git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/hw/mthca/mthca_cq.c

   1 /*
   2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2005 Cisco Systems, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  *
  34  * $Id: mthca_cq.c 1369 2004-12-20 16:17:07Z roland $
  35  */
  36
  37 #include <linux/init.h>
  38 #include <linux/hardirq.h>
  39
  40 #include <ib_pack.h>
  41
  42 #include "mthca_dev.h"
  43 #include "mthca_cmd.h"
  44 #include "mthca_memfree.h"
  45
  46 enum {
  47         MTHCA_MAX_DIRECT_CQ_SIZE = 4 * PAGE_SIZE
  48 };
  49
  50 enum {
  51         MTHCA_CQ_ENTRY_SIZE = 0x20
  52 };
  53
  54 /*
  55  * Must be packed because start is 64 bits but only aligned to 32 bits.
  56  */
  57 struct mthca_cq_context {
  58         u32 flags;
  59         u64 start;
  60         u32 logsize_usrpage;
  61         u32 error_eqn;          /* Tavor only */
  62         u32 comp_eqn;
  63         u32 pd;
  64         u32 lkey;
  65         u32 last_notified_index;
  66         u32 solicit_producer_index;
  67         u32 consumer_index;
  68         u32 producer_index;
  69         u32 cqn;
  70         u32 ci_db;              /* Arbel only */
  71         u32 state_db;           /* Arbel only */
  72         u32 reserved;
  73 } __attribute__((packed));
  74
  75 #define MTHCA_CQ_STATUS_OK          ( 0 << 28)
  76 #define MTHCA_CQ_STATUS_OVERFLOW    ( 9 << 28)
  77 #define MTHCA_CQ_STATUS_WRITE_FAIL  (10 << 28)
  78 #define MTHCA_CQ_FLAG_TR            ( 1 << 18)
  79 #define MTHCA_CQ_FLAG_OI            ( 1 << 17)
  80 #define MTHCA_CQ_STATE_DISARMED     ( 0 <<  8)
  81 #define MTHCA_CQ_STATE_ARMED        ( 1 <<  8)
  82 #define MTHCA_CQ_STATE_ARMED_SOL    ( 4 <<  8)
  83 #define MTHCA_EQ_STATE_FIRED        (10 <<  8)
  84
  85 enum {
  86         MTHCA_ERROR_CQE_OPCODE_MASK = 0xfe
  87 };
  88
  89 enum {
  90         SYNDROME_LOCAL_LENGTH_ERR        = 0x01,
  91         SYNDROME_LOCAL_QP_OP_ERR         = 0x02,
  92         SYNDROME_LOCAL_EEC_OP_ERR        = 0x03,
  93         SYNDROME_LOCAL_PROT_ERR          = 0x04,
  94         SYNDROME_WR_FLUSH_ERR            = 0x05,
  95         SYNDROME_MW_BIND_ERR             = 0x06,
  96         SYNDROME_BAD_RESP_ERR            = 0x10,
  97         SYNDROME_LOCAL_ACCESS_ERR        = 0x11,
  98         SYNDROME_REMOTE_INVAL_REQ_ERR    = 0x12,
  99         SYNDROME_REMOTE_ACCESS_ERR       = 0x13,
 100         SYNDROME_REMOTE_OP_ERR           = 0x14,
 101         SYNDROME_RETRY_EXC_ERR           = 0x15,
 102         SYNDROME_RNR_RETRY_EXC_ERR       = 0x16,
 103         SYNDROME_LOCAL_RDD_VIOL_ERR      = 0x20,
 104         SYNDROME_REMOTE_INVAL_RD_REQ_ERR = 0x21,
 105         SYNDROME_REMOTE_ABORTED_ERR      = 0x22,
 106         SYNDROME_INVAL_EECN_ERR          = 0x23,
 107         SYNDROME_INVAL_EEC_STATE_ERR     = 0x24
 108 };
 109
 110 struct mthca_cqe {
 111         u32 my_qpn;
 112         u32 my_ee;
 113         u32 rqpn;
 114         u16 sl_g_mlpath;
 115         u16 rlid;
 116         u32 imm_etype_pkey_eec;
 117         u32 byte_cnt;
 118         u32 wqe;
 119         u8  opcode;
 120         u8  is_send;
 121         u8  reserved;
 122         u8  owner;
 123 };
 124
 125 struct mthca_err_cqe {
 126         u32 my_qpn;
 127         u32 reserved1[3];
 128         u8  syndrome;
 129         u8  reserved2;
 130         u16 db_cnt;
 131         u32 reserved3;
 132         u32 wqe;
 133         u8  opcode;
 134         u8  reserved4[2];
 135         u8  owner;
 136 };
 137
 138 #define MTHCA_CQ_ENTRY_OWNER_SW      (0 << 7)
 139 #define MTHCA_CQ_ENTRY_OWNER_HW      (1 << 7)
 140
 141 #define MTHCA_TAVOR_CQ_DB_INC_CI       (1 << 24)
 142 #define MTHCA_TAVOR_CQ_DB_REQ_NOT      (2 << 24)
 143 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL  (3 << 24)
 144 #define MTHCA_TAVOR_CQ_DB_SET_CI       (4 << 24)
 145 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_MULT (5 << 24)
 146
 147 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL  (1 << 24)
 148 #define MTHCA_ARBEL_CQ_DB_REQ_NOT      (2 << 24)
 149 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
 150
 151 static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
 152 {
 153         if (cq->is_direct)
 154                 return cq->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
 155         else
 156                 return cq->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
 157                         + (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
 158 }
 159
 160 static inline struct mthca_cqe *cqe_sw(struct mthca_cq *cq, int i)
 161 {
 162         struct mthca_cqe *cqe = get_cqe(cq, i);
 163         return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
 164 }
 165
 166 static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
 167 {
 168         return cqe_sw(cq, cq->cons_index & cq->ibcq.cqe);
 169 }
 170
 171 static inline void set_cqe_hw(struct mthca_cqe *cqe)
 172 {
 173         cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
 174 }
 175
 176 static void dump_cqe(struct mthca_dev *dev, void *cqe_ptr)
 177 {
 178         __be32 *cqe = cqe_ptr;
 179
 180         (void) cqe;     /* avoid warning if mthca_dbg compiled away... */
 181         mthca_dbg(dev, "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
 182                   be32_to_cpu(cqe[0]), be32_to_cpu(cqe[1]), be32_to_cpu(cqe[2]),
 183                   be32_to_cpu(cqe[3]), be32_to_cpu(cqe[4]), be32_to_cpu(cqe[5]),
 184                   be32_to_cpu(cqe[6]), be32_to_cpu(cqe[7]));
 185 }
 186
 187 /*
 188  * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
 189  * should be correct before calling update_cons_index().
 190  */
 191 static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
 192                                      int incr)
 193 {
 194         u32 doorbell[2];
 195
 196         if (mthca_is_memfree(dev)) {
 197                 *cq->set_ci_db = cpu_to_be32(cq->cons_index);
 198                 wmb();
 199         } else {
 200                 doorbell[0] = cpu_to_be32(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn);
 201                 doorbell[1] = cpu_to_be32(incr - 1);
 202
 203                 mthca_write64(doorbell,
 204                               dev->kar + MTHCA_CQ_DOORBELL,
 205                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
 206         }
 207 }
 208
 209 void mthca_cq_event(struct mthca_dev *dev, u32 cqn)
 210 {
 211         struct mthca_cq *cq;
 212
 213         cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
 214
 215         if (!cq) {
 216                 mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
 217                 return;
 218         }
 219
 220         ++cq->arm_sn;
 221
 222         cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
 223 }
 224
 225 void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn)
 226 {
 227         struct mthca_cq *cq;
 228         struct mthca_cqe *cqe;
 229         int prod_index;
 230         int nfreed = 0;
 231
 232         spin_lock_irq(&dev->cq_table.lock);
 233         cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
 234         if (cq)
 235                 atomic_inc(&cq->refcount);
 236         spin_unlock_irq(&dev->cq_table.lock);
 237
 238         if (!cq)
 239                 return;
 240
 241         spin_lock_irq(&cq->lock);
 242
 243         /*
 244          * First we need to find the current producer index, so we
 245          * know where to start cleaning from.  It doesn't matter if HW
 246          * adds new entries after this loop -- the QP we're worried
 247          * about is already in RESET, so the new entries won't come
 248          * from our QP and therefore don't need to be checked.
 249          */
 250         for (prod_index = cq->cons_index;
 251              cqe_sw(cq, prod_index & cq->ibcq.cqe);
 252              ++prod_index)
 253                 if (prod_index == cq->cons_index + cq->ibcq.cqe)
 254                         break;
 255
 256         if (0)
 257                 mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
 258                           qpn, cqn, cq->cons_index, prod_index);
 259
 260         /*
 261          * Now sweep backwards through the CQ, removing CQ entries
 262          * that match our QP by copying older entries on top of them.
 263          */
 264         while (prod_index > cq->cons_index) {
 265                 cqe = get_cqe(cq, (prod_index - 1) & cq->ibcq.cqe);
 266                 if (cqe->my_qpn == cpu_to_be32(qpn))
 267                         ++nfreed;
 268                 else if (nfreed)
 269                         memcpy(get_cqe(cq, (prod_index - 1 + nfreed) &
 270                                        cq->ibcq.cqe),
 271                                cqe,
 272                                MTHCA_CQ_ENTRY_SIZE);
 273                 --prod_index;
 274         }
 275
 276         if (nfreed) {
 277                 wmb();
 278                 cq->cons_index += nfreed;
 279                 update_cons_index(dev, cq, nfreed);
 280         }
 281
 282         spin_unlock_irq(&cq->lock);
 283         if (atomic_dec_and_test(&cq->refcount))
 284                 wake_up(&cq->wait);
 285 }
 286
 287 static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
 288                             struct mthca_qp *qp, int wqe_index, int is_send,
 289                             struct mthca_err_cqe *cqe,
 290                             struct ib_wc *entry, int *free_cqe)
 291 {
 292         int err;
 293         int dbd;
 294         u32 new_wqe;
 295
 296         if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
 297                 mthca_dbg(dev, "local QP operation err "
 298                           "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
 299                           be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
 300                           cq->cqn, cq->cons_index);
 301                 dump_cqe(dev, cqe);
 302         }
 303
 304         /*
 305          * For completions in error, only work request ID, status (and
 306          * freed resource count for RD) have to be set.
 307          */
 308         switch (cqe->syndrome) {
 309         case SYNDROME_LOCAL_LENGTH_ERR:
 310                 entry->status = IB_WC_LOC_LEN_ERR;
 311                 break;
 312         case SYNDROME_LOCAL_QP_OP_ERR:
 313                 entry->status = IB_WC_LOC_QP_OP_ERR;
 314                 break;
 315         case SYNDROME_LOCAL_EEC_OP_ERR:
 316                 entry->status = IB_WC_LOC_EEC_OP_ERR;
 317                 break;
 318         case SYNDROME_LOCAL_PROT_ERR:
 319                 entry->status = IB_WC_LOC_PROT_ERR;
 320                 break;
 321         case SYNDROME_WR_FLUSH_ERR:
 322                 entry->status = IB_WC_WR_FLUSH_ERR;
 323                 break;
 324         case SYNDROME_MW_BIND_ERR:
 325                 entry->status = IB_WC_MW_BIND_ERR;
 326                 break;
 327         case SYNDROME_BAD_RESP_ERR:
 328                 entry->status = IB_WC_BAD_RESP_ERR;
 329                 break;
 330         case SYNDROME_LOCAL_ACCESS_ERR:
 331                 entry->status = IB_WC_LOC_ACCESS_ERR;
 332                 break;
 333         case SYNDROME_REMOTE_INVAL_REQ_ERR:
 334                 entry->status = IB_WC_REM_INV_REQ_ERR;
 335                 break;
 336         case SYNDROME_REMOTE_ACCESS_ERR:
 337                 entry->status = IB_WC_REM_ACCESS_ERR;
 338                 break;
 339         case SYNDROME_REMOTE_OP_ERR:
 340                 entry->status = IB_WC_REM_OP_ERR;
 341                 break;
 342         case SYNDROME_RETRY_EXC_ERR:
 343                 entry->status = IB_WC_RETRY_EXC_ERR;
 344                 break;
 345         case SYNDROME_RNR_RETRY_EXC_ERR:
 346                 entry->status = IB_WC_RNR_RETRY_EXC_ERR;
 347                 break;
 348         case SYNDROME_LOCAL_RDD_VIOL_ERR:
 349                 entry->status = IB_WC_LOC_RDD_VIOL_ERR;
 350                 break;
 351         case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
 352                 entry->status = IB_WC_REM_INV_RD_REQ_ERR;
 353                 break;
 354         case SYNDROME_REMOTE_ABORTED_ERR:
 355                 entry->status = IB_WC_REM_ABORT_ERR;
 356                 break;
 357         case SYNDROME_INVAL_EECN_ERR:
 358                 entry->status = IB_WC_INV_EECN_ERR;
 359                 break;
 360         case SYNDROME_INVAL_EEC_STATE_ERR:
 361                 entry->status = IB_WC_INV_EEC_STATE_ERR;
 362                 break;
 363         default:
 364                 entry->status = IB_WC_GENERAL_ERR;
 365                 break;
 366         }
 367
 368         err = mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
 369         if (err)
 370                 return err;
 371
 372         /*
 373          * If we're at the end of the WQE chain, or we've used up our
 374          * doorbell count, free the CQE.  Otherwise just update it for
 375          * the next poll operation.
 376          *
 377          * This does not apply to mem-free HCAs: they don't use the
 378          * doorbell count field, and so we should always free the CQE.
 379          */
 380         if (mthca_is_memfree(dev) ||
 381             !(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
 382                 return 0;
 383
 384         cqe->db_cnt   = cpu_to_be16(be16_to_cpu(cqe->db_cnt) - dbd);
 385         cqe->wqe      = new_wqe;
 386         cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
 387
 388         *free_cqe = 0;
 389
 390         return 0;
 391 }
 392
 393 static inline int mthca_poll_one(struct mthca_dev *dev,
 394                                  struct mthca_cq *cq,
 395                                  struct mthca_qp **cur_qp,
 396                                  int *freed,
 397                                  struct ib_wc *entry)
 398 {
 399         struct mthca_wq *wq;
 400         struct mthca_cqe *cqe;
 401         int wqe_index;
 402         int is_error;
 403         int is_send;
 404         int free_cqe = 1;
 405         int err = 0;
 406
 407         cqe = next_cqe_sw(cq);
 408         if (!cqe)
 409                 return -EAGAIN;
 410
 411         /*
 412          * Make sure we read CQ entry contents after we've checked the
 413          * ownership bit.
 414          */
 415         rmb();
 416
 417         if (0) {
 418                 mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
 419                           cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
 420                           be32_to_cpu(cqe->wqe));
 421                 dump_cqe(dev, cqe);
 422         }
 423
 424         is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
 425                 MTHCA_ERROR_CQE_OPCODE_MASK;
 426         is_send  = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
 427
 428         if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
 429                 /*
 430                  * We do not have to take the QP table lock here,
 431                  * because CQs will be locked while QPs are removed
 432                  * from the table.
 433                  */
 434                 *cur_qp = mthca_array_get(&dev->qp_table.qp,
 435                                           be32_to_cpu(cqe->my_qpn) &
 436                                           (dev->limits.num_qps - 1));
 437                 if (!*cur_qp) {
 438                         mthca_warn(dev, "CQ entry for unknown QP %06x\n",
 439                                    be32_to_cpu(cqe->my_qpn) & 0xffffff);
 440                         err = -EINVAL;
 441                         goto out;
 442                 }
 443         }
 444
 445         entry->qp_num = (*cur_qp)->qpn;
 446
 447         if (is_send) {
 448                 wq = &(*cur_qp)->sq;
 449                 wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
 450                              >> wq->wqe_shift);
 451                 entry->wr_id = (*cur_qp)->wrid[wqe_index +
 452                                                (*cur_qp)->rq.max];
 453         } else {
 454                 wq = &(*cur_qp)->rq;
 455                 wqe_index = be32_to_cpu(cqe->wqe) >> wq->wqe_shift;
 456                 entry->wr_id = (*cur_qp)->wrid[wqe_index];
 457         }
 458
 459         if (wq->last_comp < wqe_index)
 460                 wq->tail += wqe_index - wq->last_comp;
 461         else
 462                 wq->tail += wqe_index + wq->max - wq->last_comp;
 463
 464         wq->last_comp = wqe_index;
 465
 466         if (0)
 467                 mthca_dbg(dev, "%s completion for QP %06x, index %d (nr %d)\n",
 468                           is_send ? "Send" : "Receive",
 469                           (*cur_qp)->qpn, wqe_index, wq->max);
 470
 471         if (is_error) {
 472                 err = handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
 473                                        (struct mthca_err_cqe *) cqe,
 474                                        entry, &free_cqe);
 475                 goto out;
 476         }
 477
 478         if (is_send) {
 479                 entry->wc_flags = 0;
 480                 switch (cqe->opcode) {
 481                 case MTHCA_OPCODE_RDMA_WRITE:
 482                         entry->opcode    = IB_WC_RDMA_WRITE;
 483                         break;
 484                 case MTHCA_OPCODE_RDMA_WRITE_IMM:
 485                         entry->opcode    = IB_WC_RDMA_WRITE;
 486                         entry->wc_flags |= IB_WC_WITH_IMM;
 487                         break;
 488                 case MTHCA_OPCODE_SEND:
 489                         entry->opcode    = IB_WC_SEND;
 490                         break;
 491                 case MTHCA_OPCODE_SEND_IMM:
 492                         entry->opcode    = IB_WC_SEND;
 493                         entry->wc_flags |= IB_WC_WITH_IMM;
 494                         break;
 495                 case MTHCA_OPCODE_RDMA_READ:
 496                         entry->opcode    = IB_WC_RDMA_READ;
 497                         entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
 498                         break;
 499                 case MTHCA_OPCODE_ATOMIC_CS:
 500                         entry->opcode    = IB_WC_COMP_SWAP;
 501                         entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
 502                         break;
 503                 case MTHCA_OPCODE_ATOMIC_FA:
 504                         entry->opcode    = IB_WC_FETCH_ADD;
 505                         entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
 506                         break;
 507                 case MTHCA_OPCODE_BIND_MW:
 508                         entry->opcode    = IB_WC_BIND_MW;
 509                         break;
 510                 default:
 511                         entry->opcode    = MTHCA_OPCODE_INVALID;
 512                         break;
 513                 }
 514         } else {
 515                 entry->byte_len = be32_to_cpu(cqe->byte_cnt);
 516                 switch (cqe->opcode & 0x1f) {
 517                 case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
 518                 case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
 519                         entry->wc_flags = IB_WC_WITH_IMM;
 520                         entry->imm_data = cqe->imm_etype_pkey_eec;
 521                         entry->opcode = IB_WC_RECV;
 522                         break;
 523                 case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
 524                 case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
 525                         entry->wc_flags = IB_WC_WITH_IMM;
 526                         entry->imm_data = cqe->imm_etype_pkey_eec;
 527                         entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
 528                         break;
 529                 default:
 530                         entry->wc_flags = 0;
 531                         entry->opcode = IB_WC_RECV;
 532                         break;
 533                 }
 534                 entry->slid        = be16_to_cpu(cqe->rlid);
 535                 entry->sl          = be16_to_cpu(cqe->sl_g_mlpath) >> 12;
 536                 entry->src_qp      = be32_to_cpu(cqe->rqpn) & 0xffffff;
 537                 entry->dlid_path_bits = be16_to_cpu(cqe->sl_g_mlpath) & 0x7f;
 538                 entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
 539                 entry->wc_flags   |= be16_to_cpu(cqe->sl_g_mlpath) & 0x80 ?
 540                                         IB_WC_GRH : 0;
 541         }
 542
 543         entry->status = IB_WC_SUCCESS;
 544
 545  out:
 546         if (likely(free_cqe)) {
 547                 set_cqe_hw(cqe);
 548                 ++(*freed);
 549                 ++cq->cons_index;
 550         }
 551
 552         return err;
 553 }
 554
 555 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
 556                   struct ib_wc *entry)
 557 {
 558         struct mthca_dev *dev = to_mdev(ibcq->device);
 559         struct mthca_cq *cq = to_mcq(ibcq);
 560         struct mthca_qp *qp = NULL;
 561         unsigned long flags;
 562         int err = 0;
 563         int freed = 0;
 564         int npolled;
 565
 566         spin_lock_irqsave(&cq->lock, flags);
 567
 568         for (npolled = 0; npolled < num_entries; ++npolled) {
 569                 err = mthca_poll_one(dev, cq, &qp,
 570                                      &freed, entry + npolled);
 571                 if (err)
 572                         break;
 573         }
 574
 575         if (freed) {
 576                 wmb();
 577                 update_cons_index(dev, cq, freed);
 578         }
 579
 580         spin_unlock_irqrestore(&cq->lock, flags);
 581
 582         return err == 0 || err == -EAGAIN ? npolled : err;
 583 }
 584
 585 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify notify)
 586 {
 587         u32 doorbell[2];
 588
 589         doorbell[0] = cpu_to_be32((notify == IB_CQ_SOLICITED ?
 590                                    MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
 591                                    MTHCA_TAVOR_CQ_DB_REQ_NOT)      |
 592                                   to_mcq(cq)->cqn);
 593         doorbell[1] = 0xffffffff;
 594
 595         mthca_write64(doorbell,
 596                       to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
 597                       MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
 598
 599         return 0;
 600 }
 601
 602 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
 603 {
 604         struct mthca_cq *cq = to_mcq(ibcq);
 605         u32 doorbell[2];
 606         u32 sn;
 607         u32 ci;
 608
 609         sn = cq->arm_sn & 3;
 610         ci = cpu_to_be32(cq->cons_index);
 611
 612         doorbell[0] = ci;
 613         doorbell[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
 614                                   (notify == IB_CQ_SOLICITED ? 1 : 2));
 615
 616         mthca_write_db_rec(doorbell, cq->arm_db);
 617
 618         /*
 619          * Make sure that the doorbell record in host memory is
 620          * written before ringing the doorbell via PCI MMIO.
 621          */
 622         wmb();
 623
 624         doorbell[0] = cpu_to_be32((sn << 28)                       |
 625                                   (notify == IB_CQ_SOLICITED ?
 626                                    MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
 627                                    MTHCA_ARBEL_CQ_DB_REQ_NOT)      |
 628                                   cq->cqn);
 629         doorbell[1] = ci;
 630
 631         mthca_write64(doorbell,
 632                       to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
 633                       MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
 634
 635         return 0;
 636 }
 637
 638 static void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq *cq)
 639 {
 640         int i;
 641         int size;
 642
 643         if (cq->is_direct)
 644                 dma_free_coherent(&dev->pdev->dev,
 645                                   (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE,
 646                                   cq->queue.direct.buf,
 647                                   pci_unmap_addr(&cq->queue.direct,
 648                                                  mapping));
 649         else {
 650                 size = (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE;
 651                 for (i = 0; i < (size + PAGE_SIZE - 1) / PAGE_SIZE; ++i)
 652                         if (cq->queue.page_list[i].buf)
 653                                 dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
 654                                                   cq->queue.page_list[i].buf,
 655                                                   pci_unmap_addr(&cq->queue.page_list[i],
 656                                                                  mapping));
 657
 658                 kfree(cq->queue.page_list);
 659         }
 660 }
 661
 662 static int mthca_alloc_cq_buf(struct mthca_dev *dev, int size,
 663                               struct mthca_cq *cq)
 664 {
 665         int err = -ENOMEM;
 666         int npages, shift;
 667         u64 *dma_list = NULL;
 668         dma_addr_t t;
 669         int i;
 670
 671         if (size <= MTHCA_MAX_DIRECT_CQ_SIZE) {
 672                 cq->is_direct = 1;
 673                 npages        = 1;
 674                 shift         = get_order(size) + PAGE_SHIFT;
 675
 676                 cq->queue.direct.buf = dma_alloc_coherent(&dev->pdev->dev,
 677                                                           size, &t, GFP_KERNEL);
 678                 if (!cq->queue.direct.buf)
 679                         return -ENOMEM;
 680
 681                 pci_unmap_addr_set(&cq->queue.direct, mapping, t);
 682
 683                 memset(cq->queue.direct.buf, 0, size);
 684
 685                 while (t & ((1 << shift) - 1)) {
 686                         --shift;
 687                         npages *= 2;
 688                 }
 689
 690                 dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
 691                 if (!dma_list)
 692                         goto err_free;
 693
 694                 for (i = 0; i < npages; ++i)
 695                         dma_list[i] = t + i * (1 << shift);
 696         } else {
 697                 cq->is_direct = 0;
 698                 npages        = (size + PAGE_SIZE - 1) / PAGE_SIZE;
 699                 shift         = PAGE_SHIFT;
 700
 701                 dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
 702                 if (!dma_list)
 703                         return -ENOMEM;
 704
 705                 cq->queue.page_list = kmalloc(npages * sizeof *cq->queue.page_list,
 706                                               GFP_KERNEL);
 707                 if (!cq->queue.page_list)
 708                         goto err_out;
 709
 710                 for (i = 0; i < npages; ++i)
 711                         cq->queue.page_list[i].buf = NULL;
 712
 713                 for (i = 0; i < npages; ++i) {
 714                         cq->queue.page_list[i].buf =
 715                                 dma_alloc_coherent(&dev->pdev->dev, PAGE_SIZE,
 716                                                    &t, GFP_KERNEL);
 717                         if (!cq->queue.page_list[i].buf)
 718                                 goto err_free;
 719
 720                         dma_list[i] = t;
 721                         pci_unmap_addr_set(&cq->queue.page_list[i], mapping, t);
 722
 723                         memset(cq->queue.page_list[i].buf, 0, PAGE_SIZE);
 724                 }
 725         }
 726
 727         err = mthca_mr_alloc_phys(dev, dev->driver_pd.pd_num,
 728                                   dma_list, shift, npages,
 729                                   0, size,
 730                                   MTHCA_MPT_FLAG_LOCAL_WRITE |
 731                                   MTHCA_MPT_FLAG_LOCAL_READ,
 732                                   &cq->mr);
 733         if (err)
 734                 goto err_free;
 735
 736         kfree(dma_list);
 737
 738         return 0;
 739
 740 err_free:
 741         mthca_free_cq_buf(dev, cq);
 742
 743 err_out:
 744         kfree(dma_list);
 745
 746         return err;
 747 }
 748
 749 int mthca_init_cq(struct mthca_dev *dev, int nent,
 750                   struct mthca_ucontext *ctx, u32 pdn,
 751                   struct mthca_cq *cq)
 752 {
 753         int size = nent * MTHCA_CQ_ENTRY_SIZE;
 754         struct mthca_mailbox *mailbox;
 755         struct mthca_cq_context *cq_context;
 756         int err = -ENOMEM;
 757         u8 status;
 758         int i;
 759
 760         might_sleep();
 761
 762         cq->ibcq.cqe  = nent - 1;
 763         cq->is_kernel = !ctx;
 764
 765         cq->cqn = mthca_alloc(&dev->cq_table.alloc);
 766         if (cq->cqn == -1)
 767                 return -ENOMEM;
 768
 769         if (mthca_is_memfree(dev)) {
 770                 err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
 771                 if (err)
 772                         goto err_out;
 773
 774                 if (cq->is_kernel) {
 775                         cq->arm_sn = 1;
 776
 777                         err = -ENOMEM;
 778
 779                         cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
 780                                                              cq->cqn, &cq->set_ci_db);
 781                         if (cq->set_ci_db_index < 0)
 782                                 goto err_out_icm;
 783
 784                         cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
 785                                                           cq->cqn, &cq->arm_db);
 786                         if (cq->arm_db_index < 0)
 787                                 goto err_out_ci;
 788                 }
 789         }
 790
 791         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 792         if (IS_ERR(mailbox))
 793                 goto err_out_arm;
 794
 795         cq_context = mailbox->buf;
 796
 797         if (cq->is_kernel) {
 798                 err = mthca_alloc_cq_buf(dev, size, cq);
 799                 if (err)
 800                         goto err_out_mailbox;
 801
 802                 for (i = 0; i < nent; ++i)
 803                         set_cqe_hw(get_cqe(cq, i));
 804         }
 805
 806         spin_lock_init(&cq->lock);
 807         atomic_set(&cq->refcount, 1);
 808         init_waitqueue_head(&cq->wait);
 809
 810         memset(cq_context, 0, sizeof *cq_context);
 811         cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
 812                                                   MTHCA_CQ_STATE_DISARMED |
 813                                                   MTHCA_CQ_FLAG_TR);
 814         cq_context->start           = cpu_to_be64(0);
 815         cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
 816         if (ctx)
 817                 cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
 818         else
 819                 cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
 820         cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
 821         cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
 822         cq_context->pd              = cpu_to_be32(pdn);
 823         cq_context->lkey            = cpu_to_be32(cq->mr.ibmr.lkey);
 824         cq_context->cqn             = cpu_to_be32(cq->cqn);
 825
 826         if (mthca_is_memfree(dev)) {
 827                 cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
 828                 cq_context->state_db = cpu_to_be32(cq->arm_db_index);
 829         }
 830
 831         err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn, &status);
 832         if (err) {
 833                 mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
 834                 goto err_out_free_mr;
 835         }
 836
 837         if (status) {
 838                 mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
 839                            status);
 840                 err = -EINVAL;
 841                 goto err_out_free_mr;
 842         }
 843
 844         spin_lock_irq(&dev->cq_table.lock);
 845         if (mthca_array_set(&dev->cq_table.cq,
 846                             cq->cqn & (dev->limits.num_cqs - 1),
 847                             cq)) {
 848                 spin_unlock_irq(&dev->cq_table.lock);
 849                 goto err_out_free_mr;
 850         }
 851         spin_unlock_irq(&dev->cq_table.lock);
 852
 853         cq->cons_index = 0;
 854
 855         mthca_free_mailbox(dev, mailbox);
 856
 857         return 0;
 858
 859 err_out_free_mr:
 860         if (cq->is_kernel) {
 861                 mthca_free_mr(dev, &cq->mr);
 862                 mthca_free_cq_buf(dev, cq);
 863         }
 864
 865 err_out_mailbox:
 866         mthca_free_mailbox(dev, mailbox);
 867
 868 err_out_arm:
 869         if (cq->is_kernel && mthca_is_memfree(dev))
 870                 mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
 871
 872 err_out_ci:
 873         if (cq->is_kernel && mthca_is_memfree(dev))
 874                 mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
 875
 876 err_out_icm:
 877         mthca_table_put(dev, dev->cq_table.table, cq->cqn);
 878
 879 err_out:
 880         mthca_free(&dev->cq_table.alloc, cq->cqn);
 881
 882         return err;
 883 }
 884
 885 void mthca_free_cq(struct mthca_dev *dev,
 886                    struct mthca_cq *cq)
 887 {
 888         struct mthca_mailbox *mailbox;
 889         int err;
 890         u8 status;
 891
 892         might_sleep();
 893
 894         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 895         if (IS_ERR(mailbox)) {
 896                 mthca_warn(dev, "No memory for mailbox to free CQ.\n");
 897                 return;
 898         }
 899
 900         err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn, &status);
 901         if (err)
 902                 mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
 903         else if (status)
 904                 mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n", status);
 905
 906         if (0) {
 907                 u32 *ctx = mailbox->buf;
 908                 int j;
 909
 910                 printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
 911                        cq->cqn, cq->cons_index,
 912                        cq->is_kernel ? !!next_cqe_sw(cq) : 0);
 913                 for (j = 0; j < 16; ++j)
 914                         printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
 915         }
 916
 917         spin_lock_irq(&dev->cq_table.lock);
 918         mthca_array_clear(&dev->cq_table.cq,
 919                           cq->cqn & (dev->limits.num_cqs - 1));
 920         spin_unlock_irq(&dev->cq_table.lock);
 921
 922         if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
 923                 synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
 924         else
 925                 synchronize_irq(dev->pdev->irq);
 926
 927         atomic_dec(&cq->refcount);
 928         wait_event(cq->wait, !atomic_read(&cq->refcount));
 929
 930         if (cq->is_kernel) {
 931                 mthca_free_mr(dev, &cq->mr);
 932                 mthca_free_cq_buf(dev, cq);
 933                 if (mthca_is_memfree(dev)) {
 934                         mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
 935                         mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
 936                 }
 937         }
 938
 939         mthca_table_put(dev, dev->cq_table.table, cq->cqn);
 940         mthca_free(&dev->cq_table.alloc, cq->cqn);
 941         mthca_free_mailbox(dev, mailbox);
 942 }
 943
 944 int __devinit mthca_init_cq_table(struct mthca_dev *dev)
 945 {
 946         int err;
 947
 948         spin_lock_init(&dev->cq_table.lock);
 949
 950         err = mthca_alloc_init(&dev->cq_table.alloc,
 951                                dev->limits.num_cqs,
 952                                (1 << 24) - 1,
 953                                dev->limits.reserved_cqs);
 954         if (err)
 955                 return err;
 956
 957         err = mthca_array_init(&dev->cq_table.cq,
 958                                dev->limits.num_cqs);
 959         if (err)
 960                 mthca_alloc_cleanup(&dev->cq_table.alloc);
 961
 962         return err;
 963 }
 964
 965 void __devexit mthca_cleanup_cq_table(struct mthca_dev *dev)
 966 {
 967         mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
 968         mthca_alloc_cleanup(&dev->cq_table.alloc);
 969 }