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>
48 #include <rdma/ib_verbs.h>
49 #include <rdma/ib_cache.h>
50 #include <rdma/ib_addr.h>
53 #include "core_priv.h"
55 static const char * const ib_events[] = {
56 [IB_EVENT_CQ_ERR] = "CQ error",
57 [IB_EVENT_QP_FATAL] = "QP fatal error",
58 [IB_EVENT_QP_REQ_ERR] = "QP request error",
59 [IB_EVENT_QP_ACCESS_ERR] = "QP access error",
60 [IB_EVENT_COMM_EST] = "communication established",
61 [IB_EVENT_SQ_DRAINED] = "send queue drained",
62 [IB_EVENT_PATH_MIG] = "path migration successful",
63 [IB_EVENT_PATH_MIG_ERR] = "path migration error",
64 [IB_EVENT_DEVICE_FATAL] = "device fatal error",
65 [IB_EVENT_PORT_ACTIVE] = "port active",
66 [IB_EVENT_PORT_ERR] = "port error",
67 [IB_EVENT_LID_CHANGE] = "LID change",
68 [IB_EVENT_PKEY_CHANGE] = "P_key change",
69 [IB_EVENT_SM_CHANGE] = "SM change",
70 [IB_EVENT_SRQ_ERR] = "SRQ error",
71 [IB_EVENT_SRQ_LIMIT_REACHED] = "SRQ limit reached",
72 [IB_EVENT_QP_LAST_WQE_REACHED] = "last WQE reached",
73 [IB_EVENT_CLIENT_REREGISTER] = "client reregister",
74 [IB_EVENT_GID_CHANGE] = "GID changed",
77 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
81 return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
82 ib_events[index] : "unrecognized event";
84 EXPORT_SYMBOL(ib_event_msg);
86 static const char * const wc_statuses[] = {
87 [IB_WC_SUCCESS] = "success",
88 [IB_WC_LOC_LEN_ERR] = "local length error",
89 [IB_WC_LOC_QP_OP_ERR] = "local QP operation error",
90 [IB_WC_LOC_EEC_OP_ERR] = "local EE context operation error",
91 [IB_WC_LOC_PROT_ERR] = "local protection error",
92 [IB_WC_WR_FLUSH_ERR] = "WR flushed",
93 [IB_WC_MW_BIND_ERR] = "memory management operation error",
94 [IB_WC_BAD_RESP_ERR] = "bad response error",
95 [IB_WC_LOC_ACCESS_ERR] = "local access error",
96 [IB_WC_REM_INV_REQ_ERR] = "invalid request error",
97 [IB_WC_REM_ACCESS_ERR] = "remote access error",
98 [IB_WC_REM_OP_ERR] = "remote operation error",
99 [IB_WC_RETRY_EXC_ERR] = "transport retry counter exceeded",
100 [IB_WC_RNR_RETRY_EXC_ERR] = "RNR retry counter exceeded",
101 [IB_WC_LOC_RDD_VIOL_ERR] = "local RDD violation error",
102 [IB_WC_REM_INV_RD_REQ_ERR] = "remote invalid RD request",
103 [IB_WC_REM_ABORT_ERR] = "operation aborted",
104 [IB_WC_INV_EECN_ERR] = "invalid EE context number",
105 [IB_WC_INV_EEC_STATE_ERR] = "invalid EE context state",
106 [IB_WC_FATAL_ERR] = "fatal error",
107 [IB_WC_RESP_TIMEOUT_ERR] = "response timeout error",
108 [IB_WC_GENERAL_ERR] = "general error",
111 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
113 size_t index = status;
115 return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
116 wc_statuses[index] : "unrecognized status";
118 EXPORT_SYMBOL(ib_wc_status_msg);
120 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
123 case IB_RATE_2_5_GBPS: return 1;
124 case IB_RATE_5_GBPS: return 2;
125 case IB_RATE_10_GBPS: return 4;
126 case IB_RATE_20_GBPS: return 8;
127 case IB_RATE_30_GBPS: return 12;
128 case IB_RATE_40_GBPS: return 16;
129 case IB_RATE_60_GBPS: return 24;
130 case IB_RATE_80_GBPS: return 32;
131 case IB_RATE_120_GBPS: return 48;
135 EXPORT_SYMBOL(ib_rate_to_mult);
137 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
140 case 1: return IB_RATE_2_5_GBPS;
141 case 2: return IB_RATE_5_GBPS;
142 case 4: return IB_RATE_10_GBPS;
143 case 8: return IB_RATE_20_GBPS;
144 case 12: return IB_RATE_30_GBPS;
145 case 16: return IB_RATE_40_GBPS;
146 case 24: return IB_RATE_60_GBPS;
147 case 32: return IB_RATE_80_GBPS;
148 case 48: return IB_RATE_120_GBPS;
149 default: return IB_RATE_PORT_CURRENT;
152 EXPORT_SYMBOL(mult_to_ib_rate);
154 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
157 case IB_RATE_2_5_GBPS: return 2500;
158 case IB_RATE_5_GBPS: return 5000;
159 case IB_RATE_10_GBPS: return 10000;
160 case IB_RATE_20_GBPS: return 20000;
161 case IB_RATE_30_GBPS: return 30000;
162 case IB_RATE_40_GBPS: return 40000;
163 case IB_RATE_60_GBPS: return 60000;
164 case IB_RATE_80_GBPS: return 80000;
165 case IB_RATE_120_GBPS: return 120000;
166 case IB_RATE_14_GBPS: return 14062;
167 case IB_RATE_56_GBPS: return 56250;
168 case IB_RATE_112_GBPS: return 112500;
169 case IB_RATE_168_GBPS: return 168750;
170 case IB_RATE_25_GBPS: return 25781;
171 case IB_RATE_100_GBPS: return 103125;
172 case IB_RATE_200_GBPS: return 206250;
173 case IB_RATE_300_GBPS: return 309375;
177 EXPORT_SYMBOL(ib_rate_to_mbps);
179 __attribute_const__ enum rdma_transport_type
180 rdma_node_get_transport(enum rdma_node_type node_type)
183 case RDMA_NODE_IB_CA:
184 case RDMA_NODE_IB_SWITCH:
185 case RDMA_NODE_IB_ROUTER:
186 return RDMA_TRANSPORT_IB;
188 return RDMA_TRANSPORT_IWARP;
189 case RDMA_NODE_USNIC:
190 return RDMA_TRANSPORT_USNIC;
191 case RDMA_NODE_USNIC_UDP:
192 return RDMA_TRANSPORT_USNIC_UDP;
198 EXPORT_SYMBOL(rdma_node_get_transport);
200 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
202 if (device->get_link_layer)
203 return device->get_link_layer(device, port_num);
205 switch (rdma_node_get_transport(device->node_type)) {
206 case RDMA_TRANSPORT_IB:
207 return IB_LINK_LAYER_INFINIBAND;
208 case RDMA_TRANSPORT_IWARP:
209 case RDMA_TRANSPORT_USNIC:
210 case RDMA_TRANSPORT_USNIC_UDP:
211 return IB_LINK_LAYER_ETHERNET;
213 return IB_LINK_LAYER_UNSPECIFIED;
216 EXPORT_SYMBOL(rdma_port_get_link_layer);
218 /* Protection domains */
221 * ib_alloc_pd - Allocates an unused protection domain.
222 * @device: The device on which to allocate the protection domain.
224 * A protection domain object provides an association between QPs, shared
225 * receive queues, address handles, memory regions, and memory windows.
227 * Every PD has a local_dma_lkey which can be used as the lkey value for local
230 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
234 int mr_access_flags = 0;
236 pd = device->alloc_pd(device, NULL, NULL);
242 pd->__internal_mr = NULL;
243 atomic_set(&pd->usecnt, 0);
246 if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
247 pd->local_dma_lkey = device->local_dma_lkey;
249 mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
251 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
252 pr_warn("%s: enabling unsafe global rkey\n", caller);
253 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
256 if (mr_access_flags) {
259 mr = pd->device->get_dma_mr(pd, mr_access_flags);
265 mr->device = pd->device;
268 mr->need_inval = false;
270 pd->__internal_mr = mr;
272 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
273 pd->local_dma_lkey = pd->__internal_mr->lkey;
275 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
276 pd->unsafe_global_rkey = pd->__internal_mr->rkey;
281 EXPORT_SYMBOL(__ib_alloc_pd);
284 * ib_dealloc_pd - Deallocates a protection domain.
285 * @pd: The protection domain to deallocate.
287 * It is an error to call this function while any resources in the pd still
288 * exist. The caller is responsible to synchronously destroy them and
289 * guarantee no new allocations will happen.
291 void ib_dealloc_pd(struct ib_pd *pd)
295 if (pd->__internal_mr) {
296 ret = pd->device->dereg_mr(pd->__internal_mr);
298 pd->__internal_mr = NULL;
301 /* uverbs manipulates usecnt with proper locking, while the kabi
302 requires the caller to guarantee we can't race here. */
303 WARN_ON(atomic_read(&pd->usecnt));
305 /* Making delalloc_pd a void return is a WIP, no driver should return
307 ret = pd->device->dealloc_pd(pd);
308 WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd");
310 EXPORT_SYMBOL(ib_dealloc_pd);
312 /* Address handles */
314 struct ib_ah *ib_create_ah(struct ib_pd *pd, struct ib_ah_attr *ah_attr)
318 ah = pd->device->create_ah(pd, ah_attr);
321 ah->device = pd->device;
324 atomic_inc(&pd->usecnt);
329 EXPORT_SYMBOL(ib_create_ah);
331 static int ib_get_header_version(const union rdma_network_hdr *hdr)
333 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
334 struct iphdr ip4h_checked;
335 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
337 /* If it's IPv6, the version must be 6, otherwise, the first
338 * 20 bytes (before the IPv4 header) are garbled.
340 if (ip6h->version != 6)
341 return (ip4h->version == 4) ? 4 : 0;
342 /* version may be 6 or 4 because the first 20 bytes could be garbled */
344 /* RoCE v2 requires no options, thus header length
351 * We can't write on scattered buffers so we need to copy to
354 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
355 ip4h_checked.check = 0;
356 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
357 /* if IPv4 header checksum is OK, believe it */
358 if (ip4h->check == ip4h_checked.check)
363 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
365 const struct ib_grh *grh)
369 if (rdma_protocol_ib(device, port_num))
370 return RDMA_NETWORK_IB;
372 grh_version = ib_get_header_version((union rdma_network_hdr *)grh);
374 if (grh_version == 4)
375 return RDMA_NETWORK_IPV4;
377 if (grh->next_hdr == IPPROTO_UDP)
378 return RDMA_NETWORK_IPV6;
380 return RDMA_NETWORK_ROCE_V1;
383 struct find_gid_index_context {
385 enum ib_gid_type gid_type;
388 static bool find_gid_index(const union ib_gid *gid,
389 const struct ib_gid_attr *gid_attr,
392 struct find_gid_index_context *ctx =
393 (struct find_gid_index_context *)context;
395 if (ctx->gid_type != gid_attr->gid_type)
398 if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) ||
399 (is_vlan_dev(gid_attr->ndev) &&
400 vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id))
406 static int get_sgid_index_from_eth(struct ib_device *device, u8 port_num,
407 u16 vlan_id, const union ib_gid *sgid,
408 enum ib_gid_type gid_type,
411 struct find_gid_index_context context = {.vlan_id = vlan_id,
412 .gid_type = gid_type};
414 return ib_find_gid_by_filter(device, sgid, port_num, find_gid_index,
415 &context, gid_index);
418 static int get_gids_from_rdma_hdr(union rdma_network_hdr *hdr,
419 enum rdma_network_type net_type,
420 union ib_gid *sgid, union ib_gid *dgid)
422 struct sockaddr_in src_in;
423 struct sockaddr_in dst_in;
424 __be32 src_saddr, dst_saddr;
429 if (net_type == RDMA_NETWORK_IPV4) {
430 memcpy(&src_in.sin_addr.s_addr,
431 &hdr->roce4grh.saddr, 4);
432 memcpy(&dst_in.sin_addr.s_addr,
433 &hdr->roce4grh.daddr, 4);
434 src_saddr = src_in.sin_addr.s_addr;
435 dst_saddr = dst_in.sin_addr.s_addr;
436 ipv6_addr_set_v4mapped(src_saddr,
437 (struct in6_addr *)sgid);
438 ipv6_addr_set_v4mapped(dst_saddr,
439 (struct in6_addr *)dgid);
441 } else if (net_type == RDMA_NETWORK_IPV6 ||
442 net_type == RDMA_NETWORK_IB) {
443 *dgid = hdr->ibgrh.dgid;
444 *sgid = hdr->ibgrh.sgid;
451 int ib_init_ah_from_wc(struct ib_device *device, u8 port_num,
452 const struct ib_wc *wc, const struct ib_grh *grh,
453 struct ib_ah_attr *ah_attr)
458 enum rdma_network_type net_type = RDMA_NETWORK_IB;
459 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
464 memset(ah_attr, 0, sizeof *ah_attr);
465 if (rdma_cap_eth_ah(device, port_num)) {
466 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
467 net_type = wc->network_hdr_type;
469 net_type = ib_get_net_type_by_grh(device, port_num, grh);
470 gid_type = ib_network_to_gid_type(net_type);
472 ret = get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
477 if (rdma_protocol_roce(device, port_num)) {
479 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
480 wc->vlan_id : 0xffff;
481 struct net_device *idev;
482 struct net_device *resolved_dev;
484 if (!(wc->wc_flags & IB_WC_GRH))
487 if (!device->get_netdev)
490 idev = device->get_netdev(device, port_num);
494 ret = rdma_addr_find_l2_eth_by_grh(&dgid, &sgid,
496 wc->wc_flags & IB_WC_WITH_VLAN ?
498 &if_index, &hoplimit);
504 resolved_dev = dev_get_by_index(&init_net, if_index);
505 if (resolved_dev->flags & IFF_LOOPBACK) {
506 dev_put(resolved_dev);
508 dev_hold(resolved_dev);
511 if (resolved_dev != idev && !rdma_is_upper_dev_rcu(idev,
516 dev_put(resolved_dev);
520 ret = get_sgid_index_from_eth(device, port_num, vlan_id,
521 &dgid, gid_type, &gid_index);
526 ah_attr->dlid = wc->slid;
527 ah_attr->sl = wc->sl;
528 ah_attr->src_path_bits = wc->dlid_path_bits;
529 ah_attr->port_num = port_num;
531 if (wc->wc_flags & IB_WC_GRH) {
532 ah_attr->ah_flags = IB_AH_GRH;
533 ah_attr->grh.dgid = sgid;
535 if (!rdma_cap_eth_ah(device, port_num)) {
536 if (dgid.global.interface_id != cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
537 ret = ib_find_cached_gid_by_port(device, &dgid,
548 ah_attr->grh.sgid_index = (u8) gid_index;
549 flow_class = be32_to_cpu(grh->version_tclass_flow);
550 ah_attr->grh.flow_label = flow_class & 0xFFFFF;
551 ah_attr->grh.hop_limit = hoplimit;
552 ah_attr->grh.traffic_class = (flow_class >> 20) & 0xFF;
556 EXPORT_SYMBOL(ib_init_ah_from_wc);
558 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
559 const struct ib_grh *grh, u8 port_num)
561 struct ib_ah_attr ah_attr;
564 ret = ib_init_ah_from_wc(pd->device, port_num, wc, grh, &ah_attr);
568 return ib_create_ah(pd, &ah_attr);
570 EXPORT_SYMBOL(ib_create_ah_from_wc);
572 int ib_modify_ah(struct ib_ah *ah, struct ib_ah_attr *ah_attr)
574 return ah->device->modify_ah ?
575 ah->device->modify_ah(ah, ah_attr) :
578 EXPORT_SYMBOL(ib_modify_ah);
580 int ib_query_ah(struct ib_ah *ah, struct ib_ah_attr *ah_attr)
582 return ah->device->query_ah ?
583 ah->device->query_ah(ah, ah_attr) :
586 EXPORT_SYMBOL(ib_query_ah);
588 int ib_destroy_ah(struct ib_ah *ah)
594 ret = ah->device->destroy_ah(ah);
596 atomic_dec(&pd->usecnt);
600 EXPORT_SYMBOL(ib_destroy_ah);
602 /* Shared receive queues */
604 struct ib_srq *ib_create_srq(struct ib_pd *pd,
605 struct ib_srq_init_attr *srq_init_attr)
609 if (!pd->device->create_srq)
610 return ERR_PTR(-ENOSYS);
612 srq = pd->device->create_srq(pd, srq_init_attr, NULL);
615 srq->device = pd->device;
618 srq->event_handler = srq_init_attr->event_handler;
619 srq->srq_context = srq_init_attr->srq_context;
620 srq->srq_type = srq_init_attr->srq_type;
621 if (srq->srq_type == IB_SRQT_XRC) {
622 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
623 srq->ext.xrc.cq = srq_init_attr->ext.xrc.cq;
624 atomic_inc(&srq->ext.xrc.xrcd->usecnt);
625 atomic_inc(&srq->ext.xrc.cq->usecnt);
627 atomic_inc(&pd->usecnt);
628 atomic_set(&srq->usecnt, 0);
633 EXPORT_SYMBOL(ib_create_srq);
635 int ib_modify_srq(struct ib_srq *srq,
636 struct ib_srq_attr *srq_attr,
637 enum ib_srq_attr_mask srq_attr_mask)
639 return srq->device->modify_srq ?
640 srq->device->modify_srq(srq, srq_attr, srq_attr_mask, NULL) :
643 EXPORT_SYMBOL(ib_modify_srq);
645 int ib_query_srq(struct ib_srq *srq,
646 struct ib_srq_attr *srq_attr)
648 return srq->device->query_srq ?
649 srq->device->query_srq(srq, srq_attr) : -ENOSYS;
651 EXPORT_SYMBOL(ib_query_srq);
653 int ib_destroy_srq(struct ib_srq *srq)
656 enum ib_srq_type srq_type;
657 struct ib_xrcd *uninitialized_var(xrcd);
658 struct ib_cq *uninitialized_var(cq);
661 if (atomic_read(&srq->usecnt))
665 srq_type = srq->srq_type;
666 if (srq_type == IB_SRQT_XRC) {
667 xrcd = srq->ext.xrc.xrcd;
668 cq = srq->ext.xrc.cq;
671 ret = srq->device->destroy_srq(srq);
673 atomic_dec(&pd->usecnt);
674 if (srq_type == IB_SRQT_XRC) {
675 atomic_dec(&xrcd->usecnt);
676 atomic_dec(&cq->usecnt);
682 EXPORT_SYMBOL(ib_destroy_srq);
686 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
688 struct ib_qp *qp = context;
691 spin_lock_irqsave(&qp->device->event_handler_lock, flags);
692 list_for_each_entry(event->element.qp, &qp->open_list, open_list)
693 if (event->element.qp->event_handler)
694 event->element.qp->event_handler(event, event->element.qp->qp_context);
695 spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
698 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
700 mutex_lock(&xrcd->tgt_qp_mutex);
701 list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
702 mutex_unlock(&xrcd->tgt_qp_mutex);
705 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
706 void (*event_handler)(struct ib_event *, void *),
712 qp = kzalloc(sizeof *qp, GFP_KERNEL);
714 return ERR_PTR(-ENOMEM);
716 qp->real_qp = real_qp;
717 atomic_inc(&real_qp->usecnt);
718 qp->device = real_qp->device;
719 qp->event_handler = event_handler;
720 qp->qp_context = qp_context;
721 qp->qp_num = real_qp->qp_num;
722 qp->qp_type = real_qp->qp_type;
724 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
725 list_add(&qp->open_list, &real_qp->open_list);
726 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
731 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
732 struct ib_qp_open_attr *qp_open_attr)
734 struct ib_qp *qp, *real_qp;
736 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
737 return ERR_PTR(-EINVAL);
739 qp = ERR_PTR(-EINVAL);
740 mutex_lock(&xrcd->tgt_qp_mutex);
741 list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
742 if (real_qp->qp_num == qp_open_attr->qp_num) {
743 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
744 qp_open_attr->qp_context);
748 mutex_unlock(&xrcd->tgt_qp_mutex);
751 EXPORT_SYMBOL(ib_open_qp);
753 static struct ib_qp *ib_create_xrc_qp(struct ib_qp *qp,
754 struct ib_qp_init_attr *qp_init_attr)
756 struct ib_qp *real_qp = qp;
758 qp->event_handler = __ib_shared_qp_event_handler;
761 qp->send_cq = qp->recv_cq = NULL;
763 qp->xrcd = qp_init_attr->xrcd;
764 atomic_inc(&qp_init_attr->xrcd->usecnt);
765 INIT_LIST_HEAD(&qp->open_list);
767 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
768 qp_init_attr->qp_context);
770 __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
772 real_qp->device->destroy_qp(real_qp);
776 struct ib_qp *ib_create_qp(struct ib_pd *pd,
777 struct ib_qp_init_attr *qp_init_attr)
779 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
783 if (qp_init_attr->rwq_ind_tbl &&
784 (qp_init_attr->recv_cq ||
785 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
786 qp_init_attr->cap.max_recv_sge))
787 return ERR_PTR(-EINVAL);
790 * If the callers is using the RDMA API calculate the resources
791 * needed for the RDMA READ/WRITE operations.
793 * Note that these callers need to pass in a port number.
795 if (qp_init_attr->cap.max_rdma_ctxs)
796 rdma_rw_init_qp(device, qp_init_attr);
798 qp = device->create_qp(pd, qp_init_attr, NULL);
805 qp->qp_type = qp_init_attr->qp_type;
806 qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
808 atomic_set(&qp->usecnt, 0);
810 spin_lock_init(&qp->mr_lock);
811 INIT_LIST_HEAD(&qp->rdma_mrs);
812 INIT_LIST_HEAD(&qp->sig_mrs);
814 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT)
815 return ib_create_xrc_qp(qp, qp_init_attr);
817 qp->event_handler = qp_init_attr->event_handler;
818 qp->qp_context = qp_init_attr->qp_context;
819 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
823 qp->recv_cq = qp_init_attr->recv_cq;
824 if (qp_init_attr->recv_cq)
825 atomic_inc(&qp_init_attr->recv_cq->usecnt);
826 qp->srq = qp_init_attr->srq;
828 atomic_inc(&qp_init_attr->srq->usecnt);
832 qp->send_cq = qp_init_attr->send_cq;
835 atomic_inc(&pd->usecnt);
836 if (qp_init_attr->send_cq)
837 atomic_inc(&qp_init_attr->send_cq->usecnt);
838 if (qp_init_attr->rwq_ind_tbl)
839 atomic_inc(&qp->rwq_ind_tbl->usecnt);
841 if (qp_init_attr->cap.max_rdma_ctxs) {
842 ret = rdma_rw_init_mrs(qp, qp_init_attr);
844 pr_err("failed to init MR pool ret= %d\n", ret);
851 * Note: all hw drivers guarantee that max_send_sge is lower than
852 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
853 * max_send_sge <= max_sge_rd.
855 qp->max_write_sge = qp_init_attr->cap.max_send_sge;
856 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
857 device->attrs.max_sge_rd);
861 EXPORT_SYMBOL(ib_create_qp);
863 static const struct {
865 enum ib_qp_attr_mask req_param[IB_QPT_MAX];
866 enum ib_qp_attr_mask opt_param[IB_QPT_MAX];
867 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
869 [IB_QPS_RESET] = { .valid = 1 },
873 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
876 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
877 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
880 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
883 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
886 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
889 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
891 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
897 [IB_QPS_RESET] = { .valid = 1 },
898 [IB_QPS_ERR] = { .valid = 1 },
902 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
905 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
908 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
911 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
914 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
917 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
919 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
926 [IB_QPT_UC] = (IB_QP_AV |
930 [IB_QPT_RC] = (IB_QP_AV |
934 IB_QP_MAX_DEST_RD_ATOMIC |
935 IB_QP_MIN_RNR_TIMER),
936 [IB_QPT_XRC_INI] = (IB_QP_AV |
940 [IB_QPT_XRC_TGT] = (IB_QP_AV |
944 IB_QP_MAX_DEST_RD_ATOMIC |
945 IB_QP_MIN_RNR_TIMER),
948 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
950 [IB_QPT_UC] = (IB_QP_ALT_PATH |
953 [IB_QPT_RC] = (IB_QP_ALT_PATH |
956 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH |
959 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH |
962 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
964 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
970 [IB_QPS_RESET] = { .valid = 1 },
971 [IB_QPS_ERR] = { .valid = 1 },
975 [IB_QPT_UD] = IB_QP_SQ_PSN,
976 [IB_QPT_UC] = IB_QP_SQ_PSN,
977 [IB_QPT_RC] = (IB_QP_TIMEOUT |
981 IB_QP_MAX_QP_RD_ATOMIC),
982 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT |
986 IB_QP_MAX_QP_RD_ATOMIC),
987 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT |
989 [IB_QPT_SMI] = IB_QP_SQ_PSN,
990 [IB_QPT_GSI] = IB_QP_SQ_PSN,
993 [IB_QPT_UD] = (IB_QP_CUR_STATE |
995 [IB_QPT_UC] = (IB_QP_CUR_STATE |
998 IB_QP_PATH_MIG_STATE),
999 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1001 IB_QP_ACCESS_FLAGS |
1002 IB_QP_MIN_RNR_TIMER |
1003 IB_QP_PATH_MIG_STATE),
1004 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1006 IB_QP_ACCESS_FLAGS |
1007 IB_QP_PATH_MIG_STATE),
1008 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1010 IB_QP_ACCESS_FLAGS |
1011 IB_QP_MIN_RNR_TIMER |
1012 IB_QP_PATH_MIG_STATE),
1013 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1015 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1021 [IB_QPS_RESET] = { .valid = 1 },
1022 [IB_QPS_ERR] = { .valid = 1 },
1026 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1028 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1029 IB_QP_ACCESS_FLAGS |
1031 IB_QP_PATH_MIG_STATE),
1032 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1033 IB_QP_ACCESS_FLAGS |
1035 IB_QP_PATH_MIG_STATE |
1036 IB_QP_MIN_RNR_TIMER),
1037 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1038 IB_QP_ACCESS_FLAGS |
1040 IB_QP_PATH_MIG_STATE),
1041 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1042 IB_QP_ACCESS_FLAGS |
1044 IB_QP_PATH_MIG_STATE |
1045 IB_QP_MIN_RNR_TIMER),
1046 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1048 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1055 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1056 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1057 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1058 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1059 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1060 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1061 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1066 [IB_QPS_RESET] = { .valid = 1 },
1067 [IB_QPS_ERR] = { .valid = 1 },
1071 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1073 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1075 IB_QP_ACCESS_FLAGS |
1076 IB_QP_PATH_MIG_STATE),
1077 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1079 IB_QP_ACCESS_FLAGS |
1080 IB_QP_MIN_RNR_TIMER |
1081 IB_QP_PATH_MIG_STATE),
1082 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1084 IB_QP_ACCESS_FLAGS |
1085 IB_QP_PATH_MIG_STATE),
1086 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1088 IB_QP_ACCESS_FLAGS |
1089 IB_QP_MIN_RNR_TIMER |
1090 IB_QP_PATH_MIG_STATE),
1091 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1093 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1100 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1102 [IB_QPT_UC] = (IB_QP_AV |
1104 IB_QP_ACCESS_FLAGS |
1106 IB_QP_PATH_MIG_STATE),
1107 [IB_QPT_RC] = (IB_QP_PORT |
1112 IB_QP_MAX_QP_RD_ATOMIC |
1113 IB_QP_MAX_DEST_RD_ATOMIC |
1115 IB_QP_ACCESS_FLAGS |
1117 IB_QP_MIN_RNR_TIMER |
1118 IB_QP_PATH_MIG_STATE),
1119 [IB_QPT_XRC_INI] = (IB_QP_PORT |
1124 IB_QP_MAX_QP_RD_ATOMIC |
1126 IB_QP_ACCESS_FLAGS |
1128 IB_QP_PATH_MIG_STATE),
1129 [IB_QPT_XRC_TGT] = (IB_QP_PORT |
1132 IB_QP_MAX_DEST_RD_ATOMIC |
1134 IB_QP_ACCESS_FLAGS |
1136 IB_QP_MIN_RNR_TIMER |
1137 IB_QP_PATH_MIG_STATE),
1138 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1140 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1146 [IB_QPS_RESET] = { .valid = 1 },
1147 [IB_QPS_ERR] = { .valid = 1 },
1151 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1153 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1154 IB_QP_ACCESS_FLAGS),
1155 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1157 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1163 [IB_QPS_RESET] = { .valid = 1 },
1164 [IB_QPS_ERR] = { .valid = 1 }
1168 int ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1169 enum ib_qp_type type, enum ib_qp_attr_mask mask,
1170 enum rdma_link_layer ll)
1172 enum ib_qp_attr_mask req_param, opt_param;
1174 if (cur_state < 0 || cur_state > IB_QPS_ERR ||
1175 next_state < 0 || next_state > IB_QPS_ERR)
1178 if (mask & IB_QP_CUR_STATE &&
1179 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1180 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1183 if (!qp_state_table[cur_state][next_state].valid)
1186 req_param = qp_state_table[cur_state][next_state].req_param[type];
1187 opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1189 if ((mask & req_param) != req_param)
1192 if (mask & ~(req_param | opt_param | IB_QP_STATE))
1197 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1199 int ib_resolve_eth_dmac(struct ib_qp *qp,
1200 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1204 if (*qp_attr_mask & IB_QP_AV) {
1205 if (qp_attr->ah_attr.port_num < rdma_start_port(qp->device) ||
1206 qp_attr->ah_attr.port_num > rdma_end_port(qp->device))
1209 if (!rdma_cap_eth_ah(qp->device, qp_attr->ah_attr.port_num))
1212 if (rdma_link_local_addr((struct in6_addr *)qp_attr->ah_attr.grh.dgid.raw)) {
1213 rdma_get_ll_mac((struct in6_addr *)qp_attr->ah_attr.grh.dgid.raw,
1214 qp_attr->ah_attr.dmac);
1217 struct ib_gid_attr sgid_attr;
1221 ret = ib_query_gid(qp->device,
1222 qp_attr->ah_attr.port_num,
1223 qp_attr->ah_attr.grh.sgid_index,
1226 if (ret || !sgid_attr.ndev) {
1232 ifindex = sgid_attr.ndev->ifindex;
1234 ret = rdma_addr_find_l2_eth_by_grh(&sgid,
1235 &qp_attr->ah_attr.grh.dgid,
1236 qp_attr->ah_attr.dmac,
1237 NULL, &ifindex, &hop_limit);
1239 dev_put(sgid_attr.ndev);
1241 qp_attr->ah_attr.grh.hop_limit = hop_limit;
1247 EXPORT_SYMBOL(ib_resolve_eth_dmac);
1250 int ib_modify_qp(struct ib_qp *qp,
1251 struct ib_qp_attr *qp_attr,
1256 ret = ib_resolve_eth_dmac(qp, qp_attr, &qp_attr_mask);
1260 return qp->device->modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1262 EXPORT_SYMBOL(ib_modify_qp);
1264 int ib_query_qp(struct ib_qp *qp,
1265 struct ib_qp_attr *qp_attr,
1267 struct ib_qp_init_attr *qp_init_attr)
1269 return qp->device->query_qp ?
1270 qp->device->query_qp(qp->real_qp, qp_attr, qp_attr_mask, qp_init_attr) :
1273 EXPORT_SYMBOL(ib_query_qp);
1275 int ib_close_qp(struct ib_qp *qp)
1277 struct ib_qp *real_qp;
1278 unsigned long flags;
1280 real_qp = qp->real_qp;
1284 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1285 list_del(&qp->open_list);
1286 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1288 atomic_dec(&real_qp->usecnt);
1293 EXPORT_SYMBOL(ib_close_qp);
1295 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1297 struct ib_xrcd *xrcd;
1298 struct ib_qp *real_qp;
1301 real_qp = qp->real_qp;
1302 xrcd = real_qp->xrcd;
1304 mutex_lock(&xrcd->tgt_qp_mutex);
1306 if (atomic_read(&real_qp->usecnt) == 0)
1307 list_del(&real_qp->xrcd_list);
1310 mutex_unlock(&xrcd->tgt_qp_mutex);
1313 ret = ib_destroy_qp(real_qp);
1315 atomic_dec(&xrcd->usecnt);
1317 __ib_insert_xrcd_qp(xrcd, real_qp);
1323 int ib_destroy_qp(struct ib_qp *qp)
1326 struct ib_cq *scq, *rcq;
1328 struct ib_rwq_ind_table *ind_tbl;
1331 WARN_ON_ONCE(qp->mrs_used > 0);
1333 if (atomic_read(&qp->usecnt))
1336 if (qp->real_qp != qp)
1337 return __ib_destroy_shared_qp(qp);
1343 ind_tbl = qp->rwq_ind_tbl;
1346 rdma_rw_cleanup_mrs(qp);
1348 ret = qp->device->destroy_qp(qp);
1351 atomic_dec(&pd->usecnt);
1353 atomic_dec(&scq->usecnt);
1355 atomic_dec(&rcq->usecnt);
1357 atomic_dec(&srq->usecnt);
1359 atomic_dec(&ind_tbl->usecnt);
1364 EXPORT_SYMBOL(ib_destroy_qp);
1366 /* Completion queues */
1368 struct ib_cq *ib_create_cq(struct ib_device *device,
1369 ib_comp_handler comp_handler,
1370 void (*event_handler)(struct ib_event *, void *),
1372 const struct ib_cq_init_attr *cq_attr)
1376 cq = device->create_cq(device, cq_attr, NULL, NULL);
1379 cq->device = device;
1381 cq->comp_handler = comp_handler;
1382 cq->event_handler = event_handler;
1383 cq->cq_context = cq_context;
1384 atomic_set(&cq->usecnt, 0);
1389 EXPORT_SYMBOL(ib_create_cq);
1391 int ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1393 return cq->device->modify_cq ?
1394 cq->device->modify_cq(cq, cq_count, cq_period) : -ENOSYS;
1396 EXPORT_SYMBOL(ib_modify_cq);
1398 int ib_destroy_cq(struct ib_cq *cq)
1400 if (atomic_read(&cq->usecnt))
1403 return cq->device->destroy_cq(cq);
1405 EXPORT_SYMBOL(ib_destroy_cq);
1407 int ib_resize_cq(struct ib_cq *cq, int cqe)
1409 return cq->device->resize_cq ?
1410 cq->device->resize_cq(cq, cqe, NULL) : -ENOSYS;
1412 EXPORT_SYMBOL(ib_resize_cq);
1414 /* Memory regions */
1416 int ib_dereg_mr(struct ib_mr *mr)
1418 struct ib_pd *pd = mr->pd;
1421 ret = mr->device->dereg_mr(mr);
1423 atomic_dec(&pd->usecnt);
1427 EXPORT_SYMBOL(ib_dereg_mr);
1430 * ib_alloc_mr() - Allocates a memory region
1431 * @pd: protection domain associated with the region
1432 * @mr_type: memory region type
1433 * @max_num_sg: maximum sg entries available for registration.
1436 * Memory registeration page/sg lists must not exceed max_num_sg.
1437 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
1438 * max_num_sg * used_page_size.
1441 struct ib_mr *ib_alloc_mr(struct ib_pd *pd,
1442 enum ib_mr_type mr_type,
1447 if (!pd->device->alloc_mr)
1448 return ERR_PTR(-ENOSYS);
1450 mr = pd->device->alloc_mr(pd, mr_type, max_num_sg);
1452 mr->device = pd->device;
1455 atomic_inc(&pd->usecnt);
1456 mr->need_inval = false;
1461 EXPORT_SYMBOL(ib_alloc_mr);
1463 /* "Fast" memory regions */
1465 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
1466 int mr_access_flags,
1467 struct ib_fmr_attr *fmr_attr)
1471 if (!pd->device->alloc_fmr)
1472 return ERR_PTR(-ENOSYS);
1474 fmr = pd->device->alloc_fmr(pd, mr_access_flags, fmr_attr);
1476 fmr->device = pd->device;
1478 atomic_inc(&pd->usecnt);
1483 EXPORT_SYMBOL(ib_alloc_fmr);
1485 int ib_unmap_fmr(struct list_head *fmr_list)
1489 if (list_empty(fmr_list))
1492 fmr = list_entry(fmr_list->next, struct ib_fmr, list);
1493 return fmr->device->unmap_fmr(fmr_list);
1495 EXPORT_SYMBOL(ib_unmap_fmr);
1497 int ib_dealloc_fmr(struct ib_fmr *fmr)
1503 ret = fmr->device->dealloc_fmr(fmr);
1505 atomic_dec(&pd->usecnt);
1509 EXPORT_SYMBOL(ib_dealloc_fmr);
1511 /* Multicast groups */
1513 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
1515 struct ib_qp_init_attr init_attr = {};
1516 struct ib_qp_attr attr = {};
1517 int num_eth_ports = 0;
1520 /* If QP state >= init, it is assigned to a port and we can check this
1523 if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
1524 if (attr.qp_state >= IB_QPS_INIT) {
1525 if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
1526 IB_LINK_LAYER_INFINIBAND)
1532 /* Can't get a quick answer, iterate over all ports */
1533 for (port = 0; port < qp->device->phys_port_cnt; port++)
1534 if (rdma_port_get_link_layer(qp->device, port) !=
1535 IB_LINK_LAYER_INFINIBAND)
1538 /* If we have at lease one Ethernet port, RoCE annex declares that
1539 * multicast LID should be ignored. We can't tell at this step if the
1540 * QP belongs to an IB or Ethernet port.
1545 /* If all the ports are IB, we can check according to IB spec. */
1547 return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
1548 lid == be16_to_cpu(IB_LID_PERMISSIVE));
1551 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1555 if (!qp->device->attach_mcast)
1557 if (gid->raw[0] != 0xff || qp->qp_type != IB_QPT_UD ||
1558 !is_valid_mcast_lid(qp, lid))
1561 ret = qp->device->attach_mcast(qp, gid, lid);
1563 atomic_inc(&qp->usecnt);
1566 EXPORT_SYMBOL(ib_attach_mcast);
1568 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1572 if (!qp->device->detach_mcast)
1574 if (gid->raw[0] != 0xff || qp->qp_type != IB_QPT_UD ||
1575 !is_valid_mcast_lid(qp, lid))
1578 ret = qp->device->detach_mcast(qp, gid, lid);
1580 atomic_dec(&qp->usecnt);
1583 EXPORT_SYMBOL(ib_detach_mcast);
1585 struct ib_xrcd *ib_alloc_xrcd(struct ib_device *device)
1587 struct ib_xrcd *xrcd;
1589 if (!device->alloc_xrcd)
1590 return ERR_PTR(-ENOSYS);
1592 xrcd = device->alloc_xrcd(device, NULL, NULL);
1593 if (!IS_ERR(xrcd)) {
1594 xrcd->device = device;
1596 atomic_set(&xrcd->usecnt, 0);
1597 mutex_init(&xrcd->tgt_qp_mutex);
1598 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
1603 EXPORT_SYMBOL(ib_alloc_xrcd);
1605 int ib_dealloc_xrcd(struct ib_xrcd *xrcd)
1610 if (atomic_read(&xrcd->usecnt))
1613 while (!list_empty(&xrcd->tgt_qp_list)) {
1614 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
1615 ret = ib_destroy_qp(qp);
1620 return xrcd->device->dealloc_xrcd(xrcd);
1622 EXPORT_SYMBOL(ib_dealloc_xrcd);
1625 * ib_create_wq - Creates a WQ associated with the specified protection
1627 * @pd: The protection domain associated with the WQ.
1628 * @wq_init_attr: A list of initial attributes required to create the
1629 * WQ. If WQ creation succeeds, then the attributes are updated to
1630 * the actual capabilities of the created WQ.
1632 * wq_init_attr->max_wr and wq_init_attr->max_sge determine
1633 * the requested size of the WQ, and set to the actual values allocated
1635 * If ib_create_wq() succeeds, then max_wr and max_sge will always be
1636 * at least as large as the requested values.
1638 struct ib_wq *ib_create_wq(struct ib_pd *pd,
1639 struct ib_wq_init_attr *wq_attr)
1643 if (!pd->device->create_wq)
1644 return ERR_PTR(-ENOSYS);
1646 wq = pd->device->create_wq(pd, wq_attr, NULL);
1648 wq->event_handler = wq_attr->event_handler;
1649 wq->wq_context = wq_attr->wq_context;
1650 wq->wq_type = wq_attr->wq_type;
1651 wq->cq = wq_attr->cq;
1652 wq->device = pd->device;
1655 atomic_inc(&pd->usecnt);
1656 atomic_inc(&wq_attr->cq->usecnt);
1657 atomic_set(&wq->usecnt, 0);
1661 EXPORT_SYMBOL(ib_create_wq);
1664 * ib_destroy_wq - Destroys the specified WQ.
1665 * @wq: The WQ to destroy.
1667 int ib_destroy_wq(struct ib_wq *wq)
1670 struct ib_cq *cq = wq->cq;
1671 struct ib_pd *pd = wq->pd;
1673 if (atomic_read(&wq->usecnt))
1676 err = wq->device->destroy_wq(wq);
1678 atomic_dec(&pd->usecnt);
1679 atomic_dec(&cq->usecnt);
1683 EXPORT_SYMBOL(ib_destroy_wq);
1686 * ib_modify_wq - Modifies the specified WQ.
1687 * @wq: The WQ to modify.
1688 * @wq_attr: On input, specifies the WQ attributes to modify.
1689 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
1690 * are being modified.
1691 * On output, the current values of selected WQ attributes are returned.
1693 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
1698 if (!wq->device->modify_wq)
1701 err = wq->device->modify_wq(wq, wq_attr, wq_attr_mask, NULL);
1704 EXPORT_SYMBOL(ib_modify_wq);
1707 * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
1708 * @device: The device on which to create the rwq indirection table.
1709 * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
1710 * create the Indirection Table.
1712 * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
1713 * than the created ib_rwq_ind_table object and the caller is responsible
1714 * for its memory allocation/free.
1716 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
1717 struct ib_rwq_ind_table_init_attr *init_attr)
1719 struct ib_rwq_ind_table *rwq_ind_table;
1723 if (!device->create_rwq_ind_table)
1724 return ERR_PTR(-ENOSYS);
1726 table_size = (1 << init_attr->log_ind_tbl_size);
1727 rwq_ind_table = device->create_rwq_ind_table(device,
1729 if (IS_ERR(rwq_ind_table))
1730 return rwq_ind_table;
1732 rwq_ind_table->ind_tbl = init_attr->ind_tbl;
1733 rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
1734 rwq_ind_table->device = device;
1735 rwq_ind_table->uobject = NULL;
1736 atomic_set(&rwq_ind_table->usecnt, 0);
1738 for (i = 0; i < table_size; i++)
1739 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
1741 return rwq_ind_table;
1743 EXPORT_SYMBOL(ib_create_rwq_ind_table);
1746 * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
1747 * @wq_ind_table: The Indirection Table to destroy.
1749 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
1752 u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
1753 struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
1755 if (atomic_read(&rwq_ind_table->usecnt))
1758 err = rwq_ind_table->device->destroy_rwq_ind_table(rwq_ind_table);
1760 for (i = 0; i < table_size; i++)
1761 atomic_dec(&ind_tbl[i]->usecnt);
1766 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
1768 struct ib_flow *ib_create_flow(struct ib_qp *qp,
1769 struct ib_flow_attr *flow_attr,
1772 struct ib_flow *flow_id;
1773 if (!qp->device->create_flow)
1774 return ERR_PTR(-ENOSYS);
1776 flow_id = qp->device->create_flow(qp, flow_attr, domain);
1777 if (!IS_ERR(flow_id))
1778 atomic_inc(&qp->usecnt);
1781 EXPORT_SYMBOL(ib_create_flow);
1783 int ib_destroy_flow(struct ib_flow *flow_id)
1786 struct ib_qp *qp = flow_id->qp;
1788 err = qp->device->destroy_flow(flow_id);
1790 atomic_dec(&qp->usecnt);
1793 EXPORT_SYMBOL(ib_destroy_flow);
1795 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
1796 struct ib_mr_status *mr_status)
1798 return mr->device->check_mr_status ?
1799 mr->device->check_mr_status(mr, check_mask, mr_status) : -ENOSYS;
1801 EXPORT_SYMBOL(ib_check_mr_status);
1803 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
1806 if (!device->set_vf_link_state)
1809 return device->set_vf_link_state(device, vf, port, state);
1811 EXPORT_SYMBOL(ib_set_vf_link_state);
1813 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
1814 struct ifla_vf_info *info)
1816 if (!device->get_vf_config)
1819 return device->get_vf_config(device, vf, port, info);
1821 EXPORT_SYMBOL(ib_get_vf_config);
1823 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
1824 struct ifla_vf_stats *stats)
1826 if (!device->get_vf_stats)
1829 return device->get_vf_stats(device, vf, port, stats);
1831 EXPORT_SYMBOL(ib_get_vf_stats);
1833 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
1836 if (!device->set_vf_guid)
1839 return device->set_vf_guid(device, vf, port, guid, type);
1841 EXPORT_SYMBOL(ib_set_vf_guid);
1844 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
1845 * and set it the memory region.
1846 * @mr: memory region
1847 * @sg: dma mapped scatterlist
1848 * @sg_nents: number of entries in sg
1849 * @sg_offset: offset in bytes into sg
1850 * @page_size: page vector desired page size
1853 * - The first sg element is allowed to have an offset.
1854 * - Each sg element must either be aligned to page_size or virtually
1855 * contiguous to the previous element. In case an sg element has a
1856 * non-contiguous offset, the mapping prefix will not include it.
1857 * - The last sg element is allowed to have length less than page_size.
1858 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
1859 * then only max_num_sg entries will be mapped.
1860 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
1861 * constraints holds and the page_size argument is ignored.
1863 * Returns the number of sg elements that were mapped to the memory region.
1865 * After this completes successfully, the memory region
1866 * is ready for registration.
1868 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
1869 unsigned int *sg_offset, unsigned int page_size)
1871 if (unlikely(!mr->device->map_mr_sg))
1874 mr->page_size = page_size;
1876 return mr->device->map_mr_sg(mr, sg, sg_nents, sg_offset);
1878 EXPORT_SYMBOL(ib_map_mr_sg);
1881 * ib_sg_to_pages() - Convert the largest prefix of a sg list
1883 * @mr: memory region
1884 * @sgl: dma mapped scatterlist
1885 * @sg_nents: number of entries in sg
1886 * @sg_offset_p: IN: start offset in bytes into sg
1887 * OUT: offset in bytes for element n of the sg of the first
1888 * byte that has not been processed where n is the return
1889 * value of this function.
1890 * @set_page: driver page assignment function pointer
1892 * Core service helper for drivers to convert the largest
1893 * prefix of given sg list to a page vector. The sg list
1894 * prefix converted is the prefix that meet the requirements
1897 * Returns the number of sg elements that were assigned to
1900 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
1901 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
1903 struct scatterlist *sg;
1904 u64 last_end_dma_addr = 0;
1905 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
1906 unsigned int last_page_off = 0;
1907 u64 page_mask = ~((u64)mr->page_size - 1);
1910 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
1913 mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
1916 for_each_sg(sgl, sg, sg_nents, i) {
1917 u64 dma_addr = sg_dma_address(sg) + sg_offset;
1918 u64 prev_addr = dma_addr;
1919 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
1920 u64 end_dma_addr = dma_addr + dma_len;
1921 u64 page_addr = dma_addr & page_mask;
1924 * For the second and later elements, check whether either the
1925 * end of element i-1 or the start of element i is not aligned
1926 * on a page boundary.
1928 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
1929 /* Stop mapping if there is a gap. */
1930 if (last_end_dma_addr != dma_addr)
1934 * Coalesce this element with the last. If it is small
1935 * enough just update mr->length. Otherwise start
1936 * mapping from the next page.
1942 ret = set_page(mr, page_addr);
1943 if (unlikely(ret < 0)) {
1944 sg_offset = prev_addr - sg_dma_address(sg);
1945 mr->length += prev_addr - dma_addr;
1947 *sg_offset_p = sg_offset;
1948 return i || sg_offset ? i : ret;
1950 prev_addr = page_addr;
1952 page_addr += mr->page_size;
1953 } while (page_addr < end_dma_addr);
1955 mr->length += dma_len;
1956 last_end_dma_addr = end_dma_addr;
1957 last_page_off = end_dma_addr & ~page_mask;
1966 EXPORT_SYMBOL(ib_sg_to_pages);
1968 struct ib_drain_cqe {
1970 struct completion done;
1973 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
1975 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
1978 complete(&cqe->done);
1982 * Post a WR and block until its completion is reaped for the SQ.
1984 static void __ib_drain_sq(struct ib_qp *qp)
1986 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
1987 struct ib_drain_cqe sdrain;
1988 struct ib_send_wr swr = {}, *bad_swr;
1991 if (qp->send_cq->poll_ctx == IB_POLL_DIRECT) {
1992 WARN_ONCE(qp->send_cq->poll_ctx == IB_POLL_DIRECT,
1993 "IB_POLL_DIRECT poll_ctx not supported for drain\n");
1997 swr.wr_cqe = &sdrain.cqe;
1998 sdrain.cqe.done = ib_drain_qp_done;
1999 init_completion(&sdrain.done);
2001 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2003 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2007 ret = ib_post_send(qp, &swr, &bad_swr);
2009 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2013 wait_for_completion(&sdrain.done);
2017 * Post a WR and block until its completion is reaped for the RQ.
2019 static void __ib_drain_rq(struct ib_qp *qp)
2021 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2022 struct ib_drain_cqe rdrain;
2023 struct ib_recv_wr rwr = {}, *bad_rwr;
2026 if (qp->recv_cq->poll_ctx == IB_POLL_DIRECT) {
2027 WARN_ONCE(qp->recv_cq->poll_ctx == IB_POLL_DIRECT,
2028 "IB_POLL_DIRECT poll_ctx not supported for drain\n");
2032 rwr.wr_cqe = &rdrain.cqe;
2033 rdrain.cqe.done = ib_drain_qp_done;
2034 init_completion(&rdrain.done);
2036 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2038 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2042 ret = ib_post_recv(qp, &rwr, &bad_rwr);
2044 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2048 wait_for_completion(&rdrain.done);
2052 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2054 * @qp: queue pair to drain
2056 * If the device has a provider-specific drain function, then
2057 * call that. Otherwise call the generic drain function
2062 * ensure there is room in the CQ and SQ for the drain work request and
2065 * allocate the CQ using ib_alloc_cq() and the CQ poll context cannot be
2068 * ensure that there are no other contexts that are posting WRs concurrently.
2069 * Otherwise the drain is not guaranteed.
2071 void ib_drain_sq(struct ib_qp *qp)
2073 if (qp->device->drain_sq)
2074 qp->device->drain_sq(qp);
2078 EXPORT_SYMBOL(ib_drain_sq);
2081 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2083 * @qp: queue pair to drain
2085 * If the device has a provider-specific drain function, then
2086 * call that. Otherwise call the generic drain function
2091 * ensure there is room in the CQ and RQ for the drain work request and
2094 * allocate the CQ using ib_alloc_cq() and the CQ poll context cannot be
2097 * ensure that there are no other contexts that are posting WRs concurrently.
2098 * Otherwise the drain is not guaranteed.
2100 void ib_drain_rq(struct ib_qp *qp)
2102 if (qp->device->drain_rq)
2103 qp->device->drain_rq(qp);
2107 EXPORT_SYMBOL(ib_drain_rq);
2110 * ib_drain_qp() - Block until all CQEs have been consumed by the
2111 * application on both the RQ and SQ.
2112 * @qp: queue pair to drain
2116 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2119 * allocate the CQs using ib_alloc_cq() and the CQ poll context cannot be
2122 * ensure that there are no other contexts that are posting WRs concurrently.
2123 * Otherwise the drain is not guaranteed.
2125 void ib_drain_qp(struct ib_qp *qp)
2131 EXPORT_SYMBOL(ib_drain_qp);