drivers/infiniband/hw/mlx5/qp.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/module.h>
  34 #include <rdma/ib_umem.h>
  35 #include "mlx5_ib.h"
  36 #include "user.h"
  37
  38 /* not supported currently */
  39 static int wq_signature;
  40
  41 enum {
  42         MLX5_IB_ACK_REQ_FREQ    = 8,
  43 };
  44
  45 enum {
  46         MLX5_IB_DEFAULT_SCHED_QUEUE     = 0x83,
  47         MLX5_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f,
  48         MLX5_IB_LINK_TYPE_IB            = 0,
  49         MLX5_IB_LINK_TYPE_ETH           = 1
  50 };
  51
  52 enum {
  53         MLX5_IB_SQ_STRIDE       = 6,
  54         MLX5_IB_CACHE_LINE_SIZE = 64,
  55 };
  56
  57 static const u32 mlx5_ib_opcode[] = {
  58         [IB_WR_SEND]                            = MLX5_OPCODE_SEND,
  59         [IB_WR_SEND_WITH_IMM]                   = MLX5_OPCODE_SEND_IMM,
  60         [IB_WR_RDMA_WRITE]                      = MLX5_OPCODE_RDMA_WRITE,
  61         [IB_WR_RDMA_WRITE_WITH_IMM]             = MLX5_OPCODE_RDMA_WRITE_IMM,
  62         [IB_WR_RDMA_READ]                       = MLX5_OPCODE_RDMA_READ,
  63         [IB_WR_ATOMIC_CMP_AND_SWP]              = MLX5_OPCODE_ATOMIC_CS,
  64         [IB_WR_ATOMIC_FETCH_AND_ADD]            = MLX5_OPCODE_ATOMIC_FA,
  65         [IB_WR_SEND_WITH_INV]                   = MLX5_OPCODE_SEND_INVAL,
  66         [IB_WR_LOCAL_INV]                       = MLX5_OPCODE_UMR,
  67         [IB_WR_FAST_REG_MR]                     = MLX5_OPCODE_UMR,
  68         [IB_WR_MASKED_ATOMIC_CMP_AND_SWP]       = MLX5_OPCODE_ATOMIC_MASKED_CS,
  69         [IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]     = MLX5_OPCODE_ATOMIC_MASKED_FA,
  70         [MLX5_IB_WR_UMR]                        = MLX5_OPCODE_UMR,
  71 };
  72
  73
  74 static int is_qp0(enum ib_qp_type qp_type)
  75 {
  76         return qp_type == IB_QPT_SMI;
  77 }
  78
  79 static int is_sqp(enum ib_qp_type qp_type)
  80 {
  81         return is_qp0(qp_type) || is_qp1(qp_type);
  82 }
  83
  84 static void *get_wqe(struct mlx5_ib_qp *qp, int offset)
  85 {
  86         return mlx5_buf_offset(&qp->buf, offset);
  87 }
  88
  89 static void *get_recv_wqe(struct mlx5_ib_qp *qp, int n)
  90 {
  91         return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
  92 }
  93
  94 void *mlx5_get_send_wqe(struct mlx5_ib_qp *qp, int n)
  95 {
  96         return get_wqe(qp, qp->sq.offset + (n << MLX5_IB_SQ_STRIDE));
  97 }
  98
  99 /**
 100  * mlx5_ib_read_user_wqe() - Copy a user-space WQE to kernel space.
 101  *
 102  * @qp: QP to copy from.
 103  * @send: copy from the send queue when non-zero, use the receive queue
 104  *        otherwise.
 105  * @wqe_index:  index to start copying from. For send work queues, the
 106  *              wqe_index is in units of MLX5_SEND_WQE_BB.
 107  *              For receive work queue, it is the number of work queue
 108  *              element in the queue.
 109  * @buffer: destination buffer.
 110  * @length: maximum number of bytes to copy.
 111  *
 112  * Copies at least a single WQE, but may copy more data.
 113  *
 114  * Return: the number of bytes copied, or an error code.
 115  */
 116 int mlx5_ib_read_user_wqe(struct mlx5_ib_qp *qp, int send, int wqe_index,
 117                           void *buffer, u32 length)
 118 {
 119         struct ib_device *ibdev = qp->ibqp.device;
 120         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 121         struct mlx5_ib_wq *wq = send ? &qp->sq : &qp->rq;
 122         size_t offset;
 123         size_t wq_end;
 124         struct ib_umem *umem = qp->umem;
 125         u32 first_copy_length;
 126         int wqe_length;
 127         int ret;
 128
 129         if (wq->wqe_cnt == 0) {
 130                 mlx5_ib_dbg(dev, "mlx5_ib_read_user_wqe for a QP with wqe_cnt == 0. qp_type: 0x%x\n",
 131                             qp->ibqp.qp_type);
 132                 return -EINVAL;
 133         }
 134
 135         offset = wq->offset + ((wqe_index % wq->wqe_cnt) << wq->wqe_shift);
 136         wq_end = wq->offset + (wq->wqe_cnt << wq->wqe_shift);
 137
 138         if (send && length < sizeof(struct mlx5_wqe_ctrl_seg))
 139                 return -EINVAL;
 140
 141         if (offset > umem->length ||
 142             (send && offset + sizeof(struct mlx5_wqe_ctrl_seg) > umem->length))
 143                 return -EINVAL;
 144
 145         first_copy_length = min_t(u32, offset + length, wq_end) - offset;
 146         ret = ib_umem_copy_from(buffer, umem, offset, first_copy_length);
 147         if (ret)
 148                 return ret;
 149
 150         if (send) {
 151                 struct mlx5_wqe_ctrl_seg *ctrl = buffer;
 152                 int ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
 153
 154                 wqe_length = ds * MLX5_WQE_DS_UNITS;
 155         } else {
 156                 wqe_length = 1 << wq->wqe_shift;
 157         }
 158
 159         if (wqe_length <= first_copy_length)
 160                 return first_copy_length;
 161
 162         ret = ib_umem_copy_from(buffer + first_copy_length, umem, wq->offset,
 163                                 wqe_length - first_copy_length);
 164         if (ret)
 165                 return ret;
 166
 167         return wqe_length;
 168 }
 169
 170 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
 171 {
 172         struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
 173         struct ib_event event;
 174
 175         if (type == MLX5_EVENT_TYPE_PATH_MIG)
 176                 to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
 177
 178         if (ibqp->event_handler) {
 179                 event.device     = ibqp->device;
 180                 event.element.qp = ibqp;
 181                 switch (type) {
 182                 case MLX5_EVENT_TYPE_PATH_MIG:
 183                         event.event = IB_EVENT_PATH_MIG;
 184                         break;
 185                 case MLX5_EVENT_TYPE_COMM_EST:
 186                         event.event = IB_EVENT_COMM_EST;
 187                         break;
 188                 case MLX5_EVENT_TYPE_SQ_DRAINED:
 189                         event.event = IB_EVENT_SQ_DRAINED;
 190                         break;
 191                 case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
 192                         event.event = IB_EVENT_QP_LAST_WQE_REACHED;
 193                         break;
 194                 case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
 195                         event.event = IB_EVENT_QP_FATAL;
 196                         break;
 197                 case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
 198                         event.event = IB_EVENT_PATH_MIG_ERR;
 199                         break;
 200                 case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
 201                         event.event = IB_EVENT_QP_REQ_ERR;
 202                         break;
 203                 case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
 204                         event.event = IB_EVENT_QP_ACCESS_ERR;
 205                         break;
 206                 default:
 207                         pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
 208                         return;
 209                 }
 210
 211                 ibqp->event_handler(&event, ibqp->qp_context);
 212         }
 213 }
 214
 215 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
 216                        int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
 217 {
 218         int wqe_size;
 219         int wq_size;
 220
 221         /* Sanity check RQ size before proceeding */
 222         if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
 223                 return -EINVAL;
 224
 225         if (!has_rq) {
 226                 qp->rq.max_gs = 0;
 227                 qp->rq.wqe_cnt = 0;
 228                 qp->rq.wqe_shift = 0;
 229         } else {
 230                 if (ucmd) {
 231                         qp->rq.wqe_cnt = ucmd->rq_wqe_count;
 232                         qp->rq.wqe_shift = ucmd->rq_wqe_shift;
 233                         qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
 234                         qp->rq.max_post = qp->rq.wqe_cnt;
 235                 } else {
 236                         wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
 237                         wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
 238                         wqe_size = roundup_pow_of_two(wqe_size);
 239                         wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
 240                         wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
 241                         qp->rq.wqe_cnt = wq_size / wqe_size;
 242                         if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
 243                                 mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
 244                                             wqe_size,
 245                                             MLX5_CAP_GEN(dev->mdev,
 246                                                          max_wqe_sz_rq));
 247                                 return -EINVAL;
 248                         }
 249                         qp->rq.wqe_shift = ilog2(wqe_size);
 250                         qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
 251                         qp->rq.max_post = qp->rq.wqe_cnt;
 252                 }
 253         }
 254
 255         return 0;
 256 }
 257
 258 static int sq_overhead(enum ib_qp_type qp_type)
 259 {
 260         int size = 0;
 261
 262         switch (qp_type) {
 263         case IB_QPT_XRC_INI:
 264                 size += sizeof(struct mlx5_wqe_xrc_seg);
 265                 /* fall through */
 266         case IB_QPT_RC:
 267                 size += sizeof(struct mlx5_wqe_ctrl_seg) +
 268                         sizeof(struct mlx5_wqe_atomic_seg) +
 269                         sizeof(struct mlx5_wqe_raddr_seg);
 270                 break;
 271
 272         case IB_QPT_XRC_TGT:
 273                 return 0;
 274
 275         case IB_QPT_UC:
 276                 size += sizeof(struct mlx5_wqe_ctrl_seg) +
 277                         sizeof(struct mlx5_wqe_raddr_seg) +
 278                         sizeof(struct mlx5_wqe_umr_ctrl_seg) +
 279                         sizeof(struct mlx5_mkey_seg);
 280                 break;
 281
 282         case IB_QPT_UD:
 283         case IB_QPT_SMI:
 284         case IB_QPT_GSI:
 285                 size += sizeof(struct mlx5_wqe_ctrl_seg) +
 286                         sizeof(struct mlx5_wqe_datagram_seg);
 287                 break;
 288
 289         case MLX5_IB_QPT_REG_UMR:
 290                 size += sizeof(struct mlx5_wqe_ctrl_seg) +
 291                         sizeof(struct mlx5_wqe_umr_ctrl_seg) +
 292                         sizeof(struct mlx5_mkey_seg);
 293                 break;
 294
 295         default:
 296                 return -EINVAL;
 297         }
 298
 299         return size;
 300 }
 301
 302 static int calc_send_wqe(struct ib_qp_init_attr *attr)
 303 {
 304         int inl_size = 0;
 305         int size;
 306
 307         size = sq_overhead(attr->qp_type);
 308         if (size < 0)
 309                 return size;
 310
 311         if (attr->cap.max_inline_data) {
 312                 inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
 313                         attr->cap.max_inline_data;
 314         }
 315
 316         size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
 317         if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN &&
 318             ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
 319                         return MLX5_SIG_WQE_SIZE;
 320         else
 321                 return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
 322 }
 323
 324 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
 325                         struct mlx5_ib_qp *qp)
 326 {
 327         int wqe_size;
 328         int wq_size;
 329
 330         if (!attr->cap.max_send_wr)
 331                 return 0;
 332
 333         wqe_size = calc_send_wqe(attr);
 334         mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
 335         if (wqe_size < 0)
 336                 return wqe_size;
 337
 338         if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
 339                 mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
 340                             wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
 341                 return -EINVAL;
 342         }
 343
 344         qp->max_inline_data = wqe_size - sq_overhead(attr->qp_type) -
 345                 sizeof(struct mlx5_wqe_inline_seg);
 346         attr->cap.max_inline_data = qp->max_inline_data;
 347
 348         if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN)
 349                 qp->signature_en = true;
 350
 351         wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
 352         qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
 353         if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
 354                 mlx5_ib_dbg(dev, "wqe count(%d) exceeds limits(%d)\n",
 355                             qp->sq.wqe_cnt,
 356                             1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
 357                 return -ENOMEM;
 358         }
 359         qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
 360         qp->sq.max_gs = attr->cap.max_send_sge;
 361         qp->sq.max_post = wq_size / wqe_size;
 362         attr->cap.max_send_wr = qp->sq.max_post;
 363
 364         return wq_size;
 365 }
 366
 367 static int set_user_buf_size(struct mlx5_ib_dev *dev,
 368                             struct mlx5_ib_qp *qp,
 369                             struct mlx5_ib_create_qp *ucmd)
 370 {
 371         int desc_sz = 1 << qp->sq.wqe_shift;
 372
 373         if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
 374                 mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
 375                              desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
 376                 return -EINVAL;
 377         }
 378
 379         if (ucmd->sq_wqe_count && ((1 << ilog2(ucmd->sq_wqe_count)) != ucmd->sq_wqe_count)) {
 380                 mlx5_ib_warn(dev, "sq_wqe_count %d, sq_wqe_count %d\n",
 381                              ucmd->sq_wqe_count, ucmd->sq_wqe_count);
 382                 return -EINVAL;
 383         }
 384
 385         qp->sq.wqe_cnt = ucmd->sq_wqe_count;
 386
 387         if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
 388                 mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
 389                              qp->sq.wqe_cnt,
 390                              1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
 391                 return -EINVAL;
 392         }
 393
 394         qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
 395                 (qp->sq.wqe_cnt << 6);
 396
 397         return 0;
 398 }
 399
 400 static int qp_has_rq(struct ib_qp_init_attr *attr)
 401 {
 402         if (attr->qp_type == IB_QPT_XRC_INI ||
 403             attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
 404             attr->qp_type == MLX5_IB_QPT_REG_UMR ||
 405             !attr->cap.max_recv_wr)
 406                 return 0;
 407
 408         return 1;
 409 }
 410
 411 static int first_med_uuar(void)
 412 {
 413         return 1;
 414 }
 415
 416 static int next_uuar(int n)
 417 {
 418         n++;
 419
 420         while (((n % 4) & 2))
 421                 n++;
 422
 423         return n;
 424 }
 425
 426 static int num_med_uuar(struct mlx5_uuar_info *uuari)
 427 {
 428         int n;
 429
 430         n = uuari->num_uars * MLX5_NON_FP_BF_REGS_PER_PAGE -
 431                 uuari->num_low_latency_uuars - 1;
 432
 433         return n >= 0 ? n : 0;
 434 }
 435
 436 static int max_uuari(struct mlx5_uuar_info *uuari)
 437 {
 438         return uuari->num_uars * 4;
 439 }
 440
 441 static int first_hi_uuar(struct mlx5_uuar_info *uuari)
 442 {
 443         int med;
 444         int i;
 445         int t;
 446
 447         med = num_med_uuar(uuari);
 448         for (t = 0, i = first_med_uuar();; i = next_uuar(i)) {
 449                 t++;
 450                 if (t == med)
 451                         return next_uuar(i);
 452         }
 453
 454         return 0;
 455 }
 456
 457 static int alloc_high_class_uuar(struct mlx5_uuar_info *uuari)
 458 {
 459         int i;
 460
 461         for (i = first_hi_uuar(uuari); i < max_uuari(uuari); i = next_uuar(i)) {
 462                 if (!test_bit(i, uuari->bitmap)) {
 463                         set_bit(i, uuari->bitmap);
 464                         uuari->count[i]++;
 465                         return i;
 466                 }
 467         }
 468
 469         return -ENOMEM;
 470 }
 471
 472 static int alloc_med_class_uuar(struct mlx5_uuar_info *uuari)
 473 {
 474         int minidx = first_med_uuar();
 475         int i;
 476
 477         for (i = first_med_uuar(); i < first_hi_uuar(uuari); i = next_uuar(i)) {
 478                 if (uuari->count[i] < uuari->count[minidx])
 479                         minidx = i;
 480         }
 481
 482         uuari->count[minidx]++;
 483         return minidx;
 484 }
 485
 486 static int alloc_uuar(struct mlx5_uuar_info *uuari,
 487                       enum mlx5_ib_latency_class lat)
 488 {
 489         int uuarn = -EINVAL;
 490
 491         mutex_lock(&uuari->lock);
 492         switch (lat) {
 493         case MLX5_IB_LATENCY_CLASS_LOW:
 494                 uuarn = 0;
 495                 uuari->count[uuarn]++;
 496                 break;
 497
 498         case MLX5_IB_LATENCY_CLASS_MEDIUM:
 499                 if (uuari->ver < 2)
 500                         uuarn = -ENOMEM;
 501                 else
 502                         uuarn = alloc_med_class_uuar(uuari);
 503                 break;
 504
 505         case MLX5_IB_LATENCY_CLASS_HIGH:
 506                 if (uuari->ver < 2)
 507                         uuarn = -ENOMEM;
 508                 else
 509                         uuarn = alloc_high_class_uuar(uuari);
 510                 break;
 511
 512         case MLX5_IB_LATENCY_CLASS_FAST_PATH:
 513                 uuarn = 2;
 514                 break;
 515         }
 516         mutex_unlock(&uuari->lock);
 517
 518         return uuarn;
 519 }
 520
 521 static void free_med_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
 522 {
 523         clear_bit(uuarn, uuari->bitmap);
 524         --uuari->count[uuarn];
 525 }
 526
 527 static void free_high_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
 528 {
 529         clear_bit(uuarn, uuari->bitmap);
 530         --uuari->count[uuarn];
 531 }
 532
 533 static void free_uuar(struct mlx5_uuar_info *uuari, int uuarn)
 534 {
 535         int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
 536         int high_uuar = nuuars - uuari->num_low_latency_uuars;
 537
 538         mutex_lock(&uuari->lock);
 539         if (uuarn == 0) {
 540                 --uuari->count[uuarn];
 541                 goto out;
 542         }
 543
 544         if (uuarn < high_uuar) {
 545                 free_med_class_uuar(uuari, uuarn);
 546                 goto out;
 547         }
 548
 549         free_high_class_uuar(uuari, uuarn);
 550
 551 out:
 552         mutex_unlock(&uuari->lock);
 553 }
 554
 555 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
 556 {
 557         switch (state) {
 558         case IB_QPS_RESET:      return MLX5_QP_STATE_RST;
 559         case IB_QPS_INIT:       return MLX5_QP_STATE_INIT;
 560         case IB_QPS_RTR:        return MLX5_QP_STATE_RTR;
 561         case IB_QPS_RTS:        return MLX5_QP_STATE_RTS;
 562         case IB_QPS_SQD:        return MLX5_QP_STATE_SQD;
 563         case IB_QPS_SQE:        return MLX5_QP_STATE_SQER;
 564         case IB_QPS_ERR:        return MLX5_QP_STATE_ERR;
 565         default:                return -1;
 566         }
 567 }
 568
 569 static int to_mlx5_st(enum ib_qp_type type)
 570 {
 571         switch (type) {
 572         case IB_QPT_RC:                 return MLX5_QP_ST_RC;
 573         case IB_QPT_UC:                 return MLX5_QP_ST_UC;
 574         case IB_QPT_UD:                 return MLX5_QP_ST_UD;
 575         case MLX5_IB_QPT_REG_UMR:       return MLX5_QP_ST_REG_UMR;
 576         case IB_QPT_XRC_INI:
 577         case IB_QPT_XRC_TGT:            return MLX5_QP_ST_XRC;
 578         case IB_QPT_SMI:                return MLX5_QP_ST_QP0;
 579         case IB_QPT_GSI:                return MLX5_QP_ST_QP1;
 580         case IB_QPT_RAW_IPV6:           return MLX5_QP_ST_RAW_IPV6;
 581         case IB_QPT_RAW_ETHERTYPE:      return MLX5_QP_ST_RAW_ETHERTYPE;
 582         case IB_QPT_RAW_PACKET:
 583         case IB_QPT_MAX:
 584         default:                return -EINVAL;
 585         }
 586 }
 587
 588 static int uuarn_to_uar_index(struct mlx5_uuar_info *uuari, int uuarn)
 589 {
 590         return uuari->uars[uuarn / MLX5_BF_REGS_PER_PAGE].index;
 591 }
 592
 593 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
 594                           struct mlx5_ib_qp *qp, struct ib_udata *udata,
 595                           struct mlx5_create_qp_mbox_in **in,
 596                           struct mlx5_ib_create_qp_resp *resp, int *inlen)
 597 {
 598         struct mlx5_ib_ucontext *context;
 599         struct mlx5_ib_create_qp ucmd;
 600         int page_shift = 0;
 601         int uar_index;
 602         int npages;
 603         u32 offset = 0;
 604         int uuarn;
 605         int ncont = 0;
 606         int err;
 607
 608         err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
 609         if (err) {
 610                 mlx5_ib_dbg(dev, "copy failed\n");
 611                 return err;
 612         }
 613
 614         context = to_mucontext(pd->uobject->context);
 615         /*
 616          * TBD: should come from the verbs when we have the API
 617          */
 618         uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_HIGH);
 619         if (uuarn < 0) {
 620                 mlx5_ib_dbg(dev, "failed to allocate low latency UUAR\n");
 621                 mlx5_ib_dbg(dev, "reverting to medium latency\n");
 622                 uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_MEDIUM);
 623                 if (uuarn < 0) {
 624                         mlx5_ib_dbg(dev, "failed to allocate medium latency UUAR\n");
 625                         mlx5_ib_dbg(dev, "reverting to high latency\n");
 626                         uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_LOW);
 627                         if (uuarn < 0) {
 628                                 mlx5_ib_warn(dev, "uuar allocation failed\n");
 629                                 return uuarn;
 630                         }
 631                 }
 632         }
 633
 634         uar_index = uuarn_to_uar_index(&context->uuari, uuarn);
 635         mlx5_ib_dbg(dev, "uuarn 0x%x, uar_index 0x%x\n", uuarn, uar_index);
 636
 637         qp->rq.offset = 0;
 638         qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
 639         qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
 640
 641         err = set_user_buf_size(dev, qp, &ucmd);
 642         if (err)
 643                 goto err_uuar;
 644
 645         if (ucmd.buf_addr && qp->buf_size) {
 646                 qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
 647                                        qp->buf_size, 0, 0);
 648                 if (IS_ERR(qp->umem)) {
 649                         mlx5_ib_dbg(dev, "umem_get failed\n");
 650                         err = PTR_ERR(qp->umem);
 651                         goto err_uuar;
 652                 }
 653         } else {
 654                 qp->umem = NULL;
 655         }
 656
 657         if (qp->umem) {
 658                 mlx5_ib_cont_pages(qp->umem, ucmd.buf_addr, &npages, &page_shift,
 659                                    &ncont, NULL);
 660                 err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift, &offset);
 661                 if (err) {
 662                         mlx5_ib_warn(dev, "bad offset\n");
 663                         goto err_umem;
 664                 }
 665                 mlx5_ib_dbg(dev, "addr 0x%llx, size %d, npages %d, page_shift %d, ncont %d, offset %d\n",
 666                             ucmd.buf_addr, qp->buf_size, npages, page_shift, ncont, offset);
 667         }
 668
 669         *inlen = sizeof(**in) + sizeof(*(*in)->pas) * ncont;
 670         *in = mlx5_vzalloc(*inlen);
 671         if (!*in) {
 672                 err = -ENOMEM;
 673                 goto err_umem;
 674         }
 675         if (qp->umem)
 676                 mlx5_ib_populate_pas(dev, qp->umem, page_shift, (*in)->pas, 0);
 677         (*in)->ctx.log_pg_sz_remote_qpn =
 678                 cpu_to_be32((page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
 679         (*in)->ctx.params2 = cpu_to_be32(offset << 6);
 680
 681         (*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
 682         resp->uuar_index = uuarn;
 683         qp->uuarn = uuarn;
 684
 685         err = mlx5_ib_db_map_user(context, ucmd.db_addr, &qp->db);
 686         if (err) {
 687                 mlx5_ib_dbg(dev, "map failed\n");
 688                 goto err_free;
 689         }
 690
 691         err = ib_copy_to_udata(udata, resp, sizeof(*resp));
 692         if (err) {
 693                 mlx5_ib_dbg(dev, "copy failed\n");
 694                 goto err_unmap;
 695         }
 696         qp->create_type = MLX5_QP_USER;
 697
 698         return 0;
 699
 700 err_unmap:
 701         mlx5_ib_db_unmap_user(context, &qp->db);
 702
 703 err_free:
 704         kvfree(*in);
 705
 706 err_umem:
 707         if (qp->umem)
 708                 ib_umem_release(qp->umem);
 709
 710 err_uuar:
 711         free_uuar(&context->uuari, uuarn);
 712         return err;
 713 }
 714
 715 static void destroy_qp_user(struct ib_pd *pd, struct mlx5_ib_qp *qp)
 716 {
 717         struct mlx5_ib_ucontext *context;
 718
 719         context = to_mucontext(pd->uobject->context);
 720         mlx5_ib_db_unmap_user(context, &qp->db);
 721         if (qp->umem)
 722                 ib_umem_release(qp->umem);
 723         free_uuar(&context->uuari, qp->uuarn);
 724 }
 725
 726 static int create_kernel_qp(struct mlx5_ib_dev *dev,
 727                             struct ib_qp_init_attr *init_attr,
 728                             struct mlx5_ib_qp *qp,
 729                             struct mlx5_create_qp_mbox_in **in, int *inlen)
 730 {
 731         enum mlx5_ib_latency_class lc = MLX5_IB_LATENCY_CLASS_LOW;
 732         struct mlx5_uuar_info *uuari;
 733         int uar_index;
 734         int uuarn;
 735         int err;
 736
 737         uuari = &dev->mdev->priv.uuari;
 738         if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN | IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK))
 739                 return -EINVAL;
 740
 741         if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
 742                 lc = MLX5_IB_LATENCY_CLASS_FAST_PATH;
 743
 744         uuarn = alloc_uuar(uuari, lc);
 745         if (uuarn < 0) {
 746                 mlx5_ib_dbg(dev, "\n");
 747                 return -ENOMEM;
 748         }
 749
 750         qp->bf = &uuari->bfs[uuarn];
 751         uar_index = qp->bf->uar->index;
 752
 753         err = calc_sq_size(dev, init_attr, qp);
 754         if (err < 0) {
 755                 mlx5_ib_dbg(dev, "err %d\n", err);
 756                 goto err_uuar;
 757         }
 758
 759         qp->rq.offset = 0;
 760         qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
 761         qp->buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
 762
 763         err = mlx5_buf_alloc(dev->mdev, qp->buf_size, &qp->buf);
 764         if (err) {
 765                 mlx5_ib_dbg(dev, "err %d\n", err);
 766                 goto err_uuar;
 767         }
 768
 769         qp->sq.qend = mlx5_get_send_wqe(qp, qp->sq.wqe_cnt);
 770         *inlen = sizeof(**in) + sizeof(*(*in)->pas) * qp->buf.npages;
 771         *in = mlx5_vzalloc(*inlen);
 772         if (!*in) {
 773                 err = -ENOMEM;
 774                 goto err_buf;
 775         }
 776         (*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
 777         (*in)->ctx.log_pg_sz_remote_qpn =
 778                 cpu_to_be32((qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
 779         /* Set "fast registration enabled" for all kernel QPs */
 780         (*in)->ctx.params1 |= cpu_to_be32(1 << 11);
 781         (*in)->ctx.sq_crq_size |= cpu_to_be16(1 << 4);
 782
 783         mlx5_fill_page_array(&qp->buf, (*in)->pas);
 784
 785         err = mlx5_db_alloc(dev->mdev, &qp->db);
 786         if (err) {
 787                 mlx5_ib_dbg(dev, "err %d\n", err);
 788                 goto err_free;
 789         }
 790
 791         qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
 792         qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
 793         qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
 794         qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
 795         qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
 796
 797         if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
 798             !qp->sq.w_list || !qp->sq.wqe_head) {
 799                 err = -ENOMEM;
 800                 goto err_wrid;
 801         }
 802         qp->create_type = MLX5_QP_KERNEL;
 803
 804         return 0;
 805
 806 err_wrid:
 807         mlx5_db_free(dev->mdev, &qp->db);
 808         kfree(qp->sq.wqe_head);
 809         kfree(qp->sq.w_list);
 810         kfree(qp->sq.wrid);
 811         kfree(qp->sq.wr_data);
 812         kfree(qp->rq.wrid);
 813
 814 err_free:
 815         kvfree(*in);
 816
 817 err_buf:
 818         mlx5_buf_free(dev->mdev, &qp->buf);
 819
 820 err_uuar:
 821         free_uuar(&dev->mdev->priv.uuari, uuarn);
 822         return err;
 823 }
 824
 825 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
 826 {
 827         mlx5_db_free(dev->mdev, &qp->db);
 828         kfree(qp->sq.wqe_head);
 829         kfree(qp->sq.w_list);
 830         kfree(qp->sq.wrid);
 831         kfree(qp->sq.wr_data);
 832         kfree(qp->rq.wrid);
 833         mlx5_buf_free(dev->mdev, &qp->buf);
 834         free_uuar(&dev->mdev->priv.uuari, qp->bf->uuarn);
 835 }
 836
 837 static __be32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
 838 {
 839         if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
 840             (attr->qp_type == IB_QPT_XRC_INI))
 841                 return cpu_to_be32(MLX5_SRQ_RQ);
 842         else if (!qp->has_rq)
 843                 return cpu_to_be32(MLX5_ZERO_LEN_RQ);
 844         else
 845                 return cpu_to_be32(MLX5_NON_ZERO_RQ);
 846 }
 847
 848 static int is_connected(enum ib_qp_type qp_type)
 849 {
 850         if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC)
 851                 return 1;
 852
 853         return 0;
 854 }
 855
 856 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
 857                             struct ib_qp_init_attr *init_attr,
 858                             struct ib_udata *udata, struct mlx5_ib_qp *qp)
 859 {
 860         struct mlx5_ib_resources *devr = &dev->devr;
 861         struct mlx5_core_dev *mdev = dev->mdev;
 862         struct mlx5_ib_create_qp_resp resp;
 863         struct mlx5_create_qp_mbox_in *in;
 864         struct mlx5_ib_create_qp ucmd;
 865         int inlen = sizeof(*in);
 866         int err;
 867
 868         mlx5_ib_odp_create_qp(qp);
 869
 870         mutex_init(&qp->mutex);
 871         spin_lock_init(&qp->sq.lock);
 872         spin_lock_init(&qp->rq.lock);
 873
 874         if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
 875                 if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
 876                         mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
 877                         return -EINVAL;
 878                 } else {
 879                         qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
 880                 }
 881         }
 882
 883         if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
 884                 qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
 885
 886         if (pd && pd->uobject) {
 887                 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
 888                         mlx5_ib_dbg(dev, "copy failed\n");
 889                         return -EFAULT;
 890                 }
 891
 892                 qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
 893                 qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
 894         } else {
 895                 qp->wq_sig = !!wq_signature;
 896         }
 897
 898         qp->has_rq = qp_has_rq(init_attr);
 899         err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
 900                           qp, (pd && pd->uobject) ? &ucmd : NULL);
 901         if (err) {
 902                 mlx5_ib_dbg(dev, "err %d\n", err);
 903                 return err;
 904         }
 905
 906         if (pd) {
 907                 if (pd->uobject) {
 908                         __u32 max_wqes =
 909                                 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
 910                         mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
 911                         if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
 912                             ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
 913                                 mlx5_ib_dbg(dev, "invalid rq params\n");
 914                                 return -EINVAL;
 915                         }
 916                         if (ucmd.sq_wqe_count > max_wqes) {
 917                                 mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
 918                                             ucmd.sq_wqe_count, max_wqes);
 919                                 return -EINVAL;
 920                         }
 921                         err = create_user_qp(dev, pd, qp, udata, &in, &resp, &inlen);
 922                         if (err)
 923                                 mlx5_ib_dbg(dev, "err %d\n", err);
 924                 } else {
 925                         err = create_kernel_qp(dev, init_attr, qp, &in, &inlen);
 926                         if (err)
 927                                 mlx5_ib_dbg(dev, "err %d\n", err);
 928                         else
 929                                 qp->pa_lkey = to_mpd(pd)->pa_lkey;
 930                 }
 931
 932                 if (err)
 933                         return err;
 934         } else {
 935                 in = mlx5_vzalloc(sizeof(*in));
 936                 if (!in)
 937                         return -ENOMEM;
 938
 939                 qp->create_type = MLX5_QP_EMPTY;
 940         }
 941
 942         if (is_sqp(init_attr->qp_type))
 943                 qp->port = init_attr->port_num;
 944
 945         in->ctx.flags = cpu_to_be32(to_mlx5_st(init_attr->qp_type) << 16 |
 946                                     MLX5_QP_PM_MIGRATED << 11);
 947
 948         if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
 949                 in->ctx.flags_pd = cpu_to_be32(to_mpd(pd ? pd : devr->p0)->pdn);
 950         else
 951                 in->ctx.flags_pd = cpu_to_be32(MLX5_QP_LAT_SENSITIVE);
 952
 953         if (qp->wq_sig)
 954                 in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_ENABLE_SIG);
 955
 956         if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
 957                 in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_BLOCK_MCAST);
 958
 959         if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
 960                 int rcqe_sz;
 961                 int scqe_sz;
 962
 963                 rcqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->recv_cq);
 964                 scqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->send_cq);
 965
 966                 if (rcqe_sz == 128)
 967                         in->ctx.cs_res = MLX5_RES_SCAT_DATA64_CQE;
 968                 else
 969                         in->ctx.cs_res = MLX5_RES_SCAT_DATA32_CQE;
 970
 971                 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) {
 972                         if (scqe_sz == 128)
 973                                 in->ctx.cs_req = MLX5_REQ_SCAT_DATA64_CQE;
 974                         else
 975                                 in->ctx.cs_req = MLX5_REQ_SCAT_DATA32_CQE;
 976                 }
 977         }
 978
 979         if (qp->rq.wqe_cnt) {
 980                 in->ctx.rq_size_stride = (qp->rq.wqe_shift - 4);
 981                 in->ctx.rq_size_stride |= ilog2(qp->rq.wqe_cnt) << 3;
 982         }
 983
 984         in->ctx.rq_type_srqn = get_rx_type(qp, init_attr);
 985
 986         if (qp->sq.wqe_cnt)
 987                 in->ctx.sq_crq_size |= cpu_to_be16(ilog2(qp->sq.wqe_cnt) << 11);
 988         else
 989                 in->ctx.sq_crq_size |= cpu_to_be16(0x8000);
 990
 991         /* Set default resources */
 992         switch (init_attr->qp_type) {
 993         case IB_QPT_XRC_TGT:
 994                 in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
 995                 in->ctx.cqn_send = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
 996                 in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
 997                 in->ctx.xrcd = cpu_to_be32(to_mxrcd(init_attr->xrcd)->xrcdn);
 998                 break;
 999         case IB_QPT_XRC_INI:
1000                 in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
1001                 in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
1002                 in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
1003                 break;
1004         default:
1005                 if (init_attr->srq) {
1006                         in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x0)->xrcdn);
1007                         in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(init_attr->srq)->msrq.srqn);
1008                 } else {
1009                         in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
1010                         in->ctx.rq_type_srqn |=
1011                                 cpu_to_be32(to_msrq(devr->s1)->msrq.srqn);
1012                 }
1013         }
1014
1015         if (init_attr->send_cq)
1016                 in->ctx.cqn_send = cpu_to_be32(to_mcq(init_attr->send_cq)->mcq.cqn);
1017
1018         if (init_attr->recv_cq)
1019                 in->ctx.cqn_recv = cpu_to_be32(to_mcq(init_attr->recv_cq)->mcq.cqn);
1020
1021         in->ctx.db_rec_addr = cpu_to_be64(qp->db.dma);
1022
1023         err = mlx5_core_create_qp(dev->mdev, &qp->mqp, in, inlen);
1024         if (err) {
1025                 mlx5_ib_dbg(dev, "create qp failed\n");
1026                 goto err_create;
1027         }
1028
1029         kvfree(in);
1030         /* Hardware wants QPN written in big-endian order (after
1031          * shifting) for send doorbell.  Precompute this value to save
1032          * a little bit when posting sends.
1033          */
1034         qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1035
1036         qp->mqp.event = mlx5_ib_qp_event;
1037
1038         return 0;
1039
1040 err_create:
1041         if (qp->create_type == MLX5_QP_USER)
1042                 destroy_qp_user(pd, qp);
1043         else if (qp->create_type == MLX5_QP_KERNEL)
1044                 destroy_qp_kernel(dev, qp);
1045
1046         kvfree(in);
1047         return err;
1048 }
1049
1050 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1051         __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1052 {
1053         if (send_cq) {
1054                 if (recv_cq) {
1055                         if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1056                                 spin_lock_irq(&send_cq->lock);
1057                                 spin_lock_nested(&recv_cq->lock,
1058                                                  SINGLE_DEPTH_NESTING);
1059                         } else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1060                                 spin_lock_irq(&send_cq->lock);
1061                                 __acquire(&recv_cq->lock);
1062                         } else {
1063                                 spin_lock_irq(&recv_cq->lock);
1064                                 spin_lock_nested(&send_cq->lock,
1065                                                  SINGLE_DEPTH_NESTING);
1066                         }
1067                 } else {
1068                         spin_lock_irq(&send_cq->lock);
1069                         __acquire(&recv_cq->lock);
1070                 }
1071         } else if (recv_cq) {
1072                 spin_lock_irq(&recv_cq->lock);
1073                 __acquire(&send_cq->lock);
1074         } else {
1075                 __acquire(&send_cq->lock);
1076                 __acquire(&recv_cq->lock);
1077         }
1078 }
1079
1080 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1081         __releases(&send_cq->lock) __releases(&recv_cq->lock)
1082 {
1083         if (send_cq) {
1084                 if (recv_cq) {
1085                         if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1086                                 spin_unlock(&recv_cq->lock);
1087                                 spin_unlock_irq(&send_cq->lock);
1088                         } else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1089                                 __release(&recv_cq->lock);
1090                                 spin_unlock_irq(&send_cq->lock);
1091                         } else {
1092                                 spin_unlock(&send_cq->lock);
1093                                 spin_unlock_irq(&recv_cq->lock);
1094                         }
1095                 } else {
1096                         __release(&recv_cq->lock);
1097                         spin_unlock_irq(&send_cq->lock);
1098                 }
1099         } else if (recv_cq) {
1100                 __release(&send_cq->lock);
1101                 spin_unlock_irq(&recv_cq->lock);
1102         } else {
1103                 __release(&recv_cq->lock);
1104                 __release(&send_cq->lock);
1105         }
1106 }
1107
1108 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
1109 {
1110         return to_mpd(qp->ibqp.pd);
1111 }
1112
1113 static void get_cqs(struct mlx5_ib_qp *qp,
1114                     struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
1115 {
1116         switch (qp->ibqp.qp_type) {
1117         case IB_QPT_XRC_TGT:
1118                 *send_cq = NULL;
1119                 *recv_cq = NULL;
1120                 break;
1121         case MLX5_IB_QPT_REG_UMR:
1122         case IB_QPT_XRC_INI:
1123                 *send_cq = to_mcq(qp->ibqp.send_cq);
1124                 *recv_cq = NULL;
1125                 break;
1126
1127         case IB_QPT_SMI:
1128         case IB_QPT_GSI:
1129         case IB_QPT_RC:
1130         case IB_QPT_UC:
1131         case IB_QPT_UD:
1132         case IB_QPT_RAW_IPV6:
1133         case IB_QPT_RAW_ETHERTYPE:
1134                 *send_cq = to_mcq(qp->ibqp.send_cq);
1135                 *recv_cq = to_mcq(qp->ibqp.recv_cq);
1136                 break;
1137
1138         case IB_QPT_RAW_PACKET:
1139         case IB_QPT_MAX:
1140         default:
1141                 *send_cq = NULL;
1142                 *recv_cq = NULL;
1143                 break;
1144         }
1145 }
1146
1147 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1148 {
1149         struct mlx5_ib_cq *send_cq, *recv_cq;
1150         struct mlx5_modify_qp_mbox_in *in;
1151         int err;
1152
1153         in = kzalloc(sizeof(*in), GFP_KERNEL);
1154         if (!in)
1155                 return;
1156
1157         if (qp->state != IB_QPS_RESET) {
1158                 mlx5_ib_qp_disable_pagefaults(qp);
1159                 if (mlx5_core_qp_modify(dev->mdev, to_mlx5_state(qp->state),
1160                                         MLX5_QP_STATE_RST, in, 0, &qp->mqp))
1161                         mlx5_ib_warn(dev, "mlx5_ib: modify QP %06x to RESET failed\n",
1162                                      qp->mqp.qpn);
1163         }
1164
1165         get_cqs(qp, &send_cq, &recv_cq);
1166
1167         if (qp->create_type == MLX5_QP_KERNEL) {
1168                 mlx5_ib_lock_cqs(send_cq, recv_cq);
1169                 __mlx5_ib_cq_clean(recv_cq, qp->mqp.qpn,
1170                                    qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1171                 if (send_cq != recv_cq)
1172                         __mlx5_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1173                 mlx5_ib_unlock_cqs(send_cq, recv_cq);
1174         }
1175
1176         err = mlx5_core_destroy_qp(dev->mdev, &qp->mqp);
1177         if (err)
1178                 mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n", qp->mqp.qpn);
1179         kfree(in);
1180
1181
1182         if (qp->create_type == MLX5_QP_KERNEL)
1183                 destroy_qp_kernel(dev, qp);
1184         else if (qp->create_type == MLX5_QP_USER)
1185                 destroy_qp_user(&get_pd(qp)->ibpd, qp);
1186 }
1187
1188 static const char *ib_qp_type_str(enum ib_qp_type type)
1189 {
1190         switch (type) {
1191         case IB_QPT_SMI:
1192                 return "IB_QPT_SMI";
1193         case IB_QPT_GSI:
1194                 return "IB_QPT_GSI";
1195         case IB_QPT_RC:
1196                 return "IB_QPT_RC";
1197         case IB_QPT_UC:
1198                 return "IB_QPT_UC";
1199         case IB_QPT_UD:
1200                 return "IB_QPT_UD";
1201         case IB_QPT_RAW_IPV6:
1202                 return "IB_QPT_RAW_IPV6";
1203         case IB_QPT_RAW_ETHERTYPE:
1204                 return "IB_QPT_RAW_ETHERTYPE";
1205         case IB_QPT_XRC_INI:
1206                 return "IB_QPT_XRC_INI";
1207         case IB_QPT_XRC_TGT:
1208                 return "IB_QPT_XRC_TGT";
1209         case IB_QPT_RAW_PACKET:
1210                 return "IB_QPT_RAW_PACKET";
1211         case MLX5_IB_QPT_REG_UMR:
1212                 return "MLX5_IB_QPT_REG_UMR";
1213         case IB_QPT_MAX:
1214         default:
1215                 return "Invalid QP type";
1216         }
1217 }
1218
1219 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
1220                                 struct ib_qp_init_attr *init_attr,
1221                                 struct ib_udata *udata)
1222 {
1223         struct mlx5_ib_dev *dev;
1224         struct mlx5_ib_qp *qp;
1225         u16 xrcdn = 0;
1226         int err;
1227
1228         if (pd) {
1229                 dev = to_mdev(pd->device);
1230         } else {
1231                 /* being cautious here */
1232                 if (init_attr->qp_type != IB_QPT_XRC_TGT &&
1233                     init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
1234                         pr_warn("%s: no PD for transport %s\n", __func__,
1235                                 ib_qp_type_str(init_attr->qp_type));
1236                         return ERR_PTR(-EINVAL);
1237                 }
1238                 dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
1239         }
1240
1241         switch (init_attr->qp_type) {
1242         case IB_QPT_XRC_TGT:
1243         case IB_QPT_XRC_INI:
1244                 if (!MLX5_CAP_GEN(dev->mdev, xrc)) {
1245                         mlx5_ib_dbg(dev, "XRC not supported\n");
1246                         return ERR_PTR(-ENOSYS);
1247                 }
1248                 init_attr->recv_cq = NULL;
1249                 if (init_attr->qp_type == IB_QPT_XRC_TGT) {
1250                         xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1251                         init_attr->send_cq = NULL;
1252                 }
1253
1254                 /* fall through */
1255         case IB_QPT_RC:
1256         case IB_QPT_UC:
1257         case IB_QPT_UD:
1258         case IB_QPT_SMI:
1259         case IB_QPT_GSI:
1260         case MLX5_IB_QPT_REG_UMR:
1261                 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1262                 if (!qp)
1263                         return ERR_PTR(-ENOMEM);
1264
1265                 err = create_qp_common(dev, pd, init_attr, udata, qp);
1266                 if (err) {
1267                         mlx5_ib_dbg(dev, "create_qp_common failed\n");
1268                         kfree(qp);
1269                         return ERR_PTR(err);
1270                 }
1271
1272                 if (is_qp0(init_attr->qp_type))
1273                         qp->ibqp.qp_num = 0;
1274                 else if (is_qp1(init_attr->qp_type))
1275                         qp->ibqp.qp_num = 1;
1276                 else
1277                         qp->ibqp.qp_num = qp->mqp.qpn;
1278
1279                 mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
1280                             qp->ibqp.qp_num, qp->mqp.qpn, to_mcq(init_attr->recv_cq)->mcq.cqn,
1281                             to_mcq(init_attr->send_cq)->mcq.cqn);
1282
1283                 qp->xrcdn = xrcdn;
1284
1285                 break;
1286
1287         case IB_QPT_RAW_IPV6:
1288         case IB_QPT_RAW_ETHERTYPE:
1289         case IB_QPT_RAW_PACKET:
1290         case IB_QPT_MAX:
1291         default:
1292                 mlx5_ib_dbg(dev, "unsupported qp type %d\n",
1293                             init_attr->qp_type);
1294                 /* Don't support raw QPs */
1295                 return ERR_PTR(-EINVAL);
1296         }
1297
1298         return &qp->ibqp;
1299 }
1300
1301 int mlx5_ib_destroy_qp(struct ib_qp *qp)
1302 {
1303         struct mlx5_ib_dev *dev = to_mdev(qp->device);
1304         struct mlx5_ib_qp *mqp = to_mqp(qp);
1305
1306         destroy_qp_common(dev, mqp);
1307
1308         kfree(mqp);
1309
1310         return 0;
1311 }
1312
1313 static __be32 to_mlx5_access_flags(struct mlx5_ib_qp *qp, const struct ib_qp_attr *attr,
1314                                    int attr_mask)
1315 {
1316         u32 hw_access_flags = 0;
1317         u8 dest_rd_atomic;
1318         u32 access_flags;
1319
1320         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1321                 dest_rd_atomic = attr->max_dest_rd_atomic;
1322         else
1323                 dest_rd_atomic = qp->resp_depth;
1324
1325         if (attr_mask & IB_QP_ACCESS_FLAGS)
1326                 access_flags = attr->qp_access_flags;
1327         else
1328                 access_flags = qp->atomic_rd_en;
1329
1330         if (!dest_rd_atomic)
1331                 access_flags &= IB_ACCESS_REMOTE_WRITE;
1332
1333         if (access_flags & IB_ACCESS_REMOTE_READ)
1334                 hw_access_flags |= MLX5_QP_BIT_RRE;
1335         if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1336                 hw_access_flags |= (MLX5_QP_BIT_RAE | MLX5_ATOMIC_MODE_CX);
1337         if (access_flags & IB_ACCESS_REMOTE_WRITE)
1338                 hw_access_flags |= MLX5_QP_BIT_RWE;
1339
1340         return cpu_to_be32(hw_access_flags);
1341 }
1342
1343 enum {
1344         MLX5_PATH_FLAG_FL       = 1 << 0,
1345         MLX5_PATH_FLAG_FREE_AR  = 1 << 1,
1346         MLX5_PATH_FLAG_COUNTER  = 1 << 2,
1347 };
1348
1349 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
1350 {
1351         if (rate == IB_RATE_PORT_CURRENT) {
1352                 return 0;
1353         } else if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_300_GBPS) {
1354                 return -EINVAL;
1355         } else {
1356                 while (rate != IB_RATE_2_5_GBPS &&
1357                        !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
1358                          MLX5_CAP_GEN(dev->mdev, stat_rate_support)))
1359                         --rate;
1360         }
1361
1362         return rate + MLX5_STAT_RATE_OFFSET;
1363 }
1364
1365 static int mlx5_set_path(struct mlx5_ib_dev *dev, const struct ib_ah_attr *ah,
1366                          struct mlx5_qp_path *path, u8 port, int attr_mask,
1367                          u32 path_flags, const struct ib_qp_attr *attr)
1368 {
1369         int err;
1370
1371         path->fl = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
1372         path->free_ar = (path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x80 : 0;
1373
1374         if (attr_mask & IB_QP_PKEY_INDEX)
1375                 path->pkey_index = attr->pkey_index;
1376
1377         path->grh_mlid  = ah->src_path_bits & 0x7f;
1378         path->rlid      = cpu_to_be16(ah->dlid);
1379
1380         if (ah->ah_flags & IB_AH_GRH) {
1381                 if (ah->grh.sgid_index >=
1382                     dev->mdev->port_caps[port - 1].gid_table_len) {
1383                         pr_err("sgid_index (%u) too large. max is %d\n",
1384                                ah->grh.sgid_index,
1385                                dev->mdev->port_caps[port - 1].gid_table_len);
1386                         return -EINVAL;
1387                 }
1388                 path->grh_mlid |= 1 << 7;
1389                 path->mgid_index = ah->grh.sgid_index;
1390                 path->hop_limit  = ah->grh.hop_limit;
1391                 path->tclass_flowlabel =
1392                         cpu_to_be32((ah->grh.traffic_class << 20) |
1393                                     (ah->grh.flow_label));
1394                 memcpy(path->rgid, ah->grh.dgid.raw, 16);
1395         }
1396
1397         err = ib_rate_to_mlx5(dev, ah->static_rate);
1398         if (err < 0)
1399                 return err;
1400         path->static_rate = err;
1401         path->port = port;
1402
1403         if (attr_mask & IB_QP_TIMEOUT)
1404                 path->ackto_lt = attr->timeout << 3;
1405
1406         path->sl = ah->sl & 0xf;
1407
1408         return 0;
1409 }
1410
1411 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
1412         [MLX5_QP_STATE_INIT] = {
1413                 [MLX5_QP_STATE_INIT] = {
1414                         [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE            |
1415                                           MLX5_QP_OPTPAR_RAE            |
1416                                           MLX5_QP_OPTPAR_RWE            |
1417                                           MLX5_QP_OPTPAR_PKEY_INDEX     |
1418                                           MLX5_QP_OPTPAR_PRI_PORT,
1419                         [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE            |
1420                                           MLX5_QP_OPTPAR_PKEY_INDEX     |
1421                                           MLX5_QP_OPTPAR_PRI_PORT,
1422                         [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
1423                                           MLX5_QP_OPTPAR_Q_KEY          |
1424                                           MLX5_QP_OPTPAR_PRI_PORT,
1425                 },
1426                 [MLX5_QP_STATE_RTR] = {
1427                         [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
1428                                           MLX5_QP_OPTPAR_RRE            |
1429                                           MLX5_QP_OPTPAR_RAE            |
1430                                           MLX5_QP_OPTPAR_RWE            |
1431                                           MLX5_QP_OPTPAR_PKEY_INDEX,
1432                         [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
1433                                           MLX5_QP_OPTPAR_RWE            |
1434                                           MLX5_QP_OPTPAR_PKEY_INDEX,
1435                         [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
1436                                           MLX5_QP_OPTPAR_Q_KEY,
1437                         [MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX    |
1438                                            MLX5_QP_OPTPAR_Q_KEY,
1439                         [MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
1440                                           MLX5_QP_OPTPAR_RRE            |
1441                                           MLX5_QP_OPTPAR_RAE            |
1442                                           MLX5_QP_OPTPAR_RWE            |
1443                                           MLX5_QP_OPTPAR_PKEY_INDEX,
1444                 },
1445         },
1446         [MLX5_QP_STATE_RTR] = {
1447                 [MLX5_QP_STATE_RTS] = {
1448                         [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
1449                                           MLX5_QP_OPTPAR_RRE            |
1450                                           MLX5_QP_OPTPAR_RAE            |
1451                                           MLX5_QP_OPTPAR_RWE            |
1452                                           MLX5_QP_OPTPAR_PM_STATE       |
1453                                           MLX5_QP_OPTPAR_RNR_TIMEOUT,
1454                         [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
1455                                           MLX5_QP_OPTPAR_RWE            |
1456                                           MLX5_QP_OPTPAR_PM_STATE,
1457                         [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
1458                 },
1459         },
1460         [MLX5_QP_STATE_RTS] = {
1461                 [MLX5_QP_STATE_RTS] = {
1462                         [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE            |
1463                                           MLX5_QP_OPTPAR_RAE            |
1464                                           MLX5_QP_OPTPAR_RWE            |
1465                                           MLX5_QP_OPTPAR_RNR_TIMEOUT    |
1466                                           MLX5_QP_OPTPAR_PM_STATE       |
1467                                           MLX5_QP_OPTPAR_ALT_ADDR_PATH,
1468                         [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE            |
1469                                           MLX5_QP_OPTPAR_PM_STATE       |
1470                                           MLX5_QP_OPTPAR_ALT_ADDR_PATH,
1471                         [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY          |
1472                                           MLX5_QP_OPTPAR_SRQN           |
1473                                           MLX5_QP_OPTPAR_CQN_RCV,
1474                 },
1475         },
1476         [MLX5_QP_STATE_SQER] = {
1477                 [MLX5_QP_STATE_RTS] = {
1478                         [MLX5_QP_ST_UD]  = MLX5_QP_OPTPAR_Q_KEY,
1479                         [MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
1480                         [MLX5_QP_ST_UC]  = MLX5_QP_OPTPAR_RWE,
1481                         [MLX5_QP_ST_RC]  = MLX5_QP_OPTPAR_RNR_TIMEOUT   |
1482                                            MLX5_QP_OPTPAR_RWE           |
1483                                            MLX5_QP_OPTPAR_RAE           |
1484                                            MLX5_QP_OPTPAR_RRE,
1485                 },
1486         },
1487 };
1488
1489 static int ib_nr_to_mlx5_nr(int ib_mask)
1490 {
1491         switch (ib_mask) {
1492         case IB_QP_STATE:
1493                 return 0;
1494         case IB_QP_CUR_STATE:
1495                 return 0;
1496         case IB_QP_EN_SQD_ASYNC_NOTIFY:
1497                 return 0;
1498         case IB_QP_ACCESS_FLAGS:
1499                 return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
1500                         MLX5_QP_OPTPAR_RAE;
1501         case IB_QP_PKEY_INDEX:
1502                 return MLX5_QP_OPTPAR_PKEY_INDEX;
1503         case IB_QP_PORT:
1504                 return MLX5_QP_OPTPAR_PRI_PORT;
1505         case IB_QP_QKEY:
1506                 return MLX5_QP_OPTPAR_Q_KEY;
1507         case IB_QP_AV:
1508                 return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
1509                         MLX5_QP_OPTPAR_PRI_PORT;
1510         case IB_QP_PATH_MTU:
1511                 return 0;
1512         case IB_QP_TIMEOUT:
1513                 return MLX5_QP_OPTPAR_ACK_TIMEOUT;
1514         case IB_QP_RETRY_CNT:
1515                 return MLX5_QP_OPTPAR_RETRY_COUNT;
1516         case IB_QP_RNR_RETRY:
1517                 return MLX5_QP_OPTPAR_RNR_RETRY;
1518         case IB_QP_RQ_PSN:
1519                 return 0;
1520         case IB_QP_MAX_QP_RD_ATOMIC:
1521                 return MLX5_QP_OPTPAR_SRA_MAX;
1522         case IB_QP_ALT_PATH:
1523                 return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
1524         case IB_QP_MIN_RNR_TIMER:
1525                 return MLX5_QP_OPTPAR_RNR_TIMEOUT;
1526         case IB_QP_SQ_PSN:
1527                 return 0;
1528         case IB_QP_MAX_DEST_RD_ATOMIC:
1529                 return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
1530                         MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
1531         case IB_QP_PATH_MIG_STATE:
1532                 return MLX5_QP_OPTPAR_PM_STATE;
1533         case IB_QP_CAP:
1534                 return 0;
1535         case IB_QP_DEST_QPN:
1536                 return 0;
1537         }
1538         return 0;
1539 }
1540
1541 static int ib_mask_to_mlx5_opt(int ib_mask)
1542 {
1543         int result = 0;
1544         int i;
1545
1546         for (i = 0; i < 8 * sizeof(int); i++) {
1547                 if ((1 << i) & ib_mask)
1548                         result |= ib_nr_to_mlx5_nr(1 << i);
1549         }
1550
1551         return result;
1552 }
1553
1554 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
1555                                const struct ib_qp_attr *attr, int attr_mask,
1556                                enum ib_qp_state cur_state, enum ib_qp_state new_state)
1557 {
1558         struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1559         struct mlx5_ib_qp *qp = to_mqp(ibqp);
1560         struct mlx5_ib_cq *send_cq, *recv_cq;
1561         struct mlx5_qp_context *context;
1562         struct mlx5_modify_qp_mbox_in *in;
1563         struct mlx5_ib_pd *pd;
1564         enum mlx5_qp_state mlx5_cur, mlx5_new;
1565         enum mlx5_qp_optpar optpar;
1566         int sqd_event;
1567         int mlx5_st;
1568         int err;
1569
1570         in = kzalloc(sizeof(*in), GFP_KERNEL);
1571         if (!in)
1572                 return -ENOMEM;
1573
1574         context = &in->ctx;
1575         err = to_mlx5_st(ibqp->qp_type);
1576         if (err < 0)
1577                 goto out;
1578
1579         context->flags = cpu_to_be32(err << 16);
1580
1581         if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
1582                 context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
1583         } else {
1584                 switch (attr->path_mig_state) {
1585                 case IB_MIG_MIGRATED:
1586                         context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
1587                         break;
1588                 case IB_MIG_REARM:
1589                         context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
1590                         break;
1591                 case IB_MIG_ARMED:
1592                         context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
1593                         break;
1594                 }
1595         }
1596
1597         if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI) {
1598                 context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
1599         } else if (ibqp->qp_type == IB_QPT_UD ||
1600                    ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
1601                 context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1602         } else if (attr_mask & IB_QP_PATH_MTU) {
1603                 if (attr->path_mtu < IB_MTU_256 ||
1604                     attr->path_mtu > IB_MTU_4096) {
1605                         mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
1606                         err = -EINVAL;
1607                         goto out;
1608                 }
1609                 context->mtu_msgmax = (attr->path_mtu << 5) |
1610                                       (u8)MLX5_CAP_GEN(dev->mdev, log_max_msg);
1611         }
1612
1613         if (attr_mask & IB_QP_DEST_QPN)
1614                 context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
1615
1616         if (attr_mask & IB_QP_PKEY_INDEX)
1617                 context->pri_path.pkey_index = attr->pkey_index;
1618
1619         /* todo implement counter_index functionality */
1620
1621         if (is_sqp(ibqp->qp_type))
1622                 context->pri_path.port = qp->port;
1623
1624         if (attr_mask & IB_QP_PORT)
1625                 context->pri_path.port = attr->port_num;
1626
1627         if (attr_mask & IB_QP_AV) {
1628                 err = mlx5_set_path(dev, &attr->ah_attr, &context->pri_path,
1629                                     attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
1630                                     attr_mask, 0, attr);
1631                 if (err)
1632                         goto out;
1633         }
1634
1635         if (attr_mask & IB_QP_TIMEOUT)
1636                 context->pri_path.ackto_lt |= attr->timeout << 3;
1637
1638         if (attr_mask & IB_QP_ALT_PATH) {
1639                 err = mlx5_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
1640                                     attr->alt_port_num, attr_mask, 0, attr);
1641                 if (err)
1642                         goto out;
1643         }
1644
1645         pd = get_pd(qp);
1646         get_cqs(qp, &send_cq, &recv_cq);
1647
1648         context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
1649         context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
1650         context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
1651         context->params1  = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
1652
1653         if (attr_mask & IB_QP_RNR_RETRY)
1654                 context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1655
1656         if (attr_mask & IB_QP_RETRY_CNT)
1657                 context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1658
1659         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1660                 if (attr->max_rd_atomic)
1661                         context->params1 |=
1662                                 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1663         }
1664
1665         if (attr_mask & IB_QP_SQ_PSN)
1666                 context->next_send_psn = cpu_to_be32(attr->sq_psn);
1667
1668         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1669                 if (attr->max_dest_rd_atomic)
1670                         context->params2 |=
1671                                 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1672         }
1673
1674         if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
1675                 context->params2 |= to_mlx5_access_flags(qp, attr, attr_mask);
1676
1677         if (attr_mask & IB_QP_MIN_RNR_TIMER)
1678                 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1679
1680         if (attr_mask & IB_QP_RQ_PSN)
1681                 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1682
1683         if (attr_mask & IB_QP_QKEY)
1684                 context->qkey = cpu_to_be32(attr->qkey);
1685
1686         if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1687                 context->db_rec_addr = cpu_to_be64(qp->db.dma);
1688
1689         if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD  &&
1690             attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1691                 sqd_event = 1;
1692         else
1693                 sqd_event = 0;
1694
1695         if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1696                 context->sq_crq_size |= cpu_to_be16(1 << 4);
1697
1698
1699         mlx5_cur = to_mlx5_state(cur_state);
1700         mlx5_new = to_mlx5_state(new_state);
1701         mlx5_st = to_mlx5_st(ibqp->qp_type);
1702         if (mlx5_st < 0)
1703                 goto out;
1704
1705         /* If moving to a reset or error state, we must disable page faults on
1706          * this QP and flush all current page faults. Otherwise a stale page
1707          * fault may attempt to work on this QP after it is reset and moved
1708          * again to RTS, and may cause the driver and the device to get out of
1709          * sync. */
1710         if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1711             (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1712                 mlx5_ib_qp_disable_pagefaults(qp);
1713
1714         optpar = ib_mask_to_mlx5_opt(attr_mask);
1715         optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
1716         in->optparam = cpu_to_be32(optpar);
1717         err = mlx5_core_qp_modify(dev->mdev, to_mlx5_state(cur_state),
1718                                   to_mlx5_state(new_state), in, sqd_event,
1719                                   &qp->mqp);
1720         if (err)
1721                 goto out;
1722
1723         if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1724                 mlx5_ib_qp_enable_pagefaults(qp);
1725
1726         qp->state = new_state;
1727
1728         if (attr_mask & IB_QP_ACCESS_FLAGS)
1729                 qp->atomic_rd_en = attr->qp_access_flags;
1730         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1731                 qp->resp_depth = attr->max_dest_rd_atomic;
1732         if (attr_mask & IB_QP_PORT)
1733                 qp->port = attr->port_num;
1734         if (attr_mask & IB_QP_ALT_PATH)
1735                 qp->alt_port = attr->alt_port_num;
1736
1737         /*
1738          * If we moved a kernel QP to RESET, clean up all old CQ
1739          * entries and reinitialize the QP.
1740          */
1741         if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1742                 mlx5_ib_cq_clean(recv_cq, qp->mqp.qpn,
1743                                  ibqp->srq ? to_msrq(ibqp->srq) : NULL);
1744                 if (send_cq != recv_cq)
1745                         mlx5_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1746
1747                 qp->rq.head = 0;
1748                 qp->rq.tail = 0;
1749                 qp->sq.head = 0;
1750                 qp->sq.tail = 0;
1751                 qp->sq.cur_post = 0;
1752                 qp->sq.last_poll = 0;
1753                 qp->db.db[MLX5_RCV_DBR] = 0;
1754                 qp->db.db[MLX5_SND_DBR] = 0;
1755         }
1756
1757 out:
1758         kfree(in);
1759         return err;
1760 }
1761
1762 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1763                       int attr_mask, struct ib_udata *udata)
1764 {
1765         struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1766         struct mlx5_ib_qp *qp = to_mqp(ibqp);
1767         enum ib_qp_state cur_state, new_state;
1768         int err = -EINVAL;
1769         int port;
1770
1771         mutex_lock(&qp->mutex);
1772
1773         cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1774         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1775
1776         if (ibqp->qp_type != MLX5_IB_QPT_REG_UMR &&
1777             !ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask,
1778                                 IB_LINK_LAYER_UNSPECIFIED))
1779                 goto out;
1780
1781         if ((attr_mask & IB_QP_PORT) &&
1782             (attr->port_num == 0 ||
1783              attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports)))
1784                 goto out;
1785
1786         if (attr_mask & IB_QP_PKEY_INDEX) {
1787                 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1788                 if (attr->pkey_index >=
1789                     dev->mdev->port_caps[port - 1].pkey_table_len)
1790                         goto out;
1791         }
1792
1793         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1794             attr->max_rd_atomic >
1795             (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp)))
1796                 goto out;
1797
1798         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1799             attr->max_dest_rd_atomic >
1800             (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp)))
1801                 goto out;
1802
1803         if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1804                 err = 0;
1805                 goto out;
1806         }
1807
1808         err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1809
1810 out:
1811         mutex_unlock(&qp->mutex);
1812         return err;
1813 }
1814
1815 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1816 {
1817         struct mlx5_ib_cq *cq;
1818         unsigned cur;
1819
1820         cur = wq->head - wq->tail;
1821         if (likely(cur + nreq < wq->max_post))
1822                 return 0;
1823
1824         cq = to_mcq(ib_cq);
1825         spin_lock(&cq->lock);
1826         cur = wq->head - wq->tail;
1827         spin_unlock(&cq->lock);
1828
1829         return cur + nreq >= wq->max_post;
1830 }
1831
1832 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
1833                                           u64 remote_addr, u32 rkey)
1834 {
1835         rseg->raddr    = cpu_to_be64(remote_addr);
1836         rseg->rkey     = cpu_to_be32(rkey);
1837         rseg->reserved = 0;
1838 }
1839
1840 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
1841                              struct ib_send_wr *wr)
1842 {
1843         memcpy(&dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof(struct mlx5_av));
1844         dseg->av.dqp_dct = cpu_to_be32(wr->wr.ud.remote_qpn | MLX5_EXTENDED_UD_AV);
1845         dseg->av.key.qkey.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1846 }
1847
1848 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
1849 {
1850         dseg->byte_count = cpu_to_be32(sg->length);
1851         dseg->lkey       = cpu_to_be32(sg->lkey);
1852         dseg->addr       = cpu_to_be64(sg->addr);
1853 }
1854
1855 static __be16 get_klm_octo(int npages)
1856 {
1857         return cpu_to_be16(ALIGN(npages, 8) / 2);
1858 }
1859
1860 static __be64 frwr_mkey_mask(void)
1861 {
1862         u64 result;
1863
1864         result = MLX5_MKEY_MASK_LEN             |
1865                 MLX5_MKEY_MASK_PAGE_SIZE        |
1866                 MLX5_MKEY_MASK_START_ADDR       |
1867                 MLX5_MKEY_MASK_EN_RINVAL        |
1868                 MLX5_MKEY_MASK_KEY              |
1869                 MLX5_MKEY_MASK_LR               |
1870                 MLX5_MKEY_MASK_LW               |
1871                 MLX5_MKEY_MASK_RR               |
1872                 MLX5_MKEY_MASK_RW               |
1873                 MLX5_MKEY_MASK_A                |
1874                 MLX5_MKEY_MASK_SMALL_FENCE      |
1875                 MLX5_MKEY_MASK_FREE;
1876
1877         return cpu_to_be64(result);
1878 }
1879
1880 static __be64 sig_mkey_mask(void)
1881 {
1882         u64 result;
1883
1884         result = MLX5_MKEY_MASK_LEN             |
1885                 MLX5_MKEY_MASK_PAGE_SIZE        |
1886                 MLX5_MKEY_MASK_START_ADDR       |
1887                 MLX5_MKEY_MASK_EN_SIGERR        |
1888                 MLX5_MKEY_MASK_EN_RINVAL        |
1889                 MLX5_MKEY_MASK_KEY              |
1890                 MLX5_MKEY_MASK_LR               |
1891                 MLX5_MKEY_MASK_LW               |
1892                 MLX5_MKEY_MASK_RR               |
1893                 MLX5_MKEY_MASK_RW               |
1894                 MLX5_MKEY_MASK_SMALL_FENCE      |
1895                 MLX5_MKEY_MASK_FREE             |
1896                 MLX5_MKEY_MASK_BSF_EN;
1897
1898         return cpu_to_be64(result);
1899 }
1900
1901 static void set_frwr_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
1902                                  struct ib_send_wr *wr, int li)
1903 {
1904         memset(umr, 0, sizeof(*umr));
1905
1906         if (li) {
1907                 umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
1908                 umr->flags = 1 << 7;
1909                 return;
1910         }
1911
1912         umr->flags = (1 << 5); /* fail if not free */
1913         umr->klm_octowords = get_klm_octo(wr->wr.fast_reg.page_list_len);
1914         umr->mkey_mask = frwr_mkey_mask();
1915 }
1916
1917 static __be64 get_umr_reg_mr_mask(void)
1918 {
1919         u64 result;
1920
1921         result = MLX5_MKEY_MASK_LEN             |
1922                  MLX5_MKEY_MASK_PAGE_SIZE       |
1923                  MLX5_MKEY_MASK_START_ADDR      |
1924                  MLX5_MKEY_MASK_PD              |
1925                  MLX5_MKEY_MASK_LR              |
1926                  MLX5_MKEY_MASK_LW              |
1927                  MLX5_MKEY_MASK_KEY             |
1928                  MLX5_MKEY_MASK_RR              |
1929                  MLX5_MKEY_MASK_RW              |
1930                  MLX5_MKEY_MASK_A               |
1931                  MLX5_MKEY_MASK_FREE;
1932
1933         return cpu_to_be64(result);
1934 }
1935
1936 static __be64 get_umr_unreg_mr_mask(void)
1937 {
1938         u64 result;
1939
1940         result = MLX5_MKEY_MASK_FREE;
1941
1942         return cpu_to_be64(result);
1943 }
1944
1945 static __be64 get_umr_update_mtt_mask(void)
1946 {
1947         u64 result;
1948
1949         result = MLX5_MKEY_MASK_FREE;
1950
1951         return cpu_to_be64(result);
1952 }
1953
1954 static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
1955                                 struct ib_send_wr *wr)
1956 {
1957         struct mlx5_umr_wr *umrwr = (struct mlx5_umr_wr *)&wr->wr.fast_reg;
1958
1959         memset(umr, 0, sizeof(*umr));
1960
1961         if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
1962                 umr->flags = MLX5_UMR_CHECK_FREE; /* fail if free */
1963         else
1964                 umr->flags = MLX5_UMR_CHECK_NOT_FREE; /* fail if not free */
1965
1966         if (!(wr->send_flags & MLX5_IB_SEND_UMR_UNREG)) {
1967                 umr->klm_octowords = get_klm_octo(umrwr->npages);
1968                 if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT) {
1969                         umr->mkey_mask = get_umr_update_mtt_mask();
1970                         umr->bsf_octowords = get_klm_octo(umrwr->target.offset);
1971                         umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
1972                 } else {
1973                         umr->mkey_mask = get_umr_reg_mr_mask();
1974                 }
1975         } else {
1976                 umr->mkey_mask = get_umr_unreg_mr_mask();
1977         }
1978
1979         if (!wr->num_sge)
1980                 umr->flags |= MLX5_UMR_INLINE;
1981 }
1982
1983 static u8 get_umr_flags(int acc)
1984 {
1985         return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
1986                (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
1987                (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
1988                (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
1989                 MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
1990 }
1991
1992 static void set_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr,
1993                              int li, int *writ)
1994 {
1995         memset(seg, 0, sizeof(*seg));
1996         if (li) {
1997                 seg->status = MLX5_MKEY_STATUS_FREE;
1998                 return;
1999         }
2000
2001         seg->flags = get_umr_flags(wr->wr.fast_reg.access_flags) |
2002                      MLX5_ACCESS_MODE_MTT;
2003         *writ = seg->flags & (MLX5_PERM_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE);
2004         seg->qpn_mkey7_0 = cpu_to_be32((wr->wr.fast_reg.rkey & 0xff) | 0xffffff00);
2005         seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
2006         seg->start_addr = cpu_to_be64(wr->wr.fast_reg.iova_start);
2007         seg->len = cpu_to_be64(wr->wr.fast_reg.length);
2008         seg->xlt_oct_size = cpu_to_be32((wr->wr.fast_reg.page_list_len + 1) / 2);
2009         seg->log2_page_size = wr->wr.fast_reg.page_shift;
2010 }
2011
2012 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr)
2013 {
2014         struct mlx5_umr_wr *umrwr = (struct mlx5_umr_wr *)&wr->wr.fast_reg;
2015
2016         memset(seg, 0, sizeof(*seg));
2017         if (wr->send_flags & MLX5_IB_SEND_UMR_UNREG) {
2018                 seg->status = MLX5_MKEY_STATUS_FREE;
2019                 return;
2020         }
2021
2022         seg->flags = convert_access(umrwr->access_flags);
2023         if (!(wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT)) {
2024                 seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
2025                 seg->start_addr = cpu_to_be64(umrwr->target.virt_addr);
2026         }
2027         seg->len = cpu_to_be64(umrwr->length);
2028         seg->log2_page_size = umrwr->page_shift;
2029         seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
2030                                        mlx5_mkey_variant(umrwr->mkey));
2031 }
2032
2033 static void set_frwr_pages(struct mlx5_wqe_data_seg *dseg,
2034                            struct ib_send_wr *wr,
2035                            struct mlx5_core_dev *mdev,
2036                            struct mlx5_ib_pd *pd,
2037                            int writ)
2038 {
2039         struct mlx5_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
2040         u64 *page_list = wr->wr.fast_reg.page_list->page_list;
2041         u64 perm = MLX5_EN_RD | (writ ? MLX5_EN_WR : 0);
2042         int i;
2043
2044         for (i = 0; i < wr->wr.fast_reg.page_list_len; i++)
2045                 mfrpl->mapped_page_list[i] = cpu_to_be64(page_list[i] | perm);
2046         dseg->addr = cpu_to_be64(mfrpl->map);
2047         dseg->byte_count = cpu_to_be32(ALIGN(sizeof(u64) * wr->wr.fast_reg.page_list_len, 64));
2048         dseg->lkey = cpu_to_be32(pd->pa_lkey);
2049 }
2050
2051 static __be32 send_ieth(struct ib_send_wr *wr)
2052 {
2053         switch (wr->opcode) {
2054         case IB_WR_SEND_WITH_IMM:
2055         case IB_WR_RDMA_WRITE_WITH_IMM:
2056                 return wr->ex.imm_data;
2057
2058         case IB_WR_SEND_WITH_INV:
2059                 return cpu_to_be32(wr->ex.invalidate_rkey);
2060
2061         default:
2062                 return 0;
2063         }
2064 }
2065
2066 static u8 calc_sig(void *wqe, int size)
2067 {
2068         u8 *p = wqe;
2069         u8 res = 0;
2070         int i;
2071
2072         for (i = 0; i < size; i++)
2073                 res ^= p[i];
2074
2075         return ~res;
2076 }
2077
2078 static u8 wq_sig(void *wqe)
2079 {
2080         return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
2081 }
2082
2083 static int set_data_inl_seg(struct mlx5_ib_qp *qp, struct ib_send_wr *wr,
2084                             void *wqe, int *sz)
2085 {
2086         struct mlx5_wqe_inline_seg *seg;
2087         void *qend = qp->sq.qend;
2088         void *addr;
2089         int inl = 0;
2090         int copy;
2091         int len;
2092         int i;
2093
2094         seg = wqe;
2095         wqe += sizeof(*seg);
2096         for (i = 0; i < wr->num_sge; i++) {
2097                 addr = (void *)(unsigned long)(wr->sg_list[i].addr);
2098                 len  = wr->sg_list[i].length;
2099                 inl += len;
2100
2101                 if (unlikely(inl > qp->max_inline_data))
2102                         return -ENOMEM;
2103
2104                 if (unlikely(wqe + len > qend)) {
2105                         copy = qend - wqe;
2106                         memcpy(wqe, addr, copy);
2107                         addr += copy;
2108                         len -= copy;
2109                         wqe = mlx5_get_send_wqe(qp, 0);
2110                 }
2111                 memcpy(wqe, addr, len);
2112                 wqe += len;
2113         }
2114
2115         seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
2116
2117         *sz = ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
2118
2119         return 0;
2120 }
2121
2122 static u16 prot_field_size(enum ib_signature_type type)
2123 {
2124         switch (type) {
2125         case IB_SIG_TYPE_T10_DIF:
2126                 return MLX5_DIF_SIZE;
2127         default:
2128                 return 0;
2129         }
2130 }
2131
2132 static u8 bs_selector(int block_size)
2133 {
2134         switch (block_size) {
2135         case 512:           return 0x1;
2136         case 520:           return 0x2;
2137         case 4096:          return 0x3;
2138         case 4160:          return 0x4;
2139         case 1073741824:    return 0x5;
2140         default:            return 0;
2141         }
2142 }
2143
2144 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
2145                               struct mlx5_bsf_inl *inl)
2146 {
2147         /* Valid inline section and allow BSF refresh */
2148         inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
2149                                        MLX5_BSF_REFRESH_DIF);
2150         inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
2151         inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
2152         /* repeating block */
2153         inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
2154         inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
2155                         MLX5_DIF_CRC : MLX5_DIF_IPCS;
2156
2157         if (domain->sig.dif.ref_remap)
2158                 inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
2159
2160         if (domain->sig.dif.app_escape) {
2161                 if (domain->sig.dif.ref_escape)
2162                         inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
2163                 else
2164                         inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
2165         }
2166
2167         inl->dif_app_bitmask_check =
2168                 cpu_to_be16(domain->sig.dif.apptag_check_mask);
2169 }
2170
2171 static int mlx5_set_bsf(struct ib_mr *sig_mr,
2172                         struct ib_sig_attrs *sig_attrs,
2173                         struct mlx5_bsf *bsf, u32 data_size)
2174 {
2175         struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
2176         struct mlx5_bsf_basic *basic = &bsf->basic;
2177         struct ib_sig_domain *mem = &sig_attrs->mem;
2178         struct ib_sig_domain *wire = &sig_attrs->wire;
2179
2180         memset(bsf, 0, sizeof(*bsf));
2181
2182         /* Basic + Extended + Inline */
2183         basic->bsf_size_sbs = 1 << 7;
2184         /* Input domain check byte mask */
2185         basic->check_byte_mask = sig_attrs->check_mask;
2186         basic->raw_data_size = cpu_to_be32(data_size);
2187
2188         /* Memory domain */
2189         switch (sig_attrs->mem.sig_type) {
2190         case IB_SIG_TYPE_NONE:
2191                 break;
2192         case IB_SIG_TYPE_T10_DIF:
2193                 basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
2194                 basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
2195                 mlx5_fill_inl_bsf(mem, &bsf->m_inl);
2196                 break;
2197         default:
2198                 return -EINVAL;
2199         }
2200
2201         /* Wire domain */
2202         switch (sig_attrs->wire.sig_type) {
2203         case IB_SIG_TYPE_NONE:
2204                 break;
2205         case IB_SIG_TYPE_T10_DIF:
2206                 if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
2207                     mem->sig_type == wire->sig_type) {
2208                         /* Same block structure */
2209                         basic->bsf_size_sbs |= 1 << 4;
2210                         if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
2211                                 basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
2212                         if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
2213                                 basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
2214                         if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
2215                                 basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
2216                 } else
2217                         basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
2218
2219                 basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
2220                 mlx5_fill_inl_bsf(wire, &bsf->w_inl);
2221                 break;
2222         default:
2223                 return -EINVAL;
2224         }
2225
2226         return 0;
2227 }
2228
2229 static int set_sig_data_segment(struct ib_send_wr *wr, struct mlx5_ib_qp *qp,
2230                                 void **seg, int *size)
2231 {
2232         struct ib_sig_attrs *sig_attrs = wr->wr.sig_handover.sig_attrs;
2233         struct ib_mr *sig_mr = wr->wr.sig_handover.sig_mr;
2234         struct mlx5_bsf *bsf;
2235         u32 data_len = wr->sg_list->length;
2236         u32 data_key = wr->sg_list->lkey;
2237         u64 data_va = wr->sg_list->addr;
2238         int ret;
2239         int wqe_size;
2240
2241         if (!wr->wr.sig_handover.prot ||
2242             (data_key == wr->wr.sig_handover.prot->lkey &&
2243              data_va == wr->wr.sig_handover.prot->addr &&
2244              data_len == wr->wr.sig_handover.prot->length)) {
2245                 /**
2246                  * Source domain doesn't contain signature information
2247                  * or data and protection are interleaved in memory.
2248                  * So need construct:
2249                  *                  ------------------
2250                  *                 |     data_klm     |
2251                  *                  ------------------
2252                  *                 |       BSF        |
2253                  *                  ------------------
2254                  **/
2255                 struct mlx5_klm *data_klm = *seg;
2256
2257                 data_klm->bcount = cpu_to_be32(data_len);
2258                 data_klm->key = cpu_to_be32(data_key);
2259                 data_klm->va = cpu_to_be64(data_va);
2260                 wqe_size = ALIGN(sizeof(*data_klm), 64);
2261         } else {
2262                 /**
2263                  * Source domain contains signature information
2264                  * So need construct a strided block format:
2265                  *               ---------------------------
2266                  *              |     stride_block_ctrl     |
2267                  *               ---------------------------
2268                  *              |          data_klm         |
2269                  *               ---------------------------
2270                  *              |          prot_klm         |
2271                  *               ---------------------------
2272                  *              |             BSF           |
2273                  *               ---------------------------
2274                  **/
2275                 struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
2276                 struct mlx5_stride_block_entry *data_sentry;
2277                 struct mlx5_stride_block_entry *prot_sentry;
2278                 u32 prot_key = wr->wr.sig_handover.prot->lkey;
2279                 u64 prot_va = wr->wr.sig_handover.prot->addr;
2280                 u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
2281                 int prot_size;
2282
2283                 sblock_ctrl = *seg;
2284                 data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
2285                 prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
2286
2287                 prot_size = prot_field_size(sig_attrs->mem.sig_type);
2288                 if (!prot_size) {
2289                         pr_err("Bad block size given: %u\n", block_size);
2290                         return -EINVAL;
2291                 }
2292                 sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
2293                                                             prot_size);
2294                 sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
2295                 sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
2296                 sblock_ctrl->num_entries = cpu_to_be16(2);
2297
2298                 data_sentry->bcount = cpu_to_be16(block_size);
2299                 data_sentry->key = cpu_to_be32(data_key);
2300                 data_sentry->va = cpu_to_be64(data_va);
2301                 data_sentry->stride = cpu_to_be16(block_size);
2302
2303                 prot_sentry->bcount = cpu_to_be16(prot_size);
2304                 prot_sentry->key = cpu_to_be32(prot_key);
2305                 prot_sentry->va = cpu_to_be64(prot_va);
2306                 prot_sentry->stride = cpu_to_be16(prot_size);
2307
2308                 wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
2309                                  sizeof(*prot_sentry), 64);
2310         }
2311
2312         *seg += wqe_size;
2313         *size += wqe_size / 16;
2314         if (unlikely((*seg == qp->sq.qend)))
2315                 *seg = mlx5_get_send_wqe(qp, 0);
2316
2317         bsf = *seg;
2318         ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
2319         if (ret)
2320                 return -EINVAL;
2321
2322         *seg += sizeof(*bsf);
2323         *size += sizeof(*bsf) / 16;
2324         if (unlikely((*seg == qp->sq.qend)))
2325                 *seg = mlx5_get_send_wqe(qp, 0);
2326
2327         return 0;
2328 }
2329
2330 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
2331                                  struct ib_send_wr *wr, u32 nelements,
2332                                  u32 length, u32 pdn)
2333 {
2334         struct ib_mr *sig_mr = wr->wr.sig_handover.sig_mr;
2335         u32 sig_key = sig_mr->rkey;
2336         u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
2337
2338         memset(seg, 0, sizeof(*seg));
2339
2340         seg->flags = get_umr_flags(wr->wr.sig_handover.access_flags) |
2341                                    MLX5_ACCESS_MODE_KLM;
2342         seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
2343         seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
2344                                     MLX5_MKEY_BSF_EN | pdn);
2345         seg->len = cpu_to_be64(length);
2346         seg->xlt_oct_size = cpu_to_be32(be16_to_cpu(get_klm_octo(nelements)));
2347         seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
2348 }
2349
2350 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
2351                                 struct ib_send_wr *wr, u32 nelements)
2352 {
2353         memset(umr, 0, sizeof(*umr));
2354
2355         umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
2356         umr->klm_octowords = get_klm_octo(nelements);
2357         umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
2358         umr->mkey_mask = sig_mkey_mask();
2359 }
2360
2361
2362 static int set_sig_umr_wr(struct ib_send_wr *wr, struct mlx5_ib_qp *qp,
2363                           void **seg, int *size)
2364 {
2365         struct mlx5_ib_mr *sig_mr = to_mmr(wr->wr.sig_handover.sig_mr);
2366         u32 pdn = get_pd(qp)->pdn;
2367         u32 klm_oct_size;
2368         int region_len, ret;
2369
2370         if (unlikely(wr->num_sge != 1) ||
2371             unlikely(wr->wr.sig_handover.access_flags &
2372                      IB_ACCESS_REMOTE_ATOMIC) ||
2373             unlikely(!sig_mr->sig) || unlikely(!qp->signature_en) ||
2374             unlikely(!sig_mr->sig->sig_status_checked))
2375                 return -EINVAL;
2376
2377         /* length of the protected region, data + protection */
2378         region_len = wr->sg_list->length;
2379         if (wr->wr.sig_handover.prot &&
2380             (wr->wr.sig_handover.prot->lkey != wr->sg_list->lkey  ||
2381              wr->wr.sig_handover.prot->addr != wr->sg_list->addr  ||
2382              wr->wr.sig_handover.prot->length != wr->sg_list->length))
2383                 region_len += wr->wr.sig_handover.prot->length;
2384
2385         /**
2386          * KLM octoword size - if protection was provided
2387          * then we use strided block format (3 octowords),
2388          * else we use single KLM (1 octoword)
2389          **/
2390         klm_oct_size = wr->wr.sig_handover.prot ? 3 : 1;
2391
2392         set_sig_umr_segment(*seg, wr, klm_oct_size);
2393         *seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2394         *size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2395         if (unlikely((*seg == qp->sq.qend)))
2396                 *seg = mlx5_get_send_wqe(qp, 0);
2397
2398         set_sig_mkey_segment(*seg, wr, klm_oct_size, region_len, pdn);
2399         *seg += sizeof(struct mlx5_mkey_seg);
2400         *size += sizeof(struct mlx5_mkey_seg) / 16;
2401         if (unlikely((*seg == qp->sq.qend)))
2402                 *seg = mlx5_get_send_wqe(qp, 0);
2403
2404         ret = set_sig_data_segment(wr, qp, seg, size);
2405         if (ret)
2406                 return ret;
2407
2408         sig_mr->sig->sig_status_checked = false;
2409         return 0;
2410 }
2411
2412 static int set_psv_wr(struct ib_sig_domain *domain,
2413                       u32 psv_idx, void **seg, int *size)
2414 {
2415         struct mlx5_seg_set_psv *psv_seg = *seg;
2416
2417         memset(psv_seg, 0, sizeof(*psv_seg));
2418         psv_seg->psv_num = cpu_to_be32(psv_idx);
2419         switch (domain->sig_type) {
2420         case IB_SIG_TYPE_NONE:
2421                 break;
2422         case IB_SIG_TYPE_T10_DIF:
2423                 psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
2424                                                      domain->sig.dif.app_tag);
2425                 psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
2426                 break;
2427         default:
2428                 pr_err("Bad signature type given.\n");
2429                 return 1;
2430         }
2431
2432         *seg += sizeof(*psv_seg);
2433         *size += sizeof(*psv_seg) / 16;
2434
2435         return 0;
2436 }
2437
2438 static int set_frwr_li_wr(void **seg, struct ib_send_wr *wr, int *size,
2439                           struct mlx5_core_dev *mdev, struct mlx5_ib_pd *pd, struct mlx5_ib_qp *qp)
2440 {
2441         int writ = 0;
2442         int li;
2443
2444         li = wr->opcode == IB_WR_LOCAL_INV ? 1 : 0;
2445         if (unlikely(wr->send_flags & IB_SEND_INLINE))
2446                 return -EINVAL;
2447
2448         set_frwr_umr_segment(*seg, wr, li);
2449         *seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2450         *size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2451         if (unlikely((*seg == qp->sq.qend)))
2452                 *seg = mlx5_get_send_wqe(qp, 0);
2453         set_mkey_segment(*seg, wr, li, &writ);
2454         *seg += sizeof(struct mlx5_mkey_seg);
2455         *size += sizeof(struct mlx5_mkey_seg) / 16;
2456         if (unlikely((*seg == qp->sq.qend)))
2457                 *seg = mlx5_get_send_wqe(qp, 0);
2458         if (!li) {
2459                 if (unlikely(wr->wr.fast_reg.page_list_len >
2460                              wr->wr.fast_reg.page_list->max_page_list_len))
2461                         return  -ENOMEM;
2462
2463                 set_frwr_pages(*seg, wr, mdev, pd, writ);
2464                 *seg += sizeof(struct mlx5_wqe_data_seg);
2465                 *size += (sizeof(struct mlx5_wqe_data_seg) / 16);
2466         }
2467         return 0;
2468 }
2469
2470 static void dump_wqe(struct mlx5_ib_qp *qp, int idx, int size_16)
2471 {
2472         __be32 *p = NULL;
2473         int tidx = idx;
2474         int i, j;
2475
2476         pr_debug("dump wqe at %p\n", mlx5_get_send_wqe(qp, tidx));
2477         for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
2478                 if ((i & 0xf) == 0) {
2479                         void *buf = mlx5_get_send_wqe(qp, tidx);
2480                         tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
2481                         p = buf;
2482                         j = 0;
2483                 }
2484                 pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
2485                          be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
2486                          be32_to_cpu(p[j + 3]));
2487         }
2488 }
2489
2490 static void mlx5_bf_copy(u64 __iomem *dst, u64 *src,
2491                          unsigned bytecnt, struct mlx5_ib_qp *qp)
2492 {
2493         while (bytecnt > 0) {
2494                 __iowrite64_copy(dst++, src++, 8);
2495                 __iowrite64_copy(dst++, src++, 8);
2496                 __iowrite64_copy(dst++, src++, 8);
2497                 __iowrite64_copy(dst++, src++, 8);
2498                 __iowrite64_copy(dst++, src++, 8);
2499                 __iowrite64_copy(dst++, src++, 8);
2500                 __iowrite64_copy(dst++, src++, 8);
2501                 __iowrite64_copy(dst++, src++, 8);
2502                 bytecnt -= 64;
2503                 if (unlikely(src == qp->sq.qend))
2504                         src = mlx5_get_send_wqe(qp, 0);
2505         }
2506 }
2507
2508 static u8 get_fence(u8 fence, struct ib_send_wr *wr)
2509 {
2510         if (unlikely(wr->opcode == IB_WR_LOCAL_INV &&
2511                      wr->send_flags & IB_SEND_FENCE))
2512                 return MLX5_FENCE_MODE_STRONG_ORDERING;
2513
2514         if (unlikely(fence)) {
2515                 if (wr->send_flags & IB_SEND_FENCE)
2516                         return MLX5_FENCE_MODE_SMALL_AND_FENCE;
2517                 else
2518                         return fence;
2519
2520         } else {
2521                 return 0;
2522         }
2523 }
2524
2525 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
2526                      struct mlx5_wqe_ctrl_seg **ctrl,
2527                      struct ib_send_wr *wr, unsigned *idx,
2528                      int *size, int nreq)
2529 {
2530         int err = 0;
2531
2532         if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq))) {
2533                 err = -ENOMEM;
2534                 return err;
2535         }
2536
2537         *idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
2538         *seg = mlx5_get_send_wqe(qp, *idx);
2539         *ctrl = *seg;
2540         *(uint32_t *)(*seg + 8) = 0;
2541         (*ctrl)->imm = send_ieth(wr);
2542         (*ctrl)->fm_ce_se = qp->sq_signal_bits |
2543                 (wr->send_flags & IB_SEND_SIGNALED ?
2544                  MLX5_WQE_CTRL_CQ_UPDATE : 0) |
2545                 (wr->send_flags & IB_SEND_SOLICITED ?
2546                  MLX5_WQE_CTRL_SOLICITED : 0);
2547
2548         *seg += sizeof(**ctrl);
2549         *size = sizeof(**ctrl) / 16;
2550
2551         return err;
2552 }
2553
2554 static void finish_wqe(struct mlx5_ib_qp *qp,
2555                        struct mlx5_wqe_ctrl_seg *ctrl,
2556                        u8 size, unsigned idx, u64 wr_id,
2557                        int nreq, u8 fence, u8 next_fence,
2558                        u32 mlx5_opcode)
2559 {
2560         u8 opmod = 0;
2561
2562         ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
2563                                              mlx5_opcode | ((u32)opmod << 24));
2564         ctrl->qpn_ds = cpu_to_be32(size | (qp->mqp.qpn << 8));
2565         ctrl->fm_ce_se |= fence;
2566         qp->fm_cache = next_fence;
2567         if (unlikely(qp->wq_sig))
2568                 ctrl->signature = wq_sig(ctrl);
2569
2570         qp->sq.wrid[idx] = wr_id;
2571         qp->sq.w_list[idx].opcode = mlx5_opcode;
2572         qp->sq.wqe_head[idx] = qp->sq.head + nreq;
2573         qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
2574         qp->sq.w_list[idx].next = qp->sq.cur_post;
2575 }
2576
2577
2578 int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2579                       struct ib_send_wr **bad_wr)
2580 {
2581         struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
2582         struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2583         struct mlx5_core_dev *mdev = dev->mdev;
2584         struct mlx5_ib_qp *qp = to_mqp(ibqp);
2585         struct mlx5_ib_mr *mr;
2586         struct mlx5_wqe_data_seg *dpseg;
2587         struct mlx5_wqe_xrc_seg *xrc;
2588         struct mlx5_bf *bf = qp->bf;
2589         int uninitialized_var(size);
2590         void *qend = qp->sq.qend;
2591         unsigned long flags;
2592         unsigned idx;
2593         int err = 0;
2594         int inl = 0;
2595         int num_sge;
2596         void *seg;
2597         int nreq;
2598         int i;
2599         u8 next_fence = 0;
2600         u8 fence;
2601
2602         spin_lock_irqsave(&qp->sq.lock, flags);
2603
2604         for (nreq = 0; wr; nreq++, wr = wr->next) {
2605                 if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
2606                         mlx5_ib_warn(dev, "\n");
2607                         err = -EINVAL;
2608                         *bad_wr = wr;
2609                         goto out;
2610                 }
2611
2612                 fence = qp->fm_cache;
2613                 num_sge = wr->num_sge;
2614                 if (unlikely(num_sge > qp->sq.max_gs)) {
2615                         mlx5_ib_warn(dev, "\n");
2616                         err = -ENOMEM;
2617                         *bad_wr = wr;
2618                         goto out;
2619                 }
2620
2621                 err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, nreq);
2622                 if (err) {
2623                         mlx5_ib_warn(dev, "\n");
2624                         err = -ENOMEM;
2625                         *bad_wr = wr;
2626                         goto out;
2627                 }
2628
2629                 switch (ibqp->qp_type) {
2630                 case IB_QPT_XRC_INI:
2631                         xrc = seg;
2632                         xrc->xrc_srqn = htonl(wr->xrc_remote_srq_num);
2633                         seg += sizeof(*xrc);
2634                         size += sizeof(*xrc) / 16;
2635                         /* fall through */
2636                 case IB_QPT_RC:
2637                         switch (wr->opcode) {
2638                         case IB_WR_RDMA_READ:
2639                         case IB_WR_RDMA_WRITE:
2640                         case IB_WR_RDMA_WRITE_WITH_IMM:
2641                                 set_raddr_seg(seg, wr->wr.rdma.remote_addr,
2642                                               wr->wr.rdma.rkey);
2643                                 seg += sizeof(struct mlx5_wqe_raddr_seg);
2644                                 size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
2645                                 break;
2646
2647                         case IB_WR_ATOMIC_CMP_AND_SWP:
2648                         case IB_WR_ATOMIC_FETCH_AND_ADD:
2649                         case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2650                                 mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
2651                                 err = -ENOSYS;
2652                                 *bad_wr = wr;
2653                                 goto out;
2654
2655                         case IB_WR_LOCAL_INV:
2656                                 next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2657                                 qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
2658                                 ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
2659                                 err = set_frwr_li_wr(&seg, wr, &size, mdev, to_mpd(ibqp->pd), qp);
2660                                 if (err) {
2661                                         mlx5_ib_warn(dev, "\n");
2662                                         *bad_wr = wr;
2663                                         goto out;
2664                                 }
2665                                 num_sge = 0;
2666                                 break;
2667
2668                         case IB_WR_FAST_REG_MR:
2669                                 next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2670                                 qp->sq.wr_data[idx] = IB_WR_FAST_REG_MR;
2671                                 ctrl->imm = cpu_to_be32(wr->wr.fast_reg.rkey);
2672                                 err = set_frwr_li_wr(&seg, wr, &size, mdev, to_mpd(ibqp->pd), qp);
2673                                 if (err) {
2674                                         mlx5_ib_warn(dev, "\n");
2675                                         *bad_wr = wr;
2676                                         goto out;
2677                                 }
2678                                 num_sge = 0;
2679                                 break;
2680
2681                         case IB_WR_REG_SIG_MR:
2682                                 qp->sq.wr_data[idx] = IB_WR_REG_SIG_MR;
2683                                 mr = to_mmr(wr->wr.sig_handover.sig_mr);
2684
2685                                 ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
2686                                 err = set_sig_umr_wr(wr, qp, &seg, &size);
2687                                 if (err) {
2688                                         mlx5_ib_warn(dev, "\n");
2689                                         *bad_wr = wr;
2690                                         goto out;
2691                                 }
2692
2693                                 finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2694                                            nreq, get_fence(fence, wr),
2695                                            next_fence, MLX5_OPCODE_UMR);
2696                                 /*
2697                                  * SET_PSV WQEs are not signaled and solicited
2698                                  * on error
2699                                  */
2700                                 wr->send_flags &= ~IB_SEND_SIGNALED;
2701                                 wr->send_flags |= IB_SEND_SOLICITED;
2702                                 err = begin_wqe(qp, &seg, &ctrl, wr,
2703                                                 &idx, &size, nreq);
2704                                 if (err) {
2705                                         mlx5_ib_warn(dev, "\n");
2706                                         err = -ENOMEM;
2707                                         *bad_wr = wr;
2708                                         goto out;
2709                                 }
2710
2711                                 err = set_psv_wr(&wr->wr.sig_handover.sig_attrs->mem,
2712                                                  mr->sig->psv_memory.psv_idx, &seg,
2713                                                  &size);
2714                                 if (err) {
2715                                         mlx5_ib_warn(dev, "\n");
2716                                         *bad_wr = wr;
2717                                         goto out;
2718                                 }
2719
2720                                 finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2721                                            nreq, get_fence(fence, wr),
2722                                            next_fence, MLX5_OPCODE_SET_PSV);
2723                                 err = begin_wqe(qp, &seg, &ctrl, wr,
2724                                                 &idx, &size, nreq);
2725                                 if (err) {
2726                                         mlx5_ib_warn(dev, "\n");
2727                                         err = -ENOMEM;
2728                                         *bad_wr = wr;
2729                                         goto out;
2730                                 }
2731
2732                                 next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2733                                 err = set_psv_wr(&wr->wr.sig_handover.sig_attrs->wire,
2734                                                  mr->sig->psv_wire.psv_idx, &seg,
2735                                                  &size);
2736                                 if (err) {
2737                                         mlx5_ib_warn(dev, "\n");
2738                                         *bad_wr = wr;
2739                                         goto out;
2740                                 }
2741
2742                                 finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2743                                            nreq, get_fence(fence, wr),
2744                                            next_fence, MLX5_OPCODE_SET_PSV);
2745                                 num_sge = 0;
2746                                 goto skip_psv;
2747
2748                         default:
2749                                 break;
2750                         }
2751                         break;
2752
2753                 case IB_QPT_UC:
2754                         switch (wr->opcode) {
2755                         case IB_WR_RDMA_WRITE:
2756                         case IB_WR_RDMA_WRITE_WITH_IMM:
2757                                 set_raddr_seg(seg, wr->wr.rdma.remote_addr,
2758                                               wr->wr.rdma.rkey);
2759                                 seg  += sizeof(struct mlx5_wqe_raddr_seg);
2760                                 size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
2761                                 break;
2762
2763                         default:
2764                                 break;
2765                         }
2766                         break;
2767
2768                 case IB_QPT_UD:
2769                 case IB_QPT_SMI:
2770                 case IB_QPT_GSI:
2771                         set_datagram_seg(seg, wr);
2772                         seg += sizeof(struct mlx5_wqe_datagram_seg);
2773                         size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
2774                         if (unlikely((seg == qend)))
2775                                 seg = mlx5_get_send_wqe(qp, 0);
2776                         break;
2777
2778                 case MLX5_IB_QPT_REG_UMR:
2779                         if (wr->opcode != MLX5_IB_WR_UMR) {
2780                                 err = -EINVAL;
2781                                 mlx5_ib_warn(dev, "bad opcode\n");
2782                                 goto out;
2783                         }
2784                         qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
2785                         ctrl->imm = cpu_to_be32(wr->wr.fast_reg.rkey);
2786                         set_reg_umr_segment(seg, wr);
2787                         seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2788                         size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2789                         if (unlikely((seg == qend)))
2790                                 seg = mlx5_get_send_wqe(qp, 0);
2791                         set_reg_mkey_segment(seg, wr);
2792                         seg += sizeof(struct mlx5_mkey_seg);
2793                         size += sizeof(struct mlx5_mkey_seg) / 16;
2794                         if (unlikely((seg == qend)))
2795                                 seg = mlx5_get_send_wqe(qp, 0);
2796                         break;
2797
2798                 default:
2799                         break;
2800                 }
2801
2802                 if (wr->send_flags & IB_SEND_INLINE && num_sge) {
2803                         int uninitialized_var(sz);
2804
2805                         err = set_data_inl_seg(qp, wr, seg, &sz);
2806                         if (unlikely(err)) {
2807                                 mlx5_ib_warn(dev, "\n");
2808                                 *bad_wr = wr;
2809                                 goto out;
2810                         }
2811                         inl = 1;
2812                         size += sz;
2813                 } else {
2814                         dpseg = seg;
2815                         for (i = 0; i < num_sge; i++) {
2816                                 if (unlikely(dpseg == qend)) {
2817                                         seg = mlx5_get_send_wqe(qp, 0);
2818                                         dpseg = seg;
2819                                 }
2820                                 if (likely(wr->sg_list[i].length)) {
2821                                         set_data_ptr_seg(dpseg, wr->sg_list + i);
2822                                         size += sizeof(struct mlx5_wqe_data_seg) / 16;
2823                                         dpseg++;
2824                                 }
2825                         }
2826                 }
2827
2828                 finish_wqe(qp, ctrl, size, idx, wr->wr_id, nreq,
2829                            get_fence(fence, wr), next_fence,
2830                            mlx5_ib_opcode[wr->opcode]);
2831 skip_psv:
2832                 if (0)
2833                         dump_wqe(qp, idx, size);
2834         }
2835
2836 out:
2837         if (likely(nreq)) {
2838                 qp->sq.head += nreq;
2839
2840                 /* Make sure that descriptors are written before
2841                  * updating doorbell record and ringing the doorbell
2842                  */
2843                 wmb();
2844
2845                 qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
2846
2847                 /* Make sure doorbell record is visible to the HCA before
2848                  * we hit doorbell */
2849                 wmb();
2850
2851                 if (bf->need_lock)
2852                         spin_lock(&bf->lock);
2853                 else
2854                         __acquire(&bf->lock);
2855
2856                 /* TBD enable WC */
2857                 if (0 && nreq == 1 && bf->uuarn && inl && size > 1 && size <= bf->buf_size / 16) {
2858                         mlx5_bf_copy(bf->reg + bf->offset, (u64 *)ctrl, ALIGN(size * 16, 64), qp);
2859                         /* wc_wmb(); */
2860                 } else {
2861                         mlx5_write64((__be32 *)ctrl, bf->regreg + bf->offset,
2862                                      MLX5_GET_DOORBELL_LOCK(&bf->lock32));
2863                         /* Make sure doorbells don't leak out of SQ spinlock
2864                          * and reach the HCA out of order.
2865                          */
2866                         mmiowb();
2867                 }
2868                 bf->offset ^= bf->buf_size;
2869                 if (bf->need_lock)
2870                         spin_unlock(&bf->lock);
2871                 else
2872                         __release(&bf->lock);
2873         }
2874
2875         spin_unlock_irqrestore(&qp->sq.lock, flags);
2876
2877         return err;
2878 }
2879
2880 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
2881 {
2882         sig->signature = calc_sig(sig, size);
2883 }
2884
2885 int mlx5_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2886                       struct ib_recv_wr **bad_wr)
2887 {
2888         struct mlx5_ib_qp *qp = to_mqp(ibqp);
2889         struct mlx5_wqe_data_seg *scat;
2890         struct mlx5_rwqe_sig *sig;
2891         unsigned long flags;
2892         int err = 0;
2893         int nreq;
2894         int ind;
2895         int i;
2896
2897         spin_lock_irqsave(&qp->rq.lock, flags);
2898
2899         ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
2900
2901         for (nreq = 0; wr; nreq++, wr = wr->next) {
2902                 if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2903                         err = -ENOMEM;
2904                         *bad_wr = wr;
2905                         goto out;
2906                 }
2907
2908                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2909                         err = -EINVAL;
2910                         *bad_wr = wr;
2911                         goto out;
2912                 }
2913
2914                 scat = get_recv_wqe(qp, ind);
2915                 if (qp->wq_sig)
2916                         scat++;
2917
2918                 for (i = 0; i < wr->num_sge; i++)
2919                         set_data_ptr_seg(scat + i, wr->sg_list + i);
2920
2921                 if (i < qp->rq.max_gs) {
2922                         scat[i].byte_count = 0;
2923                         scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
2924                         scat[i].addr       = 0;
2925                 }
2926
2927                 if (qp->wq_sig) {
2928                         sig = (struct mlx5_rwqe_sig *)scat;
2929                         set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
2930                 }
2931
2932                 qp->rq.wrid[ind] = wr->wr_id;
2933
2934                 ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
2935         }
2936
2937 out:
2938         if (likely(nreq)) {
2939                 qp->rq.head += nreq;
2940
2941                 /* Make sure that descriptors are written before
2942                  * doorbell record.
2943                  */
2944                 wmb();
2945
2946                 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
2947         }
2948
2949         spin_unlock_irqrestore(&qp->rq.lock, flags);
2950
2951         return err;
2952 }
2953
2954 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
2955 {
2956         switch (mlx5_state) {
2957         case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
2958         case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
2959         case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
2960         case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
2961         case MLX5_QP_STATE_SQ_DRAINING:
2962         case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
2963         case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
2964         case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
2965         default:                     return -1;
2966         }
2967 }
2968
2969 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
2970 {
2971         switch (mlx5_mig_state) {
2972         case MLX5_QP_PM_ARMED:          return IB_MIG_ARMED;
2973         case MLX5_QP_PM_REARM:          return IB_MIG_REARM;
2974         case MLX5_QP_PM_MIGRATED:       return IB_MIG_MIGRATED;
2975         default: return -1;
2976         }
2977 }
2978
2979 static int to_ib_qp_access_flags(int mlx5_flags)
2980 {
2981         int ib_flags = 0;
2982
2983         if (mlx5_flags & MLX5_QP_BIT_RRE)
2984                 ib_flags |= IB_ACCESS_REMOTE_READ;
2985         if (mlx5_flags & MLX5_QP_BIT_RWE)
2986                 ib_flags |= IB_ACCESS_REMOTE_WRITE;
2987         if (mlx5_flags & MLX5_QP_BIT_RAE)
2988                 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
2989
2990         return ib_flags;
2991 }
2992
2993 static void to_ib_ah_attr(struct mlx5_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
2994                                 struct mlx5_qp_path *path)
2995 {
2996         struct mlx5_core_dev *dev = ibdev->mdev;
2997
2998         memset(ib_ah_attr, 0, sizeof(*ib_ah_attr));
2999         ib_ah_attr->port_num      = path->port;
3000
3001         if (ib_ah_attr->port_num == 0 ||
3002             ib_ah_attr->port_num > MLX5_CAP_GEN(dev, num_ports))
3003                 return;
3004
3005         ib_ah_attr->sl = path->sl & 0xf;
3006
3007         ib_ah_attr->dlid          = be16_to_cpu(path->rlid);
3008         ib_ah_attr->src_path_bits = path->grh_mlid & 0x7f;
3009         ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
3010         ib_ah_attr->ah_flags      = (path->grh_mlid & (1 << 7)) ? IB_AH_GRH : 0;
3011         if (ib_ah_attr->ah_flags) {
3012                 ib_ah_attr->grh.sgid_index = path->mgid_index;
3013                 ib_ah_attr->grh.hop_limit  = path->hop_limit;
3014                 ib_ah_attr->grh.traffic_class =
3015                         (be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3016                 ib_ah_attr->grh.flow_label =
3017                         be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3018                 memcpy(ib_ah_attr->grh.dgid.raw,
3019                        path->rgid, sizeof(ib_ah_attr->grh.dgid.raw));
3020         }
3021 }
3022
3023 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3024                      struct ib_qp_init_attr *qp_init_attr)
3025 {
3026         struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3027         struct mlx5_ib_qp *qp = to_mqp(ibqp);
3028         struct mlx5_query_qp_mbox_out *outb;
3029         struct mlx5_qp_context *context;
3030         int mlx5_state;
3031         int err = 0;
3032
3033 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
3034         /*
3035          * Wait for any outstanding page faults, in case the user frees memory
3036          * based upon this query's result.
3037          */
3038         flush_workqueue(mlx5_ib_page_fault_wq);
3039 #endif
3040
3041         mutex_lock(&qp->mutex);
3042         outb = kzalloc(sizeof(*outb), GFP_KERNEL);
3043         if (!outb) {
3044                 err = -ENOMEM;
3045                 goto out;
3046         }
3047         context = &outb->ctx;
3048         err = mlx5_core_qp_query(dev->mdev, &qp->mqp, outb, sizeof(*outb));
3049         if (err)
3050                 goto out_free;
3051
3052         mlx5_state = be32_to_cpu(context->flags) >> 28;
3053
3054         qp->state                    = to_ib_qp_state(mlx5_state);
3055         qp_attr->qp_state            = qp->state;
3056         qp_attr->path_mtu            = context->mtu_msgmax >> 5;
3057         qp_attr->path_mig_state      =
3058                 to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
3059         qp_attr->qkey                = be32_to_cpu(context->qkey);
3060         qp_attr->rq_psn              = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
3061         qp_attr->sq_psn              = be32_to_cpu(context->next_send_psn) & 0xffffff;
3062         qp_attr->dest_qp_num         = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
3063         qp_attr->qp_access_flags     =
3064                 to_ib_qp_access_flags(be32_to_cpu(context->params2));
3065
3066         if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3067                 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
3068                 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
3069                 qp_attr->alt_pkey_index = context->alt_path.pkey_index & 0x7f;
3070                 qp_attr->alt_port_num   = qp_attr->alt_ah_attr.port_num;
3071         }
3072
3073         qp_attr->pkey_index = context->pri_path.pkey_index & 0x7f;
3074         qp_attr->port_num = context->pri_path.port;
3075
3076         /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3077         qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
3078
3079         qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
3080
3081         qp_attr->max_dest_rd_atomic =
3082                 1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
3083         qp_attr->min_rnr_timer      =
3084                 (be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
3085         qp_attr->timeout            = context->pri_path.ackto_lt >> 3;
3086         qp_attr->retry_cnt          = (be32_to_cpu(context->params1) >> 16) & 0x7;
3087         qp_attr->rnr_retry          = (be32_to_cpu(context->params1) >> 13) & 0x7;
3088         qp_attr->alt_timeout        = context->alt_path.ackto_lt >> 3;
3089         qp_attr->cur_qp_state        = qp_attr->qp_state;
3090         qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
3091         qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
3092
3093         if (!ibqp->uobject) {
3094                 qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
3095                 qp_attr->cap.max_send_sge = qp->sq.max_gs;
3096         } else {
3097                 qp_attr->cap.max_send_wr  = 0;
3098                 qp_attr->cap.max_send_sge = 0;
3099         }
3100
3101         /* We don't support inline sends for kernel QPs (yet), and we
3102          * don't know what userspace's value should be.
3103          */
3104         qp_attr->cap.max_inline_data = 0;
3105
3106         qp_init_attr->cap            = qp_attr->cap;
3107
3108         qp_init_attr->create_flags = 0;
3109         if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3110                 qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3111
3112         qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
3113                 IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3114
3115 out_free:
3116         kfree(outb);
3117
3118 out:
3119         mutex_unlock(&qp->mutex);
3120         return err;
3121 }
3122
3123 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
3124                                           struct ib_ucontext *context,
3125                                           struct ib_udata *udata)
3126 {
3127         struct mlx5_ib_dev *dev = to_mdev(ibdev);
3128         struct mlx5_ib_xrcd *xrcd;
3129         int err;
3130
3131         if (!MLX5_CAP_GEN(dev->mdev, xrc))
3132                 return ERR_PTR(-ENOSYS);
3133
3134         xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
3135         if (!xrcd)
3136                 return ERR_PTR(-ENOMEM);
3137
3138         err = mlx5_core_xrcd_alloc(dev->mdev, &xrcd->xrcdn);
3139         if (err) {
3140                 kfree(xrcd);
3141                 return ERR_PTR(-ENOMEM);
3142         }
3143
3144         return &xrcd->ibxrcd;
3145 }
3146
3147 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
3148 {
3149         struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
3150         u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
3151         int err;
3152
3153         err = mlx5_core_xrcd_dealloc(dev->mdev, xrcdn);
3154         if (err) {
3155                 mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
3156                 return err;
3157         }
3158
3159         kfree(xrcd);
3160
3161         return 0;
3162 }