hw/vmw_pvrdma/pvrdma_cq.c

0001 /*
0002  * Copyright (c) 2012-2016 VMware, Inc.  All rights reserved.
0003  *
0004  * This program is free software; you can redistribute it and/or
0005  * modify it under the terms of EITHER the GNU General Public License
0006  * version 2 as published by the Free Software Foundation or the BSD
0007  * 2-Clause License. This program is distributed in the hope that it
0008  * will be useful, but WITHOUT ANY WARRANTY; WITHOUT EVEN THE IMPLIED
0009  * WARRANTY OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
0010  * See the GNU General Public License version 2 for more details at
0011  * http://www.gnu.org/licenses/old-licenses/gpl-2.0.en.html.
0012  *
0013  * You should have received a copy of the GNU General Public License
0014  * along with this program available in the file COPYING in the main
0015  * directory of this source tree.
0016  *
0017  * The BSD 2-Clause License
0018  *
0019  *     Redistribution and use in source and binary forms, with or
0020  *     without modification, are permitted provided that the following
0021  *     conditions are met:
0022  *
0023  *      - Redistributions of source code must retain the above
0024  *        copyright notice, this list of conditions and the following
0025  *        disclaimer.
0026  *
0027  *      - Redistributions in binary form must reproduce the above
0028  *        copyright notice, this list of conditions and the following
0029  *        disclaimer in the documentation and/or other materials
0030  *        provided with the distribution.
0031  *
0032  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
0033  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
0034  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
0035  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
0036  * COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
0037  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
0038  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
0039  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
0040  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
0041  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
0042  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
0043  * OF THE POSSIBILITY OF SUCH DAMAGE.
0044  */
0045
0046 #include <asm/page.h>
0047 #include <linux/io.h>
0048 #include <linux/wait.h>
0049 #include <rdma/ib_addr.h>
0050 #include <rdma/ib_smi.h>
0051 #include <rdma/ib_user_verbs.h>
0052 #include <rdma/uverbs_ioctl.h>
0053
0054 #include "pvrdma.h"
0055
0056 /**
0057  * pvrdma_req_notify_cq - request notification for a completion queue
0058  * @ibcq: the completion queue
0059  * @notify_flags: notification flags
0060  *
0061  * @return: 0 for success.
0062  */
0063 int pvrdma_req_notify_cq(struct ib_cq *ibcq,
0064              enum ib_cq_notify_flags notify_flags)
0065 {
0066     struct pvrdma_dev *dev = to_vdev(ibcq->device);
0067     struct pvrdma_cq *cq = to_vcq(ibcq);
0068     u32 val = cq->cq_handle;
0069     unsigned long flags;
0070     int has_data = 0;
0071
0072     val |= (notify_flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
0073         PVRDMA_UAR_CQ_ARM_SOL : PVRDMA_UAR_CQ_ARM;
0074
0075     spin_lock_irqsave(&cq->cq_lock, flags);
0076
0077     pvrdma_write_uar_cq(dev, val);
0078
0079     if (notify_flags & IB_CQ_REPORT_MISSED_EVENTS) {
0080         unsigned int head;
0081
0082         has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
0083                             cq->ibcq.cqe, &head);
0084         if (unlikely(has_data == PVRDMA_INVALID_IDX))
0085             dev_err(&dev->pdev->dev, "CQ ring state invalid\n");
0086     }
0087
0088     spin_unlock_irqrestore(&cq->cq_lock, flags);
0089
0090     return has_data;
0091 }
0092
0093 /**
0094  * pvrdma_create_cq - create completion queue
0095  * @ibcq: Allocated CQ
0096  * @attr: completion queue attributes
0097  * @udata: user data
0098  *
0099  * @return: 0 on success
0100  */
0101 int pvrdma_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
0102              struct ib_udata *udata)
0103 {
0104     struct ib_device *ibdev = ibcq->device;
0105     int entries = attr->cqe;
0106     struct pvrdma_dev *dev = to_vdev(ibdev);
0107     struct pvrdma_cq *cq = to_vcq(ibcq);
0108     int ret;
0109     int npages;
0110     unsigned long flags;
0111     union pvrdma_cmd_req req;
0112     union pvrdma_cmd_resp rsp;
0113     struct pvrdma_cmd_create_cq *cmd = &req.create_cq;
0114     struct pvrdma_cmd_create_cq_resp *resp = &rsp.create_cq_resp;
0115     struct pvrdma_create_cq_resp cq_resp = {};
0116     struct pvrdma_create_cq ucmd;
0117     struct pvrdma_ucontext *context = rdma_udata_to_drv_context(
0118         udata, struct pvrdma_ucontext, ibucontext);
0119
0120     BUILD_BUG_ON(sizeof(struct pvrdma_cqe) != 64);
0121
0122     if (attr->flags)
0123         return -EOPNOTSUPP;
0124
0125     entries = roundup_pow_of_two(entries);
0126     if (entries < 1 || entries > dev->dsr->caps.max_cqe)
0127         return -EINVAL;
0128
0129     if (!atomic_add_unless(&dev->num_cqs, 1, dev->dsr->caps.max_cq))
0130         return -ENOMEM;
0131
0132     cq->ibcq.cqe = entries;
0133     cq->is_kernel = !udata;
0134
0135     if (!cq->is_kernel) {
0136         if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
0137             ret = -EFAULT;
0138             goto err_cq;
0139         }
0140
0141         cq->umem = ib_umem_get(ibdev, ucmd.buf_addr, ucmd.buf_size,
0142                        IB_ACCESS_LOCAL_WRITE);
0143         if (IS_ERR(cq->umem)) {
0144             ret = PTR_ERR(cq->umem);
0145             goto err_cq;
0146         }
0147
0148         npages = ib_umem_num_dma_blocks(cq->umem, PAGE_SIZE);
0149     } else {
0150         /* One extra page for shared ring state */
0151         npages = 1 + (entries * sizeof(struct pvrdma_cqe) +
0152                   PAGE_SIZE - 1) / PAGE_SIZE;
0153
0154         /* Skip header page. */
0155         cq->offset = PAGE_SIZE;
0156     }
0157
0158     if (npages < 0 || npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
0159         dev_warn(&dev->pdev->dev,
0160              "overflow pages in completion queue\n");
0161         ret = -EINVAL;
0162         goto err_umem;
0163     }
0164
0165     ret = pvrdma_page_dir_init(dev, &cq->pdir, npages, cq->is_kernel);
0166     if (ret) {
0167         dev_warn(&dev->pdev->dev,
0168              "could not allocate page directory\n");
0169         goto err_umem;
0170     }
0171
0172     /* Ring state is always the first page. Set in library for user cq. */
0173     if (cq->is_kernel)
0174         cq->ring_state = cq->pdir.pages[0];
0175     else
0176         pvrdma_page_dir_insert_umem(&cq->pdir, cq->umem, 0);
0177
0178     refcount_set(&cq->refcnt, 1);
0179     init_completion(&cq->free);
0180     spin_lock_init(&cq->cq_lock);
0181
0182     memset(cmd, 0, sizeof(*cmd));
0183     cmd->hdr.cmd = PVRDMA_CMD_CREATE_CQ;
0184     cmd->nchunks = npages;
0185     cmd->ctx_handle = context ? context->ctx_handle : 0;
0186     cmd->cqe = entries;
0187     cmd->pdir_dma = cq->pdir.dir_dma;
0188     ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_CQ_RESP);
0189     if (ret < 0) {
0190         dev_warn(&dev->pdev->dev,
0191              "could not create completion queue, error: %d\n", ret);
0192         goto err_page_dir;
0193     }
0194
0195     cq->ibcq.cqe = resp->cqe;
0196     cq->cq_handle = resp->cq_handle;
0197     cq_resp.cqn = resp->cq_handle;
0198     spin_lock_irqsave(&dev->cq_tbl_lock, flags);
0199     dev->cq_tbl[cq->cq_handle % dev->dsr->caps.max_cq] = cq;
0200     spin_unlock_irqrestore(&dev->cq_tbl_lock, flags);
0201
0202     if (!cq->is_kernel) {
0203         cq->uar = &context->uar;
0204
0205         /* Copy udata back. */
0206         if (ib_copy_to_udata(udata, &cq_resp, sizeof(cq_resp))) {
0207             dev_warn(&dev->pdev->dev,
0208                  "failed to copy back udata\n");
0209             pvrdma_destroy_cq(&cq->ibcq, udata);
0210             return -EINVAL;
0211         }
0212     }
0213
0214     return 0;
0215
0216 err_page_dir:
0217     pvrdma_page_dir_cleanup(dev, &cq->pdir);
0218 err_umem:
0219     ib_umem_release(cq->umem);
0220 err_cq:
0221     atomic_dec(&dev->num_cqs);
0222     return ret;
0223 }
0224
0225 static void pvrdma_free_cq(struct pvrdma_dev *dev, struct pvrdma_cq *cq)
0226 {
0227     if (refcount_dec_and_test(&cq->refcnt))
0228         complete(&cq->free);
0229     wait_for_completion(&cq->free);
0230
0231     ib_umem_release(cq->umem);
0232
0233     pvrdma_page_dir_cleanup(dev, &cq->pdir);
0234 }
0235
0236 /**
0237  * pvrdma_destroy_cq - destroy completion queue
0238  * @cq: the completion queue to destroy.
0239  * @udata: user data or null for kernel object
0240  */
0241 int pvrdma_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
0242 {
0243     struct pvrdma_cq *vcq = to_vcq(cq);
0244     union pvrdma_cmd_req req;
0245     struct pvrdma_cmd_destroy_cq *cmd = &req.destroy_cq;
0246     struct pvrdma_dev *dev = to_vdev(cq->device);
0247     unsigned long flags;
0248     int ret;
0249
0250     memset(cmd, 0, sizeof(*cmd));
0251     cmd->hdr.cmd = PVRDMA_CMD_DESTROY_CQ;
0252     cmd->cq_handle = vcq->cq_handle;
0253
0254     ret = pvrdma_cmd_post(dev, &req, NULL, 0);
0255     if (ret < 0)
0256         dev_warn(&dev->pdev->dev,
0257              "could not destroy completion queue, error: %d\n",
0258              ret);
0259
0260     /* free cq's resources */
0261     spin_lock_irqsave(&dev->cq_tbl_lock, flags);
0262     dev->cq_tbl[vcq->cq_handle] = NULL;
0263     spin_unlock_irqrestore(&dev->cq_tbl_lock, flags);
0264
0265     pvrdma_free_cq(dev, vcq);
0266     atomic_dec(&dev->num_cqs);
0267     return 0;
0268 }
0269
0270 static inline struct pvrdma_cqe *get_cqe(struct pvrdma_cq *cq, int i)
0271 {
0272     return (struct pvrdma_cqe *)pvrdma_page_dir_get_ptr(
0273                     &cq->pdir,
0274                     cq->offset +
0275                     sizeof(struct pvrdma_cqe) * i);
0276 }
0277
0278 void _pvrdma_flush_cqe(struct pvrdma_qp *qp, struct pvrdma_cq *cq)
0279 {
0280     unsigned int head;
0281     int has_data;
0282
0283     if (!cq->is_kernel)
0284         return;
0285
0286     /* Lock held */
0287     has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
0288                         cq->ibcq.cqe, &head);
0289     if (unlikely(has_data > 0)) {
0290         int items;
0291         int curr;
0292         int tail = pvrdma_idx(&cq->ring_state->rx.prod_tail,
0293                       cq->ibcq.cqe);
0294         struct pvrdma_cqe *cqe;
0295         struct pvrdma_cqe *curr_cqe;
0296
0297         items = (tail > head) ? (tail - head) :
0298             (cq->ibcq.cqe - head + tail);
0299         curr = --tail;
0300         while (items-- > 0) {
0301             if (curr < 0)
0302                 curr = cq->ibcq.cqe - 1;
0303             if (tail < 0)
0304                 tail = cq->ibcq.cqe - 1;
0305             curr_cqe = get_cqe(cq, curr);
0306             if ((curr_cqe->qp & 0xFFFF) != qp->qp_handle) {
0307                 if (curr != tail) {
0308                     cqe = get_cqe(cq, tail);
0309                     *cqe = *curr_cqe;
0310                 }
0311                 tail--;
0312             } else {
0313                 pvrdma_idx_ring_inc(
0314                     &cq->ring_state->rx.cons_head,
0315                     cq->ibcq.cqe);
0316             }
0317             curr--;
0318         }
0319     }
0320 }
0321
0322 static int pvrdma_poll_one(struct pvrdma_cq *cq, struct pvrdma_qp **cur_qp,
0323                struct ib_wc *wc)
0324 {
0325     struct pvrdma_dev *dev = to_vdev(cq->ibcq.device);
0326     int has_data;
0327     unsigned int head;
0328     bool tried = false;
0329     struct pvrdma_cqe *cqe;
0330
0331 retry:
0332     has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
0333                         cq->ibcq.cqe, &head);
0334     if (has_data == 0) {
0335         if (tried)
0336             return -EAGAIN;
0337
0338         pvrdma_write_uar_cq(dev, cq->cq_handle | PVRDMA_UAR_CQ_POLL);
0339
0340         tried = true;
0341         goto retry;
0342     } else if (has_data == PVRDMA_INVALID_IDX) {
0343         dev_err(&dev->pdev->dev, "CQ ring state invalid\n");
0344         return -EAGAIN;
0345     }
0346
0347     cqe = get_cqe(cq, head);
0348
0349     /* Ensure cqe is valid. */
0350     rmb();
0351     if (dev->qp_tbl[cqe->qp & 0xffff])
0352         *cur_qp = (struct pvrdma_qp *)dev->qp_tbl[cqe->qp & 0xffff];
0353     else
0354         return -EAGAIN;
0355
0356     wc->opcode = pvrdma_wc_opcode_to_ib(cqe->opcode);
0357     wc->status = pvrdma_wc_status_to_ib(cqe->status);
0358     wc->wr_id = cqe->wr_id;
0359     wc->qp = &(*cur_qp)->ibqp;
0360     wc->byte_len = cqe->byte_len;
0361     wc->ex.imm_data = cqe->imm_data;
0362     wc->src_qp = cqe->src_qp;
0363     wc->wc_flags = pvrdma_wc_flags_to_ib(cqe->wc_flags);
0364     wc->pkey_index = cqe->pkey_index;
0365     wc->slid = cqe->slid;
0366     wc->sl = cqe->sl;
0367     wc->dlid_path_bits = cqe->dlid_path_bits;
0368     wc->port_num = cqe->port_num;
0369     wc->vendor_err = cqe->vendor_err;
0370     wc->network_hdr_type = pvrdma_network_type_to_ib(cqe->network_hdr_type);
0371
0372     /* Update shared ring state */
0373     pvrdma_idx_ring_inc(&cq->ring_state->rx.cons_head, cq->ibcq.cqe);
0374
0375     return 0;
0376 }
0377
0378 /**
0379  * pvrdma_poll_cq - poll for work completion queue entries
0380  * @ibcq: completion queue
0381  * @num_entries: the maximum number of entries
0382  * @wc: pointer to work completion array
0383  *
0384  * @return: number of polled completion entries
0385  */
0386 int pvrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
0387 {
0388     struct pvrdma_cq *cq = to_vcq(ibcq);
0389     struct pvrdma_qp *cur_qp = NULL;
0390     unsigned long flags;
0391     int npolled;
0392
0393     if (num_entries < 1 || wc == NULL)
0394         return 0;
0395
0396     spin_lock_irqsave(&cq->cq_lock, flags);
0397     for (npolled = 0; npolled < num_entries; ++npolled) {
0398         if (pvrdma_poll_one(cq, &cur_qp, wc + npolled))
0399             break;
0400     }
0401
0402     spin_unlock_irqrestore(&cq->cq_lock, flags);
0403
0404     /* Ensure we do not return errors from poll_cq */
0405     return npolled;
0406 }