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.
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:
16 * Redistribution and use in source and binary forms, with or
17 * without modification, are permitted provided that the following
20 * - Redistributions of source code must retain the above
21 * copyright notice, this list of conditions and the following
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.
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
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>
45 #include <linux/in6.h>
46 #include <net/addrconf.h>
47 #include <linux/security.h>
49 #include <rdma/ib_verbs.h>
50 #include <rdma/ib_cache.h>
51 #include <rdma/ib_addr.h>
54 #include "core_priv.h"
55 #include <trace/events/rdma_core.h>
57 #include <trace/events/rdma_core.h>
59 static int ib_resolve_eth_dmac(struct ib_device *device,
60 struct rdma_ah_attr *ah_attr);
62 static const char * const ib_events[] = {
63 [IB_EVENT_CQ_ERR] = "CQ error",
64 [IB_EVENT_QP_FATAL] = "QP fatal error",
65 [IB_EVENT_QP_REQ_ERR] = "QP request error",
66 [IB_EVENT_QP_ACCESS_ERR] = "QP access error",
67 [IB_EVENT_COMM_EST] = "communication established",
68 [IB_EVENT_SQ_DRAINED] = "send queue drained",
69 [IB_EVENT_PATH_MIG] = "path migration successful",
70 [IB_EVENT_PATH_MIG_ERR] = "path migration error",
71 [IB_EVENT_DEVICE_FATAL] = "device fatal error",
72 [IB_EVENT_PORT_ACTIVE] = "port active",
73 [IB_EVENT_PORT_ERR] = "port error",
74 [IB_EVENT_LID_CHANGE] = "LID change",
75 [IB_EVENT_PKEY_CHANGE] = "P_key change",
76 [IB_EVENT_SM_CHANGE] = "SM change",
77 [IB_EVENT_SRQ_ERR] = "SRQ error",
78 [IB_EVENT_SRQ_LIMIT_REACHED] = "SRQ limit reached",
79 [IB_EVENT_QP_LAST_WQE_REACHED] = "last WQE reached",
80 [IB_EVENT_CLIENT_REREGISTER] = "client reregister",
81 [IB_EVENT_GID_CHANGE] = "GID changed",
84 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
88 return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
89 ib_events[index] : "unrecognized event";
91 EXPORT_SYMBOL(ib_event_msg);
93 static const char * const wc_statuses[] = {
94 [IB_WC_SUCCESS] = "success",
95 [IB_WC_LOC_LEN_ERR] = "local length error",
96 [IB_WC_LOC_QP_OP_ERR] = "local QP operation error",
97 [IB_WC_LOC_EEC_OP_ERR] = "local EE context operation error",
98 [IB_WC_LOC_PROT_ERR] = "local protection error",
99 [IB_WC_WR_FLUSH_ERR] = "WR flushed",
100 [IB_WC_MW_BIND_ERR] = "memory management operation error",
101 [IB_WC_BAD_RESP_ERR] = "bad response error",
102 [IB_WC_LOC_ACCESS_ERR] = "local access error",
103 [IB_WC_REM_INV_REQ_ERR] = "invalid request error",
104 [IB_WC_REM_ACCESS_ERR] = "remote access error",
105 [IB_WC_REM_OP_ERR] = "remote operation error",
106 [IB_WC_RETRY_EXC_ERR] = "transport retry counter exceeded",
107 [IB_WC_RNR_RETRY_EXC_ERR] = "RNR retry counter exceeded",
108 [IB_WC_LOC_RDD_VIOL_ERR] = "local RDD violation error",
109 [IB_WC_REM_INV_RD_REQ_ERR] = "remote invalid RD request",
110 [IB_WC_REM_ABORT_ERR] = "operation aborted",
111 [IB_WC_INV_EECN_ERR] = "invalid EE context number",
112 [IB_WC_INV_EEC_STATE_ERR] = "invalid EE context state",
113 [IB_WC_FATAL_ERR] = "fatal error",
114 [IB_WC_RESP_TIMEOUT_ERR] = "response timeout error",
115 [IB_WC_GENERAL_ERR] = "general error",
118 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
120 size_t index = status;
122 return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
123 wc_statuses[index] : "unrecognized status";
125 EXPORT_SYMBOL(ib_wc_status_msg);
127 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
130 case IB_RATE_2_5_GBPS: return 1;
131 case IB_RATE_5_GBPS: return 2;
132 case IB_RATE_10_GBPS: return 4;
133 case IB_RATE_20_GBPS: return 8;
134 case IB_RATE_30_GBPS: return 12;
135 case IB_RATE_40_GBPS: return 16;
136 case IB_RATE_60_GBPS: return 24;
137 case IB_RATE_80_GBPS: return 32;
138 case IB_RATE_120_GBPS: return 48;
139 case IB_RATE_14_GBPS: return 6;
140 case IB_RATE_56_GBPS: return 22;
141 case IB_RATE_112_GBPS: return 45;
142 case IB_RATE_168_GBPS: return 67;
143 case IB_RATE_25_GBPS: return 10;
144 case IB_RATE_100_GBPS: return 40;
145 case IB_RATE_200_GBPS: return 80;
146 case IB_RATE_300_GBPS: return 120;
147 case IB_RATE_28_GBPS: return 11;
148 case IB_RATE_50_GBPS: return 20;
149 case IB_RATE_400_GBPS: return 160;
150 case IB_RATE_600_GBPS: return 240;
154 EXPORT_SYMBOL(ib_rate_to_mult);
156 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
159 case 1: return IB_RATE_2_5_GBPS;
160 case 2: return IB_RATE_5_GBPS;
161 case 4: return IB_RATE_10_GBPS;
162 case 8: return IB_RATE_20_GBPS;
163 case 12: return IB_RATE_30_GBPS;
164 case 16: return IB_RATE_40_GBPS;
165 case 24: return IB_RATE_60_GBPS;
166 case 32: return IB_RATE_80_GBPS;
167 case 48: return IB_RATE_120_GBPS;
168 case 6: return IB_RATE_14_GBPS;
169 case 22: return IB_RATE_56_GBPS;
170 case 45: return IB_RATE_112_GBPS;
171 case 67: return IB_RATE_168_GBPS;
172 case 10: return IB_RATE_25_GBPS;
173 case 40: return IB_RATE_100_GBPS;
174 case 80: return IB_RATE_200_GBPS;
175 case 120: return IB_RATE_300_GBPS;
176 case 11: return IB_RATE_28_GBPS;
177 case 20: return IB_RATE_50_GBPS;
178 case 160: return IB_RATE_400_GBPS;
179 case 240: return IB_RATE_600_GBPS;
180 default: return IB_RATE_PORT_CURRENT;
183 EXPORT_SYMBOL(mult_to_ib_rate);
185 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
188 case IB_RATE_2_5_GBPS: return 2500;
189 case IB_RATE_5_GBPS: return 5000;
190 case IB_RATE_10_GBPS: return 10000;
191 case IB_RATE_20_GBPS: return 20000;
192 case IB_RATE_30_GBPS: return 30000;
193 case IB_RATE_40_GBPS: return 40000;
194 case IB_RATE_60_GBPS: return 60000;
195 case IB_RATE_80_GBPS: return 80000;
196 case IB_RATE_120_GBPS: return 120000;
197 case IB_RATE_14_GBPS: return 14062;
198 case IB_RATE_56_GBPS: return 56250;
199 case IB_RATE_112_GBPS: return 112500;
200 case IB_RATE_168_GBPS: return 168750;
201 case IB_RATE_25_GBPS: return 25781;
202 case IB_RATE_100_GBPS: return 103125;
203 case IB_RATE_200_GBPS: return 206250;
204 case IB_RATE_300_GBPS: return 309375;
205 case IB_RATE_28_GBPS: return 28125;
206 case IB_RATE_50_GBPS: return 53125;
207 case IB_RATE_400_GBPS: return 425000;
208 case IB_RATE_600_GBPS: return 637500;
212 EXPORT_SYMBOL(ib_rate_to_mbps);
214 __attribute_const__ enum rdma_transport_type
215 rdma_node_get_transport(unsigned int node_type)
218 if (node_type == RDMA_NODE_USNIC)
219 return RDMA_TRANSPORT_USNIC;
220 if (node_type == RDMA_NODE_USNIC_UDP)
221 return RDMA_TRANSPORT_USNIC_UDP;
222 if (node_type == RDMA_NODE_RNIC)
223 return RDMA_TRANSPORT_IWARP;
224 if (node_type == RDMA_NODE_UNSPECIFIED)
225 return RDMA_TRANSPORT_UNSPECIFIED;
227 return RDMA_TRANSPORT_IB;
229 EXPORT_SYMBOL(rdma_node_get_transport);
231 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
233 enum rdma_transport_type lt;
234 if (device->ops.get_link_layer)
235 return device->ops.get_link_layer(device, port_num);
237 lt = rdma_node_get_transport(device->node_type);
238 if (lt == RDMA_TRANSPORT_IB)
239 return IB_LINK_LAYER_INFINIBAND;
241 return IB_LINK_LAYER_ETHERNET;
243 EXPORT_SYMBOL(rdma_port_get_link_layer);
245 /* Protection domains */
248 * ib_alloc_pd - Allocates an unused protection domain.
249 * @device: The device on which to allocate the protection domain.
250 * @flags: protection domain flags
251 * @caller: caller's build-time module name
253 * A protection domain object provides an association between QPs, shared
254 * receive queues, address handles, memory regions, and memory windows.
256 * Every PD has a local_dma_lkey which can be used as the lkey value for local
259 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
263 int mr_access_flags = 0;
266 pd = rdma_zalloc_drv_obj(device, ib_pd);
268 return ERR_PTR(-ENOMEM);
272 pd->__internal_mr = NULL;
273 atomic_set(&pd->usecnt, 0);
276 pd->res.type = RDMA_RESTRACK_PD;
277 rdma_restrack_set_task(&pd->res, caller);
279 ret = device->ops.alloc_pd(pd, NULL);
284 rdma_restrack_kadd(&pd->res);
286 if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
287 pd->local_dma_lkey = device->local_dma_lkey;
289 mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
291 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
292 pr_warn("%s: enabling unsafe global rkey\n", caller);
293 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
296 if (mr_access_flags) {
299 mr = pd->device->ops.get_dma_mr(pd, mr_access_flags);
305 mr->device = pd->device;
307 mr->type = IB_MR_TYPE_DMA;
309 mr->need_inval = false;
311 pd->__internal_mr = mr;
313 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
314 pd->local_dma_lkey = pd->__internal_mr->lkey;
316 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
317 pd->unsafe_global_rkey = pd->__internal_mr->rkey;
322 EXPORT_SYMBOL(__ib_alloc_pd);
325 * ib_dealloc_pd_user - Deallocates a protection domain.
326 * @pd: The protection domain to deallocate.
327 * @udata: Valid user data or NULL for kernel object
329 * It is an error to call this function while any resources in the pd still
330 * exist. The caller is responsible to synchronously destroy them and
331 * guarantee no new allocations will happen.
333 void ib_dealloc_pd_user(struct ib_pd *pd, struct ib_udata *udata)
337 if (pd->__internal_mr) {
338 ret = pd->device->ops.dereg_mr(pd->__internal_mr, NULL);
340 pd->__internal_mr = NULL;
343 /* uverbs manipulates usecnt with proper locking, while the kabi
344 requires the caller to guarantee we can't race here. */
345 WARN_ON(atomic_read(&pd->usecnt));
347 rdma_restrack_del(&pd->res);
348 pd->device->ops.dealloc_pd(pd, udata);
351 EXPORT_SYMBOL(ib_dealloc_pd_user);
353 /* Address handles */
356 * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
357 * @dest: Pointer to destination ah_attr. Contents of the destination
358 * pointer is assumed to be invalid and attribute are overwritten.
359 * @src: Pointer to source ah_attr.
361 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
362 const struct rdma_ah_attr *src)
365 if (dest->grh.sgid_attr)
366 rdma_hold_gid_attr(dest->grh.sgid_attr);
368 EXPORT_SYMBOL(rdma_copy_ah_attr);
371 * rdma_replace_ah_attr - Replace valid ah_attr with new new one.
372 * @old: Pointer to existing ah_attr which needs to be replaced.
373 * old is assumed to be valid or zero'd
374 * @new: Pointer to the new ah_attr.
376 * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
377 * old the ah_attr is valid; after that it copies the new attribute and holds
378 * the reference to the replaced ah_attr.
380 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
381 const struct rdma_ah_attr *new)
383 rdma_destroy_ah_attr(old);
385 if (old->grh.sgid_attr)
386 rdma_hold_gid_attr(old->grh.sgid_attr);
388 EXPORT_SYMBOL(rdma_replace_ah_attr);
391 * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
392 * @dest: Pointer to destination ah_attr to copy to.
393 * dest is assumed to be valid or zero'd
394 * @src: Pointer to the new ah_attr.
396 * rdma_move_ah_attr() first releases any reference in the destination ah_attr
397 * if it is valid. This also transfers ownership of internal references from
398 * src to dest, making src invalid in the process. No new reference of the src
401 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
403 rdma_destroy_ah_attr(dest);
405 src->grh.sgid_attr = NULL;
407 EXPORT_SYMBOL(rdma_move_ah_attr);
410 * Validate that the rdma_ah_attr is valid for the device before passing it
413 static int rdma_check_ah_attr(struct ib_device *device,
414 struct rdma_ah_attr *ah_attr)
416 if (!rdma_is_port_valid(device, ah_attr->port_num))
419 if ((rdma_is_grh_required(device, ah_attr->port_num) ||
420 ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
421 !(ah_attr->ah_flags & IB_AH_GRH))
424 if (ah_attr->grh.sgid_attr) {
426 * Make sure the passed sgid_attr is consistent with the
429 if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
430 ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
437 * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
438 * On success the caller is responsible to call rdma_unfill_sgid_attr().
440 static int rdma_fill_sgid_attr(struct ib_device *device,
441 struct rdma_ah_attr *ah_attr,
442 const struct ib_gid_attr **old_sgid_attr)
444 const struct ib_gid_attr *sgid_attr;
445 struct ib_global_route *grh;
448 *old_sgid_attr = ah_attr->grh.sgid_attr;
450 ret = rdma_check_ah_attr(device, ah_attr);
454 if (!(ah_attr->ah_flags & IB_AH_GRH))
457 grh = rdma_ah_retrieve_grh(ah_attr);
462 rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
463 if (IS_ERR(sgid_attr))
464 return PTR_ERR(sgid_attr);
466 /* Move ownerhip of the kref into the ah_attr */
467 grh->sgid_attr = sgid_attr;
471 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
472 const struct ib_gid_attr *old_sgid_attr)
475 * Fill didn't change anything, the caller retains ownership of
478 if (ah_attr->grh.sgid_attr == old_sgid_attr)
482 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
483 * doesn't see any change in the rdma_ah_attr. If we get here
484 * old_sgid_attr is NULL.
486 rdma_destroy_ah_attr(ah_attr);
489 static const struct ib_gid_attr *
490 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
491 const struct ib_gid_attr *old_attr)
494 rdma_put_gid_attr(old_attr);
495 if (ah_attr->ah_flags & IB_AH_GRH) {
496 rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
497 return ah_attr->grh.sgid_attr;
502 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
503 struct rdma_ah_attr *ah_attr,
505 struct ib_udata *udata)
507 struct ib_device *device = pd->device;
511 might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE);
513 if (!device->ops.create_ah)
514 return ERR_PTR(-EOPNOTSUPP);
516 ah = rdma_zalloc_drv_obj_gfp(
518 (flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC);
520 return ERR_PTR(-ENOMEM);
524 ah->type = ah_attr->type;
525 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
527 ret = device->ops.create_ah(ah, ah_attr, flags, udata);
533 atomic_inc(&pd->usecnt);
538 * rdma_create_ah - Creates an address handle for the
539 * given address vector.
540 * @pd: The protection domain associated with the address handle.
541 * @ah_attr: The attributes of the address vector.
542 * @flags: Create address handle flags (see enum rdma_create_ah_flags).
544 * It returns 0 on success and returns appropriate error code on error.
545 * The address handle is used to reference a local or global destination
546 * in all UD QP post sends.
548 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
551 const struct ib_gid_attr *old_sgid_attr;
555 ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
559 ah = _rdma_create_ah(pd, ah_attr, flags, NULL);
561 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
564 EXPORT_SYMBOL(rdma_create_ah);
567 * rdma_create_user_ah - Creates an address handle for the
568 * given address vector.
569 * It resolves destination mac address for ah attribute of RoCE type.
570 * @pd: The protection domain associated with the address handle.
571 * @ah_attr: The attributes of the address vector.
572 * @udata: pointer to user's input output buffer information need by
575 * It returns 0 on success and returns appropriate error code on error.
576 * The address handle is used to reference a local or global destination
577 * in all UD QP post sends.
579 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
580 struct rdma_ah_attr *ah_attr,
581 struct ib_udata *udata)
583 const struct ib_gid_attr *old_sgid_attr;
587 err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
591 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
592 err = ib_resolve_eth_dmac(pd->device, ah_attr);
599 ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE, udata);
602 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
605 EXPORT_SYMBOL(rdma_create_user_ah);
607 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
609 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
610 struct iphdr ip4h_checked;
611 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
613 /* If it's IPv6, the version must be 6, otherwise, the first
614 * 20 bytes (before the IPv4 header) are garbled.
616 if (ip6h->version != 6)
617 return (ip4h->version == 4) ? 4 : 0;
618 /* version may be 6 or 4 because the first 20 bytes could be garbled */
620 /* RoCE v2 requires no options, thus header length
627 * We can't write on scattered buffers so we need to copy to
630 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
631 ip4h_checked.check = 0;
632 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
633 /* if IPv4 header checksum is OK, believe it */
634 if (ip4h->check == ip4h_checked.check)
638 EXPORT_SYMBOL(ib_get_rdma_header_version);
640 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
642 const struct ib_grh *grh)
646 if (rdma_protocol_ib(device, port_num))
647 return RDMA_NETWORK_IB;
649 grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
651 if (grh_version == 4)
652 return RDMA_NETWORK_IPV4;
654 if (grh->next_hdr == IPPROTO_UDP)
655 return RDMA_NETWORK_IPV6;
657 return RDMA_NETWORK_ROCE_V1;
660 struct find_gid_index_context {
662 enum ib_gid_type gid_type;
665 static bool find_gid_index(const union ib_gid *gid,
666 const struct ib_gid_attr *gid_attr,
669 struct find_gid_index_context *ctx = context;
670 u16 vlan_id = 0xffff;
673 if (ctx->gid_type != gid_attr->gid_type)
676 ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
680 return ctx->vlan_id == vlan_id;
683 static const struct ib_gid_attr *
684 get_sgid_attr_from_eth(struct ib_device *device, u8 port_num,
685 u16 vlan_id, const union ib_gid *sgid,
686 enum ib_gid_type gid_type)
688 struct find_gid_index_context context = {.vlan_id = vlan_id,
689 .gid_type = gid_type};
691 return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
695 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
696 enum rdma_network_type net_type,
697 union ib_gid *sgid, union ib_gid *dgid)
699 struct sockaddr_in src_in;
700 struct sockaddr_in dst_in;
701 __be32 src_saddr, dst_saddr;
706 if (net_type == RDMA_NETWORK_IPV4) {
707 memcpy(&src_in.sin_addr.s_addr,
708 &hdr->roce4grh.saddr, 4);
709 memcpy(&dst_in.sin_addr.s_addr,
710 &hdr->roce4grh.daddr, 4);
711 src_saddr = src_in.sin_addr.s_addr;
712 dst_saddr = dst_in.sin_addr.s_addr;
713 ipv6_addr_set_v4mapped(src_saddr,
714 (struct in6_addr *)sgid);
715 ipv6_addr_set_v4mapped(dst_saddr,
716 (struct in6_addr *)dgid);
718 } else if (net_type == RDMA_NETWORK_IPV6 ||
719 net_type == RDMA_NETWORK_IB) {
720 *dgid = hdr->ibgrh.dgid;
721 *sgid = hdr->ibgrh.sgid;
727 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
729 /* Resolve destination mac address and hop limit for unicast destination
730 * GID entry, considering the source GID entry as well.
731 * ah_attribute must have have valid port_num, sgid_index.
733 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
734 struct rdma_ah_attr *ah_attr)
736 struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
737 const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
738 int hop_limit = 0xff;
741 /* If destination is link local and source GID is RoCEv1,
742 * IP stack is not used.
744 if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
745 sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
746 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
751 ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
753 sgid_attr, &hop_limit);
755 grh->hop_limit = hop_limit;
760 * This function initializes address handle attributes from the incoming packet.
761 * Incoming packet has dgid of the receiver node on which this code is
762 * getting executed and, sgid contains the GID of the sender.
764 * When resolving mac address of destination, the arrived dgid is used
765 * as sgid and, sgid is used as dgid because sgid contains destinations
766 * GID whom to respond to.
768 * On success the caller is responsible to call rdma_destroy_ah_attr on the
771 int ib_init_ah_attr_from_wc(struct ib_device *device, u8 port_num,
772 const struct ib_wc *wc, const struct ib_grh *grh,
773 struct rdma_ah_attr *ah_attr)
777 enum rdma_network_type net_type = RDMA_NETWORK_IB;
778 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
779 const struct ib_gid_attr *sgid_attr;
786 memset(ah_attr, 0, sizeof *ah_attr);
787 ah_attr->type = rdma_ah_find_type(device, port_num);
788 if (rdma_cap_eth_ah(device, port_num)) {
789 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
790 net_type = wc->network_hdr_type;
792 net_type = ib_get_net_type_by_grh(device, port_num, grh);
793 gid_type = ib_network_to_gid_type(net_type);
795 ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
800 rdma_ah_set_sl(ah_attr, wc->sl);
801 rdma_ah_set_port_num(ah_attr, port_num);
803 if (rdma_protocol_roce(device, port_num)) {
804 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
805 wc->vlan_id : 0xffff;
807 if (!(wc->wc_flags & IB_WC_GRH))
810 sgid_attr = get_sgid_attr_from_eth(device, port_num,
813 if (IS_ERR(sgid_attr))
814 return PTR_ERR(sgid_attr);
816 flow_class = be32_to_cpu(grh->version_tclass_flow);
817 rdma_move_grh_sgid_attr(ah_attr,
819 flow_class & 0xFFFFF,
821 (flow_class >> 20) & 0xFF,
824 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
826 rdma_destroy_ah_attr(ah_attr);
830 rdma_ah_set_dlid(ah_attr, wc->slid);
831 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
833 if ((wc->wc_flags & IB_WC_GRH) == 0)
836 if (dgid.global.interface_id !=
837 cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
838 sgid_attr = rdma_find_gid_by_port(
839 device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
841 sgid_attr = rdma_get_gid_attr(device, port_num, 0);
843 if (IS_ERR(sgid_attr))
844 return PTR_ERR(sgid_attr);
845 flow_class = be32_to_cpu(grh->version_tclass_flow);
846 rdma_move_grh_sgid_attr(ah_attr,
848 flow_class & 0xFFFFF,
850 (flow_class >> 20) & 0xFF,
856 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
859 * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
862 * @attr: Pointer to AH attribute structure
863 * @dgid: Destination GID
864 * @flow_label: Flow label
865 * @hop_limit: Hop limit
866 * @traffic_class: traffic class
867 * @sgid_attr: Pointer to SGID attribute
869 * This takes ownership of the sgid_attr reference. The caller must ensure
870 * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
871 * calling this function.
873 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
874 u32 flow_label, u8 hop_limit, u8 traffic_class,
875 const struct ib_gid_attr *sgid_attr)
877 rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
879 attr->grh.sgid_attr = sgid_attr;
881 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
884 * rdma_destroy_ah_attr - Release reference to SGID attribute of
886 * @ah_attr: Pointer to ah attribute
888 * Release reference to the SGID attribute of the ah attribute if it is
889 * non NULL. It is safe to call this multiple times, and safe to call it on
890 * a zero initialized ah_attr.
892 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
894 if (ah_attr->grh.sgid_attr) {
895 rdma_put_gid_attr(ah_attr->grh.sgid_attr);
896 ah_attr->grh.sgid_attr = NULL;
899 EXPORT_SYMBOL(rdma_destroy_ah_attr);
901 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
902 const struct ib_grh *grh, u8 port_num)
904 struct rdma_ah_attr ah_attr;
908 ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
912 ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE);
914 rdma_destroy_ah_attr(&ah_attr);
917 EXPORT_SYMBOL(ib_create_ah_from_wc);
919 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
921 const struct ib_gid_attr *old_sgid_attr;
924 if (ah->type != ah_attr->type)
927 ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
931 ret = ah->device->ops.modify_ah ?
932 ah->device->ops.modify_ah(ah, ah_attr) :
935 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
936 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
939 EXPORT_SYMBOL(rdma_modify_ah);
941 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
943 ah_attr->grh.sgid_attr = NULL;
945 return ah->device->ops.query_ah ?
946 ah->device->ops.query_ah(ah, ah_attr) :
949 EXPORT_SYMBOL(rdma_query_ah);
951 int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata)
953 const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
956 might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE);
960 ah->device->ops.destroy_ah(ah, flags);
961 atomic_dec(&pd->usecnt);
963 rdma_put_gid_attr(sgid_attr);
968 EXPORT_SYMBOL(rdma_destroy_ah_user);
970 /* Shared receive queues */
972 struct ib_srq *ib_create_srq(struct ib_pd *pd,
973 struct ib_srq_init_attr *srq_init_attr)
978 if (!pd->device->ops.create_srq)
979 return ERR_PTR(-EOPNOTSUPP);
981 srq = rdma_zalloc_drv_obj(pd->device, ib_srq);
983 return ERR_PTR(-ENOMEM);
985 srq->device = pd->device;
987 srq->event_handler = srq_init_attr->event_handler;
988 srq->srq_context = srq_init_attr->srq_context;
989 srq->srq_type = srq_init_attr->srq_type;
991 if (ib_srq_has_cq(srq->srq_type)) {
992 srq->ext.cq = srq_init_attr->ext.cq;
993 atomic_inc(&srq->ext.cq->usecnt);
995 if (srq->srq_type == IB_SRQT_XRC) {
996 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
997 atomic_inc(&srq->ext.xrc.xrcd->usecnt);
999 atomic_inc(&pd->usecnt);
1001 ret = pd->device->ops.create_srq(srq, srq_init_attr, NULL);
1003 atomic_dec(&srq->pd->usecnt);
1004 if (srq->srq_type == IB_SRQT_XRC)
1005 atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1006 if (ib_srq_has_cq(srq->srq_type))
1007 atomic_dec(&srq->ext.cq->usecnt);
1009 return ERR_PTR(ret);
1014 EXPORT_SYMBOL(ib_create_srq);
1016 int ib_modify_srq(struct ib_srq *srq,
1017 struct ib_srq_attr *srq_attr,
1018 enum ib_srq_attr_mask srq_attr_mask)
1020 return srq->device->ops.modify_srq ?
1021 srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask,
1022 NULL) : -EOPNOTSUPP;
1024 EXPORT_SYMBOL(ib_modify_srq);
1026 int ib_query_srq(struct ib_srq *srq,
1027 struct ib_srq_attr *srq_attr)
1029 return srq->device->ops.query_srq ?
1030 srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP;
1032 EXPORT_SYMBOL(ib_query_srq);
1034 int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata)
1036 if (atomic_read(&srq->usecnt))
1039 srq->device->ops.destroy_srq(srq, udata);
1041 atomic_dec(&srq->pd->usecnt);
1042 if (srq->srq_type == IB_SRQT_XRC)
1043 atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1044 if (ib_srq_has_cq(srq->srq_type))
1045 atomic_dec(&srq->ext.cq->usecnt);
1050 EXPORT_SYMBOL(ib_destroy_srq_user);
1054 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1056 struct ib_qp *qp = context;
1057 unsigned long flags;
1059 spin_lock_irqsave(&qp->device->event_handler_lock, flags);
1060 list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1061 if (event->element.qp->event_handler)
1062 event->element.qp->event_handler(event, event->element.qp->qp_context);
1063 spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
1066 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
1068 mutex_lock(&xrcd->tgt_qp_mutex);
1069 list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
1070 mutex_unlock(&xrcd->tgt_qp_mutex);
1073 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1074 void (*event_handler)(struct ib_event *, void *),
1078 unsigned long flags;
1081 qp = kzalloc(sizeof *qp, GFP_KERNEL);
1083 return ERR_PTR(-ENOMEM);
1085 qp->real_qp = real_qp;
1086 err = ib_open_shared_qp_security(qp, real_qp->device);
1089 return ERR_PTR(err);
1092 qp->real_qp = real_qp;
1093 atomic_inc(&real_qp->usecnt);
1094 qp->device = real_qp->device;
1095 qp->event_handler = event_handler;
1096 qp->qp_context = qp_context;
1097 qp->qp_num = real_qp->qp_num;
1098 qp->qp_type = real_qp->qp_type;
1100 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1101 list_add(&qp->open_list, &real_qp->open_list);
1102 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1107 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1108 struct ib_qp_open_attr *qp_open_attr)
1110 struct ib_qp *qp, *real_qp;
1112 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1113 return ERR_PTR(-EINVAL);
1115 qp = ERR_PTR(-EINVAL);
1116 mutex_lock(&xrcd->tgt_qp_mutex);
1117 list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
1118 if (real_qp->qp_num == qp_open_attr->qp_num) {
1119 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1120 qp_open_attr->qp_context);
1124 mutex_unlock(&xrcd->tgt_qp_mutex);
1127 EXPORT_SYMBOL(ib_open_qp);
1129 static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp,
1130 struct ib_qp_init_attr *qp_init_attr,
1131 struct ib_udata *udata)
1133 struct ib_qp *real_qp = qp;
1135 qp->event_handler = __ib_shared_qp_event_handler;
1136 qp->qp_context = qp;
1138 qp->send_cq = qp->recv_cq = NULL;
1140 qp->xrcd = qp_init_attr->xrcd;
1141 atomic_inc(&qp_init_attr->xrcd->usecnt);
1142 INIT_LIST_HEAD(&qp->open_list);
1144 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1145 qp_init_attr->qp_context);
1149 __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
1153 struct ib_qp *ib_create_qp_user(struct ib_pd *pd,
1154 struct ib_qp_init_attr *qp_init_attr,
1155 struct ib_udata *udata)
1157 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
1161 if (qp_init_attr->rwq_ind_tbl &&
1162 (qp_init_attr->recv_cq ||
1163 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
1164 qp_init_attr->cap.max_recv_sge))
1165 return ERR_PTR(-EINVAL);
1167 if ((qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN) &&
1168 !(device->attrs.device_cap_flags & IB_DEVICE_INTEGRITY_HANDOVER))
1169 return ERR_PTR(-EINVAL);
1172 * If the callers is using the RDMA API calculate the resources
1173 * needed for the RDMA READ/WRITE operations.
1175 * Note that these callers need to pass in a port number.
1177 if (qp_init_attr->cap.max_rdma_ctxs)
1178 rdma_rw_init_qp(device, qp_init_attr);
1180 qp = _ib_create_qp(device, pd, qp_init_attr, NULL, NULL);
1184 ret = ib_create_qp_security(qp, device);
1188 qp->qp_type = qp_init_attr->qp_type;
1189 qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
1191 atomic_set(&qp->usecnt, 0);
1193 spin_lock_init(&qp->mr_lock);
1194 INIT_LIST_HEAD(&qp->rdma_mrs);
1195 INIT_LIST_HEAD(&qp->sig_mrs);
1198 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT) {
1199 struct ib_qp *xrc_qp =
1200 create_xrc_qp_user(qp, qp_init_attr, udata);
1202 if (IS_ERR(xrc_qp)) {
1203 ret = PTR_ERR(xrc_qp);
1209 qp->event_handler = qp_init_attr->event_handler;
1210 qp->qp_context = qp_init_attr->qp_context;
1211 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
1215 qp->recv_cq = qp_init_attr->recv_cq;
1216 if (qp_init_attr->recv_cq)
1217 atomic_inc(&qp_init_attr->recv_cq->usecnt);
1218 qp->srq = qp_init_attr->srq;
1220 atomic_inc(&qp_init_attr->srq->usecnt);
1223 qp->send_cq = qp_init_attr->send_cq;
1226 atomic_inc(&pd->usecnt);
1227 if (qp_init_attr->send_cq)
1228 atomic_inc(&qp_init_attr->send_cq->usecnt);
1229 if (qp_init_attr->rwq_ind_tbl)
1230 atomic_inc(&qp->rwq_ind_tbl->usecnt);
1232 if (qp_init_attr->cap.max_rdma_ctxs) {
1233 ret = rdma_rw_init_mrs(qp, qp_init_attr);
1239 * Note: all hw drivers guarantee that max_send_sge is lower than
1240 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1241 * max_send_sge <= max_sge_rd.
1243 qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1244 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1245 device->attrs.max_sge_rd);
1246 if (qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN)
1247 qp->integrity_en = true;
1253 return ERR_PTR(ret);
1256 EXPORT_SYMBOL(ib_create_qp_user);
1258 static const struct {
1260 enum ib_qp_attr_mask req_param[IB_QPT_MAX];
1261 enum ib_qp_attr_mask opt_param[IB_QPT_MAX];
1262 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1264 [IB_QPS_RESET] = { .valid = 1 },
1268 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1271 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
1272 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1274 IB_QP_ACCESS_FLAGS),
1275 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1277 IB_QP_ACCESS_FLAGS),
1278 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1280 IB_QP_ACCESS_FLAGS),
1281 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1283 IB_QP_ACCESS_FLAGS),
1284 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1286 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1292 [IB_QPS_RESET] = { .valid = 1 },
1293 [IB_QPS_ERR] = { .valid = 1 },
1297 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1300 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1302 IB_QP_ACCESS_FLAGS),
1303 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1305 IB_QP_ACCESS_FLAGS),
1306 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1308 IB_QP_ACCESS_FLAGS),
1309 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1311 IB_QP_ACCESS_FLAGS),
1312 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1314 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1321 [IB_QPT_UC] = (IB_QP_AV |
1325 [IB_QPT_RC] = (IB_QP_AV |
1329 IB_QP_MAX_DEST_RD_ATOMIC |
1330 IB_QP_MIN_RNR_TIMER),
1331 [IB_QPT_XRC_INI] = (IB_QP_AV |
1335 [IB_QPT_XRC_TGT] = (IB_QP_AV |
1339 IB_QP_MAX_DEST_RD_ATOMIC |
1340 IB_QP_MIN_RNR_TIMER),
1343 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1345 [IB_QPT_UC] = (IB_QP_ALT_PATH |
1346 IB_QP_ACCESS_FLAGS |
1348 [IB_QPT_RC] = (IB_QP_ALT_PATH |
1349 IB_QP_ACCESS_FLAGS |
1351 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH |
1352 IB_QP_ACCESS_FLAGS |
1354 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH |
1355 IB_QP_ACCESS_FLAGS |
1357 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1359 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1365 [IB_QPS_RESET] = { .valid = 1 },
1366 [IB_QPS_ERR] = { .valid = 1 },
1370 [IB_QPT_UD] = IB_QP_SQ_PSN,
1371 [IB_QPT_UC] = IB_QP_SQ_PSN,
1372 [IB_QPT_RC] = (IB_QP_TIMEOUT |
1376 IB_QP_MAX_QP_RD_ATOMIC),
1377 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT |
1381 IB_QP_MAX_QP_RD_ATOMIC),
1382 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT |
1384 [IB_QPT_SMI] = IB_QP_SQ_PSN,
1385 [IB_QPT_GSI] = IB_QP_SQ_PSN,
1388 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1390 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1392 IB_QP_ACCESS_FLAGS |
1393 IB_QP_PATH_MIG_STATE),
1394 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1396 IB_QP_ACCESS_FLAGS |
1397 IB_QP_MIN_RNR_TIMER |
1398 IB_QP_PATH_MIG_STATE),
1399 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1401 IB_QP_ACCESS_FLAGS |
1402 IB_QP_PATH_MIG_STATE),
1403 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1405 IB_QP_ACCESS_FLAGS |
1406 IB_QP_MIN_RNR_TIMER |
1407 IB_QP_PATH_MIG_STATE),
1408 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1410 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1412 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1417 [IB_QPS_RESET] = { .valid = 1 },
1418 [IB_QPS_ERR] = { .valid = 1 },
1422 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1424 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1425 IB_QP_ACCESS_FLAGS |
1427 IB_QP_PATH_MIG_STATE),
1428 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1429 IB_QP_ACCESS_FLAGS |
1431 IB_QP_PATH_MIG_STATE |
1432 IB_QP_MIN_RNR_TIMER),
1433 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1434 IB_QP_ACCESS_FLAGS |
1436 IB_QP_PATH_MIG_STATE),
1437 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1438 IB_QP_ACCESS_FLAGS |
1440 IB_QP_PATH_MIG_STATE |
1441 IB_QP_MIN_RNR_TIMER),
1442 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1444 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1446 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1452 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1453 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1454 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1455 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1456 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1457 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1458 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1463 [IB_QPS_RESET] = { .valid = 1 },
1464 [IB_QPS_ERR] = { .valid = 1 },
1468 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1470 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1472 IB_QP_ACCESS_FLAGS |
1473 IB_QP_PATH_MIG_STATE),
1474 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1476 IB_QP_ACCESS_FLAGS |
1477 IB_QP_MIN_RNR_TIMER |
1478 IB_QP_PATH_MIG_STATE),
1479 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1481 IB_QP_ACCESS_FLAGS |
1482 IB_QP_PATH_MIG_STATE),
1483 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1485 IB_QP_ACCESS_FLAGS |
1486 IB_QP_MIN_RNR_TIMER |
1487 IB_QP_PATH_MIG_STATE),
1488 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1490 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1497 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1499 [IB_QPT_UC] = (IB_QP_AV |
1501 IB_QP_ACCESS_FLAGS |
1503 IB_QP_PATH_MIG_STATE),
1504 [IB_QPT_RC] = (IB_QP_PORT |
1509 IB_QP_MAX_QP_RD_ATOMIC |
1510 IB_QP_MAX_DEST_RD_ATOMIC |
1512 IB_QP_ACCESS_FLAGS |
1514 IB_QP_MIN_RNR_TIMER |
1515 IB_QP_PATH_MIG_STATE),
1516 [IB_QPT_XRC_INI] = (IB_QP_PORT |
1521 IB_QP_MAX_QP_RD_ATOMIC |
1523 IB_QP_ACCESS_FLAGS |
1525 IB_QP_PATH_MIG_STATE),
1526 [IB_QPT_XRC_TGT] = (IB_QP_PORT |
1529 IB_QP_MAX_DEST_RD_ATOMIC |
1531 IB_QP_ACCESS_FLAGS |
1533 IB_QP_MIN_RNR_TIMER |
1534 IB_QP_PATH_MIG_STATE),
1535 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1537 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1543 [IB_QPS_RESET] = { .valid = 1 },
1544 [IB_QPS_ERR] = { .valid = 1 },
1548 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1550 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1551 IB_QP_ACCESS_FLAGS),
1552 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1554 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1560 [IB_QPS_RESET] = { .valid = 1 },
1561 [IB_QPS_ERR] = { .valid = 1 }
1565 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1566 enum ib_qp_type type, enum ib_qp_attr_mask mask)
1568 enum ib_qp_attr_mask req_param, opt_param;
1570 if (mask & IB_QP_CUR_STATE &&
1571 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1572 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1575 if (!qp_state_table[cur_state][next_state].valid)
1578 req_param = qp_state_table[cur_state][next_state].req_param[type];
1579 opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1581 if ((mask & req_param) != req_param)
1584 if (mask & ~(req_param | opt_param | IB_QP_STATE))
1589 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1592 * ib_resolve_eth_dmac - Resolve destination mac address
1593 * @device: Device to consider
1594 * @ah_attr: address handle attribute which describes the
1595 * source and destination parameters
1596 * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1597 * returns 0 on success or appropriate error code. It initializes the
1598 * necessary ah_attr fields when call is successful.
1600 static int ib_resolve_eth_dmac(struct ib_device *device,
1601 struct rdma_ah_attr *ah_attr)
1605 if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1606 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1609 memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1610 ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1612 ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1613 (char *)ah_attr->roce.dmac);
1616 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1621 static bool is_qp_type_connected(const struct ib_qp *qp)
1623 return (qp->qp_type == IB_QPT_UC ||
1624 qp->qp_type == IB_QPT_RC ||
1625 qp->qp_type == IB_QPT_XRC_INI ||
1626 qp->qp_type == IB_QPT_XRC_TGT);
1630 * IB core internal function to perform QP attributes modification.
1632 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1633 int attr_mask, struct ib_udata *udata)
1635 u8 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1636 const struct ib_gid_attr *old_sgid_attr_av;
1637 const struct ib_gid_attr *old_sgid_attr_alt_av;
1640 if (attr_mask & IB_QP_AV) {
1641 ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1646 if (attr_mask & IB_QP_ALT_PATH) {
1648 * FIXME: This does not track the migration state, so if the
1649 * user loads a new alternate path after the HW has migrated
1650 * from primary->alternate we will keep the wrong
1651 * references. This is OK for IB because the reference
1652 * counting does not serve any functional purpose.
1654 ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1655 &old_sgid_attr_alt_av);
1660 * Today the core code can only handle alternate paths and APM
1661 * for IB. Ban them in roce mode.
1663 if (!(rdma_protocol_ib(qp->device,
1664 attr->alt_ah_attr.port_num) &&
1665 rdma_protocol_ib(qp->device, port))) {
1672 * If the user provided the qp_attr then we have to resolve it. Kernel
1673 * users have to provide already resolved rdma_ah_attr's
1675 if (udata && (attr_mask & IB_QP_AV) &&
1676 attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1677 is_qp_type_connected(qp)) {
1678 ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
1683 if (rdma_ib_or_roce(qp->device, port)) {
1684 if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1685 dev_warn(&qp->device->dev,
1686 "%s rq_psn overflow, masking to 24 bits\n",
1688 attr->rq_psn &= 0xffffff;
1691 if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1692 dev_warn(&qp->device->dev,
1693 " %s sq_psn overflow, masking to 24 bits\n",
1695 attr->sq_psn &= 0xffffff;
1700 * Bind this qp to a counter automatically based on the rdma counter
1701 * rules. This only set in RST2INIT with port specified
1703 if (!qp->counter && (attr_mask & IB_QP_PORT) &&
1704 ((attr_mask & IB_QP_STATE) && attr->qp_state == IB_QPS_INIT))
1705 rdma_counter_bind_qp_auto(qp, attr->port_num);
1707 ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1711 if (attr_mask & IB_QP_PORT)
1712 qp->port = attr->port_num;
1713 if (attr_mask & IB_QP_AV)
1715 rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1716 if (attr_mask & IB_QP_ALT_PATH)
1717 qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1718 &attr->alt_ah_attr, qp->alt_path_sgid_attr);
1721 if (attr_mask & IB_QP_ALT_PATH)
1722 rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1724 if (attr_mask & IB_QP_AV)
1725 rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1730 * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1731 * @ib_qp: The QP to modify.
1732 * @attr: On input, specifies the QP attributes to modify. On output,
1733 * the current values of selected QP attributes are returned.
1734 * @attr_mask: A bit-mask used to specify which attributes of the QP
1735 * are being modified.
1736 * @udata: pointer to user's input output buffer information
1737 * are being modified.
1738 * It returns 0 on success and returns appropriate error code on error.
1740 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1741 int attr_mask, struct ib_udata *udata)
1743 return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1745 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1747 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u8 *speed, u8 *width)
1751 struct net_device *netdev;
1752 struct ethtool_link_ksettings lksettings;
1754 if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1757 netdev = ib_device_get_netdev(dev, port_num);
1762 rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1768 netdev_speed = lksettings.base.speed;
1770 netdev_speed = SPEED_1000;
1771 pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name,
1775 if (netdev_speed <= SPEED_1000) {
1776 *width = IB_WIDTH_1X;
1777 *speed = IB_SPEED_SDR;
1778 } else if (netdev_speed <= SPEED_10000) {
1779 *width = IB_WIDTH_1X;
1780 *speed = IB_SPEED_FDR10;
1781 } else if (netdev_speed <= SPEED_20000) {
1782 *width = IB_WIDTH_4X;
1783 *speed = IB_SPEED_DDR;
1784 } else if (netdev_speed <= SPEED_25000) {
1785 *width = IB_WIDTH_1X;
1786 *speed = IB_SPEED_EDR;
1787 } else if (netdev_speed <= SPEED_40000) {
1788 *width = IB_WIDTH_4X;
1789 *speed = IB_SPEED_FDR10;
1791 *width = IB_WIDTH_4X;
1792 *speed = IB_SPEED_EDR;
1797 EXPORT_SYMBOL(ib_get_eth_speed);
1799 int ib_modify_qp(struct ib_qp *qp,
1800 struct ib_qp_attr *qp_attr,
1803 return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1805 EXPORT_SYMBOL(ib_modify_qp);
1807 int ib_query_qp(struct ib_qp *qp,
1808 struct ib_qp_attr *qp_attr,
1810 struct ib_qp_init_attr *qp_init_attr)
1812 qp_attr->ah_attr.grh.sgid_attr = NULL;
1813 qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
1815 return qp->device->ops.query_qp ?
1816 qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask,
1817 qp_init_attr) : -EOPNOTSUPP;
1819 EXPORT_SYMBOL(ib_query_qp);
1821 int ib_close_qp(struct ib_qp *qp)
1823 struct ib_qp *real_qp;
1824 unsigned long flags;
1826 real_qp = qp->real_qp;
1830 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1831 list_del(&qp->open_list);
1832 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1834 atomic_dec(&real_qp->usecnt);
1836 ib_close_shared_qp_security(qp->qp_sec);
1841 EXPORT_SYMBOL(ib_close_qp);
1843 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1845 struct ib_xrcd *xrcd;
1846 struct ib_qp *real_qp;
1849 real_qp = qp->real_qp;
1850 xrcd = real_qp->xrcd;
1852 mutex_lock(&xrcd->tgt_qp_mutex);
1854 if (atomic_read(&real_qp->usecnt) == 0)
1855 list_del(&real_qp->xrcd_list);
1858 mutex_unlock(&xrcd->tgt_qp_mutex);
1861 ret = ib_destroy_qp(real_qp);
1863 atomic_dec(&xrcd->usecnt);
1865 __ib_insert_xrcd_qp(xrcd, real_qp);
1871 int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata)
1873 const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
1874 const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
1876 struct ib_cq *scq, *rcq;
1878 struct ib_rwq_ind_table *ind_tbl;
1879 struct ib_qp_security *sec;
1882 WARN_ON_ONCE(qp->mrs_used > 0);
1884 if (atomic_read(&qp->usecnt))
1887 if (qp->real_qp != qp)
1888 return __ib_destroy_shared_qp(qp);
1894 ind_tbl = qp->rwq_ind_tbl;
1897 ib_destroy_qp_security_begin(sec);
1900 rdma_rw_cleanup_mrs(qp);
1902 rdma_counter_unbind_qp(qp, true);
1903 rdma_restrack_del(&qp->res);
1904 ret = qp->device->ops.destroy_qp(qp, udata);
1906 if (alt_path_sgid_attr)
1907 rdma_put_gid_attr(alt_path_sgid_attr);
1909 rdma_put_gid_attr(av_sgid_attr);
1911 atomic_dec(&pd->usecnt);
1913 atomic_dec(&scq->usecnt);
1915 atomic_dec(&rcq->usecnt);
1917 atomic_dec(&srq->usecnt);
1919 atomic_dec(&ind_tbl->usecnt);
1921 ib_destroy_qp_security_end(sec);
1924 ib_destroy_qp_security_abort(sec);
1929 EXPORT_SYMBOL(ib_destroy_qp_user);
1931 /* Completion queues */
1933 struct ib_cq *__ib_create_cq(struct ib_device *device,
1934 ib_comp_handler comp_handler,
1935 void (*event_handler)(struct ib_event *, void *),
1937 const struct ib_cq_init_attr *cq_attr,
1943 cq = rdma_zalloc_drv_obj(device, ib_cq);
1945 return ERR_PTR(-ENOMEM);
1947 cq->device = device;
1949 cq->comp_handler = comp_handler;
1950 cq->event_handler = event_handler;
1951 cq->cq_context = cq_context;
1952 atomic_set(&cq->usecnt, 0);
1953 cq->res.type = RDMA_RESTRACK_CQ;
1954 rdma_restrack_set_task(&cq->res, caller);
1956 ret = device->ops.create_cq(cq, cq_attr, NULL);
1959 return ERR_PTR(ret);
1962 rdma_restrack_kadd(&cq->res);
1965 EXPORT_SYMBOL(__ib_create_cq);
1967 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1969 return cq->device->ops.modify_cq ?
1970 cq->device->ops.modify_cq(cq, cq_count,
1971 cq_period) : -EOPNOTSUPP;
1973 EXPORT_SYMBOL(rdma_set_cq_moderation);
1975 int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata)
1977 if (atomic_read(&cq->usecnt))
1980 rdma_restrack_del(&cq->res);
1981 cq->device->ops.destroy_cq(cq, udata);
1985 EXPORT_SYMBOL(ib_destroy_cq_user);
1987 int ib_resize_cq(struct ib_cq *cq, int cqe)
1989 return cq->device->ops.resize_cq ?
1990 cq->device->ops.resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
1992 EXPORT_SYMBOL(ib_resize_cq);
1994 /* Memory regions */
1996 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata)
1998 struct ib_pd *pd = mr->pd;
1999 struct ib_dm *dm = mr->dm;
2000 struct ib_sig_attrs *sig_attrs = mr->sig_attrs;
2004 rdma_restrack_del(&mr->res);
2005 ret = mr->device->ops.dereg_mr(mr, udata);
2007 atomic_dec(&pd->usecnt);
2009 atomic_dec(&dm->usecnt);
2015 EXPORT_SYMBOL(ib_dereg_mr_user);
2018 * ib_alloc_mr_user() - Allocates a memory region
2019 * @pd: protection domain associated with the region
2020 * @mr_type: memory region type
2021 * @max_num_sg: maximum sg entries available for registration.
2022 * @udata: user data or null for kernel objects
2025 * Memory registeration page/sg lists must not exceed max_num_sg.
2026 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
2027 * max_num_sg * used_page_size.
2030 struct ib_mr *ib_alloc_mr_user(struct ib_pd *pd, enum ib_mr_type mr_type,
2031 u32 max_num_sg, struct ib_udata *udata)
2035 if (!pd->device->ops.alloc_mr) {
2036 mr = ERR_PTR(-EOPNOTSUPP);
2040 if (mr_type == IB_MR_TYPE_INTEGRITY) {
2042 mr = ERR_PTR(-EINVAL);
2046 mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg, udata);
2048 mr->device = pd->device;
2052 atomic_inc(&pd->usecnt);
2053 mr->need_inval = false;
2054 mr->res.type = RDMA_RESTRACK_MR;
2055 rdma_restrack_kadd(&mr->res);
2057 mr->sig_attrs = NULL;
2061 trace_mr_alloc(pd, mr_type, max_num_sg, mr);
2064 EXPORT_SYMBOL(ib_alloc_mr_user);
2067 * ib_alloc_mr_integrity() - Allocates an integrity memory region
2068 * @pd: protection domain associated with the region
2069 * @max_num_data_sg: maximum data sg entries available for registration
2070 * @max_num_meta_sg: maximum metadata sg entries available for
2074 * Memory registration page/sg lists must not exceed max_num_sg,
2075 * also the integrity page/sg lists must not exceed max_num_meta_sg.
2078 struct ib_mr *ib_alloc_mr_integrity(struct ib_pd *pd,
2079 u32 max_num_data_sg,
2080 u32 max_num_meta_sg)
2083 struct ib_sig_attrs *sig_attrs;
2085 if (!pd->device->ops.alloc_mr_integrity ||
2086 !pd->device->ops.map_mr_sg_pi) {
2087 mr = ERR_PTR(-EOPNOTSUPP);
2091 if (!max_num_meta_sg) {
2092 mr = ERR_PTR(-EINVAL);
2096 sig_attrs = kzalloc(sizeof(struct ib_sig_attrs), GFP_KERNEL);
2098 mr = ERR_PTR(-ENOMEM);
2102 mr = pd->device->ops.alloc_mr_integrity(pd, max_num_data_sg,
2109 mr->device = pd->device;
2113 atomic_inc(&pd->usecnt);
2114 mr->need_inval = false;
2115 mr->res.type = RDMA_RESTRACK_MR;
2116 rdma_restrack_kadd(&mr->res);
2117 mr->type = IB_MR_TYPE_INTEGRITY;
2118 mr->sig_attrs = sig_attrs;
2121 trace_mr_integ_alloc(pd, max_num_data_sg, max_num_meta_sg, mr);
2124 EXPORT_SYMBOL(ib_alloc_mr_integrity);
2126 /* "Fast" memory regions */
2128 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
2129 int mr_access_flags,
2130 struct ib_fmr_attr *fmr_attr)
2134 if (!pd->device->ops.alloc_fmr)
2135 return ERR_PTR(-EOPNOTSUPP);
2137 fmr = pd->device->ops.alloc_fmr(pd, mr_access_flags, fmr_attr);
2139 fmr->device = pd->device;
2141 atomic_inc(&pd->usecnt);
2146 EXPORT_SYMBOL(ib_alloc_fmr);
2148 int ib_unmap_fmr(struct list_head *fmr_list)
2152 if (list_empty(fmr_list))
2155 fmr = list_entry(fmr_list->next, struct ib_fmr, list);
2156 return fmr->device->ops.unmap_fmr(fmr_list);
2158 EXPORT_SYMBOL(ib_unmap_fmr);
2160 int ib_dealloc_fmr(struct ib_fmr *fmr)
2166 ret = fmr->device->ops.dealloc_fmr(fmr);
2168 atomic_dec(&pd->usecnt);
2172 EXPORT_SYMBOL(ib_dealloc_fmr);
2174 /* Multicast groups */
2176 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2178 struct ib_qp_init_attr init_attr = {};
2179 struct ib_qp_attr attr = {};
2180 int num_eth_ports = 0;
2183 /* If QP state >= init, it is assigned to a port and we can check this
2186 if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2187 if (attr.qp_state >= IB_QPS_INIT) {
2188 if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2189 IB_LINK_LAYER_INFINIBAND)
2195 /* Can't get a quick answer, iterate over all ports */
2196 for (port = 0; port < qp->device->phys_port_cnt; port++)
2197 if (rdma_port_get_link_layer(qp->device, port) !=
2198 IB_LINK_LAYER_INFINIBAND)
2201 /* If we have at lease one Ethernet port, RoCE annex declares that
2202 * multicast LID should be ignored. We can't tell at this step if the
2203 * QP belongs to an IB or Ethernet port.
2208 /* If all the ports are IB, we can check according to IB spec. */
2210 return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2211 lid == be16_to_cpu(IB_LID_PERMISSIVE));
2214 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2218 if (!qp->device->ops.attach_mcast)
2221 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2222 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2225 ret = qp->device->ops.attach_mcast(qp, gid, lid);
2227 atomic_inc(&qp->usecnt);
2230 EXPORT_SYMBOL(ib_attach_mcast);
2232 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2236 if (!qp->device->ops.detach_mcast)
2239 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2240 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2243 ret = qp->device->ops.detach_mcast(qp, gid, lid);
2245 atomic_dec(&qp->usecnt);
2248 EXPORT_SYMBOL(ib_detach_mcast);
2250 struct ib_xrcd *__ib_alloc_xrcd(struct ib_device *device, const char *caller)
2252 struct ib_xrcd *xrcd;
2254 if (!device->ops.alloc_xrcd)
2255 return ERR_PTR(-EOPNOTSUPP);
2257 xrcd = device->ops.alloc_xrcd(device, NULL);
2258 if (!IS_ERR(xrcd)) {
2259 xrcd->device = device;
2261 atomic_set(&xrcd->usecnt, 0);
2262 mutex_init(&xrcd->tgt_qp_mutex);
2263 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
2268 EXPORT_SYMBOL(__ib_alloc_xrcd);
2270 int ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata)
2275 if (atomic_read(&xrcd->usecnt))
2278 while (!list_empty(&xrcd->tgt_qp_list)) {
2279 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
2280 ret = ib_destroy_qp(qp);
2284 mutex_destroy(&xrcd->tgt_qp_mutex);
2286 return xrcd->device->ops.dealloc_xrcd(xrcd, udata);
2288 EXPORT_SYMBOL(ib_dealloc_xrcd);
2291 * ib_create_wq - Creates a WQ associated with the specified protection
2293 * @pd: The protection domain associated with the WQ.
2294 * @wq_attr: A list of initial attributes required to create the
2295 * WQ. If WQ creation succeeds, then the attributes are updated to
2296 * the actual capabilities of the created WQ.
2298 * wq_attr->max_wr and wq_attr->max_sge determine
2299 * the requested size of the WQ, and set to the actual values allocated
2301 * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2302 * at least as large as the requested values.
2304 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2305 struct ib_wq_init_attr *wq_attr)
2309 if (!pd->device->ops.create_wq)
2310 return ERR_PTR(-EOPNOTSUPP);
2312 wq = pd->device->ops.create_wq(pd, wq_attr, NULL);
2314 wq->event_handler = wq_attr->event_handler;
2315 wq->wq_context = wq_attr->wq_context;
2316 wq->wq_type = wq_attr->wq_type;
2317 wq->cq = wq_attr->cq;
2318 wq->device = pd->device;
2321 atomic_inc(&pd->usecnt);
2322 atomic_inc(&wq_attr->cq->usecnt);
2323 atomic_set(&wq->usecnt, 0);
2327 EXPORT_SYMBOL(ib_create_wq);
2330 * ib_destroy_wq - Destroys the specified user WQ.
2331 * @wq: The WQ to destroy.
2332 * @udata: Valid user data
2334 int ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata)
2336 struct ib_cq *cq = wq->cq;
2337 struct ib_pd *pd = wq->pd;
2339 if (atomic_read(&wq->usecnt))
2342 wq->device->ops.destroy_wq(wq, udata);
2343 atomic_dec(&pd->usecnt);
2344 atomic_dec(&cq->usecnt);
2348 EXPORT_SYMBOL(ib_destroy_wq);
2351 * ib_modify_wq - Modifies the specified WQ.
2352 * @wq: The WQ to modify.
2353 * @wq_attr: On input, specifies the WQ attributes to modify.
2354 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
2355 * are being modified.
2356 * On output, the current values of selected WQ attributes are returned.
2358 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
2363 if (!wq->device->ops.modify_wq)
2366 err = wq->device->ops.modify_wq(wq, wq_attr, wq_attr_mask, NULL);
2369 EXPORT_SYMBOL(ib_modify_wq);
2372 * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
2373 * @device: The device on which to create the rwq indirection table.
2374 * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
2375 * create the Indirection Table.
2377 * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
2378 * than the created ib_rwq_ind_table object and the caller is responsible
2379 * for its memory allocation/free.
2381 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
2382 struct ib_rwq_ind_table_init_attr *init_attr)
2384 struct ib_rwq_ind_table *rwq_ind_table;
2388 if (!device->ops.create_rwq_ind_table)
2389 return ERR_PTR(-EOPNOTSUPP);
2391 table_size = (1 << init_attr->log_ind_tbl_size);
2392 rwq_ind_table = device->ops.create_rwq_ind_table(device,
2394 if (IS_ERR(rwq_ind_table))
2395 return rwq_ind_table;
2397 rwq_ind_table->ind_tbl = init_attr->ind_tbl;
2398 rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
2399 rwq_ind_table->device = device;
2400 rwq_ind_table->uobject = NULL;
2401 atomic_set(&rwq_ind_table->usecnt, 0);
2403 for (i = 0; i < table_size; i++)
2404 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
2406 return rwq_ind_table;
2408 EXPORT_SYMBOL(ib_create_rwq_ind_table);
2411 * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
2412 * @wq_ind_table: The Indirection Table to destroy.
2414 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
2417 u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
2418 struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
2420 if (atomic_read(&rwq_ind_table->usecnt))
2423 err = rwq_ind_table->device->ops.destroy_rwq_ind_table(rwq_ind_table);
2425 for (i = 0; i < table_size; i++)
2426 atomic_dec(&ind_tbl[i]->usecnt);
2431 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
2433 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2434 struct ib_mr_status *mr_status)
2436 if (!mr->device->ops.check_mr_status)
2439 return mr->device->ops.check_mr_status(mr, check_mask, mr_status);
2441 EXPORT_SYMBOL(ib_check_mr_status);
2443 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
2446 if (!device->ops.set_vf_link_state)
2449 return device->ops.set_vf_link_state(device, vf, port, state);
2451 EXPORT_SYMBOL(ib_set_vf_link_state);
2453 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
2454 struct ifla_vf_info *info)
2456 if (!device->ops.get_vf_config)
2459 return device->ops.get_vf_config(device, vf, port, info);
2461 EXPORT_SYMBOL(ib_get_vf_config);
2463 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
2464 struct ifla_vf_stats *stats)
2466 if (!device->ops.get_vf_stats)
2469 return device->ops.get_vf_stats(device, vf, port, stats);
2471 EXPORT_SYMBOL(ib_get_vf_stats);
2473 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
2476 if (!device->ops.set_vf_guid)
2479 return device->ops.set_vf_guid(device, vf, port, guid, type);
2481 EXPORT_SYMBOL(ib_set_vf_guid);
2483 int ib_get_vf_guid(struct ib_device *device, int vf, u8 port,
2484 struct ifla_vf_guid *node_guid,
2485 struct ifla_vf_guid *port_guid)
2487 if (!device->ops.get_vf_guid)
2490 return device->ops.get_vf_guid(device, vf, port, node_guid, port_guid);
2492 EXPORT_SYMBOL(ib_get_vf_guid);
2494 * ib_map_mr_sg_pi() - Map the dma mapped SG lists for PI (protection
2495 * information) and set an appropriate memory region for registration.
2496 * @mr: memory region
2497 * @data_sg: dma mapped scatterlist for data
2498 * @data_sg_nents: number of entries in data_sg
2499 * @data_sg_offset: offset in bytes into data_sg
2500 * @meta_sg: dma mapped scatterlist for metadata
2501 * @meta_sg_nents: number of entries in meta_sg
2502 * @meta_sg_offset: offset in bytes into meta_sg
2503 * @page_size: page vector desired page size
2506 * - The MR must be allocated with type IB_MR_TYPE_INTEGRITY.
2508 * Return: 0 on success.
2510 * After this completes successfully, the memory region
2511 * is ready for registration.
2513 int ib_map_mr_sg_pi(struct ib_mr *mr, struct scatterlist *data_sg,
2514 int data_sg_nents, unsigned int *data_sg_offset,
2515 struct scatterlist *meta_sg, int meta_sg_nents,
2516 unsigned int *meta_sg_offset, unsigned int page_size)
2518 if (unlikely(!mr->device->ops.map_mr_sg_pi ||
2519 WARN_ON_ONCE(mr->type != IB_MR_TYPE_INTEGRITY)))
2522 mr->page_size = page_size;
2524 return mr->device->ops.map_mr_sg_pi(mr, data_sg, data_sg_nents,
2525 data_sg_offset, meta_sg,
2526 meta_sg_nents, meta_sg_offset);
2528 EXPORT_SYMBOL(ib_map_mr_sg_pi);
2531 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2532 * and set it the memory region.
2533 * @mr: memory region
2534 * @sg: dma mapped scatterlist
2535 * @sg_nents: number of entries in sg
2536 * @sg_offset: offset in bytes into sg
2537 * @page_size: page vector desired page size
2540 * - The first sg element is allowed to have an offset.
2541 * - Each sg element must either be aligned to page_size or virtually
2542 * contiguous to the previous element. In case an sg element has a
2543 * non-contiguous offset, the mapping prefix will not include it.
2544 * - The last sg element is allowed to have length less than page_size.
2545 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2546 * then only max_num_sg entries will be mapped.
2547 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2548 * constraints holds and the page_size argument is ignored.
2550 * Returns the number of sg elements that were mapped to the memory region.
2552 * After this completes successfully, the memory region
2553 * is ready for registration.
2555 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2556 unsigned int *sg_offset, unsigned int page_size)
2558 if (unlikely(!mr->device->ops.map_mr_sg))
2561 mr->page_size = page_size;
2563 return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset);
2565 EXPORT_SYMBOL(ib_map_mr_sg);
2568 * ib_sg_to_pages() - Convert the largest prefix of a sg list
2570 * @mr: memory region
2571 * @sgl: dma mapped scatterlist
2572 * @sg_nents: number of entries in sg
2573 * @sg_offset_p: IN: start offset in bytes into sg
2574 * OUT: offset in bytes for element n of the sg of the first
2575 * byte that has not been processed where n is the return
2576 * value of this function.
2577 * @set_page: driver page assignment function pointer
2579 * Core service helper for drivers to convert the largest
2580 * prefix of given sg list to a page vector. The sg list
2581 * prefix converted is the prefix that meet the requirements
2584 * Returns the number of sg elements that were assigned to
2587 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2588 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2590 struct scatterlist *sg;
2591 u64 last_end_dma_addr = 0;
2592 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2593 unsigned int last_page_off = 0;
2594 u64 page_mask = ~((u64)mr->page_size - 1);
2597 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2600 mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2603 for_each_sg(sgl, sg, sg_nents, i) {
2604 u64 dma_addr = sg_dma_address(sg) + sg_offset;
2605 u64 prev_addr = dma_addr;
2606 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2607 u64 end_dma_addr = dma_addr + dma_len;
2608 u64 page_addr = dma_addr & page_mask;
2611 * For the second and later elements, check whether either the
2612 * end of element i-1 or the start of element i is not aligned
2613 * on a page boundary.
2615 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2616 /* Stop mapping if there is a gap. */
2617 if (last_end_dma_addr != dma_addr)
2621 * Coalesce this element with the last. If it is small
2622 * enough just update mr->length. Otherwise start
2623 * mapping from the next page.
2629 ret = set_page(mr, page_addr);
2630 if (unlikely(ret < 0)) {
2631 sg_offset = prev_addr - sg_dma_address(sg);
2632 mr->length += prev_addr - dma_addr;
2634 *sg_offset_p = sg_offset;
2635 return i || sg_offset ? i : ret;
2637 prev_addr = page_addr;
2639 page_addr += mr->page_size;
2640 } while (page_addr < end_dma_addr);
2642 mr->length += dma_len;
2643 last_end_dma_addr = end_dma_addr;
2644 last_page_off = end_dma_addr & ~page_mask;
2653 EXPORT_SYMBOL(ib_sg_to_pages);
2655 struct ib_drain_cqe {
2657 struct completion done;
2660 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2662 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2665 complete(&cqe->done);
2669 * Post a WR and block until its completion is reaped for the SQ.
2671 static void __ib_drain_sq(struct ib_qp *qp)
2673 struct ib_cq *cq = qp->send_cq;
2674 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2675 struct ib_drain_cqe sdrain;
2676 struct ib_rdma_wr swr = {
2679 { .wr_cqe = &sdrain.cqe, },
2680 .opcode = IB_WR_RDMA_WRITE,
2685 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2687 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2691 sdrain.cqe.done = ib_drain_qp_done;
2692 init_completion(&sdrain.done);
2694 ret = ib_post_send(qp, &swr.wr, NULL);
2696 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2700 if (cq->poll_ctx == IB_POLL_DIRECT)
2701 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2702 ib_process_cq_direct(cq, -1);
2704 wait_for_completion(&sdrain.done);
2708 * Post a WR and block until its completion is reaped for the RQ.
2710 static void __ib_drain_rq(struct ib_qp *qp)
2712 struct ib_cq *cq = qp->recv_cq;
2713 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2714 struct ib_drain_cqe rdrain;
2715 struct ib_recv_wr rwr = {};
2718 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2720 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2724 rwr.wr_cqe = &rdrain.cqe;
2725 rdrain.cqe.done = ib_drain_qp_done;
2726 init_completion(&rdrain.done);
2728 ret = ib_post_recv(qp, &rwr, NULL);
2730 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2734 if (cq->poll_ctx == IB_POLL_DIRECT)
2735 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2736 ib_process_cq_direct(cq, -1);
2738 wait_for_completion(&rdrain.done);
2742 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2744 * @qp: queue pair to drain
2746 * If the device has a provider-specific drain function, then
2747 * call that. Otherwise call the generic drain function
2752 * ensure there is room in the CQ and SQ for the drain work request and
2755 * allocate the CQ using ib_alloc_cq().
2757 * ensure that there are no other contexts that are posting WRs concurrently.
2758 * Otherwise the drain is not guaranteed.
2760 void ib_drain_sq(struct ib_qp *qp)
2762 if (qp->device->ops.drain_sq)
2763 qp->device->ops.drain_sq(qp);
2766 trace_cq_drain_complete(qp->send_cq);
2768 EXPORT_SYMBOL(ib_drain_sq);
2771 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2773 * @qp: queue pair to drain
2775 * If the device has a provider-specific drain function, then
2776 * call that. Otherwise call the generic drain function
2781 * ensure there is room in the CQ and RQ for the drain work request and
2784 * allocate the CQ using ib_alloc_cq().
2786 * ensure that there are no other contexts that are posting WRs concurrently.
2787 * Otherwise the drain is not guaranteed.
2789 void ib_drain_rq(struct ib_qp *qp)
2791 if (qp->device->ops.drain_rq)
2792 qp->device->ops.drain_rq(qp);
2795 trace_cq_drain_complete(qp->recv_cq);
2797 EXPORT_SYMBOL(ib_drain_rq);
2800 * ib_drain_qp() - Block until all CQEs have been consumed by the
2801 * application on both the RQ and SQ.
2802 * @qp: queue pair to drain
2806 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2809 * allocate the CQs using ib_alloc_cq().
2811 * ensure that there are no other contexts that are posting WRs concurrently.
2812 * Otherwise the drain is not guaranteed.
2814 void ib_drain_qp(struct ib_qp *qp)
2820 EXPORT_SYMBOL(ib_drain_qp);
2822 struct net_device *rdma_alloc_netdev(struct ib_device *device, u8 port_num,
2823 enum rdma_netdev_t type, const char *name,
2824 unsigned char name_assign_type,
2825 void (*setup)(struct net_device *))
2827 struct rdma_netdev_alloc_params params;
2828 struct net_device *netdev;
2831 if (!device->ops.rdma_netdev_get_params)
2832 return ERR_PTR(-EOPNOTSUPP);
2834 rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2839 netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type,
2840 setup, params.txqs, params.rxqs);
2842 return ERR_PTR(-ENOMEM);
2846 EXPORT_SYMBOL(rdma_alloc_netdev);
2848 int rdma_init_netdev(struct ib_device *device, u8 port_num,
2849 enum rdma_netdev_t type, const char *name,
2850 unsigned char name_assign_type,
2851 void (*setup)(struct net_device *),
2852 struct net_device *netdev)
2854 struct rdma_netdev_alloc_params params;
2857 if (!device->ops.rdma_netdev_get_params)
2860 rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2865 return params.initialize_rdma_netdev(device, port_num,
2866 netdev, params.param);
2868 EXPORT_SYMBOL(rdma_init_netdev);
2870 void __rdma_block_iter_start(struct ib_block_iter *biter,
2871 struct scatterlist *sglist, unsigned int nents,
2874 memset(biter, 0, sizeof(struct ib_block_iter));
2875 biter->__sg = sglist;
2876 biter->__sg_nents = nents;
2878 /* Driver provides best block size to use */
2879 biter->__pg_bit = __fls(pgsz);
2881 EXPORT_SYMBOL(__rdma_block_iter_start);
2883 bool __rdma_block_iter_next(struct ib_block_iter *biter)
2885 unsigned int block_offset;
2887 if (!biter->__sg_nents || !biter->__sg)
2890 biter->__dma_addr = sg_dma_address(biter->__sg) + biter->__sg_advance;
2891 block_offset = biter->__dma_addr & (BIT_ULL(biter->__pg_bit) - 1);
2892 biter->__sg_advance += BIT_ULL(biter->__pg_bit) - block_offset;
2894 if (biter->__sg_advance >= sg_dma_len(biter->__sg)) {
2895 biter->__sg_advance = 0;
2896 biter->__sg = sg_next(biter->__sg);
2897 biter->__sg_nents--;
2902 EXPORT_SYMBOL(__rdma_block_iter_next);