2 * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0
4 * Copyright (c) 2004 Mellanox Technologies Ltd. All rights reserved.
5 * Copyright (c) 2004 Infinicon Corporation. All rights reserved.
6 * Copyright (c) 2004 Intel Corporation. All rights reserved.
7 * Copyright (c) 2004 Topspin Corporation. All rights reserved.
8 * Copyright (c) 2004 Voltaire Corporation. All rights reserved.
9 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
10 * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved.
12 * This software is available to you under a choice of one of two
13 * licenses. You may choose to be licensed under the terms of the GNU
14 * General Public License (GPL) Version 2, available from the file
15 * COPYING in the main directory of this source tree, or the
16 * OpenIB.org BSD license below:
18 * Redistribution and use in source and binary forms, with or
19 * without modification, are permitted provided that the following
22 * - Redistributions of source code must retain the above
23 * copyright notice, this list of conditions and the following
26 * - Redistributions in binary form must reproduce the above
27 * copyright notice, this list of conditions and the following
28 * disclaimer in the documentation and/or other materials
29 * provided with the distribution.
31 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
32 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
33 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
34 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
35 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
36 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
37 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
43 #include <linux/errno.h>
44 #include <linux/err.h>
45 #include <linux/string.h>
46 #include <linux/slab.h>
48 #include <linux/in6.h>
50 #include <rdma/ib_verbs.h>
51 #include <rdma/ib_cache.h>
52 #include <rdma/ib_addr.h>
54 #include <netinet/ip.h>
55 #include <netinet/ip6.h>
57 #include <machine/in_cksum.h>
59 #include "core_priv.h"
61 static const char * const ib_events[] = {
62 [IB_EVENT_CQ_ERR] = "CQ error",
63 [IB_EVENT_QP_FATAL] = "QP fatal error",
64 [IB_EVENT_QP_REQ_ERR] = "QP request error",
65 [IB_EVENT_QP_ACCESS_ERR] = "QP access error",
66 [IB_EVENT_COMM_EST] = "communication established",
67 [IB_EVENT_SQ_DRAINED] = "send queue drained",
68 [IB_EVENT_PATH_MIG] = "path migration successful",
69 [IB_EVENT_PATH_MIG_ERR] = "path migration error",
70 [IB_EVENT_DEVICE_FATAL] = "device fatal error",
71 [IB_EVENT_PORT_ACTIVE] = "port active",
72 [IB_EVENT_PORT_ERR] = "port error",
73 [IB_EVENT_LID_CHANGE] = "LID change",
74 [IB_EVENT_PKEY_CHANGE] = "P_key change",
75 [IB_EVENT_SM_CHANGE] = "SM change",
76 [IB_EVENT_SRQ_ERR] = "SRQ error",
77 [IB_EVENT_SRQ_LIMIT_REACHED] = "SRQ limit reached",
78 [IB_EVENT_QP_LAST_WQE_REACHED] = "last WQE reached",
79 [IB_EVENT_CLIENT_REREGISTER] = "client reregister",
80 [IB_EVENT_GID_CHANGE] = "GID changed",
83 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
87 return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
88 ib_events[index] : "unrecognized event";
90 EXPORT_SYMBOL(ib_event_msg);
92 static const char * const wc_statuses[] = {
93 [IB_WC_SUCCESS] = "success",
94 [IB_WC_LOC_LEN_ERR] = "local length error",
95 [IB_WC_LOC_QP_OP_ERR] = "local QP operation error",
96 [IB_WC_LOC_EEC_OP_ERR] = "local EE context operation error",
97 [IB_WC_LOC_PROT_ERR] = "local protection error",
98 [IB_WC_WR_FLUSH_ERR] = "WR flushed",
99 [IB_WC_MW_BIND_ERR] = "memory management operation error",
100 [IB_WC_BAD_RESP_ERR] = "bad response error",
101 [IB_WC_LOC_ACCESS_ERR] = "local access error",
102 [IB_WC_REM_INV_REQ_ERR] = "invalid request error",
103 [IB_WC_REM_ACCESS_ERR] = "remote access error",
104 [IB_WC_REM_OP_ERR] = "remote operation error",
105 [IB_WC_RETRY_EXC_ERR] = "transport retry counter exceeded",
106 [IB_WC_RNR_RETRY_EXC_ERR] = "RNR retry counter exceeded",
107 [IB_WC_LOC_RDD_VIOL_ERR] = "local RDD violation error",
108 [IB_WC_REM_INV_RD_REQ_ERR] = "remote invalid RD request",
109 [IB_WC_REM_ABORT_ERR] = "operation aborted",
110 [IB_WC_INV_EECN_ERR] = "invalid EE context number",
111 [IB_WC_INV_EEC_STATE_ERR] = "invalid EE context state",
112 [IB_WC_FATAL_ERR] = "fatal error",
113 [IB_WC_RESP_TIMEOUT_ERR] = "response timeout error",
114 [IB_WC_GENERAL_ERR] = "general error",
117 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
119 size_t index = status;
121 return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
122 wc_statuses[index] : "unrecognized status";
124 EXPORT_SYMBOL(ib_wc_status_msg);
126 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
129 case IB_RATE_2_5_GBPS: return 1;
130 case IB_RATE_5_GBPS: return 2;
131 case IB_RATE_10_GBPS: return 4;
132 case IB_RATE_20_GBPS: return 8;
133 case IB_RATE_30_GBPS: return 12;
134 case IB_RATE_40_GBPS: return 16;
135 case IB_RATE_60_GBPS: return 24;
136 case IB_RATE_80_GBPS: return 32;
137 case IB_RATE_120_GBPS: return 48;
141 EXPORT_SYMBOL(ib_rate_to_mult);
143 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
146 case 1: return IB_RATE_2_5_GBPS;
147 case 2: return IB_RATE_5_GBPS;
148 case 4: return IB_RATE_10_GBPS;
149 case 8: return IB_RATE_20_GBPS;
150 case 12: return IB_RATE_30_GBPS;
151 case 16: return IB_RATE_40_GBPS;
152 case 24: return IB_RATE_60_GBPS;
153 case 32: return IB_RATE_80_GBPS;
154 case 48: return IB_RATE_120_GBPS;
155 default: return IB_RATE_PORT_CURRENT;
158 EXPORT_SYMBOL(mult_to_ib_rate);
160 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
163 case IB_RATE_2_5_GBPS: return 2500;
164 case IB_RATE_5_GBPS: return 5000;
165 case IB_RATE_10_GBPS: return 10000;
166 case IB_RATE_20_GBPS: return 20000;
167 case IB_RATE_30_GBPS: return 30000;
168 case IB_RATE_40_GBPS: return 40000;
169 case IB_RATE_60_GBPS: return 60000;
170 case IB_RATE_80_GBPS: return 80000;
171 case IB_RATE_120_GBPS: return 120000;
172 case IB_RATE_14_GBPS: return 14062;
173 case IB_RATE_56_GBPS: return 56250;
174 case IB_RATE_112_GBPS: return 112500;
175 case IB_RATE_168_GBPS: return 168750;
176 case IB_RATE_25_GBPS: return 25781;
177 case IB_RATE_100_GBPS: return 103125;
178 case IB_RATE_200_GBPS: return 206250;
179 case IB_RATE_300_GBPS: return 309375;
183 EXPORT_SYMBOL(ib_rate_to_mbps);
185 __attribute_const__ enum rdma_transport_type
186 rdma_node_get_transport(enum rdma_node_type node_type)
189 case RDMA_NODE_IB_CA:
190 case RDMA_NODE_IB_SWITCH:
191 case RDMA_NODE_IB_ROUTER:
192 return RDMA_TRANSPORT_IB;
194 return RDMA_TRANSPORT_IWARP;
195 case RDMA_NODE_USNIC:
196 return RDMA_TRANSPORT_USNIC;
197 case RDMA_NODE_USNIC_UDP:
198 return RDMA_TRANSPORT_USNIC_UDP;
204 EXPORT_SYMBOL(rdma_node_get_transport);
206 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
208 if (device->get_link_layer)
209 return device->get_link_layer(device, port_num);
211 switch (rdma_node_get_transport(device->node_type)) {
212 case RDMA_TRANSPORT_IB:
213 return IB_LINK_LAYER_INFINIBAND;
214 case RDMA_TRANSPORT_IWARP:
215 case RDMA_TRANSPORT_USNIC:
216 case RDMA_TRANSPORT_USNIC_UDP:
217 return IB_LINK_LAYER_ETHERNET;
219 return IB_LINK_LAYER_UNSPECIFIED;
222 EXPORT_SYMBOL(rdma_port_get_link_layer);
224 /* Protection domains */
227 * ib_alloc_pd - Allocates an unused protection domain.
228 * @device: The device on which to allocate the protection domain.
230 * A protection domain object provides an association between QPs, shared
231 * receive queues, address handles, memory regions, and memory windows.
233 * Every PD has a local_dma_lkey which can be used as the lkey value for local
236 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
240 int mr_access_flags = 0;
242 pd = device->alloc_pd(device, NULL, NULL);
248 pd->__internal_mr = NULL;
249 atomic_set(&pd->usecnt, 0);
252 if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
253 pd->local_dma_lkey = device->local_dma_lkey;
255 mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
257 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
258 pr_warn("%s: enabling unsafe global rkey\n", caller);
259 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
262 if (mr_access_flags) {
265 mr = pd->device->get_dma_mr(pd, mr_access_flags);
271 mr->device = pd->device;
274 mr->need_inval = false;
276 pd->__internal_mr = mr;
278 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
279 pd->local_dma_lkey = pd->__internal_mr->lkey;
281 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
282 pd->unsafe_global_rkey = pd->__internal_mr->rkey;
287 EXPORT_SYMBOL(__ib_alloc_pd);
290 * ib_dealloc_pd - Deallocates a protection domain.
291 * @pd: The protection domain to deallocate.
293 * It is an error to call this function while any resources in the pd still
294 * exist. The caller is responsible to synchronously destroy them and
295 * guarantee no new allocations will happen.
297 void ib_dealloc_pd(struct ib_pd *pd)
301 if (pd->__internal_mr) {
302 ret = pd->device->dereg_mr(pd->__internal_mr);
304 pd->__internal_mr = NULL;
307 /* uverbs manipulates usecnt with proper locking, while the kabi
308 requires the caller to guarantee we can't race here. */
309 WARN_ON(atomic_read(&pd->usecnt));
311 /* Making delalloc_pd a void return is a WIP, no driver should return
313 ret = pd->device->dealloc_pd(pd);
314 WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd");
316 EXPORT_SYMBOL(ib_dealloc_pd);
318 /* Address handles */
320 struct ib_ah *ib_create_ah(struct ib_pd *pd, struct ib_ah_attr *ah_attr)
324 ah = pd->device->create_ah(pd, ah_attr);
327 ah->device = pd->device;
330 atomic_inc(&pd->usecnt);
335 EXPORT_SYMBOL(ib_create_ah);
337 static int ib_get_header_version(const union rdma_network_hdr *hdr)
339 const struct ip *ip4h = (const struct ip *)&hdr->roce4grh;
340 struct ip ip4h_checked;
341 const struct ip6_hdr *ip6h = (const struct ip6_hdr *)&hdr->ibgrh;
343 /* If it's IPv6, the version must be 6, otherwise, the first
344 * 20 bytes (before the IPv4 header) are garbled.
346 if ((ip6h->ip6_vfc & IPV6_VERSION_MASK) != IPV6_VERSION)
347 return (ip4h->ip_v == 4) ? 4 : 0;
348 /* version may be 6 or 4 because the first 20 bytes could be garbled */
350 /* RoCE v2 requires no options, thus header length
353 if (ip4h->ip_hl != 5)
357 * We can't write on scattered buffers so we need to copy to
360 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
361 ip4h_checked.ip_sum = 0;
362 #if defined(INET) || defined(INET6)
363 ip4h_checked.ip_sum = in_cksum_hdr(&ip4h_checked);
365 /* if IPv4 header checksum is OK, believe it */
366 if (ip4h->ip_sum == ip4h_checked.ip_sum)
371 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
373 const struct ib_grh *grh)
377 if (rdma_protocol_ib(device, port_num))
378 return RDMA_NETWORK_IB;
380 grh_version = ib_get_header_version((const union rdma_network_hdr *)grh);
382 if (grh_version == 4)
383 return RDMA_NETWORK_IPV4;
385 if (grh->next_hdr == IPPROTO_UDP)
386 return RDMA_NETWORK_IPV6;
388 return RDMA_NETWORK_ROCE_V1;
391 struct find_gid_index_context {
393 enum ib_gid_type gid_type;
396 static bool find_gid_index(const union ib_gid *gid,
397 const struct ib_gid_attr *gid_attr,
400 struct find_gid_index_context *ctx =
401 (struct find_gid_index_context *)context;
403 if (ctx->gid_type != gid_attr->gid_type)
406 if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) ||
407 (is_vlan_dev(gid_attr->ndev) &&
408 vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id))
414 static int get_sgid_index_from_eth(struct ib_device *device, u8 port_num,
415 u16 vlan_id, const union ib_gid *sgid,
416 enum ib_gid_type gid_type,
419 struct find_gid_index_context context = {.vlan_id = vlan_id,
420 .gid_type = gid_type};
422 return ib_find_gid_by_filter(device, sgid, port_num, find_gid_index,
423 &context, gid_index);
426 static int get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
427 enum rdma_network_type net_type,
428 union ib_gid *sgid, union ib_gid *dgid)
430 struct sockaddr_in src_in;
431 struct sockaddr_in dst_in;
432 __be32 src_saddr, dst_saddr;
437 if (net_type == RDMA_NETWORK_IPV4) {
438 memcpy(&src_in.sin_addr.s_addr,
439 &hdr->roce4grh.ip_src, 4);
440 memcpy(&dst_in.sin_addr.s_addr,
441 &hdr->roce4grh.ip_dst, 4);
442 src_saddr = src_in.sin_addr.s_addr;
443 dst_saddr = dst_in.sin_addr.s_addr;
444 ipv6_addr_set_v4mapped(src_saddr,
445 (struct in6_addr *)sgid);
446 ipv6_addr_set_v4mapped(dst_saddr,
447 (struct in6_addr *)dgid);
449 } else if (net_type == RDMA_NETWORK_IPV6 ||
450 net_type == RDMA_NETWORK_IB) {
451 *dgid = hdr->ibgrh.dgid;
452 *sgid = hdr->ibgrh.sgid;
459 int ib_init_ah_from_wc(struct ib_device *device, u8 port_num,
460 const struct ib_wc *wc, const struct ib_grh *grh,
461 struct ib_ah_attr *ah_attr)
466 enum rdma_network_type net_type = RDMA_NETWORK_IB;
467 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
472 memset(ah_attr, 0, sizeof *ah_attr);
473 if (rdma_cap_eth_ah(device, port_num)) {
474 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
475 net_type = wc->network_hdr_type;
477 net_type = ib_get_net_type_by_grh(device, port_num, grh);
478 gid_type = ib_network_to_gid_type(net_type);
480 ret = get_gids_from_rdma_hdr((const union rdma_network_hdr *)grh, net_type,
485 if (rdma_protocol_roce(device, port_num)) {
486 struct ib_gid_attr dgid_attr;
487 const u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
488 wc->vlan_id : 0xffff;
490 if (!(wc->wc_flags & IB_WC_GRH))
493 ret = get_sgid_index_from_eth(device, port_num, vlan_id,
494 &dgid, gid_type, &gid_index);
498 ret = ib_get_cached_gid(device, port_num, gid_index, &dgid, &dgid_attr);
502 if (dgid_attr.ndev == NULL)
505 ret = rdma_addr_find_l2_eth_by_grh(&dgid, &sgid, ah_attr->dmac,
506 dgid_attr.ndev, &hoplimit);
508 dev_put(dgid_attr.ndev);
513 ah_attr->dlid = wc->slid;
514 ah_attr->sl = wc->sl;
515 ah_attr->src_path_bits = wc->dlid_path_bits;
516 ah_attr->port_num = port_num;
518 if (wc->wc_flags & IB_WC_GRH) {
519 ah_attr->ah_flags = IB_AH_GRH;
520 ah_attr->grh.dgid = sgid;
522 if (!rdma_cap_eth_ah(device, port_num)) {
523 if (dgid.global.interface_id != cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
524 ret = ib_find_cached_gid_by_port(device, &dgid,
535 ah_attr->grh.sgid_index = (u8) gid_index;
536 flow_class = be32_to_cpu(grh->version_tclass_flow);
537 ah_attr->grh.flow_label = flow_class & 0xFFFFF;
538 ah_attr->grh.hop_limit = hoplimit;
539 ah_attr->grh.traffic_class = (flow_class >> 20) & 0xFF;
543 EXPORT_SYMBOL(ib_init_ah_from_wc);
545 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
546 const struct ib_grh *grh, u8 port_num)
548 struct ib_ah_attr ah_attr;
551 ret = ib_init_ah_from_wc(pd->device, port_num, wc, grh, &ah_attr);
555 return ib_create_ah(pd, &ah_attr);
557 EXPORT_SYMBOL(ib_create_ah_from_wc);
559 int ib_modify_ah(struct ib_ah *ah, struct ib_ah_attr *ah_attr)
561 return ah->device->modify_ah ?
562 ah->device->modify_ah(ah, ah_attr) :
565 EXPORT_SYMBOL(ib_modify_ah);
567 int ib_query_ah(struct ib_ah *ah, struct ib_ah_attr *ah_attr)
569 return ah->device->query_ah ?
570 ah->device->query_ah(ah, ah_attr) :
573 EXPORT_SYMBOL(ib_query_ah);
575 int ib_destroy_ah(struct ib_ah *ah)
581 ret = ah->device->destroy_ah(ah);
583 atomic_dec(&pd->usecnt);
587 EXPORT_SYMBOL(ib_destroy_ah);
589 /* Shared receive queues */
591 struct ib_srq *ib_create_srq(struct ib_pd *pd,
592 struct ib_srq_init_attr *srq_init_attr)
596 if (!pd->device->create_srq)
597 return ERR_PTR(-ENOSYS);
599 srq = pd->device->create_srq(pd, srq_init_attr, NULL);
602 srq->device = pd->device;
605 srq->event_handler = srq_init_attr->event_handler;
606 srq->srq_context = srq_init_attr->srq_context;
607 srq->srq_type = srq_init_attr->srq_type;
608 if (srq->srq_type == IB_SRQT_XRC) {
609 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
610 srq->ext.xrc.cq = srq_init_attr->ext.xrc.cq;
611 atomic_inc(&srq->ext.xrc.xrcd->usecnt);
612 atomic_inc(&srq->ext.xrc.cq->usecnt);
614 atomic_inc(&pd->usecnt);
615 atomic_set(&srq->usecnt, 0);
620 EXPORT_SYMBOL(ib_create_srq);
622 int ib_modify_srq(struct ib_srq *srq,
623 struct ib_srq_attr *srq_attr,
624 enum ib_srq_attr_mask srq_attr_mask)
626 return srq->device->modify_srq ?
627 srq->device->modify_srq(srq, srq_attr, srq_attr_mask, NULL) :
630 EXPORT_SYMBOL(ib_modify_srq);
632 int ib_query_srq(struct ib_srq *srq,
633 struct ib_srq_attr *srq_attr)
635 return srq->device->query_srq ?
636 srq->device->query_srq(srq, srq_attr) : -ENOSYS;
638 EXPORT_SYMBOL(ib_query_srq);
640 int ib_destroy_srq(struct ib_srq *srq)
643 enum ib_srq_type srq_type;
644 struct ib_xrcd *uninitialized_var(xrcd);
645 struct ib_cq *uninitialized_var(cq);
648 if (atomic_read(&srq->usecnt))
652 srq_type = srq->srq_type;
653 if (srq_type == IB_SRQT_XRC) {
654 xrcd = srq->ext.xrc.xrcd;
655 cq = srq->ext.xrc.cq;
658 ret = srq->device->destroy_srq(srq);
660 atomic_dec(&pd->usecnt);
661 if (srq_type == IB_SRQT_XRC) {
662 atomic_dec(&xrcd->usecnt);
663 atomic_dec(&cq->usecnt);
669 EXPORT_SYMBOL(ib_destroy_srq);
673 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
675 struct ib_qp *qp = context;
678 spin_lock_irqsave(&qp->device->event_handler_lock, flags);
679 list_for_each_entry(event->element.qp, &qp->open_list, open_list)
680 if (event->element.qp->event_handler)
681 event->element.qp->event_handler(event, event->element.qp->qp_context);
682 spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
685 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
687 mutex_lock(&xrcd->tgt_qp_mutex);
688 list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
689 mutex_unlock(&xrcd->tgt_qp_mutex);
692 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
693 void (*event_handler)(struct ib_event *, void *),
699 qp = kzalloc(sizeof *qp, GFP_KERNEL);
701 return ERR_PTR(-ENOMEM);
703 qp->real_qp = real_qp;
704 atomic_inc(&real_qp->usecnt);
705 qp->device = real_qp->device;
706 qp->event_handler = event_handler;
707 qp->qp_context = qp_context;
708 qp->qp_num = real_qp->qp_num;
709 qp->qp_type = real_qp->qp_type;
711 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
712 list_add(&qp->open_list, &real_qp->open_list);
713 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
718 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
719 struct ib_qp_open_attr *qp_open_attr)
721 struct ib_qp *qp, *real_qp;
723 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
724 return ERR_PTR(-EINVAL);
726 qp = ERR_PTR(-EINVAL);
727 mutex_lock(&xrcd->tgt_qp_mutex);
728 list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
729 if (real_qp->qp_num == qp_open_attr->qp_num) {
730 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
731 qp_open_attr->qp_context);
735 mutex_unlock(&xrcd->tgt_qp_mutex);
738 EXPORT_SYMBOL(ib_open_qp);
740 static struct ib_qp *ib_create_xrc_qp(struct ib_qp *qp,
741 struct ib_qp_init_attr *qp_init_attr)
743 struct ib_qp *real_qp = qp;
745 qp->event_handler = __ib_shared_qp_event_handler;
748 qp->send_cq = qp->recv_cq = NULL;
750 qp->xrcd = qp_init_attr->xrcd;
751 atomic_inc(&qp_init_attr->xrcd->usecnt);
752 INIT_LIST_HEAD(&qp->open_list);
754 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
755 qp_init_attr->qp_context);
757 __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
759 real_qp->device->destroy_qp(real_qp);
763 struct ib_qp *ib_create_qp(struct ib_pd *pd,
764 struct ib_qp_init_attr *qp_init_attr)
766 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
769 if (qp_init_attr->rwq_ind_tbl &&
770 (qp_init_attr->recv_cq ||
771 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
772 qp_init_attr->cap.max_recv_sge))
773 return ERR_PTR(-EINVAL);
775 qp = device->create_qp(pd, qp_init_attr, NULL);
782 qp->qp_type = qp_init_attr->qp_type;
783 qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
785 atomic_set(&qp->usecnt, 0);
786 spin_lock_init(&qp->mr_lock);
788 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT)
789 return ib_create_xrc_qp(qp, qp_init_attr);
791 qp->event_handler = qp_init_attr->event_handler;
792 qp->qp_context = qp_init_attr->qp_context;
793 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
797 qp->recv_cq = qp_init_attr->recv_cq;
798 if (qp_init_attr->recv_cq)
799 atomic_inc(&qp_init_attr->recv_cq->usecnt);
800 qp->srq = qp_init_attr->srq;
802 atomic_inc(&qp_init_attr->srq->usecnt);
806 qp->send_cq = qp_init_attr->send_cq;
809 atomic_inc(&pd->usecnt);
810 if (qp_init_attr->send_cq)
811 atomic_inc(&qp_init_attr->send_cq->usecnt);
812 if (qp_init_attr->rwq_ind_tbl)
813 atomic_inc(&qp->rwq_ind_tbl->usecnt);
816 * Note: all hw drivers guarantee that max_send_sge is lower than
817 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
818 * max_send_sge <= max_sge_rd.
820 qp->max_write_sge = qp_init_attr->cap.max_send_sge;
821 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
822 device->attrs.max_sge_rd);
826 EXPORT_SYMBOL(ib_create_qp);
828 static const struct {
830 enum ib_qp_attr_mask req_param[IB_QPT_MAX];
831 enum ib_qp_attr_mask opt_param[IB_QPT_MAX];
832 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
834 [IB_QPS_RESET] = { .valid = 1 },
838 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
841 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
842 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
845 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
848 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
851 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
854 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
856 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
862 [IB_QPS_RESET] = { .valid = 1 },
863 [IB_QPS_ERR] = { .valid = 1 },
867 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
870 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
873 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
876 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
879 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
882 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
884 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
891 [IB_QPT_UC] = (IB_QP_AV |
895 [IB_QPT_RC] = (IB_QP_AV |
899 IB_QP_MAX_DEST_RD_ATOMIC |
900 IB_QP_MIN_RNR_TIMER),
901 [IB_QPT_XRC_INI] = (IB_QP_AV |
905 [IB_QPT_XRC_TGT] = (IB_QP_AV |
909 IB_QP_MAX_DEST_RD_ATOMIC |
910 IB_QP_MIN_RNR_TIMER),
913 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
915 [IB_QPT_UC] = (IB_QP_ALT_PATH |
918 [IB_QPT_RC] = (IB_QP_ALT_PATH |
921 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH |
924 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH |
927 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
929 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
935 [IB_QPS_RESET] = { .valid = 1 },
936 [IB_QPS_ERR] = { .valid = 1 },
940 [IB_QPT_UD] = IB_QP_SQ_PSN,
941 [IB_QPT_UC] = IB_QP_SQ_PSN,
942 [IB_QPT_RC] = (IB_QP_TIMEOUT |
946 IB_QP_MAX_QP_RD_ATOMIC),
947 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT |
951 IB_QP_MAX_QP_RD_ATOMIC),
952 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT |
954 [IB_QPT_SMI] = IB_QP_SQ_PSN,
955 [IB_QPT_GSI] = IB_QP_SQ_PSN,
958 [IB_QPT_UD] = (IB_QP_CUR_STATE |
960 [IB_QPT_UC] = (IB_QP_CUR_STATE |
963 IB_QP_PATH_MIG_STATE),
964 [IB_QPT_RC] = (IB_QP_CUR_STATE |
967 IB_QP_MIN_RNR_TIMER |
968 IB_QP_PATH_MIG_STATE),
969 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
972 IB_QP_PATH_MIG_STATE),
973 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
976 IB_QP_MIN_RNR_TIMER |
977 IB_QP_PATH_MIG_STATE),
978 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
980 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
986 [IB_QPS_RESET] = { .valid = 1 },
987 [IB_QPS_ERR] = { .valid = 1 },
991 [IB_QPT_UD] = (IB_QP_CUR_STATE |
993 [IB_QPT_UC] = (IB_QP_CUR_STATE |
996 IB_QP_PATH_MIG_STATE),
997 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1000 IB_QP_PATH_MIG_STATE |
1001 IB_QP_MIN_RNR_TIMER),
1002 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1003 IB_QP_ACCESS_FLAGS |
1005 IB_QP_PATH_MIG_STATE),
1006 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1007 IB_QP_ACCESS_FLAGS |
1009 IB_QP_PATH_MIG_STATE |
1010 IB_QP_MIN_RNR_TIMER),
1011 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1013 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1020 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1021 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1022 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1023 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1024 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1025 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1026 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1031 [IB_QPS_RESET] = { .valid = 1 },
1032 [IB_QPS_ERR] = { .valid = 1 },
1036 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1038 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1040 IB_QP_ACCESS_FLAGS |
1041 IB_QP_PATH_MIG_STATE),
1042 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1044 IB_QP_ACCESS_FLAGS |
1045 IB_QP_MIN_RNR_TIMER |
1046 IB_QP_PATH_MIG_STATE),
1047 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1049 IB_QP_ACCESS_FLAGS |
1050 IB_QP_PATH_MIG_STATE),
1051 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1053 IB_QP_ACCESS_FLAGS |
1054 IB_QP_MIN_RNR_TIMER |
1055 IB_QP_PATH_MIG_STATE),
1056 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1058 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1065 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1067 [IB_QPT_UC] = (IB_QP_AV |
1069 IB_QP_ACCESS_FLAGS |
1071 IB_QP_PATH_MIG_STATE),
1072 [IB_QPT_RC] = (IB_QP_PORT |
1077 IB_QP_MAX_QP_RD_ATOMIC |
1078 IB_QP_MAX_DEST_RD_ATOMIC |
1080 IB_QP_ACCESS_FLAGS |
1082 IB_QP_MIN_RNR_TIMER |
1083 IB_QP_PATH_MIG_STATE),
1084 [IB_QPT_XRC_INI] = (IB_QP_PORT |
1089 IB_QP_MAX_QP_RD_ATOMIC |
1091 IB_QP_ACCESS_FLAGS |
1093 IB_QP_PATH_MIG_STATE),
1094 [IB_QPT_XRC_TGT] = (IB_QP_PORT |
1097 IB_QP_MAX_DEST_RD_ATOMIC |
1099 IB_QP_ACCESS_FLAGS |
1101 IB_QP_MIN_RNR_TIMER |
1102 IB_QP_PATH_MIG_STATE),
1103 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1105 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1111 [IB_QPS_RESET] = { .valid = 1 },
1112 [IB_QPS_ERR] = { .valid = 1 },
1116 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1118 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1119 IB_QP_ACCESS_FLAGS),
1120 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1122 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1128 [IB_QPS_RESET] = { .valid = 1 },
1129 [IB_QPS_ERR] = { .valid = 1 }
1133 int ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1134 enum ib_qp_type type, enum ib_qp_attr_mask mask,
1135 enum rdma_link_layer ll)
1137 enum ib_qp_attr_mask req_param, opt_param;
1139 if (cur_state < 0 || cur_state > IB_QPS_ERR ||
1140 next_state < 0 || next_state > IB_QPS_ERR)
1143 if (mask & IB_QP_CUR_STATE &&
1144 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1145 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1148 if (!qp_state_table[cur_state][next_state].valid)
1151 req_param = qp_state_table[cur_state][next_state].req_param[type];
1152 opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1154 if ((mask & req_param) != req_param)
1157 if (mask & ~(req_param | opt_param | IB_QP_STATE))
1162 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1164 int ib_resolve_eth_dmac(struct ib_qp *qp,
1165 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1169 if (*qp_attr_mask & IB_QP_AV) {
1170 if (qp_attr->ah_attr.port_num < rdma_start_port(qp->device) ||
1171 qp_attr->ah_attr.port_num > rdma_end_port(qp->device))
1174 if (!rdma_cap_eth_ah(qp->device, qp_attr->ah_attr.port_num))
1177 if (rdma_link_local_addr((struct in6_addr *)qp_attr->ah_attr.grh.dgid.raw)) {
1178 rdma_get_ll_mac((struct in6_addr *)qp_attr->ah_attr.grh.dgid.raw,
1179 qp_attr->ah_attr.dmac);
1182 struct ib_gid_attr sgid_attr;
1185 ret = ib_query_gid(qp->device,
1186 qp_attr->ah_attr.port_num,
1187 qp_attr->ah_attr.grh.sgid_index,
1190 if (ret || !sgid_attr.ndev) {
1196 ret = rdma_addr_find_l2_eth_by_grh(&sgid,
1197 &qp_attr->ah_attr.grh.dgid,
1198 qp_attr->ah_attr.dmac,
1199 sgid_attr.ndev, &hop_limit);
1201 dev_put(sgid_attr.ndev);
1203 qp_attr->ah_attr.grh.hop_limit = hop_limit;
1209 EXPORT_SYMBOL(ib_resolve_eth_dmac);
1212 int ib_modify_qp(struct ib_qp *qp,
1213 struct ib_qp_attr *qp_attr,
1218 ret = ib_resolve_eth_dmac(qp, qp_attr, &qp_attr_mask);
1222 return qp->device->modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1224 EXPORT_SYMBOL(ib_modify_qp);
1226 int ib_query_qp(struct ib_qp *qp,
1227 struct ib_qp_attr *qp_attr,
1229 struct ib_qp_init_attr *qp_init_attr)
1231 return qp->device->query_qp ?
1232 qp->device->query_qp(qp->real_qp, qp_attr, qp_attr_mask, qp_init_attr) :
1235 EXPORT_SYMBOL(ib_query_qp);
1237 int ib_close_qp(struct ib_qp *qp)
1239 struct ib_qp *real_qp;
1240 unsigned long flags;
1242 real_qp = qp->real_qp;
1246 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1247 list_del(&qp->open_list);
1248 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1250 atomic_dec(&real_qp->usecnt);
1255 EXPORT_SYMBOL(ib_close_qp);
1257 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1259 struct ib_xrcd *xrcd;
1260 struct ib_qp *real_qp;
1263 real_qp = qp->real_qp;
1264 xrcd = real_qp->xrcd;
1266 mutex_lock(&xrcd->tgt_qp_mutex);
1268 if (atomic_read(&real_qp->usecnt) == 0)
1269 list_del(&real_qp->xrcd_list);
1272 mutex_unlock(&xrcd->tgt_qp_mutex);
1275 ret = ib_destroy_qp(real_qp);
1277 atomic_dec(&xrcd->usecnt);
1279 __ib_insert_xrcd_qp(xrcd, real_qp);
1285 int ib_destroy_qp(struct ib_qp *qp)
1288 struct ib_cq *scq, *rcq;
1290 struct ib_rwq_ind_table *ind_tbl;
1293 if (atomic_read(&qp->usecnt))
1296 if (qp->real_qp != qp)
1297 return __ib_destroy_shared_qp(qp);
1303 ind_tbl = qp->rwq_ind_tbl;
1305 ret = qp->device->destroy_qp(qp);
1308 atomic_dec(&pd->usecnt);
1310 atomic_dec(&scq->usecnt);
1312 atomic_dec(&rcq->usecnt);
1314 atomic_dec(&srq->usecnt);
1316 atomic_dec(&ind_tbl->usecnt);
1321 EXPORT_SYMBOL(ib_destroy_qp);
1323 /* Completion queues */
1325 struct ib_cq *ib_create_cq(struct ib_device *device,
1326 ib_comp_handler comp_handler,
1327 void (*event_handler)(struct ib_event *, void *),
1329 const struct ib_cq_init_attr *cq_attr)
1333 cq = device->create_cq(device, cq_attr, NULL, NULL);
1336 cq->device = device;
1338 cq->comp_handler = comp_handler;
1339 cq->event_handler = event_handler;
1340 cq->cq_context = cq_context;
1341 atomic_set(&cq->usecnt, 0);
1346 EXPORT_SYMBOL(ib_create_cq);
1348 int ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1350 return cq->device->modify_cq ?
1351 cq->device->modify_cq(cq, cq_count, cq_period) : -ENOSYS;
1353 EXPORT_SYMBOL(ib_modify_cq);
1355 int ib_destroy_cq(struct ib_cq *cq)
1357 if (atomic_read(&cq->usecnt))
1360 return cq->device->destroy_cq(cq);
1362 EXPORT_SYMBOL(ib_destroy_cq);
1364 int ib_resize_cq(struct ib_cq *cq, int cqe)
1366 return cq->device->resize_cq ?
1367 cq->device->resize_cq(cq, cqe, NULL) : -ENOSYS;
1369 EXPORT_SYMBOL(ib_resize_cq);
1371 /* Memory regions */
1373 int ib_dereg_mr(struct ib_mr *mr)
1375 struct ib_pd *pd = mr->pd;
1378 ret = mr->device->dereg_mr(mr);
1380 atomic_dec(&pd->usecnt);
1384 EXPORT_SYMBOL(ib_dereg_mr);
1387 * ib_alloc_mr() - Allocates a memory region
1388 * @pd: protection domain associated with the region
1389 * @mr_type: memory region type
1390 * @max_num_sg: maximum sg entries available for registration.
1393 * Memory registeration page/sg lists must not exceed max_num_sg.
1394 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
1395 * max_num_sg * used_page_size.
1398 struct ib_mr *ib_alloc_mr(struct ib_pd *pd,
1399 enum ib_mr_type mr_type,
1404 if (!pd->device->alloc_mr)
1405 return ERR_PTR(-ENOSYS);
1407 mr = pd->device->alloc_mr(pd, mr_type, max_num_sg);
1409 mr->device = pd->device;
1412 atomic_inc(&pd->usecnt);
1413 mr->need_inval = false;
1418 EXPORT_SYMBOL(ib_alloc_mr);
1420 /* "Fast" memory regions */
1422 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
1423 int mr_access_flags,
1424 struct ib_fmr_attr *fmr_attr)
1428 if (!pd->device->alloc_fmr)
1429 return ERR_PTR(-ENOSYS);
1431 fmr = pd->device->alloc_fmr(pd, mr_access_flags, fmr_attr);
1433 fmr->device = pd->device;
1435 atomic_inc(&pd->usecnt);
1440 EXPORT_SYMBOL(ib_alloc_fmr);
1442 int ib_unmap_fmr(struct list_head *fmr_list)
1446 if (list_empty(fmr_list))
1449 fmr = list_entry(fmr_list->next, struct ib_fmr, list);
1450 return fmr->device->unmap_fmr(fmr_list);
1452 EXPORT_SYMBOL(ib_unmap_fmr);
1454 int ib_dealloc_fmr(struct ib_fmr *fmr)
1460 ret = fmr->device->dealloc_fmr(fmr);
1462 atomic_dec(&pd->usecnt);
1466 EXPORT_SYMBOL(ib_dealloc_fmr);
1468 /* Multicast groups */
1470 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1474 if (!qp->device->attach_mcast)
1476 if (gid->raw[0] != 0xff || qp->qp_type != IB_QPT_UD)
1479 ret = qp->device->attach_mcast(qp, gid, lid);
1481 atomic_inc(&qp->usecnt);
1484 EXPORT_SYMBOL(ib_attach_mcast);
1486 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1490 if (!qp->device->detach_mcast)
1492 if (gid->raw[0] != 0xff || qp->qp_type != IB_QPT_UD)
1495 ret = qp->device->detach_mcast(qp, gid, lid);
1497 atomic_dec(&qp->usecnt);
1500 EXPORT_SYMBOL(ib_detach_mcast);
1502 struct ib_xrcd *ib_alloc_xrcd(struct ib_device *device)
1504 struct ib_xrcd *xrcd;
1506 if (!device->alloc_xrcd)
1507 return ERR_PTR(-ENOSYS);
1509 xrcd = device->alloc_xrcd(device, NULL, NULL);
1510 if (!IS_ERR(xrcd)) {
1511 xrcd->device = device;
1513 atomic_set(&xrcd->usecnt, 0);
1514 mutex_init(&xrcd->tgt_qp_mutex);
1515 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
1520 EXPORT_SYMBOL(ib_alloc_xrcd);
1522 int ib_dealloc_xrcd(struct ib_xrcd *xrcd)
1527 if (atomic_read(&xrcd->usecnt))
1530 while (!list_empty(&xrcd->tgt_qp_list)) {
1531 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
1532 ret = ib_destroy_qp(qp);
1537 return xrcd->device->dealloc_xrcd(xrcd);
1539 EXPORT_SYMBOL(ib_dealloc_xrcd);
1542 * ib_create_wq - Creates a WQ associated with the specified protection
1544 * @pd: The protection domain associated with the WQ.
1545 * @wq_init_attr: A list of initial attributes required to create the
1546 * WQ. If WQ creation succeeds, then the attributes are updated to
1547 * the actual capabilities of the created WQ.
1549 * wq_init_attr->max_wr and wq_init_attr->max_sge determine
1550 * the requested size of the WQ, and set to the actual values allocated
1552 * If ib_create_wq() succeeds, then max_wr and max_sge will always be
1553 * at least as large as the requested values.
1555 struct ib_wq *ib_create_wq(struct ib_pd *pd,
1556 struct ib_wq_init_attr *wq_attr)
1560 if (!pd->device->create_wq)
1561 return ERR_PTR(-ENOSYS);
1563 wq = pd->device->create_wq(pd, wq_attr, NULL);
1565 wq->event_handler = wq_attr->event_handler;
1566 wq->wq_context = wq_attr->wq_context;
1567 wq->wq_type = wq_attr->wq_type;
1568 wq->cq = wq_attr->cq;
1569 wq->device = pd->device;
1572 atomic_inc(&pd->usecnt);
1573 atomic_inc(&wq_attr->cq->usecnt);
1574 atomic_set(&wq->usecnt, 0);
1578 EXPORT_SYMBOL(ib_create_wq);
1581 * ib_destroy_wq - Destroys the specified WQ.
1582 * @wq: The WQ to destroy.
1584 int ib_destroy_wq(struct ib_wq *wq)
1587 struct ib_cq *cq = wq->cq;
1588 struct ib_pd *pd = wq->pd;
1590 if (atomic_read(&wq->usecnt))
1593 err = wq->device->destroy_wq(wq);
1595 atomic_dec(&pd->usecnt);
1596 atomic_dec(&cq->usecnt);
1600 EXPORT_SYMBOL(ib_destroy_wq);
1603 * ib_modify_wq - Modifies the specified WQ.
1604 * @wq: The WQ to modify.
1605 * @wq_attr: On input, specifies the WQ attributes to modify.
1606 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
1607 * are being modified.
1608 * On output, the current values of selected WQ attributes are returned.
1610 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
1615 if (!wq->device->modify_wq)
1618 err = wq->device->modify_wq(wq, wq_attr, wq_attr_mask, NULL);
1621 EXPORT_SYMBOL(ib_modify_wq);
1624 * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
1625 * @device: The device on which to create the rwq indirection table.
1626 * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
1627 * create the Indirection Table.
1629 * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
1630 * than the created ib_rwq_ind_table object and the caller is responsible
1631 * for its memory allocation/free.
1633 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
1634 struct ib_rwq_ind_table_init_attr *init_attr)
1636 struct ib_rwq_ind_table *rwq_ind_table;
1640 if (!device->create_rwq_ind_table)
1641 return ERR_PTR(-ENOSYS);
1643 table_size = (1 << init_attr->log_ind_tbl_size);
1644 rwq_ind_table = device->create_rwq_ind_table(device,
1646 if (IS_ERR(rwq_ind_table))
1647 return rwq_ind_table;
1649 rwq_ind_table->ind_tbl = init_attr->ind_tbl;
1650 rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
1651 rwq_ind_table->device = device;
1652 rwq_ind_table->uobject = NULL;
1653 atomic_set(&rwq_ind_table->usecnt, 0);
1655 for (i = 0; i < table_size; i++)
1656 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
1658 return rwq_ind_table;
1660 EXPORT_SYMBOL(ib_create_rwq_ind_table);
1663 * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
1664 * @wq_ind_table: The Indirection Table to destroy.
1666 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
1669 u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
1670 struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
1672 if (atomic_read(&rwq_ind_table->usecnt))
1675 err = rwq_ind_table->device->destroy_rwq_ind_table(rwq_ind_table);
1677 for (i = 0; i < table_size; i++)
1678 atomic_dec(&ind_tbl[i]->usecnt);
1683 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
1685 struct ib_flow *ib_create_flow(struct ib_qp *qp,
1686 struct ib_flow_attr *flow_attr,
1689 struct ib_flow *flow_id;
1690 if (!qp->device->create_flow)
1691 return ERR_PTR(-ENOSYS);
1693 flow_id = qp->device->create_flow(qp, flow_attr, domain);
1694 if (!IS_ERR(flow_id))
1695 atomic_inc(&qp->usecnt);
1698 EXPORT_SYMBOL(ib_create_flow);
1700 int ib_destroy_flow(struct ib_flow *flow_id)
1703 struct ib_qp *qp = flow_id->qp;
1705 err = qp->device->destroy_flow(flow_id);
1707 atomic_dec(&qp->usecnt);
1710 EXPORT_SYMBOL(ib_destroy_flow);
1712 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
1713 struct ib_mr_status *mr_status)
1715 return mr->device->check_mr_status ?
1716 mr->device->check_mr_status(mr, check_mask, mr_status) : -ENOSYS;
1718 EXPORT_SYMBOL(ib_check_mr_status);
1720 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
1723 if (!device->set_vf_link_state)
1726 return device->set_vf_link_state(device, vf, port, state);
1728 EXPORT_SYMBOL(ib_set_vf_link_state);
1730 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
1731 struct ifla_vf_info *info)
1733 if (!device->get_vf_config)
1736 return device->get_vf_config(device, vf, port, info);
1738 EXPORT_SYMBOL(ib_get_vf_config);
1740 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
1741 struct ifla_vf_stats *stats)
1743 if (!device->get_vf_stats)
1746 return device->get_vf_stats(device, vf, port, stats);
1748 EXPORT_SYMBOL(ib_get_vf_stats);
1750 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
1753 if (!device->set_vf_guid)
1756 return device->set_vf_guid(device, vf, port, guid, type);
1758 EXPORT_SYMBOL(ib_set_vf_guid);
1761 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
1762 * and set it the memory region.
1763 * @mr: memory region
1764 * @sg: dma mapped scatterlist
1765 * @sg_nents: number of entries in sg
1766 * @sg_offset: offset in bytes into sg
1767 * @page_size: page vector desired page size
1770 * - The first sg element is allowed to have an offset.
1771 * - Each sg element must either be aligned to page_size or virtually
1772 * contiguous to the previous element. In case an sg element has a
1773 * non-contiguous offset, the mapping prefix will not include it.
1774 * - The last sg element is allowed to have length less than page_size.
1775 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
1776 * then only max_num_sg entries will be mapped.
1777 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
1778 * constraints holds and the page_size argument is ignored.
1780 * Returns the number of sg elements that were mapped to the memory region.
1782 * After this completes successfully, the memory region
1783 * is ready for registration.
1785 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
1786 unsigned int *sg_offset, unsigned int page_size)
1788 if (unlikely(!mr->device->map_mr_sg))
1791 mr->page_size = page_size;
1793 return mr->device->map_mr_sg(mr, sg, sg_nents, sg_offset);
1795 EXPORT_SYMBOL(ib_map_mr_sg);
1798 * ib_sg_to_pages() - Convert the largest prefix of a sg list
1800 * @mr: memory region
1801 * @sgl: dma mapped scatterlist
1802 * @sg_nents: number of entries in sg
1803 * @sg_offset_p: IN: start offset in bytes into sg
1804 * OUT: offset in bytes for element n of the sg of the first
1805 * byte that has not been processed where n is the return
1806 * value of this function.
1807 * @set_page: driver page assignment function pointer
1809 * Core service helper for drivers to convert the largest
1810 * prefix of given sg list to a page vector. The sg list
1811 * prefix converted is the prefix that meet the requirements
1814 * Returns the number of sg elements that were assigned to
1817 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
1818 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
1820 struct scatterlist *sg;
1821 u64 last_end_dma_addr = 0;
1822 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
1823 unsigned int last_page_off = 0;
1824 u64 page_mask = ~((u64)mr->page_size - 1);
1827 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
1830 mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
1833 for_each_sg(sgl, sg, sg_nents, i) {
1834 u64 dma_addr = sg_dma_address(sg) + sg_offset;
1835 u64 prev_addr = dma_addr;
1836 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
1837 u64 end_dma_addr = dma_addr + dma_len;
1838 u64 page_addr = dma_addr & page_mask;
1841 * For the second and later elements, check whether either the
1842 * end of element i-1 or the start of element i is not aligned
1843 * on a page boundary.
1845 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
1846 /* Stop mapping if there is a gap. */
1847 if (last_end_dma_addr != dma_addr)
1851 * Coalesce this element with the last. If it is small
1852 * enough just update mr->length. Otherwise start
1853 * mapping from the next page.
1859 ret = set_page(mr, page_addr);
1860 if (unlikely(ret < 0)) {
1861 sg_offset = prev_addr - sg_dma_address(sg);
1862 mr->length += prev_addr - dma_addr;
1864 *sg_offset_p = sg_offset;
1865 return i || sg_offset ? i : ret;
1867 prev_addr = page_addr;
1869 page_addr += mr->page_size;
1870 } while (page_addr < end_dma_addr);
1872 mr->length += dma_len;
1873 last_end_dma_addr = end_dma_addr;
1874 last_page_off = end_dma_addr & ~page_mask;
1883 EXPORT_SYMBOL(ib_sg_to_pages);
1885 struct ib_drain_cqe {
1887 struct completion done;
1890 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
1892 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
1895 complete(&cqe->done);
1899 * Post a WR and block until its completion is reaped for the SQ.
1901 static void __ib_drain_sq(struct ib_qp *qp)
1903 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
1904 struct ib_drain_cqe sdrain;
1905 struct ib_send_wr swr = {}, *bad_swr;
1908 if (qp->send_cq->poll_ctx == IB_POLL_DIRECT) {
1909 WARN_ONCE(qp->send_cq->poll_ctx == IB_POLL_DIRECT,
1910 "IB_POLL_DIRECT poll_ctx not supported for drain\n");
1914 swr.wr_cqe = &sdrain.cqe;
1915 sdrain.cqe.done = ib_drain_qp_done;
1916 init_completion(&sdrain.done);
1918 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
1920 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
1924 ret = ib_post_send(qp, &swr, &bad_swr);
1926 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
1930 wait_for_completion(&sdrain.done);
1934 * Post a WR and block until its completion is reaped for the RQ.
1936 static void __ib_drain_rq(struct ib_qp *qp)
1938 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
1939 struct ib_drain_cqe rdrain;
1940 struct ib_recv_wr rwr = {}, *bad_rwr;
1943 if (qp->recv_cq->poll_ctx == IB_POLL_DIRECT) {
1944 WARN_ONCE(qp->recv_cq->poll_ctx == IB_POLL_DIRECT,
1945 "IB_POLL_DIRECT poll_ctx not supported for drain\n");
1949 rwr.wr_cqe = &rdrain.cqe;
1950 rdrain.cqe.done = ib_drain_qp_done;
1951 init_completion(&rdrain.done);
1953 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
1955 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
1959 ret = ib_post_recv(qp, &rwr, &bad_rwr);
1961 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
1965 wait_for_completion(&rdrain.done);
1969 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
1971 * @qp: queue pair to drain
1973 * If the device has a provider-specific drain function, then
1974 * call that. Otherwise call the generic drain function
1979 * ensure there is room in the CQ and SQ for the drain work request and
1982 * allocate the CQ using ib_alloc_cq() and the CQ poll context cannot be
1985 * ensure that there are no other contexts that are posting WRs concurrently.
1986 * Otherwise the drain is not guaranteed.
1988 void ib_drain_sq(struct ib_qp *qp)
1990 if (qp->device->drain_sq)
1991 qp->device->drain_sq(qp);
1995 EXPORT_SYMBOL(ib_drain_sq);
1998 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2000 * @qp: queue pair to drain
2002 * If the device has a provider-specific drain function, then
2003 * call that. Otherwise call the generic drain function
2008 * ensure there is room in the CQ and RQ for the drain work request and
2011 * allocate the CQ using ib_alloc_cq() and the CQ poll context cannot be
2014 * ensure that there are no other contexts that are posting WRs concurrently.
2015 * Otherwise the drain is not guaranteed.
2017 void ib_drain_rq(struct ib_qp *qp)
2019 if (qp->device->drain_rq)
2020 qp->device->drain_rq(qp);
2024 EXPORT_SYMBOL(ib_drain_rq);
2027 * ib_drain_qp() - Block until all CQEs have been consumed by the
2028 * application on both the RQ and SQ.
2029 * @qp: queue pair to drain
2033 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2036 * allocate the CQs using ib_alloc_cq() and the CQ poll context cannot be
2039 * ensure that there are no other contexts that are posting WRs concurrently.
2040 * Otherwise the drain is not guaranteed.
2042 void ib_drain_qp(struct ib_qp *qp)
2048 EXPORT_SYMBOL(ib_drain_qp);