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>
55 #include "core_priv.h"
56 #include <trace/events/rdma_core.h>
58 static int ib_resolve_eth_dmac(struct ib_device *device,
59 struct rdma_ah_attr *ah_attr);
61 static const char * const ib_events[] = {
62 [IB_EVENT_CQ_ERR] = "CQ error",
63 [IB_EVENT_QP_FATAL] = "QP fatal error",
64 [IB_EVENT_QP_REQ_ERR] = "QP request error",
65 [IB_EVENT_QP_ACCESS_ERR] = "QP access error",
66 [IB_EVENT_COMM_EST] = "communication established",
67 [IB_EVENT_SQ_DRAINED] = "send queue drained",
68 [IB_EVENT_PATH_MIG] = "path migration successful",
69 [IB_EVENT_PATH_MIG_ERR] = "path migration error",
70 [IB_EVENT_DEVICE_FATAL] = "device fatal error",
71 [IB_EVENT_PORT_ACTIVE] = "port active",
72 [IB_EVENT_PORT_ERR] = "port error",
73 [IB_EVENT_LID_CHANGE] = "LID change",
74 [IB_EVENT_PKEY_CHANGE] = "P_key change",
75 [IB_EVENT_SM_CHANGE] = "SM change",
76 [IB_EVENT_SRQ_ERR] = "SRQ error",
77 [IB_EVENT_SRQ_LIMIT_REACHED] = "SRQ limit reached",
78 [IB_EVENT_QP_LAST_WQE_REACHED] = "last WQE reached",
79 [IB_EVENT_CLIENT_REREGISTER] = "client reregister",
80 [IB_EVENT_GID_CHANGE] = "GID changed",
83 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
87 return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
88 ib_events[index] : "unrecognized event";
90 EXPORT_SYMBOL(ib_event_msg);
92 static const char * const wc_statuses[] = {
93 [IB_WC_SUCCESS] = "success",
94 [IB_WC_LOC_LEN_ERR] = "local length error",
95 [IB_WC_LOC_QP_OP_ERR] = "local QP operation error",
96 [IB_WC_LOC_EEC_OP_ERR] = "local EE context operation error",
97 [IB_WC_LOC_PROT_ERR] = "local protection error",
98 [IB_WC_WR_FLUSH_ERR] = "WR flushed",
99 [IB_WC_MW_BIND_ERR] = "memory management operation error",
100 [IB_WC_BAD_RESP_ERR] = "bad response error",
101 [IB_WC_LOC_ACCESS_ERR] = "local access error",
102 [IB_WC_REM_INV_REQ_ERR] = "invalid request error",
103 [IB_WC_REM_ACCESS_ERR] = "remote access error",
104 [IB_WC_REM_OP_ERR] = "remote operation error",
105 [IB_WC_RETRY_EXC_ERR] = "transport retry counter exceeded",
106 [IB_WC_RNR_RETRY_EXC_ERR] = "RNR retry counter exceeded",
107 [IB_WC_LOC_RDD_VIOL_ERR] = "local RDD violation error",
108 [IB_WC_REM_INV_RD_REQ_ERR] = "remote invalid RD request",
109 [IB_WC_REM_ABORT_ERR] = "operation aborted",
110 [IB_WC_INV_EECN_ERR] = "invalid EE context number",
111 [IB_WC_INV_EEC_STATE_ERR] = "invalid EE context state",
112 [IB_WC_FATAL_ERR] = "fatal error",
113 [IB_WC_RESP_TIMEOUT_ERR] = "response timeout error",
114 [IB_WC_GENERAL_ERR] = "general error",
117 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
119 size_t index = status;
121 return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
122 wc_statuses[index] : "unrecognized status";
124 EXPORT_SYMBOL(ib_wc_status_msg);
126 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
129 case IB_RATE_2_5_GBPS: return 1;
130 case IB_RATE_5_GBPS: return 2;
131 case IB_RATE_10_GBPS: return 4;
132 case IB_RATE_20_GBPS: return 8;
133 case IB_RATE_30_GBPS: return 12;
134 case IB_RATE_40_GBPS: return 16;
135 case IB_RATE_60_GBPS: return 24;
136 case IB_RATE_80_GBPS: return 32;
137 case IB_RATE_120_GBPS: return 48;
138 case IB_RATE_14_GBPS: return 6;
139 case IB_RATE_56_GBPS: return 22;
140 case IB_RATE_112_GBPS: return 45;
141 case IB_RATE_168_GBPS: return 67;
142 case IB_RATE_25_GBPS: return 10;
143 case IB_RATE_100_GBPS: return 40;
144 case IB_RATE_200_GBPS: return 80;
145 case IB_RATE_300_GBPS: return 120;
146 case IB_RATE_28_GBPS: return 11;
147 case IB_RATE_50_GBPS: return 20;
148 case IB_RATE_400_GBPS: return 160;
149 case IB_RATE_600_GBPS: return 240;
153 EXPORT_SYMBOL(ib_rate_to_mult);
155 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
158 case 1: return IB_RATE_2_5_GBPS;
159 case 2: return IB_RATE_5_GBPS;
160 case 4: return IB_RATE_10_GBPS;
161 case 8: return IB_RATE_20_GBPS;
162 case 12: return IB_RATE_30_GBPS;
163 case 16: return IB_RATE_40_GBPS;
164 case 24: return IB_RATE_60_GBPS;
165 case 32: return IB_RATE_80_GBPS;
166 case 48: return IB_RATE_120_GBPS;
167 case 6: return IB_RATE_14_GBPS;
168 case 22: return IB_RATE_56_GBPS;
169 case 45: return IB_RATE_112_GBPS;
170 case 67: return IB_RATE_168_GBPS;
171 case 10: return IB_RATE_25_GBPS;
172 case 40: return IB_RATE_100_GBPS;
173 case 80: return IB_RATE_200_GBPS;
174 case 120: return IB_RATE_300_GBPS;
175 case 11: return IB_RATE_28_GBPS;
176 case 20: return IB_RATE_50_GBPS;
177 case 160: return IB_RATE_400_GBPS;
178 case 240: return IB_RATE_600_GBPS;
179 default: return IB_RATE_PORT_CURRENT;
182 EXPORT_SYMBOL(mult_to_ib_rate);
184 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
187 case IB_RATE_2_5_GBPS: return 2500;
188 case IB_RATE_5_GBPS: return 5000;
189 case IB_RATE_10_GBPS: return 10000;
190 case IB_RATE_20_GBPS: return 20000;
191 case IB_RATE_30_GBPS: return 30000;
192 case IB_RATE_40_GBPS: return 40000;
193 case IB_RATE_60_GBPS: return 60000;
194 case IB_RATE_80_GBPS: return 80000;
195 case IB_RATE_120_GBPS: return 120000;
196 case IB_RATE_14_GBPS: return 14062;
197 case IB_RATE_56_GBPS: return 56250;
198 case IB_RATE_112_GBPS: return 112500;
199 case IB_RATE_168_GBPS: return 168750;
200 case IB_RATE_25_GBPS: return 25781;
201 case IB_RATE_100_GBPS: return 103125;
202 case IB_RATE_200_GBPS: return 206250;
203 case IB_RATE_300_GBPS: return 309375;
204 case IB_RATE_28_GBPS: return 28125;
205 case IB_RATE_50_GBPS: return 53125;
206 case IB_RATE_400_GBPS: return 425000;
207 case IB_RATE_600_GBPS: return 637500;
211 EXPORT_SYMBOL(ib_rate_to_mbps);
213 __attribute_const__ enum rdma_transport_type
214 rdma_node_get_transport(unsigned int node_type)
217 if (node_type == RDMA_NODE_USNIC)
218 return RDMA_TRANSPORT_USNIC;
219 if (node_type == RDMA_NODE_USNIC_UDP)
220 return RDMA_TRANSPORT_USNIC_UDP;
221 if (node_type == RDMA_NODE_RNIC)
222 return RDMA_TRANSPORT_IWARP;
223 if (node_type == RDMA_NODE_UNSPECIFIED)
224 return RDMA_TRANSPORT_UNSPECIFIED;
226 return RDMA_TRANSPORT_IB;
228 EXPORT_SYMBOL(rdma_node_get_transport);
230 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
232 enum rdma_transport_type lt;
233 if (device->ops.get_link_layer)
234 return device->ops.get_link_layer(device, port_num);
236 lt = rdma_node_get_transport(device->node_type);
237 if (lt == RDMA_TRANSPORT_IB)
238 return IB_LINK_LAYER_INFINIBAND;
240 return IB_LINK_LAYER_ETHERNET;
242 EXPORT_SYMBOL(rdma_port_get_link_layer);
244 /* Protection domains */
247 * ib_alloc_pd - Allocates an unused protection domain.
248 * @device: The device on which to allocate the protection domain.
249 * @flags: protection domain flags
250 * @caller: caller's build-time module name
252 * A protection domain object provides an association between QPs, shared
253 * receive queues, address handles, memory regions, and memory windows.
255 * Every PD has a local_dma_lkey which can be used as the lkey value for local
258 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
262 int mr_access_flags = 0;
265 pd = rdma_zalloc_drv_obj(device, ib_pd);
267 return ERR_PTR(-ENOMEM);
271 pd->__internal_mr = NULL;
272 atomic_set(&pd->usecnt, 0);
275 rdma_restrack_new(&pd->res, RDMA_RESTRACK_PD);
276 rdma_restrack_set_name(&pd->res, caller);
278 ret = device->ops.alloc_pd(pd, NULL);
280 rdma_restrack_put(&pd->res);
284 rdma_restrack_add(&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 int 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 ret = pd->device->ops.dealloc_pd(pd, udata);
351 rdma_restrack_del(&pd->res);
355 EXPORT_SYMBOL(ib_dealloc_pd_user);
357 /* Address handles */
360 * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
361 * @dest: Pointer to destination ah_attr. Contents of the destination
362 * pointer is assumed to be invalid and attribute are overwritten.
363 * @src: Pointer to source ah_attr.
365 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
366 const struct rdma_ah_attr *src)
369 if (dest->grh.sgid_attr)
370 rdma_hold_gid_attr(dest->grh.sgid_attr);
372 EXPORT_SYMBOL(rdma_copy_ah_attr);
375 * rdma_replace_ah_attr - Replace valid ah_attr with new new one.
376 * @old: Pointer to existing ah_attr which needs to be replaced.
377 * old is assumed to be valid or zero'd
378 * @new: Pointer to the new ah_attr.
380 * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
381 * old the ah_attr is valid; after that it copies the new attribute and holds
382 * the reference to the replaced ah_attr.
384 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
385 const struct rdma_ah_attr *new)
387 rdma_destroy_ah_attr(old);
389 if (old->grh.sgid_attr)
390 rdma_hold_gid_attr(old->grh.sgid_attr);
392 EXPORT_SYMBOL(rdma_replace_ah_attr);
395 * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
396 * @dest: Pointer to destination ah_attr to copy to.
397 * dest is assumed to be valid or zero'd
398 * @src: Pointer to the new ah_attr.
400 * rdma_move_ah_attr() first releases any reference in the destination ah_attr
401 * if it is valid. This also transfers ownership of internal references from
402 * src to dest, making src invalid in the process. No new reference of the src
405 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
407 rdma_destroy_ah_attr(dest);
409 src->grh.sgid_attr = NULL;
411 EXPORT_SYMBOL(rdma_move_ah_attr);
414 * Validate that the rdma_ah_attr is valid for the device before passing it
417 static int rdma_check_ah_attr(struct ib_device *device,
418 struct rdma_ah_attr *ah_attr)
420 if (!rdma_is_port_valid(device, ah_attr->port_num))
423 if ((rdma_is_grh_required(device, ah_attr->port_num) ||
424 ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
425 !(ah_attr->ah_flags & IB_AH_GRH))
428 if (ah_attr->grh.sgid_attr) {
430 * Make sure the passed sgid_attr is consistent with the
433 if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
434 ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
441 * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
442 * On success the caller is responsible to call rdma_unfill_sgid_attr().
444 static int rdma_fill_sgid_attr(struct ib_device *device,
445 struct rdma_ah_attr *ah_attr,
446 const struct ib_gid_attr **old_sgid_attr)
448 const struct ib_gid_attr *sgid_attr;
449 struct ib_global_route *grh;
452 *old_sgid_attr = ah_attr->grh.sgid_attr;
454 ret = rdma_check_ah_attr(device, ah_attr);
458 if (!(ah_attr->ah_flags & IB_AH_GRH))
461 grh = rdma_ah_retrieve_grh(ah_attr);
466 rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
467 if (IS_ERR(sgid_attr))
468 return PTR_ERR(sgid_attr);
470 /* Move ownerhip of the kref into the ah_attr */
471 grh->sgid_attr = sgid_attr;
475 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
476 const struct ib_gid_attr *old_sgid_attr)
479 * Fill didn't change anything, the caller retains ownership of
482 if (ah_attr->grh.sgid_attr == old_sgid_attr)
486 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
487 * doesn't see any change in the rdma_ah_attr. If we get here
488 * old_sgid_attr is NULL.
490 rdma_destroy_ah_attr(ah_attr);
493 static const struct ib_gid_attr *
494 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
495 const struct ib_gid_attr *old_attr)
498 rdma_put_gid_attr(old_attr);
499 if (ah_attr->ah_flags & IB_AH_GRH) {
500 rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
501 return ah_attr->grh.sgid_attr;
506 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
507 struct rdma_ah_attr *ah_attr,
509 struct ib_udata *udata,
510 struct net_device *xmit_slave)
512 struct rdma_ah_init_attr init_attr = {};
513 struct ib_device *device = pd->device;
517 might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE);
519 if (!device->ops.create_ah)
520 return ERR_PTR(-EOPNOTSUPP);
522 ah = rdma_zalloc_drv_obj_gfp(
524 (flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC);
526 return ERR_PTR(-ENOMEM);
530 ah->type = ah_attr->type;
531 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
532 init_attr.ah_attr = ah_attr;
533 init_attr.flags = flags;
534 init_attr.xmit_slave = xmit_slave;
536 ret = device->ops.create_ah(ah, &init_attr, udata);
542 atomic_inc(&pd->usecnt);
547 * rdma_create_ah - Creates an address handle for the
548 * given address vector.
549 * @pd: The protection domain associated with the address handle.
550 * @ah_attr: The attributes of the address vector.
551 * @flags: Create address handle flags (see enum rdma_create_ah_flags).
553 * It returns 0 on success and returns appropriate error code on error.
554 * The address handle is used to reference a local or global destination
555 * in all UD QP post sends.
557 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
560 const struct ib_gid_attr *old_sgid_attr;
561 struct net_device *slave;
565 ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
568 slave = rdma_lag_get_ah_roce_slave(pd->device, ah_attr,
569 (flags & RDMA_CREATE_AH_SLEEPABLE) ?
570 GFP_KERNEL : GFP_ATOMIC);
572 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
573 return (void *)slave;
575 ah = _rdma_create_ah(pd, ah_attr, flags, NULL, slave);
576 rdma_lag_put_ah_roce_slave(slave);
577 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
580 EXPORT_SYMBOL(rdma_create_ah);
583 * rdma_create_user_ah - Creates an address handle for the
584 * given address vector.
585 * It resolves destination mac address for ah attribute of RoCE type.
586 * @pd: The protection domain associated with the address handle.
587 * @ah_attr: The attributes of the address vector.
588 * @udata: pointer to user's input output buffer information need by
591 * It returns 0 on success and returns appropriate error code on error.
592 * The address handle is used to reference a local or global destination
593 * in all UD QP post sends.
595 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
596 struct rdma_ah_attr *ah_attr,
597 struct ib_udata *udata)
599 const struct ib_gid_attr *old_sgid_attr;
603 err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
607 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
608 err = ib_resolve_eth_dmac(pd->device, ah_attr);
615 ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE,
619 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
622 EXPORT_SYMBOL(rdma_create_user_ah);
624 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
626 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
627 struct iphdr ip4h_checked;
628 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
630 /* If it's IPv6, the version must be 6, otherwise, the first
631 * 20 bytes (before the IPv4 header) are garbled.
633 if (ip6h->version != 6)
634 return (ip4h->version == 4) ? 4 : 0;
635 /* version may be 6 or 4 because the first 20 bytes could be garbled */
637 /* RoCE v2 requires no options, thus header length
644 * We can't write on scattered buffers so we need to copy to
647 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
648 ip4h_checked.check = 0;
649 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
650 /* if IPv4 header checksum is OK, believe it */
651 if (ip4h->check == ip4h_checked.check)
655 EXPORT_SYMBOL(ib_get_rdma_header_version);
657 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
659 const struct ib_grh *grh)
663 if (rdma_protocol_ib(device, port_num))
664 return RDMA_NETWORK_IB;
666 grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
668 if (grh_version == 4)
669 return RDMA_NETWORK_IPV4;
671 if (grh->next_hdr == IPPROTO_UDP)
672 return RDMA_NETWORK_IPV6;
674 return RDMA_NETWORK_ROCE_V1;
677 struct find_gid_index_context {
679 enum ib_gid_type gid_type;
682 static bool find_gid_index(const union ib_gid *gid,
683 const struct ib_gid_attr *gid_attr,
686 struct find_gid_index_context *ctx = context;
687 u16 vlan_id = 0xffff;
690 if (ctx->gid_type != gid_attr->gid_type)
693 ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
697 return ctx->vlan_id == vlan_id;
700 static const struct ib_gid_attr *
701 get_sgid_attr_from_eth(struct ib_device *device, u8 port_num,
702 u16 vlan_id, const union ib_gid *sgid,
703 enum ib_gid_type gid_type)
705 struct find_gid_index_context context = {.vlan_id = vlan_id,
706 .gid_type = gid_type};
708 return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
712 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
713 enum rdma_network_type net_type,
714 union ib_gid *sgid, union ib_gid *dgid)
716 struct sockaddr_in src_in;
717 struct sockaddr_in dst_in;
718 __be32 src_saddr, dst_saddr;
723 if (net_type == RDMA_NETWORK_IPV4) {
724 memcpy(&src_in.sin_addr.s_addr,
725 &hdr->roce4grh.saddr, 4);
726 memcpy(&dst_in.sin_addr.s_addr,
727 &hdr->roce4grh.daddr, 4);
728 src_saddr = src_in.sin_addr.s_addr;
729 dst_saddr = dst_in.sin_addr.s_addr;
730 ipv6_addr_set_v4mapped(src_saddr,
731 (struct in6_addr *)sgid);
732 ipv6_addr_set_v4mapped(dst_saddr,
733 (struct in6_addr *)dgid);
735 } else if (net_type == RDMA_NETWORK_IPV6 ||
736 net_type == RDMA_NETWORK_IB || RDMA_NETWORK_ROCE_V1) {
737 *dgid = hdr->ibgrh.dgid;
738 *sgid = hdr->ibgrh.sgid;
744 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
746 /* Resolve destination mac address and hop limit for unicast destination
747 * GID entry, considering the source GID entry as well.
748 * ah_attribute must have have valid port_num, sgid_index.
750 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
751 struct rdma_ah_attr *ah_attr)
753 struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
754 const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
755 int hop_limit = 0xff;
758 /* If destination is link local and source GID is RoCEv1,
759 * IP stack is not used.
761 if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
762 sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
763 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
768 ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
770 sgid_attr, &hop_limit);
772 grh->hop_limit = hop_limit;
777 * This function initializes address handle attributes from the incoming packet.
778 * Incoming packet has dgid of the receiver node on which this code is
779 * getting executed and, sgid contains the GID of the sender.
781 * When resolving mac address of destination, the arrived dgid is used
782 * as sgid and, sgid is used as dgid because sgid contains destinations
783 * GID whom to respond to.
785 * On success the caller is responsible to call rdma_destroy_ah_attr on the
788 int ib_init_ah_attr_from_wc(struct ib_device *device, u8 port_num,
789 const struct ib_wc *wc, const struct ib_grh *grh,
790 struct rdma_ah_attr *ah_attr)
794 enum rdma_network_type net_type = RDMA_NETWORK_IB;
795 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
796 const struct ib_gid_attr *sgid_attr;
803 memset(ah_attr, 0, sizeof *ah_attr);
804 ah_attr->type = rdma_ah_find_type(device, port_num);
805 if (rdma_cap_eth_ah(device, port_num)) {
806 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
807 net_type = wc->network_hdr_type;
809 net_type = ib_get_net_type_by_grh(device, port_num, grh);
810 gid_type = ib_network_to_gid_type(net_type);
812 ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
817 rdma_ah_set_sl(ah_attr, wc->sl);
818 rdma_ah_set_port_num(ah_attr, port_num);
820 if (rdma_protocol_roce(device, port_num)) {
821 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
822 wc->vlan_id : 0xffff;
824 if (!(wc->wc_flags & IB_WC_GRH))
827 sgid_attr = get_sgid_attr_from_eth(device, port_num,
830 if (IS_ERR(sgid_attr))
831 return PTR_ERR(sgid_attr);
833 flow_class = be32_to_cpu(grh->version_tclass_flow);
834 rdma_move_grh_sgid_attr(ah_attr,
836 flow_class & 0xFFFFF,
838 (flow_class >> 20) & 0xFF,
841 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
843 rdma_destroy_ah_attr(ah_attr);
847 rdma_ah_set_dlid(ah_attr, wc->slid);
848 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
850 if ((wc->wc_flags & IB_WC_GRH) == 0)
853 if (dgid.global.interface_id !=
854 cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
855 sgid_attr = rdma_find_gid_by_port(
856 device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
858 sgid_attr = rdma_get_gid_attr(device, port_num, 0);
860 if (IS_ERR(sgid_attr))
861 return PTR_ERR(sgid_attr);
862 flow_class = be32_to_cpu(grh->version_tclass_flow);
863 rdma_move_grh_sgid_attr(ah_attr,
865 flow_class & 0xFFFFF,
867 (flow_class >> 20) & 0xFF,
873 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
876 * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
879 * @attr: Pointer to AH attribute structure
880 * @dgid: Destination GID
881 * @flow_label: Flow label
882 * @hop_limit: Hop limit
883 * @traffic_class: traffic class
884 * @sgid_attr: Pointer to SGID attribute
886 * This takes ownership of the sgid_attr reference. The caller must ensure
887 * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
888 * calling this function.
890 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
891 u32 flow_label, u8 hop_limit, u8 traffic_class,
892 const struct ib_gid_attr *sgid_attr)
894 rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
896 attr->grh.sgid_attr = sgid_attr;
898 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
901 * rdma_destroy_ah_attr - Release reference to SGID attribute of
903 * @ah_attr: Pointer to ah attribute
905 * Release reference to the SGID attribute of the ah attribute if it is
906 * non NULL. It is safe to call this multiple times, and safe to call it on
907 * a zero initialized ah_attr.
909 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
911 if (ah_attr->grh.sgid_attr) {
912 rdma_put_gid_attr(ah_attr->grh.sgid_attr);
913 ah_attr->grh.sgid_attr = NULL;
916 EXPORT_SYMBOL(rdma_destroy_ah_attr);
918 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
919 const struct ib_grh *grh, u8 port_num)
921 struct rdma_ah_attr ah_attr;
925 ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
929 ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE);
931 rdma_destroy_ah_attr(&ah_attr);
934 EXPORT_SYMBOL(ib_create_ah_from_wc);
936 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
938 const struct ib_gid_attr *old_sgid_attr;
941 if (ah->type != ah_attr->type)
944 ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
948 ret = ah->device->ops.modify_ah ?
949 ah->device->ops.modify_ah(ah, ah_attr) :
952 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
953 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
956 EXPORT_SYMBOL(rdma_modify_ah);
958 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
960 ah_attr->grh.sgid_attr = NULL;
962 return ah->device->ops.query_ah ?
963 ah->device->ops.query_ah(ah, ah_attr) :
966 EXPORT_SYMBOL(rdma_query_ah);
968 int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata)
970 const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
974 might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE);
978 ret = ah->device->ops.destroy_ah(ah, flags);
982 atomic_dec(&pd->usecnt);
984 rdma_put_gid_attr(sgid_attr);
989 EXPORT_SYMBOL(rdma_destroy_ah_user);
991 /* Shared receive queues */
994 * ib_create_srq_user - Creates a SRQ associated with the specified protection
996 * @pd: The protection domain associated with the SRQ.
997 * @srq_init_attr: A list of initial attributes required to create the
998 * SRQ. If SRQ creation succeeds, then the attributes are updated to
999 * the actual capabilities of the created SRQ.
1000 * @uobject: uobject pointer if this is not a kernel SRQ
1001 * @udata: udata pointer if this is not a kernel SRQ
1003 * srq_attr->max_wr and srq_attr->max_sge are read the determine the
1004 * requested size of the SRQ, and set to the actual values allocated
1005 * on return. If ib_create_srq() succeeds, then max_wr and max_sge
1006 * will always be at least as large as the requested values.
1008 struct ib_srq *ib_create_srq_user(struct ib_pd *pd,
1009 struct ib_srq_init_attr *srq_init_attr,
1010 struct ib_usrq_object *uobject,
1011 struct ib_udata *udata)
1016 srq = rdma_zalloc_drv_obj(pd->device, ib_srq);
1018 return ERR_PTR(-ENOMEM);
1020 srq->device = pd->device;
1022 srq->event_handler = srq_init_attr->event_handler;
1023 srq->srq_context = srq_init_attr->srq_context;
1024 srq->srq_type = srq_init_attr->srq_type;
1025 srq->uobject = uobject;
1027 if (ib_srq_has_cq(srq->srq_type)) {
1028 srq->ext.cq = srq_init_attr->ext.cq;
1029 atomic_inc(&srq->ext.cq->usecnt);
1031 if (srq->srq_type == IB_SRQT_XRC) {
1032 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
1033 atomic_inc(&srq->ext.xrc.xrcd->usecnt);
1035 atomic_inc(&pd->usecnt);
1037 ret = pd->device->ops.create_srq(srq, srq_init_attr, udata);
1039 atomic_dec(&srq->pd->usecnt);
1040 if (srq->srq_type == IB_SRQT_XRC)
1041 atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1042 if (ib_srq_has_cq(srq->srq_type))
1043 atomic_dec(&srq->ext.cq->usecnt);
1045 return ERR_PTR(ret);
1050 EXPORT_SYMBOL(ib_create_srq_user);
1052 int ib_modify_srq(struct ib_srq *srq,
1053 struct ib_srq_attr *srq_attr,
1054 enum ib_srq_attr_mask srq_attr_mask)
1056 return srq->device->ops.modify_srq ?
1057 srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask,
1058 NULL) : -EOPNOTSUPP;
1060 EXPORT_SYMBOL(ib_modify_srq);
1062 int ib_query_srq(struct ib_srq *srq,
1063 struct ib_srq_attr *srq_attr)
1065 return srq->device->ops.query_srq ?
1066 srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP;
1068 EXPORT_SYMBOL(ib_query_srq);
1070 int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata)
1074 if (atomic_read(&srq->usecnt))
1077 ret = srq->device->ops.destroy_srq(srq, udata);
1081 atomic_dec(&srq->pd->usecnt);
1082 if (srq->srq_type == IB_SRQT_XRC)
1083 atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1084 if (ib_srq_has_cq(srq->srq_type))
1085 atomic_dec(&srq->ext.cq->usecnt);
1090 EXPORT_SYMBOL(ib_destroy_srq_user);
1094 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1096 struct ib_qp *qp = context;
1097 unsigned long flags;
1099 spin_lock_irqsave(&qp->device->qp_open_list_lock, flags);
1100 list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1101 if (event->element.qp->event_handler)
1102 event->element.qp->event_handler(event, event->element.qp->qp_context);
1103 spin_unlock_irqrestore(&qp->device->qp_open_list_lock, flags);
1106 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1107 void (*event_handler)(struct ib_event *, void *),
1111 unsigned long flags;
1114 qp = kzalloc(sizeof *qp, GFP_KERNEL);
1116 return ERR_PTR(-ENOMEM);
1118 qp->real_qp = real_qp;
1119 err = ib_open_shared_qp_security(qp, real_qp->device);
1122 return ERR_PTR(err);
1125 qp->real_qp = real_qp;
1126 atomic_inc(&real_qp->usecnt);
1127 qp->device = real_qp->device;
1128 qp->event_handler = event_handler;
1129 qp->qp_context = qp_context;
1130 qp->qp_num = real_qp->qp_num;
1131 qp->qp_type = real_qp->qp_type;
1133 spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1134 list_add(&qp->open_list, &real_qp->open_list);
1135 spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1140 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1141 struct ib_qp_open_attr *qp_open_attr)
1143 struct ib_qp *qp, *real_qp;
1145 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1146 return ERR_PTR(-EINVAL);
1148 down_read(&xrcd->tgt_qps_rwsem);
1149 real_qp = xa_load(&xrcd->tgt_qps, qp_open_attr->qp_num);
1151 up_read(&xrcd->tgt_qps_rwsem);
1152 return ERR_PTR(-EINVAL);
1154 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1155 qp_open_attr->qp_context);
1156 up_read(&xrcd->tgt_qps_rwsem);
1159 EXPORT_SYMBOL(ib_open_qp);
1161 static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp,
1162 struct ib_qp_init_attr *qp_init_attr)
1164 struct ib_qp *real_qp = qp;
1167 qp->event_handler = __ib_shared_qp_event_handler;
1168 qp->qp_context = qp;
1170 qp->send_cq = qp->recv_cq = NULL;
1172 qp->xrcd = qp_init_attr->xrcd;
1173 atomic_inc(&qp_init_attr->xrcd->usecnt);
1174 INIT_LIST_HEAD(&qp->open_list);
1176 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1177 qp_init_attr->qp_context);
1181 err = xa_err(xa_store(&qp_init_attr->xrcd->tgt_qps, real_qp->qp_num,
1182 real_qp, GFP_KERNEL));
1185 return ERR_PTR(err);
1191 * ib_create_qp - Creates a kernel QP associated with the specified protection
1193 * @pd: The protection domain associated with the QP.
1194 * @qp_init_attr: A list of initial attributes required to create the
1195 * QP. If QP creation succeeds, then the attributes are updated to
1196 * the actual capabilities of the created QP.
1198 * NOTE: for user qp use ib_create_qp_user with valid udata!
1200 struct ib_qp *ib_create_qp(struct ib_pd *pd,
1201 struct ib_qp_init_attr *qp_init_attr)
1203 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
1207 if (qp_init_attr->rwq_ind_tbl &&
1208 (qp_init_attr->recv_cq ||
1209 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
1210 qp_init_attr->cap.max_recv_sge))
1211 return ERR_PTR(-EINVAL);
1213 if ((qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN) &&
1214 !(device->attrs.device_cap_flags & IB_DEVICE_INTEGRITY_HANDOVER))
1215 return ERR_PTR(-EINVAL);
1218 * If the callers is using the RDMA API calculate the resources
1219 * needed for the RDMA READ/WRITE operations.
1221 * Note that these callers need to pass in a port number.
1223 if (qp_init_attr->cap.max_rdma_ctxs)
1224 rdma_rw_init_qp(device, qp_init_attr);
1226 qp = _ib_create_qp(device, pd, qp_init_attr, NULL, NULL);
1230 ret = ib_create_qp_security(qp, device);
1234 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT) {
1235 struct ib_qp *xrc_qp =
1236 create_xrc_qp_user(qp, qp_init_attr);
1238 if (IS_ERR(xrc_qp)) {
1239 ret = PTR_ERR(xrc_qp);
1245 qp->event_handler = qp_init_attr->event_handler;
1246 qp->qp_context = qp_init_attr->qp_context;
1247 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
1251 qp->recv_cq = qp_init_attr->recv_cq;
1252 if (qp_init_attr->recv_cq)
1253 atomic_inc(&qp_init_attr->recv_cq->usecnt);
1254 qp->srq = qp_init_attr->srq;
1256 atomic_inc(&qp_init_attr->srq->usecnt);
1259 qp->send_cq = qp_init_attr->send_cq;
1262 atomic_inc(&pd->usecnt);
1263 if (qp_init_attr->send_cq)
1264 atomic_inc(&qp_init_attr->send_cq->usecnt);
1265 if (qp_init_attr->rwq_ind_tbl)
1266 atomic_inc(&qp->rwq_ind_tbl->usecnt);
1268 if (qp_init_attr->cap.max_rdma_ctxs) {
1269 ret = rdma_rw_init_mrs(qp, qp_init_attr);
1275 * Note: all hw drivers guarantee that max_send_sge is lower than
1276 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1277 * max_send_sge <= max_sge_rd.
1279 qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1280 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1281 device->attrs.max_sge_rd);
1282 if (qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN)
1283 qp->integrity_en = true;
1289 return ERR_PTR(ret);
1292 EXPORT_SYMBOL(ib_create_qp);
1294 static const struct {
1296 enum ib_qp_attr_mask req_param[IB_QPT_MAX];
1297 enum ib_qp_attr_mask opt_param[IB_QPT_MAX];
1298 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1300 [IB_QPS_RESET] = { .valid = 1 },
1304 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1307 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
1308 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1310 IB_QP_ACCESS_FLAGS),
1311 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1313 IB_QP_ACCESS_FLAGS),
1314 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1316 IB_QP_ACCESS_FLAGS),
1317 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1319 IB_QP_ACCESS_FLAGS),
1320 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1322 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1328 [IB_QPS_RESET] = { .valid = 1 },
1329 [IB_QPS_ERR] = { .valid = 1 },
1333 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1336 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1338 IB_QP_ACCESS_FLAGS),
1339 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1341 IB_QP_ACCESS_FLAGS),
1342 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1344 IB_QP_ACCESS_FLAGS),
1345 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1347 IB_QP_ACCESS_FLAGS),
1348 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1350 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1357 [IB_QPT_UC] = (IB_QP_AV |
1361 [IB_QPT_RC] = (IB_QP_AV |
1365 IB_QP_MAX_DEST_RD_ATOMIC |
1366 IB_QP_MIN_RNR_TIMER),
1367 [IB_QPT_XRC_INI] = (IB_QP_AV |
1371 [IB_QPT_XRC_TGT] = (IB_QP_AV |
1375 IB_QP_MAX_DEST_RD_ATOMIC |
1376 IB_QP_MIN_RNR_TIMER),
1379 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1381 [IB_QPT_UC] = (IB_QP_ALT_PATH |
1382 IB_QP_ACCESS_FLAGS |
1384 [IB_QPT_RC] = (IB_QP_ALT_PATH |
1385 IB_QP_ACCESS_FLAGS |
1387 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH |
1388 IB_QP_ACCESS_FLAGS |
1390 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH |
1391 IB_QP_ACCESS_FLAGS |
1393 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1395 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1401 [IB_QPS_RESET] = { .valid = 1 },
1402 [IB_QPS_ERR] = { .valid = 1 },
1406 [IB_QPT_UD] = IB_QP_SQ_PSN,
1407 [IB_QPT_UC] = IB_QP_SQ_PSN,
1408 [IB_QPT_RC] = (IB_QP_TIMEOUT |
1412 IB_QP_MAX_QP_RD_ATOMIC),
1413 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT |
1417 IB_QP_MAX_QP_RD_ATOMIC),
1418 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT |
1420 [IB_QPT_SMI] = IB_QP_SQ_PSN,
1421 [IB_QPT_GSI] = IB_QP_SQ_PSN,
1424 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1426 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1428 IB_QP_ACCESS_FLAGS |
1429 IB_QP_PATH_MIG_STATE),
1430 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1432 IB_QP_ACCESS_FLAGS |
1433 IB_QP_MIN_RNR_TIMER |
1434 IB_QP_PATH_MIG_STATE),
1435 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1437 IB_QP_ACCESS_FLAGS |
1438 IB_QP_PATH_MIG_STATE),
1439 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1441 IB_QP_ACCESS_FLAGS |
1442 IB_QP_MIN_RNR_TIMER |
1443 IB_QP_PATH_MIG_STATE),
1444 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1446 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1448 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1453 [IB_QPS_RESET] = { .valid = 1 },
1454 [IB_QPS_ERR] = { .valid = 1 },
1458 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1460 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1461 IB_QP_ACCESS_FLAGS |
1463 IB_QP_PATH_MIG_STATE),
1464 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1465 IB_QP_ACCESS_FLAGS |
1467 IB_QP_PATH_MIG_STATE |
1468 IB_QP_MIN_RNR_TIMER),
1469 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1470 IB_QP_ACCESS_FLAGS |
1472 IB_QP_PATH_MIG_STATE),
1473 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1474 IB_QP_ACCESS_FLAGS |
1476 IB_QP_PATH_MIG_STATE |
1477 IB_QP_MIN_RNR_TIMER),
1478 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1480 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1482 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1488 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1489 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1490 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1491 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1492 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1493 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1494 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1499 [IB_QPS_RESET] = { .valid = 1 },
1500 [IB_QPS_ERR] = { .valid = 1 },
1504 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1506 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1508 IB_QP_ACCESS_FLAGS |
1509 IB_QP_PATH_MIG_STATE),
1510 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1512 IB_QP_ACCESS_FLAGS |
1513 IB_QP_MIN_RNR_TIMER |
1514 IB_QP_PATH_MIG_STATE),
1515 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1517 IB_QP_ACCESS_FLAGS |
1518 IB_QP_PATH_MIG_STATE),
1519 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1521 IB_QP_ACCESS_FLAGS |
1522 IB_QP_MIN_RNR_TIMER |
1523 IB_QP_PATH_MIG_STATE),
1524 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1526 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1533 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1535 [IB_QPT_UC] = (IB_QP_AV |
1537 IB_QP_ACCESS_FLAGS |
1539 IB_QP_PATH_MIG_STATE),
1540 [IB_QPT_RC] = (IB_QP_PORT |
1545 IB_QP_MAX_QP_RD_ATOMIC |
1546 IB_QP_MAX_DEST_RD_ATOMIC |
1548 IB_QP_ACCESS_FLAGS |
1550 IB_QP_MIN_RNR_TIMER |
1551 IB_QP_PATH_MIG_STATE),
1552 [IB_QPT_XRC_INI] = (IB_QP_PORT |
1557 IB_QP_MAX_QP_RD_ATOMIC |
1559 IB_QP_ACCESS_FLAGS |
1561 IB_QP_PATH_MIG_STATE),
1562 [IB_QPT_XRC_TGT] = (IB_QP_PORT |
1565 IB_QP_MAX_DEST_RD_ATOMIC |
1567 IB_QP_ACCESS_FLAGS |
1569 IB_QP_MIN_RNR_TIMER |
1570 IB_QP_PATH_MIG_STATE),
1571 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1573 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1579 [IB_QPS_RESET] = { .valid = 1 },
1580 [IB_QPS_ERR] = { .valid = 1 },
1584 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1586 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1587 IB_QP_ACCESS_FLAGS),
1588 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1590 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1596 [IB_QPS_RESET] = { .valid = 1 },
1597 [IB_QPS_ERR] = { .valid = 1 }
1601 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1602 enum ib_qp_type type, enum ib_qp_attr_mask mask)
1604 enum ib_qp_attr_mask req_param, opt_param;
1606 if (mask & IB_QP_CUR_STATE &&
1607 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1608 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1611 if (!qp_state_table[cur_state][next_state].valid)
1614 req_param = qp_state_table[cur_state][next_state].req_param[type];
1615 opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1617 if ((mask & req_param) != req_param)
1620 if (mask & ~(req_param | opt_param | IB_QP_STATE))
1625 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1628 * ib_resolve_eth_dmac - Resolve destination mac address
1629 * @device: Device to consider
1630 * @ah_attr: address handle attribute which describes the
1631 * source and destination parameters
1632 * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1633 * returns 0 on success or appropriate error code. It initializes the
1634 * necessary ah_attr fields when call is successful.
1636 static int ib_resolve_eth_dmac(struct ib_device *device,
1637 struct rdma_ah_attr *ah_attr)
1641 if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1642 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1645 memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1646 ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1648 ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1649 (char *)ah_attr->roce.dmac);
1652 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1657 static bool is_qp_type_connected(const struct ib_qp *qp)
1659 return (qp->qp_type == IB_QPT_UC ||
1660 qp->qp_type == IB_QPT_RC ||
1661 qp->qp_type == IB_QPT_XRC_INI ||
1662 qp->qp_type == IB_QPT_XRC_TGT);
1666 * IB core internal function to perform QP attributes modification.
1668 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1669 int attr_mask, struct ib_udata *udata)
1671 u8 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1672 const struct ib_gid_attr *old_sgid_attr_av;
1673 const struct ib_gid_attr *old_sgid_attr_alt_av;
1676 attr->xmit_slave = NULL;
1677 if (attr_mask & IB_QP_AV) {
1678 ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1683 if (attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1684 is_qp_type_connected(qp)) {
1685 struct net_device *slave;
1688 * If the user provided the qp_attr then we have to
1689 * resolve it. Kerne users have to provide already
1690 * resolved rdma_ah_attr's.
1693 ret = ib_resolve_eth_dmac(qp->device,
1698 slave = rdma_lag_get_ah_roce_slave(qp->device,
1703 attr->xmit_slave = slave;
1706 if (attr_mask & IB_QP_ALT_PATH) {
1708 * FIXME: This does not track the migration state, so if the
1709 * user loads a new alternate path after the HW has migrated
1710 * from primary->alternate we will keep the wrong
1711 * references. This is OK for IB because the reference
1712 * counting does not serve any functional purpose.
1714 ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1715 &old_sgid_attr_alt_av);
1720 * Today the core code can only handle alternate paths and APM
1721 * for IB. Ban them in roce mode.
1723 if (!(rdma_protocol_ib(qp->device,
1724 attr->alt_ah_attr.port_num) &&
1725 rdma_protocol_ib(qp->device, port))) {
1731 if (rdma_ib_or_roce(qp->device, port)) {
1732 if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1733 dev_warn(&qp->device->dev,
1734 "%s rq_psn overflow, masking to 24 bits\n",
1736 attr->rq_psn &= 0xffffff;
1739 if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1740 dev_warn(&qp->device->dev,
1741 " %s sq_psn overflow, masking to 24 bits\n",
1743 attr->sq_psn &= 0xffffff;
1748 * Bind this qp to a counter automatically based on the rdma counter
1749 * rules. This only set in RST2INIT with port specified
1751 if (!qp->counter && (attr_mask & IB_QP_PORT) &&
1752 ((attr_mask & IB_QP_STATE) && attr->qp_state == IB_QPS_INIT))
1753 rdma_counter_bind_qp_auto(qp, attr->port_num);
1755 ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1759 if (attr_mask & IB_QP_PORT)
1760 qp->port = attr->port_num;
1761 if (attr_mask & IB_QP_AV)
1763 rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1764 if (attr_mask & IB_QP_ALT_PATH)
1765 qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1766 &attr->alt_ah_attr, qp->alt_path_sgid_attr);
1769 if (attr_mask & IB_QP_ALT_PATH)
1770 rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1772 if (attr_mask & IB_QP_AV) {
1773 rdma_lag_put_ah_roce_slave(attr->xmit_slave);
1774 rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1780 * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1781 * @ib_qp: The QP to modify.
1782 * @attr: On input, specifies the QP attributes to modify. On output,
1783 * the current values of selected QP attributes are returned.
1784 * @attr_mask: A bit-mask used to specify which attributes of the QP
1785 * are being modified.
1786 * @udata: pointer to user's input output buffer information
1787 * are being modified.
1788 * It returns 0 on success and returns appropriate error code on error.
1790 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1791 int attr_mask, struct ib_udata *udata)
1793 return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1795 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1797 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u16 *speed, u8 *width)
1801 struct net_device *netdev;
1802 struct ethtool_link_ksettings lksettings;
1804 if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1807 netdev = ib_device_get_netdev(dev, port_num);
1812 rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1817 if (!rc && lksettings.base.speed != (u32)SPEED_UNKNOWN) {
1818 netdev_speed = lksettings.base.speed;
1820 netdev_speed = SPEED_1000;
1821 pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name,
1825 if (netdev_speed <= SPEED_1000) {
1826 *width = IB_WIDTH_1X;
1827 *speed = IB_SPEED_SDR;
1828 } else if (netdev_speed <= SPEED_10000) {
1829 *width = IB_WIDTH_1X;
1830 *speed = IB_SPEED_FDR10;
1831 } else if (netdev_speed <= SPEED_20000) {
1832 *width = IB_WIDTH_4X;
1833 *speed = IB_SPEED_DDR;
1834 } else if (netdev_speed <= SPEED_25000) {
1835 *width = IB_WIDTH_1X;
1836 *speed = IB_SPEED_EDR;
1837 } else if (netdev_speed <= SPEED_40000) {
1838 *width = IB_WIDTH_4X;
1839 *speed = IB_SPEED_FDR10;
1841 *width = IB_WIDTH_4X;
1842 *speed = IB_SPEED_EDR;
1847 EXPORT_SYMBOL(ib_get_eth_speed);
1849 int ib_modify_qp(struct ib_qp *qp,
1850 struct ib_qp_attr *qp_attr,
1853 return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1855 EXPORT_SYMBOL(ib_modify_qp);
1857 int ib_query_qp(struct ib_qp *qp,
1858 struct ib_qp_attr *qp_attr,
1860 struct ib_qp_init_attr *qp_init_attr)
1862 qp_attr->ah_attr.grh.sgid_attr = NULL;
1863 qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
1865 return qp->device->ops.query_qp ?
1866 qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask,
1867 qp_init_attr) : -EOPNOTSUPP;
1869 EXPORT_SYMBOL(ib_query_qp);
1871 int ib_close_qp(struct ib_qp *qp)
1873 struct ib_qp *real_qp;
1874 unsigned long flags;
1876 real_qp = qp->real_qp;
1880 spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1881 list_del(&qp->open_list);
1882 spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1884 atomic_dec(&real_qp->usecnt);
1886 ib_close_shared_qp_security(qp->qp_sec);
1891 EXPORT_SYMBOL(ib_close_qp);
1893 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1895 struct ib_xrcd *xrcd;
1896 struct ib_qp *real_qp;
1899 real_qp = qp->real_qp;
1900 xrcd = real_qp->xrcd;
1901 down_write(&xrcd->tgt_qps_rwsem);
1903 if (atomic_read(&real_qp->usecnt) == 0)
1904 xa_erase(&xrcd->tgt_qps, real_qp->qp_num);
1907 up_write(&xrcd->tgt_qps_rwsem);
1910 ret = ib_destroy_qp(real_qp);
1912 atomic_dec(&xrcd->usecnt);
1918 int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata)
1920 const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
1921 const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
1923 struct ib_cq *scq, *rcq;
1925 struct ib_rwq_ind_table *ind_tbl;
1926 struct ib_qp_security *sec;
1929 WARN_ON_ONCE(qp->mrs_used > 0);
1931 if (atomic_read(&qp->usecnt))
1934 if (qp->real_qp != qp)
1935 return __ib_destroy_shared_qp(qp);
1941 ind_tbl = qp->rwq_ind_tbl;
1944 ib_destroy_qp_security_begin(sec);
1947 rdma_rw_cleanup_mrs(qp);
1949 rdma_counter_unbind_qp(qp, true);
1950 rdma_restrack_del(&qp->res);
1951 ret = qp->device->ops.destroy_qp(qp, udata);
1953 if (alt_path_sgid_attr)
1954 rdma_put_gid_attr(alt_path_sgid_attr);
1956 rdma_put_gid_attr(av_sgid_attr);
1958 atomic_dec(&pd->usecnt);
1960 atomic_dec(&scq->usecnt);
1962 atomic_dec(&rcq->usecnt);
1964 atomic_dec(&srq->usecnt);
1966 atomic_dec(&ind_tbl->usecnt);
1968 ib_destroy_qp_security_end(sec);
1971 ib_destroy_qp_security_abort(sec);
1976 EXPORT_SYMBOL(ib_destroy_qp_user);
1978 /* Completion queues */
1980 struct ib_cq *__ib_create_cq(struct ib_device *device,
1981 ib_comp_handler comp_handler,
1982 void (*event_handler)(struct ib_event *, void *),
1984 const struct ib_cq_init_attr *cq_attr,
1990 cq = rdma_zalloc_drv_obj(device, ib_cq);
1992 return ERR_PTR(-ENOMEM);
1994 cq->device = device;
1996 cq->comp_handler = comp_handler;
1997 cq->event_handler = event_handler;
1998 cq->cq_context = cq_context;
1999 atomic_set(&cq->usecnt, 0);
2001 rdma_restrack_new(&cq->res, RDMA_RESTRACK_CQ);
2002 rdma_restrack_set_name(&cq->res, caller);
2004 ret = device->ops.create_cq(cq, cq_attr, NULL);
2006 rdma_restrack_put(&cq->res);
2008 return ERR_PTR(ret);
2011 rdma_restrack_add(&cq->res);
2014 EXPORT_SYMBOL(__ib_create_cq);
2016 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
2021 return cq->device->ops.modify_cq ?
2022 cq->device->ops.modify_cq(cq, cq_count,
2023 cq_period) : -EOPNOTSUPP;
2025 EXPORT_SYMBOL(rdma_set_cq_moderation);
2027 int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata)
2031 if (WARN_ON_ONCE(cq->shared))
2034 if (atomic_read(&cq->usecnt))
2037 ret = cq->device->ops.destroy_cq(cq, udata);
2041 rdma_restrack_del(&cq->res);
2045 EXPORT_SYMBOL(ib_destroy_cq_user);
2047 int ib_resize_cq(struct ib_cq *cq, int cqe)
2052 return cq->device->ops.resize_cq ?
2053 cq->device->ops.resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
2055 EXPORT_SYMBOL(ib_resize_cq);
2057 /* Memory regions */
2059 struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
2060 u64 virt_addr, int access_flags)
2064 if (access_flags & IB_ACCESS_ON_DEMAND) {
2065 if (!(pd->device->attrs.device_cap_flags &
2066 IB_DEVICE_ON_DEMAND_PAGING)) {
2067 pr_debug("ODP support not available\n");
2068 return ERR_PTR(-EINVAL);
2072 mr = pd->device->ops.reg_user_mr(pd, start, length, virt_addr,
2073 access_flags, NULL);
2078 mr->device = pd->device;
2081 atomic_inc(&pd->usecnt);
2083 rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2084 rdma_restrack_parent_name(&mr->res, &pd->res);
2085 rdma_restrack_add(&mr->res);
2089 EXPORT_SYMBOL(ib_reg_user_mr);
2091 int ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
2092 u32 flags, struct ib_sge *sg_list, u32 num_sge)
2094 if (!pd->device->ops.advise_mr)
2100 return pd->device->ops.advise_mr(pd, advice, flags, sg_list, num_sge,
2103 EXPORT_SYMBOL(ib_advise_mr);
2105 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata)
2107 struct ib_pd *pd = mr->pd;
2108 struct ib_dm *dm = mr->dm;
2109 struct ib_sig_attrs *sig_attrs = mr->sig_attrs;
2113 rdma_restrack_del(&mr->res);
2114 ret = mr->device->ops.dereg_mr(mr, udata);
2116 atomic_dec(&pd->usecnt);
2118 atomic_dec(&dm->usecnt);
2124 EXPORT_SYMBOL(ib_dereg_mr_user);
2127 * ib_alloc_mr() - Allocates a memory region
2128 * @pd: protection domain associated with the region
2129 * @mr_type: memory region type
2130 * @max_num_sg: maximum sg entries available for registration.
2133 * Memory registeration page/sg lists must not exceed max_num_sg.
2134 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
2135 * max_num_sg * used_page_size.
2138 struct ib_mr *ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2143 if (!pd->device->ops.alloc_mr) {
2144 mr = ERR_PTR(-EOPNOTSUPP);
2148 if (mr_type == IB_MR_TYPE_INTEGRITY) {
2150 mr = ERR_PTR(-EINVAL);
2154 mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg);
2158 mr->device = pd->device;
2162 atomic_inc(&pd->usecnt);
2163 mr->need_inval = false;
2165 mr->sig_attrs = NULL;
2167 rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2168 rdma_restrack_parent_name(&mr->res, &pd->res);
2169 rdma_restrack_add(&mr->res);
2171 trace_mr_alloc(pd, mr_type, max_num_sg, mr);
2174 EXPORT_SYMBOL(ib_alloc_mr);
2177 * ib_alloc_mr_integrity() - Allocates an integrity memory region
2178 * @pd: protection domain associated with the region
2179 * @max_num_data_sg: maximum data sg entries available for registration
2180 * @max_num_meta_sg: maximum metadata sg entries available for
2184 * Memory registration page/sg lists must not exceed max_num_sg,
2185 * also the integrity page/sg lists must not exceed max_num_meta_sg.
2188 struct ib_mr *ib_alloc_mr_integrity(struct ib_pd *pd,
2189 u32 max_num_data_sg,
2190 u32 max_num_meta_sg)
2193 struct ib_sig_attrs *sig_attrs;
2195 if (!pd->device->ops.alloc_mr_integrity ||
2196 !pd->device->ops.map_mr_sg_pi) {
2197 mr = ERR_PTR(-EOPNOTSUPP);
2201 if (!max_num_meta_sg) {
2202 mr = ERR_PTR(-EINVAL);
2206 sig_attrs = kzalloc(sizeof(struct ib_sig_attrs), GFP_KERNEL);
2208 mr = ERR_PTR(-ENOMEM);
2212 mr = pd->device->ops.alloc_mr_integrity(pd, max_num_data_sg,
2219 mr->device = pd->device;
2223 atomic_inc(&pd->usecnt);
2224 mr->need_inval = false;
2225 mr->type = IB_MR_TYPE_INTEGRITY;
2226 mr->sig_attrs = sig_attrs;
2228 rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2229 rdma_restrack_parent_name(&mr->res, &pd->res);
2230 rdma_restrack_add(&mr->res);
2232 trace_mr_integ_alloc(pd, max_num_data_sg, max_num_meta_sg, mr);
2235 EXPORT_SYMBOL(ib_alloc_mr_integrity);
2237 /* Multicast groups */
2239 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2241 struct ib_qp_init_attr init_attr = {};
2242 struct ib_qp_attr attr = {};
2243 int num_eth_ports = 0;
2246 /* If QP state >= init, it is assigned to a port and we can check this
2249 if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2250 if (attr.qp_state >= IB_QPS_INIT) {
2251 if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2252 IB_LINK_LAYER_INFINIBAND)
2258 /* Can't get a quick answer, iterate over all ports */
2259 for (port = 0; port < qp->device->phys_port_cnt; port++)
2260 if (rdma_port_get_link_layer(qp->device, port) !=
2261 IB_LINK_LAYER_INFINIBAND)
2264 /* If we have at lease one Ethernet port, RoCE annex declares that
2265 * multicast LID should be ignored. We can't tell at this step if the
2266 * QP belongs to an IB or Ethernet port.
2271 /* If all the ports are IB, we can check according to IB spec. */
2273 return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2274 lid == be16_to_cpu(IB_LID_PERMISSIVE));
2277 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2281 if (!qp->device->ops.attach_mcast)
2284 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2285 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2288 ret = qp->device->ops.attach_mcast(qp, gid, lid);
2290 atomic_inc(&qp->usecnt);
2293 EXPORT_SYMBOL(ib_attach_mcast);
2295 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2299 if (!qp->device->ops.detach_mcast)
2302 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2303 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2306 ret = qp->device->ops.detach_mcast(qp, gid, lid);
2308 atomic_dec(&qp->usecnt);
2311 EXPORT_SYMBOL(ib_detach_mcast);
2314 * ib_alloc_xrcd_user - Allocates an XRC domain.
2315 * @device: The device on which to allocate the XRC domain.
2316 * @inode: inode to connect XRCD
2317 * @udata: Valid user data or NULL for kernel object
2319 struct ib_xrcd *ib_alloc_xrcd_user(struct ib_device *device,
2320 struct inode *inode, struct ib_udata *udata)
2322 struct ib_xrcd *xrcd;
2325 if (!device->ops.alloc_xrcd)
2326 return ERR_PTR(-EOPNOTSUPP);
2328 xrcd = rdma_zalloc_drv_obj(device, ib_xrcd);
2330 return ERR_PTR(-ENOMEM);
2332 xrcd->device = device;
2333 xrcd->inode = inode;
2334 atomic_set(&xrcd->usecnt, 0);
2335 init_rwsem(&xrcd->tgt_qps_rwsem);
2336 xa_init(&xrcd->tgt_qps);
2338 ret = device->ops.alloc_xrcd(xrcd, udata);
2344 return ERR_PTR(ret);
2346 EXPORT_SYMBOL(ib_alloc_xrcd_user);
2349 * ib_dealloc_xrcd_user - Deallocates an XRC domain.
2350 * @xrcd: The XRC domain to deallocate.
2351 * @udata: Valid user data or NULL for kernel object
2353 int ib_dealloc_xrcd_user(struct ib_xrcd *xrcd, struct ib_udata *udata)
2357 if (atomic_read(&xrcd->usecnt))
2360 WARN_ON(!xa_empty(&xrcd->tgt_qps));
2361 ret = xrcd->device->ops.dealloc_xrcd(xrcd, udata);
2367 EXPORT_SYMBOL(ib_dealloc_xrcd_user);
2370 * ib_create_wq - Creates a WQ associated with the specified protection
2372 * @pd: The protection domain associated with the WQ.
2373 * @wq_attr: A list of initial attributes required to create the
2374 * WQ. If WQ creation succeeds, then the attributes are updated to
2375 * the actual capabilities of the created WQ.
2377 * wq_attr->max_wr and wq_attr->max_sge determine
2378 * the requested size of the WQ, and set to the actual values allocated
2380 * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2381 * at least as large as the requested values.
2383 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2384 struct ib_wq_init_attr *wq_attr)
2388 if (!pd->device->ops.create_wq)
2389 return ERR_PTR(-EOPNOTSUPP);
2391 wq = pd->device->ops.create_wq(pd, wq_attr, NULL);
2393 wq->event_handler = wq_attr->event_handler;
2394 wq->wq_context = wq_attr->wq_context;
2395 wq->wq_type = wq_attr->wq_type;
2396 wq->cq = wq_attr->cq;
2397 wq->device = pd->device;
2400 atomic_inc(&pd->usecnt);
2401 atomic_inc(&wq_attr->cq->usecnt);
2402 atomic_set(&wq->usecnt, 0);
2406 EXPORT_SYMBOL(ib_create_wq);
2409 * ib_destroy_wq_user - Destroys the specified user WQ.
2410 * @wq: The WQ to destroy.
2411 * @udata: Valid user data
2413 int ib_destroy_wq_user(struct ib_wq *wq, struct ib_udata *udata)
2415 struct ib_cq *cq = wq->cq;
2416 struct ib_pd *pd = wq->pd;
2419 if (atomic_read(&wq->usecnt))
2422 ret = wq->device->ops.destroy_wq(wq, udata);
2426 atomic_dec(&pd->usecnt);
2427 atomic_dec(&cq->usecnt);
2430 EXPORT_SYMBOL(ib_destroy_wq_user);
2433 * ib_modify_wq - Modifies the specified WQ.
2434 * @wq: The WQ to modify.
2435 * @wq_attr: On input, specifies the WQ attributes to modify.
2436 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
2437 * are being modified.
2438 * On output, the current values of selected WQ attributes are returned.
2440 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
2445 if (!wq->device->ops.modify_wq)
2448 err = wq->device->ops.modify_wq(wq, wq_attr, wq_attr_mask, NULL);
2451 EXPORT_SYMBOL(ib_modify_wq);
2453 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2454 struct ib_mr_status *mr_status)
2456 if (!mr->device->ops.check_mr_status)
2459 return mr->device->ops.check_mr_status(mr, check_mask, mr_status);
2461 EXPORT_SYMBOL(ib_check_mr_status);
2463 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
2466 if (!device->ops.set_vf_link_state)
2469 return device->ops.set_vf_link_state(device, vf, port, state);
2471 EXPORT_SYMBOL(ib_set_vf_link_state);
2473 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
2474 struct ifla_vf_info *info)
2476 if (!device->ops.get_vf_config)
2479 return device->ops.get_vf_config(device, vf, port, info);
2481 EXPORT_SYMBOL(ib_get_vf_config);
2483 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
2484 struct ifla_vf_stats *stats)
2486 if (!device->ops.get_vf_stats)
2489 return device->ops.get_vf_stats(device, vf, port, stats);
2491 EXPORT_SYMBOL(ib_get_vf_stats);
2493 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
2496 if (!device->ops.set_vf_guid)
2499 return device->ops.set_vf_guid(device, vf, port, guid, type);
2501 EXPORT_SYMBOL(ib_set_vf_guid);
2503 int ib_get_vf_guid(struct ib_device *device, int vf, u8 port,
2504 struct ifla_vf_guid *node_guid,
2505 struct ifla_vf_guid *port_guid)
2507 if (!device->ops.get_vf_guid)
2510 return device->ops.get_vf_guid(device, vf, port, node_guid, port_guid);
2512 EXPORT_SYMBOL(ib_get_vf_guid);
2514 * ib_map_mr_sg_pi() - Map the dma mapped SG lists for PI (protection
2515 * information) and set an appropriate memory region for registration.
2516 * @mr: memory region
2517 * @data_sg: dma mapped scatterlist for data
2518 * @data_sg_nents: number of entries in data_sg
2519 * @data_sg_offset: offset in bytes into data_sg
2520 * @meta_sg: dma mapped scatterlist for metadata
2521 * @meta_sg_nents: number of entries in meta_sg
2522 * @meta_sg_offset: offset in bytes into meta_sg
2523 * @page_size: page vector desired page size
2526 * - The MR must be allocated with type IB_MR_TYPE_INTEGRITY.
2528 * Return: 0 on success.
2530 * After this completes successfully, the memory region
2531 * is ready for registration.
2533 int ib_map_mr_sg_pi(struct ib_mr *mr, struct scatterlist *data_sg,
2534 int data_sg_nents, unsigned int *data_sg_offset,
2535 struct scatterlist *meta_sg, int meta_sg_nents,
2536 unsigned int *meta_sg_offset, unsigned int page_size)
2538 if (unlikely(!mr->device->ops.map_mr_sg_pi ||
2539 WARN_ON_ONCE(mr->type != IB_MR_TYPE_INTEGRITY)))
2542 mr->page_size = page_size;
2544 return mr->device->ops.map_mr_sg_pi(mr, data_sg, data_sg_nents,
2545 data_sg_offset, meta_sg,
2546 meta_sg_nents, meta_sg_offset);
2548 EXPORT_SYMBOL(ib_map_mr_sg_pi);
2551 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2552 * and set it the memory region.
2553 * @mr: memory region
2554 * @sg: dma mapped scatterlist
2555 * @sg_nents: number of entries in sg
2556 * @sg_offset: offset in bytes into sg
2557 * @page_size: page vector desired page size
2561 * - The first sg element is allowed to have an offset.
2562 * - Each sg element must either be aligned to page_size or virtually
2563 * contiguous to the previous element. In case an sg element has a
2564 * non-contiguous offset, the mapping prefix will not include it.
2565 * - The last sg element is allowed to have length less than page_size.
2566 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2567 * then only max_num_sg entries will be mapped.
2568 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2569 * constraints holds and the page_size argument is ignored.
2571 * Returns the number of sg elements that were mapped to the memory region.
2573 * After this completes successfully, the memory region
2574 * is ready for registration.
2576 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2577 unsigned int *sg_offset, unsigned int page_size)
2579 if (unlikely(!mr->device->ops.map_mr_sg))
2582 mr->page_size = page_size;
2584 return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset);
2586 EXPORT_SYMBOL(ib_map_mr_sg);
2589 * ib_sg_to_pages() - Convert the largest prefix of a sg list
2591 * @mr: memory region
2592 * @sgl: dma mapped scatterlist
2593 * @sg_nents: number of entries in sg
2594 * @sg_offset_p: ==== =======================================================
2595 * IN start offset in bytes into sg
2596 * OUT offset in bytes for element n of the sg of the first
2597 * byte that has not been processed where n is the return
2598 * value of this function.
2599 * ==== =======================================================
2600 * @set_page: driver page assignment function pointer
2602 * Core service helper for drivers to convert the largest
2603 * prefix of given sg list to a page vector. The sg list
2604 * prefix converted is the prefix that meet the requirements
2607 * Returns the number of sg elements that were assigned to
2610 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2611 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2613 struct scatterlist *sg;
2614 u64 last_end_dma_addr = 0;
2615 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2616 unsigned int last_page_off = 0;
2617 u64 page_mask = ~((u64)mr->page_size - 1);
2620 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2623 mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2626 for_each_sg(sgl, sg, sg_nents, i) {
2627 u64 dma_addr = sg_dma_address(sg) + sg_offset;
2628 u64 prev_addr = dma_addr;
2629 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2630 u64 end_dma_addr = dma_addr + dma_len;
2631 u64 page_addr = dma_addr & page_mask;
2634 * For the second and later elements, check whether either the
2635 * end of element i-1 or the start of element i is not aligned
2636 * on a page boundary.
2638 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2639 /* Stop mapping if there is a gap. */
2640 if (last_end_dma_addr != dma_addr)
2644 * Coalesce this element with the last. If it is small
2645 * enough just update mr->length. Otherwise start
2646 * mapping from the next page.
2652 ret = set_page(mr, page_addr);
2653 if (unlikely(ret < 0)) {
2654 sg_offset = prev_addr - sg_dma_address(sg);
2655 mr->length += prev_addr - dma_addr;
2657 *sg_offset_p = sg_offset;
2658 return i || sg_offset ? i : ret;
2660 prev_addr = page_addr;
2662 page_addr += mr->page_size;
2663 } while (page_addr < end_dma_addr);
2665 mr->length += dma_len;
2666 last_end_dma_addr = end_dma_addr;
2667 last_page_off = end_dma_addr & ~page_mask;
2676 EXPORT_SYMBOL(ib_sg_to_pages);
2678 struct ib_drain_cqe {
2680 struct completion done;
2683 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2685 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2688 complete(&cqe->done);
2692 * Post a WR and block until its completion is reaped for the SQ.
2694 static void __ib_drain_sq(struct ib_qp *qp)
2696 struct ib_cq *cq = qp->send_cq;
2697 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2698 struct ib_drain_cqe sdrain;
2699 struct ib_rdma_wr swr = {
2702 { .wr_cqe = &sdrain.cqe, },
2703 .opcode = IB_WR_RDMA_WRITE,
2708 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2710 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2714 sdrain.cqe.done = ib_drain_qp_done;
2715 init_completion(&sdrain.done);
2717 ret = ib_post_send(qp, &swr.wr, NULL);
2719 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2723 if (cq->poll_ctx == IB_POLL_DIRECT)
2724 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2725 ib_process_cq_direct(cq, -1);
2727 wait_for_completion(&sdrain.done);
2731 * Post a WR and block until its completion is reaped for the RQ.
2733 static void __ib_drain_rq(struct ib_qp *qp)
2735 struct ib_cq *cq = qp->recv_cq;
2736 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2737 struct ib_drain_cqe rdrain;
2738 struct ib_recv_wr rwr = {};
2741 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2743 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2747 rwr.wr_cqe = &rdrain.cqe;
2748 rdrain.cqe.done = ib_drain_qp_done;
2749 init_completion(&rdrain.done);
2751 ret = ib_post_recv(qp, &rwr, NULL);
2753 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2757 if (cq->poll_ctx == IB_POLL_DIRECT)
2758 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2759 ib_process_cq_direct(cq, -1);
2761 wait_for_completion(&rdrain.done);
2765 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2767 * @qp: queue pair to drain
2769 * If the device has a provider-specific drain function, then
2770 * call that. Otherwise call the generic drain function
2775 * ensure there is room in the CQ and SQ for the drain work request and
2778 * allocate the CQ using ib_alloc_cq().
2780 * ensure that there are no other contexts that are posting WRs concurrently.
2781 * Otherwise the drain is not guaranteed.
2783 void ib_drain_sq(struct ib_qp *qp)
2785 if (qp->device->ops.drain_sq)
2786 qp->device->ops.drain_sq(qp);
2789 trace_cq_drain_complete(qp->send_cq);
2791 EXPORT_SYMBOL(ib_drain_sq);
2794 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2796 * @qp: queue pair to drain
2798 * If the device has a provider-specific drain function, then
2799 * call that. Otherwise call the generic drain function
2804 * ensure there is room in the CQ and RQ for the drain work request and
2807 * allocate the CQ using ib_alloc_cq().
2809 * ensure that there are no other contexts that are posting WRs concurrently.
2810 * Otherwise the drain is not guaranteed.
2812 void ib_drain_rq(struct ib_qp *qp)
2814 if (qp->device->ops.drain_rq)
2815 qp->device->ops.drain_rq(qp);
2818 trace_cq_drain_complete(qp->recv_cq);
2820 EXPORT_SYMBOL(ib_drain_rq);
2823 * ib_drain_qp() - Block until all CQEs have been consumed by the
2824 * application on both the RQ and SQ.
2825 * @qp: queue pair to drain
2829 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2832 * allocate the CQs using ib_alloc_cq().
2834 * ensure that there are no other contexts that are posting WRs concurrently.
2835 * Otherwise the drain is not guaranteed.
2837 void ib_drain_qp(struct ib_qp *qp)
2843 EXPORT_SYMBOL(ib_drain_qp);
2845 struct net_device *rdma_alloc_netdev(struct ib_device *device, u8 port_num,
2846 enum rdma_netdev_t type, const char *name,
2847 unsigned char name_assign_type,
2848 void (*setup)(struct net_device *))
2850 struct rdma_netdev_alloc_params params;
2851 struct net_device *netdev;
2854 if (!device->ops.rdma_netdev_get_params)
2855 return ERR_PTR(-EOPNOTSUPP);
2857 rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2862 netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type,
2863 setup, params.txqs, params.rxqs);
2865 return ERR_PTR(-ENOMEM);
2869 EXPORT_SYMBOL(rdma_alloc_netdev);
2871 int rdma_init_netdev(struct ib_device *device, u8 port_num,
2872 enum rdma_netdev_t type, const char *name,
2873 unsigned char name_assign_type,
2874 void (*setup)(struct net_device *),
2875 struct net_device *netdev)
2877 struct rdma_netdev_alloc_params params;
2880 if (!device->ops.rdma_netdev_get_params)
2883 rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2888 return params.initialize_rdma_netdev(device, port_num,
2889 netdev, params.param);
2891 EXPORT_SYMBOL(rdma_init_netdev);
2893 void __rdma_block_iter_start(struct ib_block_iter *biter,
2894 struct scatterlist *sglist, unsigned int nents,
2897 memset(biter, 0, sizeof(struct ib_block_iter));
2898 biter->__sg = sglist;
2899 biter->__sg_nents = nents;
2901 /* Driver provides best block size to use */
2902 biter->__pg_bit = __fls(pgsz);
2904 EXPORT_SYMBOL(__rdma_block_iter_start);
2906 bool __rdma_block_iter_next(struct ib_block_iter *biter)
2908 unsigned int block_offset;
2910 if (!biter->__sg_nents || !biter->__sg)
2913 biter->__dma_addr = sg_dma_address(biter->__sg) + biter->__sg_advance;
2914 block_offset = biter->__dma_addr & (BIT_ULL(biter->__pg_bit) - 1);
2915 biter->__sg_advance += BIT_ULL(biter->__pg_bit) - block_offset;
2917 if (biter->__sg_advance >= sg_dma_len(biter->__sg)) {
2918 biter->__sg_advance = 0;
2919 biter->__sg = sg_next(biter->__sg);
2920 biter->__sg_nents--;
2925 EXPORT_SYMBOL(__rdma_block_iter_next);