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