drivers/infiniband/hw/vmw_pvrdma/pvrdma_qp.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
  53 #include "pvrdma.h"
  54
  55 static inline void get_cqs(struct pvrdma_qp *qp, struct pvrdma_cq **send_cq,
  56                            struct pvrdma_cq **recv_cq)
  57 {
  58         *send_cq = to_vcq(qp->ibqp.send_cq);
  59         *recv_cq = to_vcq(qp->ibqp.recv_cq);
  60 }
  61
  62 static void pvrdma_lock_cqs(struct pvrdma_cq *scq, struct pvrdma_cq *rcq,
  63                             unsigned long *scq_flags,
  64                             unsigned long *rcq_flags)
  65         __acquires(scq->cq_lock) __acquires(rcq->cq_lock)
  66 {
  67         if (scq == rcq) {
  68                 spin_lock_irqsave(&scq->cq_lock, *scq_flags);
  69                 __acquire(rcq->cq_lock);
  70         } else if (scq->cq_handle < rcq->cq_handle) {
  71                 spin_lock_irqsave(&scq->cq_lock, *scq_flags);
  72                 spin_lock_irqsave_nested(&rcq->cq_lock, *rcq_flags,
  73                                          SINGLE_DEPTH_NESTING);
  74         } else {
  75                 spin_lock_irqsave(&rcq->cq_lock, *rcq_flags);
  76                 spin_lock_irqsave_nested(&scq->cq_lock, *scq_flags,
  77                                          SINGLE_DEPTH_NESTING);
  78         }
  79 }
  80
  81 static void pvrdma_unlock_cqs(struct pvrdma_cq *scq, struct pvrdma_cq *rcq,
  82                               unsigned long *scq_flags,
  83                               unsigned long *rcq_flags)
  84         __releases(scq->cq_lock) __releases(rcq->cq_lock)
  85 {
  86         if (scq == rcq) {
  87                 __release(rcq->cq_lock);
  88                 spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
  89         } else if (scq->cq_handle < rcq->cq_handle) {
  90                 spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags);
  91                 spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
  92         } else {
  93                 spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
  94                 spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags);
  95         }
  96 }
  97
  98 static void pvrdma_reset_qp(struct pvrdma_qp *qp)
  99 {
 100         struct pvrdma_cq *scq, *rcq;
 101         unsigned long scq_flags, rcq_flags;
 102
 103         /* Clean up cqes */
 104         get_cqs(qp, &scq, &rcq);
 105         pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
 106
 107         _pvrdma_flush_cqe(qp, scq);
 108         if (scq != rcq)
 109                 _pvrdma_flush_cqe(qp, rcq);
 110
 111         pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
 112
 113         /*
 114          * Reset queuepair. The checks are because usermode queuepairs won't
 115          * have kernel ringstates.
 116          */
 117         if (qp->rq.ring) {
 118                 atomic_set(&qp->rq.ring->cons_head, 0);
 119                 atomic_set(&qp->rq.ring->prod_tail, 0);
 120         }
 121         if (qp->sq.ring) {
 122                 atomic_set(&qp->sq.ring->cons_head, 0);
 123                 atomic_set(&qp->sq.ring->prod_tail, 0);
 124         }
 125 }
 126
 127 static int pvrdma_set_rq_size(struct pvrdma_dev *dev,
 128                               struct ib_qp_cap *req_cap,
 129                               struct pvrdma_qp *qp)
 130 {
 131         if (req_cap->max_recv_wr > dev->dsr->caps.max_qp_wr ||
 132             req_cap->max_recv_sge > dev->dsr->caps.max_sge) {
 133                 dev_warn(&dev->pdev->dev, "recv queue size invalid\n");
 134                 return -EINVAL;
 135         }
 136
 137         qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, req_cap->max_recv_wr));
 138         qp->rq.max_sg = roundup_pow_of_two(max(1U, req_cap->max_recv_sge));
 139
 140         /* Write back */
 141         req_cap->max_recv_wr = qp->rq.wqe_cnt;
 142         req_cap->max_recv_sge = qp->rq.max_sg;
 143
 144         qp->rq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_rq_wqe_hdr) +
 145                                              sizeof(struct pvrdma_sge) *
 146                                              qp->rq.max_sg);
 147         qp->npages_recv = (qp->rq.wqe_cnt * qp->rq.wqe_size + PAGE_SIZE - 1) /
 148                           PAGE_SIZE;
 149
 150         return 0;
 151 }
 152
 153 static int pvrdma_set_sq_size(struct pvrdma_dev *dev, struct ib_qp_cap *req_cap,
 154                               struct pvrdma_qp *qp)
 155 {
 156         if (req_cap->max_send_wr > dev->dsr->caps.max_qp_wr ||
 157             req_cap->max_send_sge > dev->dsr->caps.max_sge) {
 158                 dev_warn(&dev->pdev->dev, "send queue size invalid\n");
 159                 return -EINVAL;
 160         }
 161
 162         qp->sq.wqe_cnt = roundup_pow_of_two(max(1U, req_cap->max_send_wr));
 163         qp->sq.max_sg = roundup_pow_of_two(max(1U, req_cap->max_send_sge));
 164
 165         /* Write back */
 166         req_cap->max_send_wr = qp->sq.wqe_cnt;
 167         req_cap->max_send_sge = qp->sq.max_sg;
 168
 169         qp->sq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_sq_wqe_hdr) +
 170                                              sizeof(struct pvrdma_sge) *
 171                                              qp->sq.max_sg);
 172         /* Note: one extra page for the header. */
 173         qp->npages_send = PVRDMA_QP_NUM_HEADER_PAGES +
 174                           (qp->sq.wqe_cnt * qp->sq.wqe_size + PAGE_SIZE - 1) /
 175                                                                 PAGE_SIZE;
 176
 177         return 0;
 178 }
 179
 180 /**
 181  * pvrdma_create_qp - create queue pair
 182  * @pd: protection domain
 183  * @init_attr: queue pair attributes
 184  * @udata: user data
 185  *
 186  * @return: the ib_qp pointer on success, otherwise returns an errno.
 187  */
 188 struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
 189                                struct ib_qp_init_attr *init_attr,
 190                                struct ib_udata *udata)
 191 {
 192         struct pvrdma_qp *qp = NULL;
 193         struct pvrdma_dev *dev = to_vdev(pd->device);
 194         union pvrdma_cmd_req req;
 195         union pvrdma_cmd_resp rsp;
 196         struct pvrdma_cmd_create_qp *cmd = &req.create_qp;
 197         struct pvrdma_cmd_create_qp_resp *resp = &rsp.create_qp_resp;
 198         struct pvrdma_create_qp ucmd;
 199         unsigned long flags;
 200         int ret;
 201         bool is_srq = !!init_attr->srq;
 202
 203         if (init_attr->create_flags) {
 204                 dev_warn(&dev->pdev->dev,
 205                          "invalid create queuepair flags %#x\n",
 206                          init_attr->create_flags);
 207                 return ERR_PTR(-EINVAL);
 208         }
 209
 210         if (init_attr->qp_type != IB_QPT_RC &&
 211             init_attr->qp_type != IB_QPT_UD &&
 212             init_attr->qp_type != IB_QPT_GSI) {
 213                 dev_warn(&dev->pdev->dev, "queuepair type %d not supported\n",
 214                          init_attr->qp_type);
 215                 return ERR_PTR(-EINVAL);
 216         }
 217
 218         if (is_srq && !dev->dsr->caps.max_srq) {
 219                 dev_warn(&dev->pdev->dev,
 220                          "SRQs not supported by device\n");
 221                 return ERR_PTR(-EINVAL);
 222         }
 223
 224         if (!atomic_add_unless(&dev->num_qps, 1, dev->dsr->caps.max_qp))
 225                 return ERR_PTR(-ENOMEM);
 226
 227         switch (init_attr->qp_type) {
 228         case IB_QPT_GSI:
 229                 if (init_attr->port_num == 0 ||
 230                     init_attr->port_num > pd->device->phys_port_cnt ||
 231                     udata) {
 232                         dev_warn(&dev->pdev->dev, "invalid queuepair attrs\n");
 233                         ret = -EINVAL;
 234                         goto err_qp;
 235                 }
 236                 /* fall through */
 237         case IB_QPT_RC:
 238         case IB_QPT_UD:
 239                 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
 240                 if (!qp) {
 241                         ret = -ENOMEM;
 242                         goto err_qp;
 243                 }
 244
 245                 spin_lock_init(&qp->sq.lock);
 246                 spin_lock_init(&qp->rq.lock);
 247                 mutex_init(&qp->mutex);
 248                 atomic_set(&qp->refcnt, 1);
 249                 init_completion(&qp->free);
 250
 251                 qp->state = IB_QPS_RESET;
 252
 253                 if (pd->uobject && udata) {
 254                         dev_dbg(&dev->pdev->dev,
 255                                 "create queuepair from user space\n");
 256
 257                         if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
 258                                 ret = -EFAULT;
 259                                 goto err_qp;
 260                         }
 261
 262                         if (!is_srq) {
 263                                 /* set qp->sq.wqe_cnt, shift, buf_size.. */
 264                                 qp->rumem = ib_umem_get(pd->uobject->context,
 265                                                         ucmd.rbuf_addr,
 266                                                         ucmd.rbuf_size, 0, 0);
 267                                 if (IS_ERR(qp->rumem)) {
 268                                         ret = PTR_ERR(qp->rumem);
 269                                         goto err_qp;
 270                                 }
 271                                 qp->srq = NULL;
 272                         } else {
 273                                 qp->rumem = NULL;
 274                                 qp->srq = to_vsrq(init_attr->srq);
 275                         }
 276
 277                         qp->sumem = ib_umem_get(pd->uobject->context,
 278                                                 ucmd.sbuf_addr,
 279                                                 ucmd.sbuf_size, 0, 0);
 280                         if (IS_ERR(qp->sumem)) {
 281                                 if (!is_srq)
 282                                         ib_umem_release(qp->rumem);
 283                                 ret = PTR_ERR(qp->sumem);
 284                                 goto err_qp;
 285                         }
 286
 287                         qp->npages_send = ib_umem_page_count(qp->sumem);
 288                         if (!is_srq)
 289                                 qp->npages_recv = ib_umem_page_count(qp->rumem);
 290                         else
 291                                 qp->npages_recv = 0;
 292                         qp->npages = qp->npages_send + qp->npages_recv;
 293                 } else {
 294                         qp->is_kernel = true;
 295
 296                         ret = pvrdma_set_sq_size(to_vdev(pd->device),
 297                                                  &init_attr->cap, qp);
 298                         if (ret)
 299                                 goto err_qp;
 300
 301                         ret = pvrdma_set_rq_size(to_vdev(pd->device),
 302                                                  &init_attr->cap, qp);
 303                         if (ret)
 304                                 goto err_qp;
 305
 306                         qp->npages = qp->npages_send + qp->npages_recv;
 307
 308                         /* Skip header page. */
 309                         qp->sq.offset = PVRDMA_QP_NUM_HEADER_PAGES * PAGE_SIZE;
 310
 311                         /* Recv queue pages are after send pages. */
 312                         qp->rq.offset = qp->npages_send * PAGE_SIZE;
 313                 }
 314
 315                 if (qp->npages < 0 || qp->npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
 316                         dev_warn(&dev->pdev->dev,
 317                                  "overflow pages in queuepair\n");
 318                         ret = -EINVAL;
 319                         goto err_umem;
 320                 }
 321
 322                 ret = pvrdma_page_dir_init(dev, &qp->pdir, qp->npages,
 323                                            qp->is_kernel);
 324                 if (ret) {
 325                         dev_warn(&dev->pdev->dev,
 326                                  "could not allocate page directory\n");
 327                         goto err_umem;
 328                 }
 329
 330                 if (!qp->is_kernel) {
 331                         pvrdma_page_dir_insert_umem(&qp->pdir, qp->sumem, 0);
 332                         if (!is_srq)
 333                                 pvrdma_page_dir_insert_umem(&qp->pdir,
 334                                                             qp->rumem,
 335                                                             qp->npages_send);
 336                 } else {
 337                         /* Ring state is always the first page. */
 338                         qp->sq.ring = qp->pdir.pages[0];
 339                         qp->rq.ring = is_srq ? NULL : &qp->sq.ring[1];
 340                 }
 341                 break;
 342         default:
 343                 ret = -EINVAL;
 344                 goto err_qp;
 345         }
 346
 347         /* Not supported */
 348         init_attr->cap.max_inline_data = 0;
 349
 350         memset(cmd, 0, sizeof(*cmd));
 351         cmd->hdr.cmd = PVRDMA_CMD_CREATE_QP;
 352         cmd->pd_handle = to_vpd(pd)->pd_handle;
 353         cmd->send_cq_handle = to_vcq(init_attr->send_cq)->cq_handle;
 354         cmd->recv_cq_handle = to_vcq(init_attr->recv_cq)->cq_handle;
 355         if (is_srq)
 356                 cmd->srq_handle = to_vsrq(init_attr->srq)->srq_handle;
 357         else
 358                 cmd->srq_handle = 0;
 359         cmd->max_send_wr = init_attr->cap.max_send_wr;
 360         cmd->max_recv_wr = init_attr->cap.max_recv_wr;
 361         cmd->max_send_sge = init_attr->cap.max_send_sge;
 362         cmd->max_recv_sge = init_attr->cap.max_recv_sge;
 363         cmd->max_inline_data = init_attr->cap.max_inline_data;
 364         cmd->sq_sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
 365         cmd->qp_type = ib_qp_type_to_pvrdma(init_attr->qp_type);
 366         cmd->is_srq = is_srq;
 367         cmd->lkey = 0;
 368         cmd->access_flags = IB_ACCESS_LOCAL_WRITE;
 369         cmd->total_chunks = qp->npages;
 370         cmd->send_chunks = qp->npages_send - PVRDMA_QP_NUM_HEADER_PAGES;
 371         cmd->pdir_dma = qp->pdir.dir_dma;
 372
 373         dev_dbg(&dev->pdev->dev, "create queuepair with %d, %d, %d, %d\n",
 374                 cmd->max_send_wr, cmd->max_recv_wr, cmd->max_send_sge,
 375                 cmd->max_recv_sge);
 376
 377         ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_QP_RESP);
 378         if (ret < 0) {
 379                 dev_warn(&dev->pdev->dev,
 380                          "could not create queuepair, error: %d\n", ret);
 381                 goto err_pdir;
 382         }
 383
 384         /* max_send_wr/_recv_wr/_send_sge/_recv_sge/_inline_data */
 385         qp->qp_handle = resp->qpn;
 386         qp->port = init_attr->port_num;
 387         qp->ibqp.qp_num = resp->qpn;
 388         spin_lock_irqsave(&dev->qp_tbl_lock, flags);
 389         dev->qp_tbl[qp->qp_handle % dev->dsr->caps.max_qp] = qp;
 390         spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
 391
 392         return &qp->ibqp;
 393
 394 err_pdir:
 395         pvrdma_page_dir_cleanup(dev, &qp->pdir);
 396 err_umem:
 397         if (pd->uobject && udata) {
 398                 if (qp->rumem)
 399                         ib_umem_release(qp->rumem);
 400                 if (qp->sumem)
 401                         ib_umem_release(qp->sumem);
 402         }
 403 err_qp:
 404         kfree(qp);
 405         atomic_dec(&dev->num_qps);
 406
 407         return ERR_PTR(ret);
 408 }
 409
 410 static void pvrdma_free_qp(struct pvrdma_qp *qp)
 411 {
 412         struct pvrdma_dev *dev = to_vdev(qp->ibqp.device);
 413         struct pvrdma_cq *scq;
 414         struct pvrdma_cq *rcq;
 415         unsigned long flags, scq_flags, rcq_flags;
 416
 417         /* In case cq is polling */
 418         get_cqs(qp, &scq, &rcq);
 419         pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
 420
 421         _pvrdma_flush_cqe(qp, scq);
 422         if (scq != rcq)
 423                 _pvrdma_flush_cqe(qp, rcq);
 424
 425         spin_lock_irqsave(&dev->qp_tbl_lock, flags);
 426         dev->qp_tbl[qp->qp_handle] = NULL;
 427         spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
 428
 429         pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
 430
 431         if (atomic_dec_and_test(&qp->refcnt))
 432                 complete(&qp->free);
 433         wait_for_completion(&qp->free);
 434
 435         if (!qp->is_kernel) {
 436                 if (qp->rumem)
 437                         ib_umem_release(qp->rumem);
 438                 if (qp->sumem)
 439                         ib_umem_release(qp->sumem);
 440         }
 441
 442         pvrdma_page_dir_cleanup(dev, &qp->pdir);
 443
 444         kfree(qp);
 445
 446         atomic_dec(&dev->num_qps);
 447 }
 448
 449 /**
 450  * pvrdma_destroy_qp - destroy a queue pair
 451  * @qp: the queue pair to destroy
 452  *
 453  * @return: 0 on success.
 454  */
 455 int pvrdma_destroy_qp(struct ib_qp *qp)
 456 {
 457         struct pvrdma_qp *vqp = to_vqp(qp);
 458         union pvrdma_cmd_req req;
 459         struct pvrdma_cmd_destroy_qp *cmd = &req.destroy_qp;
 460         int ret;
 461
 462         memset(cmd, 0, sizeof(*cmd));
 463         cmd->hdr.cmd = PVRDMA_CMD_DESTROY_QP;
 464         cmd->qp_handle = vqp->qp_handle;
 465
 466         ret = pvrdma_cmd_post(to_vdev(qp->device), &req, NULL, 0);
 467         if (ret < 0)
 468                 dev_warn(&to_vdev(qp->device)->pdev->dev,
 469                          "destroy queuepair failed, error: %d\n", ret);
 470
 471         pvrdma_free_qp(vqp);
 472
 473         return 0;
 474 }
 475
 476 /**
 477  * pvrdma_modify_qp - modify queue pair attributes
 478  * @ibqp: the queue pair
 479  * @attr: the new queue pair's attributes
 480  * @attr_mask: attributes mask
 481  * @udata: user data
 482  *
 483  * @returns 0 on success, otherwise returns an errno.
 484  */
 485 int pvrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 486                      int attr_mask, struct ib_udata *udata)
 487 {
 488         struct pvrdma_dev *dev = to_vdev(ibqp->device);
 489         struct pvrdma_qp *qp = to_vqp(ibqp);
 490         union pvrdma_cmd_req req;
 491         union pvrdma_cmd_resp rsp;
 492         struct pvrdma_cmd_modify_qp *cmd = &req.modify_qp;
 493         int cur_state, next_state;
 494         int ret;
 495
 496         /* Sanity checking. Should need lock here */
 497         mutex_lock(&qp->mutex);
 498         cur_state = (attr_mask & IB_QP_CUR_STATE) ? attr->cur_qp_state :
 499                 qp->state;
 500         next_state = (attr_mask & IB_QP_STATE) ? attr->qp_state : cur_state;
 501
 502         if (!ib_modify_qp_is_ok(cur_state, next_state, ibqp->qp_type,
 503                                 attr_mask, IB_LINK_LAYER_ETHERNET)) {
 504                 ret = -EINVAL;
 505                 goto out;
 506         }
 507
 508         if (attr_mask & IB_QP_PORT) {
 509                 if (attr->port_num == 0 ||
 510                     attr->port_num > ibqp->device->phys_port_cnt) {
 511                         ret = -EINVAL;
 512                         goto out;
 513                 }
 514         }
 515
 516         if (attr_mask & IB_QP_MIN_RNR_TIMER) {
 517                 if (attr->min_rnr_timer > 31) {
 518                         ret = -EINVAL;
 519                         goto out;
 520                 }
 521         }
 522
 523         if (attr_mask & IB_QP_PKEY_INDEX) {
 524                 if (attr->pkey_index >= dev->dsr->caps.max_pkeys) {
 525                         ret = -EINVAL;
 526                         goto out;
 527                 }
 528         }
 529
 530         if (attr_mask & IB_QP_QKEY)
 531                 qp->qkey = attr->qkey;
 532
 533         if (cur_state == next_state && cur_state == IB_QPS_RESET) {
 534                 ret = 0;
 535                 goto out;
 536         }
 537
 538         qp->state = next_state;
 539         memset(cmd, 0, sizeof(*cmd));
 540         cmd->hdr.cmd = PVRDMA_CMD_MODIFY_QP;
 541         cmd->qp_handle = qp->qp_handle;
 542         cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
 543         cmd->attrs.qp_state = ib_qp_state_to_pvrdma(attr->qp_state);
 544         cmd->attrs.cur_qp_state =
 545                 ib_qp_state_to_pvrdma(attr->cur_qp_state);
 546         cmd->attrs.path_mtu = ib_mtu_to_pvrdma(attr->path_mtu);
 547         cmd->attrs.path_mig_state =
 548                 ib_mig_state_to_pvrdma(attr->path_mig_state);
 549         cmd->attrs.qkey = attr->qkey;
 550         cmd->attrs.rq_psn = attr->rq_psn;
 551         cmd->attrs.sq_psn = attr->sq_psn;
 552         cmd->attrs.dest_qp_num = attr->dest_qp_num;
 553         cmd->attrs.qp_access_flags =
 554                 ib_access_flags_to_pvrdma(attr->qp_access_flags);
 555         cmd->attrs.pkey_index = attr->pkey_index;
 556         cmd->attrs.alt_pkey_index = attr->alt_pkey_index;
 557         cmd->attrs.en_sqd_async_notify = attr->en_sqd_async_notify;
 558         cmd->attrs.sq_draining = attr->sq_draining;
 559         cmd->attrs.max_rd_atomic = attr->max_rd_atomic;
 560         cmd->attrs.max_dest_rd_atomic = attr->max_dest_rd_atomic;
 561         cmd->attrs.min_rnr_timer = attr->min_rnr_timer;
 562         cmd->attrs.port_num = attr->port_num;
 563         cmd->attrs.timeout = attr->timeout;
 564         cmd->attrs.retry_cnt = attr->retry_cnt;
 565         cmd->attrs.rnr_retry = attr->rnr_retry;
 566         cmd->attrs.alt_port_num = attr->alt_port_num;
 567         cmd->attrs.alt_timeout = attr->alt_timeout;
 568         ib_qp_cap_to_pvrdma(&cmd->attrs.cap, &attr->cap);
 569         rdma_ah_attr_to_pvrdma(&cmd->attrs.ah_attr, &attr->ah_attr);
 570         rdma_ah_attr_to_pvrdma(&cmd->attrs.alt_ah_attr, &attr->alt_ah_attr);
 571
 572         ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_MODIFY_QP_RESP);
 573         if (ret < 0) {
 574                 dev_warn(&dev->pdev->dev,
 575                          "could not modify queuepair, error: %d\n", ret);
 576         } else if (rsp.hdr.err > 0) {
 577                 dev_warn(&dev->pdev->dev,
 578                          "cannot modify queuepair, error: %d\n", rsp.hdr.err);
 579                 ret = -EINVAL;
 580         }
 581
 582         if (ret == 0 && next_state == IB_QPS_RESET)
 583                 pvrdma_reset_qp(qp);
 584
 585 out:
 586         mutex_unlock(&qp->mutex);
 587
 588         return ret;
 589 }
 590
 591 static inline void *get_sq_wqe(struct pvrdma_qp *qp, unsigned int n)
 592 {
 593         return pvrdma_page_dir_get_ptr(&qp->pdir,
 594                                        qp->sq.offset + n * qp->sq.wqe_size);
 595 }
 596
 597 static inline void *get_rq_wqe(struct pvrdma_qp *qp, unsigned int n)
 598 {
 599         return pvrdma_page_dir_get_ptr(&qp->pdir,
 600                                        qp->rq.offset + n * qp->rq.wqe_size);
 601 }
 602
 603 static int set_reg_seg(struct pvrdma_sq_wqe_hdr *wqe_hdr, struct ib_reg_wr *wr)
 604 {
 605         struct pvrdma_user_mr *mr = to_vmr(wr->mr);
 606
 607         wqe_hdr->wr.fast_reg.iova_start = mr->ibmr.iova;
 608         wqe_hdr->wr.fast_reg.pl_pdir_dma = mr->pdir.dir_dma;
 609         wqe_hdr->wr.fast_reg.page_shift = mr->page_shift;
 610         wqe_hdr->wr.fast_reg.page_list_len = mr->npages;
 611         wqe_hdr->wr.fast_reg.length = mr->ibmr.length;
 612         wqe_hdr->wr.fast_reg.access_flags = wr->access;
 613         wqe_hdr->wr.fast_reg.rkey = wr->key;
 614
 615         return pvrdma_page_dir_insert_page_list(&mr->pdir, mr->pages,
 616                                                 mr->npages);
 617 }
 618
 619 /**
 620  * pvrdma_post_send - post send work request entries on a QP
 621  * @ibqp: the QP
 622  * @wr: work request list to post
 623  * @bad_wr: the first bad WR returned
 624  *
 625  * @return: 0 on success, otherwise errno returned.
 626  */
 627 int pvrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 628                      struct ib_send_wr **bad_wr)
 629 {
 630         struct pvrdma_qp *qp = to_vqp(ibqp);
 631         struct pvrdma_dev *dev = to_vdev(ibqp->device);
 632         unsigned long flags;
 633         struct pvrdma_sq_wqe_hdr *wqe_hdr;
 634         struct pvrdma_sge *sge;
 635         int i, ret;
 636
 637         /*
 638          * In states lower than RTS, we can fail immediately. In other states,
 639          * just post and let the device figure it out.
 640          */
 641         if (qp->state < IB_QPS_RTS) {
 642                 *bad_wr = wr;
 643                 return -EINVAL;
 644         }
 645
 646         spin_lock_irqsave(&qp->sq.lock, flags);
 647
 648         while (wr) {
 649                 unsigned int tail = 0;
 650
 651                 if (unlikely(!pvrdma_idx_ring_has_space(
 652                                 qp->sq.ring, qp->sq.wqe_cnt, &tail))) {
 653                         dev_warn_ratelimited(&dev->pdev->dev,
 654                                              "send queue is full\n");
 655                         *bad_wr = wr;
 656                         ret = -ENOMEM;
 657                         goto out;
 658                 }
 659
 660                 if (unlikely(wr->num_sge > qp->sq.max_sg || wr->num_sge < 0)) {
 661                         dev_warn_ratelimited(&dev->pdev->dev,
 662                                              "send SGE overflow\n");
 663                         *bad_wr = wr;
 664                         ret = -EINVAL;
 665                         goto out;
 666                 }
 667
 668                 if (unlikely(wr->opcode < 0)) {
 669                         dev_warn_ratelimited(&dev->pdev->dev,
 670                                              "invalid send opcode\n");
 671                         *bad_wr = wr;
 672                         ret = -EINVAL;
 673                         goto out;
 674                 }
 675
 676                 /*
 677                  * Only support UD, RC.
 678                  * Need to check opcode table for thorough checking.
 679                  * opcode               _UD     _UC     _RC
 680                  * _SEND                x       x       x
 681                  * _SEND_WITH_IMM       x       x       x
 682                  * _RDMA_WRITE                  x       x
 683                  * _RDMA_WRITE_WITH_IMM         x       x
 684                  * _LOCAL_INV                   x       x
 685                  * _SEND_WITH_INV               x       x
 686                  * _RDMA_READ                           x
 687                  * _ATOMIC_CMP_AND_SWP                  x
 688                  * _ATOMIC_FETCH_AND_ADD                x
 689                  * _MASK_ATOMIC_CMP_AND_SWP             x
 690                  * _MASK_ATOMIC_FETCH_AND_ADD           x
 691                  * _REG_MR                              x
 692                  *
 693                  */
 694                 if (qp->ibqp.qp_type != IB_QPT_UD &&
 695                     qp->ibqp.qp_type != IB_QPT_RC &&
 696                         wr->opcode != IB_WR_SEND) {
 697                         dev_warn_ratelimited(&dev->pdev->dev,
 698                                              "unsupported queuepair type\n");
 699                         *bad_wr = wr;
 700                         ret = -EINVAL;
 701                         goto out;
 702                 } else if (qp->ibqp.qp_type == IB_QPT_UD ||
 703                            qp->ibqp.qp_type == IB_QPT_GSI) {
 704                         if (wr->opcode != IB_WR_SEND &&
 705                             wr->opcode != IB_WR_SEND_WITH_IMM) {
 706                                 dev_warn_ratelimited(&dev->pdev->dev,
 707                                                      "invalid send opcode\n");
 708                                 *bad_wr = wr;
 709                                 ret = -EINVAL;
 710                                 goto out;
 711                         }
 712                 }
 713
 714                 wqe_hdr = (struct pvrdma_sq_wqe_hdr *)get_sq_wqe(qp, tail);
 715                 memset(wqe_hdr, 0, sizeof(*wqe_hdr));
 716                 wqe_hdr->wr_id = wr->wr_id;
 717                 wqe_hdr->num_sge = wr->num_sge;
 718                 wqe_hdr->opcode = ib_wr_opcode_to_pvrdma(wr->opcode);
 719                 wqe_hdr->send_flags = ib_send_flags_to_pvrdma(wr->send_flags);
 720                 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
 721                     wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
 722                         wqe_hdr->ex.imm_data = wr->ex.imm_data;
 723
 724                 switch (qp->ibqp.qp_type) {
 725                 case IB_QPT_GSI:
 726                 case IB_QPT_UD:
 727                         if (unlikely(!ud_wr(wr)->ah)) {
 728                                 dev_warn_ratelimited(&dev->pdev->dev,
 729                                                      "invalid address handle\n");
 730                                 *bad_wr = wr;
 731                                 ret = -EINVAL;
 732                                 goto out;
 733                         }
 734
 735                         /*
 736                          * Use qkey from qp context if high order bit set,
 737                          * otherwise from work request.
 738                          */
 739                         wqe_hdr->wr.ud.remote_qpn = ud_wr(wr)->remote_qpn;
 740                         wqe_hdr->wr.ud.remote_qkey =
 741                                 ud_wr(wr)->remote_qkey & 0x80000000 ?
 742                                 qp->qkey : ud_wr(wr)->remote_qkey;
 743                         wqe_hdr->wr.ud.av = to_vah(ud_wr(wr)->ah)->av;
 744
 745                         break;
 746                 case IB_QPT_RC:
 747                         switch (wr->opcode) {
 748                         case IB_WR_RDMA_READ:
 749                         case IB_WR_RDMA_WRITE:
 750                         case IB_WR_RDMA_WRITE_WITH_IMM:
 751                                 wqe_hdr->wr.rdma.remote_addr =
 752                                         rdma_wr(wr)->remote_addr;
 753                                 wqe_hdr->wr.rdma.rkey = rdma_wr(wr)->rkey;
 754                                 break;
 755                         case IB_WR_LOCAL_INV:
 756                         case IB_WR_SEND_WITH_INV:
 757                                 wqe_hdr->ex.invalidate_rkey =
 758                                         wr->ex.invalidate_rkey;
 759                                 break;
 760                         case IB_WR_ATOMIC_CMP_AND_SWP:
 761                         case IB_WR_ATOMIC_FETCH_AND_ADD:
 762                                 wqe_hdr->wr.atomic.remote_addr =
 763                                         atomic_wr(wr)->remote_addr;
 764                                 wqe_hdr->wr.atomic.rkey = atomic_wr(wr)->rkey;
 765                                 wqe_hdr->wr.atomic.compare_add =
 766                                         atomic_wr(wr)->compare_add;
 767                                 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP)
 768                                         wqe_hdr->wr.atomic.swap =
 769                                                 atomic_wr(wr)->swap;
 770                                 break;
 771                         case IB_WR_REG_MR:
 772                                 ret = set_reg_seg(wqe_hdr, reg_wr(wr));
 773                                 if (ret < 0) {
 774                                         dev_warn_ratelimited(&dev->pdev->dev,
 775                                                              "Failed to set fast register work request\n");
 776                                         *bad_wr = wr;
 777                                         goto out;
 778                                 }
 779                                 break;
 780                         default:
 781                                 break;
 782                         }
 783
 784                         break;
 785                 default:
 786                         dev_warn_ratelimited(&dev->pdev->dev,
 787                                              "invalid queuepair type\n");
 788                         ret = -EINVAL;
 789                         *bad_wr = wr;
 790                         goto out;
 791                 }
 792
 793                 sge = (struct pvrdma_sge *)(wqe_hdr + 1);
 794                 for (i = 0; i < wr->num_sge; i++) {
 795                         /* Need to check wqe_size 0 or max size */
 796                         sge->addr = wr->sg_list[i].addr;
 797                         sge->length = wr->sg_list[i].length;
 798                         sge->lkey = wr->sg_list[i].lkey;
 799                         sge++;
 800                 }
 801
 802                 /* Make sure wqe is written before index update */
 803                 smp_wmb();
 804
 805                 /* Update shared sq ring */
 806                 pvrdma_idx_ring_inc(&qp->sq.ring->prod_tail,
 807                                     qp->sq.wqe_cnt);
 808
 809                 wr = wr->next;
 810         }
 811
 812         ret = 0;
 813
 814 out:
 815         spin_unlock_irqrestore(&qp->sq.lock, flags);
 816
 817         if (!ret)
 818                 pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_SEND | qp->qp_handle);
 819
 820         return ret;
 821 }
 822
 823 /**
 824  * pvrdma_post_receive - post receive work request entries on a QP
 825  * @ibqp: the QP
 826  * @wr: the work request list to post
 827  * @bad_wr: the first bad WR returned
 828  *
 829  * @return: 0 on success, otherwise errno returned.
 830  */
 831 int pvrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 832                      struct ib_recv_wr **bad_wr)
 833 {
 834         struct pvrdma_dev *dev = to_vdev(ibqp->device);
 835         unsigned long flags;
 836         struct pvrdma_qp *qp = to_vqp(ibqp);
 837         struct pvrdma_rq_wqe_hdr *wqe_hdr;
 838         struct pvrdma_sge *sge;
 839         int ret = 0;
 840         int i;
 841
 842         /*
 843          * In the RESET state, we can fail immediately. For other states,
 844          * just post and let the device figure it out.
 845          */
 846         if (qp->state == IB_QPS_RESET) {
 847                 *bad_wr = wr;
 848                 return -EINVAL;
 849         }
 850
 851         if (qp->srq) {
 852                 dev_warn(&dev->pdev->dev, "QP associated with SRQ\n");
 853                 *bad_wr = wr;
 854                 return -EINVAL;
 855         }
 856
 857         spin_lock_irqsave(&qp->rq.lock, flags);
 858
 859         while (wr) {
 860                 unsigned int tail = 0;
 861
 862                 if (unlikely(wr->num_sge > qp->rq.max_sg ||
 863                              wr->num_sge < 0)) {
 864                         ret = -EINVAL;
 865                         *bad_wr = wr;
 866                         dev_warn_ratelimited(&dev->pdev->dev,
 867                                              "recv SGE overflow\n");
 868                         goto out;
 869                 }
 870
 871                 if (unlikely(!pvrdma_idx_ring_has_space(
 872                                 qp->rq.ring, qp->rq.wqe_cnt, &tail))) {
 873                         ret = -ENOMEM;
 874                         *bad_wr = wr;
 875                         dev_warn_ratelimited(&dev->pdev->dev,
 876                                              "recv queue full\n");
 877                         goto out;
 878                 }
 879
 880                 wqe_hdr = (struct pvrdma_rq_wqe_hdr *)get_rq_wqe(qp, tail);
 881                 wqe_hdr->wr_id = wr->wr_id;
 882                 wqe_hdr->num_sge = wr->num_sge;
 883                 wqe_hdr->total_len = 0;
 884
 885                 sge = (struct pvrdma_sge *)(wqe_hdr + 1);
 886                 for (i = 0; i < wr->num_sge; i++) {
 887                         sge->addr = wr->sg_list[i].addr;
 888                         sge->length = wr->sg_list[i].length;
 889                         sge->lkey = wr->sg_list[i].lkey;
 890                         sge++;
 891                 }
 892
 893                 /* Make sure wqe is written before index update */
 894                 smp_wmb();
 895
 896                 /* Update shared rq ring */
 897                 pvrdma_idx_ring_inc(&qp->rq.ring->prod_tail,
 898                                     qp->rq.wqe_cnt);
 899
 900                 wr = wr->next;
 901         }
 902
 903         spin_unlock_irqrestore(&qp->rq.lock, flags);
 904
 905         pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_RECV | qp->qp_handle);
 906
 907         return ret;
 908
 909 out:
 910         spin_unlock_irqrestore(&qp->rq.lock, flags);
 911
 912         return ret;
 913 }
 914
 915 /**
 916  * pvrdma_query_qp - query a queue pair's attributes
 917  * @ibqp: the queue pair to query
 918  * @attr: the queue pair's attributes
 919  * @attr_mask: attributes mask
 920  * @init_attr: initial queue pair attributes
 921  *
 922  * @returns 0 on success, otherwise returns an errno.
 923  */
 924 int pvrdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 925                     int attr_mask, struct ib_qp_init_attr *init_attr)
 926 {
 927         struct pvrdma_dev *dev = to_vdev(ibqp->device);
 928         struct pvrdma_qp *qp = to_vqp(ibqp);
 929         union pvrdma_cmd_req req;
 930         union pvrdma_cmd_resp rsp;
 931         struct pvrdma_cmd_query_qp *cmd = &req.query_qp;
 932         struct pvrdma_cmd_query_qp_resp *resp = &rsp.query_qp_resp;
 933         int ret = 0;
 934
 935         mutex_lock(&qp->mutex);
 936
 937         if (qp->state == IB_QPS_RESET) {
 938                 attr->qp_state = IB_QPS_RESET;
 939                 goto out;
 940         }
 941
 942         memset(cmd, 0, sizeof(*cmd));
 943         cmd->hdr.cmd = PVRDMA_CMD_QUERY_QP;
 944         cmd->qp_handle = qp->qp_handle;
 945         cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
 946
 947         ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_QUERY_QP_RESP);
 948         if (ret < 0) {
 949                 dev_warn(&dev->pdev->dev,
 950                          "could not query queuepair, error: %d\n", ret);
 951                 goto out;
 952         }
 953
 954         attr->qp_state = pvrdma_qp_state_to_ib(resp->attrs.qp_state);
 955         attr->cur_qp_state =
 956                 pvrdma_qp_state_to_ib(resp->attrs.cur_qp_state);
 957         attr->path_mtu = pvrdma_mtu_to_ib(resp->attrs.path_mtu);
 958         attr->path_mig_state =
 959                 pvrdma_mig_state_to_ib(resp->attrs.path_mig_state);
 960         attr->qkey = resp->attrs.qkey;
 961         attr->rq_psn = resp->attrs.rq_psn;
 962         attr->sq_psn = resp->attrs.sq_psn;
 963         attr->dest_qp_num = resp->attrs.dest_qp_num;
 964         attr->qp_access_flags =
 965                 pvrdma_access_flags_to_ib(resp->attrs.qp_access_flags);
 966         attr->pkey_index = resp->attrs.pkey_index;
 967         attr->alt_pkey_index = resp->attrs.alt_pkey_index;
 968         attr->en_sqd_async_notify = resp->attrs.en_sqd_async_notify;
 969         attr->sq_draining = resp->attrs.sq_draining;
 970         attr->max_rd_atomic = resp->attrs.max_rd_atomic;
 971         attr->max_dest_rd_atomic = resp->attrs.max_dest_rd_atomic;
 972         attr->min_rnr_timer = resp->attrs.min_rnr_timer;
 973         attr->port_num = resp->attrs.port_num;
 974         attr->timeout = resp->attrs.timeout;
 975         attr->retry_cnt = resp->attrs.retry_cnt;
 976         attr->rnr_retry = resp->attrs.rnr_retry;
 977         attr->alt_port_num = resp->attrs.alt_port_num;
 978         attr->alt_timeout = resp->attrs.alt_timeout;
 979         pvrdma_qp_cap_to_ib(&attr->cap, &resp->attrs.cap);
 980         pvrdma_ah_attr_to_rdma(&attr->ah_attr, &resp->attrs.ah_attr);
 981         pvrdma_ah_attr_to_rdma(&attr->alt_ah_attr, &resp->attrs.alt_ah_attr);
 982
 983         qp->state = attr->qp_state;
 984
 985         ret = 0;
 986
 987 out:
 988         attr->cur_qp_state = attr->qp_state;
 989
 990         init_attr->event_handler = qp->ibqp.event_handler;
 991         init_attr->qp_context = qp->ibqp.qp_context;
 992         init_attr->send_cq = qp->ibqp.send_cq;
 993         init_attr->recv_cq = qp->ibqp.recv_cq;
 994         init_attr->srq = qp->ibqp.srq;
 995         init_attr->xrcd = NULL;
 996         init_attr->cap = attr->cap;
 997         init_attr->sq_sig_type = 0;
 998         init_attr->qp_type = qp->ibqp.qp_type;
 999         init_attr->create_flags = 0;
1000         init_attr->port_num = qp->port;
1001
1002         mutex_unlock(&qp->mutex);
1003         return ret;
1004 }