?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
dm writecache: add cond_resched to loop in persistent_memory_claim()
[linux/fpc-iii.git] / drivers / infiniband / core / verbs.c
blob56a71337112c59c97fee9cd1f9d6ec5095cf77ec
1 /*
2 * Copyright (c) 2004 Mellanox Technologies Ltd. All rights reserved.
3 * Copyright (c) 2004 Infinicon Corporation. All rights reserved.
4 * Copyright (c) 2004 Intel Corporation. All rights reserved.
5 * Copyright (c) 2004 Topspin Corporation. All rights reserved.
6 * Copyright (c) 2004 Voltaire Corporation. All rights reserved.
7 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
8 * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved.
9 *
10 * This software is available to you under a choice of one of two
11 * licenses. You may choose to be licensed under the terms of the GNU
12 * General Public License (GPL) Version 2, available from the file
13 * COPYING in the main directory of this source tree, or the
14 * OpenIB.org BSD license below:
15 *
16 * Redistribution and use in source and binary forms, with or
17 * without modification, are permitted provided that the following
18 * conditions are met:
19 *
20 * - Redistributions of source code must retain the above
21 * copyright notice, this list of conditions and the following
22 * disclaimer.
23 *
24 * - Redistributions in binary form must reproduce the above
25 * copyright notice, this list of conditions and the following
26 * disclaimer in the documentation and/or other materials
27 * provided with the distribution.
28 *
29 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
30 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
31 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
32 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
33 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
34 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
35 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 * SOFTWARE.
37 */
38
39 #include <linux/errno.h>
40 #include <linux/err.h>
41 #include <linux/export.h>
42 #include <linux/string.h>
43 #include <linux/slab.h>
44 #include <linux/in.h>
45 #include <linux/in6.h>
46 #include <net/addrconf.h>
47 #include <linux/security.h>
48
49 #include <rdma/ib_verbs.h>
50 #include <rdma/ib_cache.h>
51 #include <rdma/ib_addr.h>
52 #include <rdma/rw.h>
53
54 #include "core_priv.h"
55 #include <trace/events/rdma_core.h>
56
57 static int ib_resolve_eth_dmac(struct ib_device *device,
58 struct rdma_ah_attr *ah_attr);
59
60 static const char * const ib_events[] = {
61 [IB_EVENT_CQ_ERR] = "CQ error",
62 [IB_EVENT_QP_FATAL] = "QP fatal error",
63 [IB_EVENT_QP_REQ_ERR] = "QP request error",
64 [IB_EVENT_QP_ACCESS_ERR] = "QP access error",
65 [IB_EVENT_COMM_EST] = "communication established",
66 [IB_EVENT_SQ_DRAINED] = "send queue drained",
67 [IB_EVENT_PATH_MIG] = "path migration successful",
68 [IB_EVENT_PATH_MIG_ERR] = "path migration error",
69 [IB_EVENT_DEVICE_FATAL] = "device fatal error",
70 [IB_EVENT_PORT_ACTIVE] = "port active",
71 [IB_EVENT_PORT_ERR] = "port error",
72 [IB_EVENT_LID_CHANGE] = "LID change",
73 [IB_EVENT_PKEY_CHANGE] = "P_key change",
74 [IB_EVENT_SM_CHANGE] = "SM change",
75 [IB_EVENT_SRQ_ERR] = "SRQ error",
76 [IB_EVENT_SRQ_LIMIT_REACHED] = "SRQ limit reached",
77 [IB_EVENT_QP_LAST_WQE_REACHED] = "last WQE reached",
78 [IB_EVENT_CLIENT_REREGISTER] = "client reregister",
79 [IB_EVENT_GID_CHANGE] = "GID changed",
80 };
81
82 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
83 {
84 size_t index = event;
85
86 return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
87 ib_events[index] : "unrecognized event";
88 }
89 EXPORT_SYMBOL(ib_event_msg);
90
91 static const char * const wc_statuses[] = {
92 [IB_WC_SUCCESS] = "success",
93 [IB_WC_LOC_LEN_ERR] = "local length error",
94 [IB_WC_LOC_QP_OP_ERR] = "local QP operation error",
95 [IB_WC_LOC_EEC_OP_ERR] = "local EE context operation error",
96 [IB_WC_LOC_PROT_ERR] = "local protection error",
97 [IB_WC_WR_FLUSH_ERR] = "WR flushed",
98 [IB_WC_MW_BIND_ERR] = "memory management operation error",
99 [IB_WC_BAD_RESP_ERR] = "bad response error",
100 [IB_WC_LOC_ACCESS_ERR] = "local access error",
101 [IB_WC_REM_INV_REQ_ERR] = "invalid request error",
102 [IB_WC_REM_ACCESS_ERR] = "remote access error",
103 [IB_WC_REM_OP_ERR] = "remote operation error",
104 [IB_WC_RETRY_EXC_ERR] = "transport retry counter exceeded",
105 [IB_WC_RNR_RETRY_EXC_ERR] = "RNR retry counter exceeded",
106 [IB_WC_LOC_RDD_VIOL_ERR] = "local RDD violation error",
107 [IB_WC_REM_INV_RD_REQ_ERR] = "remote invalid RD request",
108 [IB_WC_REM_ABORT_ERR] = "operation aborted",
109 [IB_WC_INV_EECN_ERR] = "invalid EE context number",
110 [IB_WC_INV_EEC_STATE_ERR] = "invalid EE context state",
111 [IB_WC_FATAL_ERR] = "fatal error",
112 [IB_WC_RESP_TIMEOUT_ERR] = "response timeout error",
113 [IB_WC_GENERAL_ERR] = "general error",
114 };
115
116 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
117 {
118 size_t index = status;
119
120 return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
121 wc_statuses[index] : "unrecognized status";
122 }
123 EXPORT_SYMBOL(ib_wc_status_msg);
124
125 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
126 {
127 switch (rate) {
128 case IB_RATE_2_5_GBPS: return 1;
129 case IB_RATE_5_GBPS: return 2;
130 case IB_RATE_10_GBPS: return 4;
131 case IB_RATE_20_GBPS: return 8;
132 case IB_RATE_30_GBPS: return 12;
133 case IB_RATE_40_GBPS: return 16;
134 case IB_RATE_60_GBPS: return 24;
135 case IB_RATE_80_GBPS: return 32;
136 case IB_RATE_120_GBPS: return 48;
137 case IB_RATE_14_GBPS: return 6;
138 case IB_RATE_56_GBPS: return 22;
139 case IB_RATE_112_GBPS: return 45;
140 case IB_RATE_168_GBPS: return 67;
141 case IB_RATE_25_GBPS: return 10;
142 case IB_RATE_100_GBPS: return 40;
143 case IB_RATE_200_GBPS: return 80;
144 case IB_RATE_300_GBPS: return 120;
145 case IB_RATE_28_GBPS: return 11;
146 case IB_RATE_50_GBPS: return 20;
147 case IB_RATE_400_GBPS: return 160;
148 case IB_RATE_600_GBPS: return 240;
149 default: return -1;
150 }
151 }
152 EXPORT_SYMBOL(ib_rate_to_mult);
153
154 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
155 {
156 switch (mult) {
157 case 1: return IB_RATE_2_5_GBPS;
158 case 2: return IB_RATE_5_GBPS;
159 case 4: return IB_RATE_10_GBPS;
160 case 8: return IB_RATE_20_GBPS;
161 case 12: return IB_RATE_30_GBPS;
162 case 16: return IB_RATE_40_GBPS;
163 case 24: return IB_RATE_60_GBPS;
164 case 32: return IB_RATE_80_GBPS;
165 case 48: return IB_RATE_120_GBPS;
166 case 6: return IB_RATE_14_GBPS;
167 case 22: return IB_RATE_56_GBPS;
168 case 45: return IB_RATE_112_GBPS;
169 case 67: return IB_RATE_168_GBPS;
170 case 10: return IB_RATE_25_GBPS;
171 case 40: return IB_RATE_100_GBPS;
172 case 80: return IB_RATE_200_GBPS;
173 case 120: return IB_RATE_300_GBPS;
174 case 11: return IB_RATE_28_GBPS;
175 case 20: return IB_RATE_50_GBPS;
176 case 160: return IB_RATE_400_GBPS;
177 case 240: return IB_RATE_600_GBPS;
178 default: return IB_RATE_PORT_CURRENT;
179 }
180 }
181 EXPORT_SYMBOL(mult_to_ib_rate);
182
183 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
184 {
185 switch (rate) {
186 case IB_RATE_2_5_GBPS: return 2500;
187 case IB_RATE_5_GBPS: return 5000;
188 case IB_RATE_10_GBPS: return 10000;
189 case IB_RATE_20_GBPS: return 20000;
190 case IB_RATE_30_GBPS: return 30000;
191 case IB_RATE_40_GBPS: return 40000;
192 case IB_RATE_60_GBPS: return 60000;
193 case IB_RATE_80_GBPS: return 80000;
194 case IB_RATE_120_GBPS: return 120000;
195 case IB_RATE_14_GBPS: return 14062;
196 case IB_RATE_56_GBPS: return 56250;
197 case IB_RATE_112_GBPS: return 112500;
198 case IB_RATE_168_GBPS: return 168750;
199 case IB_RATE_25_GBPS: return 25781;
200 case IB_RATE_100_GBPS: return 103125;
201 case IB_RATE_200_GBPS: return 206250;
202 case IB_RATE_300_GBPS: return 309375;
203 case IB_RATE_28_GBPS: return 28125;
204 case IB_RATE_50_GBPS: return 53125;
205 case IB_RATE_400_GBPS: return 425000;
206 case IB_RATE_600_GBPS: return 637500;
207 default: return -1;
208 }
209 }
210 EXPORT_SYMBOL(ib_rate_to_mbps);
211
212 __attribute_const__ enum rdma_transport_type
213 rdma_node_get_transport(unsigned int node_type)
214 {
215
216 if (node_type == RDMA_NODE_USNIC)
217 return RDMA_TRANSPORT_USNIC;
218 if (node_type == RDMA_NODE_USNIC_UDP)
219 return RDMA_TRANSPORT_USNIC_UDP;
220 if (node_type == RDMA_NODE_RNIC)
221 return RDMA_TRANSPORT_IWARP;
222 if (node_type == RDMA_NODE_UNSPECIFIED)
223 return RDMA_TRANSPORT_UNSPECIFIED;
224
225 return RDMA_TRANSPORT_IB;
226 }
227 EXPORT_SYMBOL(rdma_node_get_transport);
228
229 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
230 {
231 enum rdma_transport_type lt;
232 if (device->ops.get_link_layer)
233 return device->ops.get_link_layer(device, port_num);
234
235 lt = rdma_node_get_transport(device->node_type);
236 if (lt == RDMA_TRANSPORT_IB)
237 return IB_LINK_LAYER_INFINIBAND;
238
239 return IB_LINK_LAYER_ETHERNET;
240 }
241 EXPORT_SYMBOL(rdma_port_get_link_layer);
242
243 /* Protection domains */
244
245 /**
246 * ib_alloc_pd - Allocates an unused protection domain.
247 * @device: The device on which to allocate the protection domain.
248 * @flags: protection domain flags
249 * @caller: caller's build-time module name
250 *
251 * A protection domain object provides an association between QPs, shared
252 * receive queues, address handles, memory regions, and memory windows.
253 *
254 * Every PD has a local_dma_lkey which can be used as the lkey value for local
255 * memory operations.
256 */
257 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
258 const char *caller)
259 {
260 struct ib_pd *pd;
261 int mr_access_flags = 0;
262 int ret;
263
264 pd = rdma_zalloc_drv_obj(device, ib_pd);
265 if (!pd)
266 return ERR_PTR(-ENOMEM);
267
268 pd->device = device;
269 pd->uobject = NULL;
270 pd->__internal_mr = NULL;
271 atomic_set(&pd->usecnt, 0);
272 pd->flags = flags;
273
274 pd->res.type = RDMA_RESTRACK_PD;
275 rdma_restrack_set_task(&pd->res, caller);
276
277 ret = device->ops.alloc_pd(pd, NULL);
278 if (ret) {
279 kfree(pd);
280 return ERR_PTR(ret);
281 }
282 rdma_restrack_kadd(&pd->res);
283
284 if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
285 pd->local_dma_lkey = device->local_dma_lkey;
286 else
287 mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
288
289 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
290 pr_warn("%s: enabling unsafe global rkey\n", caller);
291 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
292 }
293
294 if (mr_access_flags) {
295 struct ib_mr *mr;
296
297 mr = pd->device->ops.get_dma_mr(pd, mr_access_flags);
298 if (IS_ERR(mr)) {
299 ib_dealloc_pd(pd);
300 return ERR_CAST(mr);
301 }
302
303 mr->device = pd->device;
304 mr->pd = pd;
305 mr->type = IB_MR_TYPE_DMA;
306 mr->uobject = NULL;
307 mr->need_inval = false;
308
309 pd->__internal_mr = mr;
310
311 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
312 pd->local_dma_lkey = pd->__internal_mr->lkey;
313
314 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
315 pd->unsafe_global_rkey = pd->__internal_mr->rkey;
316 }
317
318 return pd;
319 }
320 EXPORT_SYMBOL(__ib_alloc_pd);
321
322 /**
323 * ib_dealloc_pd_user - Deallocates a protection domain.
324 * @pd: The protection domain to deallocate.
325 * @udata: Valid user data or NULL for kernel object
326 *
327 * It is an error to call this function while any resources in the pd still
328 * exist. The caller is responsible to synchronously destroy them and
329 * guarantee no new allocations will happen.
330 */
331 void ib_dealloc_pd_user(struct ib_pd *pd, struct ib_udata *udata)
332 {
333 int ret;
334
335 if (pd->__internal_mr) {
336 ret = pd->device->ops.dereg_mr(pd->__internal_mr, NULL);
337 WARN_ON(ret);
338 pd->__internal_mr = NULL;
339 }
340
341 /* uverbs manipulates usecnt with proper locking, while the kabi
342 requires the caller to guarantee we can't race here. */
343 WARN_ON(atomic_read(&pd->usecnt));
344
345 rdma_restrack_del(&pd->res);
346 pd->device->ops.dealloc_pd(pd, udata);
347 kfree(pd);
348 }
349 EXPORT_SYMBOL(ib_dealloc_pd_user);
350
351 /* Address handles */
352
353 /**
354 * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
355 * @dest: Pointer to destination ah_attr. Contents of the destination
356 * pointer is assumed to be invalid and attribute are overwritten.
357 * @src: Pointer to source ah_attr.
358 */
359 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
360 const struct rdma_ah_attr *src)
361 {
362 *dest = *src;
363 if (dest->grh.sgid_attr)
364 rdma_hold_gid_attr(dest->grh.sgid_attr);
365 }
366 EXPORT_SYMBOL(rdma_copy_ah_attr);
367
368 /**
369 * rdma_replace_ah_attr - Replace valid ah_attr with new new one.
370 * @old: Pointer to existing ah_attr which needs to be replaced.
371 * old is assumed to be valid or zero'd
372 * @new: Pointer to the new ah_attr.
373 *
374 * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
375 * old the ah_attr is valid; after that it copies the new attribute and holds
376 * the reference to the replaced ah_attr.
377 */
378 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
379 const struct rdma_ah_attr *new)
380 {
381 rdma_destroy_ah_attr(old);
382 *old = *new;
383 if (old->grh.sgid_attr)
384 rdma_hold_gid_attr(old->grh.sgid_attr);
385 }
386 EXPORT_SYMBOL(rdma_replace_ah_attr);
387
388 /**
389 * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
390 * @dest: Pointer to destination ah_attr to copy to.
391 * dest is assumed to be valid or zero'd
392 * @src: Pointer to the new ah_attr.
393 *
394 * rdma_move_ah_attr() first releases any reference in the destination ah_attr
395 * if it is valid. This also transfers ownership of internal references from
396 * src to dest, making src invalid in the process. No new reference of the src
397 * ah_attr is taken.
398 */
399 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
400 {
401 rdma_destroy_ah_attr(dest);
402 *dest = *src;
403 src->grh.sgid_attr = NULL;
404 }
405 EXPORT_SYMBOL(rdma_move_ah_attr);
406
407 /*
408 * Validate that the rdma_ah_attr is valid for the device before passing it
409 * off to the driver.
410 */
411 static int rdma_check_ah_attr(struct ib_device *device,
412 struct rdma_ah_attr *ah_attr)
413 {
414 if (!rdma_is_port_valid(device, ah_attr->port_num))
415 return -EINVAL;
416
417 if ((rdma_is_grh_required(device, ah_attr->port_num) ||
418 ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
419 !(ah_attr->ah_flags & IB_AH_GRH))
420 return -EINVAL;
421
422 if (ah_attr->grh.sgid_attr) {
423 /*
424 * Make sure the passed sgid_attr is consistent with the
425 * parameters
426 */
427 if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
428 ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
429 return -EINVAL;
430 }
431 return 0;
432 }
433
434 /*
435 * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
436 * On success the caller is responsible to call rdma_unfill_sgid_attr().
437 */
438 static int rdma_fill_sgid_attr(struct ib_device *device,
439 struct rdma_ah_attr *ah_attr,
440 const struct ib_gid_attr **old_sgid_attr)
441 {
442 const struct ib_gid_attr *sgid_attr;
443 struct ib_global_route *grh;
444 int ret;
445
446 *old_sgid_attr = ah_attr->grh.sgid_attr;
447
448 ret = rdma_check_ah_attr(device, ah_attr);
449 if (ret)
450 return ret;
451
452 if (!(ah_attr->ah_flags & IB_AH_GRH))
453 return 0;
454
455 grh = rdma_ah_retrieve_grh(ah_attr);
456 if (grh->sgid_attr)
457 return 0;
458
459 sgid_attr =
460 rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
461 if (IS_ERR(sgid_attr))
462 return PTR_ERR(sgid_attr);
463
464 /* Move ownerhip of the kref into the ah_attr */
465 grh->sgid_attr = sgid_attr;
466 return 0;
467 }
468
469 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
470 const struct ib_gid_attr *old_sgid_attr)
471 {
472 /*
473 * Fill didn't change anything, the caller retains ownership of
474 * whatever it passed
475 */
476 if (ah_attr->grh.sgid_attr == old_sgid_attr)
477 return;
478
479 /*
480 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
481 * doesn't see any change in the rdma_ah_attr. If we get here
482 * old_sgid_attr is NULL.
483 */
484 rdma_destroy_ah_attr(ah_attr);
485 }
486
487 static const struct ib_gid_attr *
488 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
489 const struct ib_gid_attr *old_attr)
490 {
491 if (old_attr)
492 rdma_put_gid_attr(old_attr);
493 if (ah_attr->ah_flags & IB_AH_GRH) {
494 rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
495 return ah_attr->grh.sgid_attr;
496 }
497 return NULL;
498 }
499
500 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
501 struct rdma_ah_attr *ah_attr,
502 u32 flags,
503 struct ib_udata *udata)
504 {
505 struct ib_device *device = pd->device;
506 struct ib_ah *ah;
507 int ret;
508
509 might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE);
510
511 if (!device->ops.create_ah)
512 return ERR_PTR(-EOPNOTSUPP);
513
514 ah = rdma_zalloc_drv_obj_gfp(
515 device, ib_ah,
516 (flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC);
517 if (!ah)
518 return ERR_PTR(-ENOMEM);
519
520 ah->device = device;
521 ah->pd = pd;
522 ah->type = ah_attr->type;
523 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
524
525 ret = device->ops.create_ah(ah, ah_attr, flags, udata);
526 if (ret) {
527 kfree(ah);
528 return ERR_PTR(ret);
529 }
530
531 atomic_inc(&pd->usecnt);
532 return ah;
533 }
534
535 /**
536 * rdma_create_ah - Creates an address handle for the
537 * given address vector.
538 * @pd: The protection domain associated with the address handle.
539 * @ah_attr: The attributes of the address vector.
540 * @flags: Create address handle flags (see enum rdma_create_ah_flags).
541 *
542 * It returns 0 on success and returns appropriate error code on error.
543 * The address handle is used to reference a local or global destination
544 * in all UD QP post sends.
545 */
546 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
547 u32 flags)
548 {
549 const struct ib_gid_attr *old_sgid_attr;
550 struct ib_ah *ah;
551 int ret;
552
553 ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
554 if (ret)
555 return ERR_PTR(ret);
556
557 ah = _rdma_create_ah(pd, ah_attr, flags, NULL);
558
559 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
560 return ah;
561 }
562 EXPORT_SYMBOL(rdma_create_ah);
563
564 /**
565 * rdma_create_user_ah - Creates an address handle for the
566 * given address vector.
567 * It resolves destination mac address for ah attribute of RoCE type.
568 * @pd: The protection domain associated with the address handle.
569 * @ah_attr: The attributes of the address vector.
570 * @udata: pointer to user's input output buffer information need by
571 * provider driver.
572 *
573 * It returns 0 on success and returns appropriate error code on error.
574 * The address handle is used to reference a local or global destination
575 * in all UD QP post sends.
576 */
577 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
578 struct rdma_ah_attr *ah_attr,
579 struct ib_udata *udata)
580 {
581 const struct ib_gid_attr *old_sgid_attr;
582 struct ib_ah *ah;
583 int err;
584
585 err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
586 if (err)
587 return ERR_PTR(err);
588
589 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
590 err = ib_resolve_eth_dmac(pd->device, ah_attr);
591 if (err) {
592 ah = ERR_PTR(err);
593 goto out;
594 }
595 }
596
597 ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE, udata);
598
599 out:
600 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
601 return ah;
602 }
603 EXPORT_SYMBOL(rdma_create_user_ah);
604
605 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
606 {
607 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
608 struct iphdr ip4h_checked;
609 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
610
611 /* If it's IPv6, the version must be 6, otherwise, the first
612 * 20 bytes (before the IPv4 header) are garbled.
613 */
614 if (ip6h->version != 6)
615 return (ip4h->version == 4) ? 4 : 0;
616 /* version may be 6 or 4 because the first 20 bytes could be garbled */
617
618 /* RoCE v2 requires no options, thus header length
619 * must be 5 words
620 */
621 if (ip4h->ihl != 5)
622 return 6;
623
624 /* Verify checksum.
625 * We can't write on scattered buffers so we need to copy to
626 * temp buffer.
627 */
628 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
629 ip4h_checked.check = 0;
630 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
631 /* if IPv4 header checksum is OK, believe it */
632 if (ip4h->check == ip4h_checked.check)
633 return 4;
634 return 6;
635 }
636 EXPORT_SYMBOL(ib_get_rdma_header_version);
637
638 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
639 u8 port_num,
640 const struct ib_grh *grh)
641 {
642 int grh_version;
643
644 if (rdma_protocol_ib(device, port_num))
645 return RDMA_NETWORK_IB;
646
647 grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
648
649 if (grh_version == 4)
650 return RDMA_NETWORK_IPV4;
651
652 if (grh->next_hdr == IPPROTO_UDP)
653 return RDMA_NETWORK_IPV6;
654
655 return RDMA_NETWORK_ROCE_V1;
656 }
657
658 struct find_gid_index_context {
659 u16 vlan_id;
660 enum ib_gid_type gid_type;
661 };
662
663 static bool find_gid_index(const union ib_gid *gid,
664 const struct ib_gid_attr *gid_attr,
665 void *context)
666 {
667 struct find_gid_index_context *ctx = context;
668 u16 vlan_id = 0xffff;
669 int ret;
670
671 if (ctx->gid_type != gid_attr->gid_type)
672 return false;
673
674 ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
675 if (ret)
676 return false;
677
678 return ctx->vlan_id == vlan_id;
679 }
680
681 static const struct ib_gid_attr *
682 get_sgid_attr_from_eth(struct ib_device *device, u8 port_num,
683 u16 vlan_id, const union ib_gid *sgid,
684 enum ib_gid_type gid_type)
685 {
686 struct find_gid_index_context context = {.vlan_id = vlan_id,
687 .gid_type = gid_type};
688
689 return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
690 &context);
691 }
692
693 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
694 enum rdma_network_type net_type,
695 union ib_gid *sgid, union ib_gid *dgid)
696 {
697 struct sockaddr_in src_in;
698 struct sockaddr_in dst_in;
699 __be32 src_saddr, dst_saddr;
700
701 if (!sgid || !dgid)
702 return -EINVAL;
703
704 if (net_type == RDMA_NETWORK_IPV4) {
705 memcpy(&src_in.sin_addr.s_addr,
706 &hdr->roce4grh.saddr, 4);
707 memcpy(&dst_in.sin_addr.s_addr,
708 &hdr->roce4grh.daddr, 4);
709 src_saddr = src_in.sin_addr.s_addr;
710 dst_saddr = dst_in.sin_addr.s_addr;
711 ipv6_addr_set_v4mapped(src_saddr,
712 (struct in6_addr *)sgid);
713 ipv6_addr_set_v4mapped(dst_saddr,
714 (struct in6_addr *)dgid);
715 return 0;
716 } else if (net_type == RDMA_NETWORK_IPV6 ||
717 net_type == RDMA_NETWORK_IB) {
718 *dgid = hdr->ibgrh.dgid;
719 *sgid = hdr->ibgrh.sgid;
720 return 0;
721 } else {
722 return -EINVAL;
723 }
724 }
725 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
726
727 /* Resolve destination mac address and hop limit for unicast destination
728 * GID entry, considering the source GID entry as well.
729 * ah_attribute must have have valid port_num, sgid_index.
730 */
731 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
732 struct rdma_ah_attr *ah_attr)
733 {
734 struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
735 const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
736 int hop_limit = 0xff;
737 int ret = 0;
738
739 /* If destination is link local and source GID is RoCEv1,
740 * IP stack is not used.
741 */
742 if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
743 sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
744 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
745 ah_attr->roce.dmac);
746 return ret;
747 }
748
749 ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
750 ah_attr->roce.dmac,
751 sgid_attr, &hop_limit);
752
753 grh->hop_limit = hop_limit;
754 return ret;
755 }
756
757 /*
758 * This function initializes address handle attributes from the incoming packet.
759 * Incoming packet has dgid of the receiver node on which this code is
760 * getting executed and, sgid contains the GID of the sender.
761 *
762 * When resolving mac address of destination, the arrived dgid is used
763 * as sgid and, sgid is used as dgid because sgid contains destinations
764 * GID whom to respond to.
765 *
766 * On success the caller is responsible to call rdma_destroy_ah_attr on the
767 * attr.
768 */
769 int ib_init_ah_attr_from_wc(struct ib_device *device, u8 port_num,
770 const struct ib_wc *wc, const struct ib_grh *grh,
771 struct rdma_ah_attr *ah_attr)
772 {
773 u32 flow_class;
774 int ret;
775 enum rdma_network_type net_type = RDMA_NETWORK_IB;
776 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
777 const struct ib_gid_attr *sgid_attr;
778 int hoplimit = 0xff;
779 union ib_gid dgid;
780 union ib_gid sgid;
781
782 might_sleep();
783
784 memset(ah_attr, 0, sizeof *ah_attr);
785 ah_attr->type = rdma_ah_find_type(device, port_num);
786 if (rdma_cap_eth_ah(device, port_num)) {
787 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
788 net_type = wc->network_hdr_type;
789 else
790 net_type = ib_get_net_type_by_grh(device, port_num, grh);
791 gid_type = ib_network_to_gid_type(net_type);
792 }
793 ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
794 &sgid, &dgid);
795 if (ret)
796 return ret;
797
798 rdma_ah_set_sl(ah_attr, wc->sl);
799 rdma_ah_set_port_num(ah_attr, port_num);
800
801 if (rdma_protocol_roce(device, port_num)) {
802 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
803 wc->vlan_id : 0xffff;
804
805 if (!(wc->wc_flags & IB_WC_GRH))
806 return -EPROTOTYPE;
807
808 sgid_attr = get_sgid_attr_from_eth(device, port_num,
809 vlan_id, &dgid,
810 gid_type);
811 if (IS_ERR(sgid_attr))
812 return PTR_ERR(sgid_attr);
813
814 flow_class = be32_to_cpu(grh->version_tclass_flow);
815 rdma_move_grh_sgid_attr(ah_attr,
816 &sgid,
817 flow_class & 0xFFFFF,
818 hoplimit,
819 (flow_class >> 20) & 0xFF,
820 sgid_attr);
821
822 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
823 if (ret)
824 rdma_destroy_ah_attr(ah_attr);
825
826 return ret;
827 } else {
828 rdma_ah_set_dlid(ah_attr, wc->slid);
829 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
830
831 if ((wc->wc_flags & IB_WC_GRH) == 0)
832 return 0;
833
834 if (dgid.global.interface_id !=
835 cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
836 sgid_attr = rdma_find_gid_by_port(
837 device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
838 } else
839 sgid_attr = rdma_get_gid_attr(device, port_num, 0);
840
841 if (IS_ERR(sgid_attr))
842 return PTR_ERR(sgid_attr);
843 flow_class = be32_to_cpu(grh->version_tclass_flow);
844 rdma_move_grh_sgid_attr(ah_attr,
845 &sgid,
846 flow_class & 0xFFFFF,
847 hoplimit,
848 (flow_class >> 20) & 0xFF,
849 sgid_attr);
850
851 return 0;
852 }
853 }
854 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
855
856 /**
857 * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
858 * of the reference
859 *
860 * @attr: Pointer to AH attribute structure
861 * @dgid: Destination GID
862 * @flow_label: Flow label
863 * @hop_limit: Hop limit
864 * @traffic_class: traffic class
865 * @sgid_attr: Pointer to SGID attribute
866 *
867 * This takes ownership of the sgid_attr reference. The caller must ensure
868 * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
869 * calling this function.
870 */
871 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
872 u32 flow_label, u8 hop_limit, u8 traffic_class,
873 const struct ib_gid_attr *sgid_attr)
874 {
875 rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
876 traffic_class);
877 attr->grh.sgid_attr = sgid_attr;
878 }
879 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
880
881 /**
882 * rdma_destroy_ah_attr - Release reference to SGID attribute of
883 * ah attribute.
884 * @ah_attr: Pointer to ah attribute
885 *
886 * Release reference to the SGID attribute of the ah attribute if it is
887 * non NULL. It is safe to call this multiple times, and safe to call it on
888 * a zero initialized ah_attr.
889 */
890 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
891 {
892 if (ah_attr->grh.sgid_attr) {
893 rdma_put_gid_attr(ah_attr->grh.sgid_attr);
894 ah_attr->grh.sgid_attr = NULL;
895 }
896 }
897 EXPORT_SYMBOL(rdma_destroy_ah_attr);
898
899 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
900 const struct ib_grh *grh, u8 port_num)
901 {
902 struct rdma_ah_attr ah_attr;
903 struct ib_ah *ah;
904 int ret;
905
906 ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
907 if (ret)
908 return ERR_PTR(ret);
909
910 ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE);
911
912 rdma_destroy_ah_attr(&ah_attr);
913 return ah;
914 }
915 EXPORT_SYMBOL(ib_create_ah_from_wc);
916
917 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
918 {
919 const struct ib_gid_attr *old_sgid_attr;
920 int ret;
921
922 if (ah->type != ah_attr->type)
923 return -EINVAL;
924
925 ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
926 if (ret)
927 return ret;
928
929 ret = ah->device->ops.modify_ah ?
930 ah->device->ops.modify_ah(ah, ah_attr) :
931 -EOPNOTSUPP;
932
933 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
934 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
935 return ret;
936 }
937 EXPORT_SYMBOL(rdma_modify_ah);
938
939 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
940 {
941 ah_attr->grh.sgid_attr = NULL;
942
943 return ah->device->ops.query_ah ?
944 ah->device->ops.query_ah(ah, ah_attr) :
945 -EOPNOTSUPP;
946 }
947 EXPORT_SYMBOL(rdma_query_ah);
948
949 int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata)
950 {
951 const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
952 struct ib_pd *pd;
953
954 might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE);
955
956 pd = ah->pd;
957
958 ah->device->ops.destroy_ah(ah, flags);
959 atomic_dec(&pd->usecnt);
960 if (sgid_attr)
961 rdma_put_gid_attr(sgid_attr);
962
963 kfree(ah);
964 return 0;
965 }
966 EXPORT_SYMBOL(rdma_destroy_ah_user);
967
968 /* Shared receive queues */
969
970 struct ib_srq *ib_create_srq(struct ib_pd *pd,
971 struct ib_srq_init_attr *srq_init_attr)
972 {
973 struct ib_srq *srq;
974 int ret;
975
976 if (!pd->device->ops.create_srq)
977 return ERR_PTR(-EOPNOTSUPP);
978
979 srq = rdma_zalloc_drv_obj(pd->device, ib_srq);
980 if (!srq)
981 return ERR_PTR(-ENOMEM);
982
983 srq->device = pd->device;
984 srq->pd = pd;
985 srq->event_handler = srq_init_attr->event_handler;
986 srq->srq_context = srq_init_attr->srq_context;
987 srq->srq_type = srq_init_attr->srq_type;
988
989 if (ib_srq_has_cq(srq->srq_type)) {
990 srq->ext.cq = srq_init_attr->ext.cq;
991 atomic_inc(&srq->ext.cq->usecnt);
992 }
993 if (srq->srq_type == IB_SRQT_XRC) {
994 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
995 atomic_inc(&srq->ext.xrc.xrcd->usecnt);
996 }
997 atomic_inc(&pd->usecnt);
998
999 ret = pd->device->ops.create_srq(srq, srq_init_attr, NULL);
1000 if (ret) {
1001 atomic_dec(&srq->pd->usecnt);
1002 if (srq->srq_type == IB_SRQT_XRC)
1003 atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1004 if (ib_srq_has_cq(srq->srq_type))
1005 atomic_dec(&srq->ext.cq->usecnt);
1006 kfree(srq);
1007 return ERR_PTR(ret);
1008 }
1009
1010 return srq;
1011 }
1012 EXPORT_SYMBOL(ib_create_srq);
1013
1014 int ib_modify_srq(struct ib_srq *srq,
1015 struct ib_srq_attr *srq_attr,
1016 enum ib_srq_attr_mask srq_attr_mask)
1017 {
1018 return srq->device->ops.modify_srq ?
1019 srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask,
1020 NULL) : -EOPNOTSUPP;
1021 }
1022 EXPORT_SYMBOL(ib_modify_srq);
1023
1024 int ib_query_srq(struct ib_srq *srq,
1025 struct ib_srq_attr *srq_attr)
1026 {
1027 return srq->device->ops.query_srq ?
1028 srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP;
1029 }
1030 EXPORT_SYMBOL(ib_query_srq);
1031
1032 int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata)
1033 {
1034 if (atomic_read(&srq->usecnt))
1035 return -EBUSY;
1036
1037 srq->device->ops.destroy_srq(srq, udata);
1038
1039 atomic_dec(&srq->pd->usecnt);
1040 if (srq->srq_type == IB_SRQT_XRC)
1041 atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1042 if (ib_srq_has_cq(srq->srq_type))
1043 atomic_dec(&srq->ext.cq->usecnt);
1044 kfree(srq);
1045
1046 return 0;
1047 }
1048 EXPORT_SYMBOL(ib_destroy_srq_user);
1049
1050 /* Queue pairs */
1051
1052 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1053 {
1054 struct ib_qp *qp = context;
1055 unsigned long flags;
1056
1057 spin_lock_irqsave(&qp->device->qp_open_list_lock, flags);
1058 list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1059 if (event->element.qp->event_handler)
1060 event->element.qp->event_handler(event, event->element.qp->qp_context);
1061 spin_unlock_irqrestore(&qp->device->qp_open_list_lock, flags);
1062 }
1063
1064 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
1065 {
1066 mutex_lock(&xrcd->tgt_qp_mutex);
1067 list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
1068 mutex_unlock(&xrcd->tgt_qp_mutex);
1069 }
1070
1071 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1072 void (*event_handler)(struct ib_event *, void *),
1073 void *qp_context)
1074 {
1075 struct ib_qp *qp;
1076 unsigned long flags;
1077 int err;
1078
1079 qp = kzalloc(sizeof *qp, GFP_KERNEL);
1080 if (!qp)
1081 return ERR_PTR(-ENOMEM);
1082
1083 qp->real_qp = real_qp;
1084 err = ib_open_shared_qp_security(qp, real_qp->device);
1085 if (err) {
1086 kfree(qp);
1087 return ERR_PTR(err);
1088 }
1089
1090 qp->real_qp = real_qp;
1091 atomic_inc(&real_qp->usecnt);
1092 qp->device = real_qp->device;
1093 qp->event_handler = event_handler;
1094 qp->qp_context = qp_context;
1095 qp->qp_num = real_qp->qp_num;
1096 qp->qp_type = real_qp->qp_type;
1097
1098 spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1099 list_add(&qp->open_list, &real_qp->open_list);
1100 spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1101
1102 return qp;
1103 }
1104
1105 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1106 struct ib_qp_open_attr *qp_open_attr)
1107 {
1108 struct ib_qp *qp, *real_qp;
1109
1110 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1111 return ERR_PTR(-EINVAL);
1112
1113 qp = ERR_PTR(-EINVAL);
1114 mutex_lock(&xrcd->tgt_qp_mutex);
1115 list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
1116 if (real_qp->qp_num == qp_open_attr->qp_num) {
1117 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1118 qp_open_attr->qp_context);
1119 break;
1120 }
1121 }
1122 mutex_unlock(&xrcd->tgt_qp_mutex);
1123 return qp;
1124 }
1125 EXPORT_SYMBOL(ib_open_qp);
1126
1127 static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp,
1128 struct ib_qp_init_attr *qp_init_attr)
1129 {
1130 struct ib_qp *real_qp = qp;
1131
1132 qp->event_handler = __ib_shared_qp_event_handler;
1133 qp->qp_context = qp;
1134 qp->pd = NULL;
1135 qp->send_cq = qp->recv_cq = NULL;
1136 qp->srq = NULL;
1137 qp->xrcd = qp_init_attr->xrcd;
1138 atomic_inc(&qp_init_attr->xrcd->usecnt);
1139 INIT_LIST_HEAD(&qp->open_list);
1140
1141 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1142 qp_init_attr->qp_context);
1143 if (IS_ERR(qp))
1144 return qp;
1145
1146 __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
1147 return qp;
1148 }
1149
1150 /**
1151 * ib_create_qp - Creates a kernel QP associated with the specified protection
1152 * domain.
1153 * @pd: The protection domain associated with the QP.
1154 * @qp_init_attr: A list of initial attributes required to create the
1155 * QP. If QP creation succeeds, then the attributes are updated to
1156 * the actual capabilities of the created QP.
1157 *
1158 * NOTE: for user qp use ib_create_qp_user with valid udata!
1159 */
1160 struct ib_qp *ib_create_qp(struct ib_pd *pd,
1161 struct ib_qp_init_attr *qp_init_attr)
1162 {
1163 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
1164 struct ib_qp *qp;
1165 int ret;
1166
1167 if (qp_init_attr->rwq_ind_tbl &&
1168 (qp_init_attr->recv_cq ||
1169 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
1170 qp_init_attr->cap.max_recv_sge))
1171 return ERR_PTR(-EINVAL);
1172
1173 if ((qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN) &&
1174 !(device->attrs.device_cap_flags & IB_DEVICE_INTEGRITY_HANDOVER))
1175 return ERR_PTR(-EINVAL);
1176
1177 /*
1178 * If the callers is using the RDMA API calculate the resources
1179 * needed for the RDMA READ/WRITE operations.
1180 *
1181 * Note that these callers need to pass in a port number.
1182 */
1183 if (qp_init_attr->cap.max_rdma_ctxs)
1184 rdma_rw_init_qp(device, qp_init_attr);
1185
1186 qp = _ib_create_qp(device, pd, qp_init_attr, NULL, NULL);
1187 if (IS_ERR(qp))
1188 return qp;
1189
1190 ret = ib_create_qp_security(qp, device);
1191 if (ret)
1192 goto err;
1193
1194 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT) {
1195 struct ib_qp *xrc_qp =
1196 create_xrc_qp_user(qp, qp_init_attr);
1197
1198 if (IS_ERR(xrc_qp)) {
1199 ret = PTR_ERR(xrc_qp);
1200 goto err;
1201 }
1202 return xrc_qp;
1203 }
1204
1205 qp->event_handler = qp_init_attr->event_handler;
1206 qp->qp_context = qp_init_attr->qp_context;
1207 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
1208 qp->recv_cq = NULL;
1209 qp->srq = NULL;
1210 } else {
1211 qp->recv_cq = qp_init_attr->recv_cq;
1212 if (qp_init_attr->recv_cq)
1213 atomic_inc(&qp_init_attr->recv_cq->usecnt);
1214 qp->srq = qp_init_attr->srq;
1215 if (qp->srq)
1216 atomic_inc(&qp_init_attr->srq->usecnt);
1217 }
1218
1219 qp->send_cq = qp_init_attr->send_cq;
1220 qp->xrcd = NULL;
1221
1222 atomic_inc(&pd->usecnt);
1223 if (qp_init_attr->send_cq)
1224 atomic_inc(&qp_init_attr->send_cq->usecnt);
1225 if (qp_init_attr->rwq_ind_tbl)
1226 atomic_inc(&qp->rwq_ind_tbl->usecnt);
1227
1228 if (qp_init_attr->cap.max_rdma_ctxs) {
1229 ret = rdma_rw_init_mrs(qp, qp_init_attr);
1230 if (ret)
1231 goto err;
1232 }
1233
1234 /*
1235 * Note: all hw drivers guarantee that max_send_sge is lower than
1236 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1237 * max_send_sge <= max_sge_rd.
1238 */
1239 qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1240 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1241 device->attrs.max_sge_rd);
1242 if (qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN)
1243 qp->integrity_en = true;
1244
1245 return qp;
1246
1247 err:
1248 ib_destroy_qp(qp);
1249 return ERR_PTR(ret);
1250
1251 }
1252 EXPORT_SYMBOL(ib_create_qp);
1253
1254 static const struct {
1255 int valid;
1256 enum ib_qp_attr_mask req_param[IB_QPT_MAX];
1257 enum ib_qp_attr_mask opt_param[IB_QPT_MAX];
1258 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1259 [IB_QPS_RESET] = {
1260 [IB_QPS_RESET] = { .valid = 1 },
1261 [IB_QPS_INIT] = {
1262 .valid = 1,
1263 .req_param = {
1264 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1265 IB_QP_PORT |
1266 IB_QP_QKEY),
1267 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
1268 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1269 IB_QP_PORT |
1270 IB_QP_ACCESS_FLAGS),
1271 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1272 IB_QP_PORT |
1273 IB_QP_ACCESS_FLAGS),
1274 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1275 IB_QP_PORT |
1276 IB_QP_ACCESS_FLAGS),
1277 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1278 IB_QP_PORT |
1279 IB_QP_ACCESS_FLAGS),
1280 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1281 IB_QP_QKEY),
1282 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1283 IB_QP_QKEY),
1284 }
1285 },
1286 },
1287 [IB_QPS_INIT] = {
1288 [IB_QPS_RESET] = { .valid = 1 },
1289 [IB_QPS_ERR] = { .valid = 1 },
1290 [IB_QPS_INIT] = {
1291 .valid = 1,
1292 .opt_param = {
1293 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1294 IB_QP_PORT |
1295 IB_QP_QKEY),
1296 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1297 IB_QP_PORT |
1298 IB_QP_ACCESS_FLAGS),
1299 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1300 IB_QP_PORT |
1301 IB_QP_ACCESS_FLAGS),
1302 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1303 IB_QP_PORT |
1304 IB_QP_ACCESS_FLAGS),
1305 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1306 IB_QP_PORT |
1307 IB_QP_ACCESS_FLAGS),
1308 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1309 IB_QP_QKEY),
1310 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1311 IB_QP_QKEY),
1312 }
1313 },
1314 [IB_QPS_RTR] = {
1315 .valid = 1,
1316 .req_param = {
1317 [IB_QPT_UC] = (IB_QP_AV |
1318 IB_QP_PATH_MTU |
1319 IB_QP_DEST_QPN |
1320 IB_QP_RQ_PSN),
1321 [IB_QPT_RC] = (IB_QP_AV |
1322 IB_QP_PATH_MTU |
1323 IB_QP_DEST_QPN |
1324 IB_QP_RQ_PSN |
1325 IB_QP_MAX_DEST_RD_ATOMIC |
1326 IB_QP_MIN_RNR_TIMER),
1327 [IB_QPT_XRC_INI] = (IB_QP_AV |
1328 IB_QP_PATH_MTU |
1329 IB_QP_DEST_QPN |
1330 IB_QP_RQ_PSN),
1331 [IB_QPT_XRC_TGT] = (IB_QP_AV |
1332 IB_QP_PATH_MTU |
1333 IB_QP_DEST_QPN |
1334 IB_QP_RQ_PSN |
1335 IB_QP_MAX_DEST_RD_ATOMIC |
1336 IB_QP_MIN_RNR_TIMER),
1337 },
1338 .opt_param = {
1339 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1340 IB_QP_QKEY),
1341 [IB_QPT_UC] = (IB_QP_ALT_PATH |
1342 IB_QP_ACCESS_FLAGS |
1343 IB_QP_PKEY_INDEX),
1344 [IB_QPT_RC] = (IB_QP_ALT_PATH |
1345 IB_QP_ACCESS_FLAGS |
1346 IB_QP_PKEY_INDEX),
1347 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH |
1348 IB_QP_ACCESS_FLAGS |
1349 IB_QP_PKEY_INDEX),
1350 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH |
1351 IB_QP_ACCESS_FLAGS |
1352 IB_QP_PKEY_INDEX),
1353 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1354 IB_QP_QKEY),
1355 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1356 IB_QP_QKEY),
1357 },
1358 },
1359 },
1360 [IB_QPS_RTR] = {
1361 [IB_QPS_RESET] = { .valid = 1 },
1362 [IB_QPS_ERR] = { .valid = 1 },
1363 [IB_QPS_RTS] = {
1364 .valid = 1,
1365 .req_param = {
1366 [IB_QPT_UD] = IB_QP_SQ_PSN,
1367 [IB_QPT_UC] = IB_QP_SQ_PSN,
1368 [IB_QPT_RC] = (IB_QP_TIMEOUT |
1369 IB_QP_RETRY_CNT |
1370 IB_QP_RNR_RETRY |
1371 IB_QP_SQ_PSN |
1372 IB_QP_MAX_QP_RD_ATOMIC),
1373 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT |
1374 IB_QP_RETRY_CNT |
1375 IB_QP_RNR_RETRY |
1376 IB_QP_SQ_PSN |
1377 IB_QP_MAX_QP_RD_ATOMIC),
1378 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT |
1379 IB_QP_SQ_PSN),
1380 [IB_QPT_SMI] = IB_QP_SQ_PSN,
1381 [IB_QPT_GSI] = IB_QP_SQ_PSN,
1382 },
1383 .opt_param = {
1384 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1385 IB_QP_QKEY),
1386 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1387 IB_QP_ALT_PATH |
1388 IB_QP_ACCESS_FLAGS |
1389 IB_QP_PATH_MIG_STATE),
1390 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1391 IB_QP_ALT_PATH |
1392 IB_QP_ACCESS_FLAGS |
1393 IB_QP_MIN_RNR_TIMER |
1394 IB_QP_PATH_MIG_STATE),
1395 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1396 IB_QP_ALT_PATH |
1397 IB_QP_ACCESS_FLAGS |
1398 IB_QP_PATH_MIG_STATE),
1399 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1400 IB_QP_ALT_PATH |
1401 IB_QP_ACCESS_FLAGS |
1402 IB_QP_MIN_RNR_TIMER |
1403 IB_QP_PATH_MIG_STATE),
1404 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1405 IB_QP_QKEY),
1406 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1407 IB_QP_QKEY),
1408 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1409 }
1410 }
1411 },
1412 [IB_QPS_RTS] = {
1413 [IB_QPS_RESET] = { .valid = 1 },
1414 [IB_QPS_ERR] = { .valid = 1 },
1415 [IB_QPS_RTS] = {
1416 .valid = 1,
1417 .opt_param = {
1418 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1419 IB_QP_QKEY),
1420 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1421 IB_QP_ACCESS_FLAGS |
1422 IB_QP_ALT_PATH |
1423 IB_QP_PATH_MIG_STATE),
1424 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1425 IB_QP_ACCESS_FLAGS |
1426 IB_QP_ALT_PATH |
1427 IB_QP_PATH_MIG_STATE |
1428 IB_QP_MIN_RNR_TIMER),
1429 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1430 IB_QP_ACCESS_FLAGS |
1431 IB_QP_ALT_PATH |
1432 IB_QP_PATH_MIG_STATE),
1433 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1434 IB_QP_ACCESS_FLAGS |
1435 IB_QP_ALT_PATH |
1436 IB_QP_PATH_MIG_STATE |
1437 IB_QP_MIN_RNR_TIMER),
1438 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1439 IB_QP_QKEY),
1440 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1441 IB_QP_QKEY),
1442 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1443 }
1444 },
1445 [IB_QPS_SQD] = {
1446 .valid = 1,
1447 .opt_param = {
1448 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1449 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1450 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1451 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1452 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1453 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1454 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1455 }
1456 },
1457 },
1458 [IB_QPS_SQD] = {
1459 [IB_QPS_RESET] = { .valid = 1 },
1460 [IB_QPS_ERR] = { .valid = 1 },
1461 [IB_QPS_RTS] = {
1462 .valid = 1,
1463 .opt_param = {
1464 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1465 IB_QP_QKEY),
1466 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1467 IB_QP_ALT_PATH |
1468 IB_QP_ACCESS_FLAGS |
1469 IB_QP_PATH_MIG_STATE),
1470 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1471 IB_QP_ALT_PATH |
1472 IB_QP_ACCESS_FLAGS |
1473 IB_QP_MIN_RNR_TIMER |
1474 IB_QP_PATH_MIG_STATE),
1475 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1476 IB_QP_ALT_PATH |
1477 IB_QP_ACCESS_FLAGS |
1478 IB_QP_PATH_MIG_STATE),
1479 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1480 IB_QP_ALT_PATH |
1481 IB_QP_ACCESS_FLAGS |
1482 IB_QP_MIN_RNR_TIMER |
1483 IB_QP_PATH_MIG_STATE),
1484 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1485 IB_QP_QKEY),
1486 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1487 IB_QP_QKEY),
1488 }
1489 },
1490 [IB_QPS_SQD] = {
1491 .valid = 1,
1492 .opt_param = {
1493 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1494 IB_QP_QKEY),
1495 [IB_QPT_UC] = (IB_QP_AV |
1496 IB_QP_ALT_PATH |
1497 IB_QP_ACCESS_FLAGS |
1498 IB_QP_PKEY_INDEX |
1499 IB_QP_PATH_MIG_STATE),
1500 [IB_QPT_RC] = (IB_QP_PORT |
1501 IB_QP_AV |
1502 IB_QP_TIMEOUT |
1503 IB_QP_RETRY_CNT |
1504 IB_QP_RNR_RETRY |
1505 IB_QP_MAX_QP_RD_ATOMIC |
1506 IB_QP_MAX_DEST_RD_ATOMIC |
1507 IB_QP_ALT_PATH |
1508 IB_QP_ACCESS_FLAGS |
1509 IB_QP_PKEY_INDEX |
1510 IB_QP_MIN_RNR_TIMER |
1511 IB_QP_PATH_MIG_STATE),
1512 [IB_QPT_XRC_INI] = (IB_QP_PORT |
1513 IB_QP_AV |
1514 IB_QP_TIMEOUT |
1515 IB_QP_RETRY_CNT |
1516 IB_QP_RNR_RETRY |
1517 IB_QP_MAX_QP_RD_ATOMIC |
1518 IB_QP_ALT_PATH |
1519 IB_QP_ACCESS_FLAGS |
1520 IB_QP_PKEY_INDEX |
1521 IB_QP_PATH_MIG_STATE),
1522 [IB_QPT_XRC_TGT] = (IB_QP_PORT |
1523 IB_QP_AV |
1524 IB_QP_TIMEOUT |
1525 IB_QP_MAX_DEST_RD_ATOMIC |
1526 IB_QP_ALT_PATH |
1527 IB_QP_ACCESS_FLAGS |
1528 IB_QP_PKEY_INDEX |
1529 IB_QP_MIN_RNR_TIMER |
1530 IB_QP_PATH_MIG_STATE),
1531 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1532 IB_QP_QKEY),
1533 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1534 IB_QP_QKEY),
1535 }
1536 }
1537 },
1538 [IB_QPS_SQE] = {
1539 [IB_QPS_RESET] = { .valid = 1 },
1540 [IB_QPS_ERR] = { .valid = 1 },
1541 [IB_QPS_RTS] = {
1542 .valid = 1,
1543 .opt_param = {
1544 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1545 IB_QP_QKEY),
1546 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1547 IB_QP_ACCESS_FLAGS),
1548 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1549 IB_QP_QKEY),
1550 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1551 IB_QP_QKEY),
1552 }
1553 }
1554 },
1555 [IB_QPS_ERR] = {
1556 [IB_QPS_RESET] = { .valid = 1 },
1557 [IB_QPS_ERR] = { .valid = 1 }
1558 }
1559 };
1560
1561 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1562 enum ib_qp_type type, enum ib_qp_attr_mask mask)
1563 {
1564 enum ib_qp_attr_mask req_param, opt_param;
1565
1566 if (mask & IB_QP_CUR_STATE &&
1567 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1568 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1569 return false;
1570
1571 if (!qp_state_table[cur_state][next_state].valid)
1572 return false;
1573
1574 req_param = qp_state_table[cur_state][next_state].req_param[type];
1575 opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1576
1577 if ((mask & req_param) != req_param)
1578 return false;
1579
1580 if (mask & ~(req_param | opt_param | IB_QP_STATE))
1581 return false;
1582
1583 return true;
1584 }
1585 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1586
1587 /**
1588 * ib_resolve_eth_dmac - Resolve destination mac address
1589 * @device: Device to consider
1590 * @ah_attr: address handle attribute which describes the
1591 * source and destination parameters
1592 * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1593 * returns 0 on success or appropriate error code. It initializes the
1594 * necessary ah_attr fields when call is successful.
1595 */
1596 static int ib_resolve_eth_dmac(struct ib_device *device,
1597 struct rdma_ah_attr *ah_attr)
1598 {
1599 int ret = 0;
1600
1601 if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1602 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1603 __be32 addr = 0;
1604
1605 memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1606 ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1607 } else {
1608 ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1609 (char *)ah_attr->roce.dmac);
1610 }
1611 } else {
1612 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1613 }
1614 return ret;
1615 }
1616
1617 static bool is_qp_type_connected(const struct ib_qp *qp)
1618 {
1619 return (qp->qp_type == IB_QPT_UC ||
1620 qp->qp_type == IB_QPT_RC ||
1621 qp->qp_type == IB_QPT_XRC_INI ||
1622 qp->qp_type == IB_QPT_XRC_TGT);
1623 }
1624
1625 /**
1626 * IB core internal function to perform QP attributes modification.
1627 */
1628 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1629 int attr_mask, struct ib_udata *udata)
1630 {
1631 u8 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1632 const struct ib_gid_attr *old_sgid_attr_av;
1633 const struct ib_gid_attr *old_sgid_attr_alt_av;
1634 int ret;
1635
1636 if (attr_mask & IB_QP_AV) {
1637 ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1638 &old_sgid_attr_av);
1639 if (ret)
1640 return ret;
1641 }
1642 if (attr_mask & IB_QP_ALT_PATH) {
1643 /*
1644 * FIXME: This does not track the migration state, so if the
1645 * user loads a new alternate path after the HW has migrated
1646 * from primary->alternate we will keep the wrong
1647 * references. This is OK for IB because the reference
1648 * counting does not serve any functional purpose.
1649 */
1650 ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1651 &old_sgid_attr_alt_av);
1652 if (ret)
1653 goto out_av;
1654
1655 /*
1656 * Today the core code can only handle alternate paths and APM
1657 * for IB. Ban them in roce mode.
1658 */
1659 if (!(rdma_protocol_ib(qp->device,
1660 attr->alt_ah_attr.port_num) &&
1661 rdma_protocol_ib(qp->device, port))) {
1662 ret = EINVAL;
1663 goto out;
1664 }
1665 }
1666
1667 /*
1668 * If the user provided the qp_attr then we have to resolve it. Kernel
1669 * users have to provide already resolved rdma_ah_attr's
1670 */
1671 if (udata && (attr_mask & IB_QP_AV) &&
1672 attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1673 is_qp_type_connected(qp)) {
1674 ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
1675 if (ret)
1676 goto out;
1677 }
1678
1679 if (rdma_ib_or_roce(qp->device, port)) {
1680 if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1681 dev_warn(&qp->device->dev,
1682 "%s rq_psn overflow, masking to 24 bits\n",
1683 __func__);
1684 attr->rq_psn &= 0xffffff;
1685 }
1686
1687 if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1688 dev_warn(&qp->device->dev,
1689 " %s sq_psn overflow, masking to 24 bits\n",
1690 __func__);
1691 attr->sq_psn &= 0xffffff;
1692 }
1693 }
1694
1695 /*
1696 * Bind this qp to a counter automatically based on the rdma counter
1697 * rules. This only set in RST2INIT with port specified
1698 */
1699 if (!qp->counter && (attr_mask & IB_QP_PORT) &&
1700 ((attr_mask & IB_QP_STATE) && attr->qp_state == IB_QPS_INIT))
1701 rdma_counter_bind_qp_auto(qp, attr->port_num);
1702
1703 ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1704 if (ret)
1705 goto out;
1706
1707 if (attr_mask & IB_QP_PORT)
1708 qp->port = attr->port_num;
1709 if (attr_mask & IB_QP_AV)
1710 qp->av_sgid_attr =
1711 rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1712 if (attr_mask & IB_QP_ALT_PATH)
1713 qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1714 &attr->alt_ah_attr, qp->alt_path_sgid_attr);
1715
1716 out:
1717 if (attr_mask & IB_QP_ALT_PATH)
1718 rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1719 out_av:
1720 if (attr_mask & IB_QP_AV)
1721 rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1722 return ret;
1723 }
1724
1725 /**
1726 * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1727 * @ib_qp: The QP to modify.
1728 * @attr: On input, specifies the QP attributes to modify. On output,
1729 * the current values of selected QP attributes are returned.
1730 * @attr_mask: A bit-mask used to specify which attributes of the QP
1731 * are being modified.
1732 * @udata: pointer to user's input output buffer information
1733 * are being modified.
1734 * It returns 0 on success and returns appropriate error code on error.
1735 */
1736 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1737 int attr_mask, struct ib_udata *udata)
1738 {
1739 return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1740 }
1741 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1742
1743 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u8 *speed, u8 *width)
1744 {
1745 int rc;
1746 u32 netdev_speed;
1747 struct net_device *netdev;
1748 struct ethtool_link_ksettings lksettings;
1749
1750 if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1751 return -EINVAL;
1752
1753 netdev = ib_device_get_netdev(dev, port_num);
1754 if (!netdev)
1755 return -ENODEV;
1756
1757 rtnl_lock();
1758 rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1759 rtnl_unlock();
1760
1761 dev_put(netdev);
1762
1763 if (!rc) {
1764 netdev_speed = lksettings.base.speed;
1765 } else {
1766 netdev_speed = SPEED_1000;
1767 pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name,
1768 netdev_speed);
1769 }
1770
1771 if (netdev_speed <= SPEED_1000) {
1772 *width = IB_WIDTH_1X;
1773 *speed = IB_SPEED_SDR;
1774 } else if (netdev_speed <= SPEED_10000) {
1775 *width = IB_WIDTH_1X;
1776 *speed = IB_SPEED_FDR10;
1777 } else if (netdev_speed <= SPEED_20000) {
1778 *width = IB_WIDTH_4X;
1779 *speed = IB_SPEED_DDR;
1780 } else if (netdev_speed <= SPEED_25000) {
1781 *width = IB_WIDTH_1X;
1782 *speed = IB_SPEED_EDR;
1783 } else if (netdev_speed <= SPEED_40000) {
1784 *width = IB_WIDTH_4X;
1785 *speed = IB_SPEED_FDR10;
1786 } else {
1787 *width = IB_WIDTH_4X;
1788 *speed = IB_SPEED_EDR;
1789 }
1790
1791 return 0;
1792 }
1793 EXPORT_SYMBOL(ib_get_eth_speed);
1794
1795 int ib_modify_qp(struct ib_qp *qp,
1796 struct ib_qp_attr *qp_attr,
1797 int qp_attr_mask)
1798 {
1799 return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1800 }
1801 EXPORT_SYMBOL(ib_modify_qp);
1802
1803 int ib_query_qp(struct ib_qp *qp,
1804 struct ib_qp_attr *qp_attr,
1805 int qp_attr_mask,
1806 struct ib_qp_init_attr *qp_init_attr)
1807 {
1808 qp_attr->ah_attr.grh.sgid_attr = NULL;
1809 qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
1810
1811 return qp->device->ops.query_qp ?
1812 qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask,
1813 qp_init_attr) : -EOPNOTSUPP;
1814 }
1815 EXPORT_SYMBOL(ib_query_qp);
1816
1817 int ib_close_qp(struct ib_qp *qp)
1818 {
1819 struct ib_qp *real_qp;
1820 unsigned long flags;
1821
1822 real_qp = qp->real_qp;
1823 if (real_qp == qp)
1824 return -EINVAL;
1825
1826 spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1827 list_del(&qp->open_list);
1828 spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1829
1830 atomic_dec(&real_qp->usecnt);
1831 if (qp->qp_sec)
1832 ib_close_shared_qp_security(qp->qp_sec);
1833 kfree(qp);
1834
1835 return 0;
1836 }
1837 EXPORT_SYMBOL(ib_close_qp);
1838
1839 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1840 {
1841 struct ib_xrcd *xrcd;
1842 struct ib_qp *real_qp;
1843 int ret;
1844
1845 real_qp = qp->real_qp;
1846 xrcd = real_qp->xrcd;
1847
1848 mutex_lock(&xrcd->tgt_qp_mutex);
1849 ib_close_qp(qp);
1850 if (atomic_read(&real_qp->usecnt) == 0)
1851 list_del(&real_qp->xrcd_list);
1852 else
1853 real_qp = NULL;
1854 mutex_unlock(&xrcd->tgt_qp_mutex);
1855
1856 if (real_qp) {
1857 ret = ib_destroy_qp(real_qp);
1858 if (!ret)
1859 atomic_dec(&xrcd->usecnt);
1860 else
1861 __ib_insert_xrcd_qp(xrcd, real_qp);
1862 }
1863
1864 return 0;
1865 }
1866
1867 int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata)
1868 {
1869 const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
1870 const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
1871 struct ib_pd *pd;
1872 struct ib_cq *scq, *rcq;
1873 struct ib_srq *srq;
1874 struct ib_rwq_ind_table *ind_tbl;
1875 struct ib_qp_security *sec;
1876 int ret;
1877
1878 WARN_ON_ONCE(qp->mrs_used > 0);
1879
1880 if (atomic_read(&qp->usecnt))
1881 return -EBUSY;
1882
1883 if (qp->real_qp != qp)
1884 return __ib_destroy_shared_qp(qp);
1885
1886 pd = qp->pd;
1887 scq = qp->send_cq;
1888 rcq = qp->recv_cq;
1889 srq = qp->srq;
1890 ind_tbl = qp->rwq_ind_tbl;
1891 sec = qp->qp_sec;
1892 if (sec)
1893 ib_destroy_qp_security_begin(sec);
1894
1895 if (!qp->uobject)
1896 rdma_rw_cleanup_mrs(qp);
1897
1898 rdma_counter_unbind_qp(qp, true);
1899 rdma_restrack_del(&qp->res);
1900 ret = qp->device->ops.destroy_qp(qp, udata);
1901 if (!ret) {
1902 if (alt_path_sgid_attr)
1903 rdma_put_gid_attr(alt_path_sgid_attr);
1904 if (av_sgid_attr)
1905 rdma_put_gid_attr(av_sgid_attr);
1906 if (pd)
1907 atomic_dec(&pd->usecnt);
1908 if (scq)
1909 atomic_dec(&scq->usecnt);
1910 if (rcq)
1911 atomic_dec(&rcq->usecnt);
1912 if (srq)
1913 atomic_dec(&srq->usecnt);
1914 if (ind_tbl)
1915 atomic_dec(&ind_tbl->usecnt);
1916 if (sec)
1917 ib_destroy_qp_security_end(sec);
1918 } else {
1919 if (sec)
1920 ib_destroy_qp_security_abort(sec);
1921 }
1922
1923 return ret;
1924 }
1925 EXPORT_SYMBOL(ib_destroy_qp_user);
1926
1927 /* Completion queues */
1928
1929 struct ib_cq *__ib_create_cq(struct ib_device *device,
1930 ib_comp_handler comp_handler,
1931 void (*event_handler)(struct ib_event *, void *),
1932 void *cq_context,
1933 const struct ib_cq_init_attr *cq_attr,
1934 const char *caller)
1935 {
1936 struct ib_cq *cq;
1937 int ret;
1938
1939 cq = rdma_zalloc_drv_obj(device, ib_cq);
1940 if (!cq)
1941 return ERR_PTR(-ENOMEM);
1942
1943 cq->device = device;
1944 cq->uobject = NULL;
1945 cq->comp_handler = comp_handler;
1946 cq->event_handler = event_handler;
1947 cq->cq_context = cq_context;
1948 atomic_set(&cq->usecnt, 0);
1949 cq->res.type = RDMA_RESTRACK_CQ;
1950 rdma_restrack_set_task(&cq->res, caller);
1951
1952 ret = device->ops.create_cq(cq, cq_attr, NULL);
1953 if (ret) {
1954 kfree(cq);
1955 return ERR_PTR(ret);
1956 }
1957
1958 rdma_restrack_kadd(&cq->res);
1959 return cq;
1960 }
1961 EXPORT_SYMBOL(__ib_create_cq);
1962
1963 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1964 {
1965 return cq->device->ops.modify_cq ?
1966 cq->device->ops.modify_cq(cq, cq_count,
1967 cq_period) : -EOPNOTSUPP;
1968 }
1969 EXPORT_SYMBOL(rdma_set_cq_moderation);
1970
1971 int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata)
1972 {
1973 if (atomic_read(&cq->usecnt))
1974 return -EBUSY;
1975
1976 rdma_restrack_del(&cq->res);
1977 cq->device->ops.destroy_cq(cq, udata);
1978 kfree(cq);
1979 return 0;
1980 }
1981 EXPORT_SYMBOL(ib_destroy_cq_user);
1982
1983 int ib_resize_cq(struct ib_cq *cq, int cqe)
1984 {
1985 return cq->device->ops.resize_cq ?
1986 cq->device->ops.resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
1987 }
1988 EXPORT_SYMBOL(ib_resize_cq);
1989
1990 /* Memory regions */
1991
1992 struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
1993 u64 virt_addr, int access_flags)
1994 {
1995 struct ib_mr *mr;
1996
1997 if (access_flags & IB_ACCESS_ON_DEMAND) {
1998 if (!(pd->device->attrs.device_cap_flags &
1999 IB_DEVICE_ON_DEMAND_PAGING)) {
2000 pr_debug("ODP support not available\n");
2001 return ERR_PTR(-EINVAL);
2002 }
2003 }
2004
2005 mr = pd->device->ops.reg_user_mr(pd, start, length, virt_addr,
2006 access_flags, NULL);
2007
2008 if (IS_ERR(mr))
2009 return mr;
2010
2011 mr->device = pd->device;
2012 mr->pd = pd;
2013 mr->dm = NULL;
2014 atomic_inc(&pd->usecnt);
2015 mr->res.type = RDMA_RESTRACK_MR;
2016 rdma_restrack_kadd(&mr->res);
2017
2018 return mr;
2019 }
2020 EXPORT_SYMBOL(ib_reg_user_mr);
2021
2022 int ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
2023 u32 flags, struct ib_sge *sg_list, u32 num_sge)
2024 {
2025 if (!pd->device->ops.advise_mr)
2026 return -EOPNOTSUPP;
2027
2028 return pd->device->ops.advise_mr(pd, advice, flags, sg_list, num_sge,
2029 NULL);
2030 }
2031 EXPORT_SYMBOL(ib_advise_mr);
2032
2033 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata)
2034 {
2035 struct ib_pd *pd = mr->pd;
2036 struct ib_dm *dm = mr->dm;
2037 struct ib_sig_attrs *sig_attrs = mr->sig_attrs;
2038 int ret;
2039
2040 trace_mr_dereg(mr);
2041 rdma_restrack_del(&mr->res);
2042 ret = mr->device->ops.dereg_mr(mr, udata);
2043 if (!ret) {
2044 atomic_dec(&pd->usecnt);
2045 if (dm)
2046 atomic_dec(&dm->usecnt);
2047 kfree(sig_attrs);
2048 }
2049
2050 return ret;
2051 }
2052 EXPORT_SYMBOL(ib_dereg_mr_user);
2053
2054 /**
2055 * ib_alloc_mr_user() - Allocates a memory region
2056 * @pd: protection domain associated with the region
2057 * @mr_type: memory region type
2058 * @max_num_sg: maximum sg entries available for registration.
2059 * @udata: user data or null for kernel objects
2060 *
2061 * Notes:
2062 * Memory registeration page/sg lists must not exceed max_num_sg.
2063 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
2064 * max_num_sg * used_page_size.
2065 *
2066 */
2067 struct ib_mr *ib_alloc_mr_user(struct ib_pd *pd, enum ib_mr_type mr_type,
2068 u32 max_num_sg, struct ib_udata *udata)
2069 {
2070 struct ib_mr *mr;
2071
2072 if (!pd->device->ops.alloc_mr) {
2073 mr = ERR_PTR(-EOPNOTSUPP);
2074 goto out;
2075 }
2076
2077 if (mr_type == IB_MR_TYPE_INTEGRITY) {
2078 WARN_ON_ONCE(1);
2079 mr = ERR_PTR(-EINVAL);
2080 goto out;
2081 }
2082
2083 mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg, udata);
2084 if (!IS_ERR(mr)) {
2085 mr->device = pd->device;
2086 mr->pd = pd;
2087 mr->dm = NULL;
2088 mr->uobject = NULL;
2089 atomic_inc(&pd->usecnt);
2090 mr->need_inval = false;
2091 mr->res.type = RDMA_RESTRACK_MR;
2092 rdma_restrack_kadd(&mr->res);
2093 mr->type = mr_type;
2094 mr->sig_attrs = NULL;
2095 }
2096
2097 out:
2098 trace_mr_alloc(pd, mr_type, max_num_sg, mr);
2099 return mr;
2100 }
2101 EXPORT_SYMBOL(ib_alloc_mr_user);
2102
2103 /**
2104 * ib_alloc_mr_integrity() - Allocates an integrity memory region
2105 * @pd: protection domain associated with the region
2106 * @max_num_data_sg: maximum data sg entries available for registration
2107 * @max_num_meta_sg: maximum metadata sg entries available for
2108 * registration
2109 *
2110 * Notes:
2111 * Memory registration page/sg lists must not exceed max_num_sg,
2112 * also the integrity page/sg lists must not exceed max_num_meta_sg.
2113 *
2114 */
2115 struct ib_mr *ib_alloc_mr_integrity(struct ib_pd *pd,
2116 u32 max_num_data_sg,
2117 u32 max_num_meta_sg)
2118 {
2119 struct ib_mr *mr;
2120 struct ib_sig_attrs *sig_attrs;
2121
2122 if (!pd->device->ops.alloc_mr_integrity ||
2123 !pd->device->ops.map_mr_sg_pi) {
2124 mr = ERR_PTR(-EOPNOTSUPP);
2125 goto out;
2126 }
2127
2128 if (!max_num_meta_sg) {
2129 mr = ERR_PTR(-EINVAL);
2130 goto out;
2131 }
2132
2133 sig_attrs = kzalloc(sizeof(struct ib_sig_attrs), GFP_KERNEL);
2134 if (!sig_attrs) {
2135 mr = ERR_PTR(-ENOMEM);
2136 goto out;
2137 }
2138
2139 mr = pd->device->ops.alloc_mr_integrity(pd, max_num_data_sg,
2140 max_num_meta_sg);
2141 if (IS_ERR(mr)) {
2142 kfree(sig_attrs);
2143 goto out;
2144 }
2145
2146 mr->device = pd->device;
2147 mr->pd = pd;
2148 mr->dm = NULL;
2149 mr->uobject = NULL;
2150 atomic_inc(&pd->usecnt);
2151 mr->need_inval = false;
2152 mr->res.type = RDMA_RESTRACK_MR;
2153 rdma_restrack_kadd(&mr->res);
2154 mr->type = IB_MR_TYPE_INTEGRITY;
2155 mr->sig_attrs = sig_attrs;
2156
2157 out:
2158 trace_mr_integ_alloc(pd, max_num_data_sg, max_num_meta_sg, mr);
2159 return mr;
2160 }
2161 EXPORT_SYMBOL(ib_alloc_mr_integrity);
2162
2163 /* "Fast" memory regions */
2164
2165 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
2166 int mr_access_flags,
2167 struct ib_fmr_attr *fmr_attr)
2168 {
2169 struct ib_fmr *fmr;
2170
2171 if (!pd->device->ops.alloc_fmr)
2172 return ERR_PTR(-EOPNOTSUPP);
2173
2174 fmr = pd->device->ops.alloc_fmr(pd, mr_access_flags, fmr_attr);
2175 if (!IS_ERR(fmr)) {
2176 fmr->device = pd->device;
2177 fmr->pd = pd;
2178 atomic_inc(&pd->usecnt);
2179 }
2180
2181 return fmr;
2182 }
2183 EXPORT_SYMBOL(ib_alloc_fmr);
2184
2185 int ib_unmap_fmr(struct list_head *fmr_list)
2186 {
2187 struct ib_fmr *fmr;
2188
2189 if (list_empty(fmr_list))
2190 return 0;
2191
2192 fmr = list_entry(fmr_list->next, struct ib_fmr, list);
2193 return fmr->device->ops.unmap_fmr(fmr_list);
2194 }
2195 EXPORT_SYMBOL(ib_unmap_fmr);
2196
2197 int ib_dealloc_fmr(struct ib_fmr *fmr)
2198 {
2199 struct ib_pd *pd;
2200 int ret;
2201
2202 pd = fmr->pd;
2203 ret = fmr->device->ops.dealloc_fmr(fmr);
2204 if (!ret)
2205 atomic_dec(&pd->usecnt);
2206
2207 return ret;
2208 }
2209 EXPORT_SYMBOL(ib_dealloc_fmr);
2210
2211 /* Multicast groups */
2212
2213 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2214 {
2215 struct ib_qp_init_attr init_attr = {};
2216 struct ib_qp_attr attr = {};
2217 int num_eth_ports = 0;
2218 int port;
2219
2220 /* If QP state >= init, it is assigned to a port and we can check this
2221 * port only.
2222 */
2223 if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2224 if (attr.qp_state >= IB_QPS_INIT) {
2225 if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2226 IB_LINK_LAYER_INFINIBAND)
2227 return true;
2228 goto lid_check;
2229 }
2230 }
2231
2232 /* Can't get a quick answer, iterate over all ports */
2233 for (port = 0; port < qp->device->phys_port_cnt; port++)
2234 if (rdma_port_get_link_layer(qp->device, port) !=
2235 IB_LINK_LAYER_INFINIBAND)
2236 num_eth_ports++;
2237
2238 /* If we have at lease one Ethernet port, RoCE annex declares that
2239 * multicast LID should be ignored. We can't tell at this step if the
2240 * QP belongs to an IB or Ethernet port.
2241 */
2242 if (num_eth_ports)
2243 return true;
2244
2245 /* If all the ports are IB, we can check according to IB spec. */
2246 lid_check:
2247 return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2248 lid == be16_to_cpu(IB_LID_PERMISSIVE));
2249 }
2250
2251 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2252 {
2253 int ret;
2254
2255 if (!qp->device->ops.attach_mcast)
2256 return -EOPNOTSUPP;
2257
2258 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2259 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2260 return -EINVAL;
2261
2262 ret = qp->device->ops.attach_mcast(qp, gid, lid);
2263 if (!ret)
2264 atomic_inc(&qp->usecnt);
2265 return ret;
2266 }
2267 EXPORT_SYMBOL(ib_attach_mcast);
2268
2269 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2270 {
2271 int ret;
2272
2273 if (!qp->device->ops.detach_mcast)
2274 return -EOPNOTSUPP;
2275
2276 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2277 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2278 return -EINVAL;
2279
2280 ret = qp->device->ops.detach_mcast(qp, gid, lid);
2281 if (!ret)
2282 atomic_dec(&qp->usecnt);
2283 return ret;
2284 }
2285 EXPORT_SYMBOL(ib_detach_mcast);
2286
2287 struct ib_xrcd *__ib_alloc_xrcd(struct ib_device *device, const char *caller)
2288 {
2289 struct ib_xrcd *xrcd;
2290
2291 if (!device->ops.alloc_xrcd)
2292 return ERR_PTR(-EOPNOTSUPP);
2293
2294 xrcd = device->ops.alloc_xrcd(device, NULL);
2295 if (!IS_ERR(xrcd)) {
2296 xrcd->device = device;
2297 xrcd->inode = NULL;
2298 atomic_set(&xrcd->usecnt, 0);
2299 mutex_init(&xrcd->tgt_qp_mutex);
2300 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
2301 }
2302
2303 return xrcd;
2304 }
2305 EXPORT_SYMBOL(__ib_alloc_xrcd);
2306
2307 int ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata)
2308 {
2309 struct ib_qp *qp;
2310 int ret;
2311
2312 if (atomic_read(&xrcd->usecnt))
2313 return -EBUSY;
2314
2315 while (!list_empty(&xrcd->tgt_qp_list)) {
2316 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
2317 ret = ib_destroy_qp(qp);
2318 if (ret)
2319 return ret;
2320 }
2321 mutex_destroy(&xrcd->tgt_qp_mutex);
2322
2323 return xrcd->device->ops.dealloc_xrcd(xrcd, udata);
2324 }
2325 EXPORT_SYMBOL(ib_dealloc_xrcd);
2326
2327 /**
2328 * ib_create_wq - Creates a WQ associated with the specified protection
2329 * domain.
2330 * @pd: The protection domain associated with the WQ.
2331 * @wq_attr: A list of initial attributes required to create the
2332 * WQ. If WQ creation succeeds, then the attributes are updated to
2333 * the actual capabilities of the created WQ.
2334 *
2335 * wq_attr->max_wr and wq_attr->max_sge determine
2336 * the requested size of the WQ, and set to the actual values allocated
2337 * on return.
2338 * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2339 * at least as large as the requested values.
2340 */
2341 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2342 struct ib_wq_init_attr *wq_attr)
2343 {
2344 struct ib_wq *wq;
2345
2346 if (!pd->device->ops.create_wq)
2347 return ERR_PTR(-EOPNOTSUPP);
2348
2349 wq = pd->device->ops.create_wq(pd, wq_attr, NULL);
2350 if (!IS_ERR(wq)) {
2351 wq->event_handler = wq_attr->event_handler;
2352 wq->wq_context = wq_attr->wq_context;
2353 wq->wq_type = wq_attr->wq_type;
2354 wq->cq = wq_attr->cq;
2355 wq->device = pd->device;
2356 wq->pd = pd;
2357 wq->uobject = NULL;
2358 atomic_inc(&pd->usecnt);
2359 atomic_inc(&wq_attr->cq->usecnt);
2360 atomic_set(&wq->usecnt, 0);
2361 }
2362 return wq;
2363 }
2364 EXPORT_SYMBOL(ib_create_wq);
2365
2366 /**
2367 * ib_destroy_wq - Destroys the specified user WQ.
2368 * @wq: The WQ to destroy.
2369 * @udata: Valid user data
2370 */
2371 int ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata)
2372 {
2373 struct ib_cq *cq = wq->cq;
2374 struct ib_pd *pd = wq->pd;
2375
2376 if (atomic_read(&wq->usecnt))
2377 return -EBUSY;
2378
2379 wq->device->ops.destroy_wq(wq, udata);
2380 atomic_dec(&pd->usecnt);
2381 atomic_dec(&cq->usecnt);
2382
2383 return 0;
2384 }
2385 EXPORT_SYMBOL(ib_destroy_wq);
2386
2387 /**
2388 * ib_modify_wq - Modifies the specified WQ.
2389 * @wq: The WQ to modify.
2390 * @wq_attr: On input, specifies the WQ attributes to modify.
2391 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
2392 * are being modified.
2393 * On output, the current values of selected WQ attributes are returned.
2394 */
2395 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
2396 u32 wq_attr_mask)
2397 {
2398 int err;
2399
2400 if (!wq->device->ops.modify_wq)
2401 return -EOPNOTSUPP;
2402
2403 err = wq->device->ops.modify_wq(wq, wq_attr, wq_attr_mask, NULL);
2404 return err;
2405 }
2406 EXPORT_SYMBOL(ib_modify_wq);
2407
2408 /*
2409 * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
2410 * @device: The device on which to create the rwq indirection table.
2411 * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
2412 * create the Indirection Table.
2413 *
2414 * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
2415 * than the created ib_rwq_ind_table object and the caller is responsible
2416 * for its memory allocation/free.
2417 */
2418 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
2419 struct ib_rwq_ind_table_init_attr *init_attr)
2420 {
2421 struct ib_rwq_ind_table *rwq_ind_table;
2422 int i;
2423 u32 table_size;
2424
2425 if (!device->ops.create_rwq_ind_table)
2426 return ERR_PTR(-EOPNOTSUPP);
2427
2428 table_size = (1 << init_attr->log_ind_tbl_size);
2429 rwq_ind_table = device->ops.create_rwq_ind_table(device,
2430 init_attr, NULL);
2431 if (IS_ERR(rwq_ind_table))
2432 return rwq_ind_table;
2433
2434 rwq_ind_table->ind_tbl = init_attr->ind_tbl;
2435 rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
2436 rwq_ind_table->device = device;
2437 rwq_ind_table->uobject = NULL;
2438 atomic_set(&rwq_ind_table->usecnt, 0);
2439
2440 for (i = 0; i < table_size; i++)
2441 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
2442
2443 return rwq_ind_table;
2444 }
2445 EXPORT_SYMBOL(ib_create_rwq_ind_table);
2446
2447 /*
2448 * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
2449 * @wq_ind_table: The Indirection Table to destroy.
2450 */
2451 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
2452 {
2453 int err, i;
2454 u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
2455 struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
2456
2457 if (atomic_read(&rwq_ind_table->usecnt))
2458 return -EBUSY;
2459
2460 err = rwq_ind_table->device->ops.destroy_rwq_ind_table(rwq_ind_table);
2461 if (!err) {
2462 for (i = 0; i < table_size; i++)
2463 atomic_dec(&ind_tbl[i]->usecnt);
2464 }
2465
2466 return err;
2467 }
2468 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
2469
2470 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2471 struct ib_mr_status *mr_status)
2472 {
2473 if (!mr->device->ops.check_mr_status)
2474 return -EOPNOTSUPP;
2475
2476 return mr->device->ops.check_mr_status(mr, check_mask, mr_status);
2477 }
2478 EXPORT_SYMBOL(ib_check_mr_status);
2479
2480 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
2481 int state)
2482 {
2483 if (!device->ops.set_vf_link_state)
2484 return -EOPNOTSUPP;
2485
2486 return device->ops.set_vf_link_state(device, vf, port, state);
2487 }
2488 EXPORT_SYMBOL(ib_set_vf_link_state);
2489
2490 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
2491 struct ifla_vf_info *info)
2492 {
2493 if (!device->ops.get_vf_config)
2494 return -EOPNOTSUPP;
2495
2496 return device->ops.get_vf_config(device, vf, port, info);
2497 }
2498 EXPORT_SYMBOL(ib_get_vf_config);
2499
2500 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
2501 struct ifla_vf_stats *stats)
2502 {
2503 if (!device->ops.get_vf_stats)
2504 return -EOPNOTSUPP;
2505
2506 return device->ops.get_vf_stats(device, vf, port, stats);
2507 }
2508 EXPORT_SYMBOL(ib_get_vf_stats);
2509
2510 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
2511 int type)
2512 {
2513 if (!device->ops.set_vf_guid)
2514 return -EOPNOTSUPP;
2515
2516 return device->ops.set_vf_guid(device, vf, port, guid, type);
2517 }
2518 EXPORT_SYMBOL(ib_set_vf_guid);
2519
2520 int ib_get_vf_guid(struct ib_device *device, int vf, u8 port,
2521 struct ifla_vf_guid *node_guid,
2522 struct ifla_vf_guid *port_guid)
2523 {
2524 if (!device->ops.get_vf_guid)
2525 return -EOPNOTSUPP;
2526
2527 return device->ops.get_vf_guid(device, vf, port, node_guid, port_guid);
2528 }
2529 EXPORT_SYMBOL(ib_get_vf_guid);
2530 /**
2531 * ib_map_mr_sg_pi() - Map the dma mapped SG lists for PI (protection
2532 * information) and set an appropriate memory region for registration.
2533 * @mr: memory region
2534 * @data_sg: dma mapped scatterlist for data
2535 * @data_sg_nents: number of entries in data_sg
2536 * @data_sg_offset: offset in bytes into data_sg
2537 * @meta_sg: dma mapped scatterlist for metadata
2538 * @meta_sg_nents: number of entries in meta_sg
2539 * @meta_sg_offset: offset in bytes into meta_sg
2540 * @page_size: page vector desired page size
2541 *
2542 * Constraints:
2543 * - The MR must be allocated with type IB_MR_TYPE_INTEGRITY.
2544 *
2545 * Return: 0 on success.
2546 *
2547 * After this completes successfully, the memory region
2548 * is ready for registration.
2549 */
2550 int ib_map_mr_sg_pi(struct ib_mr *mr, struct scatterlist *data_sg,
2551 int data_sg_nents, unsigned int *data_sg_offset,
2552 struct scatterlist *meta_sg, int meta_sg_nents,
2553 unsigned int *meta_sg_offset, unsigned int page_size)
2554 {
2555 if (unlikely(!mr->device->ops.map_mr_sg_pi ||
2556 WARN_ON_ONCE(mr->type != IB_MR_TYPE_INTEGRITY)))
2557 return -EOPNOTSUPP;
2558
2559 mr->page_size = page_size;
2560
2561 return mr->device->ops.map_mr_sg_pi(mr, data_sg, data_sg_nents,
2562 data_sg_offset, meta_sg,
2563 meta_sg_nents, meta_sg_offset);
2564 }
2565 EXPORT_SYMBOL(ib_map_mr_sg_pi);
2566
2567 /**
2568 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2569 * and set it the memory region.
2570 * @mr: memory region
2571 * @sg: dma mapped scatterlist
2572 * @sg_nents: number of entries in sg
2573 * @sg_offset: offset in bytes into sg
2574 * @page_size: page vector desired page size
2575 *
2576 * Constraints:
2577 * - The first sg element is allowed to have an offset.
2578 * - Each sg element must either be aligned to page_size or virtually
2579 * contiguous to the previous element. In case an sg element has a
2580 * non-contiguous offset, the mapping prefix will not include it.
2581 * - The last sg element is allowed to have length less than page_size.
2582 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2583 * then only max_num_sg entries will be mapped.
2584 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2585 * constraints holds and the page_size argument is ignored.
2586 *
2587 * Returns the number of sg elements that were mapped to the memory region.
2588 *
2589 * After this completes successfully, the memory region
2590 * is ready for registration.
2591 */
2592 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2593 unsigned int *sg_offset, unsigned int page_size)
2594 {
2595 if (unlikely(!mr->device->ops.map_mr_sg))
2596 return -EOPNOTSUPP;
2597
2598 mr->page_size = page_size;
2599
2600 return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset);
2601 }
2602 EXPORT_SYMBOL(ib_map_mr_sg);
2603
2604 /**
2605 * ib_sg_to_pages() - Convert the largest prefix of a sg list
2606 * to a page vector
2607 * @mr: memory region
2608 * @sgl: dma mapped scatterlist
2609 * @sg_nents: number of entries in sg
2610 * @sg_offset_p: IN: start offset in bytes into sg
2611 * OUT: offset in bytes for element n of the sg of the first
2612 * byte that has not been processed where n is the return
2613 * value of this function.
2614 * @set_page: driver page assignment function pointer
2615 *
2616 * Core service helper for drivers to convert the largest
2617 * prefix of given sg list to a page vector. The sg list
2618 * prefix converted is the prefix that meet the requirements
2619 * of ib_map_mr_sg.
2620 *
2621 * Returns the number of sg elements that were assigned to
2622 * a page vector.
2623 */
2624 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2625 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2626 {
2627 struct scatterlist *sg;
2628 u64 last_end_dma_addr = 0;
2629 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2630 unsigned int last_page_off = 0;
2631 u64 page_mask = ~((u64)mr->page_size - 1);
2632 int i, ret;
2633
2634 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2635 return -EINVAL;
2636
2637 mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2638 mr->length = 0;
2639
2640 for_each_sg(sgl, sg, sg_nents, i) {
2641 u64 dma_addr = sg_dma_address(sg) + sg_offset;
2642 u64 prev_addr = dma_addr;
2643 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2644 u64 end_dma_addr = dma_addr + dma_len;
2645 u64 page_addr = dma_addr & page_mask;
2646
2647 /*
2648 * For the second and later elements, check whether either the
2649 * end of element i-1 or the start of element i is not aligned
2650 * on a page boundary.
2651 */
2652 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2653 /* Stop mapping if there is a gap. */
2654 if (last_end_dma_addr != dma_addr)
2655 break;
2656
2657 /*
2658 * Coalesce this element with the last. If it is small
2659 * enough just update mr->length. Otherwise start
2660 * mapping from the next page.
2661 */
2662 goto next_page;
2663 }
2664
2665 do {
2666 ret = set_page(mr, page_addr);
2667 if (unlikely(ret < 0)) {
2668 sg_offset = prev_addr - sg_dma_address(sg);
2669 mr->length += prev_addr - dma_addr;
2670 if (sg_offset_p)
2671 *sg_offset_p = sg_offset;
2672 return i || sg_offset ? i : ret;
2673 }
2674 prev_addr = page_addr;
2675 next_page:
2676 page_addr += mr->page_size;
2677 } while (page_addr < end_dma_addr);
2678
2679 mr->length += dma_len;
2680 last_end_dma_addr = end_dma_addr;
2681 last_page_off = end_dma_addr & ~page_mask;
2682
2683 sg_offset = 0;
2684 }
2685
2686 if (sg_offset_p)
2687 *sg_offset_p = 0;
2688 return i;
2689 }
2690 EXPORT_SYMBOL(ib_sg_to_pages);
2691
2692 struct ib_drain_cqe {
2693 struct ib_cqe cqe;
2694 struct completion done;
2695 };
2696
2697 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2698 {
2699 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2700 cqe);
2701
2702 complete(&cqe->done);
2703 }
2704
2705 /*
2706 * Post a WR and block until its completion is reaped for the SQ.
2707 */
2708 static void __ib_drain_sq(struct ib_qp *qp)
2709 {
2710 struct ib_cq *cq = qp->send_cq;
2711 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2712 struct ib_drain_cqe sdrain;
2713 struct ib_rdma_wr swr = {
2714 .wr = {
2715 .next = NULL,
2716 { .wr_cqe = &sdrain.cqe, },
2717 .opcode = IB_WR_RDMA_WRITE,
2718 },
2719 };
2720 int ret;
2721
2722 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2723 if (ret) {
2724 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2725 return;
2726 }
2727
2728 sdrain.cqe.done = ib_drain_qp_done;
2729 init_completion(&sdrain.done);
2730
2731 ret = ib_post_send(qp, &swr.wr, NULL);
2732 if (ret) {
2733 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2734 return;
2735 }
2736
2737 if (cq->poll_ctx == IB_POLL_DIRECT)
2738 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2739 ib_process_cq_direct(cq, -1);
2740 else
2741 wait_for_completion(&sdrain.done);
2742 }
2743
2744 /*
2745 * Post a WR and block until its completion is reaped for the RQ.
2746 */
2747 static void __ib_drain_rq(struct ib_qp *qp)
2748 {
2749 struct ib_cq *cq = qp->recv_cq;
2750 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2751 struct ib_drain_cqe rdrain;
2752 struct ib_recv_wr rwr = {};
2753 int ret;
2754
2755 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2756 if (ret) {
2757 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2758 return;
2759 }
2760
2761 rwr.wr_cqe = &rdrain.cqe;
2762 rdrain.cqe.done = ib_drain_qp_done;
2763 init_completion(&rdrain.done);
2764
2765 ret = ib_post_recv(qp, &rwr, NULL);
2766 if (ret) {
2767 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2768 return;
2769 }
2770
2771 if (cq->poll_ctx == IB_POLL_DIRECT)
2772 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2773 ib_process_cq_direct(cq, -1);
2774 else
2775 wait_for_completion(&rdrain.done);
2776 }
2777
2778 /**
2779 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2780 * application.
2781 * @qp: queue pair to drain
2782 *
2783 * If the device has a provider-specific drain function, then
2784 * call that. Otherwise call the generic drain function
2785 * __ib_drain_sq().
2786 *
2787 * The caller must:
2788 *
2789 * ensure there is room in the CQ and SQ for the drain work request and
2790 * completion.
2791 *
2792 * allocate the CQ using ib_alloc_cq().
2793 *
2794 * ensure that there are no other contexts that are posting WRs concurrently.
2795 * Otherwise the drain is not guaranteed.
2796 */
2797 void ib_drain_sq(struct ib_qp *qp)
2798 {
2799 if (qp->device->ops.drain_sq)
2800 qp->device->ops.drain_sq(qp);
2801 else
2802 __ib_drain_sq(qp);
2803 trace_cq_drain_complete(qp->send_cq);
2804 }
2805 EXPORT_SYMBOL(ib_drain_sq);
2806
2807 /**
2808 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2809 * application.
2810 * @qp: queue pair to drain
2811 *
2812 * If the device has a provider-specific drain function, then
2813 * call that. Otherwise call the generic drain function
2814 * __ib_drain_rq().
2815 *
2816 * The caller must:
2817 *
2818 * ensure there is room in the CQ and RQ for the drain work request and
2819 * completion.
2820 *
2821 * allocate the CQ using ib_alloc_cq().
2822 *
2823 * ensure that there are no other contexts that are posting WRs concurrently.
2824 * Otherwise the drain is not guaranteed.
2825 */
2826 void ib_drain_rq(struct ib_qp *qp)
2827 {
2828 if (qp->device->ops.drain_rq)
2829 qp->device->ops.drain_rq(qp);
2830 else
2831 __ib_drain_rq(qp);
2832 trace_cq_drain_complete(qp->recv_cq);
2833 }
2834 EXPORT_SYMBOL(ib_drain_rq);
2835
2836 /**
2837 * ib_drain_qp() - Block until all CQEs have been consumed by the
2838 * application on both the RQ and SQ.
2839 * @qp: queue pair to drain
2840 *
2841 * The caller must:
2842 *
2843 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2844 * and completions.
2845 *
2846 * allocate the CQs using ib_alloc_cq().
2847 *
2848 * ensure that there are no other contexts that are posting WRs concurrently.
2849 * Otherwise the drain is not guaranteed.
2850 */
2851 void ib_drain_qp(struct ib_qp *qp)
2852 {
2853 ib_drain_sq(qp);
2854 if (!qp->srq)
2855 ib_drain_rq(qp);
2856 }
2857 EXPORT_SYMBOL(ib_drain_qp);
2858
2859 struct net_device *rdma_alloc_netdev(struct ib_device *device, u8 port_num,
2860 enum rdma_netdev_t type, const char *name,
2861 unsigned char name_assign_type,
2862 void (*setup)(struct net_device *))
2863 {
2864 struct rdma_netdev_alloc_params params;
2865 struct net_device *netdev;
2866 int rc;
2867
2868 if (!device->ops.rdma_netdev_get_params)
2869 return ERR_PTR(-EOPNOTSUPP);
2870
2871 rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2872 &params);
2873 if (rc)
2874 return ERR_PTR(rc);
2875
2876 netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type,
2877 setup, params.txqs, params.rxqs);
2878 if (!netdev)
2879 return ERR_PTR(-ENOMEM);
2880
2881 return netdev;
2882 }
2883 EXPORT_SYMBOL(rdma_alloc_netdev);
2884
2885 int rdma_init_netdev(struct ib_device *device, u8 port_num,
2886 enum rdma_netdev_t type, const char *name,
2887 unsigned char name_assign_type,
2888 void (*setup)(struct net_device *),
2889 struct net_device *netdev)
2890 {
2891 struct rdma_netdev_alloc_params params;
2892 int rc;
2893
2894 if (!device->ops.rdma_netdev_get_params)
2895 return -EOPNOTSUPP;
2896
2897 rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2898 &params);
2899 if (rc)
2900 return rc;
2901
2902 return params.initialize_rdma_netdev(device, port_num,
2903 netdev, params.param);
2904 }
2905 EXPORT_SYMBOL(rdma_init_netdev);
2906
2907 void __rdma_block_iter_start(struct ib_block_iter *biter,
2908 struct scatterlist *sglist, unsigned int nents,
2909 unsigned long pgsz)
2910 {
2911 memset(biter, 0, sizeof(struct ib_block_iter));
2912 biter->__sg = sglist;
2913 biter->__sg_nents = nents;
2914
2915 /* Driver provides best block size to use */
2916 biter->__pg_bit = __fls(pgsz);
2917 }
2918 EXPORT_SYMBOL(__rdma_block_iter_start);
2919
2920 bool __rdma_block_iter_next(struct ib_block_iter *biter)
2921 {
2922 unsigned int block_offset;
2923
2924 if (!biter->__sg_nents || !biter->__sg)
2925 return false;
2926
2927 biter->__dma_addr = sg_dma_address(biter->__sg) + biter->__sg_advance;
2928 block_offset = biter->__dma_addr & (BIT_ULL(biter->__pg_bit) - 1);
2929 biter->__sg_advance += BIT_ULL(biter->__pg_bit) - block_offset;
2930
2931 if (biter->__sg_advance >= sg_dma_len(biter->__sg)) {
2932 biter->__sg_advance = 0;
2933 biter->__sg = sg_next(biter->__sg);
2934 biter->__sg_nents--;
2935 }
2936
2937 return true;
2938 }
2939 EXPORT_SYMBOL(__rdma_block_iter_next);
Linux with added FPC-III support