From patchwork Thu Jun 11 23:10:59 2015 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: "Marciniszyn, Mike" X-Patchwork-Id: 6593211 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 A4F5BC0020 for ; Thu, 11 Jun 2015 23:11:27 +0000 (UTC) Received: from mail.kernel.org (localhost [127.0.0.1]) by mail.kernel.org (Postfix) with ESMTP id 7B54A2052D for ; Thu, 11 Jun 2015 23:11:25 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by mail.kernel.org (Postfix) with ESMTP id C98342062B for ; Thu, 11 Jun 2015 23:11:21 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752961AbbFKXLP (ORCPT ); Thu, 11 Jun 2015 19:11:15 -0400 Received: from mga02.intel.com ([134.134.136.20]:2283 "EHLO mga02.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752875AbbFKXLN (ORCPT ); Thu, 11 Jun 2015 19:11:13 -0400 Received: from fmsmga003.fm.intel.com ([10.253.24.29]) by orsmga101.jf.intel.com with ESMTP; 11 Jun 2015 16:11:01 -0700 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.13,598,1427785200"; d="scan'208";a="506937395" Received: from sedona.ch.intel.com ([143.182.228.65]) by FMSMGA003.fm.intel.com with ESMTP; 11 Jun 2015 16:11:00 -0700 Received: from phlsvsles11.ph.intel.com (phlsvsles11.ph.intel.com [10.228.195.43]) by sedona.ch.intel.com (8.13.6/8.14.3/Standard MailSET/Hub) with ESMTP id t5BNB0DM003715; Thu, 11 Jun 2015 16:11:00 -0700 Received: from phlsvslse11.ph.intel.com (localhost [127.0.0.1]) by phlsvsles11.ph.intel.com with ESMTP id t5BNAxqi017222; Thu, 11 Jun 2015 19:10:59 -0400 Subject: [PATCH 28/41] IB/hfi1: add routines for RC/UC To: dledford@redhat.com From: Mike Marciniszyn Cc: linux-rdma@vger.kernel.org Date: Thu, 11 Jun 2015 19:10:59 -0400 Message-ID: <20150611231059.16479.58100.stgit@phlsvslse11.ph.intel.com> In-Reply-To: <20150611230710.16479.62955.stgit@phlsvslse11.ph.intel.com> References: <20150611230710.16479.62955.stgit@phlsvslse11.ph.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, T_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 Signed-off-by: Andrew Friedley Signed-off-by: Arthur Kepner Signed-off-by: Brendan Cunningham Signed-off-by: Brian Welty Signed-off-by: Caz Yokoyama Signed-off-by: Dean Luick Signed-off-by: Dennis Dalessandro Signed-off-by: Easwar Hariharan Signed-off-by: Harish Chegondi Signed-off-by: Ira Weiny Signed-off-by: Jim Snow Signed-off-by: John Gregor Signed-off-by: Jubin John Signed-off-by: Kaike Wan Signed-off-by: Kevin Pine Signed-off-by: Kyle Liddell Signed-off-by: Mike Marciniszyn Signed-off-by: Mitko Haralanov Signed-off-by: Ravi Krishnaswamy Signed-off-by: Sadanand Warrier Signed-off-by: Sanath Kumar Signed-off-by: Sudeep Dutt Signed-off-by: Vlad Danushevsky --- drivers/infiniband/hw/hfi1/ruc.c | 944 ++++++++++++++++++++++++++++++++++++++ 1 file changed, 944 insertions(+) create mode 100644 drivers/infiniband/hw/hfi1/ruc.c -- 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/hw/hfi1/ruc.c b/drivers/infiniband/hw/hfi1/ruc.c new file mode 100644 index 0000000..1811e5e --- /dev/null +++ b/drivers/infiniband/hw/hfi1/ruc.c @@ -0,0 +1,944 @@ +/* + * + * This file is provided under a dual BSD/GPLv2 license. When using or + * redistributing this file, you may do so under either license. + * + * GPL LICENSE SUMMARY + * + * Copyright(c) 2015 Intel Corporation. + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of version 2 of the GNU General Public License as + * published by the Free Software Foundation. + * + * 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 for more details. + * + * BSD LICENSE + * + * Copyright(c) 2015 Intel Corporation. + * + * 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. + * - Neither the name of Intel Corporation nor the names of its + * contributors may be used to endorse or promote products derived + * from this software without specific prior written permission. + * + * 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 + * OWNER 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 "hfi.h" +#include "mad.h" +#include "qp.h" +#include "sdma.h" + +/* + * Convert the AETH RNR timeout code into the number of microseconds. + */ +const u32 ib_hfi1_rnr_table[32] = { + 655360, /* 00: 655.36 */ + 10, /* 01: .01 */ + 20, /* 02 .02 */ + 30, /* 03: .03 */ + 40, /* 04: .04 */ + 60, /* 05: .06 */ + 80, /* 06: .08 */ + 120, /* 07: .12 */ + 160, /* 08: .16 */ + 240, /* 09: .24 */ + 320, /* 0A: .32 */ + 480, /* 0B: .48 */ + 640, /* 0C: .64 */ + 960, /* 0D: .96 */ + 1280, /* 0E: 1.28 */ + 1920, /* 0F: 1.92 */ + 2560, /* 10: 2.56 */ + 3840, /* 11: 3.84 */ + 5120, /* 12: 5.12 */ + 7680, /* 13: 7.68 */ + 10240, /* 14: 10.24 */ + 15360, /* 15: 15.36 */ + 20480, /* 16: 20.48 */ + 30720, /* 17: 30.72 */ + 40960, /* 18: 40.96 */ + 61440, /* 19: 61.44 */ + 81920, /* 1A: 81.92 */ + 122880, /* 1B: 122.88 */ + 163840, /* 1C: 163.84 */ + 245760, /* 1D: 245.76 */ + 327680, /* 1E: 327.68 */ + 491520 /* 1F: 491.52 */ +}; + +/* + * Validate a RWQE and fill in the SGE state. + * Return 1 if OK. + */ +static int init_sge(struct hfi1_qp *qp, struct hfi1_rwqe *wqe) +{ + int i, j, ret; + struct ib_wc wc; + struct hfi1_lkey_table *rkt; + struct hfi1_pd *pd; + struct hfi1_sge_state *ss; + + rkt = &to_idev(qp->ibqp.device)->lk_table; + pd = to_ipd(qp->ibqp.srq ? qp->ibqp.srq->pd : qp->ibqp.pd); + ss = &qp->r_sge; + ss->sg_list = qp->r_sg_list; + qp->r_len = 0; + for (i = j = 0; i < wqe->num_sge; i++) { + if (wqe->sg_list[i].length == 0) + continue; + /* Check LKEY */ + if (!hfi1_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge, + &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE)) + goto bad_lkey; + qp->r_len += wqe->sg_list[i].length; + j++; + } + ss->num_sge = j; + ss->total_len = qp->r_len; + ret = 1; + goto bail; + +bad_lkey: + while (j) { + struct hfi1_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge; + + hfi1_put_mr(sge->mr); + } + ss->num_sge = 0; + memset(&wc, 0, sizeof(wc)); + wc.wr_id = wqe->wr_id; + wc.status = IB_WC_LOC_PROT_ERR; + wc.opcode = IB_WC_RECV; + wc.qp = &qp->ibqp; + /* Signal solicited completion event. */ + hfi1_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1); + ret = 0; +bail: + return ret; +} + +/** + * hfi1_get_rwqe - copy the next RWQE into the QP's RWQE + * @qp: the QP + * @wr_id_only: update qp->r_wr_id only, not qp->r_sge + * + * Return -1 if there is a local error, 0 if no RWQE is available, + * otherwise return 1. + * + * Can be called from interrupt level. + */ +int hfi1_get_rwqe(struct hfi1_qp *qp, int wr_id_only) +{ + unsigned long flags; + struct hfi1_rq *rq; + struct hfi1_rwq *wq; + struct hfi1_srq *srq; + struct hfi1_rwqe *wqe; + void (*handler)(struct ib_event *, void *); + u32 tail; + int ret; + + if (qp->ibqp.srq) { + srq = to_isrq(qp->ibqp.srq); + handler = srq->ibsrq.event_handler; + rq = &srq->rq; + } else { + srq = NULL; + handler = NULL; + rq = &qp->r_rq; + } + + spin_lock_irqsave(&rq->lock, flags); + if (!(ib_hfi1_state_ops[qp->state] & HFI1_PROCESS_RECV_OK)) { + ret = 0; + goto unlock; + } + + wq = rq->wq; + tail = wq->tail; + /* Validate tail before using it since it is user writable. */ + if (tail >= rq->size) + tail = 0; + if (unlikely(tail == wq->head)) { + ret = 0; + goto unlock; + } + /* Make sure entry is read after head index is read. */ + smp_rmb(); + wqe = get_rwqe_ptr(rq, tail); + /* + * Even though we update the tail index in memory, the verbs + * consumer is not supposed to post more entries until a + * completion is generated. + */ + if (++tail >= rq->size) + tail = 0; + wq->tail = tail; + if (!wr_id_only && !init_sge(qp, wqe)) { + ret = -1; + goto unlock; + } + qp->r_wr_id = wqe->wr_id; + + ret = 1; + set_bit(HFI1_R_WRID_VALID, &qp->r_aflags); + if (handler) { + u32 n; + + /* + * Validate head pointer value and compute + * the number of remaining WQEs. + */ + n = wq->head; + if (n >= rq->size) + n = 0; + if (n < tail) + n += rq->size - tail; + else + n -= tail; + if (n < srq->limit) { + struct ib_event ev; + + srq->limit = 0; + spin_unlock_irqrestore(&rq->lock, flags); + ev.device = qp->ibqp.device; + ev.element.srq = qp->ibqp.srq; + ev.event = IB_EVENT_SRQ_LIMIT_REACHED; + handler(&ev, srq->ibsrq.srq_context); + goto bail; + } + } +unlock: + spin_unlock_irqrestore(&rq->lock, flags); +bail: + return ret; +} + +/* + * Switch to alternate path. + * The QP s_lock should be held and interrupts disabled. + */ +void hfi1_migrate_qp(struct hfi1_qp *qp) +{ + struct ib_event ev; + + qp->s_mig_state = IB_MIG_MIGRATED; + qp->remote_ah_attr = qp->alt_ah_attr; + qp->port_num = qp->alt_ah_attr.port_num; + qp->s_pkey_index = qp->s_alt_pkey_index; + qp->s_flags |= HFI1_S_AHG_CLEAR; + + ev.device = qp->ibqp.device; + ev.element.qp = &qp->ibqp; + ev.event = IB_EVENT_PATH_MIG; + qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); +} + +static __be64 get_sguid(struct hfi1_ibport *ibp, unsigned index) +{ + if (!index) { + struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); + + return ppd->guid; + } + return ibp->guids[index - 1]; +} + +static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id) +{ + return (gid->global.interface_id == id && + (gid->global.subnet_prefix == gid_prefix || + gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX)); +} + +/* + * + * This should be called with the QP r_lock held. + * + * The s_lock will be acquired around the hfi1_migrate_qp() call. + */ +int hfi1_ruc_check_hdr(struct hfi1_ibport *ibp, struct hfi1_ib_header *hdr, + int has_grh, struct hfi1_qp *qp, u32 bth0) +{ + __be64 guid; + unsigned long flags; + u8 sc5 = ibp->sl_to_sc[qp->remote_ah_attr.sl]; + + if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) { + if (!has_grh) { + if (qp->alt_ah_attr.ah_flags & IB_AH_GRH) + goto err; + } else { + if (!(qp->alt_ah_attr.ah_flags & IB_AH_GRH)) + goto err; + guid = get_sguid(ibp, qp->alt_ah_attr.grh.sgid_index); + if (!gid_ok(&hdr->u.l.grh.dgid, ibp->gid_prefix, guid)) + goto err; + if (!gid_ok(&hdr->u.l.grh.sgid, + qp->alt_ah_attr.grh.dgid.global.subnet_prefix, + qp->alt_ah_attr.grh.dgid.global.interface_id)) + goto err; + } + if (unlikely(ingress_pkey_check(ppd_from_ibp(ibp), (u16)bth0, + sc5, qp->s_alt_pkey_index, + be16_to_cpu(hdr->lrh[3])))) { + hfi1_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY, + (u16)bth0, + (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF, + 0, qp->ibqp.qp_num, + hdr->lrh[3], hdr->lrh[1]); + goto err; + } + /* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */ + if (be16_to_cpu(hdr->lrh[3]) != qp->alt_ah_attr.dlid || + ppd_from_ibp(ibp)->port != qp->alt_ah_attr.port_num) + goto err; + spin_lock_irqsave(&qp->s_lock, flags); + hfi1_migrate_qp(qp); + spin_unlock_irqrestore(&qp->s_lock, flags); + } else { + if (!has_grh) { + if (qp->remote_ah_attr.ah_flags & IB_AH_GRH) + goto err; + } else { + if (!(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) + goto err; + guid = get_sguid(ibp, + qp->remote_ah_attr.grh.sgid_index); + if (!gid_ok(&hdr->u.l.grh.dgid, ibp->gid_prefix, guid)) + goto err; + if (!gid_ok(&hdr->u.l.grh.sgid, + qp->remote_ah_attr.grh.dgid.global.subnet_prefix, + qp->remote_ah_attr.grh.dgid.global.interface_id)) + goto err; + } + if (unlikely(ingress_pkey_check(ppd_from_ibp(ibp), (u16)bth0, + sc5, qp->s_pkey_index, + be16_to_cpu(hdr->lrh[3])))) { + hfi1_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY, + (u16)bth0, + (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF, + 0, qp->ibqp.qp_num, + hdr->lrh[3], hdr->lrh[1]); + goto err; + } + /* Validate the SLID. See Ch. 9.6.1.5 */ + if (be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid || + ppd_from_ibp(ibp)->port != qp->port_num) + goto err; + if (qp->s_mig_state == IB_MIG_REARM && + !(bth0 & IB_BTH_MIG_REQ)) + qp->s_mig_state = IB_MIG_ARMED; + } + + return 0; + +err: + return 1; +} + +/** + * ruc_loopback - handle UC and RC loopback requests + * @sqp: the sending QP + * + * This is called from hfi1_do_send() to + * forward a WQE addressed to the same HCA. + * Note that although we are single threaded due to the tasklet, we still + * have to protect against post_send(). We don't have to worry about + * receive interrupts since this is a connected protocol and all packets + * will pass through here. + */ +static void ruc_loopback(struct hfi1_qp *sqp) +{ + struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num); + struct hfi1_qp *qp; + struct hfi1_swqe *wqe; + struct hfi1_sge *sge; + unsigned long flags; + struct ib_wc wc; + u64 sdata; + atomic64_t *maddr; + enum ib_wc_status send_status; + int release; + int ret; + + /* + * Note that we check the responder QP state after + * checking the requester's state. + */ + qp = hfi1_lookup_qpn(ibp, sqp->remote_qpn); + + spin_lock_irqsave(&sqp->s_lock, flags); + + /* Return if we are already busy processing a work request. */ + if ((sqp->s_flags & (HFI1_S_BUSY | HFI1_S_ANY_WAIT)) || + !(ib_hfi1_state_ops[sqp->state] & HFI1_PROCESS_OR_FLUSH_SEND)) + goto unlock; + + sqp->s_flags |= HFI1_S_BUSY; + +again: + if (sqp->s_last == sqp->s_head) + goto clr_busy; + wqe = get_swqe_ptr(sqp, sqp->s_last); + + /* Return if it is not OK to start a new work request. */ + if (!(ib_hfi1_state_ops[sqp->state] & HFI1_PROCESS_NEXT_SEND_OK)) { + if (!(ib_hfi1_state_ops[sqp->state] & HFI1_FLUSH_SEND)) + goto clr_busy; + /* We are in the error state, flush the work request. */ + send_status = IB_WC_WR_FLUSH_ERR; + goto flush_send; + } + + /* + * We can rely on the entry not changing without the s_lock + * being held until we update s_last. + * We increment s_cur to indicate s_last is in progress. + */ + if (sqp->s_last == sqp->s_cur) { + if (++sqp->s_cur >= sqp->s_size) + sqp->s_cur = 0; + } + spin_unlock_irqrestore(&sqp->s_lock, flags); + + if (!qp || !(ib_hfi1_state_ops[qp->state] & HFI1_PROCESS_RECV_OK) || + qp->ibqp.qp_type != sqp->ibqp.qp_type) { + ibp->n_pkt_drops++; + /* + * For RC, the requester would timeout and retry so + * shortcut the timeouts and just signal too many retries. + */ + if (sqp->ibqp.qp_type == IB_QPT_RC) + send_status = IB_WC_RETRY_EXC_ERR; + else + send_status = IB_WC_SUCCESS; + goto serr; + } + + memset(&wc, 0, sizeof(wc)); + send_status = IB_WC_SUCCESS; + + release = 1; + sqp->s_sge.sge = wqe->sg_list[0]; + sqp->s_sge.sg_list = wqe->sg_list + 1; + sqp->s_sge.num_sge = wqe->wr.num_sge; + sqp->s_len = wqe->length; + switch (wqe->wr.opcode) { + case IB_WR_SEND_WITH_IMM: + wc.wc_flags = IB_WC_WITH_IMM; + wc.ex.imm_data = wqe->wr.ex.imm_data; + /* FALLTHROUGH */ + case IB_WR_SEND: + ret = hfi1_get_rwqe(qp, 0); + if (ret < 0) + goto op_err; + if (!ret) + goto rnr_nak; + break; + + case IB_WR_RDMA_WRITE_WITH_IMM: + if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE))) + goto inv_err; + wc.wc_flags = IB_WC_WITH_IMM; + wc.ex.imm_data = wqe->wr.ex.imm_data; + ret = hfi1_get_rwqe(qp, 1); + if (ret < 0) + goto op_err; + if (!ret) + goto rnr_nak; + /* FALLTHROUGH */ + case IB_WR_RDMA_WRITE: + if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE))) + goto inv_err; + if (wqe->length == 0) + break; + if (unlikely(!hfi1_rkey_ok(qp, &qp->r_sge.sge, wqe->length, + wqe->wr.wr.rdma.remote_addr, + wqe->wr.wr.rdma.rkey, + IB_ACCESS_REMOTE_WRITE))) + goto acc_err; + qp->r_sge.sg_list = NULL; + qp->r_sge.num_sge = 1; + qp->r_sge.total_len = wqe->length; + break; + + case IB_WR_RDMA_READ: + if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ))) + goto inv_err; + if (unlikely(!hfi1_rkey_ok(qp, &sqp->s_sge.sge, wqe->length, + wqe->wr.wr.rdma.remote_addr, + wqe->wr.wr.rdma.rkey, + IB_ACCESS_REMOTE_READ))) + goto acc_err; + release = 0; + sqp->s_sge.sg_list = NULL; + sqp->s_sge.num_sge = 1; + qp->r_sge.sge = wqe->sg_list[0]; + qp->r_sge.sg_list = wqe->sg_list + 1; + qp->r_sge.num_sge = wqe->wr.num_sge; + qp->r_sge.total_len = wqe->length; + break; + + case IB_WR_ATOMIC_CMP_AND_SWP: + case IB_WR_ATOMIC_FETCH_AND_ADD: + if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC))) + goto inv_err; + if (unlikely(!hfi1_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64), + wqe->wr.wr.atomic.remote_addr, + wqe->wr.wr.atomic.rkey, + IB_ACCESS_REMOTE_ATOMIC))) + goto acc_err; + /* Perform atomic OP and save result. */ + maddr = (atomic64_t *) qp->r_sge.sge.vaddr; + sdata = wqe->wr.wr.atomic.compare_add; + *(u64 *) sqp->s_sge.sge.vaddr = + (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ? + (u64) atomic64_add_return(sdata, maddr) - sdata : + (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr, + sdata, wqe->wr.wr.atomic.swap); + hfi1_put_mr(qp->r_sge.sge.mr); + qp->r_sge.num_sge = 0; + goto send_comp; + + default: + send_status = IB_WC_LOC_QP_OP_ERR; + goto serr; + } + + sge = &sqp->s_sge.sge; + while (sqp->s_len) { + u32 len = sqp->s_len; + + if (len > sge->length) + len = sge->length; + if (len > sge->sge_length) + len = sge->sge_length; + BUG_ON(len == 0); + hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, release); + sge->vaddr += len; + sge->length -= len; + sge->sge_length -= len; + if (sge->sge_length == 0) { + if (!release) + hfi1_put_mr(sge->mr); + if (--sqp->s_sge.num_sge) + *sge = *sqp->s_sge.sg_list++; + } else if (sge->length == 0 && sge->mr->lkey) { + if (++sge->n >= HFI1_SEGSZ) { + if (++sge->m >= sge->mr->mapsz) + break; + sge->n = 0; + } + sge->vaddr = + sge->mr->map[sge->m]->segs[sge->n].vaddr; + sge->length = + sge->mr->map[sge->m]->segs[sge->n].length; + } + sqp->s_len -= len; + } + if (release) + hfi1_put_ss(&qp->r_sge); + + if (!test_and_clear_bit(HFI1_R_WRID_VALID, &qp->r_aflags)) + goto send_comp; + + if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM) + wc.opcode = IB_WC_RECV_RDMA_WITH_IMM; + else + wc.opcode = IB_WC_RECV; + wc.wr_id = qp->r_wr_id; + wc.status = IB_WC_SUCCESS; + wc.byte_len = wqe->length; + wc.qp = &qp->ibqp; + wc.src_qp = qp->remote_qpn; + wc.slid = qp->remote_ah_attr.dlid; + wc.sl = qp->remote_ah_attr.sl; + wc.port_num = 1; + /* Signal completion event if the solicited bit is set. */ + hfi1_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, + wqe->wr.send_flags & IB_SEND_SOLICITED); + +send_comp: + spin_lock_irqsave(&sqp->s_lock, flags); + ibp->n_loop_pkts++; +flush_send: + sqp->s_rnr_retry = sqp->s_rnr_retry_cnt; + hfi1_send_complete(sqp, wqe, send_status); + goto again; + +rnr_nak: + /* Handle RNR NAK */ + if (qp->ibqp.qp_type == IB_QPT_UC) + goto send_comp; + ibp->n_rnr_naks++; + /* + * Note: we don't need the s_lock held since the BUSY flag + * makes this single threaded. + */ + if (sqp->s_rnr_retry == 0) { + send_status = IB_WC_RNR_RETRY_EXC_ERR; + goto serr; + } + if (sqp->s_rnr_retry_cnt < 7) + sqp->s_rnr_retry--; + spin_lock_irqsave(&sqp->s_lock, flags); + if (!(ib_hfi1_state_ops[sqp->state] & HFI1_PROCESS_RECV_OK)) + goto clr_busy; + sqp->s_flags |= HFI1_S_WAIT_RNR; + sqp->s_timer.function = hfi1_rc_rnr_retry; + sqp->s_timer.expires = jiffies + + usecs_to_jiffies(ib_hfi1_rnr_table[qp->r_min_rnr_timer]); + add_timer(&sqp->s_timer); + goto clr_busy; + +op_err: + send_status = IB_WC_REM_OP_ERR; + wc.status = IB_WC_LOC_QP_OP_ERR; + goto err; + +inv_err: + send_status = IB_WC_REM_INV_REQ_ERR; + wc.status = IB_WC_LOC_QP_OP_ERR; + goto err; + +acc_err: + send_status = IB_WC_REM_ACCESS_ERR; + wc.status = IB_WC_LOC_PROT_ERR; +err: + /* responder goes to error state */ + hfi1_rc_error(qp, wc.status); + +serr: + spin_lock_irqsave(&sqp->s_lock, flags); + hfi1_send_complete(sqp, wqe, send_status); + if (sqp->ibqp.qp_type == IB_QPT_RC) { + int lastwqe = hfi1_error_qp(sqp, IB_WC_WR_FLUSH_ERR); + + sqp->s_flags &= ~HFI1_S_BUSY; + spin_unlock_irqrestore(&sqp->s_lock, flags); + if (lastwqe) { + struct ib_event ev; + + ev.device = sqp->ibqp.device; + ev.element.qp = &sqp->ibqp; + ev.event = IB_EVENT_QP_LAST_WQE_REACHED; + sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context); + } + goto done; + } +clr_busy: + sqp->s_flags &= ~HFI1_S_BUSY; +unlock: + spin_unlock_irqrestore(&sqp->s_lock, flags); +done: + if (qp && atomic_dec_and_test(&qp->refcount)) + wake_up(&qp->wait); +} + +/** + * hfi1_make_grh - construct a GRH header + * @ibp: a pointer to the IB port + * @hdr: a pointer to the GRH header being constructed + * @grh: the global route address to send to + * @hwords: the number of 32 bit words of header being sent + * @nwords: the number of 32 bit words of data being sent + * + * Return the size of the header in 32 bit words. + */ +u32 hfi1_make_grh(struct hfi1_ibport *ibp, struct ib_grh *hdr, + struct ib_global_route *grh, u32 hwords, u32 nwords) +{ + hdr->version_tclass_flow = + cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) | + (grh->traffic_class << IB_GRH_TCLASS_SHIFT) | + (grh->flow_label << IB_GRH_FLOW_SHIFT)); + hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2); + /* next_hdr is defined by C8-7 in ch. 8.4.1 */ + hdr->next_hdr = IB_GRH_NEXT_HDR; + hdr->hop_limit = grh->hop_limit; + /* The SGID is 32-bit aligned. */ + hdr->sgid.global.subnet_prefix = ibp->gid_prefix; + hdr->sgid.global.interface_id = + grh->sgid_index && grh->sgid_index < ARRAY_SIZE(ibp->guids) ? + ibp->guids[grh->sgid_index - 1] : ppd_from_ibp(ibp)->guid; + hdr->dgid = grh->dgid; + + /* GRH header size in 32-bit words. */ + return sizeof(struct ib_grh) / sizeof(u32); +} + +/* + * free_ahg - clear ahg from QP + */ +void clear_ahg(struct hfi1_qp *qp) +{ + qp->s_hdr->ahgcount = 0; + qp->s_flags &= ~(HFI1_S_AHG_VALID | HFI1_S_AHG_CLEAR); + if (qp->s_sde) + sdma_ahg_free(qp->s_sde, qp->s_ahgidx); + qp->s_ahgidx = -1; + qp->s_sde = NULL; +} + +#define BTH2_OFFSET (offsetof(struct hfi1_pio_header, hdr.u.oth.bth[2]) / 4) + +/** + * build_ahg - create ahg in s_hdr + * @qp: a pointer to QP + * @npsn: the next PSN for the request/response + * + * This routine handles the AHG by allocating an ahg entry and causing the + * copy of the first middle. + * + * Subsequent middles use the copied entry, editing the + * PSN with 1 or 2 edits. + */ +static inline void build_ahg(struct hfi1_qp *qp, u32 npsn) +{ + if (unlikely(qp->s_flags & HFI1_S_AHG_CLEAR)) + clear_ahg(qp); + if (!(qp->s_flags & HFI1_S_AHG_VALID)) { + /* first middle that needs copy */ + if (qp->s_ahgidx < 0) { + if (!qp->s_sde) + qp->s_sde = qp_to_sdma_engine(qp, qp->s_sc); + qp->s_ahgidx = sdma_ahg_alloc(qp->s_sde); + } + if (qp->s_ahgidx >= 0) { + qp->s_ahgpsn = npsn; + qp->s_hdr->tx_flags |= SDMA_TXREQ_F_AHG_COPY; + /* save to protect a change in another thread */ + qp->s_hdr->sde = qp->s_sde; + qp->s_hdr->ahgidx = qp->s_ahgidx; + qp->s_flags |= HFI1_S_AHG_VALID; + } + } else { + /* subsequent middle after valid */ + if (qp->s_ahgidx >= 0) { + qp->s_hdr->tx_flags |= SDMA_TXREQ_F_USE_AHG; + qp->s_hdr->ahgidx = qp->s_ahgidx; + qp->s_hdr->ahgcount++; + qp->s_hdr->ahgdesc[0] = + sdma_build_ahg_descriptor( + (__force u16)cpu_to_be16((u16)npsn), + BTH2_OFFSET, + 16, + 16); + if ((npsn & 0xffff0000) != + (qp->s_ahgpsn & 0xffff0000)) { + qp->s_hdr->ahgcount++; + qp->s_hdr->ahgdesc[1] = + sdma_build_ahg_descriptor( + (__force u16)cpu_to_be16( + (u16)(npsn >> 16)), + BTH2_OFFSET, + 0, + 16); + } + } + } +} + +void hfi1_make_ruc_header(struct hfi1_qp *qp, struct hfi1_other_headers *ohdr, + u32 bth0, u32 bth2, int middle) +{ + struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); + u16 lrh0; + u32 nwords; + u32 extra_bytes; + u8 sc5; + + /* Construct the header. */ + extra_bytes = -qp->s_cur_size & 3; + nwords = (qp->s_cur_size + extra_bytes) >> 2; + lrh0 = HFI1_LRH_BTH; + if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) { + qp->s_hdrwords += hfi1_make_grh(ibp, &qp->s_hdr->ibh.u.l.grh, + &qp->remote_ah_attr.grh, + qp->s_hdrwords, nwords); + lrh0 = HFI1_LRH_GRH; + middle = 0; + } + sc5 = ibp->sl_to_sc[qp->remote_ah_attr.sl]; + lrh0 |= (sc5 & 0xf) << 12 | (qp->remote_ah_attr.sl & 0xf) << 4; + qp->s_sc = sc5; + /* + * reset s_hdr/AHG fields + * + * This insures that the ahgentry/ahgcount + * are at a non-AHG default to protect + * build_verbs_tx_desc() from using + * an include ahgidx. + * + * build_ahg() will modify as appropriate + * to use the AHG feature. + */ + qp->s_hdr->tx_flags = 0; + qp->s_hdr->ahgcount = 0; + qp->s_hdr->ahgidx = 0; + qp->s_hdr->sde = NULL; + if (qp->s_mig_state == IB_MIG_MIGRATED) + bth0 |= IB_BTH_MIG_REQ; + else + middle = 0; + if (middle) + build_ahg(qp, bth2); + else + qp->s_flags &= ~HFI1_S_AHG_VALID; + qp->s_hdr->ibh.lrh[0] = cpu_to_be16(lrh0); + qp->s_hdr->ibh.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid); + qp->s_hdr->ibh.lrh[2] = + cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC); + qp->s_hdr->ibh.lrh[3] = cpu_to_be16(ppd_from_ibp(ibp)->lid | + qp->remote_ah_attr.src_path_bits); + bth0 |= hfi1_get_pkey(ibp, qp->s_pkey_index); + bth0 |= extra_bytes << 20; + ohdr->bth[0] = cpu_to_be32(bth0); + ohdr->bth[1] = cpu_to_be32(qp->remote_qpn); + if (test_and_clear_bit(HFI1_S_ECN, &qp->s_aflags)) + /* we recently received a FECN, so return a BECN */ + ohdr->bth[1] |= cpu_to_be32(HFI1_BECN_MASK << HFI1_BECN_SHIFT); + ohdr->bth[2] = cpu_to_be32(bth2); +} + +/** + * hfi1_do_send - perform a send on a QP + * @work: contains a pointer to the QP + * + * Process entries in the send work queue until credit or queue is + * exhausted. Only allow one CPU to send a packet per QP (tasklet). + * Otherwise, two threads could send packets out of order. + */ +void hfi1_do_send(struct work_struct *work) +{ + struct iowait *wait = container_of(work, struct iowait, iowork); + struct hfi1_qp *qp = container_of(wait, struct hfi1_qp, s_iowait); + struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); + struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); + int (*make_req)(struct hfi1_qp *qp); + unsigned long flags; + + if ((qp->ibqp.qp_type == IB_QPT_RC || + qp->ibqp.qp_type == IB_QPT_UC) && + !loopback && + (qp->remote_ah_attr.dlid & ~((1 << ppd->lmc) - 1)) == ppd->lid) { + ruc_loopback(qp); + return; + } + + if (qp->ibqp.qp_type == IB_QPT_RC) + make_req = hfi1_make_rc_req; + else if (qp->ibqp.qp_type == IB_QPT_UC) + make_req = hfi1_make_uc_req; + else + make_req = hfi1_make_ud_req; + + spin_lock_irqsave(&qp->s_lock, flags); + + /* Return if we are already busy processing a work request. */ + if (!hfi1_send_ok(qp)) { + spin_unlock_irqrestore(&qp->s_lock, flags); + return; + } + + qp->s_flags |= HFI1_S_BUSY; + + spin_unlock_irqrestore(&qp->s_lock, flags); + + do { + /* Check for a constructed packet to be sent. */ + if (qp->s_hdrwords != 0) { + /* + * If the packet cannot be sent now, return and + * the send tasklet will be woken up later. + */ + if (hfi1_verbs_send(qp, qp->s_hdr, qp->s_hdrwords, + qp->s_cur_sge, qp->s_cur_size)) + break; + /* Record that s_hdr is empty. */ + qp->s_hdrwords = 0; + } + } while (make_req(qp)); +} + +/* + * This should be called with s_lock held. + */ +void hfi1_send_complete(struct hfi1_qp *qp, struct hfi1_swqe *wqe, + enum ib_wc_status status) +{ + u32 old_last, last; + unsigned i; + + if (!(ib_hfi1_state_ops[qp->state] & HFI1_PROCESS_OR_FLUSH_SEND)) + return; + + for (i = 0; i < wqe->wr.num_sge; i++) { + struct hfi1_sge *sge = &wqe->sg_list[i]; + + hfi1_put_mr(sge->mr); + } + if (qp->ibqp.qp_type == IB_QPT_UD || + qp->ibqp.qp_type == IB_QPT_SMI || + qp->ibqp.qp_type == IB_QPT_GSI) + atomic_dec(&to_iah(wqe->wr.wr.ud.ah)->refcount); + + /* See ch. 11.2.4.1 and 10.7.3.1 */ + if (!(qp->s_flags & HFI1_S_SIGNAL_REQ_WR) || + (wqe->wr.send_flags & IB_SEND_SIGNALED) || + status != IB_WC_SUCCESS) { + struct ib_wc wc; + + memset(&wc, 0, sizeof(wc)); + wc.wr_id = wqe->wr.wr_id; + wc.status = status; + wc.opcode = ib_hfi1_wc_opcode[wqe->wr.opcode]; + wc.qp = &qp->ibqp; + if (status == IB_WC_SUCCESS) + wc.byte_len = wqe->length; + hfi1_cq_enter(to_icq(qp->ibqp.send_cq), &wc, + status != IB_WC_SUCCESS); + } + + last = qp->s_last; + old_last = last; + if (++last >= qp->s_size) + last = 0; + qp->s_last = last; + if (qp->s_acked == old_last) + qp->s_acked = last; + if (qp->s_cur == old_last) + qp->s_cur = last; + if (qp->s_tail == old_last) + qp->s_tail = last; + if (qp->state == IB_QPS_SQD && last == qp->s_cur) + qp->s_draining = 0; +}