From mboxrd@z Thu Jan 1 00:00:00 1970 From: Leon Romanovsky Subject: Re: [PATCH v3 11/15] IB/pvrdma: Add Queue Pair support Date: Wed, 10 Aug 2016 07:16:40 +0300 Message-ID: <20160810041640.GG23921@leon.nu> References: <1470266864-16888-1-git-send-email-aditr@vmware.com> <1470266864-16888-12-git-send-email-aditr@vmware.com> Mime-Version: 1.0 Content-Type: multipart/signed; micalg=pgp-sha1; protocol="application/pgp-signature"; boundary="XRI2XbIfl/05pQwm" Return-path: Content-Disposition: inline In-Reply-To: <1470266864-16888-12-git-send-email-aditr-pghWNbHTmq7QT0dZR+AlfA@public.gmane.org> Sender: linux-rdma-owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org To: Adit Ranadive Cc: dledford-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org, linux-rdma-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, pv-drivers-pghWNbHTmq7QT0dZR+AlfA@public.gmane.org, jhansen-pghWNbHTmq7QT0dZR+AlfA@public.gmane.org, asarwade-pghWNbHTmq7QT0dZR+AlfA@public.gmane.org, georgezhang-pghWNbHTmq7QT0dZR+AlfA@public.gmane.org, bryantan-pghWNbHTmq7QT0dZR+AlfA@public.gmane.org List-Id: linux-rdma@vger.kernel.org --XRI2XbIfl/05pQwm Content-Type: text/plain; charset=us-ascii Content-Disposition: inline Content-Transfer-Encoding: quoted-printable On Wed, Aug 03, 2016 at 04:27:40PM -0700, Adit Ranadive wrote: > This patch adds the ability to create, modify, query and destroy QPs. The > PVRDMA device supports RC, UD and GSI QPs. >=20 > Changes v2->v3: > - Removed boolean in pvrdma_cmd_post. >=20 > Reviewed-by: Jorgen Hansen > Reviewed-by: George Zhang > Reviewed-by: Aditya Sarwade > Reviewed-by: Bryan Tan > Signed-off-by: Adit Ranadive > --- > drivers/infiniband/hw/pvrdma/pvrdma_qp.c | 975 +++++++++++++++++++++++++= ++++++ > 1 file changed, 975 insertions(+) > create mode 100644 drivers/infiniband/hw/pvrdma/pvrdma_qp.c >=20 > diff --git a/drivers/infiniband/hw/pvrdma/pvrdma_qp.c b/drivers/infiniban= d/hw/pvrdma/pvrdma_qp.c > new file mode 100644 > index 0000000..59acb5c > --- /dev/null > +++ b/drivers/infiniband/hw/pvrdma/pvrdma_qp.c > @@ -0,0 +1,975 @@ > +/* > + * Copyright (c) 2012-2016 VMware, Inc. All rights reserved. > + * > + * This program is free software; you can redistribute it and/or > + * modify it under the terms of EITHER the GNU General Public License > + * version 2 as published by the Free Software Foundation or the BSD > + * 2-Clause License. This program is distributed in the hope that it > + * will be useful, but WITHOUT ANY WARRANTY; WITHOUT EVEN THE IMPLIED > + * WARRANTY OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. > + * See the GNU General Public License version 2 for more details at > + * http://www.gnu.org/licenses/old-licenses/gpl-2.0.en.html. > + * > + * You should have received a copy of the GNU General Public License > + * along with this program available in the file COPYING in the main > + * directory of this source tree. > + * > + * The BSD 2-Clause License > + * > + * Redistribution and use in source and binary forms, with or > + * without modification, are permitted provided that the following > + * conditions are met: > + * > + * - Redistributions of source code must retain the above > + * copyright notice, this list of conditions and the following > + * disclaimer. > + * > + * - Redistributions in binary form must reproduce the above > + * copyright notice, this list of conditions and the following > + * disclaimer in the documentation and/or other materials > + * provided with the distribution. > + * > + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS > + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT > + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS > + * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE > + * COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, > + * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES > + * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR > + * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) > + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, > + * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) > + * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED > + * OF THE POSSIBILITY OF SUCH DAMAGE. > + */ > + > +#include > +#include > +#include > +#include > +#include > +#include > + > +#include "pvrdma.h" > +#include "pvrdma_user.h" > + > +static inline void get_cqs(struct pvrdma_qp *qp, struct pvrdma_cq **send= _cq, > + struct pvrdma_cq **recv_cq) > +{ > + *send_cq =3D to_vcq(qp->ibqp.send_cq); > + *recv_cq =3D to_vcq(qp->ibqp.recv_cq); > +} > + > +static void pvrdma_reset_qp(struct pvrdma_qp *qp) > +{ > + struct pvrdma_cq *scq, *rcq; > + unsigned long flags; > + > + /* Clean up cqes */ > + get_cqs(qp, &scq, &rcq); > + spin_lock_irqsave(&scq->cq_lock, flags); > + pvrdma_flush_cqe(qp, scq); The preference is to call functions which require locks to be held before executing them in _func_name() format and not as you called func_nam= e(). > + spin_unlock_irqrestore(&scq->cq_lock, flags); > + if (rcq !=3D scq) { > + spin_lock_irqsave(&rcq->cq_lock, flags); > + pvrdma_flush_cqe(qp, rcq); > + spin_unlock_irqrestore(&rcq->cq_lock, flags); > + } > + > + /* > + * Reset queuepair. The checks are because usermode queuepairs won't > + * have kernel ringstates. > + */ > + if (qp->rq.ring) { > + atomic_set(&qp->rq.ring->cons_head, 0); > + atomic_set(&qp->rq.ring->prod_tail, 0); > + } > + if (qp->sq.ring) { > + atomic_set(&qp->sq.ring->cons_head, 0); > + atomic_set(&qp->sq.ring->prod_tail, 0); > + } > +} > + > +static int pvrdma_set_rq_size(struct pvrdma_dev *dev, > + struct ib_qp_cap *cap, > + struct pvrdma_qp *qp) > +{ > + if (cap->max_recv_wr > dev->dsr->caps.max_qp_wr || > + cap->max_recv_sge > dev->dsr->caps.max_sge) { > + dev_warn(&dev->pdev->dev, "recv queue size invalid\n"); > + return -EINVAL; > + } > + > + qp->rq.wqe_cnt =3D roundup_pow_of_two(max(1U, cap->max_recv_wr)); > + qp->rq.max_gs =3D roundup_pow_of_two(max(1U, cap->max_recv_sge)); > + > + /* Write back */ > + cap->max_recv_wr =3D qp->rq.wqe_cnt; > + cap->max_recv_sge =3D qp->rq.max_gs; > + > + qp->rq.wqe_size =3D roundup_pow_of_two(sizeof(struct pvrdma_rq_wqe_hdr)= + > + sizeof(struct ib_sge) * > + qp->rq.max_gs); > + qp->npages_recv =3D (qp->rq.wqe_cnt * qp->rq.wqe_size + PAGE_SIZE - 1) / > + PAGE_SIZE; > + > + return 0; > +} > + > +static int pvrdma_set_sq_size(struct pvrdma_dev *dev, struct ib_qp_cap *= cap, > + enum ib_qp_type type, struct pvrdma_qp *qp) > +{ > + if (cap->max_send_wr > dev->dsr->caps.max_qp_wr || > + cap->max_send_sge > dev->dsr->caps.max_sge) { > + dev_warn(&dev->pdev->dev, "send queue size invalid\n"); > + return -EINVAL; > + } > + > + qp->sq.wqe_cnt =3D roundup_pow_of_two(max(1U, cap->max_send_wr)); > + qp->sq.max_gs =3D roundup_pow_of_two(max(1U, cap->max_send_sge)); > + > + /* Write back */ > + cap->max_send_wr =3D qp->sq.wqe_cnt; > + cap->max_send_sge =3D qp->sq.max_gs; > + > + qp->sq.wqe_size =3D roundup_pow_of_two(sizeof(struct pvrdma_sq_wqe_hdr)= + > + sizeof(struct ib_sge) * > + qp->sq.max_gs); > + /* Note: one extra page for the header. */ > + qp->npages_send =3D 1 + (qp->sq.wqe_cnt * qp->sq.wqe_size + > + PAGE_SIZE - 1) / PAGE_SIZE; > + > + return 0; > +} > + > +/** > + * pvrdma_create_qp - create queue pair > + * @pd: protection domain > + * @init_attr: queue pair attributes > + * @udata: user data > + * > + * @return: the ib_qp pointer on success, otherwise returns an errno. > + */ > +struct ib_qp *pvrdma_create_qp(struct ib_pd *pd, > + struct ib_qp_init_attr *init_attr, > + struct ib_udata *udata) > +{ > + struct pvrdma_qp *qp; > + struct pvrdma_dev *dev =3D to_vdev(pd->device); > + struct pvrdma_cq *send_cq, *recv_cq; > + union pvrdma_cmd_req req; > + union pvrdma_cmd_resp rsp; > + struct pvrdma_cmd_create_qp *cmd =3D &req.create_qp; > + struct pvrdma_cmd_create_qp_resp *resp =3D &rsp.create_qp_resp; > + struct pvrdma_create_qp ucmd; > + unsigned long flags; > + int ret; > + > + if (!atomic_add_unless(&dev->num_qps, 1, dev->dsr->caps.max_qp)) > + return ERR_PTR(-EINVAL); > + > + if (init_attr->create_flags) { > + dev_warn(&dev->pdev->dev, > + "invalid create queuepair flags %#x\n", > + init_attr->create_flags); > + atomic_dec(&dev->num_qps); > + return ERR_PTR(-EINVAL); Declare qp =3D=3D NULL and use one exit point (goto err_qp), you are doing the same for all returns. > + } > + > + if (init_attr->qp_type !=3D IB_QPT_RC && > + init_attr->qp_type !=3D IB_QPT_UD && > + init_attr->qp_type !=3D IB_QPT_GSI) { > + dev_warn(&dev->pdev->dev, "queuepair type %d not supported\n", > + init_attr->qp_type); > + atomic_dec(&dev->num_qps); > + return ERR_PTR(-EINVAL); > + } > + > + switch (init_attr->qp_type) { > + case IB_QPT_GSI: > + if (init_attr->port_num =3D=3D 0 || > + init_attr->port_num > pd->device->phys_port_cnt || > + udata) { > + dev_warn(&dev->pdev->dev, "invalid queuepair attrs\n"); > + atomic_dec(&dev->num_qps); > + return ERR_PTR(-EINVAL); > + } > + /* fall through */ > + case IB_QPT_RC: > + case IB_QPT_UD: > + qp =3D kzalloc(sizeof(*qp), GFP_KERNEL); > + if (!qp) { > + atomic_dec(&dev->num_qps); > + return ERR_PTR(-ENOMEM); > + } > + > + spin_lock_init(&qp->sq.lock); > + spin_lock_init(&qp->rq.lock); > + mutex_init(&qp->mutex); > + atomic_set(&qp->refcnt, 1); > + init_waitqueue_head(&qp->wait); > + > + qp->state =3D IB_QPS_RESET; > + > + if (pd->uobject && udata) { > + dev_dbg(&dev->pdev->dev, > + "create queuepair from user space\n"); > + > + if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) { > + ret =3D -EFAULT; > + goto err_qp; > + } > + > + /* set qp->sq.wqe_cnt, shift, buf_size.. */ > + qp->rumem =3D ib_umem_get(pd->uobject->context, > + ucmd.rbuf_addr, > + ucmd.rbuf_size, 0, 0); > + if (IS_ERR(qp->rumem)) { > + ret =3D PTR_ERR(qp->rumem); > + goto err_qp; > + } > + > + qp->sumem =3D ib_umem_get(pd->uobject->context, > + ucmd.sbuf_addr, > + ucmd.sbuf_size, 0, 0); > + if (IS_ERR(qp->sumem)) { > + ib_umem_release(qp->rumem); > + ret =3D PTR_ERR(qp->sumem); > + goto err_qp; > + } > + > + qp->npages_send =3D ib_umem_page_count(qp->sumem); > + qp->npages_recv =3D ib_umem_page_count(qp->rumem); > + qp->npages =3D qp->npages_send + qp->npages_recv; > + } else { > + qp->is_kernel =3D true; > + > + send_cq =3D to_vcq(init_attr->send_cq); > + recv_cq =3D to_vcq(init_attr->recv_cq); > + > + ret =3D pvrdma_set_sq_size(to_vdev(pd->device), > + &init_attr->cap, > + init_attr->qp_type, qp); > + if (ret) > + goto err_qp; > + > + ret =3D pvrdma_set_rq_size(to_vdev(pd->device), > + &init_attr->cap, qp); > + if (ret) > + goto err_qp; > + > + qp->npages =3D qp->npages_send + qp->npages_recv; > + > + /* Skip header page. */ > + qp->sq.offset =3D PAGE_SIZE; > + > + /* Recv queue pages are after send pages. */ > + qp->rq.offset =3D qp->npages_send * PAGE_SIZE; > + } > + > + if (qp->npages < 0 || qp->npages > PVRDMA_PAGE_DIR_MAX_PAGES) { > + dev_warn(&dev->pdev->dev, > + "overflow pages in queuepair\n"); > + goto err_umem; > + } > + > + ret =3D pvrdma_page_dir_init(dev, &qp->pdir, qp->npages, > + qp->is_kernel); > + if (ret) { > + dev_warn(&dev->pdev->dev, > + "could not allocate page directory\n"); > + goto err_umem; > + } > + > + if (!qp->is_kernel) { > + pvrdma_page_dir_insert_umem(&qp->pdir, qp->sumem, 0); > + pvrdma_page_dir_insert_umem(&qp->pdir, qp->rumem, > + qp->npages_send); > + } else { > + /* Ring state is always the first page. */ > + qp->sq.ring =3D qp->pdir.pages[0]; > + qp->rq.ring =3D &qp->sq.ring[1]; > + } > + break; > + default: > + atomic_dec(&dev->num_qps); > + return ERR_PTR(-EINVAL); > + } > + > + /* Not supported */ > + init_attr->cap.max_inline_data =3D 0; > + > + memset(cmd, 0, sizeof(*cmd)); > + cmd->hdr.cmd =3D PVRDMA_CMD_CREATE_QP; > + cmd->pd_handle =3D to_vpd(pd)->pd_handle; > + cmd->send_cq_handle =3D to_vcq(init_attr->send_cq)->cq_handle; > + cmd->recv_cq_handle =3D to_vcq(init_attr->recv_cq)->cq_handle; > + cmd->srq_handle =3D 0; > + cmd->max_send_wr =3D init_attr->cap.max_send_wr; > + cmd->max_recv_wr =3D init_attr->cap.max_recv_wr; > + cmd->max_send_sge =3D init_attr->cap.max_send_sge; > + cmd->max_recv_sge =3D init_attr->cap.max_recv_sge; > + cmd->max_inline_data =3D init_attr->cap.max_inline_data; > + cmd->sq_sig_all =3D (init_attr->sq_sig_type =3D=3D IB_SIGNAL_ALL_WR) ? = 1 : 0; > + cmd->qp_type =3D ib_qp_type_to_pvrdma(init_attr->qp_type); > + cmd->is_srq =3D 0; /* No support */ > + cmd->lkey =3D 0; > + cmd->access_flags =3D IB_ACCESS_LOCAL_WRITE; > + cmd->total_chunks =3D qp->npages; > + cmd->send_chunks =3D qp->npages_send - 1; > + cmd->pdir_dma =3D qp->pdir.dir_dma; > + > + dev_dbg(&dev->pdev->dev, "create queuepair with %d, %d, %d, %d\n", > + cmd->max_send_wr, cmd->max_recv_wr, cmd->max_send_sge, > + cmd->max_recv_sge); > + > + ret =3D pvrdma_cmd_post(dev, &req, &rsp); > + > + if (ret < 0 || resp->hdr.ack !=3D PVRDMA_CMD_CREATE_QP_RESP) { > + dev_warn(&dev->pdev->dev, "could not create queuepair\n"); > + goto err_pdir; > + } > + > + /* max_send_wr/_recv_wr/_send_sge/_recv_sge/_inline_data */ > + qp->qp_handle =3D resp->qpn; > + qp->port =3D init_attr->port_num; > + qp->ibqp.qp_num =3D resp->qpn; > + spin_lock_irqsave(&dev->qp_tbl_lock, flags); > + dev->qp_tbl[qp->qp_handle % dev->dsr->caps.max_qp] =3D qp; > + spin_unlock_irqrestore(&dev->qp_tbl_lock, flags); > + > + return &qp->ibqp; > + > +err_pdir: > + pvrdma_page_dir_cleanup(dev, &qp->pdir); > +err_umem: > + if (pd->uobject && udata) { > + if (qp->rumem) > + ib_umem_release(qp->rumem); > + if (qp->sumem) > + ib_umem_release(qp->sumem); > + } > +err_qp: > + kfree(qp); > + atomic_dec(&dev->num_qps); > + > + return ERR_PTR(ret); > +} > + > +static void pvrdma_free_qp(struct pvrdma_qp *qp) > +{ > + struct pvrdma_dev *dev =3D to_vdev(qp->ibqp.device); > + struct pvrdma_cq *scq; > + struct pvrdma_cq *rcq; > + unsigned long flags, flags1, flags2; > + > + /* In case cq is polling */ > + get_cqs(qp, &scq, &rcq); > + if ((uintptr_t)scq <=3D (uintptr_t)rcq) { > + spin_lock_irqsave(&scq->cq_lock, flags1); > + pvrdma_flush_cqe(qp, scq); > + if (rcq !=3D scq) { > + spin_lock_irqsave(&rcq->cq_lock, flags2); > + pvrdma_flush_cqe(qp, rcq); > + } > + } else { > + spin_lock_irqsave(&rcq->cq_lock, flags2); > + pvrdma_flush_cqe(qp, rcq); > + spin_lock_irqsave(&scq->cq_lock, flags1); > + pvrdma_flush_cqe(qp, scq); > + } > + > + spin_lock_irqsave(&dev->qp_tbl_lock, flags); > + dev->qp_tbl[qp->qp_handle] =3D NULL; > + spin_unlock_irqrestore(&dev->qp_tbl_lock, flags); > + > + if ((uintptr_t)scq <=3D (uintptr_t)rcq) { > + if (rcq !=3D scq) > + spin_unlock_irqrestore(&rcq->cq_lock, flags2); > + spin_unlock_irqrestore(&scq->cq_lock, flags1); > + } else { > + spin_unlock_irqrestore(&scq->cq_lock, flags1); > + spin_unlock_irqrestore(&rcq->cq_lock, flags2); > + } > + > + atomic_dec(&qp->refcnt); > + wait_event(qp->wait, !atomic_read(&qp->refcnt)); > + > + pvrdma_page_dir_cleanup(dev, &qp->pdir); > + > + kfree(qp); > + > + atomic_dec(&dev->num_qps); > +} > + > +/** > + * pvrdma_destroy_qp - destroy a queue pair > + * @qp: the queue pair to destroy > + * > + * @return: 0 on success. > + */ > +int pvrdma_destroy_qp(struct ib_qp *qp) > +{ > + struct pvrdma_qp *vqp =3D to_vqp(qp); > + union pvrdma_cmd_req req; > + struct pvrdma_cmd_destroy_qp *cmd =3D &req.destroy_qp; > + int ret; > + > + memset(cmd, 0, sizeof(*cmd)); > + cmd->hdr.cmd =3D PVRDMA_CMD_DESTROY_QP; > + cmd->qp_handle =3D vqp->qp_handle; > + > + ret =3D pvrdma_cmd_post(to_vdev(qp->device), &req, NULL); > + if (ret < 0) { > + struct pvrdma_dev *dev =3D to_vdev(qp->device); > + > + dev_warn(&dev->pdev->dev, "destroy queuepair failed\n"); > + } > + > + pvrdma_free_qp(vqp); > + > + return 0; > +} > + > +/** > + * pvrdma_modify_qp - modify queue pair attributes > + * @ibqp: the queue pair > + * @attr: the new queue pair's attributes > + * @attr_mask: attributes mask > + * @udata: user data > + * > + * @returns 0 on success, otherwise returns an errno. > + */ > +int pvrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, > + int attr_mask, struct ib_udata *udata) > +{ > + struct pvrdma_dev *dev =3D to_vdev(ibqp->device); > + struct pvrdma_qp *qp =3D to_vqp(ibqp); > + union pvrdma_cmd_req req; > + union pvrdma_cmd_resp rsp; > + struct pvrdma_cmd_modify_qp *cmd =3D &req.modify_qp; > + int cur_state, next_state; > + int ret; > + > + /* Sanity checking. Should need lock here */ > + mutex_lock(&qp->mutex); > + cur_state =3D (attr_mask & IB_QP_CUR_STATE) ? attr->cur_qp_state : > + qp->state; > + next_state =3D (attr_mask & IB_QP_STATE) ? attr->qp_state : cur_state; > + > + if (!ib_modify_qp_is_ok(cur_state, next_state, ibqp->qp_type, > + attr_mask, IB_LINK_LAYER_ETHERNET)) { > + ret =3D -EINVAL; > + goto out; > + } > + > + if (attr_mask & IB_QP_PORT) { > + if (attr->port_num =3D=3D 0 || > + attr->port_num > ibqp->device->phys_port_cnt) { > + ret =3D -EINVAL; > + goto out; > + } > + } > + > + if (attr_mask & IB_QP_MIN_RNR_TIMER) { > + if (attr->min_rnr_timer > 31) { > + ret =3D -EINVAL; > + goto out; > + } > + } > + > + if (attr_mask & IB_QP_PKEY_INDEX) { > + if (attr->pkey_index >=3D dev->dsr->caps.max_pkeys) { > + ret =3D -EINVAL; > + goto out; > + } > + } > + > + if (attr_mask & IB_QP_QKEY) > + qp->qkey =3D attr->qkey; > + > + if (cur_state =3D=3D next_state && cur_state =3D=3D IB_QPS_RESET) { > + ret =3D 0; > + goto out; > + } > + > + switch (next_state) { > + case IB_QPS_RESET: > + case IB_QPS_INIT: > + case IB_QPS_RTR: > + case IB_QPS_RTS: > + case IB_QPS_SQD: > + case IB_QPS_SQE: > + case IB_QPS_ERR: > + default: Why do you need such switch-case? > + qp->state =3D next_state; > + memset(cmd, 0, sizeof(*cmd)); > + cmd->hdr.cmd =3D PVRDMA_CMD_MODIFY_QP; > + cmd->qp_handle =3D qp->qp_handle; > + cmd->attr_mask =3D ib_qp_attr_mask_to_pvrdma(attr_mask); > + cmd->attrs.qp_state =3D ib_qp_state_to_pvrdma(attr->qp_state); > + cmd->attrs.cur_qp_state =3D > + ib_qp_state_to_pvrdma(attr->cur_qp_state); > + cmd->attrs.path_mtu =3D ib_mtu_to_pvrdma(attr->path_mtu); > + cmd->attrs.path_mig_state =3D > + ib_mig_state_to_pvrdma(attr->path_mig_state); > + cmd->attrs.qkey =3D attr->qkey; > + cmd->attrs.rq_psn =3D attr->rq_psn; > + cmd->attrs.sq_psn =3D attr->sq_psn; > + cmd->attrs.dest_qp_num =3D attr->dest_qp_num; > + cmd->attrs.qp_access_flags =3D > + ib_access_flags_to_pvrdma(attr->qp_access_flags); > + cmd->attrs.pkey_index =3D attr->pkey_index; > + cmd->attrs.alt_pkey_index =3D attr->alt_pkey_index; > + cmd->attrs.en_sqd_async_notify =3D attr->en_sqd_async_notify; > + cmd->attrs.sq_draining =3D attr->sq_draining; > + cmd->attrs.max_rd_atomic =3D attr->max_rd_atomic; > + cmd->attrs.max_dest_rd_atomic =3D attr->max_dest_rd_atomic; > + cmd->attrs.min_rnr_timer =3D attr->min_rnr_timer; > + cmd->attrs.port_num =3D attr->port_num; > + cmd->attrs.timeout =3D attr->timeout; > + cmd->attrs.retry_cnt =3D attr->retry_cnt; > + cmd->attrs.rnr_retry =3D attr->rnr_retry; > + cmd->attrs.alt_port_num =3D attr->alt_port_num; > + cmd->attrs.alt_timeout =3D attr->alt_timeout; > + ib_qp_cap_to_pvrdma(&attr->cap, &cmd->attrs.cap); > + ib_ah_attr_to_pvrdma(&attr->ah_attr, &cmd->attrs.ah_attr); > + ib_ah_attr_to_pvrdma(&attr->alt_ah_attr, > + &cmd->attrs.alt_ah_attr); > + ret =3D pvrdma_cmd_post(dev, &req, &rsp); > + if (ret < 0 || rsp.hdr.ack !=3D PVRDMA_CMD_MODIFY_QP_RESP || > + rsp.hdr.err > 0) { > + dev_warn(&dev->pdev->dev, > + "could not modify queuepair\n"); > + if (ret =3D=3D 0) { > + if (rsp.hdr.ack !=3D PVRDMA_CMD_MODIFY_QP_RESP) > + ret =3D -EINVAL; > + else > + ret =3D rsp.hdr.err; > + } > + } > + > + /* Fall through */ > + break; > + } > + > + if (ret =3D=3D 0 && next_state =3D=3D IB_QPS_RESET) > + pvrdma_reset_qp(qp); > + > +out: > + mutex_unlock(&qp->mutex); > + > + return ret; > +} > + > +static inline void *get_sq_wqe(struct pvrdma_qp *qp, int n) > +{ > + return pvrdma_page_dir_get_ptr(&qp->pdir, > + qp->sq.offset + n * qp->sq.wqe_size); > +} > + > +static inline void *get_rq_wqe(struct pvrdma_qp *qp, int n) > +{ > + return pvrdma_page_dir_get_ptr(&qp->pdir, > + qp->rq.offset + n * qp->rq.wqe_size); > +} > + > +static int set_reg_seg(struct pvrdma_sq_wqe_hdr *wqe_hdr, struct ib_reg_= wr *wr) > +{ > + struct pvrdma_user_mr *mr =3D to_vmr(wr->mr); > + > + wqe_hdr->wr.fast_reg.iova_start =3D mr->ibmr.iova; > + wqe_hdr->wr.fast_reg.pl_pdir_dma =3D mr->pdir.dir_dma; > + wqe_hdr->wr.fast_reg.page_shift =3D mr->page_shift; > + wqe_hdr->wr.fast_reg.page_list_len =3D mr->npages; > + wqe_hdr->wr.fast_reg.length =3D mr->ibmr.length; > + wqe_hdr->wr.fast_reg.access_flags =3D wr->access; > + wqe_hdr->wr.fast_reg.rkey =3D wr->key; > + > + return pvrdma_page_dir_insert_page_list(&mr->pdir, mr->pages, > + mr->npages); > +} > + > +/** > + * pvrdma_post_send - post send work request entries on a QP > + * @ibqp: the QP > + * @wr: work request list to post > + * @bad_wr: the first bad WR returned > + * > + * @return: 0 on success, otherwise errno returned. > + */ > +int pvrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, > + struct ib_send_wr **bad_wr) > +{ > + struct pvrdma_qp *qp =3D to_vqp(ibqp); > + struct pvrdma_dev *dev =3D to_vdev(ibqp->device); > + unsigned long flags; > + struct pvrdma_sq_wqe_hdr *wqe_hdr; > + struct ib_sge *sge; > + int i, index; > + int nreq; > + int ret; > + > + /* > + * In states lower than RTS, we can fail immediately. In other states, > + * just post and let the device figure it out. > + */ > + if (qp->state < IB_QPS_RTS) { > + *bad_wr =3D wr; > + return -EINVAL; > + } > + > + spin_lock_irqsave(&qp->sq.lock, flags); > + > + index =3D pvrdma_idx(&qp->sq.ring->prod_tail, qp->sq.wqe_cnt); > + for (nreq =3D 0; wr; nreq++, wr =3D wr->next) { > + unsigned int tail; > + > + if (unlikely(!pvrdma_idx_ring_has_space( > + qp->sq.ring, qp->sq.wqe_cnt, &tail))) { > + dev_warn(&dev->pdev->dev, "send queue is full\n"); > + *bad_wr =3D wr; > + ret =3D -ENOMEM; > + goto out; > + } > + > + if (unlikely(wr->num_sge > qp->sq.max_gs || wr->num_sge < 0)) { > + dev_warn(&dev->pdev->dev, "send SGE overflow\n"); > + *bad_wr =3D wr; > + ret =3D -EINVAL; > + goto out; > + } > + > + if (unlikely(wr->opcode < 0)) { > + dev_warn(&dev->pdev->dev, "invalid send opcode\n"); > + *bad_wr =3D wr; > + ret =3D -EINVAL; > + goto out; > + } > + > + /* > + * Only support UD, RC. > + * Need to check opcode table for thorough checking. > + * opcode _UD _UC _RC > + * _SEND x x x > + * _SEND_WITH_IMM x x x > + * _RDMA_WRITE x x > + * _RDMA_WRITE_WITH_IMM x x > + * _LOCAL_INV x x > + * _SEND_WITH_INV x x > + * _RDMA_READ x > + * _ATOMIC_CMP_AND_SWP x > + * _ATOMIC_FETCH_AND_ADD x > + * _MASK_ATOMIC_CMP_AND_SWP x > + * _MASK_ATOMIC_FETCH_AND_ADD x > + * _REG_MR x > + * > + */ > + if (qp->ibqp.qp_type !=3D IB_QPT_UD && > + qp->ibqp.qp_type !=3D IB_QPT_RC && > + wr->opcode !=3D IB_WR_SEND) { > + dev_warn(&dev->pdev->dev, > + "unsupported queuepair type\n"); > + *bad_wr =3D wr; > + ret =3D -EINVAL; > + goto out; > + } else if (qp->ibqp.qp_type =3D=3D IB_QPT_UD || > + qp->ibqp.qp_type =3D=3D IB_QPT_GSI) { > + if (wr->opcode !=3D IB_WR_SEND && > + wr->opcode !=3D IB_WR_SEND_WITH_IMM) { > + dev_warn(&dev->pdev->dev, > + "invalid send opcode\n"); > + *bad_wr =3D wr; > + ret =3D -EINVAL; > + goto out; > + } > + } > + > + wqe_hdr =3D (struct pvrdma_sq_wqe_hdr *)get_sq_wqe(qp, index); > + memset(wqe_hdr, 0, sizeof(*wqe_hdr)); > + wqe_hdr->wr_id =3D wr->wr_id; > + wqe_hdr->num_sge =3D wr->num_sge; > + wqe_hdr->opcode =3D ib_wr_opcode_to_pvrdma(wr->opcode); > + wqe_hdr->send_flags =3D ib_send_flags_to_pvrdma(wr->send_flags); > + if (wr->opcode =3D=3D IB_WR_SEND_WITH_IMM || > + wr->opcode =3D=3D IB_WR_RDMA_WRITE_WITH_IMM) > + wqe_hdr->ex.imm_data =3D wr->ex.imm_data; > + > + switch (qp->ibqp.qp_type) { > + case IB_QPT_GSI: > + case IB_QPT_UD: > + if (unlikely(!ud_wr(wr)->ah)) { > + dev_warn(&dev->pdev->dev, > + "invalid address handle\n"); > + *bad_wr =3D wr; > + ret =3D -EINVAL; > + goto out; > + } > + > + /* > + * Use qkey from qp context if high order bit set, > + * otherwise from work request. > + */ > + wqe_hdr->wr.ud.remote_qpn =3D ud_wr(wr)->remote_qpn; > + wqe_hdr->wr.ud.remote_qkey =3D > + ud_wr(wr)->remote_qkey & 0x80000000 ? > + qp->qkey : ud_wr(wr)->remote_qkey; > + wqe_hdr->wr.ud.av =3D to_vah(ud_wr(wr)->ah)->av; > + > + break; > + case IB_QPT_RC: > + switch (wr->opcode) { > + case IB_WR_RDMA_READ: > + case IB_WR_RDMA_WRITE: > + case IB_WR_RDMA_WRITE_WITH_IMM: > + wqe_hdr->wr.rdma.remote_addr =3D > + rdma_wr(wr)->remote_addr; > + wqe_hdr->wr.rdma.rkey =3D rdma_wr(wr)->rkey; > + break; > + case IB_WR_LOCAL_INV: > + case IB_WR_SEND_WITH_INV: > + wqe_hdr->ex.invalidate_rkey =3D > + wr->ex.invalidate_rkey; > + break; > + case IB_WR_ATOMIC_CMP_AND_SWP: > + case IB_WR_ATOMIC_FETCH_AND_ADD: > + wqe_hdr->wr.atomic.remote_addr =3D > + atomic_wr(wr)->remote_addr; > + wqe_hdr->wr.atomic.rkey =3D atomic_wr(wr)->rkey; > + wqe_hdr->wr.atomic.compare_add =3D > + atomic_wr(wr)->compare_add; > + if (wr->opcode =3D=3D IB_WR_ATOMIC_CMP_AND_SWP) > + wqe_hdr->wr.atomic.swap =3D > + atomic_wr(wr)->swap; > + break; > + case IB_WR_REG_MR: > + ret =3D set_reg_seg(wqe_hdr, reg_wr(wr)); > + if (ret < 0) { > + dev_warn(&dev->pdev->dev, > + "Failed to set fast register work request\n"); > + *bad_wr =3D wr; > + goto out; > + } > + break; > + default: > + break; > + } > + > + break; > + default: > + dev_warn(&dev->pdev->dev, "invalid queuepair type\n"); > + ret =3D -EINVAL; > + *bad_wr =3D wr; > + goto out; > + } > + > + sge =3D (struct ib_sge *)(wqe_hdr + 1); > + for (i =3D 0; i < wr->num_sge; i++) { > + /* Need to check wqe_size 0 or max size */ > + sge->addr =3D wr->sg_list[i].addr; > + sge->length =3D wr->sg_list[i].length; > + sge->lkey =3D wr->sg_list[i].lkey; > + sge++; > + } > + > + /* Make sure wqe is written before index update */ > + smp_wmb(); > + > + index++; > + if (unlikely(index >=3D qp->sq.wqe_cnt)) > + index =3D 0; > + /* Update shared sq ring */ > + pvrdma_idx_ring_inc(&qp->sq.ring->prod_tail, > + qp->sq.wqe_cnt); > + } > + > + ret =3D 0; > + > +out: > + spin_unlock_irqrestore(&qp->sq.lock, flags); > + > + if (!ret) > + writel(cpu_to_le32(PVRDMA_UAR_QP_SEND | qp->qp_handle), > + dev->driver_uar.map + PVRDMA_UAR_QP_OFFSET); > + > + return ret; > +} > + > +/** > + * pvrdma_post_receive - post receive work request entries on a QP > + * @ibqp: the QP > + * @wr: the work request list to post > + * @bad_wr: the first bad WR returned > + * > + * @return: 0 on success, otherwise errno returned. > + */ > +int pvrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr, > + struct ib_recv_wr **bad_wr) > +{ > + struct pvrdma_dev *dev =3D to_vdev(ibqp->device); > + unsigned long flags; > + struct pvrdma_qp *qp =3D to_vqp(ibqp); > + struct pvrdma_rq_wqe_hdr *wqe_hdr; > + struct ib_sge *sge; > + int index, nreq; > + int ret =3D 0; > + int i; > + > + /* > + * In the RESET state, we can fail immediately. For other states, > + * just post and let the device figure it out. > + */ > + if (qp->state =3D=3D IB_QPS_RESET) { > + *bad_wr =3D wr; > + return -EINVAL; > + } > + > + spin_lock_irqsave(&qp->rq.lock, flags); > + > + index =3D pvrdma_idx(&qp->rq.ring->prod_tail, qp->rq.wqe_cnt); > + for (nreq =3D 0; wr; nreq++, wr =3D wr->next) { > + unsigned int tail; > + > + if (unlikely(wr->num_sge > qp->rq.max_gs || > + wr->num_sge < 0)) { > + ret =3D -EINVAL; > + *bad_wr =3D wr; > + dev_warn(&dev->pdev->dev, "recv SGE overflow\n"); > + goto out; > + } > + > + if (unlikely(!pvrdma_idx_ring_has_space( > + qp->rq.ring, qp->rq.wqe_cnt, &tail))) { > + ret =3D -ENOMEM; > + *bad_wr =3D wr; > + dev_warn(&dev->pdev->dev, "recv queue full\n"); > + goto out; > + } > + > + wqe_hdr =3D (struct pvrdma_rq_wqe_hdr *)get_rq_wqe(qp, index); > + wqe_hdr->wr_id =3D wr->wr_id; > + wqe_hdr->num_sge =3D wr->num_sge; > + wqe_hdr->total_len =3D 0; > + > + sge =3D (struct ib_sge *)(wqe_hdr + 1); > + for (i =3D 0; i < wr->num_sge; i++) { > + sge->addr =3D wr->sg_list[i].addr; > + sge->length =3D wr->sg_list[i].length; > + sge->lkey =3D wr->sg_list[i].lkey; > + sge++; > + } > + > + /* Make sure wqe is written before index update */ > + smp_wmb(); > + > + index++; > + if (unlikely(index >=3D qp->rq.wqe_cnt)) > + index =3D 0; > + /* Update shared rq ring */ > + pvrdma_idx_ring_inc(&qp->rq.ring->prod_tail, > + qp->rq.wqe_cnt); > + } > + > + spin_unlock_irqrestore(&qp->rq.lock, flags); > + > + writel(cpu_to_le32(PVRDMA_UAR_QP_RECV | qp->qp_handle), > + dev->driver_uar.map + PVRDMA_UAR_QP_OFFSET); > + > + return ret; > + > +out: > + spin_unlock_irqrestore(&qp->rq.lock, flags); > + > + return ret; > +} > + > +/** > + * pvrdma_query_qp - query a queue pair's attributes > + * @ibqp: the queue pair to query > + * @attr: the queue pair's attributes > + * @attr_mask: attributes mask > + * @init_attr: initial queue pair attributes > + * > + * @returns 0 on success, otherwise returns an errno. > + */ > +int pvrdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, > + int attr_mask, struct ib_qp_init_attr *init_attr) > +{ > + struct pvrdma_dev *dev =3D to_vdev(ibqp->device); > + struct pvrdma_qp *qp =3D to_vqp(ibqp); > + union pvrdma_cmd_req req; > + union pvrdma_cmd_resp rsp; > + struct pvrdma_cmd_query_qp *cmd =3D &req.query_qp; > + struct pvrdma_cmd_query_qp_resp *resp =3D &rsp.query_qp_resp; > + int ret =3D 0; > + > + mutex_lock(&qp->mutex); > + > + if (qp->state =3D=3D IB_QPS_RESET) { > + attr->qp_state =3D IB_QPS_RESET; > + goto out; > + } > + > + memset(cmd, 0, sizeof(*cmd)); > + cmd->hdr.cmd =3D PVRDMA_CMD_QUERY_QP; > + cmd->qp_handle =3D qp->qp_handle; > + cmd->attr_mask =3D ib_qp_attr_mask_to_pvrdma(attr_mask); > + > + ret =3D pvrdma_cmd_post(dev, &req, &rsp); > + if (ret < 0 || resp->hdr.ack !=3D PVRDMA_CMD_QUERY_QP_RESP) { > + dev_warn(&dev->pdev->dev, "could not query queuepair\n"); > + ret =3D -EINVAL; > + goto out; > + } > + > + attr->qp_state =3D pvrdma_qp_state_to_ib(resp->attrs.qp_state); > + attr->cur_qp_state =3D > + pvrdma_qp_state_to_ib(resp->attrs.cur_qp_state); > + attr->path_mtu =3D pvrdma_mtu_to_ib(resp->attrs.path_mtu); > + attr->path_mig_state =3D > + pvrdma_mig_state_to_ib(resp->attrs.path_mig_state); > + attr->qkey =3D resp->attrs.qkey; > + attr->rq_psn =3D resp->attrs.rq_psn; > + attr->sq_psn =3D resp->attrs.sq_psn; > + attr->dest_qp_num =3D resp->attrs.dest_qp_num; > + attr->qp_access_flags =3D > + pvrdma_access_flags_to_ib(resp->attrs.qp_access_flags); > + attr->pkey_index =3D resp->attrs.pkey_index; > + attr->alt_pkey_index =3D resp->attrs.alt_pkey_index; > + attr->en_sqd_async_notify =3D resp->attrs.en_sqd_async_notify; > + attr->sq_draining =3D resp->attrs.sq_draining; > + attr->max_rd_atomic =3D resp->attrs.max_rd_atomic; > + attr->max_dest_rd_atomic =3D resp->attrs.max_dest_rd_atomic; > + attr->min_rnr_timer =3D resp->attrs.min_rnr_timer; > + attr->port_num =3D resp->attrs.port_num; > + attr->timeout =3D resp->attrs.timeout; > + attr->retry_cnt =3D resp->attrs.retry_cnt; > + attr->rnr_retry =3D resp->attrs.rnr_retry; > + attr->alt_port_num =3D resp->attrs.alt_port_num; > + attr->alt_timeout =3D resp->attrs.alt_timeout; > + pvrdma_qp_cap_to_ib(&resp->attrs.cap, &attr->cap); > + pvrdma_ah_attr_to_ib(&resp->attrs.ah_attr, &attr->ah_attr); > + pvrdma_ah_attr_to_ib(&resp->attrs.alt_ah_attr, &attr->alt_ah_attr); > + > + qp->state =3D attr->qp_state; > + > + ret =3D 0; > + > +out: > + attr->cur_qp_state =3D attr->qp_state; > + > + init_attr->event_handler =3D qp->ibqp.event_handler; > + init_attr->qp_context =3D qp->ibqp.qp_context; > + init_attr->send_cq =3D qp->ibqp.send_cq; > + init_attr->recv_cq =3D qp->ibqp.recv_cq; > + init_attr->srq =3D qp->ibqp.srq; > + init_attr->xrcd =3D NULL; > + init_attr->cap =3D attr->cap; > + init_attr->sq_sig_type =3D 0; > + init_attr->qp_type =3D qp->ibqp.qp_type; > + init_attr->create_flags =3D 0; > + init_attr->port_num =3D qp->port; > + > + mutex_unlock(&qp->mutex); > + return ret; > +} > --=20 > 2.7.4 >=20 > -- > To unsubscribe from this list: send the line "unsubscribe linux-rdma" in > the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org > More majordomo info at http://vger.kernel.org/majordomo-info.html --XRI2XbIfl/05pQwm Content-Type: application/pgp-signature; name="signature.asc" Content-Description: Digital signature -----BEGIN PGP SIGNATURE----- Version: GnuPG v1 iQIcBAEBAgAGBQJXqqqoAAoJEORje4g2clinEhUQAKeMcjDztsE9atdU78nAnE6v HOZEVS3xH41ikGg2S7fvstj7MX+p8Lr98JzKGalOduES5uUgzKRyg+x8HA2VnClm bhDU/df+bR0aunm8/02aKHE481+rl4evHfd+IxuXqPLfCMvNZeGLrx/7MiyDoRkU SIM2jr+44XokHoPPebZ0lWDbf+lO9e2MX4yEEBDa4F90RGwWY5tZMOp6eGPfYRTj YplZpEUJPfFmee4SciXM0q1oq87HGY+qv6u+tzS9AfLhhA4PKaGYJPAfjfphyfV/ vwMjtHXbuQNg7+BwP/vy1OjExEBmpemtjqtSoACDGe0+PiyIWTYFCcfkL5jQ9hNO zOlvg19qgkpvQVaM3l942otPoRlVHo74invndJvJcXu+4fD6u8kuIMMJRbCYHY+J 1w3dGelkUe28PhgE0OFDAhWX3RMoASU8THu2nkgeHmEDfRr7CF8s4ngAA6q89Nqb Ll+c26FKmqFS2jqxEn/AVJIhrpoXdfB2nPJFyF9nAzShIhJKh6vYGtji0pJSCzIT Ok1FqR4gVVOa3N8qcwwqiOxa8CNEs6Ky18YeNlLzPZRRStheUStdIhcZqWRIVxVR cO/Fm1u3GbYf30xbQ4v7wlrRX//hzc3NDAGHi184qoK+84JbqqN9v5oj3Sl7AIDy wvmIYQ+Q39QjNlsR7cit =xC3H -----END PGP SIGNATURE----- --XRI2XbIfl/05pQwm-- -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org More majordomo info at http://vger.kernel.org/majordomo-info.html