drivers/infiniband/hw/vmw_pvrdma/pvrdma_cq.c

   1 /*
   2  * Copyright (c) 2012-2016 VMware, Inc.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of EITHER the GNU General Public License
   6  * version 2 as published by the Free Software Foundation or the BSD
   7  * 2-Clause License. This program is distributed in the hope that it
   8  * will be useful, but WITHOUT ANY WARRANTY; WITHOUT EVEN THE IMPLIED
   9  * WARRANTY OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
  10  * See the GNU General Public License version 2 for more details at
  11  * http://www.gnu.org/licenses/old-licenses/gpl-2.0.en.html.
  12  *
  13  * You should have received a copy of the GNU General Public License
  14  * along with this program available in the file COPYING in the main
  15  * directory of this source tree.
  16  *
  17  * The BSD 2-Clause License
  18  *
  19  *     Redistribution and use in source and binary forms, with or
  20  *     without modification, are permitted provided that the following
  21  *     conditions are met:
  22  *
  23  *      - Redistributions of source code must retain the above
  24  *        copyright notice, this list of conditions and the following
  25  *        disclaimer.
  26  *
  27  *      - Redistributions in binary form must reproduce the above
  28  *        copyright notice, this list of conditions and the following
  29  *        disclaimer in the documentation and/or other materials
  30  *        provided with the distribution.
  31  *
  32  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  33  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  34  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  35  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
  36  * COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
  37  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  38  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  39  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  40  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  41  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
  42  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
  43  * OF THE POSSIBILITY OF SUCH DAMAGE.
  44  */
  45
  46 #include <asm/page.h>
  47 #include <linux/io.h>
  48 #include <linux/wait.h>
  49 #include <rdma/ib_addr.h>
  50 #include <rdma/ib_smi.h>
  51 #include <rdma/ib_user_verbs.h>
  52 #include <rdma/uverbs_ioctl.h>
  53
  54 #include "pvrdma.h"
  55
  56 /**
  57  * pvrdma_req_notify_cq - request notification for a completion queue
  58  * @ibcq: the completion queue
  59  * @notify_flags: notification flags
  60  *
  61  * @return: 0 for success.
  62  */
  63 int pvrdma_req_notify_cq(struct ib_cq *ibcq,
  64                          enum ib_cq_notify_flags notify_flags)
  65 {
  66         struct pvrdma_dev *dev = to_vdev(ibcq->device);
  67         struct pvrdma_cq *cq = to_vcq(ibcq);
  68         u32 val = cq->cq_handle;
  69         unsigned long flags;
  70         int has_data = 0;
  71
  72         val |= (notify_flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
  73                 PVRDMA_UAR_CQ_ARM_SOL : PVRDMA_UAR_CQ_ARM;
  74
  75         spin_lock_irqsave(&cq->cq_lock, flags);
  76
  77         pvrdma_write_uar_cq(dev, val);
  78
  79         if (notify_flags & IB_CQ_REPORT_MISSED_EVENTS) {
  80                 unsigned int head;
  81
  82                 has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
  83                                                     cq->ibcq.cqe, &head);
  84                 if (unlikely(has_data == PVRDMA_INVALID_IDX))
  85                         dev_err(&dev->pdev->dev, "CQ ring state invalid\n");
  86         }
  87
  88         spin_unlock_irqrestore(&cq->cq_lock, flags);
  89
  90         return has_data;
  91 }
  92
  93 /**
  94  * pvrdma_create_cq - create completion queue
  95  * @ibdev: the device
  96  * @attr: completion queue attributes
  97  * @udata: user data
  98  *
  99  * @return: ib_cq completion queue pointer on success,
 100  *          otherwise returns negative errno.
 101  */
 102 struct ib_cq *pvrdma_create_cq(struct ib_device *ibdev,
 103                                const struct ib_cq_init_attr *attr,
 104                                struct ib_udata *udata)
 105 {
 106         int entries = attr->cqe;
 107         struct pvrdma_dev *dev = to_vdev(ibdev);
 108         struct pvrdma_cq *cq;
 109         int ret;
 110         int npages;
 111         unsigned long flags;
 112         union pvrdma_cmd_req req;
 113         union pvrdma_cmd_resp rsp;
 114         struct pvrdma_cmd_create_cq *cmd = &req.create_cq;
 115         struct pvrdma_cmd_create_cq_resp *resp = &rsp.create_cq_resp;
 116         struct pvrdma_create_cq_resp cq_resp = {0};
 117         struct pvrdma_create_cq ucmd;
 118         struct pvrdma_ucontext *context = rdma_udata_to_drv_context(
 119                 udata, struct pvrdma_ucontext, ibucontext);
 120
 121         BUILD_BUG_ON(sizeof(struct pvrdma_cqe) != 64);
 122
 123         entries = roundup_pow_of_two(entries);
 124         if (entries < 1 || entries > dev->dsr->caps.max_cqe)
 125                 return ERR_PTR(-EINVAL);
 126
 127         if (!atomic_add_unless(&dev->num_cqs, 1, dev->dsr->caps.max_cq))
 128                 return ERR_PTR(-ENOMEM);
 129
 130         cq = kzalloc(sizeof(*cq), GFP_KERNEL);
 131         if (!cq) {
 132                 atomic_dec(&dev->num_cqs);
 133                 return ERR_PTR(-ENOMEM);
 134         }
 135
 136         cq->ibcq.cqe = entries;
 137         cq->is_kernel = !udata;
 138
 139         if (!cq->is_kernel) {
 140                 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
 141                         ret = -EFAULT;
 142                         goto err_cq;
 143                 }
 144
 145                 cq->umem = ib_umem_get(udata, ucmd.buf_addr, ucmd.buf_size,
 146                                        IB_ACCESS_LOCAL_WRITE, 1);
 147                 if (IS_ERR(cq->umem)) {
 148                         ret = PTR_ERR(cq->umem);
 149                         goto err_cq;
 150                 }
 151
 152                 npages = ib_umem_page_count(cq->umem);
 153         } else {
 154                 /* One extra page for shared ring state */
 155                 npages = 1 + (entries * sizeof(struct pvrdma_cqe) +
 156                               PAGE_SIZE - 1) / PAGE_SIZE;
 157
 158                 /* Skip header page. */
 159                 cq->offset = PAGE_SIZE;
 160         }
 161
 162         if (npages < 0 || npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
 163                 dev_warn(&dev->pdev->dev,
 164                          "overflow pages in completion queue\n");
 165                 ret = -EINVAL;
 166                 goto err_umem;
 167         }
 168
 169         ret = pvrdma_page_dir_init(dev, &cq->pdir, npages, cq->is_kernel);
 170         if (ret) {
 171                 dev_warn(&dev->pdev->dev,
 172                          "could not allocate page directory\n");
 173                 goto err_umem;
 174         }
 175
 176         /* Ring state is always the first page. Set in library for user cq. */
 177         if (cq->is_kernel)
 178                 cq->ring_state = cq->pdir.pages[0];
 179         else
 180                 pvrdma_page_dir_insert_umem(&cq->pdir, cq->umem, 0);
 181
 182         refcount_set(&cq->refcnt, 1);
 183         init_completion(&cq->free);
 184         spin_lock_init(&cq->cq_lock);
 185
 186         memset(cmd, 0, sizeof(*cmd));
 187         cmd->hdr.cmd = PVRDMA_CMD_CREATE_CQ;
 188         cmd->nchunks = npages;
 189         cmd->ctx_handle = context ? context->ctx_handle : 0;
 190         cmd->cqe = entries;
 191         cmd->pdir_dma = cq->pdir.dir_dma;
 192         ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_CQ_RESP);
 193         if (ret < 0) {
 194                 dev_warn(&dev->pdev->dev,
 195                          "could not create completion queue, error: %d\n", ret);
 196                 goto err_page_dir;
 197         }
 198
 199         cq->ibcq.cqe = resp->cqe;
 200         cq->cq_handle = resp->cq_handle;
 201         cq_resp.cqn = resp->cq_handle;
 202         spin_lock_irqsave(&dev->cq_tbl_lock, flags);
 203         dev->cq_tbl[cq->cq_handle % dev->dsr->caps.max_cq] = cq;
 204         spin_unlock_irqrestore(&dev->cq_tbl_lock, flags);
 205
 206         if (!cq->is_kernel) {
 207                 cq->uar = &context->uar;
 208
 209                 /* Copy udata back. */
 210                 if (ib_copy_to_udata(udata, &cq_resp, sizeof(cq_resp))) {
 211                         dev_warn(&dev->pdev->dev,
 212                                  "failed to copy back udata\n");
 213                         pvrdma_destroy_cq(&cq->ibcq, udata);
 214                         return ERR_PTR(-EINVAL);
 215                 }
 216         }
 217
 218         return &cq->ibcq;
 219
 220 err_page_dir:
 221         pvrdma_page_dir_cleanup(dev, &cq->pdir);
 222 err_umem:
 223         if (!cq->is_kernel)
 224                 ib_umem_release(cq->umem);
 225 err_cq:
 226         atomic_dec(&dev->num_cqs);
 227         kfree(cq);
 228
 229         return ERR_PTR(ret);
 230 }
 231
 232 static void pvrdma_free_cq(struct pvrdma_dev *dev, struct pvrdma_cq *cq)
 233 {
 234         if (refcount_dec_and_test(&cq->refcnt))
 235                 complete(&cq->free);
 236         wait_for_completion(&cq->free);
 237
 238         if (!cq->is_kernel)
 239                 ib_umem_release(cq->umem);
 240
 241         pvrdma_page_dir_cleanup(dev, &cq->pdir);
 242         kfree(cq);
 243 }
 244
 245 /**
 246  * pvrdma_destroy_cq - destroy completion queue
 247  * @cq: the completion queue to destroy.
 248  * @udata: user data or null for kernel object
 249  *
 250  * @return: 0 for success.
 251  */
 252 int pvrdma_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
 253 {
 254         struct pvrdma_cq *vcq = to_vcq(cq);
 255         union pvrdma_cmd_req req;
 256         struct pvrdma_cmd_destroy_cq *cmd = &req.destroy_cq;
 257         struct pvrdma_dev *dev = to_vdev(cq->device);
 258         unsigned long flags;
 259         int ret;
 260
 261         memset(cmd, 0, sizeof(*cmd));
 262         cmd->hdr.cmd = PVRDMA_CMD_DESTROY_CQ;
 263         cmd->cq_handle = vcq->cq_handle;
 264
 265         ret = pvrdma_cmd_post(dev, &req, NULL, 0);
 266         if (ret < 0)
 267                 dev_warn(&dev->pdev->dev,
 268                          "could not destroy completion queue, error: %d\n",
 269                          ret);
 270
 271         /* free cq's resources */
 272         spin_lock_irqsave(&dev->cq_tbl_lock, flags);
 273         dev->cq_tbl[vcq->cq_handle] = NULL;
 274         spin_unlock_irqrestore(&dev->cq_tbl_lock, flags);
 275
 276         pvrdma_free_cq(dev, vcq);
 277         atomic_dec(&dev->num_cqs);
 278
 279         return ret;
 280 }
 281
 282 static inline struct pvrdma_cqe *get_cqe(struct pvrdma_cq *cq, int i)
 283 {
 284         return (struct pvrdma_cqe *)pvrdma_page_dir_get_ptr(
 285                                         &cq->pdir,
 286                                         cq->offset +
 287                                         sizeof(struct pvrdma_cqe) * i);
 288 }
 289
 290 void _pvrdma_flush_cqe(struct pvrdma_qp *qp, struct pvrdma_cq *cq)
 291 {
 292         unsigned int head;
 293         int has_data;
 294
 295         if (!cq->is_kernel)
 296                 return;
 297
 298         /* Lock held */
 299         has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
 300                                             cq->ibcq.cqe, &head);
 301         if (unlikely(has_data > 0)) {
 302                 int items;
 303                 int curr;
 304                 int tail = pvrdma_idx(&cq->ring_state->rx.prod_tail,
 305                                       cq->ibcq.cqe);
 306                 struct pvrdma_cqe *cqe;
 307                 struct pvrdma_cqe *curr_cqe;
 308
 309                 items = (tail > head) ? (tail - head) :
 310                         (cq->ibcq.cqe - head + tail);
 311                 curr = --tail;
 312                 while (items-- > 0) {
 313                         if (curr < 0)
 314                                 curr = cq->ibcq.cqe - 1;
 315                         if (tail < 0)
 316                                 tail = cq->ibcq.cqe - 1;
 317                         curr_cqe = get_cqe(cq, curr);
 318                         if ((curr_cqe->qp & 0xFFFF) != qp->qp_handle) {
 319                                 if (curr != tail) {
 320                                         cqe = get_cqe(cq, tail);
 321                                         *cqe = *curr_cqe;
 322                                 }
 323                                 tail--;
 324                         } else {
 325                                 pvrdma_idx_ring_inc(
 326                                         &cq->ring_state->rx.cons_head,
 327                                         cq->ibcq.cqe);
 328                         }
 329                         curr--;
 330                 }
 331         }
 332 }
 333
 334 static int pvrdma_poll_one(struct pvrdma_cq *cq, struct pvrdma_qp **cur_qp,
 335                            struct ib_wc *wc)
 336 {
 337         struct pvrdma_dev *dev = to_vdev(cq->ibcq.device);
 338         int has_data;
 339         unsigned int head;
 340         bool tried = false;
 341         struct pvrdma_cqe *cqe;
 342
 343 retry:
 344         has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
 345                                             cq->ibcq.cqe, &head);
 346         if (has_data == 0) {
 347                 if (tried)
 348                         return -EAGAIN;
 349
 350                 pvrdma_write_uar_cq(dev, cq->cq_handle | PVRDMA_UAR_CQ_POLL);
 351
 352                 tried = true;
 353                 goto retry;
 354         } else if (has_data == PVRDMA_INVALID_IDX) {
 355                 dev_err(&dev->pdev->dev, "CQ ring state invalid\n");
 356                 return -EAGAIN;
 357         }
 358
 359         cqe = get_cqe(cq, head);
 360
 361         /* Ensure cqe is valid. */
 362         rmb();
 363         if (dev->qp_tbl[cqe->qp & 0xffff])
 364                 *cur_qp = (struct pvrdma_qp *)dev->qp_tbl[cqe->qp & 0xffff];
 365         else
 366                 return -EAGAIN;
 367
 368         wc->opcode = pvrdma_wc_opcode_to_ib(cqe->opcode);
 369         wc->status = pvrdma_wc_status_to_ib(cqe->status);
 370         wc->wr_id = cqe->wr_id;
 371         wc->qp = &(*cur_qp)->ibqp;
 372         wc->byte_len = cqe->byte_len;
 373         wc->ex.imm_data = cqe->imm_data;
 374         wc->src_qp = cqe->src_qp;
 375         wc->wc_flags = pvrdma_wc_flags_to_ib(cqe->wc_flags);
 376         wc->pkey_index = cqe->pkey_index;
 377         wc->slid = cqe->slid;
 378         wc->sl = cqe->sl;
 379         wc->dlid_path_bits = cqe->dlid_path_bits;
 380         wc->port_num = cqe->port_num;
 381         wc->vendor_err = cqe->vendor_err;
 382         wc->network_hdr_type = cqe->network_hdr_type;
 383
 384         /* Update shared ring state */
 385         pvrdma_idx_ring_inc(&cq->ring_state->rx.cons_head, cq->ibcq.cqe);
 386
 387         return 0;
 388 }
 389
 390 /**
 391  * pvrdma_poll_cq - poll for work completion queue entries
 392  * @ibcq: completion queue
 393  * @num_entries: the maximum number of entries
 394  * @entry: pointer to work completion array
 395  *
 396  * @return: number of polled completion entries
 397  */
 398 int pvrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
 399 {
 400         struct pvrdma_cq *cq = to_vcq(ibcq);
 401         struct pvrdma_qp *cur_qp = NULL;
 402         unsigned long flags;
 403         int npolled;
 404
 405         if (num_entries < 1 || wc == NULL)
 406                 return 0;
 407
 408         spin_lock_irqsave(&cq->cq_lock, flags);
 409         for (npolled = 0; npolled < num_entries; ++npolled) {
 410                 if (pvrdma_poll_one(cq, &cur_qp, wc + npolled))
 411                         break;
 412         }
 413
 414         spin_unlock_irqrestore(&cq->cq_lock, flags);
 415
 416         /* Ensure we do not return errors from poll_cq */
 417         return npolled;
 418 }