drivers/infiniband/hw/mthca/mthca_qp.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Cisco Systems. All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
   5  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  *
  35  * $Id: mthca_qp.c 1355 2004-12-17 15:23:43Z roland $
  36  */
  37
  38 #include <linux/init.h>
  39 #include <linux/string.h>
  40 #include <linux/slab.h>
  41
  42 #include <asm/io.h>
  43
  44 #include <rdma/ib_verbs.h>
  45 #include <rdma/ib_cache.h>
  46 #include <rdma/ib_pack.h>
  47
  48 #include "mthca_dev.h"
  49 #include "mthca_cmd.h"
  50 #include "mthca_memfree.h"
  51 #include "mthca_wqe.h"
  52
  53 enum {
  54         MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
  55         MTHCA_ACK_REQ_FREQ       = 10,
  56         MTHCA_FLIGHT_LIMIT       = 9,
  57         MTHCA_UD_HEADER_SIZE     = 72, /* largest UD header possible */
  58         MTHCA_INLINE_HEADER_SIZE = 4,  /* data segment overhead for inline */
  59         MTHCA_INLINE_CHUNK_SIZE  = 16  /* inline data segment chunk */
  60 };
  61
  62 enum {
  63         MTHCA_QP_STATE_RST  = 0,
  64         MTHCA_QP_STATE_INIT = 1,
  65         MTHCA_QP_STATE_RTR  = 2,
  66         MTHCA_QP_STATE_RTS  = 3,
  67         MTHCA_QP_STATE_SQE  = 4,
  68         MTHCA_QP_STATE_SQD  = 5,
  69         MTHCA_QP_STATE_ERR  = 6,
  70         MTHCA_QP_STATE_DRAINING = 7
  71 };
  72
  73 enum {
  74         MTHCA_QP_ST_RC  = 0x0,
  75         MTHCA_QP_ST_UC  = 0x1,
  76         MTHCA_QP_ST_RD  = 0x2,
  77         MTHCA_QP_ST_UD  = 0x3,
  78         MTHCA_QP_ST_MLX = 0x7
  79 };
  80
  81 enum {
  82         MTHCA_QP_PM_MIGRATED = 0x3,
  83         MTHCA_QP_PM_ARMED    = 0x0,
  84         MTHCA_QP_PM_REARM    = 0x1
  85 };
  86
  87 enum {
  88         /* qp_context flags */
  89         MTHCA_QP_BIT_DE  = 1 <<  8,
  90         /* params1 */
  91         MTHCA_QP_BIT_SRE = 1 << 15,
  92         MTHCA_QP_BIT_SWE = 1 << 14,
  93         MTHCA_QP_BIT_SAE = 1 << 13,
  94         MTHCA_QP_BIT_SIC = 1 <<  4,
  95         MTHCA_QP_BIT_SSC = 1 <<  3,
  96         /* params2 */
  97         MTHCA_QP_BIT_RRE = 1 << 15,
  98         MTHCA_QP_BIT_RWE = 1 << 14,
  99         MTHCA_QP_BIT_RAE = 1 << 13,
 100         MTHCA_QP_BIT_RIC = 1 <<  4,
 101         MTHCA_QP_BIT_RSC = 1 <<  3
 102 };
 103
 104 enum {
 105         MTHCA_SEND_DOORBELL_FENCE = 1 << 5
 106 };
 107
 108 struct mthca_qp_path {
 109         __be32 port_pkey;
 110         u8     rnr_retry;
 111         u8     g_mylmc;
 112         __be16 rlid;
 113         u8     ackto;
 114         u8     mgid_index;
 115         u8     static_rate;
 116         u8     hop_limit;
 117         __be32 sl_tclass_flowlabel;
 118         u8     rgid[16];
 119 } __attribute__((packed));
 120
 121 struct mthca_qp_context {
 122         __be32 flags;
 123         __be32 tavor_sched_queue; /* Reserved on Arbel */
 124         u8     mtu_msgmax;
 125         u8     rq_size_stride;  /* Reserved on Tavor */
 126         u8     sq_size_stride;  /* Reserved on Tavor */
 127         u8     rlkey_arbel_sched_queue; /* Reserved on Tavor */
 128         __be32 usr_page;
 129         __be32 local_qpn;
 130         __be32 remote_qpn;
 131         u32    reserved1[2];
 132         struct mthca_qp_path pri_path;
 133         struct mthca_qp_path alt_path;
 134         __be32 rdd;
 135         __be32 pd;
 136         __be32 wqe_base;
 137         __be32 wqe_lkey;
 138         __be32 params1;
 139         __be32 reserved2;
 140         __be32 next_send_psn;
 141         __be32 cqn_snd;
 142         __be32 snd_wqe_base_l;  /* Next send WQE on Tavor */
 143         __be32 snd_db_index;    /* (debugging only entries) */
 144         __be32 last_acked_psn;
 145         __be32 ssn;
 146         __be32 params2;
 147         __be32 rnr_nextrecvpsn;
 148         __be32 ra_buff_indx;
 149         __be32 cqn_rcv;
 150         __be32 rcv_wqe_base_l;  /* Next recv WQE on Tavor */
 151         __be32 rcv_db_index;    /* (debugging only entries) */
 152         __be32 qkey;
 153         __be32 srqn;
 154         __be32 rmsn;
 155         __be16 rq_wqe_counter;  /* reserved on Tavor */
 156         __be16 sq_wqe_counter;  /* reserved on Tavor */
 157         u32    reserved3[18];
 158 } __attribute__((packed));
 159
 160 struct mthca_qp_param {
 161         __be32 opt_param_mask;
 162         u32    reserved1;
 163         struct mthca_qp_context context;
 164         u32    reserved2[62];
 165 } __attribute__((packed));
 166
 167 enum {
 168         MTHCA_QP_OPTPAR_ALT_ADDR_PATH     = 1 << 0,
 169         MTHCA_QP_OPTPAR_RRE               = 1 << 1,
 170         MTHCA_QP_OPTPAR_RAE               = 1 << 2,
 171         MTHCA_QP_OPTPAR_RWE               = 1 << 3,
 172         MTHCA_QP_OPTPAR_PKEY_INDEX        = 1 << 4,
 173         MTHCA_QP_OPTPAR_Q_KEY             = 1 << 5,
 174         MTHCA_QP_OPTPAR_RNR_TIMEOUT       = 1 << 6,
 175         MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
 176         MTHCA_QP_OPTPAR_SRA_MAX           = 1 << 8,
 177         MTHCA_QP_OPTPAR_RRA_MAX           = 1 << 9,
 178         MTHCA_QP_OPTPAR_PM_STATE          = 1 << 10,
 179         MTHCA_QP_OPTPAR_PORT_NUM          = 1 << 11,
 180         MTHCA_QP_OPTPAR_RETRY_COUNT       = 1 << 12,
 181         MTHCA_QP_OPTPAR_ALT_RNR_RETRY     = 1 << 13,
 182         MTHCA_QP_OPTPAR_ACK_TIMEOUT       = 1 << 14,
 183         MTHCA_QP_OPTPAR_RNR_RETRY         = 1 << 15,
 184         MTHCA_QP_OPTPAR_SCHED_QUEUE       = 1 << 16
 185 };
 186
 187 static const u8 mthca_opcode[] = {
 188         [IB_WR_SEND]                 = MTHCA_OPCODE_SEND,
 189         [IB_WR_SEND_WITH_IMM]        = MTHCA_OPCODE_SEND_IMM,
 190         [IB_WR_RDMA_WRITE]           = MTHCA_OPCODE_RDMA_WRITE,
 191         [IB_WR_RDMA_WRITE_WITH_IMM]  = MTHCA_OPCODE_RDMA_WRITE_IMM,
 192         [IB_WR_RDMA_READ]            = MTHCA_OPCODE_RDMA_READ,
 193         [IB_WR_ATOMIC_CMP_AND_SWP]   = MTHCA_OPCODE_ATOMIC_CS,
 194         [IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
 195 };
 196
 197 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
 198 {
 199         return qp->qpn >= dev->qp_table.sqp_start &&
 200                 qp->qpn <= dev->qp_table.sqp_start + 3;
 201 }
 202
 203 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
 204 {
 205         return qp->qpn >= dev->qp_table.sqp_start &&
 206                 qp->qpn <= dev->qp_table.sqp_start + 1;
 207 }
 208
 209 static void *get_recv_wqe(struct mthca_qp *qp, int n)
 210 {
 211         if (qp->is_direct)
 212                 return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
 213         else
 214                 return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
 215                         ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
 216 }
 217
 218 static void *get_send_wqe(struct mthca_qp *qp, int n)
 219 {
 220         if (qp->is_direct)
 221                 return qp->queue.direct.buf + qp->send_wqe_offset +
 222                         (n << qp->sq.wqe_shift);
 223         else
 224                 return qp->queue.page_list[(qp->send_wqe_offset +
 225                                             (n << qp->sq.wqe_shift)) >>
 226                                            PAGE_SHIFT].buf +
 227                         ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
 228                          (PAGE_SIZE - 1));
 229 }
 230
 231 static void mthca_wq_reset(struct mthca_wq *wq)
 232 {
 233         wq->next_ind  = 0;
 234         wq->last_comp = wq->max - 1;
 235         wq->head      = 0;
 236         wq->tail      = 0;
 237 }
 238
 239 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
 240                     enum ib_event_type event_type)
 241 {
 242         struct mthca_qp *qp;
 243         struct ib_event event;
 244
 245         spin_lock(&dev->qp_table.lock);
 246         qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
 247         if (qp)
 248                 ++qp->refcount;
 249         spin_unlock(&dev->qp_table.lock);
 250
 251         if (!qp) {
 252                 mthca_warn(dev, "Async event for bogus QP %08x\n", qpn);
 253                 return;
 254         }
 255
 256         if (event_type == IB_EVENT_PATH_MIG)
 257                 qp->port = qp->alt_port;
 258
 259         event.device      = &dev->ib_dev;
 260         event.event       = event_type;
 261         event.element.qp  = &qp->ibqp;
 262         if (qp->ibqp.event_handler)
 263                 qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
 264
 265         spin_lock(&dev->qp_table.lock);
 266         if (!--qp->refcount)
 267                 wake_up(&qp->wait);
 268         spin_unlock(&dev->qp_table.lock);
 269 }
 270
 271 static int to_mthca_state(enum ib_qp_state ib_state)
 272 {
 273         switch (ib_state) {
 274         case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
 275         case IB_QPS_INIT:  return MTHCA_QP_STATE_INIT;
 276         case IB_QPS_RTR:   return MTHCA_QP_STATE_RTR;
 277         case IB_QPS_RTS:   return MTHCA_QP_STATE_RTS;
 278         case IB_QPS_SQD:   return MTHCA_QP_STATE_SQD;
 279         case IB_QPS_SQE:   return MTHCA_QP_STATE_SQE;
 280         case IB_QPS_ERR:   return MTHCA_QP_STATE_ERR;
 281         default:                return -1;
 282         }
 283 }
 284
 285 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
 286
 287 static int to_mthca_st(int transport)
 288 {
 289         switch (transport) {
 290         case RC:  return MTHCA_QP_ST_RC;
 291         case UC:  return MTHCA_QP_ST_UC;
 292         case UD:  return MTHCA_QP_ST_UD;
 293         case RD:  return MTHCA_QP_ST_RD;
 294         case MLX: return MTHCA_QP_ST_MLX;
 295         default:  return -1;
 296         }
 297 }
 298
 299 static void store_attrs(struct mthca_sqp *sqp, struct ib_qp_attr *attr,
 300                         int attr_mask)
 301 {
 302         if (attr_mask & IB_QP_PKEY_INDEX)
 303                 sqp->pkey_index = attr->pkey_index;
 304         if (attr_mask & IB_QP_QKEY)
 305                 sqp->qkey = attr->qkey;
 306         if (attr_mask & IB_QP_SQ_PSN)
 307                 sqp->send_psn = attr->sq_psn;
 308 }
 309
 310 static void init_port(struct mthca_dev *dev, int port)
 311 {
 312         int err;
 313         u8 status;
 314         struct mthca_init_ib_param param;
 315
 316         memset(&param, 0, sizeof param);
 317
 318         param.port_width = dev->limits.port_width_cap;
 319         param.vl_cap     = dev->limits.vl_cap;
 320         param.mtu_cap    = dev->limits.mtu_cap;
 321         param.gid_cap    = dev->limits.gid_table_len;
 322         param.pkey_cap   = dev->limits.pkey_table_len;
 323
 324         err = mthca_INIT_IB(dev, &param, port, &status);
 325         if (err)
 326                 mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
 327         if (status)
 328                 mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
 329 }
 330
 331 static __be32 get_hw_access_flags(struct mthca_qp *qp, struct ib_qp_attr *attr,
 332                                   int attr_mask)
 333 {
 334         u8 dest_rd_atomic;
 335         u32 access_flags;
 336         u32 hw_access_flags = 0;
 337
 338         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 339                 dest_rd_atomic = attr->max_dest_rd_atomic;
 340         else
 341                 dest_rd_atomic = qp->resp_depth;
 342
 343         if (attr_mask & IB_QP_ACCESS_FLAGS)
 344                 access_flags = attr->qp_access_flags;
 345         else
 346                 access_flags = qp->atomic_rd_en;
 347
 348         if (!dest_rd_atomic)
 349                 access_flags &= IB_ACCESS_REMOTE_WRITE;
 350
 351         if (access_flags & IB_ACCESS_REMOTE_READ)
 352                 hw_access_flags |= MTHCA_QP_BIT_RRE;
 353         if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
 354                 hw_access_flags |= MTHCA_QP_BIT_RAE;
 355         if (access_flags & IB_ACCESS_REMOTE_WRITE)
 356                 hw_access_flags |= MTHCA_QP_BIT_RWE;
 357
 358         return cpu_to_be32(hw_access_flags);
 359 }
 360
 361 static inline enum ib_qp_state to_ib_qp_state(int mthca_state)
 362 {
 363         switch (mthca_state) {
 364         case MTHCA_QP_STATE_RST:      return IB_QPS_RESET;
 365         case MTHCA_QP_STATE_INIT:     return IB_QPS_INIT;
 366         case MTHCA_QP_STATE_RTR:      return IB_QPS_RTR;
 367         case MTHCA_QP_STATE_RTS:      return IB_QPS_RTS;
 368         case MTHCA_QP_STATE_DRAINING:
 369         case MTHCA_QP_STATE_SQD:      return IB_QPS_SQD;
 370         case MTHCA_QP_STATE_SQE:      return IB_QPS_SQE;
 371         case MTHCA_QP_STATE_ERR:      return IB_QPS_ERR;
 372         default:                      return -1;
 373         }
 374 }
 375
 376 static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state)
 377 {
 378         switch (mthca_mig_state) {
 379         case 0:  return IB_MIG_ARMED;
 380         case 1:  return IB_MIG_REARM;
 381         case 3:  return IB_MIG_MIGRATED;
 382         default: return -1;
 383         }
 384 }
 385
 386 static int to_ib_qp_access_flags(int mthca_flags)
 387 {
 388         int ib_flags = 0;
 389
 390         if (mthca_flags & MTHCA_QP_BIT_RRE)
 391                 ib_flags |= IB_ACCESS_REMOTE_READ;
 392         if (mthca_flags & MTHCA_QP_BIT_RWE)
 393                 ib_flags |= IB_ACCESS_REMOTE_WRITE;
 394         if (mthca_flags & MTHCA_QP_BIT_RAE)
 395                 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
 396
 397         return ib_flags;
 398 }
 399
 400 static void to_ib_ah_attr(struct mthca_dev *dev, struct ib_ah_attr *ib_ah_attr,
 401                                 struct mthca_qp_path *path)
 402 {
 403         memset(ib_ah_attr, 0, sizeof *path);
 404         ib_ah_attr->port_num      = (be32_to_cpu(path->port_pkey) >> 24) & 0x3;
 405
 406         if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->limits.num_ports)
 407                 return;
 408
 409         ib_ah_attr->dlid          = be16_to_cpu(path->rlid);
 410         ib_ah_attr->sl            = be32_to_cpu(path->sl_tclass_flowlabel) >> 28;
 411         ib_ah_attr->src_path_bits = path->g_mylmc & 0x7f;
 412         ib_ah_attr->static_rate   = mthca_rate_to_ib(dev,
 413                                                      path->static_rate & 0x7,
 414                                                      ib_ah_attr->port_num);
 415         ib_ah_attr->ah_flags      = (path->g_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
 416         if (ib_ah_attr->ah_flags) {
 417                 ib_ah_attr->grh.sgid_index = path->mgid_index & (dev->limits.gid_table_len - 1);
 418                 ib_ah_attr->grh.hop_limit  = path->hop_limit;
 419                 ib_ah_attr->grh.traffic_class =
 420                         (be32_to_cpu(path->sl_tclass_flowlabel) >> 20) & 0xff;
 421                 ib_ah_attr->grh.flow_label =
 422                         be32_to_cpu(path->sl_tclass_flowlabel) & 0xfffff;
 423                 memcpy(ib_ah_attr->grh.dgid.raw,
 424                         path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
 425         }
 426 }
 427
 428 int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
 429                    struct ib_qp_init_attr *qp_init_attr)
 430 {
 431         struct mthca_dev *dev = to_mdev(ibqp->device);
 432         struct mthca_qp *qp = to_mqp(ibqp);
 433         int err;
 434         struct mthca_mailbox *mailbox;
 435         struct mthca_qp_param *qp_param;
 436         struct mthca_qp_context *context;
 437         int mthca_state;
 438         u8 status;
 439
 440         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 441         if (IS_ERR(mailbox))
 442                 return PTR_ERR(mailbox);
 443
 444         err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox, &status);
 445         if (err)
 446                 goto out;
 447         if (status) {
 448                 mthca_warn(dev, "QUERY_QP returned status %02x\n", status);
 449                 err = -EINVAL;
 450                 goto out;
 451         }
 452
 453         qp_param    = mailbox->buf;
 454         context     = &qp_param->context;
 455         mthca_state = be32_to_cpu(context->flags) >> 28;
 456
 457         qp_attr->qp_state            = to_ib_qp_state(mthca_state);
 458         qp_attr->cur_qp_state        = qp_attr->qp_state;
 459         qp_attr->path_mtu            = context->mtu_msgmax >> 5;
 460         qp_attr->path_mig_state      =
 461                 to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
 462         qp_attr->qkey                = be32_to_cpu(context->qkey);
 463         qp_attr->rq_psn              = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
 464         qp_attr->sq_psn              = be32_to_cpu(context->next_send_psn) & 0xffffff;
 465         qp_attr->dest_qp_num         = be32_to_cpu(context->remote_qpn) & 0xffffff;
 466         qp_attr->qp_access_flags     =
 467                 to_ib_qp_access_flags(be32_to_cpu(context->params2));
 468         qp_attr->cap.max_send_wr     = qp->sq.max;
 469         qp_attr->cap.max_recv_wr     = qp->rq.max;
 470         qp_attr->cap.max_send_sge    = qp->sq.max_gs;
 471         qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
 472         qp_attr->cap.max_inline_data = qp->max_inline_data;
 473
 474         if (qp->transport == RC || qp->transport == UC) {
 475                 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
 476                 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
 477         }
 478
 479         qp_attr->pkey_index     = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
 480         qp_attr->alt_pkey_index = be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
 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->port_num           = qp_attr->ah_attr.port_num;
 492         qp_attr->timeout            = context->pri_path.ackto >> 3;
 493         qp_attr->retry_cnt          = (be32_to_cpu(context->params1) >> 16) & 0x7;
 494         qp_attr->rnr_retry          = context->pri_path.rnr_retry >> 5;
 495         qp_attr->alt_port_num       = qp_attr->alt_ah_attr.port_num;
 496         qp_attr->alt_timeout        = context->alt_path.ackto >> 3;
 497         qp_init_attr->cap           = qp_attr->cap;
 498
 499 out:
 500         mthca_free_mailbox(dev, mailbox);
 501         return err;
 502 }
 503
 504 static int mthca_path_set(struct mthca_dev *dev, struct ib_ah_attr *ah,
 505                           struct mthca_qp_path *path, u8 port)
 506 {
 507         path->g_mylmc     = ah->src_path_bits & 0x7f;
 508         path->rlid        = cpu_to_be16(ah->dlid);
 509         path->static_rate = mthca_get_rate(dev, ah->static_rate, port);
 510
 511         if (ah->ah_flags & IB_AH_GRH) {
 512                 if (ah->grh.sgid_index >= dev->limits.gid_table_len) {
 513                         mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
 514                                   ah->grh.sgid_index, dev->limits.gid_table_len-1);
 515                         return -1;
 516                 }
 517
 518                 path->g_mylmc   |= 1 << 7;
 519                 path->mgid_index = ah->grh.sgid_index;
 520                 path->hop_limit  = ah->grh.hop_limit;
 521                 path->sl_tclass_flowlabel =
 522                         cpu_to_be32((ah->sl << 28)                |
 523                                     (ah->grh.traffic_class << 20) |
 524                                     (ah->grh.flow_label));
 525                 memcpy(path->rgid, ah->grh.dgid.raw, 16);
 526         } else
 527                 path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28);
 528
 529         return 0;
 530 }
 531
 532 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 533 {
 534         struct mthca_dev *dev = to_mdev(ibqp->device);
 535         struct mthca_qp *qp = to_mqp(ibqp);
 536         enum ib_qp_state cur_state, new_state;
 537         struct mthca_mailbox *mailbox;
 538         struct mthca_qp_param *qp_param;
 539         struct mthca_qp_context *qp_context;
 540         u32 sqd_event = 0;
 541         u8 status;
 542         int err = -EINVAL;
 543
 544         mutex_lock(&qp->mutex);
 545
 546         if (attr_mask & IB_QP_CUR_STATE) {
 547                 cur_state = attr->cur_qp_state;
 548         } else {
 549                 spin_lock_irq(&qp->sq.lock);
 550                 spin_lock(&qp->rq.lock);
 551                 cur_state = qp->state;
 552                 spin_unlock(&qp->rq.lock);
 553                 spin_unlock_irq(&qp->sq.lock);
 554         }
 555
 556         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
 557
 558         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
 559                 mthca_dbg(dev, "Bad QP transition (transport %d) "
 560                           "%d->%d with attr 0x%08x\n",
 561                           qp->transport, cur_state, new_state,
 562                           attr_mask);
 563                 goto out;
 564         }
 565
 566         if ((attr_mask & IB_QP_PKEY_INDEX) &&
 567              attr->pkey_index >= dev->limits.pkey_table_len) {
 568                 mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
 569                           attr->pkey_index, dev->limits.pkey_table_len-1);
 570                 goto out;
 571         }
 572
 573         if ((attr_mask & IB_QP_PORT) &&
 574             (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
 575                 mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
 576                 goto out;
 577         }
 578
 579         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
 580             attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
 581                 mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
 582                           attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
 583                 goto out;
 584         }
 585
 586         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
 587             attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
 588                 mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
 589                           attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
 590                 goto out;
 591         }
 592
 593         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 594         if (IS_ERR(mailbox)) {
 595                 err = PTR_ERR(mailbox);
 596                 goto out;
 597         }
 598         qp_param = mailbox->buf;
 599         qp_context = &qp_param->context;
 600         memset(qp_param, 0, sizeof *qp_param);
 601
 602         qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
 603                                              (to_mthca_st(qp->transport) << 16));
 604         qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
 605         if (!(attr_mask & IB_QP_PATH_MIG_STATE))
 606                 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
 607         else {
 608                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
 609                 switch (attr->path_mig_state) {
 610                 case IB_MIG_MIGRATED:
 611                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
 612                         break;
 613                 case IB_MIG_REARM:
 614                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
 615                         break;
 616                 case IB_MIG_ARMED:
 617                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
 618                         break;
 619                 }
 620         }
 621
 622         /* leave tavor_sched_queue as 0 */
 623
 624         if (qp->transport == MLX || qp->transport == UD)
 625                 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
 626         else if (attr_mask & IB_QP_PATH_MTU) {
 627                 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
 628                         mthca_dbg(dev, "path MTU (%u) is invalid\n",
 629                                   attr->path_mtu);
 630                         goto out_mailbox;
 631                 }
 632                 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
 633         }
 634
 635         if (mthca_is_memfree(dev)) {
 636                 if (qp->rq.max)
 637                         qp_context->rq_size_stride = long_log2(qp->rq.max) << 3;
 638                 qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
 639
 640                 if (qp->sq.max)
 641                         qp_context->sq_size_stride = long_log2(qp->sq.max) << 3;
 642                 qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
 643         }
 644
 645         /* leave arbel_sched_queue as 0 */
 646
 647         if (qp->ibqp.uobject)
 648                 qp_context->usr_page =
 649                         cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
 650         else
 651                 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
 652         qp_context->local_qpn  = cpu_to_be32(qp->qpn);
 653         if (attr_mask & IB_QP_DEST_QPN) {
 654                 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
 655         }
 656
 657         if (qp->transport == MLX)
 658                 qp_context->pri_path.port_pkey |=
 659                         cpu_to_be32(qp->port << 24);
 660         else {
 661                 if (attr_mask & IB_QP_PORT) {
 662                         qp_context->pri_path.port_pkey |=
 663                                 cpu_to_be32(attr->port_num << 24);
 664                         qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
 665                 }
 666         }
 667
 668         if (attr_mask & IB_QP_PKEY_INDEX) {
 669                 qp_context->pri_path.port_pkey |=
 670                         cpu_to_be32(attr->pkey_index);
 671                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
 672         }
 673
 674         if (attr_mask & IB_QP_RNR_RETRY) {
 675                 qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
 676                         attr->rnr_retry << 5;
 677                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
 678                                                         MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
 679         }
 680
 681         if (attr_mask & IB_QP_AV) {
 682                 if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
 683                                    attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
 684                         goto out_mailbox;
 685
 686                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
 687         }
 688
 689         if (attr_mask & IB_QP_TIMEOUT) {
 690                 qp_context->pri_path.ackto = attr->timeout << 3;
 691                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
 692         }
 693
 694         if (attr_mask & IB_QP_ALT_PATH) {
 695                 if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
 696                         mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
 697                                   attr->alt_pkey_index, dev->limits.pkey_table_len-1);
 698                         goto out_mailbox;
 699                 }
 700
 701                 if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
 702                         mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
 703                                 attr->alt_port_num);
 704                         goto out_mailbox;
 705                 }
 706
 707                 if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
 708                                    attr->alt_ah_attr.port_num))
 709                         goto out_mailbox;
 710
 711                 qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
 712                                                               attr->alt_port_num << 24);
 713                 qp_context->alt_path.ackto = attr->alt_timeout << 3;
 714                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
 715         }
 716
 717         /* leave rdd as 0 */
 718         qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
 719         /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
 720         qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
 721         qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
 722                                              (MTHCA_FLIGHT_LIMIT << 24) |
 723                                              MTHCA_QP_BIT_SWE);
 724         if (qp->sq_policy == IB_SIGNAL_ALL_WR)
 725                 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
 726         if (attr_mask & IB_QP_RETRY_CNT) {
 727                 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
 728                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
 729         }
 730
 731         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
 732                 if (attr->max_rd_atomic) {
 733                         qp_context->params1 |=
 734                                 cpu_to_be32(MTHCA_QP_BIT_SRE |
 735                                             MTHCA_QP_BIT_SAE);
 736                         qp_context->params1 |=
 737                                 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
 738                 }
 739                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
 740         }
 741
 742         if (attr_mask & IB_QP_SQ_PSN)
 743                 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
 744         qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
 745
 746         if (mthca_is_memfree(dev)) {
 747                 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
 748                 qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
 749         }
 750
 751         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
 752                 if (attr->max_dest_rd_atomic)
 753                         qp_context->params2 |=
 754                                 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
 755
 756                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
 757         }
 758
 759         if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
 760                 qp_context->params2      |= get_hw_access_flags(qp, attr, attr_mask);
 761                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
 762                                                         MTHCA_QP_OPTPAR_RRE |
 763                                                         MTHCA_QP_OPTPAR_RAE);
 764         }
 765
 766         qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
 767
 768         if (ibqp->srq)
 769                 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
 770
 771         if (attr_mask & IB_QP_MIN_RNR_TIMER) {
 772                 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
 773                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
 774         }
 775         if (attr_mask & IB_QP_RQ_PSN)
 776                 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
 777
 778         qp_context->ra_buff_indx =
 779                 cpu_to_be32(dev->qp_table.rdb_base +
 780                             ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
 781                              dev->qp_table.rdb_shift));
 782
 783         qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
 784
 785         if (mthca_is_memfree(dev))
 786                 qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
 787
 788         if (attr_mask & IB_QP_QKEY) {
 789                 qp_context->qkey = cpu_to_be32(attr->qkey);
 790                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
 791         }
 792
 793         if (ibqp->srq)
 794                 qp_context->srqn = cpu_to_be32(1 << 24 |
 795                                                to_msrq(ibqp->srq)->srqn);
 796
 797         if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD  &&
 798             attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY               &&
 799             attr->en_sqd_async_notify)
 800                 sqd_event = 1 << 31;
 801
 802         err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
 803                               mailbox, sqd_event, &status);
 804         if (err)
 805                 goto out_mailbox;
 806         if (status) {
 807                 mthca_warn(dev, "modify QP %d->%d returned status %02x.\n",
 808                            cur_state, new_state, status);
 809                 err = -EINVAL;
 810                 goto out_mailbox;
 811         }
 812
 813         qp->state = new_state;
 814         if (attr_mask & IB_QP_ACCESS_FLAGS)
 815                 qp->atomic_rd_en = attr->qp_access_flags;
 816         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 817                 qp->resp_depth = attr->max_dest_rd_atomic;
 818         if (attr_mask & IB_QP_PORT)
 819                 qp->port = attr->port_num;
 820         if (attr_mask & IB_QP_ALT_PATH)
 821                 qp->alt_port = attr->alt_port_num;
 822
 823         if (is_sqp(dev, qp))
 824                 store_attrs(to_msqp(qp), attr, attr_mask);
 825
 826         /*
 827          * If we moved QP0 to RTR, bring the IB link up; if we moved
 828          * QP0 to RESET or ERROR, bring the link back down.
 829          */
 830         if (is_qp0(dev, qp)) {
 831                 if (cur_state != IB_QPS_RTR &&
 832                     new_state == IB_QPS_RTR)
 833                         init_port(dev, qp->port);
 834
 835                 if (cur_state != IB_QPS_RESET &&
 836                     cur_state != IB_QPS_ERR &&
 837                     (new_state == IB_QPS_RESET ||
 838                      new_state == IB_QPS_ERR))
 839                         mthca_CLOSE_IB(dev, qp->port, &status);
 840         }
 841
 842         /*
 843          * If we moved a kernel QP to RESET, clean up all old CQ
 844          * entries and reinitialize the QP.
 845          */
 846         if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
 847                 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn,
 848                                qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
 849                 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
 850                         mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
 851                                        qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
 852
 853                 mthca_wq_reset(&qp->sq);
 854                 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
 855
 856                 mthca_wq_reset(&qp->rq);
 857                 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
 858
 859                 if (mthca_is_memfree(dev)) {
 860                         *qp->sq.db = 0;
 861                         *qp->rq.db = 0;
 862                 }
 863         }
 864
 865 out_mailbox:
 866         mthca_free_mailbox(dev, mailbox);
 867
 868 out:
 869         mutex_unlock(&qp->mutex);
 870         return err;
 871 }
 872
 873 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
 874 {
 875         /*
 876          * Calculate the maximum size of WQE s/g segments, excluding
 877          * the next segment and other non-data segments.
 878          */
 879         int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
 880
 881         switch (qp->transport) {
 882         case MLX:
 883                 max_data_size -= 2 * sizeof (struct mthca_data_seg);
 884                 break;
 885
 886         case UD:
 887                 if (mthca_is_memfree(dev))
 888                         max_data_size -= sizeof (struct mthca_arbel_ud_seg);
 889                 else
 890                         max_data_size -= sizeof (struct mthca_tavor_ud_seg);
 891                 break;
 892
 893         default:
 894                 max_data_size -= sizeof (struct mthca_raddr_seg);
 895                 break;
 896         }
 897
 898         return max_data_size;
 899 }
 900
 901 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
 902 {
 903         /* We don't support inline data for kernel QPs (yet). */
 904         return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
 905 }
 906
 907 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
 908                                  struct mthca_pd *pd,
 909                                  struct mthca_qp *qp)
 910 {
 911         int max_data_size = mthca_max_data_size(dev, qp,
 912                                                 min(dev->limits.max_desc_sz,
 913                                                     1 << qp->sq.wqe_shift));
 914
 915         qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
 916
 917         qp->sq.max_gs = min_t(int, dev->limits.max_sg,
 918                               max_data_size / sizeof (struct mthca_data_seg));
 919         qp->rq.max_gs = min_t(int, dev->limits.max_sg,
 920                                (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
 921                                 sizeof (struct mthca_next_seg)) /
 922                                sizeof (struct mthca_data_seg));
 923 }
 924
 925 /*
 926  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
 927  * rq.max_gs and sq.max_gs must all be assigned.
 928  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
 929  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
 930  * queue)
 931  */
 932 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
 933                                struct mthca_pd *pd,
 934                                struct mthca_qp *qp)
 935 {
 936         int size;
 937         int err = -ENOMEM;
 938
 939         size = sizeof (struct mthca_next_seg) +
 940                 qp->rq.max_gs * sizeof (struct mthca_data_seg);
 941
 942         if (size > dev->limits.max_desc_sz)
 943                 return -EINVAL;
 944
 945         for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
 946              qp->rq.wqe_shift++)
 947                 ; /* nothing */
 948
 949         size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
 950         switch (qp->transport) {
 951         case MLX:
 952                 size += 2 * sizeof (struct mthca_data_seg);
 953                 break;
 954
 955         case UD:
 956                 size += mthca_is_memfree(dev) ?
 957                         sizeof (struct mthca_arbel_ud_seg) :
 958                         sizeof (struct mthca_tavor_ud_seg);
 959                 break;
 960
 961         case UC:
 962                 size += sizeof (struct mthca_raddr_seg);
 963                 break;
 964
 965         case RC:
 966                 size += sizeof (struct mthca_raddr_seg);
 967                 /*
 968                  * An atomic op will require an atomic segment, a
 969                  * remote address segment and one scatter entry.
 970                  */
 971                 size = max_t(int, size,
 972                              sizeof (struct mthca_atomic_seg) +
 973                              sizeof (struct mthca_raddr_seg) +
 974                              sizeof (struct mthca_data_seg));
 975                 break;
 976
 977         default:
 978                 break;
 979         }
 980
 981         /* Make sure that we have enough space for a bind request */
 982         size = max_t(int, size, sizeof (struct mthca_bind_seg));
 983
 984         size += sizeof (struct mthca_next_seg);
 985
 986         if (size > dev->limits.max_desc_sz)
 987                 return -EINVAL;
 988
 989         for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
 990              qp->sq.wqe_shift++)
 991                 ; /* nothing */
 992
 993         qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
 994                                     1 << qp->sq.wqe_shift);
 995
 996         /*
 997          * If this is a userspace QP, we don't actually have to
 998          * allocate anything.  All we need is to calculate the WQE
 999          * sizes and the send_wqe_offset, so we're done now.
1000          */
1001         if (pd->ibpd.uobject)
1002                 return 0;
1003
1004         size = PAGE_ALIGN(qp->send_wqe_offset +
1005                           (qp->sq.max << qp->sq.wqe_shift));
1006
1007         qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1008                            GFP_KERNEL);
1009         if (!qp->wrid)
1010                 goto err_out;
1011
1012         err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1013                               &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1014         if (err)
1015                 goto err_out;
1016
1017         return 0;
1018
1019 err_out:
1020         kfree(qp->wrid);
1021         return err;
1022 }
1023
1024 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1025                                struct mthca_qp *qp)
1026 {
1027         mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1028                                        (qp->sq.max << qp->sq.wqe_shift)),
1029                        &qp->queue, qp->is_direct, &qp->mr);
1030         kfree(qp->wrid);
1031 }
1032
1033 static int mthca_map_memfree(struct mthca_dev *dev,
1034                              struct mthca_qp *qp)
1035 {
1036         int ret;
1037
1038         if (mthca_is_memfree(dev)) {
1039                 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1040                 if (ret)
1041                         return ret;
1042
1043                 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1044                 if (ret)
1045                         goto err_qpc;
1046
1047                 ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1048                                       qp->qpn << dev->qp_table.rdb_shift);
1049                 if (ret)
1050                         goto err_eqpc;
1051
1052         }
1053
1054         return 0;
1055
1056 err_eqpc:
1057         mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1058
1059 err_qpc:
1060         mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1061
1062         return ret;
1063 }
1064
1065 static void mthca_unmap_memfree(struct mthca_dev *dev,
1066                                 struct mthca_qp *qp)
1067 {
1068         mthca_table_put(dev, dev->qp_table.rdb_table,
1069                         qp->qpn << dev->qp_table.rdb_shift);
1070         mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1071         mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1072 }
1073
1074 static int mthca_alloc_memfree(struct mthca_dev *dev,
1075                                struct mthca_qp *qp)
1076 {
1077         int ret = 0;
1078
1079         if (mthca_is_memfree(dev)) {
1080                 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1081                                                  qp->qpn, &qp->rq.db);
1082                 if (qp->rq.db_index < 0)
1083                         return ret;
1084
1085                 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1086                                                  qp->qpn, &qp->sq.db);
1087                 if (qp->sq.db_index < 0)
1088                         mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1089         }
1090
1091         return ret;
1092 }
1093
1094 static void mthca_free_memfree(struct mthca_dev *dev,
1095                                struct mthca_qp *qp)
1096 {
1097         if (mthca_is_memfree(dev)) {
1098                 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1099                 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1100         }
1101 }
1102
1103 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1104                                  struct mthca_pd *pd,
1105                                  struct mthca_cq *send_cq,
1106                                  struct mthca_cq *recv_cq,
1107                                  enum ib_sig_type send_policy,
1108                                  struct mthca_qp *qp)
1109 {
1110         int ret;
1111         int i;
1112
1113         qp->refcount = 1;
1114         init_waitqueue_head(&qp->wait);
1115         mutex_init(&qp->mutex);
1116         qp->state        = IB_QPS_RESET;
1117         qp->atomic_rd_en = 0;
1118         qp->resp_depth   = 0;
1119         qp->sq_policy    = send_policy;
1120         mthca_wq_reset(&qp->sq);
1121         mthca_wq_reset(&qp->rq);
1122
1123         spin_lock_init(&qp->sq.lock);
1124         spin_lock_init(&qp->rq.lock);
1125
1126         ret = mthca_map_memfree(dev, qp);
1127         if (ret)
1128                 return ret;
1129
1130         ret = mthca_alloc_wqe_buf(dev, pd, qp);
1131         if (ret) {
1132                 mthca_unmap_memfree(dev, qp);
1133                 return ret;
1134         }
1135
1136         mthca_adjust_qp_caps(dev, pd, qp);
1137
1138         /*
1139          * If this is a userspace QP, we're done now.  The doorbells
1140          * will be allocated and buffers will be initialized in
1141          * userspace.
1142          */
1143         if (pd->ibpd.uobject)
1144                 return 0;
1145
1146         ret = mthca_alloc_memfree(dev, qp);
1147         if (ret) {
1148                 mthca_free_wqe_buf(dev, qp);
1149                 mthca_unmap_memfree(dev, qp);
1150                 return ret;
1151         }
1152
1153         if (mthca_is_memfree(dev)) {
1154                 struct mthca_next_seg *next;
1155                 struct mthca_data_seg *scatter;
1156                 int size = (sizeof (struct mthca_next_seg) +
1157                             qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1158
1159                 for (i = 0; i < qp->rq.max; ++i) {
1160                         next = get_recv_wqe(qp, i);
1161                         next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1162                                                    qp->rq.wqe_shift);
1163                         next->ee_nds = cpu_to_be32(size);
1164
1165                         for (scatter = (void *) (next + 1);
1166                              (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1167                              ++scatter)
1168                                 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1169                 }
1170
1171                 for (i = 0; i < qp->sq.max; ++i) {
1172                         next = get_send_wqe(qp, i);
1173                         next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1174                                                     qp->sq.wqe_shift) +
1175                                                    qp->send_wqe_offset);
1176                 }
1177         }
1178
1179         qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1180         qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1181
1182         return 0;
1183 }
1184
1185 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1186                              struct mthca_pd *pd, struct mthca_qp *qp)
1187 {
1188         int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1189
1190         /* Sanity check QP size before proceeding */
1191         if (cap->max_send_wr     > dev->limits.max_wqes ||
1192             cap->max_recv_wr     > dev->limits.max_wqes ||
1193             cap->max_send_sge    > dev->limits.max_sg   ||
1194             cap->max_recv_sge    > dev->limits.max_sg   ||
1195             cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1196                 return -EINVAL;
1197
1198         /*
1199          * For MLX transport we need 2 extra S/G entries:
1200          * one for the header and one for the checksum at the end
1201          */
1202         if (qp->transport == MLX && cap->max_recv_sge + 2 > dev->limits.max_sg)
1203                 return -EINVAL;
1204
1205         if (mthca_is_memfree(dev)) {
1206                 qp->rq.max = cap->max_recv_wr ?
1207                         roundup_pow_of_two(cap->max_recv_wr) : 0;
1208                 qp->sq.max = cap->max_send_wr ?
1209                         roundup_pow_of_two(cap->max_send_wr) : 0;
1210         } else {
1211                 qp->rq.max = cap->max_recv_wr;
1212                 qp->sq.max = cap->max_send_wr;
1213         }
1214
1215         qp->rq.max_gs = cap->max_recv_sge;
1216         qp->sq.max_gs = max_t(int, cap->max_send_sge,
1217                               ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1218                                     MTHCA_INLINE_CHUNK_SIZE) /
1219                               sizeof (struct mthca_data_seg));
1220
1221         return 0;
1222 }
1223
1224 int mthca_alloc_qp(struct mthca_dev *dev,
1225                    struct mthca_pd *pd,
1226                    struct mthca_cq *send_cq,
1227                    struct mthca_cq *recv_cq,
1228                    enum ib_qp_type type,
1229                    enum ib_sig_type send_policy,
1230                    struct ib_qp_cap *cap,
1231                    struct mthca_qp *qp)
1232 {
1233         int err;
1234
1235         switch (type) {
1236         case IB_QPT_RC: qp->transport = RC; break;
1237         case IB_QPT_UC: qp->transport = UC; break;
1238         case IB_QPT_UD: qp->transport = UD; break;
1239         default: return -EINVAL;
1240         }
1241
1242         err = mthca_set_qp_size(dev, cap, pd, qp);
1243         if (err)
1244                 return err;
1245
1246         qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1247         if (qp->qpn == -1)
1248                 return -ENOMEM;
1249
1250         /* initialize port to zero for error-catching. */
1251         qp->port = 0;
1252
1253         err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1254                                     send_policy, qp);
1255         if (err) {
1256                 mthca_free(&dev->qp_table.alloc, qp->qpn);
1257                 return err;
1258         }
1259
1260         spin_lock_irq(&dev->qp_table.lock);
1261         mthca_array_set(&dev->qp_table.qp,
1262                         qp->qpn & (dev->limits.num_qps - 1), qp);
1263         spin_unlock_irq(&dev->qp_table.lock);
1264
1265         return 0;
1266 }
1267
1268 static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1269 {
1270         if (send_cq == recv_cq)
1271                 spin_lock_irq(&send_cq->lock);
1272         else if (send_cq->cqn < recv_cq->cqn) {
1273                 spin_lock_irq(&send_cq->lock);
1274                 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1275         } else {
1276                 spin_lock_irq(&recv_cq->lock);
1277                 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1278         }
1279 }
1280
1281 static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1282 {
1283         if (send_cq == recv_cq)
1284                 spin_unlock_irq(&send_cq->lock);
1285         else if (send_cq->cqn < recv_cq->cqn) {
1286                 spin_unlock(&recv_cq->lock);
1287                 spin_unlock_irq(&send_cq->lock);
1288         } else {
1289                 spin_unlock(&send_cq->lock);
1290                 spin_unlock_irq(&recv_cq->lock);
1291         }
1292 }
1293
1294 int mthca_alloc_sqp(struct mthca_dev *dev,
1295                     struct mthca_pd *pd,
1296                     struct mthca_cq *send_cq,
1297                     struct mthca_cq *recv_cq,
1298                     enum ib_sig_type send_policy,
1299                     struct ib_qp_cap *cap,
1300                     int qpn,
1301                     int port,
1302                     struct mthca_sqp *sqp)
1303 {
1304         u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1305         int err;
1306
1307         sqp->qp.transport = MLX;
1308         err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1309         if (err)
1310                 return err;
1311
1312         sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1313         sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1314                                              &sqp->header_dma, GFP_KERNEL);
1315         if (!sqp->header_buf)
1316                 return -ENOMEM;
1317
1318         spin_lock_irq(&dev->qp_table.lock);
1319         if (mthca_array_get(&dev->qp_table.qp, mqpn))
1320                 err = -EBUSY;
1321         else
1322                 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1323         spin_unlock_irq(&dev->qp_table.lock);
1324
1325         if (err)
1326                 goto err_out;
1327
1328         sqp->qp.port      = port;
1329         sqp->qp.qpn       = mqpn;
1330         sqp->qp.transport = MLX;
1331
1332         err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1333                                     send_policy, &sqp->qp);
1334         if (err)
1335                 goto err_out_free;
1336
1337         atomic_inc(&pd->sqp_count);
1338
1339         return 0;
1340
1341  err_out_free:
1342         /*
1343          * Lock CQs here, so that CQ polling code can do QP lookup
1344          * without taking a lock.
1345          */
1346         mthca_lock_cqs(send_cq, recv_cq);
1347
1348         spin_lock(&dev->qp_table.lock);
1349         mthca_array_clear(&dev->qp_table.qp, mqpn);
1350         spin_unlock(&dev->qp_table.lock);
1351
1352         mthca_unlock_cqs(send_cq, recv_cq);
1353
1354  err_out:
1355         dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1356                           sqp->header_buf, sqp->header_dma);
1357
1358         return err;
1359 }
1360
1361 static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp)
1362 {
1363         int c;
1364
1365         spin_lock_irq(&dev->qp_table.lock);
1366         c = qp->refcount;
1367         spin_unlock_irq(&dev->qp_table.lock);
1368
1369         return c;
1370 }
1371
1372 void mthca_free_qp(struct mthca_dev *dev,
1373                    struct mthca_qp *qp)
1374 {
1375         u8 status;
1376         struct mthca_cq *send_cq;
1377         struct mthca_cq *recv_cq;
1378
1379         send_cq = to_mcq(qp->ibqp.send_cq);
1380         recv_cq = to_mcq(qp->ibqp.recv_cq);
1381
1382         /*
1383          * Lock CQs here, so that CQ polling code can do QP lookup
1384          * without taking a lock.
1385          */
1386         mthca_lock_cqs(send_cq, recv_cq);
1387
1388         spin_lock(&dev->qp_table.lock);
1389         mthca_array_clear(&dev->qp_table.qp,
1390                           qp->qpn & (dev->limits.num_qps - 1));
1391         --qp->refcount;
1392         spin_unlock(&dev->qp_table.lock);
1393
1394         mthca_unlock_cqs(send_cq, recv_cq);
1395
1396         wait_event(qp->wait, !get_qp_refcount(dev, qp));
1397
1398         if (qp->state != IB_QPS_RESET)
1399                 mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1400                                 NULL, 0, &status);
1401
1402         /*
1403          * If this is a userspace QP, the buffers, MR, CQs and so on
1404          * will be cleaned up in userspace, so all we have to do is
1405          * unref the mem-free tables and free the QPN in our table.
1406          */
1407         if (!qp->ibqp.uobject) {
1408                 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn,
1409                                qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1410                 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
1411                         mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
1412                                        qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1413
1414                 mthca_free_memfree(dev, qp);
1415                 mthca_free_wqe_buf(dev, qp);
1416         }
1417
1418         mthca_unmap_memfree(dev, qp);
1419
1420         if (is_sqp(dev, qp)) {
1421                 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1422                 dma_free_coherent(&dev->pdev->dev,
1423                                   to_msqp(qp)->header_buf_size,
1424                                   to_msqp(qp)->header_buf,
1425                                   to_msqp(qp)->header_dma);
1426         } else
1427                 mthca_free(&dev->qp_table.alloc, qp->qpn);
1428 }
1429
1430 /* Create UD header for an MLX send and build a data segment for it */
1431 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1432                             int ind, struct ib_send_wr *wr,
1433                             struct mthca_mlx_seg *mlx,
1434                             struct mthca_data_seg *data)
1435 {
1436         int header_size;
1437         int err;
1438         u16 pkey;
1439
1440         ib_ud_header_init(256, /* assume a MAD */
1441                           mthca_ah_grh_present(to_mah(wr->wr.ud.ah)),
1442                           &sqp->ud_header);
1443
1444         err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1445         if (err)
1446                 return err;
1447         mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1448         mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1449                                   (sqp->ud_header.lrh.destination_lid ==
1450                                    IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1451                                   (sqp->ud_header.lrh.service_level << 8));
1452         mlx->rlid = sqp->ud_header.lrh.destination_lid;
1453         mlx->vcrc = 0;
1454
1455         switch (wr->opcode) {
1456         case IB_WR_SEND:
1457                 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1458                 sqp->ud_header.immediate_present = 0;
1459                 break;
1460         case IB_WR_SEND_WITH_IMM:
1461                 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1462                 sqp->ud_header.immediate_present = 1;
1463                 sqp->ud_header.immediate_data = wr->imm_data;
1464                 break;
1465         default:
1466                 return -EINVAL;
1467         }
1468
1469         sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1470         if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1471                 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1472         sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1473         if (!sqp->qp.ibqp.qp_num)
1474                 ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1475                                    sqp->pkey_index, &pkey);
1476         else
1477                 ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1478                                    wr->wr.ud.pkey_index, &pkey);
1479         sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1480         sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1481         sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1482         sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1483                                                sqp->qkey : wr->wr.ud.remote_qkey);
1484         sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1485
1486         header_size = ib_ud_header_pack(&sqp->ud_header,
1487                                         sqp->header_buf +
1488                                         ind * MTHCA_UD_HEADER_SIZE);
1489
1490         data->byte_count = cpu_to_be32(header_size);
1491         data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1492         data->addr       = cpu_to_be64(sqp->header_dma +
1493                                        ind * MTHCA_UD_HEADER_SIZE);
1494
1495         return 0;
1496 }
1497
1498 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1499                                     struct ib_cq *ib_cq)
1500 {
1501         unsigned cur;
1502         struct mthca_cq *cq;
1503
1504         cur = wq->head - wq->tail;
1505         if (likely(cur + nreq < wq->max))
1506                 return 0;
1507
1508         cq = to_mcq(ib_cq);
1509         spin_lock(&cq->lock);
1510         cur = wq->head - wq->tail;
1511         spin_unlock(&cq->lock);
1512
1513         return cur + nreq >= wq->max;
1514 }
1515
1516 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1517                           struct ib_send_wr **bad_wr)
1518 {
1519         struct mthca_dev *dev = to_mdev(ibqp->device);
1520         struct mthca_qp *qp = to_mqp(ibqp);
1521         void *wqe;
1522         void *prev_wqe;
1523         unsigned long flags;
1524         int err = 0;
1525         int nreq;
1526         int i;
1527         int size;
1528         int size0 = 0;
1529         u32 f0;
1530         int ind;
1531         u8 op0 = 0;
1532
1533         spin_lock_irqsave(&qp->sq.lock, flags);
1534
1535         /* XXX check that state is OK to post send */
1536
1537         ind = qp->sq.next_ind;
1538
1539         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1540                 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1541                         mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1542                                         " %d max, %d nreq)\n", qp->qpn,
1543                                         qp->sq.head, qp->sq.tail,
1544                                         qp->sq.max, nreq);
1545                         err = -ENOMEM;
1546                         *bad_wr = wr;
1547                         goto out;
1548                 }
1549
1550                 wqe = get_send_wqe(qp, ind);
1551                 prev_wqe = qp->sq.last;
1552                 qp->sq.last = wqe;
1553
1554                 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1555                 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1556                 ((struct mthca_next_seg *) wqe)->flags =
1557                         ((wr->send_flags & IB_SEND_SIGNALED) ?
1558                          cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1559                         ((wr->send_flags & IB_SEND_SOLICITED) ?
1560                          cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1561                         cpu_to_be32(1);
1562                 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1563                     wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1564                         ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1565
1566                 wqe += sizeof (struct mthca_next_seg);
1567                 size = sizeof (struct mthca_next_seg) / 16;
1568
1569                 switch (qp->transport) {
1570                 case RC:
1571                         switch (wr->opcode) {
1572                         case IB_WR_ATOMIC_CMP_AND_SWP:
1573                         case IB_WR_ATOMIC_FETCH_AND_ADD:
1574                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1575                                         cpu_to_be64(wr->wr.atomic.remote_addr);
1576                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1577                                         cpu_to_be32(wr->wr.atomic.rkey);
1578                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1579
1580                                 wqe += sizeof (struct mthca_raddr_seg);
1581
1582                                 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1583                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1584                                                 cpu_to_be64(wr->wr.atomic.swap);
1585                                         ((struct mthca_atomic_seg *) wqe)->compare =
1586                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1587                                 } else {
1588                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1589                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1590                                         ((struct mthca_atomic_seg *) wqe)->compare = 0;
1591                                 }
1592
1593                                 wqe += sizeof (struct mthca_atomic_seg);
1594                                 size += (sizeof (struct mthca_raddr_seg) +
1595                                          sizeof (struct mthca_atomic_seg)) / 16;
1596                                 break;
1597
1598                         case IB_WR_RDMA_WRITE:
1599                         case IB_WR_RDMA_WRITE_WITH_IMM:
1600                         case IB_WR_RDMA_READ:
1601                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1602                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1603                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1604                                         cpu_to_be32(wr->wr.rdma.rkey);
1605                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1606                                 wqe += sizeof (struct mthca_raddr_seg);
1607                                 size += sizeof (struct mthca_raddr_seg) / 16;
1608                                 break;
1609
1610                         default:
1611                                 /* No extra segments required for sends */
1612                                 break;
1613                         }
1614
1615                         break;
1616
1617                 case UC:
1618                         switch (wr->opcode) {
1619                         case IB_WR_RDMA_WRITE:
1620                         case IB_WR_RDMA_WRITE_WITH_IMM:
1621                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1622                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1623                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1624                                         cpu_to_be32(wr->wr.rdma.rkey);
1625                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1626                                 wqe += sizeof (struct mthca_raddr_seg);
1627                                 size += sizeof (struct mthca_raddr_seg) / 16;
1628                                 break;
1629
1630                         default:
1631                                 /* No extra segments required for sends */
1632                                 break;
1633                         }
1634
1635                         break;
1636
1637                 case UD:
1638                         ((struct mthca_tavor_ud_seg *) wqe)->lkey =
1639                                 cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1640                         ((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1641                                 cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1642                         ((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1643                                 cpu_to_be32(wr->wr.ud.remote_qpn);
1644                         ((struct mthca_tavor_ud_seg *) wqe)->qkey =
1645                                 cpu_to_be32(wr->wr.ud.remote_qkey);
1646
1647                         wqe += sizeof (struct mthca_tavor_ud_seg);
1648                         size += sizeof (struct mthca_tavor_ud_seg) / 16;
1649                         break;
1650
1651                 case MLX:
1652                         err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1653                                                wqe - sizeof (struct mthca_next_seg),
1654                                                wqe);
1655                         if (err) {
1656                                 *bad_wr = wr;
1657                                 goto out;
1658                         }
1659                         wqe += sizeof (struct mthca_data_seg);
1660                         size += sizeof (struct mthca_data_seg) / 16;
1661                         break;
1662                 }
1663
1664                 if (wr->num_sge > qp->sq.max_gs) {
1665                         mthca_err(dev, "too many gathers\n");
1666                         err = -EINVAL;
1667                         *bad_wr = wr;
1668                         goto out;
1669                 }
1670
1671                 for (i = 0; i < wr->num_sge; ++i) {
1672                         ((struct mthca_data_seg *) wqe)->byte_count =
1673                                 cpu_to_be32(wr->sg_list[i].length);
1674                         ((struct mthca_data_seg *) wqe)->lkey =
1675                                 cpu_to_be32(wr->sg_list[i].lkey);
1676                         ((struct mthca_data_seg *) wqe)->addr =
1677                                 cpu_to_be64(wr->sg_list[i].addr);
1678                         wqe += sizeof (struct mthca_data_seg);
1679                         size += sizeof (struct mthca_data_seg) / 16;
1680                 }
1681
1682                 /* Add one more inline data segment for ICRC */
1683                 if (qp->transport == MLX) {
1684                         ((struct mthca_data_seg *) wqe)->byte_count =
1685                                 cpu_to_be32((1 << 31) | 4);
1686                         ((u32 *) wqe)[1] = 0;
1687                         wqe += sizeof (struct mthca_data_seg);
1688                         size += sizeof (struct mthca_data_seg) / 16;
1689                 }
1690
1691                 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1692
1693                 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1694                         mthca_err(dev, "opcode invalid\n");
1695                         err = -EINVAL;
1696                         *bad_wr = wr;
1697                         goto out;
1698                 }
1699
1700                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1701                         cpu_to_be32(((ind << qp->sq.wqe_shift) +
1702                                      qp->send_wqe_offset) |
1703                                     mthca_opcode[wr->opcode]);
1704                 wmb();
1705                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1706                         cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size |
1707                                     ((wr->send_flags & IB_SEND_FENCE) ?
1708                                     MTHCA_NEXT_FENCE : 0));
1709
1710                 if (!size0) {
1711                         size0 = size;
1712                         op0   = mthca_opcode[wr->opcode];
1713                         f0    = wr->send_flags & IB_SEND_FENCE ?
1714                                 MTHCA_SEND_DOORBELL_FENCE : 0;
1715                 }
1716
1717                 ++ind;
1718                 if (unlikely(ind >= qp->sq.max))
1719                         ind -= qp->sq.max;
1720         }
1721
1722 out:
1723         if (likely(nreq)) {
1724                 __be32 doorbell[2];
1725
1726                 doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1727                                            qp->send_wqe_offset) | f0 | op0);
1728                 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1729
1730                 wmb();
1731
1732                 mthca_write64(doorbell,
1733                               dev->kar + MTHCA_SEND_DOORBELL,
1734                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1735                 /*
1736                  * Make sure doorbells don't leak out of SQ spinlock
1737                  * and reach the HCA out of order:
1738                  */
1739                 mmiowb();
1740         }
1741
1742         qp->sq.next_ind = ind;
1743         qp->sq.head    += nreq;
1744
1745         spin_unlock_irqrestore(&qp->sq.lock, flags);
1746         return err;
1747 }
1748
1749 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1750                              struct ib_recv_wr **bad_wr)
1751 {
1752         struct mthca_dev *dev = to_mdev(ibqp->device);
1753         struct mthca_qp *qp = to_mqp(ibqp);
1754         __be32 doorbell[2];
1755         unsigned long flags;
1756         int err = 0;
1757         int nreq;
1758         int i;
1759         int size;
1760         int size0 = 0;
1761         int ind;
1762         void *wqe;
1763         void *prev_wqe;
1764
1765         spin_lock_irqsave(&qp->rq.lock, flags);
1766
1767         /* XXX check that state is OK to post receive */
1768
1769         ind = qp->rq.next_ind;
1770
1771         for (nreq = 0; wr; wr = wr->next) {
1772                 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1773                         mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1774                                         " %d max, %d nreq)\n", qp->qpn,
1775                                         qp->rq.head, qp->rq.tail,
1776                                         qp->rq.max, nreq);
1777                         err = -ENOMEM;
1778                         *bad_wr = wr;
1779                         goto out;
1780                 }
1781
1782                 wqe = get_recv_wqe(qp, ind);
1783                 prev_wqe = qp->rq.last;
1784                 qp->rq.last = wqe;
1785
1786                 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1787                 ((struct mthca_next_seg *) wqe)->ee_nds =
1788                         cpu_to_be32(MTHCA_NEXT_DBD);
1789                 ((struct mthca_next_seg *) wqe)->flags = 0;
1790
1791                 wqe += sizeof (struct mthca_next_seg);
1792                 size = sizeof (struct mthca_next_seg) / 16;
1793
1794                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1795                         err = -EINVAL;
1796                         *bad_wr = wr;
1797                         goto out;
1798                 }
1799
1800                 for (i = 0; i < wr->num_sge; ++i) {
1801                         ((struct mthca_data_seg *) wqe)->byte_count =
1802                                 cpu_to_be32(wr->sg_list[i].length);
1803                         ((struct mthca_data_seg *) wqe)->lkey =
1804                                 cpu_to_be32(wr->sg_list[i].lkey);
1805                         ((struct mthca_data_seg *) wqe)->addr =
1806                                 cpu_to_be64(wr->sg_list[i].addr);
1807                         wqe += sizeof (struct mthca_data_seg);
1808                         size += sizeof (struct mthca_data_seg) / 16;
1809                 }
1810
1811                 qp->wrid[ind] = wr->wr_id;
1812
1813                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1814                         cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1815                 wmb();
1816                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1817                         cpu_to_be32(MTHCA_NEXT_DBD | size);
1818
1819                 if (!size0)
1820                         size0 = size;
1821
1822                 ++ind;
1823                 if (unlikely(ind >= qp->rq.max))
1824                         ind -= qp->rq.max;
1825
1826                 ++nreq;
1827                 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1828                         nreq = 0;
1829
1830                         doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1831                         doorbell[1] = cpu_to_be32(qp->qpn << 8);
1832
1833                         wmb();
1834
1835                         mthca_write64(doorbell,
1836                                       dev->kar + MTHCA_RECEIVE_DOORBELL,
1837                                       MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1838
1839                         qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1840                         size0 = 0;
1841                 }
1842         }
1843
1844 out:
1845         if (likely(nreq)) {
1846                 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1847                 doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
1848
1849                 wmb();
1850
1851                 mthca_write64(doorbell,
1852                               dev->kar + MTHCA_RECEIVE_DOORBELL,
1853                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1854         }
1855
1856         qp->rq.next_ind = ind;
1857         qp->rq.head    += nreq;
1858
1859         /*
1860          * Make sure doorbells don't leak out of RQ spinlock and reach
1861          * the HCA out of order:
1862          */
1863         mmiowb();
1864
1865         spin_unlock_irqrestore(&qp->rq.lock, flags);
1866         return err;
1867 }
1868
1869 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1870                           struct ib_send_wr **bad_wr)
1871 {
1872         struct mthca_dev *dev = to_mdev(ibqp->device);
1873         struct mthca_qp *qp = to_mqp(ibqp);
1874         __be32 doorbell[2];
1875         void *wqe;
1876         void *prev_wqe;
1877         unsigned long flags;
1878         int err = 0;
1879         int nreq;
1880         int i;
1881         int size;
1882         int size0 = 0;
1883         u32 f0;
1884         int ind;
1885         u8 op0 = 0;
1886
1887         spin_lock_irqsave(&qp->sq.lock, flags);
1888
1889         /* XXX check that state is OK to post send */
1890
1891         ind = qp->sq.head & (qp->sq.max - 1);
1892
1893         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1894                 if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1895                         nreq = 0;
1896
1897                         doorbell[0] = cpu_to_be32((MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1898                                                   ((qp->sq.head & 0xffff) << 8) |
1899                                                   f0 | op0);
1900                         doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1901
1902                         qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1903                         size0 = 0;
1904
1905                         /*
1906                          * Make sure that descriptors are written before
1907                          * doorbell record.
1908                          */
1909                         wmb();
1910                         *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1911
1912                         /*
1913                          * Make sure doorbell record is written before we
1914                          * write MMIO send doorbell.
1915                          */
1916                         wmb();
1917                         mthca_write64(doorbell,
1918                                       dev->kar + MTHCA_SEND_DOORBELL,
1919                                       MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1920                 }
1921
1922                 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1923                         mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1924                                         " %d max, %d nreq)\n", qp->qpn,
1925                                         qp->sq.head, qp->sq.tail,
1926                                         qp->sq.max, nreq);
1927                         err = -ENOMEM;
1928                         *bad_wr = wr;
1929                         goto out;
1930                 }
1931
1932                 wqe = get_send_wqe(qp, ind);
1933                 prev_wqe = qp->sq.last;
1934                 qp->sq.last = wqe;
1935
1936                 ((struct mthca_next_seg *) wqe)->flags =
1937                         ((wr->send_flags & IB_SEND_SIGNALED) ?
1938                          cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1939                         ((wr->send_flags & IB_SEND_SOLICITED) ?
1940                          cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1941                         cpu_to_be32(1);
1942                 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1943                     wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1944                         ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1945
1946                 wqe += sizeof (struct mthca_next_seg);
1947                 size = sizeof (struct mthca_next_seg) / 16;
1948
1949                 switch (qp->transport) {
1950                 case RC:
1951                         switch (wr->opcode) {
1952                         case IB_WR_ATOMIC_CMP_AND_SWP:
1953                         case IB_WR_ATOMIC_FETCH_AND_ADD:
1954                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1955                                         cpu_to_be64(wr->wr.atomic.remote_addr);
1956                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1957                                         cpu_to_be32(wr->wr.atomic.rkey);
1958                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1959
1960                                 wqe += sizeof (struct mthca_raddr_seg);
1961
1962                                 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1963                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1964                                                 cpu_to_be64(wr->wr.atomic.swap);
1965                                         ((struct mthca_atomic_seg *) wqe)->compare =
1966                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1967                                 } else {
1968                                         ((struct mthca_atomic_seg *) wqe)->swap_add =
1969                                                 cpu_to_be64(wr->wr.atomic.compare_add);
1970                                         ((struct mthca_atomic_seg *) wqe)->compare = 0;
1971                                 }
1972
1973                                 wqe += sizeof (struct mthca_atomic_seg);
1974                                 size += (sizeof (struct mthca_raddr_seg) +
1975                                          sizeof (struct mthca_atomic_seg)) / 16;
1976                                 break;
1977
1978                         case IB_WR_RDMA_READ:
1979                         case IB_WR_RDMA_WRITE:
1980                         case IB_WR_RDMA_WRITE_WITH_IMM:
1981                                 ((struct mthca_raddr_seg *) wqe)->raddr =
1982                                         cpu_to_be64(wr->wr.rdma.remote_addr);
1983                                 ((struct mthca_raddr_seg *) wqe)->rkey =
1984                                         cpu_to_be32(wr->wr.rdma.rkey);
1985                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1986                                 wqe += sizeof (struct mthca_raddr_seg);
1987                                 size += sizeof (struct mthca_raddr_seg) / 16;
1988                                 break;
1989
1990                         default:
1991                                 /* No extra segments required for sends */
1992                                 break;
1993                         }
1994
1995                         break;
1996
1997                 case UC:
1998                         switch (wr->opcode) {
1999                         case IB_WR_RDMA_WRITE:
2000                         case IB_WR_RDMA_WRITE_WITH_IMM:
2001                                 ((struct mthca_raddr_seg *) wqe)->raddr =
2002                                         cpu_to_be64(wr->wr.rdma.remote_addr);
2003                                 ((struct mthca_raddr_seg *) wqe)->rkey =
2004                                         cpu_to_be32(wr->wr.rdma.rkey);
2005                                 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
2006                                 wqe += sizeof (struct mthca_raddr_seg);
2007                                 size += sizeof (struct mthca_raddr_seg) / 16;
2008                                 break;
2009
2010                         default:
2011                                 /* No extra segments required for sends */
2012                                 break;
2013                         }
2014
2015                         break;
2016
2017                 case UD:
2018                         memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
2019                                to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
2020                         ((struct mthca_arbel_ud_seg *) wqe)->dqpn =
2021                                 cpu_to_be32(wr->wr.ud.remote_qpn);
2022                         ((struct mthca_arbel_ud_seg *) wqe)->qkey =
2023                                 cpu_to_be32(wr->wr.ud.remote_qkey);
2024
2025                         wqe += sizeof (struct mthca_arbel_ud_seg);
2026                         size += sizeof (struct mthca_arbel_ud_seg) / 16;
2027                         break;
2028
2029                 case MLX:
2030                         err = build_mlx_header(dev, to_msqp(qp), ind, wr,
2031                                                wqe - sizeof (struct mthca_next_seg),
2032                                                wqe);
2033                         if (err) {
2034                                 *bad_wr = wr;
2035                                 goto out;
2036                         }
2037                         wqe += sizeof (struct mthca_data_seg);
2038                         size += sizeof (struct mthca_data_seg) / 16;
2039                         break;
2040                 }
2041
2042                 if (wr->num_sge > qp->sq.max_gs) {
2043                         mthca_err(dev, "too many gathers\n");
2044                         err = -EINVAL;
2045                         *bad_wr = wr;
2046                         goto out;
2047                 }
2048
2049                 for (i = 0; i < wr->num_sge; ++i) {
2050                         ((struct mthca_data_seg *) wqe)->byte_count =
2051                                 cpu_to_be32(wr->sg_list[i].length);
2052                         ((struct mthca_data_seg *) wqe)->lkey =
2053                                 cpu_to_be32(wr->sg_list[i].lkey);
2054                         ((struct mthca_data_seg *) wqe)->addr =
2055                                 cpu_to_be64(wr->sg_list[i].addr);
2056                         wqe += sizeof (struct mthca_data_seg);
2057                         size += sizeof (struct mthca_data_seg) / 16;
2058                 }
2059
2060                 /* Add one more inline data segment for ICRC */
2061                 if (qp->transport == MLX) {
2062                         ((struct mthca_data_seg *) wqe)->byte_count =
2063                                 cpu_to_be32((1 << 31) | 4);
2064                         ((u32 *) wqe)[1] = 0;
2065                         wqe += sizeof (struct mthca_data_seg);
2066                         size += sizeof (struct mthca_data_seg) / 16;
2067                 }
2068
2069                 qp->wrid[ind + qp->rq.max] = wr->wr_id;
2070
2071                 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2072                         mthca_err(dev, "opcode invalid\n");
2073                         err = -EINVAL;
2074                         *bad_wr = wr;
2075                         goto out;
2076                 }
2077
2078                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
2079                         cpu_to_be32(((ind << qp->sq.wqe_shift) +
2080                                      qp->send_wqe_offset) |
2081                                     mthca_opcode[wr->opcode]);
2082                 wmb();
2083                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
2084                         cpu_to_be32(MTHCA_NEXT_DBD | size |
2085                                     ((wr->send_flags & IB_SEND_FENCE) ?
2086                                      MTHCA_NEXT_FENCE : 0));
2087
2088                 if (!size0) {
2089                         size0 = size;
2090                         op0   = mthca_opcode[wr->opcode];
2091                         f0    = wr->send_flags & IB_SEND_FENCE ?
2092                                 MTHCA_SEND_DOORBELL_FENCE : 0;
2093                 }
2094
2095                 ++ind;
2096                 if (unlikely(ind >= qp->sq.max))
2097                         ind -= qp->sq.max;
2098         }
2099
2100 out:
2101         if (likely(nreq)) {
2102                 doorbell[0] = cpu_to_be32((nreq << 24)                  |
2103                                           ((qp->sq.head & 0xffff) << 8) |
2104                                           f0 | op0);
2105                 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
2106
2107                 qp->sq.head += nreq;
2108
2109                 /*
2110                  * Make sure that descriptors are written before
2111                  * doorbell record.
2112                  */
2113                 wmb();
2114                 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2115
2116                 /*
2117                  * Make sure doorbell record is written before we
2118                  * write MMIO send doorbell.
2119                  */
2120                 wmb();
2121                 mthca_write64(doorbell,
2122                               dev->kar + MTHCA_SEND_DOORBELL,
2123                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2124         }
2125
2126         /*
2127          * Make sure doorbells don't leak out of SQ spinlock and reach
2128          * the HCA out of order:
2129          */
2130         mmiowb();
2131
2132         spin_unlock_irqrestore(&qp->sq.lock, flags);
2133         return err;
2134 }
2135
2136 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2137                              struct ib_recv_wr **bad_wr)
2138 {
2139         struct mthca_dev *dev = to_mdev(ibqp->device);
2140         struct mthca_qp *qp = to_mqp(ibqp);
2141         unsigned long flags;
2142         int err = 0;
2143         int nreq;
2144         int ind;
2145         int i;
2146         void *wqe;
2147
2148         spin_lock_irqsave(&qp->rq.lock, flags);
2149
2150         /* XXX check that state is OK to post receive */
2151
2152         ind = qp->rq.head & (qp->rq.max - 1);
2153
2154         for (nreq = 0; wr; ++nreq, wr = wr->next) {
2155                 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2156                         mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2157                                         " %d max, %d nreq)\n", qp->qpn,
2158                                         qp->rq.head, qp->rq.tail,
2159                                         qp->rq.max, nreq);
2160                         err = -ENOMEM;
2161                         *bad_wr = wr;
2162                         goto out;
2163                 }
2164
2165                 wqe = get_recv_wqe(qp, ind);
2166
2167                 ((struct mthca_next_seg *) wqe)->flags = 0;
2168
2169                 wqe += sizeof (struct mthca_next_seg);
2170
2171                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2172                         err = -EINVAL;
2173                         *bad_wr = wr;
2174                         goto out;
2175                 }
2176
2177                 for (i = 0; i < wr->num_sge; ++i) {
2178                         ((struct mthca_data_seg *) wqe)->byte_count =
2179                                 cpu_to_be32(wr->sg_list[i].length);
2180                         ((struct mthca_data_seg *) wqe)->lkey =
2181                                 cpu_to_be32(wr->sg_list[i].lkey);
2182                         ((struct mthca_data_seg *) wqe)->addr =
2183                                 cpu_to_be64(wr->sg_list[i].addr);
2184                         wqe += sizeof (struct mthca_data_seg);
2185                 }
2186
2187                 if (i < qp->rq.max_gs) {
2188                         ((struct mthca_data_seg *) wqe)->byte_count = 0;
2189                         ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
2190                         ((struct mthca_data_seg *) wqe)->addr = 0;
2191                 }
2192
2193                 qp->wrid[ind] = wr->wr_id;
2194
2195                 ++ind;
2196                 if (unlikely(ind >= qp->rq.max))
2197                         ind -= qp->rq.max;
2198         }
2199 out:
2200         if (likely(nreq)) {
2201                 qp->rq.head += nreq;
2202
2203                 /*
2204                  * Make sure that descriptors are written before
2205                  * doorbell record.
2206                  */
2207                 wmb();
2208                 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2209         }
2210
2211         spin_unlock_irqrestore(&qp->rq.lock, flags);
2212         return err;
2213 }
2214
2215 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2216                         int index, int *dbd, __be32 *new_wqe)
2217 {
2218         struct mthca_next_seg *next;
2219
2220         /*
2221          * For SRQs, all WQEs generate a CQE, so we're always at the
2222          * end of the doorbell chain.
2223          */
2224         if (qp->ibqp.srq) {
2225                 *new_wqe = 0;
2226                 return;
2227         }
2228
2229         if (is_send)
2230                 next = get_send_wqe(qp, index);
2231         else
2232                 next = get_recv_wqe(qp, index);
2233
2234         *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2235         if (next->ee_nds & cpu_to_be32(0x3f))
2236                 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2237                         (next->ee_nds & cpu_to_be32(0x3f));
2238         else
2239                 *new_wqe = 0;
2240 }
2241
2242 int __devinit mthca_init_qp_table(struct mthca_dev *dev)
2243 {
2244         int err;
2245         u8 status;
2246         int i;
2247
2248         spin_lock_init(&dev->qp_table.lock);
2249
2250         /*
2251          * We reserve 2 extra QPs per port for the special QPs.  The
2252          * special QP for port 1 has to be even, so round up.
2253          */
2254         dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2255         err = mthca_alloc_init(&dev->qp_table.alloc,
2256                                dev->limits.num_qps,
2257                                (1 << 24) - 1,
2258                                dev->qp_table.sqp_start +
2259                                MTHCA_MAX_PORTS * 2);
2260         if (err)
2261                 return err;
2262
2263         err = mthca_array_init(&dev->qp_table.qp,
2264                                dev->limits.num_qps);
2265         if (err) {
2266                 mthca_alloc_cleanup(&dev->qp_table.alloc);
2267                 return err;
2268         }
2269
2270         for (i = 0; i < 2; ++i) {
2271                 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2272                                             dev->qp_table.sqp_start + i * 2,
2273                                             &status);
2274                 if (err)
2275                         goto err_out;
2276                 if (status) {
2277                         mthca_warn(dev, "CONF_SPECIAL_QP returned "
2278                                    "status %02x, aborting.\n",
2279                                    status);
2280                         err = -EINVAL;
2281                         goto err_out;
2282                 }
2283         }
2284         return 0;
2285
2286  err_out:
2287         for (i = 0; i < 2; ++i)
2288                 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2289
2290         mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2291         mthca_alloc_cleanup(&dev->qp_table.alloc);
2292
2293         return err;
2294 }
2295
2296 void mthca_cleanup_qp_table(struct mthca_dev *dev)
2297 {
2298         int i;
2299         u8 status;
2300
2301         for (i = 0; i < 2; ++i)
2302                 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2303
2304         mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2305         mthca_alloc_cleanup(&dev->qp_table.alloc);
2306 }