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                 *cur_qp = to_mibqp(mqp);
 557         }
 558
 559         wc->qp  = &(*cur_qp)->ibqp;
 560         switch (opcode) {
 561         case MLX5_CQE_REQ:
 562                 wq = &(*cur_qp)->sq;
 563                 wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
 564                 idx = wqe_ctr & (wq->wqe_cnt - 1);
 565                 handle_good_req(wc, cqe64, wq, idx);
 566                 handle_atomics(*cur_qp, cqe64, wq->last_poll, idx);
 567                 wc->wr_id = wq->wrid[idx];
 568                 wq->tail = wq->wqe_head[idx] + 1;
 569                 wc->status = IB_WC_SUCCESS;
 570                 break;
 571         case MLX5_CQE_RESP_WR_IMM:
 572         case MLX5_CQE_RESP_SEND:
 573         case MLX5_CQE_RESP_SEND_IMM:
 574         case MLX5_CQE_RESP_SEND_INV:
 575                 handle_responder(wc, cqe64, *cur_qp);
 576                 wc->status = IB_WC_SUCCESS;
 577                 break;
 578         case MLX5_CQE_RESIZE_CQ:
 579                 break;
 580         case MLX5_CQE_REQ_ERR:
 581         case MLX5_CQE_RESP_ERR:
 582                 err_cqe = (struct mlx5_err_cqe *)cqe64;
 583                 mlx5_handle_error_cqe(dev, err_cqe, wc);
 584                 mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
 585                             opcode == MLX5_CQE_REQ_ERR ?
 586                             "Requestor" : "Responder", cq->mcq.cqn);
 587                 mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
 588                             err_cqe->syndrome, err_cqe->vendor_err_synd);
 589                 if (opcode == MLX5_CQE_REQ_ERR) {
 590                         wq = &(*cur_qp)->sq;
 591                         wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
 592                         idx = wqe_ctr & (wq->wqe_cnt - 1);
 593                         wc->wr_id = wq->wrid[idx];
 594                         wq->tail = wq->wqe_head[idx] + 1;
 595                 } else {
 596                         struct mlx5_ib_srq *srq;
 597
 598                         if ((*cur_qp)->ibqp.srq) {
 599                                 srq = to_msrq((*cur_qp)->ibqp.srq);
 600                                 wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
 601                                 wc->wr_id = srq->wrid[wqe_ctr];
 602                                 mlx5_ib_free_srq_wqe(srq, wqe_ctr);
 603                         } else {
 604                                 wq = &(*cur_qp)->rq;
 605                                 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
 606                                 ++wq->tail;
 607                         }
 608                 }
 609                 break;
 610         case MLX5_CQE_SIG_ERR:
 611                 sig_err_cqe = (struct mlx5_sig_err_cqe *)cqe64;
 612
 613                 read_lock(&dev->mdev->priv.mkey_table.lock);
 614                 mmkey = __mlx5_mr_lookup(dev->mdev,
 615                                          mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey)));
 616                 mr = to_mibmr(mmkey);
 617                 get_sig_err_item(sig_err_cqe, &mr->sig->err_item);
 618                 mr->sig->sig_err_exists = true;
 619                 mr->sig->sigerr_count++;
 620
 621                 mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n",
 622                              cq->mcq.cqn, mr->sig->err_item.key,
 623                              mr->sig->err_item.err_type,
 624                              mr->sig->err_item.sig_err_offset,
 625                              mr->sig->err_item.expected,
 626                              mr->sig->err_item.actual);
 627
 628                 read_unlock(&dev->mdev->priv.mkey_table.lock);
 629                 goto repoll;
 630         }
 631
 632         return 0;
 633 }
 634
 635 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries,
 636                         struct ib_wc *wc)
 637 {
 638         struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
 639         struct mlx5_ib_wc *soft_wc, *next;
 640         int npolled = 0;
 641
 642         list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
 643                 if (npolled >= num_entries)
 644                         break;
 645
 646                 mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n",
 647                             cq->mcq.cqn);
 648
 649                 wc[npolled++] = soft_wc->wc;
 650                 list_del(&soft_wc->list);
 651                 kfree(soft_wc);
 652         }
 653
 654         return npolled;
 655 }
 656
 657 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
 658 {
 659         struct mlx5_ib_cq *cq = to_mcq(ibcq);
 660         struct mlx5_ib_qp *cur_qp = NULL;
 661         struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
 662         struct mlx5_core_dev *mdev = dev->mdev;
 663         unsigned long flags;
 664         int soft_polled = 0;
 665         int npolled;
 666
 667         spin_lock_irqsave(&cq->lock, flags);
 668         if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
 669                 mlx5_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
 670                 goto out;
 671         }
 672
 673         if (unlikely(!list_empty(&cq->wc_list)))
 674                 soft_polled = poll_soft_wc(cq, num_entries, wc);
 675
 676         for (npolled = 0; npolled < num_entries - soft_polled; npolled++) {
 677                 if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled))
 678                         break;
 679         }
 680
 681         if (npolled)
 682                 mlx5_cq_set_ci(&cq->mcq);
 683 out:
 684         spin_unlock_irqrestore(&cq->lock, flags);
 685
 686         return soft_polled + npolled;
 687 }
 688
 689 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
 690 {
 691         struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
 692         struct mlx5_ib_cq *cq = to_mcq(ibcq);
 693         void __iomem *uar_page = mdev->priv.uuari.uars[0].map;
 694         unsigned long irq_flags;
 695         int ret = 0;
 696
 697         spin_lock_irqsave(&cq->lock, irq_flags);
 698         if (cq->notify_flags != IB_CQ_NEXT_COMP)
 699                 cq->notify_flags = flags & IB_CQ_SOLICITED_MASK;
 700
 701         if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list))
 702                 ret = 1;
 703         spin_unlock_irqrestore(&cq->lock, irq_flags);
 704
 705         mlx5_cq_arm(&cq->mcq,
 706                     (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
 707                     MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
 708                     uar_page,
 709                     MLX5_GET_DOORBELL_LOCK(&mdev->priv.cq_uar_lock),
 710                     to_mcq(ibcq)->mcq.cons_index);
 711
 712         return ret;
 713 }
 714
 715 static int alloc_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf,
 716                         int nent, int cqe_size)
 717 {
 718         int err;
 719
 720         err = mlx5_buf_alloc(dev->mdev, nent * cqe_size, &buf->buf);
 721         if (err)
 722                 return err;
 723
 724         buf->cqe_size = cqe_size;
 725         buf->nent = nent;
 726
 727         return 0;
 728 }
 729
 730 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
 731                           struct ib_ucontext *context, struct mlx5_ib_cq *cq,
 732                           int entries, struct mlx5_create_cq_mbox_in **cqb,
 733                           int *cqe_size, int *index, int *inlen)
 734 {
 735         struct mlx5_ib_create_cq ucmd;
 736         size_t ucmdlen;
 737         int page_shift;
 738         int npages;
 739         int ncont;
 740         int err;
 741
 742         ucmdlen =
 743                 (udata->inlen - sizeof(struct ib_uverbs_cmd_hdr) <
 744                  sizeof(ucmd)) ? (sizeof(ucmd) -
 745                                   sizeof(ucmd.reserved)) : sizeof(ucmd);
 746
 747         if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
 748                 return -EFAULT;
 749
 750         if (ucmdlen == sizeof(ucmd) &&
 751             ucmd.reserved != 0)
 752                 return -EINVAL;
 753
 754         if (ucmd.cqe_size != 64 && ucmd.cqe_size != 128)
 755                 return -EINVAL;
 756
 757         *cqe_size = ucmd.cqe_size;
 758
 759         cq->buf.umem = ib_umem_get(context, ucmd.buf_addr,
 760                                    entries * ucmd.cqe_size,
 761                                    IB_ACCESS_LOCAL_WRITE, 1);
 762         if (IS_ERR(cq->buf.umem)) {
 763                 err = PTR_ERR(cq->buf.umem);
 764                 return err;
 765         }
 766
 767         err = mlx5_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
 768                                   &cq->db);
 769         if (err)
 770                 goto err_umem;
 771
 772         mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, &npages, &page_shift,
 773                            &ncont, NULL);
 774         mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n",
 775                     ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont);
 776
 777         *inlen = sizeof(**cqb) + sizeof(*(*cqb)->pas) * ncont;
 778         *cqb = mlx5_vzalloc(*inlen);
 779         if (!*cqb) {
 780                 err = -ENOMEM;
 781                 goto err_db;
 782         }
 783         mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, (*cqb)->pas, 0);
 784         (*cqb)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
 785
 786         *index = to_mucontext(context)->uuari.uars[0].index;
 787
 788         return 0;
 789
 790 err_db:
 791         mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
 792
 793 err_umem:
 794         ib_umem_release(cq->buf.umem);
 795         return err;
 796 }
 797
 798 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_ucontext *context)
 799 {
 800         mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
 801         ib_umem_release(cq->buf.umem);
 802 }
 803
 804 static void init_cq_buf(struct mlx5_ib_cq *cq, struct mlx5_ib_cq_buf *buf)
 805 {
 806         int i;
 807         void *cqe;
 808         struct mlx5_cqe64 *cqe64;
 809
 810         for (i = 0; i < buf->nent; i++) {
 811                 cqe = get_cqe_from_buf(buf, i, buf->cqe_size);
 812                 cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
 813                 cqe64->op_own = MLX5_CQE_INVALID << 4;
 814         }
 815 }
 816
 817 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
 818                             int entries, int cqe_size,
 819                             struct mlx5_create_cq_mbox_in **cqb,
 820                             int *index, int *inlen)
 821 {
 822         int err;
 823
 824         err = mlx5_db_alloc(dev->mdev, &cq->db);
 825         if (err)
 826                 return err;
 827
 828         cq->mcq.set_ci_db  = cq->db.db;
 829         cq->mcq.arm_db     = cq->db.db + 1;
 830         cq->mcq.cqe_sz = cqe_size;
 831
 832         err = alloc_cq_buf(dev, &cq->buf, entries, cqe_size);
 833         if (err)
 834                 goto err_db;
 835
 836         init_cq_buf(cq, &cq->buf);
 837
 838         *inlen = sizeof(**cqb) + sizeof(*(*cqb)->pas) * cq->buf.buf.npages;
 839         *cqb = mlx5_vzalloc(*inlen);
 840         if (!*cqb) {
 841                 err = -ENOMEM;
 842                 goto err_buf;
 843         }
 844         mlx5_fill_page_array(&cq->buf.buf, (*cqb)->pas);
 845
 846         (*cqb)->ctx.log_pg_sz = cq->buf.buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT;
 847         *index = dev->mdev->priv.uuari.uars[0].index;
 848
 849         return 0;
 850
 851 err_buf:
 852         free_cq_buf(dev, &cq->buf);
 853
 854 err_db:
 855         mlx5_db_free(dev->mdev, &cq->db);
 856         return err;
 857 }
 858
 859 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
 860 {
 861         free_cq_buf(dev, &cq->buf);
 862         mlx5_db_free(dev->mdev, &cq->db);
 863 }
 864
 865 static void notify_soft_wc_handler(struct work_struct *work)
 866 {
 867         struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq,
 868                                              notify_work);
 869
 870         cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
 871 }
 872
 873 struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev,
 874                                 const struct ib_cq_init_attr *attr,
 875                                 struct ib_ucontext *context,
 876                                 struct ib_udata *udata)
 877 {
 878         int entries = attr->cqe;
 879         int vector = attr->comp_vector;
 880         struct mlx5_create_cq_mbox_in *cqb = NULL;
 881         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 882         struct mlx5_ib_cq *cq;
 883         int uninitialized_var(index);
 884         int uninitialized_var(inlen);
 885         int cqe_size;
 886         unsigned int irqn;
 887         int eqn;
 888         int err;
 889
 890         if (entries < 0 ||
 891             (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))))
 892                 return ERR_PTR(-EINVAL);
 893
 894         if (check_cq_create_flags(attr->flags))
 895                 return ERR_PTR(-EOPNOTSUPP);
 896
 897         entries = roundup_pow_of_two(entries + 1);
 898         if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)))
 899                 return ERR_PTR(-EINVAL);
 900
 901         cq = kzalloc(sizeof(*cq), GFP_KERNEL);
 902         if (!cq)
 903                 return ERR_PTR(-ENOMEM);
 904
 905         cq->ibcq.cqe = entries - 1;
 906         mutex_init(&cq->resize_mutex);
 907         spin_lock_init(&cq->lock);
 908         cq->resize_buf = NULL;
 909         cq->resize_umem = NULL;
 910         cq->create_flags = attr->flags;
 911         INIT_LIST_HEAD(&cq->list_send_qp);
 912         INIT_LIST_HEAD(&cq->list_recv_qp);
 913
 914         if (context) {
 915                 err = create_cq_user(dev, udata, context, cq, entries,
 916                                      &cqb, &cqe_size, &index, &inlen);
 917                 if (err)
 918                         goto err_create;
 919         } else {
 920                 /* for now choose 64 bytes till we have a proper interface */
 921                 cqe_size = 64;
 922                 err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
 923                                        &index, &inlen);
 924                 if (err)
 925                         goto err_create;
 926
 927                 INIT_WORK(&cq->notify_work, notify_soft_wc_handler);
 928         }
 929
 930         cq->cqe_size = cqe_size;
 931         cqb->ctx.cqe_sz_flags = cqe_sz_to_mlx_sz(cqe_size) << 5;
 932
 933         if (cq->create_flags & IB_CQ_FLAGS_IGNORE_OVERRUN)
 934                 cqb->ctx.cqe_sz_flags |= (1 << 1);
 935
 936         cqb->ctx.log_sz_usr_page = cpu_to_be32((ilog2(entries) << 24) | index);
 937         err = mlx5_vector2eqn(dev->mdev, vector, &eqn, &irqn);
 938         if (err)
 939                 goto err_cqb;
 940
 941         cqb->ctx.c_eqn = cpu_to_be16(eqn);
 942         cqb->ctx.db_record_addr = cpu_to_be64(cq->db.dma);
 943
 944         err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen);
 945         if (err)
 946                 goto err_cqb;
 947
 948         mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
 949         cq->mcq.irqn = irqn;
 950         if (context)
 951                 cq->mcq.tasklet_ctx.comp = mlx5_ib_cq_comp;
 952         else
 953                 cq->mcq.comp  = mlx5_ib_cq_comp;
 954         cq->mcq.event = mlx5_ib_cq_event;
 955
 956         INIT_LIST_HEAD(&cq->wc_list);
 957
 958         if (context)
 959                 if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
 960                         err = -EFAULT;
 961                         goto err_cmd;
 962                 }
 963
 964
 965         kvfree(cqb);
 966         return &cq->ibcq;
 967
 968 err_cmd:
 969         mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
 970
 971 err_cqb:
 972         kvfree(cqb);
 973         if (context)
 974                 destroy_cq_user(cq, context);
 975         else
 976                 destroy_cq_kernel(dev, cq);
 977
 978 err_create:
 979         kfree(cq);
 980
 981         return ERR_PTR(err);
 982 }
 983
 984
 985 int mlx5_ib_destroy_cq(struct ib_cq *cq)
 986 {
 987         struct mlx5_ib_dev *dev = to_mdev(cq->device);
 988         struct mlx5_ib_cq *mcq = to_mcq(cq);
 989         struct ib_ucontext *context = NULL;
 990
 991         if (cq->uobject)
 992                 context = cq->uobject->context;
 993
 994         mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
 995         if (context)
 996                 destroy_cq_user(mcq, context);
 997         else
 998                 destroy_cq_kernel(dev, mcq);
 999
1000         kfree(mcq);
1001
1002         return 0;
1003 }
1004
1005 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
1006 {
1007         return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
1008 }
1009
1010 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
1011 {
1012         struct mlx5_cqe64 *cqe64, *dest64;
1013         void *cqe, *dest;
1014         u32 prod_index;
1015         int nfreed = 0;
1016         u8 owner_bit;
1017
1018         if (!cq)
1019                 return;
1020
1021         /* First we need to find the current producer index, so we
1022          * know where to start cleaning from.  It doesn't matter if HW
1023          * adds new entries after this loop -- the QP we're worried
1024          * about is already in RESET, so the new entries won't come
1025          * from our QP and therefore don't need to be checked.
1026          */
1027         for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
1028                 if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
1029                         break;
1030
1031         /* Now sweep backwards through the CQ, removing CQ entries
1032          * that match our QP by copying older entries on top of them.
1033          */
1034         while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
1035                 cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
1036                 cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
1037                 if (is_equal_rsn(cqe64, rsn)) {
1038                         if (srq && (ntohl(cqe64->srqn) & 0xffffff))
1039                                 mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
1040                         ++nfreed;
1041                 } else if (nfreed) {
1042                         dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
1043                         dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
1044                         owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
1045                         memcpy(dest, cqe, cq->mcq.cqe_sz);
1046                         dest64->op_own = owner_bit |
1047                                 (dest64->op_own & ~MLX5_CQE_OWNER_MASK);
1048                 }
1049         }
1050
1051         if (nfreed) {
1052                 cq->mcq.cons_index += nfreed;
1053                 /* Make sure update of buffer contents is done before
1054                  * updating consumer index.
1055                  */
1056                 wmb();
1057                 mlx5_cq_set_ci(&cq->mcq);
1058         }
1059 }
1060
1061 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
1062 {
1063         if (!cq)
1064                 return;
1065
1066         spin_lock_irq(&cq->lock);
1067         __mlx5_ib_cq_clean(cq, qpn, srq);
1068         spin_unlock_irq(&cq->lock);
1069 }
1070
1071 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1072 {
1073         struct mlx5_modify_cq_mbox_in *in;
1074         struct mlx5_ib_dev *dev = to_mdev(cq->device);
1075         struct mlx5_ib_cq *mcq = to_mcq(cq);
1076         int err;
1077         u32 fsel;
1078
1079         if (!MLX5_CAP_GEN(dev->mdev, cq_moderation))
1080                 return -ENOSYS;
1081
1082         in = kzalloc(sizeof(*in), GFP_KERNEL);
1083         if (!in)
1084                 return -ENOMEM;
1085
1086         in->cqn = cpu_to_be32(mcq->mcq.cqn);
1087         fsel = (MLX5_CQ_MODIFY_PERIOD | MLX5_CQ_MODIFY_COUNT);
1088         in->ctx.cq_period = cpu_to_be16(cq_period);
1089         in->ctx.cq_max_count = cpu_to_be16(cq_count);
1090         in->field_select = cpu_to_be32(fsel);
1091         err = mlx5_core_modify_cq(dev->mdev, &mcq->mcq, in, sizeof(*in));
1092         kfree(in);
1093
1094         if (err)
1095                 mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
1096
1097         return err;
1098 }
1099
1100 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1101                        int entries, struct ib_udata *udata, int *npas,
1102                        int *page_shift, int *cqe_size)
1103 {
1104         struct mlx5_ib_resize_cq ucmd;
1105         struct ib_umem *umem;
1106         int err;
1107         int npages;
1108         struct ib_ucontext *context = cq->buf.umem->context;
1109
1110         err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
1111         if (err)
1112                 return err;
1113
1114         if (ucmd.reserved0 || ucmd.reserved1)
1115                 return -EINVAL;
1116
1117         umem = ib_umem_get(context, ucmd.buf_addr, entries * ucmd.cqe_size,
1118                            IB_ACCESS_LOCAL_WRITE, 1);
1119         if (IS_ERR(umem)) {
1120                 err = PTR_ERR(umem);
1121                 return err;
1122         }
1123
1124         mlx5_ib_cont_pages(umem, ucmd.buf_addr, &npages, page_shift,
1125                            npas, NULL);
1126
1127         cq->resize_umem = umem;
1128         *cqe_size = ucmd.cqe_size;
1129
1130         return 0;
1131 }
1132
1133 static void un_resize_user(struct mlx5_ib_cq *cq)
1134 {
1135         ib_umem_release(cq->resize_umem);
1136 }
1137
1138 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1139                          int entries, int cqe_size)
1140 {
1141         int err;
1142
1143         cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
1144         if (!cq->resize_buf)
1145                 return -ENOMEM;
1146
1147         err = alloc_cq_buf(dev, cq->resize_buf, entries, cqe_size);
1148         if (err)
1149                 goto ex;
1150
1151         init_cq_buf(cq, cq->resize_buf);
1152
1153         return 0;
1154
1155 ex:
1156         kfree(cq->resize_buf);
1157         return err;
1158 }
1159
1160 static void un_resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
1161 {
1162         free_cq_buf(dev, cq->resize_buf);
1163         cq->resize_buf = NULL;
1164 }
1165
1166 static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1167 {
1168         struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1169         struct mlx5_cqe64 *scqe64;
1170         struct mlx5_cqe64 *dcqe64;
1171         void *start_cqe;
1172         void *scqe;
1173         void *dcqe;
1174         int ssize;
1175         int dsize;
1176         int i;
1177         u8 sw_own;
1178
1179         ssize = cq->buf.cqe_size;
1180         dsize = cq->resize_buf->cqe_size;
1181         if (ssize != dsize) {
1182                 mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1183                 return -EINVAL;
1184         }
1185
1186         i = cq->mcq.cons_index;
1187         scqe = get_sw_cqe(cq, i);
1188         scqe64 = ssize == 64 ? scqe : scqe + 64;
1189         start_cqe = scqe;
1190         if (!scqe) {
1191                 mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1192                 return -EINVAL;
1193         }
1194
1195         while ((scqe64->op_own >> 4) != MLX5_CQE_RESIZE_CQ) {
1196                 dcqe = get_cqe_from_buf(cq->resize_buf,
1197                                         (i + 1) & (cq->resize_buf->nent),
1198                                         dsize);
1199                 dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1200                 sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1201                 memcpy(dcqe, scqe, dsize);
1202                 dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1203
1204                 ++i;
1205                 scqe = get_sw_cqe(cq, i);
1206                 scqe64 = ssize == 64 ? scqe : scqe + 64;
1207                 if (!scqe) {
1208                         mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1209                         return -EINVAL;
1210                 }
1211
1212                 if (scqe == start_cqe) {
1213                         pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1214                                 cq->mcq.cqn);
1215                         return -ENOMEM;
1216                 }
1217         }
1218         ++cq->mcq.cons_index;
1219         return 0;
1220 }
1221
1222 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1223 {
1224         struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1225         struct mlx5_ib_cq *cq = to_mcq(ibcq);
1226         struct mlx5_modify_cq_mbox_in *in;
1227         int err;
1228         int npas;
1229         int page_shift;
1230         int inlen;
1231         int uninitialized_var(cqe_size);
1232         unsigned long flags;
1233
1234         if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) {
1235                 pr_info("Firmware does not support resize CQ\n");
1236                 return -ENOSYS;
1237         }
1238
1239         if (entries < 1 ||
1240             entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) {
1241                 mlx5_ib_warn(dev, "wrong entries number %d, max %d\n",
1242                              entries,
1243                              1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz));
1244                 return -EINVAL;
1245         }
1246
1247         entries = roundup_pow_of_two(entries + 1);
1248         if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1)
1249                 return -EINVAL;
1250
1251         if (entries == ibcq->cqe + 1)
1252                 return 0;
1253
1254         mutex_lock(&cq->resize_mutex);
1255         if (udata) {
1256                 err = resize_user(dev, cq, entries, udata, &npas, &page_shift,
1257                                   &cqe_size);
1258         } else {
1259                 cqe_size = 64;
1260                 err = resize_kernel(dev, cq, entries, cqe_size);
1261                 if (!err) {
1262                         npas = cq->resize_buf->buf.npages;
1263                         page_shift = cq->resize_buf->buf.page_shift;
1264                 }
1265         }
1266
1267         if (err)
1268                 goto ex;
1269
1270         inlen = sizeof(*in) + npas * sizeof(in->pas[0]);
1271         in = mlx5_vzalloc(inlen);
1272         if (!in) {
1273                 err = -ENOMEM;
1274                 goto ex_resize;
1275         }
1276
1277         if (udata)
1278                 mlx5_ib_populate_pas(dev, cq->resize_umem, page_shift,
1279                                      in->pas, 0);
1280         else
1281                 mlx5_fill_page_array(&cq->resize_buf->buf, in->pas);
1282
1283         in->field_select = cpu_to_be32(MLX5_MODIFY_CQ_MASK_LOG_SIZE  |
1284                                        MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1285                                        MLX5_MODIFY_CQ_MASK_PG_SIZE);
1286         in->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
1287         in->ctx.cqe_sz_flags = cqe_sz_to_mlx_sz(cqe_size) << 5;
1288         in->ctx.page_offset = 0;
1289         in->ctx.log_sz_usr_page = cpu_to_be32(ilog2(entries) << 24);
1290         in->hdr.opmod = cpu_to_be16(MLX5_CQ_OPMOD_RESIZE);
1291         in->cqn = cpu_to_be32(cq->mcq.cqn);
1292
1293         err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen);
1294         if (err)
1295                 goto ex_alloc;
1296
1297         if (udata) {
1298                 cq->ibcq.cqe = entries - 1;
1299                 ib_umem_release(cq->buf.umem);
1300                 cq->buf.umem = cq->resize_umem;
1301                 cq->resize_umem = NULL;
1302         } else {
1303                 struct mlx5_ib_cq_buf tbuf;
1304                 int resized = 0;
1305
1306                 spin_lock_irqsave(&cq->lock, flags);
1307                 if (cq->resize_buf) {
1308                         err = copy_resize_cqes(cq);
1309                         if (!err) {
1310                                 tbuf = cq->buf;
1311                                 cq->buf = *cq->resize_buf;
1312                                 kfree(cq->resize_buf);
1313                                 cq->resize_buf = NULL;
1314                                 resized = 1;
1315                         }
1316                 }
1317                 cq->ibcq.cqe = entries - 1;
1318                 spin_unlock_irqrestore(&cq->lock, flags);
1319                 if (resized)
1320                         free_cq_buf(dev, &tbuf);
1321         }
1322         mutex_unlock(&cq->resize_mutex);
1323
1324         kvfree(in);
1325         return 0;
1326
1327 ex_alloc:
1328         kvfree(in);
1329
1330 ex_resize:
1331         if (udata)
1332                 un_resize_user(cq);
1333         else
1334                 un_resize_kernel(dev, cq);
1335 ex:
1336         mutex_unlock(&cq->resize_mutex);
1337         return err;
1338 }
1339
1340 int mlx5_ib_get_cqe_size(struct mlx5_ib_dev *dev, struct ib_cq *ibcq)
1341 {
1342         struct mlx5_ib_cq *cq;
1343
1344         if (!ibcq)
1345                 return 128;
1346
1347         cq = to_mcq(ibcq);
1348         return cq->cqe_size;
1349 }
1350
1351 /* Called from atomic context */
1352 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc)
1353 {
1354         struct mlx5_ib_wc *soft_wc;
1355         struct mlx5_ib_cq *cq = to_mcq(ibcq);
1356         unsigned long flags;
1357
1358         soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC);
1359         if (!soft_wc)
1360                 return -ENOMEM;
1361
1362         soft_wc->wc = *wc;
1363         spin_lock_irqsave(&cq->lock, flags);
1364         list_add_tail(&soft_wc->list, &cq->wc_list);
1365         if (cq->notify_flags == IB_CQ_NEXT_COMP ||
1366             wc->status != IB_WC_SUCCESS) {
1367                 cq->notify_flags = 0;
1368                 schedule_work(&cq->notify_work);
1369         }
1370         spin_unlock_irqrestore(&cq->lock, flags);
1371
1372         return 0;
1373 }