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
  36 #include <linux/string.h>
  37 #include <linux/slab.h>
  38 #include <linux/sched.h>
  39
  40 #include <asm/io.h>
  41
  42 #include <rdma/ib_verbs.h>
  43 #include <rdma/ib_cache.h>
  44 #include <rdma/ib_pack.h>
  45
  46 #include "mthca_dev.h"
  47 #include "mthca_cmd.h"
  48 #include "mthca_memfree.h"
  49 #include "mthca_wqe.h"
  50
  51 enum {
  52         MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
  53         MTHCA_ACK_REQ_FREQ       = 10,
  54         MTHCA_FLIGHT_LIMIT       = 9,
  55         MTHCA_UD_HEADER_SIZE     = 72, /* largest UD header possible */
  56         MTHCA_INLINE_HEADER_SIZE = 4,  /* data segment overhead for inline */
  57         MTHCA_INLINE_CHUNK_SIZE  = 16  /* inline data segment chunk */
  58 };
  59
  60 enum {
  61         MTHCA_QP_STATE_RST  = 0,
  62         MTHCA_QP_STATE_INIT = 1,
  63         MTHCA_QP_STATE_RTR  = 2,
  64         MTHCA_QP_STATE_RTS  = 3,
  65         MTHCA_QP_STATE_SQE  = 4,
  66         MTHCA_QP_STATE_SQD  = 5,
  67         MTHCA_QP_STATE_ERR  = 6,
  68         MTHCA_QP_STATE_DRAINING = 7
  69 };
  70
  71 enum {
  72         MTHCA_QP_ST_RC  = 0x0,
  73         MTHCA_QP_ST_UC  = 0x1,
  74         MTHCA_QP_ST_RD  = 0x2,
  75         MTHCA_QP_ST_UD  = 0x3,
  76         MTHCA_QP_ST_MLX = 0x7
  77 };
  78
  79 enum {
  80         MTHCA_QP_PM_MIGRATED = 0x3,
  81         MTHCA_QP_PM_ARMED    = 0x0,
  82         MTHCA_QP_PM_REARM    = 0x1
  83 };
  84
  85 enum {
  86         /* qp_context flags */
  87         MTHCA_QP_BIT_DE  = 1 <<  8,
  88         /* params1 */
  89         MTHCA_QP_BIT_SRE = 1 << 15,
  90         MTHCA_QP_BIT_SWE = 1 << 14,
  91         MTHCA_QP_BIT_SAE = 1 << 13,
  92         MTHCA_QP_BIT_SIC = 1 <<  4,
  93         MTHCA_QP_BIT_SSC = 1 <<  3,
  94         /* params2 */
  95         MTHCA_QP_BIT_RRE = 1 << 15,
  96         MTHCA_QP_BIT_RWE = 1 << 14,
  97         MTHCA_QP_BIT_RAE = 1 << 13,
  98         MTHCA_QP_BIT_RIC = 1 <<  4,
  99         MTHCA_QP_BIT_RSC = 1 <<  3
 100 };
 101
 102 enum {
 103         MTHCA_SEND_DOORBELL_FENCE = 1 << 5
 104 };
 105
 106 struct mthca_qp_path {
 107         __be32 port_pkey;
 108         u8     rnr_retry;
 109         u8     g_mylmc;
 110         __be16 rlid;
 111         u8     ackto;
 112         u8     mgid_index;
 113         u8     static_rate;
 114         u8     hop_limit;
 115         __be32 sl_tclass_flowlabel;
 116         u8     rgid[16];
 117 } __attribute__((packed));
 118
 119 struct mthca_qp_context {
 120         __be32 flags;
 121         __be32 tavor_sched_queue; /* Reserved on Arbel */
 122         u8     mtu_msgmax;
 123         u8     rq_size_stride;  /* Reserved on Tavor */
 124         u8     sq_size_stride;  /* Reserved on Tavor */
 125         u8     rlkey_arbel_sched_queue; /* Reserved on Tavor */
 126         __be32 usr_page;
 127         __be32 local_qpn;
 128         __be32 remote_qpn;
 129         u32    reserved1[2];
 130         struct mthca_qp_path pri_path;
 131         struct mthca_qp_path alt_path;
 132         __be32 rdd;
 133         __be32 pd;
 134         __be32 wqe_base;
 135         __be32 wqe_lkey;
 136         __be32 params1;
 137         __be32 reserved2;
 138         __be32 next_send_psn;
 139         __be32 cqn_snd;
 140         __be32 snd_wqe_base_l;  /* Next send WQE on Tavor */
 141         __be32 snd_db_index;    /* (debugging only entries) */
 142         __be32 last_acked_psn;
 143         __be32 ssn;
 144         __be32 params2;
 145         __be32 rnr_nextrecvpsn;
 146         __be32 ra_buff_indx;
 147         __be32 cqn_rcv;
 148         __be32 rcv_wqe_base_l;  /* Next recv WQE on Tavor */
 149         __be32 rcv_db_index;    /* (debugging only entries) */
 150         __be32 qkey;
 151         __be32 srqn;
 152         __be32 rmsn;
 153         __be16 rq_wqe_counter;  /* reserved on Tavor */
 154         __be16 sq_wqe_counter;  /* reserved on Tavor */
 155         u32    reserved3[18];
 156 } __attribute__((packed));
 157
 158 struct mthca_qp_param {
 159         __be32 opt_param_mask;
 160         u32    reserved1;
 161         struct mthca_qp_context context;
 162         u32    reserved2[62];
 163 } __attribute__((packed));
 164
 165 enum {
 166         MTHCA_QP_OPTPAR_ALT_ADDR_PATH     = 1 << 0,
 167         MTHCA_QP_OPTPAR_RRE               = 1 << 1,
 168         MTHCA_QP_OPTPAR_RAE               = 1 << 2,
 169         MTHCA_QP_OPTPAR_RWE               = 1 << 3,
 170         MTHCA_QP_OPTPAR_PKEY_INDEX        = 1 << 4,
 171         MTHCA_QP_OPTPAR_Q_KEY             = 1 << 5,
 172         MTHCA_QP_OPTPAR_RNR_TIMEOUT       = 1 << 6,
 173         MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
 174         MTHCA_QP_OPTPAR_SRA_MAX           = 1 << 8,
 175         MTHCA_QP_OPTPAR_RRA_MAX           = 1 << 9,
 176         MTHCA_QP_OPTPAR_PM_STATE          = 1 << 10,
 177         MTHCA_QP_OPTPAR_PORT_NUM          = 1 << 11,
 178         MTHCA_QP_OPTPAR_RETRY_COUNT       = 1 << 12,
 179         MTHCA_QP_OPTPAR_ALT_RNR_RETRY     = 1 << 13,
 180         MTHCA_QP_OPTPAR_ACK_TIMEOUT       = 1 << 14,
 181         MTHCA_QP_OPTPAR_RNR_RETRY         = 1 << 15,
 182         MTHCA_QP_OPTPAR_SCHED_QUEUE       = 1 << 16
 183 };
 184
 185 static const u8 mthca_opcode[] = {
 186         [IB_WR_SEND]                 = MTHCA_OPCODE_SEND,
 187         [IB_WR_SEND_WITH_IMM]        = MTHCA_OPCODE_SEND_IMM,
 188         [IB_WR_RDMA_WRITE]           = MTHCA_OPCODE_RDMA_WRITE,
 189         [IB_WR_RDMA_WRITE_WITH_IMM]  = MTHCA_OPCODE_RDMA_WRITE_IMM,
 190         [IB_WR_RDMA_READ]            = MTHCA_OPCODE_RDMA_READ,
 191         [IB_WR_ATOMIC_CMP_AND_SWP]   = MTHCA_OPCODE_ATOMIC_CS,
 192         [IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
 193 };
 194
 195 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
 196 {
 197         return qp->qpn >= dev->qp_table.sqp_start &&
 198                 qp->qpn <= dev->qp_table.sqp_start + 3;
 199 }
 200
 201 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
 202 {
 203         return qp->qpn >= dev->qp_table.sqp_start &&
 204                 qp->qpn <= dev->qp_table.sqp_start + 1;
 205 }
 206
 207 static void *get_recv_wqe(struct mthca_qp *qp, int n)
 208 {
 209         if (qp->is_direct)
 210                 return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
 211         else
 212                 return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
 213                         ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
 214 }
 215
 216 static void *get_send_wqe(struct mthca_qp *qp, int n)
 217 {
 218         if (qp->is_direct)
 219                 return qp->queue.direct.buf + qp->send_wqe_offset +
 220                         (n << qp->sq.wqe_shift);
 221         else
 222                 return qp->queue.page_list[(qp->send_wqe_offset +
 223                                             (n << qp->sq.wqe_shift)) >>
 224                                            PAGE_SHIFT].buf +
 225                         ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
 226                          (PAGE_SIZE - 1));
 227 }
 228
 229 static void mthca_wq_reset(struct mthca_wq *wq)
 230 {
 231         wq->next_ind  = 0;
 232         wq->last_comp = wq->max - 1;
 233         wq->head      = 0;
 234         wq->tail      = 0;
 235 }
 236
 237 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
 238                     enum ib_event_type event_type)
 239 {
 240         struct mthca_qp *qp;
 241         struct ib_event event;
 242
 243         spin_lock(&dev->qp_table.lock);
 244         qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
 245         if (qp)
 246                 ++qp->refcount;
 247         spin_unlock(&dev->qp_table.lock);
 248
 249         if (!qp) {
 250                 mthca_warn(dev, "Async event for bogus QP %08x\n", qpn);
 251                 return;
 252         }
 253
 254         if (event_type == IB_EVENT_PATH_MIG)
 255                 qp->port = qp->alt_port;
 256
 257         event.device      = &dev->ib_dev;
 258         event.event       = event_type;
 259         event.element.qp  = &qp->ibqp;
 260         if (qp->ibqp.event_handler)
 261                 qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
 262
 263         spin_lock(&dev->qp_table.lock);
 264         if (!--qp->refcount)
 265                 wake_up(&qp->wait);
 266         spin_unlock(&dev->qp_table.lock);
 267 }
 268
 269 static int to_mthca_state(enum ib_qp_state ib_state)
 270 {
 271         switch (ib_state) {
 272         case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
 273         case IB_QPS_INIT:  return MTHCA_QP_STATE_INIT;
 274         case IB_QPS_RTR:   return MTHCA_QP_STATE_RTR;
 275         case IB_QPS_RTS:   return MTHCA_QP_STATE_RTS;
 276         case IB_QPS_SQD:   return MTHCA_QP_STATE_SQD;
 277         case IB_QPS_SQE:   return MTHCA_QP_STATE_SQE;
 278         case IB_QPS_ERR:   return MTHCA_QP_STATE_ERR;
 279         default:                return -1;
 280         }
 281 }
 282
 283 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
 284
 285 static int to_mthca_st(int transport)
 286 {
 287         switch (transport) {
 288         case RC:  return MTHCA_QP_ST_RC;
 289         case UC:  return MTHCA_QP_ST_UC;
 290         case UD:  return MTHCA_QP_ST_UD;
 291         case RD:  return MTHCA_QP_ST_RD;
 292         case MLX: return MTHCA_QP_ST_MLX;
 293         default:  return -1;
 294         }
 295 }
 296
 297 static void store_attrs(struct mthca_sqp *sqp, const struct ib_qp_attr *attr,
 298                         int attr_mask)
 299 {
 300         if (attr_mask & IB_QP_PKEY_INDEX)
 301                 sqp->pkey_index = attr->pkey_index;
 302         if (attr_mask & IB_QP_QKEY)
 303                 sqp->qkey = attr->qkey;
 304         if (attr_mask & IB_QP_SQ_PSN)
 305                 sqp->send_psn = attr->sq_psn;
 306 }
 307
 308 static void init_port(struct mthca_dev *dev, int port)
 309 {
 310         int err;
 311         u8 status;
 312         struct mthca_init_ib_param param;
 313
 314         memset(&param, 0, sizeof param);
 315
 316         param.port_width = dev->limits.port_width_cap;
 317         param.vl_cap     = dev->limits.vl_cap;
 318         param.mtu_cap    = dev->limits.mtu_cap;
 319         param.gid_cap    = dev->limits.gid_table_len;
 320         param.pkey_cap   = dev->limits.pkey_table_len;
 321
 322         err = mthca_INIT_IB(dev, &param, port, &status);
 323         if (err)
 324                 mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
 325         if (status)
 326                 mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
 327 }
 328
 329 static __be32 get_hw_access_flags(struct mthca_qp *qp, const struct ib_qp_attr *attr,
 330                                   int attr_mask)
 331 {
 332         u8 dest_rd_atomic;
 333         u32 access_flags;
 334         u32 hw_access_flags = 0;
 335
 336         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 337                 dest_rd_atomic = attr->max_dest_rd_atomic;
 338         else
 339                 dest_rd_atomic = qp->resp_depth;
 340
 341         if (attr_mask & IB_QP_ACCESS_FLAGS)
 342                 access_flags = attr->qp_access_flags;
 343         else
 344                 access_flags = qp->atomic_rd_en;
 345
 346         if (!dest_rd_atomic)
 347                 access_flags &= IB_ACCESS_REMOTE_WRITE;
 348
 349         if (access_flags & IB_ACCESS_REMOTE_READ)
 350                 hw_access_flags |= MTHCA_QP_BIT_RRE;
 351         if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
 352                 hw_access_flags |= MTHCA_QP_BIT_RAE;
 353         if (access_flags & IB_ACCESS_REMOTE_WRITE)
 354                 hw_access_flags |= MTHCA_QP_BIT_RWE;
 355
 356         return cpu_to_be32(hw_access_flags);
 357 }
 358
 359 static inline enum ib_qp_state to_ib_qp_state(int mthca_state)
 360 {
 361         switch (mthca_state) {
 362         case MTHCA_QP_STATE_RST:      return IB_QPS_RESET;
 363         case MTHCA_QP_STATE_INIT:     return IB_QPS_INIT;
 364         case MTHCA_QP_STATE_RTR:      return IB_QPS_RTR;
 365         case MTHCA_QP_STATE_RTS:      return IB_QPS_RTS;
 366         case MTHCA_QP_STATE_DRAINING:
 367         case MTHCA_QP_STATE_SQD:      return IB_QPS_SQD;
 368         case MTHCA_QP_STATE_SQE:      return IB_QPS_SQE;
 369         case MTHCA_QP_STATE_ERR:      return IB_QPS_ERR;
 370         default:                      return -1;
 371         }
 372 }
 373
 374 static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state)
 375 {
 376         switch (mthca_mig_state) {
 377         case 0:  return IB_MIG_ARMED;
 378         case 1:  return IB_MIG_REARM;
 379         case 3:  return IB_MIG_MIGRATED;
 380         default: return -1;
 381         }
 382 }
 383
 384 static int to_ib_qp_access_flags(int mthca_flags)
 385 {
 386         int ib_flags = 0;
 387
 388         if (mthca_flags & MTHCA_QP_BIT_RRE)
 389                 ib_flags |= IB_ACCESS_REMOTE_READ;
 390         if (mthca_flags & MTHCA_QP_BIT_RWE)
 391                 ib_flags |= IB_ACCESS_REMOTE_WRITE;
 392         if (mthca_flags & MTHCA_QP_BIT_RAE)
 393                 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
 394
 395         return ib_flags;
 396 }
 397
 398 static void to_ib_ah_attr(struct mthca_dev *dev, struct ib_ah_attr *ib_ah_attr,
 399                                 struct mthca_qp_path *path)
 400 {
 401         memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
 402         ib_ah_attr->port_num      = (be32_to_cpu(path->port_pkey) >> 24) & 0x3;
 403
 404         if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->limits.num_ports)
 405                 return;
 406
 407         ib_ah_attr->dlid          = be16_to_cpu(path->rlid);
 408         ib_ah_attr->sl            = be32_to_cpu(path->sl_tclass_flowlabel) >> 28;
 409         ib_ah_attr->src_path_bits = path->g_mylmc & 0x7f;
 410         ib_ah_attr->static_rate   = mthca_rate_to_ib(dev,
 411                                                      path->static_rate & 0xf,
 412                                                      ib_ah_attr->port_num);
 413         ib_ah_attr->ah_flags      = (path->g_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
 414         if (ib_ah_attr->ah_flags) {
 415                 ib_ah_attr->grh.sgid_index = path->mgid_index & (dev->limits.gid_table_len - 1);
 416                 ib_ah_attr->grh.hop_limit  = path->hop_limit;
 417                 ib_ah_attr->grh.traffic_class =
 418                         (be32_to_cpu(path->sl_tclass_flowlabel) >> 20) & 0xff;
 419                 ib_ah_attr->grh.flow_label =
 420                         be32_to_cpu(path->sl_tclass_flowlabel) & 0xfffff;
 421                 memcpy(ib_ah_attr->grh.dgid.raw,
 422                         path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
 423         }
 424 }
 425
 426 int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
 427                    struct ib_qp_init_attr *qp_init_attr)
 428 {
 429         struct mthca_dev *dev = to_mdev(ibqp->device);
 430         struct mthca_qp *qp = to_mqp(ibqp);
 431         int err = 0;
 432         struct mthca_mailbox *mailbox = NULL;
 433         struct mthca_qp_param *qp_param;
 434         struct mthca_qp_context *context;
 435         int mthca_state;
 436         u8 status;
 437
 438         mutex_lock(&qp->mutex);
 439
 440         if (qp->state == IB_QPS_RESET) {
 441                 qp_attr->qp_state = IB_QPS_RESET;
 442                 goto done;
 443         }
 444
 445         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 446         if (IS_ERR(mailbox)) {
 447                 err = PTR_ERR(mailbox);
 448                 goto out;
 449         }
 450
 451         err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox, &status);
 452         if (err)
 453                 goto out_mailbox;
 454         if (status) {
 455                 mthca_warn(dev, "QUERY_QP returned status %02x\n", status);
 456                 err = -EINVAL;
 457                 goto out_mailbox;
 458         }
 459
 460         qp_param    = mailbox->buf;
 461         context     = &qp_param->context;
 462         mthca_state = be32_to_cpu(context->flags) >> 28;
 463
 464         qp->state                    = to_ib_qp_state(mthca_state);
 465         qp_attr->qp_state            = qp->state;
 466         qp_attr->path_mtu            = context->mtu_msgmax >> 5;
 467         qp_attr->path_mig_state      =
 468                 to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
 469         qp_attr->qkey                = be32_to_cpu(context->qkey);
 470         qp_attr->rq_psn              = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
 471         qp_attr->sq_psn              = be32_to_cpu(context->next_send_psn) & 0xffffff;
 472         qp_attr->dest_qp_num         = be32_to_cpu(context->remote_qpn) & 0xffffff;
 473         qp_attr->qp_access_flags     =
 474                 to_ib_qp_access_flags(be32_to_cpu(context->params2));
 475
 476         if (qp->transport == RC || qp->transport == UC) {
 477                 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
 478                 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
 479                 qp_attr->alt_pkey_index =
 480                         be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
 481                 qp_attr->alt_port_num   = qp_attr->alt_ah_attr.port_num;
 482         }
 483
 484         qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
 485         qp_attr->port_num   =
 486                 (be32_to_cpu(context->pri_path.port_pkey) >> 24) & 0x3;
 487
 488         /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
 489         qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
 490
 491         qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
 492
 493         qp_attr->max_dest_rd_atomic =
 494                 1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
 495         qp_attr->min_rnr_timer      =
 496                 (be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
 497         qp_attr->timeout            = context->pri_path.ackto >> 3;
 498         qp_attr->retry_cnt          = (be32_to_cpu(context->params1) >> 16) & 0x7;
 499         qp_attr->rnr_retry          = context->pri_path.rnr_retry >> 5;
 500         qp_attr->alt_timeout        = context->alt_path.ackto >> 3;
 501
 502 done:
 503         qp_attr->cur_qp_state        = qp_attr->qp_state;
 504         qp_attr->cap.max_send_wr     = qp->sq.max;
 505         qp_attr->cap.max_recv_wr     = qp->rq.max;
 506         qp_attr->cap.max_send_sge    = qp->sq.max_gs;
 507         qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
 508         qp_attr->cap.max_inline_data = qp->max_inline_data;
 509
 510         qp_init_attr->cap            = qp_attr->cap;
 511
 512 out_mailbox:
 513         mthca_free_mailbox(dev, mailbox);
 514
 515 out:
 516         mutex_unlock(&qp->mutex);
 517         return err;
 518 }
 519
 520 static int mthca_path_set(struct mthca_dev *dev, const struct ib_ah_attr *ah,
 521                           struct mthca_qp_path *path, u8 port)
 522 {
 523         path->g_mylmc     = ah->src_path_bits & 0x7f;
 524         path->rlid        = cpu_to_be16(ah->dlid);
 525         path->static_rate = mthca_get_rate(dev, ah->static_rate, port);
 526
 527         if (ah->ah_flags & IB_AH_GRH) {
 528                 if (ah->grh.sgid_index >= dev->limits.gid_table_len) {
 529                         mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
 530                                   ah->grh.sgid_index, dev->limits.gid_table_len-1);
 531                         return -1;
 532                 }
 533
 534                 path->g_mylmc   |= 1 << 7;
 535                 path->mgid_index = ah->grh.sgid_index;
 536                 path->hop_limit  = ah->grh.hop_limit;
 537                 path->sl_tclass_flowlabel =
 538                         cpu_to_be32((ah->sl << 28)                |
 539                                     (ah->grh.traffic_class << 20) |
 540                                     (ah->grh.flow_label));
 541                 memcpy(path->rgid, ah->grh.dgid.raw, 16);
 542         } else
 543                 path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28);
 544
 545         return 0;
 546 }
 547
 548 static int __mthca_modify_qp(struct ib_qp *ibqp,
 549                              const struct ib_qp_attr *attr, int attr_mask,
 550                              enum ib_qp_state cur_state, enum ib_qp_state new_state)
 551 {
 552         struct mthca_dev *dev = to_mdev(ibqp->device);
 553         struct mthca_qp *qp = to_mqp(ibqp);
 554         struct mthca_mailbox *mailbox;
 555         struct mthca_qp_param *qp_param;
 556         struct mthca_qp_context *qp_context;
 557         u32 sqd_event = 0;
 558         u8 status;
 559         int err = -EINVAL;
 560
 561         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 562         if (IS_ERR(mailbox)) {
 563                 err = PTR_ERR(mailbox);
 564                 goto out;
 565         }
 566         qp_param = mailbox->buf;
 567         qp_context = &qp_param->context;
 568         memset(qp_param, 0, sizeof *qp_param);
 569
 570         qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
 571                                              (to_mthca_st(qp->transport) << 16));
 572         qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
 573         if (!(attr_mask & IB_QP_PATH_MIG_STATE))
 574                 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
 575         else {
 576                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
 577                 switch (attr->path_mig_state) {
 578                 case IB_MIG_MIGRATED:
 579                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
 580                         break;
 581                 case IB_MIG_REARM:
 582                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
 583                         break;
 584                 case IB_MIG_ARMED:
 585                         qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
 586                         break;
 587                 }
 588         }
 589
 590         /* leave tavor_sched_queue as 0 */
 591
 592         if (qp->transport == MLX || qp->transport == UD)
 593                 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
 594         else if (attr_mask & IB_QP_PATH_MTU) {
 595                 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
 596                         mthca_dbg(dev, "path MTU (%u) is invalid\n",
 597                                   attr->path_mtu);
 598                         goto out_mailbox;
 599                 }
 600                 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
 601         }
 602
 603         if (mthca_is_memfree(dev)) {
 604                 if (qp->rq.max)
 605                         qp_context->rq_size_stride = ilog2(qp->rq.max) << 3;
 606                 qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
 607
 608                 if (qp->sq.max)
 609                         qp_context->sq_size_stride = ilog2(qp->sq.max) << 3;
 610                 qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
 611         }
 612
 613         /* leave arbel_sched_queue as 0 */
 614
 615         if (qp->ibqp.uobject)
 616                 qp_context->usr_page =
 617                         cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
 618         else
 619                 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
 620         qp_context->local_qpn  = cpu_to_be32(qp->qpn);
 621         if (attr_mask & IB_QP_DEST_QPN) {
 622                 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
 623         }
 624
 625         if (qp->transport == MLX)
 626                 qp_context->pri_path.port_pkey |=
 627                         cpu_to_be32(qp->port << 24);
 628         else {
 629                 if (attr_mask & IB_QP_PORT) {
 630                         qp_context->pri_path.port_pkey |=
 631                                 cpu_to_be32(attr->port_num << 24);
 632                         qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
 633                 }
 634         }
 635
 636         if (attr_mask & IB_QP_PKEY_INDEX) {
 637                 qp_context->pri_path.port_pkey |=
 638                         cpu_to_be32(attr->pkey_index);
 639                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
 640         }
 641
 642         if (attr_mask & IB_QP_RNR_RETRY) {
 643                 qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
 644                         attr->rnr_retry << 5;
 645                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
 646                                                         MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
 647         }
 648
 649         if (attr_mask & IB_QP_AV) {
 650                 if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
 651                                    attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
 652                         goto out_mailbox;
 653
 654                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
 655         }
 656
 657         if (ibqp->qp_type == IB_QPT_RC &&
 658             cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
 659                 u8 sched_queue = ibqp->uobject ? 0x2 : 0x1;
 660
 661                 if (mthca_is_memfree(dev))
 662                         qp_context->rlkey_arbel_sched_queue |= sched_queue;
 663                 else
 664                         qp_context->tavor_sched_queue |= cpu_to_be32(sched_queue);
 665
 666                 qp_param->opt_param_mask |=
 667                         cpu_to_be32(MTHCA_QP_OPTPAR_SCHED_QUEUE);
 668         }
 669
 670         if (attr_mask & IB_QP_TIMEOUT) {
 671                 qp_context->pri_path.ackto = attr->timeout << 3;
 672                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
 673         }
 674
 675         if (attr_mask & IB_QP_ALT_PATH) {
 676                 if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
 677                         mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
 678                                   attr->alt_pkey_index, dev->limits.pkey_table_len-1);
 679                         goto out_mailbox;
 680                 }
 681
 682                 if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
 683                         mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
 684                                 attr->alt_port_num);
 685                         goto out_mailbox;
 686                 }
 687
 688                 if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
 689                                    attr->alt_ah_attr.port_num))
 690                         goto out_mailbox;
 691
 692                 qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
 693                                                               attr->alt_port_num << 24);
 694                 qp_context->alt_path.ackto = attr->alt_timeout << 3;
 695                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
 696         }
 697
 698         /* leave rdd as 0 */
 699         qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
 700         /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
 701         qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
 702         qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
 703                                              (MTHCA_FLIGHT_LIMIT << 24) |
 704                                              MTHCA_QP_BIT_SWE);
 705         if (qp->sq_policy == IB_SIGNAL_ALL_WR)
 706                 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
 707         if (attr_mask & IB_QP_RETRY_CNT) {
 708                 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
 709                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
 710         }
 711
 712         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
 713                 if (attr->max_rd_atomic) {
 714                         qp_context->params1 |=
 715                                 cpu_to_be32(MTHCA_QP_BIT_SRE |
 716                                             MTHCA_QP_BIT_SAE);
 717                         qp_context->params1 |=
 718                                 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
 719                 }
 720                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
 721         }
 722
 723         if (attr_mask & IB_QP_SQ_PSN)
 724                 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
 725         qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
 726
 727         if (mthca_is_memfree(dev)) {
 728                 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
 729                 qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
 730         }
 731
 732         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
 733                 if (attr->max_dest_rd_atomic)
 734                         qp_context->params2 |=
 735                                 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
 736
 737                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
 738         }
 739
 740         if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
 741                 qp_context->params2      |= get_hw_access_flags(qp, attr, attr_mask);
 742                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
 743                                                         MTHCA_QP_OPTPAR_RRE |
 744                                                         MTHCA_QP_OPTPAR_RAE);
 745         }
 746
 747         qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
 748
 749         if (ibqp->srq)
 750                 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
 751
 752         if (attr_mask & IB_QP_MIN_RNR_TIMER) {
 753                 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
 754                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
 755         }
 756         if (attr_mask & IB_QP_RQ_PSN)
 757                 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
 758
 759         qp_context->ra_buff_indx =
 760                 cpu_to_be32(dev->qp_table.rdb_base +
 761                             ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
 762                              dev->qp_table.rdb_shift));
 763
 764         qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
 765
 766         if (mthca_is_memfree(dev))
 767                 qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
 768
 769         if (attr_mask & IB_QP_QKEY) {
 770                 qp_context->qkey = cpu_to_be32(attr->qkey);
 771                 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
 772         }
 773
 774         if (ibqp->srq)
 775                 qp_context->srqn = cpu_to_be32(1 << 24 |
 776                                                to_msrq(ibqp->srq)->srqn);
 777
 778         if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD  &&
 779             attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY               &&
 780             attr->en_sqd_async_notify)
 781                 sqd_event = 1 << 31;
 782
 783         err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
 784                               mailbox, sqd_event, &status);
 785         if (err)
 786                 goto out_mailbox;
 787         if (status) {
 788                 mthca_warn(dev, "modify QP %d->%d returned status %02x.\n",
 789                            cur_state, new_state, status);
 790                 err = -EINVAL;
 791                 goto out_mailbox;
 792         }
 793
 794         qp->state = new_state;
 795         if (attr_mask & IB_QP_ACCESS_FLAGS)
 796                 qp->atomic_rd_en = attr->qp_access_flags;
 797         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 798                 qp->resp_depth = attr->max_dest_rd_atomic;
 799         if (attr_mask & IB_QP_PORT)
 800                 qp->port = attr->port_num;
 801         if (attr_mask & IB_QP_ALT_PATH)
 802                 qp->alt_port = attr->alt_port_num;
 803
 804         if (is_sqp(dev, qp))
 805                 store_attrs(to_msqp(qp), attr, attr_mask);
 806
 807         /*
 808          * If we moved QP0 to RTR, bring the IB link up; if we moved
 809          * QP0 to RESET or ERROR, bring the link back down.
 810          */
 811         if (is_qp0(dev, qp)) {
 812                 if (cur_state != IB_QPS_RTR &&
 813                     new_state == IB_QPS_RTR)
 814                         init_port(dev, qp->port);
 815
 816                 if (cur_state != IB_QPS_RESET &&
 817                     cur_state != IB_QPS_ERR &&
 818                     (new_state == IB_QPS_RESET ||
 819                      new_state == IB_QPS_ERR))
 820                         mthca_CLOSE_IB(dev, qp->port, &status);
 821         }
 822
 823         /*
 824          * If we moved a kernel QP to RESET, clean up all old CQ
 825          * entries and reinitialize the QP.
 826          */
 827         if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
 828                 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
 829                                qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
 830                 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
 831                         mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn, NULL);
 832
 833                 mthca_wq_reset(&qp->sq);
 834                 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
 835
 836                 mthca_wq_reset(&qp->rq);
 837                 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
 838
 839                 if (mthca_is_memfree(dev)) {
 840                         *qp->sq.db = 0;
 841                         *qp->rq.db = 0;
 842                 }
 843         }
 844
 845 out_mailbox:
 846         mthca_free_mailbox(dev, mailbox);
 847 out:
 848         return err;
 849 }
 850
 851 static const struct ib_qp_attr dummy_init_attr = { .port_num = 1 };
 852 static const int dummy_init_attr_mask[] = {
 853         [IB_QPT_UD]  = (IB_QP_PKEY_INDEX                |
 854                         IB_QP_PORT                      |
 855                         IB_QP_QKEY),
 856         [IB_QPT_UC]  = (IB_QP_PKEY_INDEX                |
 857                         IB_QP_PORT                      |
 858                         IB_QP_ACCESS_FLAGS),
 859         [IB_QPT_RC]  = (IB_QP_PKEY_INDEX                |
 860                         IB_QP_PORT                      |
 861                         IB_QP_ACCESS_FLAGS),
 862         [IB_QPT_SMI] = (IB_QP_PKEY_INDEX                |
 863                         IB_QP_QKEY),
 864         [IB_QPT_GSI] = (IB_QP_PKEY_INDEX                |
 865                         IB_QP_QKEY),
 866 };
 867
 868 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
 869                     struct ib_udata *udata)
 870 {
 871         struct mthca_dev *dev = to_mdev(ibqp->device);
 872         struct mthca_qp *qp = to_mqp(ibqp);
 873         enum ib_qp_state cur_state, new_state;
 874         int err = -EINVAL;
 875
 876         mutex_lock(&qp->mutex);
 877         if (attr_mask & IB_QP_CUR_STATE) {
 878                 cur_state = attr->cur_qp_state;
 879         } else {
 880                 spin_lock_irq(&qp->sq.lock);
 881                 spin_lock(&qp->rq.lock);
 882                 cur_state = qp->state;
 883                 spin_unlock(&qp->rq.lock);
 884                 spin_unlock_irq(&qp->sq.lock);
 885         }
 886
 887         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
 888
 889         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
 890                 mthca_dbg(dev, "Bad QP transition (transport %d) "
 891                           "%d->%d with attr 0x%08x\n",
 892                           qp->transport, cur_state, new_state,
 893                           attr_mask);
 894                 goto out;
 895         }
 896
 897         if ((attr_mask & IB_QP_PKEY_INDEX) &&
 898              attr->pkey_index >= dev->limits.pkey_table_len) {
 899                 mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
 900                           attr->pkey_index, dev->limits.pkey_table_len-1);
 901                 goto out;
 902         }
 903
 904         if ((attr_mask & IB_QP_PORT) &&
 905             (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
 906                 mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
 907                 goto out;
 908         }
 909
 910         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
 911             attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
 912                 mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
 913                           attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
 914                 goto out;
 915         }
 916
 917         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
 918             attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
 919                 mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
 920                           attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
 921                 goto out;
 922         }
 923
 924         if (cur_state == new_state && cur_state == IB_QPS_RESET) {
 925                 err = 0;
 926                 goto out;
 927         }
 928
 929         if (cur_state == IB_QPS_RESET && new_state == IB_QPS_ERR) {
 930                 err = __mthca_modify_qp(ibqp, &dummy_init_attr,
 931                                         dummy_init_attr_mask[ibqp->qp_type],
 932                                         IB_QPS_RESET, IB_QPS_INIT);
 933                 if (err)
 934                         goto out;
 935                 cur_state = IB_QPS_INIT;
 936         }
 937
 938         err = __mthca_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
 939
 940 out:
 941         mutex_unlock(&qp->mutex);
 942         return err;
 943 }
 944
 945 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
 946 {
 947         /*
 948          * Calculate the maximum size of WQE s/g segments, excluding
 949          * the next segment and other non-data segments.
 950          */
 951         int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
 952
 953         switch (qp->transport) {
 954         case MLX:
 955                 max_data_size -= 2 * sizeof (struct mthca_data_seg);
 956                 break;
 957
 958         case UD:
 959                 if (mthca_is_memfree(dev))
 960                         max_data_size -= sizeof (struct mthca_arbel_ud_seg);
 961                 else
 962                         max_data_size -= sizeof (struct mthca_tavor_ud_seg);
 963                 break;
 964
 965         default:
 966                 max_data_size -= sizeof (struct mthca_raddr_seg);
 967                 break;
 968         }
 969
 970         return max_data_size;
 971 }
 972
 973 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
 974 {
 975         /* We don't support inline data for kernel QPs (yet). */
 976         return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
 977 }
 978
 979 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
 980                                  struct mthca_pd *pd,
 981                                  struct mthca_qp *qp)
 982 {
 983         int max_data_size = mthca_max_data_size(dev, qp,
 984                                                 min(dev->limits.max_desc_sz,
 985                                                     1 << qp->sq.wqe_shift));
 986
 987         qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
 988
 989         qp->sq.max_gs = min_t(int, dev->limits.max_sg,
 990                               max_data_size / sizeof (struct mthca_data_seg));
 991         qp->rq.max_gs = min_t(int, dev->limits.max_sg,
 992                                (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
 993                                 sizeof (struct mthca_next_seg)) /
 994                                sizeof (struct mthca_data_seg));
 995 }
 996
 997 /*
 998  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
 999  * rq.max_gs and sq.max_gs must all be assigned.
1000  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
1001  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
1002  * queue)
1003  */
1004 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
1005                                struct mthca_pd *pd,
1006                                struct mthca_qp *qp)
1007 {
1008         int size;
1009         int err = -ENOMEM;
1010
1011         size = sizeof (struct mthca_next_seg) +
1012                 qp->rq.max_gs * sizeof (struct mthca_data_seg);
1013
1014         if (size > dev->limits.max_desc_sz)
1015                 return -EINVAL;
1016
1017         for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
1018              qp->rq.wqe_shift++)
1019                 ; /* nothing */
1020
1021         size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
1022         switch (qp->transport) {
1023         case MLX:
1024                 size += 2 * sizeof (struct mthca_data_seg);
1025                 break;
1026
1027         case UD:
1028                 size += mthca_is_memfree(dev) ?
1029                         sizeof (struct mthca_arbel_ud_seg) :
1030                         sizeof (struct mthca_tavor_ud_seg);
1031                 break;
1032
1033         case UC:
1034                 size += sizeof (struct mthca_raddr_seg);
1035                 break;
1036
1037         case RC:
1038                 size += sizeof (struct mthca_raddr_seg);
1039                 /*
1040                  * An atomic op will require an atomic segment, a
1041                  * remote address segment and one scatter entry.
1042                  */
1043                 size = max_t(int, size,
1044                              sizeof (struct mthca_atomic_seg) +
1045                              sizeof (struct mthca_raddr_seg) +
1046                              sizeof (struct mthca_data_seg));
1047                 break;
1048
1049         default:
1050                 break;
1051         }
1052
1053         /* Make sure that we have enough space for a bind request */
1054         size = max_t(int, size, sizeof (struct mthca_bind_seg));
1055
1056         size += sizeof (struct mthca_next_seg);
1057
1058         if (size > dev->limits.max_desc_sz)
1059                 return -EINVAL;
1060
1061         for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
1062              qp->sq.wqe_shift++)
1063                 ; /* nothing */
1064
1065         qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
1066                                     1 << qp->sq.wqe_shift);
1067
1068         /*
1069          * If this is a userspace QP, we don't actually have to
1070          * allocate anything.  All we need is to calculate the WQE
1071          * sizes and the send_wqe_offset, so we're done now.
1072          */
1073         if (pd->ibpd.uobject)
1074                 return 0;
1075
1076         size = PAGE_ALIGN(qp->send_wqe_offset +
1077                           (qp->sq.max << qp->sq.wqe_shift));
1078
1079         qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1080                            GFP_KERNEL);
1081         if (!qp->wrid)
1082                 goto err_out;
1083
1084         err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1085                               &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1086         if (err)
1087                 goto err_out;
1088
1089         return 0;
1090
1091 err_out:
1092         kfree(qp->wrid);
1093         return err;
1094 }
1095
1096 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1097                                struct mthca_qp *qp)
1098 {
1099         mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1100                                        (qp->sq.max << qp->sq.wqe_shift)),
1101                        &qp->queue, qp->is_direct, &qp->mr);
1102         kfree(qp->wrid);
1103 }
1104
1105 static int mthca_map_memfree(struct mthca_dev *dev,
1106                              struct mthca_qp *qp)
1107 {
1108         int ret;
1109
1110         if (mthca_is_memfree(dev)) {
1111                 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1112                 if (ret)
1113                         return ret;
1114
1115                 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1116                 if (ret)
1117                         goto err_qpc;
1118
1119                 ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1120                                       qp->qpn << dev->qp_table.rdb_shift);
1121                 if (ret)
1122                         goto err_eqpc;
1123
1124         }
1125
1126         return 0;
1127
1128 err_eqpc:
1129         mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1130
1131 err_qpc:
1132         mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1133
1134         return ret;
1135 }
1136
1137 static void mthca_unmap_memfree(struct mthca_dev *dev,
1138                                 struct mthca_qp *qp)
1139 {
1140         mthca_table_put(dev, dev->qp_table.rdb_table,
1141                         qp->qpn << dev->qp_table.rdb_shift);
1142         mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1143         mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1144 }
1145
1146 static int mthca_alloc_memfree(struct mthca_dev *dev,
1147                                struct mthca_qp *qp)
1148 {
1149         if (mthca_is_memfree(dev)) {
1150                 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1151                                                  qp->qpn, &qp->rq.db);
1152                 if (qp->rq.db_index < 0)
1153                         return -ENOMEM;
1154
1155                 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1156                                                  qp->qpn, &qp->sq.db);
1157                 if (qp->sq.db_index < 0) {
1158                         mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1159                         return -ENOMEM;
1160                 }
1161         }
1162
1163         return 0;
1164 }
1165
1166 static void mthca_free_memfree(struct mthca_dev *dev,
1167                                struct mthca_qp *qp)
1168 {
1169         if (mthca_is_memfree(dev)) {
1170                 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1171                 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1172         }
1173 }
1174
1175 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1176                                  struct mthca_pd *pd,
1177                                  struct mthca_cq *send_cq,
1178                                  struct mthca_cq *recv_cq,
1179                                  enum ib_sig_type send_policy,
1180                                  struct mthca_qp *qp)
1181 {
1182         int ret;
1183         int i;
1184         struct mthca_next_seg *next;
1185
1186         qp->refcount = 1;
1187         init_waitqueue_head(&qp->wait);
1188         mutex_init(&qp->mutex);
1189         qp->state        = IB_QPS_RESET;
1190         qp->atomic_rd_en = 0;
1191         qp->resp_depth   = 0;
1192         qp->sq_policy    = send_policy;
1193         mthca_wq_reset(&qp->sq);
1194         mthca_wq_reset(&qp->rq);
1195
1196         spin_lock_init(&qp->sq.lock);
1197         spin_lock_init(&qp->rq.lock);
1198
1199         ret = mthca_map_memfree(dev, qp);
1200         if (ret)
1201                 return ret;
1202
1203         ret = mthca_alloc_wqe_buf(dev, pd, qp);
1204         if (ret) {
1205                 mthca_unmap_memfree(dev, qp);
1206                 return ret;
1207         }
1208
1209         mthca_adjust_qp_caps(dev, pd, qp);
1210
1211         /*
1212          * If this is a userspace QP, we're done now.  The doorbells
1213          * will be allocated and buffers will be initialized in
1214          * userspace.
1215          */
1216         if (pd->ibpd.uobject)
1217                 return 0;
1218
1219         ret = mthca_alloc_memfree(dev, qp);
1220         if (ret) {
1221                 mthca_free_wqe_buf(dev, qp);
1222                 mthca_unmap_memfree(dev, qp);
1223                 return ret;
1224         }
1225
1226         if (mthca_is_memfree(dev)) {
1227                 struct mthca_data_seg *scatter;
1228                 int size = (sizeof (struct mthca_next_seg) +
1229                             qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1230
1231                 for (i = 0; i < qp->rq.max; ++i) {
1232                         next = get_recv_wqe(qp, i);
1233                         next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1234                                                    qp->rq.wqe_shift);
1235                         next->ee_nds = cpu_to_be32(size);
1236
1237                         for (scatter = (void *) (next + 1);
1238                              (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1239                              ++scatter)
1240                                 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1241                 }
1242
1243                 for (i = 0; i < qp->sq.max; ++i) {
1244                         next = get_send_wqe(qp, i);
1245                         next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1246                                                     qp->sq.wqe_shift) +
1247                                                    qp->send_wqe_offset);
1248                 }
1249         } else {
1250                 for (i = 0; i < qp->rq.max; ++i) {
1251                         next = get_recv_wqe(qp, i);
1252                         next->nda_op = htonl((((i + 1) % qp->rq.max) <<
1253                                               qp->rq.wqe_shift) | 1);
1254                 }
1255
1256         }
1257
1258         qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1259         qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1260
1261         return 0;
1262 }
1263
1264 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1265                              struct mthca_pd *pd, struct mthca_qp *qp)
1266 {
1267         int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1268
1269         /* Sanity check QP size before proceeding */
1270         if (cap->max_send_wr     > dev->limits.max_wqes ||
1271             cap->max_recv_wr     > dev->limits.max_wqes ||
1272             cap->max_send_sge    > dev->limits.max_sg   ||
1273             cap->max_recv_sge    > dev->limits.max_sg   ||
1274             cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1275                 return -EINVAL;
1276
1277         /*
1278          * For MLX transport we need 2 extra S/G entries:
1279          * one for the header and one for the checksum at the end
1280          */
1281         if (qp->transport == MLX && cap->max_recv_sge + 2 > dev->limits.max_sg)
1282                 return -EINVAL;
1283
1284         if (mthca_is_memfree(dev)) {
1285                 qp->rq.max = cap->max_recv_wr ?
1286                         roundup_pow_of_two(cap->max_recv_wr) : 0;
1287                 qp->sq.max = cap->max_send_wr ?
1288                         roundup_pow_of_two(cap->max_send_wr) : 0;
1289         } else {
1290                 qp->rq.max = cap->max_recv_wr;
1291                 qp->sq.max = cap->max_send_wr;
1292         }
1293
1294         qp->rq.max_gs = cap->max_recv_sge;
1295         qp->sq.max_gs = max_t(int, cap->max_send_sge,
1296                               ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1297                                     MTHCA_INLINE_CHUNK_SIZE) /
1298                               sizeof (struct mthca_data_seg));
1299
1300         return 0;
1301 }
1302
1303 int mthca_alloc_qp(struct mthca_dev *dev,
1304                    struct mthca_pd *pd,
1305                    struct mthca_cq *send_cq,
1306                    struct mthca_cq *recv_cq,
1307                    enum ib_qp_type type,
1308                    enum ib_sig_type send_policy,
1309                    struct ib_qp_cap *cap,
1310                    struct mthca_qp *qp)
1311 {
1312         int err;
1313
1314         switch (type) {
1315         case IB_QPT_RC: qp->transport = RC; break;
1316         case IB_QPT_UC: qp->transport = UC; break;
1317         case IB_QPT_UD: qp->transport = UD; break;
1318         default: return -EINVAL;
1319         }
1320
1321         err = mthca_set_qp_size(dev, cap, pd, qp);
1322         if (err)
1323                 return err;
1324
1325         qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1326         if (qp->qpn == -1)
1327                 return -ENOMEM;
1328
1329         /* initialize port to zero for error-catching. */
1330         qp->port = 0;
1331
1332         err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1333                                     send_policy, qp);
1334         if (err) {
1335                 mthca_free(&dev->qp_table.alloc, qp->qpn);
1336                 return err;
1337         }
1338
1339         spin_lock_irq(&dev->qp_table.lock);
1340         mthca_array_set(&dev->qp_table.qp,
1341                         qp->qpn & (dev->limits.num_qps - 1), qp);
1342         spin_unlock_irq(&dev->qp_table.lock);
1343
1344         return 0;
1345 }
1346
1347 static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1348 {
1349         if (send_cq == recv_cq)
1350                 spin_lock_irq(&send_cq->lock);
1351         else if (send_cq->cqn < recv_cq->cqn) {
1352                 spin_lock_irq(&send_cq->lock);
1353                 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1354         } else {
1355                 spin_lock_irq(&recv_cq->lock);
1356                 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1357         }
1358 }
1359
1360 static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1361 {
1362         if (send_cq == recv_cq)
1363                 spin_unlock_irq(&send_cq->lock);
1364         else if (send_cq->cqn < recv_cq->cqn) {
1365                 spin_unlock(&recv_cq->lock);
1366                 spin_unlock_irq(&send_cq->lock);
1367         } else {
1368                 spin_unlock(&send_cq->lock);
1369                 spin_unlock_irq(&recv_cq->lock);
1370         }
1371 }
1372
1373 int mthca_alloc_sqp(struct mthca_dev *dev,
1374                     struct mthca_pd *pd,
1375                     struct mthca_cq *send_cq,
1376                     struct mthca_cq *recv_cq,
1377                     enum ib_sig_type send_policy,
1378                     struct ib_qp_cap *cap,
1379                     int qpn,
1380                     int port,
1381                     struct mthca_sqp *sqp)
1382 {
1383         u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1384         int err;
1385
1386         sqp->qp.transport = MLX;
1387         err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1388         if (err)
1389                 return err;
1390
1391         sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1392         sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1393                                              &sqp->header_dma, GFP_KERNEL);
1394         if (!sqp->header_buf)
1395                 return -ENOMEM;
1396
1397         spin_lock_irq(&dev->qp_table.lock);
1398         if (mthca_array_get(&dev->qp_table.qp, mqpn))
1399                 err = -EBUSY;
1400         else
1401                 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1402         spin_unlock_irq(&dev->qp_table.lock);
1403
1404         if (err)
1405                 goto err_out;
1406
1407         sqp->qp.port      = port;
1408         sqp->qp.qpn       = mqpn;
1409         sqp->qp.transport = MLX;
1410
1411         err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1412                                     send_policy, &sqp->qp);
1413         if (err)
1414                 goto err_out_free;
1415
1416         atomic_inc(&pd->sqp_count);
1417
1418         return 0;
1419
1420  err_out_free:
1421         /*
1422          * Lock CQs here, so that CQ polling code can do QP lookup
1423          * without taking a lock.
1424          */
1425         mthca_lock_cqs(send_cq, recv_cq);
1426
1427         spin_lock(&dev->qp_table.lock);
1428         mthca_array_clear(&dev->qp_table.qp, mqpn);
1429         spin_unlock(&dev->qp_table.lock);
1430
1431         mthca_unlock_cqs(send_cq, recv_cq);
1432
1433  err_out:
1434         dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1435                           sqp->header_buf, sqp->header_dma);
1436
1437         return err;
1438 }
1439
1440 static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp)
1441 {
1442         int c;
1443
1444         spin_lock_irq(&dev->qp_table.lock);
1445         c = qp->refcount;
1446         spin_unlock_irq(&dev->qp_table.lock);
1447
1448         return c;
1449 }
1450
1451 void mthca_free_qp(struct mthca_dev *dev,
1452                    struct mthca_qp *qp)
1453 {
1454         u8 status;
1455         struct mthca_cq *send_cq;
1456         struct mthca_cq *recv_cq;
1457
1458         send_cq = to_mcq(qp->ibqp.send_cq);
1459         recv_cq = to_mcq(qp->ibqp.recv_cq);
1460
1461         /*
1462          * Lock CQs here, so that CQ polling code can do QP lookup
1463          * without taking a lock.
1464          */
1465         mthca_lock_cqs(send_cq, recv_cq);
1466
1467         spin_lock(&dev->qp_table.lock);
1468         mthca_array_clear(&dev->qp_table.qp,
1469                           qp->qpn & (dev->limits.num_qps - 1));
1470         --qp->refcount;
1471         spin_unlock(&dev->qp_table.lock);
1472
1473         mthca_unlock_cqs(send_cq, recv_cq);
1474
1475         wait_event(qp->wait, !get_qp_refcount(dev, qp));
1476
1477         if (qp->state != IB_QPS_RESET)
1478                 mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1479                                 NULL, 0, &status);
1480
1481         /*
1482          * If this is a userspace QP, the buffers, MR, CQs and so on
1483          * will be cleaned up in userspace, so all we have to do is
1484          * unref the mem-free tables and free the QPN in our table.
1485          */
1486         if (!qp->ibqp.uobject) {
1487                 mthca_cq_clean(dev, recv_cq, qp->qpn,
1488                                qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1489                 if (send_cq != recv_cq)
1490                         mthca_cq_clean(dev, send_cq, qp->qpn, NULL);
1491
1492                 mthca_free_memfree(dev, qp);
1493                 mthca_free_wqe_buf(dev, qp);
1494         }
1495
1496         mthca_unmap_memfree(dev, qp);
1497
1498         if (is_sqp(dev, qp)) {
1499                 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1500                 dma_free_coherent(&dev->pdev->dev,
1501                                   to_msqp(qp)->header_buf_size,
1502                                   to_msqp(qp)->header_buf,
1503                                   to_msqp(qp)->header_dma);
1504         } else
1505                 mthca_free(&dev->qp_table.alloc, qp->qpn);
1506 }
1507
1508 /* Create UD header for an MLX send and build a data segment for it */
1509 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1510                             int ind, struct ib_send_wr *wr,
1511                             struct mthca_mlx_seg *mlx,
1512                             struct mthca_data_seg *data)
1513 {
1514         int header_size;
1515         int err;
1516         u16 pkey;
1517
1518         ib_ud_header_init(256, /* assume a MAD */
1519                           mthca_ah_grh_present(to_mah(wr->wr.ud.ah)),
1520                           &sqp->ud_header);
1521
1522         err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1523         if (err)
1524                 return err;
1525         mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1526         mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1527                                   (sqp->ud_header.lrh.destination_lid ==
1528                                    IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1529                                   (sqp->ud_header.lrh.service_level << 8));
1530         mlx->rlid = sqp->ud_header.lrh.destination_lid;
1531         mlx->vcrc = 0;
1532
1533         switch (wr->opcode) {
1534         case IB_WR_SEND:
1535                 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1536                 sqp->ud_header.immediate_present = 0;
1537                 break;
1538         case IB_WR_SEND_WITH_IMM:
1539                 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1540                 sqp->ud_header.immediate_present = 1;
1541                 sqp->ud_header.immediate_data = wr->ex.imm_data;
1542                 break;
1543         default:
1544                 return -EINVAL;
1545         }
1546
1547         sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1548         if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1549                 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1550         sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1551         if (!sqp->qp.ibqp.qp_num)
1552                 ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1553                                    sqp->pkey_index, &pkey);
1554         else
1555                 ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1556                                    wr->wr.ud.pkey_index, &pkey);
1557         sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1558         sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1559         sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1560         sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1561                                                sqp->qkey : wr->wr.ud.remote_qkey);
1562         sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1563
1564         header_size = ib_ud_header_pack(&sqp->ud_header,
1565                                         sqp->header_buf +
1566                                         ind * MTHCA_UD_HEADER_SIZE);
1567
1568         data->byte_count = cpu_to_be32(header_size);
1569         data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1570         data->addr       = cpu_to_be64(sqp->header_dma +
1571                                        ind * MTHCA_UD_HEADER_SIZE);
1572
1573         return 0;
1574 }
1575
1576 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1577                                     struct ib_cq *ib_cq)
1578 {
1579         unsigned cur;
1580         struct mthca_cq *cq;
1581
1582         cur = wq->head - wq->tail;
1583         if (likely(cur + nreq < wq->max))
1584                 return 0;
1585
1586         cq = to_mcq(ib_cq);
1587         spin_lock(&cq->lock);
1588         cur = wq->head - wq->tail;
1589         spin_unlock(&cq->lock);
1590
1591         return cur + nreq >= wq->max;
1592 }
1593
1594 static __always_inline void set_raddr_seg(struct mthca_raddr_seg *rseg,
1595                                           u64 remote_addr, u32 rkey)
1596 {
1597         rseg->raddr    = cpu_to_be64(remote_addr);
1598         rseg->rkey     = cpu_to_be32(rkey);
1599         rseg->reserved = 0;
1600 }
1601
1602 static __always_inline void set_atomic_seg(struct mthca_atomic_seg *aseg,
1603                                            struct ib_send_wr *wr)
1604 {
1605         if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1606                 aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
1607                 aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
1608         } else {
1609                 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
1610                 aseg->compare  = 0;
1611         }
1612
1613 }
1614
1615 static void set_tavor_ud_seg(struct mthca_tavor_ud_seg *useg,
1616                              struct ib_send_wr *wr)
1617 {
1618         useg->lkey    = cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1619         useg->av_addr = cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1620         useg->dqpn    = cpu_to_be32(wr->wr.ud.remote_qpn);
1621         useg->qkey    = cpu_to_be32(wr->wr.ud.remote_qkey);
1622
1623 }
1624
1625 static void set_arbel_ud_seg(struct mthca_arbel_ud_seg *useg,
1626                              struct ib_send_wr *wr)
1627 {
1628         memcpy(useg->av, to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1629         useg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1630         useg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1631 }
1632
1633 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1634                           struct ib_send_wr **bad_wr)
1635 {
1636         struct mthca_dev *dev = to_mdev(ibqp->device);
1637         struct mthca_qp *qp = to_mqp(ibqp);
1638         void *wqe;
1639         void *prev_wqe;
1640         unsigned long flags;
1641         int err = 0;
1642         int nreq;
1643         int i;
1644         int size;
1645         /*
1646          * f0 and size0 are only used if nreq != 0, and they will
1647          * always be initialized the first time through the main loop
1648          * before nreq is incremented.  So nreq cannot become non-zero
1649          * without initializing f0 and size0, and they are in fact
1650          * never used uninitialized.
1651          */
1652         int uninitialized_var(size0);
1653         u32 uninitialized_var(f0);
1654         int ind;
1655         u8 op0 = 0;
1656
1657         spin_lock_irqsave(&qp->sq.lock, flags);
1658
1659         /* XXX check that state is OK to post send */
1660
1661         ind = qp->sq.next_ind;
1662
1663         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1664                 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1665                         mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1666                                         " %d max, %d nreq)\n", qp->qpn,
1667                                         qp->sq.head, qp->sq.tail,
1668                                         qp->sq.max, nreq);
1669                         err = -ENOMEM;
1670                         *bad_wr = wr;
1671                         goto out;
1672                 }
1673
1674                 wqe = get_send_wqe(qp, ind);
1675                 prev_wqe = qp->sq.last;
1676                 qp->sq.last = wqe;
1677
1678                 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1679                 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1680                 ((struct mthca_next_seg *) wqe)->flags =
1681                         ((wr->send_flags & IB_SEND_SIGNALED) ?
1682                          cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1683                         ((wr->send_flags & IB_SEND_SOLICITED) ?
1684                          cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1685                         cpu_to_be32(1);
1686                 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1687                     wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1688                         ((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
1689
1690                 wqe += sizeof (struct mthca_next_seg);
1691                 size = sizeof (struct mthca_next_seg) / 16;
1692
1693                 switch (qp->transport) {
1694                 case RC:
1695                         switch (wr->opcode) {
1696                         case IB_WR_ATOMIC_CMP_AND_SWP:
1697                         case IB_WR_ATOMIC_FETCH_AND_ADD:
1698                                 set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
1699                                               wr->wr.atomic.rkey);
1700                                 wqe += sizeof (struct mthca_raddr_seg);
1701
1702                                 set_atomic_seg(wqe, wr);
1703                                 wqe += sizeof (struct mthca_atomic_seg);
1704                                 size += (sizeof (struct mthca_raddr_seg) +
1705                                          sizeof (struct mthca_atomic_seg)) / 16;
1706                                 break;
1707
1708                         case IB_WR_RDMA_WRITE:
1709                         case IB_WR_RDMA_WRITE_WITH_IMM:
1710                         case IB_WR_RDMA_READ:
1711                                 set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1712                                               wr->wr.rdma.rkey);
1713                                 wqe  += sizeof (struct mthca_raddr_seg);
1714                                 size += sizeof (struct mthca_raddr_seg) / 16;
1715                                 break;
1716
1717                         default:
1718                                 /* No extra segments required for sends */
1719                                 break;
1720                         }
1721
1722                         break;
1723
1724                 case UC:
1725                         switch (wr->opcode) {
1726                         case IB_WR_RDMA_WRITE:
1727                         case IB_WR_RDMA_WRITE_WITH_IMM:
1728                                 set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1729                                               wr->wr.rdma.rkey);
1730                                 wqe  += sizeof (struct mthca_raddr_seg);
1731                                 size += sizeof (struct mthca_raddr_seg) / 16;
1732                                 break;
1733
1734                         default:
1735                                 /* No extra segments required for sends */
1736                                 break;
1737                         }
1738
1739                         break;
1740
1741                 case UD:
1742                         set_tavor_ud_seg(wqe, wr);
1743                         wqe  += sizeof (struct mthca_tavor_ud_seg);
1744                         size += sizeof (struct mthca_tavor_ud_seg) / 16;
1745                         break;
1746
1747                 case MLX:
1748                         err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1749                                                wqe - sizeof (struct mthca_next_seg),
1750                                                wqe);
1751                         if (err) {
1752                                 *bad_wr = wr;
1753                                 goto out;
1754                         }
1755                         wqe += sizeof (struct mthca_data_seg);
1756                         size += sizeof (struct mthca_data_seg) / 16;
1757                         break;
1758                 }
1759
1760                 if (wr->num_sge > qp->sq.max_gs) {
1761                         mthca_err(dev, "too many gathers\n");
1762                         err = -EINVAL;
1763                         *bad_wr = wr;
1764                         goto out;
1765                 }
1766
1767                 for (i = 0; i < wr->num_sge; ++i) {
1768                         mthca_set_data_seg(wqe, wr->sg_list + i);
1769                         wqe  += sizeof (struct mthca_data_seg);
1770                         size += sizeof (struct mthca_data_seg) / 16;
1771                 }
1772
1773                 /* Add one more inline data segment for ICRC */
1774                 if (qp->transport == MLX) {
1775                         ((struct mthca_data_seg *) wqe)->byte_count =
1776                                 cpu_to_be32((1 << 31) | 4);
1777                         ((u32 *) wqe)[1] = 0;
1778                         wqe += sizeof (struct mthca_data_seg);
1779                         size += sizeof (struct mthca_data_seg) / 16;
1780                 }
1781
1782                 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1783
1784                 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1785                         mthca_err(dev, "opcode invalid\n");
1786                         err = -EINVAL;
1787                         *bad_wr = wr;
1788                         goto out;
1789                 }
1790
1791                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1792                         cpu_to_be32(((ind << qp->sq.wqe_shift) +
1793                                      qp->send_wqe_offset) |
1794                                     mthca_opcode[wr->opcode]);
1795                 wmb();
1796                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1797                         cpu_to_be32((nreq ? 0 : MTHCA_NEXT_DBD) | size |
1798                                     ((wr->send_flags & IB_SEND_FENCE) ?
1799                                     MTHCA_NEXT_FENCE : 0));
1800
1801                 if (!nreq) {
1802                         size0 = size;
1803                         op0   = mthca_opcode[wr->opcode];
1804                         f0    = wr->send_flags & IB_SEND_FENCE ?
1805                                 MTHCA_SEND_DOORBELL_FENCE : 0;
1806                 }
1807
1808                 ++ind;
1809                 if (unlikely(ind >= qp->sq.max))
1810                         ind -= qp->sq.max;
1811         }
1812
1813 out:
1814         if (likely(nreq)) {
1815                 wmb();
1816
1817                 mthca_write64(((qp->sq.next_ind << qp->sq.wqe_shift) +
1818                                qp->send_wqe_offset) | f0 | op0,
1819                               (qp->qpn << 8) | size0,
1820                               dev->kar + MTHCA_SEND_DOORBELL,
1821                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1822                 /*
1823                  * Make sure doorbells don't leak out of SQ spinlock
1824                  * and reach the HCA out of order:
1825                  */
1826                 mmiowb();
1827         }
1828
1829         qp->sq.next_ind = ind;
1830         qp->sq.head    += nreq;
1831
1832         spin_unlock_irqrestore(&qp->sq.lock, flags);
1833         return err;
1834 }
1835
1836 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1837                              struct ib_recv_wr **bad_wr)
1838 {
1839         struct mthca_dev *dev = to_mdev(ibqp->device);
1840         struct mthca_qp *qp = to_mqp(ibqp);
1841         unsigned long flags;
1842         int err = 0;
1843         int nreq;
1844         int i;
1845         int size;
1846         /*
1847          * size0 is only used if nreq != 0, and it will always be
1848          * initialized the first time through the main loop before
1849          * nreq is incremented.  So nreq cannot become non-zero
1850          * without initializing size0, and it is in fact never used
1851          * uninitialized.
1852          */
1853         int uninitialized_var(size0);
1854         int ind;
1855         void *wqe;
1856         void *prev_wqe;
1857
1858         spin_lock_irqsave(&qp->rq.lock, flags);
1859
1860         /* XXX check that state is OK to post receive */
1861
1862         ind = qp->rq.next_ind;
1863
1864         for (nreq = 0; wr; wr = wr->next) {
1865                 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1866                         mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1867                                         " %d max, %d nreq)\n", qp->qpn,
1868                                         qp->rq.head, qp->rq.tail,
1869                                         qp->rq.max, nreq);
1870                         err = -ENOMEM;
1871                         *bad_wr = wr;
1872                         goto out;
1873                 }
1874
1875                 wqe = get_recv_wqe(qp, ind);
1876                 prev_wqe = qp->rq.last;
1877                 qp->rq.last = wqe;
1878
1879                 ((struct mthca_next_seg *) wqe)->ee_nds =
1880                         cpu_to_be32(MTHCA_NEXT_DBD);
1881                 ((struct mthca_next_seg *) wqe)->flags = 0;
1882
1883                 wqe += sizeof (struct mthca_next_seg);
1884                 size = sizeof (struct mthca_next_seg) / 16;
1885
1886                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1887                         err = -EINVAL;
1888                         *bad_wr = wr;
1889                         goto out;
1890                 }
1891
1892                 for (i = 0; i < wr->num_sge; ++i) {
1893                         mthca_set_data_seg(wqe, wr->sg_list + i);
1894                         wqe  += sizeof (struct mthca_data_seg);
1895                         size += sizeof (struct mthca_data_seg) / 16;
1896                 }
1897
1898                 qp->wrid[ind] = wr->wr_id;
1899
1900                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1901                         cpu_to_be32(MTHCA_NEXT_DBD | size);
1902
1903                 if (!nreq)
1904                         size0 = size;
1905
1906                 ++ind;
1907                 if (unlikely(ind >= qp->rq.max))
1908                         ind -= qp->rq.max;
1909
1910                 ++nreq;
1911                 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1912                         nreq = 0;
1913
1914                         wmb();
1915
1916                         mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1917                                       qp->qpn << 8, dev->kar + MTHCA_RECEIVE_DOORBELL,
1918                                       MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1919
1920                         qp->rq.next_ind = ind;
1921                         qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1922                 }
1923         }
1924
1925 out:
1926         if (likely(nreq)) {
1927                 wmb();
1928
1929                 mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1930                               qp->qpn << 8 | nreq, dev->kar + MTHCA_RECEIVE_DOORBELL,
1931                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1932         }
1933
1934         qp->rq.next_ind = ind;
1935         qp->rq.head    += nreq;
1936
1937         /*
1938          * Make sure doorbells don't leak out of RQ spinlock and reach
1939          * the HCA out of order:
1940          */
1941         mmiowb();
1942
1943         spin_unlock_irqrestore(&qp->rq.lock, flags);
1944         return err;
1945 }
1946
1947 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1948                           struct ib_send_wr **bad_wr)
1949 {
1950         struct mthca_dev *dev = to_mdev(ibqp->device);
1951         struct mthca_qp *qp = to_mqp(ibqp);
1952         u32 dbhi;
1953         void *wqe;
1954         void *prev_wqe;
1955         unsigned long flags;
1956         int err = 0;
1957         int nreq;
1958         int i;
1959         int size;
1960         /*
1961          * f0 and size0 are only used if nreq != 0, and they will
1962          * always be initialized the first time through the main loop
1963          * before nreq is incremented.  So nreq cannot become non-zero
1964          * without initializing f0 and size0, and they are in fact
1965          * never used uninitialized.
1966          */
1967         int uninitialized_var(size0);
1968         u32 uninitialized_var(f0);
1969         int ind;
1970         u8 op0 = 0;
1971
1972         spin_lock_irqsave(&qp->sq.lock, flags);
1973
1974         /* XXX check that state is OK to post send */
1975
1976         ind = qp->sq.head & (qp->sq.max - 1);
1977
1978         for (nreq = 0; wr; ++nreq, wr = wr->next) {
1979                 if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1980                         nreq = 0;
1981
1982                         dbhi = (MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1983                                 ((qp->sq.head & 0xffff) << 8) | f0 | op0;
1984
1985                         qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1986
1987                         /*
1988                          * Make sure that descriptors are written before
1989                          * doorbell record.
1990                          */
1991                         wmb();
1992                         *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1993
1994                         /*
1995                          * Make sure doorbell record is written before we
1996                          * write MMIO send doorbell.
1997                          */
1998                         wmb();
1999
2000                         mthca_write64(dbhi, (qp->qpn << 8) | size0,
2001                                       dev->kar + MTHCA_SEND_DOORBELL,
2002                                       MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2003                 }
2004
2005                 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2006                         mthca_err(dev, "SQ %06x full (%u head, %u tail,"
2007                                         " %d max, %d nreq)\n", qp->qpn,
2008                                         qp->sq.head, qp->sq.tail,
2009                                         qp->sq.max, nreq);
2010                         err = -ENOMEM;
2011                         *bad_wr = wr;
2012                         goto out;
2013                 }
2014
2015                 wqe = get_send_wqe(qp, ind);
2016                 prev_wqe = qp->sq.last;
2017                 qp->sq.last = wqe;
2018
2019                 ((struct mthca_next_seg *) wqe)->flags =
2020                         ((wr->send_flags & IB_SEND_SIGNALED) ?
2021                          cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
2022                         ((wr->send_flags & IB_SEND_SOLICITED) ?
2023                          cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
2024                         ((wr->send_flags & IB_SEND_IP_CSUM) ?
2025                          cpu_to_be32(MTHCA_NEXT_IP_CSUM | MTHCA_NEXT_TCP_UDP_CSUM) : 0) |
2026                         cpu_to_be32(1);
2027                 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
2028                     wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
2029                         ((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
2030
2031                 wqe += sizeof (struct mthca_next_seg);
2032                 size = sizeof (struct mthca_next_seg) / 16;
2033
2034                 switch (qp->transport) {
2035                 case RC:
2036                         switch (wr->opcode) {
2037                         case IB_WR_ATOMIC_CMP_AND_SWP:
2038                         case IB_WR_ATOMIC_FETCH_AND_ADD:
2039                                 set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2040                                               wr->wr.atomic.rkey);
2041                                 wqe += sizeof (struct mthca_raddr_seg);
2042
2043                                 set_atomic_seg(wqe, wr);
2044                                 wqe  += sizeof (struct mthca_atomic_seg);
2045                                 size += (sizeof (struct mthca_raddr_seg) +
2046                                          sizeof (struct mthca_atomic_seg)) / 16;
2047                                 break;
2048
2049                         case IB_WR_RDMA_READ:
2050                         case IB_WR_RDMA_WRITE:
2051                         case IB_WR_RDMA_WRITE_WITH_IMM:
2052                                 set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2053                                               wr->wr.rdma.rkey);
2054                                 wqe  += sizeof (struct mthca_raddr_seg);
2055                                 size += sizeof (struct mthca_raddr_seg) / 16;
2056                                 break;
2057
2058                         default:
2059                                 /* No extra segments required for sends */
2060                                 break;
2061                         }
2062
2063                         break;
2064
2065                 case UC:
2066                         switch (wr->opcode) {
2067                         case IB_WR_RDMA_WRITE:
2068                         case IB_WR_RDMA_WRITE_WITH_IMM:
2069                                 set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2070                                               wr->wr.rdma.rkey);
2071                                 wqe  += sizeof (struct mthca_raddr_seg);
2072                                 size += sizeof (struct mthca_raddr_seg) / 16;
2073                                 break;
2074
2075                         default:
2076                                 /* No extra segments required for sends */
2077                                 break;
2078                         }
2079
2080                         break;
2081
2082                 case UD:
2083                         set_arbel_ud_seg(wqe, wr);
2084                         wqe  += sizeof (struct mthca_arbel_ud_seg);
2085                         size += sizeof (struct mthca_arbel_ud_seg) / 16;
2086                         break;
2087
2088                 case MLX:
2089                         err = build_mlx_header(dev, to_msqp(qp), ind, wr,
2090                                                wqe - sizeof (struct mthca_next_seg),
2091                                                wqe);
2092                         if (err) {
2093                                 *bad_wr = wr;
2094                                 goto out;
2095                         }
2096                         wqe += sizeof (struct mthca_data_seg);
2097                         size += sizeof (struct mthca_data_seg) / 16;
2098                         break;
2099                 }
2100
2101                 if (wr->num_sge > qp->sq.max_gs) {
2102                         mthca_err(dev, "too many gathers\n");
2103                         err = -EINVAL;
2104                         *bad_wr = wr;
2105                         goto out;
2106                 }
2107
2108                 for (i = 0; i < wr->num_sge; ++i) {
2109                         mthca_set_data_seg(wqe, wr->sg_list + i);
2110                         wqe  += sizeof (struct mthca_data_seg);
2111                         size += sizeof (struct mthca_data_seg) / 16;
2112                 }
2113
2114                 /* Add one more inline data segment for ICRC */
2115                 if (qp->transport == MLX) {
2116                         ((struct mthca_data_seg *) wqe)->byte_count =
2117                                 cpu_to_be32((1 << 31) | 4);
2118                         ((u32 *) wqe)[1] = 0;
2119                         wqe += sizeof (struct mthca_data_seg);
2120                         size += sizeof (struct mthca_data_seg) / 16;
2121                 }
2122
2123                 qp->wrid[ind + qp->rq.max] = wr->wr_id;
2124
2125                 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2126                         mthca_err(dev, "opcode invalid\n");
2127                         err = -EINVAL;
2128                         *bad_wr = wr;
2129                         goto out;
2130                 }
2131
2132                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
2133                         cpu_to_be32(((ind << qp->sq.wqe_shift) +
2134                                      qp->send_wqe_offset) |
2135                                     mthca_opcode[wr->opcode]);
2136                 wmb();
2137                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
2138                         cpu_to_be32(MTHCA_NEXT_DBD | size |
2139                                     ((wr->send_flags & IB_SEND_FENCE) ?
2140                                      MTHCA_NEXT_FENCE : 0));
2141
2142                 if (!nreq) {
2143                         size0 = size;
2144                         op0   = mthca_opcode[wr->opcode];
2145                         f0    = wr->send_flags & IB_SEND_FENCE ?
2146                                 MTHCA_SEND_DOORBELL_FENCE : 0;
2147                 }
2148
2149                 ++ind;
2150                 if (unlikely(ind >= qp->sq.max))
2151                         ind -= qp->sq.max;
2152         }
2153
2154 out:
2155         if (likely(nreq)) {
2156                 dbhi = (nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0;
2157
2158                 qp->sq.head += nreq;
2159
2160                 /*
2161                  * Make sure that descriptors are written before
2162                  * doorbell record.
2163                  */
2164                 wmb();
2165                 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2166
2167                 /*
2168                  * Make sure doorbell record is written before we
2169                  * write MMIO send doorbell.
2170                  */
2171                 wmb();
2172
2173                 mthca_write64(dbhi, (qp->qpn << 8) | size0, dev->kar + MTHCA_SEND_DOORBELL,
2174                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2175         }
2176
2177         /*
2178          * Make sure doorbells don't leak out of SQ spinlock and reach
2179          * the HCA out of order:
2180          */
2181         mmiowb();
2182
2183         spin_unlock_irqrestore(&qp->sq.lock, flags);
2184         return err;
2185 }
2186
2187 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2188                              struct ib_recv_wr **bad_wr)
2189 {
2190         struct mthca_dev *dev = to_mdev(ibqp->device);
2191         struct mthca_qp *qp = to_mqp(ibqp);
2192         unsigned long flags;
2193         int err = 0;
2194         int nreq;
2195         int ind;
2196         int i;
2197         void *wqe;
2198
2199         spin_lock_irqsave(&qp->rq.lock, flags);
2200
2201         /* XXX check that state is OK to post receive */
2202
2203         ind = qp->rq.head & (qp->rq.max - 1);
2204
2205         for (nreq = 0; wr; ++nreq, wr = wr->next) {
2206                 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2207                         mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2208                                         " %d max, %d nreq)\n", qp->qpn,
2209                                         qp->rq.head, qp->rq.tail,
2210                                         qp->rq.max, nreq);
2211                         err = -ENOMEM;
2212                         *bad_wr = wr;
2213                         goto out;
2214                 }
2215
2216                 wqe = get_recv_wqe(qp, ind);
2217
2218                 ((struct mthca_next_seg *) wqe)->flags = 0;
2219
2220                 wqe += sizeof (struct mthca_next_seg);
2221
2222                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2223                         err = -EINVAL;
2224                         *bad_wr = wr;
2225                         goto out;
2226                 }
2227
2228                 for (i = 0; i < wr->num_sge; ++i) {
2229                         mthca_set_data_seg(wqe, wr->sg_list + i);
2230                         wqe += sizeof (struct mthca_data_seg);
2231                 }
2232
2233                 if (i < qp->rq.max_gs)
2234                         mthca_set_data_seg_inval(wqe);
2235
2236                 qp->wrid[ind] = wr->wr_id;
2237
2238                 ++ind;
2239                 if (unlikely(ind >= qp->rq.max))
2240                         ind -= qp->rq.max;
2241         }
2242 out:
2243         if (likely(nreq)) {
2244                 qp->rq.head += nreq;
2245
2246                 /*
2247                  * Make sure that descriptors are written before
2248                  * doorbell record.
2249                  */
2250                 wmb();
2251                 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2252         }
2253
2254         spin_unlock_irqrestore(&qp->rq.lock, flags);
2255         return err;
2256 }
2257
2258 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2259                         int index, int *dbd, __be32 *new_wqe)
2260 {
2261         struct mthca_next_seg *next;
2262
2263         /*
2264          * For SRQs, all receive WQEs generate a CQE, so we're always
2265          * at the end of the doorbell chain.
2266          */
2267         if (qp->ibqp.srq && !is_send) {
2268                 *new_wqe = 0;
2269                 return;
2270         }
2271
2272         if (is_send)
2273                 next = get_send_wqe(qp, index);
2274         else
2275                 next = get_recv_wqe(qp, index);
2276
2277         *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2278         if (next->ee_nds & cpu_to_be32(0x3f))
2279                 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2280                         (next->ee_nds & cpu_to_be32(0x3f));
2281         else
2282                 *new_wqe = 0;
2283 }
2284
2285 int mthca_init_qp_table(struct mthca_dev *dev)
2286 {
2287         int err;
2288         u8 status;
2289         int i;
2290
2291         spin_lock_init(&dev->qp_table.lock);
2292
2293         /*
2294          * We reserve 2 extra QPs per port for the special QPs.  The
2295          * special QP for port 1 has to be even, so round up.
2296          */
2297         dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2298         err = mthca_alloc_init(&dev->qp_table.alloc,
2299                                dev->limits.num_qps,
2300                                (1 << 24) - 1,
2301                                dev->qp_table.sqp_start +
2302                                MTHCA_MAX_PORTS * 2);
2303         if (err)
2304                 return err;
2305
2306         err = mthca_array_init(&dev->qp_table.qp,
2307                                dev->limits.num_qps);
2308         if (err) {
2309                 mthca_alloc_cleanup(&dev->qp_table.alloc);
2310                 return err;
2311         }
2312
2313         for (i = 0; i < 2; ++i) {
2314                 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2315                                             dev->qp_table.sqp_start + i * 2,
2316                                             &status);
2317                 if (err)
2318                         goto err_out;
2319                 if (status) {
2320                         mthca_warn(dev, "CONF_SPECIAL_QP returned "
2321                                    "status %02x, aborting.\n",
2322                                    status);
2323                         err = -EINVAL;
2324                         goto err_out;
2325                 }
2326         }
2327         return 0;
2328
2329  err_out:
2330         for (i = 0; i < 2; ++i)
2331                 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2332
2333         mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2334         mthca_alloc_cleanup(&dev->qp_table.alloc);
2335
2336         return err;
2337 }
2338
2339 void mthca_cleanup_qp_table(struct mthca_dev *dev)
2340 {
2341         int i;
2342         u8 status;
2343
2344         for (i = 0; i < 2; ++i)
2345                 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2346
2347         mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2348         mthca_alloc_cleanup(&dev->qp_table.alloc);
2349 }