xref: /openbmc/linux/drivers/infiniband/hw/mlx5/main.c (revision d2999e1b)
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 	if (flags & MLX5_DEV_CAP_FLAG_SIG_HAND_OVER) {
277 		props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
278 		/* At this stage no support for signature handover */
279 		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
280 				      IB_PROT_T10DIF_TYPE_2 |
281 				      IB_PROT_T10DIF_TYPE_3;
282 		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
283 				       IB_GUARD_T10DIF_CSUM;
284 	}
285 	if (flags & MLX5_DEV_CAP_FLAG_BLOCK_MCAST)
286 		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
287 
288 	props->vendor_id	   = be32_to_cpup((__be32 *)(out_mad->data + 36)) &
289 		0xffffff;
290 	props->vendor_part_id	   = be16_to_cpup((__be16 *)(out_mad->data + 30));
291 	props->hw_ver		   = be32_to_cpup((__be32 *)(out_mad->data + 32));
292 	memcpy(&props->sys_image_guid, out_mad->data +	4, 8);
293 
294 	props->max_mr_size	   = ~0ull;
295 	props->page_size_cap	   = dev->mdev.caps.min_page_sz;
296 	props->max_qp		   = 1 << dev->mdev.caps.log_max_qp;
297 	props->max_qp_wr	   = dev->mdev.caps.max_wqes;
298 	max_rq_sg = dev->mdev.caps.max_rq_desc_sz / sizeof(struct mlx5_wqe_data_seg);
299 	max_sq_sg = (dev->mdev.caps.max_sq_desc_sz - sizeof(struct mlx5_wqe_ctrl_seg)) /
300 		sizeof(struct mlx5_wqe_data_seg);
301 	props->max_sge = min(max_rq_sg, max_sq_sg);
302 	props->max_cq		   = 1 << dev->mdev.caps.log_max_cq;
303 	props->max_cqe		   = dev->mdev.caps.max_cqes - 1;
304 	props->max_mr		   = 1 << dev->mdev.caps.log_max_mkey;
305 	props->max_pd		   = 1 << dev->mdev.caps.log_max_pd;
306 	props->max_qp_rd_atom	   = dev->mdev.caps.max_ra_req_qp;
307 	props->max_qp_init_rd_atom = dev->mdev.caps.max_ra_res_qp;
308 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
309 	props->max_srq		   = 1 << dev->mdev.caps.log_max_srq;
310 	props->max_srq_wr	   = dev->mdev.caps.max_srq_wqes - 1;
311 	props->max_srq_sge	   = max_rq_sg - 1;
312 	props->max_fast_reg_page_list_len = (unsigned int)-1;
313 	props->local_ca_ack_delay  = dev->mdev.caps.local_ca_ack_delay;
314 	props->atomic_cap	   = IB_ATOMIC_NONE;
315 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
316 	props->max_pkeys	   = be16_to_cpup((__be16 *)(out_mad->data + 28));
317 	props->max_mcast_grp	   = 1 << dev->mdev.caps.log_max_mcg;
318 	props->max_mcast_qp_attach = dev->mdev.caps.max_qp_mcg;
319 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
320 					   props->max_mcast_grp;
321 	props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
322 
323 out:
324 	kfree(in_mad);
325 	kfree(out_mad);
326 
327 	return err;
328 }
329 
330 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
331 		       struct ib_port_attr *props)
332 {
333 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
334 	struct ib_smp *in_mad  = NULL;
335 	struct ib_smp *out_mad = NULL;
336 	int ext_active_speed;
337 	int err = -ENOMEM;
338 
339 	if (port < 1 || port > dev->mdev.caps.num_ports) {
340 		mlx5_ib_warn(dev, "invalid port number %d\n", port);
341 		return -EINVAL;
342 	}
343 
344 	in_mad  = kzalloc(sizeof(*in_mad), GFP_KERNEL);
345 	out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
346 	if (!in_mad || !out_mad)
347 		goto out;
348 
349 	memset(props, 0, sizeof(*props));
350 
351 	init_query_mad(in_mad);
352 	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
353 	in_mad->attr_mod = cpu_to_be32(port);
354 
355 	err = mlx5_MAD_IFC(dev, 1, 1, port, NULL, NULL, in_mad, out_mad);
356 	if (err) {
357 		mlx5_ib_warn(dev, "err %d\n", err);
358 		goto out;
359 	}
360 
361 
362 	props->lid		= be16_to_cpup((__be16 *)(out_mad->data + 16));
363 	props->lmc		= out_mad->data[34] & 0x7;
364 	props->sm_lid		= be16_to_cpup((__be16 *)(out_mad->data + 18));
365 	props->sm_sl		= out_mad->data[36] & 0xf;
366 	props->state		= out_mad->data[32] & 0xf;
367 	props->phys_state	= out_mad->data[33] >> 4;
368 	props->port_cap_flags	= be32_to_cpup((__be32 *)(out_mad->data + 20));
369 	props->gid_tbl_len	= out_mad->data[50];
370 	props->max_msg_sz	= 1 << to_mdev(ibdev)->mdev.caps.log_max_msg;
371 	props->pkey_tbl_len	= to_mdev(ibdev)->mdev.caps.port[port - 1].pkey_table_len;
372 	props->bad_pkey_cntr	= be16_to_cpup((__be16 *)(out_mad->data + 46));
373 	props->qkey_viol_cntr	= be16_to_cpup((__be16 *)(out_mad->data + 48));
374 	props->active_width	= out_mad->data[31] & 0xf;
375 	props->active_speed	= out_mad->data[35] >> 4;
376 	props->max_mtu		= out_mad->data[41] & 0xf;
377 	props->active_mtu	= out_mad->data[36] >> 4;
378 	props->subnet_timeout	= out_mad->data[51] & 0x1f;
379 	props->max_vl_num	= out_mad->data[37] >> 4;
380 	props->init_type_reply	= out_mad->data[41] >> 4;
381 
382 	/* Check if extended speeds (EDR/FDR/...) are supported */
383 	if (props->port_cap_flags & IB_PORT_EXTENDED_SPEEDS_SUP) {
384 		ext_active_speed = out_mad->data[62] >> 4;
385 
386 		switch (ext_active_speed) {
387 		case 1:
388 			props->active_speed = 16; /* FDR */
389 			break;
390 		case 2:
391 			props->active_speed = 32; /* EDR */
392 			break;
393 		}
394 	}
395 
396 	/* If reported active speed is QDR, check if is FDR-10 */
397 	if (props->active_speed == 4) {
398 		if (dev->mdev.caps.ext_port_cap[port - 1] &
399 		    MLX_EXT_PORT_CAP_FLAG_EXTENDED_PORT_INFO) {
400 			init_query_mad(in_mad);
401 			in_mad->attr_id = MLX5_ATTR_EXTENDED_PORT_INFO;
402 			in_mad->attr_mod = cpu_to_be32(port);
403 
404 			err = mlx5_MAD_IFC(dev, 1, 1, port,
405 					   NULL, NULL, in_mad, out_mad);
406 			if (err)
407 				goto out;
408 
409 			/* Checking LinkSpeedActive for FDR-10 */
410 			if (out_mad->data[15] & 0x1)
411 				props->active_speed = 8;
412 		}
413 	}
414 
415 out:
416 	kfree(in_mad);
417 	kfree(out_mad);
418 
419 	return err;
420 }
421 
422 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
423 			     union ib_gid *gid)
424 {
425 	struct ib_smp *in_mad  = NULL;
426 	struct ib_smp *out_mad = NULL;
427 	int err = -ENOMEM;
428 
429 	in_mad  = kzalloc(sizeof(*in_mad), GFP_KERNEL);
430 	out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
431 	if (!in_mad || !out_mad)
432 		goto out;
433 
434 	init_query_mad(in_mad);
435 	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
436 	in_mad->attr_mod = cpu_to_be32(port);
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, out_mad->data + 8, 8);
443 
444 	init_query_mad(in_mad);
445 	in_mad->attr_id  = IB_SMP_ATTR_GUID_INFO;
446 	in_mad->attr_mod = cpu_to_be32(index / 8);
447 
448 	err = mlx5_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
449 	if (err)
450 		goto out;
451 
452 	memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
453 
454 out:
455 	kfree(in_mad);
456 	kfree(out_mad);
457 	return err;
458 }
459 
460 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
461 			      u16 *pkey)
462 {
463 	struct ib_smp *in_mad  = NULL;
464 	struct ib_smp *out_mad = NULL;
465 	int err = -ENOMEM;
466 
467 	in_mad  = kzalloc(sizeof(*in_mad), GFP_KERNEL);
468 	out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
469 	if (!in_mad || !out_mad)
470 		goto out;
471 
472 	init_query_mad(in_mad);
473 	in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
474 	in_mad->attr_mod = cpu_to_be32(index / 32);
475 
476 	err = mlx5_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
477 	if (err)
478 		goto out;
479 
480 	*pkey = be16_to_cpu(((__be16 *)out_mad->data)[index % 32]);
481 
482 out:
483 	kfree(in_mad);
484 	kfree(out_mad);
485 	return err;
486 }
487 
488 struct mlx5_reg_node_desc {
489 	u8	desc[64];
490 };
491 
492 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
493 				 struct ib_device_modify *props)
494 {
495 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
496 	struct mlx5_reg_node_desc in;
497 	struct mlx5_reg_node_desc out;
498 	int err;
499 
500 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
501 		return -EOPNOTSUPP;
502 
503 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
504 		return 0;
505 
506 	/*
507 	 * If possible, pass node desc to FW, so it can generate
508 	 * a 144 trap.  If cmd fails, just ignore.
509 	 */
510 	memcpy(&in, props->node_desc, 64);
511 	err = mlx5_core_access_reg(&dev->mdev, &in, sizeof(in), &out,
512 				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
513 	if (err)
514 		return err;
515 
516 	memcpy(ibdev->node_desc, props->node_desc, 64);
517 
518 	return err;
519 }
520 
521 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
522 			       struct ib_port_modify *props)
523 {
524 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
525 	struct ib_port_attr attr;
526 	u32 tmp;
527 	int err;
528 
529 	mutex_lock(&dev->cap_mask_mutex);
530 
531 	err = mlx5_ib_query_port(ibdev, port, &attr);
532 	if (err)
533 		goto out;
534 
535 	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
536 		~props->clr_port_cap_mask;
537 
538 	err = mlx5_set_port_caps(&dev->mdev, port, tmp);
539 
540 out:
541 	mutex_unlock(&dev->cap_mask_mutex);
542 	return err;
543 }
544 
545 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
546 						  struct ib_udata *udata)
547 {
548 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
549 	struct mlx5_ib_alloc_ucontext_req_v2 req;
550 	struct mlx5_ib_alloc_ucontext_resp resp;
551 	struct mlx5_ib_ucontext *context;
552 	struct mlx5_uuar_info *uuari;
553 	struct mlx5_uar *uars;
554 	int gross_uuars;
555 	int num_uars;
556 	int ver;
557 	int uuarn;
558 	int err;
559 	int i;
560 	int reqlen;
561 
562 	if (!dev->ib_active)
563 		return ERR_PTR(-EAGAIN);
564 
565 	memset(&req, 0, sizeof(req));
566 	reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
567 	if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
568 		ver = 0;
569 	else if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req_v2))
570 		ver = 2;
571 	else
572 		return ERR_PTR(-EINVAL);
573 
574 	err = ib_copy_from_udata(&req, udata, reqlen);
575 	if (err)
576 		return ERR_PTR(err);
577 
578 	if (req.flags || req.reserved)
579 		return ERR_PTR(-EINVAL);
580 
581 	if (req.total_num_uuars > MLX5_MAX_UUARS)
582 		return ERR_PTR(-ENOMEM);
583 
584 	if (req.total_num_uuars == 0)
585 		return ERR_PTR(-EINVAL);
586 
587 	req.total_num_uuars = ALIGN(req.total_num_uuars,
588 				    MLX5_NON_FP_BF_REGS_PER_PAGE);
589 	if (req.num_low_latency_uuars > req.total_num_uuars - 1)
590 		return ERR_PTR(-EINVAL);
591 
592 	num_uars = req.total_num_uuars / MLX5_NON_FP_BF_REGS_PER_PAGE;
593 	gross_uuars = num_uars * MLX5_BF_REGS_PER_PAGE;
594 	resp.qp_tab_size      = 1 << dev->mdev.caps.log_max_qp;
595 	resp.bf_reg_size      = dev->mdev.caps.bf_reg_size;
596 	resp.cache_line_size  = L1_CACHE_BYTES;
597 	resp.max_sq_desc_sz = dev->mdev.caps.max_sq_desc_sz;
598 	resp.max_rq_desc_sz = dev->mdev.caps.max_rq_desc_sz;
599 	resp.max_send_wqebb = dev->mdev.caps.max_wqes;
600 	resp.max_recv_wr = dev->mdev.caps.max_wqes;
601 	resp.max_srq_recv_wr = dev->mdev.caps.max_srq_wqes;
602 
603 	context = kzalloc(sizeof(*context), GFP_KERNEL);
604 	if (!context)
605 		return ERR_PTR(-ENOMEM);
606 
607 	uuari = &context->uuari;
608 	mutex_init(&uuari->lock);
609 	uars = kcalloc(num_uars, sizeof(*uars), GFP_KERNEL);
610 	if (!uars) {
611 		err = -ENOMEM;
612 		goto out_ctx;
613 	}
614 
615 	uuari->bitmap = kcalloc(BITS_TO_LONGS(gross_uuars),
616 				sizeof(*uuari->bitmap),
617 				GFP_KERNEL);
618 	if (!uuari->bitmap) {
619 		err = -ENOMEM;
620 		goto out_uar_ctx;
621 	}
622 	/*
623 	 * clear all fast path uuars
624 	 */
625 	for (i = 0; i < gross_uuars; i++) {
626 		uuarn = i & 3;
627 		if (uuarn == 2 || uuarn == 3)
628 			set_bit(i, uuari->bitmap);
629 	}
630 
631 	uuari->count = kcalloc(gross_uuars, sizeof(*uuari->count), GFP_KERNEL);
632 	if (!uuari->count) {
633 		err = -ENOMEM;
634 		goto out_bitmap;
635 	}
636 
637 	for (i = 0; i < num_uars; i++) {
638 		err = mlx5_cmd_alloc_uar(&dev->mdev, &uars[i].index);
639 		if (err)
640 			goto out_count;
641 	}
642 
643 	INIT_LIST_HEAD(&context->db_page_list);
644 	mutex_init(&context->db_page_mutex);
645 
646 	resp.tot_uuars = req.total_num_uuars;
647 	resp.num_ports = dev->mdev.caps.num_ports;
648 	err = ib_copy_to_udata(udata, &resp,
649 			       sizeof(resp) - sizeof(resp.reserved));
650 	if (err)
651 		goto out_uars;
652 
653 	uuari->ver = ver;
654 	uuari->num_low_latency_uuars = req.num_low_latency_uuars;
655 	uuari->uars = uars;
656 	uuari->num_uars = num_uars;
657 	return &context->ibucontext;
658 
659 out_uars:
660 	for (i--; i >= 0; i--)
661 		mlx5_cmd_free_uar(&dev->mdev, uars[i].index);
662 out_count:
663 	kfree(uuari->count);
664 
665 out_bitmap:
666 	kfree(uuari->bitmap);
667 
668 out_uar_ctx:
669 	kfree(uars);
670 
671 out_ctx:
672 	kfree(context);
673 	return ERR_PTR(err);
674 }
675 
676 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
677 {
678 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
679 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
680 	struct mlx5_uuar_info *uuari = &context->uuari;
681 	int i;
682 
683 	for (i = 0; i < uuari->num_uars; i++) {
684 		if (mlx5_cmd_free_uar(&dev->mdev, uuari->uars[i].index))
685 			mlx5_ib_warn(dev, "failed to free UAR 0x%x\n", uuari->uars[i].index);
686 	}
687 
688 	kfree(uuari->count);
689 	kfree(uuari->bitmap);
690 	kfree(uuari->uars);
691 	kfree(context);
692 
693 	return 0;
694 }
695 
696 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, int index)
697 {
698 	return (pci_resource_start(dev->mdev.pdev, 0) >> PAGE_SHIFT) + index;
699 }
700 
701 static int get_command(unsigned long offset)
702 {
703 	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
704 }
705 
706 static int get_arg(unsigned long offset)
707 {
708 	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
709 }
710 
711 static int get_index(unsigned long offset)
712 {
713 	return get_arg(offset);
714 }
715 
716 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
717 {
718 	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
719 	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
720 	struct mlx5_uuar_info *uuari = &context->uuari;
721 	unsigned long command;
722 	unsigned long idx;
723 	phys_addr_t pfn;
724 
725 	command = get_command(vma->vm_pgoff);
726 	switch (command) {
727 	case MLX5_IB_MMAP_REGULAR_PAGE:
728 		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
729 			return -EINVAL;
730 
731 		idx = get_index(vma->vm_pgoff);
732 		pfn = uar_index2pfn(dev, uuari->uars[idx].index);
733 		mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn 0x%llx\n", idx,
734 			    (unsigned long long)pfn);
735 
736 		if (idx >= uuari->num_uars)
737 			return -EINVAL;
738 
739 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
740 		if (io_remap_pfn_range(vma, vma->vm_start, pfn,
741 				       PAGE_SIZE, vma->vm_page_prot))
742 			return -EAGAIN;
743 
744 		mlx5_ib_dbg(dev, "mapped WC at 0x%lx, PA 0x%llx\n",
745 			    vma->vm_start,
746 			    (unsigned long long)pfn << PAGE_SHIFT);
747 		break;
748 
749 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
750 		return -ENOSYS;
751 
752 	default:
753 		return -EINVAL;
754 	}
755 
756 	return 0;
757 }
758 
759 static int alloc_pa_mkey(struct mlx5_ib_dev *dev, u32 *key, u32 pdn)
760 {
761 	struct mlx5_create_mkey_mbox_in *in;
762 	struct mlx5_mkey_seg *seg;
763 	struct mlx5_core_mr mr;
764 	int err;
765 
766 	in = kzalloc(sizeof(*in), GFP_KERNEL);
767 	if (!in)
768 		return -ENOMEM;
769 
770 	seg = &in->seg;
771 	seg->flags = MLX5_PERM_LOCAL_READ | MLX5_ACCESS_MODE_PA;
772 	seg->flags_pd = cpu_to_be32(pdn | MLX5_MKEY_LEN64);
773 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff << 8);
774 	seg->start_addr = 0;
775 
776 	err = mlx5_core_create_mkey(&dev->mdev, &mr, in, sizeof(*in),
777 				    NULL, NULL, NULL);
778 	if (err) {
779 		mlx5_ib_warn(dev, "failed to create mkey, %d\n", err);
780 		goto err_in;
781 	}
782 
783 	kfree(in);
784 	*key = mr.key;
785 
786 	return 0;
787 
788 err_in:
789 	kfree(in);
790 
791 	return err;
792 }
793 
794 static void free_pa_mkey(struct mlx5_ib_dev *dev, u32 key)
795 {
796 	struct mlx5_core_mr mr;
797 	int err;
798 
799 	memset(&mr, 0, sizeof(mr));
800 	mr.key = key;
801 	err = mlx5_core_destroy_mkey(&dev->mdev, &mr);
802 	if (err)
803 		mlx5_ib_warn(dev, "failed to destroy mkey 0x%x\n", key);
804 }
805 
806 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
807 				      struct ib_ucontext *context,
808 				      struct ib_udata *udata)
809 {
810 	struct mlx5_ib_alloc_pd_resp resp;
811 	struct mlx5_ib_pd *pd;
812 	int err;
813 
814 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
815 	if (!pd)
816 		return ERR_PTR(-ENOMEM);
817 
818 	err = mlx5_core_alloc_pd(&to_mdev(ibdev)->mdev, &pd->pdn);
819 	if (err) {
820 		kfree(pd);
821 		return ERR_PTR(err);
822 	}
823 
824 	if (context) {
825 		resp.pdn = pd->pdn;
826 		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
827 			mlx5_core_dealloc_pd(&to_mdev(ibdev)->mdev, pd->pdn);
828 			kfree(pd);
829 			return ERR_PTR(-EFAULT);
830 		}
831 	} else {
832 		err = alloc_pa_mkey(to_mdev(ibdev), &pd->pa_lkey, pd->pdn);
833 		if (err) {
834 			mlx5_core_dealloc_pd(&to_mdev(ibdev)->mdev, pd->pdn);
835 			kfree(pd);
836 			return ERR_PTR(err);
837 		}
838 	}
839 
840 	return &pd->ibpd;
841 }
842 
843 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
844 {
845 	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
846 	struct mlx5_ib_pd *mpd = to_mpd(pd);
847 
848 	if (!pd->uobject)
849 		free_pa_mkey(mdev, mpd->pa_lkey);
850 
851 	mlx5_core_dealloc_pd(&mdev->mdev, mpd->pdn);
852 	kfree(mpd);
853 
854 	return 0;
855 }
856 
857 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
858 {
859 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
860 	int err;
861 
862 	err = mlx5_core_attach_mcg(&dev->mdev, gid, ibqp->qp_num);
863 	if (err)
864 		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
865 			     ibqp->qp_num, gid->raw);
866 
867 	return err;
868 }
869 
870 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
871 {
872 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
873 	int err;
874 
875 	err = mlx5_core_detach_mcg(&dev->mdev, gid, ibqp->qp_num);
876 	if (err)
877 		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
878 			     ibqp->qp_num, gid->raw);
879 
880 	return err;
881 }
882 
883 static int init_node_data(struct mlx5_ib_dev *dev)
884 {
885 	struct ib_smp *in_mad  = NULL;
886 	struct ib_smp *out_mad = NULL;
887 	int err = -ENOMEM;
888 
889 	in_mad  = kzalloc(sizeof(*in_mad), GFP_KERNEL);
890 	out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
891 	if (!in_mad || !out_mad)
892 		goto out;
893 
894 	init_query_mad(in_mad);
895 	in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
896 
897 	err = mlx5_MAD_IFC(dev, 1, 1, 1, NULL, NULL, in_mad, out_mad);
898 	if (err)
899 		goto out;
900 
901 	memcpy(dev->ib_dev.node_desc, out_mad->data, 64);
902 
903 	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
904 
905 	err = mlx5_MAD_IFC(dev, 1, 1, 1, NULL, NULL, in_mad, out_mad);
906 	if (err)
907 		goto out;
908 
909 	dev->mdev.rev_id = be32_to_cpup((__be32 *)(out_mad->data + 32));
910 	memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
911 
912 out:
913 	kfree(in_mad);
914 	kfree(out_mad);
915 	return err;
916 }
917 
918 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
919 			     char *buf)
920 {
921 	struct mlx5_ib_dev *dev =
922 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
923 
924 	return sprintf(buf, "%d\n", dev->mdev.priv.fw_pages);
925 }
926 
927 static ssize_t show_reg_pages(struct device *device,
928 			      struct device_attribute *attr, char *buf)
929 {
930 	struct mlx5_ib_dev *dev =
931 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
932 
933 	return sprintf(buf, "%d\n", dev->mdev.priv.reg_pages);
934 }
935 
936 static ssize_t show_hca(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, "MT%d\n", dev->mdev.pdev->device);
942 }
943 
944 static ssize_t show_fw_ver(struct device *device, struct device_attribute *attr,
945 			   char *buf)
946 {
947 	struct mlx5_ib_dev *dev =
948 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
949 	return sprintf(buf, "%d.%d.%d\n", fw_rev_maj(&dev->mdev),
950 		       fw_rev_min(&dev->mdev), fw_rev_sub(&dev->mdev));
951 }
952 
953 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
954 			char *buf)
955 {
956 	struct mlx5_ib_dev *dev =
957 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
958 	return sprintf(buf, "%x\n", dev->mdev.rev_id);
959 }
960 
961 static ssize_t show_board(struct device *device, struct device_attribute *attr,
962 			  char *buf)
963 {
964 	struct mlx5_ib_dev *dev =
965 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
966 	return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
967 		       dev->mdev.board_id);
968 }
969 
970 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
971 static DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
972 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
973 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
974 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
975 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
976 
977 static struct device_attribute *mlx5_class_attributes[] = {
978 	&dev_attr_hw_rev,
979 	&dev_attr_fw_ver,
980 	&dev_attr_hca_type,
981 	&dev_attr_board_id,
982 	&dev_attr_fw_pages,
983 	&dev_attr_reg_pages,
984 };
985 
986 static void mlx5_ib_event(struct mlx5_core_dev *dev, enum mlx5_dev_event event,
987 			  void *data)
988 {
989 	struct mlx5_ib_dev *ibdev = container_of(dev, struct mlx5_ib_dev, mdev);
990 	struct ib_event ibev;
991 	u8 port = 0;
992 
993 	switch (event) {
994 	case MLX5_DEV_EVENT_SYS_ERROR:
995 		ibdev->ib_active = false;
996 		ibev.event = IB_EVENT_DEVICE_FATAL;
997 		break;
998 
999 	case MLX5_DEV_EVENT_PORT_UP:
1000 		ibev.event = IB_EVENT_PORT_ACTIVE;
1001 		port = *(u8 *)data;
1002 		break;
1003 
1004 	case MLX5_DEV_EVENT_PORT_DOWN:
1005 		ibev.event = IB_EVENT_PORT_ERR;
1006 		port = *(u8 *)data;
1007 		break;
1008 
1009 	case MLX5_DEV_EVENT_PORT_INITIALIZED:
1010 		/* not used by ULPs */
1011 		return;
1012 
1013 	case MLX5_DEV_EVENT_LID_CHANGE:
1014 		ibev.event = IB_EVENT_LID_CHANGE;
1015 		port = *(u8 *)data;
1016 		break;
1017 
1018 	case MLX5_DEV_EVENT_PKEY_CHANGE:
1019 		ibev.event = IB_EVENT_PKEY_CHANGE;
1020 		port = *(u8 *)data;
1021 		break;
1022 
1023 	case MLX5_DEV_EVENT_GUID_CHANGE:
1024 		ibev.event = IB_EVENT_GID_CHANGE;
1025 		port = *(u8 *)data;
1026 		break;
1027 
1028 	case MLX5_DEV_EVENT_CLIENT_REREG:
1029 		ibev.event = IB_EVENT_CLIENT_REREGISTER;
1030 		port = *(u8 *)data;
1031 		break;
1032 	}
1033 
1034 	ibev.device	      = &ibdev->ib_dev;
1035 	ibev.element.port_num = port;
1036 
1037 	if (port < 1 || port > ibdev->num_ports) {
1038 		mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
1039 		return;
1040 	}
1041 
1042 	if (ibdev->ib_active)
1043 		ib_dispatch_event(&ibev);
1044 }
1045 
1046 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
1047 {
1048 	int port;
1049 
1050 	for (port = 1; port <= dev->mdev.caps.num_ports; port++)
1051 		mlx5_query_ext_port_caps(dev, port);
1052 }
1053 
1054 static int get_port_caps(struct mlx5_ib_dev *dev)
1055 {
1056 	struct ib_device_attr *dprops = NULL;
1057 	struct ib_port_attr *pprops = NULL;
1058 	int err = 0;
1059 	int port;
1060 
1061 	pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
1062 	if (!pprops)
1063 		goto out;
1064 
1065 	dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
1066 	if (!dprops)
1067 		goto out;
1068 
1069 	err = mlx5_ib_query_device(&dev->ib_dev, dprops);
1070 	if (err) {
1071 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
1072 		goto out;
1073 	}
1074 
1075 	for (port = 1; port <= dev->mdev.caps.num_ports; port++) {
1076 		err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
1077 		if (err) {
1078 			mlx5_ib_warn(dev, "query_port %d failed %d\n", port, err);
1079 			break;
1080 		}
1081 		dev->mdev.caps.port[port - 1].pkey_table_len = dprops->max_pkeys;
1082 		dev->mdev.caps.port[port - 1].gid_table_len = pprops->gid_tbl_len;
1083 		mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
1084 			    dprops->max_pkeys, pprops->gid_tbl_len);
1085 	}
1086 
1087 out:
1088 	kfree(pprops);
1089 	kfree(dprops);
1090 
1091 	return err;
1092 }
1093 
1094 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
1095 {
1096 	int err;
1097 
1098 	err = mlx5_mr_cache_cleanup(dev);
1099 	if (err)
1100 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
1101 
1102 	mlx5_ib_destroy_qp(dev->umrc.qp);
1103 	ib_destroy_cq(dev->umrc.cq);
1104 	ib_dereg_mr(dev->umrc.mr);
1105 	ib_dealloc_pd(dev->umrc.pd);
1106 }
1107 
1108 enum {
1109 	MAX_UMR_WR = 128,
1110 };
1111 
1112 static int create_umr_res(struct mlx5_ib_dev *dev)
1113 {
1114 	struct ib_qp_init_attr *init_attr = NULL;
1115 	struct ib_qp_attr *attr = NULL;
1116 	struct ib_pd *pd;
1117 	struct ib_cq *cq;
1118 	struct ib_qp *qp;
1119 	struct ib_mr *mr;
1120 	int ret;
1121 
1122 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
1123 	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
1124 	if (!attr || !init_attr) {
1125 		ret = -ENOMEM;
1126 		goto error_0;
1127 	}
1128 
1129 	pd = ib_alloc_pd(&dev->ib_dev);
1130 	if (IS_ERR(pd)) {
1131 		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
1132 		ret = PTR_ERR(pd);
1133 		goto error_0;
1134 	}
1135 
1136 	mr = ib_get_dma_mr(pd,  IB_ACCESS_LOCAL_WRITE);
1137 	if (IS_ERR(mr)) {
1138 		mlx5_ib_dbg(dev, "Couldn't create DMA MR for sync UMR QP\n");
1139 		ret = PTR_ERR(mr);
1140 		goto error_1;
1141 	}
1142 
1143 	cq = ib_create_cq(&dev->ib_dev, mlx5_umr_cq_handler, NULL, NULL, 128,
1144 			  0);
1145 	if (IS_ERR(cq)) {
1146 		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
1147 		ret = PTR_ERR(cq);
1148 		goto error_2;
1149 	}
1150 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1151 
1152 	init_attr->send_cq = cq;
1153 	init_attr->recv_cq = cq;
1154 	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
1155 	init_attr->cap.max_send_wr = MAX_UMR_WR;
1156 	init_attr->cap.max_send_sge = 1;
1157 	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
1158 	init_attr->port_num = 1;
1159 	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
1160 	if (IS_ERR(qp)) {
1161 		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
1162 		ret = PTR_ERR(qp);
1163 		goto error_3;
1164 	}
1165 	qp->device     = &dev->ib_dev;
1166 	qp->real_qp    = qp;
1167 	qp->uobject    = NULL;
1168 	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
1169 
1170 	attr->qp_state = IB_QPS_INIT;
1171 	attr->port_num = 1;
1172 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
1173 				IB_QP_PORT, NULL);
1174 	if (ret) {
1175 		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
1176 		goto error_4;
1177 	}
1178 
1179 	memset(attr, 0, sizeof(*attr));
1180 	attr->qp_state = IB_QPS_RTR;
1181 	attr->path_mtu = IB_MTU_256;
1182 
1183 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
1184 	if (ret) {
1185 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
1186 		goto error_4;
1187 	}
1188 
1189 	memset(attr, 0, sizeof(*attr));
1190 	attr->qp_state = IB_QPS_RTS;
1191 	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
1192 	if (ret) {
1193 		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
1194 		goto error_4;
1195 	}
1196 
1197 	dev->umrc.qp = qp;
1198 	dev->umrc.cq = cq;
1199 	dev->umrc.mr = mr;
1200 	dev->umrc.pd = pd;
1201 
1202 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
1203 	ret = mlx5_mr_cache_init(dev);
1204 	if (ret) {
1205 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
1206 		goto error_4;
1207 	}
1208 
1209 	kfree(attr);
1210 	kfree(init_attr);
1211 
1212 	return 0;
1213 
1214 error_4:
1215 	mlx5_ib_destroy_qp(qp);
1216 
1217 error_3:
1218 	ib_destroy_cq(cq);
1219 
1220 error_2:
1221 	ib_dereg_mr(mr);
1222 
1223 error_1:
1224 	ib_dealloc_pd(pd);
1225 
1226 error_0:
1227 	kfree(attr);
1228 	kfree(init_attr);
1229 	return ret;
1230 }
1231 
1232 static int create_dev_resources(struct mlx5_ib_resources *devr)
1233 {
1234 	struct ib_srq_init_attr attr;
1235 	struct mlx5_ib_dev *dev;
1236 	int ret = 0;
1237 
1238 	dev = container_of(devr, struct mlx5_ib_dev, devr);
1239 
1240 	devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
1241 	if (IS_ERR(devr->p0)) {
1242 		ret = PTR_ERR(devr->p0);
1243 		goto error0;
1244 	}
1245 	devr->p0->device  = &dev->ib_dev;
1246 	devr->p0->uobject = NULL;
1247 	atomic_set(&devr->p0->usecnt, 0);
1248 
1249 	devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, 1, 0, NULL, NULL);
1250 	if (IS_ERR(devr->c0)) {
1251 		ret = PTR_ERR(devr->c0);
1252 		goto error1;
1253 	}
1254 	devr->c0->device        = &dev->ib_dev;
1255 	devr->c0->uobject       = NULL;
1256 	devr->c0->comp_handler  = NULL;
1257 	devr->c0->event_handler = NULL;
1258 	devr->c0->cq_context    = NULL;
1259 	atomic_set(&devr->c0->usecnt, 0);
1260 
1261 	devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
1262 	if (IS_ERR(devr->x0)) {
1263 		ret = PTR_ERR(devr->x0);
1264 		goto error2;
1265 	}
1266 	devr->x0->device = &dev->ib_dev;
1267 	devr->x0->inode = NULL;
1268 	atomic_set(&devr->x0->usecnt, 0);
1269 	mutex_init(&devr->x0->tgt_qp_mutex);
1270 	INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
1271 
1272 	devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
1273 	if (IS_ERR(devr->x1)) {
1274 		ret = PTR_ERR(devr->x1);
1275 		goto error3;
1276 	}
1277 	devr->x1->device = &dev->ib_dev;
1278 	devr->x1->inode = NULL;
1279 	atomic_set(&devr->x1->usecnt, 0);
1280 	mutex_init(&devr->x1->tgt_qp_mutex);
1281 	INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
1282 
1283 	memset(&attr, 0, sizeof(attr));
1284 	attr.attr.max_sge = 1;
1285 	attr.attr.max_wr = 1;
1286 	attr.srq_type = IB_SRQT_XRC;
1287 	attr.ext.xrc.cq = devr->c0;
1288 	attr.ext.xrc.xrcd = devr->x0;
1289 
1290 	devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
1291 	if (IS_ERR(devr->s0)) {
1292 		ret = PTR_ERR(devr->s0);
1293 		goto error4;
1294 	}
1295 	devr->s0->device	= &dev->ib_dev;
1296 	devr->s0->pd		= devr->p0;
1297 	devr->s0->uobject       = NULL;
1298 	devr->s0->event_handler = NULL;
1299 	devr->s0->srq_context   = NULL;
1300 	devr->s0->srq_type      = IB_SRQT_XRC;
1301 	devr->s0->ext.xrc.xrcd	= devr->x0;
1302 	devr->s0->ext.xrc.cq	= devr->c0;
1303 	atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
1304 	atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
1305 	atomic_inc(&devr->p0->usecnt);
1306 	atomic_set(&devr->s0->usecnt, 0);
1307 
1308 	return 0;
1309 
1310 error4:
1311 	mlx5_ib_dealloc_xrcd(devr->x1);
1312 error3:
1313 	mlx5_ib_dealloc_xrcd(devr->x0);
1314 error2:
1315 	mlx5_ib_destroy_cq(devr->c0);
1316 error1:
1317 	mlx5_ib_dealloc_pd(devr->p0);
1318 error0:
1319 	return ret;
1320 }
1321 
1322 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
1323 {
1324 	mlx5_ib_destroy_srq(devr->s0);
1325 	mlx5_ib_dealloc_xrcd(devr->x0);
1326 	mlx5_ib_dealloc_xrcd(devr->x1);
1327 	mlx5_ib_destroy_cq(devr->c0);
1328 	mlx5_ib_dealloc_pd(devr->p0);
1329 }
1330 
1331 static int init_one(struct pci_dev *pdev,
1332 		    const struct pci_device_id *id)
1333 {
1334 	struct mlx5_core_dev *mdev;
1335 	struct mlx5_ib_dev *dev;
1336 	int err;
1337 	int i;
1338 
1339 	printk_once(KERN_INFO "%s", mlx5_version);
1340 
1341 	dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
1342 	if (!dev)
1343 		return -ENOMEM;
1344 
1345 	mdev = &dev->mdev;
1346 	mdev->event = mlx5_ib_event;
1347 	if (prof_sel >= ARRAY_SIZE(profile)) {
1348 		pr_warn("selected pofile out of range, selceting default\n");
1349 		prof_sel = 0;
1350 	}
1351 	mdev->profile = &profile[prof_sel];
1352 	err = mlx5_dev_init(mdev, pdev);
1353 	if (err)
1354 		goto err_free;
1355 
1356 	err = get_port_caps(dev);
1357 	if (err)
1358 		goto err_cleanup;
1359 
1360 	get_ext_port_caps(dev);
1361 
1362 	err = alloc_comp_eqs(dev);
1363 	if (err)
1364 		goto err_cleanup;
1365 
1366 	MLX5_INIT_DOORBELL_LOCK(&dev->uar_lock);
1367 
1368 	strlcpy(dev->ib_dev.name, "mlx5_%d", IB_DEVICE_NAME_MAX);
1369 	dev->ib_dev.owner		= THIS_MODULE;
1370 	dev->ib_dev.node_type		= RDMA_NODE_IB_CA;
1371 	dev->ib_dev.local_dma_lkey	= mdev->caps.reserved_lkey;
1372 	dev->num_ports		= mdev->caps.num_ports;
1373 	dev->ib_dev.phys_port_cnt     = dev->num_ports;
1374 	dev->ib_dev.num_comp_vectors	= dev->num_comp_vectors;
1375 	dev->ib_dev.dma_device	= &mdev->pdev->dev;
1376 
1377 	dev->ib_dev.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION;
1378 	dev->ib_dev.uverbs_cmd_mask	=
1379 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
1380 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
1381 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
1382 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
1383 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
1384 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
1385 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
1386 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
1387 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
1388 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
1389 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
1390 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
1391 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
1392 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
1393 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
1394 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
1395 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
1396 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
1397 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
1398 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
1399 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
1400 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
1401 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
1402 
1403 	dev->ib_dev.query_device	= mlx5_ib_query_device;
1404 	dev->ib_dev.query_port		= mlx5_ib_query_port;
1405 	dev->ib_dev.query_gid		= mlx5_ib_query_gid;
1406 	dev->ib_dev.query_pkey		= mlx5_ib_query_pkey;
1407 	dev->ib_dev.modify_device	= mlx5_ib_modify_device;
1408 	dev->ib_dev.modify_port		= mlx5_ib_modify_port;
1409 	dev->ib_dev.alloc_ucontext	= mlx5_ib_alloc_ucontext;
1410 	dev->ib_dev.dealloc_ucontext	= mlx5_ib_dealloc_ucontext;
1411 	dev->ib_dev.mmap		= mlx5_ib_mmap;
1412 	dev->ib_dev.alloc_pd		= mlx5_ib_alloc_pd;
1413 	dev->ib_dev.dealloc_pd		= mlx5_ib_dealloc_pd;
1414 	dev->ib_dev.create_ah		= mlx5_ib_create_ah;
1415 	dev->ib_dev.query_ah		= mlx5_ib_query_ah;
1416 	dev->ib_dev.destroy_ah		= mlx5_ib_destroy_ah;
1417 	dev->ib_dev.create_srq		= mlx5_ib_create_srq;
1418 	dev->ib_dev.modify_srq		= mlx5_ib_modify_srq;
1419 	dev->ib_dev.query_srq		= mlx5_ib_query_srq;
1420 	dev->ib_dev.destroy_srq		= mlx5_ib_destroy_srq;
1421 	dev->ib_dev.post_srq_recv	= mlx5_ib_post_srq_recv;
1422 	dev->ib_dev.create_qp		= mlx5_ib_create_qp;
1423 	dev->ib_dev.modify_qp		= mlx5_ib_modify_qp;
1424 	dev->ib_dev.query_qp		= mlx5_ib_query_qp;
1425 	dev->ib_dev.destroy_qp		= mlx5_ib_destroy_qp;
1426 	dev->ib_dev.post_send		= mlx5_ib_post_send;
1427 	dev->ib_dev.post_recv		= mlx5_ib_post_recv;
1428 	dev->ib_dev.create_cq		= mlx5_ib_create_cq;
1429 	dev->ib_dev.modify_cq		= mlx5_ib_modify_cq;
1430 	dev->ib_dev.resize_cq		= mlx5_ib_resize_cq;
1431 	dev->ib_dev.destroy_cq		= mlx5_ib_destroy_cq;
1432 	dev->ib_dev.poll_cq		= mlx5_ib_poll_cq;
1433 	dev->ib_dev.req_notify_cq	= mlx5_ib_arm_cq;
1434 	dev->ib_dev.get_dma_mr		= mlx5_ib_get_dma_mr;
1435 	dev->ib_dev.reg_user_mr		= mlx5_ib_reg_user_mr;
1436 	dev->ib_dev.dereg_mr		= mlx5_ib_dereg_mr;
1437 	dev->ib_dev.destroy_mr		= mlx5_ib_destroy_mr;
1438 	dev->ib_dev.attach_mcast	= mlx5_ib_mcg_attach;
1439 	dev->ib_dev.detach_mcast	= mlx5_ib_mcg_detach;
1440 	dev->ib_dev.process_mad		= mlx5_ib_process_mad;
1441 	dev->ib_dev.create_mr		= mlx5_ib_create_mr;
1442 	dev->ib_dev.alloc_fast_reg_mr	= mlx5_ib_alloc_fast_reg_mr;
1443 	dev->ib_dev.alloc_fast_reg_page_list = mlx5_ib_alloc_fast_reg_page_list;
1444 	dev->ib_dev.free_fast_reg_page_list  = mlx5_ib_free_fast_reg_page_list;
1445 	dev->ib_dev.check_mr_status	= mlx5_ib_check_mr_status;
1446 
1447 	if (mdev->caps.flags & MLX5_DEV_CAP_FLAG_XRC) {
1448 		dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
1449 		dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
1450 		dev->ib_dev.uverbs_cmd_mask |=
1451 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
1452 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
1453 	}
1454 
1455 	err = init_node_data(dev);
1456 	if (err)
1457 		goto err_eqs;
1458 
1459 	mutex_init(&dev->cap_mask_mutex);
1460 	spin_lock_init(&dev->mr_lock);
1461 
1462 	err = create_dev_resources(&dev->devr);
1463 	if (err)
1464 		goto err_eqs;
1465 
1466 	err = ib_register_device(&dev->ib_dev, NULL);
1467 	if (err)
1468 		goto err_rsrc;
1469 
1470 	err = create_umr_res(dev);
1471 	if (err)
1472 		goto err_dev;
1473 
1474 	for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
1475 		err = device_create_file(&dev->ib_dev.dev,
1476 					 mlx5_class_attributes[i]);
1477 		if (err)
1478 			goto err_umrc;
1479 	}
1480 
1481 	dev->ib_active = true;
1482 
1483 	return 0;
1484 
1485 err_umrc:
1486 	destroy_umrc_res(dev);
1487 
1488 err_dev:
1489 	ib_unregister_device(&dev->ib_dev);
1490 
1491 err_rsrc:
1492 	destroy_dev_resources(&dev->devr);
1493 
1494 err_eqs:
1495 	free_comp_eqs(dev);
1496 
1497 err_cleanup:
1498 	mlx5_dev_cleanup(mdev);
1499 
1500 err_free:
1501 	ib_dealloc_device((struct ib_device *)dev);
1502 
1503 	return err;
1504 }
1505 
1506 static void remove_one(struct pci_dev *pdev)
1507 {
1508 	struct mlx5_ib_dev *dev = mlx5_pci2ibdev(pdev);
1509 
1510 	destroy_umrc_res(dev);
1511 	ib_unregister_device(&dev->ib_dev);
1512 	destroy_dev_resources(&dev->devr);
1513 	free_comp_eqs(dev);
1514 	mlx5_dev_cleanup(&dev->mdev);
1515 	ib_dealloc_device(&dev->ib_dev);
1516 }
1517 
1518 static DEFINE_PCI_DEVICE_TABLE(mlx5_ib_pci_table) = {
1519 	{ PCI_VDEVICE(MELLANOX, 4113) }, /* MT4113 Connect-IB */
1520 	{ 0, }
1521 };
1522 
1523 MODULE_DEVICE_TABLE(pci, mlx5_ib_pci_table);
1524 
1525 static struct pci_driver mlx5_ib_driver = {
1526 	.name		= DRIVER_NAME,
1527 	.id_table	= mlx5_ib_pci_table,
1528 	.probe		= init_one,
1529 	.remove		= remove_one
1530 };
1531 
1532 static int __init mlx5_ib_init(void)
1533 {
1534 	return pci_register_driver(&mlx5_ib_driver);
1535 }
1536 
1537 static void __exit mlx5_ib_cleanup(void)
1538 {
1539 	pci_unregister_driver(&mlx5_ib_driver);
1540 }
1541 
1542 module_init(mlx5_ib_init);
1543 module_exit(mlx5_ib_cleanup);
1544