sys/dev/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
  36 #include <linux/string.h>
  37 #include <linux/slab.h>
  38 #include <linux/sched.h>
  39
  40 #include <asm/io.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 enum {
 103         MTHCA_SEND_DOORBELL_FENCE = 1 << 5
 104 };
 105
 106 struct mthca_qp_path {
 107         __be32 port_pkey;
 108         u8     rnr_retry;
 109         u8     g_mylmc;
 110         __be16 rlid;
 111         u8     ackto;
 112         u8     mgid_index;
 113         u8     static_rate;
 114         u8     hop_limit;
 115         __be32 sl_tclass_flowlabel;
 116         u8     rgid[16];
 117 } __attribute__((packed));
 118
 119 struct mthca_qp_context {
 120         __be32 flags;
 121         __be32 tavor_sched_queue; /* Reserved on Arbel */
 122         u8     mtu_msgmax;
 123         u8     rq_size_stride;  /* Reserved on Tavor */
 124         u8     sq_size_stride;  /* Reserved on Tavor */
 125         u8     rlkey_arbel_sched_queue; /* Reserved on Tavor */
 126         __be32 usr_page;
 127         __be32 local_qpn;
 128         __be32 remote_qpn;
 129         u32    reserved1[2];
 130         struct mthca_qp_path pri_path;
 131         struct mthca_qp_path alt_path;
 132         __be32 rdd;
 133         __be32 pd;
 134         __be32 wqe_base;
 135         __be32 wqe_lkey;
 136         __be32 params1;
 137         __be32 reserved2;
 138         __be32 next_send_psn;
 139         __be32 cqn_snd;
 140         __be32 snd_wqe_base_l;  /* Next send WQE on Tavor */
 141         __be32 snd_db_index;    /* (debugging only entries) */
 142         __be32 last_acked_psn;
 143         __be32 ssn;
 144         __be32 params2;
 145         __be32 rnr_nextrecvpsn;
 146         __be32 ra_buff_indx;
 147         __be32 cqn_rcv;
 148         __be32 rcv_wqe_base_l;  /* Next recv WQE on Tavor */
 149         __be32 rcv_db_index;    /* (debugging only entries) */
 150         __be32 qkey;
 151         __be32 srqn;
 152         __be32 rmsn;
 153         __be16 rq_wqe_counter;  /* reserved on Tavor */
 154         __be16 sq_wqe_counter;  /* reserved on Tavor */
 155         u32    reserved3[18];
 156 } __attribute__((packed));
 157
 158 struct mthca_qp_param {
 159         __be32 opt_param_mask;
 160         u32    reserved1;
 161         struct mthca_qp_context context;
 162         u32    reserved2[62];
 163 } __attribute__((packed));
 164
 165 enum {
 166         MTHCA_QP_OPTPAR_ALT_ADDR_PATH     = 1 << 0,
 167         MTHCA_QP_OPTPAR_RRE               = 1 << 1,
 168         MTHCA_QP_OPTPAR_RAE               = 1 << 2,
 169         MTHCA_QP_OPTPAR_RWE               = 1 << 3,
 170         MTHCA_QP_OPTPAR_PKEY_INDEX        = 1 << 4,
 171         MTHCA_QP_OPTPAR_Q_KEY             = 1 << 5,
 172         MTHCA_QP_OPTPAR_RNR_TIMEOUT       = 1 << 6,
 173         MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
 174         MTHCA_QP_OPTPAR_SRA_MAX           = 1 << 8,
 175         MTHCA_QP_OPTPAR_RRA_MAX           = 1 << 9,
 176         MTHCA_QP_OPTPAR_PM_STATE          = 1 << 10,
 177         MTHCA_QP_OPTPAR_PORT_NUM          = 1 << 11,
 178         MTHCA_QP_OPTPAR_RETRY_COUNT       = 1 << 12,
 179         MTHCA_QP_OPTPAR_ALT_RNR_RETRY     = 1 << 13,
 180         MTHCA_QP_OPTPAR_ACK_TIMEOUT       = 1 << 14,
 181         MTHCA_QP_OPTPAR_RNR_RETRY         = 1 << 15,
 182         MTHCA_QP_OPTPAR_SCHED_QUEUE       = 1 << 16
 183 };
 184
 185 static const u8 mthca_opcode[] = {
 186         [IB_WR_SEND]                 = MTHCA_OPCODE_SEND,
 187         [IB_WR_SEND_WITH_IMM]        = MTHCA_OPCODE_SEND_IMM,
 188         [IB_WR_RDMA_WRITE]           = MTHCA_OPCODE_RDMA_WRITE,
 189         [IB_WR_RDMA_WRITE_WITH_IMM]  = MTHCA_OPCODE_RDMA_WRITE_IMM,
 190         [IB_WR_RDMA_READ]            = MTHCA_OPCODE_RDMA_READ,
 191         [IB_WR_ATOMIC_CMP_AND_SWP]   = MTHCA_OPCODE_ATOMIC_CS,
 192         [IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
 193 };
 194
 195 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
 196 {
 197         return qp->qpn >= dev->qp_table.sqp_start &&
 198                 qp->qpn <= dev->qp_table.sqp_start + 3;
 199 }
 200
 201 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
 202 {
 203         return qp->qpn >= dev->qp_table.sqp_start &&
 204                 qp->qpn <= dev->qp_table.sqp_start + 1;
 205 }
 206
 207 static void *get_recv_wqe(struct mthca_qp *qp, int n)
 208 {
 209         if (qp->is_direct)
 210                 return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
 211         else
 212                 return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
 213                         ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
 214 }
 215
 216 static void *get_send_wqe(struct mthca_qp *qp, int n)
 217 {
 218         if (qp->is_direct)
 219                 return qp->queue.direct.buf + qp->send_wqe_offset +
 220                         (n << qp->sq.wqe_shift);
 221         else
 222                 return qp->queue.page_list[(qp->send_wqe_offset +
 223                                             (n << qp->sq.wqe_shift)) >>
 224                                            PAGE_SHIFT].buf +
 225                         ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
 226                          (PAGE_SIZE - 1));
 227 }
 228
 229 static void mthca_wq_reset(struct mthca_wq *wq)
 230 {
 231         wq->next_ind  = 0;
 232         wq->last_comp = wq->max - 1;
 233         wq->head      = 0;
 234         wq->tail      = 0;
 235 }
 236
 237 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
 238                     enum ib_event_type event_type)
 239 {
 240         struct mthca_qp *qp;
 241         struct ib_event event;
 242
 243         spin_lock(&dev->qp_table.lock);
 244         qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
 245         if (qp)
 246                 ++qp->refcount;
 247         spin_unlock(&dev->qp_table.lock);
 248
 249         if (!qp) {
 250                 mthca_warn(dev, "Async event %d for bogus QP %08x\n",
 251                            event_type, qpn);
 252                 return;
 253         }
 254
 255         if (event_type == IB_EVENT_PATH_MIG)
 256                 qp->port = qp->alt_port;
 257
 258         event.device      = &dev->ib_dev;
 259         event.event       = event_type;
 260         event.element.qp  = &qp->ibqp;
 261         if (qp->ibqp.event_handler)
 262                 qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
 263
 264         spin_lock(&dev->qp_table.lock);
 265         if (!--qp->refcount)
 266                 wake_up(&qp->wait);
 267         spin_unlock(&dev->qp_table.lock);
 268 }
 269
 270 static int to_mthca_state(enum ib_qp_state ib_state)
 271 {
 272         switch (ib_state) {
 273         case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
 274         case IB_QPS_INIT:  return MTHCA_QP_STATE_INIT;
 275         case IB_QPS_RTR:   return MTHCA_QP_STATE_RTR;
 276         case IB_QPS_RTS:   return MTHCA_QP_STATE_RTS;
 277         case IB_QPS_SQD:   return MTHCA_QP_STATE_SQD;
 278         case IB_QPS_SQE:   return MTHCA_QP_STATE_SQE;
 279         case IB_QPS_ERR:   return MTHCA_QP_STATE_ERR;
 280         default:                return -1;
 281         }
 282 }
 283
 284 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
 285
 286 static int to_mthca_st(int transport)
 287 {
 288         switch (transport) {
 289         case RC:  return MTHCA_QP_ST_RC;
 290         case UC:  return MTHCA_QP_ST_UC;
 291         case UD:  return MTHCA_QP_ST_UD;
 292         case RD:  return MTHCA_QP_ST_RD;
 293         case MLX: return MTHCA_QP_ST_MLX;
 294         default:  return -1;
 295         }
 296 }
 297
 298 static void store_attrs(struct mthca_sqp *sqp, const struct ib_qp_attr *attr,
 299                         int attr_mask)
 300 {
 301         if (attr_mask & IB_QP_PKEY_INDEX)
 302                 sqp->pkey_index = attr->pkey_index;
 303         if (attr_mask & IB_QP_QKEY)
 304                 sqp->qkey = attr->qkey;
 305         if (attr_mask & IB_QP_SQ_PSN)
 306                 sqp->send_psn = attr->sq_psn;
 307 }
 308
 309 static void init_port(struct mthca_dev *dev, int port)
 310 {
 311         int err;
 312         struct mthca_init_ib_param param;
 313
 314         memset(&param, 0, sizeof param);
 315
 316         param.port_width = dev->limits.port_width_cap;
 317         param.vl_cap     = dev->limits.vl_cap;
 318         param.mtu_cap    = dev->limits.mtu_cap;
 319         param.gid_cap    = dev->limits.gid_table_len;
 320         param.pkey_cap   = dev->limits.pkey_table_len;
 321
 322         err = mthca_INIT_IB(dev, &param, port);
 323         if (err)
 324                 mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
 325 }
 326
 327 static __be32 get_hw_access_flags(struct mthca_qp *qp, const struct ib_qp_attr *attr,
 328                                   int attr_mask)
 329 {
 330         u8 dest_rd_atomic;
 331         u32 access_flags;
 332         u32 hw_access_flags = 0;
 333
 334         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 335                 dest_rd_atomic = attr->max_dest_rd_atomic;
 336         else
 337                 dest_rd_atomic = qp->resp_depth;
 338
 339         if (attr_mask & IB_QP_ACCESS_FLAGS)
 340                 access_flags = attr->qp_access_flags;
 341         else
 342                 access_flags = qp->atomic_rd_en;
 343
 344         if (!dest_rd_atomic)
 345                 access_flags &= IB_ACCESS_REMOTE_WRITE;
 346
 347         if (access_flags & IB_ACCESS_REMOTE_READ)
 348                 hw_access_flags |= MTHCA_QP_BIT_RRE;
 349         if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
 350                 hw_access_flags |= MTHCA_QP_BIT_RAE;
 351         if (access_flags & IB_ACCESS_REMOTE_WRITE)
 352                 hw_access_flags |= MTHCA_QP_BIT_RWE;
 353
 354         return cpu_to_be32(hw_access_flags);
 355 }
 356
 357 static inline enum ib_qp_state to_ib_qp_state(int mthca_state)
 358 {
 359         switch (mthca_state) {
 360         case MTHCA_QP_STATE_RST:      return IB_QPS_RESET;
 361         case MTHCA_QP_STATE_INIT:     return IB_QPS_INIT;
 362         case MTHCA_QP_STATE_RTR:      return IB_QPS_RTR;
 363         case MTHCA_QP_STATE_RTS:      return IB_QPS_RTS;
 364         case MTHCA_QP_STATE_DRAINING:
 365         case MTHCA_QP_STATE_SQD:      return IB_QPS_SQD;
 366         case MTHCA_QP_STATE_SQE:      return IB_QPS_SQE;
 367         case MTHCA_QP_STATE_ERR:      return IB_QPS_ERR;
 368         default:                      return -1;
 369         }
 370 }
 371
 372 static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state)
 373 {
 374         switch (mthca_mig_state) {
 375         case 0:  return IB_MIG_ARMED;
 376         case 1:  return IB_MIG_REARM;
 377         case 3:  return IB_MIG_MIGRATED;
 378         default: return -1;
 379         }
 380 }
 381
 382 static int to_ib_qp_access_flags(int mthca_flags)
 383 {
 384         int ib_flags = 0;
 385
 386         if (mthca_flags & MTHCA_QP_BIT_RRE)
 387                 ib_flags |= IB_ACCESS_REMOTE_READ;
 388         if (mthca_flags & MTHCA_QP_BIT_RWE)
 389                 ib_flags |= IB_ACCESS_REMOTE_WRITE;
 390         if (mthca_flags & MTHCA_QP_BIT_RAE)
 391                 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
 392
 393         return ib_flags;
 394 }
 395
 396 static void to_ib_ah_attr(struct mthca_dev *dev, struct ib_ah_attr *ib_ah_attr,
 397                                 struct mthca_qp_path *path)
 398 {
 399         memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
 400         ib_ah_attr->port_num      = (be32_to_cpu(path->port_pkey) >> 24) & 0x3;
 401
 402         if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->limits.num_ports)
 403                 return;
 404
 405         ib_ah_attr->dlid          = be16_to_cpu(path->rlid);
 406         ib_ah_attr->sl            = be32_to_cpu(path->sl_tclass_flowlabel) >> 28;
 407         ib_ah_attr->src_path_bits = path->g_mylmc & 0x7f;
 408         ib_ah_attr->static_rate   = mthca_rate_to_ib(dev,
 409                                                      path->static_rate & 0xf,
 410                                                      ib_ah_attr->port_num);
 411         ib_ah_attr->ah_flags      = (path->g_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
 412         if (ib_ah_attr->ah_flags) {
 413                 ib_ah_attr->grh.sgid_index = path->mgid_index & (dev->limits.gid_table_len - 1);
 414                 ib_ah_attr->grh.hop_limit  = path->hop_limit;
 415                 ib_ah_attr->grh.traffic_class =
 416                         (be32_to_cpu(path->sl_tclass_flowlabel) >> 20) & 0xff;
 417                 ib_ah_attr->grh.flow_label =
 418                         be32_to_cpu(path->sl_tclass_flowlabel) & 0xfffff;
 419                 memcpy(ib_ah_attr->grh.dgid.raw,
 420                         path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
 421         }
 422 }
 423
 424 int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
 425                    struct ib_qp_init_attr *qp_init_attr)
 426 {
 427         struct mthca_dev *dev = to_mdev(ibqp->device);
 428         struct mthca_qp *qp = to_mqp(ibqp);
 429         int err = 0;
 430         struct mthca_mailbox *mailbox = NULL;
 431         struct mthca_qp_param *qp_param;
 432         struct mthca_qp_context *context;
 433         int mthca_state;
 434
 435         mutex_lock(&qp->mutex);
 436
 437         if (qp->state == IB_QPS_RESET) {
 438                 qp_attr->qp_state = IB_QPS_RESET;
 439                 goto done;
 440         }
 441
 442         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 443         if (IS_ERR(mailbox)) {
 444                 err = PTR_ERR(mailbox);
 445                 goto out;
 446         }
 447
 448         err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox);
 449         if (err) {
 450                 mthca_warn(dev, "QUERY_QP failed (%d)\n", err);
 451                 goto out_mailbox;
 452         }
 453
 454         qp_param    = mailbox->buf;
 455         context     = &qp_param->context;
 456         mthca_state = be32_to_cpu(context->flags) >> 28;
 457
 458         qp->state                    = to_ib_qp_state(mthca_state);
 459         qp_attr->qp_state            = qp->state;
 460         qp_attr->path_mtu            = context->mtu_msgmax >> 5;
 461         qp_attr->path_mig_state      =
 462                 to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
 463         qp_attr->qkey                = be32_to_cpu(context->qkey);
 464         qp_attr->rq_psn              = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
 465         qp_attr->sq_psn              = be32_to_cpu(context->next_send_psn) & 0xffffff;
 466         qp_attr->dest_qp_num         = be32_to_cpu(context->remote_qpn) & 0xffffff;
 467         qp_attr->qp_access_flags     =
 468                 to_ib_qp_access_flags(be32_to_cpu(context->params2));
 469
 470         if (qp->transport == RC || qp->transport == UC) {
 471                 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
 472                 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
 473                 qp_attr->alt_pkey_index =
 474                         be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
 475                 qp_attr->alt_port_num   = qp_attr->alt_ah_attr.port_num;
 476         }
 477
 478         qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
 479         qp_attr->port_num   =
 480                 (be32_to_cpu(context->pri_path.port_pkey) >> 24) & 0x3;
 481
 482         /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
 483         qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
 484
 485         qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
 486
 487         qp_attr->max_dest_rd_atomic =
 488                 1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
 489         qp_attr->min_rnr_timer      =
 490                 (be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
 491         qp_attr->timeout            = context->pri_path.ackto >> 3;
 492         qp_attr->retry_cnt          = (be32_to_cpu(context->params1) >> 16) & 0x7;
 493         qp_attr->rnr_retry          = context->pri_path.rnr_retry >> 5;
 494         qp_attr->alt_timeout        = context->alt_path.ackto >> 3;
 495
 496 done:
 497         qp_attr->cur_qp_state        = qp_attr->qp_state;
 498         qp_attr->cap.max_send_wr     = qp->sq.max;
 499         qp_attr->cap.max_recv_wr     = qp->rq.max;
 500         qp_attr->cap.max_send_sge    = qp->sq.max_gs;
 501         qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
 502         qp_attr->cap.max_inline_data = qp->max_inline_data;
 503
 504         qp_init_attr->cap            = qp_attr->cap;
 505         qp_init_attr->sq_sig_type    = qp->sq_policy;
 506
 507 out_mailbox:
 508         mthca_free_mailbox(dev, mailbox);
 509
 510 out:
 511         mutex_unlock(&qp->mutex);
 512         return err;
 513 }
 514
 515 static int mthca_path_set(struct mthca_dev *dev, const struct ib_ah_attr *ah,
 516                           struct mthca_qp_path *path, u8 port)
 517 {
 518         path->g_mylmc     = ah->src_path_bits & 0x7f;
 519         path->rlid        = cpu_to_be16(ah->dlid);
 520         path->static_rate = mthca_get_rate(dev, ah->static_rate, port);
 521
 522         if (ah->ah_flags & IB_AH_GRH) {
 523                 if (ah->grh.sgid_index >= dev->limits.gid_table_len) {
 524                         mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
 525                                   ah->grh.sgid_index, dev->limits.gid_table_len-1);
 526                         return -1;
 527                 }
 528
 529                 path->g_mylmc   |= 1 << 7;
 530                 path->mgid_index = ah->grh.sgid_index;
 531                 path->hop_limit  = ah->grh.hop_limit;
 532                 path->sl_tclass_flowlabel =
 533                         cpu_to_be32((ah->sl << 28)                |
 534                                     (ah->grh.traffic_class << 20) |
 535                                     (ah->grh.flow_label));
 536                 memcpy(path->rgid, ah->grh.dgid.raw, 16);
 537         } else
 538                 path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28);
 539
 540         return 0;
 541 }
 542
 543 static int __mthca_modify_qp(struct ib_qp *ibqp,
 544                              const struct ib_qp_attr *attr, int attr_mask,
 545                              enum ib_qp_state cur_state, enum ib_qp_state new_state)
 546 {
 547         struct mthca_dev *dev = to_mdev(ibqp->device);
 548         struct mthca_qp *qp = to_mqp(ibqp);
 549         struct mthca_mailbox *mailbox;
 550         struct mthca_qp_param *qp_param;
 551         struct mthca_qp_context *qp_context;
 552         u32 sqd_event = 0;
 553         int err = -EINVAL;
 554
 555         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 556         if (IS_ERR(mailbox)) {
 557                 err = PTR_ERR(mailbox);
 558                 goto out;
 559         }
 560         qp_param = mailbox->buf;
 561         qp_context = &qp_param->context;
 562         memset(qp_param, 0, sizeof *qp_param);
 563
 564         qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
 565                                              (to_mthca_st(qp->transport) << 16));
 566         qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
 567         if (!(attr_mask & IB_QP_PATH_MIG_STATE))
 568                 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
 569         else {
 570                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
 571                 switch (attr->path_mig_state) {
 572                 case IB_MIG_MIGRATED:
 573                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
 574                         break;
 575                 case IB_MIG_REARM:
 576                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
 577                         break;
 578                 case IB_MIG_ARMED:
 579                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
 580                         break;
 581                 }
 582         }
 583
 584         /* leave tavor_sched_queue as 0 */
 585
 586         if (qp->transport == MLX || qp->transport == UD)
 587                 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
 588         else if (attr_mask & IB_QP_PATH_MTU) {
 589                 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
 590                         mthca_dbg(dev, "path MTU (%u) is invalid\n",
 591                                   attr->path_mtu);
 592                         goto out_mailbox;
 593                 }
 594                 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
 595         }
 596
 597         if (mthca_is_memfree(dev)) {
 598                 if (qp->rq.max)
 599                         qp_context->rq_size_stride = ilog2(qp->rq.max) << 3;
 600                 qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
 601
 602                 if (qp->sq.max)
 603                         qp_context->sq_size_stride = ilog2(qp->sq.max) << 3;
 604                 qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
 605         }
 606
 607         /* leave arbel_sched_queue as 0 */
 608
 609         if (qp->ibqp.uobject)
 610                 qp_context->usr_page =
 611                         cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
 612         else
 613                 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
 614         qp_context->local_qpn  = cpu_to_be32(qp->qpn);
 615         if (attr_mask & IB_QP_DEST_QPN) {
 616                 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
 617         }
 618
 619         if (qp->transport == MLX)
 620                 qp_context->pri_path.port_pkey |=
 621                         cpu_to_be32(qp->port << 24);
 622         else {
 623                 if (attr_mask & IB_QP_PORT) {
 624                         qp_context->pri_path.port_pkey |=
 625                                 cpu_to_be32(attr->port_num << 24);
 626                         qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
 627                 }
 628         }
 629
 630         if (attr_mask & IB_QP_PKEY_INDEX) {
 631                 qp_context->pri_path.port_pkey |=
 632                         cpu_to_be32(attr->pkey_index);
 633                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
 634         }
 635
 636         if (attr_mask & IB_QP_RNR_RETRY) {
 637                 qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
 638                         attr->rnr_retry << 5;
 639                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
 640                                                         MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
 641         }
 642
 643         if (attr_mask & IB_QP_AV) {
 644                 if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
 645                                    attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
 646                         goto out_mailbox;
 647
 648                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
 649         }
 650
 651         if (ibqp->qp_type == IB_QPT_RC &&
 652             cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
 653                 u8 sched_queue = ibqp->uobject ? 0x2 : 0x1;
 654
 655                 if (mthca_is_memfree(dev))
 656                         qp_context->rlkey_arbel_sched_queue |= sched_queue;
 657                 else
 658                         qp_context->tavor_sched_queue |= cpu_to_be32(sched_queue);
 659
 660                 qp_param->opt_param_mask |=
 661                         cpu_to_be32(MTHCA_QP_OPTPAR_SCHED_QUEUE);
 662         }
 663
 664         if (attr_mask & IB_QP_TIMEOUT) {
 665                 qp_context->pri_path.ackto = attr->timeout << 3;
 666                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
 667         }
 668
 669         if (attr_mask & IB_QP_ALT_PATH) {
 670                 if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
 671                         mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
 672                                   attr->alt_pkey_index, dev->limits.pkey_table_len-1);
 673                         goto out_mailbox;
 674                 }
 675
 676                 if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
 677                         mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
 678                                 attr->alt_port_num);
 679                         goto out_mailbox;
 680                 }
 681
 682                 if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
 683                                    attr->alt_ah_attr.port_num))
 684                         goto out_mailbox;
 685
 686                 qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
 687                                                               attr->alt_port_num << 24);
 688                 qp_context->alt_path.ackto = attr->alt_timeout << 3;
 689                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
 690         }
 691
 692         /* leave rdd as 0 */
 693         qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
 694         /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
 695         qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
 696         qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
 697                                              (MTHCA_FLIGHT_LIMIT << 24) |
 698                                              MTHCA_QP_BIT_SWE);
 699         if (qp->sq_policy == IB_SIGNAL_ALL_WR)
 700                 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
 701         if (attr_mask & IB_QP_RETRY_CNT) {
 702                 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
 703                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
 704         }
 705
 706         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
 707                 if (attr->max_rd_atomic) {
 708                         qp_context->params1 |=
 709                                 cpu_to_be32(MTHCA_QP_BIT_SRE |
 710                                             MTHCA_QP_BIT_SAE);
 711                         qp_context->params1 |=
 712                                 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
 713                 }
 714                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
 715         }
 716
 717         if (attr_mask & IB_QP_SQ_PSN)
 718                 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
 719         qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
 720
 721         if (mthca_is_memfree(dev)) {
 722                 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
 723                 qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
 724         }
 725
 726         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
 727                 if (attr->max_dest_rd_atomic)
 728                         qp_context->params2 |=
 729                                 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
 730
 731                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
 732         }
 733
 734         if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
 735                 qp_context->params2      |= get_hw_access_flags(qp, attr, attr_mask);
 736                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
 737                                                         MTHCA_QP_OPTPAR_RRE |
 738                                                         MTHCA_QP_OPTPAR_RAE);
 739         }
 740
 741         qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
 742
 743         if (ibqp->srq)
 744                 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
 745
 746         if (attr_mask & IB_QP_MIN_RNR_TIMER) {
 747                 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
 748                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
 749         }
 750         if (attr_mask & IB_QP_RQ_PSN)
 751                 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
 752
 753         qp_context->ra_buff_indx =
 754                 cpu_to_be32(dev->qp_table.rdb_base +
 755                             ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
 756                              dev->qp_table.rdb_shift));
 757
 758         qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
 759
 760         if (mthca_is_memfree(dev))
 761                 qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
 762
 763         if (attr_mask & IB_QP_QKEY) {
 764                 qp_context->qkey = cpu_to_be32(attr->qkey);
 765                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
 766         }
 767
 768         if (ibqp->srq)
 769                 qp_context->srqn = cpu_to_be32(1 << 24 |
 770                                                to_msrq(ibqp->srq)->srqn);
 771
 772         if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD  &&
 773             attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY               &&
 774             attr->en_sqd_async_notify)
 775                 sqd_event = 1 << 31;
 776
 777         err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
 778                               mailbox, sqd_event);
 779         if (err) {
 780                 mthca_warn(dev, "modify QP %d->%d returned %d.\n",
 781                            cur_state, new_state, err);
 782                 goto out_mailbox;
 783         }
 784
 785         qp->state = new_state;
 786         if (attr_mask & IB_QP_ACCESS_FLAGS)
 787                 qp->atomic_rd_en = attr->qp_access_flags;
 788         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 789                 qp->resp_depth = attr->max_dest_rd_atomic;
 790         if (attr_mask & IB_QP_PORT)
 791                 qp->port = attr->port_num;
 792         if (attr_mask & IB_QP_ALT_PATH)
 793                 qp->alt_port = attr->alt_port_num;
 794
 795         if (is_sqp(dev, qp))
 796                 store_attrs(to_msqp(qp), attr, attr_mask);
 797
 798         /*
 799          * If we moved QP0 to RTR, bring the IB link up; if we moved
 800          * QP0 to RESET or ERROR, bring the link back down.
 801          */
 802         if (is_qp0(dev, qp)) {
 803                 if (cur_state != IB_QPS_RTR &&
 804                     new_state == IB_QPS_RTR)
 805                         init_port(dev, qp->port);
 806
 807                 if (cur_state != IB_QPS_RESET &&
 808                     cur_state != IB_QPS_ERR &&
 809                     (new_state == IB_QPS_RESET ||
 810                      new_state == IB_QPS_ERR))
 811                         mthca_CLOSE_IB(dev, qp->port);
 812         }
 813
 814         /*
 815          * If we moved a kernel QP to RESET, clean up all old CQ
 816          * entries and reinitialize the QP.
 817          */
 818         if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
 819                 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
 820                                qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
 821                 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
 822                         mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn, NULL);
 823
 824                 mthca_wq_reset(&qp->sq);
 825                 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
 826
 827                 mthca_wq_reset(&qp->rq);
 828                 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
 829
 830                 if (mthca_is_memfree(dev)) {
 831                         *qp->sq.db = 0;
 832                         *qp->rq.db = 0;
 833                 }
 834         }
 835
 836 out_mailbox:
 837         mthca_free_mailbox(dev, mailbox);
 838 out:
 839         return err;
 840 }
 841
 842 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
 843                     struct ib_udata *udata)
 844 {
 845         struct mthca_dev *dev = to_mdev(ibqp->device);
 846         struct mthca_qp *qp = to_mqp(ibqp);
 847         enum ib_qp_state cur_state, new_state;
 848         int err = -EINVAL;
 849
 850         mutex_lock(&qp->mutex);
 851         if (attr_mask & IB_QP_CUR_STATE) {
 852                 cur_state = attr->cur_qp_state;
 853         } else {
 854                 spin_lock_irq(&qp->sq.lock);
 855                 spin_lock(&qp->rq.lock);
 856                 cur_state = qp->state;
 857                 spin_unlock(&qp->rq.lock);
 858                 spin_unlock_irq(&qp->sq.lock);
 859         }
 860
 861         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
 862
 863         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask,
 864                                 IB_LINK_LAYER_UNSPECIFIED)) {
 865                 mthca_dbg(dev, "Bad QP transition (transport %d) "
 866                           "%d->%d with attr 0x%08x\n",
 867                           qp->transport, cur_state, new_state,
 868                           attr_mask);
 869                 goto out;
 870         }
 871
 872         if ((attr_mask & IB_QP_PKEY_INDEX) &&
 873              attr->pkey_index >= dev->limits.pkey_table_len) {
 874                 mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
 875                           attr->pkey_index, dev->limits.pkey_table_len-1);
 876                 goto out;
 877         }
 878
 879         if ((attr_mask & IB_QP_PORT) &&
 880             (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
 881                 mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
 882                 goto out;
 883         }
 884
 885         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
 886             attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
 887                 mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
 888                           attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
 889                 goto out;
 890         }
 891
 892         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
 893             attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
 894                 mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
 895                           attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
 896                 goto out;
 897         }
 898
 899         if (cur_state == new_state && cur_state == IB_QPS_RESET) {
 900                 err = 0;
 901                 goto out;
 902         }
 903
 904         err = __mthca_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
 905
 906 out:
 907         mutex_unlock(&qp->mutex);
 908         return err;
 909 }
 910
 911 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
 912 {
 913         /*
 914          * Calculate the maximum size of WQE s/g segments, excluding
 915          * the next segment and other non-data segments.
 916          */
 917         int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
 918
 919         switch (qp->transport) {
 920         case MLX:
 921                 max_data_size -= 2 * sizeof (struct mthca_data_seg);
 922                 break;
 923
 924         case UD:
 925                 if (mthca_is_memfree(dev))
 926                         max_data_size -= sizeof (struct mthca_arbel_ud_seg);
 927                 else
 928                         max_data_size -= sizeof (struct mthca_tavor_ud_seg);
 929                 break;
 930
 931         default:
 932                 max_data_size -= sizeof (struct mthca_raddr_seg);
 933                 break;
 934         }
 935
 936         return max_data_size;
 937 }
 938
 939 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
 940 {
 941         /* We don't support inline data for kernel QPs (yet). */
 942         return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
 943 }
 944
 945 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
 946                                  struct mthca_pd *pd,
 947                                  struct mthca_qp *qp)
 948 {
 949         int max_data_size = mthca_max_data_size(dev, qp,
 950                                                 min(dev->limits.max_desc_sz,
 951                                                     1 << qp->sq.wqe_shift));
 952
 953         qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
 954
 955         qp->sq.max_gs = min_t(int, dev->limits.max_sg,
 956                               max_data_size / sizeof (struct mthca_data_seg));
 957         qp->rq.max_gs = min_t(int, dev->limits.max_sg,
 958                                (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
 959                                 sizeof (struct mthca_next_seg)) /
 960                                sizeof (struct mthca_data_seg));
 961 }
 962
 963 /*
 964  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
 965  * rq.max_gs and sq.max_gs must all be assigned.
 966  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
 967  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
 968  * queue)
 969  */
 970 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
 971                                struct mthca_pd *pd,
 972                                struct mthca_qp *qp)
 973 {
 974         int size;
 975         int err = -ENOMEM;
 976
 977         size = sizeof (struct mthca_next_seg) +
 978                 qp->rq.max_gs * sizeof (struct mthca_data_seg);
 979
 980         if (size > dev->limits.max_desc_sz)
 981                 return -EINVAL;
 982
 983         for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
 984              qp->rq.wqe_shift++)
 985                 ; /* nothing */
 986
 987         size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
 988         switch (qp->transport) {
 989         case MLX:
 990                 size += 2 * sizeof (struct mthca_data_seg);
 991                 break;
 992
 993         case UD:
 994                 size += mthca_is_memfree(dev) ?
 995                         sizeof (struct mthca_arbel_ud_seg) :
 996                         sizeof (struct mthca_tavor_ud_seg);
 997                 break;
 998
 999         case UC:
1000                 size += sizeof (struct mthca_raddr_seg);
1001                 break;
1002
1003         case RC:
1004                 size += sizeof (struct mthca_raddr_seg);
1005                 /*
1006                  * An atomic op will require an atomic segment, a
1007                  * remote address segment and one scatter entry.
1008                  */
1009                 size = max_t(int, size,
1010                              sizeof (struct mthca_atomic_seg) +
1011                              sizeof (struct mthca_raddr_seg) +
1012                              sizeof (struct mthca_data_seg));
1013                 break;
1014
1015         default:
1016                 break;
1017         }
1018
1019         /* Make sure that we have enough space for a bind request */
1020         size = max_t(int, size, sizeof (struct mthca_bind_seg));
1021
1022         size += sizeof (struct mthca_next_seg);
1023
1024         if (size > dev->limits.max_desc_sz)
1025                 return -EINVAL;
1026
1027         for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
1028              qp->sq.wqe_shift++)
1029                 ; /* nothing */
1030
1031         qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
1032                                     1 << qp->sq.wqe_shift);
1033
1034         /*
1035          * If this is a userspace QP, we don't actually have to
1036          * allocate anything.  All we need is to calculate the WQE
1037          * sizes and the send_wqe_offset, so we're done now.
1038          */
1039         if (pd->ibpd.uobject)
1040                 return 0;
1041
1042         size = PAGE_ALIGN(qp->send_wqe_offset +
1043                           (qp->sq.max << qp->sq.wqe_shift));
1044
1045         qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1046                            GFP_KERNEL);
1047         if (!qp->wrid)
1048                 goto err_out;
1049
1050         err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1051                               &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1052         if (err)
1053                 goto err_out;
1054
1055         return 0;
1056
1057 err_out:
1058         kfree(qp->wrid);
1059         return err;
1060 }
1061
1062 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1063                                struct mthca_qp *qp)
1064 {
1065         mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1066                                        (qp->sq.max << qp->sq.wqe_shift)),
1067                        &qp->queue, qp->is_direct, &qp->mr);
1068         kfree(qp->wrid);
1069 }
1070
1071 static int mthca_map_memfree(struct mthca_dev *dev,
1072                              struct mthca_qp *qp)
1073 {
1074         int ret;
1075
1076         if (mthca_is_memfree(dev)) {
1077                 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1078                 if (ret)
1079                         return ret;
1080
1081                 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1082                 if (ret)
1083                         goto err_qpc;
1084
1085                 ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1086                                       qp->qpn << dev->qp_table.rdb_shift);
1087                 if (ret)
1088                         goto err_eqpc;
1089         }
1090
1091         return 0;
1092
1093 err_eqpc:
1094         mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1095
1096 err_qpc:
1097         mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1098
1099         return ret;
1100 }
1101
1102 static void mthca_unmap_memfree(struct mthca_dev *dev,
1103                                 struct mthca_qp *qp)
1104 {
1105         mthca_table_put(dev, dev->qp_table.rdb_table,
1106                         qp->qpn << dev->qp_table.rdb_shift);
1107         mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1108         mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1109 }
1110
1111 static int mthca_alloc_memfree(struct mthca_dev *dev,
1112                                struct mthca_qp *qp)
1113 {
1114         if (mthca_is_memfree(dev)) {
1115                 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1116                                                  qp->qpn, &qp->rq.db);
1117                 if (qp->rq.db_index < 0)
1118                         return -ENOMEM;
1119
1120                 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1121                                                  qp->qpn, &qp->sq.db);
1122                 if (qp->sq.db_index < 0) {
1123                         mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1124                         return -ENOMEM;
1125                 }
1126         }
1127
1128         return 0;
1129 }
1130
1131 static void mthca_free_memfree(struct mthca_dev *dev,
1132                                struct mthca_qp *qp)
1133 {
1134         if (mthca_is_memfree(dev)) {
1135                 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1136                 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1137         }
1138 }
1139
1140 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1141                                  struct mthca_pd *pd,
1142                                  struct mthca_cq *send_cq,
1143                                  struct mthca_cq *recv_cq,
1144                                  enum ib_sig_type send_policy,
1145                                  struct mthca_qp *qp)
1146 {
1147         int ret;
1148         int i;
1149         struct mthca_next_seg *next;
1150
1151         qp->refcount = 1;
1152         init_waitqueue_head(&qp->wait);
1153         mutex_init(&qp->mutex);
1154         qp->state        = IB_QPS_RESET;
1155         qp->atomic_rd_en = 0;
1156         qp->resp_depth   = 0;
1157         qp->sq_policy    = send_policy;
1158         mthca_wq_reset(&qp->sq);
1159         mthca_wq_reset(&qp->rq);
1160
1161         spin_lock_init(&qp->sq.lock);
1162         spin_lock_init(&qp->rq.lock);
1163
1164         ret = mthca_map_memfree(dev, qp);
1165         if (ret)
1166                 return ret;
1167
1168         ret = mthca_alloc_wqe_buf(dev, pd, qp);
1169         if (ret) {
1170                 mthca_unmap_memfree(dev, qp);
1171                 return ret;
1172         }
1173
1174         mthca_adjust_qp_caps(dev, pd, qp);
1175
1176         /*
1177          * If this is a userspace QP, we're done now.  The doorbells
1178          * will be allocated and buffers will be initialized in
1179          * userspace.
1180          */
1181         if (pd->ibpd.uobject)
1182                 return 0;
1183
1184         ret = mthca_alloc_memfree(dev, qp);
1185         if (ret) {
1186                 mthca_free_wqe_buf(dev, qp);
1187                 mthca_unmap_memfree(dev, qp);
1188                 return ret;
1189         }
1190
1191         if (mthca_is_memfree(dev)) {
1192                 struct mthca_data_seg *scatter;
1193                 int size = (sizeof (struct mthca_next_seg) +
1194                             qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1195
1196                 for (i = 0; i < qp->rq.max; ++i) {
1197                         next = get_recv_wqe(qp, i);
1198                         next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1199                                                    qp->rq.wqe_shift);
1200                         next->ee_nds = cpu_to_be32(size);
1201
1202                         for (scatter = (void *) (next + 1);
1203                              (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1204                              ++scatter)
1205                                 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1206                 }
1207
1208                 for (i = 0; i < qp->sq.max; ++i) {
1209                         next = get_send_wqe(qp, i);
1210                         next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1211                                                     qp->sq.wqe_shift) +
1212                                                    qp->send_wqe_offset);
1213                 }
1214         } else {
1215                 for (i = 0; i < qp->rq.max; ++i) {
1216                         next = get_recv_wqe(qp, i);
1217                         next->nda_op = htonl((((i + 1) % qp->rq.max) <<
1218                                               qp->rq.wqe_shift) | 1);
1219                 }
1220         }
1221
1222         qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1223         qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1224
1225         return 0;
1226 }
1227
1228 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1229                              struct mthca_pd *pd, struct mthca_qp *qp)
1230 {
1231         int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1232
1233         /* Sanity check QP size before proceeding */
1234         if (cap->max_send_wr     > dev->limits.max_wqes ||
1235             cap->max_recv_wr     > dev->limits.max_wqes ||
1236             cap->max_send_sge    > dev->limits.max_sg   ||
1237             cap->max_recv_sge    > dev->limits.max_sg   ||
1238             cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1239                 return -EINVAL;
1240
1241         /*
1242          * For MLX transport we need 2 extra send gather entries:
1243          * one for the header and one for the checksum at the end
1244          */
1245         if (qp->transport == MLX && cap->max_send_sge + 2 > dev->limits.max_sg)
1246                 return -EINVAL;
1247
1248         if (mthca_is_memfree(dev)) {
1249                 qp->rq.max = cap->max_recv_wr ?
1250                         roundup_pow_of_two(cap->max_recv_wr) : 0;
1251                 qp->sq.max = cap->max_send_wr ?
1252                         roundup_pow_of_two(cap->max_send_wr) : 0;
1253         } else {
1254                 qp->rq.max = cap->max_recv_wr;
1255                 qp->sq.max = cap->max_send_wr;
1256         }
1257
1258         qp->rq.max_gs = cap->max_recv_sge;
1259         qp->sq.max_gs = max_t(int, cap->max_send_sge,
1260                               ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1261                                     MTHCA_INLINE_CHUNK_SIZE) /
1262                               sizeof (struct mthca_data_seg));
1263
1264         return 0;
1265 }
1266
1267 int mthca_alloc_qp(struct mthca_dev *dev,
1268                    struct mthca_pd *pd,
1269                    struct mthca_cq *send_cq,
1270                    struct mthca_cq *recv_cq,
1271                    enum ib_qp_type type,
1272                    enum ib_sig_type send_policy,
1273                    struct ib_qp_cap *cap,
1274                    struct mthca_qp *qp)
1275 {
1276         int err;
1277
1278         switch (type) {
1279         case IB_QPT_RC: qp->transport = RC; break;
1280         case IB_QPT_UC: qp->transport = UC; break;
1281         case IB_QPT_UD: qp->transport = UD; break;
1282         default: return -EINVAL;
1283         }
1284
1285         err = mthca_set_qp_size(dev, cap, pd, qp);
1286         if (err)
1287                 return err;
1288
1289         qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1290         if (qp->qpn == -1)
1291                 return -ENOMEM;
1292
1293         /* initialize port to zero for error-catching. */
1294         qp->port = 0;
1295
1296         err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1297                                     send_policy, qp);
1298         if (err) {
1299                 mthca_free(&dev->qp_table.alloc, qp->qpn);
1300                 return err;
1301         }
1302
1303         spin_lock_irq(&dev->qp_table.lock);
1304         mthca_array_set(&dev->qp_table.qp,
1305                         qp->qpn & (dev->limits.num_qps - 1), qp);
1306         spin_unlock_irq(&dev->qp_table.lock);
1307
1308         return 0;
1309 }
1310
1311 static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1312         __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1313 {
1314         if (send_cq == recv_cq) {
1315                 spin_lock_irq(&send_cq->lock);
1316                 __acquire(&recv_cq->lock);
1317         } else if (send_cq->cqn < recv_cq->cqn) {
1318                 spin_lock_irq(&send_cq->lock);
1319                 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1320         } else {
1321                 spin_lock_irq(&recv_cq->lock);
1322                 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1323         }
1324 }
1325
1326 static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1327         __releases(&send_cq->lock) __releases(&recv_cq->lock)
1328 {
1329         if (send_cq == recv_cq) {
1330                 __release(&recv_cq->lock);
1331                 spin_unlock_irq(&send_cq->lock);
1332         } else if (send_cq->cqn < recv_cq->cqn) {
1333                 spin_unlock(&recv_cq->lock);
1334                 spin_unlock_irq(&send_cq->lock);
1335         } else {
1336                 spin_unlock(&send_cq->lock);
1337                 spin_unlock_irq(&recv_cq->lock);
1338         }
1339 }
1340
1341 int mthca_alloc_sqp(struct mthca_dev *dev,
1342                     struct mthca_pd *pd,
1343                     struct mthca_cq *send_cq,
1344                     struct mthca_cq *recv_cq,
1345                     enum ib_sig_type send_policy,
1346                     struct ib_qp_cap *cap,
1347                     int qpn,
1348                     int port,
1349                     struct mthca_sqp *sqp)
1350 {
1351         u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1352         int err;
1353
1354         sqp->qp.transport = MLX;
1355         err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1356         if (err)
1357                 return err;
1358
1359         sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1360         sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1361                                              &sqp->header_dma, GFP_KERNEL);
1362         if (!sqp->header_buf)
1363                 return -ENOMEM;
1364
1365         spin_lock_irq(&dev->qp_table.lock);
1366         if (mthca_array_get(&dev->qp_table.qp, mqpn))
1367                 err = -EBUSY;
1368         else
1369                 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1370         spin_unlock_irq(&dev->qp_table.lock);
1371
1372         if (err)
1373                 goto err_out;
1374
1375         sqp->qp.port      = port;
1376         sqp->qp.qpn       = mqpn;
1377         sqp->qp.transport = MLX;
1378
1379         err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1380                                     send_policy, &sqp->qp);
1381         if (err)
1382                 goto err_out_free;
1383
1384         atomic_inc(&pd->sqp_count);
1385
1386         return 0;
1387
1388  err_out_free:
1389         /*
1390          * Lock CQs here, so that CQ polling code can do QP lookup
1391          * without taking a lock.
1392          */
1393         mthca_lock_cqs(send_cq, recv_cq);
1394
1395         spin_lock(&dev->qp_table.lock);
1396         mthca_array_clear(&dev->qp_table.qp, mqpn);
1397         spin_unlock(&dev->qp_table.lock);
1398
1399         mthca_unlock_cqs(send_cq, recv_cq);
1400
1401  err_out:
1402         dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1403                           sqp->header_buf, sqp->header_dma);
1404
1405         return err;
1406 }
1407
1408 static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp)
1409 {
1410         int c;
1411
1412         spin_lock_irq(&dev->qp_table.lock);
1413         c = qp->refcount;
1414         spin_unlock_irq(&dev->qp_table.lock);
1415
1416         return c;
1417 }
1418
1419 void mthca_free_qp(struct mthca_dev *dev,
1420                    struct mthca_qp *qp)
1421 {
1422         struct mthca_cq *send_cq;
1423         struct mthca_cq *recv_cq;
1424
1425         send_cq = to_mcq(qp->ibqp.send_cq);
1426         recv_cq = to_mcq(qp->ibqp.recv_cq);
1427
1428         /*
1429          * Lock CQs here, so that CQ polling code can do QP lookup
1430          * without taking a lock.
1431          */
1432         mthca_lock_cqs(send_cq, recv_cq);
1433
1434         spin_lock(&dev->qp_table.lock);
1435         mthca_array_clear(&dev->qp_table.qp,
1436                           qp->qpn & (dev->limits.num_qps - 1));
1437         --qp->refcount;
1438         spin_unlock(&dev->qp_table.lock);
1439
1440         mthca_unlock_cqs(send_cq, recv_cq);
1441
1442         wait_event(qp->wait, !get_qp_refcount(dev, qp));
1443
1444         if (qp->state != IB_QPS_RESET)
1445                 mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1446                                 NULL, 0);
1447
1448         /*
1449          * If this is a userspace QP, the buffers, MR, CQs and so on
1450          * will be cleaned up in userspace, so all we have to do is
1451          * unref the mem-free tables and free the QPN in our table.
1452          */
1453         if (!qp->ibqp.uobject) {
1454                 mthca_cq_clean(dev, recv_cq, qp->qpn,
1455                                qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1456                 if (send_cq != recv_cq)
1457                         mthca_cq_clean(dev, send_cq, qp->qpn, NULL);
1458
1459                 mthca_free_memfree(dev, qp);
1460                 mthca_free_wqe_buf(dev, qp);
1461         }
1462
1463         mthca_unmap_memfree(dev, qp);
1464
1465         if (is_sqp(dev, qp)) {
1466                 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1467                 dma_free_coherent(&dev->pdev->dev,
1468                                   to_msqp(qp)->header_buf_size,
1469                                   to_msqp(qp)->header_buf,
1470                                   to_msqp(qp)->header_dma);
1471         } else
1472                 mthca_free(&dev->qp_table.alloc, qp->qpn);
1473 }
1474
1475 /* Create UD header for an MLX send and build a data segment for it */
1476 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1477                             int ind, struct ib_ud_wr *wr,
1478                             struct mthca_mlx_seg *mlx,
1479                             struct mthca_data_seg *data)
1480 {
1481         int header_size;
1482         int err;
1483         u16 pkey;
1484
1485         ib_ud_header_init(256, /* assume a MAD */ 1, 0, 0,
1486                           mthca_ah_grh_present(to_mah(wr->ah)), 0, 0, 0,
1487                           &sqp->ud_header);
1488
1489         err = mthca_read_ah(dev, to_mah(wr->ah), &sqp->ud_header);
1490         if (err)
1491                 return err;
1492         mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1493         mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1494                                   (sqp->ud_header.lrh.destination_lid ==
1495                                    IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1496                                   (sqp->ud_header.lrh.service_level << 8));
1497         mlx->rlid = sqp->ud_header.lrh.destination_lid;
1498         mlx->vcrc = 0;
1499
1500         switch (wr->wr.opcode) {
1501         case IB_WR_SEND:
1502                 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1503                 sqp->ud_header.immediate_present = 0;
1504                 break;
1505         case IB_WR_SEND_WITH_IMM:
1506                 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1507                 sqp->ud_header.immediate_present = 1;
1508                 sqp->ud_header.immediate_data = wr->wr.ex.imm_data;
1509                 break;
1510         default:
1511                 return -EINVAL;
1512         }
1513
1514         sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1515         if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1516                 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1517         sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
1518         if (!sqp->qp.ibqp.qp_num)
1519                 ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1520                                    sqp->pkey_index, &pkey);
1521         else
1522                 ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1523                                    wr->pkey_index, &pkey);
1524         sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1525         sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
1526         sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1527         sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
1528                                                sqp->qkey : wr->remote_qkey);
1529         sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1530
1531         header_size = ib_ud_header_pack(&sqp->ud_header,
1532                                         sqp->header_buf +
1533                                         ind * MTHCA_UD_HEADER_SIZE);
1534
1535         data->byte_count = cpu_to_be32(header_size);
1536         data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1537         data->addr       = cpu_to_be64(sqp->header_dma +
1538                                        ind * MTHCA_UD_HEADER_SIZE);
1539
1540         return 0;
1541 }
1542
1543 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1544                                     struct ib_cq *ib_cq)
1545 {
1546         unsigned cur;
1547         struct mthca_cq *cq;
1548
1549         cur = wq->head - wq->tail;
1550         if (likely(cur + nreq < wq->max))
1551                 return 0;
1552
1553         cq = to_mcq(ib_cq);
1554         spin_lock(&cq->lock);
1555         cur = wq->head - wq->tail;
1556         spin_unlock(&cq->lock);
1557
1558         return cur + nreq >= wq->max;
1559 }
1560
1561 static __always_inline void set_raddr_seg(struct mthca_raddr_seg *rseg,
1562                                           u64 remote_addr, u32 rkey)
1563 {
1564         rseg->raddr    = cpu_to_be64(remote_addr);
1565         rseg->rkey     = cpu_to_be32(rkey);
1566         rseg->reserved = 0;
1567 }
1568
1569 static __always_inline void set_atomic_seg(struct mthca_atomic_seg *aseg,
1570                                            struct ib_atomic_wr *wr)
1571 {
1572         if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1573                 aseg->swap_add = cpu_to_be64(wr->swap);
1574                 aseg->compare  = cpu_to_be64(wr->compare_add);
1575         } else {
1576                 aseg->swap_add = cpu_to_be64(wr->compare_add);
1577                 aseg->compare  = 0;
1578         }
1579
1580 }
1581
1582 static void set_tavor_ud_seg(struct mthca_tavor_ud_seg *useg,
1583                              struct ib_ud_wr *wr)
1584 {
1585         useg->lkey    = cpu_to_be32(to_mah(wr->ah)->key);
1586         useg->av_addr = cpu_to_be64(to_mah(wr->ah)->avdma);
1587         useg->dqpn    = cpu_to_be32(wr->remote_qpn);
1588         useg->qkey    = cpu_to_be32(wr->remote_qkey);
1589
1590 }
1591
1592 static void set_arbel_ud_seg(struct mthca_arbel_ud_seg *useg,
1593                              struct ib_ud_wr *wr)
1594 {
1595         memcpy(useg->av, to_mah(wr->ah)->av, MTHCA_AV_SIZE);
1596         useg->dqpn = cpu_to_be32(wr->remote_qpn);
1597         useg->qkey = cpu_to_be32(wr->remote_qkey);
1598 }
1599
1600 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1601                           struct ib_send_wr **bad_wr)
1602 {
1603         struct mthca_dev *dev = to_mdev(ibqp->device);
1604         struct mthca_qp *qp = to_mqp(ibqp);
1605         void *wqe;
1606         void *prev_wqe;
1607         unsigned long flags;
1608         int err = 0;
1609         int nreq;
1610         int i;
1611         int size;
1612         /*
1613          * f0 and size0 are only used if nreq != 0, and they will
1614          * always be initialized the first time through the main loop
1615          * before nreq is incremented.  So nreq cannot become non-zero
1616          * without initializing f0 and size0, and they are in fact
1617          * never used uninitialized.
1618          */
1619         int uninitialized_var(size0);
1620         u32 uninitialized_var(f0);
1621         int ind;
1622         u8 op0 = 0;
1623
1624         spin_lock_irqsave(&qp->sq.lock, flags);
1625
1626         /* XXX check that state is OK to post send */
1627
1628         ind = qp->sq.next_ind;
1629
1630         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1631                 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1632                         mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1633                                         " %d max, %d nreq)\n", qp->qpn,
1634                                         qp->sq.head, qp->sq.tail,
1635                                         qp->sq.max, nreq);
1636                         err = -ENOMEM;
1637                         *bad_wr = wr;
1638                         goto out;
1639                 }
1640
1641                 wqe = get_send_wqe(qp, ind);
1642                 prev_wqe = qp->sq.last;
1643                 qp->sq.last = wqe;
1644
1645                 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1646                 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1647                 ((struct mthca_next_seg *) wqe)->flags =
1648                         ((wr->send_flags & IB_SEND_SIGNALED) ?
1649                          cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1650                         ((wr->send_flags & IB_SEND_SOLICITED) ?
1651                          cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1652                         cpu_to_be32(1);
1653                 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1654                     wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1655                         ((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
1656
1657                 wqe += sizeof (struct mthca_next_seg);
1658                 size = sizeof (struct mthca_next_seg) / 16;
1659
1660                 switch (qp->transport) {
1661                 case RC:
1662                         switch (wr->opcode) {
1663                         case IB_WR_ATOMIC_CMP_AND_SWP:
1664                         case IB_WR_ATOMIC_FETCH_AND_ADD:
1665                                 set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
1666                                               atomic_wr(wr)->rkey);
1667                                 wqe += sizeof (struct mthca_raddr_seg);
1668
1669                                 set_atomic_seg(wqe, atomic_wr(wr));
1670                                 wqe += sizeof (struct mthca_atomic_seg);
1671                                 size += (sizeof (struct mthca_raddr_seg) +
1672                                          sizeof (struct mthca_atomic_seg)) / 16;
1673                                 break;
1674
1675                         case IB_WR_RDMA_WRITE:
1676                         case IB_WR_RDMA_WRITE_WITH_IMM:
1677                         case IB_WR_RDMA_READ:
1678                                 set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
1679                                               rdma_wr(wr)->rkey);
1680                                 wqe  += sizeof (struct mthca_raddr_seg);
1681                                 size += sizeof (struct mthca_raddr_seg) / 16;
1682                                 break;
1683
1684                         default:
1685                                 /* No extra segments required for sends */
1686                                 break;
1687                         }
1688
1689                         break;
1690
1691                 case UC:
1692                         switch (wr->opcode) {
1693                         case IB_WR_RDMA_WRITE:
1694                         case IB_WR_RDMA_WRITE_WITH_IMM:
1695                                 set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
1696                                               rdma_wr(wr)->rkey);
1697                                 wqe  += sizeof (struct mthca_raddr_seg);
1698                                 size += sizeof (struct mthca_raddr_seg) / 16;
1699                                 break;
1700
1701                         default:
1702                                 /* No extra segments required for sends */
1703                                 break;
1704                         }
1705
1706                         break;
1707
1708                 case UD:
1709                         set_tavor_ud_seg(wqe, ud_wr(wr));
1710                         wqe  += sizeof (struct mthca_tavor_ud_seg);
1711                         size += sizeof (struct mthca_tavor_ud_seg) / 16;
1712                         break;
1713
1714                 case MLX:
1715                         err = build_mlx_header(dev, to_msqp(qp), ind, ud_wr(wr),
1716                                                wqe - sizeof (struct mthca_next_seg),
1717                                                wqe);
1718                         if (err) {
1719                                 *bad_wr = wr;
1720                                 goto out;
1721                         }
1722                         wqe += sizeof (struct mthca_data_seg);
1723                         size += sizeof (struct mthca_data_seg) / 16;
1724                         break;
1725                 }
1726
1727                 if (wr->num_sge > qp->sq.max_gs) {
1728                         mthca_err(dev, "too many gathers\n");
1729                         err = -EINVAL;
1730                         *bad_wr = wr;
1731                         goto out;
1732                 }
1733
1734                 for (i = 0; i < wr->num_sge; ++i) {
1735                         mthca_set_data_seg(wqe, wr->sg_list + i);
1736                         wqe  += sizeof (struct mthca_data_seg);
1737                         size += sizeof (struct mthca_data_seg) / 16;
1738                 }
1739
1740                 /* Add one more inline data segment for ICRC */
1741                 if (qp->transport == MLX) {
1742                         ((struct mthca_data_seg *) wqe)->byte_count =
1743                                 cpu_to_be32((1 << 31) | 4);
1744                         ((u32 *) wqe)[1] = 0;
1745                         wqe += sizeof (struct mthca_data_seg);
1746                         size += sizeof (struct mthca_data_seg) / 16;
1747                 }
1748
1749                 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1750
1751                 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1752                         mthca_err(dev, "opcode invalid\n");
1753                         err = -EINVAL;
1754                         *bad_wr = wr;
1755                         goto out;
1756                 }
1757
1758                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1759                         cpu_to_be32(((ind << qp->sq.wqe_shift) +
1760                                      qp->send_wqe_offset) |
1761                                     mthca_opcode[wr->opcode]);
1762                 wmb();
1763                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1764                         cpu_to_be32((nreq ? 0 : MTHCA_NEXT_DBD) | size |
1765                                     ((wr->send_flags & IB_SEND_FENCE) ?
1766                                     MTHCA_NEXT_FENCE : 0));
1767
1768                 if (!nreq) {
1769                         size0 = size;
1770                         op0   = mthca_opcode[wr->opcode];
1771                         f0    = wr->send_flags & IB_SEND_FENCE ?
1772                                 MTHCA_SEND_DOORBELL_FENCE : 0;
1773                 }
1774
1775                 ++ind;
1776                 if (unlikely(ind >= qp->sq.max))
1777                         ind -= qp->sq.max;
1778         }
1779
1780 out:
1781         if (likely(nreq)) {
1782                 wmb();
1783
1784                 mthca_write64(((qp->sq.next_ind << qp->sq.wqe_shift) +
1785                                qp->send_wqe_offset) | f0 | op0,
1786                               (qp->qpn << 8) | size0,
1787                               dev->kar + MTHCA_SEND_DOORBELL,
1788                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1789                 /*
1790                  * Make sure doorbells don't leak out of SQ spinlock
1791                  * and reach the HCA out of order:
1792                  */
1793                 mmiowb();
1794         }
1795
1796         qp->sq.next_ind = ind;
1797         qp->sq.head    += nreq;
1798
1799         spin_unlock_irqrestore(&qp->sq.lock, flags);
1800         return err;
1801 }
1802
1803 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1804                              struct ib_recv_wr **bad_wr)
1805 {
1806         struct mthca_dev *dev = to_mdev(ibqp->device);
1807         struct mthca_qp *qp = to_mqp(ibqp);
1808         unsigned long flags;
1809         int err = 0;
1810         int nreq;
1811         int i;
1812         int size;
1813         /*
1814          * size0 is only used if nreq != 0, and it will always be
1815          * initialized the first time through the main loop before
1816          * nreq is incremented.  So nreq cannot become non-zero
1817          * without initializing size0, and it is in fact never used
1818          * uninitialized.
1819          */
1820         int uninitialized_var(size0);
1821         int ind;
1822         void *wqe;
1823         void *prev_wqe;
1824
1825         spin_lock_irqsave(&qp->rq.lock, flags);
1826
1827         /* XXX check that state is OK to post receive */
1828
1829         ind = qp->rq.next_ind;
1830
1831         for (nreq = 0; wr; wr = wr->next) {
1832                 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1833                         mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1834                                         " %d max, %d nreq)\n", qp->qpn,
1835                                         qp->rq.head, qp->rq.tail,
1836                                         qp->rq.max, nreq);
1837                         err = -ENOMEM;
1838                         *bad_wr = wr;
1839                         goto out;
1840                 }
1841
1842                 wqe = get_recv_wqe(qp, ind);
1843                 prev_wqe = qp->rq.last;
1844                 qp->rq.last = wqe;
1845
1846                 ((struct mthca_next_seg *) wqe)->ee_nds =
1847                         cpu_to_be32(MTHCA_NEXT_DBD);
1848                 ((struct mthca_next_seg *) wqe)->flags = 0;
1849
1850                 wqe += sizeof (struct mthca_next_seg);
1851                 size = sizeof (struct mthca_next_seg) / 16;
1852
1853                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1854                         err = -EINVAL;
1855                         *bad_wr = wr;
1856                         goto out;
1857                 }
1858
1859                 for (i = 0; i < wr->num_sge; ++i) {
1860                         mthca_set_data_seg(wqe, wr->sg_list + i);
1861                         wqe  += sizeof (struct mthca_data_seg);
1862                         size += sizeof (struct mthca_data_seg) / 16;
1863                 }
1864
1865                 qp->wrid[ind] = wr->wr_id;
1866
1867                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1868                         cpu_to_be32(MTHCA_NEXT_DBD | size);
1869
1870                 if (!nreq)
1871                         size0 = size;
1872
1873                 ++ind;
1874                 if (unlikely(ind >= qp->rq.max))
1875                         ind -= qp->rq.max;
1876
1877                 ++nreq;
1878                 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1879                         nreq = 0;
1880
1881                         wmb();
1882
1883                         mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1884                                       qp->qpn << 8, dev->kar + MTHCA_RECEIVE_DOORBELL,
1885                                       MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1886
1887                         qp->rq.next_ind = ind;
1888                         qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1889                 }
1890         }
1891
1892 out:
1893         if (likely(nreq)) {
1894                 wmb();
1895
1896                 mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1897                               qp->qpn << 8 | nreq, dev->kar + MTHCA_RECEIVE_DOORBELL,
1898                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1899         }
1900
1901         qp->rq.next_ind = ind;
1902         qp->rq.head    += nreq;
1903
1904         /*
1905          * Make sure doorbells don't leak out of RQ spinlock and reach
1906          * the HCA out of order:
1907          */
1908         mmiowb();
1909
1910         spin_unlock_irqrestore(&qp->rq.lock, flags);
1911         return err;
1912 }
1913
1914 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1915                           struct ib_send_wr **bad_wr)
1916 {
1917         struct mthca_dev *dev = to_mdev(ibqp->device);
1918         struct mthca_qp *qp = to_mqp(ibqp);
1919         u32 dbhi;
1920         void *wqe;
1921         void *prev_wqe;
1922         unsigned long flags;
1923         int err = 0;
1924         int nreq;
1925         int i;
1926         int size;
1927         /*
1928          * f0 and size0 are only used if nreq != 0, and they will
1929          * always be initialized the first time through the main loop
1930          * before nreq is incremented.  So nreq cannot become non-zero
1931          * without initializing f0 and size0, and they are in fact
1932          * never used uninitialized.
1933          */
1934         int uninitialized_var(size0);
1935         u32 uninitialized_var(f0);
1936         int ind;
1937         u8 op0 = 0;
1938
1939         spin_lock_irqsave(&qp->sq.lock, flags);
1940
1941         /* XXX check that state is OK to post send */
1942
1943         ind = qp->sq.head & (qp->sq.max - 1);
1944
1945         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1946                 if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1947                         nreq = 0;
1948
1949                         dbhi = (MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1950                                 ((qp->sq.head & 0xffff) << 8) | f0 | op0;
1951
1952                         qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1953
1954                         /*
1955                          * Make sure that descriptors are written before
1956                          * doorbell record.
1957                          */
1958                         wmb();
1959                         *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1960
1961                         /*
1962                          * Make sure doorbell record is written before we
1963                          * write MMIO send doorbell.
1964                          */
1965                         wmb();
1966
1967                         mthca_write64(dbhi, (qp->qpn << 8) | size0,
1968                                       dev->kar + MTHCA_SEND_DOORBELL,
1969                                       MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1970                 }
1971
1972                 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1973                         mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1974                                         " %d max, %d nreq)\n", qp->qpn,
1975                                         qp->sq.head, qp->sq.tail,
1976                                         qp->sq.max, nreq);
1977                         err = -ENOMEM;
1978                         *bad_wr = wr;
1979                         goto out;
1980                 }
1981
1982                 wqe = get_send_wqe(qp, ind);
1983                 prev_wqe = qp->sq.last;
1984                 qp->sq.last = wqe;
1985
1986                 ((struct mthca_next_seg *) wqe)->flags =
1987                         ((wr->send_flags & IB_SEND_SIGNALED) ?
1988                          cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1989                         ((wr->send_flags & IB_SEND_SOLICITED) ?
1990                          cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1991                         ((wr->send_flags & IB_SEND_IP_CSUM) ?
1992                          cpu_to_be32(MTHCA_NEXT_IP_CSUM | MTHCA_NEXT_TCP_UDP_CSUM) : 0) |
1993                         cpu_to_be32(1);
1994                 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1995                     wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1996                         ((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
1997
1998                 wqe += sizeof (struct mthca_next_seg);
1999                 size = sizeof (struct mthca_next_seg) / 16;
2000
2001                 switch (qp->transport) {
2002                 case RC:
2003                         switch (wr->opcode) {
2004                         case IB_WR_ATOMIC_CMP_AND_SWP:
2005                         case IB_WR_ATOMIC_FETCH_AND_ADD:
2006                                 set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
2007                                               atomic_wr(wr)->rkey);
2008                                 wqe += sizeof (struct mthca_raddr_seg);
2009
2010                                 set_atomic_seg(wqe, atomic_wr(wr));
2011                                 wqe  += sizeof (struct mthca_atomic_seg);
2012                                 size += (sizeof (struct mthca_raddr_seg) +
2013                                          sizeof (struct mthca_atomic_seg)) / 16;
2014                                 break;
2015
2016                         case IB_WR_RDMA_READ:
2017                         case IB_WR_RDMA_WRITE:
2018                         case IB_WR_RDMA_WRITE_WITH_IMM:
2019                                 set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
2020                                               rdma_wr(wr)->rkey);
2021                                 wqe  += sizeof (struct mthca_raddr_seg);
2022                                 size += sizeof (struct mthca_raddr_seg) / 16;
2023                                 break;
2024
2025                         default:
2026                                 /* No extra segments required for sends */
2027                                 break;
2028                         }
2029
2030                         break;
2031
2032                 case UC:
2033                         switch (wr->opcode) {
2034                         case IB_WR_RDMA_WRITE:
2035                         case IB_WR_RDMA_WRITE_WITH_IMM:
2036                                 set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
2037                                               rdma_wr(wr)->rkey);
2038                                 wqe  += sizeof (struct mthca_raddr_seg);
2039                                 size += sizeof (struct mthca_raddr_seg) / 16;
2040                                 break;
2041
2042                         default:
2043                                 /* No extra segments required for sends */
2044                                 break;
2045                         }
2046
2047                         break;
2048
2049                 case UD:
2050                         set_arbel_ud_seg(wqe, ud_wr(wr));
2051                         wqe  += sizeof (struct mthca_arbel_ud_seg);
2052                         size += sizeof (struct mthca_arbel_ud_seg) / 16;
2053                         break;
2054
2055                 case MLX:
2056                         err = build_mlx_header(dev, to_msqp(qp), ind, ud_wr(wr),
2057                                                wqe - sizeof (struct mthca_next_seg),
2058                                                wqe);
2059                         if (err) {
2060                                 *bad_wr = wr;
2061                                 goto out;
2062                         }
2063                         wqe += sizeof (struct mthca_data_seg);
2064                         size += sizeof (struct mthca_data_seg) / 16;
2065                         break;
2066                 }
2067
2068                 if (wr->num_sge > qp->sq.max_gs) {
2069                         mthca_err(dev, "too many gathers\n");
2070                         err = -EINVAL;
2071                         *bad_wr = wr;
2072                         goto out;
2073                 }
2074
2075                 for (i = 0; i < wr->num_sge; ++i) {
2076                         mthca_set_data_seg(wqe, wr->sg_list + i);
2077                         wqe  += sizeof (struct mthca_data_seg);
2078                         size += sizeof (struct mthca_data_seg) / 16;
2079                 }
2080
2081                 /* Add one more inline data segment for ICRC */
2082                 if (qp->transport == MLX) {
2083                         ((struct mthca_data_seg *) wqe)->byte_count =
2084                                 cpu_to_be32((1 << 31) | 4);
2085                         ((u32 *) wqe)[1] = 0;
2086                         wqe += sizeof (struct mthca_data_seg);
2087                         size += sizeof (struct mthca_data_seg) / 16;
2088                 }
2089
2090                 qp->wrid[ind + qp->rq.max] = wr->wr_id;
2091
2092                 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2093                         mthca_err(dev, "opcode invalid\n");
2094                         err = -EINVAL;
2095                         *bad_wr = wr;
2096                         goto out;
2097                 }
2098
2099                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
2100                         cpu_to_be32(((ind << qp->sq.wqe_shift) +
2101                                      qp->send_wqe_offset) |
2102                                     mthca_opcode[wr->opcode]);
2103                 wmb();
2104                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
2105                         cpu_to_be32(MTHCA_NEXT_DBD | size |
2106                                     ((wr->send_flags & IB_SEND_FENCE) ?
2107                                      MTHCA_NEXT_FENCE : 0));
2108
2109                 if (!nreq) {
2110                         size0 = size;
2111                         op0   = mthca_opcode[wr->opcode];
2112                         f0    = wr->send_flags & IB_SEND_FENCE ?
2113                                 MTHCA_SEND_DOORBELL_FENCE : 0;
2114                 }
2115
2116                 ++ind;
2117                 if (unlikely(ind >= qp->sq.max))
2118                         ind -= qp->sq.max;
2119         }
2120
2121 out:
2122         if (likely(nreq)) {
2123                 dbhi = (nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0;
2124
2125                 qp->sq.head += nreq;
2126
2127                 /*
2128                  * Make sure that descriptors are written before
2129                  * doorbell record.
2130                  */
2131                 wmb();
2132                 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2133
2134                 /*
2135                  * Make sure doorbell record is written before we
2136                  * write MMIO send doorbell.
2137                  */
2138                 wmb();
2139
2140                 mthca_write64(dbhi, (qp->qpn << 8) | size0, dev->kar + MTHCA_SEND_DOORBELL,
2141                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2142         }
2143
2144         /*
2145          * Make sure doorbells don't leak out of SQ spinlock and reach
2146          * the HCA out of order:
2147          */
2148         mmiowb();
2149
2150         spin_unlock_irqrestore(&qp->sq.lock, flags);
2151         return err;
2152 }
2153
2154 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2155                              struct ib_recv_wr **bad_wr)
2156 {
2157         struct mthca_dev *dev = to_mdev(ibqp->device);
2158         struct mthca_qp *qp = to_mqp(ibqp);
2159         unsigned long flags;
2160         int err = 0;
2161         int nreq;
2162         int ind;
2163         int i;
2164         void *wqe;
2165
2166         spin_lock_irqsave(&qp->rq.lock, flags);
2167
2168         /* XXX check that state is OK to post receive */
2169
2170         ind = qp->rq.head & (qp->rq.max - 1);
2171
2172         for (nreq = 0; wr; ++nreq, wr = wr->next) {
2173                 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2174                         mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2175                                         " %d max, %d nreq)\n", qp->qpn,
2176                                         qp->rq.head, qp->rq.tail,
2177                                         qp->rq.max, nreq);
2178                         err = -ENOMEM;
2179                         *bad_wr = wr;
2180                         goto out;
2181                 }
2182
2183                 wqe = get_recv_wqe(qp, ind);
2184
2185                 ((struct mthca_next_seg *) wqe)->flags = 0;
2186
2187                 wqe += sizeof (struct mthca_next_seg);
2188
2189                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2190                         err = -EINVAL;
2191                         *bad_wr = wr;
2192                         goto out;
2193                 }
2194
2195                 for (i = 0; i < wr->num_sge; ++i) {
2196                         mthca_set_data_seg(wqe, wr->sg_list + i);
2197                         wqe += sizeof (struct mthca_data_seg);
2198                 }
2199
2200                 if (i < qp->rq.max_gs)
2201                         mthca_set_data_seg_inval(wqe);
2202
2203                 qp->wrid[ind] = wr->wr_id;
2204
2205                 ++ind;
2206                 if (unlikely(ind >= qp->rq.max))
2207                         ind -= qp->rq.max;
2208         }
2209 out:
2210         if (likely(nreq)) {
2211                 qp->rq.head += nreq;
2212
2213                 /*
2214                  * Make sure that descriptors are written before
2215                  * doorbell record.
2216                  */
2217                 wmb();
2218                 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2219         }
2220
2221         spin_unlock_irqrestore(&qp->rq.lock, flags);
2222         return err;
2223 }
2224
2225 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2226                         int index, int *dbd, __be32 *new_wqe)
2227 {
2228         struct mthca_next_seg *next;
2229
2230         /*
2231          * For SRQs, all receive WQEs generate a CQE, so we're always
2232          * at the end of the doorbell chain.
2233          */
2234         if (qp->ibqp.srq && !is_send) {
2235                 *new_wqe = 0;
2236                 return;
2237         }
2238
2239         if (is_send)
2240                 next = get_send_wqe(qp, index);
2241         else
2242                 next = get_recv_wqe(qp, index);
2243
2244         *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2245         if (next->ee_nds & cpu_to_be32(0x3f))
2246                 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2247                         (next->ee_nds & cpu_to_be32(0x3f));
2248         else
2249                 *new_wqe = 0;
2250 }
2251
2252 int mthca_init_qp_table(struct mthca_dev *dev)
2253 {
2254         int err;
2255         int i;
2256
2257         spin_lock_init(&dev->qp_table.lock);
2258
2259         /*
2260          * We reserve 2 extra QPs per port for the special QPs.  The
2261          * special QP for port 1 has to be even, so round up.
2262          */
2263         dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2264         err = mthca_alloc_init(&dev->qp_table.alloc,
2265                                dev->limits.num_qps,
2266                                (1 << 24) - 1,
2267                                dev->qp_table.sqp_start +
2268                                MTHCA_MAX_PORTS * 2);
2269         if (err)
2270                 return err;
2271
2272         err = mthca_array_init(&dev->qp_table.qp,
2273                                dev->limits.num_qps);
2274         if (err) {
2275                 mthca_alloc_cleanup(&dev->qp_table.alloc);
2276                 return err;
2277         }
2278
2279         for (i = 0; i < 2; ++i) {
2280                 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2281                                     dev->qp_table.sqp_start + i * 2);
2282                 if (err) {
2283                         mthca_warn(dev, "CONF_SPECIAL_QP returned "
2284                                    "%d, aborting.\n", err);
2285                         goto err_out;
2286                 }
2287         }
2288         return 0;
2289
2290  err_out:
2291         for (i = 0; i < 2; ++i)
2292                 mthca_CONF_SPECIAL_QP(dev, i, 0);
2293
2294         mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2295         mthca_alloc_cleanup(&dev->qp_table.alloc);
2296
2297         return err;
2298 }
2299
2300 void mthca_cleanup_qp_table(struct mthca_dev *dev)
2301 {
2302         int i;
2303
2304         for (i = 0; i < 2; ++i)
2305                 mthca_CONF_SPECIAL_QP(dev, i, 0);
2306
2307         mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2308         mthca_alloc_cleanup(&dev->qp_table.alloc);
2309 }