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/ib_umem.h>
53 #include <rdma/rw.h>
54 #include <rdma/lag.h>
55
56 #include "rdma_core.h"
57 #include "core_priv.h"
58 #include <trace/events/rdma_core.h>
59
60 static int ib_resolve_eth_dmac(struct ib_device *device,
61 struct rdma_ah_attr *ah_attr);
62
63 static const char * const ib_events[] = {
64 [IB_EVENT_CQ_ERR] = "CQ error",
65 [IB_EVENT_QP_FATAL] = "QP fatal error",
66 [IB_EVENT_QP_REQ_ERR] = "QP request error",
67 [IB_EVENT_QP_ACCESS_ERR] = "QP access error",
68 [IB_EVENT_COMM_EST] = "communication established",
69 [IB_EVENT_SQ_DRAINED] = "send queue drained",
70 [IB_EVENT_PATH_MIG] = "path migration successful",
71 [IB_EVENT_PATH_MIG_ERR] = "path migration error",
72 [IB_EVENT_DEVICE_FATAL] = "device fatal error",
73 [IB_EVENT_PORT_ACTIVE] = "port active",
74 [IB_EVENT_PORT_ERR] = "port error",
75 [IB_EVENT_LID_CHANGE] = "LID change",
76 [IB_EVENT_PKEY_CHANGE] = "P_key change",
77 [IB_EVENT_SM_CHANGE] = "SM change",
78 [IB_EVENT_SRQ_ERR] = "SRQ error",
79 [IB_EVENT_SRQ_LIMIT_REACHED] = "SRQ limit reached",
80 [IB_EVENT_QP_LAST_WQE_REACHED] = "last WQE reached",
81 [IB_EVENT_CLIENT_REREGISTER] = "client reregister",
82 [IB_EVENT_GID_CHANGE] = "GID changed",
83 [IB_EVENT_DEVICE_SPEED_CHANGE] = "device speed change"
84 };
85
ib_event_msg(enum ib_event_type event)86 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
87 {
88 size_t index = event;
89
90 return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
91 ib_events[index] : "unrecognized event";
92 }
93 EXPORT_SYMBOL(ib_event_msg);
94
95 static const char * const wc_statuses[] = {
96 [IB_WC_SUCCESS] = "success",
97 [IB_WC_LOC_LEN_ERR] = "local length error",
98 [IB_WC_LOC_QP_OP_ERR] = "local QP operation error",
99 [IB_WC_LOC_EEC_OP_ERR] = "local EE context operation error",
100 [IB_WC_LOC_PROT_ERR] = "local protection error",
101 [IB_WC_WR_FLUSH_ERR] = "WR flushed",
102 [IB_WC_MW_BIND_ERR] = "memory bind operation error",
103 [IB_WC_BAD_RESP_ERR] = "bad response error",
104 [IB_WC_LOC_ACCESS_ERR] = "local access error",
105 [IB_WC_REM_INV_REQ_ERR] = "remote invalid request error",
106 [IB_WC_REM_ACCESS_ERR] = "remote access error",
107 [IB_WC_REM_OP_ERR] = "remote operation error",
108 [IB_WC_RETRY_EXC_ERR] = "transport retry counter exceeded",
109 [IB_WC_RNR_RETRY_EXC_ERR] = "RNR retry counter exceeded",
110 [IB_WC_LOC_RDD_VIOL_ERR] = "local RDD violation error",
111 [IB_WC_REM_INV_RD_REQ_ERR] = "remote invalid RD request",
112 [IB_WC_REM_ABORT_ERR] = "operation aborted",
113 [IB_WC_INV_EECN_ERR] = "invalid EE context number",
114 [IB_WC_INV_EEC_STATE_ERR] = "invalid EE context state",
115 [IB_WC_FATAL_ERR] = "fatal error",
116 [IB_WC_RESP_TIMEOUT_ERR] = "response timeout error",
117 [IB_WC_GENERAL_ERR] = "general error",
118 };
119
ib_wc_status_msg(enum ib_wc_status status)120 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
121 {
122 size_t index = status;
123
124 return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
125 wc_statuses[index] : "unrecognized status";
126 }
127 EXPORT_SYMBOL(ib_wc_status_msg);
128
ib_rate_to_mult(enum ib_rate rate)129 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
130 {
131 switch (rate) {
132 case IB_RATE_2_5_GBPS: return 1;
133 case IB_RATE_5_GBPS: return 2;
134 case IB_RATE_10_GBPS: return 4;
135 case IB_RATE_20_GBPS: return 8;
136 case IB_RATE_30_GBPS: return 12;
137 case IB_RATE_40_GBPS: return 16;
138 case IB_RATE_60_GBPS: return 24;
139 case IB_RATE_80_GBPS: return 32;
140 case IB_RATE_120_GBPS: return 48;
141 case IB_RATE_14_GBPS: return 6;
142 case IB_RATE_56_GBPS: return 22;
143 case IB_RATE_112_GBPS: return 45;
144 case IB_RATE_168_GBPS: return 67;
145 case IB_RATE_25_GBPS: return 10;
146 case IB_RATE_100_GBPS: return 40;
147 case IB_RATE_200_GBPS: return 80;
148 case IB_RATE_300_GBPS: return 120;
149 case IB_RATE_28_GBPS: return 11;
150 case IB_RATE_50_GBPS: return 20;
151 case IB_RATE_400_GBPS: return 160;
152 case IB_RATE_600_GBPS: return 240;
153 case IB_RATE_800_GBPS: return 320;
154 case IB_RATE_1600_GBPS: return 640;
155 default: return -1;
156 }
157 }
158 EXPORT_SYMBOL(ib_rate_to_mult);
159
mult_to_ib_rate(int mult)160 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
161 {
162 switch (mult) {
163 case 1: return IB_RATE_2_5_GBPS;
164 case 2: return IB_RATE_5_GBPS;
165 case 4: return IB_RATE_10_GBPS;
166 case 8: return IB_RATE_20_GBPS;
167 case 12: return IB_RATE_30_GBPS;
168 case 16: return IB_RATE_40_GBPS;
169 case 24: return IB_RATE_60_GBPS;
170 case 32: return IB_RATE_80_GBPS;
171 case 48: return IB_RATE_120_GBPS;
172 case 6: return IB_RATE_14_GBPS;
173 case 22: return IB_RATE_56_GBPS;
174 case 45: return IB_RATE_112_GBPS;
175 case 67: return IB_RATE_168_GBPS;
176 case 10: return IB_RATE_25_GBPS;
177 case 40: return IB_RATE_100_GBPS;
178 case 80: return IB_RATE_200_GBPS;
179 case 120: return IB_RATE_300_GBPS;
180 case 11: return IB_RATE_28_GBPS;
181 case 20: return IB_RATE_50_GBPS;
182 case 160: return IB_RATE_400_GBPS;
183 case 240: return IB_RATE_600_GBPS;
184 case 320: return IB_RATE_800_GBPS;
185 case 640: return IB_RATE_1600_GBPS;
186 default: return IB_RATE_PORT_CURRENT;
187 }
188 }
189 EXPORT_SYMBOL(mult_to_ib_rate);
190
ib_rate_to_mbps(enum ib_rate rate)191 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
192 {
193 switch (rate) {
194 case IB_RATE_2_5_GBPS: return 2500;
195 case IB_RATE_5_GBPS: return 5000;
196 case IB_RATE_10_GBPS: return 10000;
197 case IB_RATE_20_GBPS: return 20000;
198 case IB_RATE_30_GBPS: return 30000;
199 case IB_RATE_40_GBPS: return 40000;
200 case IB_RATE_60_GBPS: return 60000;
201 case IB_RATE_80_GBPS: return 80000;
202 case IB_RATE_120_GBPS: return 120000;
203 case IB_RATE_14_GBPS: return 14062;
204 case IB_RATE_56_GBPS: return 56250;
205 case IB_RATE_112_GBPS: return 112500;
206 case IB_RATE_168_GBPS: return 168750;
207 case IB_RATE_25_GBPS: return 25781;
208 case IB_RATE_100_GBPS: return 103125;
209 case IB_RATE_200_GBPS: return 206250;
210 case IB_RATE_300_GBPS: return 309375;
211 case IB_RATE_28_GBPS: return 28125;
212 case IB_RATE_50_GBPS: return 53125;
213 case IB_RATE_400_GBPS: return 425000;
214 case IB_RATE_600_GBPS: return 637500;
215 case IB_RATE_800_GBPS: return 850000;
216 case IB_RATE_1600_GBPS: return 1700000;
217 default: return -1;
218 }
219 }
220 EXPORT_SYMBOL(ib_rate_to_mbps);
221
222 struct ib_speed_attr {
223 const char *str;
224 int speed;
225 };
226
227 #define IB_SPEED_ATTR(speed_type, _str, _speed) \
228 [speed_type] = {.str = _str, .speed = _speed}
229
230 static const struct ib_speed_attr ib_speed_attrs[] = {
231 IB_SPEED_ATTR(IB_SPEED_SDR, " SDR", 25),
232 IB_SPEED_ATTR(IB_SPEED_DDR, " DDR", 50),
233 IB_SPEED_ATTR(IB_SPEED_QDR, " QDR", 100),
234 IB_SPEED_ATTR(IB_SPEED_FDR10, " FDR10", 100),
235 IB_SPEED_ATTR(IB_SPEED_FDR, " FDR", 140),
236 IB_SPEED_ATTR(IB_SPEED_EDR, " EDR", 250),
237 IB_SPEED_ATTR(IB_SPEED_HDR, " HDR", 500),
238 IB_SPEED_ATTR(IB_SPEED_NDR, " NDR", 1000),
239 IB_SPEED_ATTR(IB_SPEED_XDR, " XDR", 2000),
240 };
241
ib_port_attr_to_speed_info(struct ib_port_attr * attr,struct ib_port_speed_info * speed_info)242 int ib_port_attr_to_speed_info(struct ib_port_attr *attr,
243 struct ib_port_speed_info *speed_info)
244 {
245 int speed_idx = attr->active_speed;
246
247 switch (attr->active_speed) {
248 case IB_SPEED_DDR:
249 case IB_SPEED_QDR:
250 case IB_SPEED_FDR10:
251 case IB_SPEED_FDR:
252 case IB_SPEED_EDR:
253 case IB_SPEED_HDR:
254 case IB_SPEED_NDR:
255 case IB_SPEED_XDR:
256 case IB_SPEED_SDR:
257 break;
258 default:
259 speed_idx = IB_SPEED_SDR; /* Default to SDR for invalid rates */
260 break;
261 }
262
263 speed_info->str = ib_speed_attrs[speed_idx].str;
264 speed_info->rate = ib_speed_attrs[speed_idx].speed;
265 speed_info->rate *= ib_width_enum_to_int(attr->active_width);
266 if (speed_info->rate < 0)
267 return -EINVAL;
268
269 return 0;
270 }
271 EXPORT_SYMBOL(ib_port_attr_to_speed_info);
272
273 __attribute_const__ enum rdma_transport_type
rdma_node_get_transport(unsigned int node_type)274 rdma_node_get_transport(unsigned int node_type)
275 {
276
277 if (node_type == RDMA_NODE_USNIC)
278 return RDMA_TRANSPORT_USNIC;
279 if (node_type == RDMA_NODE_USNIC_UDP)
280 return RDMA_TRANSPORT_USNIC_UDP;
281 if (node_type == RDMA_NODE_RNIC)
282 return RDMA_TRANSPORT_IWARP;
283 if (node_type == RDMA_NODE_UNSPECIFIED)
284 return RDMA_TRANSPORT_UNSPECIFIED;
285
286 return RDMA_TRANSPORT_IB;
287 }
288 EXPORT_SYMBOL(rdma_node_get_transport);
289
rdma_port_get_link_layer(struct ib_device * device,u32 port_num)290 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device,
291 u32 port_num)
292 {
293 enum rdma_transport_type lt;
294 if (device->ops.get_link_layer)
295 return device->ops.get_link_layer(device, port_num);
296
297 lt = rdma_node_get_transport(device->node_type);
298 if (lt == RDMA_TRANSPORT_IB)
299 return IB_LINK_LAYER_INFINIBAND;
300
301 return IB_LINK_LAYER_ETHERNET;
302 }
303 EXPORT_SYMBOL(rdma_port_get_link_layer);
304
305 /* Protection domains */
306
307 /**
308 * __ib_alloc_pd - Allocates an unused protection domain.
309 * @device: The device on which to allocate the protection domain.
310 * @flags: protection domain flags
311 * @caller: caller's build-time module name
312 *
313 * A protection domain object provides an association between QPs, shared
314 * receive queues, address handles, memory regions, and memory windows.
315 *
316 * Every PD has a local_dma_lkey which can be used as the lkey value for local
317 * memory operations.
318 */
__ib_alloc_pd(struct ib_device * device,unsigned int flags,const char * caller)319 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
320 const char *caller)
321 {
322 struct ib_pd *pd;
323 int mr_access_flags = 0;
324 int ret;
325
326 pd = rdma_zalloc_drv_obj(device, ib_pd);
327 if (!pd)
328 return ERR_PTR(-ENOMEM);
329
330 pd->device = device;
331 pd->flags = flags;
332
333 rdma_restrack_new(&pd->res, RDMA_RESTRACK_PD);
334 rdma_restrack_set_name(&pd->res, caller);
335
336 ret = device->ops.alloc_pd(pd, NULL);
337 if (ret) {
338 rdma_restrack_put(&pd->res);
339 kfree(pd);
340 return ERR_PTR(ret);
341 }
342 rdma_restrack_add(&pd->res);
343
344 if (device->attrs.kernel_cap_flags & IBK_LOCAL_DMA_LKEY)
345 pd->local_dma_lkey = device->local_dma_lkey;
346 else
347 mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
348
349 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
350 pr_warn("%s: enabling unsafe global rkey\n", caller);
351 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
352 }
353
354 if (mr_access_flags) {
355 struct ib_mr *mr;
356
357 mr = pd->device->ops.get_dma_mr(pd, mr_access_flags);
358 if (IS_ERR(mr)) {
359 ib_dealloc_pd(pd);
360 return ERR_CAST(mr);
361 }
362
363 mr->device = pd->device;
364 mr->pd = pd;
365 mr->type = IB_MR_TYPE_DMA;
366 mr->uobject = NULL;
367 mr->need_inval = false;
368
369 pd->__internal_mr = mr;
370
371 if (!(device->attrs.kernel_cap_flags & IBK_LOCAL_DMA_LKEY))
372 pd->local_dma_lkey = pd->__internal_mr->lkey;
373
374 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
375 pd->unsafe_global_rkey = pd->__internal_mr->rkey;
376 }
377
378 return pd;
379 }
380 EXPORT_SYMBOL(__ib_alloc_pd);
381
382 /**
383 * ib_dealloc_pd_user - Deallocates a protection domain.
384 * @pd: The protection domain to deallocate.
385 * @udata: Valid user data or NULL for kernel object
386 *
387 * It is an error to call this function while any resources in the pd still
388 * exist. The caller is responsible to synchronously destroy them and
389 * guarantee no new allocations will happen.
390 */
ib_dealloc_pd_user(struct ib_pd * pd,struct ib_udata * udata)391 int ib_dealloc_pd_user(struct ib_pd *pd, struct ib_udata *udata)
392 {
393 int ret;
394
395 rdma_restrack_begin_del(&pd->res);
396 if (pd->__internal_mr) {
397 ret = pd->device->ops.dereg_mr(pd->__internal_mr, NULL);
398 WARN_ON(ret);
399 pd->__internal_mr = NULL;
400 }
401
402 ret = pd->device->ops.dealloc_pd(pd, udata);
403 if (ret) {
404 rdma_restrack_abort_del(&pd->res);
405 return ret;
406 }
407
408 rdma_restrack_commit_del(&pd->res);
409 kfree(pd);
410 return ret;
411 }
412 EXPORT_SYMBOL(ib_dealloc_pd_user);
413
414 /* Address handles */
415
416 /**
417 * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
418 * @dest: Pointer to destination ah_attr. Contents of the destination
419 * pointer is assumed to be invalid and attribute are overwritten.
420 * @src: Pointer to source ah_attr.
421 */
rdma_copy_ah_attr(struct rdma_ah_attr * dest,const struct rdma_ah_attr * src)422 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
423 const struct rdma_ah_attr *src)
424 {
425 *dest = *src;
426 if (dest->grh.sgid_attr)
427 rdma_hold_gid_attr(dest->grh.sgid_attr);
428 }
429 EXPORT_SYMBOL(rdma_copy_ah_attr);
430
431 /**
432 * rdma_replace_ah_attr - Replace valid ah_attr with new one.
433 * @old: Pointer to existing ah_attr which needs to be replaced.
434 * old is assumed to be valid or zero'd
435 * @new: Pointer to the new ah_attr.
436 *
437 * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
438 * old the ah_attr is valid; after that it copies the new attribute and holds
439 * the reference to the replaced ah_attr.
440 */
rdma_replace_ah_attr(struct rdma_ah_attr * old,const struct rdma_ah_attr * new)441 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
442 const struct rdma_ah_attr *new)
443 {
444 rdma_destroy_ah_attr(old);
445 *old = *new;
446 if (old->grh.sgid_attr)
447 rdma_hold_gid_attr(old->grh.sgid_attr);
448 }
449 EXPORT_SYMBOL(rdma_replace_ah_attr);
450
451 /**
452 * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
453 * @dest: Pointer to destination ah_attr to copy to.
454 * dest is assumed to be valid or zero'd
455 * @src: Pointer to the new ah_attr.
456 *
457 * rdma_move_ah_attr() first releases any reference in the destination ah_attr
458 * if it is valid. This also transfers ownership of internal references from
459 * src to dest, making src invalid in the process. No new reference of the src
460 * ah_attr is taken.
461 */
rdma_move_ah_attr(struct rdma_ah_attr * dest,struct rdma_ah_attr * src)462 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
463 {
464 rdma_destroy_ah_attr(dest);
465 *dest = *src;
466 src->grh.sgid_attr = NULL;
467 }
468 EXPORT_SYMBOL(rdma_move_ah_attr);
469
470 /*
471 * Validate that the rdma_ah_attr is valid for the device before passing it
472 * off to the driver.
473 */
rdma_check_ah_attr(struct ib_device * device,struct rdma_ah_attr * ah_attr)474 static int rdma_check_ah_attr(struct ib_device *device,
475 struct rdma_ah_attr *ah_attr)
476 {
477 if (!rdma_is_port_valid(device, ah_attr->port_num))
478 return -EINVAL;
479
480 if ((rdma_is_grh_required(device, ah_attr->port_num) ||
481 ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
482 !(ah_attr->ah_flags & IB_AH_GRH))
483 return -EINVAL;
484
485 if (ah_attr->grh.sgid_attr) {
486 /*
487 * Make sure the passed sgid_attr is consistent with the
488 * parameters
489 */
490 if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
491 ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
492 return -EINVAL;
493 }
494 return 0;
495 }
496
497 /*
498 * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
499 * On success the caller is responsible to call rdma_unfill_sgid_attr().
500 */
rdma_fill_sgid_attr(struct ib_device * device,struct rdma_ah_attr * ah_attr,const struct ib_gid_attr ** old_sgid_attr)501 static int rdma_fill_sgid_attr(struct ib_device *device,
502 struct rdma_ah_attr *ah_attr,
503 const struct ib_gid_attr **old_sgid_attr)
504 {
505 const struct ib_gid_attr *sgid_attr;
506 struct ib_global_route *grh;
507 int ret;
508
509 *old_sgid_attr = ah_attr->grh.sgid_attr;
510
511 ret = rdma_check_ah_attr(device, ah_attr);
512 if (ret)
513 return ret;
514
515 if (!(ah_attr->ah_flags & IB_AH_GRH))
516 return 0;
517
518 grh = rdma_ah_retrieve_grh(ah_attr);
519 if (grh->sgid_attr)
520 return 0;
521
522 sgid_attr =
523 rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
524 if (IS_ERR(sgid_attr))
525 return PTR_ERR(sgid_attr);
526
527 /* Move ownerhip of the kref into the ah_attr */
528 grh->sgid_attr = sgid_attr;
529 return 0;
530 }
531
rdma_unfill_sgid_attr(struct rdma_ah_attr * ah_attr,const struct ib_gid_attr * old_sgid_attr)532 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
533 const struct ib_gid_attr *old_sgid_attr)
534 {
535 /*
536 * Fill didn't change anything, the caller retains ownership of
537 * whatever it passed
538 */
539 if (ah_attr->grh.sgid_attr == old_sgid_attr)
540 return;
541
542 /*
543 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
544 * doesn't see any change in the rdma_ah_attr. If we get here
545 * old_sgid_attr is NULL.
546 */
547 rdma_destroy_ah_attr(ah_attr);
548 }
549
550 static const struct ib_gid_attr *
rdma_update_sgid_attr(struct rdma_ah_attr * ah_attr,const struct ib_gid_attr * old_attr)551 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
552 const struct ib_gid_attr *old_attr)
553 {
554 if (old_attr)
555 rdma_put_gid_attr(old_attr);
556 if (ah_attr->ah_flags & IB_AH_GRH) {
557 rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
558 return ah_attr->grh.sgid_attr;
559 }
560 return NULL;
561 }
562
_rdma_create_ah(struct ib_pd * pd,struct rdma_ah_attr * ah_attr,u32 flags,struct ib_udata * udata,struct net_device * xmit_slave)563 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
564 struct rdma_ah_attr *ah_attr,
565 u32 flags,
566 struct ib_udata *udata,
567 struct net_device *xmit_slave)
568 {
569 struct rdma_ah_init_attr init_attr = {};
570 struct ib_device *device = pd->device;
571 struct ib_ah *ah;
572 int ret;
573
574 might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE);
575
576 if (!udata && !device->ops.create_ah)
577 return ERR_PTR(-EOPNOTSUPP);
578
579 ah = rdma_zalloc_drv_obj_gfp(
580 device, ib_ah,
581 (flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC);
582 if (!ah)
583 return ERR_PTR(-ENOMEM);
584
585 ah->device = device;
586 ah->pd = pd;
587 ah->type = ah_attr->type;
588 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
589 init_attr.ah_attr = ah_attr;
590 init_attr.flags = flags;
591 init_attr.xmit_slave = xmit_slave;
592
593 if (udata)
594 ret = device->ops.create_user_ah(ah, &init_attr, udata);
595 else
596 ret = device->ops.create_ah(ah, &init_attr, NULL);
597 if (ret) {
598 if (ah->sgid_attr)
599 rdma_put_gid_attr(ah->sgid_attr);
600 kfree(ah);
601 return ERR_PTR(ret);
602 }
603
604 atomic_inc(&pd->usecnt);
605 return ah;
606 }
607
608 /**
609 * rdma_create_ah - Creates an address handle for the
610 * given address vector.
611 * @pd: The protection domain associated with the address handle.
612 * @ah_attr: The attributes of the address vector.
613 * @flags: Create address handle flags (see enum rdma_create_ah_flags).
614 *
615 * It returns 0 on success and returns appropriate error code on error.
616 * The address handle is used to reference a local or global destination
617 * in all UD QP post sends.
618 */
rdma_create_ah(struct ib_pd * pd,struct rdma_ah_attr * ah_attr,u32 flags)619 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
620 u32 flags)
621 {
622 const struct ib_gid_attr *old_sgid_attr;
623 struct net_device *slave;
624 struct ib_ah *ah;
625 int ret;
626
627 ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
628 if (ret)
629 return ERR_PTR(ret);
630 slave = rdma_lag_get_ah_roce_slave(pd->device, ah_attr,
631 (flags & RDMA_CREATE_AH_SLEEPABLE) ?
632 GFP_KERNEL : GFP_ATOMIC);
633 if (IS_ERR(slave)) {
634 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
635 return ERR_CAST(slave);
636 }
637 ah = _rdma_create_ah(pd, ah_attr, flags, NULL, slave);
638 rdma_lag_put_ah_roce_slave(slave);
639 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
640 return ah;
641 }
642 EXPORT_SYMBOL(rdma_create_ah);
643
644 /**
645 * rdma_create_user_ah - Creates an address handle for the
646 * given address vector.
647 * It resolves destination mac address for ah attribute of RoCE type.
648 * @pd: The protection domain associated with the address handle.
649 * @ah_attr: The attributes of the address vector.
650 * @udata: pointer to user's input output buffer information need by
651 * provider driver.
652 *
653 * It returns 0 on success and returns appropriate error code on error.
654 * The address handle is used to reference a local or global destination
655 * in all UD QP post sends.
656 */
rdma_create_user_ah(struct ib_pd * pd,struct rdma_ah_attr * ah_attr,struct ib_udata * udata)657 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
658 struct rdma_ah_attr *ah_attr,
659 struct ib_udata *udata)
660 {
661 const struct ib_gid_attr *old_sgid_attr;
662 struct ib_ah *ah;
663 int err;
664
665 err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
666 if (err)
667 return ERR_PTR(err);
668
669 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
670 err = ib_resolve_eth_dmac(pd->device, ah_attr);
671 if (err) {
672 ah = ERR_PTR(err);
673 goto out;
674 }
675 }
676
677 ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE,
678 udata, NULL);
679
680 out:
681 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
682 return ah;
683 }
684 EXPORT_SYMBOL(rdma_create_user_ah);
685
ib_get_rdma_header_version(const union rdma_network_hdr * hdr)686 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
687 {
688 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
689 struct iphdr ip4h_checked;
690 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
691
692 /* If it's IPv6, the version must be 6, otherwise, the first
693 * 20 bytes (before the IPv4 header) are garbled.
694 */
695 if (ip6h->version != 6)
696 return (ip4h->version == 4) ? 4 : 0;
697 /* version may be 6 or 4 because the first 20 bytes could be garbled */
698
699 /* RoCE v2 requires no options, thus header length
700 * must be 5 words
701 */
702 if (ip4h->ihl != 5)
703 return 6;
704
705 /* Verify checksum.
706 * We can't write on scattered buffers so we need to copy to
707 * temp buffer.
708 */
709 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
710 ip4h_checked.check = 0;
711 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
712 /* if IPv4 header checksum is OK, believe it */
713 if (ip4h->check == ip4h_checked.check)
714 return 4;
715 return 6;
716 }
717 EXPORT_SYMBOL(ib_get_rdma_header_version);
718
ib_get_net_type_by_grh(struct ib_device * device,u32 port_num,const struct ib_grh * grh)719 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
720 u32 port_num,
721 const struct ib_grh *grh)
722 {
723 int grh_version;
724
725 if (rdma_protocol_ib(device, port_num))
726 return RDMA_NETWORK_IB;
727
728 grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
729
730 if (grh_version == 4)
731 return RDMA_NETWORK_IPV4;
732
733 if (grh->next_hdr == IPPROTO_UDP)
734 return RDMA_NETWORK_IPV6;
735
736 return RDMA_NETWORK_ROCE_V1;
737 }
738
739 struct find_gid_index_context {
740 u16 vlan_id;
741 enum ib_gid_type gid_type;
742 };
743
find_gid_index(const union ib_gid * gid,const struct ib_gid_attr * gid_attr,void * context)744 static bool find_gid_index(const union ib_gid *gid,
745 const struct ib_gid_attr *gid_attr,
746 void *context)
747 {
748 struct find_gid_index_context *ctx = context;
749 u16 vlan_id = 0xffff;
750 int ret;
751
752 if (ctx->gid_type != gid_attr->gid_type)
753 return false;
754
755 ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
756 if (ret)
757 return false;
758
759 return ctx->vlan_id == vlan_id;
760 }
761
762 static const struct ib_gid_attr *
get_sgid_attr_from_eth(struct ib_device * device,u32 port_num,u16 vlan_id,const union ib_gid * sgid,enum ib_gid_type gid_type)763 get_sgid_attr_from_eth(struct ib_device *device, u32 port_num,
764 u16 vlan_id, const union ib_gid *sgid,
765 enum ib_gid_type gid_type)
766 {
767 struct find_gid_index_context context = {.vlan_id = vlan_id,
768 .gid_type = gid_type};
769
770 return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
771 &context);
772 }
773
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)774 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
775 enum rdma_network_type net_type,
776 union ib_gid *sgid, union ib_gid *dgid)
777 {
778 struct sockaddr_in src_in;
779 struct sockaddr_in dst_in;
780 __be32 src_saddr, dst_saddr;
781
782 if (!sgid || !dgid)
783 return -EINVAL;
784
785 if (net_type == RDMA_NETWORK_IPV4) {
786 memcpy(&src_in.sin_addr.s_addr,
787 &hdr->roce4grh.saddr, 4);
788 memcpy(&dst_in.sin_addr.s_addr,
789 &hdr->roce4grh.daddr, 4);
790 src_saddr = src_in.sin_addr.s_addr;
791 dst_saddr = dst_in.sin_addr.s_addr;
792 ipv6_addr_set_v4mapped(src_saddr,
793 (struct in6_addr *)sgid);
794 ipv6_addr_set_v4mapped(dst_saddr,
795 (struct in6_addr *)dgid);
796 return 0;
797 } else if (net_type == RDMA_NETWORK_IPV6 ||
798 net_type == RDMA_NETWORK_IB || net_type == RDMA_NETWORK_ROCE_V1) {
799 *dgid = hdr->ibgrh.dgid;
800 *sgid = hdr->ibgrh.sgid;
801 return 0;
802 } else {
803 return -EINVAL;
804 }
805 }
806 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
807
808 /* Resolve destination mac address and hop limit for unicast destination
809 * GID entry, considering the source GID entry as well.
810 * ah_attribute must have valid port_num, sgid_index.
811 */
ib_resolve_unicast_gid_dmac(struct ib_device * device,struct rdma_ah_attr * ah_attr)812 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
813 struct rdma_ah_attr *ah_attr)
814 {
815 struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
816 const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
817 int hop_limit = 0xff;
818 int ret = 0;
819
820 /* If destination is link local and source GID is RoCEv1,
821 * IP stack is not used.
822 */
823 if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
824 sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
825 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
826 ah_attr->roce.dmac);
827 return ret;
828 }
829
830 ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
831 ah_attr->roce.dmac,
832 sgid_attr, &hop_limit);
833
834 grh->hop_limit = hop_limit;
835 return ret;
836 }
837
838 /*
839 * This function initializes address handle attributes from the incoming packet.
840 * Incoming packet has dgid of the receiver node on which this code is
841 * getting executed and, sgid contains the GID of the sender.
842 *
843 * When resolving mac address of destination, the arrived dgid is used
844 * as sgid and, sgid is used as dgid because sgid contains destinations
845 * GID whom to respond to.
846 *
847 * On success the caller is responsible to call rdma_destroy_ah_attr on the
848 * attr.
849 */
ib_init_ah_attr_from_wc(struct ib_device * device,u32 port_num,const struct ib_wc * wc,const struct ib_grh * grh,struct rdma_ah_attr * ah_attr)850 int ib_init_ah_attr_from_wc(struct ib_device *device, u32 port_num,
851 const struct ib_wc *wc, const struct ib_grh *grh,
852 struct rdma_ah_attr *ah_attr)
853 {
854 u32 flow_class;
855 int ret;
856 enum rdma_network_type net_type = RDMA_NETWORK_IB;
857 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
858 const struct ib_gid_attr *sgid_attr;
859 int hoplimit = 0xff;
860 union ib_gid dgid;
861 union ib_gid sgid;
862
863 might_sleep();
864
865 memset(ah_attr, 0, sizeof *ah_attr);
866 ah_attr->type = rdma_ah_find_type(device, port_num);
867 if (rdma_cap_eth_ah(device, port_num)) {
868 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
869 net_type = wc->network_hdr_type;
870 else
871 net_type = ib_get_net_type_by_grh(device, port_num, grh);
872 gid_type = ib_network_to_gid_type(net_type);
873 }
874 ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
875 &sgid, &dgid);
876 if (ret)
877 return ret;
878
879 rdma_ah_set_sl(ah_attr, wc->sl);
880 rdma_ah_set_port_num(ah_attr, port_num);
881
882 if (rdma_protocol_roce(device, port_num)) {
883 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
884 wc->vlan_id : 0xffff;
885
886 if (!(wc->wc_flags & IB_WC_GRH))
887 return -EPROTOTYPE;
888
889 sgid_attr = get_sgid_attr_from_eth(device, port_num,
890 vlan_id, &dgid,
891 gid_type);
892 if (IS_ERR(sgid_attr))
893 return PTR_ERR(sgid_attr);
894
895 flow_class = be32_to_cpu(grh->version_tclass_flow);
896 rdma_move_grh_sgid_attr(ah_attr,
897 &sgid,
898 flow_class & 0xFFFFF,
899 hoplimit,
900 (flow_class >> 20) & 0xFF,
901 sgid_attr);
902
903 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
904 if (ret)
905 rdma_destroy_ah_attr(ah_attr);
906
907 return ret;
908 } else {
909 rdma_ah_set_dlid(ah_attr, wc->slid);
910 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
911
912 if ((wc->wc_flags & IB_WC_GRH) == 0)
913 return 0;
914
915 if (dgid.global.interface_id !=
916 cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
917 sgid_attr = rdma_find_gid_by_port(
918 device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
919 } else
920 sgid_attr = rdma_get_gid_attr(device, port_num, 0);
921
922 if (IS_ERR(sgid_attr))
923 return PTR_ERR(sgid_attr);
924 flow_class = be32_to_cpu(grh->version_tclass_flow);
925 rdma_move_grh_sgid_attr(ah_attr,
926 &sgid,
927 flow_class & 0xFFFFF,
928 hoplimit,
929 (flow_class >> 20) & 0xFF,
930 sgid_attr);
931
932 return 0;
933 }
934 }
935 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
936
937 /**
938 * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
939 * of the reference
940 *
941 * @attr: Pointer to AH attribute structure
942 * @dgid: Destination GID
943 * @flow_label: Flow label
944 * @hop_limit: Hop limit
945 * @traffic_class: traffic class
946 * @sgid_attr: Pointer to SGID attribute
947 *
948 * This takes ownership of the sgid_attr reference. The caller must ensure
949 * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
950 * calling this function.
951 */
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)952 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
953 u32 flow_label, u8 hop_limit, u8 traffic_class,
954 const struct ib_gid_attr *sgid_attr)
955 {
956 rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
957 traffic_class);
958 attr->grh.sgid_attr = sgid_attr;
959 }
960 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
961
962 /**
963 * rdma_destroy_ah_attr - Release reference to SGID attribute of
964 * ah attribute.
965 * @ah_attr: Pointer to ah attribute
966 *
967 * Release reference to the SGID attribute of the ah attribute if it is
968 * non NULL. It is safe to call this multiple times, and safe to call it on
969 * a zero initialized ah_attr.
970 */
rdma_destroy_ah_attr(struct rdma_ah_attr * ah_attr)971 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
972 {
973 if (ah_attr->grh.sgid_attr) {
974 rdma_put_gid_attr(ah_attr->grh.sgid_attr);
975 ah_attr->grh.sgid_attr = NULL;
976 }
977 }
978 EXPORT_SYMBOL(rdma_destroy_ah_attr);
979
ib_create_ah_from_wc(struct ib_pd * pd,const struct ib_wc * wc,const struct ib_grh * grh,u32 port_num)980 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
981 const struct ib_grh *grh, u32 port_num)
982 {
983 struct rdma_ah_attr ah_attr;
984 struct ib_ah *ah;
985 int ret;
986
987 ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
988 if (ret)
989 return ERR_PTR(ret);
990
991 ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE);
992
993 rdma_destroy_ah_attr(&ah_attr);
994 return ah;
995 }
996 EXPORT_SYMBOL(ib_create_ah_from_wc);
997
rdma_modify_ah(struct ib_ah * ah,struct rdma_ah_attr * ah_attr)998 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
999 {
1000 const struct ib_gid_attr *old_sgid_attr;
1001 int ret;
1002
1003 if (ah->type != ah_attr->type)
1004 return -EINVAL;
1005
1006 ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
1007 if (ret)
1008 return ret;
1009
1010 ret = ah->device->ops.modify_ah ?
1011 ah->device->ops.modify_ah(ah, ah_attr) :
1012 -EOPNOTSUPP;
1013
1014 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
1015 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
1016 return ret;
1017 }
1018 EXPORT_SYMBOL(rdma_modify_ah);
1019
rdma_query_ah(struct ib_ah * ah,struct rdma_ah_attr * ah_attr)1020 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
1021 {
1022 ah_attr->grh.sgid_attr = NULL;
1023
1024 return ah->device->ops.query_ah ?
1025 ah->device->ops.query_ah(ah, ah_attr) :
1026 -EOPNOTSUPP;
1027 }
1028 EXPORT_SYMBOL(rdma_query_ah);
1029
rdma_destroy_ah_user(struct ib_ah * ah,u32 flags,struct ib_udata * udata)1030 int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata)
1031 {
1032 const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
1033 struct ib_pd *pd;
1034 int ret;
1035
1036 might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE);
1037
1038 pd = ah->pd;
1039
1040 ret = ah->device->ops.destroy_ah(ah, flags);
1041 if (ret)
1042 return ret;
1043
1044 atomic_dec(&pd->usecnt);
1045 if (sgid_attr)
1046 rdma_put_gid_attr(sgid_attr);
1047
1048 kfree(ah);
1049 return ret;
1050 }
1051 EXPORT_SYMBOL(rdma_destroy_ah_user);
1052
1053 /* Shared receive queues */
1054
1055 /**
1056 * ib_create_srq_user - Creates a SRQ associated with the specified protection
1057 * domain.
1058 * @pd: The protection domain associated with the SRQ.
1059 * @srq_init_attr: A list of initial attributes required to create the
1060 * SRQ. If SRQ creation succeeds, then the attributes are updated to
1061 * the actual capabilities of the created SRQ.
1062 * @uobject: uobject pointer if this is not a kernel SRQ
1063 * @udata: udata pointer if this is not a kernel SRQ
1064 *
1065 * srq_attr->max_wr and srq_attr->max_sge are read the determine the
1066 * requested size of the SRQ, and set to the actual values allocated
1067 * on return. If ib_create_srq() succeeds, then max_wr and max_sge
1068 * will always be at least as large as the requested values.
1069 */
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)1070 struct ib_srq *ib_create_srq_user(struct ib_pd *pd,
1071 struct ib_srq_init_attr *srq_init_attr,
1072 struct ib_usrq_object *uobject,
1073 struct ib_udata *udata)
1074 {
1075 struct ib_srq *srq;
1076 int ret;
1077
1078 srq = rdma_zalloc_drv_obj(pd->device, ib_srq);
1079 if (!srq)
1080 return ERR_PTR(-ENOMEM);
1081
1082 srq->device = pd->device;
1083 srq->pd = pd;
1084 srq->event_handler = srq_init_attr->event_handler;
1085 srq->srq_context = srq_init_attr->srq_context;
1086 srq->srq_type = srq_init_attr->srq_type;
1087 srq->uobject = uobject;
1088
1089 if (ib_srq_has_cq(srq->srq_type)) {
1090 srq->ext.cq = srq_init_attr->ext.cq;
1091 atomic_inc(&srq->ext.cq->usecnt);
1092 }
1093 if (srq->srq_type == IB_SRQT_XRC) {
1094 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
1095 if (srq->ext.xrc.xrcd)
1096 atomic_inc(&srq->ext.xrc.xrcd->usecnt);
1097 }
1098 atomic_inc(&pd->usecnt);
1099
1100 rdma_restrack_new(&srq->res, RDMA_RESTRACK_SRQ);
1101 rdma_restrack_parent_name(&srq->res, &pd->res);
1102
1103 ret = pd->device->ops.create_srq(srq, srq_init_attr, udata);
1104 if (ret) {
1105 rdma_restrack_put(&srq->res);
1106 atomic_dec(&pd->usecnt);
1107 if (srq->srq_type == IB_SRQT_XRC && srq->ext.xrc.xrcd)
1108 atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1109 if (ib_srq_has_cq(srq->srq_type))
1110 atomic_dec(&srq->ext.cq->usecnt);
1111 kfree(srq);
1112 return ERR_PTR(ret);
1113 }
1114
1115 rdma_restrack_add(&srq->res);
1116
1117 return srq;
1118 }
1119 EXPORT_SYMBOL(ib_create_srq_user);
1120
ib_modify_srq(struct ib_srq * srq,struct ib_srq_attr * srq_attr,enum ib_srq_attr_mask srq_attr_mask)1121 int ib_modify_srq(struct ib_srq *srq,
1122 struct ib_srq_attr *srq_attr,
1123 enum ib_srq_attr_mask srq_attr_mask)
1124 {
1125 return srq->device->ops.modify_srq ?
1126 srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask,
1127 NULL) : -EOPNOTSUPP;
1128 }
1129 EXPORT_SYMBOL(ib_modify_srq);
1130
ib_query_srq(struct ib_srq * srq,struct ib_srq_attr * srq_attr)1131 int ib_query_srq(struct ib_srq *srq,
1132 struct ib_srq_attr *srq_attr)
1133 {
1134 return srq->device->ops.query_srq ?
1135 srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP;
1136 }
1137 EXPORT_SYMBOL(ib_query_srq);
1138
ib_destroy_srq_user(struct ib_srq * srq,struct ib_udata * udata)1139 int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata)
1140 {
1141 int ret;
1142
1143 if (atomic_read(&srq->usecnt))
1144 return -EBUSY;
1145
1146 rdma_restrack_begin_del(&srq->res);
1147
1148 ret = srq->device->ops.destroy_srq(srq, udata);
1149 if (ret) {
1150 rdma_restrack_abort_del(&srq->res);
1151 return ret;
1152 }
1153
1154 atomic_dec(&srq->pd->usecnt);
1155 if (srq->srq_type == IB_SRQT_XRC && srq->ext.xrc.xrcd)
1156 atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1157 if (ib_srq_has_cq(srq->srq_type))
1158 atomic_dec(&srq->ext.cq->usecnt);
1159 rdma_restrack_commit_del(&srq->res);
1160 kfree(srq);
1161
1162 return ret;
1163 }
1164 EXPORT_SYMBOL(ib_destroy_srq_user);
1165
1166 /* Queue pairs */
1167
__ib_qp_event_handler(struct ib_event * event,void * context)1168 static void __ib_qp_event_handler(struct ib_event *event, void *context)
1169 {
1170 struct ib_qp *qp = event->element.qp;
1171
1172 if (event->event == IB_EVENT_QP_LAST_WQE_REACHED)
1173 complete(&qp->srq_completion);
1174 if (qp->registered_event_handler)
1175 qp->registered_event_handler(event, qp->qp_context);
1176 }
1177
__ib_shared_qp_event_handler(struct ib_event * event,void * context)1178 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1179 {
1180 struct ib_qp *qp = context;
1181 unsigned long flags;
1182
1183 spin_lock_irqsave(&qp->device->qp_open_list_lock, flags);
1184 list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1185 if (event->element.qp->event_handler)
1186 event->element.qp->event_handler(event, event->element.qp->qp_context);
1187 spin_unlock_irqrestore(&qp->device->qp_open_list_lock, flags);
1188 }
1189
__ib_open_qp(struct ib_qp * real_qp,void (* event_handler)(struct ib_event *,void *),void * qp_context)1190 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1191 void (*event_handler)(struct ib_event *, void *),
1192 void *qp_context)
1193 {
1194 struct ib_qp *qp;
1195 unsigned long flags;
1196 int err;
1197
1198 qp = kzalloc_obj(*qp);
1199 if (!qp)
1200 return ERR_PTR(-ENOMEM);
1201
1202 qp->real_qp = real_qp;
1203 err = ib_open_shared_qp_security(qp, real_qp->device);
1204 if (err) {
1205 kfree(qp);
1206 return ERR_PTR(err);
1207 }
1208
1209 qp->real_qp = real_qp;
1210 atomic_inc(&real_qp->usecnt);
1211 qp->device = real_qp->device;
1212 qp->event_handler = event_handler;
1213 qp->qp_context = qp_context;
1214 qp->qp_num = real_qp->qp_num;
1215 qp->qp_type = real_qp->qp_type;
1216
1217 spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1218 list_add(&qp->open_list, &real_qp->open_list);
1219 spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1220
1221 return qp;
1222 }
1223
ib_open_qp(struct ib_xrcd * xrcd,struct ib_qp_open_attr * qp_open_attr)1224 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1225 struct ib_qp_open_attr *qp_open_attr)
1226 {
1227 struct ib_qp *qp, *real_qp;
1228
1229 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1230 return ERR_PTR(-EINVAL);
1231
1232 down_read(&xrcd->tgt_qps_rwsem);
1233 real_qp = xa_load(&xrcd->tgt_qps, qp_open_attr->qp_num);
1234 if (!real_qp) {
1235 up_read(&xrcd->tgt_qps_rwsem);
1236 return ERR_PTR(-EINVAL);
1237 }
1238 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1239 qp_open_attr->qp_context);
1240 up_read(&xrcd->tgt_qps_rwsem);
1241 return qp;
1242 }
1243 EXPORT_SYMBOL(ib_open_qp);
1244
create_xrc_qp_user(struct ib_qp * qp,struct ib_qp_init_attr * qp_init_attr)1245 static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp,
1246 struct ib_qp_init_attr *qp_init_attr)
1247 {
1248 struct ib_qp *real_qp = qp;
1249 int err;
1250
1251 qp->event_handler = __ib_shared_qp_event_handler;
1252 qp->qp_context = qp;
1253 qp->pd = NULL;
1254 qp->send_cq = qp->recv_cq = NULL;
1255 qp->srq = NULL;
1256 qp->xrcd = qp_init_attr->xrcd;
1257 atomic_inc(&qp_init_attr->xrcd->usecnt);
1258 INIT_LIST_HEAD(&qp->open_list);
1259
1260 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1261 qp_init_attr->qp_context);
1262 if (IS_ERR(qp))
1263 return qp;
1264
1265 err = xa_err(xa_store(&qp_init_attr->xrcd->tgt_qps, real_qp->qp_num,
1266 real_qp, GFP_KERNEL));
1267 if (err) {
1268 ib_close_qp(qp);
1269 return ERR_PTR(err);
1270 }
1271 return qp;
1272 }
1273
create_qp(struct ib_device * dev,struct ib_pd * pd,struct ib_qp_init_attr * attr,struct uverbs_attr_bundle * uattrs,struct ib_uqp_object * uobj,const char * caller)1274 static struct ib_qp *create_qp(struct ib_device *dev, struct ib_pd *pd,
1275 struct ib_qp_init_attr *attr,
1276 struct uverbs_attr_bundle *uattrs,
1277 struct ib_uqp_object *uobj, const char *caller)
1278 {
1279 struct ib_qp *qp;
1280 int ret;
1281
1282 if (!dev->ops.create_qp)
1283 return ERR_PTR(-EOPNOTSUPP);
1284
1285 qp = rdma_zalloc_drv_obj_numa(dev, ib_qp);
1286 if (!qp)
1287 return ERR_PTR(-ENOMEM);
1288
1289 qp->device = dev;
1290 qp->pd = pd;
1291 qp->uobject = uobj;
1292 qp->real_qp = qp;
1293
1294 qp->qp_type = attr->qp_type;
1295 qp->rwq_ind_tbl = attr->rwq_ind_tbl;
1296 qp->srq = attr->srq;
1297 qp->event_handler = __ib_qp_event_handler;
1298 qp->registered_event_handler = attr->event_handler;
1299 qp->port = attr->port_num;
1300 qp->qp_context = attr->qp_context;
1301
1302 spin_lock_init(&qp->mr_lock);
1303 xa_init(&qp->comp_cntrs);
1304 INIT_LIST_HEAD(&qp->rdma_mrs);
1305 INIT_LIST_HEAD(&qp->sig_mrs);
1306 init_completion(&qp->srq_completion);
1307
1308 qp->send_cq = attr->send_cq;
1309 qp->recv_cq = attr->recv_cq;
1310
1311 rdma_restrack_new(&qp->res, RDMA_RESTRACK_QP);
1312 WARN_ONCE(!uattrs && !caller, "Missing kernel QP owner");
1313 rdma_restrack_set_name(&qp->res, uattrs ? NULL : caller);
1314 ret = dev->ops.create_qp(qp, attr,
1315 uattrs ? &uattrs->driver_udata : NULL);
1316 if (ret)
1317 goto err_create;
1318
1319 /*
1320 * TODO: The mlx4 internally overwrites send_cq and recv_cq.
1321 * Unfortunately, it is not an easy task to fix that driver.
1322 */
1323 qp->send_cq = attr->send_cq;
1324 qp->recv_cq = attr->recv_cq;
1325
1326 ret = ib_create_qp_security(qp, dev);
1327 if (ret)
1328 goto err_security;
1329
1330 rdma_restrack_add(&qp->res);
1331 return qp;
1332
1333 err_security:
1334 qp->device->ops.destroy_qp(
1335 qp, uattrs ? uverbs_get_cleared_udata(uattrs) : NULL);
1336 err_create:
1337 rdma_restrack_put(&qp->res);
1338 xa_destroy(&qp->comp_cntrs);
1339 kfree(qp);
1340 return ERR_PTR(ret);
1341
1342 }
1343
1344 /**
1345 * ib_create_qp_user - Creates a QP associated with the specified protection
1346 * domain.
1347 * @dev: IB device
1348 * @pd: The protection domain associated with the QP.
1349 * @attr: A list of initial attributes required to create the
1350 * QP. If QP creation succeeds, then the attributes are updated to
1351 * the actual capabilities of the created QP.
1352 * @uattrs: User ioctl attributes and udata
1353 * @uobj: uverbs obect
1354 * @caller: caller's build-time module name
1355 */
ib_create_qp_user(struct ib_device * dev,struct ib_pd * pd,struct ib_qp_init_attr * attr,struct uverbs_attr_bundle * uattrs,struct ib_uqp_object * uobj,const char * caller)1356 struct ib_qp *ib_create_qp_user(struct ib_device *dev, struct ib_pd *pd,
1357 struct ib_qp_init_attr *attr,
1358 struct uverbs_attr_bundle *uattrs,
1359 struct ib_uqp_object *uobj, const char *caller)
1360 {
1361 struct ib_qp *qp, *xrc_qp;
1362
1363 if (attr->qp_type == IB_QPT_XRC_TGT)
1364 qp = create_qp(dev, pd, attr, NULL, NULL, caller);
1365 else
1366 qp = create_qp(dev, pd, attr, uattrs, uobj, NULL);
1367 if (attr->qp_type != IB_QPT_XRC_TGT || IS_ERR(qp))
1368 return qp;
1369
1370 xrc_qp = create_xrc_qp_user(qp, attr);
1371 if (IS_ERR(xrc_qp)) {
1372 ib_destroy_qp(qp);
1373 return xrc_qp;
1374 }
1375
1376 xrc_qp->uobject = uobj;
1377 return xrc_qp;
1378 }
1379 EXPORT_SYMBOL(ib_create_qp_user);
1380
ib_qp_usecnt_inc(struct ib_qp * qp)1381 void ib_qp_usecnt_inc(struct ib_qp *qp)
1382 {
1383 if (qp->pd)
1384 atomic_inc(&qp->pd->usecnt);
1385 if (qp->send_cq)
1386 atomic_inc(&qp->send_cq->usecnt);
1387 if (qp->recv_cq)
1388 atomic_inc(&qp->recv_cq->usecnt);
1389 if (qp->srq)
1390 atomic_inc(&qp->srq->usecnt);
1391 if (qp->rwq_ind_tbl)
1392 atomic_inc(&qp->rwq_ind_tbl->usecnt);
1393 }
1394 EXPORT_SYMBOL(ib_qp_usecnt_inc);
1395
ib_qp_usecnt_dec(struct ib_qp * qp)1396 void ib_qp_usecnt_dec(struct ib_qp *qp)
1397 {
1398 if (qp->rwq_ind_tbl)
1399 atomic_dec(&qp->rwq_ind_tbl->usecnt);
1400 if (qp->srq)
1401 atomic_dec(&qp->srq->usecnt);
1402 if (qp->recv_cq)
1403 atomic_dec(&qp->recv_cq->usecnt);
1404 if (qp->send_cq)
1405 atomic_dec(&qp->send_cq->usecnt);
1406 if (qp->pd)
1407 atomic_dec(&qp->pd->usecnt);
1408 }
1409 EXPORT_SYMBOL(ib_qp_usecnt_dec);
1410
ib_create_qp_kernel(struct ib_pd * pd,struct ib_qp_init_attr * qp_init_attr,const char * caller)1411 struct ib_qp *ib_create_qp_kernel(struct ib_pd *pd,
1412 struct ib_qp_init_attr *qp_init_attr,
1413 const char *caller)
1414 {
1415 struct ib_device *device = pd->device;
1416 struct ib_qp *qp;
1417 int ret;
1418
1419 /*
1420 * If the callers is using the RDMA API calculate the resources
1421 * needed for the RDMA READ/WRITE operations.
1422 *
1423 * Note that these callers need to pass in a port number.
1424 */
1425 if (qp_init_attr->cap.max_rdma_ctxs)
1426 rdma_rw_init_qp(device, qp_init_attr);
1427
1428 qp = create_qp(device, pd, qp_init_attr, NULL, NULL, caller);
1429 if (IS_ERR(qp))
1430 return qp;
1431
1432 ib_qp_usecnt_inc(qp);
1433
1434 if (qp_init_attr->cap.max_rdma_ctxs) {
1435 ret = rdma_rw_init_mrs(qp, qp_init_attr);
1436 if (ret)
1437 goto err;
1438 }
1439
1440 /*
1441 * Note: all hw drivers guarantee that max_send_sge is lower than
1442 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1443 * max_send_sge <= max_sge_rd.
1444 */
1445 qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1446 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1447 device->attrs.max_sge_rd);
1448 if (qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN)
1449 qp->integrity_en = true;
1450
1451 return qp;
1452
1453 err:
1454 ib_destroy_qp(qp);
1455 return ERR_PTR(ret);
1456
1457 }
1458 EXPORT_SYMBOL(ib_create_qp_kernel);
1459
1460 static const struct {
1461 int valid;
1462 enum ib_qp_attr_mask req_param[IB_QPT_MAX];
1463 enum ib_qp_attr_mask opt_param[IB_QPT_MAX];
1464 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1465 [IB_QPS_RESET] = {
1466 [IB_QPS_RESET] = { .valid = 1 },
1467 [IB_QPS_INIT] = {
1468 .valid = 1,
1469 .req_param = {
1470 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1471 IB_QP_PORT |
1472 IB_QP_QKEY),
1473 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
1474 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1475 IB_QP_PORT |
1476 IB_QP_ACCESS_FLAGS),
1477 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1478 IB_QP_PORT |
1479 IB_QP_ACCESS_FLAGS),
1480 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1481 IB_QP_PORT |
1482 IB_QP_ACCESS_FLAGS),
1483 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1484 IB_QP_PORT |
1485 IB_QP_ACCESS_FLAGS),
1486 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1487 IB_QP_QKEY),
1488 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1489 IB_QP_QKEY),
1490 }
1491 },
1492 },
1493 [IB_QPS_INIT] = {
1494 [IB_QPS_RESET] = { .valid = 1 },
1495 [IB_QPS_ERR] = { .valid = 1 },
1496 [IB_QPS_INIT] = {
1497 .valid = 1,
1498 .opt_param = {
1499 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1500 IB_QP_PORT |
1501 IB_QP_QKEY),
1502 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1503 IB_QP_PORT |
1504 IB_QP_ACCESS_FLAGS),
1505 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1506 IB_QP_PORT |
1507 IB_QP_ACCESS_FLAGS),
1508 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1509 IB_QP_PORT |
1510 IB_QP_ACCESS_FLAGS),
1511 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1512 IB_QP_PORT |
1513 IB_QP_ACCESS_FLAGS),
1514 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1515 IB_QP_QKEY),
1516 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1517 IB_QP_QKEY),
1518 }
1519 },
1520 [IB_QPS_RTR] = {
1521 .valid = 1,
1522 .req_param = {
1523 [IB_QPT_UC] = (IB_QP_AV |
1524 IB_QP_PATH_MTU |
1525 IB_QP_DEST_QPN |
1526 IB_QP_RQ_PSN),
1527 [IB_QPT_RC] = (IB_QP_AV |
1528 IB_QP_PATH_MTU |
1529 IB_QP_DEST_QPN |
1530 IB_QP_RQ_PSN |
1531 IB_QP_MAX_DEST_RD_ATOMIC |
1532 IB_QP_MIN_RNR_TIMER),
1533 [IB_QPT_XRC_INI] = (IB_QP_AV |
1534 IB_QP_PATH_MTU |
1535 IB_QP_DEST_QPN |
1536 IB_QP_RQ_PSN),
1537 [IB_QPT_XRC_TGT] = (IB_QP_AV |
1538 IB_QP_PATH_MTU |
1539 IB_QP_DEST_QPN |
1540 IB_QP_RQ_PSN |
1541 IB_QP_MAX_DEST_RD_ATOMIC |
1542 IB_QP_MIN_RNR_TIMER),
1543 },
1544 .opt_param = {
1545 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1546 IB_QP_QKEY),
1547 [IB_QPT_UC] = (IB_QP_ALT_PATH |
1548 IB_QP_ACCESS_FLAGS |
1549 IB_QP_PKEY_INDEX),
1550 [IB_QPT_RC] = (IB_QP_ALT_PATH |
1551 IB_QP_ACCESS_FLAGS |
1552 IB_QP_PKEY_INDEX),
1553 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH |
1554 IB_QP_ACCESS_FLAGS |
1555 IB_QP_PKEY_INDEX),
1556 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH |
1557 IB_QP_ACCESS_FLAGS |
1558 IB_QP_PKEY_INDEX),
1559 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1560 IB_QP_QKEY),
1561 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1562 IB_QP_QKEY),
1563 },
1564 },
1565 },
1566 [IB_QPS_RTR] = {
1567 [IB_QPS_RESET] = { .valid = 1 },
1568 [IB_QPS_ERR] = { .valid = 1 },
1569 [IB_QPS_RTS] = {
1570 .valid = 1,
1571 .req_param = {
1572 [IB_QPT_UD] = IB_QP_SQ_PSN,
1573 [IB_QPT_UC] = IB_QP_SQ_PSN,
1574 [IB_QPT_RC] = (IB_QP_TIMEOUT |
1575 IB_QP_RETRY_CNT |
1576 IB_QP_RNR_RETRY |
1577 IB_QP_SQ_PSN |
1578 IB_QP_MAX_QP_RD_ATOMIC),
1579 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT |
1580 IB_QP_RETRY_CNT |
1581 IB_QP_RNR_RETRY |
1582 IB_QP_SQ_PSN |
1583 IB_QP_MAX_QP_RD_ATOMIC),
1584 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT |
1585 IB_QP_SQ_PSN),
1586 [IB_QPT_SMI] = IB_QP_SQ_PSN,
1587 [IB_QPT_GSI] = IB_QP_SQ_PSN,
1588 },
1589 .opt_param = {
1590 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1591 IB_QP_QKEY),
1592 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1593 IB_QP_ALT_PATH |
1594 IB_QP_ACCESS_FLAGS |
1595 IB_QP_PATH_MIG_STATE),
1596 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1597 IB_QP_ALT_PATH |
1598 IB_QP_ACCESS_FLAGS |
1599 IB_QP_MIN_RNR_TIMER |
1600 IB_QP_PATH_MIG_STATE),
1601 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1602 IB_QP_ALT_PATH |
1603 IB_QP_ACCESS_FLAGS |
1604 IB_QP_PATH_MIG_STATE),
1605 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1606 IB_QP_ALT_PATH |
1607 IB_QP_ACCESS_FLAGS |
1608 IB_QP_MIN_RNR_TIMER |
1609 IB_QP_PATH_MIG_STATE),
1610 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1611 IB_QP_QKEY),
1612 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1613 IB_QP_QKEY),
1614 }
1615 }
1616 },
1617 [IB_QPS_RTS] = {
1618 [IB_QPS_RESET] = { .valid = 1 },
1619 [IB_QPS_ERR] = { .valid = 1 },
1620 [IB_QPS_RTS] = {
1621 .valid = 1,
1622 .opt_param = {
1623 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1624 IB_QP_QKEY),
1625 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1626 IB_QP_ACCESS_FLAGS |
1627 IB_QP_ALT_PATH |
1628 IB_QP_PATH_MIG_STATE),
1629 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1630 IB_QP_ACCESS_FLAGS |
1631 IB_QP_ALT_PATH |
1632 IB_QP_PATH_MIG_STATE |
1633 IB_QP_MIN_RNR_TIMER),
1634 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1635 IB_QP_ACCESS_FLAGS |
1636 IB_QP_ALT_PATH |
1637 IB_QP_PATH_MIG_STATE),
1638 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1639 IB_QP_ACCESS_FLAGS |
1640 IB_QP_ALT_PATH |
1641 IB_QP_PATH_MIG_STATE |
1642 IB_QP_MIN_RNR_TIMER),
1643 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1644 IB_QP_QKEY),
1645 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1646 IB_QP_QKEY),
1647 }
1648 },
1649 [IB_QPS_SQD] = {
1650 .valid = 1,
1651 .opt_param = {
1652 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1653 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1654 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1655 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1656 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1657 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1658 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1659 }
1660 },
1661 },
1662 [IB_QPS_SQD] = {
1663 [IB_QPS_RESET] = { .valid = 1 },
1664 [IB_QPS_ERR] = { .valid = 1 },
1665 [IB_QPS_RTS] = {
1666 .valid = 1,
1667 .opt_param = {
1668 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1669 IB_QP_QKEY),
1670 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1671 IB_QP_ALT_PATH |
1672 IB_QP_ACCESS_FLAGS |
1673 IB_QP_PATH_MIG_STATE),
1674 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1675 IB_QP_ALT_PATH |
1676 IB_QP_ACCESS_FLAGS |
1677 IB_QP_MIN_RNR_TIMER |
1678 IB_QP_PATH_MIG_STATE),
1679 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1680 IB_QP_ALT_PATH |
1681 IB_QP_ACCESS_FLAGS |
1682 IB_QP_PATH_MIG_STATE),
1683 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1684 IB_QP_ALT_PATH |
1685 IB_QP_ACCESS_FLAGS |
1686 IB_QP_MIN_RNR_TIMER |
1687 IB_QP_PATH_MIG_STATE),
1688 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1689 IB_QP_QKEY),
1690 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1691 IB_QP_QKEY),
1692 }
1693 },
1694 [IB_QPS_SQD] = {
1695 .valid = 1,
1696 .opt_param = {
1697 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1698 IB_QP_QKEY),
1699 [IB_QPT_UC] = (IB_QP_AV |
1700 IB_QP_ALT_PATH |
1701 IB_QP_ACCESS_FLAGS |
1702 IB_QP_PKEY_INDEX |
1703 IB_QP_PATH_MIG_STATE),
1704 [IB_QPT_RC] = (IB_QP_PORT |
1705 IB_QP_AV |
1706 IB_QP_TIMEOUT |
1707 IB_QP_RETRY_CNT |
1708 IB_QP_RNR_RETRY |
1709 IB_QP_MAX_QP_RD_ATOMIC |
1710 IB_QP_MAX_DEST_RD_ATOMIC |
1711 IB_QP_ALT_PATH |
1712 IB_QP_ACCESS_FLAGS |
1713 IB_QP_PKEY_INDEX |
1714 IB_QP_MIN_RNR_TIMER |
1715 IB_QP_PATH_MIG_STATE),
1716 [IB_QPT_XRC_INI] = (IB_QP_PORT |
1717 IB_QP_AV |
1718 IB_QP_TIMEOUT |
1719 IB_QP_RETRY_CNT |
1720 IB_QP_RNR_RETRY |
1721 IB_QP_MAX_QP_RD_ATOMIC |
1722 IB_QP_ALT_PATH |
1723 IB_QP_ACCESS_FLAGS |
1724 IB_QP_PKEY_INDEX |
1725 IB_QP_PATH_MIG_STATE),
1726 [IB_QPT_XRC_TGT] = (IB_QP_PORT |
1727 IB_QP_AV |
1728 IB_QP_TIMEOUT |
1729 IB_QP_MAX_DEST_RD_ATOMIC |
1730 IB_QP_ALT_PATH |
1731 IB_QP_ACCESS_FLAGS |
1732 IB_QP_PKEY_INDEX |
1733 IB_QP_MIN_RNR_TIMER |
1734 IB_QP_PATH_MIG_STATE),
1735 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1736 IB_QP_QKEY),
1737 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1738 IB_QP_QKEY),
1739 }
1740 }
1741 },
1742 [IB_QPS_SQE] = {
1743 [IB_QPS_RESET] = { .valid = 1 },
1744 [IB_QPS_ERR] = { .valid = 1 },
1745 [IB_QPS_RTS] = {
1746 .valid = 1,
1747 .opt_param = {
1748 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1749 IB_QP_QKEY),
1750 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1751 IB_QP_ACCESS_FLAGS),
1752 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1753 IB_QP_QKEY),
1754 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1755 IB_QP_QKEY),
1756 }
1757 }
1758 },
1759 [IB_QPS_ERR] = {
1760 [IB_QPS_RESET] = { .valid = 1 },
1761 [IB_QPS_ERR] = { .valid = 1 }
1762 }
1763 };
1764
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)1765 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1766 enum ib_qp_type type, enum ib_qp_attr_mask mask)
1767 {
1768 enum ib_qp_attr_mask req_param, opt_param;
1769
1770 if (mask & IB_QP_CUR_STATE &&
1771 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1772 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1773 return false;
1774
1775 if (!qp_state_table[cur_state][next_state].valid)
1776 return false;
1777
1778 req_param = qp_state_table[cur_state][next_state].req_param[type];
1779 opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1780
1781 if ((mask & req_param) != req_param)
1782 return false;
1783
1784 if (mask & ~(req_param | opt_param | IB_QP_STATE | IB_QP_RATE_LIMIT))
1785 return false;
1786
1787 return true;
1788 }
1789 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1790
1791 /**
1792 * ib_resolve_eth_dmac - Resolve destination mac address
1793 * @device: Device to consider
1794 * @ah_attr: address handle attribute which describes the
1795 * source and destination parameters
1796 * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1797 * returns 0 on success or appropriate error code. It initializes the
1798 * necessary ah_attr fields when call is successful.
1799 */
ib_resolve_eth_dmac(struct ib_device * device,struct rdma_ah_attr * ah_attr)1800 static int ib_resolve_eth_dmac(struct ib_device *device,
1801 struct rdma_ah_attr *ah_attr)
1802 {
1803 int ret = 0;
1804
1805 if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1806 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1807 __be32 addr = 0;
1808
1809 memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1810 ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1811 } else {
1812 ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1813 (char *)ah_attr->roce.dmac);
1814 }
1815 } else {
1816 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1817 }
1818 return ret;
1819 }
1820
is_qp_type_connected(const struct ib_qp * qp)1821 static bool is_qp_type_connected(const struct ib_qp *qp)
1822 {
1823 return (qp->qp_type == IB_QPT_UC ||
1824 qp->qp_type == IB_QPT_RC ||
1825 qp->qp_type == IB_QPT_XRC_INI ||
1826 qp->qp_type == IB_QPT_XRC_TGT);
1827 }
1828
1829 /*
1830 * IB core internal function to perform QP attributes modification.
1831 */
_ib_modify_qp(struct ib_qp * qp,struct ib_qp_attr * attr,int attr_mask,struct ib_udata * udata)1832 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1833 int attr_mask, struct ib_udata *udata)
1834 {
1835 u32 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1836 const struct ib_gid_attr *old_sgid_attr_av;
1837 const struct ib_gid_attr *old_sgid_attr_alt_av;
1838 int ret;
1839
1840 attr->xmit_slave = NULL;
1841 if (attr_mask & IB_QP_AV) {
1842 ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1843 &old_sgid_attr_av);
1844 if (ret)
1845 return ret;
1846
1847 if (attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1848 is_qp_type_connected(qp)) {
1849 struct net_device *slave;
1850
1851 /*
1852 * If the user provided the qp_attr then we have to
1853 * resolve it. Kerne users have to provide already
1854 * resolved rdma_ah_attr's.
1855 */
1856 if (udata) {
1857 ret = ib_resolve_eth_dmac(qp->device,
1858 &attr->ah_attr);
1859 if (ret)
1860 goto out_av;
1861 }
1862 slave = rdma_lag_get_ah_roce_slave(qp->device,
1863 &attr->ah_attr,
1864 GFP_KERNEL);
1865 if (IS_ERR(slave)) {
1866 ret = PTR_ERR(slave);
1867 goto out_av;
1868 }
1869 attr->xmit_slave = slave;
1870 }
1871 }
1872 if (attr_mask & IB_QP_ALT_PATH) {
1873 /*
1874 * FIXME: This does not track the migration state, so if the
1875 * user loads a new alternate path after the HW has migrated
1876 * from primary->alternate we will keep the wrong
1877 * references. This is OK for IB because the reference
1878 * counting does not serve any functional purpose.
1879 */
1880 ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1881 &old_sgid_attr_alt_av);
1882 if (ret)
1883 goto out_av;
1884
1885 /*
1886 * Today the core code can only handle alternate paths and APM
1887 * for IB. Ban them in roce mode.
1888 */
1889 if (!(rdma_protocol_ib(qp->device,
1890 attr->alt_ah_attr.port_num) &&
1891 rdma_protocol_ib(qp->device, port))) {
1892 ret = -EINVAL;
1893 goto out;
1894 }
1895 }
1896
1897 if (rdma_ib_or_roce(qp->device, port)) {
1898 if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1899 dev_warn(&qp->device->dev,
1900 "%s rq_psn overflow, masking to 24 bits\n",
1901 __func__);
1902 attr->rq_psn &= 0xffffff;
1903 }
1904
1905 if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1906 dev_warn(&qp->device->dev,
1907 " %s sq_psn overflow, masking to 24 bits\n",
1908 __func__);
1909 attr->sq_psn &= 0xffffff;
1910 }
1911 }
1912
1913 /*
1914 * Bind this qp to a counter automatically based on the rdma counter
1915 * rules. This only set in RST2INIT with port specified
1916 */
1917 if (!qp->counter && (attr_mask & IB_QP_PORT) &&
1918 ((attr_mask & IB_QP_STATE) && attr->qp_state == IB_QPS_INIT))
1919 rdma_counter_bind_qp_auto(qp, attr->port_num);
1920
1921 ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1922 if (ret)
1923 goto out;
1924
1925 if (attr_mask & IB_QP_PORT)
1926 qp->port = attr->port_num;
1927 if (attr_mask & IB_QP_AV)
1928 qp->av_sgid_attr =
1929 rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1930 if (attr_mask & IB_QP_ALT_PATH)
1931 qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1932 &attr->alt_ah_attr, qp->alt_path_sgid_attr);
1933
1934 out:
1935 if (attr_mask & IB_QP_ALT_PATH)
1936 rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1937 out_av:
1938 if (attr_mask & IB_QP_AV) {
1939 rdma_lag_put_ah_roce_slave(attr->xmit_slave);
1940 rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1941 }
1942 return ret;
1943 }
1944
1945 /**
1946 * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1947 * @ib_qp: The QP to modify.
1948 * @attr: On input, specifies the QP attributes to modify. On output,
1949 * the current values of selected QP attributes are returned.
1950 * @attr_mask: A bit-mask used to specify which attributes of the QP
1951 * are being modified.
1952 * @udata: pointer to user's input output buffer information
1953 * are being modified.
1954 * It returns 0 on success and returns appropriate error code on error.
1955 */
ib_modify_qp_with_udata(struct ib_qp * ib_qp,struct ib_qp_attr * attr,int attr_mask,struct ib_udata * udata)1956 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1957 int attr_mask, struct ib_udata *udata)
1958 {
1959 return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1960 }
1961 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1962
ib_get_width_and_speed(u32 netdev_speed,u32 lanes,u16 * speed,u8 * width)1963 static void ib_get_width_and_speed(u32 netdev_speed, u32 lanes,
1964 u16 *speed, u8 *width)
1965 {
1966 if (!lanes) {
1967 if (netdev_speed <= SPEED_1000) {
1968 *width = IB_WIDTH_1X;
1969 *speed = IB_SPEED_SDR;
1970 } else if (netdev_speed <= SPEED_10000) {
1971 *width = IB_WIDTH_1X;
1972 *speed = IB_SPEED_FDR10;
1973 } else if (netdev_speed <= SPEED_20000) {
1974 *width = IB_WIDTH_4X;
1975 *speed = IB_SPEED_DDR;
1976 } else if (netdev_speed <= SPEED_25000) {
1977 *width = IB_WIDTH_1X;
1978 *speed = IB_SPEED_EDR;
1979 } else if (netdev_speed <= SPEED_40000) {
1980 *width = IB_WIDTH_4X;
1981 *speed = IB_SPEED_FDR10;
1982 } else if (netdev_speed <= SPEED_50000) {
1983 *width = IB_WIDTH_2X;
1984 *speed = IB_SPEED_EDR;
1985 } else if (netdev_speed <= SPEED_100000) {
1986 *width = IB_WIDTH_4X;
1987 *speed = IB_SPEED_EDR;
1988 } else if (netdev_speed <= SPEED_200000) {
1989 *width = IB_WIDTH_4X;
1990 *speed = IB_SPEED_HDR;
1991 } else {
1992 *width = IB_WIDTH_4X;
1993 *speed = IB_SPEED_NDR;
1994 }
1995
1996 return;
1997 }
1998
1999 switch (lanes) {
2000 case 1:
2001 *width = IB_WIDTH_1X;
2002 break;
2003 case 2:
2004 *width = IB_WIDTH_2X;
2005 break;
2006 case 4:
2007 *width = IB_WIDTH_4X;
2008 break;
2009 case 8:
2010 *width = IB_WIDTH_8X;
2011 break;
2012 case 12:
2013 *width = IB_WIDTH_12X;
2014 break;
2015 default:
2016 *width = IB_WIDTH_1X;
2017 }
2018
2019 switch (netdev_speed / lanes) {
2020 case SPEED_2500:
2021 *speed = IB_SPEED_SDR;
2022 break;
2023 case SPEED_5000:
2024 *speed = IB_SPEED_DDR;
2025 break;
2026 case SPEED_10000:
2027 *speed = IB_SPEED_FDR10;
2028 break;
2029 case SPEED_14000:
2030 *speed = IB_SPEED_FDR;
2031 break;
2032 case SPEED_25000:
2033 *speed = IB_SPEED_EDR;
2034 break;
2035 case SPEED_50000:
2036 *speed = IB_SPEED_HDR;
2037 break;
2038 case SPEED_100000:
2039 *speed = IB_SPEED_NDR;
2040 break;
2041 default:
2042 *speed = IB_SPEED_SDR;
2043 }
2044 }
2045
ib_get_eth_speed(struct ib_device * dev,u32 port_num,u16 * speed,u8 * width)2046 int ib_get_eth_speed(struct ib_device *dev, u32 port_num, u16 *speed, u8 *width)
2047 {
2048 int rc;
2049 u32 netdev_speed;
2050 struct net_device *netdev;
2051 struct ethtool_link_ksettings lksettings = {};
2052
2053 if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
2054 return -EINVAL;
2055
2056 netdev = ib_device_get_netdev(dev, port_num);
2057 if (!netdev)
2058 return -ENODEV;
2059
2060 rtnl_lock();
2061 if (READ_ONCE(netdev->reg_state) != NETREG_REGISTERED) {
2062 dev_put(netdev);
2063 rtnl_unlock();
2064 return -ENODEV;
2065 }
2066
2067 rc = __ethtool_get_link_ksettings(netdev, &lksettings);
2068 if (!rc && lksettings.base.speed != (u32)SPEED_UNKNOWN) {
2069 netdev_speed = lksettings.base.speed;
2070 } else {
2071 netdev_speed = SPEED_1000;
2072 if (rc)
2073 pr_warn("%s speed is unknown, defaulting to %u\n",
2074 netdev->name, netdev_speed);
2075 }
2076 dev_put(netdev);
2077 rtnl_unlock();
2078
2079 ib_get_width_and_speed(netdev_speed, lksettings.lanes,
2080 speed, width);
2081
2082 return 0;
2083 }
2084 EXPORT_SYMBOL(ib_get_eth_speed);
2085
ib_modify_qp(struct ib_qp * qp,struct ib_qp_attr * qp_attr,int qp_attr_mask)2086 int ib_modify_qp(struct ib_qp *qp,
2087 struct ib_qp_attr *qp_attr,
2088 int qp_attr_mask)
2089 {
2090 return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
2091 }
2092 EXPORT_SYMBOL(ib_modify_qp);
2093
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)2094 int ib_query_qp(struct ib_qp *qp,
2095 struct ib_qp_attr *qp_attr,
2096 int qp_attr_mask,
2097 struct ib_qp_init_attr *qp_init_attr)
2098 {
2099 qp_attr->ah_attr.grh.sgid_attr = NULL;
2100 qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
2101
2102 return qp->device->ops.query_qp ?
2103 qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask,
2104 qp_init_attr) : -EOPNOTSUPP;
2105 }
2106 EXPORT_SYMBOL(ib_query_qp);
2107
ib_close_qp(struct ib_qp * qp)2108 int ib_close_qp(struct ib_qp *qp)
2109 {
2110 struct ib_qp *real_qp;
2111 unsigned long flags;
2112
2113 real_qp = qp->real_qp;
2114 if (real_qp == qp)
2115 return -EINVAL;
2116
2117 spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
2118 list_del(&qp->open_list);
2119 spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
2120
2121 atomic_dec(&real_qp->usecnt);
2122 if (qp->qp_sec)
2123 ib_close_shared_qp_security(qp->qp_sec);
2124 kfree(qp);
2125
2126 return 0;
2127 }
2128 EXPORT_SYMBOL(ib_close_qp);
2129
__ib_destroy_shared_qp(struct ib_qp * qp)2130 static int __ib_destroy_shared_qp(struct ib_qp *qp)
2131 {
2132 struct ib_xrcd *xrcd;
2133 struct ib_qp *real_qp;
2134 int ret;
2135
2136 real_qp = qp->real_qp;
2137 xrcd = real_qp->xrcd;
2138 down_write(&xrcd->tgt_qps_rwsem);
2139 ib_close_qp(qp);
2140 if (atomic_read(&real_qp->usecnt) == 0)
2141 xa_erase(&xrcd->tgt_qps, real_qp->qp_num);
2142 else
2143 real_qp = NULL;
2144 up_write(&xrcd->tgt_qps_rwsem);
2145
2146 if (real_qp) {
2147 ret = ib_destroy_qp(real_qp);
2148 if (!ret)
2149 atomic_dec(&xrcd->usecnt);
2150 }
2151
2152 return 0;
2153 }
2154
ib_destroy_qp_user(struct ib_qp * qp,struct ib_udata * udata)2155 int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata)
2156 {
2157 const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
2158 const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
2159 struct ib_qp_security *sec;
2160 struct ib_comp_cntr *cc;
2161 unsigned long index;
2162 int ret;
2163
2164 WARN_ON_ONCE(qp->mrs_used > 0);
2165
2166 if (atomic_read(&qp->usecnt))
2167 return -EBUSY;
2168
2169 if (qp->real_qp != qp)
2170 return __ib_destroy_shared_qp(qp);
2171
2172 rdma_restrack_begin_del(&qp->res);
2173
2174 sec = qp->qp_sec;
2175 if (sec)
2176 ib_destroy_qp_security_begin(sec);
2177
2178 if (!qp->uobject)
2179 rdma_rw_cleanup_mrs(qp);
2180
2181 rdma_counter_unbind_qp(qp, qp->port, true);
2182 ret = qp->device->ops.destroy_qp(qp, udata);
2183 if (ret) {
2184 if (sec)
2185 ib_destroy_qp_security_abort(sec);
2186 rdma_restrack_abort_del(&qp->res);
2187 return ret;
2188 }
2189
2190 if (alt_path_sgid_attr)
2191 rdma_put_gid_attr(alt_path_sgid_attr);
2192 if (av_sgid_attr)
2193 rdma_put_gid_attr(av_sgid_attr);
2194
2195 xa_for_each(&qp->comp_cntrs, index, cc)
2196 atomic_dec(&cc->usecnt);
2197 xa_destroy(&qp->comp_cntrs);
2198
2199 ib_qp_usecnt_dec(qp);
2200 if (sec)
2201 ib_destroy_qp_security_end(sec);
2202
2203 rdma_restrack_commit_del(&qp->res);
2204 kfree(qp);
2205 return ret;
2206 }
2207 EXPORT_SYMBOL(ib_destroy_qp_user);
2208
2209 /* Completion queues */
2210
__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)2211 struct ib_cq *__ib_create_cq(struct ib_device *device,
2212 ib_comp_handler comp_handler,
2213 void (*event_handler)(struct ib_event *, void *),
2214 void *cq_context,
2215 const struct ib_cq_init_attr *cq_attr,
2216 const char *caller)
2217 {
2218 struct ib_cq *cq;
2219 int ret;
2220
2221 if (WARN_ON_ONCE(!cq_attr->cqe))
2222 return ERR_PTR(-EINVAL);
2223
2224 cq = rdma_zalloc_drv_obj(device, ib_cq);
2225 if (!cq)
2226 return ERR_PTR(-ENOMEM);
2227
2228 cq->device = device;
2229 cq->comp_handler = comp_handler;
2230 cq->event_handler = event_handler;
2231 cq->cq_context = cq_context;
2232 atomic_set(&cq->usecnt, 0);
2233
2234 rdma_restrack_new(&cq->res, RDMA_RESTRACK_CQ);
2235 rdma_restrack_set_name(&cq->res, caller);
2236
2237 ret = device->ops.create_cq(cq, cq_attr, NULL);
2238 if (ret) {
2239 rdma_restrack_put(&cq->res);
2240 kfree(cq);
2241 return ERR_PTR(ret);
2242 }
2243 rdma_restrack_add(&cq->res);
2244 return cq;
2245 }
2246 EXPORT_SYMBOL(__ib_create_cq);
2247
rdma_set_cq_moderation(struct ib_cq * cq,u16 cq_count,u16 cq_period)2248 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
2249 {
2250 if (cq->shared)
2251 return -EOPNOTSUPP;
2252
2253 return cq->device->ops.modify_cq ?
2254 cq->device->ops.modify_cq(cq, cq_count,
2255 cq_period) : -EOPNOTSUPP;
2256 }
2257 EXPORT_SYMBOL(rdma_set_cq_moderation);
2258
ib_destroy_cq_user(struct ib_cq * cq,struct ib_udata * udata)2259 int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata)
2260 {
2261 int ret;
2262
2263 if (WARN_ON_ONCE(cq->shared))
2264 return -EOPNOTSUPP;
2265
2266 if (atomic_read(&cq->usecnt))
2267 return -EBUSY;
2268
2269 rdma_restrack_begin_del(&cq->res);
2270
2271 ret = cq->device->ops.destroy_cq(cq, udata);
2272 if (ret) {
2273 rdma_restrack_abort_del(&cq->res);
2274 return ret;
2275 }
2276
2277 rdma_restrack_commit_del(&cq->res);
2278 kfree(cq);
2279 return ret;
2280 }
2281 EXPORT_SYMBOL(ib_destroy_cq_user);
2282
2283 /* Memory regions */
2284
ib_reg_user_mr(struct ib_pd * pd,u64 start,u64 length,u64 virt_addr,int access_flags)2285 struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
2286 u64 virt_addr, int access_flags)
2287 {
2288 struct ib_mr *mr;
2289
2290 if (access_flags & IB_ACCESS_ON_DEMAND) {
2291 if (!(pd->device->attrs.kernel_cap_flags &
2292 IBK_ON_DEMAND_PAGING)) {
2293 pr_debug("ODP support not available\n");
2294 return ERR_PTR(-EINVAL);
2295 }
2296 }
2297
2298 mr = pd->device->ops.reg_user_mr(pd, start, length, virt_addr,
2299 access_flags, NULL, NULL);
2300
2301 if (IS_ERR(mr))
2302 return mr;
2303
2304 mr->device = pd->device;
2305 mr->type = IB_MR_TYPE_USER;
2306 mr->pd = pd;
2307 mr->dm = NULL;
2308 atomic_inc(&pd->usecnt);
2309 mr->iova = virt_addr;
2310 mr->length = length;
2311
2312 rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2313 rdma_restrack_parent_name(&mr->res, &pd->res);
2314 rdma_restrack_add(&mr->res);
2315
2316 return mr;
2317 }
2318 EXPORT_SYMBOL(ib_reg_user_mr);
2319
ib_advise_mr(struct ib_pd * pd,enum ib_uverbs_advise_mr_advice advice,u32 flags,struct ib_sge * sg_list,u32 num_sge)2320 int ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
2321 u32 flags, struct ib_sge *sg_list, u32 num_sge)
2322 {
2323 if (!pd->device->ops.advise_mr)
2324 return -EOPNOTSUPP;
2325
2326 if (!num_sge)
2327 return 0;
2328
2329 return pd->device->ops.advise_mr(pd, advice, flags, sg_list, num_sge,
2330 NULL);
2331 }
2332 EXPORT_SYMBOL(ib_advise_mr);
2333
ib_dereg_mr_user(struct ib_mr * mr,struct ib_udata * udata)2334 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata)
2335 {
2336 struct ib_pd *pd = mr->pd;
2337 struct ib_dm *dm = mr->dm;
2338 struct ib_dmah *dmah = mr->dmah;
2339 struct ib_sig_attrs *sig_attrs = mr->sig_attrs;
2340 int ret;
2341
2342 trace_mr_dereg(mr);
2343 rdma_restrack_del(&mr->res);
2344 ret = mr->device->ops.dereg_mr(mr, udata);
2345 if (!ret) {
2346 atomic_dec(&pd->usecnt);
2347 if (dm)
2348 atomic_dec(&dm->usecnt);
2349 if (dmah)
2350 atomic_dec(&dmah->usecnt);
2351 kfree(sig_attrs);
2352 }
2353
2354 return ret;
2355 }
2356 EXPORT_SYMBOL(ib_dereg_mr_user);
2357
2358 /**
2359 * ib_alloc_mr() - Allocates a memory region
2360 * @pd: protection domain associated with the region
2361 * @mr_type: memory region type
2362 * @max_num_sg: maximum sg entries available for registration.
2363 *
2364 * Notes:
2365 * Memory registeration page/sg lists must not exceed max_num_sg.
2366 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
2367 * max_num_sg * used_page_size.
2368 *
2369 */
ib_alloc_mr(struct ib_pd * pd,enum ib_mr_type mr_type,u32 max_num_sg)2370 struct ib_mr *ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2371 u32 max_num_sg)
2372 {
2373 struct ib_mr *mr;
2374
2375 if (!pd->device->ops.alloc_mr) {
2376 mr = ERR_PTR(-EOPNOTSUPP);
2377 goto out;
2378 }
2379
2380 if (mr_type == IB_MR_TYPE_INTEGRITY) {
2381 WARN_ON_ONCE(1);
2382 mr = ERR_PTR(-EINVAL);
2383 goto out;
2384 }
2385
2386 mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg);
2387 if (IS_ERR(mr))
2388 goto out;
2389
2390 mr->device = pd->device;
2391 mr->pd = pd;
2392 mr->dm = NULL;
2393 mr->uobject = NULL;
2394 atomic_inc(&pd->usecnt);
2395 mr->need_inval = false;
2396 mr->type = mr_type;
2397 mr->sig_attrs = NULL;
2398
2399 rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2400 rdma_restrack_parent_name(&mr->res, &pd->res);
2401 rdma_restrack_add(&mr->res);
2402 out:
2403 trace_mr_alloc(pd, mr_type, max_num_sg, mr);
2404 return mr;
2405 }
2406 EXPORT_SYMBOL(ib_alloc_mr);
2407
2408 /**
2409 * ib_alloc_mr_integrity() - Allocates an integrity memory region
2410 * @pd: protection domain associated with the region
2411 * @max_num_data_sg: maximum data sg entries available for registration
2412 * @max_num_meta_sg: maximum metadata sg entries available for
2413 * registration
2414 *
2415 * Notes:
2416 * Memory registration page/sg lists must not exceed max_num_sg,
2417 * also the integrity page/sg lists must not exceed max_num_meta_sg.
2418 *
2419 */
ib_alloc_mr_integrity(struct ib_pd * pd,u32 max_num_data_sg,u32 max_num_meta_sg)2420 struct ib_mr *ib_alloc_mr_integrity(struct ib_pd *pd,
2421 u32 max_num_data_sg,
2422 u32 max_num_meta_sg)
2423 {
2424 struct ib_mr *mr;
2425 struct ib_sig_attrs *sig_attrs;
2426
2427 if (!pd->device->ops.alloc_mr_integrity ||
2428 !pd->device->ops.map_mr_sg_pi) {
2429 mr = ERR_PTR(-EOPNOTSUPP);
2430 goto out;
2431 }
2432
2433 if (!max_num_meta_sg) {
2434 mr = ERR_PTR(-EINVAL);
2435 goto out;
2436 }
2437
2438 sig_attrs = kzalloc_obj(struct ib_sig_attrs);
2439 if (!sig_attrs) {
2440 mr = ERR_PTR(-ENOMEM);
2441 goto out;
2442 }
2443
2444 mr = pd->device->ops.alloc_mr_integrity(pd, max_num_data_sg,
2445 max_num_meta_sg);
2446 if (IS_ERR(mr)) {
2447 kfree(sig_attrs);
2448 goto out;
2449 }
2450
2451 mr->device = pd->device;
2452 mr->pd = pd;
2453 mr->dm = NULL;
2454 mr->uobject = NULL;
2455 atomic_inc(&pd->usecnt);
2456 mr->need_inval = false;
2457 mr->type = IB_MR_TYPE_INTEGRITY;
2458 mr->sig_attrs = sig_attrs;
2459
2460 rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2461 rdma_restrack_parent_name(&mr->res, &pd->res);
2462 rdma_restrack_add(&mr->res);
2463 out:
2464 trace_mr_integ_alloc(pd, max_num_data_sg, max_num_meta_sg, mr);
2465 return mr;
2466 }
2467 EXPORT_SYMBOL(ib_alloc_mr_integrity);
2468
2469 /* Multicast groups */
2470
is_valid_mcast_lid(struct ib_qp * qp,u16 lid)2471 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2472 {
2473 struct ib_qp_init_attr init_attr = {};
2474 struct ib_qp_attr attr = {};
2475 int num_eth_ports = 0;
2476 unsigned int port;
2477
2478 /* If QP state >= init, it is assigned to a port and we can check this
2479 * port only.
2480 */
2481 if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2482 if (attr.qp_state >= IB_QPS_INIT) {
2483 if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2484 IB_LINK_LAYER_INFINIBAND)
2485 return true;
2486 goto lid_check;
2487 }
2488 }
2489
2490 /* Can't get a quick answer, iterate over all ports */
2491 rdma_for_each_port(qp->device, port)
2492 if (rdma_port_get_link_layer(qp->device, port) !=
2493 IB_LINK_LAYER_INFINIBAND)
2494 num_eth_ports++;
2495
2496 /* If we have at lease one Ethernet port, RoCE annex declares that
2497 * multicast LID should be ignored. We can't tell at this step if the
2498 * QP belongs to an IB or Ethernet port.
2499 */
2500 if (num_eth_ports)
2501 return true;
2502
2503 /* If all the ports are IB, we can check according to IB spec. */
2504 lid_check:
2505 return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2506 lid == be16_to_cpu(IB_LID_PERMISSIVE));
2507 }
2508
ib_attach_mcast(struct ib_qp * qp,union ib_gid * gid,u16 lid)2509 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2510 {
2511 int ret;
2512
2513 if (!qp->device->ops.attach_mcast)
2514 return -EOPNOTSUPP;
2515
2516 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2517 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2518 return -EINVAL;
2519
2520 ret = qp->device->ops.attach_mcast(qp, gid, lid);
2521 if (!ret)
2522 atomic_inc(&qp->usecnt);
2523 return ret;
2524 }
2525 EXPORT_SYMBOL(ib_attach_mcast);
2526
ib_detach_mcast(struct ib_qp * qp,union ib_gid * gid,u16 lid)2527 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2528 {
2529 int ret;
2530
2531 if (!qp->device->ops.detach_mcast)
2532 return -EOPNOTSUPP;
2533
2534 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2535 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2536 return -EINVAL;
2537
2538 ret = qp->device->ops.detach_mcast(qp, gid, lid);
2539 if (!ret)
2540 atomic_dec(&qp->usecnt);
2541 return ret;
2542 }
2543 EXPORT_SYMBOL(ib_detach_mcast);
2544
2545 /**
2546 * ib_alloc_xrcd_user - Allocates an XRC domain.
2547 * @device: The device on which to allocate the XRC domain.
2548 * @inode: inode to connect XRCD
2549 * @udata: Valid user data or NULL for kernel object
2550 */
ib_alloc_xrcd_user(struct ib_device * device,struct inode * inode,struct ib_udata * udata)2551 struct ib_xrcd *ib_alloc_xrcd_user(struct ib_device *device,
2552 struct inode *inode, struct ib_udata *udata)
2553 {
2554 struct ib_xrcd *xrcd;
2555 int ret;
2556
2557 if (!device->ops.alloc_xrcd)
2558 return ERR_PTR(-EOPNOTSUPP);
2559
2560 xrcd = rdma_zalloc_drv_obj(device, ib_xrcd);
2561 if (!xrcd)
2562 return ERR_PTR(-ENOMEM);
2563
2564 xrcd->device = device;
2565 xrcd->inode = inode;
2566 atomic_set(&xrcd->usecnt, 0);
2567 init_rwsem(&xrcd->tgt_qps_rwsem);
2568 xa_init(&xrcd->tgt_qps);
2569
2570 ret = device->ops.alloc_xrcd(xrcd, udata);
2571 if (ret)
2572 goto err;
2573 return xrcd;
2574 err:
2575 kfree(xrcd);
2576 return ERR_PTR(ret);
2577 }
2578 EXPORT_SYMBOL(ib_alloc_xrcd_user);
2579
2580 /**
2581 * ib_dealloc_xrcd_user - Deallocates an XRC domain.
2582 * @xrcd: The XRC domain to deallocate.
2583 * @udata: Valid user data or NULL for kernel object
2584 */
ib_dealloc_xrcd_user(struct ib_xrcd * xrcd,struct ib_udata * udata)2585 int ib_dealloc_xrcd_user(struct ib_xrcd *xrcd, struct ib_udata *udata)
2586 {
2587 int ret;
2588
2589 if (atomic_read(&xrcd->usecnt))
2590 return -EBUSY;
2591
2592 WARN_ON(!xa_empty(&xrcd->tgt_qps));
2593 ret = xrcd->device->ops.dealloc_xrcd(xrcd, udata);
2594 if (ret)
2595 return ret;
2596 kfree(xrcd);
2597 return ret;
2598 }
2599 EXPORT_SYMBOL(ib_dealloc_xrcd_user);
2600
2601 /**
2602 * ib_create_wq - Creates a WQ associated with the specified protection
2603 * domain.
2604 * @pd: The protection domain associated with the WQ.
2605 * @wq_attr: A list of initial attributes required to create the
2606 * WQ. If WQ creation succeeds, then the attributes are updated to
2607 * the actual capabilities of the created WQ.
2608 *
2609 * wq_attr->max_wr and wq_attr->max_sge determine
2610 * the requested size of the WQ, and set to the actual values allocated
2611 * on return.
2612 * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2613 * at least as large as the requested values.
2614 */
ib_create_wq(struct ib_pd * pd,struct ib_wq_init_attr * wq_attr)2615 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2616 struct ib_wq_init_attr *wq_attr)
2617 {
2618 struct ib_wq *wq;
2619
2620 if (!pd->device->ops.create_wq)
2621 return ERR_PTR(-EOPNOTSUPP);
2622
2623 wq = pd->device->ops.create_wq(pd, wq_attr, NULL);
2624 if (!IS_ERR(wq)) {
2625 wq->event_handler = wq_attr->event_handler;
2626 wq->wq_context = wq_attr->wq_context;
2627 wq->wq_type = wq_attr->wq_type;
2628 wq->cq = wq_attr->cq;
2629 wq->device = pd->device;
2630 wq->pd = pd;
2631 wq->uobject = NULL;
2632 atomic_inc(&pd->usecnt);
2633 atomic_inc(&wq_attr->cq->usecnt);
2634 atomic_set(&wq->usecnt, 0);
2635 }
2636 return wq;
2637 }
2638 EXPORT_SYMBOL(ib_create_wq);
2639
2640 /**
2641 * ib_destroy_wq_user - Destroys the specified user WQ.
2642 * @wq: The WQ to destroy.
2643 * @udata: Valid user data
2644 */
ib_destroy_wq_user(struct ib_wq * wq,struct ib_udata * udata)2645 int ib_destroy_wq_user(struct ib_wq *wq, struct ib_udata *udata)
2646 {
2647 struct ib_cq *cq = wq->cq;
2648 struct ib_pd *pd = wq->pd;
2649 int ret;
2650
2651 if (atomic_read(&wq->usecnt))
2652 return -EBUSY;
2653
2654 ret = wq->device->ops.destroy_wq(wq, udata);
2655 if (ret)
2656 return ret;
2657
2658 atomic_dec(&pd->usecnt);
2659 atomic_dec(&cq->usecnt);
2660 return ret;
2661 }
2662 EXPORT_SYMBOL(ib_destroy_wq_user);
2663
ib_check_mr_status(struct ib_mr * mr,u32 check_mask,struct ib_mr_status * mr_status)2664 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2665 struct ib_mr_status *mr_status)
2666 {
2667 if (!mr->device->ops.check_mr_status)
2668 return -EOPNOTSUPP;
2669
2670 return mr->device->ops.check_mr_status(mr, check_mask, mr_status);
2671 }
2672 EXPORT_SYMBOL(ib_check_mr_status);
2673
ib_set_vf_link_state(struct ib_device * device,int vf,u32 port,int state)2674 int ib_set_vf_link_state(struct ib_device *device, int vf, u32 port,
2675 int state)
2676 {
2677 if (!device->ops.set_vf_link_state)
2678 return -EOPNOTSUPP;
2679
2680 return device->ops.set_vf_link_state(device, vf, port, state);
2681 }
2682 EXPORT_SYMBOL(ib_set_vf_link_state);
2683
ib_get_vf_config(struct ib_device * device,int vf,u32 port,struct ifla_vf_info * info)2684 int ib_get_vf_config(struct ib_device *device, int vf, u32 port,
2685 struct ifla_vf_info *info)
2686 {
2687 if (!device->ops.get_vf_config)
2688 return -EOPNOTSUPP;
2689
2690 return device->ops.get_vf_config(device, vf, port, info);
2691 }
2692 EXPORT_SYMBOL(ib_get_vf_config);
2693
ib_get_vf_stats(struct ib_device * device,int vf,u32 port,struct ifla_vf_stats * stats)2694 int ib_get_vf_stats(struct ib_device *device, int vf, u32 port,
2695 struct ifla_vf_stats *stats)
2696 {
2697 if (!device->ops.get_vf_stats)
2698 return -EOPNOTSUPP;
2699
2700 return device->ops.get_vf_stats(device, vf, port, stats);
2701 }
2702 EXPORT_SYMBOL(ib_get_vf_stats);
2703
ib_set_vf_guid(struct ib_device * device,int vf,u32 port,u64 guid,int type)2704 int ib_set_vf_guid(struct ib_device *device, int vf, u32 port, u64 guid,
2705 int type)
2706 {
2707 if (!device->ops.set_vf_guid)
2708 return -EOPNOTSUPP;
2709
2710 return device->ops.set_vf_guid(device, vf, port, guid, type);
2711 }
2712 EXPORT_SYMBOL(ib_set_vf_guid);
2713
ib_get_vf_guid(struct ib_device * device,int vf,u32 port,struct ifla_vf_guid * node_guid,struct ifla_vf_guid * port_guid)2714 int ib_get_vf_guid(struct ib_device *device, int vf, u32 port,
2715 struct ifla_vf_guid *node_guid,
2716 struct ifla_vf_guid *port_guid)
2717 {
2718 if (!device->ops.get_vf_guid)
2719 return -EOPNOTSUPP;
2720
2721 return device->ops.get_vf_guid(device, vf, port, node_guid, port_guid);
2722 }
2723 EXPORT_SYMBOL(ib_get_vf_guid);
2724 /**
2725 * ib_map_mr_sg_pi() - Map the dma mapped SG lists for PI (protection
2726 * information) and set an appropriate memory region for registration.
2727 * @mr: memory region
2728 * @data_sg: dma mapped scatterlist for data
2729 * @data_sg_nents: number of entries in data_sg
2730 * @data_sg_offset: offset in bytes into data_sg
2731 * @meta_sg: dma mapped scatterlist for metadata
2732 * @meta_sg_nents: number of entries in meta_sg
2733 * @meta_sg_offset: offset in bytes into meta_sg
2734 * @page_size: page vector desired page size
2735 *
2736 * Constraints:
2737 * - The MR must be allocated with type IB_MR_TYPE_INTEGRITY.
2738 *
2739 * Return: 0 on success.
2740 *
2741 * After this completes successfully, the memory region
2742 * is ready for registration.
2743 */
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)2744 int ib_map_mr_sg_pi(struct ib_mr *mr, struct scatterlist *data_sg,
2745 int data_sg_nents, unsigned int *data_sg_offset,
2746 struct scatterlist *meta_sg, int meta_sg_nents,
2747 unsigned int *meta_sg_offset, unsigned int page_size)
2748 {
2749 if (unlikely(!mr->device->ops.map_mr_sg_pi ||
2750 WARN_ON_ONCE(mr->type != IB_MR_TYPE_INTEGRITY)))
2751 return -EOPNOTSUPP;
2752
2753 mr->page_size = page_size;
2754
2755 return mr->device->ops.map_mr_sg_pi(mr, data_sg, data_sg_nents,
2756 data_sg_offset, meta_sg,
2757 meta_sg_nents, meta_sg_offset);
2758 }
2759 EXPORT_SYMBOL(ib_map_mr_sg_pi);
2760
2761 /**
2762 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2763 * and set it the memory region.
2764 * @mr: memory region
2765 * @sg: dma mapped scatterlist
2766 * @sg_nents: number of entries in sg
2767 * @sg_offset: offset in bytes into sg
2768 * @page_size: page vector desired page size
2769 *
2770 * Constraints:
2771 *
2772 * - The first sg element is allowed to have an offset.
2773 * - Each sg element must either be aligned to page_size or virtually
2774 * contiguous to the previous element. In case an sg element has a
2775 * non-contiguous offset, the mapping prefix will not include it.
2776 * - The last sg element is allowed to have length less than page_size.
2777 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2778 * then only max_num_sg entries will be mapped.
2779 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2780 * constraints holds and the page_size argument is ignored.
2781 *
2782 * Returns the number of sg elements that were mapped to the memory region.
2783 *
2784 * After this completes successfully, the memory region
2785 * is ready for registration.
2786 */
ib_map_mr_sg(struct ib_mr * mr,struct scatterlist * sg,int sg_nents,unsigned int * sg_offset,unsigned int page_size)2787 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2788 unsigned int *sg_offset, unsigned int page_size)
2789 {
2790 if (unlikely(!mr->device->ops.map_mr_sg))
2791 return -EOPNOTSUPP;
2792
2793 mr->page_size = page_size;
2794
2795 return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset);
2796 }
2797 EXPORT_SYMBOL(ib_map_mr_sg);
2798
2799 /**
2800 * ib_sg_to_pages() - Convert the largest prefix of a sg list
2801 * to a page vector
2802 * @mr: memory region
2803 * @sgl: dma mapped scatterlist
2804 * @sg_nents: number of entries in sg
2805 * @sg_offset_p: ==== =======================================================
2806 * IN start offset in bytes into sg
2807 * OUT offset in bytes for element n of the sg of the first
2808 * byte that has not been processed where n is the return
2809 * value of this function.
2810 * ==== =======================================================
2811 * @set_page: driver page assignment function pointer
2812 *
2813 * Core service helper for drivers to convert the largest
2814 * prefix of given sg list to a page vector. The sg list
2815 * prefix converted is the prefix that meet the requirements
2816 * of ib_map_mr_sg.
2817 *
2818 * Returns the number of sg elements that were assigned to
2819 * a page vector.
2820 */
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))2821 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2822 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2823 {
2824 struct scatterlist *sg;
2825 u64 last_end_dma_addr = 0;
2826 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2827 unsigned int last_page_off = 0;
2828 u64 page_mask = ~((u64)mr->page_size - 1);
2829 int i, ret;
2830
2831 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2832 return -EINVAL;
2833
2834 mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2835 mr->length = 0;
2836
2837 for_each_sg(sgl, sg, sg_nents, i) {
2838 u64 dma_addr = sg_dma_address(sg) + sg_offset;
2839 u64 prev_addr = dma_addr;
2840 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2841 u64 end_dma_addr = dma_addr + dma_len;
2842 u64 page_addr = dma_addr & page_mask;
2843
2844 /*
2845 * For the second and later elements, check whether either the
2846 * end of element i-1 or the start of element i is not aligned
2847 * on a page boundary.
2848 */
2849 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2850 /* Stop mapping if there is a gap. */
2851 if (last_end_dma_addr != dma_addr)
2852 break;
2853
2854 /*
2855 * Coalesce this element with the last. If it is small
2856 * enough just update mr->length. Otherwise start
2857 * mapping from the next page.
2858 */
2859 goto next_page;
2860 }
2861
2862 do {
2863 ret = set_page(mr, page_addr);
2864 if (unlikely(ret < 0)) {
2865 sg_offset = prev_addr - sg_dma_address(sg);
2866 mr->length += prev_addr - dma_addr;
2867 if (sg_offset_p)
2868 *sg_offset_p = sg_offset;
2869 return i || sg_offset ? i : ret;
2870 }
2871 prev_addr = page_addr;
2872 next_page:
2873 page_addr += mr->page_size;
2874 } while (page_addr < end_dma_addr);
2875
2876 mr->length += dma_len;
2877 last_end_dma_addr = end_dma_addr;
2878 last_page_off = end_dma_addr & ~page_mask;
2879
2880 sg_offset = 0;
2881 }
2882
2883 if (sg_offset_p)
2884 *sg_offset_p = 0;
2885 return i;
2886 }
2887 EXPORT_SYMBOL(ib_sg_to_pages);
2888
2889 struct ib_drain_cqe {
2890 struct ib_cqe cqe;
2891 struct completion done;
2892 };
2893
ib_drain_qp_done(struct ib_cq * cq,struct ib_wc * wc)2894 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2895 {
2896 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2897 cqe);
2898
2899 complete(&cqe->done);
2900 }
2901
2902 /*
2903 * Post a WR and block until its completion is reaped for the SQ.
2904 */
__ib_drain_sq(struct ib_qp * qp)2905 static void __ib_drain_sq(struct ib_qp *qp)
2906 {
2907 struct ib_cq *cq = qp->send_cq;
2908 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2909 struct ib_drain_cqe sdrain;
2910 struct ib_rdma_wr swr = {
2911 .wr = {
2912 .next = NULL,
2913 { .wr_cqe = &sdrain.cqe, },
2914 .opcode = IB_WR_RDMA_WRITE,
2915 },
2916 };
2917 int ret;
2918
2919 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2920 if (ret) {
2921 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2922 return;
2923 }
2924
2925 sdrain.cqe.done = ib_drain_qp_done;
2926 init_completion(&sdrain.done);
2927
2928 ret = ib_post_send(qp, &swr.wr, NULL);
2929 if (ret) {
2930 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2931 return;
2932 }
2933
2934 if (cq->poll_ctx == IB_POLL_DIRECT)
2935 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2936 ib_process_cq_direct(cq, -1);
2937 else
2938 wait_for_completion(&sdrain.done);
2939 }
2940
2941 /*
2942 * Post a WR and block until its completion is reaped for the RQ.
2943 */
__ib_drain_rq(struct ib_qp * qp)2944 static void __ib_drain_rq(struct ib_qp *qp)
2945 {
2946 struct ib_cq *cq = qp->recv_cq;
2947 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2948 struct ib_drain_cqe rdrain;
2949 struct ib_recv_wr rwr = {};
2950 int ret;
2951
2952 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2953 if (ret) {
2954 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2955 return;
2956 }
2957
2958 rwr.wr_cqe = &rdrain.cqe;
2959 rdrain.cqe.done = ib_drain_qp_done;
2960 init_completion(&rdrain.done);
2961
2962 ret = ib_post_recv(qp, &rwr, NULL);
2963 if (ret) {
2964 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2965 return;
2966 }
2967
2968 if (cq->poll_ctx == IB_POLL_DIRECT)
2969 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2970 ib_process_cq_direct(cq, -1);
2971 else
2972 wait_for_completion(&rdrain.done);
2973 }
2974
2975 /*
2976 * __ib_drain_srq() - Block until Last WQE Reached event arrives, or timeout
2977 * expires.
2978 * @qp: queue pair associated with SRQ to drain
2979 *
2980 * Quoting 10.3.1 Queue Pair and EE Context States:
2981 *
2982 * Note, for QPs that are associated with an SRQ, the Consumer should take the
2983 * QP through the Error State before invoking a Destroy QP or a Modify QP to the
2984 * Reset State. The Consumer may invoke the Destroy QP without first performing
2985 * a Modify QP to the Error State and waiting for the Affiliated Asynchronous
2986 * Last WQE Reached Event. However, if the Consumer does not wait for the
2987 * Affiliated Asynchronous Last WQE Reached Event, then WQE and Data Segment
2988 * leakage may occur. Therefore, it is good programming practice to tear down a
2989 * QP that is associated with an SRQ by using the following process:
2990 *
2991 * - Put the QP in the Error State
2992 * - Wait for the Affiliated Asynchronous Last WQE Reached Event;
2993 * - either:
2994 * drain the CQ by invoking the Poll CQ verb and either wait for CQ
2995 * to be empty or the number of Poll CQ operations has exceeded
2996 * CQ capacity size;
2997 * - or
2998 * post another WR that completes on the same CQ and wait for this
2999 * WR to return as a WC;
3000 * - and then invoke a Destroy QP or Reset QP.
3001 *
3002 * We use the first option.
3003 */
__ib_drain_srq(struct ib_qp * qp)3004 static void __ib_drain_srq(struct ib_qp *qp)
3005 {
3006 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
3007 struct ib_cq *cq;
3008 int n, polled = 0;
3009 int ret;
3010
3011 if (!qp->srq) {
3012 WARN_ONCE(1, "QP 0x%p is not associated with SRQ\n", qp);
3013 return;
3014 }
3015
3016 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
3017 if (ret) {
3018 WARN_ONCE(ret, "failed to drain shared recv queue: %d\n", ret);
3019 return;
3020 }
3021
3022 if (ib_srq_has_cq(qp->srq->srq_type)) {
3023 cq = qp->srq->ext.cq;
3024 } else if (qp->recv_cq) {
3025 cq = qp->recv_cq;
3026 } else {
3027 WARN_ONCE(1, "QP 0x%p has no CQ associated with SRQ\n", qp);
3028 return;
3029 }
3030
3031 if (wait_for_completion_timeout(&qp->srq_completion, 60 * HZ) > 0) {
3032 while (polled != cq->cqe) {
3033 n = ib_process_cq_direct(cq, cq->cqe - polled);
3034 if (!n)
3035 return;
3036 polled += n;
3037 }
3038 }
3039 }
3040
3041 /**
3042 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
3043 * application.
3044 * @qp: queue pair to drain
3045 *
3046 * If the device has a provider-specific drain function, then
3047 * call that. Otherwise call the generic drain function
3048 * __ib_drain_sq().
3049 *
3050 * The caller must:
3051 *
3052 * ensure there is room in the CQ and SQ for the drain work request and
3053 * completion.
3054 *
3055 * allocate the CQ using ib_alloc_cq().
3056 *
3057 * ensure that there are no other contexts that are posting WRs concurrently.
3058 * Otherwise the drain is not guaranteed.
3059 */
ib_drain_sq(struct ib_qp * qp)3060 void ib_drain_sq(struct ib_qp *qp)
3061 {
3062 if (qp->device->ops.drain_sq)
3063 qp->device->ops.drain_sq(qp);
3064 else
3065 __ib_drain_sq(qp);
3066 trace_cq_drain_complete(qp->send_cq);
3067 }
3068 EXPORT_SYMBOL(ib_drain_sq);
3069
3070 /**
3071 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
3072 * application.
3073 * @qp: queue pair to drain
3074 *
3075 * If the device has a provider-specific drain function, then
3076 * call that. Otherwise call the generic drain function
3077 * __ib_drain_rq().
3078 *
3079 * The caller must:
3080 *
3081 * ensure there is room in the CQ and RQ for the drain work request and
3082 * completion.
3083 *
3084 * allocate the CQ using ib_alloc_cq().
3085 *
3086 * ensure that there are no other contexts that are posting WRs concurrently.
3087 * Otherwise the drain is not guaranteed.
3088 */
ib_drain_rq(struct ib_qp * qp)3089 void ib_drain_rq(struct ib_qp *qp)
3090 {
3091 if (qp->device->ops.drain_rq)
3092 qp->device->ops.drain_rq(qp);
3093 else
3094 __ib_drain_rq(qp);
3095 trace_cq_drain_complete(qp->recv_cq);
3096 }
3097 EXPORT_SYMBOL(ib_drain_rq);
3098
3099 /**
3100 * ib_drain_qp() - Block until all CQEs have been consumed by the
3101 * application on both the RQ and SQ.
3102 * @qp: queue pair to drain
3103 *
3104 * The caller must:
3105 *
3106 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
3107 * and completions.
3108 *
3109 * allocate the CQs using ib_alloc_cq().
3110 *
3111 * ensure that there are no other contexts that are posting WRs concurrently.
3112 * Otherwise the drain is not guaranteed.
3113 */
ib_drain_qp(struct ib_qp * qp)3114 void ib_drain_qp(struct ib_qp *qp)
3115 {
3116 ib_drain_sq(qp);
3117 if (!qp->srq)
3118 ib_drain_rq(qp);
3119 else
3120 __ib_drain_srq(qp);
3121 }
3122 EXPORT_SYMBOL(ib_drain_qp);
3123
rdma_alloc_netdev(struct ib_device * device,u32 port_num,enum rdma_netdev_t type,const char * name,unsigned char name_assign_type,void (* setup)(struct net_device *))3124 struct net_device *rdma_alloc_netdev(struct ib_device *device, u32 port_num,
3125 enum rdma_netdev_t type, const char *name,
3126 unsigned char name_assign_type,
3127 void (*setup)(struct net_device *))
3128 {
3129 struct rdma_netdev_alloc_params params;
3130 struct net_device *netdev;
3131 int rc;
3132
3133 if (!device->ops.rdma_netdev_get_params)
3134 return ERR_PTR(-EOPNOTSUPP);
3135
3136 rc = device->ops.rdma_netdev_get_params(device, port_num, type,
3137 ¶ms);
3138 if (rc)
3139 return ERR_PTR(rc);
3140
3141 netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type,
3142 setup, params.txqs, params.rxqs);
3143 if (!netdev)
3144 return ERR_PTR(-ENOMEM);
3145
3146 return netdev;
3147 }
3148 EXPORT_SYMBOL(rdma_alloc_netdev);
3149
rdma_init_netdev(struct ib_device * device,u32 port_num,enum rdma_netdev_t type,const char * name,unsigned char name_assign_type,void (* setup)(struct net_device *),struct net_device * netdev)3150 int rdma_init_netdev(struct ib_device *device, u32 port_num,
3151 enum rdma_netdev_t type, const char *name,
3152 unsigned char name_assign_type,
3153 void (*setup)(struct net_device *),
3154 struct net_device *netdev)
3155 {
3156 struct rdma_netdev_alloc_params params;
3157 int rc;
3158
3159 if (!device->ops.rdma_netdev_get_params)
3160 return -EOPNOTSUPP;
3161
3162 rc = device->ops.rdma_netdev_get_params(device, port_num, type,
3163 ¶ms);
3164 if (rc)
3165 return rc;
3166
3167 return params.initialize_rdma_netdev(device, port_num,
3168 netdev, params.param);
3169 }
3170 EXPORT_SYMBOL(rdma_init_netdev);
3171
3172 /**
3173 * rdma_alloc_hw_stats_struct - Helper function to allocate dynamic struct
3174 * for the drivers.
3175 * @descs: array of static descriptors
3176 * @num_counters: number of elements in array
3177 * @lifespan: milliseconds between updates
3178 */
rdma_alloc_hw_stats_struct(const struct rdma_stat_desc * descs,int num_counters,unsigned long lifespan)3179 struct rdma_hw_stats *rdma_alloc_hw_stats_struct(
3180 const struct rdma_stat_desc *descs, int num_counters,
3181 unsigned long lifespan)
3182 {
3183 struct rdma_hw_stats *stats;
3184
3185 stats = kzalloc_flex(*stats, value, num_counters);
3186 if (!stats)
3187 return NULL;
3188
3189 stats->is_disabled = kcalloc(BITS_TO_LONGS(num_counters),
3190 sizeof(*stats->is_disabled), GFP_KERNEL);
3191 if (!stats->is_disabled)
3192 goto err;
3193
3194 stats->descs = descs;
3195 stats->num_counters = num_counters;
3196 stats->lifespan = msecs_to_jiffies(lifespan);
3197 mutex_init(&stats->lock);
3198
3199 return stats;
3200
3201 err:
3202 kfree(stats);
3203 return NULL;
3204 }
3205 EXPORT_SYMBOL(rdma_alloc_hw_stats_struct);
3206
3207 /**
3208 * rdma_free_hw_stats_struct - Helper function to release rdma_hw_stats
3209 * @stats: statistics to release
3210 */
rdma_free_hw_stats_struct(struct rdma_hw_stats * stats)3211 void rdma_free_hw_stats_struct(struct rdma_hw_stats *stats)
3212 {
3213 if (!stats)
3214 return;
3215
3216 kfree(stats->is_disabled);
3217 kfree(stats);
3218 }
3219 EXPORT_SYMBOL(rdma_free_hw_stats_struct);
3220