xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision a2fb4d78)
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 "2.2-1"
50 #define DRIVER_RELDATE	"Feb 2014"
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 	flags = dev->mdev.caps.flags;
266 	if (flags & MLX5_DEV_CAP_FLAG_BAD_PKEY_CNTR)
267 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
268 	if (flags & MLX5_DEV_CAP_FLAG_BAD_QKEY_CNTR)
269 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
270 	if (flags & MLX5_DEV_CAP_FLAG_APM)
271 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
272 	props->device_cap_flags |= IB_DEVICE_LOCAL_DMA_LKEY;
273 	if (flags & MLX5_DEV_CAP_FLAG_XRC)
274 		props->device_cap_flags |= IB_DEVICE_XRC;
275 	props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
276 
277 	props->vendor_id	   = be32_to_cpup((__be32 *)(out_mad->data + 36)) &
278 		0xffffff;
279 	props->vendor_part_id	   = be16_to_cpup((__be16 *)(out_mad->data + 30));
280 	props->hw_ver		   = be32_to_cpup((__be32 *)(out_mad->data + 32));
281 	memcpy(&props->sys_image_guid, out_mad->data +	4, 8);
282 
283 	props->max_mr_size	   = ~0ull;
284 	props->page_size_cap	   = dev->mdev.caps.min_page_sz;
285 	props->max_qp		   = 1 << dev->mdev.caps.log_max_qp;
286 	props->max_qp_wr	   = dev->mdev.caps.max_wqes;
287 	max_rq_sg = dev->mdev.caps.max_rq_desc_sz / sizeof(struct mlx5_wqe_data_seg);
288 	max_sq_sg = (dev->mdev.caps.max_sq_desc_sz - sizeof(struct mlx5_wqe_ctrl_seg)) /
289 		sizeof(struct mlx5_wqe_data_seg);
290 	props->max_sge = min(max_rq_sg, max_sq_sg);
291 	props->max_cq		   = 1 << dev->mdev.caps.log_max_cq;
292 	props->max_cqe		   = dev->mdev.caps.max_cqes - 1;
293 	props->max_mr		   = 1 << dev->mdev.caps.log_max_mkey;
294 	props->max_pd		   = 1 << dev->mdev.caps.log_max_pd;
295 	props->max_qp_rd_atom	   = dev->mdev.caps.max_ra_req_qp;
296 	props->max_qp_init_rd_atom = dev->mdev.caps.max_ra_res_qp;
297 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
298 	props->max_srq		   = 1 << dev->mdev.caps.log_max_srq;
299 	props->max_srq_wr	   = dev->mdev.caps.max_srq_wqes - 1;
300 	props->max_srq_sge	   = max_rq_sg - 1;
301 	props->max_fast_reg_page_list_len = (unsigned int)-1;
302 	props->local_ca_ack_delay  = dev->mdev.caps.local_ca_ack_delay;
303 	props->atomic_cap	   = IB_ATOMIC_NONE;
304 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
305 	props->max_pkeys	   = be16_to_cpup((__be16 *)(out_mad->data + 28));
306 	props->max_mcast_grp	   = 1 << dev->mdev.caps.log_max_mcg;
307 	props->max_mcast_qp_attach = dev->mdev.caps.max_qp_mcg;
308 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
309 					   props->max_mcast_grp;
310 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
311 
312 out:
313 	kfree(in_mad);
314 	kfree(out_mad);
315 
316 	return err;
317 }
318 
319 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
320 		       struct ib_port_attr *props)
321 {
322 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
323 	struct ib_smp *in_mad  = NULL;
324 	struct ib_smp *out_mad = NULL;
325 	int ext_active_speed;
326 	int err = -ENOMEM;
327 
328 	if (port < 1 || port > dev->mdev.caps.num_ports) {
329 		mlx5_ib_warn(dev, "invalid port number %d\n", port);
330 		return -EINVAL;
331 	}
332 
333 	in_mad  = kzalloc(sizeof(*in_mad), GFP_KERNEL);
334 	out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
335 	if (!in_mad || !out_mad)
336 		goto out;
337 
338 	memset(props, 0, sizeof(*props));
339 
340 	init_query_mad(in_mad);
341 	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
342 	in_mad->attr_mod = cpu_to_be32(port);
343 
344 	err = mlx5_MAD_IFC(dev, 1, 1, port, NULL, NULL, in_mad, out_mad);
345 	if (err) {
346 		mlx5_ib_warn(dev, "err %d\n", err);
347 		goto out;
348 	}
349 
350 
351 	props->lid		= be16_to_cpup((__be16 *)(out_mad->data + 16));
352 	props->lmc		= out_mad->data[34] & 0x7;
353 	props->sm_lid		= be16_to_cpup((__be16 *)(out_mad->data + 18));
354 	props->sm_sl		= out_mad->data[36] & 0xf;
355 	props->state		= out_mad->data[32] & 0xf;
356 	props->phys_state	= out_mad->data[33] >> 4;
357 	props->port_cap_flags	= be32_to_cpup((__be32 *)(out_mad->data + 20));
358 	props->gid_tbl_len	= out_mad->data[50];
359 	props->max_msg_sz	= 1 << to_mdev(ibdev)->mdev.caps.log_max_msg;
360 	props->pkey_tbl_len	= to_mdev(ibdev)->mdev.caps.port[port - 1].pkey_table_len;
361 	props->bad_pkey_cntr	= be16_to_cpup((__be16 *)(out_mad->data + 46));
362 	props->qkey_viol_cntr	= be16_to_cpup((__be16 *)(out_mad->data + 48));
363 	props->active_width	= out_mad->data[31] & 0xf;
364 	props->active_speed	= out_mad->data[35] >> 4;
365 	props->max_mtu		= out_mad->data[41] & 0xf;
366 	props->active_mtu	= out_mad->data[36] >> 4;
367 	props->subnet_timeout	= out_mad->data[51] & 0x1f;
368 	props->max_vl_num	= out_mad->data[37] >> 4;
369 	props->init_type_reply	= out_mad->data[41] >> 4;
370 
371 	/* Check if extended speeds (EDR/FDR/...) are supported */
372 	if (props->port_cap_flags & IB_PORT_EXTENDED_SPEEDS_SUP) {
373 		ext_active_speed = out_mad->data[62] >> 4;
374 
375 		switch (ext_active_speed) {
376 		case 1:
377 			props->active_speed = 16; /* FDR */
378 			break;
379 		case 2:
380 			props->active_speed = 32; /* EDR */
381 			break;
382 		}
383 	}
384 
385 	/* If reported active speed is QDR, check if is FDR-10 */
386 	if (props->active_speed == 4) {
387 		if (dev->mdev.caps.ext_port_cap[port - 1] &
388 		    MLX_EXT_PORT_CAP_FLAG_EXTENDED_PORT_INFO) {
389 			init_query_mad(in_mad);
390 			in_mad->attr_id = MLX5_ATTR_EXTENDED_PORT_INFO;
391 			in_mad->attr_mod = cpu_to_be32(port);
392 
393 			err = mlx5_MAD_IFC(dev, 1, 1, port,
394 					   NULL, NULL, in_mad, out_mad);
395 			if (err)
396 				goto out;
397 
398 			/* Checking LinkSpeedActive for FDR-10 */
399 			if (out_mad->data[15] & 0x1)
400 				props->active_speed = 8;
401 		}
402 	}
403 
404 out:
405 	kfree(in_mad);
406 	kfree(out_mad);
407 
408 	return err;
409 }
410 
411 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
412 			     union ib_gid *gid)
413 {
414 	struct ib_smp *in_mad  = NULL;
415 	struct ib_smp *out_mad = NULL;
416 	int err = -ENOMEM;
417 
418 	in_mad  = kzalloc(sizeof(*in_mad), GFP_KERNEL);
419 	out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
420 	if (!in_mad || !out_mad)
421 		goto out;
422 
423 	init_query_mad(in_mad);
424 	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
425 	in_mad->attr_mod = cpu_to_be32(port);
426 
427 	err = mlx5_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
428 	if (err)
429 		goto out;
430 
431 	memcpy(gid->raw, out_mad->data + 8, 8);
432 
433 	init_query_mad(in_mad);
434 	in_mad->attr_id  = IB_SMP_ATTR_GUID_INFO;
435 	in_mad->attr_mod = cpu_to_be32(index / 8);
436 
437 	err = mlx5_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
438 	if (err)
439 		goto out;
440 
441 	memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
442 
443 out:
444 	kfree(in_mad);
445 	kfree(out_mad);
446 	return err;
447 }
448 
449 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
450 			      u16 *pkey)
451 {
452 	struct ib_smp *in_mad  = NULL;
453 	struct ib_smp *out_mad = NULL;
454 	int err = -ENOMEM;
455 
456 	in_mad  = kzalloc(sizeof(*in_mad), GFP_KERNEL);
457 	out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
458 	if (!in_mad || !out_mad)
459 		goto out;
460 
461 	init_query_mad(in_mad);
462 	in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
463 	in_mad->attr_mod = cpu_to_be32(index / 32);
464 
465 	err = mlx5_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
466 	if (err)
467 		goto out;
468 
469 	*pkey = be16_to_cpu(((__be16 *)out_mad->data)[index % 32]);
470 
471 out:
472 	kfree(in_mad);
473 	kfree(out_mad);
474 	return err;
475 }
476 
477 struct mlx5_reg_node_desc {
478 	u8	desc[64];
479 };
480 
481 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
482 				 struct ib_device_modify *props)
483 {
484 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
485 	struct mlx5_reg_node_desc in;
486 	struct mlx5_reg_node_desc out;
487 	int err;
488 
489 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
490 		return -EOPNOTSUPP;
491 
492 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
493 		return 0;
494 
495 	/*
496 	 * If possible, pass node desc to FW, so it can generate
497 	 * a 144 trap.  If cmd fails, just ignore.
498 	 */
499 	memcpy(&in, props->node_desc, 64);
500 	err = mlx5_core_access_reg(&dev->mdev, &in, sizeof(in), &out,
501 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
502 	if (err)
503 		return err;
504 
505 	memcpy(ibdev->node_desc, props->node_desc, 64);
506 
507 	return err;
508 }
509 
510 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
511 			       struct ib_port_modify *props)
512 {
513 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
514 	struct ib_port_attr attr;
515 	u32 tmp;
516 	int err;
517 
518 	mutex_lock(&dev->cap_mask_mutex);
519 
520 	err = mlx5_ib_query_port(ibdev, port, &attr);
521 	if (err)
522 		goto out;
523 
524 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
525 		~props->clr_port_cap_mask;
526 
527 	err = mlx5_set_port_caps(&dev->mdev, port, tmp);
528 
529 out:
530 	mutex_unlock(&dev->cap_mask_mutex);
531 	return err;
532 }
533 
534 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
535 						  struct ib_udata *udata)
536 {
537 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
538 	struct mlx5_ib_alloc_ucontext_req_v2 req;
539 	struct mlx5_ib_alloc_ucontext_resp resp;
540 	struct mlx5_ib_ucontext *context;
541 	struct mlx5_uuar_info *uuari;
542 	struct mlx5_uar *uars;
543 	int gross_uuars;
544 	int num_uars;
545 	int ver;
546 	int uuarn;
547 	int err;
548 	int i;
549 	int reqlen;
550 
551 	if (!dev->ib_active)
552 		return ERR_PTR(-EAGAIN);
553 
554 	memset(&req, 0, sizeof(req));
555 	reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
556 	if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
557 		ver = 0;
558 	else if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req_v2))
559 		ver = 2;
560 	else
561 		return ERR_PTR(-EINVAL);
562 
563 	err = ib_copy_from_udata(&req, udata, reqlen);
564 	if (err)
565 		return ERR_PTR(err);
566 
567 	if (req.flags || req.reserved)
568 		return ERR_PTR(-EINVAL);
569 
570 	if (req.total_num_uuars > MLX5_MAX_UUARS)
571 		return ERR_PTR(-ENOMEM);
572 
573 	if (req.total_num_uuars == 0)
574 		return ERR_PTR(-EINVAL);
575 
576 	req.total_num_uuars = ALIGN(req.total_num_uuars,
577 				    MLX5_NON_FP_BF_REGS_PER_PAGE);
578 	if (req.num_low_latency_uuars > req.total_num_uuars - 1)
579 		return ERR_PTR(-EINVAL);
580 
581 	num_uars = req.total_num_uuars / MLX5_NON_FP_BF_REGS_PER_PAGE;
582 	gross_uuars = num_uars * MLX5_BF_REGS_PER_PAGE;
583 	resp.qp_tab_size      = 1 << dev->mdev.caps.log_max_qp;
584 	resp.bf_reg_size      = dev->mdev.caps.bf_reg_size;
585 	resp.cache_line_size  = L1_CACHE_BYTES;
586 	resp.max_sq_desc_sz = dev->mdev.caps.max_sq_desc_sz;
587 	resp.max_rq_desc_sz = dev->mdev.caps.max_rq_desc_sz;
588 	resp.max_send_wqebb = dev->mdev.caps.max_wqes;
589 	resp.max_recv_wr = dev->mdev.caps.max_wqes;
590 	resp.max_srq_recv_wr = dev->mdev.caps.max_srq_wqes;
591 
592 	context = kzalloc(sizeof(*context), GFP_KERNEL);
593 	if (!context)
594 		return ERR_PTR(-ENOMEM);
595 
596 	uuari = &context->uuari;
597 	mutex_init(&uuari->lock);
598 	uars = kcalloc(num_uars, sizeof(*uars), GFP_KERNEL);
599 	if (!uars) {
600 		err = -ENOMEM;
601 		goto out_ctx;
602 	}
603 
604 	uuari->bitmap = kcalloc(BITS_TO_LONGS(gross_uuars),
605 				sizeof(*uuari->bitmap),
606 				GFP_KERNEL);
607 	if (!uuari->bitmap) {
608 		err = -ENOMEM;
609 		goto out_uar_ctx;
610 	}
611 	/*
612 	 * clear all fast path uuars
613 	 */
614 	for (i = 0; i < gross_uuars; i++) {
615 		uuarn = i & 3;
616 		if (uuarn == 2 || uuarn == 3)
617 			set_bit(i, uuari->bitmap);
618 	}
619 
620 	uuari->count = kcalloc(gross_uuars, sizeof(*uuari->count), GFP_KERNEL);
621 	if (!uuari->count) {
622 		err = -ENOMEM;
623 		goto out_bitmap;
624 	}
625 
626 	for (i = 0; i < num_uars; i++) {
627 		err = mlx5_cmd_alloc_uar(&dev->mdev, &uars[i].index);
628 		if (err)
629 			goto out_count;
630 	}
631 
632 	INIT_LIST_HEAD(&context->db_page_list);
633 	mutex_init(&context->db_page_mutex);
634 
635 	resp.tot_uuars = req.total_num_uuars;
636 	resp.num_ports = dev->mdev.caps.num_ports;
637 	err = ib_copy_to_udata(udata, &resp,
638 			       sizeof(resp) - sizeof(resp.reserved));
639 	if (err)
640 		goto out_uars;
641 
642 	uuari->ver = ver;
643 	uuari->num_low_latency_uuars = req.num_low_latency_uuars;
644 	uuari->uars = uars;
645 	uuari->num_uars = num_uars;
646 	return &context->ibucontext;
647 
648 out_uars:
649 	for (i--; i >= 0; i--)
650 		mlx5_cmd_free_uar(&dev->mdev, uars[i].index);
651 out_count:
652 	kfree(uuari->count);
653 
654 out_bitmap:
655 	kfree(uuari->bitmap);
656 
657 out_uar_ctx:
658 	kfree(uars);
659 
660 out_ctx:
661 	kfree(context);
662 	return ERR_PTR(err);
663 }
664 
665 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
666 {
667 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
668 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
669 	struct mlx5_uuar_info *uuari = &context->uuari;
670 	int i;
671 
672 	for (i = 0; i < uuari->num_uars; i++) {
673 		if (mlx5_cmd_free_uar(&dev->mdev, uuari->uars[i].index))
674 			mlx5_ib_warn(dev, "failed to free UAR 0x%x\n", uuari->uars[i].index);
675 	}
676 
677 	kfree(uuari->count);
678 	kfree(uuari->bitmap);
679 	kfree(uuari->uars);
680 	kfree(context);
681 
682 	return 0;
683 }
684 
685 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, int index)
686 {
687 	return (pci_resource_start(dev->mdev.pdev, 0) >> PAGE_SHIFT) + index;
688 }
689 
690 static int get_command(unsigned long offset)
691 {
692 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
693 }
694 
695 static int get_arg(unsigned long offset)
696 {
697 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
698 }
699 
700 static int get_index(unsigned long offset)
701 {
702 	return get_arg(offset);
703 }
704 
705 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
706 {
707 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
708 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
709 	struct mlx5_uuar_info *uuari = &context->uuari;
710 	unsigned long command;
711 	unsigned long idx;
712 	phys_addr_t pfn;
713 
714 	command = get_command(vma->vm_pgoff);
715 	switch (command) {
716 	case MLX5_IB_MMAP_REGULAR_PAGE:
717 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
718 			return -EINVAL;
719 
720 		idx = get_index(vma->vm_pgoff);
721 		pfn = uar_index2pfn(dev, uuari->uars[idx].index);
722 		mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn 0x%llx\n", idx,
723 			    (unsigned long long)pfn);
724 
725 		if (idx >= uuari->num_uars)
726 			return -EINVAL;
727 
728 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
729 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
730 				       PAGE_SIZE, vma->vm_page_prot))
731 			return -EAGAIN;
732 
733 		mlx5_ib_dbg(dev, "mapped WC at 0x%lx, PA 0x%llx\n",
734 			    vma->vm_start,
735 			    (unsigned long long)pfn << PAGE_SHIFT);
736 		break;
737 
738 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
739 		return -ENOSYS;
740 
741 	default:
742 		return -EINVAL;
743 	}
744 
745 	return 0;
746 }
747 
748 static int alloc_pa_mkey(struct mlx5_ib_dev *dev, u32 *key, u32 pdn)
749 {
750 	struct mlx5_create_mkey_mbox_in *in;
751 	struct mlx5_mkey_seg *seg;
752 	struct mlx5_core_mr mr;
753 	int err;
754 
755 	in = kzalloc(sizeof(*in), GFP_KERNEL);
756 	if (!in)
757 		return -ENOMEM;
758 
759 	seg = &in->seg;
760 	seg->flags = MLX5_PERM_LOCAL_READ | MLX5_ACCESS_MODE_PA;
761 	seg->flags_pd = cpu_to_be32(pdn | MLX5_MKEY_LEN64);
762 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff << 8);
763 	seg->start_addr = 0;
764 
765 	err = mlx5_core_create_mkey(&dev->mdev, &mr, in, sizeof(*in),
766 				    NULL, NULL, NULL);
767 	if (err) {
768 		mlx5_ib_warn(dev, "failed to create mkey, %d\n", err);
769 		goto err_in;
770 	}
771 
772 	kfree(in);
773 	*key = mr.key;
774 
775 	return 0;
776 
777 err_in:
778 	kfree(in);
779 
780 	return err;
781 }
782 
783 static void free_pa_mkey(struct mlx5_ib_dev *dev, u32 key)
784 {
785 	struct mlx5_core_mr mr;
786 	int err;
787 
788 	memset(&mr, 0, sizeof(mr));
789 	mr.key = key;
790 	err = mlx5_core_destroy_mkey(&dev->mdev, &mr);
791 	if (err)
792 		mlx5_ib_warn(dev, "failed to destroy mkey 0x%x\n", key);
793 }
794 
795 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
796 				      struct ib_ucontext *context,
797 				      struct ib_udata *udata)
798 {
799 	struct mlx5_ib_alloc_pd_resp resp;
800 	struct mlx5_ib_pd *pd;
801 	int err;
802 
803 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
804 	if (!pd)
805 		return ERR_PTR(-ENOMEM);
806 
807 	err = mlx5_core_alloc_pd(&to_mdev(ibdev)->mdev, &pd->pdn);
808 	if (err) {
809 		kfree(pd);
810 		return ERR_PTR(err);
811 	}
812 
813 	if (context) {
814 		resp.pdn = pd->pdn;
815 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
816 			mlx5_core_dealloc_pd(&to_mdev(ibdev)->mdev, pd->pdn);
817 			kfree(pd);
818 			return ERR_PTR(-EFAULT);
819 		}
820 	} else {
821 		err = alloc_pa_mkey(to_mdev(ibdev), &pd->pa_lkey, pd->pdn);
822 		if (err) {
823 			mlx5_core_dealloc_pd(&to_mdev(ibdev)->mdev, pd->pdn);
824 			kfree(pd);
825 			return ERR_PTR(err);
826 		}
827 	}
828 
829 	return &pd->ibpd;
830 }
831 
832 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
833 {
834 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
835 	struct mlx5_ib_pd *mpd = to_mpd(pd);
836 
837 	if (!pd->uobject)
838 		free_pa_mkey(mdev, mpd->pa_lkey);
839 
840 	mlx5_core_dealloc_pd(&mdev->mdev, mpd->pdn);
841 	kfree(mpd);
842 
843 	return 0;
844 }
845 
846 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
847 {
848 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
849 	int err;
850 
851 	err = mlx5_core_attach_mcg(&dev->mdev, gid, ibqp->qp_num);
852 	if (err)
853 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
854 			     ibqp->qp_num, gid->raw);
855 
856 	return err;
857 }
858 
859 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
860 {
861 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
862 	int err;
863 
864 	err = mlx5_core_detach_mcg(&dev->mdev, gid, ibqp->qp_num);
865 	if (err)
866 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
867 			     ibqp->qp_num, gid->raw);
868 
869 	return err;
870 }
871 
872 static int init_node_data(struct mlx5_ib_dev *dev)
873 {
874 	struct ib_smp *in_mad  = NULL;
875 	struct ib_smp *out_mad = NULL;
876 	int err = -ENOMEM;
877 
878 	in_mad  = kzalloc(sizeof(*in_mad), GFP_KERNEL);
879 	out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
880 	if (!in_mad || !out_mad)
881 		goto out;
882 
883 	init_query_mad(in_mad);
884 	in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
885 
886 	err = mlx5_MAD_IFC(dev, 1, 1, 1, NULL, NULL, in_mad, out_mad);
887 	if (err)
888 		goto out;
889 
890 	memcpy(dev->ib_dev.node_desc, out_mad->data, 64);
891 
892 	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
893 
894 	err = mlx5_MAD_IFC(dev, 1, 1, 1, NULL, NULL, in_mad, out_mad);
895 	if (err)
896 		goto out;
897 
898 	dev->mdev.rev_id = be32_to_cpup((__be32 *)(out_mad->data + 32));
899 	memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
900 
901 out:
902 	kfree(in_mad);
903 	kfree(out_mad);
904 	return err;
905 }
906 
907 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
908 			     char *buf)
909 {
910 	struct mlx5_ib_dev *dev =
911 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
912 
913 	return sprintf(buf, "%d\n", dev->mdev.priv.fw_pages);
914 }
915 
916 static ssize_t show_reg_pages(struct device *device,
917 			      struct device_attribute *attr, char *buf)
918 {
919 	struct mlx5_ib_dev *dev =
920 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
921 
922 	return sprintf(buf, "%d\n", dev->mdev.priv.reg_pages);
923 }
924 
925 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
926 			char *buf)
927 {
928 	struct mlx5_ib_dev *dev =
929 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
930 	return sprintf(buf, "MT%d\n", dev->mdev.pdev->device);
931 }
932 
933 static ssize_t show_fw_ver(struct device *device, struct device_attribute *attr,
934 			   char *buf)
935 {
936 	struct mlx5_ib_dev *dev =
937 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
938 	return sprintf(buf, "%d.%d.%d\n", fw_rev_maj(&dev->mdev),
939 		       fw_rev_min(&dev->mdev), fw_rev_sub(&dev->mdev));
940 }
941 
942 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
943 			char *buf)
944 {
945 	struct mlx5_ib_dev *dev =
946 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
947 	return sprintf(buf, "%x\n", dev->mdev.rev_id);
948 }
949 
950 static ssize_t show_board(struct device *device, struct device_attribute *attr,
951 			  char *buf)
952 {
953 	struct mlx5_ib_dev *dev =
954 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
955 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
956 		       dev->mdev.board_id);
957 }
958 
959 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
960 static DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
961 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
962 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
963 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
964 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
965 
966 static struct device_attribute *mlx5_class_attributes[] = {
967 	&dev_attr_hw_rev,
968 	&dev_attr_fw_ver,
969 	&dev_attr_hca_type,
970 	&dev_attr_board_id,
971 	&dev_attr_fw_pages,
972 	&dev_attr_reg_pages,
973 };
974 
975 static void mlx5_ib_event(struct mlx5_core_dev *dev, enum mlx5_dev_event event,
976 			  void *data)
977 {
978 	struct mlx5_ib_dev *ibdev = container_of(dev, struct mlx5_ib_dev, mdev);
979 	struct ib_event ibev;
980 	u8 port = 0;
981 
982 	switch (event) {
983 	case MLX5_DEV_EVENT_SYS_ERROR:
984 		ibdev->ib_active = false;
985 		ibev.event = IB_EVENT_DEVICE_FATAL;
986 		break;
987 
988 	case MLX5_DEV_EVENT_PORT_UP:
989 		ibev.event = IB_EVENT_PORT_ACTIVE;
990 		port = *(u8 *)data;
991 		break;
992 
993 	case MLX5_DEV_EVENT_PORT_DOWN:
994 		ibev.event = IB_EVENT_PORT_ERR;
995 		port = *(u8 *)data;
996 		break;
997 
998 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
999 		/* not used by ULPs */
1000 		return;
1001 
1002 	case MLX5_DEV_EVENT_LID_CHANGE:
1003 		ibev.event = IB_EVENT_LID_CHANGE;
1004 		port = *(u8 *)data;
1005 		break;
1006 
1007 	case MLX5_DEV_EVENT_PKEY_CHANGE:
1008 		ibev.event = IB_EVENT_PKEY_CHANGE;
1009 		port = *(u8 *)data;
1010 		break;
1011 
1012 	case MLX5_DEV_EVENT_GUID_CHANGE:
1013 		ibev.event = IB_EVENT_GID_CHANGE;
1014 		port = *(u8 *)data;
1015 		break;
1016 
1017 	case MLX5_DEV_EVENT_CLIENT_REREG:
1018 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
1019 		port = *(u8 *)data;
1020 		break;
1021 	}
1022 
1023 	ibev.device	      = &ibdev->ib_dev;
1024 	ibev.element.port_num = port;
1025 
1026 	if (port < 1 || port > ibdev->num_ports) {
1027 		mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
1028 		return;
1029 	}
1030 
1031 	if (ibdev->ib_active)
1032 		ib_dispatch_event(&ibev);
1033 }
1034 
1035 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
1036 {
1037 	int port;
1038 
1039 	for (port = 1; port <= dev->mdev.caps.num_ports; port++)
1040 		mlx5_query_ext_port_caps(dev, port);
1041 }
1042 
1043 static int get_port_caps(struct mlx5_ib_dev *dev)
1044 {
1045 	struct ib_device_attr *dprops = NULL;
1046 	struct ib_port_attr *pprops = NULL;
1047 	int err = 0;
1048 	int port;
1049 
1050 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
1051 	if (!pprops)
1052 		goto out;
1053 
1054 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
1055 	if (!dprops)
1056 		goto out;
1057 
1058 	err = mlx5_ib_query_device(&dev->ib_dev, dprops);
1059 	if (err) {
1060 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
1061 		goto out;
1062 	}
1063 
1064 	for (port = 1; port <= dev->mdev.caps.num_ports; port++) {
1065 		err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
1066 		if (err) {
1067 			mlx5_ib_warn(dev, "query_port %d failed %d\n", port, err);
1068 			break;
1069 		}
1070 		dev->mdev.caps.port[port - 1].pkey_table_len = dprops->max_pkeys;
1071 		dev->mdev.caps.port[port - 1].gid_table_len = pprops->gid_tbl_len;
1072 		mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
1073 			    dprops->max_pkeys, pprops->gid_tbl_len);
1074 	}
1075 
1076 out:
1077 	kfree(pprops);
1078 	kfree(dprops);
1079 
1080 	return err;
1081 }
1082 
1083 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
1084 {
1085 	int err;
1086 
1087 	err = mlx5_mr_cache_cleanup(dev);
1088 	if (err)
1089 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
1090 
1091 	mlx5_ib_destroy_qp(dev->umrc.qp);
1092 	ib_destroy_cq(dev->umrc.cq);
1093 	ib_dereg_mr(dev->umrc.mr);
1094 	ib_dealloc_pd(dev->umrc.pd);
1095 }
1096 
1097 enum {
1098 	MAX_UMR_WR = 128,
1099 };
1100 
1101 static int create_umr_res(struct mlx5_ib_dev *dev)
1102 {
1103 	struct ib_qp_init_attr *init_attr = NULL;
1104 	struct ib_qp_attr *attr = NULL;
1105 	struct ib_pd *pd;
1106 	struct ib_cq *cq;
1107 	struct ib_qp *qp;
1108 	struct ib_mr *mr;
1109 	int ret;
1110 
1111 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
1112 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
1113 	if (!attr || !init_attr) {
1114 		ret = -ENOMEM;
1115 		goto error_0;
1116 	}
1117 
1118 	pd = ib_alloc_pd(&dev->ib_dev);
1119 	if (IS_ERR(pd)) {
1120 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
1121 		ret = PTR_ERR(pd);
1122 		goto error_0;
1123 	}
1124 
1125 	mr = ib_get_dma_mr(pd,  IB_ACCESS_LOCAL_WRITE);
1126 	if (IS_ERR(mr)) {
1127 		mlx5_ib_dbg(dev, "Couldn't create DMA MR for sync UMR QP\n");
1128 		ret = PTR_ERR(mr);
1129 		goto error_1;
1130 	}
1131 
1132 	cq = ib_create_cq(&dev->ib_dev, mlx5_umr_cq_handler, NULL, NULL, 128,
1133 			  0);
1134 	if (IS_ERR(cq)) {
1135 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
1136 		ret = PTR_ERR(cq);
1137 		goto error_2;
1138 	}
1139 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1140 
1141 	init_attr->send_cq = cq;
1142 	init_attr->recv_cq = cq;
1143 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
1144 	init_attr->cap.max_send_wr = MAX_UMR_WR;
1145 	init_attr->cap.max_send_sge = 1;
1146 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
1147 	init_attr->port_num = 1;
1148 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
1149 	if (IS_ERR(qp)) {
1150 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
1151 		ret = PTR_ERR(qp);
1152 		goto error_3;
1153 	}
1154 	qp->device     = &dev->ib_dev;
1155 	qp->real_qp    = qp;
1156 	qp->uobject    = NULL;
1157 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
1158 
1159 	attr->qp_state = IB_QPS_INIT;
1160 	attr->port_num = 1;
1161 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
1162 				IB_QP_PORT, NULL);
1163 	if (ret) {
1164 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
1165 		goto error_4;
1166 	}
1167 
1168 	memset(attr, 0, sizeof(*attr));
1169 	attr->qp_state = IB_QPS_RTR;
1170 	attr->path_mtu = IB_MTU_256;
1171 
1172 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
1173 	if (ret) {
1174 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
1175 		goto error_4;
1176 	}
1177 
1178 	memset(attr, 0, sizeof(*attr));
1179 	attr->qp_state = IB_QPS_RTS;
1180 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
1181 	if (ret) {
1182 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
1183 		goto error_4;
1184 	}
1185 
1186 	dev->umrc.qp = qp;
1187 	dev->umrc.cq = cq;
1188 	dev->umrc.mr = mr;
1189 	dev->umrc.pd = pd;
1190 
1191 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
1192 	ret = mlx5_mr_cache_init(dev);
1193 	if (ret) {
1194 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
1195 		goto error_4;
1196 	}
1197 
1198 	kfree(attr);
1199 	kfree(init_attr);
1200 
1201 	return 0;
1202 
1203 error_4:
1204 	mlx5_ib_destroy_qp(qp);
1205 
1206 error_3:
1207 	ib_destroy_cq(cq);
1208 
1209 error_2:
1210 	ib_dereg_mr(mr);
1211 
1212 error_1:
1213 	ib_dealloc_pd(pd);
1214 
1215 error_0:
1216 	kfree(attr);
1217 	kfree(init_attr);
1218 	return ret;
1219 }
1220 
1221 static int create_dev_resources(struct mlx5_ib_resources *devr)
1222 {
1223 	struct ib_srq_init_attr attr;
1224 	struct mlx5_ib_dev *dev;
1225 	int ret = 0;
1226 
1227 	dev = container_of(devr, struct mlx5_ib_dev, devr);
1228 
1229 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
1230 	if (IS_ERR(devr->p0)) {
1231 		ret = PTR_ERR(devr->p0);
1232 		goto error0;
1233 	}
1234 	devr->p0->device  = &dev->ib_dev;
1235 	devr->p0->uobject = NULL;
1236 	atomic_set(&devr->p0->usecnt, 0);
1237 
1238 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, 1, 0, NULL, NULL);
1239 	if (IS_ERR(devr->c0)) {
1240 		ret = PTR_ERR(devr->c0);
1241 		goto error1;
1242 	}
1243 	devr->c0->device        = &dev->ib_dev;
1244 	devr->c0->uobject       = NULL;
1245 	devr->c0->comp_handler  = NULL;
1246 	devr->c0->event_handler = NULL;
1247 	devr->c0->cq_context    = NULL;
1248 	atomic_set(&devr->c0->usecnt, 0);
1249 
1250 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
1251 	if (IS_ERR(devr->x0)) {
1252 		ret = PTR_ERR(devr->x0);
1253 		goto error2;
1254 	}
1255 	devr->x0->device = &dev->ib_dev;
1256 	devr->x0->inode = NULL;
1257 	atomic_set(&devr->x0->usecnt, 0);
1258 	mutex_init(&devr->x0->tgt_qp_mutex);
1259 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
1260 
1261 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
1262 	if (IS_ERR(devr->x1)) {
1263 		ret = PTR_ERR(devr->x1);
1264 		goto error3;
1265 	}
1266 	devr->x1->device = &dev->ib_dev;
1267 	devr->x1->inode = NULL;
1268 	atomic_set(&devr->x1->usecnt, 0);
1269 	mutex_init(&devr->x1->tgt_qp_mutex);
1270 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
1271 
1272 	memset(&attr, 0, sizeof(attr));
1273 	attr.attr.max_sge = 1;
1274 	attr.attr.max_wr = 1;
1275 	attr.srq_type = IB_SRQT_XRC;
1276 	attr.ext.xrc.cq = devr->c0;
1277 	attr.ext.xrc.xrcd = devr->x0;
1278 
1279 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
1280 	if (IS_ERR(devr->s0)) {
1281 		ret = PTR_ERR(devr->s0);
1282 		goto error4;
1283 	}
1284 	devr->s0->device	= &dev->ib_dev;
1285 	devr->s0->pd		= devr->p0;
1286 	devr->s0->uobject       = NULL;
1287 	devr->s0->event_handler = NULL;
1288 	devr->s0->srq_context   = NULL;
1289 	devr->s0->srq_type      = IB_SRQT_XRC;
1290 	devr->s0->ext.xrc.xrcd	= devr->x0;
1291 	devr->s0->ext.xrc.cq	= devr->c0;
1292 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
1293 	atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
1294 	atomic_inc(&devr->p0->usecnt);
1295 	atomic_set(&devr->s0->usecnt, 0);
1296 
1297 	return 0;
1298 
1299 error4:
1300 	mlx5_ib_dealloc_xrcd(devr->x1);
1301 error3:
1302 	mlx5_ib_dealloc_xrcd(devr->x0);
1303 error2:
1304 	mlx5_ib_destroy_cq(devr->c0);
1305 error1:
1306 	mlx5_ib_dealloc_pd(devr->p0);
1307 error0:
1308 	return ret;
1309 }
1310 
1311 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
1312 {
1313 	mlx5_ib_destroy_srq(devr->s0);
1314 	mlx5_ib_dealloc_xrcd(devr->x0);
1315 	mlx5_ib_dealloc_xrcd(devr->x1);
1316 	mlx5_ib_destroy_cq(devr->c0);
1317 	mlx5_ib_dealloc_pd(devr->p0);
1318 }
1319 
1320 static int init_one(struct pci_dev *pdev,
1321 		    const struct pci_device_id *id)
1322 {
1323 	struct mlx5_core_dev *mdev;
1324 	struct mlx5_ib_dev *dev;
1325 	int err;
1326 	int i;
1327 
1328 	printk_once(KERN_INFO "%s", mlx5_version);
1329 
1330 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
1331 	if (!dev)
1332 		return -ENOMEM;
1333 
1334 	mdev = &dev->mdev;
1335 	mdev->event = mlx5_ib_event;
1336 	if (prof_sel >= ARRAY_SIZE(profile)) {
1337 		pr_warn("selected pofile out of range, selceting default\n");
1338 		prof_sel = 0;
1339 	}
1340 	mdev->profile = &profile[prof_sel];
1341 	err = mlx5_dev_init(mdev, pdev);
1342 	if (err)
1343 		goto err_free;
1344 
1345 	err = get_port_caps(dev);
1346 	if (err)
1347 		goto err_cleanup;
1348 
1349 	get_ext_port_caps(dev);
1350 
1351 	err = alloc_comp_eqs(dev);
1352 	if (err)
1353 		goto err_cleanup;
1354 
1355 	MLX5_INIT_DOORBELL_LOCK(&dev->uar_lock);
1356 
1357 	strlcpy(dev->ib_dev.name, "mlx5_%d", IB_DEVICE_NAME_MAX);
1358 	dev->ib_dev.owner		= THIS_MODULE;
1359 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
1360 	dev->ib_dev.local_dma_lkey	= mdev->caps.reserved_lkey;
1361 	dev->num_ports		= mdev->caps.num_ports;
1362 	dev->ib_dev.phys_port_cnt     = dev->num_ports;
1363 	dev->ib_dev.num_comp_vectors	= dev->num_comp_vectors;
1364 	dev->ib_dev.dma_device	= &mdev->pdev->dev;
1365 
1366 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
1367 	dev->ib_dev.uverbs_cmd_mask	=
1368 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
1369 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
1370 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
1371 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
1372 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
1373 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
1374 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
1375 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
1376 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
1377 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
1378 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
1379 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
1380 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
1381 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
1382 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
1383 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
1384 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
1385 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
1386 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
1387 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
1388 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
1389 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
1390 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
1391 
1392 	dev->ib_dev.query_device	= mlx5_ib_query_device;
1393 	dev->ib_dev.query_port		= mlx5_ib_query_port;
1394 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
1395 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
1396 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
1397 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
1398 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
1399 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
1400 	dev->ib_dev.mmap		= mlx5_ib_mmap;
1401 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
1402 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
1403 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
1404 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
1405 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
1406 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
1407 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
1408 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
1409 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
1410 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
1411 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
1412 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
1413 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
1414 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
1415 	dev->ib_dev.post_send		= mlx5_ib_post_send;
1416 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
1417 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
1418 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
1419 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
1420 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
1421 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
1422 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
1423 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
1424 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
1425 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
1426 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
1427 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
1428 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
1429 	dev->ib_dev.alloc_fast_reg_mr	= mlx5_ib_alloc_fast_reg_mr;
1430 	dev->ib_dev.alloc_fast_reg_page_list = mlx5_ib_alloc_fast_reg_page_list;
1431 	dev->ib_dev.free_fast_reg_page_list  = mlx5_ib_free_fast_reg_page_list;
1432 
1433 	if (mdev->caps.flags & MLX5_DEV_CAP_FLAG_XRC) {
1434 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
1435 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
1436 		dev->ib_dev.uverbs_cmd_mask |=
1437 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
1438 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
1439 	}
1440 
1441 	err = init_node_data(dev);
1442 	if (err)
1443 		goto err_eqs;
1444 
1445 	mutex_init(&dev->cap_mask_mutex);
1446 	spin_lock_init(&dev->mr_lock);
1447 
1448 	err = create_dev_resources(&dev->devr);
1449 	if (err)
1450 		goto err_eqs;
1451 
1452 	err = ib_register_device(&dev->ib_dev, NULL);
1453 	if (err)
1454 		goto err_rsrc;
1455 
1456 	err = create_umr_res(dev);
1457 	if (err)
1458 		goto err_dev;
1459 
1460 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
1461 		err = device_create_file(&dev->ib_dev.dev,
1462 					 mlx5_class_attributes[i]);
1463 		if (err)
1464 			goto err_umrc;
1465 	}
1466 
1467 	dev->ib_active = true;
1468 
1469 	return 0;
1470 
1471 err_umrc:
1472 	destroy_umrc_res(dev);
1473 
1474 err_dev:
1475 	ib_unregister_device(&dev->ib_dev);
1476 
1477 err_rsrc:
1478 	destroy_dev_resources(&dev->devr);
1479 
1480 err_eqs:
1481 	free_comp_eqs(dev);
1482 
1483 err_cleanup:
1484 	mlx5_dev_cleanup(mdev);
1485 
1486 err_free:
1487 	ib_dealloc_device((struct ib_device *)dev);
1488 
1489 	return err;
1490 }
1491 
1492 static void remove_one(struct pci_dev *pdev)
1493 {
1494 	struct mlx5_ib_dev *dev = mlx5_pci2ibdev(pdev);
1495 
1496 	destroy_umrc_res(dev);
1497 	ib_unregister_device(&dev->ib_dev);
1498 	destroy_dev_resources(&dev->devr);
1499 	free_comp_eqs(dev);
1500 	mlx5_dev_cleanup(&dev->mdev);
1501 	ib_dealloc_device(&dev->ib_dev);
1502 }
1503 
1504 static DEFINE_PCI_DEVICE_TABLE(mlx5_ib_pci_table) = {
1505 	{ PCI_VDEVICE(MELLANOX, 4113) }, /* MT4113 Connect-IB */
1506 	{ 0, }
1507 };
1508 
1509 MODULE_DEVICE_TABLE(pci, mlx5_ib_pci_table);
1510 
1511 static struct pci_driver mlx5_ib_driver = {
1512 	.name		= DRIVER_NAME,
1513 	.id_table	= mlx5_ib_pci_table,
1514 	.probe		= init_one,
1515 	.remove		= remove_one
1516 };
1517 
1518 static int __init mlx5_ib_init(void)
1519 {
1520 	return pci_register_driver(&mlx5_ib_driver);
1521 }
1522 
1523 static void __exit mlx5_ib_cleanup(void)
1524 {
1525 	pci_unregister_driver(&mlx5_ib_driver);
1526 }
1527 
1528 module_init(mlx5_ib_init);
1529 module_exit(mlx5_ib_cleanup);
1530