git.oblomov.eu Git - linux-2.6/blob - 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                 /*
 734                  * Only enable RDMA/atomics if we have responder
 735                  * resources set to a non-zero value.
 736                  */
 737                 if (qp->resp_depth) {
 738                         qp_context->params2 |=
 739                                 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE ?
 740                                             MTHCA_QP_BIT_RWE : 0);
 741                         qp_context->params2 |=
 742                                 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_READ ?
 743                                             MTHCA_QP_BIT_RRE : 0);
 744                         qp_context->params2 |=
 745                                 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC ?
 746                                             MTHCA_QP_BIT_RAE : 0);
 747                 }
 748
 749                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
 750                                                         MTHCA_QP_OPTPAR_RRE |
 751                                                         MTHCA_QP_OPTPAR_RAE);
 752
 753                 qp->atomic_rd_en = attr->qp_access_flags;
 754         }
 755
 756         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
 757                 u8 rra_max;
 758
 759                 if (qp->resp_depth && !attr->max_dest_rd_atomic) {
 760                         /*
 761                          * Lowering our responder resources to zero.
 762                          * Turn off RDMA/atomics as responder.
 763                          * (RWE/RRE/RAE in params2 already zero)
 764                          */
 765                         qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
 766                                                                 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/atomics as appropriate.
 774                          */
 775                         qp_context->params2 |=
 776                                 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_WRITE ?
 777                                             MTHCA_QP_BIT_RWE : 0);
 778                         qp_context->params2 |=
 779                                 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_READ ?
 780                                             MTHCA_QP_BIT_RRE : 0);
 781                         qp_context->params2 |=
 782                                 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_ATOMIC ?
 783                                             MTHCA_QP_BIT_RAE : 0);
 784
 785                         qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
 786                                                                 MTHCA_QP_OPTPAR_RRE |
 787                                                                 MTHCA_QP_OPTPAR_RAE);
 788                 }
 789
 790                 for (rra_max = 0;
 791                      1 << rra_max < attr->max_dest_rd_atomic &&
 792                              rra_max < dev->qp_table.rdb_shift;
 793                      ++rra_max)
 794                         ; /* nothing */
 795
 796                 qp_context->params2      |= cpu_to_be32(rra_max << 21);
 797                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
 798
 799                 qp->resp_depth = attr->max_dest_rd_atomic;
 800         }
 801
 802         qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
 803
 804         if (ibqp->srq)
 805                 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
 806
 807         if (attr_mask & IB_QP_MIN_RNR_TIMER) {
 808                 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
 809                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
 810         }
 811         if (attr_mask & IB_QP_RQ_PSN)
 812                 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
 813
 814         qp_context->ra_buff_indx =
 815                 cpu_to_be32(dev->qp_table.rdb_base +
 816                             ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
 817                              dev->qp_table.rdb_shift));
 818
 819         qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
 820
 821         if (mthca_is_memfree(dev))
 822                 qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
 823
 824         if (attr_mask & IB_QP_QKEY) {
 825                 qp_context->qkey = cpu_to_be32(attr->qkey);
 826                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
 827         }
 828
 829         if (ibqp->srq)
 830                 qp_context->srqn = cpu_to_be32(1 << 24 |
 831                                                to_msrq(ibqp->srq)->srqn);
 832
 833         err = mthca_MODIFY_QP(dev, state_table[cur_state][new_state].trans,
 834                               qp->qpn, 0, mailbox, 0, &status);
 835         if (status) {
 836                 mthca_warn(dev, "modify QP %d returned status %02x.\n",
 837                            state_table[cur_state][new_state].trans, status);
 838                 err = -EINVAL;
 839         }
 840
 841         if (!err)
 842                 qp->state = new_state;
 843
 844         mthca_free_mailbox(dev, mailbox);
 845
 846         if (is_sqp(dev, qp))
 847                 store_attrs(to_msqp(qp), attr, attr_mask);
 848
 849         /*
 850          * If we moved QP0 to RTR, bring the IB link up; if we moved
 851          * QP0 to RESET or ERROR, bring the link back down.
 852          */
 853         if (is_qp0(dev, qp)) {
 854                 if (cur_state != IB_QPS_RTR &&
 855                     new_state == IB_QPS_RTR)
 856                         init_port(dev, to_msqp(qp)->port);
 857
 858                 if (cur_state != IB_QPS_RESET &&
 859                     cur_state != IB_QPS_ERR &&
 860                     (new_state == IB_QPS_RESET ||
 861                      new_state == IB_QPS_ERR))
 862                         mthca_CLOSE_IB(dev, to_msqp(qp)->port, &status);
 863         }
 864
 865         /*
 866          * If we moved a kernel QP to RESET, clean up all old CQ
 867          * entries and reinitialize the QP.
 868          */
 869         if (!err && new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
 870                 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
 871                                qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
 872                 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
 873                         mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
 874                                        qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
 875
 876                 mthca_wq_init(&qp->sq);
 877                 mthca_wq_init(&qp->rq);
 878
 879                 if (mthca_is_memfree(dev)) {
 880                         *qp->sq.db = 0;
 881                         *qp->rq.db = 0;
 882                 }
 883         }
 884
 885         return err;
 886 }
 887
 888 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
 889                                  struct mthca_pd *pd,
 890                                  struct mthca_qp *qp)
 891 {
 892         int max_data_size;
 893
 894         /*
 895          * Calculate the maximum size of WQE s/g segments, excluding
 896          * the next segment and other non-data segments.
 897          */
 898         max_data_size = min(dev->limits.max_desc_sz, 1 << qp->sq.wqe_shift) -
 899                 sizeof (struct mthca_next_seg);
 900
 901         switch (qp->transport) {
 902         case MLX:
 903                 max_data_size -= 2 * sizeof (struct mthca_data_seg);
 904                 break;
 905
 906         case UD:
 907                 if (mthca_is_memfree(dev))
 908                         max_data_size -= sizeof (struct mthca_arbel_ud_seg);
 909                 else
 910                         max_data_size -= sizeof (struct mthca_tavor_ud_seg);
 911                 break;
 912
 913         default:
 914                 max_data_size -= sizeof (struct mthca_raddr_seg);
 915                 break;
 916         }
 917
 918         /* We don't support inline data for kernel QPs (yet). */
 919         if (!pd->ibpd.uobject)
 920                 qp->max_inline_data = 0;
 921         else
 922                 qp->max_inline_data = max_data_size - MTHCA_INLINE_HEADER_SIZE;
 923
 924         qp->sq.max_gs = max_data_size / sizeof (struct mthca_data_seg);
 925         qp->rq.max_gs = (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
 926                         sizeof (struct mthca_next_seg)) /
 927                         sizeof (struct mthca_data_seg);
 928 }
 929
 930 /*
 931  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
 932  * rq.max_gs and sq.max_gs must all be assigned.
 933  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
 934  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
 935  * queue)
 936  */
 937 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
 938                                struct mthca_pd *pd,
 939                                struct mthca_qp *qp)
 940 {
 941         int size;
 942         int err = -ENOMEM;
 943
 944         size = sizeof (struct mthca_next_seg) +
 945                 qp->rq.max_gs * sizeof (struct mthca_data_seg);
 946
 947         if (size > dev->limits.max_desc_sz)
 948                 return -EINVAL;
 949
 950         for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
 951              qp->rq.wqe_shift++)
 952                 ; /* nothing */
 953
 954         size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
 955         switch (qp->transport) {
 956         case MLX:
 957                 size += 2 * sizeof (struct mthca_data_seg);
 958                 break;
 959
 960         case UD:
 961                 size += mthca_is_memfree(dev) ?
 962                         sizeof (struct mthca_arbel_ud_seg) :
 963                         sizeof (struct mthca_tavor_ud_seg);
 964                 break;
 965
 966         case UC:
 967                 size += sizeof (struct mthca_raddr_seg);
 968                 break;
 969
 970         case RC:
 971                 size += sizeof (struct mthca_raddr_seg);
 972                 /*
 973                  * An atomic op will require an atomic segment, a
 974                  * remote address segment and one scatter entry.
 975                  */
 976                 size = max_t(int, size,
 977                              sizeof (struct mthca_atomic_seg) +
 978                              sizeof (struct mthca_raddr_seg) +
 979                              sizeof (struct mthca_data_seg));
 980                 break;
 981
 982         default:
 983                 break;
 984         }
 985
 986         /* Make sure that we have enough space for a bind request */
 987         size = max_t(int, size, sizeof (struct mthca_bind_seg));
 988
 989         size += sizeof (struct mthca_next_seg);
 990
 991         if (size > dev->limits.max_desc_sz)
 992                 return -EINVAL;
 993
 994         for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
 995              qp->sq.wqe_shift++)
 996                 ; /* nothing */
 997
 998         qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
 999                                     1 << qp->sq.wqe_shift);
1000
1001         /*
1002          * If this is a userspace QP, we don't actually have to
1003          * allocate anything.  All we need is to calculate the WQE
1004          * sizes and the send_wqe_offset, so we're done now.
1005          */
1006         if (pd->ibpd.uobject)
1007                 return 0;
1008
1009         size = PAGE_ALIGN(qp->send_wqe_offset +
1010                           (qp->sq.max << qp->sq.wqe_shift));
1011
1012         qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1013                            GFP_KERNEL);
1014         if (!qp->wrid)
1015                 goto err_out;
1016
1017         err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1018                               &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1019         if (err)
1020                 goto err_out;
1021
1022         return 0;
1023
1024 err_out:
1025         kfree(qp->wrid);
1026         return err;
1027 }
1028
1029 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1030                                struct mthca_qp *qp)
1031 {
1032         mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1033                                        (qp->sq.max << qp->sq.wqe_shift)),
1034                        &qp->queue, qp->is_direct, &qp->mr);
1035         kfree(qp->wrid);
1036 }
1037
1038 static int mthca_map_memfree(struct mthca_dev *dev,
1039                              struct mthca_qp *qp)
1040 {
1041         int ret;
1042
1043         if (mthca_is_memfree(dev)) {
1044                 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1045                 if (ret)
1046                         return ret;
1047
1048                 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1049                 if (ret)
1050                         goto err_qpc;
1051
1052                 ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1053                                       qp->qpn << dev->qp_table.rdb_shift);
1054                 if (ret)
1055                         goto err_eqpc;
1056
1057         }
1058
1059         return 0;
1060
1061 err_eqpc:
1062         mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1063
1064 err_qpc:
1065         mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1066
1067         return ret;
1068 }
1069
1070 static void mthca_unmap_memfree(struct mthca_dev *dev,
1071                                 struct mthca_qp *qp)
1072 {
1073         mthca_table_put(dev, dev->qp_table.rdb_table,
1074                         qp->qpn << dev->qp_table.rdb_shift);
1075         mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1076         mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1077 }
1078
1079 static int mthca_alloc_memfree(struct mthca_dev *dev,
1080                                struct mthca_qp *qp)
1081 {
1082         int ret = 0;
1083
1084         if (mthca_is_memfree(dev)) {
1085                 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1086                                                  qp->qpn, &qp->rq.db);
1087                 if (qp->rq.db_index < 0)
1088                         return ret;
1089
1090                 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1091                                                  qp->qpn, &qp->sq.db);
1092                 if (qp->sq.db_index < 0)
1093                         mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1094         }
1095
1096         return ret;
1097 }
1098
1099 static void mthca_free_memfree(struct mthca_dev *dev,
1100                                struct mthca_qp *qp)
1101 {
1102         if (mthca_is_memfree(dev)) {
1103                 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1104                 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1105         }
1106 }
1107
1108 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1109                                  struct mthca_pd *pd,
1110                                  struct mthca_cq *send_cq,
1111                                  struct mthca_cq *recv_cq,
1112                                  enum ib_sig_type send_policy,
1113                                  struct mthca_qp *qp)
1114 {
1115         int ret;
1116         int i;
1117
1118         atomic_set(&qp->refcount, 1);
1119         init_waitqueue_head(&qp->wait);
1120         qp->state        = IB_QPS_RESET;
1121         qp->atomic_rd_en = 0;
1122         qp->resp_depth   = 0;
1123         qp->sq_policy    = send_policy;
1124         mthca_wq_init(&qp->sq);
1125         mthca_wq_init(&qp->rq);
1126
1127         ret = mthca_map_memfree(dev, qp);
1128         if (ret)
1129                 return ret;
1130
1131         ret = mthca_alloc_wqe_buf(dev, pd, qp);
1132         if (ret) {
1133                 mthca_unmap_memfree(dev, qp);
1134                 return ret;
1135         }
1136
1137         mthca_adjust_qp_caps(dev, pd, qp);
1138
1139         /*
1140          * If this is a userspace QP, we're done now.  The doorbells
1141          * will be allocated and buffers will be initialized in
1142          * userspace.
1143          */
1144         if (pd->ibpd.uobject)
1145                 return 0;
1146
1147         ret = mthca_alloc_memfree(dev, qp);
1148         if (ret) {
1149                 mthca_free_wqe_buf(dev, qp);
1150                 mthca_unmap_memfree(dev, qp);
1151                 return ret;
1152         }
1153
1154         if (mthca_is_memfree(dev)) {
1155                 struct mthca_next_seg *next;
1156                 struct mthca_data_seg *scatter;
1157                 int size = (sizeof (struct mthca_next_seg) +
1158                             qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1159
1160                 for (i = 0; i < qp->rq.max; ++i) {
1161                         next = get_recv_wqe(qp, i);
1162                         next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1163                                                    qp->rq.wqe_shift);
1164                         next->ee_nds = cpu_to_be32(size);
1165
1166                         for (scatter = (void *) (next + 1);
1167                              (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1168                              ++scatter)
1169                                 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1170                 }
1171
1172                 for (i = 0; i < qp->sq.max; ++i) {
1173                         next = get_send_wqe(qp, i);
1174                         next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1175                                                     qp->sq.wqe_shift) +
1176                                                    qp->send_wqe_offset);
1177                 }
1178         }
1179
1180         qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1181         qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1182
1183         return 0;
1184 }
1185
1186 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1187                              struct mthca_qp *qp)
1188 {
1189         /* Sanity check QP size before proceeding */
1190         if (cap->max_send_wr  > dev->limits.max_wqes ||
1191             cap->max_recv_wr  > dev->limits.max_wqes ||
1192             cap->max_send_sge > dev->limits.max_sg   ||
1193             cap->max_recv_sge > dev->limits.max_sg)
1194                 return -EINVAL;
1195
1196         if (mthca_is_memfree(dev)) {
1197                 qp->rq.max = cap->max_recv_wr ?
1198                         roundup_pow_of_two(cap->max_recv_wr) : 0;
1199                 qp->sq.max = cap->max_send_wr ?
1200                         roundup_pow_of_two(cap->max_send_wr) : 0;
1201         } else {
1202                 qp->rq.max = cap->max_recv_wr;
1203                 qp->sq.max = cap->max_send_wr;
1204         }
1205
1206         qp->rq.max_gs = cap->max_recv_sge;
1207         qp->sq.max_gs = max_t(int, cap->max_send_sge,
1208                               ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1209                                     MTHCA_INLINE_CHUNK_SIZE) /
1210                               sizeof (struct mthca_data_seg));
1211
1212         /*
1213          * For MLX transport we need 2 extra S/G entries:
1214          * one for the header and one for the checksum at the end
1215          */
1216         if ((qp->transport == MLX && qp->sq.max_gs + 2 > dev->limits.max_sg) ||
1217             qp->sq.max_gs > dev->limits.max_sg || qp->rq.max_gs > dev->limits.max_sg)
1218                 return -EINVAL;
1219
1220         return 0;
1221 }
1222
1223 int mthca_alloc_qp(struct mthca_dev *dev,
1224                    struct mthca_pd *pd,
1225                    struct mthca_cq *send_cq,
1226                    struct mthca_cq *recv_cq,
1227                    enum ib_qp_type type,
1228                    enum ib_sig_type send_policy,
1229                    struct ib_qp_cap *cap,
1230                    struct mthca_qp *qp)
1231 {
1232         int err;
1233
1234         err = mthca_set_qp_size(dev, cap, qp);
1235         if (err)
1236                 return err;
1237
1238         switch (type) {
1239         case IB_QPT_RC: qp->transport = RC; break;
1240         case IB_QPT_UC: qp->transport = UC; break;
1241         case IB_QPT_UD: qp->transport = UD; break;
1242         default: return -EINVAL;
1243         }
1244
1245         qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1246         if (qp->qpn == -1)
1247                 return -ENOMEM;
1248
1249         err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1250                                     send_policy, qp);
1251         if (err) {
1252                 mthca_free(&dev->qp_table.alloc, qp->qpn);
1253                 return err;
1254         }
1255
1256         spin_lock_irq(&dev->qp_table.lock);
1257         mthca_array_set(&dev->qp_table.qp,
1258                         qp->qpn & (dev->limits.num_qps - 1), qp);
1259         spin_unlock_irq(&dev->qp_table.lock);
1260
1261         return 0;
1262 }
1263
1264 int mthca_alloc_sqp(struct mthca_dev *dev,
1265                     struct mthca_pd *pd,
1266                     struct mthca_cq *send_cq,
1267                     struct mthca_cq *recv_cq,
1268                     enum ib_sig_type send_policy,
1269                     struct ib_qp_cap *cap,
1270                     int qpn,
1271                     int port,
1272                     struct mthca_sqp *sqp)
1273 {
1274         u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1275         int err;
1276
1277         err = mthca_set_qp_size(dev, cap, &sqp->qp);
1278         if (err)
1279                 return err;
1280
1281         sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1282         sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1283                                              &sqp->header_dma, GFP_KERNEL);
1284         if (!sqp->header_buf)
1285                 return -ENOMEM;
1286
1287         spin_lock_irq(&dev->qp_table.lock);
1288         if (mthca_array_get(&dev->qp_table.qp, mqpn))
1289                 err = -EBUSY;
1290         else
1291                 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1292         spin_unlock_irq(&dev->qp_table.lock);
1293
1294         if (err)
1295                 goto err_out;
1296
1297         sqp->port = port;
1298         sqp->qp.qpn       = mqpn;
1299         sqp->qp.transport = MLX;
1300
1301         err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1302                                     send_policy, &sqp->qp);
1303         if (err)
1304                 goto err_out_free;
1305
1306         atomic_inc(&pd->sqp_count);
1307
1308         return 0;
1309
1310  err_out_free:
1311         /*
1312          * Lock CQs here, so that CQ polling code can do QP lookup
1313          * without taking a lock.
1314          */
1315         spin_lock_irq(&send_cq->lock);
1316         if (send_cq != recv_cq)
1317                 spin_lock(&recv_cq->lock);
1318
1319         spin_lock(&dev->qp_table.lock);
1320         mthca_array_clear(&dev->qp_table.qp, mqpn);
1321         spin_unlock(&dev->qp_table.lock);
1322
1323         if (send_cq != recv_cq)
1324                 spin_unlock(&recv_cq->lock);
1325         spin_unlock_irq(&send_cq->lock);
1326
1327  err_out:
1328         dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1329                           sqp->header_buf, sqp->header_dma);
1330
1331         return err;
1332 }
1333
1334 void mthca_free_qp(struct mthca_dev *dev,
1335                    struct mthca_qp *qp)
1336 {
1337         u8 status;
1338         struct mthca_cq *send_cq;
1339         struct mthca_cq *recv_cq;
1340
1341         send_cq = to_mcq(qp->ibqp.send_cq);
1342         recv_cq = to_mcq(qp->ibqp.recv_cq);
1343
1344         /*
1345          * Lock CQs here, so that CQ polling code can do QP lookup
1346          * without taking a lock.
1347          */
1348         spin_lock_irq(&send_cq->lock);
1349         if (send_cq != recv_cq)
1350                 spin_lock(&recv_cq->lock);
1351
1352         spin_lock(&dev->qp_table.lock);
1353         mthca_array_clear(&dev->qp_table.qp,
1354                           qp->qpn & (dev->limits.num_qps - 1));
1355         spin_unlock(&dev->qp_table.lock);
1356
1357         if (send_cq != recv_cq)
1358                 spin_unlock(&recv_cq->lock);
1359         spin_unlock_irq(&send_cq->lock);
1360
1361         atomic_dec(&qp->refcount);
1362         wait_event(qp->wait, !atomic_read(&qp->refcount));
1363
1364         if (qp->state != IB_QPS_RESET)
1365                 mthca_MODIFY_QP(dev, MTHCA_TRANS_ANY2RST, qp->qpn, 0, NULL, 0, &status);
1366
1367         /*
1368          * If this is a userspace QP, the buffers, MR, CQs and so on
1369          * will be cleaned up in userspace, so all we have to do is
1370          * unref the mem-free tables and free the QPN in our table.
1371          */
1372         if (!qp->ibqp.uobject) {
1373                 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
1374                                qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1375                 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
1376                         mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
1377                                        qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1378
1379                 mthca_free_memfree(dev, qp);
1380                 mthca_free_wqe_buf(dev, qp);
1381         }
1382
1383         mthca_unmap_memfree(dev, qp);
1384
1385         if (is_sqp(dev, qp)) {
1386                 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1387                 dma_free_coherent(&dev->pdev->dev,
1388                                   to_msqp(qp)->header_buf_size,
1389                                   to_msqp(qp)->header_buf,
1390                                   to_msqp(qp)->header_dma);
1391         } else
1392                 mthca_free(&dev->qp_table.alloc, qp->qpn);
1393 }
1394
1395 /* Create UD header for an MLX send and build a data segment for it */
1396 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1397                             int ind, struct ib_send_wr *wr,
1398                             struct mthca_mlx_seg *mlx,
1399                             struct mthca_data_seg *data)
1400 {
1401         int header_size;
1402         int err;
1403         u16 pkey;
1404
1405         ib_ud_header_init(256, /* assume a MAD */
1406                           sqp->ud_header.grh_present,
1407                           &sqp->ud_header);
1408
1409         err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1410         if (err)
1411                 return err;
1412         mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1413         mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1414                                   (sqp->ud_header.lrh.destination_lid ==
1415                                    IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1416                                   (sqp->ud_header.lrh.service_level << 8));
1417         mlx->rlid = sqp->ud_header.lrh.destination_lid;
1418         mlx->vcrc = 0;
1419
1420         switch (wr->opcode) {
1421         case IB_WR_SEND:
1422                 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1423                 sqp->ud_header.immediate_present = 0;
1424                 break;
1425         case IB_WR_SEND_WITH_IMM:
1426                 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1427                 sqp->ud_header.immediate_present = 1;
1428                 sqp->ud_header.immediate_data = wr->imm_data;
1429                 break;
1430         default:
1431                 return -EINVAL;
1432         }
1433
1434         sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1435         if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1436                 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1437         sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1438         if (!sqp->qp.ibqp.qp_num)
1439                 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1440                                    sqp->pkey_index, &pkey);
1441         else
1442                 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1443                                    wr->wr.ud.pkey_index, &pkey);
1444         sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1445         sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1446         sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1447         sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1448                                                sqp->qkey : wr->wr.ud.remote_qkey);
1449         sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1450
1451         header_size = ib_ud_header_pack(&sqp->ud_header,
1452                                         sqp->header_buf +
1453                                         ind * MTHCA_UD_HEADER_SIZE);
1454
1455         data->byte_count = cpu_to_be32(header_size);
1456         data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1457         data->addr       = cpu_to_be64(sqp->header_dma +
1458                                        ind * MTHCA_UD_HEADER_SIZE);
1459
1460         return 0;
1461 }
1462
1463 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1464                                     struct ib_cq *ib_cq)
1465 {
1466         unsigned cur;
1467         struct mthca_cq *cq;
1468
1469         cur = wq->head - wq->tail;
1470         if (likely(cur + nreq < wq->max))
1471                 return 0;
1472
1473         cq = to_mcq(ib_cq);
1474         spin_lock(&cq->lock);
1475         cur = wq->head - wq->tail;
1476         spin_unlock(&cq->lock);
1477
1478         return cur + nreq >= wq->max;
1479 }
1480
1481 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1482                           struct ib_send_wr **bad_wr)
1483 {
1484         struct mthca_dev *dev = to_mdev(ibqp->device);
1485         struct mthca_qp *qp = to_mqp(ibqp);
1486         void *wqe;
1487         void *prev_wqe;
1488         unsigned long flags;
1489         int err = 0;
1490         int nreq;
1491         int i;
1492         int size;
1493         int size0 = 0;
1494         u32 f0 = 0;
1495         int ind;
1496         u8 op0 = 0;
1497
1498         spin_lock_irqsave(&qp->sq.lock, flags);
1499
1500         /* XXX check that state is OK to post send */
1501
1502         ind = qp->sq.next_ind;
1503
1504         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1505                 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1506                         mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1507                                         " %d max, %d nreq)\n", qp->qpn,
1508                                         qp->sq.head, qp->sq.tail,
1509                                         qp->sq.max, nreq);
1510                         err = -ENOMEM;
1511                         *bad_wr = wr;
1512                         goto out;
1513                 }
1514
1515                 wqe = get_send_wqe(qp, ind);
1516                 prev_wqe = qp->sq.last;
1517                 qp->sq.last = wqe;
1518
1519                 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1520                 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1521                 ((struct mthca_next_seg *) wqe)->flags =
1522                         ((wr->send_flags & IB_SEND_SIGNALED) ?
1523                          cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1524                         ((wr->send_flags & IB_SEND_SOLICITED) ?
1525                          cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1526                         cpu_to_be32(1);
1527                 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1528                     wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1529                         ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1530
1531                 wqe += sizeof (struct mthca_next_seg);
1532                 size = sizeof (struct mthca_next_seg) / 16;
1533
1534                 switch (qp->transport) {
1535                 case RC:
1536                         switch (wr->opcode) {
1537                         case IB_WR_ATOMIC_CMP_AND_SWP:
1538                         case IB_WR_ATOMIC_FETCH_AND_ADD:
1539                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1540                                         cpu_to_be64(wr->wr.atomic.remote_addr);
1541                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1542                                         cpu_to_be32(wr->wr.atomic.rkey);
1543                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1544
1545                                 wqe += sizeof (struct mthca_raddr_seg);
1546
1547                                 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1548                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1549                                                 cpu_to_be64(wr->wr.atomic.swap);
1550                                         ((struct mthca_atomic_seg *) wqe)->compare =
1551                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1552                                 } else {
1553                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1554                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1555                                         ((struct mthca_atomic_seg *) wqe)->compare = 0;
1556                                 }
1557
1558                                 wqe += sizeof (struct mthca_atomic_seg);
1559                                 size += (sizeof (struct mthca_raddr_seg) +
1560                                          sizeof (struct mthca_atomic_seg)) / 16;
1561                                 break;
1562
1563                         case IB_WR_RDMA_WRITE:
1564                         case IB_WR_RDMA_WRITE_WITH_IMM:
1565                         case IB_WR_RDMA_READ:
1566                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1567                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1568                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1569                                         cpu_to_be32(wr->wr.rdma.rkey);
1570                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1571                                 wqe += sizeof (struct mthca_raddr_seg);
1572                                 size += sizeof (struct mthca_raddr_seg) / 16;
1573                                 break;
1574
1575                         default:
1576                                 /* No extra segments required for sends */
1577                                 break;
1578                         }
1579
1580                         break;
1581
1582                 case UC:
1583                         switch (wr->opcode) {
1584                         case IB_WR_RDMA_WRITE:
1585                         case IB_WR_RDMA_WRITE_WITH_IMM:
1586                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1587                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1588                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1589                                         cpu_to_be32(wr->wr.rdma.rkey);
1590                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1591                                 wqe += sizeof (struct mthca_raddr_seg);
1592                                 size += sizeof (struct mthca_raddr_seg) / 16;
1593                                 break;
1594
1595                         default:
1596                                 /* No extra segments required for sends */
1597                                 break;
1598                         }
1599
1600                         break;
1601
1602                 case UD:
1603                         ((struct mthca_tavor_ud_seg *) wqe)->lkey =
1604                                 cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1605                         ((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1606                                 cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1607                         ((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1608                                 cpu_to_be32(wr->wr.ud.remote_qpn);
1609                         ((struct mthca_tavor_ud_seg *) wqe)->qkey =
1610                                 cpu_to_be32(wr->wr.ud.remote_qkey);
1611
1612                         wqe += sizeof (struct mthca_tavor_ud_seg);
1613                         size += sizeof (struct mthca_tavor_ud_seg) / 16;
1614                         break;
1615
1616                 case MLX:
1617                         err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1618                                                wqe - sizeof (struct mthca_next_seg),
1619                                                wqe);
1620                         if (err) {
1621                                 *bad_wr = wr;
1622                                 goto out;
1623                         }
1624                         wqe += sizeof (struct mthca_data_seg);
1625                         size += sizeof (struct mthca_data_seg) / 16;
1626                         break;
1627                 }
1628
1629                 if (wr->num_sge > qp->sq.max_gs) {
1630                         mthca_err(dev, "too many gathers\n");
1631                         err = -EINVAL;
1632                         *bad_wr = wr;
1633                         goto out;
1634                 }
1635
1636                 for (i = 0; i < wr->num_sge; ++i) {
1637                         ((struct mthca_data_seg *) wqe)->byte_count =
1638                                 cpu_to_be32(wr->sg_list[i].length);
1639                         ((struct mthca_data_seg *) wqe)->lkey =
1640                                 cpu_to_be32(wr->sg_list[i].lkey);
1641                         ((struct mthca_data_seg *) wqe)->addr =
1642                                 cpu_to_be64(wr->sg_list[i].addr);
1643                         wqe += sizeof (struct mthca_data_seg);
1644                         size += sizeof (struct mthca_data_seg) / 16;
1645                 }
1646
1647                 /* Add one more inline data segment for ICRC */
1648                 if (qp->transport == MLX) {
1649                         ((struct mthca_data_seg *) wqe)->byte_count =
1650                                 cpu_to_be32((1 << 31) | 4);
1651                         ((u32 *) wqe)[1] = 0;
1652                         wqe += sizeof (struct mthca_data_seg);
1653                         size += sizeof (struct mthca_data_seg) / 16;
1654                 }
1655
1656                 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1657
1658                 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1659                         mthca_err(dev, "opcode invalid\n");
1660                         err = -EINVAL;
1661                         *bad_wr = wr;
1662                         goto out;
1663                 }
1664
1665                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1666                         cpu_to_be32(((ind << qp->sq.wqe_shift) +
1667                                      qp->send_wqe_offset) |
1668                                     mthca_opcode[wr->opcode]);
1669                 wmb();
1670                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1671                         cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size);
1672
1673                 if (!size0) {
1674                         size0 = size;
1675                         op0   = mthca_opcode[wr->opcode];
1676                 }
1677
1678                 ++ind;
1679                 if (unlikely(ind >= qp->sq.max))
1680                         ind -= qp->sq.max;
1681         }
1682
1683 out:
1684         if (likely(nreq)) {
1685                 __be32 doorbell[2];
1686
1687                 doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1688                                            qp->send_wqe_offset) | f0 | op0);
1689                 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1690
1691                 wmb();
1692
1693                 mthca_write64(doorbell,
1694                               dev->kar + MTHCA_SEND_DOORBELL,
1695                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1696         }
1697
1698         qp->sq.next_ind = ind;
1699         qp->sq.head    += nreq;
1700
1701         spin_unlock_irqrestore(&qp->sq.lock, flags);
1702         return err;
1703 }
1704
1705 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1706                              struct ib_recv_wr **bad_wr)
1707 {
1708         struct mthca_dev *dev = to_mdev(ibqp->device);
1709         struct mthca_qp *qp = to_mqp(ibqp);
1710         __be32 doorbell[2];
1711         unsigned long flags;
1712         int err = 0;
1713         int nreq;
1714         int i;
1715         int size;
1716         int size0 = 0;
1717         int ind;
1718         void *wqe;
1719         void *prev_wqe;
1720
1721         spin_lock_irqsave(&qp->rq.lock, flags);
1722
1723         /* XXX check that state is OK to post receive */
1724
1725         ind = qp->rq.next_ind;
1726
1727         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1728                 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1729                         nreq = 0;
1730
1731                         doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1732                         doorbell[1] = cpu_to_be32(qp->qpn << 8);
1733
1734                         wmb();
1735
1736                         mthca_write64(doorbell,
1737                                       dev->kar + MTHCA_RECEIVE_DOORBELL,
1738                                       MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1739
1740                         qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1741                         size0 = 0;
1742                 }
1743
1744                 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1745                         mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1746                                         " %d max, %d nreq)\n", qp->qpn,
1747                                         qp->rq.head, qp->rq.tail,
1748                                         qp->rq.max, nreq);
1749                         err = -ENOMEM;
1750                         *bad_wr = wr;
1751                         goto out;
1752                 }
1753
1754                 wqe = get_recv_wqe(qp, ind);
1755                 prev_wqe = qp->rq.last;
1756                 qp->rq.last = wqe;
1757
1758                 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1759                 ((struct mthca_next_seg *) wqe)->ee_nds =
1760                         cpu_to_be32(MTHCA_NEXT_DBD);
1761                 ((struct mthca_next_seg *) wqe)->flags = 0;
1762
1763                 wqe += sizeof (struct mthca_next_seg);
1764                 size = sizeof (struct mthca_next_seg) / 16;
1765
1766                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1767                         err = -EINVAL;
1768                         *bad_wr = wr;
1769                         goto out;
1770                 }
1771
1772                 for (i = 0; i < wr->num_sge; ++i) {
1773                         ((struct mthca_data_seg *) wqe)->byte_count =
1774                                 cpu_to_be32(wr->sg_list[i].length);
1775                         ((struct mthca_data_seg *) wqe)->lkey =
1776                                 cpu_to_be32(wr->sg_list[i].lkey);
1777                         ((struct mthca_data_seg *) wqe)->addr =
1778                                 cpu_to_be64(wr->sg_list[i].addr);
1779                         wqe += sizeof (struct mthca_data_seg);
1780                         size += sizeof (struct mthca_data_seg) / 16;
1781                 }
1782
1783                 qp->wrid[ind] = wr->wr_id;
1784
1785                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1786                         cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1787                 wmb();
1788                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1789                         cpu_to_be32(MTHCA_NEXT_DBD | size);
1790
1791                 if (!size0)
1792                         size0 = size;
1793
1794                 ++ind;
1795                 if (unlikely(ind >= qp->rq.max))
1796                         ind -= qp->rq.max;
1797         }
1798
1799 out:
1800         if (likely(nreq)) {
1801                 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1802                 doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
1803
1804                 wmb();
1805
1806                 mthca_write64(doorbell,
1807                               dev->kar + MTHCA_RECEIVE_DOORBELL,
1808                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1809         }
1810
1811         qp->rq.next_ind = ind;
1812         qp->rq.head    += nreq;
1813
1814         spin_unlock_irqrestore(&qp->rq.lock, flags);
1815         return err;
1816 }
1817
1818 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1819                           struct ib_send_wr **bad_wr)
1820 {
1821         struct mthca_dev *dev = to_mdev(ibqp->device);
1822         struct mthca_qp *qp = to_mqp(ibqp);
1823         void *wqe;
1824         void *prev_wqe;
1825         unsigned long flags;
1826         int err = 0;
1827         int nreq;
1828         int i;
1829         int size;
1830         int size0 = 0;
1831         u32 f0 = 0;
1832         int ind;
1833         u8 op0 = 0;
1834
1835         spin_lock_irqsave(&qp->sq.lock, flags);
1836
1837         /* XXX check that state is OK to post send */
1838
1839         ind = qp->sq.head & (qp->sq.max - 1);
1840
1841         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1842                 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1843                         mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1844                                         " %d max, %d nreq)\n", qp->qpn,
1845                                         qp->sq.head, qp->sq.tail,
1846                                         qp->sq.max, nreq);
1847                         err = -ENOMEM;
1848                         *bad_wr = wr;
1849                         goto out;
1850                 }
1851
1852                 wqe = get_send_wqe(qp, ind);
1853                 prev_wqe = qp->sq.last;
1854                 qp->sq.last = wqe;
1855
1856                 ((struct mthca_next_seg *) wqe)->flags =
1857                         ((wr->send_flags & IB_SEND_SIGNALED) ?
1858                          cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1859                         ((wr->send_flags & IB_SEND_SOLICITED) ?
1860                          cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1861                         cpu_to_be32(1);
1862                 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1863                     wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1864                         ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1865
1866                 wqe += sizeof (struct mthca_next_seg);
1867                 size = sizeof (struct mthca_next_seg) / 16;
1868
1869                 switch (qp->transport) {
1870                 case RC:
1871                         switch (wr->opcode) {
1872                         case IB_WR_ATOMIC_CMP_AND_SWP:
1873                         case IB_WR_ATOMIC_FETCH_AND_ADD:
1874                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1875                                         cpu_to_be64(wr->wr.atomic.remote_addr);
1876                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1877                                         cpu_to_be32(wr->wr.atomic.rkey);
1878                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1879
1880                                 wqe += sizeof (struct mthca_raddr_seg);
1881
1882                                 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1883                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1884                                                 cpu_to_be64(wr->wr.atomic.swap);
1885                                         ((struct mthca_atomic_seg *) wqe)->compare =
1886                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1887                                 } else {
1888                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1889                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1890                                         ((struct mthca_atomic_seg *) wqe)->compare = 0;
1891                                 }
1892
1893                                 wqe += sizeof (struct mthca_atomic_seg);
1894                                 size += (sizeof (struct mthca_raddr_seg) +
1895                                          sizeof (struct mthca_atomic_seg)) / 16;
1896                                 break;
1897
1898                         case IB_WR_RDMA_READ:
1899                         case IB_WR_RDMA_WRITE:
1900                         case IB_WR_RDMA_WRITE_WITH_IMM:
1901                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1902                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1903                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1904                                         cpu_to_be32(wr->wr.rdma.rkey);
1905                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1906                                 wqe += sizeof (struct mthca_raddr_seg);
1907                                 size += sizeof (struct mthca_raddr_seg) / 16;
1908                                 break;
1909
1910                         default:
1911                                 /* No extra segments required for sends */
1912                                 break;
1913                         }
1914
1915                         break;
1916
1917                 case UC:
1918                         switch (wr->opcode) {
1919                         case IB_WR_RDMA_WRITE:
1920                         case IB_WR_RDMA_WRITE_WITH_IMM:
1921                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1922                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1923                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1924                                         cpu_to_be32(wr->wr.rdma.rkey);
1925                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1926                                 wqe += sizeof (struct mthca_raddr_seg);
1927                                 size += sizeof (struct mthca_raddr_seg) / 16;
1928                                 break;
1929
1930                         default:
1931                                 /* No extra segments required for sends */
1932                                 break;
1933                         }
1934
1935                         break;
1936
1937                 case UD:
1938                         memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
1939                                to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1940                         ((struct mthca_arbel_ud_seg *) wqe)->dqpn =
1941                                 cpu_to_be32(wr->wr.ud.remote_qpn);
1942                         ((struct mthca_arbel_ud_seg *) wqe)->qkey =
1943                                 cpu_to_be32(wr->wr.ud.remote_qkey);
1944
1945                         wqe += sizeof (struct mthca_arbel_ud_seg);
1946                         size += sizeof (struct mthca_arbel_ud_seg) / 16;
1947                         break;
1948
1949                 case MLX:
1950                         err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1951                                                wqe - sizeof (struct mthca_next_seg),
1952                                                wqe);
1953                         if (err) {
1954                                 *bad_wr = wr;
1955                                 goto out;
1956                         }
1957                         wqe += sizeof (struct mthca_data_seg);
1958                         size += sizeof (struct mthca_data_seg) / 16;
1959                         break;
1960                 }
1961
1962                 if (wr->num_sge > qp->sq.max_gs) {
1963                         mthca_err(dev, "too many gathers\n");
1964                         err = -EINVAL;
1965                         *bad_wr = wr;
1966                         goto out;
1967                 }
1968
1969                 for (i = 0; i < wr->num_sge; ++i) {
1970                         ((struct mthca_data_seg *) wqe)->byte_count =
1971                                 cpu_to_be32(wr->sg_list[i].length);
1972                         ((struct mthca_data_seg *) wqe)->lkey =
1973                                 cpu_to_be32(wr->sg_list[i].lkey);
1974                         ((struct mthca_data_seg *) wqe)->addr =
1975                                 cpu_to_be64(wr->sg_list[i].addr);
1976                         wqe += sizeof (struct mthca_data_seg);
1977                         size += sizeof (struct mthca_data_seg) / 16;
1978                 }
1979
1980                 /* Add one more inline data segment for ICRC */
1981                 if (qp->transport == MLX) {
1982                         ((struct mthca_data_seg *) wqe)->byte_count =
1983                                 cpu_to_be32((1 << 31) | 4);
1984                         ((u32 *) wqe)[1] = 0;
1985                         wqe += sizeof (struct mthca_data_seg);
1986                         size += sizeof (struct mthca_data_seg) / 16;
1987                 }
1988
1989                 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1990
1991                 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1992                         mthca_err(dev, "opcode invalid\n");
1993                         err = -EINVAL;
1994                         *bad_wr = wr;
1995                         goto out;
1996                 }
1997
1998                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1999                         cpu_to_be32(((ind << qp->sq.wqe_shift) +
2000                                      qp->send_wqe_offset) |
2001                                     mthca_opcode[wr->opcode]);
2002                 wmb();
2003                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
2004                         cpu_to_be32(MTHCA_NEXT_DBD | size);
2005
2006                 if (!size0) {
2007                         size0 = size;
2008                         op0   = mthca_opcode[wr->opcode];
2009                 }
2010
2011                 ++ind;
2012                 if (unlikely(ind >= qp->sq.max))
2013                         ind -= qp->sq.max;
2014         }
2015
2016 out:
2017         if (likely(nreq)) {
2018                 __be32 doorbell[2];
2019
2020                 doorbell[0] = cpu_to_be32((nreq << 24)                  |
2021                                           ((qp->sq.head & 0xffff) << 8) |
2022                                           f0 | op0);
2023                 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
2024
2025                 qp->sq.head += nreq;
2026
2027                 /*
2028                  * Make sure that descriptors are written before
2029                  * doorbell record.
2030                  */
2031                 wmb();
2032                 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2033
2034                 /*
2035                  * Make sure doorbell record is written before we
2036                  * write MMIO send doorbell.
2037                  */
2038                 wmb();
2039                 mthca_write64(doorbell,
2040                               dev->kar + MTHCA_SEND_DOORBELL,
2041                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2042         }
2043
2044         spin_unlock_irqrestore(&qp->sq.lock, flags);
2045         return err;
2046 }
2047
2048 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2049                              struct ib_recv_wr **bad_wr)
2050 {
2051         struct mthca_dev *dev = to_mdev(ibqp->device);
2052         struct mthca_qp *qp = to_mqp(ibqp);
2053         unsigned long flags;
2054         int err = 0;
2055         int nreq;
2056         int ind;
2057         int i;
2058         void *wqe;
2059
2060         spin_lock_irqsave(&qp->rq.lock, flags);
2061
2062         /* XXX check that state is OK to post receive */
2063
2064         ind = qp->rq.head & (qp->rq.max - 1);
2065
2066         for (nreq = 0; wr; ++nreq, wr = wr->next) {
2067                 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2068                         mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2069                                         " %d max, %d nreq)\n", qp->qpn,
2070                                         qp->rq.head, qp->rq.tail,
2071                                         qp->rq.max, nreq);
2072                         err = -ENOMEM;
2073                         *bad_wr = wr;
2074                         goto out;
2075                 }
2076
2077                 wqe = get_recv_wqe(qp, ind);
2078
2079                 ((struct mthca_next_seg *) wqe)->flags = 0;
2080
2081                 wqe += sizeof (struct mthca_next_seg);
2082
2083                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2084                         err = -EINVAL;
2085                         *bad_wr = wr;
2086                         goto out;
2087                 }
2088
2089                 for (i = 0; i < wr->num_sge; ++i) {
2090                         ((struct mthca_data_seg *) wqe)->byte_count =
2091                                 cpu_to_be32(wr->sg_list[i].length);
2092                         ((struct mthca_data_seg *) wqe)->lkey =
2093                                 cpu_to_be32(wr->sg_list[i].lkey);
2094                         ((struct mthca_data_seg *) wqe)->addr =
2095                                 cpu_to_be64(wr->sg_list[i].addr);
2096                         wqe += sizeof (struct mthca_data_seg);
2097                 }
2098
2099                 if (i < qp->rq.max_gs) {
2100                         ((struct mthca_data_seg *) wqe)->byte_count = 0;
2101                         ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
2102                         ((struct mthca_data_seg *) wqe)->addr = 0;
2103                 }
2104
2105                 qp->wrid[ind] = wr->wr_id;
2106
2107                 ++ind;
2108                 if (unlikely(ind >= qp->rq.max))
2109                         ind -= qp->rq.max;
2110         }
2111 out:
2112         if (likely(nreq)) {
2113                 qp->rq.head += nreq;
2114
2115                 /*
2116                  * Make sure that descriptors are written before
2117                  * doorbell record.
2118                  */
2119                 wmb();
2120                 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2121         }
2122
2123         spin_unlock_irqrestore(&qp->rq.lock, flags);
2124         return err;
2125 }
2126
2127 int mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2128                        int index, int *dbd, __be32 *new_wqe)
2129 {
2130         struct mthca_next_seg *next;
2131
2132         /*
2133          * For SRQs, all WQEs generate a CQE, so we're always at the
2134          * end of the doorbell chain.
2135          */
2136         if (qp->ibqp.srq) {
2137                 *new_wqe = 0;
2138                 return 0;
2139         }
2140
2141         if (is_send)
2142                 next = get_send_wqe(qp, index);
2143         else
2144                 next = get_recv_wqe(qp, index);
2145
2146         *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2147         if (next->ee_nds & cpu_to_be32(0x3f))
2148                 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2149                         (next->ee_nds & cpu_to_be32(0x3f));
2150         else
2151                 *new_wqe = 0;
2152
2153         return 0;
2154 }
2155
2156 int __devinit mthca_init_qp_table(struct mthca_dev *dev)
2157 {
2158         int err;
2159         u8 status;
2160         int i;
2161
2162         spin_lock_init(&dev->qp_table.lock);
2163
2164         /*
2165          * We reserve 2 extra QPs per port for the special QPs.  The
2166          * special QP for port 1 has to be even, so round up.
2167          */
2168         dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2169         err = mthca_alloc_init(&dev->qp_table.alloc,
2170                                dev->limits.num_qps,
2171                                (1 << 24) - 1,
2172                                dev->qp_table.sqp_start +
2173                                MTHCA_MAX_PORTS * 2);
2174         if (err)
2175                 return err;
2176
2177         err = mthca_array_init(&dev->qp_table.qp,
2178                                dev->limits.num_qps);
2179         if (err) {
2180                 mthca_alloc_cleanup(&dev->qp_table.alloc);
2181                 return err;
2182         }
2183
2184         for (i = 0; i < 2; ++i) {
2185                 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2186                                             dev->qp_table.sqp_start + i * 2,
2187                                             &status);
2188                 if (err)
2189                         goto err_out;
2190                 if (status) {
2191                         mthca_warn(dev, "CONF_SPECIAL_QP returned "
2192                                    "status %02x, aborting.\n",
2193                                    status);
2194                         err = -EINVAL;
2195                         goto err_out;
2196                 }
2197         }
2198         return 0;
2199
2200  err_out:
2201         for (i = 0; i < 2; ++i)
2202                 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2203
2204         mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2205         mthca_alloc_cleanup(&dev->qp_table.alloc);
2206
2207         return err;
2208 }
2209
2210 void __devexit mthca_cleanup_qp_table(struct mthca_dev *dev)
2211 {
2212         int i;
2213         u8 status;
2214
2215         for (i = 0; i < 2; ++i)
2216                 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2217
2218         mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2219         mthca_alloc_cleanup(&dev->qp_table.alloc);
2220 }