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"
56 static int ib_resolve_eth_dmac(struct ib_device *device,
57 struct rdma_ah_attr *ah_attr);
59 static const char * const ib_events[] = {
60 [IB_EVENT_CQ_ERR] = "CQ error",
61 [IB_EVENT_QP_FATAL] = "QP fatal error",
62 [IB_EVENT_QP_REQ_ERR] = "QP request error",
63 [IB_EVENT_QP_ACCESS_ERR] = "QP access error",
64 [IB_EVENT_COMM_EST] = "communication established",
65 [IB_EVENT_SQ_DRAINED] = "send queue drained",
66 [IB_EVENT_PATH_MIG] = "path migration successful",
67 [IB_EVENT_PATH_MIG_ERR] = "path migration error",
68 [IB_EVENT_DEVICE_FATAL] = "device fatal error",
69 [IB_EVENT_PORT_ACTIVE] = "port active",
70 [IB_EVENT_PORT_ERR] = "port error",
71 [IB_EVENT_LID_CHANGE] = "LID change",
72 [IB_EVENT_PKEY_CHANGE] = "P_key change",
73 [IB_EVENT_SM_CHANGE] = "SM change",
74 [IB_EVENT_SRQ_ERR] = "SRQ error",
75 [IB_EVENT_SRQ_LIMIT_REACHED] = "SRQ limit reached",
76 [IB_EVENT_QP_LAST_WQE_REACHED] = "last WQE reached",
77 [IB_EVENT_CLIENT_REREGISTER] = "client reregister",
78 [IB_EVENT_GID_CHANGE] = "GID changed",
81 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
85 return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
86 ib_events[index] : "unrecognized event";
88 EXPORT_SYMBOL(ib_event_msg);
90 static const char * const wc_statuses[] = {
91 [IB_WC_SUCCESS] = "success",
92 [IB_WC_LOC_LEN_ERR] = "local length error",
93 [IB_WC_LOC_QP_OP_ERR] = "local QP operation error",
94 [IB_WC_LOC_EEC_OP_ERR] = "local EE context operation error",
95 [IB_WC_LOC_PROT_ERR] = "local protection error",
96 [IB_WC_WR_FLUSH_ERR] = "WR flushed",
97 [IB_WC_MW_BIND_ERR] = "memory management operation error",
98 [IB_WC_BAD_RESP_ERR] = "bad response error",
99 [IB_WC_LOC_ACCESS_ERR] = "local access error",
100 [IB_WC_REM_INV_REQ_ERR] = "invalid request error",
101 [IB_WC_REM_ACCESS_ERR] = "remote access error",
102 [IB_WC_REM_OP_ERR] = "remote operation error",
103 [IB_WC_RETRY_EXC_ERR] = "transport retry counter exceeded",
104 [IB_WC_RNR_RETRY_EXC_ERR] = "RNR retry counter exceeded",
105 [IB_WC_LOC_RDD_VIOL_ERR] = "local RDD violation error",
106 [IB_WC_REM_INV_RD_REQ_ERR] = "remote invalid RD request",
107 [IB_WC_REM_ABORT_ERR] = "operation aborted",
108 [IB_WC_INV_EECN_ERR] = "invalid EE context number",
109 [IB_WC_INV_EEC_STATE_ERR] = "invalid EE context state",
110 [IB_WC_FATAL_ERR] = "fatal error",
111 [IB_WC_RESP_TIMEOUT_ERR] = "response timeout error",
112 [IB_WC_GENERAL_ERR] = "general error",
115 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
117 size_t index = status;
119 return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
120 wc_statuses[index] : "unrecognized status";
122 EXPORT_SYMBOL(ib_wc_status_msg);
124 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
127 case IB_RATE_2_5_GBPS: return 1;
128 case IB_RATE_5_GBPS: return 2;
129 case IB_RATE_10_GBPS: return 4;
130 case IB_RATE_20_GBPS: return 8;
131 case IB_RATE_30_GBPS: return 12;
132 case IB_RATE_40_GBPS: return 16;
133 case IB_RATE_60_GBPS: return 24;
134 case IB_RATE_80_GBPS: return 32;
135 case IB_RATE_120_GBPS: return 48;
139 EXPORT_SYMBOL(ib_rate_to_mult);
141 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
144 case 1: return IB_RATE_2_5_GBPS;
145 case 2: return IB_RATE_5_GBPS;
146 case 4: return IB_RATE_10_GBPS;
147 case 8: return IB_RATE_20_GBPS;
148 case 12: return IB_RATE_30_GBPS;
149 case 16: return IB_RATE_40_GBPS;
150 case 24: return IB_RATE_60_GBPS;
151 case 32: return IB_RATE_80_GBPS;
152 case 48: return IB_RATE_120_GBPS;
153 default: return IB_RATE_PORT_CURRENT;
156 EXPORT_SYMBOL(mult_to_ib_rate);
158 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
161 case IB_RATE_2_5_GBPS: return 2500;
162 case IB_RATE_5_GBPS: return 5000;
163 case IB_RATE_10_GBPS: return 10000;
164 case IB_RATE_20_GBPS: return 20000;
165 case IB_RATE_30_GBPS: return 30000;
166 case IB_RATE_40_GBPS: return 40000;
167 case IB_RATE_60_GBPS: return 60000;
168 case IB_RATE_80_GBPS: return 80000;
169 case IB_RATE_120_GBPS: return 120000;
170 case IB_RATE_14_GBPS: return 14062;
171 case IB_RATE_56_GBPS: return 56250;
172 case IB_RATE_112_GBPS: return 112500;
173 case IB_RATE_168_GBPS: return 168750;
174 case IB_RATE_25_GBPS: return 25781;
175 case IB_RATE_100_GBPS: return 103125;
176 case IB_RATE_200_GBPS: return 206250;
177 case IB_RATE_300_GBPS: return 309375;
181 EXPORT_SYMBOL(ib_rate_to_mbps);
183 __attribute_const__ enum rdma_transport_type
184 rdma_node_get_transport(enum rdma_node_type node_type)
187 if (node_type == RDMA_NODE_USNIC)
188 return RDMA_TRANSPORT_USNIC;
189 if (node_type == RDMA_NODE_USNIC_UDP)
190 return RDMA_TRANSPORT_USNIC_UDP;
191 if (node_type == RDMA_NODE_RNIC)
192 return RDMA_TRANSPORT_IWARP;
194 return RDMA_TRANSPORT_IB;
196 EXPORT_SYMBOL(rdma_node_get_transport);
198 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
200 enum rdma_transport_type lt;
201 if (device->get_link_layer)
202 return device->get_link_layer(device, port_num);
204 lt = rdma_node_get_transport(device->node_type);
205 if (lt == RDMA_TRANSPORT_IB)
206 return IB_LINK_LAYER_INFINIBAND;
208 return IB_LINK_LAYER_ETHERNET;
210 EXPORT_SYMBOL(rdma_port_get_link_layer);
212 /* Protection domains */
215 * ib_alloc_pd - Allocates an unused protection domain.
216 * @device: The device on which to allocate the protection domain.
218 * A protection domain object provides an association between QPs, shared
219 * receive queues, address handles, memory regions, and memory windows.
221 * Every PD has a local_dma_lkey which can be used as the lkey value for local
224 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
228 int mr_access_flags = 0;
230 pd = device->alloc_pd(device, NULL, NULL);
236 pd->__internal_mr = NULL;
237 atomic_set(&pd->usecnt, 0);
240 if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
241 pd->local_dma_lkey = device->local_dma_lkey;
243 mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
245 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
246 pr_warn("%s: enabling unsafe global rkey\n", caller);
247 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
250 if (mr_access_flags) {
253 mr = pd->device->get_dma_mr(pd, mr_access_flags);
259 mr->device = pd->device;
262 mr->need_inval = false;
264 pd->__internal_mr = mr;
266 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
267 pd->local_dma_lkey = pd->__internal_mr->lkey;
269 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
270 pd->unsafe_global_rkey = pd->__internal_mr->rkey;
275 EXPORT_SYMBOL(__ib_alloc_pd);
278 * ib_dealloc_pd - Deallocates a protection domain.
279 * @pd: The protection domain to deallocate.
281 * It is an error to call this function while any resources in the pd still
282 * exist. The caller is responsible to synchronously destroy them and
283 * guarantee no new allocations will happen.
285 void ib_dealloc_pd(struct ib_pd *pd)
289 if (pd->__internal_mr) {
290 ret = pd->device->dereg_mr(pd->__internal_mr);
292 pd->__internal_mr = NULL;
295 /* uverbs manipulates usecnt with proper locking, while the kabi
296 requires the caller to guarantee we can't race here. */
297 WARN_ON(atomic_read(&pd->usecnt));
299 /* Making delalloc_pd a void return is a WIP, no driver should return
301 ret = pd->device->dealloc_pd(pd);
302 WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd");
304 EXPORT_SYMBOL(ib_dealloc_pd);
306 /* Address handles */
308 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
309 struct rdma_ah_attr *ah_attr,
310 struct ib_udata *udata)
314 ah = pd->device->create_ah(pd, ah_attr, udata);
317 ah->device = pd->device;
320 ah->type = ah_attr->type;
321 atomic_inc(&pd->usecnt);
327 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr)
329 return _rdma_create_ah(pd, ah_attr, NULL);
331 EXPORT_SYMBOL(rdma_create_ah);
334 * rdma_create_user_ah - Creates an address handle for the
335 * given address vector.
336 * It resolves destination mac address for ah attribute of RoCE type.
337 * @pd: The protection domain associated with the address handle.
338 * @ah_attr: The attributes of the address vector.
339 * @udata: pointer to user's input output buffer information need by
342 * It returns 0 on success and returns appropriate error code on error.
343 * The address handle is used to reference a local or global destination
344 * in all UD QP post sends.
346 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
347 struct rdma_ah_attr *ah_attr,
348 struct ib_udata *udata)
352 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
353 err = ib_resolve_eth_dmac(pd->device, ah_attr);
358 return _rdma_create_ah(pd, ah_attr, udata);
360 EXPORT_SYMBOL(rdma_create_user_ah);
362 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
364 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
365 struct iphdr ip4h_checked;
366 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
368 /* If it's IPv6, the version must be 6, otherwise, the first
369 * 20 bytes (before the IPv4 header) are garbled.
371 if (ip6h->version != 6)
372 return (ip4h->version == 4) ? 4 : 0;
373 /* version may be 6 or 4 because the first 20 bytes could be garbled */
375 /* RoCE v2 requires no options, thus header length
382 * We can't write on scattered buffers so we need to copy to
385 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
386 ip4h_checked.check = 0;
387 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
388 /* if IPv4 header checksum is OK, believe it */
389 if (ip4h->check == ip4h_checked.check)
393 EXPORT_SYMBOL(ib_get_rdma_header_version);
395 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
397 const struct ib_grh *grh)
401 if (rdma_protocol_ib(device, port_num))
402 return RDMA_NETWORK_IB;
404 grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
406 if (grh_version == 4)
407 return RDMA_NETWORK_IPV4;
409 if (grh->next_hdr == IPPROTO_UDP)
410 return RDMA_NETWORK_IPV6;
412 return RDMA_NETWORK_ROCE_V1;
415 struct find_gid_index_context {
417 enum ib_gid_type gid_type;
420 static bool find_gid_index(const union ib_gid *gid,
421 const struct ib_gid_attr *gid_attr,
424 struct find_gid_index_context *ctx =
425 (struct find_gid_index_context *)context;
427 if (ctx->gid_type != gid_attr->gid_type)
430 if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) ||
431 (is_vlan_dev(gid_attr->ndev) &&
432 vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id))
438 static int get_sgid_index_from_eth(struct ib_device *device, u8 port_num,
439 u16 vlan_id, const union ib_gid *sgid,
440 enum ib_gid_type gid_type,
443 struct find_gid_index_context context = {.vlan_id = vlan_id,
444 .gid_type = gid_type};
446 return ib_find_gid_by_filter(device, sgid, port_num, find_gid_index,
447 &context, gid_index);
450 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
451 enum rdma_network_type net_type,
452 union ib_gid *sgid, union ib_gid *dgid)
454 struct sockaddr_in src_in;
455 struct sockaddr_in dst_in;
456 __be32 src_saddr, dst_saddr;
461 if (net_type == RDMA_NETWORK_IPV4) {
462 memcpy(&src_in.sin_addr.s_addr,
463 &hdr->roce4grh.saddr, 4);
464 memcpy(&dst_in.sin_addr.s_addr,
465 &hdr->roce4grh.daddr, 4);
466 src_saddr = src_in.sin_addr.s_addr;
467 dst_saddr = dst_in.sin_addr.s_addr;
468 ipv6_addr_set_v4mapped(src_saddr,
469 (struct in6_addr *)sgid);
470 ipv6_addr_set_v4mapped(dst_saddr,
471 (struct in6_addr *)dgid);
473 } else if (net_type == RDMA_NETWORK_IPV6 ||
474 net_type == RDMA_NETWORK_IB) {
475 *dgid = hdr->ibgrh.dgid;
476 *sgid = hdr->ibgrh.sgid;
482 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
485 * This function creates ah from the incoming packet.
486 * Incoming packet has dgid of the receiver node on which this code is
487 * getting executed and, sgid contains the GID of the sender.
489 * When resolving mac address of destination, the arrived dgid is used
490 * as sgid and, sgid is used as dgid because sgid contains destinations
491 * GID whom to respond to.
493 * This is why when calling rdma_addr_find_l2_eth_by_grh() function, the
494 * position of arguments dgid and sgid do not match the order of the
497 int ib_init_ah_from_wc(struct ib_device *device, u8 port_num,
498 const struct ib_wc *wc, const struct ib_grh *grh,
499 struct rdma_ah_attr *ah_attr)
504 enum rdma_network_type net_type = RDMA_NETWORK_IB;
505 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
512 memset(ah_attr, 0, sizeof *ah_attr);
513 ah_attr->type = rdma_ah_find_type(device, port_num);
514 if (rdma_cap_eth_ah(device, port_num)) {
515 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
516 net_type = wc->network_hdr_type;
518 net_type = ib_get_net_type_by_grh(device, port_num, grh);
519 gid_type = ib_network_to_gid_type(net_type);
521 ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
526 if (rdma_protocol_roce(device, port_num)) {
528 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
529 wc->vlan_id : 0xffff;
530 struct net_device *idev;
531 struct net_device *resolved_dev;
533 if (!(wc->wc_flags & IB_WC_GRH))
536 if (!device->get_netdev)
539 idev = device->get_netdev(device, port_num);
543 ret = rdma_addr_find_l2_eth_by_grh(&dgid, &sgid,
545 wc->wc_flags & IB_WC_WITH_VLAN ?
547 &if_index, &hoplimit);
553 resolved_dev = dev_get_by_index(&init_net, if_index);
555 if (resolved_dev != idev && !rdma_is_upper_dev_rcu(idev,
560 dev_put(resolved_dev);
564 ret = get_sgid_index_from_eth(device, port_num, vlan_id,
565 &dgid, gid_type, &gid_index);
570 rdma_ah_set_dlid(ah_attr, wc->slid);
571 rdma_ah_set_sl(ah_attr, wc->sl);
572 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
573 rdma_ah_set_port_num(ah_attr, port_num);
575 if (wc->wc_flags & IB_WC_GRH) {
576 if (!rdma_cap_eth_ah(device, port_num)) {
577 if (dgid.global.interface_id != cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
578 ret = ib_find_cached_gid_by_port(device, &dgid,
589 flow_class = be32_to_cpu(grh->version_tclass_flow);
590 rdma_ah_set_grh(ah_attr, &sgid,
591 flow_class & 0xFFFFF,
592 (u8)gid_index, hoplimit,
593 (flow_class >> 20) & 0xFF);
598 EXPORT_SYMBOL(ib_init_ah_from_wc);
600 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
601 const struct ib_grh *grh, u8 port_num)
603 struct rdma_ah_attr ah_attr;
606 ret = ib_init_ah_from_wc(pd->device, port_num, wc, grh, &ah_attr);
610 return rdma_create_ah(pd, &ah_attr);
612 EXPORT_SYMBOL(ib_create_ah_from_wc);
614 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
616 if (ah->type != ah_attr->type)
619 return ah->device->modify_ah ?
620 ah->device->modify_ah(ah, ah_attr) :
623 EXPORT_SYMBOL(rdma_modify_ah);
625 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
627 return ah->device->query_ah ?
628 ah->device->query_ah(ah, ah_attr) :
631 EXPORT_SYMBOL(rdma_query_ah);
633 int rdma_destroy_ah(struct ib_ah *ah)
639 ret = ah->device->destroy_ah(ah);
641 atomic_dec(&pd->usecnt);
645 EXPORT_SYMBOL(rdma_destroy_ah);
647 /* Shared receive queues */
649 struct ib_srq *ib_create_srq(struct ib_pd *pd,
650 struct ib_srq_init_attr *srq_init_attr)
654 if (!pd->device->create_srq)
655 return ERR_PTR(-ENOSYS);
657 srq = pd->device->create_srq(pd, srq_init_attr, NULL);
660 srq->device = pd->device;
663 srq->event_handler = srq_init_attr->event_handler;
664 srq->srq_context = srq_init_attr->srq_context;
665 srq->srq_type = srq_init_attr->srq_type;
666 if (ib_srq_has_cq(srq->srq_type)) {
667 srq->ext.cq = srq_init_attr->ext.cq;
668 atomic_inc(&srq->ext.cq->usecnt);
670 if (srq->srq_type == IB_SRQT_XRC) {
671 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
672 atomic_inc(&srq->ext.xrc.xrcd->usecnt);
674 atomic_inc(&pd->usecnt);
675 atomic_set(&srq->usecnt, 0);
680 EXPORT_SYMBOL(ib_create_srq);
682 int ib_modify_srq(struct ib_srq *srq,
683 struct ib_srq_attr *srq_attr,
684 enum ib_srq_attr_mask srq_attr_mask)
686 return srq->device->modify_srq ?
687 srq->device->modify_srq(srq, srq_attr, srq_attr_mask, NULL) :
690 EXPORT_SYMBOL(ib_modify_srq);
692 int ib_query_srq(struct ib_srq *srq,
693 struct ib_srq_attr *srq_attr)
695 return srq->device->query_srq ?
696 srq->device->query_srq(srq, srq_attr) : -ENOSYS;
698 EXPORT_SYMBOL(ib_query_srq);
700 int ib_destroy_srq(struct ib_srq *srq)
703 enum ib_srq_type srq_type;
704 struct ib_xrcd *uninitialized_var(xrcd);
705 struct ib_cq *uninitialized_var(cq);
708 if (atomic_read(&srq->usecnt))
712 srq_type = srq->srq_type;
713 if (ib_srq_has_cq(srq_type))
715 if (srq_type == IB_SRQT_XRC)
716 xrcd = srq->ext.xrc.xrcd;
718 ret = srq->device->destroy_srq(srq);
720 atomic_dec(&pd->usecnt);
721 if (srq_type == IB_SRQT_XRC)
722 atomic_dec(&xrcd->usecnt);
723 if (ib_srq_has_cq(srq_type))
724 atomic_dec(&cq->usecnt);
729 EXPORT_SYMBOL(ib_destroy_srq);
733 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
735 struct ib_qp *qp = context;
738 spin_lock_irqsave(&qp->device->event_handler_lock, flags);
739 list_for_each_entry(event->element.qp, &qp->open_list, open_list)
740 if (event->element.qp->event_handler)
741 event->element.qp->event_handler(event, event->element.qp->qp_context);
742 spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
745 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
747 mutex_lock(&xrcd->tgt_qp_mutex);
748 list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
749 mutex_unlock(&xrcd->tgt_qp_mutex);
752 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
753 void (*event_handler)(struct ib_event *, void *),
760 qp = kzalloc(sizeof *qp, GFP_KERNEL);
762 return ERR_PTR(-ENOMEM);
764 qp->real_qp = real_qp;
765 err = ib_open_shared_qp_security(qp, real_qp->device);
771 qp->real_qp = real_qp;
772 atomic_inc(&real_qp->usecnt);
773 qp->device = real_qp->device;
774 qp->event_handler = event_handler;
775 qp->qp_context = qp_context;
776 qp->qp_num = real_qp->qp_num;
777 qp->qp_type = real_qp->qp_type;
779 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
780 list_add(&qp->open_list, &real_qp->open_list);
781 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
786 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
787 struct ib_qp_open_attr *qp_open_attr)
789 struct ib_qp *qp, *real_qp;
791 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
792 return ERR_PTR(-EINVAL);
794 qp = ERR_PTR(-EINVAL);
795 mutex_lock(&xrcd->tgt_qp_mutex);
796 list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
797 if (real_qp->qp_num == qp_open_attr->qp_num) {
798 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
799 qp_open_attr->qp_context);
803 mutex_unlock(&xrcd->tgt_qp_mutex);
806 EXPORT_SYMBOL(ib_open_qp);
808 static struct ib_qp *ib_create_xrc_qp(struct ib_qp *qp,
809 struct ib_qp_init_attr *qp_init_attr)
811 struct ib_qp *real_qp = qp;
813 qp->event_handler = __ib_shared_qp_event_handler;
816 qp->send_cq = qp->recv_cq = NULL;
818 qp->xrcd = qp_init_attr->xrcd;
819 atomic_inc(&qp_init_attr->xrcd->usecnt);
820 INIT_LIST_HEAD(&qp->open_list);
822 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
823 qp_init_attr->qp_context);
825 __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
827 real_qp->device->destroy_qp(real_qp);
831 struct ib_qp *ib_create_qp(struct ib_pd *pd,
832 struct ib_qp_init_attr *qp_init_attr)
834 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
838 if (qp_init_attr->rwq_ind_tbl &&
839 (qp_init_attr->recv_cq ||
840 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
841 qp_init_attr->cap.max_recv_sge))
842 return ERR_PTR(-EINVAL);
845 * If the callers is using the RDMA API calculate the resources
846 * needed for the RDMA READ/WRITE operations.
848 * Note that these callers need to pass in a port number.
850 if (qp_init_attr->cap.max_rdma_ctxs)
851 rdma_rw_init_qp(device, qp_init_attr);
853 qp = device->create_qp(pd, qp_init_attr, NULL);
857 ret = ib_create_qp_security(qp, device);
866 qp->qp_type = qp_init_attr->qp_type;
867 qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
869 atomic_set(&qp->usecnt, 0);
871 spin_lock_init(&qp->mr_lock);
872 INIT_LIST_HEAD(&qp->rdma_mrs);
873 INIT_LIST_HEAD(&qp->sig_mrs);
876 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT)
877 return ib_create_xrc_qp(qp, qp_init_attr);
879 qp->event_handler = qp_init_attr->event_handler;
880 qp->qp_context = qp_init_attr->qp_context;
881 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
885 qp->recv_cq = qp_init_attr->recv_cq;
886 if (qp_init_attr->recv_cq)
887 atomic_inc(&qp_init_attr->recv_cq->usecnt);
888 qp->srq = qp_init_attr->srq;
890 atomic_inc(&qp_init_attr->srq->usecnt);
894 qp->send_cq = qp_init_attr->send_cq;
897 atomic_inc(&pd->usecnt);
898 if (qp_init_attr->send_cq)
899 atomic_inc(&qp_init_attr->send_cq->usecnt);
900 if (qp_init_attr->rwq_ind_tbl)
901 atomic_inc(&qp->rwq_ind_tbl->usecnt);
903 if (qp_init_attr->cap.max_rdma_ctxs) {
904 ret = rdma_rw_init_mrs(qp, qp_init_attr);
906 pr_err("failed to init MR pool ret= %d\n", ret);
913 * Note: all hw drivers guarantee that max_send_sge is lower than
914 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
915 * max_send_sge <= max_sge_rd.
917 qp->max_write_sge = qp_init_attr->cap.max_send_sge;
918 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
919 device->attrs.max_sge_rd);
923 EXPORT_SYMBOL(ib_create_qp);
925 static const struct {
927 enum ib_qp_attr_mask req_param[IB_QPT_MAX];
928 enum ib_qp_attr_mask opt_param[IB_QPT_MAX];
929 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
931 [IB_QPS_RESET] = { .valid = 1 },
935 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
938 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
939 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
942 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
945 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
948 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
951 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
953 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
959 [IB_QPS_RESET] = { .valid = 1 },
960 [IB_QPS_ERR] = { .valid = 1 },
964 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
967 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
970 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
973 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
976 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
979 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
981 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
988 [IB_QPT_UC] = (IB_QP_AV |
992 [IB_QPT_RC] = (IB_QP_AV |
996 IB_QP_MAX_DEST_RD_ATOMIC |
997 IB_QP_MIN_RNR_TIMER),
998 [IB_QPT_XRC_INI] = (IB_QP_AV |
1002 [IB_QPT_XRC_TGT] = (IB_QP_AV |
1006 IB_QP_MAX_DEST_RD_ATOMIC |
1007 IB_QP_MIN_RNR_TIMER),
1010 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1012 [IB_QPT_UC] = (IB_QP_ALT_PATH |
1013 IB_QP_ACCESS_FLAGS |
1015 [IB_QPT_RC] = (IB_QP_ALT_PATH |
1016 IB_QP_ACCESS_FLAGS |
1018 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH |
1019 IB_QP_ACCESS_FLAGS |
1021 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH |
1022 IB_QP_ACCESS_FLAGS |
1024 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1026 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1032 [IB_QPS_RESET] = { .valid = 1 },
1033 [IB_QPS_ERR] = { .valid = 1 },
1037 [IB_QPT_UD] = IB_QP_SQ_PSN,
1038 [IB_QPT_UC] = IB_QP_SQ_PSN,
1039 [IB_QPT_RC] = (IB_QP_TIMEOUT |
1043 IB_QP_MAX_QP_RD_ATOMIC),
1044 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT |
1048 IB_QP_MAX_QP_RD_ATOMIC),
1049 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT |
1051 [IB_QPT_SMI] = IB_QP_SQ_PSN,
1052 [IB_QPT_GSI] = IB_QP_SQ_PSN,
1055 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1057 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1059 IB_QP_ACCESS_FLAGS |
1060 IB_QP_PATH_MIG_STATE),
1061 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1063 IB_QP_ACCESS_FLAGS |
1064 IB_QP_MIN_RNR_TIMER |
1065 IB_QP_PATH_MIG_STATE),
1066 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1068 IB_QP_ACCESS_FLAGS |
1069 IB_QP_PATH_MIG_STATE),
1070 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1072 IB_QP_ACCESS_FLAGS |
1073 IB_QP_MIN_RNR_TIMER |
1074 IB_QP_PATH_MIG_STATE),
1075 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1077 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1079 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1084 [IB_QPS_RESET] = { .valid = 1 },
1085 [IB_QPS_ERR] = { .valid = 1 },
1089 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1091 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1092 IB_QP_ACCESS_FLAGS |
1094 IB_QP_PATH_MIG_STATE),
1095 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1096 IB_QP_ACCESS_FLAGS |
1098 IB_QP_PATH_MIG_STATE |
1099 IB_QP_MIN_RNR_TIMER),
1100 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1101 IB_QP_ACCESS_FLAGS |
1103 IB_QP_PATH_MIG_STATE),
1104 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1105 IB_QP_ACCESS_FLAGS |
1107 IB_QP_PATH_MIG_STATE |
1108 IB_QP_MIN_RNR_TIMER),
1109 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1111 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1113 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1119 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1120 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1121 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1122 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1123 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1124 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1125 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1130 [IB_QPS_RESET] = { .valid = 1 },
1131 [IB_QPS_ERR] = { .valid = 1 },
1135 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1137 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1139 IB_QP_ACCESS_FLAGS |
1140 IB_QP_PATH_MIG_STATE),
1141 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1143 IB_QP_ACCESS_FLAGS |
1144 IB_QP_MIN_RNR_TIMER |
1145 IB_QP_PATH_MIG_STATE),
1146 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1148 IB_QP_ACCESS_FLAGS |
1149 IB_QP_PATH_MIG_STATE),
1150 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1152 IB_QP_ACCESS_FLAGS |
1153 IB_QP_MIN_RNR_TIMER |
1154 IB_QP_PATH_MIG_STATE),
1155 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1157 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1164 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1166 [IB_QPT_UC] = (IB_QP_AV |
1168 IB_QP_ACCESS_FLAGS |
1170 IB_QP_PATH_MIG_STATE),
1171 [IB_QPT_RC] = (IB_QP_PORT |
1176 IB_QP_MAX_QP_RD_ATOMIC |
1177 IB_QP_MAX_DEST_RD_ATOMIC |
1179 IB_QP_ACCESS_FLAGS |
1181 IB_QP_MIN_RNR_TIMER |
1182 IB_QP_PATH_MIG_STATE),
1183 [IB_QPT_XRC_INI] = (IB_QP_PORT |
1188 IB_QP_MAX_QP_RD_ATOMIC |
1190 IB_QP_ACCESS_FLAGS |
1192 IB_QP_PATH_MIG_STATE),
1193 [IB_QPT_XRC_TGT] = (IB_QP_PORT |
1196 IB_QP_MAX_DEST_RD_ATOMIC |
1198 IB_QP_ACCESS_FLAGS |
1200 IB_QP_MIN_RNR_TIMER |
1201 IB_QP_PATH_MIG_STATE),
1202 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1204 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1210 [IB_QPS_RESET] = { .valid = 1 },
1211 [IB_QPS_ERR] = { .valid = 1 },
1215 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1217 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1218 IB_QP_ACCESS_FLAGS),
1219 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1221 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1227 [IB_QPS_RESET] = { .valid = 1 },
1228 [IB_QPS_ERR] = { .valid = 1 }
1232 int ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1233 enum ib_qp_type type, enum ib_qp_attr_mask mask,
1234 enum rdma_link_layer ll)
1236 enum ib_qp_attr_mask req_param, opt_param;
1238 if (cur_state < 0 || cur_state > IB_QPS_ERR ||
1239 next_state < 0 || next_state > IB_QPS_ERR)
1242 if (mask & IB_QP_CUR_STATE &&
1243 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1244 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1247 if (!qp_state_table[cur_state][next_state].valid)
1250 req_param = qp_state_table[cur_state][next_state].req_param[type];
1251 opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1253 if ((mask & req_param) != req_param)
1256 if (mask & ~(req_param | opt_param | IB_QP_STATE))
1261 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1263 static int ib_resolve_eth_dmac(struct ib_device *device,
1264 struct rdma_ah_attr *ah_attr)
1267 struct ib_global_route *grh;
1269 if (!rdma_is_port_valid(device, rdma_ah_get_port_num(ah_attr)))
1272 if (ah_attr->type != RDMA_AH_ATTR_TYPE_ROCE)
1275 grh = rdma_ah_retrieve_grh(ah_attr);
1277 if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw)) {
1278 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
1279 ah_attr->roce.dmac);
1282 if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1283 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1286 memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1287 ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1289 ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1290 (char *)ah_attr->roce.dmac);
1294 struct ib_gid_attr sgid_attr;
1298 ret = ib_query_gid(device,
1299 rdma_ah_get_port_num(ah_attr),
1303 if (ret || !sgid_attr.ndev) {
1309 ifindex = sgid_attr.ndev->ifindex;
1312 rdma_addr_find_l2_eth_by_grh(&sgid, &grh->dgid,
1314 NULL, &ifindex, &hop_limit);
1316 dev_put(sgid_attr.ndev);
1318 grh->hop_limit = hop_limit;
1325 * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1326 * @qp: The QP to modify.
1327 * @attr: On input, specifies the QP attributes to modify. On output,
1328 * the current values of selected QP attributes are returned.
1329 * @attr_mask: A bit-mask used to specify which attributes of the QP
1330 * are being modified.
1331 * @udata: pointer to user's input output buffer information
1332 * are being modified.
1333 * It returns 0 on success and returns appropriate error code on error.
1335 int ib_modify_qp_with_udata(struct ib_qp *qp, struct ib_qp_attr *attr,
1336 int attr_mask, struct ib_udata *udata)
1340 if (attr_mask & IB_QP_AV) {
1341 ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
1345 ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1346 if (!ret && (attr_mask & IB_QP_PORT))
1347 qp->port = attr->port_num;
1351 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1353 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u8 *speed, u8 *width)
1357 struct net_device *netdev;
1358 struct ethtool_link_ksettings lksettings;
1360 if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1363 if (!dev->get_netdev)
1366 netdev = dev->get_netdev(dev, port_num);
1371 rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1377 netdev_speed = lksettings.base.speed;
1379 netdev_speed = SPEED_1000;
1380 pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name,
1384 if (netdev_speed <= SPEED_1000) {
1385 *width = IB_WIDTH_1X;
1386 *speed = IB_SPEED_SDR;
1387 } else if (netdev_speed <= SPEED_10000) {
1388 *width = IB_WIDTH_1X;
1389 *speed = IB_SPEED_FDR10;
1390 } else if (netdev_speed <= SPEED_20000) {
1391 *width = IB_WIDTH_4X;
1392 *speed = IB_SPEED_DDR;
1393 } else if (netdev_speed <= SPEED_25000) {
1394 *width = IB_WIDTH_1X;
1395 *speed = IB_SPEED_EDR;
1396 } else if (netdev_speed <= SPEED_40000) {
1397 *width = IB_WIDTH_4X;
1398 *speed = IB_SPEED_FDR10;
1400 *width = IB_WIDTH_4X;
1401 *speed = IB_SPEED_EDR;
1406 EXPORT_SYMBOL(ib_get_eth_speed);
1408 int ib_modify_qp(struct ib_qp *qp,
1409 struct ib_qp_attr *qp_attr,
1412 return ib_modify_qp_with_udata(qp, qp_attr, qp_attr_mask, NULL);
1414 EXPORT_SYMBOL(ib_modify_qp);
1416 int ib_query_qp(struct ib_qp *qp,
1417 struct ib_qp_attr *qp_attr,
1419 struct ib_qp_init_attr *qp_init_attr)
1421 return qp->device->query_qp ?
1422 qp->device->query_qp(qp->real_qp, qp_attr, qp_attr_mask, qp_init_attr) :
1425 EXPORT_SYMBOL(ib_query_qp);
1427 int ib_close_qp(struct ib_qp *qp)
1429 struct ib_qp *real_qp;
1430 unsigned long flags;
1432 real_qp = qp->real_qp;
1436 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1437 list_del(&qp->open_list);
1438 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1440 atomic_dec(&real_qp->usecnt);
1441 ib_close_shared_qp_security(qp->qp_sec);
1446 EXPORT_SYMBOL(ib_close_qp);
1448 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1450 struct ib_xrcd *xrcd;
1451 struct ib_qp *real_qp;
1454 real_qp = qp->real_qp;
1455 xrcd = real_qp->xrcd;
1457 mutex_lock(&xrcd->tgt_qp_mutex);
1459 if (atomic_read(&real_qp->usecnt) == 0)
1460 list_del(&real_qp->xrcd_list);
1463 mutex_unlock(&xrcd->tgt_qp_mutex);
1466 ret = ib_destroy_qp(real_qp);
1468 atomic_dec(&xrcd->usecnt);
1470 __ib_insert_xrcd_qp(xrcd, real_qp);
1476 int ib_destroy_qp(struct ib_qp *qp)
1479 struct ib_cq *scq, *rcq;
1481 struct ib_rwq_ind_table *ind_tbl;
1482 struct ib_qp_security *sec;
1485 WARN_ON_ONCE(qp->mrs_used > 0);
1487 if (atomic_read(&qp->usecnt))
1490 if (qp->real_qp != qp)
1491 return __ib_destroy_shared_qp(qp);
1497 ind_tbl = qp->rwq_ind_tbl;
1500 ib_destroy_qp_security_begin(sec);
1503 rdma_rw_cleanup_mrs(qp);
1505 ret = qp->device->destroy_qp(qp);
1508 atomic_dec(&pd->usecnt);
1510 atomic_dec(&scq->usecnt);
1512 atomic_dec(&rcq->usecnt);
1514 atomic_dec(&srq->usecnt);
1516 atomic_dec(&ind_tbl->usecnt);
1518 ib_destroy_qp_security_end(sec);
1521 ib_destroy_qp_security_abort(sec);
1526 EXPORT_SYMBOL(ib_destroy_qp);
1528 /* Completion queues */
1530 struct ib_cq *ib_create_cq(struct ib_device *device,
1531 ib_comp_handler comp_handler,
1532 void (*event_handler)(struct ib_event *, void *),
1534 const struct ib_cq_init_attr *cq_attr)
1538 cq = device->create_cq(device, cq_attr, NULL, NULL);
1541 cq->device = device;
1543 cq->comp_handler = comp_handler;
1544 cq->event_handler = event_handler;
1545 cq->cq_context = cq_context;
1546 atomic_set(&cq->usecnt, 0);
1551 EXPORT_SYMBOL(ib_create_cq);
1553 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1555 return cq->device->modify_cq ?
1556 cq->device->modify_cq(cq, cq_count, cq_period) : -ENOSYS;
1558 EXPORT_SYMBOL(rdma_set_cq_moderation);
1560 int ib_destroy_cq(struct ib_cq *cq)
1562 if (atomic_read(&cq->usecnt))
1565 return cq->device->destroy_cq(cq);
1567 EXPORT_SYMBOL(ib_destroy_cq);
1569 int ib_resize_cq(struct ib_cq *cq, int cqe)
1571 return cq->device->resize_cq ?
1572 cq->device->resize_cq(cq, cqe, NULL) : -ENOSYS;
1574 EXPORT_SYMBOL(ib_resize_cq);
1576 /* Memory regions */
1578 int ib_dereg_mr(struct ib_mr *mr)
1580 struct ib_pd *pd = mr->pd;
1583 ret = mr->device->dereg_mr(mr);
1585 atomic_dec(&pd->usecnt);
1589 EXPORT_SYMBOL(ib_dereg_mr);
1592 * ib_alloc_mr() - Allocates a memory region
1593 * @pd: protection domain associated with the region
1594 * @mr_type: memory region type
1595 * @max_num_sg: maximum sg entries available for registration.
1598 * Memory registeration page/sg lists must not exceed max_num_sg.
1599 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
1600 * max_num_sg * used_page_size.
1603 struct ib_mr *ib_alloc_mr(struct ib_pd *pd,
1604 enum ib_mr_type mr_type,
1609 if (!pd->device->alloc_mr)
1610 return ERR_PTR(-ENOSYS);
1612 mr = pd->device->alloc_mr(pd, mr_type, max_num_sg);
1614 mr->device = pd->device;
1617 atomic_inc(&pd->usecnt);
1618 mr->need_inval = false;
1623 EXPORT_SYMBOL(ib_alloc_mr);
1625 /* "Fast" memory regions */
1627 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
1628 int mr_access_flags,
1629 struct ib_fmr_attr *fmr_attr)
1633 if (!pd->device->alloc_fmr)
1634 return ERR_PTR(-ENOSYS);
1636 fmr = pd->device->alloc_fmr(pd, mr_access_flags, fmr_attr);
1638 fmr->device = pd->device;
1640 atomic_inc(&pd->usecnt);
1645 EXPORT_SYMBOL(ib_alloc_fmr);
1647 int ib_unmap_fmr(struct list_head *fmr_list)
1651 if (list_empty(fmr_list))
1654 fmr = list_entry(fmr_list->next, struct ib_fmr, list);
1655 return fmr->device->unmap_fmr(fmr_list);
1657 EXPORT_SYMBOL(ib_unmap_fmr);
1659 int ib_dealloc_fmr(struct ib_fmr *fmr)
1665 ret = fmr->device->dealloc_fmr(fmr);
1667 atomic_dec(&pd->usecnt);
1671 EXPORT_SYMBOL(ib_dealloc_fmr);
1673 /* Multicast groups */
1675 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
1677 struct ib_qp_init_attr init_attr = {};
1678 struct ib_qp_attr attr = {};
1679 int num_eth_ports = 0;
1682 /* If QP state >= init, it is assigned to a port and we can check this
1685 if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
1686 if (attr.qp_state >= IB_QPS_INIT) {
1687 if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
1688 IB_LINK_LAYER_INFINIBAND)
1694 /* Can't get a quick answer, iterate over all ports */
1695 for (port = 0; port < qp->device->phys_port_cnt; port++)
1696 if (rdma_port_get_link_layer(qp->device, port) !=
1697 IB_LINK_LAYER_INFINIBAND)
1700 /* If we have at lease one Ethernet port, RoCE annex declares that
1701 * multicast LID should be ignored. We can't tell at this step if the
1702 * QP belongs to an IB or Ethernet port.
1707 /* If all the ports are IB, we can check according to IB spec. */
1709 return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
1710 lid == be16_to_cpu(IB_LID_PERMISSIVE));
1713 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1717 if (!qp->device->attach_mcast)
1720 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
1721 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
1724 ret = qp->device->attach_mcast(qp, gid, lid);
1726 atomic_inc(&qp->usecnt);
1729 EXPORT_SYMBOL(ib_attach_mcast);
1731 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1735 if (!qp->device->detach_mcast)
1738 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
1739 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
1742 ret = qp->device->detach_mcast(qp, gid, lid);
1744 atomic_dec(&qp->usecnt);
1747 EXPORT_SYMBOL(ib_detach_mcast);
1749 struct ib_xrcd *ib_alloc_xrcd(struct ib_device *device)
1751 struct ib_xrcd *xrcd;
1753 if (!device->alloc_xrcd)
1754 return ERR_PTR(-ENOSYS);
1756 xrcd = device->alloc_xrcd(device, NULL, NULL);
1757 if (!IS_ERR(xrcd)) {
1758 xrcd->device = device;
1760 atomic_set(&xrcd->usecnt, 0);
1761 mutex_init(&xrcd->tgt_qp_mutex);
1762 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
1767 EXPORT_SYMBOL(ib_alloc_xrcd);
1769 int ib_dealloc_xrcd(struct ib_xrcd *xrcd)
1774 if (atomic_read(&xrcd->usecnt))
1777 while (!list_empty(&xrcd->tgt_qp_list)) {
1778 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
1779 ret = ib_destroy_qp(qp);
1784 return xrcd->device->dealloc_xrcd(xrcd);
1786 EXPORT_SYMBOL(ib_dealloc_xrcd);
1789 * ib_create_wq - Creates a WQ associated with the specified protection
1791 * @pd: The protection domain associated with the WQ.
1792 * @wq_init_attr: A list of initial attributes required to create the
1793 * WQ. If WQ creation succeeds, then the attributes are updated to
1794 * the actual capabilities of the created WQ.
1796 * wq_init_attr->max_wr and wq_init_attr->max_sge determine
1797 * the requested size of the WQ, and set to the actual values allocated
1799 * If ib_create_wq() succeeds, then max_wr and max_sge will always be
1800 * at least as large as the requested values.
1802 struct ib_wq *ib_create_wq(struct ib_pd *pd,
1803 struct ib_wq_init_attr *wq_attr)
1807 if (!pd->device->create_wq)
1808 return ERR_PTR(-ENOSYS);
1810 wq = pd->device->create_wq(pd, wq_attr, NULL);
1812 wq->event_handler = wq_attr->event_handler;
1813 wq->wq_context = wq_attr->wq_context;
1814 wq->wq_type = wq_attr->wq_type;
1815 wq->cq = wq_attr->cq;
1816 wq->device = pd->device;
1819 atomic_inc(&pd->usecnt);
1820 atomic_inc(&wq_attr->cq->usecnt);
1821 atomic_set(&wq->usecnt, 0);
1825 EXPORT_SYMBOL(ib_create_wq);
1828 * ib_destroy_wq - Destroys the specified WQ.
1829 * @wq: The WQ to destroy.
1831 int ib_destroy_wq(struct ib_wq *wq)
1834 struct ib_cq *cq = wq->cq;
1835 struct ib_pd *pd = wq->pd;
1837 if (atomic_read(&wq->usecnt))
1840 err = wq->device->destroy_wq(wq);
1842 atomic_dec(&pd->usecnt);
1843 atomic_dec(&cq->usecnt);
1847 EXPORT_SYMBOL(ib_destroy_wq);
1850 * ib_modify_wq - Modifies the specified WQ.
1851 * @wq: The WQ to modify.
1852 * @wq_attr: On input, specifies the WQ attributes to modify.
1853 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
1854 * are being modified.
1855 * On output, the current values of selected WQ attributes are returned.
1857 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
1862 if (!wq->device->modify_wq)
1865 err = wq->device->modify_wq(wq, wq_attr, wq_attr_mask, NULL);
1868 EXPORT_SYMBOL(ib_modify_wq);
1871 * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
1872 * @device: The device on which to create the rwq indirection table.
1873 * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
1874 * create the Indirection Table.
1876 * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
1877 * than the created ib_rwq_ind_table object and the caller is responsible
1878 * for its memory allocation/free.
1880 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
1881 struct ib_rwq_ind_table_init_attr *init_attr)
1883 struct ib_rwq_ind_table *rwq_ind_table;
1887 if (!device->create_rwq_ind_table)
1888 return ERR_PTR(-ENOSYS);
1890 table_size = (1 << init_attr->log_ind_tbl_size);
1891 rwq_ind_table = device->create_rwq_ind_table(device,
1893 if (IS_ERR(rwq_ind_table))
1894 return rwq_ind_table;
1896 rwq_ind_table->ind_tbl = init_attr->ind_tbl;
1897 rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
1898 rwq_ind_table->device = device;
1899 rwq_ind_table->uobject = NULL;
1900 atomic_set(&rwq_ind_table->usecnt, 0);
1902 for (i = 0; i < table_size; i++)
1903 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
1905 return rwq_ind_table;
1907 EXPORT_SYMBOL(ib_create_rwq_ind_table);
1910 * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
1911 * @wq_ind_table: The Indirection Table to destroy.
1913 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
1916 u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
1917 struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
1919 if (atomic_read(&rwq_ind_table->usecnt))
1922 err = rwq_ind_table->device->destroy_rwq_ind_table(rwq_ind_table);
1924 for (i = 0; i < table_size; i++)
1925 atomic_dec(&ind_tbl[i]->usecnt);
1930 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
1932 struct ib_flow *ib_create_flow(struct ib_qp *qp,
1933 struct ib_flow_attr *flow_attr,
1936 struct ib_flow *flow_id;
1937 if (!qp->device->create_flow)
1938 return ERR_PTR(-ENOSYS);
1940 flow_id = qp->device->create_flow(qp, flow_attr, domain);
1941 if (!IS_ERR(flow_id)) {
1942 atomic_inc(&qp->usecnt);
1947 EXPORT_SYMBOL(ib_create_flow);
1949 int ib_destroy_flow(struct ib_flow *flow_id)
1952 struct ib_qp *qp = flow_id->qp;
1954 err = qp->device->destroy_flow(flow_id);
1956 atomic_dec(&qp->usecnt);
1959 EXPORT_SYMBOL(ib_destroy_flow);
1961 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
1962 struct ib_mr_status *mr_status)
1964 return mr->device->check_mr_status ?
1965 mr->device->check_mr_status(mr, check_mask, mr_status) : -ENOSYS;
1967 EXPORT_SYMBOL(ib_check_mr_status);
1969 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
1972 if (!device->set_vf_link_state)
1975 return device->set_vf_link_state(device, vf, port, state);
1977 EXPORT_SYMBOL(ib_set_vf_link_state);
1979 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
1980 struct ifla_vf_info *info)
1982 if (!device->get_vf_config)
1985 return device->get_vf_config(device, vf, port, info);
1987 EXPORT_SYMBOL(ib_get_vf_config);
1989 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
1990 struct ifla_vf_stats *stats)
1992 if (!device->get_vf_stats)
1995 return device->get_vf_stats(device, vf, port, stats);
1997 EXPORT_SYMBOL(ib_get_vf_stats);
1999 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
2002 if (!device->set_vf_guid)
2005 return device->set_vf_guid(device, vf, port, guid, type);
2007 EXPORT_SYMBOL(ib_set_vf_guid);
2010 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2011 * and set it the memory region.
2012 * @mr: memory region
2013 * @sg: dma mapped scatterlist
2014 * @sg_nents: number of entries in sg
2015 * @sg_offset: offset in bytes into sg
2016 * @page_size: page vector desired page size
2019 * - The first sg element is allowed to have an offset.
2020 * - Each sg element must either be aligned to page_size or virtually
2021 * contiguous to the previous element. In case an sg element has a
2022 * non-contiguous offset, the mapping prefix will not include it.
2023 * - The last sg element is allowed to have length less than page_size.
2024 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2025 * then only max_num_sg entries will be mapped.
2026 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2027 * constraints holds and the page_size argument is ignored.
2029 * Returns the number of sg elements that were mapped to the memory region.
2031 * After this completes successfully, the memory region
2032 * is ready for registration.
2034 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2035 unsigned int *sg_offset, unsigned int page_size)
2037 if (unlikely(!mr->device->map_mr_sg))
2040 mr->page_size = page_size;
2042 return mr->device->map_mr_sg(mr, sg, sg_nents, sg_offset);
2044 EXPORT_SYMBOL(ib_map_mr_sg);
2047 * ib_sg_to_pages() - Convert the largest prefix of a sg list
2049 * @mr: memory region
2050 * @sgl: dma mapped scatterlist
2051 * @sg_nents: number of entries in sg
2052 * @sg_offset_p: IN: start offset in bytes into sg
2053 * OUT: offset in bytes for element n of the sg of the first
2054 * byte that has not been processed where n is the return
2055 * value of this function.
2056 * @set_page: driver page assignment function pointer
2058 * Core service helper for drivers to convert the largest
2059 * prefix of given sg list to a page vector. The sg list
2060 * prefix converted is the prefix that meet the requirements
2063 * Returns the number of sg elements that were assigned to
2066 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2067 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2069 struct scatterlist *sg;
2070 u64 last_end_dma_addr = 0;
2071 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2072 unsigned int last_page_off = 0;
2073 u64 page_mask = ~((u64)mr->page_size - 1);
2076 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2079 mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2082 for_each_sg(sgl, sg, sg_nents, i) {
2083 u64 dma_addr = sg_dma_address(sg) + sg_offset;
2084 u64 prev_addr = dma_addr;
2085 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2086 u64 end_dma_addr = dma_addr + dma_len;
2087 u64 page_addr = dma_addr & page_mask;
2090 * For the second and later elements, check whether either the
2091 * end of element i-1 or the start of element i is not aligned
2092 * on a page boundary.
2094 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2095 /* Stop mapping if there is a gap. */
2096 if (last_end_dma_addr != dma_addr)
2100 * Coalesce this element with the last. If it is small
2101 * enough just update mr->length. Otherwise start
2102 * mapping from the next page.
2108 ret = set_page(mr, page_addr);
2109 if (unlikely(ret < 0)) {
2110 sg_offset = prev_addr - sg_dma_address(sg);
2111 mr->length += prev_addr - dma_addr;
2113 *sg_offset_p = sg_offset;
2114 return i || sg_offset ? i : ret;
2116 prev_addr = page_addr;
2118 page_addr += mr->page_size;
2119 } while (page_addr < end_dma_addr);
2121 mr->length += dma_len;
2122 last_end_dma_addr = end_dma_addr;
2123 last_page_off = end_dma_addr & ~page_mask;
2132 EXPORT_SYMBOL(ib_sg_to_pages);
2134 struct ib_drain_cqe {
2136 struct completion done;
2139 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2141 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2144 complete(&cqe->done);
2148 * Post a WR and block until its completion is reaped for the SQ.
2150 static void __ib_drain_sq(struct ib_qp *qp)
2152 struct ib_cq *cq = qp->send_cq;
2153 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2154 struct ib_drain_cqe sdrain;
2155 struct ib_send_wr swr = {}, *bad_swr;
2158 swr.wr_cqe = &sdrain.cqe;
2159 sdrain.cqe.done = ib_drain_qp_done;
2160 init_completion(&sdrain.done);
2162 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2164 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2168 ret = ib_post_send(qp, &swr, &bad_swr);
2170 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2174 if (cq->poll_ctx == IB_POLL_DIRECT)
2175 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2176 ib_process_cq_direct(cq, -1);
2178 wait_for_completion(&sdrain.done);
2182 * Post a WR and block until its completion is reaped for the RQ.
2184 static void __ib_drain_rq(struct ib_qp *qp)
2186 struct ib_cq *cq = qp->recv_cq;
2187 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2188 struct ib_drain_cqe rdrain;
2189 struct ib_recv_wr rwr = {}, *bad_rwr;
2192 rwr.wr_cqe = &rdrain.cqe;
2193 rdrain.cqe.done = ib_drain_qp_done;
2194 init_completion(&rdrain.done);
2196 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2198 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2202 ret = ib_post_recv(qp, &rwr, &bad_rwr);
2204 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2208 if (cq->poll_ctx == IB_POLL_DIRECT)
2209 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2210 ib_process_cq_direct(cq, -1);
2212 wait_for_completion(&rdrain.done);
2216 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2218 * @qp: queue pair to drain
2220 * If the device has a provider-specific drain function, then
2221 * call that. Otherwise call the generic drain function
2226 * ensure there is room in the CQ and SQ for the drain work request and
2229 * allocate the CQ using ib_alloc_cq().
2231 * ensure that there are no other contexts that are posting WRs concurrently.
2232 * Otherwise the drain is not guaranteed.
2234 void ib_drain_sq(struct ib_qp *qp)
2236 if (qp->device->drain_sq)
2237 qp->device->drain_sq(qp);
2241 EXPORT_SYMBOL(ib_drain_sq);
2244 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2246 * @qp: queue pair to drain
2248 * If the device has a provider-specific drain function, then
2249 * call that. Otherwise call the generic drain function
2254 * ensure there is room in the CQ and RQ for the drain work request and
2257 * allocate the CQ using ib_alloc_cq().
2259 * ensure that there are no other contexts that are posting WRs concurrently.
2260 * Otherwise the drain is not guaranteed.
2262 void ib_drain_rq(struct ib_qp *qp)
2264 if (qp->device->drain_rq)
2265 qp->device->drain_rq(qp);
2269 EXPORT_SYMBOL(ib_drain_rq);
2272 * ib_drain_qp() - Block until all CQEs have been consumed by the
2273 * application on both the RQ and SQ.
2274 * @qp: queue pair to drain
2278 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2281 * allocate the CQs using ib_alloc_cq().
2283 * ensure that there are no other contexts that are posting WRs concurrently.
2284 * Otherwise the drain is not guaranteed.
2286 void ib_drain_qp(struct ib_qp *qp)
2292 EXPORT_SYMBOL(ib_drain_qp);