2 * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
3 * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <linux/dma-mapping.h>
35 #include <net/addrconf.h>
36 #include <rdma/uverbs_ioctl.h>
39 #include "rxe_queue.h"
40 #include "rxe_hw_counters.h"
42 static int rxe_query_device(struct ib_device *dev,
43 struct ib_device_attr *attr,
46 struct rxe_dev *rxe = to_rdev(dev);
48 if (uhw->inlen || uhw->outlen)
55 static int rxe_query_port(struct ib_device *dev,
56 u8 port_num, struct ib_port_attr *attr)
58 struct rxe_dev *rxe = to_rdev(dev);
59 struct rxe_port *port;
64 /* *attr being zeroed by the caller, avoid zeroing it here */
67 mutex_lock(&rxe->usdev_lock);
68 rc = ib_get_eth_speed(dev, port_num, &attr->active_speed,
71 if (attr->state == IB_PORT_ACTIVE)
72 attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
73 else if (dev_get_flags(rxe->ndev) & IFF_UP)
74 attr->phys_state = IB_PORT_PHYS_STATE_POLLING;
76 attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
78 mutex_unlock(&rxe->usdev_lock);
83 static int rxe_query_pkey(struct ib_device *device,
84 u8 port_num, u16 index, u16 *pkey)
86 struct rxe_dev *rxe = to_rdev(device);
87 struct rxe_port *port;
91 if (unlikely(index >= port->attr.pkey_tbl_len)) {
92 dev_warn(device->dev.parent, "invalid index = %d\n",
97 *pkey = port->pkey_tbl[index];
104 static int rxe_modify_device(struct ib_device *dev,
105 int mask, struct ib_device_modify *attr)
107 struct rxe_dev *rxe = to_rdev(dev);
109 if (mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
110 IB_DEVICE_MODIFY_NODE_DESC))
113 if (mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
114 rxe->attr.sys_image_guid = cpu_to_be64(attr->sys_image_guid);
116 if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
117 memcpy(rxe->ib_dev.node_desc,
118 attr->node_desc, sizeof(rxe->ib_dev.node_desc));
124 static int rxe_modify_port(struct ib_device *dev,
125 u8 port_num, int mask, struct ib_port_modify *attr)
127 struct rxe_dev *rxe = to_rdev(dev);
128 struct rxe_port *port;
132 port->attr.port_cap_flags |= attr->set_port_cap_mask;
133 port->attr.port_cap_flags &= ~attr->clr_port_cap_mask;
135 if (mask & IB_PORT_RESET_QKEY_CNTR)
136 port->attr.qkey_viol_cntr = 0;
141 static enum rdma_link_layer rxe_get_link_layer(struct ib_device *dev,
144 struct rxe_dev *rxe = to_rdev(dev);
146 return rxe_link_layer(rxe, port_num);
149 static int rxe_alloc_ucontext(struct ib_ucontext *uctx, struct ib_udata *udata)
151 struct rxe_dev *rxe = to_rdev(uctx->device);
152 struct rxe_ucontext *uc = to_ruc(uctx);
154 return rxe_add_to_pool(&rxe->uc_pool, &uc->pelem);
157 static void rxe_dealloc_ucontext(struct ib_ucontext *ibuc)
159 struct rxe_ucontext *uc = to_ruc(ibuc);
164 static int rxe_port_immutable(struct ib_device *dev, u8 port_num,
165 struct ib_port_immutable *immutable)
168 struct ib_port_attr attr;
170 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
172 err = ib_query_port(dev, port_num, &attr);
176 immutable->pkey_tbl_len = attr.pkey_tbl_len;
177 immutable->gid_tbl_len = attr.gid_tbl_len;
178 immutable->max_mad_size = IB_MGMT_MAD_SIZE;
183 static int rxe_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
185 struct rxe_dev *rxe = to_rdev(ibpd->device);
186 struct rxe_pd *pd = to_rpd(ibpd);
188 return rxe_add_to_pool(&rxe->pd_pool, &pd->pelem);
191 static void rxe_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
193 struct rxe_pd *pd = to_rpd(ibpd);
198 static int rxe_create_ah(struct ib_ah *ibah,
199 struct rdma_ah_init_attr *init_attr,
200 struct ib_udata *udata)
204 struct rxe_dev *rxe = to_rdev(ibah->device);
205 struct rxe_ah *ah = to_rah(ibah);
207 err = rxe_av_chk_attr(rxe, init_attr->ah_attr);
211 err = rxe_add_to_pool(&rxe->ah_pool, &ah->pelem);
215 rxe_init_av(init_attr->ah_attr, &ah->av);
219 static int rxe_modify_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr)
222 struct rxe_dev *rxe = to_rdev(ibah->device);
223 struct rxe_ah *ah = to_rah(ibah);
225 err = rxe_av_chk_attr(rxe, attr);
229 rxe_init_av(attr, &ah->av);
233 static int rxe_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr)
235 struct rxe_ah *ah = to_rah(ibah);
237 memset(attr, 0, sizeof(*attr));
238 attr->type = ibah->type;
239 rxe_av_to_attr(&ah->av, attr);
243 static void rxe_destroy_ah(struct ib_ah *ibah, u32 flags)
245 struct rxe_ah *ah = to_rah(ibah);
250 static int post_one_recv(struct rxe_rq *rq, const struct ib_recv_wr *ibwr)
255 struct rxe_recv_wqe *recv_wqe;
256 int num_sge = ibwr->num_sge;
258 if (unlikely(queue_full(rq->queue))) {
263 if (unlikely(num_sge > rq->max_sge)) {
269 for (i = 0; i < num_sge; i++)
270 length += ibwr->sg_list[i].length;
272 recv_wqe = producer_addr(rq->queue);
273 recv_wqe->wr_id = ibwr->wr_id;
274 recv_wqe->num_sge = num_sge;
276 memcpy(recv_wqe->dma.sge, ibwr->sg_list,
277 num_sge * sizeof(struct ib_sge));
279 recv_wqe->dma.length = length;
280 recv_wqe->dma.resid = length;
281 recv_wqe->dma.num_sge = num_sge;
282 recv_wqe->dma.cur_sge = 0;
283 recv_wqe->dma.sge_offset = 0;
285 /* make sure all changes to the work queue are written before we
286 * update the producer pointer
290 advance_producer(rq->queue);
297 static int rxe_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init,
298 struct ib_udata *udata)
301 struct rxe_dev *rxe = to_rdev(ibsrq->device);
302 struct rxe_pd *pd = to_rpd(ibsrq->pd);
303 struct rxe_srq *srq = to_rsrq(ibsrq);
304 struct rxe_create_srq_resp __user *uresp = NULL;
307 if (udata->outlen < sizeof(*uresp))
309 uresp = udata->outbuf;
312 err = rxe_srq_chk_attr(rxe, NULL, &init->attr, IB_SRQ_INIT_MASK);
316 err = rxe_add_to_pool(&rxe->srq_pool, &srq->pelem);
323 err = rxe_srq_from_init(rxe, srq, init, udata, uresp);
336 static int rxe_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
337 enum ib_srq_attr_mask mask,
338 struct ib_udata *udata)
341 struct rxe_srq *srq = to_rsrq(ibsrq);
342 struct rxe_dev *rxe = to_rdev(ibsrq->device);
343 struct rxe_modify_srq_cmd ucmd = {};
346 if (udata->inlen < sizeof(ucmd))
349 err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
354 err = rxe_srq_chk_attr(rxe, srq, attr, mask);
358 err = rxe_srq_from_attr(rxe, srq, attr, mask, &ucmd, udata);
368 static int rxe_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
370 struct rxe_srq *srq = to_rsrq(ibsrq);
375 attr->max_wr = srq->rq.queue->buf->index_mask;
376 attr->max_sge = srq->rq.max_sge;
377 attr->srq_limit = srq->limit;
381 static void rxe_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
383 struct rxe_srq *srq = to_rsrq(ibsrq);
386 rxe_queue_cleanup(srq->rq.queue);
388 rxe_drop_ref(srq->pd);
392 static int rxe_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
393 const struct ib_recv_wr **bad_wr)
397 struct rxe_srq *srq = to_rsrq(ibsrq);
399 spin_lock_irqsave(&srq->rq.producer_lock, flags);
402 err = post_one_recv(&srq->rq, wr);
408 spin_unlock_irqrestore(&srq->rq.producer_lock, flags);
416 static struct ib_qp *rxe_create_qp(struct ib_pd *ibpd,
417 struct ib_qp_init_attr *init,
418 struct ib_udata *udata)
421 struct rxe_dev *rxe = to_rdev(ibpd->device);
422 struct rxe_pd *pd = to_rpd(ibpd);
424 struct rxe_create_qp_resp __user *uresp = NULL;
427 if (udata->outlen < sizeof(*uresp))
428 return ERR_PTR(-EINVAL);
429 uresp = udata->outbuf;
432 err = rxe_qp_chk_init(rxe, init);
436 qp = rxe_alloc(&rxe->qp_pool);
452 err = rxe_qp_from_init(rxe, qp, pd, init, uresp, ibpd, udata);
466 static int rxe_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
467 int mask, struct ib_udata *udata)
470 struct rxe_dev *rxe = to_rdev(ibqp->device);
471 struct rxe_qp *qp = to_rqp(ibqp);
473 err = rxe_qp_chk_attr(rxe, qp, attr, mask);
477 err = rxe_qp_from_attr(qp, attr, mask, udata);
487 static int rxe_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
488 int mask, struct ib_qp_init_attr *init)
490 struct rxe_qp *qp = to_rqp(ibqp);
492 rxe_qp_to_init(qp, init);
493 rxe_qp_to_attr(qp, attr, mask);
498 static int rxe_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
500 struct rxe_qp *qp = to_rqp(ibqp);
508 static int validate_send_wr(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
509 unsigned int mask, unsigned int length)
511 int num_sge = ibwr->num_sge;
512 struct rxe_sq *sq = &qp->sq;
514 if (unlikely(num_sge > sq->max_sge))
517 if (unlikely(mask & WR_ATOMIC_MASK)) {
521 if (atomic_wr(ibwr)->remote_addr & 0x7)
525 if (unlikely((ibwr->send_flags & IB_SEND_INLINE) &&
526 (length > sq->max_inline)))
535 static void init_send_wr(struct rxe_qp *qp, struct rxe_send_wr *wr,
536 const struct ib_send_wr *ibwr)
538 wr->wr_id = ibwr->wr_id;
539 wr->num_sge = ibwr->num_sge;
540 wr->opcode = ibwr->opcode;
541 wr->send_flags = ibwr->send_flags;
543 if (qp_type(qp) == IB_QPT_UD ||
544 qp_type(qp) == IB_QPT_SMI ||
545 qp_type(qp) == IB_QPT_GSI) {
546 wr->wr.ud.remote_qpn = ud_wr(ibwr)->remote_qpn;
547 wr->wr.ud.remote_qkey = ud_wr(ibwr)->remote_qkey;
548 if (qp_type(qp) == IB_QPT_GSI)
549 wr->wr.ud.pkey_index = ud_wr(ibwr)->pkey_index;
550 if (wr->opcode == IB_WR_SEND_WITH_IMM)
551 wr->ex.imm_data = ibwr->ex.imm_data;
553 switch (wr->opcode) {
554 case IB_WR_RDMA_WRITE_WITH_IMM:
555 wr->ex.imm_data = ibwr->ex.imm_data;
557 case IB_WR_RDMA_READ:
558 case IB_WR_RDMA_WRITE:
559 wr->wr.rdma.remote_addr = rdma_wr(ibwr)->remote_addr;
560 wr->wr.rdma.rkey = rdma_wr(ibwr)->rkey;
562 case IB_WR_SEND_WITH_IMM:
563 wr->ex.imm_data = ibwr->ex.imm_data;
565 case IB_WR_SEND_WITH_INV:
566 wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey;
568 case IB_WR_ATOMIC_CMP_AND_SWP:
569 case IB_WR_ATOMIC_FETCH_AND_ADD:
570 wr->wr.atomic.remote_addr =
571 atomic_wr(ibwr)->remote_addr;
572 wr->wr.atomic.compare_add =
573 atomic_wr(ibwr)->compare_add;
574 wr->wr.atomic.swap = atomic_wr(ibwr)->swap;
575 wr->wr.atomic.rkey = atomic_wr(ibwr)->rkey;
577 case IB_WR_LOCAL_INV:
578 wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey;
581 wr->wr.reg.mr = reg_wr(ibwr)->mr;
582 wr->wr.reg.key = reg_wr(ibwr)->key;
583 wr->wr.reg.access = reg_wr(ibwr)->access;
591 static int init_send_wqe(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
592 unsigned int mask, unsigned int length,
593 struct rxe_send_wqe *wqe)
595 int num_sge = ibwr->num_sge;
600 init_send_wr(qp, &wqe->wr, ibwr);
602 if (qp_type(qp) == IB_QPT_UD ||
603 qp_type(qp) == IB_QPT_SMI ||
604 qp_type(qp) == IB_QPT_GSI)
605 memcpy(&wqe->av, &to_rah(ud_wr(ibwr)->ah)->av, sizeof(wqe->av));
607 if (unlikely(ibwr->send_flags & IB_SEND_INLINE)) {
608 p = wqe->dma.inline_data;
611 for (i = 0; i < num_sge; i++, sge++) {
612 memcpy(p, (void *)(uintptr_t)sge->addr,
617 } else if (mask & WR_REG_MASK) {
619 wqe->state = wqe_state_posted;
622 memcpy(wqe->dma.sge, ibwr->sg_list,
623 num_sge * sizeof(struct ib_sge));
625 wqe->iova = mask & WR_ATOMIC_MASK ? atomic_wr(ibwr)->remote_addr :
626 mask & WR_READ_OR_WRITE_MASK ? rdma_wr(ibwr)->remote_addr : 0;
628 wqe->dma.length = length;
629 wqe->dma.resid = length;
630 wqe->dma.num_sge = num_sge;
631 wqe->dma.cur_sge = 0;
632 wqe->dma.sge_offset = 0;
633 wqe->state = wqe_state_posted;
634 wqe->ssn = atomic_add_return(1, &qp->ssn);
639 static int post_one_send(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
640 unsigned int mask, u32 length)
643 struct rxe_sq *sq = &qp->sq;
644 struct rxe_send_wqe *send_wqe;
647 err = validate_send_wr(qp, ibwr, mask, length);
651 spin_lock_irqsave(&qp->sq.sq_lock, flags);
653 if (unlikely(queue_full(sq->queue))) {
658 send_wqe = producer_addr(sq->queue);
660 err = init_send_wqe(qp, ibwr, mask, length, send_wqe);
665 * make sure all changes to the work queue are
666 * written before we update the producer pointer
670 advance_producer(sq->queue);
671 spin_unlock_irqrestore(&qp->sq.sq_lock, flags);
676 spin_unlock_irqrestore(&qp->sq.sq_lock, flags);
680 static int rxe_post_send_kernel(struct rxe_qp *qp, const struct ib_send_wr *wr,
681 const struct ib_send_wr **bad_wr)
685 unsigned int length = 0;
689 mask = wr_opcode_mask(wr->opcode, qp);
690 if (unlikely(!mask)) {
696 if (unlikely((wr->send_flags & IB_SEND_INLINE) &&
697 !(mask & WR_INLINE_MASK))) {
704 for (i = 0; i < wr->num_sge; i++)
705 length += wr->sg_list[i].length;
707 err = post_one_send(qp, wr, mask, length);
716 rxe_run_task(&qp->req.task, 1);
717 if (unlikely(qp->req.state == QP_STATE_ERROR))
718 rxe_run_task(&qp->comp.task, 1);
723 static int rxe_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
724 const struct ib_send_wr **bad_wr)
726 struct rxe_qp *qp = to_rqp(ibqp);
728 if (unlikely(!qp->valid)) {
733 if (unlikely(qp->req.state < QP_STATE_READY)) {
739 /* Utilize process context to do protocol processing */
740 rxe_run_task(&qp->req.task, 0);
743 return rxe_post_send_kernel(qp, wr, bad_wr);
746 static int rxe_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
747 const struct ib_recv_wr **bad_wr)
750 struct rxe_qp *qp = to_rqp(ibqp);
751 struct rxe_rq *rq = &qp->rq;
754 if (unlikely((qp_state(qp) < IB_QPS_INIT) || !qp->valid)) {
760 if (unlikely(qp->srq)) {
766 spin_lock_irqsave(&rq->producer_lock, flags);
769 err = post_one_recv(rq, wr);
777 spin_unlock_irqrestore(&rq->producer_lock, flags);
779 if (qp->resp.state == QP_STATE_ERROR)
780 rxe_run_task(&qp->resp.task, 1);
786 static int rxe_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
787 struct ib_udata *udata)
790 struct ib_device *dev = ibcq->device;
791 struct rxe_dev *rxe = to_rdev(dev);
792 struct rxe_cq *cq = to_rcq(ibcq);
793 struct rxe_create_cq_resp __user *uresp = NULL;
796 if (udata->outlen < sizeof(*uresp))
798 uresp = udata->outbuf;
804 err = rxe_cq_chk_attr(rxe, NULL, attr->cqe, attr->comp_vector);
808 err = rxe_cq_from_init(rxe, cq, attr->cqe, attr->comp_vector, udata,
813 return rxe_add_to_pool(&rxe->cq_pool, &cq->pelem);
816 static void rxe_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
818 struct rxe_cq *cq = to_rcq(ibcq);
825 static int rxe_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
828 struct rxe_cq *cq = to_rcq(ibcq);
829 struct rxe_dev *rxe = to_rdev(ibcq->device);
830 struct rxe_resize_cq_resp __user *uresp = NULL;
833 if (udata->outlen < sizeof(*uresp))
835 uresp = udata->outbuf;
838 err = rxe_cq_chk_attr(rxe, cq, cqe, 0);
842 err = rxe_cq_resize_queue(cq, cqe, uresp, udata);
852 static int rxe_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
855 struct rxe_cq *cq = to_rcq(ibcq);
859 spin_lock_irqsave(&cq->cq_lock, flags);
860 for (i = 0; i < num_entries; i++) {
861 cqe = queue_head(cq->queue);
865 memcpy(wc++, &cqe->ibwc, sizeof(*wc));
866 advance_consumer(cq->queue);
868 spin_unlock_irqrestore(&cq->cq_lock, flags);
873 static int rxe_peek_cq(struct ib_cq *ibcq, int wc_cnt)
875 struct rxe_cq *cq = to_rcq(ibcq);
876 int count = queue_count(cq->queue);
878 return (count > wc_cnt) ? wc_cnt : count;
881 static int rxe_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
883 struct rxe_cq *cq = to_rcq(ibcq);
884 unsigned long irq_flags;
887 spin_lock_irqsave(&cq->cq_lock, irq_flags);
888 if (cq->notify != IB_CQ_NEXT_COMP)
889 cq->notify = flags & IB_CQ_SOLICITED_MASK;
891 if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !queue_empty(cq->queue))
894 spin_unlock_irqrestore(&cq->cq_lock, irq_flags);
899 static struct ib_mr *rxe_get_dma_mr(struct ib_pd *ibpd, int access)
901 struct rxe_dev *rxe = to_rdev(ibpd->device);
902 struct rxe_pd *pd = to_rpd(ibpd);
906 mr = rxe_alloc(&rxe->mr_pool);
916 err = rxe_mem_init_dma(pd, access, mr);
930 static struct ib_mr *rxe_reg_user_mr(struct ib_pd *ibpd,
934 int access, struct ib_udata *udata)
937 struct rxe_dev *rxe = to_rdev(ibpd->device);
938 struct rxe_pd *pd = to_rpd(ibpd);
941 mr = rxe_alloc(&rxe->mr_pool);
951 err = rxe_mem_init_user(pd, start, length, iova,
966 static int rxe_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
968 struct rxe_mem *mr = to_rmr(ibmr);
970 mr->state = RXE_MEM_STATE_ZOMBIE;
971 rxe_drop_ref(mr->pd);
977 static struct ib_mr *rxe_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
978 u32 max_num_sg, struct ib_udata *udata)
980 struct rxe_dev *rxe = to_rdev(ibpd->device);
981 struct rxe_pd *pd = to_rpd(ibpd);
985 if (mr_type != IB_MR_TYPE_MEM_REG)
986 return ERR_PTR(-EINVAL);
988 mr = rxe_alloc(&rxe->mr_pool);
998 err = rxe_mem_init_fast(pd, max_num_sg, mr);
1009 return ERR_PTR(err);
1012 static int rxe_set_page(struct ib_mr *ibmr, u64 addr)
1014 struct rxe_mem *mr = to_rmr(ibmr);
1015 struct rxe_map *map;
1016 struct rxe_phys_buf *buf;
1018 if (unlikely(mr->nbuf == mr->num_buf))
1021 map = mr->map[mr->nbuf / RXE_BUF_PER_MAP];
1022 buf = &map->buf[mr->nbuf % RXE_BUF_PER_MAP];
1025 buf->size = ibmr->page_size;
1031 static int rxe_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
1032 int sg_nents, unsigned int *sg_offset)
1034 struct rxe_mem *mr = to_rmr(ibmr);
1039 n = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, rxe_set_page);
1041 mr->va = ibmr->iova;
1042 mr->iova = ibmr->iova;
1043 mr->length = ibmr->length;
1044 mr->page_shift = ilog2(ibmr->page_size);
1045 mr->page_mask = ibmr->page_size - 1;
1046 mr->offset = mr->iova & mr->page_mask;
1051 static int rxe_attach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid)
1054 struct rxe_dev *rxe = to_rdev(ibqp->device);
1055 struct rxe_qp *qp = to_rqp(ibqp);
1056 struct rxe_mc_grp *grp;
1058 /* takes a ref on grp if successful */
1059 err = rxe_mcast_get_grp(rxe, mgid, &grp);
1063 err = rxe_mcast_add_grp_elem(rxe, qp, grp);
1069 static int rxe_detach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid)
1071 struct rxe_dev *rxe = to_rdev(ibqp->device);
1072 struct rxe_qp *qp = to_rqp(ibqp);
1074 return rxe_mcast_drop_grp_elem(rxe, qp, mgid);
1077 static ssize_t parent_show(struct device *device,
1078 struct device_attribute *attr, char *buf)
1080 struct rxe_dev *rxe =
1081 rdma_device_to_drv_device(device, struct rxe_dev, ib_dev);
1083 return snprintf(buf, 16, "%s\n", rxe_parent_name(rxe, 1));
1086 static DEVICE_ATTR_RO(parent);
1088 static struct attribute *rxe_dev_attributes[] = {
1089 &dev_attr_parent.attr,
1093 static const struct attribute_group rxe_attr_group = {
1094 .attrs = rxe_dev_attributes,
1097 static int rxe_enable_driver(struct ib_device *ib_dev)
1099 struct rxe_dev *rxe = container_of(ib_dev, struct rxe_dev, ib_dev);
1101 rxe_set_port_state(rxe);
1102 dev_info(&rxe->ib_dev.dev, "added %s\n", netdev_name(rxe->ndev));
1106 static const struct ib_device_ops rxe_dev_ops = {
1107 .owner = THIS_MODULE,
1108 .driver_id = RDMA_DRIVER_RXE,
1109 .uverbs_abi_ver = RXE_UVERBS_ABI_VERSION,
1111 .alloc_hw_stats = rxe_ib_alloc_hw_stats,
1112 .alloc_mr = rxe_alloc_mr,
1113 .alloc_pd = rxe_alloc_pd,
1114 .alloc_ucontext = rxe_alloc_ucontext,
1115 .attach_mcast = rxe_attach_mcast,
1116 .create_ah = rxe_create_ah,
1117 .create_cq = rxe_create_cq,
1118 .create_qp = rxe_create_qp,
1119 .create_srq = rxe_create_srq,
1120 .dealloc_driver = rxe_dealloc,
1121 .dealloc_pd = rxe_dealloc_pd,
1122 .dealloc_ucontext = rxe_dealloc_ucontext,
1123 .dereg_mr = rxe_dereg_mr,
1124 .destroy_ah = rxe_destroy_ah,
1125 .destroy_cq = rxe_destroy_cq,
1126 .destroy_qp = rxe_destroy_qp,
1127 .destroy_srq = rxe_destroy_srq,
1128 .detach_mcast = rxe_detach_mcast,
1129 .enable_driver = rxe_enable_driver,
1130 .get_dma_mr = rxe_get_dma_mr,
1131 .get_hw_stats = rxe_ib_get_hw_stats,
1132 .get_link_layer = rxe_get_link_layer,
1133 .get_port_immutable = rxe_port_immutable,
1134 .map_mr_sg = rxe_map_mr_sg,
1136 .modify_ah = rxe_modify_ah,
1137 .modify_device = rxe_modify_device,
1138 .modify_port = rxe_modify_port,
1139 .modify_qp = rxe_modify_qp,
1140 .modify_srq = rxe_modify_srq,
1141 .peek_cq = rxe_peek_cq,
1142 .poll_cq = rxe_poll_cq,
1143 .post_recv = rxe_post_recv,
1144 .post_send = rxe_post_send,
1145 .post_srq_recv = rxe_post_srq_recv,
1146 .query_ah = rxe_query_ah,
1147 .query_device = rxe_query_device,
1148 .query_pkey = rxe_query_pkey,
1149 .query_port = rxe_query_port,
1150 .query_qp = rxe_query_qp,
1151 .query_srq = rxe_query_srq,
1152 .reg_user_mr = rxe_reg_user_mr,
1153 .req_notify_cq = rxe_req_notify_cq,
1154 .resize_cq = rxe_resize_cq,
1156 INIT_RDMA_OBJ_SIZE(ib_ah, rxe_ah, ibah),
1157 INIT_RDMA_OBJ_SIZE(ib_cq, rxe_cq, ibcq),
1158 INIT_RDMA_OBJ_SIZE(ib_pd, rxe_pd, ibpd),
1159 INIT_RDMA_OBJ_SIZE(ib_srq, rxe_srq, ibsrq),
1160 INIT_RDMA_OBJ_SIZE(ib_ucontext, rxe_ucontext, ibuc),
1163 int rxe_register_device(struct rxe_dev *rxe, const char *ibdev_name)
1166 struct ib_device *dev = &rxe->ib_dev;
1167 struct crypto_shash *tfm;
1169 strlcpy(dev->node_desc, "rxe", sizeof(dev->node_desc));
1171 dev->node_type = RDMA_NODE_IB_CA;
1172 dev->phys_port_cnt = 1;
1173 dev->num_comp_vectors = num_possible_cpus();
1174 dev->dev.parent = rxe_dma_device(rxe);
1175 dev->local_dma_lkey = 0;
1176 addrconf_addr_eui48((unsigned char *)&dev->node_guid,
1177 rxe->ndev->dev_addr);
1178 dev->dev.dma_ops = &dma_virt_ops;
1179 dev->dev.dma_parms = &rxe->dma_parms;
1180 rxe->dma_parms = (struct device_dma_parameters)
1181 { .max_segment_size = SZ_2G };
1182 dma_coerce_mask_and_coherent(&dev->dev,
1183 dma_get_required_mask(&dev->dev));
1185 dev->uverbs_cmd_mask = BIT_ULL(IB_USER_VERBS_CMD_GET_CONTEXT)
1186 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)
1187 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_DEVICE)
1188 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_PORT)
1189 | BIT_ULL(IB_USER_VERBS_CMD_ALLOC_PD)
1190 | BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_PD)
1191 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_SRQ)
1192 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_SRQ)
1193 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_SRQ)
1194 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_SRQ)
1195 | BIT_ULL(IB_USER_VERBS_CMD_POST_SRQ_RECV)
1196 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_QP)
1197 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_QP)
1198 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_QP)
1199 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_QP)
1200 | BIT_ULL(IB_USER_VERBS_CMD_POST_SEND)
1201 | BIT_ULL(IB_USER_VERBS_CMD_POST_RECV)
1202 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_CQ)
1203 | BIT_ULL(IB_USER_VERBS_CMD_RESIZE_CQ)
1204 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_CQ)
1205 | BIT_ULL(IB_USER_VERBS_CMD_POLL_CQ)
1206 | BIT_ULL(IB_USER_VERBS_CMD_PEEK_CQ)
1207 | BIT_ULL(IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)
1208 | BIT_ULL(IB_USER_VERBS_CMD_REG_MR)
1209 | BIT_ULL(IB_USER_VERBS_CMD_DEREG_MR)
1210 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_AH)
1211 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_AH)
1212 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_AH)
1213 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_AH)
1214 | BIT_ULL(IB_USER_VERBS_CMD_ATTACH_MCAST)
1215 | BIT_ULL(IB_USER_VERBS_CMD_DETACH_MCAST)
1218 ib_set_device_ops(dev, &rxe_dev_ops);
1219 err = ib_device_set_netdev(&rxe->ib_dev, rxe->ndev, 1);
1223 tfm = crypto_alloc_shash("crc32", 0, 0);
1225 pr_err("failed to allocate crc algorithm err:%ld\n",
1227 return PTR_ERR(tfm);
1231 rdma_set_device_sysfs_group(dev, &rxe_attr_group);
1232 err = ib_register_device(dev, ibdev_name);
1234 pr_warn("%s failed with error %d\n", __func__, err);
1237 * Note that rxe may be invalid at this point if another thread