xref: /freebsd/contrib/ofed/libibverbs/verbs.c (revision 8aac90f18aef7c9eea906c3ff9a001ca7b94f375)
1 /*
2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2006, 2007 Cisco Systems, Inc.  All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #define _GNU_SOURCE
35 #include <config.h>
36 
37 #include <infiniband/endian.h>
38 #include <stdio.h>
39 #include <unistd.h>
40 #include <stdlib.h>
41 #include <errno.h>
42 #include <string.h>
43 #include <dirent.h>
44 #include <netinet/in.h>
45 #include <netinet/ip.h>
46 #include <sys/socket.h>
47 
48 #include "ibverbs.h"
49 #ifndef NRESOLVE_NEIGH
50 #include <net/if.h>
51 #include <net/if_arp.h>
52 #include "neigh.h"
53 #endif
54 
55 /* Hack to avoid GCC's -Wmissing-prototypes and the similar error from sparse
56    with these prototypes. Symbol versionining requires the goofy names, the
57    prototype must match the version in verbs.h.
58  */
59 int __ibv_query_device(struct ibv_context *context,
60 		       struct ibv_device_attr *device_attr);
61 int __ibv_query_port(struct ibv_context *context, uint8_t port_num,
62 		     struct ibv_port_attr *port_attr);
63 int __ibv_query_gid(struct ibv_context *context, uint8_t port_num, int index,
64 		    union ibv_gid *gid);
65 int __ibv_query_pkey(struct ibv_context *context, uint8_t port_num, int index,
66 		     __be16 *pkey);
67 struct ibv_pd *__ibv_alloc_pd(struct ibv_context *context);
68 int __ibv_dealloc_pd(struct ibv_pd *pd);
69 struct ibv_mr *__ibv_reg_mr(struct ibv_pd *pd, void *addr, size_t length,
70 			    int access);
71 int __ibv_rereg_mr(struct ibv_mr *mr, int flags, struct ibv_pd *pd, void *addr,
72 		   size_t length, int access);
73 int __ibv_dereg_mr(struct ibv_mr *mr);
74 struct ibv_cq *__ibv_create_cq(struct ibv_context *context, int cqe,
75 			       void *cq_context,
76 			       struct ibv_comp_channel *channel,
77 			       int comp_vector);
78 int __ibv_resize_cq(struct ibv_cq *cq, int cqe);
79 int __ibv_destroy_cq(struct ibv_cq *cq);
80 int __ibv_get_cq_event(struct ibv_comp_channel *channel, struct ibv_cq **cq,
81 		       void **cq_context);
82 void __ibv_ack_cq_events(struct ibv_cq *cq, unsigned int nevents);
83 struct ibv_srq *__ibv_create_srq(struct ibv_pd *pd,
84 				 struct ibv_srq_init_attr *srq_init_attr);
85 int __ibv_modify_srq(struct ibv_srq *srq, struct ibv_srq_attr *srq_attr,
86 		     int srq_attr_mask);
87 int __ibv_query_srq(struct ibv_srq *srq, struct ibv_srq_attr *srq_attr);
88 int __ibv_destroy_srq(struct ibv_srq *srq);
89 struct ibv_qp *__ibv_create_qp(struct ibv_pd *pd,
90 			       struct ibv_qp_init_attr *qp_init_attr);
91 int __ibv_query_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr, int attr_mask,
92 		   struct ibv_qp_init_attr *init_attr);
93 int __ibv_modify_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr, int attr_mask);
94 int __ibv_destroy_qp(struct ibv_qp *qp);
95 struct ibv_ah *__ibv_create_ah(struct ibv_pd *pd, struct ibv_ah_attr *attr);
96 int __ibv_destroy_ah(struct ibv_ah *ah);
97 int __ibv_attach_mcast(struct ibv_qp *qp, const union ibv_gid *gid,
98 		       uint16_t lid);
99 int __ibv_detach_mcast(struct ibv_qp *qp, const union ibv_gid *gid,
100 		       uint16_t lid);
101 
102 int __attribute__((const)) ibv_rate_to_mult(enum ibv_rate rate)
103 {
104 	switch (rate) {
105 	case IBV_RATE_2_5_GBPS: return  1;
106 	case IBV_RATE_5_GBPS:   return  2;
107 	case IBV_RATE_10_GBPS:  return  4;
108 	case IBV_RATE_20_GBPS:  return  8;
109 	case IBV_RATE_30_GBPS:  return 12;
110 	case IBV_RATE_40_GBPS:  return 16;
111 	case IBV_RATE_60_GBPS:  return 24;
112 	case IBV_RATE_80_GBPS:  return 32;
113 	case IBV_RATE_120_GBPS: return 48;
114 	case IBV_RATE_28_GBPS:  return 11;
115 	case IBV_RATE_50_GBPS:  return 20;
116 	case IBV_RATE_400_GBPS: return 160;
117 	case IBV_RATE_600_GBPS: return 240;
118 	default:           return -1;
119 	}
120 }
121 
122 enum ibv_rate __attribute__((const)) mult_to_ibv_rate(int mult)
123 {
124 	switch (mult) {
125 	case 1:  return IBV_RATE_2_5_GBPS;
126 	case 2:  return IBV_RATE_5_GBPS;
127 	case 4:  return IBV_RATE_10_GBPS;
128 	case 8:  return IBV_RATE_20_GBPS;
129 	case 12: return IBV_RATE_30_GBPS;
130 	case 16: return IBV_RATE_40_GBPS;
131 	case 24: return IBV_RATE_60_GBPS;
132 	case 32: return IBV_RATE_80_GBPS;
133 	case 48: return IBV_RATE_120_GBPS;
134 	case 11: return IBV_RATE_28_GBPS;
135 	case 20: return IBV_RATE_50_GBPS;
136 	case 160: return IBV_RATE_400_GBPS;
137 	case 240: return IBV_RATE_600_GBPS;
138 	default: return IBV_RATE_MAX;
139 	}
140 }
141 
142 int  __attribute__((const)) ibv_rate_to_mbps(enum ibv_rate rate)
143 {
144 	switch (rate) {
145 	case IBV_RATE_2_5_GBPS: return 2500;
146 	case IBV_RATE_5_GBPS:   return 5000;
147 	case IBV_RATE_10_GBPS:  return 10000;
148 	case IBV_RATE_20_GBPS:  return 20000;
149 	case IBV_RATE_30_GBPS:  return 30000;
150 	case IBV_RATE_40_GBPS:  return 40000;
151 	case IBV_RATE_60_GBPS:  return 60000;
152 	case IBV_RATE_80_GBPS:  return 80000;
153 	case IBV_RATE_120_GBPS: return 120000;
154 	case IBV_RATE_14_GBPS:  return 14062;
155 	case IBV_RATE_56_GBPS:  return 56250;
156 	case IBV_RATE_112_GBPS: return 112500;
157 	case IBV_RATE_168_GBPS: return 168750;
158 	case IBV_RATE_25_GBPS:  return 25781;
159 	case IBV_RATE_100_GBPS: return 103125;
160 	case IBV_RATE_200_GBPS: return 206250;
161 	case IBV_RATE_300_GBPS: return 309375;
162 	case IBV_RATE_28_GBPS:  return 28125;
163 	case IBV_RATE_50_GBPS:  return 53125;
164 	case IBV_RATE_400_GBPS: return 425000;
165 	case IBV_RATE_600_GBPS: return 637500;
166 	default:               return -1;
167 	}
168 }
169 
170 enum ibv_rate __attribute__((const)) mbps_to_ibv_rate(int mbps)
171 {
172 	switch (mbps) {
173 	case 2500:   return IBV_RATE_2_5_GBPS;
174 	case 5000:   return IBV_RATE_5_GBPS;
175 	case 10000:  return IBV_RATE_10_GBPS;
176 	case 20000:  return IBV_RATE_20_GBPS;
177 	case 30000:  return IBV_RATE_30_GBPS;
178 	case 40000:  return IBV_RATE_40_GBPS;
179 	case 60000:  return IBV_RATE_60_GBPS;
180 	case 80000:  return IBV_RATE_80_GBPS;
181 	case 120000: return IBV_RATE_120_GBPS;
182 	case 14062:  return IBV_RATE_14_GBPS;
183 	case 56250:  return IBV_RATE_56_GBPS;
184 	case 112500: return IBV_RATE_112_GBPS;
185 	case 168750: return IBV_RATE_168_GBPS;
186 	case 25781:  return IBV_RATE_25_GBPS;
187 	case 103125: return IBV_RATE_100_GBPS;
188 	case 206250: return IBV_RATE_200_GBPS;
189 	case 309375: return IBV_RATE_300_GBPS;
190 	case 28125:  return IBV_RATE_28_GBPS;
191 	case 53125:  return IBV_RATE_50_GBPS;
192 	case 425000: return IBV_RATE_400_GBPS;
193 	case 637500: return IBV_RATE_600_GBPS;
194 	default:     return IBV_RATE_MAX;
195 	}
196 }
197 
198 int __ibv_query_device(struct ibv_context *context,
199 		       struct ibv_device_attr *device_attr)
200 {
201 	return context->ops.query_device(context, device_attr);
202 }
203 default_symver(__ibv_query_device, ibv_query_device);
204 
205 int __ibv_query_port(struct ibv_context *context, uint8_t port_num,
206 		     struct ibv_port_attr *port_attr)
207 {
208 	return context->ops.query_port(context, port_num, port_attr);
209 }
210 default_symver(__ibv_query_port, ibv_query_port);
211 
212 int __ibv_query_gid(struct ibv_context *context, uint8_t port_num,
213 		    int index, union ibv_gid *gid)
214 {
215 	char name[24];
216 	char attr[41];
217 	uint16_t val;
218 	int i;
219 
220 	snprintf(name, sizeof name, "ports/%d/gids/%d", port_num, index);
221 
222 	if (ibv_read_sysfs_file(context->device->ibdev_path, name,
223 				attr, sizeof attr) < 0)
224 		return -1;
225 
226 	for (i = 0; i < 8; ++i) {
227 		if (sscanf(attr + i * 5, "%hx", &val) != 1)
228 			return -1;
229 		gid->raw[i * 2    ] = val >> 8;
230 		gid->raw[i * 2 + 1] = val & 0xff;
231 	}
232 
233 	return 0;
234 }
235 default_symver(__ibv_query_gid, ibv_query_gid);
236 
237 int __ibv_query_pkey(struct ibv_context *context, uint8_t port_num,
238 		     int index, __be16 *pkey)
239 {
240 	char name[24];
241 	char attr[8];
242 	uint16_t val;
243 
244 	snprintf(name, sizeof name, "ports/%d/pkeys/%d", port_num, index);
245 
246 	if (ibv_read_sysfs_file(context->device->ibdev_path, name,
247 				attr, sizeof attr) < 0)
248 		return -1;
249 
250 	if (sscanf(attr, "%hx", &val) != 1)
251 		return -1;
252 
253 	*pkey = htobe16(val);
254 	return 0;
255 }
256 default_symver(__ibv_query_pkey, ibv_query_pkey);
257 
258 struct ibv_pd *__ibv_alloc_pd(struct ibv_context *context)
259 {
260 	struct ibv_pd *pd;
261 
262 	pd = context->ops.alloc_pd(context);
263 	if (pd)
264 		pd->context = context;
265 
266 	return pd;
267 }
268 default_symver(__ibv_alloc_pd, ibv_alloc_pd);
269 
270 int __ibv_dealloc_pd(struct ibv_pd *pd)
271 {
272 	return pd->context->ops.dealloc_pd(pd);
273 }
274 default_symver(__ibv_dealloc_pd, ibv_dealloc_pd);
275 
276 struct ibv_mr *__ibv_reg_mr(struct ibv_pd *pd, void *addr,
277 			    size_t length, int access)
278 {
279 	struct ibv_mr *mr;
280 
281 	if (ibv_dontfork_range(addr, length))
282 		return NULL;
283 
284 	mr = pd->context->ops.reg_mr(pd, addr, length, access);
285 	if (mr) {
286 		mr->context = pd->context;
287 		mr->pd      = pd;
288 		mr->addr    = addr;
289 		mr->length  = length;
290 	} else
291 		ibv_dofork_range(addr, length);
292 
293 	return mr;
294 }
295 default_symver(__ibv_reg_mr, ibv_reg_mr);
296 
297 int __ibv_rereg_mr(struct ibv_mr *mr, int flags,
298 		   struct ibv_pd *pd, void *addr,
299 		   size_t length, int access)
300 {
301 	int dofork_onfail = 0;
302 	int err;
303 	void *old_addr;
304 	size_t old_len;
305 
306 	if (flags & ~IBV_REREG_MR_FLAGS_SUPPORTED) {
307 		errno = EINVAL;
308 		return IBV_REREG_MR_ERR_INPUT;
309 	}
310 
311 	if ((flags & IBV_REREG_MR_CHANGE_TRANSLATION) &&
312 	    (!length || !addr)) {
313 		errno = EINVAL;
314 		return IBV_REREG_MR_ERR_INPUT;
315 	}
316 
317 	if (access && !(flags & IBV_REREG_MR_CHANGE_ACCESS)) {
318 		errno = EINVAL;
319 		return IBV_REREG_MR_ERR_INPUT;
320 	}
321 
322 	if (!mr->context->ops.rereg_mr) {
323 		errno = ENOSYS;
324 		return IBV_REREG_MR_ERR_INPUT;
325 	}
326 
327 	if (flags & IBV_REREG_MR_CHANGE_TRANSLATION) {
328 		err = ibv_dontfork_range(addr, length);
329 		if (err)
330 			return IBV_REREG_MR_ERR_DONT_FORK_NEW;
331 		dofork_onfail = 1;
332 	}
333 
334 	old_addr = mr->addr;
335 	old_len = mr->length;
336 	err = mr->context->ops.rereg_mr(mr, flags, pd, addr, length, access);
337 	if (!err) {
338 		if (flags & IBV_REREG_MR_CHANGE_PD)
339 			mr->pd = pd;
340 		if (flags & IBV_REREG_MR_CHANGE_TRANSLATION) {
341 			mr->addr    = addr;
342 			mr->length  = length;
343 			err = ibv_dofork_range(old_addr, old_len);
344 			if (err)
345 				return IBV_REREG_MR_ERR_DO_FORK_OLD;
346 		}
347 	} else {
348 		err = IBV_REREG_MR_ERR_CMD;
349 		if (dofork_onfail) {
350 			if (ibv_dofork_range(addr, length))
351 				err = IBV_REREG_MR_ERR_CMD_AND_DO_FORK_NEW;
352 		}
353 	}
354 
355 	return err;
356 }
357 default_symver(__ibv_rereg_mr, ibv_rereg_mr);
358 
359 int __ibv_dereg_mr(struct ibv_mr *mr)
360 {
361 	int ret;
362 	void *addr	= mr->addr;
363 	size_t length	= mr->length;
364 
365 	ret = mr->context->ops.dereg_mr(mr);
366 	if (!ret)
367 		ibv_dofork_range(addr, length);
368 
369 	return ret;
370 }
371 default_symver(__ibv_dereg_mr, ibv_dereg_mr);
372 
373 static struct ibv_comp_channel *ibv_create_comp_channel_v2(struct ibv_context *context)
374 {
375 	struct ibv_abi_compat_v2 *t = context->abi_compat;
376 	static int warned;
377 
378 	if (!pthread_mutex_trylock(&t->in_use))
379 		return &t->channel;
380 
381 	if (!warned) {
382 		fprintf(stderr, PFX "Warning: kernel's ABI version %d limits capacity.\n"
383 			"    Only one completion channel can be created per context.\n",
384 			abi_ver);
385 		++warned;
386 	}
387 
388 	return NULL;
389 }
390 
391 struct ibv_comp_channel *ibv_create_comp_channel(struct ibv_context *context)
392 {
393 	struct ibv_comp_channel            *channel;
394 	struct ibv_create_comp_channel      cmd;
395 	struct ibv_create_comp_channel_resp resp;
396 
397 	if (abi_ver <= 2)
398 		return ibv_create_comp_channel_v2(context);
399 
400 	channel = malloc(sizeof *channel);
401 	if (!channel)
402 		return NULL;
403 
404 	IBV_INIT_CMD_RESP(&cmd, sizeof cmd, CREATE_COMP_CHANNEL, &resp, sizeof resp);
405 	if (write(context->cmd_fd, &cmd, sizeof cmd) != sizeof cmd) {
406 		free(channel);
407 		return NULL;
408 	}
409 
410 	(void) VALGRIND_MAKE_MEM_DEFINED(&resp, sizeof resp);
411 
412 	channel->context = context;
413 	channel->fd      = resp.fd;
414 	channel->refcnt  = 0;
415 
416 	return channel;
417 }
418 
419 static int ibv_destroy_comp_channel_v2(struct ibv_comp_channel *channel)
420 {
421 	struct ibv_abi_compat_v2 *t = (struct ibv_abi_compat_v2 *) channel;
422 	pthread_mutex_unlock(&t->in_use);
423 	return 0;
424 }
425 
426 int ibv_destroy_comp_channel(struct ibv_comp_channel *channel)
427 {
428 	struct ibv_context *context;
429 	int ret;
430 
431 	context = channel->context;
432 	pthread_mutex_lock(&context->mutex);
433 
434 	if (channel->refcnt) {
435 		ret = EBUSY;
436 		goto out;
437 	}
438 
439 	if (abi_ver <= 2) {
440 		ret = ibv_destroy_comp_channel_v2(channel);
441 		goto out;
442 	}
443 
444 	close(channel->fd);
445 	free(channel);
446 	ret = 0;
447 
448 out:
449 	pthread_mutex_unlock(&context->mutex);
450 
451 	return ret;
452 }
453 
454 struct ibv_cq *__ibv_create_cq(struct ibv_context *context, int cqe, void *cq_context,
455 			       struct ibv_comp_channel *channel, int comp_vector)
456 {
457 	struct ibv_cq *cq;
458 	int err = 0;
459 
460 	cq = context->ops.create_cq(context, cqe, channel, comp_vector);
461 
462 	if (!cq)
463 		return NULL;
464 
465 	err = verbs_init_cq(cq, context, channel, cq_context);
466 	if (err)
467 		goto err;
468 
469 	return cq;
470 
471 err:
472 	context->ops.destroy_cq(cq);
473 
474 	return NULL;
475 }
476 default_symver(__ibv_create_cq, ibv_create_cq);
477 
478 int __ibv_resize_cq(struct ibv_cq *cq, int cqe)
479 {
480 	if (!cq->context->ops.resize_cq)
481 		return ENOSYS;
482 
483 	return cq->context->ops.resize_cq(cq, cqe);
484 }
485 default_symver(__ibv_resize_cq, ibv_resize_cq);
486 
487 int __ibv_destroy_cq(struct ibv_cq *cq)
488 {
489 	struct ibv_comp_channel *channel = cq->channel;
490 	int ret;
491 
492 	ret = cq->context->ops.destroy_cq(cq);
493 
494 	if (channel) {
495 		if (!ret) {
496 			pthread_mutex_lock(&channel->context->mutex);
497 			--channel->refcnt;
498 			pthread_mutex_unlock(&channel->context->mutex);
499 		}
500 	}
501 
502 	return ret;
503 }
504 default_symver(__ibv_destroy_cq, ibv_destroy_cq);
505 
506 int __ibv_get_cq_event(struct ibv_comp_channel *channel,
507 		       struct ibv_cq **cq, void **cq_context)
508 {
509 	struct ibv_comp_event ev;
510 
511 	if (read(channel->fd, &ev, sizeof ev) != sizeof ev)
512 		return -1;
513 
514 	*cq         = (struct ibv_cq *) (uintptr_t) ev.cq_handle;
515 	*cq_context = (*cq)->cq_context;
516 
517 	if ((*cq)->context->ops.cq_event)
518 		(*cq)->context->ops.cq_event(*cq);
519 
520 	return 0;
521 }
522 default_symver(__ibv_get_cq_event, ibv_get_cq_event);
523 
524 void __ibv_ack_cq_events(struct ibv_cq *cq, unsigned int nevents)
525 {
526 	pthread_mutex_lock(&cq->mutex);
527 	cq->comp_events_completed += nevents;
528 	pthread_cond_signal(&cq->cond);
529 	pthread_mutex_unlock(&cq->mutex);
530 }
531 default_symver(__ibv_ack_cq_events, ibv_ack_cq_events);
532 
533 struct ibv_srq *__ibv_create_srq(struct ibv_pd *pd,
534 				 struct ibv_srq_init_attr *srq_init_attr)
535 {
536 	struct ibv_srq *srq;
537 
538 	if (!pd->context->ops.create_srq)
539 		return NULL;
540 
541 	srq = pd->context->ops.create_srq(pd, srq_init_attr);
542 	if (!srq)
543 		return NULL;
544 
545 	srq->context		  = pd->context;
546 	srq->srq_context	  = srq_init_attr->srq_context;
547 	srq->pd				  = pd;
548 	srq->events_completed = 0;
549 	if (pthread_mutex_init(&srq->mutex, NULL))
550 		goto err;
551 	if (pthread_cond_init(&srq->cond, NULL))
552 		goto err_mutex;
553 
554 	return srq;
555 
556 err_mutex:
557 	pthread_mutex_destroy(&srq->mutex);
558 err:
559 	pd->context->ops.destroy_srq(srq);
560 
561 	return NULL;
562 }
563 default_symver(__ibv_create_srq, ibv_create_srq);
564 
565 int __ibv_modify_srq(struct ibv_srq *srq,
566 		     struct ibv_srq_attr *srq_attr,
567 		     int srq_attr_mask)
568 {
569 	return srq->context->ops.modify_srq(srq, srq_attr, srq_attr_mask);
570 }
571 default_symver(__ibv_modify_srq, ibv_modify_srq);
572 
573 int __ibv_query_srq(struct ibv_srq *srq, struct ibv_srq_attr *srq_attr)
574 {
575 	return srq->context->ops.query_srq(srq, srq_attr);
576 }
577 default_symver(__ibv_query_srq, ibv_query_srq);
578 
579 int __ibv_destroy_srq(struct ibv_srq *srq)
580 {
581 	pthread_cond_destroy(&srq->cond);
582 	pthread_mutex_destroy(&srq->mutex);
583 	return srq->context->ops.destroy_srq(srq);
584 }
585 default_symver(__ibv_destroy_srq, ibv_destroy_srq);
586 
587 struct ibv_qp *__ibv_create_qp(struct ibv_pd *pd,
588 			       struct ibv_qp_init_attr *qp_init_attr)
589 {
590 	struct ibv_qp *qp = pd->context->ops.create_qp(pd, qp_init_attr);
591 
592 	if (qp) {
593 		qp->context    	     = pd->context;
594 		qp->qp_context 	     = qp_init_attr->qp_context;
595 		qp->pd         	     = pd;
596 		qp->send_cq    	     = qp_init_attr->send_cq;
597 		qp->recv_cq    	     = qp_init_attr->recv_cq;
598 		qp->srq        	     = qp_init_attr->srq;
599 		qp->qp_type          = qp_init_attr->qp_type;
600 		qp->state	     = IBV_QPS_RESET;
601 		qp->events_completed = 0;
602 		pthread_mutex_init(&qp->mutex, NULL);
603 		pthread_cond_init(&qp->cond, NULL);
604 	}
605 
606 	return qp;
607 }
608 default_symver(__ibv_create_qp, ibv_create_qp);
609 
610 int __ibv_query_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
611 		   int attr_mask,
612 		   struct ibv_qp_init_attr *init_attr)
613 {
614 	int ret;
615 
616 	ret = qp->context->ops.query_qp(qp, attr, attr_mask, init_attr);
617 	if (ret)
618 		return ret;
619 
620 	if (attr_mask & IBV_QP_STATE)
621 		qp->state = attr->qp_state;
622 
623 	return 0;
624 }
625 default_symver(__ibv_query_qp, ibv_query_qp);
626 
627 int __ibv_modify_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
628 		    int attr_mask)
629 {
630 	int ret;
631 
632 	ret = qp->context->ops.modify_qp(qp, attr, attr_mask);
633 	if (ret)
634 		return ret;
635 
636 	if (attr_mask & IBV_QP_STATE)
637 		qp->state = attr->qp_state;
638 
639 	return 0;
640 }
641 default_symver(__ibv_modify_qp, ibv_modify_qp);
642 
643 int __ibv_destroy_qp(struct ibv_qp *qp)
644 {
645 	return qp->context->ops.destroy_qp(qp);
646 }
647 default_symver(__ibv_destroy_qp, ibv_destroy_qp);
648 
649 struct ibv_ah *__ibv_create_ah(struct ibv_pd *pd, struct ibv_ah_attr *attr)
650 {
651 	struct ibv_ah *ah = pd->context->ops.create_ah(pd, attr);
652 
653 	if (ah) {
654 		ah->context = pd->context;
655 		ah->pd      = pd;
656 	}
657 
658 	return ah;
659 }
660 default_symver(__ibv_create_ah, ibv_create_ah);
661 
662 /* GID types as appear in sysfs, no change is expected as of ABI
663  * compatibility.
664  */
665 #define V1_TYPE "IB/RoCE v1"
666 #define V2_TYPE "RoCE v2"
667 int ibv_query_gid_type(struct ibv_context *context, uint8_t port_num,
668 		       unsigned int index, enum ibv_gid_type *type)
669 {
670 	char name[32];
671 	char buff[11];
672 
673 	snprintf(name, sizeof(name), "ports/%d/gid_attrs/types/%d", port_num,
674 		 index);
675 
676 	/* Reset errno so that we can rely on its value upon any error flow in
677 	 * ibv_read_sysfs_file.
678 	 */
679 	errno = 0;
680 	if (ibv_read_sysfs_file(context->device->ibdev_path, name, buff,
681 				sizeof(buff)) <= 0) {
682 		char *dir_path;
683 		DIR *dir;
684 
685 		if (errno == EINVAL) {
686 			/* In IB, this file doesn't exist and the kernel sets
687 			 * errno to -EINVAL.
688 			 */
689 			*type = IBV_GID_TYPE_IB_ROCE_V1;
690 			return 0;
691 		}
692 		if (asprintf(&dir_path, "%s/%s/%d/%s/",
693 			     context->device->ibdev_path, "ports", port_num,
694 			     "gid_attrs") < 0)
695 			return -1;
696 		dir = opendir(dir_path);
697 		free(dir_path);
698 		if (!dir) {
699 			if (errno == ENOENT)
700 				/* Assuming that if gid_attrs doesn't exist,
701 				 * we have an old kernel and all GIDs are
702 				 * IB/RoCE v1
703 				 */
704 				*type = IBV_GID_TYPE_IB_ROCE_V1;
705 			else
706 				return -1;
707 		} else {
708 			closedir(dir);
709 			errno = EFAULT;
710 			return -1;
711 		}
712 	} else {
713 		if (!strcmp(buff, V1_TYPE)) {
714 			*type = IBV_GID_TYPE_IB_ROCE_V1;
715 		} else if (!strcmp(buff, V2_TYPE)) {
716 			*type = IBV_GID_TYPE_ROCE_V2;
717 		} else {
718 			errno = ENOTSUP;
719 			return -1;
720 		}
721 	}
722 
723 	return 0;
724 }
725 
726 static int ibv_find_gid_index(struct ibv_context *context, uint8_t port_num,
727 			      union ibv_gid *gid, enum ibv_gid_type gid_type)
728 {
729 	enum ibv_gid_type sgid_type = 0;
730 	union ibv_gid sgid;
731 	int i = 0, ret;
732 
733 	do {
734 		ret = ibv_query_gid(context, port_num, i, &sgid);
735 		if (!ret) {
736 			ret = ibv_query_gid_type(context, port_num, i,
737 						 &sgid_type);
738 		}
739 		i++;
740 	} while (!ret && (memcmp(&sgid, gid, sizeof(*gid)) ||
741 		 (gid_type != sgid_type)));
742 
743 	return ret ? ret : i - 1;
744 }
745 
746 static inline void map_ipv4_addr_to_ipv6(__be32 ipv4, struct in6_addr *ipv6)
747 {
748 	ipv6->s6_addr32[0] = 0;
749 	ipv6->s6_addr32[1] = 0;
750 	ipv6->s6_addr32[2] = htobe32(0x0000FFFF);
751 	ipv6->s6_addr32[3] = ipv4;
752 }
753 
754 static inline __sum16 ipv4_calc_hdr_csum(uint16_t *data, unsigned int num_hwords)
755 {
756 	unsigned int i = 0;
757 	uint32_t sum = 0;
758 
759 	for (i = 0; i < num_hwords; i++)
760 		sum += *(data++);
761 
762 	sum = (sum & 0xffff) + (sum >> 16);
763 
764 	return (__sum16)~sum;
765 }
766 
767 static inline int get_grh_header_version(struct ibv_grh *grh)
768 {
769 	int ip6h_version = (be32toh(grh->version_tclass_flow) >> 28) & 0xf;
770 	struct ip *ip4h = (struct ip *)((void *)grh + 20);
771 	struct ip ip4h_checked;
772 
773 	if (ip6h_version != 6) {
774 		if (ip4h->ip_v == 4)
775 			return 4;
776 		errno = EPROTONOSUPPORT;
777 		return -1;
778 	}
779 	/* version may be 6 or 4 */
780 	if (ip4h->ip_hl != 5) /* IPv4 header length must be 5 for RoCE v2. */
781 		return 6;
782 	/*
783 	* Verify checksum.
784 	* We can't write on scattered buffers so we have to copy to temp
785 	* buffer.
786 	*/
787 	memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
788 	/* Need to set the checksum field (check) to 0 before re-calculating
789 	 * the checksum.
790 	 */
791 	ip4h_checked.ip_sum = 0;
792 	ip4h_checked.ip_sum = ipv4_calc_hdr_csum((uint16_t *)&ip4h_checked, 10);
793 	/* if IPv4 header checksum is OK, believe it */
794 	if (ip4h->ip_sum == ip4h_checked.ip_sum)
795 		return 4;
796 	return 6;
797 }
798 
799 static inline void set_ah_attr_generic_fields(struct ibv_ah_attr *ah_attr,
800 					      struct ibv_wc *wc,
801 					      struct ibv_grh *grh,
802 					      uint8_t port_num)
803 {
804 	uint32_t flow_class;
805 
806 	flow_class = be32toh(grh->version_tclass_flow);
807 	ah_attr->grh.flow_label = flow_class & 0xFFFFF;
808 	ah_attr->dlid = wc->slid;
809 	ah_attr->sl = wc->sl;
810 	ah_attr->src_path_bits = wc->dlid_path_bits;
811 	ah_attr->port_num = port_num;
812 }
813 
814 static inline int set_ah_attr_by_ipv4(struct ibv_context *context,
815 				      struct ibv_ah_attr *ah_attr,
816 				      struct ip *ip4h, uint8_t port_num)
817 {
818 	union ibv_gid sgid;
819 	int ret;
820 
821 	/* No point searching multicast GIDs in GID table */
822 	if (IN_CLASSD(be32toh(ip4h->ip_dst.s_addr))) {
823 		errno = EINVAL;
824 		return -1;
825 	}
826 
827 	map_ipv4_addr_to_ipv6(ip4h->ip_dst.s_addr, (struct in6_addr *)&sgid);
828 	ret = ibv_find_gid_index(context, port_num, &sgid,
829 				 IBV_GID_TYPE_ROCE_V2);
830 	if (ret < 0)
831 		return ret;
832 
833 	map_ipv4_addr_to_ipv6(ip4h->ip_src.s_addr,
834 			      (struct in6_addr *)&ah_attr->grh.dgid);
835 	ah_attr->grh.sgid_index = (uint8_t) ret;
836 	ah_attr->grh.hop_limit = ip4h->ip_ttl;
837 	ah_attr->grh.traffic_class = ip4h->ip_tos;
838 
839 	return 0;
840 }
841 
842 #define IB_NEXT_HDR    0x1b
843 static inline int set_ah_attr_by_ipv6(struct ibv_context *context,
844 				  struct ibv_ah_attr *ah_attr,
845 				  struct ibv_grh *grh, uint8_t port_num)
846 {
847 	uint32_t flow_class;
848 	uint32_t sgid_type;
849 	int ret;
850 
851 	/* No point searching multicast GIDs in GID table */
852 	if (grh->dgid.raw[0] == 0xFF) {
853 		errno = EINVAL;
854 		return -1;
855 	}
856 
857 	ah_attr->grh.dgid = grh->sgid;
858 	if (grh->next_hdr == IPPROTO_UDP) {
859 		sgid_type = IBV_GID_TYPE_ROCE_V2;
860 	} else if (grh->next_hdr == IB_NEXT_HDR) {
861 		sgid_type = IBV_GID_TYPE_IB_ROCE_V1;
862 	} else {
863 		errno = EPROTONOSUPPORT;
864 		return -1;
865 	}
866 
867 	ret = ibv_find_gid_index(context, port_num, &grh->dgid,
868 				 sgid_type);
869 	if (ret < 0)
870 		return ret;
871 
872 	ah_attr->grh.sgid_index = (uint8_t) ret;
873 	flow_class = be32toh(grh->version_tclass_flow);
874 	ah_attr->grh.hop_limit = grh->hop_limit;
875 	ah_attr->grh.traffic_class = (flow_class >> 20) & 0xFF;
876 
877 	return 0;
878 }
879 
880 int ibv_init_ah_from_wc(struct ibv_context *context, uint8_t port_num,
881 			struct ibv_wc *wc, struct ibv_grh *grh,
882 			struct ibv_ah_attr *ah_attr)
883 {
884 	int version;
885 	int ret = 0;
886 
887 	memset(ah_attr, 0, sizeof *ah_attr);
888 	set_ah_attr_generic_fields(ah_attr, wc, grh, port_num);
889 
890 	if (wc->wc_flags & IBV_WC_GRH) {
891 		ah_attr->is_global = 1;
892 		version = get_grh_header_version(grh);
893 
894 		if (version == 4)
895 			ret = set_ah_attr_by_ipv4(context, ah_attr,
896 						  (struct ip *)((void *)grh + 20),
897 						  port_num);
898 		else if (version == 6)
899 			ret = set_ah_attr_by_ipv6(context, ah_attr, grh,
900 						  port_num);
901 		else
902 			ret = -1;
903 	}
904 
905 	return ret;
906 }
907 
908 struct ibv_ah *ibv_create_ah_from_wc(struct ibv_pd *pd, struct ibv_wc *wc,
909 				     struct ibv_grh *grh, uint8_t port_num)
910 {
911 	struct ibv_ah_attr ah_attr;
912 	int ret;
913 
914 	ret = ibv_init_ah_from_wc(pd->context, port_num, wc, grh, &ah_attr);
915 	if (ret)
916 		return NULL;
917 
918 	return ibv_create_ah(pd, &ah_attr);
919 }
920 
921 int __ibv_destroy_ah(struct ibv_ah *ah)
922 {
923 	return ah->context->ops.destroy_ah(ah);
924 }
925 default_symver(__ibv_destroy_ah, ibv_destroy_ah);
926 
927 int __ibv_attach_mcast(struct ibv_qp *qp, const union ibv_gid *gid, uint16_t lid)
928 {
929 	return qp->context->ops.attach_mcast(qp, gid, lid);
930 }
931 default_symver(__ibv_attach_mcast, ibv_attach_mcast);
932 
933 int __ibv_detach_mcast(struct ibv_qp *qp, const union ibv_gid *gid, uint16_t lid)
934 {
935 	return qp->context->ops.detach_mcast(qp, gid, lid);
936 }
937 default_symver(__ibv_detach_mcast, ibv_detach_mcast);
938 
939 static inline int ipv6_addr_v4mapped(const struct in6_addr *a)
940 {
941 	return IN6_IS_ADDR_V4MAPPED(a) ||
942 		/* IPv4 encoded multicast addresses */
943 		(a->s6_addr32[0]  == htobe32(0xff0e0000) &&
944 		((a->s6_addr32[1] |
945 		 (a->s6_addr32[2] ^ htobe32(0x0000ffff))) == 0UL));
946 }
947 
948 struct peer_address {
949 	void *address;
950 	uint32_t size;
951 };
952 
953 static inline int create_peer_from_gid(int family, void *raw_gid,
954 				       struct peer_address *peer_address)
955 {
956 	switch (family) {
957 	case AF_INET:
958 		peer_address->address = raw_gid + 12;
959 		peer_address->size = 4;
960 		break;
961 	case AF_INET6:
962 		peer_address->address = raw_gid;
963 		peer_address->size = 16;
964 		break;
965 	default:
966 		return -1;
967 	}
968 
969 	return 0;
970 }
971 
972 #define NEIGH_GET_DEFAULT_TIMEOUT_MS 3000
973 int ibv_resolve_eth_l2_from_gid(struct ibv_context *context,
974 				struct ibv_ah_attr *attr,
975 				uint8_t eth_mac[ETHERNET_LL_SIZE],
976 				uint16_t *vid)
977 {
978 #ifndef NRESOLVE_NEIGH
979 	int dst_family;
980 	int src_family;
981 	int oif;
982 	struct get_neigh_handler neigh_handler;
983 	union ibv_gid sgid;
984 	int ether_len;
985 	struct peer_address src;
986 	struct peer_address dst;
987 	uint16_t ret_vid;
988 	int ret = -EINVAL;
989 	int err;
990 
991 	err = ibv_query_gid(context, attr->port_num,
992 			    attr->grh.sgid_index, &sgid);
993 
994 	if (err)
995 		return err;
996 
997 	err = neigh_init_resources(&neigh_handler,
998 				   NEIGH_GET_DEFAULT_TIMEOUT_MS);
999 
1000 	if (err)
1001 		return err;
1002 
1003 	dst_family = ipv6_addr_v4mapped((struct in6_addr *)attr->grh.dgid.raw) ?
1004 			AF_INET : AF_INET6;
1005 	src_family = ipv6_addr_v4mapped((struct in6_addr *)sgid.raw) ?
1006 			AF_INET : AF_INET6;
1007 
1008 	if (create_peer_from_gid(dst_family, attr->grh.dgid.raw, &dst))
1009 		goto free_resources;
1010 
1011 	if (create_peer_from_gid(src_family, &sgid.raw, &src))
1012 		goto free_resources;
1013 
1014 	if (neigh_set_dst(&neigh_handler, dst_family, dst.address,
1015 			  dst.size))
1016 		goto free_resources;
1017 
1018 	if (neigh_set_src(&neigh_handler, src_family, src.address,
1019 			  src.size))
1020 		goto free_resources;
1021 
1022 	oif = neigh_get_oif_from_src(&neigh_handler);
1023 
1024 	if (oif > 0)
1025 		neigh_set_oif(&neigh_handler, oif);
1026 	else
1027 		goto free_resources;
1028 
1029 	ret = -EHOSTUNREACH;
1030 
1031 	/* blocking call */
1032 	if (process_get_neigh(&neigh_handler))
1033 		goto free_resources;
1034 
1035 	ret_vid = neigh_get_vlan_id_from_dev(&neigh_handler);
1036 
1037 	if (ret_vid <= 0xfff)
1038 		neigh_set_vlan_id(&neigh_handler, ret_vid);
1039 
1040 	/* We are using only Ethernet here */
1041 	ether_len = neigh_get_ll(&neigh_handler,
1042 				 eth_mac,
1043 				 sizeof(uint8_t) * ETHERNET_LL_SIZE);
1044 
1045 	if (ether_len <= 0)
1046 		goto free_resources;
1047 
1048 	*vid = ret_vid;
1049 
1050 	ret = 0;
1051 
1052 free_resources:
1053 	neigh_free_resources(&neigh_handler);
1054 
1055 	return ret;
1056 #else
1057 	return -ENOSYS;
1058 #endif
1059 }
1060