drivers/infiniband/hw/mthca/mthca_qp.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Cisco Systems. All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
   5  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  *
  35  * $Id: mthca_qp.c 1355 2004-12-17 15:23:43Z roland $
  36  */
  37
  38 #include <linux/init.h>
  39 #include <linux/string.h>
  40 #include <linux/slab.h>
  41
  42 #include <rdma/ib_verbs.h>
  43 #include <rdma/ib_cache.h>
  44 #include <rdma/ib_pack.h>
  45
  46 #include "mthca_dev.h"
  47 #include "mthca_cmd.h"
  48 #include "mthca_memfree.h"
  49 #include "mthca_wqe.h"
  50
  51 enum {
  52         MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
  53         MTHCA_ACK_REQ_FREQ       = 10,
  54         MTHCA_FLIGHT_LIMIT       = 9,
  55         MTHCA_UD_HEADER_SIZE     = 72, /* largest UD header possible */
  56         MTHCA_INLINE_HEADER_SIZE = 4,  /* data segment overhead for inline */
  57         MTHCA_INLINE_CHUNK_SIZE  = 16  /* inline data segment chunk */
  58 };
  59
  60 enum {
  61         MTHCA_QP_STATE_RST  = 0,
  62         MTHCA_QP_STATE_INIT = 1,
  63         MTHCA_QP_STATE_RTR  = 2,
  64         MTHCA_QP_STATE_RTS  = 3,
  65         MTHCA_QP_STATE_SQE  = 4,
  66         MTHCA_QP_STATE_SQD  = 5,
  67         MTHCA_QP_STATE_ERR  = 6,
  68         MTHCA_QP_STATE_DRAINING = 7
  69 };
  70
  71 enum {
  72         MTHCA_QP_ST_RC  = 0x0,
  73         MTHCA_QP_ST_UC  = 0x1,
  74         MTHCA_QP_ST_RD  = 0x2,
  75         MTHCA_QP_ST_UD  = 0x3,
  76         MTHCA_QP_ST_MLX = 0x7
  77 };
  78
  79 enum {
  80         MTHCA_QP_PM_MIGRATED = 0x3,
  81         MTHCA_QP_PM_ARMED    = 0x0,
  82         MTHCA_QP_PM_REARM    = 0x1
  83 };
  84
  85 enum {
  86         /* qp_context flags */
  87         MTHCA_QP_BIT_DE  = 1 <<  8,
  88         /* params1 */
  89         MTHCA_QP_BIT_SRE = 1 << 15,
  90         MTHCA_QP_BIT_SWE = 1 << 14,
  91         MTHCA_QP_BIT_SAE = 1 << 13,
  92         MTHCA_QP_BIT_SIC = 1 <<  4,
  93         MTHCA_QP_BIT_SSC = 1 <<  3,
  94         /* params2 */
  95         MTHCA_QP_BIT_RRE = 1 << 15,
  96         MTHCA_QP_BIT_RWE = 1 << 14,
  97         MTHCA_QP_BIT_RAE = 1 << 13,
  98         MTHCA_QP_BIT_RIC = 1 <<  4,
  99         MTHCA_QP_BIT_RSC = 1 <<  3
 100 };
 101
 102 struct mthca_qp_path {
 103         __be32 port_pkey;
 104         u8     rnr_retry;
 105         u8     g_mylmc;
 106         __be16 rlid;
 107         u8     ackto;
 108         u8     mgid_index;
 109         u8     static_rate;
 110         u8     hop_limit;
 111         __be32 sl_tclass_flowlabel;
 112         u8     rgid[16];
 113 } __attribute__((packed));
 114
 115 struct mthca_qp_context {
 116         __be32 flags;
 117         __be32 tavor_sched_queue; /* Reserved on Arbel */
 118         u8     mtu_msgmax;
 119         u8     rq_size_stride;  /* Reserved on Tavor */
 120         u8     sq_size_stride;  /* Reserved on Tavor */
 121         u8     rlkey_arbel_sched_queue; /* Reserved on Tavor */
 122         __be32 usr_page;
 123         __be32 local_qpn;
 124         __be32 remote_qpn;
 125         u32    reserved1[2];
 126         struct mthca_qp_path pri_path;
 127         struct mthca_qp_path alt_path;
 128         __be32 rdd;
 129         __be32 pd;
 130         __be32 wqe_base;
 131         __be32 wqe_lkey;
 132         __be32 params1;
 133         __be32 reserved2;
 134         __be32 next_send_psn;
 135         __be32 cqn_snd;
 136         __be32 snd_wqe_base_l;  /* Next send WQE on Tavor */
 137         __be32 snd_db_index;    /* (debugging only entries) */
 138         __be32 last_acked_psn;
 139         __be32 ssn;
 140         __be32 params2;
 141         __be32 rnr_nextrecvpsn;
 142         __be32 ra_buff_indx;
 143         __be32 cqn_rcv;
 144         __be32 rcv_wqe_base_l;  /* Next recv WQE on Tavor */
 145         __be32 rcv_db_index;    /* (debugging only entries) */
 146         __be32 qkey;
 147         __be32 srqn;
 148         __be32 rmsn;
 149         __be16 rq_wqe_counter;  /* reserved on Tavor */
 150         __be16 sq_wqe_counter;  /* reserved on Tavor */
 151         u32    reserved3[18];
 152 } __attribute__((packed));
 153
 154 struct mthca_qp_param {
 155         __be32 opt_param_mask;
 156         u32    reserved1;
 157         struct mthca_qp_context context;
 158         u32    reserved2[62];
 159 } __attribute__((packed));
 160
 161 enum {
 162         MTHCA_QP_OPTPAR_ALT_ADDR_PATH     = 1 << 0,
 163         MTHCA_QP_OPTPAR_RRE               = 1 << 1,
 164         MTHCA_QP_OPTPAR_RAE               = 1 << 2,
 165         MTHCA_QP_OPTPAR_RWE               = 1 << 3,
 166         MTHCA_QP_OPTPAR_PKEY_INDEX        = 1 << 4,
 167         MTHCA_QP_OPTPAR_Q_KEY             = 1 << 5,
 168         MTHCA_QP_OPTPAR_RNR_TIMEOUT       = 1 << 6,
 169         MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
 170         MTHCA_QP_OPTPAR_SRA_MAX           = 1 << 8,
 171         MTHCA_QP_OPTPAR_RRA_MAX           = 1 << 9,
 172         MTHCA_QP_OPTPAR_PM_STATE          = 1 << 10,
 173         MTHCA_QP_OPTPAR_PORT_NUM          = 1 << 11,
 174         MTHCA_QP_OPTPAR_RETRY_COUNT       = 1 << 12,
 175         MTHCA_QP_OPTPAR_ALT_RNR_RETRY     = 1 << 13,
 176         MTHCA_QP_OPTPAR_ACK_TIMEOUT       = 1 << 14,
 177         MTHCA_QP_OPTPAR_RNR_RETRY         = 1 << 15,
 178         MTHCA_QP_OPTPAR_SCHED_QUEUE       = 1 << 16
 179 };
 180
 181 static const u8 mthca_opcode[] = {
 182         [IB_WR_SEND]                 = MTHCA_OPCODE_SEND,
 183         [IB_WR_SEND_WITH_IMM]        = MTHCA_OPCODE_SEND_IMM,
 184         [IB_WR_RDMA_WRITE]           = MTHCA_OPCODE_RDMA_WRITE,
 185         [IB_WR_RDMA_WRITE_WITH_IMM]  = MTHCA_OPCODE_RDMA_WRITE_IMM,
 186         [IB_WR_RDMA_READ]            = MTHCA_OPCODE_RDMA_READ,
 187         [IB_WR_ATOMIC_CMP_AND_SWP]   = MTHCA_OPCODE_ATOMIC_CS,
 188         [IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
 189 };
 190
 191 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
 192 {
 193         return qp->qpn >= dev->qp_table.sqp_start &&
 194                 qp->qpn <= dev->qp_table.sqp_start + 3;
 195 }
 196
 197 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
 198 {
 199         return qp->qpn >= dev->qp_table.sqp_start &&
 200                 qp->qpn <= dev->qp_table.sqp_start + 1;
 201 }
 202
 203 static void *get_recv_wqe(struct mthca_qp *qp, int n)
 204 {
 205         if (qp->is_direct)
 206                 return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
 207         else
 208                 return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
 209                         ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
 210 }
 211
 212 static void *get_send_wqe(struct mthca_qp *qp, int n)
 213 {
 214         if (qp->is_direct)
 215                 return qp->queue.direct.buf + qp->send_wqe_offset +
 216                         (n << qp->sq.wqe_shift);
 217         else
 218                 return qp->queue.page_list[(qp->send_wqe_offset +
 219                                             (n << qp->sq.wqe_shift)) >>
 220                                            PAGE_SHIFT].buf +
 221                         ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
 222                          (PAGE_SIZE - 1));
 223 }
 224
 225 static void mthca_wq_init(struct mthca_wq *wq)
 226 {
 227         spin_lock_init(&wq->lock);
 228         wq->next_ind  = 0;
 229         wq->last_comp = wq->max - 1;
 230         wq->head      = 0;
 231         wq->tail      = 0;
 232 }
 233
 234 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
 235                     enum ib_event_type event_type)
 236 {
 237         struct mthca_qp *qp;
 238         struct ib_event event;
 239
 240         spin_lock(&dev->qp_table.lock);
 241         qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
 242         if (qp)
 243                 atomic_inc(&qp->refcount);
 244         spin_unlock(&dev->qp_table.lock);
 245
 246         if (!qp) {
 247                 mthca_warn(dev, "Async event for bogus QP %08x\n", qpn);
 248                 return;
 249         }
 250
 251         event.device      = &dev->ib_dev;
 252         event.event       = event_type;
 253         event.element.qp  = &qp->ibqp;
 254         if (qp->ibqp.event_handler)
 255                 qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
 256
 257         if (atomic_dec_and_test(&qp->refcount))
 258                 wake_up(&qp->wait);
 259 }
 260
 261 static int to_mthca_state(enum ib_qp_state ib_state)
 262 {
 263         switch (ib_state) {
 264         case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
 265         case IB_QPS_INIT:  return MTHCA_QP_STATE_INIT;
 266         case IB_QPS_RTR:   return MTHCA_QP_STATE_RTR;
 267         case IB_QPS_RTS:   return MTHCA_QP_STATE_RTS;
 268         case IB_QPS_SQD:   return MTHCA_QP_STATE_SQD;
 269         case IB_QPS_SQE:   return MTHCA_QP_STATE_SQE;
 270         case IB_QPS_ERR:   return MTHCA_QP_STATE_ERR;
 271         default:                return -1;
 272         }
 273 }
 274
 275 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
 276
 277 static int to_mthca_st(int transport)
 278 {
 279         switch (transport) {
 280         case RC:  return MTHCA_QP_ST_RC;
 281         case UC:  return MTHCA_QP_ST_UC;
 282         case UD:  return MTHCA_QP_ST_UD;
 283         case RD:  return MTHCA_QP_ST_RD;
 284         case MLX: return MTHCA_QP_ST_MLX;
 285         default:  return -1;
 286         }
 287 }
 288
 289 static const struct {
 290         int trans;
 291         u32 req_param[NUM_TRANS];
 292         u32 opt_param[NUM_TRANS];
 293 } state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
 294         [IB_QPS_RESET] = {
 295                 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
 296                 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
 297                 [IB_QPS_INIT]  = {
 298                         .trans = MTHCA_TRANS_RST2INIT,
 299                         .req_param = {
 300                                 [UD]  = (IB_QP_PKEY_INDEX |
 301                                          IB_QP_PORT       |
 302                                          IB_QP_QKEY),
 303                                 [UC]  = (IB_QP_PKEY_INDEX |
 304                                          IB_QP_PORT       |
 305                                          IB_QP_ACCESS_FLAGS),
 306                                 [RC]  = (IB_QP_PKEY_INDEX |
 307                                          IB_QP_PORT       |
 308                                          IB_QP_ACCESS_FLAGS),
 309                                 [MLX] = (IB_QP_PKEY_INDEX |
 310                                          IB_QP_QKEY),
 311                         },
 312                         /* bug-for-bug compatibility with VAPI: */
 313                         .opt_param = {
 314                                 [MLX] = IB_QP_PORT
 315                         }
 316                 },
 317         },
 318         [IB_QPS_INIT]  = {
 319                 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
 320                 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
 321                 [IB_QPS_INIT]  = {
 322                         .trans = MTHCA_TRANS_INIT2INIT,
 323                         .opt_param = {
 324                                 [UD]  = (IB_QP_PKEY_INDEX |
 325                                          IB_QP_PORT       |
 326                                          IB_QP_QKEY),
 327                                 [UC]  = (IB_QP_PKEY_INDEX |
 328                                          IB_QP_PORT       |
 329                                          IB_QP_ACCESS_FLAGS),
 330                                 [RC]  = (IB_QP_PKEY_INDEX |
 331                                          IB_QP_PORT       |
 332                                          IB_QP_ACCESS_FLAGS),
 333                                 [MLX] = (IB_QP_PKEY_INDEX |
 334                                          IB_QP_QKEY),
 335                         }
 336                 },
 337                 [IB_QPS_RTR]   = {
 338                         .trans = MTHCA_TRANS_INIT2RTR,
 339                         .req_param = {
 340                                 [UC]  = (IB_QP_AV                  |
 341                                          IB_QP_PATH_MTU            |
 342                                          IB_QP_DEST_QPN            |
 343                                          IB_QP_RQ_PSN),
 344                                 [RC]  = (IB_QP_AV                  |
 345                                          IB_QP_PATH_MTU            |
 346                                          IB_QP_DEST_QPN            |
 347                                          IB_QP_RQ_PSN              |
 348                                          IB_QP_MAX_DEST_RD_ATOMIC  |
 349                                          IB_QP_MIN_RNR_TIMER),
 350                         },
 351                         .opt_param = {
 352                                 [UD]  = (IB_QP_PKEY_INDEX |
 353                                          IB_QP_QKEY),
 354                                 [UC]  = (IB_QP_ALT_PATH     |
 355                                          IB_QP_ACCESS_FLAGS |
 356                                          IB_QP_PKEY_INDEX),
 357                                 [RC]  = (IB_QP_ALT_PATH     |
 358                                          IB_QP_ACCESS_FLAGS |
 359                                          IB_QP_PKEY_INDEX),
 360                                 [MLX] = (IB_QP_PKEY_INDEX |
 361                                          IB_QP_QKEY),
 362                         }
 363                 }
 364         },
 365         [IB_QPS_RTR]   = {
 366                 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
 367                 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
 368                 [IB_QPS_RTS]   = {
 369                         .trans = MTHCA_TRANS_RTR2RTS,
 370                         .req_param = {
 371                                 [UD]  = IB_QP_SQ_PSN,
 372                                 [UC]  = IB_QP_SQ_PSN,
 373                                 [RC]  = (IB_QP_TIMEOUT           |
 374                                          IB_QP_RETRY_CNT         |
 375                                          IB_QP_RNR_RETRY         |
 376                                          IB_QP_SQ_PSN            |
 377                                          IB_QP_MAX_QP_RD_ATOMIC),
 378                                 [MLX] = IB_QP_SQ_PSN,
 379                         },
 380                         .opt_param = {
 381                                 [UD]  = (IB_QP_CUR_STATE             |
 382                                          IB_QP_QKEY),
 383                                 [UC]  = (IB_QP_CUR_STATE             |
 384                                          IB_QP_ALT_PATH              |
 385                                          IB_QP_ACCESS_FLAGS          |
 386                                          IB_QP_PKEY_INDEX            |
 387                                          IB_QP_PATH_MIG_STATE),
 388                                 [RC]  = (IB_QP_CUR_STATE             |
 389                                          IB_QP_ALT_PATH              |
 390                                          IB_QP_ACCESS_FLAGS          |
 391                                          IB_QP_PKEY_INDEX            |
 392                                          IB_QP_MIN_RNR_TIMER         |
 393                                          IB_QP_PATH_MIG_STATE),
 394                                 [MLX] = (IB_QP_CUR_STATE             |
 395                                          IB_QP_QKEY),
 396                         }
 397                 }
 398         },
 399         [IB_QPS_RTS]   = {
 400                 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
 401                 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
 402                 [IB_QPS_RTS]   = {
 403                         .trans = MTHCA_TRANS_RTS2RTS,
 404                         .opt_param = {
 405                                 [UD]  = (IB_QP_CUR_STATE             |
 406                                          IB_QP_QKEY),
 407                                 [UC]  = (IB_QP_ACCESS_FLAGS          |
 408                                          IB_QP_ALT_PATH              |
 409                                          IB_QP_PATH_MIG_STATE),
 410                                 [RC]  = (IB_QP_ACCESS_FLAGS          |
 411                                          IB_QP_ALT_PATH              |
 412                                          IB_QP_PATH_MIG_STATE        |
 413                                          IB_QP_MIN_RNR_TIMER),
 414                                 [MLX] = (IB_QP_CUR_STATE             |
 415                                          IB_QP_QKEY),
 416                         }
 417                 },
 418                 [IB_QPS_SQD]   = {
 419                         .trans = MTHCA_TRANS_RTS2SQD,
 420                 },
 421         },
 422         [IB_QPS_SQD]   = {
 423                 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
 424                 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
 425                 [IB_QPS_RTS]   = {
 426                         .trans = MTHCA_TRANS_SQD2RTS,
 427                         .opt_param = {
 428                                 [UD]  = (IB_QP_CUR_STATE             |
 429                                          IB_QP_QKEY),
 430                                 [UC]  = (IB_QP_CUR_STATE             |
 431                                          IB_QP_ALT_PATH              |
 432                                          IB_QP_ACCESS_FLAGS          |
 433                                          IB_QP_PATH_MIG_STATE),
 434                                 [RC]  = (IB_QP_CUR_STATE             |
 435                                          IB_QP_ALT_PATH              |
 436                                          IB_QP_ACCESS_FLAGS          |
 437                                          IB_QP_MIN_RNR_TIMER         |
 438                                          IB_QP_PATH_MIG_STATE),
 439                                 [MLX] = (IB_QP_CUR_STATE             |
 440                                          IB_QP_QKEY),
 441                         }
 442                 },
 443                 [IB_QPS_SQD]   = {
 444                         .trans = MTHCA_TRANS_SQD2SQD,
 445                         .opt_param = {
 446                                 [UD]  = (IB_QP_PKEY_INDEX            |
 447                                          IB_QP_QKEY),
 448                                 [UC]  = (IB_QP_AV                    |
 449                                          IB_QP_CUR_STATE             |
 450                                          IB_QP_ALT_PATH              |
 451                                          IB_QP_ACCESS_FLAGS          |
 452                                          IB_QP_PKEY_INDEX            |
 453                                          IB_QP_PATH_MIG_STATE),
 454                                 [RC]  = (IB_QP_AV                    |
 455                                          IB_QP_TIMEOUT               |
 456                                          IB_QP_RETRY_CNT             |
 457                                          IB_QP_RNR_RETRY             |
 458                                          IB_QP_MAX_QP_RD_ATOMIC      |
 459                                          IB_QP_MAX_DEST_RD_ATOMIC    |
 460                                          IB_QP_CUR_STATE             |
 461                                          IB_QP_ALT_PATH              |
 462                                          IB_QP_ACCESS_FLAGS          |
 463                                          IB_QP_PKEY_INDEX            |
 464                                          IB_QP_MIN_RNR_TIMER         |
 465                                          IB_QP_PATH_MIG_STATE),
 466                                 [MLX] = (IB_QP_PKEY_INDEX            |
 467                                          IB_QP_QKEY),
 468                         }
 469                 }
 470         },
 471         [IB_QPS_SQE]   = {
 472                 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
 473                 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
 474                 [IB_QPS_RTS]   = {
 475                         .trans = MTHCA_TRANS_SQERR2RTS,
 476                         .opt_param = {
 477                                 [UD]  = (IB_QP_CUR_STATE             |
 478                                          IB_QP_QKEY),
 479                                 [UC]  = IB_QP_CUR_STATE,
 480                                 [RC]  = (IB_QP_CUR_STATE             |
 481                                          IB_QP_MIN_RNR_TIMER),
 482                                 [MLX] = (IB_QP_CUR_STATE             |
 483                                          IB_QP_QKEY),
 484                         }
 485                 }
 486         },
 487         [IB_QPS_ERR] = {
 488                 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
 489                 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR }
 490         }
 491 };
 492
 493 static void store_attrs(struct mthca_sqp *sqp, struct ib_qp_attr *attr,
 494                         int attr_mask)
 495 {
 496         if (attr_mask & IB_QP_PKEY_INDEX)
 497                 sqp->pkey_index = attr->pkey_index;
 498         if (attr_mask & IB_QP_QKEY)
 499                 sqp->qkey = attr->qkey;
 500         if (attr_mask & IB_QP_SQ_PSN)
 501                 sqp->send_psn = attr->sq_psn;
 502 }
 503
 504 static void init_port(struct mthca_dev *dev, int port)
 505 {
 506         int err;
 507         u8 status;
 508         struct mthca_init_ib_param param;
 509
 510         memset(&param, 0, sizeof param);
 511
 512         param.port_width = dev->limits.port_width_cap;
 513         param.vl_cap     = dev->limits.vl_cap;
 514         param.mtu_cap    = dev->limits.mtu_cap;
 515         param.gid_cap    = dev->limits.gid_table_len;
 516         param.pkey_cap   = dev->limits.pkey_table_len;
 517
 518         err = mthca_INIT_IB(dev, &param, port, &status);
 519         if (err)
 520                 mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
 521         if (status)
 522                 mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
 523 }
 524
 525 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 526 {
 527         struct mthca_dev *dev = to_mdev(ibqp->device);
 528         struct mthca_qp *qp = to_mqp(ibqp);
 529         enum ib_qp_state cur_state, new_state;
 530         struct mthca_mailbox *mailbox;
 531         struct mthca_qp_param *qp_param;
 532         struct mthca_qp_context *qp_context;
 533         u32 req_param, opt_param;
 534         u8 status;
 535         int err;
 536
 537         if (attr_mask & IB_QP_CUR_STATE) {
 538                 if (attr->cur_qp_state != IB_QPS_RTR &&
 539                     attr->cur_qp_state != IB_QPS_RTS &&
 540                     attr->cur_qp_state != IB_QPS_SQD &&
 541                     attr->cur_qp_state != IB_QPS_SQE)
 542                         return -EINVAL;
 543                 else
 544                         cur_state = attr->cur_qp_state;
 545         } else {
 546                 spin_lock_irq(&qp->sq.lock);
 547                 spin_lock(&qp->rq.lock);
 548                 cur_state = qp->state;
 549                 spin_unlock(&qp->rq.lock);
 550                 spin_unlock_irq(&qp->sq.lock);
 551         }
 552
 553         if (attr_mask & IB_QP_STATE) {
 554                if (attr->qp_state < 0 || attr->qp_state > IB_QPS_ERR)
 555                         return -EINVAL;
 556                 new_state = attr->qp_state;
 557         } else
 558                 new_state = cur_state;
 559
 560         if (state_table[cur_state][new_state].trans == MTHCA_TRANS_INVALID) {
 561                 mthca_dbg(dev, "Illegal QP transition "
 562                           "%d->%d\n", cur_state, new_state);
 563                 return -EINVAL;
 564         }
 565
 566         req_param = state_table[cur_state][new_state].req_param[qp->transport];
 567         opt_param = state_table[cur_state][new_state].opt_param[qp->transport];
 568
 569         if ((req_param & attr_mask) != req_param) {
 570                 mthca_dbg(dev, "QP transition "
 571                           "%d->%d missing req attr 0x%08x\n",
 572                           cur_state, new_state,
 573                           req_param & ~attr_mask);
 574                 return -EINVAL;
 575         }
 576
 577         if (attr_mask & ~(req_param | opt_param | IB_QP_STATE)) {
 578                 mthca_dbg(dev, "QP transition (transport %d) "
 579                           "%d->%d has extra attr 0x%08x\n",
 580                           qp->transport,
 581                           cur_state, new_state,
 582                           attr_mask & ~(req_param | opt_param |
 583                                                  IB_QP_STATE));
 584                 return -EINVAL;
 585         }
 586
 587         if ((attr_mask & IB_QP_PKEY_INDEX) &&
 588              attr->pkey_index >= dev->limits.pkey_table_len) {
 589                 mthca_dbg(dev, "PKey index (%u) too large. max is %d\n",
 590                           attr->pkey_index,dev->limits.pkey_table_len-1);
 591                 return -EINVAL;
 592         }
 593
 594         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 595         if (IS_ERR(mailbox))
 596                 return PTR_ERR(mailbox);
 597         qp_param = mailbox->buf;
 598         qp_context = &qp_param->context;
 599         memset(qp_param, 0, sizeof *qp_param);
 600
 601         qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
 602                                              (to_mthca_st(qp->transport) << 16));
 603         qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
 604         if (!(attr_mask & IB_QP_PATH_MIG_STATE))
 605                 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
 606         else {
 607                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
 608                 switch (attr->path_mig_state) {
 609                 case IB_MIG_MIGRATED:
 610                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
 611                         break;
 612                 case IB_MIG_REARM:
 613                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
 614                         break;
 615                 case IB_MIG_ARMED:
 616                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
 617                         break;
 618                 }
 619         }
 620
 621         /* leave tavor_sched_queue as 0 */
 622
 623         if (qp->transport == MLX || qp->transport == UD)
 624                 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
 625         else if (attr_mask & IB_QP_PATH_MTU)
 626                 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
 627
 628         if (mthca_is_memfree(dev)) {
 629                 if (qp->rq.max)
 630                         qp_context->rq_size_stride = long_log2(qp->rq.max) << 3;
 631                 qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
 632
 633                 if (qp->sq.max)
 634                         qp_context->sq_size_stride = long_log2(qp->sq.max) << 3;
 635                 qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
 636         }
 637
 638         /* leave arbel_sched_queue as 0 */
 639
 640         if (qp->ibqp.uobject)
 641                 qp_context->usr_page =
 642                         cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
 643         else
 644                 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
 645         qp_context->local_qpn  = cpu_to_be32(qp->qpn);
 646         if (attr_mask & IB_QP_DEST_QPN) {
 647                 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
 648         }
 649
 650         if (qp->transport == MLX)
 651                 qp_context->pri_path.port_pkey |=
 652                         cpu_to_be32(to_msqp(qp)->port << 24);
 653         else {
 654                 if (attr_mask & IB_QP_PORT) {
 655                         qp_context->pri_path.port_pkey |=
 656                                 cpu_to_be32(attr->port_num << 24);
 657                         qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
 658                 }
 659         }
 660
 661         if (attr_mask & IB_QP_PKEY_INDEX) {
 662                 qp_context->pri_path.port_pkey |=
 663                         cpu_to_be32(attr->pkey_index);
 664                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
 665         }
 666
 667         if (attr_mask & IB_QP_RNR_RETRY) {
 668                 qp_context->pri_path.rnr_retry = attr->rnr_retry << 5;
 669                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY);
 670         }
 671
 672         if (attr_mask & IB_QP_AV) {
 673                 qp_context->pri_path.g_mylmc     = attr->ah_attr.src_path_bits & 0x7f;
 674                 qp_context->pri_path.rlid        = cpu_to_be16(attr->ah_attr.dlid);
 675                 qp_context->pri_path.static_rate = !!attr->ah_attr.static_rate;
 676                 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
 677                         qp_context->pri_path.g_mylmc |= 1 << 7;
 678                         qp_context->pri_path.mgid_index = attr->ah_attr.grh.sgid_index;
 679                         qp_context->pri_path.hop_limit = attr->ah_attr.grh.hop_limit;
 680                         qp_context->pri_path.sl_tclass_flowlabel =
 681                                 cpu_to_be32((attr->ah_attr.sl << 28)                |
 682                                             (attr->ah_attr.grh.traffic_class << 20) |
 683                                             (attr->ah_attr.grh.flow_label));
 684                         memcpy(qp_context->pri_path.rgid,
 685                                attr->ah_attr.grh.dgid.raw, 16);
 686                 } else {
 687                         qp_context->pri_path.sl_tclass_flowlabel =
 688                                 cpu_to_be32(attr->ah_attr.sl << 28);
 689                 }
 690                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
 691         }
 692
 693         if (attr_mask & IB_QP_TIMEOUT) {
 694                 qp_context->pri_path.ackto = attr->timeout << 3;
 695                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
 696         }
 697
 698         /* XXX alt_path */
 699
 700         /* leave rdd as 0 */
 701         qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
 702         /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
 703         qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
 704         qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
 705                                              (MTHCA_FLIGHT_LIMIT << 24) |
 706                                              MTHCA_QP_BIT_SRE           |
 707                                              MTHCA_QP_BIT_SWE           |
 708                                              MTHCA_QP_BIT_SAE);
 709         if (qp->sq_policy == IB_SIGNAL_ALL_WR)
 710                 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
 711         if (attr_mask & IB_QP_RETRY_CNT) {
 712                 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
 713                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
 714         }
 715
 716         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
 717                 qp_context->params1 |= cpu_to_be32(min(attr->max_rd_atomic ?
 718                                                        ffs(attr->max_rd_atomic) - 1 : 0,
 719                                                        7) << 21);
 720                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
 721         }
 722
 723         if (attr_mask & IB_QP_SQ_PSN)
 724                 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
 725         qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
 726
 727         if (mthca_is_memfree(dev)) {
 728                 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
 729                 qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
 730         }
 731
 732         if (attr_mask & IB_QP_ACCESS_FLAGS) {
 733                 qp_context->params2 |=
 734                         cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE ?
 735                                     MTHCA_QP_BIT_RWE : 0);
 736
 737                 /*
 738                  * Only enable RDMA reads and atomics if we have
 739                  * responder resources set to a non-zero value.
 740                  */
 741                 if (qp->resp_depth) {
 742                         qp_context->params2 |=
 743                                 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_READ ?
 744                                             MTHCA_QP_BIT_RRE : 0);
 745                         qp_context->params2 |=
 746                                 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC ?
 747                                             MTHCA_QP_BIT_RAE : 0);
 748                 }
 749
 750                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
 751                                                         MTHCA_QP_OPTPAR_RRE |
 752                                                         MTHCA_QP_OPTPAR_RAE);
 753
 754                 qp->atomic_rd_en = attr->qp_access_flags;
 755         }
 756
 757         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
 758                 u8 rra_max;
 759
 760                 if (qp->resp_depth && !attr->max_dest_rd_atomic) {
 761                         /*
 762                          * Lowering our responder resources to zero.
 763                          * Turn off reads RDMA and atomics as responder.
 764                          * (RRE/RAE in params2 already zero)
 765                          */
 766                         qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRE |
 767                                                                 MTHCA_QP_OPTPAR_RAE);
 768                 }
 769
 770                 if (!qp->resp_depth && attr->max_dest_rd_atomic) {
 771                         /*
 772                          * Increasing our responder resources from
 773                          * zero.  Turn on RDMA reads and atomics as
 774                          * appropriate.
 775                          */
 776                         qp_context->params2 |=
 777                                 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_READ ?
 778                                             MTHCA_QP_BIT_RRE : 0);
 779                         qp_context->params2 |=
 780                                 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_ATOMIC ?
 781                                             MTHCA_QP_BIT_RAE : 0);
 782
 783                         qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRE |
 784                                                                 MTHCA_QP_OPTPAR_RAE);
 785                 }
 786
 787                 for (rra_max = 0;
 788                      1 << rra_max < attr->max_dest_rd_atomic &&
 789                              rra_max < dev->qp_table.rdb_shift;
 790                      ++rra_max)
 791                         ; /* nothing */
 792
 793                 qp_context->params2      |= cpu_to_be32(rra_max << 21);
 794                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
 795
 796                 qp->resp_depth = attr->max_dest_rd_atomic;
 797         }
 798
 799         qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
 800
 801         if (ibqp->srq)
 802                 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
 803
 804         if (attr_mask & IB_QP_MIN_RNR_TIMER) {
 805                 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
 806                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
 807         }
 808         if (attr_mask & IB_QP_RQ_PSN)
 809                 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
 810
 811         qp_context->ra_buff_indx =
 812                 cpu_to_be32(dev->qp_table.rdb_base +
 813                             ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
 814                              dev->qp_table.rdb_shift));
 815
 816         qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
 817
 818         if (mthca_is_memfree(dev))
 819                 qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
 820
 821         if (attr_mask & IB_QP_QKEY) {
 822                 qp_context->qkey = cpu_to_be32(attr->qkey);
 823                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
 824         }
 825
 826         if (ibqp->srq)
 827                 qp_context->srqn = cpu_to_be32(1 << 24 |
 828                                                to_msrq(ibqp->srq)->srqn);
 829
 830         err = mthca_MODIFY_QP(dev, state_table[cur_state][new_state].trans,
 831                               qp->qpn, 0, mailbox, 0, &status);
 832         if (status) {
 833                 mthca_warn(dev, "modify QP %d returned status %02x.\n",
 834                            state_table[cur_state][new_state].trans, status);
 835                 err = -EINVAL;
 836         }
 837
 838         if (!err)
 839                 qp->state = new_state;
 840
 841         mthca_free_mailbox(dev, mailbox);
 842
 843         if (is_sqp(dev, qp))
 844                 store_attrs(to_msqp(qp), attr, attr_mask);
 845
 846         /*
 847          * If we moved QP0 to RTR, bring the IB link up; if we moved
 848          * QP0 to RESET or ERROR, bring the link back down.
 849          */
 850         if (is_qp0(dev, qp)) {
 851                 if (cur_state != IB_QPS_RTR &&
 852                     new_state == IB_QPS_RTR)
 853                         init_port(dev, to_msqp(qp)->port);
 854
 855                 if (cur_state != IB_QPS_RESET &&
 856                     cur_state != IB_QPS_ERR &&
 857                     (new_state == IB_QPS_RESET ||
 858                      new_state == IB_QPS_ERR))
 859                         mthca_CLOSE_IB(dev, to_msqp(qp)->port, &status);
 860         }
 861
 862         /*
 863          * If we moved a kernel QP to RESET, clean up all old CQ
 864          * entries and reinitialize the QP.
 865          */
 866         if (!err && new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
 867                 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
 868                                qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
 869                 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
 870                         mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
 871                                        qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
 872
 873                 mthca_wq_init(&qp->sq);
 874                 mthca_wq_init(&qp->rq);
 875
 876                 if (mthca_is_memfree(dev)) {
 877                         *qp->sq.db = 0;
 878                         *qp->rq.db = 0;
 879                 }
 880         }
 881
 882         return err;
 883 }
 884
 885 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
 886                                  struct mthca_pd *pd,
 887                                  struct mthca_qp *qp)
 888 {
 889         int max_data_size;
 890
 891         /*
 892          * Calculate the maximum size of WQE s/g segments, excluding
 893          * the next segment and other non-data segments.
 894          */
 895         max_data_size = min(dev->limits.max_desc_sz, 1 << qp->sq.wqe_shift) -
 896                 sizeof (struct mthca_next_seg);
 897
 898         switch (qp->transport) {
 899         case MLX:
 900                 max_data_size -= 2 * sizeof (struct mthca_data_seg);
 901                 break;
 902
 903         case UD:
 904                 if (mthca_is_memfree(dev))
 905                         max_data_size -= sizeof (struct mthca_arbel_ud_seg);
 906                 else
 907                         max_data_size -= sizeof (struct mthca_tavor_ud_seg);
 908                 break;
 909
 910         default:
 911                 max_data_size -= sizeof (struct mthca_raddr_seg);
 912                 break;
 913         }
 914
 915         /* We don't support inline data for kernel QPs (yet). */
 916         if (!pd->ibpd.uobject)
 917                 qp->max_inline_data = 0;
 918         else
 919                 qp->max_inline_data = max_data_size - MTHCA_INLINE_HEADER_SIZE;
 920
 921         qp->sq.max_gs = min_t(int, dev->limits.max_sg,
 922                               max_data_size / sizeof (struct mthca_data_seg));
 923         qp->rq.max_gs = min_t(int, dev->limits.max_sg,
 924                                (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
 925                                 sizeof (struct mthca_next_seg)) /
 926                                sizeof (struct mthca_data_seg));
 927 }
 928
 929 /*
 930  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
 931  * rq.max_gs and sq.max_gs must all be assigned.
 932  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
 933  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
 934  * queue)
 935  */
 936 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
 937                                struct mthca_pd *pd,
 938                                struct mthca_qp *qp)
 939 {
 940         int size;
 941         int err = -ENOMEM;
 942
 943         size = sizeof (struct mthca_next_seg) +
 944                 qp->rq.max_gs * sizeof (struct mthca_data_seg);
 945
 946         if (size > dev->limits.max_desc_sz)
 947                 return -EINVAL;
 948
 949         for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
 950              qp->rq.wqe_shift++)
 951                 ; /* nothing */
 952
 953         size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
 954         switch (qp->transport) {
 955         case MLX:
 956                 size += 2 * sizeof (struct mthca_data_seg);
 957                 break;
 958
 959         case UD:
 960                 size += mthca_is_memfree(dev) ?
 961                         sizeof (struct mthca_arbel_ud_seg) :
 962                         sizeof (struct mthca_tavor_ud_seg);
 963                 break;
 964
 965         case UC:
 966                 size += sizeof (struct mthca_raddr_seg);
 967                 break;
 968
 969         case RC:
 970                 size += sizeof (struct mthca_raddr_seg);
 971                 /*
 972                  * An atomic op will require an atomic segment, a
 973                  * remote address segment and one scatter entry.
 974                  */
 975                 size = max_t(int, size,
 976                              sizeof (struct mthca_atomic_seg) +
 977                              sizeof (struct mthca_raddr_seg) +
 978                              sizeof (struct mthca_data_seg));
 979                 break;
 980
 981         default:
 982                 break;
 983         }
 984
 985         /* Make sure that we have enough space for a bind request */
 986         size = max_t(int, size, sizeof (struct mthca_bind_seg));
 987
 988         size += sizeof (struct mthca_next_seg);
 989
 990         if (size > dev->limits.max_desc_sz)
 991                 return -EINVAL;
 992
 993         for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
 994              qp->sq.wqe_shift++)
 995                 ; /* nothing */
 996
 997         qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
 998                                     1 << qp->sq.wqe_shift);
 999
1000         /*
1001          * If this is a userspace QP, we don't actually have to
1002          * allocate anything.  All we need is to calculate the WQE
1003          * sizes and the send_wqe_offset, so we're done now.
1004          */
1005         if (pd->ibpd.uobject)
1006                 return 0;
1007
1008         size = PAGE_ALIGN(qp->send_wqe_offset +
1009                           (qp->sq.max << qp->sq.wqe_shift));
1010
1011         qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1012                            GFP_KERNEL);
1013         if (!qp->wrid)
1014                 goto err_out;
1015
1016         err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1017                               &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1018         if (err)
1019                 goto err_out;
1020
1021         return 0;
1022
1023 err_out:
1024         kfree(qp->wrid);
1025         return err;
1026 }
1027
1028 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1029                                struct mthca_qp *qp)
1030 {
1031         mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1032                                        (qp->sq.max << qp->sq.wqe_shift)),
1033                        &qp->queue, qp->is_direct, &qp->mr);
1034         kfree(qp->wrid);
1035 }
1036
1037 static int mthca_map_memfree(struct mthca_dev *dev,
1038                              struct mthca_qp *qp)
1039 {
1040         int ret;
1041
1042         if (mthca_is_memfree(dev)) {
1043                 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1044                 if (ret)
1045                         return ret;
1046
1047                 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1048                 if (ret)
1049                         goto err_qpc;
1050
1051                 ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1052                                       qp->qpn << dev->qp_table.rdb_shift);
1053                 if (ret)
1054                         goto err_eqpc;
1055
1056         }
1057
1058         return 0;
1059
1060 err_eqpc:
1061         mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1062
1063 err_qpc:
1064         mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1065
1066         return ret;
1067 }
1068
1069 static void mthca_unmap_memfree(struct mthca_dev *dev,
1070                                 struct mthca_qp *qp)
1071 {
1072         mthca_table_put(dev, dev->qp_table.rdb_table,
1073                         qp->qpn << dev->qp_table.rdb_shift);
1074         mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1075         mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1076 }
1077
1078 static int mthca_alloc_memfree(struct mthca_dev *dev,
1079                                struct mthca_qp *qp)
1080 {
1081         int ret = 0;
1082
1083         if (mthca_is_memfree(dev)) {
1084                 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1085                                                  qp->qpn, &qp->rq.db);
1086                 if (qp->rq.db_index < 0)
1087                         return ret;
1088
1089                 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1090                                                  qp->qpn, &qp->sq.db);
1091                 if (qp->sq.db_index < 0)
1092                         mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1093         }
1094
1095         return ret;
1096 }
1097
1098 static void mthca_free_memfree(struct mthca_dev *dev,
1099                                struct mthca_qp *qp)
1100 {
1101         if (mthca_is_memfree(dev)) {
1102                 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1103                 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1104         }
1105 }
1106
1107 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1108                                  struct mthca_pd *pd,
1109                                  struct mthca_cq *send_cq,
1110                                  struct mthca_cq *recv_cq,
1111                                  enum ib_sig_type send_policy,
1112                                  struct mthca_qp *qp)
1113 {
1114         int ret;
1115         int i;
1116
1117         atomic_set(&qp->refcount, 1);
1118         init_waitqueue_head(&qp->wait);
1119         qp->state        = IB_QPS_RESET;
1120         qp->atomic_rd_en = 0;
1121         qp->resp_depth   = 0;
1122         qp->sq_policy    = send_policy;
1123         mthca_wq_init(&qp->sq);
1124         mthca_wq_init(&qp->rq);
1125
1126         ret = mthca_map_memfree(dev, qp);
1127         if (ret)
1128                 return ret;
1129
1130         ret = mthca_alloc_wqe_buf(dev, pd, qp);
1131         if (ret) {
1132                 mthca_unmap_memfree(dev, qp);
1133                 return ret;
1134         }
1135
1136         mthca_adjust_qp_caps(dev, pd, qp);
1137
1138         /*
1139          * If this is a userspace QP, we're done now.  The doorbells
1140          * will be allocated and buffers will be initialized in
1141          * userspace.
1142          */
1143         if (pd->ibpd.uobject)
1144                 return 0;
1145
1146         ret = mthca_alloc_memfree(dev, qp);
1147         if (ret) {
1148                 mthca_free_wqe_buf(dev, qp);
1149                 mthca_unmap_memfree(dev, qp);
1150                 return ret;
1151         }
1152
1153         if (mthca_is_memfree(dev)) {
1154                 struct mthca_next_seg *next;
1155                 struct mthca_data_seg *scatter;
1156                 int size = (sizeof (struct mthca_next_seg) +
1157                             qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1158
1159                 for (i = 0; i < qp->rq.max; ++i) {
1160                         next = get_recv_wqe(qp, i);
1161                         next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1162                                                    qp->rq.wqe_shift);
1163                         next->ee_nds = cpu_to_be32(size);
1164
1165                         for (scatter = (void *) (next + 1);
1166                              (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1167                              ++scatter)
1168                                 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1169                 }
1170
1171                 for (i = 0; i < qp->sq.max; ++i) {
1172                         next = get_send_wqe(qp, i);
1173                         next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1174                                                     qp->sq.wqe_shift) +
1175                                                    qp->send_wqe_offset);
1176                 }
1177         }
1178
1179         qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1180         qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1181
1182         return 0;
1183 }
1184
1185 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1186                              struct mthca_qp *qp)
1187 {
1188         /* Sanity check QP size before proceeding */
1189         if (cap->max_send_wr  > dev->limits.max_wqes ||
1190             cap->max_recv_wr  > dev->limits.max_wqes ||
1191             cap->max_send_sge > dev->limits.max_sg   ||
1192             cap->max_recv_sge > dev->limits.max_sg)
1193                 return -EINVAL;
1194
1195         if (mthca_is_memfree(dev)) {
1196                 qp->rq.max = cap->max_recv_wr ?
1197                         roundup_pow_of_two(cap->max_recv_wr) : 0;
1198                 qp->sq.max = cap->max_send_wr ?
1199                         roundup_pow_of_two(cap->max_send_wr) : 0;
1200         } else {
1201                 qp->rq.max = cap->max_recv_wr;
1202                 qp->sq.max = cap->max_send_wr;
1203         }
1204
1205         qp->rq.max_gs = cap->max_recv_sge;
1206         qp->sq.max_gs = max_t(int, cap->max_send_sge,
1207                               ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1208                                     MTHCA_INLINE_CHUNK_SIZE) /
1209                               sizeof (struct mthca_data_seg));
1210
1211         /*
1212          * For MLX transport we need 2 extra S/G entries:
1213          * one for the header and one for the checksum at the end
1214          */
1215         if ((qp->transport == MLX && qp->sq.max_gs + 2 > dev->limits.max_sg) ||
1216             qp->sq.max_gs > dev->limits.max_sg || qp->rq.max_gs > dev->limits.max_sg)
1217                 return -EINVAL;
1218
1219         return 0;
1220 }
1221
1222 int mthca_alloc_qp(struct mthca_dev *dev,
1223                    struct mthca_pd *pd,
1224                    struct mthca_cq *send_cq,
1225                    struct mthca_cq *recv_cq,
1226                    enum ib_qp_type type,
1227                    enum ib_sig_type send_policy,
1228                    struct ib_qp_cap *cap,
1229                    struct mthca_qp *qp)
1230 {
1231         int err;
1232
1233         err = mthca_set_qp_size(dev, cap, qp);
1234         if (err)
1235                 return err;
1236
1237         switch (type) {
1238         case IB_QPT_RC: qp->transport = RC; break;
1239         case IB_QPT_UC: qp->transport = UC; break;
1240         case IB_QPT_UD: qp->transport = UD; break;
1241         default: return -EINVAL;
1242         }
1243
1244         qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1245         if (qp->qpn == -1)
1246                 return -ENOMEM;
1247
1248         err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1249                                     send_policy, qp);
1250         if (err) {
1251                 mthca_free(&dev->qp_table.alloc, qp->qpn);
1252                 return err;
1253         }
1254
1255         spin_lock_irq(&dev->qp_table.lock);
1256         mthca_array_set(&dev->qp_table.qp,
1257                         qp->qpn & (dev->limits.num_qps - 1), qp);
1258         spin_unlock_irq(&dev->qp_table.lock);
1259
1260         return 0;
1261 }
1262
1263 int mthca_alloc_sqp(struct mthca_dev *dev,
1264                     struct mthca_pd *pd,
1265                     struct mthca_cq *send_cq,
1266                     struct mthca_cq *recv_cq,
1267                     enum ib_sig_type send_policy,
1268                     struct ib_qp_cap *cap,
1269                     int qpn,
1270                     int port,
1271                     struct mthca_sqp *sqp)
1272 {
1273         u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1274         int err;
1275
1276         err = mthca_set_qp_size(dev, cap, &sqp->qp);
1277         if (err)
1278                 return err;
1279
1280         sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1281         sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1282                                              &sqp->header_dma, GFP_KERNEL);
1283         if (!sqp->header_buf)
1284                 return -ENOMEM;
1285
1286         spin_lock_irq(&dev->qp_table.lock);
1287         if (mthca_array_get(&dev->qp_table.qp, mqpn))
1288                 err = -EBUSY;
1289         else
1290                 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1291         spin_unlock_irq(&dev->qp_table.lock);
1292
1293         if (err)
1294                 goto err_out;
1295
1296         sqp->port = port;
1297         sqp->qp.qpn       = mqpn;
1298         sqp->qp.transport = MLX;
1299
1300         err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1301                                     send_policy, &sqp->qp);
1302         if (err)
1303                 goto err_out_free;
1304
1305         atomic_inc(&pd->sqp_count);
1306
1307         return 0;
1308
1309  err_out_free:
1310         /*
1311          * Lock CQs here, so that CQ polling code can do QP lookup
1312          * without taking a lock.
1313          */
1314         spin_lock_irq(&send_cq->lock);
1315         if (send_cq != recv_cq)
1316                 spin_lock(&recv_cq->lock);
1317
1318         spin_lock(&dev->qp_table.lock);
1319         mthca_array_clear(&dev->qp_table.qp, mqpn);
1320         spin_unlock(&dev->qp_table.lock);
1321
1322         if (send_cq != recv_cq)
1323                 spin_unlock(&recv_cq->lock);
1324         spin_unlock_irq(&send_cq->lock);
1325
1326  err_out:
1327         dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1328                           sqp->header_buf, sqp->header_dma);
1329
1330         return err;
1331 }
1332
1333 void mthca_free_qp(struct mthca_dev *dev,
1334                    struct mthca_qp *qp)
1335 {
1336         u8 status;
1337         struct mthca_cq *send_cq;
1338         struct mthca_cq *recv_cq;
1339
1340         send_cq = to_mcq(qp->ibqp.send_cq);
1341         recv_cq = to_mcq(qp->ibqp.recv_cq);
1342
1343         /*
1344          * Lock CQs here, so that CQ polling code can do QP lookup
1345          * without taking a lock.
1346          */
1347         spin_lock_irq(&send_cq->lock);
1348         if (send_cq != recv_cq)
1349                 spin_lock(&recv_cq->lock);
1350
1351         spin_lock(&dev->qp_table.lock);
1352         mthca_array_clear(&dev->qp_table.qp,
1353                           qp->qpn & (dev->limits.num_qps - 1));
1354         spin_unlock(&dev->qp_table.lock);
1355
1356         if (send_cq != recv_cq)
1357                 spin_unlock(&recv_cq->lock);
1358         spin_unlock_irq(&send_cq->lock);
1359
1360         atomic_dec(&qp->refcount);
1361         wait_event(qp->wait, !atomic_read(&qp->refcount));
1362
1363         if (qp->state != IB_QPS_RESET)
1364                 mthca_MODIFY_QP(dev, MTHCA_TRANS_ANY2RST, qp->qpn, 0, NULL, 0, &status);
1365
1366         /*
1367          * If this is a userspace QP, the buffers, MR, CQs and so on
1368          * will be cleaned up in userspace, so all we have to do is
1369          * unref the mem-free tables and free the QPN in our table.
1370          */
1371         if (!qp->ibqp.uobject) {
1372                 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
1373                                qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1374                 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
1375                         mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
1376                                        qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1377
1378                 mthca_free_memfree(dev, qp);
1379                 mthca_free_wqe_buf(dev, qp);
1380         }
1381
1382         mthca_unmap_memfree(dev, qp);
1383
1384         if (is_sqp(dev, qp)) {
1385                 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1386                 dma_free_coherent(&dev->pdev->dev,
1387                                   to_msqp(qp)->header_buf_size,
1388                                   to_msqp(qp)->header_buf,
1389                                   to_msqp(qp)->header_dma);
1390         } else
1391                 mthca_free(&dev->qp_table.alloc, qp->qpn);
1392 }
1393
1394 /* Create UD header for an MLX send and build a data segment for it */
1395 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1396                             int ind, struct ib_send_wr *wr,
1397                             struct mthca_mlx_seg *mlx,
1398                             struct mthca_data_seg *data)
1399 {
1400         int header_size;
1401         int err;
1402         u16 pkey;
1403
1404         ib_ud_header_init(256, /* assume a MAD */
1405                           sqp->ud_header.grh_present,
1406                           &sqp->ud_header);
1407
1408         err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1409         if (err)
1410                 return err;
1411         mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1412         mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1413                                   (sqp->ud_header.lrh.destination_lid ==
1414                                    IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1415                                   (sqp->ud_header.lrh.service_level << 8));
1416         mlx->rlid = sqp->ud_header.lrh.destination_lid;
1417         mlx->vcrc = 0;
1418
1419         switch (wr->opcode) {
1420         case IB_WR_SEND:
1421                 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1422                 sqp->ud_header.immediate_present = 0;
1423                 break;
1424         case IB_WR_SEND_WITH_IMM:
1425                 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1426                 sqp->ud_header.immediate_present = 1;
1427                 sqp->ud_header.immediate_data = wr->imm_data;
1428                 break;
1429         default:
1430                 return -EINVAL;
1431         }
1432
1433         sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1434         if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1435                 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1436         sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1437         if (!sqp->qp.ibqp.qp_num)
1438                 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1439                                    sqp->pkey_index, &pkey);
1440         else
1441                 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1442                                    wr->wr.ud.pkey_index, &pkey);
1443         sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1444         sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1445         sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1446         sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1447                                                sqp->qkey : wr->wr.ud.remote_qkey);
1448         sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1449
1450         header_size = ib_ud_header_pack(&sqp->ud_header,
1451                                         sqp->header_buf +
1452                                         ind * MTHCA_UD_HEADER_SIZE);
1453
1454         data->byte_count = cpu_to_be32(header_size);
1455         data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1456         data->addr       = cpu_to_be64(sqp->header_dma +
1457                                        ind * MTHCA_UD_HEADER_SIZE);
1458
1459         return 0;
1460 }
1461
1462 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1463                                     struct ib_cq *ib_cq)
1464 {
1465         unsigned cur;
1466         struct mthca_cq *cq;
1467
1468         cur = wq->head - wq->tail;
1469         if (likely(cur + nreq < wq->max))
1470                 return 0;
1471
1472         cq = to_mcq(ib_cq);
1473         spin_lock(&cq->lock);
1474         cur = wq->head - wq->tail;
1475         spin_unlock(&cq->lock);
1476
1477         return cur + nreq >= wq->max;
1478 }
1479
1480 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1481                           struct ib_send_wr **bad_wr)
1482 {
1483         struct mthca_dev *dev = to_mdev(ibqp->device);
1484         struct mthca_qp *qp = to_mqp(ibqp);
1485         void *wqe;
1486         void *prev_wqe;
1487         unsigned long flags;
1488         int err = 0;
1489         int nreq;
1490         int i;
1491         int size;
1492         int size0 = 0;
1493         u32 f0 = 0;
1494         int ind;
1495         u8 op0 = 0;
1496
1497         spin_lock_irqsave(&qp->sq.lock, flags);
1498
1499         /* XXX check that state is OK to post send */
1500
1501         ind = qp->sq.next_ind;
1502
1503         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1504                 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1505                         mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1506                                         " %d max, %d nreq)\n", qp->qpn,
1507                                         qp->sq.head, qp->sq.tail,
1508                                         qp->sq.max, nreq);
1509                         err = -ENOMEM;
1510                         *bad_wr = wr;
1511                         goto out;
1512                 }
1513
1514                 wqe = get_send_wqe(qp, ind);
1515                 prev_wqe = qp->sq.last;
1516                 qp->sq.last = wqe;
1517
1518                 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1519                 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1520                 ((struct mthca_next_seg *) wqe)->flags =
1521                         ((wr->send_flags & IB_SEND_SIGNALED) ?
1522                          cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1523                         ((wr->send_flags & IB_SEND_SOLICITED) ?
1524                          cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1525                         cpu_to_be32(1);
1526                 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1527                     wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1528                         ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1529
1530                 wqe += sizeof (struct mthca_next_seg);
1531                 size = sizeof (struct mthca_next_seg) / 16;
1532
1533                 switch (qp->transport) {
1534                 case RC:
1535                         switch (wr->opcode) {
1536                         case IB_WR_ATOMIC_CMP_AND_SWP:
1537                         case IB_WR_ATOMIC_FETCH_AND_ADD:
1538                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1539                                         cpu_to_be64(wr->wr.atomic.remote_addr);
1540                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1541                                         cpu_to_be32(wr->wr.atomic.rkey);
1542                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1543
1544                                 wqe += sizeof (struct mthca_raddr_seg);
1545
1546                                 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1547                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1548                                                 cpu_to_be64(wr->wr.atomic.swap);
1549                                         ((struct mthca_atomic_seg *) wqe)->compare =
1550                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1551                                 } else {
1552                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1553                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1554                                         ((struct mthca_atomic_seg *) wqe)->compare = 0;
1555                                 }
1556
1557                                 wqe += sizeof (struct mthca_atomic_seg);
1558                                 size += (sizeof (struct mthca_raddr_seg) +
1559                                          sizeof (struct mthca_atomic_seg)) / 16;
1560                                 break;
1561
1562                         case IB_WR_RDMA_WRITE:
1563                         case IB_WR_RDMA_WRITE_WITH_IMM:
1564                         case IB_WR_RDMA_READ:
1565                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1566                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1567                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1568                                         cpu_to_be32(wr->wr.rdma.rkey);
1569                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1570                                 wqe += sizeof (struct mthca_raddr_seg);
1571                                 size += sizeof (struct mthca_raddr_seg) / 16;
1572                                 break;
1573
1574                         default:
1575                                 /* No extra segments required for sends */
1576                                 break;
1577                         }
1578
1579                         break;
1580
1581                 case UC:
1582                         switch (wr->opcode) {
1583                         case IB_WR_RDMA_WRITE:
1584                         case IB_WR_RDMA_WRITE_WITH_IMM:
1585                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1586                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1587                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1588                                         cpu_to_be32(wr->wr.rdma.rkey);
1589                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1590                                 wqe += sizeof (struct mthca_raddr_seg);
1591                                 size += sizeof (struct mthca_raddr_seg) / 16;
1592                                 break;
1593
1594                         default:
1595                                 /* No extra segments required for sends */
1596                                 break;
1597                         }
1598
1599                         break;
1600
1601                 case UD:
1602                         ((struct mthca_tavor_ud_seg *) wqe)->lkey =
1603                                 cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1604                         ((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1605                                 cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1606                         ((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1607                                 cpu_to_be32(wr->wr.ud.remote_qpn);
1608                         ((struct mthca_tavor_ud_seg *) wqe)->qkey =
1609                                 cpu_to_be32(wr->wr.ud.remote_qkey);
1610
1611                         wqe += sizeof (struct mthca_tavor_ud_seg);
1612                         size += sizeof (struct mthca_tavor_ud_seg) / 16;
1613                         break;
1614
1615                 case MLX:
1616                         err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1617                                                wqe - sizeof (struct mthca_next_seg),
1618                                                wqe);
1619                         if (err) {
1620                                 *bad_wr = wr;
1621                                 goto out;
1622                         }
1623                         wqe += sizeof (struct mthca_data_seg);
1624                         size += sizeof (struct mthca_data_seg) / 16;
1625                         break;
1626                 }
1627
1628                 if (wr->num_sge > qp->sq.max_gs) {
1629                         mthca_err(dev, "too many gathers\n");
1630                         err = -EINVAL;
1631                         *bad_wr = wr;
1632                         goto out;
1633                 }
1634
1635                 for (i = 0; i < wr->num_sge; ++i) {
1636                         ((struct mthca_data_seg *) wqe)->byte_count =
1637                                 cpu_to_be32(wr->sg_list[i].length);
1638                         ((struct mthca_data_seg *) wqe)->lkey =
1639                                 cpu_to_be32(wr->sg_list[i].lkey);
1640                         ((struct mthca_data_seg *) wqe)->addr =
1641                                 cpu_to_be64(wr->sg_list[i].addr);
1642                         wqe += sizeof (struct mthca_data_seg);
1643                         size += sizeof (struct mthca_data_seg) / 16;
1644                 }
1645
1646                 /* Add one more inline data segment for ICRC */
1647                 if (qp->transport == MLX) {
1648                         ((struct mthca_data_seg *) wqe)->byte_count =
1649                                 cpu_to_be32((1 << 31) | 4);
1650                         ((u32 *) wqe)[1] = 0;
1651                         wqe += sizeof (struct mthca_data_seg);
1652                         size += sizeof (struct mthca_data_seg) / 16;
1653                 }
1654
1655                 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1656
1657                 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1658                         mthca_err(dev, "opcode invalid\n");
1659                         err = -EINVAL;
1660                         *bad_wr = wr;
1661                         goto out;
1662                 }
1663
1664                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1665                         cpu_to_be32(((ind << qp->sq.wqe_shift) +
1666                                      qp->send_wqe_offset) |
1667                                     mthca_opcode[wr->opcode]);
1668                 wmb();
1669                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1670                         cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size);
1671
1672                 if (!size0) {
1673                         size0 = size;
1674                         op0   = mthca_opcode[wr->opcode];
1675                 }
1676
1677                 ++ind;
1678                 if (unlikely(ind >= qp->sq.max))
1679                         ind -= qp->sq.max;
1680         }
1681
1682 out:
1683         if (likely(nreq)) {
1684                 __be32 doorbell[2];
1685
1686                 doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1687                                            qp->send_wqe_offset) | f0 | op0);
1688                 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1689
1690                 wmb();
1691
1692                 mthca_write64(doorbell,
1693                               dev->kar + MTHCA_SEND_DOORBELL,
1694                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1695         }
1696
1697         qp->sq.next_ind = ind;
1698         qp->sq.head    += nreq;
1699
1700         spin_unlock_irqrestore(&qp->sq.lock, flags);
1701         return err;
1702 }
1703
1704 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1705                              struct ib_recv_wr **bad_wr)
1706 {
1707         struct mthca_dev *dev = to_mdev(ibqp->device);
1708         struct mthca_qp *qp = to_mqp(ibqp);
1709         __be32 doorbell[2];
1710         unsigned long flags;
1711         int err = 0;
1712         int nreq;
1713         int i;
1714         int size;
1715         int size0 = 0;
1716         int ind;
1717         void *wqe;
1718         void *prev_wqe;
1719
1720         spin_lock_irqsave(&qp->rq.lock, flags);
1721
1722         /* XXX check that state is OK to post receive */
1723
1724         ind = qp->rq.next_ind;
1725
1726         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1727                 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1728                         nreq = 0;
1729
1730                         doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1731                         doorbell[1] = cpu_to_be32(qp->qpn << 8);
1732
1733                         wmb();
1734
1735                         mthca_write64(doorbell,
1736                                       dev->kar + MTHCA_RECEIVE_DOORBELL,
1737                                       MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1738
1739                         qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1740                         size0 = 0;
1741                 }
1742
1743                 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1744                         mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1745                                         " %d max, %d nreq)\n", qp->qpn,
1746                                         qp->rq.head, qp->rq.tail,
1747                                         qp->rq.max, nreq);
1748                         err = -ENOMEM;
1749                         *bad_wr = wr;
1750                         goto out;
1751                 }
1752
1753                 wqe = get_recv_wqe(qp, ind);
1754                 prev_wqe = qp->rq.last;
1755                 qp->rq.last = wqe;
1756
1757                 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1758                 ((struct mthca_next_seg *) wqe)->ee_nds =
1759                         cpu_to_be32(MTHCA_NEXT_DBD);
1760                 ((struct mthca_next_seg *) wqe)->flags = 0;
1761
1762                 wqe += sizeof (struct mthca_next_seg);
1763                 size = sizeof (struct mthca_next_seg) / 16;
1764
1765                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1766                         err = -EINVAL;
1767                         *bad_wr = wr;
1768                         goto out;
1769                 }
1770
1771                 for (i = 0; i < wr->num_sge; ++i) {
1772                         ((struct mthca_data_seg *) wqe)->byte_count =
1773                                 cpu_to_be32(wr->sg_list[i].length);
1774                         ((struct mthca_data_seg *) wqe)->lkey =
1775                                 cpu_to_be32(wr->sg_list[i].lkey);
1776                         ((struct mthca_data_seg *) wqe)->addr =
1777                                 cpu_to_be64(wr->sg_list[i].addr);
1778                         wqe += sizeof (struct mthca_data_seg);
1779                         size += sizeof (struct mthca_data_seg) / 16;
1780                 }
1781
1782                 qp->wrid[ind] = wr->wr_id;
1783
1784                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1785                         cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1786                 wmb();
1787                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1788                         cpu_to_be32(MTHCA_NEXT_DBD | size);
1789
1790                 if (!size0)
1791                         size0 = size;
1792
1793                 ++ind;
1794                 if (unlikely(ind >= qp->rq.max))
1795                         ind -= qp->rq.max;
1796         }
1797
1798 out:
1799         if (likely(nreq)) {
1800                 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1801                 doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
1802
1803                 wmb();
1804
1805                 mthca_write64(doorbell,
1806                               dev->kar + MTHCA_RECEIVE_DOORBELL,
1807                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1808         }
1809
1810         qp->rq.next_ind = ind;
1811         qp->rq.head    += nreq;
1812
1813         spin_unlock_irqrestore(&qp->rq.lock, flags);
1814         return err;
1815 }
1816
1817 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1818                           struct ib_send_wr **bad_wr)
1819 {
1820         struct mthca_dev *dev = to_mdev(ibqp->device);
1821         struct mthca_qp *qp = to_mqp(ibqp);
1822         void *wqe;
1823         void *prev_wqe;
1824         unsigned long flags;
1825         int err = 0;
1826         int nreq;
1827         int i;
1828         int size;
1829         int size0 = 0;
1830         u32 f0 = 0;
1831         int ind;
1832         u8 op0 = 0;
1833
1834         spin_lock_irqsave(&qp->sq.lock, flags);
1835
1836         /* XXX check that state is OK to post send */
1837
1838         ind = qp->sq.head & (qp->sq.max - 1);
1839
1840         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1841                 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1842                         mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1843                                         " %d max, %d nreq)\n", qp->qpn,
1844                                         qp->sq.head, qp->sq.tail,
1845                                         qp->sq.max, nreq);
1846                         err = -ENOMEM;
1847                         *bad_wr = wr;
1848                         goto out;
1849                 }
1850
1851                 wqe = get_send_wqe(qp, ind);
1852                 prev_wqe = qp->sq.last;
1853                 qp->sq.last = wqe;
1854
1855                 ((struct mthca_next_seg *) wqe)->flags =
1856                         ((wr->send_flags & IB_SEND_SIGNALED) ?
1857                          cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1858                         ((wr->send_flags & IB_SEND_SOLICITED) ?
1859                          cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1860                         cpu_to_be32(1);
1861                 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1862                     wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1863                         ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1864
1865                 wqe += sizeof (struct mthca_next_seg);
1866                 size = sizeof (struct mthca_next_seg) / 16;
1867
1868                 switch (qp->transport) {
1869                 case RC:
1870                         switch (wr->opcode) {
1871                         case IB_WR_ATOMIC_CMP_AND_SWP:
1872                         case IB_WR_ATOMIC_FETCH_AND_ADD:
1873                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1874                                         cpu_to_be64(wr->wr.atomic.remote_addr);
1875                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1876                                         cpu_to_be32(wr->wr.atomic.rkey);
1877                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1878
1879                                 wqe += sizeof (struct mthca_raddr_seg);
1880
1881                                 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1882                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1883                                                 cpu_to_be64(wr->wr.atomic.swap);
1884                                         ((struct mthca_atomic_seg *) wqe)->compare =
1885                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1886                                 } else {
1887                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1888                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1889                                         ((struct mthca_atomic_seg *) wqe)->compare = 0;
1890                                 }
1891
1892                                 wqe += sizeof (struct mthca_atomic_seg);
1893                                 size += (sizeof (struct mthca_raddr_seg) +
1894                                          sizeof (struct mthca_atomic_seg)) / 16;
1895                                 break;
1896
1897                         case IB_WR_RDMA_READ:
1898                         case IB_WR_RDMA_WRITE:
1899                         case IB_WR_RDMA_WRITE_WITH_IMM:
1900                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1901                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1902                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1903                                         cpu_to_be32(wr->wr.rdma.rkey);
1904                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1905                                 wqe += sizeof (struct mthca_raddr_seg);
1906                                 size += sizeof (struct mthca_raddr_seg) / 16;
1907                                 break;
1908
1909                         default:
1910                                 /* No extra segments required for sends */
1911                                 break;
1912                         }
1913
1914                         break;
1915
1916                 case UC:
1917                         switch (wr->opcode) {
1918                         case IB_WR_RDMA_WRITE:
1919                         case IB_WR_RDMA_WRITE_WITH_IMM:
1920                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1921                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1922                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1923                                         cpu_to_be32(wr->wr.rdma.rkey);
1924                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1925                                 wqe += sizeof (struct mthca_raddr_seg);
1926                                 size += sizeof (struct mthca_raddr_seg) / 16;
1927                                 break;
1928
1929                         default:
1930                                 /* No extra segments required for sends */
1931                                 break;
1932                         }
1933
1934                         break;
1935
1936                 case UD:
1937                         memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
1938                                to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1939                         ((struct mthca_arbel_ud_seg *) wqe)->dqpn =
1940                                 cpu_to_be32(wr->wr.ud.remote_qpn);
1941                         ((struct mthca_arbel_ud_seg *) wqe)->qkey =
1942                                 cpu_to_be32(wr->wr.ud.remote_qkey);
1943
1944                         wqe += sizeof (struct mthca_arbel_ud_seg);
1945                         size += sizeof (struct mthca_arbel_ud_seg) / 16;
1946                         break;
1947
1948                 case MLX:
1949                         err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1950                                                wqe - sizeof (struct mthca_next_seg),
1951                                                wqe);
1952                         if (err) {
1953                                 *bad_wr = wr;
1954                                 goto out;
1955                         }
1956                         wqe += sizeof (struct mthca_data_seg);
1957                         size += sizeof (struct mthca_data_seg) / 16;
1958                         break;
1959                 }
1960
1961                 if (wr->num_sge > qp->sq.max_gs) {
1962                         mthca_err(dev, "too many gathers\n");
1963                         err = -EINVAL;
1964                         *bad_wr = wr;
1965                         goto out;
1966                 }
1967
1968                 for (i = 0; i < wr->num_sge; ++i) {
1969                         ((struct mthca_data_seg *) wqe)->byte_count =
1970                                 cpu_to_be32(wr->sg_list[i].length);
1971                         ((struct mthca_data_seg *) wqe)->lkey =
1972                                 cpu_to_be32(wr->sg_list[i].lkey);
1973                         ((struct mthca_data_seg *) wqe)->addr =
1974                                 cpu_to_be64(wr->sg_list[i].addr);
1975                         wqe += sizeof (struct mthca_data_seg);
1976                         size += sizeof (struct mthca_data_seg) / 16;
1977                 }
1978
1979                 /* Add one more inline data segment for ICRC */
1980                 if (qp->transport == MLX) {
1981                         ((struct mthca_data_seg *) wqe)->byte_count =
1982                                 cpu_to_be32((1 << 31) | 4);
1983                         ((u32 *) wqe)[1] = 0;
1984                         wqe += sizeof (struct mthca_data_seg);
1985                         size += sizeof (struct mthca_data_seg) / 16;
1986                 }
1987
1988                 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1989
1990                 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1991                         mthca_err(dev, "opcode invalid\n");
1992                         err = -EINVAL;
1993                         *bad_wr = wr;
1994                         goto out;
1995                 }
1996
1997                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1998                         cpu_to_be32(((ind << qp->sq.wqe_shift) +
1999                                      qp->send_wqe_offset) |
2000                                     mthca_opcode[wr->opcode]);
2001                 wmb();
2002                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
2003                         cpu_to_be32(MTHCA_NEXT_DBD | size);
2004
2005                 if (!size0) {
2006                         size0 = size;
2007                         op0   = mthca_opcode[wr->opcode];
2008                 }
2009
2010                 ++ind;
2011                 if (unlikely(ind >= qp->sq.max))
2012                         ind -= qp->sq.max;
2013         }
2014
2015 out:
2016         if (likely(nreq)) {
2017                 __be32 doorbell[2];
2018
2019                 doorbell[0] = cpu_to_be32((nreq << 24)                  |
2020                                           ((qp->sq.head & 0xffff) << 8) |
2021                                           f0 | op0);
2022                 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
2023
2024                 qp->sq.head += nreq;
2025
2026                 /*
2027                  * Make sure that descriptors are written before
2028                  * doorbell record.
2029                  */
2030                 wmb();
2031                 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2032
2033                 /*
2034                  * Make sure doorbell record is written before we
2035                  * write MMIO send doorbell.
2036                  */
2037                 wmb();
2038                 mthca_write64(doorbell,
2039                               dev->kar + MTHCA_SEND_DOORBELL,
2040                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2041         }
2042
2043         spin_unlock_irqrestore(&qp->sq.lock, flags);
2044         return err;
2045 }
2046
2047 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2048                              struct ib_recv_wr **bad_wr)
2049 {
2050         struct mthca_dev *dev = to_mdev(ibqp->device);
2051         struct mthca_qp *qp = to_mqp(ibqp);
2052         unsigned long flags;
2053         int err = 0;
2054         int nreq;
2055         int ind;
2056         int i;
2057         void *wqe;
2058
2059         spin_lock_irqsave(&qp->rq.lock, flags);
2060
2061         /* XXX check that state is OK to post receive */
2062
2063         ind = qp->rq.head & (qp->rq.max - 1);
2064
2065         for (nreq = 0; wr; ++nreq, wr = wr->next) {
2066                 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2067                         mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2068                                         " %d max, %d nreq)\n", qp->qpn,
2069                                         qp->rq.head, qp->rq.tail,
2070                                         qp->rq.max, nreq);
2071                         err = -ENOMEM;
2072                         *bad_wr = wr;
2073                         goto out;
2074                 }
2075
2076                 wqe = get_recv_wqe(qp, ind);
2077
2078                 ((struct mthca_next_seg *) wqe)->flags = 0;
2079
2080                 wqe += sizeof (struct mthca_next_seg);
2081
2082                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2083                         err = -EINVAL;
2084                         *bad_wr = wr;
2085                         goto out;
2086                 }
2087
2088                 for (i = 0; i < wr->num_sge; ++i) {
2089                         ((struct mthca_data_seg *) wqe)->byte_count =
2090                                 cpu_to_be32(wr->sg_list[i].length);
2091                         ((struct mthca_data_seg *) wqe)->lkey =
2092                                 cpu_to_be32(wr->sg_list[i].lkey);
2093                         ((struct mthca_data_seg *) wqe)->addr =
2094                                 cpu_to_be64(wr->sg_list[i].addr);
2095                         wqe += sizeof (struct mthca_data_seg);
2096                 }
2097
2098                 if (i < qp->rq.max_gs) {
2099                         ((struct mthca_data_seg *) wqe)->byte_count = 0;
2100                         ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
2101                         ((struct mthca_data_seg *) wqe)->addr = 0;
2102                 }
2103
2104                 qp->wrid[ind] = wr->wr_id;
2105
2106                 ++ind;
2107                 if (unlikely(ind >= qp->rq.max))
2108                         ind -= qp->rq.max;
2109         }
2110 out:
2111         if (likely(nreq)) {
2112                 qp->rq.head += nreq;
2113
2114                 /*
2115                  * Make sure that descriptors are written before
2116                  * doorbell record.
2117                  */
2118                 wmb();
2119                 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2120         }
2121
2122         spin_unlock_irqrestore(&qp->rq.lock, flags);
2123         return err;
2124 }
2125
2126 int mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2127                        int index, int *dbd, __be32 *new_wqe)
2128 {
2129         struct mthca_next_seg *next;
2130
2131         /*
2132          * For SRQs, all WQEs generate a CQE, so we're always at the
2133          * end of the doorbell chain.
2134          */
2135         if (qp->ibqp.srq) {
2136                 *new_wqe = 0;
2137                 return 0;
2138         }
2139
2140         if (is_send)
2141                 next = get_send_wqe(qp, index);
2142         else
2143                 next = get_recv_wqe(qp, index);
2144
2145         *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2146         if (next->ee_nds & cpu_to_be32(0x3f))
2147                 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2148                         (next->ee_nds & cpu_to_be32(0x3f));
2149         else
2150                 *new_wqe = 0;
2151
2152         return 0;
2153 }
2154
2155 int __devinit mthca_init_qp_table(struct mthca_dev *dev)
2156 {
2157         int err;
2158         u8 status;
2159         int i;
2160
2161         spin_lock_init(&dev->qp_table.lock);
2162
2163         /*
2164          * We reserve 2 extra QPs per port for the special QPs.  The
2165          * special QP for port 1 has to be even, so round up.
2166          */
2167         dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2168         err = mthca_alloc_init(&dev->qp_table.alloc,
2169                                dev->limits.num_qps,
2170                                (1 << 24) - 1,
2171                                dev->qp_table.sqp_start +
2172                                MTHCA_MAX_PORTS * 2);
2173         if (err)
2174                 return err;
2175
2176         err = mthca_array_init(&dev->qp_table.qp,
2177                                dev->limits.num_qps);
2178         if (err) {
2179                 mthca_alloc_cleanup(&dev->qp_table.alloc);
2180                 return err;
2181         }
2182
2183         for (i = 0; i < 2; ++i) {
2184                 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2185                                             dev->qp_table.sqp_start + i * 2,
2186                                             &status);
2187                 if (err)
2188                         goto err_out;
2189                 if (status) {
2190                         mthca_warn(dev, "CONF_SPECIAL_QP returned "
2191                                    "status %02x, aborting.\n",
2192                                    status);
2193                         err = -EINVAL;
2194                         goto err_out;
2195                 }
2196         }
2197         return 0;
2198
2199  err_out:
2200         for (i = 0; i < 2; ++i)
2201                 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2202
2203         mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2204         mthca_alloc_cleanup(&dev->qp_table.alloc);
2205
2206         return err;
2207 }
2208
2209 void __devexit mthca_cleanup_qp_table(struct mthca_dev *dev)
2210 {
2211         int i;
2212         u8 status;
2213
2214         for (i = 0; i < 2; ++i)
2215                 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2216
2217         mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2218         mthca_alloc_cleanup(&dev->qp_table.alloc);
2219 }