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