From patchwork Wed Jan 6 18:15:11 2016 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Dennis Dalessandro X-Patchwork-Id: 7969791 Return-Path: X-Original-To: patchwork-linux-rdma@patchwork.kernel.org Delivered-To: patchwork-parsemail@patchwork2.web.kernel.org Received: from mail.kernel.org (mail.kernel.org [198.145.29.136]) by patchwork2.web.kernel.org (Postfix) with ESMTP id 6A590BEEE5 for ; Wed, 6 Jan 2016 18:15:30 +0000 (UTC) Received: from mail.kernel.org (localhost [127.0.0.1]) by mail.kernel.org (Postfix) with ESMTP id E8E772013D for ; Wed, 6 Jan 2016 18:15:28 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by mail.kernel.org (Postfix) with ESMTP id 35BFE20123 for ; Wed, 6 Jan 2016 18:15:27 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752067AbcAFSP0 (ORCPT ); Wed, 6 Jan 2016 13:15:26 -0500 Received: from mga09.intel.com ([134.134.136.24]:22019 "EHLO mga09.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751863AbcAFSPZ (ORCPT ); Wed, 6 Jan 2016 13:15:25 -0500 Received: from fmsmga003.fm.intel.com ([10.253.24.29]) by orsmga102.jf.intel.com with ESMTP; 06 Jan 2016 10:15:12 -0800 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.20,530,1444719600"; d="scan'208";a="629028678" Received: from scymds02.sc.intel.com ([10.82.195.37]) by FMSMGA003.fm.intel.com with ESMTP; 06 Jan 2016 10:15:11 -0800 Received: from scvm10.sc.intel.com (scvm10.sc.intel.com [10.82.195.27]) by scymds02.sc.intel.com with ESMTP id u06IFB6a028027; Wed, 6 Jan 2016 10:15:11 -0800 Received: from scvm10.sc.intel.com (localhost [127.0.0.1]) by scvm10.sc.intel.com with ESMTP id u06IFBt4004323; Wed, 6 Jan 2016 10:15:11 -0800 Subject: [PATCH 3/6] IB/rdmavt: Add create queue pair functionality To: dledford@redhat.com From: Dennis Dalessandro Cc: linux-rdma@vger.kernel.org, Harish Chegondi , Ira Weiny Date: Wed, 06 Jan 2016 10:15:11 -0800 Message-ID: <20160106181511.3238.72564.stgit@scvm10.sc.intel.com> In-Reply-To: <20160106181226.3238.98769.stgit@scvm10.sc.intel.com> References: <20160106181226.3238.98769.stgit@scvm10.sc.intel.com> User-Agent: StGit/0.16 MIME-Version: 1.0 Sender: linux-rdma-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: linux-rdma@vger.kernel.org X-Spam-Status: No, score=-6.9 required=5.0 tests=BAYES_00, RCVD_IN_DNSWL_HI, RP_MATCHES_RCVD, UNPARSEABLE_RELAY autolearn=unavailable version=3.3.1 X-Spam-Checker-Version: SpamAssassin 3.3.1 (2010-03-16) on mail.kernel.org X-Virus-Scanned: ClamAV using ClamSMTP Add create queue pair verbs call as well as supporting functions. Reviewed-by: Ira Weiny Reviewed-by: Harish Chegondi Signed-off-by: Dennis Dalessandro --- drivers/infiniband/sw/rdmavt/qp.c | 425 +++++++++++++++++++++++++++++++++++-- drivers/infiniband/sw/rdmavt/vt.c | 1 include/rdma/rdma_vt.h | 10 + 3 files changed, 413 insertions(+), 23 deletions(-) -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html diff --git a/drivers/infiniband/sw/rdmavt/qp.c b/drivers/infiniband/sw/rdmavt/qp.c index 17dd6ab..7d1f02e 100644 --- a/drivers/infiniband/sw/rdmavt/qp.c +++ b/drivers/infiniband/sw/rdmavt/qp.c @@ -47,8 +47,11 @@ #include #include -#include "vt.h" +#include +#include +#include #include "qp.h" +#include "vt.h" static void get_map_page(struct rvt_qpn_table *qpt, struct rvt_qpn_map *map) { @@ -151,7 +154,10 @@ int rvt_driver_qp_init(struct rvt_dev_info *rdi) * If driver is not doing any QP allocation then make sure it is * providing the necessary QP functions. */ - if (!rdi->driver_f.free_all_qps) + if (!rdi->driver_f.free_all_qps || + !rdi->driver_f.qp_priv_alloc || + !rdi->driver_f.qp_priv_free || + !rdi->driver_f.notify_qp_reset) return -EINVAL; /* allocate parent object */ @@ -178,7 +184,9 @@ int rvt_driver_qp_init(struct rvt_dev_info *rdi) if (init_qpn_table(rdi, &rdi->qp_dev->qpn_table)) goto fail_table; - return ret; + spin_lock_init(&rdi->n_qps_lock); + + return 0; fail_table: kfree(rdi->qp_dev->qp_table); @@ -197,31 +205,29 @@ no_qp_table: * There should not be any QPs still in use. * Free memory for table. */ -static unsigned free_all_qps(struct rvt_dev_info *rdi) +static unsigned rvt_free_all_qps(struct rvt_dev_info *rdi) { unsigned long flags; struct rvt_qp *qp; unsigned n, qp_inuse = 0; spinlock_t *ql; /* work around too long line below */ - rdi->driver_f.free_all_qps(rdi); + if (rdi->driver_f.free_all_qps) + qp_inuse = rdi->driver_f.free_all_qps(rdi); if (!rdi->qp_dev) - return 0; + return qp_inuse; ql = &rdi->qp_dev->qpt_lock; - spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags); + spin_lock_irqsave(ql, flags); for (n = 0; n < rdi->qp_dev->qp_table_size; n++) { qp = rcu_dereference_protected(rdi->qp_dev->qp_table[n], lockdep_is_held(ql)); RCU_INIT_POINTER(rdi->qp_dev->qp_table[n], NULL); - qp = rcu_dereference_protected(qp->next, - lockdep_is_held(ql)); - while (qp) { + + for (; qp; qp = rcu_dereference_protected(qp->next, + lockdep_is_held(ql))) qp_inuse++; - qp = rcu_dereference_protected(qp->next, - lockdep_is_held(ql)); - } } spin_unlock_irqrestore(ql, flags); synchronize_rcu(); @@ -230,26 +236,190 @@ static unsigned free_all_qps(struct rvt_dev_info *rdi) void rvt_qp_exit(struct rvt_dev_info *rdi) { - u32 qps_inuse = free_all_qps(rdi); + u32 qps_inuse = rvt_free_all_qps(rdi); - qps_inuse = free_all_qps(rdi); if (qps_inuse) rvt_pr_err(rdi, "QP memory leak! %u still in use\n", qps_inuse); if (!rdi->qp_dev) return; + if (rdi->flags & RVT_FLAG_QP_INIT_DRIVER) + return; /* driver did the qp init so nothing else to do */ + kfree(rdi->qp_dev->qp_table); free_qpn_table(&rdi->qp_dev->qpn_table); kfree(rdi->qp_dev); } +static inline unsigned mk_qpn(struct rvt_qpn_table *qpt, + struct rvt_qpn_map *map, unsigned off) +{ + return (map - qpt->map) * RVT_BITS_PER_PAGE + off; +} + +/* + * Allocate the next available QPN or + * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI. + */ +static int alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt, + enum ib_qp_type type, u8 port) +{ + u32 i, offset, max_scan, qpn; + struct rvt_qpn_map *map; + u32 ret; + + if (rdi->driver_f.alloc_qpn) + return rdi->driver_f.alloc_qpn(rdi, qpt, type, port); + + if (type == IB_QPT_SMI || type == IB_QPT_GSI) { + unsigned n; + + ret = type == IB_QPT_GSI; + n = 1 << (ret + 2 * (port - 1)); + spin_lock(&qpt->lock); + if (qpt->flags & n) + ret = -EINVAL; + else + qpt->flags |= n; + spin_unlock(&qpt->lock); + goto bail; + } + + qpn = qpt->last + qpt->incr; + if (qpn >= RVT_QPN_MAX) + qpn = qpt->incr | ((qpt->last & 1) ^ 1); + /* offset carries bit 0 */ + offset = qpn & RVT_BITS_PER_PAGE_MASK; + map = &qpt->map[qpn / RVT_BITS_PER_PAGE]; + max_scan = qpt->nmaps - !offset; + for (i = 0;;) { + if (unlikely(!map->page)) { + get_map_page(qpt, map); + if (unlikely(!map->page)) + break; + } + do { + if (!test_and_set_bit(offset, map->page)) { + qpt->last = qpn; + ret = qpn; + goto bail; + } + offset += qpt->incr; + /* + * This qpn might be bogus if offset >= BITS_PER_PAGE. + * That is OK. It gets re-assigned below + */ + qpn = mk_qpn(qpt, map, offset); + } while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX); + /* + * In order to keep the number of pages allocated to a + * minimum, we scan the all existing pages before increasing + * the size of the bitmap table. + */ + if (++i > max_scan) { + if (qpt->nmaps == RVT_QPNMAP_ENTRIES) + break; + map = &qpt->map[qpt->nmaps++]; + /* start at incr with current bit 0 */ + offset = qpt->incr | (offset & 1); + } else if (map < &qpt->map[qpt->nmaps]) { + ++map; + /* start at incr with current bit 0 */ + offset = qpt->incr | (offset & 1); + } else { + map = &qpt->map[0]; + /* wrap to first map page, invert bit 0 */ + offset = qpt->incr | ((offset & 1) ^ 1); + } + /* there can be no bits at shift and below */ + WARN_ON(offset & (rdi->dparms.qos_shift - 1)); + qpn = mk_qpn(qpt, map, offset); + } + + ret = -ENOMEM; + +bail: + return ret; +} + +static void free_qpn(struct rvt_qpn_table *qpt, u32 qpn) +{ + struct rvt_qpn_map *map; + + map = qpt->map + qpn / RVT_BITS_PER_PAGE; + if (map->page) + clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page); +} + +/** + * reset_qp - initialize the QP state to the reset state + * @qp: the QP to reset + * @type: the QP type + */ +static void reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, + enum ib_qp_type type) +{ + qp->remote_qpn = 0; + qp->qkey = 0; + qp->qp_access_flags = 0; + + /* + * Let driver do anything it needs to for a new/reset qp + */ + rdi->driver_f.notify_qp_reset(qp); + + qp->s_flags &= RVT_S_SIGNAL_REQ_WR; + qp->s_hdrwords = 0; + qp->s_wqe = NULL; + qp->s_draining = 0; + qp->s_next_psn = 0; + qp->s_last_psn = 0; + qp->s_sending_psn = 0; + qp->s_sending_hpsn = 0; + qp->s_psn = 0; + qp->r_psn = 0; + qp->r_msn = 0; + if (type == IB_QPT_RC) { + qp->s_state = IB_OPCODE_RC_SEND_LAST; + qp->r_state = IB_OPCODE_RC_SEND_LAST; + } else { + qp->s_state = IB_OPCODE_UC_SEND_LAST; + qp->r_state = IB_OPCODE_UC_SEND_LAST; + } + qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE; + qp->r_nak_state = 0; + qp->r_aflags = 0; + qp->r_flags = 0; + qp->s_head = 0; + qp->s_tail = 0; + qp->s_cur = 0; + qp->s_acked = 0; + qp->s_last = 0; + qp->s_ssn = 1; + qp->s_lsn = 0; + qp->s_mig_state = IB_MIG_MIGRATED; + memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue)); + qp->r_head_ack_queue = 0; + qp->s_tail_ack_queue = 0; + qp->s_num_rd_atomic = 0; + if (qp->r_rq.wq) { + qp->r_rq.wq->head = 0; + qp->r_rq.wq->tail = 0; + } + qp->r_sge.num_sge = 0; +} + /** * rvt_create_qp - create a queue pair for a device * @ibpd: the protection domain who's device we create the queue pair for * @init_attr: the attributes of the queue pair * @udata: user data for libibverbs.so * + * Queue pair creation is mostly an rvt issue. However, drivers have their own + * unique idea of what queue pair numbers mean. For instance there is a reserved + * range for PSM. + * * Returns the queue pair on success, otherwise returns an errno. * * Called by the ib_create_qp() core verbs function. @@ -258,15 +428,226 @@ struct ib_qp *rvt_create_qp(struct ib_pd *ibpd, struct ib_qp_init_attr *init_attr, struct ib_udata *udata) { + struct rvt_qp *qp; + int err; + struct rvt_swqe *swq = NULL; + size_t sz; + size_t sg_list_sz; + struct ib_qp *ret = ERR_PTR(-ENOMEM); + struct rvt_dev_info *rdi = ib_to_rvt(ibpd->device); + void *priv = NULL; + + if (!rdi) + return ERR_PTR(-EINVAL); + + if (init_attr->cap.max_send_sge > rdi->dparms.props.max_sge || + init_attr->cap.max_send_wr > rdi->dparms.props.max_qp_wr || + init_attr->create_flags) + return ERR_PTR(-EINVAL); + + /* Check receive queue parameters if no SRQ is specified. */ + if (!init_attr->srq) { + if (init_attr->cap.max_recv_sge > rdi->dparms.props.max_sge || + init_attr->cap.max_recv_wr > rdi->dparms.props.max_qp_wr) + return ERR_PTR(-EINVAL); + + if (init_attr->cap.max_send_sge + + init_attr->cap.max_send_wr + + init_attr->cap.max_recv_sge + + init_attr->cap.max_recv_wr == 0) + return ERR_PTR(-EINVAL); + } + + switch (init_attr->qp_type) { + case IB_QPT_SMI: + case IB_QPT_GSI: + if (init_attr->port_num == 0 || + init_attr->port_num > ibpd->device->phys_port_cnt) + return ERR_PTR(-EINVAL); + case IB_QPT_UC: + case IB_QPT_RC: + case IB_QPT_UD: + sz = sizeof(struct rvt_sge) * + init_attr->cap.max_send_sge + + sizeof(struct rvt_swqe); + swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz); + if (!swq) + return ERR_PTR(-ENOMEM); + + sz = sizeof(*qp); + sg_list_sz = 0; + if (init_attr->srq) { + struct rvt_srq *srq = ibsrq_to_rvtsrq(init_attr->srq); + + if (srq->rq.max_sge > 1) + sg_list_sz = sizeof(*qp->r_sg_list) * + (srq->rq.max_sge - 1); + } else if (init_attr->cap.max_recv_sge > 1) + sg_list_sz = sizeof(*qp->r_sg_list) * + (init_attr->cap.max_recv_sge - 1); + qp = kzalloc(sz + sg_list_sz, GFP_KERNEL); + if (!qp) + goto bail_swq; + + RCU_INIT_POINTER(qp->next, NULL); + + /* + * Driver needs to set up it's private QP structure and do any + * initialization that is needed. + */ + priv = rdi->driver_f.qp_priv_alloc(rdi, qp); + if (!priv) + goto bail_qp; + qp->priv = priv; + qp->timeout_jiffies = + usecs_to_jiffies((4096UL * (1UL << qp->timeout)) / + 1000UL); + if (init_attr->srq) { + sz = 0; + } else { + qp->r_rq.size = init_attr->cap.max_recv_wr + 1; + qp->r_rq.max_sge = init_attr->cap.max_recv_sge; + sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) + + sizeof(struct rvt_rwqe); + qp->r_rq.wq = vmalloc_user(sizeof(struct rvt_rwq) + + qp->r_rq.size * sz); + if (!qp->r_rq.wq) + goto bail_driver_priv; + } + + /* + * ib_create_qp() will initialize qp->ibqp + * except for qp->ibqp.qp_num. + */ + spin_lock_init(&qp->r_lock); + spin_lock_init(&qp->s_lock); + spin_lock_init(&qp->r_rq.lock); + atomic_set(&qp->refcount, 0); + init_waitqueue_head(&qp->wait); + init_timer(&qp->s_timer); + qp->s_timer.data = (unsigned long)qp; + INIT_LIST_HEAD(&qp->rspwait); + qp->state = IB_QPS_RESET; + qp->s_wq = swq; + qp->s_size = init_attr->cap.max_send_wr + 1; + qp->s_max_sge = init_attr->cap.max_send_sge; + if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR) + qp->s_flags = RVT_S_SIGNAL_REQ_WR; + + err = alloc_qpn(rdi, &rdi->qp_dev->qpn_table, + init_attr->qp_type, + init_attr->port_num); + if (err < 0) { + ret = ERR_PTR(err); + goto bail_rq_wq; + } + qp->ibqp.qp_num = err; + qp->port_num = init_attr->port_num; + reset_qp(rdi, qp, init_attr->qp_type); + break; + + default: + /* Don't support raw QPs */ + return ERR_PTR(-EINVAL); + } + + init_attr->cap.max_inline_data = 0; + /* - * Queue pair creation is mostly an rvt issue. However, drivers have - * their own unique idea of what queue pare numbers mean. For instance - * there is a reserved range for PSM. - * - * VI-DRIVER-API: make_qpn() - * Returns a valid QPN for verbs to use + * Return the address of the RWQ as the offset to mmap. + * See hfi1_mmap() for details. */ - return ERR_PTR(-EOPNOTSUPP); + if (udata && udata->outlen >= sizeof(__u64)) { + if (!qp->r_rq.wq) { + __u64 offset = 0; + + err = ib_copy_to_udata(udata, &offset, + sizeof(offset)); + if (err) { + ret = ERR_PTR(err); + goto bail_qpn; + } + } else { + u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz; + + qp->ip = rvt_create_mmap_info(rdi, s, + ibpd->uobject->context, + qp->r_rq.wq); + if (!qp->ip) { + ret = ERR_PTR(-ENOMEM); + goto bail_qpn; + } + + err = ib_copy_to_udata(udata, &qp->ip->offset, + sizeof(qp->ip->offset)); + if (err) { + ret = ERR_PTR(err); + goto bail_ip; + } + } + } + + spin_lock(&rdi->n_qps_lock); + if (rdi->n_qps_allocated == rdi->dparms.props.max_qp) { + spin_unlock(&rdi->n_qps_lock); + ret = ERR_PTR(-ENOMEM); + goto bail_ip; + } + + rdi->n_qps_allocated++; + spin_unlock(&rdi->n_qps_lock); + + if (qp->ip) { + spin_lock_irq(&rdi->pending_lock); + list_add(&qp->ip->pending_mmaps, &rdi->pending_mmaps); + spin_unlock_irq(&rdi->pending_lock); + } + + ret = &qp->ibqp; + + /* + * We have our QP and its good, now keep track of what types of opcodes + * can be processed on this QP. We do this by keeping track of what the + * 3 high order bits of the opcode are. + */ + switch (init_attr->qp_type) { + case IB_QPT_SMI: + case IB_QPT_GSI: + case IB_QPT_UD: + qp->allowed_ops = IB_OPCODE_UD_SEND_ONLY & RVT_OPCODE_QP_MASK; + break; + case IB_QPT_RC: + qp->allowed_ops = IB_OPCODE_RC_SEND_ONLY & RVT_OPCODE_QP_MASK; + break; + case IB_QPT_UC: + qp->allowed_ops = IB_OPCODE_UC_SEND_ONLY & RVT_OPCODE_QP_MASK; + break; + default: + ret = ERR_PTR(-EINVAL); + goto bail_ip; + } + + return ret; + +bail_ip: + kref_put(&qp->ip->ref, rvt_release_mmap_info); + +bail_qpn: + free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num); + +bail_rq_wq: + vfree(qp->r_rq.wq); + +bail_driver_priv: + rdi->driver_f.qp_priv_free(rdi, qp); + +bail_qp: + kfree(qp); + +bail_swq: + vfree(swq); + + return ret; } /** diff --git a/drivers/infiniband/sw/rdmavt/vt.c b/drivers/infiniband/sw/rdmavt/vt.c index df2df36..e75eb3d 100644 --- a/drivers/infiniband/sw/rdmavt/vt.c +++ b/drivers/infiniband/sw/rdmavt/vt.c @@ -362,6 +362,7 @@ void rvt_unregister_device(struct rvt_dev_info *rdi) ib_unregister_device(&rdi->ibdev); rvt_mr_exit(rdi); + rvt_qp_exit(rdi); } EXPORT_SYMBOL(rvt_unregister_device); diff --git a/include/rdma/rdma_vt.h b/include/rdma/rdma_vt.h index 3a78f20..3bdeac7 100644 --- a/include/rdma/rdma_vt.h +++ b/include/rdma/rdma_vt.h @@ -222,7 +222,10 @@ struct rvt_driver_provided { int (*port_callback)(struct ib_device *, u8, struct kobject *); const char * (*get_card_name)(struct rvt_dev_info *rdi); struct pci_dev * (*get_pci_dev)(struct rvt_dev_info *rdi); - void (*free_all_qps)(struct rvt_dev_info *rdi); + unsigned (*free_all_qps)(struct rvt_dev_info *rdi); + void * (*qp_priv_alloc)(struct rvt_dev_info *rdi, struct rvt_qp *qp); + void (*qp_priv_free)(struct rvt_dev_info *rdi, struct rvt_qp *qp); + void (*notify_qp_reset)(struct rvt_qp *qp); /*--------------------*/ /* Optional functions */ @@ -230,6 +233,8 @@ struct rvt_driver_provided { int (*check_ah)(struct ib_device *, struct ib_ah_attr *); void (*notify_new_ah)(struct ib_device *, struct ib_ah_attr *, struct rvt_ah *); + int (*alloc_qpn)(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt, + enum ib_qp_type type, u8 port); }; struct rvt_dev_info { @@ -262,7 +267,10 @@ struct rvt_dev_info { int flags; struct rvt_ibport **ports; + /* QP */ struct rvt_qp_ibdev *qp_dev; + u32 n_qps_allocated; /* number of QPs allocated for device */ + spinlock_t n_qps_lock; /* keep track of number of qps */ /* memory maps */ struct list_head pending_mmaps;