mlx5_ib_alloc_fast_reg_mr;
1416 	dev->ib_dev.alloc_fast_reg_page_list = mlx5_ib_alloc_fast_reg_page_list;
1417 	dev->ib_dev.free_fast_reg_page_list  = mlx5_ib_free_fast_reg_page_list;
1418 
1419 	if (mdev->caps.flags & MLX5_DEV_CAP_FLAG_XRC) {
1420 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
1421 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
1422 		dev->ib_dev.uverbs_cmd_mask |=
1423 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
1424 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
1425 	}
1426 
1427 	err = init_node_data(dev);
1428 	if (err)
1429 		goto err_eqs;
1430 
1431 	mutex_init(&dev->cap_mask_mutex);
1432 	spin_lock_init(&dev->mr_lock);
1433 
1434 	err = create_dev_resources(&dev->devr);
1435 	if (err)
1436 		goto err_eqs;
1437 
1438 	err = ib_register_device(&dev->ib_dev, NULL);
1439 	if (err)
1440 		goto err_rsrc;
1441 
1442 	err = create_umr_res(dev);
1443 	if (err)
1444 		goto err_dev;
1445 
1446 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
1447 		err = device_create_file(&dev->ib_dev.dev,
1448 					 mlx5_class_attributes[i]);
1449 		if (err)
1450 			goto err_umrc;
1451 	}
1452 
1453 	dev->ib_active = true;
1454 
1455 	return 0;
1456 
1457 err_umrc:
1458 	destroy_umrc_res(dev);
1459 
1460 err_dev:
1461 	ib_unregister_device(&dev->ib_dev);
1462 
1463 err_rsrc:
1464 	destroy_dev_resources(&dev->devr);
1465 
1466 err_eqs:
1467 	free_comp_eqs(dev);
1468 
1469 err_cleanup:
1470 	mlx5_dev_cleanup(mdev);
1471 
1472 err_free:
1473 	ib_dealloc_device((struct ib_device *)dev);
1474 
1475 	return err;
1476 }
1477 
1478 static void remove_one(struct pci_dev *pdev)
1479 {
1480 	struct mlx5_ib_dev *dev = mlx5_pci2ibdev(pdev);
1481 
1482 	destroy_umrc_res(dev);
1483 	ib_unregister_device(&dev->ib_dev);
1484 	destroy_dev_resources(&dev->devr);
1485 	free_comp_eqs(dev);
1486 	mlx5_dev_cleanup(&dev->mdev);
1487 	ib_dealloc_device(&dev->ib_dev);
1488 }
1489 
1490 static DEFINE_PCI_DEVICE_TABLE(mlx5_ib_pci_table) = {
1491 	{ PCI_VDEVICE(MELLANOX, 4113) }, /* MT4113 Connect-IB */
1492 	{ 0, }
1493 };
1494 
1495 MODULE_DEVICE_TABLE(pci, mlx5_ib_pci_table);
1496 
1497 static struct pci_driver mlx5_ib_driver = {
1498 	.name		= DRIVER_NAME,
1499 	.id_table	= mlx5_ib_pci_table,
1500 	.probe		= init_one,
1501 	.remove		= remove_one
1502 };
1503 
1504 static int __init mlx5_ib_init(void)
1505 {
1506 	return pci_register_driver(&mlx5_ib_driver);
1507 }
1508 
1509 static void __exit mlx5_ib_cleanup(void)
1510 {
1511 	pci_unregister_driver(&mlx5_ib_driver);
1512 }
1513 
1514 module_init(mlx5_ib_init);
1515 module_exit(mlx5_ib_cleanup);
1516