2 * Copyright (c) 2004 Mellanox Technologies Ltd. All rights reserved.
3 * Copyright (c) 2004 Infinicon Corporation. All rights reserved.
4 * Copyright (c) 2004 Intel Corporation. All rights reserved.
5 * Copyright (c) 2004 Topspin Corporation. All rights reserved.
6 * Copyright (c) 2004 Voltaire Corporation. All rights reserved.
7 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
8 * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved.
10 * This software is available to you under a choice of one of two
11 * licenses. You may choose to be licensed under the terms of the GNU
12 * General Public License (GPL) Version 2, available from the file
13 * COPYING in the main directory of this source tree, or the
14 * OpenIB.org BSD license below:
16 * Redistribution and use in source and binary forms, with or
17 * without modification, are permitted provided that the following
20 * - Redistributions of source code must retain the above
21 * copyright notice, this list of conditions and the following
24 * - Redistributions in binary form must reproduce the above
25 * copyright notice, this list of conditions and the following
26 * disclaimer in the documentation and/or other materials
27 * provided with the distribution.
29 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
30 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
31 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
32 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
33 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
34 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
35 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
39 #include <linux/errno.h>
40 #include <linux/err.h>
41 #include <linux/export.h>
42 #include <linux/string.h>
43 #include <linux/slab.h>
45 #include <linux/in6.h>
46 #include <net/addrconf.h>
47 #include <linux/security.h>
49 #include <rdma/ib_verbs.h>
50 #include <rdma/ib_cache.h>
51 #include <rdma/ib_addr.h>
54 #include "core_priv.h"
56 static const char * const ib_events[] = {
57 [IB_EVENT_CQ_ERR] = "CQ error",
58 [IB_EVENT_QP_FATAL] = "QP fatal error",
59 [IB_EVENT_QP_REQ_ERR] = "QP request error",
60 [IB_EVENT_QP_ACCESS_ERR] = "QP access error",
61 [IB_EVENT_COMM_EST] = "communication established",
62 [IB_EVENT_SQ_DRAINED] = "send queue drained",
63 [IB_EVENT_PATH_MIG] = "path migration successful",
64 [IB_EVENT_PATH_MIG_ERR] = "path migration error",
65 [IB_EVENT_DEVICE_FATAL] = "device fatal error",
66 [IB_EVENT_PORT_ACTIVE] = "port active",
67 [IB_EVENT_PORT_ERR] = "port error",
68 [IB_EVENT_LID_CHANGE] = "LID change",
69 [IB_EVENT_PKEY_CHANGE] = "P_key change",
70 [IB_EVENT_SM_CHANGE] = "SM change",
71 [IB_EVENT_SRQ_ERR] = "SRQ error",
72 [IB_EVENT_SRQ_LIMIT_REACHED] = "SRQ limit reached",
73 [IB_EVENT_QP_LAST_WQE_REACHED] = "last WQE reached",
74 [IB_EVENT_CLIENT_REREGISTER] = "client reregister",
75 [IB_EVENT_GID_CHANGE] = "GID changed",
78 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
82 return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
83 ib_events[index] : "unrecognized event";
85 EXPORT_SYMBOL(ib_event_msg);
87 static const char * const wc_statuses[] = {
88 [IB_WC_SUCCESS] = "success",
89 [IB_WC_LOC_LEN_ERR] = "local length error",
90 [IB_WC_LOC_QP_OP_ERR] = "local QP operation error",
91 [IB_WC_LOC_EEC_OP_ERR] = "local EE context operation error",
92 [IB_WC_LOC_PROT_ERR] = "local protection error",
93 [IB_WC_WR_FLUSH_ERR] = "WR flushed",
94 [IB_WC_MW_BIND_ERR] = "memory management operation error",
95 [IB_WC_BAD_RESP_ERR] = "bad response error",
96 [IB_WC_LOC_ACCESS_ERR] = "local access error",
97 [IB_WC_REM_INV_REQ_ERR] = "invalid request error",
98 [IB_WC_REM_ACCESS_ERR] = "remote access error",
99 [IB_WC_REM_OP_ERR] = "remote operation error",
100 [IB_WC_RETRY_EXC_ERR] = "transport retry counter exceeded",
101 [IB_WC_RNR_RETRY_EXC_ERR] = "RNR retry counter exceeded",
102 [IB_WC_LOC_RDD_VIOL_ERR] = "local RDD violation error",
103 [IB_WC_REM_INV_RD_REQ_ERR] = "remote invalid RD request",
104 [IB_WC_REM_ABORT_ERR] = "operation aborted",
105 [IB_WC_INV_EECN_ERR] = "invalid EE context number",
106 [IB_WC_INV_EEC_STATE_ERR] = "invalid EE context state",
107 [IB_WC_FATAL_ERR] = "fatal error",
108 [IB_WC_RESP_TIMEOUT_ERR] = "response timeout error",
109 [IB_WC_GENERAL_ERR] = "general error",
112 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
114 size_t index = status;
116 return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
117 wc_statuses[index] : "unrecognized status";
119 EXPORT_SYMBOL(ib_wc_status_msg);
121 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
124 case IB_RATE_2_5_GBPS: return 1;
125 case IB_RATE_5_GBPS: return 2;
126 case IB_RATE_10_GBPS: return 4;
127 case IB_RATE_20_GBPS: return 8;
128 case IB_RATE_30_GBPS: return 12;
129 case IB_RATE_40_GBPS: return 16;
130 case IB_RATE_60_GBPS: return 24;
131 case IB_RATE_80_GBPS: return 32;
132 case IB_RATE_120_GBPS: return 48;
136 EXPORT_SYMBOL(ib_rate_to_mult);
138 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
141 case 1: return IB_RATE_2_5_GBPS;
142 case 2: return IB_RATE_5_GBPS;
143 case 4: return IB_RATE_10_GBPS;
144 case 8: return IB_RATE_20_GBPS;
145 case 12: return IB_RATE_30_GBPS;
146 case 16: return IB_RATE_40_GBPS;
147 case 24: return IB_RATE_60_GBPS;
148 case 32: return IB_RATE_80_GBPS;
149 case 48: return IB_RATE_120_GBPS;
150 default: return IB_RATE_PORT_CURRENT;
153 EXPORT_SYMBOL(mult_to_ib_rate);
155 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
158 case IB_RATE_2_5_GBPS: return 2500;
159 case IB_RATE_5_GBPS: return 5000;
160 case IB_RATE_10_GBPS: return 10000;
161 case IB_RATE_20_GBPS: return 20000;
162 case IB_RATE_30_GBPS: return 30000;
163 case IB_RATE_40_GBPS: return 40000;
164 case IB_RATE_60_GBPS: return 60000;
165 case IB_RATE_80_GBPS: return 80000;
166 case IB_RATE_120_GBPS: return 120000;
167 case IB_RATE_14_GBPS: return 14062;
168 case IB_RATE_56_GBPS: return 56250;
169 case IB_RATE_112_GBPS: return 112500;
170 case IB_RATE_168_GBPS: return 168750;
171 case IB_RATE_25_GBPS: return 25781;
172 case IB_RATE_100_GBPS: return 103125;
173 case IB_RATE_200_GBPS: return 206250;
174 case IB_RATE_300_GBPS: return 309375;
178 EXPORT_SYMBOL(ib_rate_to_mbps);
180 __attribute_const__ enum rdma_transport_type
181 rdma_node_get_transport(enum rdma_node_type node_type)
184 case RDMA_NODE_IB_CA:
185 case RDMA_NODE_IB_SWITCH:
186 case RDMA_NODE_IB_ROUTER:
187 return RDMA_TRANSPORT_IB;
189 return RDMA_TRANSPORT_IWARP;
190 case RDMA_NODE_USNIC:
191 return RDMA_TRANSPORT_USNIC;
192 case RDMA_NODE_USNIC_UDP:
193 return RDMA_TRANSPORT_USNIC_UDP;
199 EXPORT_SYMBOL(rdma_node_get_transport);
201 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
203 if (device->get_link_layer)
204 return device->get_link_layer(device, port_num);
206 switch (rdma_node_get_transport(device->node_type)) {
207 case RDMA_TRANSPORT_IB:
208 return IB_LINK_LAYER_INFINIBAND;
209 case RDMA_TRANSPORT_IWARP:
210 case RDMA_TRANSPORT_USNIC:
211 case RDMA_TRANSPORT_USNIC_UDP:
212 return IB_LINK_LAYER_ETHERNET;
214 return IB_LINK_LAYER_UNSPECIFIED;
217 EXPORT_SYMBOL(rdma_port_get_link_layer);
219 /* Protection domains */
222 * ib_alloc_pd - Allocates an unused protection domain.
223 * @device: The device on which to allocate the protection domain.
225 * A protection domain object provides an association between QPs, shared
226 * receive queues, address handles, memory regions, and memory windows.
228 * Every PD has a local_dma_lkey which can be used as the lkey value for local
231 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
235 int mr_access_flags = 0;
237 pd = device->alloc_pd(device, NULL, NULL);
243 pd->__internal_mr = NULL;
244 atomic_set(&pd->usecnt, 0);
247 if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
248 pd->local_dma_lkey = device->local_dma_lkey;
250 mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
252 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
253 pr_warn("%s: enabling unsafe global rkey\n", caller);
254 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
257 if (mr_access_flags) {
260 mr = pd->device->get_dma_mr(pd, mr_access_flags);
266 mr->device = pd->device;
269 mr->need_inval = false;
271 pd->__internal_mr = mr;
273 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
274 pd->local_dma_lkey = pd->__internal_mr->lkey;
276 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
277 pd->unsafe_global_rkey = pd->__internal_mr->rkey;
282 EXPORT_SYMBOL(__ib_alloc_pd);
285 * ib_dealloc_pd - Deallocates a protection domain.
286 * @pd: The protection domain to deallocate.
288 * It is an error to call this function while any resources in the pd still
289 * exist. The caller is responsible to synchronously destroy them and
290 * guarantee no new allocations will happen.
292 void ib_dealloc_pd(struct ib_pd *pd)
296 if (pd->__internal_mr) {
297 ret = pd->device->dereg_mr(pd->__internal_mr);
299 pd->__internal_mr = NULL;
302 /* uverbs manipulates usecnt with proper locking, while the kabi
303 requires the caller to guarantee we can't race here. */
304 WARN_ON(atomic_read(&pd->usecnt));
306 /* Making delalloc_pd a void return is a WIP, no driver should return
308 ret = pd->device->dealloc_pd(pd);
309 WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd");
311 EXPORT_SYMBOL(ib_dealloc_pd);
313 /* Address handles */
315 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr)
319 ah = pd->device->create_ah(pd, ah_attr, NULL);
322 ah->device = pd->device;
325 ah->type = ah_attr->type;
326 atomic_inc(&pd->usecnt);
331 EXPORT_SYMBOL(rdma_create_ah);
333 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
335 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
336 struct iphdr ip4h_checked;
337 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
339 /* If it's IPv6, the version must be 6, otherwise, the first
340 * 20 bytes (before the IPv4 header) are garbled.
342 if (ip6h->version != 6)
343 return (ip4h->version == 4) ? 4 : 0;
344 /* version may be 6 or 4 because the first 20 bytes could be garbled */
346 /* RoCE v2 requires no options, thus header length
353 * We can't write on scattered buffers so we need to copy to
356 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
357 ip4h_checked.check = 0;
358 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
359 /* if IPv4 header checksum is OK, believe it */
360 if (ip4h->check == ip4h_checked.check)
364 EXPORT_SYMBOL(ib_get_rdma_header_version);
366 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
368 const struct ib_grh *grh)
372 if (rdma_protocol_ib(device, port_num))
373 return RDMA_NETWORK_IB;
375 grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
377 if (grh_version == 4)
378 return RDMA_NETWORK_IPV4;
380 if (grh->next_hdr == IPPROTO_UDP)
381 return RDMA_NETWORK_IPV6;
383 return RDMA_NETWORK_ROCE_V1;
386 struct find_gid_index_context {
388 enum ib_gid_type gid_type;
391 static bool find_gid_index(const union ib_gid *gid,
392 const struct ib_gid_attr *gid_attr,
395 struct find_gid_index_context *ctx =
396 (struct find_gid_index_context *)context;
398 if (ctx->gid_type != gid_attr->gid_type)
401 if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) ||
402 (is_vlan_dev(gid_attr->ndev) &&
403 vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id))
409 static int get_sgid_index_from_eth(struct ib_device *device, u8 port_num,
410 u16 vlan_id, const union ib_gid *sgid,
411 enum ib_gid_type gid_type,
414 struct find_gid_index_context context = {.vlan_id = vlan_id,
415 .gid_type = gid_type};
417 return ib_find_gid_by_filter(device, sgid, port_num, find_gid_index,
418 &context, gid_index);
421 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
422 enum rdma_network_type net_type,
423 union ib_gid *sgid, union ib_gid *dgid)
425 struct sockaddr_in src_in;
426 struct sockaddr_in dst_in;
427 __be32 src_saddr, dst_saddr;
432 if (net_type == RDMA_NETWORK_IPV4) {
433 memcpy(&src_in.sin_addr.s_addr,
434 &hdr->roce4grh.saddr, 4);
435 memcpy(&dst_in.sin_addr.s_addr,
436 &hdr->roce4grh.daddr, 4);
437 src_saddr = src_in.sin_addr.s_addr;
438 dst_saddr = dst_in.sin_addr.s_addr;
439 ipv6_addr_set_v4mapped(src_saddr,
440 (struct in6_addr *)sgid);
441 ipv6_addr_set_v4mapped(dst_saddr,
442 (struct in6_addr *)dgid);
444 } else if (net_type == RDMA_NETWORK_IPV6 ||
445 net_type == RDMA_NETWORK_IB) {
446 *dgid = hdr->ibgrh.dgid;
447 *sgid = hdr->ibgrh.sgid;
453 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
456 * This function creates ah from the incoming packet.
457 * Incoming packet has dgid of the receiver node on which this code is
458 * getting executed and, sgid contains the GID of the sender.
460 * When resolving mac address of destination, the arrived dgid is used
461 * as sgid and, sgid is used as dgid because sgid contains destinations
462 * GID whom to respond to.
464 * This is why when calling rdma_addr_find_l2_eth_by_grh() function, the
465 * position of arguments dgid and sgid do not match the order of the
468 int ib_init_ah_from_wc(struct ib_device *device, u8 port_num,
469 const struct ib_wc *wc, const struct ib_grh *grh,
470 struct rdma_ah_attr *ah_attr)
475 enum rdma_network_type net_type = RDMA_NETWORK_IB;
476 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
481 memset(ah_attr, 0, sizeof *ah_attr);
482 ah_attr->type = rdma_ah_find_type(device, port_num);
483 if (rdma_cap_eth_ah(device, port_num)) {
484 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
485 net_type = wc->network_hdr_type;
487 net_type = ib_get_net_type_by_grh(device, port_num, grh);
488 gid_type = ib_network_to_gid_type(net_type);
490 ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
495 if (rdma_protocol_roce(device, port_num)) {
497 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
498 wc->vlan_id : 0xffff;
499 struct net_device *idev;
500 struct net_device *resolved_dev;
502 if (!(wc->wc_flags & IB_WC_GRH))
505 if (!device->get_netdev)
508 idev = device->get_netdev(device, port_num);
512 ret = rdma_addr_find_l2_eth_by_grh(&dgid, &sgid,
514 wc->wc_flags & IB_WC_WITH_VLAN ?
516 &if_index, &hoplimit);
522 resolved_dev = dev_get_by_index(&init_net, if_index);
524 if (resolved_dev != idev && !rdma_is_upper_dev_rcu(idev,
529 dev_put(resolved_dev);
533 ret = get_sgid_index_from_eth(device, port_num, vlan_id,
534 &dgid, gid_type, &gid_index);
539 rdma_ah_set_dlid(ah_attr, wc->slid);
540 rdma_ah_set_sl(ah_attr, wc->sl);
541 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
542 rdma_ah_set_port_num(ah_attr, port_num);
544 if (wc->wc_flags & IB_WC_GRH) {
545 if (!rdma_cap_eth_ah(device, port_num)) {
546 if (dgid.global.interface_id != cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
547 ret = ib_find_cached_gid_by_port(device, &dgid,
558 flow_class = be32_to_cpu(grh->version_tclass_flow);
559 rdma_ah_set_grh(ah_attr, &sgid,
560 flow_class & 0xFFFFF,
561 (u8)gid_index, hoplimit,
562 (flow_class >> 20) & 0xFF);
567 EXPORT_SYMBOL(ib_init_ah_from_wc);
569 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
570 const struct ib_grh *grh, u8 port_num)
572 struct rdma_ah_attr ah_attr;
575 ret = ib_init_ah_from_wc(pd->device, port_num, wc, grh, &ah_attr);
579 return rdma_create_ah(pd, &ah_attr);
581 EXPORT_SYMBOL(ib_create_ah_from_wc);
583 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
585 if (ah->type != ah_attr->type)
588 return ah->device->modify_ah ?
589 ah->device->modify_ah(ah, ah_attr) :
592 EXPORT_SYMBOL(rdma_modify_ah);
594 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
596 return ah->device->query_ah ?
597 ah->device->query_ah(ah, ah_attr) :
600 EXPORT_SYMBOL(rdma_query_ah);
602 int rdma_destroy_ah(struct ib_ah *ah)
608 ret = ah->device->destroy_ah(ah);
610 atomic_dec(&pd->usecnt);
614 EXPORT_SYMBOL(rdma_destroy_ah);
616 /* Shared receive queues */
618 struct ib_srq *ib_create_srq(struct ib_pd *pd,
619 struct ib_srq_init_attr *srq_init_attr)
623 if (!pd->device->create_srq)
624 return ERR_PTR(-ENOSYS);
626 srq = pd->device->create_srq(pd, srq_init_attr, NULL);
629 srq->device = pd->device;
632 srq->event_handler = srq_init_attr->event_handler;
633 srq->srq_context = srq_init_attr->srq_context;
634 srq->srq_type = srq_init_attr->srq_type;
635 if (srq->srq_type == IB_SRQT_XRC) {
636 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
637 srq->ext.xrc.cq = srq_init_attr->ext.xrc.cq;
638 atomic_inc(&srq->ext.xrc.xrcd->usecnt);
639 atomic_inc(&srq->ext.xrc.cq->usecnt);
641 atomic_inc(&pd->usecnt);
642 atomic_set(&srq->usecnt, 0);
647 EXPORT_SYMBOL(ib_create_srq);
649 int ib_modify_srq(struct ib_srq *srq,
650 struct ib_srq_attr *srq_attr,
651 enum ib_srq_attr_mask srq_attr_mask)
653 return srq->device->modify_srq ?
654 srq->device->modify_srq(srq, srq_attr, srq_attr_mask, NULL) :
657 EXPORT_SYMBOL(ib_modify_srq);
659 int ib_query_srq(struct ib_srq *srq,
660 struct ib_srq_attr *srq_attr)
662 return srq->device->query_srq ?
663 srq->device->query_srq(srq, srq_attr) : -ENOSYS;
665 EXPORT_SYMBOL(ib_query_srq);
667 int ib_destroy_srq(struct ib_srq *srq)
670 enum ib_srq_type srq_type;
671 struct ib_xrcd *uninitialized_var(xrcd);
672 struct ib_cq *uninitialized_var(cq);
675 if (atomic_read(&srq->usecnt))
679 srq_type = srq->srq_type;
680 if (srq_type == IB_SRQT_XRC) {
681 xrcd = srq->ext.xrc.xrcd;
682 cq = srq->ext.xrc.cq;
685 ret = srq->device->destroy_srq(srq);
687 atomic_dec(&pd->usecnt);
688 if (srq_type == IB_SRQT_XRC) {
689 atomic_dec(&xrcd->usecnt);
690 atomic_dec(&cq->usecnt);
696 EXPORT_SYMBOL(ib_destroy_srq);
700 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
702 struct ib_qp *qp = context;
705 spin_lock_irqsave(&qp->device->event_handler_lock, flags);
706 list_for_each_entry(event->element.qp, &qp->open_list, open_list)
707 if (event->element.qp->event_handler)
708 event->element.qp->event_handler(event, event->element.qp->qp_context);
709 spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
712 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
714 mutex_lock(&xrcd->tgt_qp_mutex);
715 list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
716 mutex_unlock(&xrcd->tgt_qp_mutex);
719 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
720 void (*event_handler)(struct ib_event *, void *),
727 qp = kzalloc(sizeof *qp, GFP_KERNEL);
729 return ERR_PTR(-ENOMEM);
731 qp->real_qp = real_qp;
732 err = ib_open_shared_qp_security(qp, real_qp->device);
738 qp->real_qp = real_qp;
739 atomic_inc(&real_qp->usecnt);
740 qp->device = real_qp->device;
741 qp->event_handler = event_handler;
742 qp->qp_context = qp_context;
743 qp->qp_num = real_qp->qp_num;
744 qp->qp_type = real_qp->qp_type;
746 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
747 list_add(&qp->open_list, &real_qp->open_list);
748 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
753 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
754 struct ib_qp_open_attr *qp_open_attr)
756 struct ib_qp *qp, *real_qp;
758 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
759 return ERR_PTR(-EINVAL);
761 qp = ERR_PTR(-EINVAL);
762 mutex_lock(&xrcd->tgt_qp_mutex);
763 list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
764 if (real_qp->qp_num == qp_open_attr->qp_num) {
765 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
766 qp_open_attr->qp_context);
770 mutex_unlock(&xrcd->tgt_qp_mutex);
773 EXPORT_SYMBOL(ib_open_qp);
775 static struct ib_qp *ib_create_xrc_qp(struct ib_qp *qp,
776 struct ib_qp_init_attr *qp_init_attr)
778 struct ib_qp *real_qp = qp;
780 qp->event_handler = __ib_shared_qp_event_handler;
783 qp->send_cq = qp->recv_cq = NULL;
785 qp->xrcd = qp_init_attr->xrcd;
786 atomic_inc(&qp_init_attr->xrcd->usecnt);
787 INIT_LIST_HEAD(&qp->open_list);
789 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
790 qp_init_attr->qp_context);
792 __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
794 real_qp->device->destroy_qp(real_qp);
798 struct ib_qp *ib_create_qp(struct ib_pd *pd,
799 struct ib_qp_init_attr *qp_init_attr)
801 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
805 if (qp_init_attr->rwq_ind_tbl &&
806 (qp_init_attr->recv_cq ||
807 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
808 qp_init_attr->cap.max_recv_sge))
809 return ERR_PTR(-EINVAL);
812 * If the callers is using the RDMA API calculate the resources
813 * needed for the RDMA READ/WRITE operations.
815 * Note that these callers need to pass in a port number.
817 if (qp_init_attr->cap.max_rdma_ctxs)
818 rdma_rw_init_qp(device, qp_init_attr);
820 qp = device->create_qp(pd, qp_init_attr, NULL);
824 ret = ib_create_qp_security(qp, device);
833 qp->qp_type = qp_init_attr->qp_type;
834 qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
836 atomic_set(&qp->usecnt, 0);
838 spin_lock_init(&qp->mr_lock);
839 INIT_LIST_HEAD(&qp->rdma_mrs);
840 INIT_LIST_HEAD(&qp->sig_mrs);
842 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT)
843 return ib_create_xrc_qp(qp, qp_init_attr);
845 qp->event_handler = qp_init_attr->event_handler;
846 qp->qp_context = qp_init_attr->qp_context;
847 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
851 qp->recv_cq = qp_init_attr->recv_cq;
852 if (qp_init_attr->recv_cq)
853 atomic_inc(&qp_init_attr->recv_cq->usecnt);
854 qp->srq = qp_init_attr->srq;
856 atomic_inc(&qp_init_attr->srq->usecnt);
860 qp->send_cq = qp_init_attr->send_cq;
863 atomic_inc(&pd->usecnt);
864 if (qp_init_attr->send_cq)
865 atomic_inc(&qp_init_attr->send_cq->usecnt);
866 if (qp_init_attr->rwq_ind_tbl)
867 atomic_inc(&qp->rwq_ind_tbl->usecnt);
869 if (qp_init_attr->cap.max_rdma_ctxs) {
870 ret = rdma_rw_init_mrs(qp, qp_init_attr);
872 pr_err("failed to init MR pool ret= %d\n", ret);
879 * Note: all hw drivers guarantee that max_send_sge is lower than
880 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
881 * max_send_sge <= max_sge_rd.
883 qp->max_write_sge = qp_init_attr->cap.max_send_sge;
884 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
885 device->attrs.max_sge_rd);
889 EXPORT_SYMBOL(ib_create_qp);
891 static const struct {
893 enum ib_qp_attr_mask req_param[IB_QPT_MAX];
894 enum ib_qp_attr_mask opt_param[IB_QPT_MAX];
895 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
897 [IB_QPS_RESET] = { .valid = 1 },
901 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
904 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
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 |
925 [IB_QPS_RESET] = { .valid = 1 },
926 [IB_QPS_ERR] = { .valid = 1 },
930 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
933 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
936 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
939 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
942 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
945 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
947 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
954 [IB_QPT_UC] = (IB_QP_AV |
958 [IB_QPT_RC] = (IB_QP_AV |
962 IB_QP_MAX_DEST_RD_ATOMIC |
963 IB_QP_MIN_RNR_TIMER),
964 [IB_QPT_XRC_INI] = (IB_QP_AV |
968 [IB_QPT_XRC_TGT] = (IB_QP_AV |
972 IB_QP_MAX_DEST_RD_ATOMIC |
973 IB_QP_MIN_RNR_TIMER),
976 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
978 [IB_QPT_UC] = (IB_QP_ALT_PATH |
981 [IB_QPT_RC] = (IB_QP_ALT_PATH |
984 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH |
987 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH |
990 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
992 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
998 [IB_QPS_RESET] = { .valid = 1 },
999 [IB_QPS_ERR] = { .valid = 1 },
1003 [IB_QPT_UD] = IB_QP_SQ_PSN,
1004 [IB_QPT_UC] = IB_QP_SQ_PSN,
1005 [IB_QPT_RC] = (IB_QP_TIMEOUT |
1009 IB_QP_MAX_QP_RD_ATOMIC),
1010 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT |
1014 IB_QP_MAX_QP_RD_ATOMIC),
1015 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT |
1017 [IB_QPT_SMI] = IB_QP_SQ_PSN,
1018 [IB_QPT_GSI] = IB_QP_SQ_PSN,
1021 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1023 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1025 IB_QP_ACCESS_FLAGS |
1026 IB_QP_PATH_MIG_STATE),
1027 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1029 IB_QP_ACCESS_FLAGS |
1030 IB_QP_MIN_RNR_TIMER |
1031 IB_QP_PATH_MIG_STATE),
1032 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1034 IB_QP_ACCESS_FLAGS |
1035 IB_QP_PATH_MIG_STATE),
1036 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1038 IB_QP_ACCESS_FLAGS |
1039 IB_QP_MIN_RNR_TIMER |
1040 IB_QP_PATH_MIG_STATE),
1041 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1043 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1045 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1050 [IB_QPS_RESET] = { .valid = 1 },
1051 [IB_QPS_ERR] = { .valid = 1 },
1055 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1057 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1058 IB_QP_ACCESS_FLAGS |
1060 IB_QP_PATH_MIG_STATE),
1061 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1062 IB_QP_ACCESS_FLAGS |
1064 IB_QP_PATH_MIG_STATE |
1065 IB_QP_MIN_RNR_TIMER),
1066 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1067 IB_QP_ACCESS_FLAGS |
1069 IB_QP_PATH_MIG_STATE),
1070 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1071 IB_QP_ACCESS_FLAGS |
1073 IB_QP_PATH_MIG_STATE |
1074 IB_QP_MIN_RNR_TIMER),
1075 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1077 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1079 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1085 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1086 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1087 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1088 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1089 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1090 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1091 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1096 [IB_QPS_RESET] = { .valid = 1 },
1097 [IB_QPS_ERR] = { .valid = 1 },
1101 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1103 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1105 IB_QP_ACCESS_FLAGS |
1106 IB_QP_PATH_MIG_STATE),
1107 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1109 IB_QP_ACCESS_FLAGS |
1110 IB_QP_MIN_RNR_TIMER |
1111 IB_QP_PATH_MIG_STATE),
1112 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1114 IB_QP_ACCESS_FLAGS |
1115 IB_QP_PATH_MIG_STATE),
1116 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1118 IB_QP_ACCESS_FLAGS |
1119 IB_QP_MIN_RNR_TIMER |
1120 IB_QP_PATH_MIG_STATE),
1121 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1123 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1130 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1132 [IB_QPT_UC] = (IB_QP_AV |
1134 IB_QP_ACCESS_FLAGS |
1136 IB_QP_PATH_MIG_STATE),
1137 [IB_QPT_RC] = (IB_QP_PORT |
1142 IB_QP_MAX_QP_RD_ATOMIC |
1143 IB_QP_MAX_DEST_RD_ATOMIC |
1145 IB_QP_ACCESS_FLAGS |
1147 IB_QP_MIN_RNR_TIMER |
1148 IB_QP_PATH_MIG_STATE),
1149 [IB_QPT_XRC_INI] = (IB_QP_PORT |
1154 IB_QP_MAX_QP_RD_ATOMIC |
1156 IB_QP_ACCESS_FLAGS |
1158 IB_QP_PATH_MIG_STATE),
1159 [IB_QPT_XRC_TGT] = (IB_QP_PORT |
1162 IB_QP_MAX_DEST_RD_ATOMIC |
1164 IB_QP_ACCESS_FLAGS |
1166 IB_QP_MIN_RNR_TIMER |
1167 IB_QP_PATH_MIG_STATE),
1168 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1170 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1176 [IB_QPS_RESET] = { .valid = 1 },
1177 [IB_QPS_ERR] = { .valid = 1 },
1181 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1183 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1184 IB_QP_ACCESS_FLAGS),
1185 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1187 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1193 [IB_QPS_RESET] = { .valid = 1 },
1194 [IB_QPS_ERR] = { .valid = 1 }
1198 int ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1199 enum ib_qp_type type, enum ib_qp_attr_mask mask,
1200 enum rdma_link_layer ll)
1202 enum ib_qp_attr_mask req_param, opt_param;
1204 if (cur_state < 0 || cur_state > IB_QPS_ERR ||
1205 next_state < 0 || next_state > IB_QPS_ERR)
1208 if (mask & IB_QP_CUR_STATE &&
1209 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1210 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1213 if (!qp_state_table[cur_state][next_state].valid)
1216 req_param = qp_state_table[cur_state][next_state].req_param[type];
1217 opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1219 if ((mask & req_param) != req_param)
1222 if (mask & ~(req_param | opt_param | IB_QP_STATE))
1227 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1229 int ib_resolve_eth_dmac(struct ib_device *device,
1230 struct rdma_ah_attr *ah_attr)
1233 struct ib_global_route *grh;
1235 if (!rdma_is_port_valid(device, rdma_ah_get_port_num(ah_attr)))
1238 if (ah_attr->type != RDMA_AH_ATTR_TYPE_ROCE)
1241 grh = rdma_ah_retrieve_grh(ah_attr);
1243 if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw)) {
1244 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
1245 ah_attr->roce.dmac);
1248 struct ib_gid_attr sgid_attr;
1252 ret = ib_query_gid(device,
1253 rdma_ah_get_port_num(ah_attr),
1257 if (ret || !sgid_attr.ndev) {
1263 ifindex = sgid_attr.ndev->ifindex;
1266 rdma_addr_find_l2_eth_by_grh(&sgid, &grh->dgid,
1268 NULL, &ifindex, &hop_limit);
1270 dev_put(sgid_attr.ndev);
1272 grh->hop_limit = hop_limit;
1277 EXPORT_SYMBOL(ib_resolve_eth_dmac);
1280 * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1281 * @qp: The QP to modify.
1282 * @attr: On input, specifies the QP attributes to modify. On output,
1283 * the current values of selected QP attributes are returned.
1284 * @attr_mask: A bit-mask used to specify which attributes of the QP
1285 * are being modified.
1286 * @udata: pointer to user's input output buffer information
1287 * are being modified.
1288 * It returns 0 on success and returns appropriate error code on error.
1290 int ib_modify_qp_with_udata(struct ib_qp *qp, struct ib_qp_attr *attr,
1291 int attr_mask, struct ib_udata *udata)
1295 if (attr_mask & IB_QP_AV) {
1296 ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
1300 return ib_security_modify_qp(qp, attr, attr_mask, udata);
1302 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1304 int ib_modify_qp(struct ib_qp *qp,
1305 struct ib_qp_attr *qp_attr,
1308 return ib_modify_qp_with_udata(qp, qp_attr, qp_attr_mask, NULL);
1310 EXPORT_SYMBOL(ib_modify_qp);
1312 int ib_query_qp(struct ib_qp *qp,
1313 struct ib_qp_attr *qp_attr,
1315 struct ib_qp_init_attr *qp_init_attr)
1317 return qp->device->query_qp ?
1318 qp->device->query_qp(qp->real_qp, qp_attr, qp_attr_mask, qp_init_attr) :
1321 EXPORT_SYMBOL(ib_query_qp);
1323 int ib_close_qp(struct ib_qp *qp)
1325 struct ib_qp *real_qp;
1326 unsigned long flags;
1328 real_qp = qp->real_qp;
1332 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1333 list_del(&qp->open_list);
1334 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1336 atomic_dec(&real_qp->usecnt);
1337 ib_close_shared_qp_security(qp->qp_sec);
1342 EXPORT_SYMBOL(ib_close_qp);
1344 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1346 struct ib_xrcd *xrcd;
1347 struct ib_qp *real_qp;
1350 real_qp = qp->real_qp;
1351 xrcd = real_qp->xrcd;
1353 mutex_lock(&xrcd->tgt_qp_mutex);
1355 if (atomic_read(&real_qp->usecnt) == 0)
1356 list_del(&real_qp->xrcd_list);
1359 mutex_unlock(&xrcd->tgt_qp_mutex);
1362 ret = ib_destroy_qp(real_qp);
1364 atomic_dec(&xrcd->usecnt);
1366 __ib_insert_xrcd_qp(xrcd, real_qp);
1372 int ib_destroy_qp(struct ib_qp *qp)
1375 struct ib_cq *scq, *rcq;
1377 struct ib_rwq_ind_table *ind_tbl;
1378 struct ib_qp_security *sec;
1381 WARN_ON_ONCE(qp->mrs_used > 0);
1383 if (atomic_read(&qp->usecnt))
1386 if (qp->real_qp != qp)
1387 return __ib_destroy_shared_qp(qp);
1393 ind_tbl = qp->rwq_ind_tbl;
1396 ib_destroy_qp_security_begin(sec);
1399 rdma_rw_cleanup_mrs(qp);
1401 ret = qp->device->destroy_qp(qp);
1404 atomic_dec(&pd->usecnt);
1406 atomic_dec(&scq->usecnt);
1408 atomic_dec(&rcq->usecnt);
1410 atomic_dec(&srq->usecnt);
1412 atomic_dec(&ind_tbl->usecnt);
1414 ib_destroy_qp_security_end(sec);
1417 ib_destroy_qp_security_abort(sec);
1422 EXPORT_SYMBOL(ib_destroy_qp);
1424 /* Completion queues */
1426 struct ib_cq *ib_create_cq(struct ib_device *device,
1427 ib_comp_handler comp_handler,
1428 void (*event_handler)(struct ib_event *, void *),
1430 const struct ib_cq_init_attr *cq_attr)
1434 cq = device->create_cq(device, cq_attr, NULL, NULL);
1437 cq->device = device;
1439 cq->comp_handler = comp_handler;
1440 cq->event_handler = event_handler;
1441 cq->cq_context = cq_context;
1442 atomic_set(&cq->usecnt, 0);
1447 EXPORT_SYMBOL(ib_create_cq);
1449 int ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1451 return cq->device->modify_cq ?
1452 cq->device->modify_cq(cq, cq_count, cq_period) : -ENOSYS;
1454 EXPORT_SYMBOL(ib_modify_cq);
1456 int ib_destroy_cq(struct ib_cq *cq)
1458 if (atomic_read(&cq->usecnt))
1461 return cq->device->destroy_cq(cq);
1463 EXPORT_SYMBOL(ib_destroy_cq);
1465 int ib_resize_cq(struct ib_cq *cq, int cqe)
1467 return cq->device->resize_cq ?
1468 cq->device->resize_cq(cq, cqe, NULL) : -ENOSYS;
1470 EXPORT_SYMBOL(ib_resize_cq);
1472 /* Memory regions */
1474 int ib_dereg_mr(struct ib_mr *mr)
1476 struct ib_pd *pd = mr->pd;
1479 ret = mr->device->dereg_mr(mr);
1481 atomic_dec(&pd->usecnt);
1485 EXPORT_SYMBOL(ib_dereg_mr);
1488 * ib_alloc_mr() - Allocates a memory region
1489 * @pd: protection domain associated with the region
1490 * @mr_type: memory region type
1491 * @max_num_sg: maximum sg entries available for registration.
1494 * Memory registeration page/sg lists must not exceed max_num_sg.
1495 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
1496 * max_num_sg * used_page_size.
1499 struct ib_mr *ib_alloc_mr(struct ib_pd *pd,
1500 enum ib_mr_type mr_type,
1505 if (!pd->device->alloc_mr)
1506 return ERR_PTR(-ENOSYS);
1508 mr = pd->device->alloc_mr(pd, mr_type, max_num_sg);
1510 mr->device = pd->device;
1513 atomic_inc(&pd->usecnt);
1514 mr->need_inval = false;
1519 EXPORT_SYMBOL(ib_alloc_mr);
1521 /* "Fast" memory regions */
1523 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
1524 int mr_access_flags,
1525 struct ib_fmr_attr *fmr_attr)
1529 if (!pd->device->alloc_fmr)
1530 return ERR_PTR(-ENOSYS);
1532 fmr = pd->device->alloc_fmr(pd, mr_access_flags, fmr_attr);
1534 fmr->device = pd->device;
1536 atomic_inc(&pd->usecnt);
1541 EXPORT_SYMBOL(ib_alloc_fmr);
1543 int ib_unmap_fmr(struct list_head *fmr_list)
1547 if (list_empty(fmr_list))
1550 fmr = list_entry(fmr_list->next, struct ib_fmr, list);
1551 return fmr->device->unmap_fmr(fmr_list);
1553 EXPORT_SYMBOL(ib_unmap_fmr);
1555 int ib_dealloc_fmr(struct ib_fmr *fmr)
1561 ret = fmr->device->dealloc_fmr(fmr);
1563 atomic_dec(&pd->usecnt);
1567 EXPORT_SYMBOL(ib_dealloc_fmr);
1569 /* Multicast groups */
1571 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1575 if (!qp->device->attach_mcast)
1577 if (gid->raw[0] != 0xff || qp->qp_type != IB_QPT_UD ||
1578 lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
1579 lid == be16_to_cpu(IB_LID_PERMISSIVE))
1582 ret = qp->device->attach_mcast(qp, gid, lid);
1584 atomic_inc(&qp->usecnt);
1587 EXPORT_SYMBOL(ib_attach_mcast);
1589 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1593 if (!qp->device->detach_mcast)
1595 if (gid->raw[0] != 0xff || qp->qp_type != IB_QPT_UD ||
1596 lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
1597 lid == be16_to_cpu(IB_LID_PERMISSIVE))
1600 ret = qp->device->detach_mcast(qp, gid, lid);
1602 atomic_dec(&qp->usecnt);
1605 EXPORT_SYMBOL(ib_detach_mcast);
1607 struct ib_xrcd *ib_alloc_xrcd(struct ib_device *device)
1609 struct ib_xrcd *xrcd;
1611 if (!device->alloc_xrcd)
1612 return ERR_PTR(-ENOSYS);
1614 xrcd = device->alloc_xrcd(device, NULL, NULL);
1615 if (!IS_ERR(xrcd)) {
1616 xrcd->device = device;
1618 atomic_set(&xrcd->usecnt, 0);
1619 mutex_init(&xrcd->tgt_qp_mutex);
1620 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
1625 EXPORT_SYMBOL(ib_alloc_xrcd);
1627 int ib_dealloc_xrcd(struct ib_xrcd *xrcd)
1632 if (atomic_read(&xrcd->usecnt))
1635 while (!list_empty(&xrcd->tgt_qp_list)) {
1636 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
1637 ret = ib_destroy_qp(qp);
1642 return xrcd->device->dealloc_xrcd(xrcd);
1644 EXPORT_SYMBOL(ib_dealloc_xrcd);
1647 * ib_create_wq - Creates a WQ associated with the specified protection
1649 * @pd: The protection domain associated with the WQ.
1650 * @wq_init_attr: A list of initial attributes required to create the
1651 * WQ. If WQ creation succeeds, then the attributes are updated to
1652 * the actual capabilities of the created WQ.
1654 * wq_init_attr->max_wr and wq_init_attr->max_sge determine
1655 * the requested size of the WQ, and set to the actual values allocated
1657 * If ib_create_wq() succeeds, then max_wr and max_sge will always be
1658 * at least as large as the requested values.
1660 struct ib_wq *ib_create_wq(struct ib_pd *pd,
1661 struct ib_wq_init_attr *wq_attr)
1665 if (!pd->device->create_wq)
1666 return ERR_PTR(-ENOSYS);
1668 wq = pd->device->create_wq(pd, wq_attr, NULL);
1670 wq->event_handler = wq_attr->event_handler;
1671 wq->wq_context = wq_attr->wq_context;
1672 wq->wq_type = wq_attr->wq_type;
1673 wq->cq = wq_attr->cq;
1674 wq->device = pd->device;
1677 atomic_inc(&pd->usecnt);
1678 atomic_inc(&wq_attr->cq->usecnt);
1679 atomic_set(&wq->usecnt, 0);
1683 EXPORT_SYMBOL(ib_create_wq);
1686 * ib_destroy_wq - Destroys the specified WQ.
1687 * @wq: The WQ to destroy.
1689 int ib_destroy_wq(struct ib_wq *wq)
1692 struct ib_cq *cq = wq->cq;
1693 struct ib_pd *pd = wq->pd;
1695 if (atomic_read(&wq->usecnt))
1698 err = wq->device->destroy_wq(wq);
1700 atomic_dec(&pd->usecnt);
1701 atomic_dec(&cq->usecnt);
1705 EXPORT_SYMBOL(ib_destroy_wq);
1708 * ib_modify_wq - Modifies the specified WQ.
1709 * @wq: The WQ to modify.
1710 * @wq_attr: On input, specifies the WQ attributes to modify.
1711 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
1712 * are being modified.
1713 * On output, the current values of selected WQ attributes are returned.
1715 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
1720 if (!wq->device->modify_wq)
1723 err = wq->device->modify_wq(wq, wq_attr, wq_attr_mask, NULL);
1726 EXPORT_SYMBOL(ib_modify_wq);
1729 * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
1730 * @device: The device on which to create the rwq indirection table.
1731 * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
1732 * create the Indirection Table.
1734 * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
1735 * than the created ib_rwq_ind_table object and the caller is responsible
1736 * for its memory allocation/free.
1738 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
1739 struct ib_rwq_ind_table_init_attr *init_attr)
1741 struct ib_rwq_ind_table *rwq_ind_table;
1745 if (!device->create_rwq_ind_table)
1746 return ERR_PTR(-ENOSYS);
1748 table_size = (1 << init_attr->log_ind_tbl_size);
1749 rwq_ind_table = device->create_rwq_ind_table(device,
1751 if (IS_ERR(rwq_ind_table))
1752 return rwq_ind_table;
1754 rwq_ind_table->ind_tbl = init_attr->ind_tbl;
1755 rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
1756 rwq_ind_table->device = device;
1757 rwq_ind_table->uobject = NULL;
1758 atomic_set(&rwq_ind_table->usecnt, 0);
1760 for (i = 0; i < table_size; i++)
1761 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
1763 return rwq_ind_table;
1765 EXPORT_SYMBOL(ib_create_rwq_ind_table);
1768 * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
1769 * @wq_ind_table: The Indirection Table to destroy.
1771 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
1774 u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
1775 struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
1777 if (atomic_read(&rwq_ind_table->usecnt))
1780 err = rwq_ind_table->device->destroy_rwq_ind_table(rwq_ind_table);
1782 for (i = 0; i < table_size; i++)
1783 atomic_dec(&ind_tbl[i]->usecnt);
1788 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
1790 struct ib_flow *ib_create_flow(struct ib_qp *qp,
1791 struct ib_flow_attr *flow_attr,
1794 struct ib_flow *flow_id;
1795 if (!qp->device->create_flow)
1796 return ERR_PTR(-ENOSYS);
1798 flow_id = qp->device->create_flow(qp, flow_attr, domain);
1799 if (!IS_ERR(flow_id)) {
1800 atomic_inc(&qp->usecnt);
1805 EXPORT_SYMBOL(ib_create_flow);
1807 int ib_destroy_flow(struct ib_flow *flow_id)
1810 struct ib_qp *qp = flow_id->qp;
1812 err = qp->device->destroy_flow(flow_id);
1814 atomic_dec(&qp->usecnt);
1817 EXPORT_SYMBOL(ib_destroy_flow);
1819 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
1820 struct ib_mr_status *mr_status)
1822 return mr->device->check_mr_status ?
1823 mr->device->check_mr_status(mr, check_mask, mr_status) : -ENOSYS;
1825 EXPORT_SYMBOL(ib_check_mr_status);
1827 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
1830 if (!device->set_vf_link_state)
1833 return device->set_vf_link_state(device, vf, port, state);
1835 EXPORT_SYMBOL(ib_set_vf_link_state);
1837 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
1838 struct ifla_vf_info *info)
1840 if (!device->get_vf_config)
1843 return device->get_vf_config(device, vf, port, info);
1845 EXPORT_SYMBOL(ib_get_vf_config);
1847 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
1848 struct ifla_vf_stats *stats)
1850 if (!device->get_vf_stats)
1853 return device->get_vf_stats(device, vf, port, stats);
1855 EXPORT_SYMBOL(ib_get_vf_stats);
1857 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
1860 if (!device->set_vf_guid)
1863 return device->set_vf_guid(device, vf, port, guid, type);
1865 EXPORT_SYMBOL(ib_set_vf_guid);
1868 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
1869 * and set it the memory region.
1870 * @mr: memory region
1871 * @sg: dma mapped scatterlist
1872 * @sg_nents: number of entries in sg
1873 * @sg_offset: offset in bytes into sg
1874 * @page_size: page vector desired page size
1877 * - The first sg element is allowed to have an offset.
1878 * - Each sg element must either be aligned to page_size or virtually
1879 * contiguous to the previous element. In case an sg element has a
1880 * non-contiguous offset, the mapping prefix will not include it.
1881 * - The last sg element is allowed to have length less than page_size.
1882 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
1883 * then only max_num_sg entries will be mapped.
1884 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
1885 * constraints holds and the page_size argument is ignored.
1887 * Returns the number of sg elements that were mapped to the memory region.
1889 * After this completes successfully, the memory region
1890 * is ready for registration.
1892 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
1893 unsigned int *sg_offset, unsigned int page_size)
1895 if (unlikely(!mr->device->map_mr_sg))
1898 mr->page_size = page_size;
1900 return mr->device->map_mr_sg(mr, sg, sg_nents, sg_offset);
1902 EXPORT_SYMBOL(ib_map_mr_sg);
1905 * ib_sg_to_pages() - Convert the largest prefix of a sg list
1907 * @mr: memory region
1908 * @sgl: dma mapped scatterlist
1909 * @sg_nents: number of entries in sg
1910 * @sg_offset_p: IN: start offset in bytes into sg
1911 * OUT: offset in bytes for element n of the sg of the first
1912 * byte that has not been processed where n is the return
1913 * value of this function.
1914 * @set_page: driver page assignment function pointer
1916 * Core service helper for drivers to convert the largest
1917 * prefix of given sg list to a page vector. The sg list
1918 * prefix converted is the prefix that meet the requirements
1921 * Returns the number of sg elements that were assigned to
1924 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
1925 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
1927 struct scatterlist *sg;
1928 u64 last_end_dma_addr = 0;
1929 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
1930 unsigned int last_page_off = 0;
1931 u64 page_mask = ~((u64)mr->page_size - 1);
1934 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
1937 mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
1940 for_each_sg(sgl, sg, sg_nents, i) {
1941 u64 dma_addr = sg_dma_address(sg) + sg_offset;
1942 u64 prev_addr = dma_addr;
1943 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
1944 u64 end_dma_addr = dma_addr + dma_len;
1945 u64 page_addr = dma_addr & page_mask;
1948 * For the second and later elements, check whether either the
1949 * end of element i-1 or the start of element i is not aligned
1950 * on a page boundary.
1952 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
1953 /* Stop mapping if there is a gap. */
1954 if (last_end_dma_addr != dma_addr)
1958 * Coalesce this element with the last. If it is small
1959 * enough just update mr->length. Otherwise start
1960 * mapping from the next page.
1966 ret = set_page(mr, page_addr);
1967 if (unlikely(ret < 0)) {
1968 sg_offset = prev_addr - sg_dma_address(sg);
1969 mr->length += prev_addr - dma_addr;
1971 *sg_offset_p = sg_offset;
1972 return i || sg_offset ? i : ret;
1974 prev_addr = page_addr;
1976 page_addr += mr->page_size;
1977 } while (page_addr < end_dma_addr);
1979 mr->length += dma_len;
1980 last_end_dma_addr = end_dma_addr;
1981 last_page_off = end_dma_addr & ~page_mask;
1990 EXPORT_SYMBOL(ib_sg_to_pages);
1992 struct ib_drain_cqe {
1994 struct completion done;
1997 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
1999 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2002 complete(&cqe->done);
2006 * Post a WR and block until its completion is reaped for the SQ.
2008 static void __ib_drain_sq(struct ib_qp *qp)
2010 struct ib_cq *cq = qp->send_cq;
2011 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2012 struct ib_drain_cqe sdrain;
2013 struct ib_send_wr swr = {}, *bad_swr;
2016 swr.wr_cqe = &sdrain.cqe;
2017 sdrain.cqe.done = ib_drain_qp_done;
2018 init_completion(&sdrain.done);
2020 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2022 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2026 ret = ib_post_send(qp, &swr, &bad_swr);
2028 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2032 if (cq->poll_ctx == IB_POLL_DIRECT)
2033 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2034 ib_process_cq_direct(cq, -1);
2036 wait_for_completion(&sdrain.done);
2040 * Post a WR and block until its completion is reaped for the RQ.
2042 static void __ib_drain_rq(struct ib_qp *qp)
2044 struct ib_cq *cq = qp->recv_cq;
2045 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2046 struct ib_drain_cqe rdrain;
2047 struct ib_recv_wr rwr = {}, *bad_rwr;
2050 rwr.wr_cqe = &rdrain.cqe;
2051 rdrain.cqe.done = ib_drain_qp_done;
2052 init_completion(&rdrain.done);
2054 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2056 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2060 ret = ib_post_recv(qp, &rwr, &bad_rwr);
2062 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2066 if (cq->poll_ctx == IB_POLL_DIRECT)
2067 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2068 ib_process_cq_direct(cq, -1);
2070 wait_for_completion(&rdrain.done);
2074 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2076 * @qp: queue pair to drain
2078 * If the device has a provider-specific drain function, then
2079 * call that. Otherwise call the generic drain function
2084 * ensure there is room in the CQ and SQ for the drain work request and
2087 * allocate the CQ using ib_alloc_cq().
2089 * ensure that there are no other contexts that are posting WRs concurrently.
2090 * Otherwise the drain is not guaranteed.
2092 void ib_drain_sq(struct ib_qp *qp)
2094 if (qp->device->drain_sq)
2095 qp->device->drain_sq(qp);
2099 EXPORT_SYMBOL(ib_drain_sq);
2102 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2104 * @qp: queue pair to drain
2106 * If the device has a provider-specific drain function, then
2107 * call that. Otherwise call the generic drain function
2112 * ensure there is room in the CQ and RQ for the drain work request and
2115 * allocate the CQ using ib_alloc_cq().
2117 * ensure that there are no other contexts that are posting WRs concurrently.
2118 * Otherwise the drain is not guaranteed.
2120 void ib_drain_rq(struct ib_qp *qp)
2122 if (qp->device->drain_rq)
2123 qp->device->drain_rq(qp);
2127 EXPORT_SYMBOL(ib_drain_rq);
2130 * ib_drain_qp() - Block until all CQEs have been consumed by the
2131 * application on both the RQ and SQ.
2132 * @qp: queue pair to drain
2136 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2139 * allocate the CQs using ib_alloc_cq().
2141 * ensure that there are no other contexts that are posting WRs concurrently.
2142 * Otherwise the drain is not guaranteed.
2144 void ib_drain_qp(struct ib_qp *qp)
2150 EXPORT_SYMBOL(ib_drain_qp);