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;
136 case IB_RATE_14_GBPS: return 6;
137 case IB_RATE_56_GBPS: return 22;
138 case IB_RATE_112_GBPS: return 45;
139 case IB_RATE_168_GBPS: return 67;
140 case IB_RATE_25_GBPS: return 10;
141 case IB_RATE_100_GBPS: return 40;
142 case IB_RATE_200_GBPS: return 80;
143 case IB_RATE_300_GBPS: return 120;
147 EXPORT_SYMBOL(ib_rate_to_mult);
149 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
152 case 1: return IB_RATE_2_5_GBPS;
153 case 2: return IB_RATE_5_GBPS;
154 case 4: return IB_RATE_10_GBPS;
155 case 8: return IB_RATE_20_GBPS;
156 case 12: return IB_RATE_30_GBPS;
157 case 16: return IB_RATE_40_GBPS;
158 case 24: return IB_RATE_60_GBPS;
159 case 32: return IB_RATE_80_GBPS;
160 case 48: return IB_RATE_120_GBPS;
161 case 6: return IB_RATE_14_GBPS;
162 case 22: return IB_RATE_56_GBPS;
163 case 45: return IB_RATE_112_GBPS;
164 case 67: return IB_RATE_168_GBPS;
165 case 10: return IB_RATE_25_GBPS;
166 case 40: return IB_RATE_100_GBPS;
167 case 80: return IB_RATE_200_GBPS;
168 case 120: return IB_RATE_300_GBPS;
169 default: return IB_RATE_PORT_CURRENT;
172 EXPORT_SYMBOL(mult_to_ib_rate);
174 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
177 case IB_RATE_2_5_GBPS: return 2500;
178 case IB_RATE_5_GBPS: return 5000;
179 case IB_RATE_10_GBPS: return 10000;
180 case IB_RATE_20_GBPS: return 20000;
181 case IB_RATE_30_GBPS: return 30000;
182 case IB_RATE_40_GBPS: return 40000;
183 case IB_RATE_60_GBPS: return 60000;
184 case IB_RATE_80_GBPS: return 80000;
185 case IB_RATE_120_GBPS: return 120000;
186 case IB_RATE_14_GBPS: return 14062;
187 case IB_RATE_56_GBPS: return 56250;
188 case IB_RATE_112_GBPS: return 112500;
189 case IB_RATE_168_GBPS: return 168750;
190 case IB_RATE_25_GBPS: return 25781;
191 case IB_RATE_100_GBPS: return 103125;
192 case IB_RATE_200_GBPS: return 206250;
193 case IB_RATE_300_GBPS: return 309375;
197 EXPORT_SYMBOL(ib_rate_to_mbps);
199 __attribute_const__ enum rdma_transport_type
200 rdma_node_get_transport(enum rdma_node_type node_type)
203 if (node_type == RDMA_NODE_USNIC)
204 return RDMA_TRANSPORT_USNIC;
205 if (node_type == RDMA_NODE_USNIC_UDP)
206 return RDMA_TRANSPORT_USNIC_UDP;
207 if (node_type == RDMA_NODE_RNIC)
208 return RDMA_TRANSPORT_IWARP;
210 return RDMA_TRANSPORT_IB;
212 EXPORT_SYMBOL(rdma_node_get_transport);
214 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
216 enum rdma_transport_type lt;
217 if (device->get_link_layer)
218 return device->get_link_layer(device, port_num);
220 lt = rdma_node_get_transport(device->node_type);
221 if (lt == RDMA_TRANSPORT_IB)
222 return IB_LINK_LAYER_INFINIBAND;
224 return IB_LINK_LAYER_ETHERNET;
226 EXPORT_SYMBOL(rdma_port_get_link_layer);
228 /* Protection domains */
231 * ib_alloc_pd - Allocates an unused protection domain.
232 * @device: The device on which to allocate the protection domain.
234 * A protection domain object provides an association between QPs, shared
235 * receive queues, address handles, memory regions, and memory windows.
237 * Every PD has a local_dma_lkey which can be used as the lkey value for local
240 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
244 int mr_access_flags = 0;
246 pd = device->alloc_pd(device, NULL, NULL);
252 pd->__internal_mr = NULL;
253 atomic_set(&pd->usecnt, 0);
256 if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
257 pd->local_dma_lkey = device->local_dma_lkey;
259 mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
261 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
262 pr_warn("%s: enabling unsafe global rkey\n", caller);
263 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
266 pd->res.type = RDMA_RESTRACK_PD;
267 pd->res.kern_name = caller;
268 rdma_restrack_add(&pd->res);
270 if (mr_access_flags) {
273 mr = pd->device->get_dma_mr(pd, mr_access_flags);
279 mr->device = pd->device;
282 mr->need_inval = false;
284 pd->__internal_mr = mr;
286 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
287 pd->local_dma_lkey = pd->__internal_mr->lkey;
289 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
290 pd->unsafe_global_rkey = pd->__internal_mr->rkey;
295 EXPORT_SYMBOL(__ib_alloc_pd);
298 * ib_dealloc_pd - Deallocates a protection domain.
299 * @pd: The protection domain to deallocate.
301 * It is an error to call this function while any resources in the pd still
302 * exist. The caller is responsible to synchronously destroy them and
303 * guarantee no new allocations will happen.
305 void ib_dealloc_pd(struct ib_pd *pd)
309 if (pd->__internal_mr) {
310 ret = pd->device->dereg_mr(pd->__internal_mr);
312 pd->__internal_mr = NULL;
315 /* uverbs manipulates usecnt with proper locking, while the kabi
316 requires the caller to guarantee we can't race here. */
317 WARN_ON(atomic_read(&pd->usecnt));
319 rdma_restrack_del(&pd->res);
320 /* Making delalloc_pd a void return is a WIP, no driver should return
322 ret = pd->device->dealloc_pd(pd);
323 WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd");
325 EXPORT_SYMBOL(ib_dealloc_pd);
327 /* Address handles */
329 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
330 struct rdma_ah_attr *ah_attr,
331 struct ib_udata *udata)
335 ah = pd->device->create_ah(pd, ah_attr, udata);
338 ah->device = pd->device;
341 ah->type = ah_attr->type;
342 atomic_inc(&pd->usecnt);
348 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr)
350 return _rdma_create_ah(pd, ah_attr, NULL);
352 EXPORT_SYMBOL(rdma_create_ah);
355 * rdma_create_user_ah - Creates an address handle for the
356 * given address vector.
357 * It resolves destination mac address for ah attribute of RoCE type.
358 * @pd: The protection domain associated with the address handle.
359 * @ah_attr: The attributes of the address vector.
360 * @udata: pointer to user's input output buffer information need by
363 * It returns 0 on success and returns appropriate error code on error.
364 * The address handle is used to reference a local or global destination
365 * in all UD QP post sends.
367 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
368 struct rdma_ah_attr *ah_attr,
369 struct ib_udata *udata)
373 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
374 err = ib_resolve_eth_dmac(pd->device, ah_attr);
379 return _rdma_create_ah(pd, ah_attr, udata);
381 EXPORT_SYMBOL(rdma_create_user_ah);
383 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
385 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
386 struct iphdr ip4h_checked;
387 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
389 /* If it's IPv6, the version must be 6, otherwise, the first
390 * 20 bytes (before the IPv4 header) are garbled.
392 if (ip6h->version != 6)
393 return (ip4h->version == 4) ? 4 : 0;
394 /* version may be 6 or 4 because the first 20 bytes could be garbled */
396 /* RoCE v2 requires no options, thus header length
403 * We can't write on scattered buffers so we need to copy to
406 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
407 ip4h_checked.check = 0;
408 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
409 /* if IPv4 header checksum is OK, believe it */
410 if (ip4h->check == ip4h_checked.check)
414 EXPORT_SYMBOL(ib_get_rdma_header_version);
416 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
418 const struct ib_grh *grh)
422 if (rdma_protocol_ib(device, port_num))
423 return RDMA_NETWORK_IB;
425 grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
427 if (grh_version == 4)
428 return RDMA_NETWORK_IPV4;
430 if (grh->next_hdr == IPPROTO_UDP)
431 return RDMA_NETWORK_IPV6;
433 return RDMA_NETWORK_ROCE_V1;
436 struct find_gid_index_context {
438 enum ib_gid_type gid_type;
441 static bool find_gid_index(const union ib_gid *gid,
442 const struct ib_gid_attr *gid_attr,
445 struct find_gid_index_context *ctx = context;
447 if (ctx->gid_type != gid_attr->gid_type)
450 if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) ||
451 (is_vlan_dev(gid_attr->ndev) &&
452 vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id))
458 static int get_sgid_index_from_eth(struct ib_device *device, u8 port_num,
459 u16 vlan_id, const union ib_gid *sgid,
460 enum ib_gid_type gid_type,
463 struct find_gid_index_context context = {.vlan_id = vlan_id,
464 .gid_type = gid_type};
466 return ib_find_gid_by_filter(device, sgid, port_num, find_gid_index,
467 &context, gid_index);
470 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
471 enum rdma_network_type net_type,
472 union ib_gid *sgid, union ib_gid *dgid)
474 struct sockaddr_in src_in;
475 struct sockaddr_in dst_in;
476 __be32 src_saddr, dst_saddr;
481 if (net_type == RDMA_NETWORK_IPV4) {
482 memcpy(&src_in.sin_addr.s_addr,
483 &hdr->roce4grh.saddr, 4);
484 memcpy(&dst_in.sin_addr.s_addr,
485 &hdr->roce4grh.daddr, 4);
486 src_saddr = src_in.sin_addr.s_addr;
487 dst_saddr = dst_in.sin_addr.s_addr;
488 ipv6_addr_set_v4mapped(src_saddr,
489 (struct in6_addr *)sgid);
490 ipv6_addr_set_v4mapped(dst_saddr,
491 (struct in6_addr *)dgid);
493 } else if (net_type == RDMA_NETWORK_IPV6 ||
494 net_type == RDMA_NETWORK_IB) {
495 *dgid = hdr->ibgrh.dgid;
496 *sgid = hdr->ibgrh.sgid;
502 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
504 /* Resolve destination mac address and hop limit for unicast destination
505 * GID entry, considering the source GID entry as well.
506 * ah_attribute must have have valid port_num, sgid_index.
508 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
509 struct rdma_ah_attr *ah_attr)
511 struct ib_gid_attr sgid_attr;
512 struct ib_global_route *grh;
513 int hop_limit = 0xff;
517 grh = rdma_ah_retrieve_grh(ah_attr);
519 ret = ib_query_gid(device,
520 rdma_ah_get_port_num(ah_attr),
523 if (ret || !sgid_attr.ndev) {
529 /* If destination is link local and source GID is RoCEv1,
530 * IP stack is not used.
532 if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
533 sgid_attr.gid_type == IB_GID_TYPE_ROCE) {
534 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
539 ret = rdma_addr_find_l2_eth_by_grh(&sgid, &grh->dgid,
541 sgid_attr.ndev, &hop_limit);
543 dev_put(sgid_attr.ndev);
545 grh->hop_limit = hop_limit;
550 * This function initializes address handle attributes from the incoming packet.
551 * Incoming packet has dgid of the receiver node on which this code is
552 * getting executed and, sgid contains the GID of the sender.
554 * When resolving mac address of destination, the arrived dgid is used
555 * as sgid and, sgid is used as dgid because sgid contains destinations
556 * GID whom to respond to.
559 int ib_init_ah_attr_from_wc(struct ib_device *device, u8 port_num,
560 const struct ib_wc *wc, const struct ib_grh *grh,
561 struct rdma_ah_attr *ah_attr)
566 enum rdma_network_type net_type = RDMA_NETWORK_IB;
567 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
574 memset(ah_attr, 0, sizeof *ah_attr);
575 ah_attr->type = rdma_ah_find_type(device, port_num);
576 if (rdma_cap_eth_ah(device, port_num)) {
577 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
578 net_type = wc->network_hdr_type;
580 net_type = ib_get_net_type_by_grh(device, port_num, grh);
581 gid_type = ib_network_to_gid_type(net_type);
583 ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
588 rdma_ah_set_sl(ah_attr, wc->sl);
589 rdma_ah_set_port_num(ah_attr, port_num);
591 if (rdma_protocol_roce(device, port_num)) {
592 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
593 wc->vlan_id : 0xffff;
595 if (!(wc->wc_flags & IB_WC_GRH))
598 ret = get_sgid_index_from_eth(device, port_num,
600 gid_type, &gid_index);
604 flow_class = be32_to_cpu(grh->version_tclass_flow);
605 rdma_ah_set_grh(ah_attr, &sgid,
606 flow_class & 0xFFFFF,
607 (u8)gid_index, hoplimit,
608 (flow_class >> 20) & 0xFF);
609 return ib_resolve_unicast_gid_dmac(device, ah_attr);
611 rdma_ah_set_dlid(ah_attr, wc->slid);
612 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
614 if (wc->wc_flags & IB_WC_GRH) {
615 if (dgid.global.interface_id != cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
616 ret = ib_find_cached_gid_by_port(device, &dgid,
626 flow_class = be32_to_cpu(grh->version_tclass_flow);
627 rdma_ah_set_grh(ah_attr, &sgid,
628 flow_class & 0xFFFFF,
629 (u8)gid_index, hoplimit,
630 (flow_class >> 20) & 0xFF);
635 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
637 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
638 const struct ib_grh *grh, u8 port_num)
640 struct rdma_ah_attr ah_attr;
643 ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
647 return rdma_create_ah(pd, &ah_attr);
649 EXPORT_SYMBOL(ib_create_ah_from_wc);
651 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
653 if (ah->type != ah_attr->type)
656 return ah->device->modify_ah ?
657 ah->device->modify_ah(ah, ah_attr) :
660 EXPORT_SYMBOL(rdma_modify_ah);
662 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
664 return ah->device->query_ah ?
665 ah->device->query_ah(ah, ah_attr) :
668 EXPORT_SYMBOL(rdma_query_ah);
670 int rdma_destroy_ah(struct ib_ah *ah)
676 ret = ah->device->destroy_ah(ah);
678 atomic_dec(&pd->usecnt);
682 EXPORT_SYMBOL(rdma_destroy_ah);
684 /* Shared receive queues */
686 struct ib_srq *ib_create_srq(struct ib_pd *pd,
687 struct ib_srq_init_attr *srq_init_attr)
691 if (!pd->device->create_srq)
692 return ERR_PTR(-EOPNOTSUPP);
694 srq = pd->device->create_srq(pd, srq_init_attr, NULL);
697 srq->device = pd->device;
700 srq->event_handler = srq_init_attr->event_handler;
701 srq->srq_context = srq_init_attr->srq_context;
702 srq->srq_type = srq_init_attr->srq_type;
703 if (ib_srq_has_cq(srq->srq_type)) {
704 srq->ext.cq = srq_init_attr->ext.cq;
705 atomic_inc(&srq->ext.cq->usecnt);
707 if (srq->srq_type == IB_SRQT_XRC) {
708 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
709 atomic_inc(&srq->ext.xrc.xrcd->usecnt);
711 atomic_inc(&pd->usecnt);
712 atomic_set(&srq->usecnt, 0);
717 EXPORT_SYMBOL(ib_create_srq);
719 int ib_modify_srq(struct ib_srq *srq,
720 struct ib_srq_attr *srq_attr,
721 enum ib_srq_attr_mask srq_attr_mask)
723 return srq->device->modify_srq ?
724 srq->device->modify_srq(srq, srq_attr, srq_attr_mask, NULL) :
727 EXPORT_SYMBOL(ib_modify_srq);
729 int ib_query_srq(struct ib_srq *srq,
730 struct ib_srq_attr *srq_attr)
732 return srq->device->query_srq ?
733 srq->device->query_srq(srq, srq_attr) : -EOPNOTSUPP;
735 EXPORT_SYMBOL(ib_query_srq);
737 int ib_destroy_srq(struct ib_srq *srq)
740 enum ib_srq_type srq_type;
741 struct ib_xrcd *uninitialized_var(xrcd);
742 struct ib_cq *uninitialized_var(cq);
745 if (atomic_read(&srq->usecnt))
749 srq_type = srq->srq_type;
750 if (ib_srq_has_cq(srq_type))
752 if (srq_type == IB_SRQT_XRC)
753 xrcd = srq->ext.xrc.xrcd;
755 ret = srq->device->destroy_srq(srq);
757 atomic_dec(&pd->usecnt);
758 if (srq_type == IB_SRQT_XRC)
759 atomic_dec(&xrcd->usecnt);
760 if (ib_srq_has_cq(srq_type))
761 atomic_dec(&cq->usecnt);
766 EXPORT_SYMBOL(ib_destroy_srq);
770 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
772 struct ib_qp *qp = context;
775 spin_lock_irqsave(&qp->device->event_handler_lock, flags);
776 list_for_each_entry(event->element.qp, &qp->open_list, open_list)
777 if (event->element.qp->event_handler)
778 event->element.qp->event_handler(event, event->element.qp->qp_context);
779 spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
782 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
784 mutex_lock(&xrcd->tgt_qp_mutex);
785 list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
786 mutex_unlock(&xrcd->tgt_qp_mutex);
789 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
790 void (*event_handler)(struct ib_event *, void *),
797 qp = kzalloc(sizeof *qp, GFP_KERNEL);
799 return ERR_PTR(-ENOMEM);
801 qp->real_qp = real_qp;
802 err = ib_open_shared_qp_security(qp, real_qp->device);
808 qp->real_qp = real_qp;
809 atomic_inc(&real_qp->usecnt);
810 qp->device = real_qp->device;
811 qp->event_handler = event_handler;
812 qp->qp_context = qp_context;
813 qp->qp_num = real_qp->qp_num;
814 qp->qp_type = real_qp->qp_type;
816 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
817 list_add(&qp->open_list, &real_qp->open_list);
818 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
823 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
824 struct ib_qp_open_attr *qp_open_attr)
826 struct ib_qp *qp, *real_qp;
828 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
829 return ERR_PTR(-EINVAL);
831 qp = ERR_PTR(-EINVAL);
832 mutex_lock(&xrcd->tgt_qp_mutex);
833 list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
834 if (real_qp->qp_num == qp_open_attr->qp_num) {
835 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
836 qp_open_attr->qp_context);
840 mutex_unlock(&xrcd->tgt_qp_mutex);
843 EXPORT_SYMBOL(ib_open_qp);
845 static struct ib_qp *ib_create_xrc_qp(struct ib_qp *qp,
846 struct ib_qp_init_attr *qp_init_attr)
848 struct ib_qp *real_qp = qp;
850 qp->event_handler = __ib_shared_qp_event_handler;
853 qp->send_cq = qp->recv_cq = NULL;
855 qp->xrcd = qp_init_attr->xrcd;
856 atomic_inc(&qp_init_attr->xrcd->usecnt);
857 INIT_LIST_HEAD(&qp->open_list);
859 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
860 qp_init_attr->qp_context);
862 __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
864 real_qp->device->destroy_qp(real_qp);
868 struct ib_qp *ib_create_qp(struct ib_pd *pd,
869 struct ib_qp_init_attr *qp_init_attr)
871 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
875 if (qp_init_attr->rwq_ind_tbl &&
876 (qp_init_attr->recv_cq ||
877 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
878 qp_init_attr->cap.max_recv_sge))
879 return ERR_PTR(-EINVAL);
882 * If the callers is using the RDMA API calculate the resources
883 * needed for the RDMA READ/WRITE operations.
885 * Note that these callers need to pass in a port number.
887 if (qp_init_attr->cap.max_rdma_ctxs)
888 rdma_rw_init_qp(device, qp_init_attr);
890 qp = _ib_create_qp(device, pd, qp_init_attr, NULL, NULL);
894 ret = ib_create_qp_security(qp, device);
901 qp->qp_type = qp_init_attr->qp_type;
902 qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
904 atomic_set(&qp->usecnt, 0);
906 spin_lock_init(&qp->mr_lock);
907 INIT_LIST_HEAD(&qp->rdma_mrs);
908 INIT_LIST_HEAD(&qp->sig_mrs);
911 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT)
912 return ib_create_xrc_qp(qp, qp_init_attr);
914 qp->event_handler = qp_init_attr->event_handler;
915 qp->qp_context = qp_init_attr->qp_context;
916 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
920 qp->recv_cq = qp_init_attr->recv_cq;
921 if (qp_init_attr->recv_cq)
922 atomic_inc(&qp_init_attr->recv_cq->usecnt);
923 qp->srq = qp_init_attr->srq;
925 atomic_inc(&qp_init_attr->srq->usecnt);
928 qp->send_cq = qp_init_attr->send_cq;
931 atomic_inc(&pd->usecnt);
932 if (qp_init_attr->send_cq)
933 atomic_inc(&qp_init_attr->send_cq->usecnt);
934 if (qp_init_attr->rwq_ind_tbl)
935 atomic_inc(&qp->rwq_ind_tbl->usecnt);
937 if (qp_init_attr->cap.max_rdma_ctxs) {
938 ret = rdma_rw_init_mrs(qp, qp_init_attr);
940 pr_err("failed to init MR pool ret= %d\n", ret);
947 * Note: all hw drivers guarantee that max_send_sge is lower than
948 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
949 * max_send_sge <= max_sge_rd.
951 qp->max_write_sge = qp_init_attr->cap.max_send_sge;
952 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
953 device->attrs.max_sge_rd);
957 EXPORT_SYMBOL(ib_create_qp);
959 static const struct {
961 enum ib_qp_attr_mask req_param[IB_QPT_MAX];
962 enum ib_qp_attr_mask opt_param[IB_QPT_MAX];
963 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
965 [IB_QPS_RESET] = { .valid = 1 },
969 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
972 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
973 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
976 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
979 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
982 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
985 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
987 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
993 [IB_QPS_RESET] = { .valid = 1 },
994 [IB_QPS_ERR] = { .valid = 1 },
998 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1001 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1003 IB_QP_ACCESS_FLAGS),
1004 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1006 IB_QP_ACCESS_FLAGS),
1007 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1009 IB_QP_ACCESS_FLAGS),
1010 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1012 IB_QP_ACCESS_FLAGS),
1013 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1015 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1022 [IB_QPT_UC] = (IB_QP_AV |
1026 [IB_QPT_RC] = (IB_QP_AV |
1030 IB_QP_MAX_DEST_RD_ATOMIC |
1031 IB_QP_MIN_RNR_TIMER),
1032 [IB_QPT_XRC_INI] = (IB_QP_AV |
1036 [IB_QPT_XRC_TGT] = (IB_QP_AV |
1040 IB_QP_MAX_DEST_RD_ATOMIC |
1041 IB_QP_MIN_RNR_TIMER),
1044 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1046 [IB_QPT_UC] = (IB_QP_ALT_PATH |
1047 IB_QP_ACCESS_FLAGS |
1049 [IB_QPT_RC] = (IB_QP_ALT_PATH |
1050 IB_QP_ACCESS_FLAGS |
1052 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH |
1053 IB_QP_ACCESS_FLAGS |
1055 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH |
1056 IB_QP_ACCESS_FLAGS |
1058 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1060 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1066 [IB_QPS_RESET] = { .valid = 1 },
1067 [IB_QPS_ERR] = { .valid = 1 },
1071 [IB_QPT_UD] = IB_QP_SQ_PSN,
1072 [IB_QPT_UC] = IB_QP_SQ_PSN,
1073 [IB_QPT_RC] = (IB_QP_TIMEOUT |
1077 IB_QP_MAX_QP_RD_ATOMIC),
1078 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT |
1082 IB_QP_MAX_QP_RD_ATOMIC),
1083 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT |
1085 [IB_QPT_SMI] = IB_QP_SQ_PSN,
1086 [IB_QPT_GSI] = IB_QP_SQ_PSN,
1089 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1091 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1093 IB_QP_ACCESS_FLAGS |
1094 IB_QP_PATH_MIG_STATE),
1095 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1097 IB_QP_ACCESS_FLAGS |
1098 IB_QP_MIN_RNR_TIMER |
1099 IB_QP_PATH_MIG_STATE),
1100 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1102 IB_QP_ACCESS_FLAGS |
1103 IB_QP_PATH_MIG_STATE),
1104 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1106 IB_QP_ACCESS_FLAGS |
1107 IB_QP_MIN_RNR_TIMER |
1108 IB_QP_PATH_MIG_STATE),
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,
1118 [IB_QPS_RESET] = { .valid = 1 },
1119 [IB_QPS_ERR] = { .valid = 1 },
1123 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1125 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1126 IB_QP_ACCESS_FLAGS |
1128 IB_QP_PATH_MIG_STATE),
1129 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1130 IB_QP_ACCESS_FLAGS |
1132 IB_QP_PATH_MIG_STATE |
1133 IB_QP_MIN_RNR_TIMER),
1134 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1135 IB_QP_ACCESS_FLAGS |
1137 IB_QP_PATH_MIG_STATE),
1138 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1139 IB_QP_ACCESS_FLAGS |
1141 IB_QP_PATH_MIG_STATE |
1142 IB_QP_MIN_RNR_TIMER),
1143 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1145 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1147 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1153 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1154 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1155 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1156 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1157 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1158 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1159 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1164 [IB_QPS_RESET] = { .valid = 1 },
1165 [IB_QPS_ERR] = { .valid = 1 },
1169 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1171 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1173 IB_QP_ACCESS_FLAGS |
1174 IB_QP_PATH_MIG_STATE),
1175 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1177 IB_QP_ACCESS_FLAGS |
1178 IB_QP_MIN_RNR_TIMER |
1179 IB_QP_PATH_MIG_STATE),
1180 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1182 IB_QP_ACCESS_FLAGS |
1183 IB_QP_PATH_MIG_STATE),
1184 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1186 IB_QP_ACCESS_FLAGS |
1187 IB_QP_MIN_RNR_TIMER |
1188 IB_QP_PATH_MIG_STATE),
1189 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1191 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1198 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1200 [IB_QPT_UC] = (IB_QP_AV |
1202 IB_QP_ACCESS_FLAGS |
1204 IB_QP_PATH_MIG_STATE),
1205 [IB_QPT_RC] = (IB_QP_PORT |
1210 IB_QP_MAX_QP_RD_ATOMIC |
1211 IB_QP_MAX_DEST_RD_ATOMIC |
1213 IB_QP_ACCESS_FLAGS |
1215 IB_QP_MIN_RNR_TIMER |
1216 IB_QP_PATH_MIG_STATE),
1217 [IB_QPT_XRC_INI] = (IB_QP_PORT |
1222 IB_QP_MAX_QP_RD_ATOMIC |
1224 IB_QP_ACCESS_FLAGS |
1226 IB_QP_PATH_MIG_STATE),
1227 [IB_QPT_XRC_TGT] = (IB_QP_PORT |
1230 IB_QP_MAX_DEST_RD_ATOMIC |
1232 IB_QP_ACCESS_FLAGS |
1234 IB_QP_MIN_RNR_TIMER |
1235 IB_QP_PATH_MIG_STATE),
1236 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1238 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1244 [IB_QPS_RESET] = { .valid = 1 },
1245 [IB_QPS_ERR] = { .valid = 1 },
1249 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1251 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1252 IB_QP_ACCESS_FLAGS),
1253 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1255 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1261 [IB_QPS_RESET] = { .valid = 1 },
1262 [IB_QPS_ERR] = { .valid = 1 }
1266 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1267 enum ib_qp_type type, enum ib_qp_attr_mask mask,
1268 enum rdma_link_layer ll)
1270 enum ib_qp_attr_mask req_param, opt_param;
1272 if (mask & IB_QP_CUR_STATE &&
1273 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1274 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1277 if (!qp_state_table[cur_state][next_state].valid)
1280 req_param = qp_state_table[cur_state][next_state].req_param[type];
1281 opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1283 if ((mask & req_param) != req_param)
1286 if (mask & ~(req_param | opt_param | IB_QP_STATE))
1291 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1293 static int ib_resolve_eth_dmac(struct ib_device *device,
1294 struct rdma_ah_attr *ah_attr)
1297 struct ib_global_route *grh;
1299 if (!rdma_is_port_valid(device, rdma_ah_get_port_num(ah_attr)))
1302 grh = rdma_ah_retrieve_grh(ah_attr);
1304 if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1305 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1308 memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1309 ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1311 ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1312 (char *)ah_attr->roce.dmac);
1315 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1321 * IB core internal function to perform QP attributes modification.
1323 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1324 int attr_mask, struct ib_udata *udata)
1326 u8 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1329 if (rdma_ib_or_roce(qp->device, port)) {
1330 if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1331 pr_warn("%s: %s rq_psn overflow, masking to 24 bits\n",
1332 __func__, qp->device->name);
1333 attr->rq_psn &= 0xffffff;
1336 if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1337 pr_warn("%s: %s sq_psn overflow, masking to 24 bits\n",
1338 __func__, qp->device->name);
1339 attr->sq_psn &= 0xffffff;
1343 ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1344 if (!ret && (attr_mask & IB_QP_PORT))
1345 qp->port = attr->port_num;
1350 static bool is_qp_type_connected(const struct ib_qp *qp)
1352 return (qp->qp_type == IB_QPT_UC ||
1353 qp->qp_type == IB_QPT_RC ||
1354 qp->qp_type == IB_QPT_XRC_INI ||
1355 qp->qp_type == IB_QPT_XRC_TGT);
1359 * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1360 * @ib_qp: The QP to modify.
1361 * @attr: On input, specifies the QP attributes to modify. On output,
1362 * the current values of selected QP attributes are returned.
1363 * @attr_mask: A bit-mask used to specify which attributes of the QP
1364 * are being modified.
1365 * @udata: pointer to user's input output buffer information
1366 * are being modified.
1367 * It returns 0 on success and returns appropriate error code on error.
1369 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1370 int attr_mask, struct ib_udata *udata)
1372 struct ib_qp *qp = ib_qp->real_qp;
1375 if (attr_mask & IB_QP_AV &&
1376 attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1377 is_qp_type_connected(qp)) {
1378 ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
1382 return _ib_modify_qp(qp, attr, attr_mask, udata);
1384 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1386 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u8 *speed, u8 *width)
1390 struct net_device *netdev;
1391 struct ethtool_link_ksettings lksettings;
1393 if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1396 if (!dev->get_netdev)
1399 netdev = dev->get_netdev(dev, port_num);
1404 rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1410 netdev_speed = lksettings.base.speed;
1412 netdev_speed = SPEED_1000;
1413 pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name,
1417 if (netdev_speed <= SPEED_1000) {
1418 *width = IB_WIDTH_1X;
1419 *speed = IB_SPEED_SDR;
1420 } else if (netdev_speed <= SPEED_10000) {
1421 *width = IB_WIDTH_1X;
1422 *speed = IB_SPEED_FDR10;
1423 } else if (netdev_speed <= SPEED_20000) {
1424 *width = IB_WIDTH_4X;
1425 *speed = IB_SPEED_DDR;
1426 } else if (netdev_speed <= SPEED_25000) {
1427 *width = IB_WIDTH_1X;
1428 *speed = IB_SPEED_EDR;
1429 } else if (netdev_speed <= SPEED_40000) {
1430 *width = IB_WIDTH_4X;
1431 *speed = IB_SPEED_FDR10;
1433 *width = IB_WIDTH_4X;
1434 *speed = IB_SPEED_EDR;
1439 EXPORT_SYMBOL(ib_get_eth_speed);
1441 int ib_modify_qp(struct ib_qp *qp,
1442 struct ib_qp_attr *qp_attr,
1445 return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1447 EXPORT_SYMBOL(ib_modify_qp);
1449 int ib_query_qp(struct ib_qp *qp,
1450 struct ib_qp_attr *qp_attr,
1452 struct ib_qp_init_attr *qp_init_attr)
1454 return qp->device->query_qp ?
1455 qp->device->query_qp(qp->real_qp, qp_attr, qp_attr_mask, qp_init_attr) :
1458 EXPORT_SYMBOL(ib_query_qp);
1460 int ib_close_qp(struct ib_qp *qp)
1462 struct ib_qp *real_qp;
1463 unsigned long flags;
1465 real_qp = qp->real_qp;
1469 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1470 list_del(&qp->open_list);
1471 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1473 atomic_dec(&real_qp->usecnt);
1475 ib_close_shared_qp_security(qp->qp_sec);
1480 EXPORT_SYMBOL(ib_close_qp);
1482 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1484 struct ib_xrcd *xrcd;
1485 struct ib_qp *real_qp;
1488 real_qp = qp->real_qp;
1489 xrcd = real_qp->xrcd;
1491 mutex_lock(&xrcd->tgt_qp_mutex);
1493 if (atomic_read(&real_qp->usecnt) == 0)
1494 list_del(&real_qp->xrcd_list);
1497 mutex_unlock(&xrcd->tgt_qp_mutex);
1500 ret = ib_destroy_qp(real_qp);
1502 atomic_dec(&xrcd->usecnt);
1504 __ib_insert_xrcd_qp(xrcd, real_qp);
1510 int ib_destroy_qp(struct ib_qp *qp)
1513 struct ib_cq *scq, *rcq;
1515 struct ib_rwq_ind_table *ind_tbl;
1516 struct ib_qp_security *sec;
1519 WARN_ON_ONCE(qp->mrs_used > 0);
1521 if (atomic_read(&qp->usecnt))
1524 if (qp->real_qp != qp)
1525 return __ib_destroy_shared_qp(qp);
1531 ind_tbl = qp->rwq_ind_tbl;
1534 ib_destroy_qp_security_begin(sec);
1537 rdma_rw_cleanup_mrs(qp);
1539 rdma_restrack_del(&qp->res);
1540 ret = qp->device->destroy_qp(qp);
1543 atomic_dec(&pd->usecnt);
1545 atomic_dec(&scq->usecnt);
1547 atomic_dec(&rcq->usecnt);
1549 atomic_dec(&srq->usecnt);
1551 atomic_dec(&ind_tbl->usecnt);
1553 ib_destroy_qp_security_end(sec);
1556 ib_destroy_qp_security_abort(sec);
1561 EXPORT_SYMBOL(ib_destroy_qp);
1563 /* Completion queues */
1565 struct ib_cq *__ib_create_cq(struct ib_device *device,
1566 ib_comp_handler comp_handler,
1567 void (*event_handler)(struct ib_event *, void *),
1569 const struct ib_cq_init_attr *cq_attr,
1574 cq = device->create_cq(device, cq_attr, NULL, NULL);
1577 cq->device = device;
1579 cq->comp_handler = comp_handler;
1580 cq->event_handler = event_handler;
1581 cq->cq_context = cq_context;
1582 atomic_set(&cq->usecnt, 0);
1583 cq->res.type = RDMA_RESTRACK_CQ;
1584 cq->res.kern_name = caller;
1585 rdma_restrack_add(&cq->res);
1590 EXPORT_SYMBOL(__ib_create_cq);
1592 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1594 return cq->device->modify_cq ?
1595 cq->device->modify_cq(cq, cq_count, cq_period) : -EOPNOTSUPP;
1597 EXPORT_SYMBOL(rdma_set_cq_moderation);
1599 int ib_destroy_cq(struct ib_cq *cq)
1601 if (atomic_read(&cq->usecnt))
1604 rdma_restrack_del(&cq->res);
1605 return cq->device->destroy_cq(cq);
1607 EXPORT_SYMBOL(ib_destroy_cq);
1609 int ib_resize_cq(struct ib_cq *cq, int cqe)
1611 return cq->device->resize_cq ?
1612 cq->device->resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
1614 EXPORT_SYMBOL(ib_resize_cq);
1616 /* Memory regions */
1618 int ib_dereg_mr(struct ib_mr *mr)
1620 struct ib_pd *pd = mr->pd;
1621 struct ib_dm *dm = mr->dm;
1624 rdma_restrack_del(&mr->res);
1625 ret = mr->device->dereg_mr(mr);
1627 atomic_dec(&pd->usecnt);
1629 atomic_dec(&dm->usecnt);
1634 EXPORT_SYMBOL(ib_dereg_mr);
1637 * ib_alloc_mr() - Allocates a memory region
1638 * @pd: protection domain associated with the region
1639 * @mr_type: memory region type
1640 * @max_num_sg: maximum sg entries available for registration.
1643 * Memory registeration page/sg lists must not exceed max_num_sg.
1644 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
1645 * max_num_sg * used_page_size.
1648 struct ib_mr *ib_alloc_mr(struct ib_pd *pd,
1649 enum ib_mr_type mr_type,
1654 if (!pd->device->alloc_mr)
1655 return ERR_PTR(-EOPNOTSUPP);
1657 mr = pd->device->alloc_mr(pd, mr_type, max_num_sg);
1659 mr->device = pd->device;
1663 atomic_inc(&pd->usecnt);
1664 mr->need_inval = false;
1665 mr->res.type = RDMA_RESTRACK_MR;
1666 rdma_restrack_add(&mr->res);
1671 EXPORT_SYMBOL(ib_alloc_mr);
1673 /* "Fast" memory regions */
1675 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
1676 int mr_access_flags,
1677 struct ib_fmr_attr *fmr_attr)
1681 if (!pd->device->alloc_fmr)
1682 return ERR_PTR(-EOPNOTSUPP);
1684 fmr = pd->device->alloc_fmr(pd, mr_access_flags, fmr_attr);
1686 fmr->device = pd->device;
1688 atomic_inc(&pd->usecnt);
1693 EXPORT_SYMBOL(ib_alloc_fmr);
1695 int ib_unmap_fmr(struct list_head *fmr_list)
1699 if (list_empty(fmr_list))
1702 fmr = list_entry(fmr_list->next, struct ib_fmr, list);
1703 return fmr->device->unmap_fmr(fmr_list);
1705 EXPORT_SYMBOL(ib_unmap_fmr);
1707 int ib_dealloc_fmr(struct ib_fmr *fmr)
1713 ret = fmr->device->dealloc_fmr(fmr);
1715 atomic_dec(&pd->usecnt);
1719 EXPORT_SYMBOL(ib_dealloc_fmr);
1721 /* Multicast groups */
1723 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
1725 struct ib_qp_init_attr init_attr = {};
1726 struct ib_qp_attr attr = {};
1727 int num_eth_ports = 0;
1730 /* If QP state >= init, it is assigned to a port and we can check this
1733 if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
1734 if (attr.qp_state >= IB_QPS_INIT) {
1735 if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
1736 IB_LINK_LAYER_INFINIBAND)
1742 /* Can't get a quick answer, iterate over all ports */
1743 for (port = 0; port < qp->device->phys_port_cnt; port++)
1744 if (rdma_port_get_link_layer(qp->device, port) !=
1745 IB_LINK_LAYER_INFINIBAND)
1748 /* If we have at lease one Ethernet port, RoCE annex declares that
1749 * multicast LID should be ignored. We can't tell at this step if the
1750 * QP belongs to an IB or Ethernet port.
1755 /* If all the ports are IB, we can check according to IB spec. */
1757 return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
1758 lid == be16_to_cpu(IB_LID_PERMISSIVE));
1761 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1765 if (!qp->device->attach_mcast)
1768 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
1769 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
1772 ret = qp->device->attach_mcast(qp, gid, lid);
1774 atomic_inc(&qp->usecnt);
1777 EXPORT_SYMBOL(ib_attach_mcast);
1779 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1783 if (!qp->device->detach_mcast)
1786 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
1787 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
1790 ret = qp->device->detach_mcast(qp, gid, lid);
1792 atomic_dec(&qp->usecnt);
1795 EXPORT_SYMBOL(ib_detach_mcast);
1797 struct ib_xrcd *__ib_alloc_xrcd(struct ib_device *device, const char *caller)
1799 struct ib_xrcd *xrcd;
1801 if (!device->alloc_xrcd)
1802 return ERR_PTR(-EOPNOTSUPP);
1804 xrcd = device->alloc_xrcd(device, NULL, NULL);
1805 if (!IS_ERR(xrcd)) {
1806 xrcd->device = device;
1808 atomic_set(&xrcd->usecnt, 0);
1809 mutex_init(&xrcd->tgt_qp_mutex);
1810 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
1815 EXPORT_SYMBOL(__ib_alloc_xrcd);
1817 int ib_dealloc_xrcd(struct ib_xrcd *xrcd)
1822 if (atomic_read(&xrcd->usecnt))
1825 while (!list_empty(&xrcd->tgt_qp_list)) {
1826 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
1827 ret = ib_destroy_qp(qp);
1832 return xrcd->device->dealloc_xrcd(xrcd);
1834 EXPORT_SYMBOL(ib_dealloc_xrcd);
1837 * ib_create_wq - Creates a WQ associated with the specified protection
1839 * @pd: The protection domain associated with the WQ.
1840 * @wq_attr: A list of initial attributes required to create the
1841 * WQ. If WQ creation succeeds, then the attributes are updated to
1842 * the actual capabilities of the created WQ.
1844 * wq_attr->max_wr and wq_attr->max_sge determine
1845 * the requested size of the WQ, and set to the actual values allocated
1847 * If ib_create_wq() succeeds, then max_wr and max_sge will always be
1848 * at least as large as the requested values.
1850 struct ib_wq *ib_create_wq(struct ib_pd *pd,
1851 struct ib_wq_init_attr *wq_attr)
1855 if (!pd->device->create_wq)
1856 return ERR_PTR(-EOPNOTSUPP);
1858 wq = pd->device->create_wq(pd, wq_attr, NULL);
1860 wq->event_handler = wq_attr->event_handler;
1861 wq->wq_context = wq_attr->wq_context;
1862 wq->wq_type = wq_attr->wq_type;
1863 wq->cq = wq_attr->cq;
1864 wq->device = pd->device;
1867 atomic_inc(&pd->usecnt);
1868 atomic_inc(&wq_attr->cq->usecnt);
1869 atomic_set(&wq->usecnt, 0);
1873 EXPORT_SYMBOL(ib_create_wq);
1876 * ib_destroy_wq - Destroys the specified WQ.
1877 * @wq: The WQ to destroy.
1879 int ib_destroy_wq(struct ib_wq *wq)
1882 struct ib_cq *cq = wq->cq;
1883 struct ib_pd *pd = wq->pd;
1885 if (atomic_read(&wq->usecnt))
1888 err = wq->device->destroy_wq(wq);
1890 atomic_dec(&pd->usecnt);
1891 atomic_dec(&cq->usecnt);
1895 EXPORT_SYMBOL(ib_destroy_wq);
1898 * ib_modify_wq - Modifies the specified WQ.
1899 * @wq: The WQ to modify.
1900 * @wq_attr: On input, specifies the WQ attributes to modify.
1901 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
1902 * are being modified.
1903 * On output, the current values of selected WQ attributes are returned.
1905 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
1910 if (!wq->device->modify_wq)
1913 err = wq->device->modify_wq(wq, wq_attr, wq_attr_mask, NULL);
1916 EXPORT_SYMBOL(ib_modify_wq);
1919 * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
1920 * @device: The device on which to create the rwq indirection table.
1921 * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
1922 * create the Indirection Table.
1924 * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
1925 * than the created ib_rwq_ind_table object and the caller is responsible
1926 * for its memory allocation/free.
1928 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
1929 struct ib_rwq_ind_table_init_attr *init_attr)
1931 struct ib_rwq_ind_table *rwq_ind_table;
1935 if (!device->create_rwq_ind_table)
1936 return ERR_PTR(-EOPNOTSUPP);
1938 table_size = (1 << init_attr->log_ind_tbl_size);
1939 rwq_ind_table = device->create_rwq_ind_table(device,
1941 if (IS_ERR(rwq_ind_table))
1942 return rwq_ind_table;
1944 rwq_ind_table->ind_tbl = init_attr->ind_tbl;
1945 rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
1946 rwq_ind_table->device = device;
1947 rwq_ind_table->uobject = NULL;
1948 atomic_set(&rwq_ind_table->usecnt, 0);
1950 for (i = 0; i < table_size; i++)
1951 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
1953 return rwq_ind_table;
1955 EXPORT_SYMBOL(ib_create_rwq_ind_table);
1958 * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
1959 * @wq_ind_table: The Indirection Table to destroy.
1961 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
1964 u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
1965 struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
1967 if (atomic_read(&rwq_ind_table->usecnt))
1970 err = rwq_ind_table->device->destroy_rwq_ind_table(rwq_ind_table);
1972 for (i = 0; i < table_size; i++)
1973 atomic_dec(&ind_tbl[i]->usecnt);
1978 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
1980 struct ib_flow *ib_create_flow(struct ib_qp *qp,
1981 struct ib_flow_attr *flow_attr,
1984 struct ib_flow *flow_id;
1985 if (!qp->device->create_flow)
1986 return ERR_PTR(-EOPNOTSUPP);
1988 flow_id = qp->device->create_flow(qp, flow_attr, domain, NULL);
1989 if (!IS_ERR(flow_id)) {
1990 atomic_inc(&qp->usecnt);
1995 EXPORT_SYMBOL(ib_create_flow);
1997 int ib_destroy_flow(struct ib_flow *flow_id)
2000 struct ib_qp *qp = flow_id->qp;
2002 err = qp->device->destroy_flow(flow_id);
2004 atomic_dec(&qp->usecnt);
2007 EXPORT_SYMBOL(ib_destroy_flow);
2009 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2010 struct ib_mr_status *mr_status)
2012 return mr->device->check_mr_status ?
2013 mr->device->check_mr_status(mr, check_mask, mr_status) : -EOPNOTSUPP;
2015 EXPORT_SYMBOL(ib_check_mr_status);
2017 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
2020 if (!device->set_vf_link_state)
2023 return device->set_vf_link_state(device, vf, port, state);
2025 EXPORT_SYMBOL(ib_set_vf_link_state);
2027 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
2028 struct ifla_vf_info *info)
2030 if (!device->get_vf_config)
2033 return device->get_vf_config(device, vf, port, info);
2035 EXPORT_SYMBOL(ib_get_vf_config);
2037 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
2038 struct ifla_vf_stats *stats)
2040 if (!device->get_vf_stats)
2043 return device->get_vf_stats(device, vf, port, stats);
2045 EXPORT_SYMBOL(ib_get_vf_stats);
2047 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
2050 if (!device->set_vf_guid)
2053 return device->set_vf_guid(device, vf, port, guid, type);
2055 EXPORT_SYMBOL(ib_set_vf_guid);
2058 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2059 * and set it the memory region.
2060 * @mr: memory region
2061 * @sg: dma mapped scatterlist
2062 * @sg_nents: number of entries in sg
2063 * @sg_offset: offset in bytes into sg
2064 * @page_size: page vector desired page size
2067 * - The first sg element is allowed to have an offset.
2068 * - Each sg element must either be aligned to page_size or virtually
2069 * contiguous to the previous element. In case an sg element has a
2070 * non-contiguous offset, the mapping prefix will not include it.
2071 * - The last sg element is allowed to have length less than page_size.
2072 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2073 * then only max_num_sg entries will be mapped.
2074 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2075 * constraints holds and the page_size argument is ignored.
2077 * Returns the number of sg elements that were mapped to the memory region.
2079 * After this completes successfully, the memory region
2080 * is ready for registration.
2082 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2083 unsigned int *sg_offset, unsigned int page_size)
2085 if (unlikely(!mr->device->map_mr_sg))
2088 mr->page_size = page_size;
2090 return mr->device->map_mr_sg(mr, sg, sg_nents, sg_offset);
2092 EXPORT_SYMBOL(ib_map_mr_sg);
2095 * ib_sg_to_pages() - Convert the largest prefix of a sg list
2097 * @mr: memory region
2098 * @sgl: dma mapped scatterlist
2099 * @sg_nents: number of entries in sg
2100 * @sg_offset_p: IN: start offset in bytes into sg
2101 * OUT: offset in bytes for element n of the sg of the first
2102 * byte that has not been processed where n is the return
2103 * value of this function.
2104 * @set_page: driver page assignment function pointer
2106 * Core service helper for drivers to convert the largest
2107 * prefix of given sg list to a page vector. The sg list
2108 * prefix converted is the prefix that meet the requirements
2111 * Returns the number of sg elements that were assigned to
2114 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2115 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2117 struct scatterlist *sg;
2118 u64 last_end_dma_addr = 0;
2119 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2120 unsigned int last_page_off = 0;
2121 u64 page_mask = ~((u64)mr->page_size - 1);
2124 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2127 mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2130 for_each_sg(sgl, sg, sg_nents, i) {
2131 u64 dma_addr = sg_dma_address(sg) + sg_offset;
2132 u64 prev_addr = dma_addr;
2133 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2134 u64 end_dma_addr = dma_addr + dma_len;
2135 u64 page_addr = dma_addr & page_mask;
2138 * For the second and later elements, check whether either the
2139 * end of element i-1 or the start of element i is not aligned
2140 * on a page boundary.
2142 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2143 /* Stop mapping if there is a gap. */
2144 if (last_end_dma_addr != dma_addr)
2148 * Coalesce this element with the last. If it is small
2149 * enough just update mr->length. Otherwise start
2150 * mapping from the next page.
2156 ret = set_page(mr, page_addr);
2157 if (unlikely(ret < 0)) {
2158 sg_offset = prev_addr - sg_dma_address(sg);
2159 mr->length += prev_addr - dma_addr;
2161 *sg_offset_p = sg_offset;
2162 return i || sg_offset ? i : ret;
2164 prev_addr = page_addr;
2166 page_addr += mr->page_size;
2167 } while (page_addr < end_dma_addr);
2169 mr->length += dma_len;
2170 last_end_dma_addr = end_dma_addr;
2171 last_page_off = end_dma_addr & ~page_mask;
2180 EXPORT_SYMBOL(ib_sg_to_pages);
2182 struct ib_drain_cqe {
2184 struct completion done;
2187 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2189 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2192 complete(&cqe->done);
2196 * Post a WR and block until its completion is reaped for the SQ.
2198 static void __ib_drain_sq(struct ib_qp *qp)
2200 struct ib_cq *cq = qp->send_cq;
2201 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2202 struct ib_drain_cqe sdrain;
2203 struct ib_send_wr *bad_swr;
2204 struct ib_rdma_wr swr = {
2207 { .wr_cqe = &sdrain.cqe, },
2208 .opcode = IB_WR_RDMA_WRITE,
2213 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2215 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2219 sdrain.cqe.done = ib_drain_qp_done;
2220 init_completion(&sdrain.done);
2222 ret = ib_post_send(qp, &swr.wr, &bad_swr);
2224 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2228 if (cq->poll_ctx == IB_POLL_DIRECT)
2229 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2230 ib_process_cq_direct(cq, -1);
2232 wait_for_completion(&sdrain.done);
2236 * Post a WR and block until its completion is reaped for the RQ.
2238 static void __ib_drain_rq(struct ib_qp *qp)
2240 struct ib_cq *cq = qp->recv_cq;
2241 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2242 struct ib_drain_cqe rdrain;
2243 struct ib_recv_wr rwr = {}, *bad_rwr;
2246 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2248 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2252 rwr.wr_cqe = &rdrain.cqe;
2253 rdrain.cqe.done = ib_drain_qp_done;
2254 init_completion(&rdrain.done);
2256 ret = ib_post_recv(qp, &rwr, &bad_rwr);
2258 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2262 if (cq->poll_ctx == IB_POLL_DIRECT)
2263 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2264 ib_process_cq_direct(cq, -1);
2266 wait_for_completion(&rdrain.done);
2270 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2272 * @qp: queue pair to drain
2274 * If the device has a provider-specific drain function, then
2275 * call that. Otherwise call the generic drain function
2280 * ensure there is room in the CQ and SQ for the drain work request and
2283 * allocate the CQ using ib_alloc_cq().
2285 * ensure that there are no other contexts that are posting WRs concurrently.
2286 * Otherwise the drain is not guaranteed.
2288 void ib_drain_sq(struct ib_qp *qp)
2290 if (qp->device->drain_sq)
2291 qp->device->drain_sq(qp);
2295 EXPORT_SYMBOL(ib_drain_sq);
2298 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2300 * @qp: queue pair to drain
2302 * If the device has a provider-specific drain function, then
2303 * call that. Otherwise call the generic drain function
2308 * ensure there is room in the CQ and RQ for the drain work request and
2311 * allocate the CQ using ib_alloc_cq().
2313 * ensure that there are no other contexts that are posting WRs concurrently.
2314 * Otherwise the drain is not guaranteed.
2316 void ib_drain_rq(struct ib_qp *qp)
2318 if (qp->device->drain_rq)
2319 qp->device->drain_rq(qp);
2323 EXPORT_SYMBOL(ib_drain_rq);
2326 * ib_drain_qp() - Block until all CQEs have been consumed by the
2327 * application on both the RQ and SQ.
2328 * @qp: queue pair to drain
2332 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2335 * allocate the CQs using ib_alloc_cq().
2337 * ensure that there are no other contexts that are posting WRs concurrently.
2338 * Otherwise the drain is not guaranteed.
2340 void ib_drain_qp(struct ib_qp *qp)
2346 EXPORT_SYMBOL(ib_drain_qp);