drivers/infiniband/hw/ipath/ipath_verbs.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. All rights reserved.
   3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <rdma/ib_mad.h>
  35 #include <rdma/ib_user_verbs.h>
  36 #include <linux/utsname.h>
  37
  38 #include "ipath_kernel.h"
  39 #include "ipath_verbs.h"
  40 #include "ipath_common.h"
  41
  42 /* Not static, because we don't want the compiler removing it */
  43 const char ipath_verbs_version[] = "ipath_verbs " IPATH_IDSTR;
  44
  45 static unsigned int ib_ipath_qp_table_size = 251;
  46 module_param_named(qp_table_size, ib_ipath_qp_table_size, uint, S_IRUGO);
  47 MODULE_PARM_DESC(qp_table_size, "QP table size");
  48
  49 unsigned int ib_ipath_lkey_table_size = 12;
  50 module_param_named(lkey_table_size, ib_ipath_lkey_table_size, uint,
  51                    S_IRUGO);
  52 MODULE_PARM_DESC(lkey_table_size,
  53                  "LKEY table size in bits (2^n, 1 <= n <= 23)");
  54
  55 unsigned int ib_ipath_debug;    /* debug mask */
  56 module_param_named(debug, ib_ipath_debug, uint, S_IWUSR | S_IRUGO);
  57 MODULE_PARM_DESC(debug, "Verbs debug mask");
  58
  59 static unsigned int ib_ipath_max_pds = 0xFFFF;
  60 module_param_named(max_pds, ib_ipath_max_pds, uint, S_IWUSR | S_IRUGO);
  61 MODULE_PARM_DESC(max_pds,
  62                  "Maximum number of protection domains to support");
  63
  64 static unsigned int ib_ipath_max_ahs = 0xFFFF;
  65 module_param_named(max_ahs, ib_ipath_max_ahs, uint, S_IWUSR | S_IRUGO);
  66 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
  67
  68 unsigned int ib_ipath_max_cqes = 0x2FFFF;
  69 module_param_named(max_cqes, ib_ipath_max_cqes, uint, S_IWUSR | S_IRUGO);
  70 MODULE_PARM_DESC(max_cqes,
  71                  "Maximum number of completion queue entries to support");
  72
  73 unsigned int ib_ipath_max_cqs = 0x1FFFF;
  74 module_param_named(max_cqs, ib_ipath_max_cqs, uint, S_IWUSR | S_IRUGO);
  75 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
  76
  77 unsigned int ib_ipath_max_qp_wrs = 0x3FFF;
  78 module_param_named(max_qp_wrs, ib_ipath_max_qp_wrs, uint,
  79                    S_IWUSR | S_IRUGO);
  80 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
  81
  82 unsigned int ib_ipath_max_sges = 0x60;
  83 module_param_named(max_sges, ib_ipath_max_sges, uint, S_IWUSR | S_IRUGO);
  84 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
  85
  86 unsigned int ib_ipath_max_mcast_grps = 16384;
  87 module_param_named(max_mcast_grps, ib_ipath_max_mcast_grps, uint,
  88                    S_IWUSR | S_IRUGO);
  89 MODULE_PARM_DESC(max_mcast_grps,
  90                  "Maximum number of multicast groups to support");
  91
  92 unsigned int ib_ipath_max_mcast_qp_attached = 16;
  93 module_param_named(max_mcast_qp_attached, ib_ipath_max_mcast_qp_attached,
  94                    uint, S_IWUSR | S_IRUGO);
  95 MODULE_PARM_DESC(max_mcast_qp_attached,
  96                  "Maximum number of attached QPs to support");
  97
  98 unsigned int ib_ipath_max_srqs = 1024;
  99 module_param_named(max_srqs, ib_ipath_max_srqs, uint, S_IWUSR | S_IRUGO);
 100 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
 101
 102 unsigned int ib_ipath_max_srq_sges = 128;
 103 module_param_named(max_srq_sges, ib_ipath_max_srq_sges,
 104                    uint, S_IWUSR | S_IRUGO);
 105 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
 106
 107 unsigned int ib_ipath_max_srq_wrs = 0x1FFFF;
 108 module_param_named(max_srq_wrs, ib_ipath_max_srq_wrs,
 109                    uint, S_IWUSR | S_IRUGO);
 110 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
 111
 112 MODULE_LICENSE("GPL");
 113 MODULE_AUTHOR("QLogic <support@pathscale.com>");
 114 MODULE_DESCRIPTION("QLogic InfiniPath driver");
 115
 116 const int ib_ipath_state_ops[IB_QPS_ERR + 1] = {
 117         [IB_QPS_RESET] = 0,
 118         [IB_QPS_INIT] = IPATH_POST_RECV_OK,
 119         [IB_QPS_RTR] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK,
 120         [IB_QPS_RTS] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
 121             IPATH_POST_SEND_OK | IPATH_PROCESS_SEND_OK,
 122         [IB_QPS_SQD] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
 123             IPATH_POST_SEND_OK,
 124         [IB_QPS_SQE] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK,
 125         [IB_QPS_ERR] = 0,
 126 };
 127
 128 /*
 129  * Translate ib_wr_opcode into ib_wc_opcode.
 130  */
 131 const enum ib_wc_opcode ib_ipath_wc_opcode[] = {
 132         [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
 133         [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
 134         [IB_WR_SEND] = IB_WC_SEND,
 135         [IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
 136         [IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
 137         [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
 138         [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD
 139 };
 140
 141 /*
 142  * System image GUID.
 143  */
 144 static __be64 sys_image_guid;
 145
 146 /**
 147  * ipath_copy_sge - copy data to SGE memory
 148  * @ss: the SGE state
 149  * @data: the data to copy
 150  * @length: the length of the data
 151  */
 152 void ipath_copy_sge(struct ipath_sge_state *ss, void *data, u32 length)
 153 {
 154         struct ipath_sge *sge = &ss->sge;
 155
 156         while (length) {
 157                 u32 len = sge->length;
 158
 159                 BUG_ON(len == 0);
 160                 if (len > length)
 161                         len = length;
 162                 memcpy(sge->vaddr, data, len);
 163                 sge->vaddr += len;
 164                 sge->length -= len;
 165                 sge->sge_length -= len;
 166                 if (sge->sge_length == 0) {
 167                         if (--ss->num_sge)
 168                                 *sge = *ss->sg_list++;
 169                 } else if (sge->length == 0 && sge->mr != NULL) {
 170                         if (++sge->n >= IPATH_SEGSZ) {
 171                                 if (++sge->m >= sge->mr->mapsz)
 172                                         break;
 173                                 sge->n = 0;
 174                         }
 175                         sge->vaddr =
 176                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 177                         sge->length =
 178                                 sge->mr->map[sge->m]->segs[sge->n].length;
 179                 }
 180                 data += len;
 181                 length -= len;
 182         }
 183 }
 184
 185 /**
 186  * ipath_skip_sge - skip over SGE memory - XXX almost dup of prev func
 187  * @ss: the SGE state
 188  * @length: the number of bytes to skip
 189  */
 190 void ipath_skip_sge(struct ipath_sge_state *ss, u32 length)
 191 {
 192         struct ipath_sge *sge = &ss->sge;
 193
 194         while (length) {
 195                 u32 len = sge->length;
 196
 197                 BUG_ON(len == 0);
 198                 if (len > length)
 199                         len = length;
 200                 sge->vaddr += len;
 201                 sge->length -= len;
 202                 sge->sge_length -= len;
 203                 if (sge->sge_length == 0) {
 204                         if (--ss->num_sge)
 205                                 *sge = *ss->sg_list++;
 206                 } else if (sge->length == 0 && sge->mr != NULL) {
 207                         if (++sge->n >= IPATH_SEGSZ) {
 208                                 if (++sge->m >= sge->mr->mapsz)
 209                                         break;
 210                                 sge->n = 0;
 211                         }
 212                         sge->vaddr =
 213                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 214                         sge->length =
 215                                 sge->mr->map[sge->m]->segs[sge->n].length;
 216                 }
 217                 length -= len;
 218         }
 219 }
 220
 221 /**
 222  * ipath_post_send - post a send on a QP
 223  * @ibqp: the QP to post the send on
 224  * @wr: the list of work requests to post
 225  * @bad_wr: the first bad WR is put here
 226  *
 227  * This may be called from interrupt context.
 228  */
 229 static int ipath_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 230                            struct ib_send_wr **bad_wr)
 231 {
 232         struct ipath_qp *qp = to_iqp(ibqp);
 233         int err = 0;
 234
 235         /* Check that state is OK to post send. */
 236         if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_SEND_OK)) {
 237                 *bad_wr = wr;
 238                 err = -EINVAL;
 239                 goto bail;
 240         }
 241
 242         for (; wr; wr = wr->next) {
 243                 switch (qp->ibqp.qp_type) {
 244                 case IB_QPT_UC:
 245                 case IB_QPT_RC:
 246                         err = ipath_post_ruc_send(qp, wr);
 247                         break;
 248
 249                 case IB_QPT_SMI:
 250                 case IB_QPT_GSI:
 251                 case IB_QPT_UD:
 252                         err = ipath_post_ud_send(qp, wr);
 253                         break;
 254
 255                 default:
 256                         err = -EINVAL;
 257                 }
 258                 if (err) {
 259                         *bad_wr = wr;
 260                         break;
 261                 }
 262         }
 263
 264 bail:
 265         return err;
 266 }
 267
 268 /**
 269  * ipath_post_receive - post a receive on a QP
 270  * @ibqp: the QP to post the receive on
 271  * @wr: the WR to post
 272  * @bad_wr: the first bad WR is put here
 273  *
 274  * This may be called from interrupt context.
 275  */
 276 static int ipath_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 277                               struct ib_recv_wr **bad_wr)
 278 {
 279         struct ipath_qp *qp = to_iqp(ibqp);
 280         unsigned long flags;
 281         int ret;
 282
 283         /* Check that state is OK to post receive. */
 284         if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_RECV_OK)) {
 285                 *bad_wr = wr;
 286                 ret = -EINVAL;
 287                 goto bail;
 288         }
 289
 290         for (; wr; wr = wr->next) {
 291                 struct ipath_rwqe *wqe;
 292                 u32 next;
 293                 int i, j;
 294
 295                 if (wr->num_sge > qp->r_rq.max_sge) {
 296                         *bad_wr = wr;
 297                         ret = -ENOMEM;
 298                         goto bail;
 299                 }
 300
 301                 spin_lock_irqsave(&qp->r_rq.lock, flags);
 302                 next = qp->r_rq.head + 1;
 303                 if (next >= qp->r_rq.size)
 304                         next = 0;
 305                 if (next == qp->r_rq.tail) {
 306                         spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 307                         *bad_wr = wr;
 308                         ret = -ENOMEM;
 309                         goto bail;
 310                 }
 311
 312                 wqe = get_rwqe_ptr(&qp->r_rq, qp->r_rq.head);
 313                 wqe->wr_id = wr->wr_id;
 314                 wqe->sg_list[0].mr = NULL;
 315                 wqe->sg_list[0].vaddr = NULL;
 316                 wqe->sg_list[0].length = 0;
 317                 wqe->sg_list[0].sge_length = 0;
 318                 wqe->length = 0;
 319                 for (i = 0, j = 0; i < wr->num_sge; i++) {
 320                         /* Check LKEY */
 321                         if (to_ipd(qp->ibqp.pd)->user &&
 322                             wr->sg_list[i].lkey == 0) {
 323                                 spin_unlock_irqrestore(&qp->r_rq.lock,
 324                                                        flags);
 325                                 *bad_wr = wr;
 326                                 ret = -EINVAL;
 327                                 goto bail;
 328                         }
 329                         if (wr->sg_list[i].length == 0)
 330                                 continue;
 331                         if (!ipath_lkey_ok(
 332                                     &to_idev(qp->ibqp.device)->lk_table,
 333                                     &wqe->sg_list[j], &wr->sg_list[i],
 334                                     IB_ACCESS_LOCAL_WRITE)) {
 335                                 spin_unlock_irqrestore(&qp->r_rq.lock,
 336                                                        flags);
 337                                 *bad_wr = wr;
 338                                 ret = -EINVAL;
 339                                 goto bail;
 340                         }
 341                         wqe->length += wr->sg_list[i].length;
 342                         j++;
 343                 }
 344                 wqe->num_sge = j;
 345                 qp->r_rq.head = next;
 346                 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 347         }
 348         ret = 0;
 349
 350 bail:
 351         return ret;
 352 }
 353
 354 /**
 355  * ipath_qp_rcv - processing an incoming packet on a QP
 356  * @dev: the device the packet came on
 357  * @hdr: the packet header
 358  * @has_grh: true if the packet has a GRH
 359  * @data: the packet data
 360  * @tlen: the packet length
 361  * @qp: the QP the packet came on
 362  *
 363  * This is called from ipath_ib_rcv() to process an incoming packet
 364  * for the given QP.
 365  * Called at interrupt level.
 366  */
 367 static void ipath_qp_rcv(struct ipath_ibdev *dev,
 368                          struct ipath_ib_header *hdr, int has_grh,
 369                          void *data, u32 tlen, struct ipath_qp *qp)
 370 {
 371         /* Check for valid receive state. */
 372         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
 373                 dev->n_pkt_drops++;
 374                 return;
 375         }
 376
 377         switch (qp->ibqp.qp_type) {
 378         case IB_QPT_SMI:
 379         case IB_QPT_GSI:
 380         case IB_QPT_UD:
 381                 ipath_ud_rcv(dev, hdr, has_grh, data, tlen, qp);
 382                 break;
 383
 384         case IB_QPT_RC:
 385                 ipath_rc_rcv(dev, hdr, has_grh, data, tlen, qp);
 386                 break;
 387
 388         case IB_QPT_UC:
 389                 ipath_uc_rcv(dev, hdr, has_grh, data, tlen, qp);
 390                 break;
 391
 392         default:
 393                 break;
 394         }
 395 }
 396
 397 /**
 398  * ipath_ib_rcv - process and incoming packet
 399  * @arg: the device pointer
 400  * @rhdr: the header of the packet
 401  * @data: the packet data
 402  * @tlen: the packet length
 403  *
 404  * This is called from ipath_kreceive() to process an incoming packet at
 405  * interrupt level. Tlen is the length of the header + data + CRC in bytes.
 406  */
 407 static void ipath_ib_rcv(void *arg, void *rhdr, void *data, u32 tlen)
 408 {
 409         struct ipath_ibdev *dev = (struct ipath_ibdev *) arg;
 410         struct ipath_ib_header *hdr = rhdr;
 411         struct ipath_other_headers *ohdr;
 412         struct ipath_qp *qp;
 413         u32 qp_num;
 414         int lnh;
 415         u8 opcode;
 416         u16 lid;
 417
 418         if (unlikely(dev == NULL))
 419                 goto bail;
 420
 421         if (unlikely(tlen < 24)) {      /* LRH+BTH+CRC */
 422                 dev->rcv_errors++;
 423                 goto bail;
 424         }
 425
 426         /* Check for a valid destination LID (see ch. 7.11.1). */
 427         lid = be16_to_cpu(hdr->lrh[1]);
 428         if (lid < IPATH_MULTICAST_LID_BASE) {
 429                 lid &= ~((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
 430                 if (unlikely(lid != ipath_layer_get_lid(dev->dd))) {
 431                         dev->rcv_errors++;
 432                         goto bail;
 433                 }
 434         }
 435
 436         /* Check for GRH */
 437         lnh = be16_to_cpu(hdr->lrh[0]) & 3;
 438         if (lnh == IPATH_LRH_BTH)
 439                 ohdr = &hdr->u.oth;
 440         else if (lnh == IPATH_LRH_GRH)
 441                 ohdr = &hdr->u.l.oth;
 442         else {
 443                 dev->rcv_errors++;
 444                 goto bail;
 445         }
 446
 447         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
 448         dev->opstats[opcode].n_bytes += tlen;
 449         dev->opstats[opcode].n_packets++;
 450
 451         /* Get the destination QP number. */
 452         qp_num = be32_to_cpu(ohdr->bth[1]) & IPATH_QPN_MASK;
 453         if (qp_num == IPATH_MULTICAST_QPN) {
 454                 struct ipath_mcast *mcast;
 455                 struct ipath_mcast_qp *p;
 456
 457                 mcast = ipath_mcast_find(&hdr->u.l.grh.dgid);
 458                 if (mcast == NULL) {
 459                         dev->n_pkt_drops++;
 460                         goto bail;
 461                 }
 462                 dev->n_multicast_rcv++;
 463                 list_for_each_entry_rcu(p, &mcast->qp_list, list)
 464                         ipath_qp_rcv(dev, hdr, lnh == IPATH_LRH_GRH, data,
 465                                      tlen, p->qp);
 466                 /*
 467                  * Notify ipath_multicast_detach() if it is waiting for us
 468                  * to finish.
 469                  */
 470                 if (atomic_dec_return(&mcast->refcount) <= 1)
 471                         wake_up(&mcast->wait);
 472         } else {
 473                 qp = ipath_lookup_qpn(&dev->qp_table, qp_num);
 474                 if (qp) {
 475                         dev->n_unicast_rcv++;
 476                         ipath_qp_rcv(dev, hdr, lnh == IPATH_LRH_GRH, data,
 477                                      tlen, qp);
 478                         /*
 479                          * Notify ipath_destroy_qp() if it is waiting
 480                          * for us to finish.
 481                          */
 482                         if (atomic_dec_and_test(&qp->refcount))
 483                                 wake_up(&qp->wait);
 484                 } else
 485                         dev->n_pkt_drops++;
 486         }
 487
 488 bail:;
 489 }
 490
 491 /**
 492  * ipath_ib_timer - verbs timer
 493  * @arg: the device pointer
 494  *
 495  * This is called from ipath_do_rcv_timer() at interrupt level to check for
 496  * QPs which need retransmits and to collect performance numbers.
 497  */
 498 static void ipath_ib_timer(void *arg)
 499 {
 500         struct ipath_ibdev *dev = (struct ipath_ibdev *) arg;
 501         struct ipath_qp *resend = NULL;
 502         struct list_head *last;
 503         struct ipath_qp *qp;
 504         unsigned long flags;
 505
 506         if (dev == NULL)
 507                 return;
 508
 509         spin_lock_irqsave(&dev->pending_lock, flags);
 510         /* Start filling the next pending queue. */
 511         if (++dev->pending_index >= ARRAY_SIZE(dev->pending))
 512                 dev->pending_index = 0;
 513         /* Save any requests still in the new queue, they have timed out. */
 514         last = &dev->pending[dev->pending_index];
 515         while (!list_empty(last)) {
 516                 qp = list_entry(last->next, struct ipath_qp, timerwait);
 517                 list_del_init(&qp->timerwait);
 518                 qp->timer_next = resend;
 519                 resend = qp;
 520                 atomic_inc(&qp->refcount);
 521         }
 522         last = &dev->rnrwait;
 523         if (!list_empty(last)) {
 524                 qp = list_entry(last->next, struct ipath_qp, timerwait);
 525                 if (--qp->s_rnr_timeout == 0) {
 526                         do {
 527                                 list_del_init(&qp->timerwait);
 528                                 tasklet_hi_schedule(&qp->s_task);
 529                                 if (list_empty(last))
 530                                         break;
 531                                 qp = list_entry(last->next, struct ipath_qp,
 532                                                 timerwait);
 533                         } while (qp->s_rnr_timeout == 0);
 534                 }
 535         }
 536         /*
 537          * We should only be in the started state if pma_sample_start != 0
 538          */
 539         if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_STARTED &&
 540             --dev->pma_sample_start == 0) {
 541                 dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_RUNNING;
 542                 ipath_layer_snapshot_counters(dev->dd, &dev->ipath_sword,
 543                                               &dev->ipath_rword,
 544                                               &dev->ipath_spkts,
 545                                               &dev->ipath_rpkts,
 546                                               &dev->ipath_xmit_wait);
 547         }
 548         if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_RUNNING) {
 549                 if (dev->pma_sample_interval == 0) {
 550                         u64 ta, tb, tc, td, te;
 551
 552                         dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_DONE;
 553                         ipath_layer_snapshot_counters(dev->dd, &ta, &tb,
 554                                                       &tc, &td, &te);
 555
 556                         dev->ipath_sword = ta - dev->ipath_sword;
 557                         dev->ipath_rword = tb - dev->ipath_rword;
 558                         dev->ipath_spkts = tc - dev->ipath_spkts;
 559                         dev->ipath_rpkts = td - dev->ipath_rpkts;
 560                         dev->ipath_xmit_wait = te - dev->ipath_xmit_wait;
 561                 }
 562                 else
 563                         dev->pma_sample_interval--;
 564         }
 565         spin_unlock_irqrestore(&dev->pending_lock, flags);
 566
 567         /* XXX What if timer fires again while this is running? */
 568         for (qp = resend; qp != NULL; qp = qp->timer_next) {
 569                 struct ib_wc wc;
 570
 571                 spin_lock_irqsave(&qp->s_lock, flags);
 572                 if (qp->s_last != qp->s_tail && qp->state == IB_QPS_RTS) {
 573                         dev->n_timeouts++;
 574                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
 575                 }
 576                 spin_unlock_irqrestore(&qp->s_lock, flags);
 577
 578                 /* Notify ipath_destroy_qp() if it is waiting. */
 579                 if (atomic_dec_and_test(&qp->refcount))
 580                         wake_up(&qp->wait);
 581         }
 582 }
 583
 584 /**
 585  * ipath_ib_piobufavail - callback when a PIO buffer is available
 586  * @arg: the device pointer
 587  *
 588  * This is called from ipath_intr() at interrupt level when a PIO buffer is
 589  * available after ipath_verbs_send() returned an error that no buffers were
 590  * available.  Return 1 if we consumed all the PIO buffers and we still have
 591  * QPs waiting for buffers (for now, just do a tasklet_hi_schedule and
 592  * return zero).
 593  */
 594 static int ipath_ib_piobufavail(void *arg)
 595 {
 596         struct ipath_ibdev *dev = (struct ipath_ibdev *) arg;
 597         struct ipath_qp *qp;
 598         unsigned long flags;
 599
 600         if (dev == NULL)
 601                 goto bail;
 602
 603         spin_lock_irqsave(&dev->pending_lock, flags);
 604         while (!list_empty(&dev->piowait)) {
 605                 qp = list_entry(dev->piowait.next, struct ipath_qp,
 606                                 piowait);
 607                 list_del_init(&qp->piowait);
 608                 tasklet_hi_schedule(&qp->s_task);
 609         }
 610         spin_unlock_irqrestore(&dev->pending_lock, flags);
 611
 612 bail:
 613         return 0;
 614 }
 615
 616 static int ipath_query_device(struct ib_device *ibdev,
 617                               struct ib_device_attr *props)
 618 {
 619         struct ipath_ibdev *dev = to_idev(ibdev);
 620
 621         memset(props, 0, sizeof(*props));
 622
 623         props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
 624                 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
 625                 IB_DEVICE_SYS_IMAGE_GUID;
 626         props->page_size_cap = PAGE_SIZE;
 627         props->vendor_id = ipath_layer_get_vendorid(dev->dd);
 628         props->vendor_part_id = ipath_layer_get_deviceid(dev->dd);
 629         props->hw_ver = ipath_layer_get_pcirev(dev->dd);
 630
 631         props->sys_image_guid = dev->sys_image_guid;
 632
 633         props->max_mr_size = ~0ull;
 634         props->max_qp = dev->qp_table.max;
 635         props->max_qp_wr = ib_ipath_max_qp_wrs;
 636         props->max_sge = ib_ipath_max_sges;
 637         props->max_cq = ib_ipath_max_cqs;
 638         props->max_ah = ib_ipath_max_ahs;
 639         props->max_cqe = ib_ipath_max_cqes;
 640         props->max_mr = dev->lk_table.max;
 641         props->max_pd = ib_ipath_max_pds;
 642         props->max_qp_rd_atom = 1;
 643         props->max_qp_init_rd_atom = 1;
 644         /* props->max_res_rd_atom */
 645         props->max_srq = ib_ipath_max_srqs;
 646         props->max_srq_wr = ib_ipath_max_srq_wrs;
 647         props->max_srq_sge = ib_ipath_max_srq_sges;
 648         /* props->local_ca_ack_delay */
 649         props->atomic_cap = IB_ATOMIC_HCA;
 650         props->max_pkeys = ipath_layer_get_npkeys(dev->dd);
 651         props->max_mcast_grp = ib_ipath_max_mcast_grps;
 652         props->max_mcast_qp_attach = ib_ipath_max_mcast_qp_attached;
 653         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
 654                 props->max_mcast_grp;
 655
 656         return 0;
 657 }
 658
 659 const u8 ipath_cvt_physportstate[16] = {
 660         [INFINIPATH_IBCS_LT_STATE_DISABLED] = 3,
 661         [INFINIPATH_IBCS_LT_STATE_LINKUP] = 5,
 662         [INFINIPATH_IBCS_LT_STATE_POLLACTIVE] = 2,
 663         [INFINIPATH_IBCS_LT_STATE_POLLQUIET] = 2,
 664         [INFINIPATH_IBCS_LT_STATE_SLEEPDELAY] = 1,
 665         [INFINIPATH_IBCS_LT_STATE_SLEEPQUIET] = 1,
 666         [INFINIPATH_IBCS_LT_STATE_CFGDEBOUNCE] = 4,
 667         [INFINIPATH_IBCS_LT_STATE_CFGRCVFCFG] = 4,
 668         [INFINIPATH_IBCS_LT_STATE_CFGWAITRMT] = 4,
 669         [INFINIPATH_IBCS_LT_STATE_CFGIDLE] = 4,
 670         [INFINIPATH_IBCS_LT_STATE_RECOVERRETRAIN] = 6,
 671         [INFINIPATH_IBCS_LT_STATE_RECOVERWAITRMT] = 6,
 672         [INFINIPATH_IBCS_LT_STATE_RECOVERIDLE] = 6,
 673 };
 674
 675 static int ipath_query_port(struct ib_device *ibdev,
 676                             u8 port, struct ib_port_attr *props)
 677 {
 678         struct ipath_ibdev *dev = to_idev(ibdev);
 679         enum ib_mtu mtu;
 680         u16 lid = ipath_layer_get_lid(dev->dd);
 681         u64 ibcstat;
 682
 683         memset(props, 0, sizeof(*props));
 684         props->lid = lid ? lid : __constant_be16_to_cpu(IB_LID_PERMISSIVE);
 685         props->lmc = dev->mkeyprot_resv_lmc & 7;
 686         props->sm_lid = dev->sm_lid;
 687         props->sm_sl = dev->sm_sl;
 688         ibcstat = ipath_layer_get_lastibcstat(dev->dd);
 689         props->state = ((ibcstat >> 4) & 0x3) + 1;
 690         /* See phys_state_show() */
 691         props->phys_state = ipath_cvt_physportstate[
 692                 ipath_layer_get_lastibcstat(dev->dd) & 0xf];
 693         props->port_cap_flags = dev->port_cap_flags;
 694         props->gid_tbl_len = 1;
 695         props->max_msg_sz = 0x80000000;
 696         props->pkey_tbl_len = ipath_layer_get_npkeys(dev->dd);
 697         props->bad_pkey_cntr = ipath_layer_get_cr_errpkey(dev->dd) -
 698                 dev->z_pkey_violations;
 699         props->qkey_viol_cntr = dev->qkey_violations;
 700         props->active_width = IB_WIDTH_4X;
 701         /* See rate_show() */
 702         props->active_speed = 1;        /* Regular 10Mbs speed. */
 703         props->max_vl_num = 1;          /* VLCap = VL0 */
 704         props->init_type_reply = 0;
 705
 706         props->max_mtu = IB_MTU_4096;
 707         switch (ipath_layer_get_ibmtu(dev->dd)) {
 708         case 4096:
 709                 mtu = IB_MTU_4096;
 710                 break;
 711         case 2048:
 712                 mtu = IB_MTU_2048;
 713                 break;
 714         case 1024:
 715                 mtu = IB_MTU_1024;
 716                 break;
 717         case 512:
 718                 mtu = IB_MTU_512;
 719                 break;
 720         case 256:
 721                 mtu = IB_MTU_256;
 722                 break;
 723         default:
 724                 mtu = IB_MTU_2048;
 725         }
 726         props->active_mtu = mtu;
 727         props->subnet_timeout = dev->subnet_timeout;
 728
 729         return 0;
 730 }
 731
 732 static int ipath_modify_device(struct ib_device *device,
 733                                int device_modify_mask,
 734                                struct ib_device_modify *device_modify)
 735 {
 736         int ret;
 737
 738         if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
 739                                    IB_DEVICE_MODIFY_NODE_DESC)) {
 740                 ret = -EOPNOTSUPP;
 741                 goto bail;
 742         }
 743
 744         if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC)
 745                 memcpy(device->node_desc, device_modify->node_desc, 64);
 746
 747         if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
 748                 to_idev(device)->sys_image_guid =
 749                         cpu_to_be64(device_modify->sys_image_guid);
 750
 751         ret = 0;
 752
 753 bail:
 754         return ret;
 755 }
 756
 757 static int ipath_modify_port(struct ib_device *ibdev,
 758                              u8 port, int port_modify_mask,
 759                              struct ib_port_modify *props)
 760 {
 761         struct ipath_ibdev *dev = to_idev(ibdev);
 762
 763         dev->port_cap_flags |= props->set_port_cap_mask;
 764         dev->port_cap_flags &= ~props->clr_port_cap_mask;
 765         if (port_modify_mask & IB_PORT_SHUTDOWN)
 766                 ipath_layer_set_linkstate(dev->dd, IPATH_IB_LINKDOWN);
 767         if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR)
 768                 dev->qkey_violations = 0;
 769         return 0;
 770 }
 771
 772 static int ipath_query_gid(struct ib_device *ibdev, u8 port,
 773                            int index, union ib_gid *gid)
 774 {
 775         struct ipath_ibdev *dev = to_idev(ibdev);
 776         int ret;
 777
 778         if (index >= 1) {
 779                 ret = -EINVAL;
 780                 goto bail;
 781         }
 782         gid->global.subnet_prefix = dev->gid_prefix;
 783         gid->global.interface_id = ipath_layer_get_guid(dev->dd);
 784
 785         ret = 0;
 786
 787 bail:
 788         return ret;
 789 }
 790
 791 static struct ib_pd *ipath_alloc_pd(struct ib_device *ibdev,
 792                                     struct ib_ucontext *context,
 793                                     struct ib_udata *udata)
 794 {
 795         struct ipath_ibdev *dev = to_idev(ibdev);
 796         struct ipath_pd *pd;
 797         struct ib_pd *ret;
 798
 799         /*
 800          * This is actually totally arbitrary.  Some correctness tests
 801          * assume there's a maximum number of PDs that can be allocated.
 802          * We don't actually have this limit, but we fail the test if
 803          * we allow allocations of more than we report for this value.
 804          */
 805
 806         if (dev->n_pds_allocated == ib_ipath_max_pds) {
 807                 ret = ERR_PTR(-ENOMEM);
 808                 goto bail;
 809         }
 810
 811         pd = kmalloc(sizeof *pd, GFP_KERNEL);
 812         if (!pd) {
 813                 ret = ERR_PTR(-ENOMEM);
 814                 goto bail;
 815         }
 816
 817         dev->n_pds_allocated++;
 818
 819         /* ib_alloc_pd() will initialize pd->ibpd. */
 820         pd->user = udata != NULL;
 821
 822         ret = &pd->ibpd;
 823
 824 bail:
 825         return ret;
 826 }
 827
 828 static int ipath_dealloc_pd(struct ib_pd *ibpd)
 829 {
 830         struct ipath_pd *pd = to_ipd(ibpd);
 831         struct ipath_ibdev *dev = to_idev(ibpd->device);
 832
 833         dev->n_pds_allocated--;
 834
 835         kfree(pd);
 836
 837         return 0;
 838 }
 839
 840 /**
 841  * ipath_create_ah - create an address handle
 842  * @pd: the protection domain
 843  * @ah_attr: the attributes of the AH
 844  *
 845  * This may be called from interrupt context.
 846  */
 847 static struct ib_ah *ipath_create_ah(struct ib_pd *pd,
 848                                      struct ib_ah_attr *ah_attr)
 849 {
 850         struct ipath_ah *ah;
 851         struct ib_ah *ret;
 852         struct ipath_ibdev *dev = to_idev(pd->device);
 853
 854         if (dev->n_ahs_allocated == ib_ipath_max_ahs) {
 855                 ret = ERR_PTR(-ENOMEM);
 856                 goto bail;
 857         }
 858
 859         /* A multicast address requires a GRH (see ch. 8.4.1). */
 860         if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
 861             ah_attr->dlid != IPATH_PERMISSIVE_LID &&
 862             !(ah_attr->ah_flags & IB_AH_GRH)) {
 863                 ret = ERR_PTR(-EINVAL);
 864                 goto bail;
 865         }
 866
 867         if (ah_attr->dlid == 0) {
 868                 ret = ERR_PTR(-EINVAL);
 869                 goto bail;
 870         }
 871
 872         if (ah_attr->port_num < 1 ||
 873             ah_attr->port_num > pd->device->phys_port_cnt) {
 874                 ret = ERR_PTR(-EINVAL);
 875                 goto bail;
 876         }
 877
 878         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
 879         if (!ah) {
 880                 ret = ERR_PTR(-ENOMEM);
 881                 goto bail;
 882         }
 883
 884         dev->n_ahs_allocated++;
 885
 886         /* ib_create_ah() will initialize ah->ibah. */
 887         ah->attr = *ah_attr;
 888
 889         ret = &ah->ibah;
 890
 891 bail:
 892         return ret;
 893 }
 894
 895 /**
 896  * ipath_destroy_ah - destroy an address handle
 897  * @ibah: the AH to destroy
 898  *
 899  * This may be called from interrupt context.
 900  */
 901 static int ipath_destroy_ah(struct ib_ah *ibah)
 902 {
 903         struct ipath_ibdev *dev = to_idev(ibah->device);
 904         struct ipath_ah *ah = to_iah(ibah);
 905
 906         dev->n_ahs_allocated--;
 907
 908         kfree(ah);
 909
 910         return 0;
 911 }
 912
 913 static int ipath_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
 914 {
 915         struct ipath_ah *ah = to_iah(ibah);
 916
 917         *ah_attr = ah->attr;
 918
 919         return 0;
 920 }
 921
 922 static int ipath_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
 923                             u16 *pkey)
 924 {
 925         struct ipath_ibdev *dev = to_idev(ibdev);
 926         int ret;
 927
 928         if (index >= ipath_layer_get_npkeys(dev->dd)) {
 929                 ret = -EINVAL;
 930                 goto bail;
 931         }
 932
 933         *pkey = ipath_layer_get_pkey(dev->dd, index);
 934         ret = 0;
 935
 936 bail:
 937         return ret;
 938 }
 939
 940
 941 /**
 942  * ipath_alloc_ucontext - allocate a ucontest
 943  * @ibdev: the infiniband device
 944  * @udata: not used by the InfiniPath driver
 945  */
 946
 947 static struct ib_ucontext *ipath_alloc_ucontext(struct ib_device *ibdev,
 948                                                 struct ib_udata *udata)
 949 {
 950         struct ipath_ucontext *context;
 951         struct ib_ucontext *ret;
 952
 953         context = kmalloc(sizeof *context, GFP_KERNEL);
 954         if (!context) {
 955                 ret = ERR_PTR(-ENOMEM);
 956                 goto bail;
 957         }
 958
 959         ret = &context->ibucontext;
 960
 961 bail:
 962         return ret;
 963 }
 964
 965 static int ipath_dealloc_ucontext(struct ib_ucontext *context)
 966 {
 967         kfree(to_iucontext(context));
 968         return 0;
 969 }
 970
 971 static int ipath_verbs_register_sysfs(struct ib_device *dev);
 972
 973 /**
 974  * ipath_register_ib_device - register our device with the infiniband core
 975  * @unit: the device number to register
 976  * @dd: the device data structure
 977  * Return the allocated ipath_ibdev pointer or NULL on error.
 978  */
 979 static void *ipath_register_ib_device(int unit, struct ipath_devdata *dd)
 980 {
 981         struct ipath_layer_counters cntrs;
 982         struct ipath_ibdev *idev;
 983         struct ib_device *dev;
 984         int ret;
 985
 986         idev = (struct ipath_ibdev *)ib_alloc_device(sizeof *idev);
 987         if (idev == NULL)
 988                 goto bail;
 989
 990         dev = &idev->ibdev;
 991
 992         /* Only need to initialize non-zero fields. */
 993         spin_lock_init(&idev->qp_table.lock);
 994         spin_lock_init(&idev->lk_table.lock);
 995         idev->sm_lid = __constant_be16_to_cpu(IB_LID_PERMISSIVE);
 996         /* Set the prefix to the default value (see ch. 4.1.1) */
 997         idev->gid_prefix = __constant_cpu_to_be64(0xfe80000000000000ULL);
 998
 999         ret = ipath_init_qp_table(idev, ib_ipath_qp_table_size);
1000         if (ret)
1001                 goto err_qp;
1002
1003         /*
1004          * The top ib_ipath_lkey_table_size bits are used to index the
1005          * table.  The lower 8 bits can be owned by the user (copied from
1006          * the LKEY).  The remaining bits act as a generation number or tag.
1007          */
1008         idev->lk_table.max = 1 << ib_ipath_lkey_table_size;
1009         idev->lk_table.table = kzalloc(idev->lk_table.max *
1010                                        sizeof(*idev->lk_table.table),
1011                                        GFP_KERNEL);
1012         if (idev->lk_table.table == NULL) {
1013                 ret = -ENOMEM;
1014                 goto err_lk;
1015         }
1016         spin_lock_init(&idev->pending_lock);
1017         INIT_LIST_HEAD(&idev->pending[0]);
1018         INIT_LIST_HEAD(&idev->pending[1]);
1019         INIT_LIST_HEAD(&idev->pending[2]);
1020         INIT_LIST_HEAD(&idev->piowait);
1021         INIT_LIST_HEAD(&idev->rnrwait);
1022         idev->pending_index = 0;
1023         idev->port_cap_flags =
1024                 IB_PORT_SYS_IMAGE_GUID_SUP | IB_PORT_CLIENT_REG_SUP;
1025         idev->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1026         idev->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1027         idev->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1028         idev->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1029         idev->pma_counter_select[5] = IB_PMA_PORT_XMIT_WAIT;
1030         idev->link_width_enabled = 3;   /* 1x or 4x */
1031
1032         /* Snapshot current HW counters to "clear" them. */
1033         ipath_layer_get_counters(dd, &cntrs);
1034         idev->z_symbol_error_counter = cntrs.symbol_error_counter;
1035         idev->z_link_error_recovery_counter =
1036                 cntrs.link_error_recovery_counter;
1037         idev->z_link_downed_counter = cntrs.link_downed_counter;
1038         idev->z_port_rcv_errors = cntrs.port_rcv_errors;
1039         idev->z_port_rcv_remphys_errors =
1040                 cntrs.port_rcv_remphys_errors;
1041         idev->z_port_xmit_discards = cntrs.port_xmit_discards;
1042         idev->z_port_xmit_data = cntrs.port_xmit_data;
1043         idev->z_port_rcv_data = cntrs.port_rcv_data;
1044         idev->z_port_xmit_packets = cntrs.port_xmit_packets;
1045         idev->z_port_rcv_packets = cntrs.port_rcv_packets;
1046         idev->z_local_link_integrity_errors =
1047                 cntrs.local_link_integrity_errors;
1048         idev->z_excessive_buffer_overrun_errors =
1049                 cntrs.excessive_buffer_overrun_errors;
1050
1051         /*
1052          * The system image GUID is supposed to be the same for all
1053          * IB HCAs in a single system but since there can be other
1054          * device types in the system, we can't be sure this is unique.
1055          */
1056         if (!sys_image_guid)
1057                 sys_image_guid = ipath_layer_get_guid(dd);
1058         idev->sys_image_guid = sys_image_guid;
1059         idev->ib_unit = unit;
1060         idev->dd = dd;
1061
1062         strlcpy(dev->name, "ipath%d", IB_DEVICE_NAME_MAX);
1063         dev->owner = THIS_MODULE;
1064         dev->node_guid = ipath_layer_get_guid(dd);
1065         dev->uverbs_abi_ver = IPATH_UVERBS_ABI_VERSION;
1066         dev->uverbs_cmd_mask =
1067                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
1068                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
1069                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
1070                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
1071                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
1072                 (1ull << IB_USER_VERBS_CMD_CREATE_AH)           |
1073                 (1ull << IB_USER_VERBS_CMD_DESTROY_AH)          |
1074                 (1ull << IB_USER_VERBS_CMD_QUERY_AH)            |
1075                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
1076                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
1077                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1078                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
1079                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
1080                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
1081                 (1ull << IB_USER_VERBS_CMD_POLL_CQ)             |
1082                 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)       |
1083                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
1084                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
1085                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
1086                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
1087                 (1ull << IB_USER_VERBS_CMD_POST_SEND)           |
1088                 (1ull << IB_USER_VERBS_CMD_POST_RECV)           |
1089                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
1090                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
1091                 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
1092                 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
1093                 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
1094                 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
1095                 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
1096         dev->node_type = IB_NODE_CA;
1097         dev->phys_port_cnt = 1;
1098         dev->dma_device = ipath_layer_get_device(dd);
1099         dev->class_dev.dev = dev->dma_device;
1100         dev->query_device = ipath_query_device;
1101         dev->modify_device = ipath_modify_device;
1102         dev->query_port = ipath_query_port;
1103         dev->modify_port = ipath_modify_port;
1104         dev->query_pkey = ipath_query_pkey;
1105         dev->query_gid = ipath_query_gid;
1106         dev->alloc_ucontext = ipath_alloc_ucontext;
1107         dev->dealloc_ucontext = ipath_dealloc_ucontext;
1108         dev->alloc_pd = ipath_alloc_pd;
1109         dev->dealloc_pd = ipath_dealloc_pd;
1110         dev->create_ah = ipath_create_ah;
1111         dev->destroy_ah = ipath_destroy_ah;
1112         dev->query_ah = ipath_query_ah;
1113         dev->create_srq = ipath_create_srq;
1114         dev->modify_srq = ipath_modify_srq;
1115         dev->query_srq = ipath_query_srq;
1116         dev->destroy_srq = ipath_destroy_srq;
1117         dev->create_qp = ipath_create_qp;
1118         dev->modify_qp = ipath_modify_qp;
1119         dev->query_qp = ipath_query_qp;
1120         dev->destroy_qp = ipath_destroy_qp;
1121         dev->post_send = ipath_post_send;
1122         dev->post_recv = ipath_post_receive;
1123         dev->post_srq_recv = ipath_post_srq_receive;
1124         dev->create_cq = ipath_create_cq;
1125         dev->destroy_cq = ipath_destroy_cq;
1126         dev->resize_cq = ipath_resize_cq;
1127         dev->poll_cq = ipath_poll_cq;
1128         dev->req_notify_cq = ipath_req_notify_cq;
1129         dev->get_dma_mr = ipath_get_dma_mr;
1130         dev->reg_phys_mr = ipath_reg_phys_mr;
1131         dev->reg_user_mr = ipath_reg_user_mr;
1132         dev->dereg_mr = ipath_dereg_mr;
1133         dev->alloc_fmr = ipath_alloc_fmr;
1134         dev->map_phys_fmr = ipath_map_phys_fmr;
1135         dev->unmap_fmr = ipath_unmap_fmr;
1136         dev->dealloc_fmr = ipath_dealloc_fmr;
1137         dev->attach_mcast = ipath_multicast_attach;
1138         dev->detach_mcast = ipath_multicast_detach;
1139         dev->process_mad = ipath_process_mad;
1140
1141         snprintf(dev->node_desc, sizeof(dev->node_desc),
1142                  IPATH_IDSTR " %s kernel_SMA", system_utsname.nodename);
1143
1144         ret = ib_register_device(dev);
1145         if (ret)
1146                 goto err_reg;
1147
1148         if (ipath_verbs_register_sysfs(dev))
1149                 goto err_class;
1150
1151         ipath_layer_enable_timer(dd);
1152
1153         goto bail;
1154
1155 err_class:
1156         ib_unregister_device(dev);
1157 err_reg:
1158         kfree(idev->lk_table.table);
1159 err_lk:
1160         kfree(idev->qp_table.table);
1161 err_qp:
1162         ib_dealloc_device(dev);
1163         _VERBS_ERROR("ib_ipath%d cannot register verbs (%d)!\n",
1164                      unit, -ret);
1165         idev = NULL;
1166
1167 bail:
1168         return idev;
1169 }
1170
1171 static void ipath_unregister_ib_device(void *arg)
1172 {
1173         struct ipath_ibdev *dev = (struct ipath_ibdev *) arg;
1174         struct ib_device *ibdev = &dev->ibdev;
1175
1176         ipath_layer_disable_timer(dev->dd);
1177
1178         ib_unregister_device(ibdev);
1179
1180         if (!list_empty(&dev->pending[0]) ||
1181             !list_empty(&dev->pending[1]) ||
1182             !list_empty(&dev->pending[2]))
1183                 _VERBS_ERROR("ipath%d pending list not empty!\n",
1184                              dev->ib_unit);
1185         if (!list_empty(&dev->piowait))
1186                 _VERBS_ERROR("ipath%d piowait list not empty!\n",
1187                              dev->ib_unit);
1188         if (!list_empty(&dev->rnrwait))
1189                 _VERBS_ERROR("ipath%d rnrwait list not empty!\n",
1190                              dev->ib_unit);
1191         if (!ipath_mcast_tree_empty())
1192                 _VERBS_ERROR("ipath%d multicast table memory leak!\n",
1193                              dev->ib_unit);
1194         /*
1195          * Note that ipath_unregister_ib_device() can be called before all
1196          * the QPs are destroyed!
1197          */
1198         ipath_free_all_qps(&dev->qp_table);
1199         kfree(dev->qp_table.table);
1200         kfree(dev->lk_table.table);
1201         ib_dealloc_device(ibdev);
1202 }
1203
1204 static int __init ipath_verbs_init(void)
1205 {
1206         return ipath_verbs_register(ipath_register_ib_device,
1207                                     ipath_unregister_ib_device,
1208                                     ipath_ib_piobufavail, ipath_ib_rcv,
1209                                     ipath_ib_timer);
1210 }
1211
1212 static void __exit ipath_verbs_cleanup(void)
1213 {
1214         ipath_verbs_unregister();
1215 }
1216
1217 static ssize_t show_rev(struct class_device *cdev, char *buf)
1218 {
1219         struct ipath_ibdev *dev =
1220                 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1221
1222         return sprintf(buf, "%x\n", ipath_layer_get_pcirev(dev->dd));
1223 }
1224
1225 static ssize_t show_hca(struct class_device *cdev, char *buf)
1226 {
1227         struct ipath_ibdev *dev =
1228                 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1229         int ret;
1230
1231         ret = ipath_layer_get_boardname(dev->dd, buf, 128);
1232         if (ret < 0)
1233                 goto bail;
1234         strcat(buf, "\n");
1235         ret = strlen(buf);
1236
1237 bail:
1238         return ret;
1239 }
1240
1241 static ssize_t show_stats(struct class_device *cdev, char *buf)
1242 {
1243         struct ipath_ibdev *dev =
1244                 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1245         int i;
1246         int len;
1247
1248         len = sprintf(buf,
1249                       "RC resends  %d\n"
1250                       "RC no QACK  %d\n"
1251                       "RC ACKs     %d\n"
1252                       "RC SEQ NAKs %d\n"
1253                       "RC RDMA seq %d\n"
1254                       "RC RNR NAKs %d\n"
1255                       "RC OTH NAKs %d\n"
1256                       "RC timeouts %d\n"
1257                       "RC RDMA dup %d\n"
1258                       "piobuf wait %d\n"
1259                       "no piobuf   %d\n"
1260                       "PKT drops   %d\n"
1261                       "WQE errs    %d\n",
1262                       dev->n_rc_resends, dev->n_rc_qacks, dev->n_rc_acks,
1263                       dev->n_seq_naks, dev->n_rdma_seq, dev->n_rnr_naks,
1264                       dev->n_other_naks, dev->n_timeouts,
1265                       dev->n_rdma_dup_busy, dev->n_piowait,
1266                       dev->n_no_piobuf, dev->n_pkt_drops, dev->n_wqe_errs);
1267         for (i = 0; i < ARRAY_SIZE(dev->opstats); i++) {
1268                 const struct ipath_opcode_stats *si = &dev->opstats[i];
1269
1270                 if (!si->n_packets && !si->n_bytes)
1271                         continue;
1272                 len += sprintf(buf + len, "%02x %llu/%llu\n", i,
1273                                (unsigned long long) si->n_packets,
1274                                (unsigned long long) si->n_bytes);
1275         }
1276         return len;
1277 }
1278
1279 static CLASS_DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
1280 static CLASS_DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
1281 static CLASS_DEVICE_ATTR(board_id, S_IRUGO, show_hca, NULL);
1282 static CLASS_DEVICE_ATTR(stats, S_IRUGO, show_stats, NULL);
1283
1284 static struct class_device_attribute *ipath_class_attributes[] = {
1285         &class_device_attr_hw_rev,
1286         &class_device_attr_hca_type,
1287         &class_device_attr_board_id,
1288         &class_device_attr_stats
1289 };
1290
1291 static int ipath_verbs_register_sysfs(struct ib_device *dev)
1292 {
1293         int i;
1294         int ret;
1295
1296         for (i = 0; i < ARRAY_SIZE(ipath_class_attributes); ++i)
1297                 if (class_device_create_file(&dev->class_dev,
1298                                              ipath_class_attributes[i])) {
1299                         ret = 1;
1300                         goto bail;
1301                 }
1302
1303         ret = 0;
1304
1305 bail:
1306         return ret;
1307 }
1308
1309 module_init(ipath_verbs_init);
1310 module_exit(ipath_verbs_cleanup);