drivers/infiniband/hw/mlx5/cq.c

   1 /*
   2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <linux/kref.h>
  34 #include <rdma/ib_umem.h>
  35 #include <rdma/ib_user_verbs.h>
  36 #include <rdma/ib_cache.h>
  37 #include "mlx5_ib.h"
  38
  39 static void mlx5_ib_cq_comp(struct mlx5_core_cq *cq)
  40 {
  41         struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
  42
  43         ibcq->comp_handler(ibcq, ibcq->cq_context);
  44 }
  45
  46 static void mlx5_ib_cq_event(struct mlx5_core_cq *mcq, enum mlx5_event type)
  47 {
  48         struct mlx5_ib_cq *cq = container_of(mcq, struct mlx5_ib_cq, mcq);
  49         struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
  50         struct ib_cq *ibcq = &cq->ibcq;
  51         struct ib_event event;
  52
  53         if (type != MLX5_EVENT_TYPE_CQ_ERROR) {
  54                 mlx5_ib_warn(dev, "Unexpected event type %d on CQ %06x\n",
  55                              type, mcq->cqn);
  56                 return;
  57         }
  58
  59         if (ibcq->event_handler) {
  60                 event.device     = &dev->ib_dev;
  61                 event.event      = IB_EVENT_CQ_ERR;
  62                 event.element.cq = ibcq;
  63                 ibcq->event_handler(&event, ibcq->cq_context);
  64         }
  65 }
  66
  67 static void *get_cqe_from_buf(struct mlx5_ib_cq_buf *buf, int n, int size)
  68 {
  69         return mlx5_buf_offset(&buf->buf, n * size);
  70 }
  71
  72 static void *get_cqe(struct mlx5_ib_cq *cq, int n)
  73 {
  74         return get_cqe_from_buf(&cq->buf, n, cq->mcq.cqe_sz);
  75 }
  76
  77 static u8 sw_ownership_bit(int n, int nent)
  78 {
  79         return (n & nent) ? 1 : 0;
  80 }
  81
  82 static void *get_sw_cqe(struct mlx5_ib_cq *cq, int n)
  83 {
  84         void *cqe = get_cqe(cq, n & cq->ibcq.cqe);
  85         struct mlx5_cqe64 *cqe64;
  86
  87         cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
  88
  89         if (likely((cqe64->op_own) >> 4 != MLX5_CQE_INVALID) &&
  90             !((cqe64->op_own & MLX5_CQE_OWNER_MASK) ^ !!(n & (cq->ibcq.cqe + 1)))) {
  91                 return cqe;
  92         } else {
  93                 return NULL;
  94         }
  95 }
  96
  97 static void *next_cqe_sw(struct mlx5_ib_cq *cq)
  98 {
  99         return get_sw_cqe(cq, cq->mcq.cons_index);
 100 }
 101
 102 static enum ib_wc_opcode get_umr_comp(struct mlx5_ib_wq *wq, int idx)
 103 {
 104         switch (wq->wr_data[idx]) {
 105         case MLX5_IB_WR_UMR:
 106                 return 0;
 107
 108         case IB_WR_LOCAL_INV:
 109                 return IB_WC_LOCAL_INV;
 110
 111         case IB_WR_REG_MR:
 112                 return IB_WC_REG_MR;
 113
 114         default:
 115                 pr_warn("unknown completion status\n");
 116                 return 0;
 117         }
 118 }
 119
 120 static void handle_good_req(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
 121                             struct mlx5_ib_wq *wq, int idx)
 122 {
 123         wc->wc_flags = 0;
 124         switch (be32_to_cpu(cqe->sop_drop_qpn) >> 24) {
 125         case MLX5_OPCODE_RDMA_WRITE_IMM:
 126                 wc->wc_flags |= IB_WC_WITH_IMM;
 127         case MLX5_OPCODE_RDMA_WRITE:
 128                 wc->opcode    = IB_WC_RDMA_WRITE;
 129                 break;
 130         case MLX5_OPCODE_SEND_IMM:
 131                 wc->wc_flags |= IB_WC_WITH_IMM;
 132         case MLX5_OPCODE_SEND:
 133         case MLX5_OPCODE_SEND_INVAL:
 134                 wc->opcode    = IB_WC_SEND;
 135                 break;
 136         case MLX5_OPCODE_RDMA_READ:
 137                 wc->opcode    = IB_WC_RDMA_READ;
 138                 wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
 139                 break;
 140         case MLX5_OPCODE_ATOMIC_CS:
 141                 wc->opcode    = IB_WC_COMP_SWAP;
 142                 wc->byte_len  = 8;
 143                 break;
 144         case MLX5_OPCODE_ATOMIC_FA:
 145                 wc->opcode    = IB_WC_FETCH_ADD;
 146                 wc->byte_len  = 8;
 147                 break;
 148         case MLX5_OPCODE_ATOMIC_MASKED_CS:
 149                 wc->opcode    = IB_WC_MASKED_COMP_SWAP;
 150                 wc->byte_len  = 8;
 151                 break;
 152         case MLX5_OPCODE_ATOMIC_MASKED_FA:
 153                 wc->opcode    = IB_WC_MASKED_FETCH_ADD;
 154                 wc->byte_len  = 8;
 155                 break;
 156         case MLX5_OPCODE_UMR:
 157                 wc->opcode = get_umr_comp(wq, idx);
 158                 break;
 159         }
 160 }
 161
 162 enum {
 163         MLX5_GRH_IN_BUFFER = 1,
 164         MLX5_GRH_IN_CQE    = 2,
 165 };
 166
 167 static void handle_responder(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
 168                              struct mlx5_ib_qp *qp)
 169 {
 170         enum rdma_link_layer ll = rdma_port_get_link_layer(qp->ibqp.device, 1);
 171         struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
 172         struct mlx5_ib_srq *srq;
 173         struct mlx5_ib_wq *wq;
 174         u16 wqe_ctr;
 175         u8  roce_packet_type;
 176         bool vlan_present;
 177         u8 g;
 178
 179         if (qp->ibqp.srq || qp->ibqp.xrcd) {
 180                 struct mlx5_core_srq *msrq = NULL;
 181
 182                 if (qp->ibqp.xrcd) {
 183                         msrq = mlx5_core_get_srq(dev->mdev,
 184                                                  be32_to_cpu(cqe->srqn));
 185                         srq = to_mibsrq(msrq);
 186                 } else {
 187                         srq = to_msrq(qp->ibqp.srq);
 188                 }
 189                 if (srq) {
 190                         wqe_ctr = be16_to_cpu(cqe->wqe_counter);
 191                         wc->wr_id = srq->wrid[wqe_ctr];
 192                         mlx5_ib_free_srq_wqe(srq, wqe_ctr);
 193                         if (msrq && atomic_dec_and_test(&msrq->refcount))
 194                                 complete(&msrq->free);
 195                 }
 196         } else {
 197                 wq        = &qp->rq;
 198                 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
 199                 ++wq->tail;
 200         }
 201         wc->byte_len = be32_to_cpu(cqe->byte_cnt);
 202
 203         switch (cqe->op_own >> 4) {
 204         case MLX5_CQE_RESP_WR_IMM:
 205                 wc->opcode      = IB_WC_RECV_RDMA_WITH_IMM;
 206                 wc->wc_flags    = IB_WC_WITH_IMM;
 207                 wc->ex.imm_data = cqe->imm_inval_pkey;
 208                 break;
 209         case MLX5_CQE_RESP_SEND:
 210                 wc->opcode   = IB_WC_RECV;
 211                 wc->wc_flags = IB_WC_IP_CSUM_OK;
 212                 if (unlikely(!((cqe->hds_ip_ext & CQE_L3_OK) &&
 213                                (cqe->hds_ip_ext & CQE_L4_OK))))
 214                         wc->wc_flags = 0;
 215                 break;
 216         case MLX5_CQE_RESP_SEND_IMM:
 217                 wc->opcode      = IB_WC_RECV;
 218                 wc->wc_flags    = IB_WC_WITH_IMM;
 219                 wc->ex.imm_data = cqe->imm_inval_pkey;
 220                 break;
 221         case MLX5_CQE_RESP_SEND_INV:
 222                 wc->opcode      = IB_WC_RECV;
 223                 wc->wc_flags    = IB_WC_WITH_INVALIDATE;
 224                 wc->ex.invalidate_rkey = be32_to_cpu(cqe->imm_inval_pkey);
 225                 break;
 226         }
 227         wc->slid           = be16_to_cpu(cqe->slid);
 228         wc->src_qp         = be32_to_cpu(cqe->flags_rqpn) & 0xffffff;
 229         wc->dlid_path_bits = cqe->ml_path;
 230         g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
 231         wc->wc_flags |= g ? IB_WC_GRH : 0;
 232         if (unlikely(is_qp1(qp->ibqp.qp_type))) {
 233                 u16 pkey = be32_to_cpu(cqe->imm_inval_pkey) & 0xffff;
 234
 235                 ib_find_cached_pkey(&dev->ib_dev, qp->port, pkey,
 236                                     &wc->pkey_index);
 237         } else {
 238                 wc->pkey_index = 0;
 239         }
 240
 241         if (ll != IB_LINK_LAYER_ETHERNET) {
 242                 wc->sl = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0xf;
 243                 return;
 244         }
 245
 246         vlan_present = cqe->l4_l3_hdr_type & 0x1;
 247         roce_packet_type   = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0x3;
 248         if (vlan_present) {
 249                 wc->vlan_id = (be16_to_cpu(cqe->vlan_info)) & 0xfff;
 250                 wc->sl = (be16_to_cpu(cqe->vlan_info) >> 13) & 0x7;
 251                 wc->wc_flags |= IB_WC_WITH_VLAN;
 252         } else {
 253                 wc->sl = 0;
 254         }
 255
 256         switch (roce_packet_type) {
 257         case MLX5_CQE_ROCE_L3_HEADER_TYPE_GRH:
 258                 wc->network_hdr_type = RDMA_NETWORK_IB;
 259                 break;
 260         case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV6:
 261                 wc->network_hdr_type = RDMA_NETWORK_IPV6;
 262                 break;
 263         case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV4:
 264                 wc->network_hdr_type = RDMA_NETWORK_IPV4;
 265                 break;
 266         }
 267         wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
 268 }
 269
 270 static void dump_cqe(struct mlx5_ib_dev *dev, struct mlx5_err_cqe *cqe)
 271 {
 272         __be32 *p = (__be32 *)cqe;
 273         int i;
 274
 275         mlx5_ib_warn(dev, "dump error cqe\n");
 276         for (i = 0; i < sizeof(*cqe) / 16; i++, p += 4)
 277                 pr_info("%08x %08x %08x %08x\n", be32_to_cpu(p[0]),
 278                         be32_to_cpu(p[1]), be32_to_cpu(p[2]),
 279                         be32_to_cpu(p[3]));
 280 }
 281
 282 static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev,
 283                                   struct mlx5_err_cqe *cqe,
 284                                   struct ib_wc *wc)
 285 {
 286         int dump = 1;
 287
 288         switch (cqe->syndrome) {
 289         case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR:
 290                 wc->status = IB_WC_LOC_LEN_ERR;
 291                 break;
 292         case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR:
 293                 wc->status = IB_WC_LOC_QP_OP_ERR;
 294                 break;
 295         case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR:
 296                 wc->status = IB_WC_LOC_PROT_ERR;
 297                 break;
 298         case MLX5_CQE_SYNDROME_WR_FLUSH_ERR:
 299                 dump = 0;
 300                 wc->status = IB_WC_WR_FLUSH_ERR;
 301                 break;
 302         case MLX5_CQE_SYNDROME_MW_BIND_ERR:
 303                 wc->status = IB_WC_MW_BIND_ERR;
 304                 break;
 305         case MLX5_CQE_SYNDROME_BAD_RESP_ERR:
 306                 wc->status = IB_WC_BAD_RESP_ERR;
 307                 break;
 308         case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR:
 309                 wc->status = IB_WC_LOC_ACCESS_ERR;
 310                 break;
 311         case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
 312                 wc->status = IB_WC_REM_INV_REQ_ERR;
 313                 break;
 314         case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR:
 315                 wc->status = IB_WC_REM_ACCESS_ERR;
 316                 break;
 317         case MLX5_CQE_SYNDROME_REMOTE_OP_ERR:
 318                 wc->status = IB_WC_REM_OP_ERR;
 319                 break;
 320         case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
 321                 wc->status = IB_WC_RETRY_EXC_ERR;
 322                 dump = 0;
 323                 break;
 324         case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
 325                 wc->status = IB_WC_RNR_RETRY_EXC_ERR;
 326                 dump = 0;
 327                 break;
 328         case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR:
 329                 wc->status = IB_WC_REM_ABORT_ERR;
 330                 break;
 331         default:
 332                 wc->status = IB_WC_GENERAL_ERR;
 333                 break;
 334         }
 335
 336         wc->vendor_err = cqe->vendor_err_synd;
 337         if (dump)
 338                 dump_cqe(dev, cqe);
 339 }
 340
 341 static int is_atomic_response(struct mlx5_ib_qp *qp, uint16_t idx)
 342 {
 343         /* TBD: waiting decision
 344         */
 345         return 0;
 346 }
 347
 348 static void *mlx5_get_atomic_laddr(struct mlx5_ib_qp *qp, uint16_t idx)
 349 {
 350         struct mlx5_wqe_data_seg *dpseg;
 351         void *addr;
 352
 353         dpseg = mlx5_get_send_wqe(qp, idx) + sizeof(struct mlx5_wqe_ctrl_seg) +
 354                 sizeof(struct mlx5_wqe_raddr_seg) +
 355                 sizeof(struct mlx5_wqe_atomic_seg);
 356         addr = (void *)(unsigned long)be64_to_cpu(dpseg->addr);
 357         return addr;
 358 }
 359
 360 static void handle_atomic(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
 361                           uint16_t idx)
 362 {
 363         void *addr;
 364         int byte_count;
 365         int i;
 366
 367         if (!is_atomic_response(qp, idx))
 368                 return;
 369
 370         byte_count = be32_to_cpu(cqe64->byte_cnt);
 371         addr = mlx5_get_atomic_laddr(qp, idx);
 372
 373         if (byte_count == 4) {
 374                 *(uint32_t *)addr = be32_to_cpu(*((__be32 *)addr));
 375         } else {
 376                 for (i = 0; i < byte_count; i += 8) {
 377                         *(uint64_t *)addr = be64_to_cpu(*((__be64 *)addr));
 378                         addr += 8;
 379                 }
 380         }
 381
 382         return;
 383 }
 384
 385 static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
 386                            u16 tail, u16 head)
 387 {
 388         u16 idx;
 389
 390         do {
 391                 idx = tail & (qp->sq.wqe_cnt - 1);
 392                 handle_atomic(qp, cqe64, idx);
 393                 if (idx == head)
 394                         break;
 395
 396                 tail = qp->sq.w_list[idx].next;
 397         } while (1);
 398         tail = qp->sq.w_list[idx].next;
 399         qp->sq.last_poll = tail;
 400 }
 401
 402 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf)
 403 {
 404         mlx5_buf_free(dev->mdev, &buf->buf);
 405 }
 406
 407 static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe,
 408                              struct ib_sig_err *item)
 409 {
 410         u16 syndrome = be16_to_cpu(cqe->syndrome);
 411
 412 #define GUARD_ERR   (1 << 13)
 413 #define APPTAG_ERR  (1 << 12)
 414 #define REFTAG_ERR  (1 << 11)
 415
 416         if (syndrome & GUARD_ERR) {
 417                 item->err_type = IB_SIG_BAD_GUARD;
 418                 item->expected = be32_to_cpu(cqe->expected_trans_sig) >> 16;
 419                 item->actual = be32_to_cpu(cqe->actual_trans_sig) >> 16;
 420         } else
 421         if (syndrome & REFTAG_ERR) {
 422                 item->err_type = IB_SIG_BAD_REFTAG;
 423                 item->expected = be32_to_cpu(cqe->expected_reftag);
 424                 item->actual = be32_to_cpu(cqe->actual_reftag);
 425         } else
 426         if (syndrome & APPTAG_ERR) {
 427                 item->err_type = IB_SIG_BAD_APPTAG;
 428                 item->expected = be32_to_cpu(cqe->expected_trans_sig) & 0xffff;
 429                 item->actual = be32_to_cpu(cqe->actual_trans_sig) & 0xffff;
 430         } else {
 431                 pr_err("Got signature completion error with bad syndrome %04x\n",
 432                        syndrome);
 433         }
 434
 435         item->sig_err_offset = be64_to_cpu(cqe->err_offset);
 436         item->key = be32_to_cpu(cqe->mkey);
 437 }
 438
 439 static void sw_send_comp(struct mlx5_ib_qp *qp, int num_entries,
 440                          struct ib_wc *wc, int *npolled)
 441 {
 442         struct mlx5_ib_wq *wq;
 443         unsigned int cur;
 444         unsigned int idx;
 445         int np;
 446         int i;
 447
 448         wq = &qp->sq;
 449         cur = wq->head - wq->tail;
 450         np = *npolled;
 451
 452         if (cur == 0)
 453                 return;
 454
 455         for (i = 0;  i < cur && np < num_entries; i++) {
 456                 idx = wq->last_poll & (wq->wqe_cnt - 1);
 457                 wc->wr_id = wq->wrid[idx];
 458                 wc->status = IB_WC_WR_FLUSH_ERR;
 459                 wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
 460                 wq->tail++;
 461                 np++;
 462                 wc->qp = &qp->ibqp;
 463                 wc++;
 464                 wq->last_poll = wq->w_list[idx].next;
 465         }
 466         *npolled = np;
 467 }
 468
 469 static void sw_recv_comp(struct mlx5_ib_qp *qp, int num_entries,
 470                          struct ib_wc *wc, int *npolled)
 471 {
 472         struct mlx5_ib_wq *wq;
 473         unsigned int cur;
 474         int np;
 475         int i;
 476
 477         wq = &qp->rq;
 478         cur = wq->head - wq->tail;
 479         np = *npolled;
 480
 481         if (cur == 0)
 482                 return;
 483
 484         for (i = 0;  i < cur && np < num_entries; i++) {
 485                 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
 486                 wc->status = IB_WC_WR_FLUSH_ERR;
 487                 wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
 488                 wq->tail++;
 489                 np++;
 490                 wc->qp = &qp->ibqp;
 491                 wc++;
 492         }
 493         *npolled = np;
 494 }
 495
 496 static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries,
 497                                  struct ib_wc *wc, int *npolled)
 498 {
 499         struct mlx5_ib_qp *qp;
 500
 501         *npolled = 0;
 502         /* Find uncompleted WQEs belonging to that cq and retrun mmics ones */
 503         list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) {
 504                 sw_send_comp(qp, num_entries, wc + *npolled, npolled);
 505                 if (*npolled >= num_entries)
 506                         return;
 507         }
 508
 509         list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) {
 510                 sw_recv_comp(qp, num_entries, wc + *npolled, npolled);
 511                 if (*npolled >= num_entries)
 512                         return;
 513         }
 514 }
 515
 516 static int mlx5_poll_one(struct mlx5_ib_cq *cq,
 517                          struct mlx5_ib_qp **cur_qp,
 518                          struct ib_wc *wc)
 519 {
 520         struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
 521         struct mlx5_err_cqe *err_cqe;
 522         struct mlx5_cqe64 *cqe64;
 523         struct mlx5_core_qp *mqp;
 524         struct mlx5_ib_wq *wq;
 525         struct mlx5_sig_err_cqe *sig_err_cqe;
 526         struct mlx5_core_mkey *mmkey;
 527         struct mlx5_ib_mr *mr;
 528         uint8_t opcode;
 529         uint32_t qpn;
 530         u16 wqe_ctr;
 531         void *cqe;
 532         int idx;
 533
 534 repoll:
 535         cqe = next_cqe_sw(cq);
 536         if (!cqe)
 537                 return -EAGAIN;
 538
 539         cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
 540
 541         ++cq->mcq.cons_index;
 542
 543         /* Make sure we read CQ entry contents after we've checked the
 544          * ownership bit.
 545          */
 546         rmb();
 547
 548         opcode = cqe64->op_own >> 4;
 549         if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) {
 550                 if (likely(cq->resize_buf)) {
 551                         free_cq_buf(dev, &cq->buf);
 552                         cq->buf = *cq->resize_buf;
 553                         kfree(cq->resize_buf);
 554                         cq->resize_buf = NULL;
 555                         goto repoll;
 556                 } else {
 557                         mlx5_ib_warn(dev, "unexpected resize cqe\n");
 558                 }
 559         }
 560
 561         qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff;
 562         if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) {
 563                 /* We do not have to take the QP table lock here,
 564                  * because CQs will be locked while QPs are removed
 565                  * from the table.
 566                  */
 567                 mqp = __mlx5_qp_lookup(dev->mdev, qpn);
 568                 *cur_qp = to_mibqp(mqp);
 569         }
 570
 571         wc->qp  = &(*cur_qp)->ibqp;
 572         switch (opcode) {
 573         case MLX5_CQE_REQ:
 574                 wq = &(*cur_qp)->sq;
 575                 wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
 576                 idx = wqe_ctr & (wq->wqe_cnt - 1);
 577                 handle_good_req(wc, cqe64, wq, idx);
 578                 handle_atomics(*cur_qp, cqe64, wq->last_poll, idx);
 579                 wc->wr_id = wq->wrid[idx];
 580                 wq->tail = wq->wqe_head[idx] + 1;
 581                 wc->status = IB_WC_SUCCESS;
 582                 break;
 583         case MLX5_CQE_RESP_WR_IMM:
 584         case MLX5_CQE_RESP_SEND:
 585         case MLX5_CQE_RESP_SEND_IMM:
 586         case MLX5_CQE_RESP_SEND_INV:
 587                 handle_responder(wc, cqe64, *cur_qp);
 588                 wc->status = IB_WC_SUCCESS;
 589                 break;
 590         case MLX5_CQE_RESIZE_CQ:
 591                 break;
 592         case MLX5_CQE_REQ_ERR:
 593         case MLX5_CQE_RESP_ERR:
 594                 err_cqe = (struct mlx5_err_cqe *)cqe64;
 595                 mlx5_handle_error_cqe(dev, err_cqe, wc);
 596                 mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
 597                             opcode == MLX5_CQE_REQ_ERR ?
 598                             "Requestor" : "Responder", cq->mcq.cqn);
 599                 mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
 600                             err_cqe->syndrome, err_cqe->vendor_err_synd);
 601                 if (opcode == MLX5_CQE_REQ_ERR) {
 602                         wq = &(*cur_qp)->sq;
 603                         wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
 604                         idx = wqe_ctr & (wq->wqe_cnt - 1);
 605                         wc->wr_id = wq->wrid[idx];
 606                         wq->tail = wq->wqe_head[idx] + 1;
 607                 } else {
 608                         struct mlx5_ib_srq *srq;
 609
 610                         if ((*cur_qp)->ibqp.srq) {
 611                                 srq = to_msrq((*cur_qp)->ibqp.srq);
 612                                 wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
 613                                 wc->wr_id = srq->wrid[wqe_ctr];
 614                                 mlx5_ib_free_srq_wqe(srq, wqe_ctr);
 615                         } else {
 616                                 wq = &(*cur_qp)->rq;
 617                                 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
 618                                 ++wq->tail;
 619                         }
 620                 }
 621                 break;
 622         case MLX5_CQE_SIG_ERR:
 623                 sig_err_cqe = (struct mlx5_sig_err_cqe *)cqe64;
 624
 625                 read_lock(&dev->mdev->priv.mkey_table.lock);
 626                 mmkey = __mlx5_mr_lookup(dev->mdev,
 627                                          mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey)));
 628                 mr = to_mibmr(mmkey);
 629                 get_sig_err_item(sig_err_cqe, &mr->sig->err_item);
 630                 mr->sig->sig_err_exists = true;
 631                 mr->sig->sigerr_count++;
 632
 633                 mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n",
 634                              cq->mcq.cqn, mr->sig->err_item.key,
 635                              mr->sig->err_item.err_type,
 636                              mr->sig->err_item.sig_err_offset,
 637                              mr->sig->err_item.expected,
 638                              mr->sig->err_item.actual);
 639
 640                 read_unlock(&dev->mdev->priv.mkey_table.lock);
 641                 goto repoll;
 642         }
 643
 644         return 0;
 645 }
 646
 647 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries,
 648                         struct ib_wc *wc, bool is_fatal_err)
 649 {
 650         struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
 651         struct mlx5_ib_wc *soft_wc, *next;
 652         int npolled = 0;
 653
 654         list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
 655                 if (npolled >= num_entries)
 656                         break;
 657
 658                 mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n",
 659                             cq->mcq.cqn);
 660
 661                 if (unlikely(is_fatal_err)) {
 662                         soft_wc->wc.status = IB_WC_WR_FLUSH_ERR;
 663                         soft_wc->wc.vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
 664                 }
 665                 wc[npolled++] = soft_wc->wc;
 666                 list_del(&soft_wc->list);
 667                 kfree(soft_wc);
 668         }
 669
 670         return npolled;
 671 }
 672
 673 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
 674 {
 675         struct mlx5_ib_cq *cq = to_mcq(ibcq);
 676         struct mlx5_ib_qp *cur_qp = NULL;
 677         struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
 678         struct mlx5_core_dev *mdev = dev->mdev;
 679         unsigned long flags;
 680         int soft_polled = 0;
 681         int npolled;
 682
 683         spin_lock_irqsave(&cq->lock, flags);
 684         if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
 685                 /* make sure no soft wqe's are waiting */
 686                 if (unlikely(!list_empty(&cq->wc_list)))
 687                         soft_polled = poll_soft_wc(cq, num_entries, wc, true);
 688
 689                 mlx5_ib_poll_sw_comp(cq, num_entries - soft_polled,
 690                                      wc + soft_polled, &npolled);
 691                 goto out;
 692         }
 693
 694         if (unlikely(!list_empty(&cq->wc_list)))
 695                 soft_polled = poll_soft_wc(cq, num_entries, wc, false);
 696
 697         for (npolled = 0; npolled < num_entries - soft_polled; npolled++) {
 698                 if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled))
 699                         break;
 700         }
 701
 702         if (npolled)
 703                 mlx5_cq_set_ci(&cq->mcq);
 704 out:
 705         spin_unlock_irqrestore(&cq->lock, flags);
 706
 707         return soft_polled + npolled;
 708 }
 709
 710 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
 711 {
 712         struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
 713         struct mlx5_ib_cq *cq = to_mcq(ibcq);
 714         void __iomem *uar_page = mdev->priv.uuari.uars[0].map;
 715         unsigned long irq_flags;
 716         int ret = 0;
 717
 718         spin_lock_irqsave(&cq->lock, irq_flags);
 719         if (cq->notify_flags != IB_CQ_NEXT_COMP)
 720                 cq->notify_flags = flags & IB_CQ_SOLICITED_MASK;
 721
 722         if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list))
 723                 ret = 1;
 724         spin_unlock_irqrestore(&cq->lock, irq_flags);
 725
 726         mlx5_cq_arm(&cq->mcq,
 727                     (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
 728                     MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
 729                     uar_page,
 730                     MLX5_GET_DOORBELL_LOCK(&mdev->priv.cq_uar_lock),
 731                     to_mcq(ibcq)->mcq.cons_index);
 732
 733         return ret;
 734 }
 735
 736 static int alloc_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf,
 737                         int nent, int cqe_size)
 738 {
 739         int err;
 740
 741         err = mlx5_buf_alloc(dev->mdev, nent * cqe_size, &buf->buf);
 742         if (err)
 743                 return err;
 744
 745         buf->cqe_size = cqe_size;
 746         buf->nent = nent;
 747
 748         return 0;
 749 }
 750
 751 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
 752                           struct ib_ucontext *context, struct mlx5_ib_cq *cq,
 753                           int entries, u32 **cqb,
 754                           int *cqe_size, int *index, int *inlen)
 755 {
 756         struct mlx5_ib_create_cq ucmd;
 757         size_t ucmdlen;
 758         int page_shift;
 759         __be64 *pas;
 760         int npages;
 761         int ncont;
 762         void *cqc;
 763         int err;
 764
 765         ucmdlen =
 766                 (udata->inlen - sizeof(struct ib_uverbs_cmd_hdr) <
 767                  sizeof(ucmd)) ? (sizeof(ucmd) -
 768                                   sizeof(ucmd.reserved)) : sizeof(ucmd);
 769
 770         if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
 771                 return -EFAULT;
 772
 773         if (ucmdlen == sizeof(ucmd) &&
 774             ucmd.reserved != 0)
 775                 return -EINVAL;
 776
 777         if (ucmd.cqe_size != 64 && ucmd.cqe_size != 128)
 778                 return -EINVAL;
 779
 780         *cqe_size = ucmd.cqe_size;
 781
 782         cq->buf.umem = ib_umem_get(context, ucmd.buf_addr,
 783                                    entries * ucmd.cqe_size,
 784                                    IB_ACCESS_LOCAL_WRITE, 1);
 785         if (IS_ERR(cq->buf.umem)) {
 786                 err = PTR_ERR(cq->buf.umem);
 787                 return err;
 788         }
 789
 790         err = mlx5_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
 791                                   &cq->db);
 792         if (err)
 793                 goto err_umem;
 794
 795         mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, &npages, &page_shift,
 796                            &ncont, NULL);
 797         mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n",
 798                     ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont);
 799
 800         *inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
 801                  MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * ncont;
 802         *cqb = mlx5_vzalloc(*inlen);
 803         if (!*cqb) {
 804                 err = -ENOMEM;
 805                 goto err_db;
 806         }
 807
 808         pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
 809         mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, pas, 0);
 810
 811         cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
 812         MLX5_SET(cqc, cqc, log_page_size,
 813                  page_shift - MLX5_ADAPTER_PAGE_SHIFT);
 814
 815         *index = to_mucontext(context)->uuari.uars[0].index;
 816
 817         return 0;
 818
 819 err_db:
 820         mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
 821
 822 err_umem:
 823         ib_umem_release(cq->buf.umem);
 824         return err;
 825 }
 826
 827 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_ucontext *context)
 828 {
 829         mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
 830         ib_umem_release(cq->buf.umem);
 831 }
 832
 833 static void init_cq_buf(struct mlx5_ib_cq *cq, struct mlx5_ib_cq_buf *buf)
 834 {
 835         int i;
 836         void *cqe;
 837         struct mlx5_cqe64 *cqe64;
 838
 839         for (i = 0; i < buf->nent; i++) {
 840                 cqe = get_cqe_from_buf(buf, i, buf->cqe_size);
 841                 cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
 842                 cqe64->op_own = MLX5_CQE_INVALID << 4;
 843         }
 844 }
 845
 846 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
 847                             int entries, int cqe_size,
 848                             u32 **cqb, int *index, int *inlen)
 849 {
 850         __be64 *pas;
 851         void *cqc;
 852         int err;
 853
 854         err = mlx5_db_alloc(dev->mdev, &cq->db);
 855         if (err)
 856                 return err;
 857
 858         cq->mcq.set_ci_db  = cq->db.db;
 859         cq->mcq.arm_db     = cq->db.db + 1;
 860         cq->mcq.cqe_sz = cqe_size;
 861
 862         err = alloc_cq_buf(dev, &cq->buf, entries, cqe_size);
 863         if (err)
 864                 goto err_db;
 865
 866         init_cq_buf(cq, &cq->buf);
 867
 868         *inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
 869                  MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * cq->buf.buf.npages;
 870         *cqb = mlx5_vzalloc(*inlen);
 871         if (!*cqb) {
 872                 err = -ENOMEM;
 873                 goto err_buf;
 874         }
 875
 876         pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
 877         mlx5_fill_page_array(&cq->buf.buf, pas);
 878
 879         cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
 880         MLX5_SET(cqc, cqc, log_page_size,
 881                  cq->buf.buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
 882
 883         *index = dev->mdev->priv.uuari.uars[0].index;
 884
 885         return 0;
 886
 887 err_buf:
 888         free_cq_buf(dev, &cq->buf);
 889
 890 err_db:
 891         mlx5_db_free(dev->mdev, &cq->db);
 892         return err;
 893 }
 894
 895 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
 896 {
 897         free_cq_buf(dev, &cq->buf);
 898         mlx5_db_free(dev->mdev, &cq->db);
 899 }
 900
 901 static void notify_soft_wc_handler(struct work_struct *work)
 902 {
 903         struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq,
 904                                              notify_work);
 905
 906         cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
 907 }
 908
 909 struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev,
 910                                 const struct ib_cq_init_attr *attr,
 911                                 struct ib_ucontext *context,
 912                                 struct ib_udata *udata)
 913 {
 914         int entries = attr->cqe;
 915         int vector = attr->comp_vector;
 916         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 917         struct mlx5_ib_cq *cq;
 918         int uninitialized_var(index);
 919         int uninitialized_var(inlen);
 920         u32 *cqb = NULL;
 921         void *cqc;
 922         int cqe_size;
 923         unsigned int irqn;
 924         int eqn;
 925         int err;
 926
 927         if (entries < 0 ||
 928             (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))))
 929                 return ERR_PTR(-EINVAL);
 930
 931         if (check_cq_create_flags(attr->flags))
 932                 return ERR_PTR(-EOPNOTSUPP);
 933
 934         entries = roundup_pow_of_two(entries + 1);
 935         if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)))
 936                 return ERR_PTR(-EINVAL);
 937
 938         cq = kzalloc(sizeof(*cq), GFP_KERNEL);
 939         if (!cq)
 940                 return ERR_PTR(-ENOMEM);
 941
 942         cq->ibcq.cqe = entries - 1;
 943         mutex_init(&cq->resize_mutex);
 944         spin_lock_init(&cq->lock);
 945         cq->resize_buf = NULL;
 946         cq->resize_umem = NULL;
 947         cq->create_flags = attr->flags;
 948         INIT_LIST_HEAD(&cq->list_send_qp);
 949         INIT_LIST_HEAD(&cq->list_recv_qp);
 950
 951         if (context) {
 952                 err = create_cq_user(dev, udata, context, cq, entries,
 953                                      &cqb, &cqe_size, &index, &inlen);
 954                 if (err)
 955                         goto err_create;
 956         } else {
 957                 cqe_size = cache_line_size() == 128 ? 128 : 64;
 958                 err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
 959                                        &index, &inlen);
 960                 if (err)
 961                         goto err_create;
 962
 963                 INIT_WORK(&cq->notify_work, notify_soft_wc_handler);
 964         }
 965
 966         err = mlx5_vector2eqn(dev->mdev, vector, &eqn, &irqn);
 967         if (err)
 968                 goto err_cqb;
 969
 970         cq->cqe_size = cqe_size;
 971
 972         cqc = MLX5_ADDR_OF(create_cq_in, cqb, cq_context);
 973         MLX5_SET(cqc, cqc, cqe_sz, cqe_sz_to_mlx_sz(cqe_size));
 974         MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
 975         MLX5_SET(cqc, cqc, uar_page, index);
 976         MLX5_SET(cqc, cqc, c_eqn, eqn);
 977         MLX5_SET64(cqc, cqc, dbr_addr, cq->db.dma);
 978         if (cq->create_flags & IB_CQ_FLAGS_IGNORE_OVERRUN)
 979                 MLX5_SET(cqc, cqc, oi, 1);
 980
 981         err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen);
 982         if (err)
 983                 goto err_cqb;
 984
 985         mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
 986         cq->mcq.irqn = irqn;
 987         if (context)
 988                 cq->mcq.tasklet_ctx.comp = mlx5_ib_cq_comp;
 989         else
 990                 cq->mcq.comp  = mlx5_ib_cq_comp;
 991         cq->mcq.event = mlx5_ib_cq_event;
 992
 993         INIT_LIST_HEAD(&cq->wc_list);
 994
 995         if (context)
 996                 if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
 997                         err = -EFAULT;
 998                         goto err_cmd;
 999                 }
1000
1001
1002         kvfree(cqb);
1003         return &cq->ibcq;
1004
1005 err_cmd:
1006         mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
1007
1008 err_cqb:
1009         kvfree(cqb);
1010         if (context)
1011                 destroy_cq_user(cq, context);
1012         else
1013                 destroy_cq_kernel(dev, cq);
1014
1015 err_create:
1016         kfree(cq);
1017
1018         return ERR_PTR(err);
1019 }
1020
1021
1022 int mlx5_ib_destroy_cq(struct ib_cq *cq)
1023 {
1024         struct mlx5_ib_dev *dev = to_mdev(cq->device);
1025         struct mlx5_ib_cq *mcq = to_mcq(cq);
1026         struct ib_ucontext *context = NULL;
1027
1028         if (cq->uobject)
1029                 context = cq->uobject->context;
1030
1031         mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
1032         if (context)
1033                 destroy_cq_user(mcq, context);
1034         else
1035                 destroy_cq_kernel(dev, mcq);
1036
1037         kfree(mcq);
1038
1039         return 0;
1040 }
1041
1042 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
1043 {
1044         return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
1045 }
1046
1047 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
1048 {
1049         struct mlx5_cqe64 *cqe64, *dest64;
1050         void *cqe, *dest;
1051         u32 prod_index;
1052         int nfreed = 0;
1053         u8 owner_bit;
1054
1055         if (!cq)
1056                 return;
1057
1058         /* First we need to find the current producer index, so we
1059          * know where to start cleaning from.  It doesn't matter if HW
1060          * adds new entries after this loop -- the QP we're worried
1061          * about is already in RESET, so the new entries won't come
1062          * from our QP and therefore don't need to be checked.
1063          */
1064         for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
1065                 if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
1066                         break;
1067
1068         /* Now sweep backwards through the CQ, removing CQ entries
1069          * that match our QP by copying older entries on top of them.
1070          */
1071         while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
1072                 cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
1073                 cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
1074                 if (is_equal_rsn(cqe64, rsn)) {
1075                         if (srq && (ntohl(cqe64->srqn) & 0xffffff))
1076                                 mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
1077                         ++nfreed;
1078                 } else if (nfreed) {
1079                         dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
1080                         dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
1081                         owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
1082                         memcpy(dest, cqe, cq->mcq.cqe_sz);
1083                         dest64->op_own = owner_bit |
1084                                 (dest64->op_own & ~MLX5_CQE_OWNER_MASK);
1085                 }
1086         }
1087
1088         if (nfreed) {
1089                 cq->mcq.cons_index += nfreed;
1090                 /* Make sure update of buffer contents is done before
1091                  * updating consumer index.
1092                  */
1093                 wmb();
1094                 mlx5_cq_set_ci(&cq->mcq);
1095         }
1096 }
1097
1098 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
1099 {
1100         if (!cq)
1101                 return;
1102
1103         spin_lock_irq(&cq->lock);
1104         __mlx5_ib_cq_clean(cq, qpn, srq);
1105         spin_unlock_irq(&cq->lock);
1106 }
1107
1108 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1109 {
1110         struct mlx5_ib_dev *dev = to_mdev(cq->device);
1111         struct mlx5_ib_cq *mcq = to_mcq(cq);
1112         int err;
1113
1114         if (!MLX5_CAP_GEN(dev->mdev, cq_moderation))
1115                 return -ENOSYS;
1116
1117         err = mlx5_core_modify_cq_moderation(dev->mdev, &mcq->mcq,
1118                                              cq_period, cq_count);
1119         if (err)
1120                 mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
1121
1122         return err;
1123 }
1124
1125 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1126                        int entries, struct ib_udata *udata, int *npas,
1127                        int *page_shift, int *cqe_size)
1128 {
1129         struct mlx5_ib_resize_cq ucmd;
1130         struct ib_umem *umem;
1131         int err;
1132         int npages;
1133         struct ib_ucontext *context = cq->buf.umem->context;
1134
1135         err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
1136         if (err)
1137                 return err;
1138
1139         if (ucmd.reserved0 || ucmd.reserved1)
1140                 return -EINVAL;
1141
1142         /* check multiplication overflow */
1143         if (ucmd.cqe_size && SIZE_MAX / ucmd.cqe_size <= entries - 1)
1144                 return -EINVAL;
1145
1146         umem = ib_umem_get(context, ucmd.buf_addr,
1147                            (size_t)ucmd.cqe_size * entries,
1148                            IB_ACCESS_LOCAL_WRITE, 1);
1149         if (IS_ERR(umem)) {
1150                 err = PTR_ERR(umem);
1151                 return err;
1152         }
1153
1154         mlx5_ib_cont_pages(umem, ucmd.buf_addr, &npages, page_shift,
1155                            npas, NULL);
1156
1157         cq->resize_umem = umem;
1158         *cqe_size = ucmd.cqe_size;
1159
1160         return 0;
1161 }
1162
1163 static void un_resize_user(struct mlx5_ib_cq *cq)
1164 {
1165         ib_umem_release(cq->resize_umem);
1166 }
1167
1168 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1169                          int entries, int cqe_size)
1170 {
1171         int err;
1172
1173         cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
1174         if (!cq->resize_buf)
1175                 return -ENOMEM;
1176
1177         err = alloc_cq_buf(dev, cq->resize_buf, entries, cqe_size);
1178         if (err)
1179                 goto ex;
1180
1181         init_cq_buf(cq, cq->resize_buf);
1182
1183         return 0;
1184
1185 ex:
1186         kfree(cq->resize_buf);
1187         return err;
1188 }
1189
1190 static void un_resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
1191 {
1192         free_cq_buf(dev, cq->resize_buf);
1193         cq->resize_buf = NULL;
1194 }
1195
1196 static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1197 {
1198         struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1199         struct mlx5_cqe64 *scqe64;
1200         struct mlx5_cqe64 *dcqe64;
1201         void *start_cqe;
1202         void *scqe;
1203         void *dcqe;
1204         int ssize;
1205         int dsize;
1206         int i;
1207         u8 sw_own;
1208
1209         ssize = cq->buf.cqe_size;
1210         dsize = cq->resize_buf->cqe_size;
1211         if (ssize != dsize) {
1212                 mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1213                 return -EINVAL;
1214         }
1215
1216         i = cq->mcq.cons_index;
1217         scqe = get_sw_cqe(cq, i);
1218         scqe64 = ssize == 64 ? scqe : scqe + 64;
1219         start_cqe = scqe;
1220         if (!scqe) {
1221                 mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1222                 return -EINVAL;
1223         }
1224
1225         while ((scqe64->op_own >> 4) != MLX5_CQE_RESIZE_CQ) {
1226                 dcqe = get_cqe_from_buf(cq->resize_buf,
1227                                         (i + 1) & (cq->resize_buf->nent),
1228                                         dsize);
1229                 dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1230                 sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1231                 memcpy(dcqe, scqe, dsize);
1232                 dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1233
1234                 ++i;
1235                 scqe = get_sw_cqe(cq, i);
1236                 scqe64 = ssize == 64 ? scqe : scqe + 64;
1237                 if (!scqe) {
1238                         mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1239                         return -EINVAL;
1240                 }
1241
1242                 if (scqe == start_cqe) {
1243                         pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1244                                 cq->mcq.cqn);
1245                         return -ENOMEM;
1246                 }
1247         }
1248         ++cq->mcq.cons_index;
1249         return 0;
1250 }
1251
1252 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1253 {
1254         struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1255         struct mlx5_ib_cq *cq = to_mcq(ibcq);
1256         void *cqc;
1257         u32 *in;
1258         int err;
1259         int npas;
1260         __be64 *pas;
1261         int page_shift;
1262         int inlen;
1263         int uninitialized_var(cqe_size);
1264         unsigned long flags;
1265
1266         if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) {
1267                 pr_info("Firmware does not support resize CQ\n");
1268                 return -ENOSYS;
1269         }
1270
1271         if (entries < 1 ||
1272             entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) {
1273                 mlx5_ib_warn(dev, "wrong entries number %d, max %d\n",
1274                              entries,
1275                              1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz));
1276                 return -EINVAL;
1277         }
1278
1279         entries = roundup_pow_of_two(entries + 1);
1280         if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1)
1281                 return -EINVAL;
1282
1283         if (entries == ibcq->cqe + 1)
1284                 return 0;
1285
1286         mutex_lock(&cq->resize_mutex);
1287         if (udata) {
1288                 err = resize_user(dev, cq, entries, udata, &npas, &page_shift,
1289                                   &cqe_size);
1290         } else {
1291                 cqe_size = 64;
1292                 err = resize_kernel(dev, cq, entries, cqe_size);
1293                 if (!err) {
1294                         npas = cq->resize_buf->buf.npages;
1295                         page_shift = cq->resize_buf->buf.page_shift;
1296                 }
1297         }
1298
1299         if (err)
1300                 goto ex;
1301
1302         inlen = MLX5_ST_SZ_BYTES(modify_cq_in) +
1303                 MLX5_FLD_SZ_BYTES(modify_cq_in, pas[0]) * npas;
1304
1305         in = mlx5_vzalloc(inlen);
1306         if (!in) {
1307                 err = -ENOMEM;
1308                 goto ex_resize;
1309         }
1310
1311         pas = (__be64 *)MLX5_ADDR_OF(modify_cq_in, in, pas);
1312         if (udata)
1313                 mlx5_ib_populate_pas(dev, cq->resize_umem, page_shift,
1314                                      pas, 0);
1315         else
1316                 mlx5_fill_page_array(&cq->resize_buf->buf, pas);
1317
1318         MLX5_SET(modify_cq_in, in,
1319                  modify_field_select_resize_field_select.resize_field_select.resize_field_select,
1320                  MLX5_MODIFY_CQ_MASK_LOG_SIZE  |
1321                  MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1322                  MLX5_MODIFY_CQ_MASK_PG_SIZE);
1323
1324         cqc = MLX5_ADDR_OF(modify_cq_in, in, cq_context);
1325
1326         MLX5_SET(cqc, cqc, log_page_size,
1327                  page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1328         MLX5_SET(cqc, cqc, cqe_sz, cqe_sz_to_mlx_sz(cqe_size));
1329         MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1330
1331         MLX5_SET(modify_cq_in, in, op_mod, MLX5_CQ_OPMOD_RESIZE);
1332         MLX5_SET(modify_cq_in, in, cqn, cq->mcq.cqn);
1333
1334         err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen);
1335         if (err)
1336                 goto ex_alloc;
1337
1338         if (udata) {
1339                 cq->ibcq.cqe = entries - 1;
1340                 ib_umem_release(cq->buf.umem);
1341                 cq->buf.umem = cq->resize_umem;
1342                 cq->resize_umem = NULL;
1343         } else {
1344                 struct mlx5_ib_cq_buf tbuf;
1345                 int resized = 0;
1346
1347                 spin_lock_irqsave(&cq->lock, flags);
1348                 if (cq->resize_buf) {
1349                         err = copy_resize_cqes(cq);
1350                         if (!err) {
1351                                 tbuf = cq->buf;
1352                                 cq->buf = *cq->resize_buf;
1353                                 kfree(cq->resize_buf);
1354                                 cq->resize_buf = NULL;
1355                                 resized = 1;
1356                         }
1357                 }
1358                 cq->ibcq.cqe = entries - 1;
1359                 spin_unlock_irqrestore(&cq->lock, flags);
1360                 if (resized)
1361                         free_cq_buf(dev, &tbuf);
1362         }
1363         mutex_unlock(&cq->resize_mutex);
1364
1365         kvfree(in);
1366         return 0;
1367
1368 ex_alloc:
1369         kvfree(in);
1370
1371 ex_resize:
1372         if (udata)
1373                 un_resize_user(cq);
1374         else
1375                 un_resize_kernel(dev, cq);
1376 ex:
1377         mutex_unlock(&cq->resize_mutex);
1378         return err;
1379 }
1380
1381 int mlx5_ib_get_cqe_size(struct mlx5_ib_dev *dev, struct ib_cq *ibcq)
1382 {
1383         struct mlx5_ib_cq *cq;
1384
1385         if (!ibcq)
1386                 return 128;
1387
1388         cq = to_mcq(ibcq);
1389         return cq->cqe_size;
1390 }
1391
1392 /* Called from atomic context */
1393 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc)
1394 {
1395         struct mlx5_ib_wc *soft_wc;
1396         struct mlx5_ib_cq *cq = to_mcq(ibcq);
1397         unsigned long flags;
1398
1399         soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC);
1400         if (!soft_wc)
1401                 return -ENOMEM;
1402
1403         soft_wc->wc = *wc;
1404         spin_lock_irqsave(&cq->lock, flags);
1405         list_add_tail(&soft_wc->list, &cq->wc_list);
1406         if (cq->notify_flags == IB_CQ_NEXT_COMP ||
1407             wc->status != IB_WC_SUCCESS) {
1408                 cq->notify_flags = 0;
1409                 schedule_work(&cq->notify_work);
1410         }
1411         spin_unlock_irqrestore(&cq->lock, flags);
1412
1413         return 0;
1414 }