drivers/infiniband/hw/ipath/ipath_verbs.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008 QLogic Corporation. 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/io.h>
  37 #include <linux/utsname.h>
  38
  39 #include "ipath_kernel.h"
  40 #include "ipath_verbs.h"
  41 #include "ipath_common.h"
  42
  43 static unsigned int ib_ipath_qp_table_size = 251;
  44 module_param_named(qp_table_size, ib_ipath_qp_table_size, uint, S_IRUGO);
  45 MODULE_PARM_DESC(qp_table_size, "QP table size");
  46
  47 unsigned int ib_ipath_lkey_table_size = 12;
  48 module_param_named(lkey_table_size, ib_ipath_lkey_table_size, uint,
  49                    S_IRUGO);
  50 MODULE_PARM_DESC(lkey_table_size,
  51                  "LKEY table size in bits (2^n, 1 <= n <= 23)");
  52
  53 static unsigned int ib_ipath_max_pds = 0xFFFF;
  54 module_param_named(max_pds, ib_ipath_max_pds, uint, S_IWUSR | S_IRUGO);
  55 MODULE_PARM_DESC(max_pds,
  56                  "Maximum number of protection domains to support");
  57
  58 static unsigned int ib_ipath_max_ahs = 0xFFFF;
  59 module_param_named(max_ahs, ib_ipath_max_ahs, uint, S_IWUSR | S_IRUGO);
  60 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
  61
  62 unsigned int ib_ipath_max_cqes = 0x2FFFF;
  63 module_param_named(max_cqes, ib_ipath_max_cqes, uint, S_IWUSR | S_IRUGO);
  64 MODULE_PARM_DESC(max_cqes,
  65                  "Maximum number of completion queue entries to support");
  66
  67 unsigned int ib_ipath_max_cqs = 0x1FFFF;
  68 module_param_named(max_cqs, ib_ipath_max_cqs, uint, S_IWUSR | S_IRUGO);
  69 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
  70
  71 unsigned int ib_ipath_max_qp_wrs = 0x3FFF;
  72 module_param_named(max_qp_wrs, ib_ipath_max_qp_wrs, uint,
  73                    S_IWUSR | S_IRUGO);
  74 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
  75
  76 unsigned int ib_ipath_max_qps = 16384;
  77 module_param_named(max_qps, ib_ipath_max_qps, uint, S_IWUSR | S_IRUGO);
  78 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
  79
  80 unsigned int ib_ipath_max_sges = 0x60;
  81 module_param_named(max_sges, ib_ipath_max_sges, uint, S_IWUSR | S_IRUGO);
  82 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
  83
  84 unsigned int ib_ipath_max_mcast_grps = 16384;
  85 module_param_named(max_mcast_grps, ib_ipath_max_mcast_grps, uint,
  86                    S_IWUSR | S_IRUGO);
  87 MODULE_PARM_DESC(max_mcast_grps,
  88                  "Maximum number of multicast groups to support");
  89
  90 unsigned int ib_ipath_max_mcast_qp_attached = 16;
  91 module_param_named(max_mcast_qp_attached, ib_ipath_max_mcast_qp_attached,
  92                    uint, S_IWUSR | S_IRUGO);
  93 MODULE_PARM_DESC(max_mcast_qp_attached,
  94                  "Maximum number of attached QPs to support");
  95
  96 unsigned int ib_ipath_max_srqs = 1024;
  97 module_param_named(max_srqs, ib_ipath_max_srqs, uint, S_IWUSR | S_IRUGO);
  98 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
  99
 100 unsigned int ib_ipath_max_srq_sges = 128;
 101 module_param_named(max_srq_sges, ib_ipath_max_srq_sges,
 102                    uint, S_IWUSR | S_IRUGO);
 103 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
 104
 105 unsigned int ib_ipath_max_srq_wrs = 0x1FFFF;
 106 module_param_named(max_srq_wrs, ib_ipath_max_srq_wrs,
 107                    uint, S_IWUSR | S_IRUGO);
 108 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
 109
 110 static unsigned int ib_ipath_disable_sma;
 111 module_param_named(disable_sma, ib_ipath_disable_sma, uint, S_IWUSR | S_IRUGO);
 112 MODULE_PARM_DESC(disable_sma, "Disable the SMA");
 113
 114 /*
 115  * Note that it is OK to post send work requests in the SQE and ERR
 116  * states; ipath_do_send() will process them and generate error
 117  * completions as per IB 1.2 C10-96.
 118  */
 119 const int ib_ipath_state_ops[IB_QPS_ERR + 1] = {
 120         [IB_QPS_RESET] = 0,
 121         [IB_QPS_INIT] = IPATH_POST_RECV_OK,
 122         [IB_QPS_RTR] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK,
 123         [IB_QPS_RTS] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
 124             IPATH_POST_SEND_OK | IPATH_PROCESS_SEND_OK |
 125             IPATH_PROCESS_NEXT_SEND_OK,
 126         [IB_QPS_SQD] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
 127             IPATH_POST_SEND_OK | IPATH_PROCESS_SEND_OK,
 128         [IB_QPS_SQE] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
 129             IPATH_POST_SEND_OK | IPATH_FLUSH_SEND,
 130         [IB_QPS_ERR] = IPATH_POST_RECV_OK | IPATH_FLUSH_RECV |
 131             IPATH_POST_SEND_OK | IPATH_FLUSH_SEND,
 132 };
 133
 134 struct ipath_ucontext {
 135         struct ib_ucontext ibucontext;
 136 };
 137
 138 static inline struct ipath_ucontext *to_iucontext(struct ib_ucontext
 139                                                   *ibucontext)
 140 {
 141         return container_of(ibucontext, struct ipath_ucontext, ibucontext);
 142 }
 143
 144 /*
 145  * Translate ib_wr_opcode into ib_wc_opcode.
 146  */
 147 const enum ib_wc_opcode ib_ipath_wc_opcode[] = {
 148         [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
 149         [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
 150         [IB_WR_SEND] = IB_WC_SEND,
 151         [IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
 152         [IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
 153         [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
 154         [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD
 155 };
 156
 157 /*
 158  * System image GUID.
 159  */
 160 static __be64 sys_image_guid;
 161
 162 /**
 163  * ipath_copy_sge - copy data to SGE memory
 164  * @ss: the SGE state
 165  * @data: the data to copy
 166  * @length: the length of the data
 167  */
 168 void ipath_copy_sge(struct ipath_sge_state *ss, void *data, u32 length)
 169 {
 170         struct ipath_sge *sge = &ss->sge;
 171
 172         while (length) {
 173                 u32 len = sge->length;
 174
 175                 if (len > length)
 176                         len = length;
 177                 if (len > sge->sge_length)
 178                         len = sge->sge_length;
 179                 BUG_ON(len == 0);
 180                 memcpy(sge->vaddr, data, len);
 181                 sge->vaddr += len;
 182                 sge->length -= len;
 183                 sge->sge_length -= len;
 184                 if (sge->sge_length == 0) {
 185                         if (--ss->num_sge)
 186                                 *sge = *ss->sg_list++;
 187                 } else if (sge->length == 0 && sge->mr != NULL) {
 188                         if (++sge->n >= IPATH_SEGSZ) {
 189                                 if (++sge->m >= sge->mr->mapsz)
 190                                         break;
 191                                 sge->n = 0;
 192                         }
 193                         sge->vaddr =
 194                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 195                         sge->length =
 196                                 sge->mr->map[sge->m]->segs[sge->n].length;
 197                 }
 198                 data += len;
 199                 length -= len;
 200         }
 201 }
 202
 203 /**
 204  * ipath_skip_sge - skip over SGE memory - XXX almost dup of prev func
 205  * @ss: the SGE state
 206  * @length: the number of bytes to skip
 207  */
 208 void ipath_skip_sge(struct ipath_sge_state *ss, u32 length)
 209 {
 210         struct ipath_sge *sge = &ss->sge;
 211
 212         while (length) {
 213                 u32 len = sge->length;
 214
 215                 if (len > length)
 216                         len = length;
 217                 if (len > sge->sge_length)
 218                         len = sge->sge_length;
 219                 BUG_ON(len == 0);
 220                 sge->vaddr += len;
 221                 sge->length -= len;
 222                 sge->sge_length -= len;
 223                 if (sge->sge_length == 0) {
 224                         if (--ss->num_sge)
 225                                 *sge = *ss->sg_list++;
 226                 } else if (sge->length == 0 && sge->mr != NULL) {
 227                         if (++sge->n >= IPATH_SEGSZ) {
 228                                 if (++sge->m >= sge->mr->mapsz)
 229                                         break;
 230                                 sge->n = 0;
 231                         }
 232                         sge->vaddr =
 233                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 234                         sge->length =
 235                                 sge->mr->map[sge->m]->segs[sge->n].length;
 236                 }
 237                 length -= len;
 238         }
 239 }
 240
 241 /*
 242  * Count the number of DMA descriptors needed to send length bytes of data.
 243  * Don't modify the ipath_sge_state to get the count.
 244  * Return zero if any of the segments is not aligned.
 245  */
 246 static u32 ipath_count_sge(struct ipath_sge_state *ss, u32 length)
 247 {
 248         struct ipath_sge *sg_list = ss->sg_list;
 249         struct ipath_sge sge = ss->sge;
 250         u8 num_sge = ss->num_sge;
 251         u32 ndesc = 1;  /* count the header */
 252
 253         while (length) {
 254                 u32 len = sge.length;
 255
 256                 if (len > length)
 257                         len = length;
 258                 if (len > sge.sge_length)
 259                         len = sge.sge_length;
 260                 BUG_ON(len == 0);
 261                 if (((long) sge.vaddr & (sizeof(u32) - 1)) ||
 262                     (len != length && (len & (sizeof(u32) - 1)))) {
 263                         ndesc = 0;
 264                         break;
 265                 }
 266                 ndesc++;
 267                 sge.vaddr += len;
 268                 sge.length -= len;
 269                 sge.sge_length -= len;
 270                 if (sge.sge_length == 0) {
 271                         if (--num_sge)
 272                                 sge = *sg_list++;
 273                 } else if (sge.length == 0 && sge.mr != NULL) {
 274                         if (++sge.n >= IPATH_SEGSZ) {
 275                                 if (++sge.m >= sge.mr->mapsz)
 276                                         break;
 277                                 sge.n = 0;
 278                         }
 279                         sge.vaddr =
 280                                 sge.mr->map[sge.m]->segs[sge.n].vaddr;
 281                         sge.length =
 282                                 sge.mr->map[sge.m]->segs[sge.n].length;
 283                 }
 284                 length -= len;
 285         }
 286         return ndesc;
 287 }
 288
 289 /*
 290  * Copy from the SGEs to the data buffer.
 291  */
 292 static void ipath_copy_from_sge(void *data, struct ipath_sge_state *ss,
 293                                 u32 length)
 294 {
 295         struct ipath_sge *sge = &ss->sge;
 296
 297         while (length) {
 298                 u32 len = sge->length;
 299
 300                 if (len > length)
 301                         len = length;
 302                 if (len > sge->sge_length)
 303                         len = sge->sge_length;
 304                 BUG_ON(len == 0);
 305                 memcpy(data, sge->vaddr, len);
 306                 sge->vaddr += len;
 307                 sge->length -= len;
 308                 sge->sge_length -= len;
 309                 if (sge->sge_length == 0) {
 310                         if (--ss->num_sge)
 311                                 *sge = *ss->sg_list++;
 312                 } else if (sge->length == 0 && sge->mr != NULL) {
 313                         if (++sge->n >= IPATH_SEGSZ) {
 314                                 if (++sge->m >= sge->mr->mapsz)
 315                                         break;
 316                                 sge->n = 0;
 317                         }
 318                         sge->vaddr =
 319                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 320                         sge->length =
 321                                 sge->mr->map[sge->m]->segs[sge->n].length;
 322                 }
 323                 data += len;
 324                 length -= len;
 325         }
 326 }
 327
 328 /**
 329  * ipath_post_one_send - post one RC, UC, or UD send work request
 330  * @qp: the QP to post on
 331  * @wr: the work request to send
 332  */
 333 static int ipath_post_one_send(struct ipath_qp *qp, struct ib_send_wr *wr)
 334 {
 335         struct ipath_swqe *wqe;
 336         u32 next;
 337         int i;
 338         int j;
 339         int acc;
 340         int ret;
 341         unsigned long flags;
 342
 343         spin_lock_irqsave(&qp->s_lock, flags);
 344
 345         /* Check that state is OK to post send. */
 346         if (unlikely(!(ib_ipath_state_ops[qp->state] & IPATH_POST_SEND_OK)))
 347                 goto bail_inval;
 348
 349         /* IB spec says that num_sge == 0 is OK. */
 350         if (wr->num_sge > qp->s_max_sge)
 351                 goto bail_inval;
 352
 353         /*
 354          * Don't allow RDMA reads or atomic operations on UC or
 355          * undefined operations.
 356          * Make sure buffer is large enough to hold the result for atomics.
 357          */
 358         if (qp->ibqp.qp_type == IB_QPT_UC) {
 359                 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ)
 360                         goto bail_inval;
 361         } else if (qp->ibqp.qp_type == IB_QPT_UD) {
 362                 /* Check UD opcode */
 363                 if (wr->opcode != IB_WR_SEND &&
 364                     wr->opcode != IB_WR_SEND_WITH_IMM)
 365                         goto bail_inval;
 366                 /* Check UD destination address PD */
 367                 if (qp->ibqp.pd != wr->wr.ud.ah->pd)
 368                         goto bail_inval;
 369         } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD)
 370                 goto bail_inval;
 371         else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
 372                    (wr->num_sge == 0 ||
 373                     wr->sg_list[0].length < sizeof(u64) ||
 374                     wr->sg_list[0].addr & (sizeof(u64) - 1)))
 375                 goto bail_inval;
 376         else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic)
 377                 goto bail_inval;
 378
 379         next = qp->s_head + 1;
 380         if (next >= qp->s_size)
 381                 next = 0;
 382         if (next == qp->s_last) {
 383                 ret = -ENOMEM;
 384                 goto bail;
 385         }
 386
 387         wqe = get_swqe_ptr(qp, qp->s_head);
 388         wqe->wr = *wr;
 389         wqe->length = 0;
 390         if (wr->num_sge) {
 391                 acc = wr->opcode >= IB_WR_RDMA_READ ?
 392                         IB_ACCESS_LOCAL_WRITE : 0;
 393                 for (i = 0, j = 0; i < wr->num_sge; i++) {
 394                         u32 length = wr->sg_list[i].length;
 395                         int ok;
 396
 397                         if (length == 0)
 398                                 continue;
 399                         ok = ipath_lkey_ok(qp, &wqe->sg_list[j],
 400                                            &wr->sg_list[i], acc);
 401                         if (!ok)
 402                                 goto bail_inval;
 403                         wqe->length += length;
 404                         j++;
 405                 }
 406                 wqe->wr.num_sge = j;
 407         }
 408         if (qp->ibqp.qp_type == IB_QPT_UC ||
 409             qp->ibqp.qp_type == IB_QPT_RC) {
 410                 if (wqe->length > 0x80000000U)
 411                         goto bail_inval;
 412         } else if (wqe->length > to_idev(qp->ibqp.device)->dd->ipath_ibmtu)
 413                 goto bail_inval;
 414         wqe->ssn = qp->s_ssn++;
 415         qp->s_head = next;
 416
 417         ret = 0;
 418         goto bail;
 419
 420 bail_inval:
 421         ret = -EINVAL;
 422 bail:
 423         spin_unlock_irqrestore(&qp->s_lock, flags);
 424         return ret;
 425 }
 426
 427 /**
 428  * ipath_post_send - post a send on a QP
 429  * @ibqp: the QP to post the send on
 430  * @wr: the list of work requests to post
 431  * @bad_wr: the first bad WR is put here
 432  *
 433  * This may be called from interrupt context.
 434  */
 435 static int ipath_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 436                            struct ib_send_wr **bad_wr)
 437 {
 438         struct ipath_qp *qp = to_iqp(ibqp);
 439         int err = 0;
 440
 441         for (; wr; wr = wr->next) {
 442                 err = ipath_post_one_send(qp, wr);
 443                 if (err) {
 444                         *bad_wr = wr;
 445                         goto bail;
 446                 }
 447         }
 448
 449         /* Try to do the send work in the caller's context. */
 450         ipath_do_send((unsigned long) qp);
 451
 452 bail:
 453         return err;
 454 }
 455
 456 /**
 457  * ipath_post_receive - post a receive on a QP
 458  * @ibqp: the QP to post the receive on
 459  * @wr: the WR to post
 460  * @bad_wr: the first bad WR is put here
 461  *
 462  * This may be called from interrupt context.
 463  */
 464 static int ipath_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 465                               struct ib_recv_wr **bad_wr)
 466 {
 467         struct ipath_qp *qp = to_iqp(ibqp);
 468         struct ipath_rwq *wq = qp->r_rq.wq;
 469         unsigned long flags;
 470         int ret;
 471
 472         /* Check that state is OK to post receive. */
 473         if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_RECV_OK) || !wq) {
 474                 *bad_wr = wr;
 475                 ret = -EINVAL;
 476                 goto bail;
 477         }
 478
 479         for (; wr; wr = wr->next) {
 480                 struct ipath_rwqe *wqe;
 481                 u32 next;
 482                 int i;
 483
 484                 if ((unsigned) wr->num_sge > qp->r_rq.max_sge) {
 485                         *bad_wr = wr;
 486                         ret = -EINVAL;
 487                         goto bail;
 488                 }
 489
 490                 spin_lock_irqsave(&qp->r_rq.lock, flags);
 491                 next = wq->head + 1;
 492                 if (next >= qp->r_rq.size)
 493                         next = 0;
 494                 if (next == wq->tail) {
 495                         spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 496                         *bad_wr = wr;
 497                         ret = -ENOMEM;
 498                         goto bail;
 499                 }
 500
 501                 wqe = get_rwqe_ptr(&qp->r_rq, wq->head);
 502                 wqe->wr_id = wr->wr_id;
 503                 wqe->num_sge = wr->num_sge;
 504                 for (i = 0; i < wr->num_sge; i++)
 505                         wqe->sg_list[i] = wr->sg_list[i];
 506                 /* Make sure queue entry is written before the head index. */
 507                 smp_wmb();
 508                 wq->head = next;
 509                 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 510         }
 511         ret = 0;
 512
 513 bail:
 514         return ret;
 515 }
 516
 517 /**
 518  * ipath_qp_rcv - processing an incoming packet on a QP
 519  * @dev: the device the packet came on
 520  * @hdr: the packet header
 521  * @has_grh: true if the packet has a GRH
 522  * @data: the packet data
 523  * @tlen: the packet length
 524  * @qp: the QP the packet came on
 525  *
 526  * This is called from ipath_ib_rcv() to process an incoming packet
 527  * for the given QP.
 528  * Called at interrupt level.
 529  */
 530 static void ipath_qp_rcv(struct ipath_ibdev *dev,
 531                          struct ipath_ib_header *hdr, int has_grh,
 532                          void *data, u32 tlen, struct ipath_qp *qp)
 533 {
 534         /* Check for valid receive state. */
 535         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
 536                 dev->n_pkt_drops++;
 537                 return;
 538         }
 539
 540         switch (qp->ibqp.qp_type) {
 541         case IB_QPT_SMI:
 542         case IB_QPT_GSI:
 543                 if (ib_ipath_disable_sma)
 544                         break;
 545                 /* FALLTHROUGH */
 546         case IB_QPT_UD:
 547                 ipath_ud_rcv(dev, hdr, has_grh, data, tlen, qp);
 548                 break;
 549
 550         case IB_QPT_RC:
 551                 ipath_rc_rcv(dev, hdr, has_grh, data, tlen, qp);
 552                 break;
 553
 554         case IB_QPT_UC:
 555                 ipath_uc_rcv(dev, hdr, has_grh, data, tlen, qp);
 556                 break;
 557
 558         default:
 559                 break;
 560         }
 561 }
 562
 563 /**
 564  * ipath_ib_rcv - process an incoming packet
 565  * @arg: the device pointer
 566  * @rhdr: the header of the packet
 567  * @data: the packet data
 568  * @tlen: the packet length
 569  *
 570  * This is called from ipath_kreceive() to process an incoming packet at
 571  * interrupt level. Tlen is the length of the header + data + CRC in bytes.
 572  */
 573 void ipath_ib_rcv(struct ipath_ibdev *dev, void *rhdr, void *data,
 574                   u32 tlen)
 575 {
 576         struct ipath_ib_header *hdr = rhdr;
 577         struct ipath_other_headers *ohdr;
 578         struct ipath_qp *qp;
 579         u32 qp_num;
 580         int lnh;
 581         u8 opcode;
 582         u16 lid;
 583
 584         if (unlikely(dev == NULL))
 585                 goto bail;
 586
 587         if (unlikely(tlen < 24)) {      /* LRH+BTH+CRC */
 588                 dev->rcv_errors++;
 589                 goto bail;
 590         }
 591
 592         /* Check for a valid destination LID (see ch. 7.11.1). */
 593         lid = be16_to_cpu(hdr->lrh[1]);
 594         if (lid < IPATH_MULTICAST_LID_BASE) {
 595                 lid &= ~((1 << dev->dd->ipath_lmc) - 1);
 596                 if (unlikely(lid != dev->dd->ipath_lid)) {
 597                         dev->rcv_errors++;
 598                         goto bail;
 599                 }
 600         }
 601
 602         /* Check for GRH */
 603         lnh = be16_to_cpu(hdr->lrh[0]) & 3;
 604         if (lnh == IPATH_LRH_BTH)
 605                 ohdr = &hdr->u.oth;
 606         else if (lnh == IPATH_LRH_GRH)
 607                 ohdr = &hdr->u.l.oth;
 608         else {
 609                 dev->rcv_errors++;
 610                 goto bail;
 611         }
 612
 613         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
 614         dev->opstats[opcode].n_bytes += tlen;
 615         dev->opstats[opcode].n_packets++;
 616
 617         /* Get the destination QP number. */
 618         qp_num = be32_to_cpu(ohdr->bth[1]) & IPATH_QPN_MASK;
 619         if (qp_num == IPATH_MULTICAST_QPN) {
 620                 struct ipath_mcast *mcast;
 621                 struct ipath_mcast_qp *p;
 622
 623                 if (lnh != IPATH_LRH_GRH) {
 624                         dev->n_pkt_drops++;
 625                         goto bail;
 626                 }
 627                 mcast = ipath_mcast_find(&hdr->u.l.grh.dgid);
 628                 if (mcast == NULL) {
 629                         dev->n_pkt_drops++;
 630                         goto bail;
 631                 }
 632                 dev->n_multicast_rcv++;
 633                 list_for_each_entry_rcu(p, &mcast->qp_list, list)
 634                         ipath_qp_rcv(dev, hdr, 1, data, tlen, p->qp);
 635                 /*
 636                  * Notify ipath_multicast_detach() if it is waiting for us
 637                  * to finish.
 638                  */
 639                 if (atomic_dec_return(&mcast->refcount) <= 1)
 640                         wake_up(&mcast->wait);
 641         } else {
 642                 qp = ipath_lookup_qpn(&dev->qp_table, qp_num);
 643                 if (qp) {
 644                         dev->n_unicast_rcv++;
 645                         ipath_qp_rcv(dev, hdr, lnh == IPATH_LRH_GRH, data,
 646                                      tlen, qp);
 647                         /*
 648                          * Notify ipath_destroy_qp() if it is waiting
 649                          * for us to finish.
 650                          */
 651                         if (atomic_dec_and_test(&qp->refcount))
 652                                 wake_up(&qp->wait);
 653                 } else
 654                         dev->n_pkt_drops++;
 655         }
 656
 657 bail:;
 658 }
 659
 660 /**
 661  * ipath_ib_timer - verbs timer
 662  * @arg: the device pointer
 663  *
 664  * This is called from ipath_do_rcv_timer() at interrupt level to check for
 665  * QPs which need retransmits and to collect performance numbers.
 666  */
 667 static void ipath_ib_timer(struct ipath_ibdev *dev)
 668 {
 669         struct ipath_qp *resend = NULL;
 670         struct ipath_qp *rnr = NULL;
 671         struct list_head *last;
 672         struct ipath_qp *qp;
 673         unsigned long flags;
 674
 675         if (dev == NULL)
 676                 return;
 677
 678         spin_lock_irqsave(&dev->pending_lock, flags);
 679         /* Start filling the next pending queue. */
 680         if (++dev->pending_index >= ARRAY_SIZE(dev->pending))
 681                 dev->pending_index = 0;
 682         /* Save any requests still in the new queue, they have timed out. */
 683         last = &dev->pending[dev->pending_index];
 684         while (!list_empty(last)) {
 685                 qp = list_entry(last->next, struct ipath_qp, timerwait);
 686                 list_del_init(&qp->timerwait);
 687                 qp->timer_next = resend;
 688                 resend = qp;
 689                 atomic_inc(&qp->refcount);
 690         }
 691         last = &dev->rnrwait;
 692         if (!list_empty(last)) {
 693                 qp = list_entry(last->next, struct ipath_qp, timerwait);
 694                 if (--qp->s_rnr_timeout == 0) {
 695                         do {
 696                                 list_del_init(&qp->timerwait);
 697                                 qp->timer_next = rnr;
 698                                 rnr = qp;
 699                                 atomic_inc(&qp->refcount);
 700                                 if (list_empty(last))
 701                                         break;
 702                                 qp = list_entry(last->next, struct ipath_qp,
 703                                                 timerwait);
 704                         } while (qp->s_rnr_timeout == 0);
 705                 }
 706         }
 707         /*
 708          * We should only be in the started state if pma_sample_start != 0
 709          */
 710         if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_STARTED &&
 711             --dev->pma_sample_start == 0) {
 712                 dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_RUNNING;
 713                 ipath_snapshot_counters(dev->dd, &dev->ipath_sword,
 714                                         &dev->ipath_rword,
 715                                         &dev->ipath_spkts,
 716                                         &dev->ipath_rpkts,
 717                                         &dev->ipath_xmit_wait);
 718         }
 719         if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_RUNNING) {
 720                 if (dev->pma_sample_interval == 0) {
 721                         u64 ta, tb, tc, td, te;
 722
 723                         dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_DONE;
 724                         ipath_snapshot_counters(dev->dd, &ta, &tb,
 725                                                 &tc, &td, &te);
 726
 727                         dev->ipath_sword = ta - dev->ipath_sword;
 728                         dev->ipath_rword = tb - dev->ipath_rword;
 729                         dev->ipath_spkts = tc - dev->ipath_spkts;
 730                         dev->ipath_rpkts = td - dev->ipath_rpkts;
 731                         dev->ipath_xmit_wait = te - dev->ipath_xmit_wait;
 732                 }
 733                 else
 734                         dev->pma_sample_interval--;
 735         }
 736         spin_unlock_irqrestore(&dev->pending_lock, flags);
 737
 738         /* XXX What if timer fires again while this is running? */
 739         while (resend != NULL) {
 740                 qp = resend;
 741                 resend = qp->timer_next;
 742
 743                 spin_lock_irqsave(&qp->s_lock, flags);
 744                 if (qp->s_last != qp->s_tail &&
 745                     ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) {
 746                         dev->n_timeouts++;
 747                         ipath_restart_rc(qp, qp->s_last_psn + 1);
 748                 }
 749                 spin_unlock_irqrestore(&qp->s_lock, flags);
 750
 751                 /* Notify ipath_destroy_qp() if it is waiting. */
 752                 if (atomic_dec_and_test(&qp->refcount))
 753                         wake_up(&qp->wait);
 754         }
 755         while (rnr != NULL) {
 756                 qp = rnr;
 757                 rnr = qp->timer_next;
 758
 759                 spin_lock_irqsave(&qp->s_lock, flags);
 760                 if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)
 761                         ipath_schedule_send(qp);
 762                 spin_unlock_irqrestore(&qp->s_lock, flags);
 763
 764                 /* Notify ipath_destroy_qp() if it is waiting. */
 765                 if (atomic_dec_and_test(&qp->refcount))
 766                         wake_up(&qp->wait);
 767         }
 768 }
 769
 770 static void update_sge(struct ipath_sge_state *ss, u32 length)
 771 {
 772         struct ipath_sge *sge = &ss->sge;
 773
 774         sge->vaddr += length;
 775         sge->length -= length;
 776         sge->sge_length -= length;
 777         if (sge->sge_length == 0) {
 778                 if (--ss->num_sge)
 779                         *sge = *ss->sg_list++;
 780         } else if (sge->length == 0 && sge->mr != NULL) {
 781                 if (++sge->n >= IPATH_SEGSZ) {
 782                         if (++sge->m >= sge->mr->mapsz)
 783                                 return;
 784                         sge->n = 0;
 785                 }
 786                 sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
 787                 sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
 788         }
 789 }
 790
 791 #ifdef __LITTLE_ENDIAN
 792 static inline u32 get_upper_bits(u32 data, u32 shift)
 793 {
 794         return data >> shift;
 795 }
 796
 797 static inline u32 set_upper_bits(u32 data, u32 shift)
 798 {
 799         return data << shift;
 800 }
 801
 802 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
 803 {
 804         data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
 805         data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
 806         return data;
 807 }
 808 #else
 809 static inline u32 get_upper_bits(u32 data, u32 shift)
 810 {
 811         return data << shift;
 812 }
 813
 814 static inline u32 set_upper_bits(u32 data, u32 shift)
 815 {
 816         return data >> shift;
 817 }
 818
 819 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
 820 {
 821         data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
 822         data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
 823         return data;
 824 }
 825 #endif
 826
 827 static void copy_io(u32 __iomem *piobuf, struct ipath_sge_state *ss,
 828                     u32 length, unsigned flush_wc)
 829 {
 830         u32 extra = 0;
 831         u32 data = 0;
 832         u32 last;
 833
 834         while (1) {
 835                 u32 len = ss->sge.length;
 836                 u32 off;
 837
 838                 if (len > length)
 839                         len = length;
 840                 if (len > ss->sge.sge_length)
 841                         len = ss->sge.sge_length;
 842                 BUG_ON(len == 0);
 843                 /* If the source address is not aligned, try to align it. */
 844                 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
 845                 if (off) {
 846                         u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
 847                                             ~(sizeof(u32) - 1));
 848                         u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
 849                         u32 y;
 850
 851                         y = sizeof(u32) - off;
 852                         if (len > y)
 853                                 len = y;
 854                         if (len + extra >= sizeof(u32)) {
 855                                 data |= set_upper_bits(v, extra *
 856                                                        BITS_PER_BYTE);
 857                                 len = sizeof(u32) - extra;
 858                                 if (len == length) {
 859                                         last = data;
 860                                         break;
 861                                 }
 862                                 __raw_writel(data, piobuf);
 863                                 piobuf++;
 864                                 extra = 0;
 865                                 data = 0;
 866                         } else {
 867                                 /* Clear unused upper bytes */
 868                                 data |= clear_upper_bytes(v, len, extra);
 869                                 if (len == length) {
 870                                         last = data;
 871                                         break;
 872                                 }
 873                                 extra += len;
 874                         }
 875                 } else if (extra) {
 876                         /* Source address is aligned. */
 877                         u32 *addr = (u32 *) ss->sge.vaddr;
 878                         int shift = extra * BITS_PER_BYTE;
 879                         int ushift = 32 - shift;
 880                         u32 l = len;
 881
 882                         while (l >= sizeof(u32)) {
 883                                 u32 v = *addr;
 884
 885                                 data |= set_upper_bits(v, shift);
 886                                 __raw_writel(data, piobuf);
 887                                 data = get_upper_bits(v, ushift);
 888                                 piobuf++;
 889                                 addr++;
 890                                 l -= sizeof(u32);
 891                         }
 892                         /*
 893                          * We still have 'extra' number of bytes leftover.
 894                          */
 895                         if (l) {
 896                                 u32 v = *addr;
 897
 898                                 if (l + extra >= sizeof(u32)) {
 899                                         data |= set_upper_bits(v, shift);
 900                                         len -= l + extra - sizeof(u32);
 901                                         if (len == length) {
 902                                                 last = data;
 903                                                 break;
 904                                         }
 905                                         __raw_writel(data, piobuf);
 906                                         piobuf++;
 907                                         extra = 0;
 908                                         data = 0;
 909                                 } else {
 910                                         /* Clear unused upper bytes */
 911                                         data |= clear_upper_bytes(v, l,
 912                                                                   extra);
 913                                         if (len == length) {
 914                                                 last = data;
 915                                                 break;
 916                                         }
 917                                         extra += l;
 918                                 }
 919                         } else if (len == length) {
 920                                 last = data;
 921                                 break;
 922                         }
 923                 } else if (len == length) {
 924                         u32 w;
 925
 926                         /*
 927                          * Need to round up for the last dword in the
 928                          * packet.
 929                          */
 930                         w = (len + 3) >> 2;
 931                         __iowrite32_copy(piobuf, ss->sge.vaddr, w - 1);
 932                         piobuf += w - 1;
 933                         last = ((u32 *) ss->sge.vaddr)[w - 1];
 934                         break;
 935                 } else {
 936                         u32 w = len >> 2;
 937
 938                         __iowrite32_copy(piobuf, ss->sge.vaddr, w);
 939                         piobuf += w;
 940
 941                         extra = len & (sizeof(u32) - 1);
 942                         if (extra) {
 943                                 u32 v = ((u32 *) ss->sge.vaddr)[w];
 944
 945                                 /* Clear unused upper bytes */
 946                                 data = clear_upper_bytes(v, extra, 0);
 947                         }
 948                 }
 949                 update_sge(ss, len);
 950                 length -= len;
 951         }
 952         /* Update address before sending packet. */
 953         update_sge(ss, length);
 954         if (flush_wc) {
 955                 /* must flush early everything before trigger word */
 956                 ipath_flush_wc();
 957                 __raw_writel(last, piobuf);
 958                 /* be sure trigger word is written */
 959                 ipath_flush_wc();
 960         } else
 961                 __raw_writel(last, piobuf);
 962 }
 963
 964 /*
 965  * Convert IB rate to delay multiplier.
 966  */
 967 unsigned ipath_ib_rate_to_mult(enum ib_rate rate)
 968 {
 969         switch (rate) {
 970         case IB_RATE_2_5_GBPS: return 8;
 971         case IB_RATE_5_GBPS:   return 4;
 972         case IB_RATE_10_GBPS:  return 2;
 973         case IB_RATE_20_GBPS:  return 1;
 974         default:               return 0;
 975         }
 976 }
 977
 978 /*
 979  * Convert delay multiplier to IB rate
 980  */
 981 static enum ib_rate ipath_mult_to_ib_rate(unsigned mult)
 982 {
 983         switch (mult) {
 984         case 8:  return IB_RATE_2_5_GBPS;
 985         case 4:  return IB_RATE_5_GBPS;
 986         case 2:  return IB_RATE_10_GBPS;
 987         case 1:  return IB_RATE_20_GBPS;
 988         default: return IB_RATE_PORT_CURRENT;
 989         }
 990 }
 991
 992 static inline struct ipath_verbs_txreq *get_txreq(struct ipath_ibdev *dev)
 993 {
 994         struct ipath_verbs_txreq *tx = NULL;
 995         unsigned long flags;
 996
 997         spin_lock_irqsave(&dev->pending_lock, flags);
 998         if (!list_empty(&dev->txreq_free)) {
 999                 struct list_head *l = dev->txreq_free.next;
1000
1001                 list_del(l);
1002                 tx = list_entry(l, struct ipath_verbs_txreq, txreq.list);
1003         }
1004         spin_unlock_irqrestore(&dev->pending_lock, flags);
1005         return tx;
1006 }
1007
1008 static inline void put_txreq(struct ipath_ibdev *dev,
1009                              struct ipath_verbs_txreq *tx)
1010 {
1011         unsigned long flags;
1012
1013         spin_lock_irqsave(&dev->pending_lock, flags);
1014         list_add(&tx->txreq.list, &dev->txreq_free);
1015         spin_unlock_irqrestore(&dev->pending_lock, flags);
1016 }
1017
1018 static void sdma_complete(void *cookie, int status)
1019 {
1020         struct ipath_verbs_txreq *tx = cookie;
1021         struct ipath_qp *qp = tx->qp;
1022         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
1023         unsigned int flags;
1024         enum ib_wc_status ibs = status == IPATH_SDMA_TXREQ_S_OK ?
1025                 IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR;
1026
1027         if (atomic_dec_and_test(&qp->s_dma_busy)) {
1028                 spin_lock_irqsave(&qp->s_lock, flags);
1029                 if (tx->wqe)
1030                         ipath_send_complete(qp, tx->wqe, ibs);
1031                 if ((ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND &&
1032                      qp->s_last != qp->s_head) ||
1033                     (qp->s_flags & IPATH_S_WAIT_DMA))
1034                         ipath_schedule_send(qp);
1035                 spin_unlock_irqrestore(&qp->s_lock, flags);
1036                 wake_up(&qp->wait_dma);
1037         } else if (tx->wqe) {
1038                 spin_lock_irqsave(&qp->s_lock, flags);
1039                 ipath_send_complete(qp, tx->wqe, ibs);
1040                 spin_unlock_irqrestore(&qp->s_lock, flags);
1041         }
1042
1043         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEBUF)
1044                 kfree(tx->txreq.map_addr);
1045         put_txreq(dev, tx);
1046
1047         if (atomic_dec_and_test(&qp->refcount))
1048                 wake_up(&qp->wait);
1049 }
1050
1051 static void decrement_dma_busy(struct ipath_qp *qp)
1052 {
1053         unsigned int flags;
1054
1055         if (atomic_dec_and_test(&qp->s_dma_busy)) {
1056                 spin_lock_irqsave(&qp->s_lock, flags);
1057                 if ((ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND &&
1058                      qp->s_last != qp->s_head) ||
1059                     (qp->s_flags & IPATH_S_WAIT_DMA))
1060                         ipath_schedule_send(qp);
1061                 spin_unlock_irqrestore(&qp->s_lock, flags);
1062                 wake_up(&qp->wait_dma);
1063         }
1064 }
1065
1066 /*
1067  * Compute the number of clock cycles of delay before sending the next packet.
1068  * The multipliers reflect the number of clocks for the fastest rate so
1069  * one tick at 4xDDR is 8 ticks at 1xSDR.
1070  * If the destination port will take longer to receive a packet than
1071  * the outgoing link can send it, we need to delay sending the next packet
1072  * by the difference in time it takes the receiver to receive and the sender
1073  * to send this packet.
1074  * Note that this delay is always correct for UC and RC but not always
1075  * optimal for UD. For UD, the destination HCA can be different for each
1076  * packet, in which case, we could send packets to a different destination
1077  * while "waiting" for the delay. The overhead for doing this without
1078  * HW support is more than just paying the cost of delaying some packets
1079  * unnecessarily.
1080  */
1081 static inline unsigned ipath_pkt_delay(u32 plen, u8 snd_mult, u8 rcv_mult)
1082 {
1083         return (rcv_mult > snd_mult) ?
1084                 (plen * (rcv_mult - snd_mult) + 1) >> 1 : 0;
1085 }
1086
1087 static int ipath_verbs_send_dma(struct ipath_qp *qp,
1088                                 struct ipath_ib_header *hdr, u32 hdrwords,
1089                                 struct ipath_sge_state *ss, u32 len,
1090                                 u32 plen, u32 dwords)
1091 {
1092         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
1093         struct ipath_devdata *dd = dev->dd;
1094         struct ipath_verbs_txreq *tx;
1095         u32 *piobuf;
1096         u32 control;
1097         u32 ndesc;
1098         int ret;
1099
1100         tx = qp->s_tx;
1101         if (tx) {
1102                 qp->s_tx = NULL;
1103                 /* resend previously constructed packet */
1104                 atomic_inc(&qp->s_dma_busy);
1105                 ret = ipath_sdma_verbs_send(dd, tx->ss, tx->len, tx);
1106                 if (ret) {
1107                         qp->s_tx = tx;
1108                         decrement_dma_busy(qp);
1109                 }
1110                 goto bail;
1111         }
1112
1113         tx = get_txreq(dev);
1114         if (!tx) {
1115                 ret = -EBUSY;
1116                 goto bail;
1117         }
1118
1119         /*
1120          * Get the saved delay count we computed for the previous packet
1121          * and save the delay count for this packet to be used next time
1122          * we get here.
1123          */
1124         control = qp->s_pkt_delay;
1125         qp->s_pkt_delay = ipath_pkt_delay(plen, dd->delay_mult, qp->s_dmult);
1126
1127         tx->qp = qp;
1128         atomic_inc(&qp->refcount);
1129         tx->wqe = qp->s_wqe;
1130         tx->txreq.callback = sdma_complete;
1131         tx->txreq.callback_cookie = tx;
1132         tx->txreq.flags = IPATH_SDMA_TXREQ_F_HEADTOHOST |
1133                 IPATH_SDMA_TXREQ_F_INTREQ | IPATH_SDMA_TXREQ_F_FREEDESC;
1134         if (plen + 1 >= IPATH_SMALLBUF_DWORDS)
1135                 tx->txreq.flags |= IPATH_SDMA_TXREQ_F_USELARGEBUF;
1136
1137         /* VL15 packets bypass credit check */
1138         if ((be16_to_cpu(hdr->lrh[0]) >> 12) == 15) {
1139                 control |= 1ULL << 31;
1140                 tx->txreq.flags |= IPATH_SDMA_TXREQ_F_VL15;
1141         }
1142
1143         if (len) {
1144                 /*
1145                  * Don't try to DMA if it takes more descriptors than
1146                  * the queue holds.
1147                  */
1148                 ndesc = ipath_count_sge(ss, len);
1149                 if (ndesc >= dd->ipath_sdma_descq_cnt)
1150                         ndesc = 0;
1151         } else
1152                 ndesc = 1;
1153         if (ndesc) {
1154                 tx->hdr.pbc[0] = cpu_to_le32(plen);
1155                 tx->hdr.pbc[1] = cpu_to_le32(control);
1156                 memcpy(&tx->hdr.hdr, hdr, hdrwords << 2);
1157                 tx->txreq.sg_count = ndesc;
1158                 tx->map_len = (hdrwords + 2) << 2;
1159                 tx->txreq.map_addr = &tx->hdr;
1160                 atomic_inc(&qp->s_dma_busy);
1161                 ret = ipath_sdma_verbs_send(dd, ss, dwords, tx);
1162                 if (ret) {
1163                         /* save ss and length in dwords */
1164                         tx->ss = ss;
1165                         tx->len = dwords;
1166                         qp->s_tx = tx;
1167                         decrement_dma_busy(qp);
1168                 }
1169                 goto bail;
1170         }
1171
1172         /* Allocate a buffer and copy the header and payload to it. */
1173         tx->map_len = (plen + 1) << 2;
1174         piobuf = kmalloc(tx->map_len, GFP_ATOMIC);
1175         if (unlikely(piobuf == NULL)) {
1176                 ret = -EBUSY;
1177                 goto err_tx;
1178         }
1179         tx->txreq.map_addr = piobuf;
1180         tx->txreq.flags |= IPATH_SDMA_TXREQ_F_FREEBUF;
1181         tx->txreq.sg_count = 1;
1182
1183         *piobuf++ = (__force u32) cpu_to_le32(plen);
1184         *piobuf++ = (__force u32) cpu_to_le32(control);
1185         memcpy(piobuf, hdr, hdrwords << 2);
1186         ipath_copy_from_sge(piobuf + hdrwords, ss, len);
1187
1188         atomic_inc(&qp->s_dma_busy);
1189         ret = ipath_sdma_verbs_send(dd, NULL, 0, tx);
1190         /*
1191          * If we couldn't queue the DMA request, save the info
1192          * and try again later rather than destroying the
1193          * buffer and undoing the side effects of the copy.
1194          */
1195         if (ret) {
1196                 tx->ss = NULL;
1197                 tx->len = 0;
1198                 qp->s_tx = tx;
1199                 decrement_dma_busy(qp);
1200         }
1201         dev->n_unaligned++;
1202         goto bail;
1203
1204 err_tx:
1205         if (atomic_dec_and_test(&qp->refcount))
1206                 wake_up(&qp->wait);
1207         put_txreq(dev, tx);
1208 bail:
1209         return ret;
1210 }
1211
1212 static int ipath_verbs_send_pio(struct ipath_qp *qp,
1213                                 struct ipath_ib_header *ibhdr, u32 hdrwords,
1214                                 struct ipath_sge_state *ss, u32 len,
1215                                 u32 plen, u32 dwords)
1216 {
1217         struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
1218         u32 *hdr = (u32 *) ibhdr;
1219         u32 __iomem *piobuf;
1220         unsigned flush_wc;
1221         u32 control;
1222         int ret;
1223         unsigned int flags;
1224
1225         piobuf = ipath_getpiobuf(dd, plen, NULL);
1226         if (unlikely(piobuf == NULL)) {
1227                 ret = -EBUSY;
1228                 goto bail;
1229         }
1230
1231         /*
1232          * Get the saved delay count we computed for the previous packet
1233          * and save the delay count for this packet to be used next time
1234          * we get here.
1235          */
1236         control = qp->s_pkt_delay;
1237         qp->s_pkt_delay = ipath_pkt_delay(plen, dd->delay_mult, qp->s_dmult);
1238
1239         /* VL15 packets bypass credit check */
1240         if ((be16_to_cpu(ibhdr->lrh[0]) >> 12) == 15)
1241                 control |= 1ULL << 31;
1242
1243         /*
1244          * Write the length to the control qword plus any needed flags.
1245          * We have to flush after the PBC for correctness on some cpus
1246          * or WC buffer can be written out of order.
1247          */
1248         writeq(((u64) control << 32) | plen, piobuf);
1249         piobuf += 2;
1250
1251         flush_wc = dd->ipath_flags & IPATH_PIO_FLUSH_WC;
1252         if (len == 0) {
1253                 /*
1254                  * If there is just the header portion, must flush before
1255                  * writing last word of header for correctness, and after
1256                  * the last header word (trigger word).
1257                  */
1258                 if (flush_wc) {
1259                         ipath_flush_wc();
1260                         __iowrite32_copy(piobuf, hdr, hdrwords - 1);
1261                         ipath_flush_wc();
1262                         __raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
1263                         ipath_flush_wc();
1264                 } else
1265                         __iowrite32_copy(piobuf, hdr, hdrwords);
1266                 goto done;
1267         }
1268
1269         if (flush_wc)
1270                 ipath_flush_wc();
1271         __iowrite32_copy(piobuf, hdr, hdrwords);
1272         piobuf += hdrwords;
1273
1274         /* The common case is aligned and contained in one segment. */
1275         if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
1276                    !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
1277                 u32 *addr = (u32 *) ss->sge.vaddr;
1278
1279                 /* Update address before sending packet. */
1280                 update_sge(ss, len);
1281                 if (flush_wc) {
1282                         __iowrite32_copy(piobuf, addr, dwords - 1);
1283                         /* must flush early everything before trigger word */
1284                         ipath_flush_wc();
1285                         __raw_writel(addr[dwords - 1], piobuf + dwords - 1);
1286                         /* be sure trigger word is written */
1287                         ipath_flush_wc();
1288                 } else
1289                         __iowrite32_copy(piobuf, addr, dwords);
1290                 goto done;
1291         }
1292         copy_io(piobuf, ss, len, flush_wc);
1293 done:
1294         if (qp->s_wqe) {
1295                 spin_lock_irqsave(&qp->s_lock, flags);
1296                 ipath_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
1297                 spin_unlock_irqrestore(&qp->s_lock, flags);
1298         }
1299         ret = 0;
1300 bail:
1301         return ret;
1302 }
1303
1304 /**
1305  * ipath_verbs_send - send a packet
1306  * @qp: the QP to send on
1307  * @hdr: the packet header
1308  * @hdrwords: the number of 32-bit words in the header
1309  * @ss: the SGE to send
1310  * @len: the length of the packet in bytes
1311  */
1312 int ipath_verbs_send(struct ipath_qp *qp, struct ipath_ib_header *hdr,
1313                      u32 hdrwords, struct ipath_sge_state *ss, u32 len)
1314 {
1315         struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
1316         u32 plen;
1317         int ret;
1318         u32 dwords = (len + 3) >> 2;
1319
1320         /*
1321          * Calculate the send buffer trigger address.
1322          * The +1 counts for the pbc control dword following the pbc length.
1323          */
1324         plen = hdrwords + dwords + 1;
1325
1326         /*
1327          * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1328          * can defer SDMA restart until link goes ACTIVE without
1329          * worrying about just how we got there.
1330          */
1331         if (qp->ibqp.qp_type == IB_QPT_SMI ||
1332             !(dd->ipath_flags & IPATH_HAS_SEND_DMA))
1333                 ret = ipath_verbs_send_pio(qp, hdr, hdrwords, ss, len,
1334                                            plen, dwords);
1335         else
1336                 ret = ipath_verbs_send_dma(qp, hdr, hdrwords, ss, len,
1337                                            plen, dwords);
1338
1339         return ret;
1340 }
1341
1342 int ipath_snapshot_counters(struct ipath_devdata *dd, u64 *swords,
1343                             u64 *rwords, u64 *spkts, u64 *rpkts,
1344                             u64 *xmit_wait)
1345 {
1346         int ret;
1347
1348         if (!(dd->ipath_flags & IPATH_INITTED)) {
1349                 /* no hardware, freeze, etc. */
1350                 ret = -EINVAL;
1351                 goto bail;
1352         }
1353         *swords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt);
1354         *rwords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
1355         *spkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
1356         *rpkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
1357         *xmit_wait = ipath_snap_cntr(dd, dd->ipath_cregs->cr_sendstallcnt);
1358
1359         ret = 0;
1360
1361 bail:
1362         return ret;
1363 }
1364
1365 /**
1366  * ipath_get_counters - get various chip counters
1367  * @dd: the infinipath device
1368  * @cntrs: counters are placed here
1369  *
1370  * Return the counters needed by recv_pma_get_portcounters().
1371  */
1372 int ipath_get_counters(struct ipath_devdata *dd,
1373                        struct ipath_verbs_counters *cntrs)
1374 {
1375         struct ipath_cregs const *crp = dd->ipath_cregs;
1376         int ret;
1377
1378         if (!(dd->ipath_flags & IPATH_INITTED)) {
1379                 /* no hardware, freeze, etc. */
1380                 ret = -EINVAL;
1381                 goto bail;
1382         }
1383         cntrs->symbol_error_counter =
1384                 ipath_snap_cntr(dd, crp->cr_ibsymbolerrcnt);
1385         cntrs->link_error_recovery_counter =
1386                 ipath_snap_cntr(dd, crp->cr_iblinkerrrecovcnt);
1387         /*
1388          * The link downed counter counts when the other side downs the
1389          * connection.  We add in the number of times we downed the link
1390          * due to local link integrity errors to compensate.
1391          */
1392         cntrs->link_downed_counter =
1393                 ipath_snap_cntr(dd, crp->cr_iblinkdowncnt);
1394         cntrs->port_rcv_errors =
1395                 ipath_snap_cntr(dd, crp->cr_rxdroppktcnt) +
1396                 ipath_snap_cntr(dd, crp->cr_rcvovflcnt) +
1397                 ipath_snap_cntr(dd, crp->cr_portovflcnt) +
1398                 ipath_snap_cntr(dd, crp->cr_err_rlencnt) +
1399                 ipath_snap_cntr(dd, crp->cr_invalidrlencnt) +
1400                 ipath_snap_cntr(dd, crp->cr_errlinkcnt) +
1401                 ipath_snap_cntr(dd, crp->cr_erricrccnt) +
1402                 ipath_snap_cntr(dd, crp->cr_errvcrccnt) +
1403                 ipath_snap_cntr(dd, crp->cr_errlpcrccnt) +
1404                 ipath_snap_cntr(dd, crp->cr_badformatcnt) +
1405                 dd->ipath_rxfc_unsupvl_errs;
1406         if (crp->cr_rxotherlocalphyerrcnt)
1407                 cntrs->port_rcv_errors +=
1408                         ipath_snap_cntr(dd, crp->cr_rxotherlocalphyerrcnt);
1409         if (crp->cr_rxvlerrcnt)
1410                 cntrs->port_rcv_errors +=
1411                         ipath_snap_cntr(dd, crp->cr_rxvlerrcnt);
1412         cntrs->port_rcv_remphys_errors =
1413                 ipath_snap_cntr(dd, crp->cr_rcvebpcnt);
1414         cntrs->port_xmit_discards = ipath_snap_cntr(dd, crp->cr_unsupvlcnt);
1415         cntrs->port_xmit_data = ipath_snap_cntr(dd, crp->cr_wordsendcnt);
1416         cntrs->port_rcv_data = ipath_snap_cntr(dd, crp->cr_wordrcvcnt);
1417         cntrs->port_xmit_packets = ipath_snap_cntr(dd, crp->cr_pktsendcnt);
1418         cntrs->port_rcv_packets = ipath_snap_cntr(dd, crp->cr_pktrcvcnt);
1419         cntrs->local_link_integrity_errors =
1420                 crp->cr_locallinkintegrityerrcnt ?
1421                 ipath_snap_cntr(dd, crp->cr_locallinkintegrityerrcnt) :
1422                 ((dd->ipath_flags & IPATH_GPIO_ERRINTRS) ?
1423                  dd->ipath_lli_errs : dd->ipath_lli_errors);
1424         cntrs->excessive_buffer_overrun_errors =
1425                 crp->cr_excessbufferovflcnt ?
1426                 ipath_snap_cntr(dd, crp->cr_excessbufferovflcnt) :
1427                 dd->ipath_overrun_thresh_errs;
1428         cntrs->vl15_dropped = crp->cr_vl15droppedpktcnt ?
1429                 ipath_snap_cntr(dd, crp->cr_vl15droppedpktcnt) : 0;
1430
1431         ret = 0;
1432
1433 bail:
1434         return ret;
1435 }
1436
1437 /**
1438  * ipath_ib_piobufavail - callback when a PIO buffer is available
1439  * @arg: the device pointer
1440  *
1441  * This is called from ipath_intr() at interrupt level when a PIO buffer is
1442  * available after ipath_verbs_send() returned an error that no buffers were
1443  * available.  Return 1 if we consumed all the PIO buffers and we still have
1444  * QPs waiting for buffers (for now, just restart the send tasklet and
1445  * return zero).
1446  */
1447 int ipath_ib_piobufavail(struct ipath_ibdev *dev)
1448 {
1449         struct list_head *list;
1450         struct ipath_qp *qplist;
1451         struct ipath_qp *qp;
1452         unsigned long flags;
1453
1454         if (dev == NULL)
1455                 goto bail;
1456
1457         list = &dev->piowait;
1458         qplist = NULL;
1459
1460         spin_lock_irqsave(&dev->pending_lock, flags);
1461         while (!list_empty(list)) {
1462                 qp = list_entry(list->next, struct ipath_qp, piowait);
1463                 list_del_init(&qp->piowait);
1464                 qp->pio_next = qplist;
1465                 qplist = qp;
1466                 atomic_inc(&qp->refcount);
1467         }
1468         spin_unlock_irqrestore(&dev->pending_lock, flags);
1469
1470         while (qplist != NULL) {
1471                 qp = qplist;
1472                 qplist = qp->pio_next;
1473
1474                 spin_lock_irqsave(&qp->s_lock, flags);
1475                 if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)
1476                         ipath_schedule_send(qp);
1477                 spin_unlock_irqrestore(&qp->s_lock, flags);
1478
1479                 /* Notify ipath_destroy_qp() if it is waiting. */
1480                 if (atomic_dec_and_test(&qp->refcount))
1481                         wake_up(&qp->wait);
1482         }
1483
1484 bail:
1485         return 0;
1486 }
1487
1488 static int ipath_query_device(struct ib_device *ibdev,
1489                               struct ib_device_attr *props)
1490 {
1491         struct ipath_ibdev *dev = to_idev(ibdev);
1492
1493         memset(props, 0, sizeof(*props));
1494
1495         props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1496                 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1497                 IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1498                 IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE;
1499         props->page_size_cap = PAGE_SIZE;
1500         props->vendor_id =
1501                 IPATH_SRC_OUI_1 << 16 | IPATH_SRC_OUI_2 << 8 | IPATH_SRC_OUI_3;
1502         props->vendor_part_id = dev->dd->ipath_deviceid;
1503         props->hw_ver = dev->dd->ipath_pcirev;
1504
1505         props->sys_image_guid = dev->sys_image_guid;
1506
1507         props->max_mr_size = ~0ull;
1508         props->max_qp = ib_ipath_max_qps;
1509         props->max_qp_wr = ib_ipath_max_qp_wrs;
1510         props->max_sge = ib_ipath_max_sges;
1511         props->max_cq = ib_ipath_max_cqs;
1512         props->max_ah = ib_ipath_max_ahs;
1513         props->max_cqe = ib_ipath_max_cqes;
1514         props->max_mr = dev->lk_table.max;
1515         props->max_fmr = dev->lk_table.max;
1516         props->max_map_per_fmr = 32767;
1517         props->max_pd = ib_ipath_max_pds;
1518         props->max_qp_rd_atom = IPATH_MAX_RDMA_ATOMIC;
1519         props->max_qp_init_rd_atom = 255;
1520         /* props->max_res_rd_atom */
1521         props->max_srq = ib_ipath_max_srqs;
1522         props->max_srq_wr = ib_ipath_max_srq_wrs;
1523         props->max_srq_sge = ib_ipath_max_srq_sges;
1524         /* props->local_ca_ack_delay */
1525         props->atomic_cap = IB_ATOMIC_GLOB;
1526         props->max_pkeys = ipath_get_npkeys(dev->dd);
1527         props->max_mcast_grp = ib_ipath_max_mcast_grps;
1528         props->max_mcast_qp_attach = ib_ipath_max_mcast_qp_attached;
1529         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1530                 props->max_mcast_grp;
1531
1532         return 0;
1533 }
1534
1535 const u8 ipath_cvt_physportstate[32] = {
1536         [INFINIPATH_IBCS_LT_STATE_DISABLED] = IB_PHYSPORTSTATE_DISABLED,
1537         [INFINIPATH_IBCS_LT_STATE_LINKUP] = IB_PHYSPORTSTATE_LINKUP,
1538         [INFINIPATH_IBCS_LT_STATE_POLLACTIVE] = IB_PHYSPORTSTATE_POLL,
1539         [INFINIPATH_IBCS_LT_STATE_POLLQUIET] = IB_PHYSPORTSTATE_POLL,
1540         [INFINIPATH_IBCS_LT_STATE_SLEEPDELAY] = IB_PHYSPORTSTATE_SLEEP,
1541         [INFINIPATH_IBCS_LT_STATE_SLEEPQUIET] = IB_PHYSPORTSTATE_SLEEP,
1542         [INFINIPATH_IBCS_LT_STATE_CFGDEBOUNCE] =
1543                 IB_PHYSPORTSTATE_CFG_TRAIN,
1544         [INFINIPATH_IBCS_LT_STATE_CFGRCVFCFG] =
1545                 IB_PHYSPORTSTATE_CFG_TRAIN,
1546         [INFINIPATH_IBCS_LT_STATE_CFGWAITRMT] =
1547                 IB_PHYSPORTSTATE_CFG_TRAIN,
1548         [INFINIPATH_IBCS_LT_STATE_CFGIDLE] = IB_PHYSPORTSTATE_CFG_TRAIN,
1549         [INFINIPATH_IBCS_LT_STATE_RECOVERRETRAIN] =
1550                 IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1551         [INFINIPATH_IBCS_LT_STATE_RECOVERWAITRMT] =
1552                 IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1553         [INFINIPATH_IBCS_LT_STATE_RECOVERIDLE] =
1554                 IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1555         [0x10] = IB_PHYSPORTSTATE_CFG_TRAIN,
1556         [0x11] = IB_PHYSPORTSTATE_CFG_TRAIN,
1557         [0x12] = IB_PHYSPORTSTATE_CFG_TRAIN,
1558         [0x13] = IB_PHYSPORTSTATE_CFG_TRAIN,
1559         [0x14] = IB_PHYSPORTSTATE_CFG_TRAIN,
1560         [0x15] = IB_PHYSPORTSTATE_CFG_TRAIN,
1561         [0x16] = IB_PHYSPORTSTATE_CFG_TRAIN,
1562         [0x17] = IB_PHYSPORTSTATE_CFG_TRAIN
1563 };
1564
1565 u32 ipath_get_cr_errpkey(struct ipath_devdata *dd)
1566 {
1567         return ipath_read_creg32(dd, dd->ipath_cregs->cr_errpkey);
1568 }
1569
1570 static int ipath_query_port(struct ib_device *ibdev,
1571                             u8 port, struct ib_port_attr *props)
1572 {
1573         struct ipath_ibdev *dev = to_idev(ibdev);
1574         struct ipath_devdata *dd = dev->dd;
1575         enum ib_mtu mtu;
1576         u16 lid = dd->ipath_lid;
1577         u64 ibcstat;
1578
1579         memset(props, 0, sizeof(*props));
1580         props->lid = lid ? lid : __constant_be16_to_cpu(IB_LID_PERMISSIVE);
1581         props->lmc = dd->ipath_lmc;
1582         props->sm_lid = dev->sm_lid;
1583         props->sm_sl = dev->sm_sl;
1584         ibcstat = dd->ipath_lastibcstat;
1585         /* map LinkState to IB portinfo values.  */
1586         props->state = ipath_ib_linkstate(dd, ibcstat) + 1;
1587
1588         /* See phys_state_show() */
1589         props->phys_state = /* MEA: assumes shift == 0 */
1590                 ipath_cvt_physportstate[dd->ipath_lastibcstat &
1591                 dd->ibcs_lts_mask];
1592         props->port_cap_flags = dev->port_cap_flags;
1593         props->gid_tbl_len = 1;
1594         props->max_msg_sz = 0x80000000;
1595         props->pkey_tbl_len = ipath_get_npkeys(dd);
1596         props->bad_pkey_cntr = ipath_get_cr_errpkey(dd) -
1597                 dev->z_pkey_violations;
1598         props->qkey_viol_cntr = dev->qkey_violations;
1599         props->active_width = dd->ipath_link_width_active;
1600         /* See rate_show() */
1601         props->active_speed = dd->ipath_link_speed_active;
1602         props->max_vl_num = 1;          /* VLCap = VL0 */
1603         props->init_type_reply = 0;
1604
1605         props->max_mtu = ipath_mtu4096 ? IB_MTU_4096 : IB_MTU_2048;
1606         switch (dd->ipath_ibmtu) {
1607         case 4096:
1608                 mtu = IB_MTU_4096;
1609                 break;
1610         case 2048:
1611                 mtu = IB_MTU_2048;
1612                 break;
1613         case 1024:
1614                 mtu = IB_MTU_1024;
1615                 break;
1616         case 512:
1617                 mtu = IB_MTU_512;
1618                 break;
1619         case 256:
1620                 mtu = IB_MTU_256;
1621                 break;
1622         default:
1623                 mtu = IB_MTU_2048;
1624         }
1625         props->active_mtu = mtu;
1626         props->subnet_timeout = dev->subnet_timeout;
1627
1628         return 0;
1629 }
1630
1631 static int ipath_modify_device(struct ib_device *device,
1632                                int device_modify_mask,
1633                                struct ib_device_modify *device_modify)
1634 {
1635         int ret;
1636
1637         if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1638                                    IB_DEVICE_MODIFY_NODE_DESC)) {
1639                 ret = -EOPNOTSUPP;
1640                 goto bail;
1641         }
1642
1643         if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC)
1644                 memcpy(device->node_desc, device_modify->node_desc, 64);
1645
1646         if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
1647                 to_idev(device)->sys_image_guid =
1648                         cpu_to_be64(device_modify->sys_image_guid);
1649
1650         ret = 0;
1651
1652 bail:
1653         return ret;
1654 }
1655
1656 static int ipath_modify_port(struct ib_device *ibdev,
1657                              u8 port, int port_modify_mask,
1658                              struct ib_port_modify *props)
1659 {
1660         struct ipath_ibdev *dev = to_idev(ibdev);
1661
1662         dev->port_cap_flags |= props->set_port_cap_mask;
1663         dev->port_cap_flags &= ~props->clr_port_cap_mask;
1664         if (port_modify_mask & IB_PORT_SHUTDOWN)
1665                 ipath_set_linkstate(dev->dd, IPATH_IB_LINKDOWN);
1666         if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR)
1667                 dev->qkey_violations = 0;
1668         return 0;
1669 }
1670
1671 static int ipath_query_gid(struct ib_device *ibdev, u8 port,
1672                            int index, union ib_gid *gid)
1673 {
1674         struct ipath_ibdev *dev = to_idev(ibdev);
1675         int ret;
1676
1677         if (index >= 1) {
1678                 ret = -EINVAL;
1679                 goto bail;
1680         }
1681         gid->global.subnet_prefix = dev->gid_prefix;
1682         gid->global.interface_id = dev->dd->ipath_guid;
1683
1684         ret = 0;
1685
1686 bail:
1687         return ret;
1688 }
1689
1690 static struct ib_pd *ipath_alloc_pd(struct ib_device *ibdev,
1691                                     struct ib_ucontext *context,
1692                                     struct ib_udata *udata)
1693 {
1694         struct ipath_ibdev *dev = to_idev(ibdev);
1695         struct ipath_pd *pd;
1696         struct ib_pd *ret;
1697
1698         /*
1699          * This is actually totally arbitrary.  Some correctness tests
1700          * assume there's a maximum number of PDs that can be allocated.
1701          * We don't actually have this limit, but we fail the test if
1702          * we allow allocations of more than we report for this value.
1703          */
1704
1705         pd = kmalloc(sizeof *pd, GFP_KERNEL);
1706         if (!pd) {
1707                 ret = ERR_PTR(-ENOMEM);
1708                 goto bail;
1709         }
1710
1711         spin_lock(&dev->n_pds_lock);
1712         if (dev->n_pds_allocated == ib_ipath_max_pds) {
1713                 spin_unlock(&dev->n_pds_lock);
1714                 kfree(pd);
1715                 ret = ERR_PTR(-ENOMEM);
1716                 goto bail;
1717         }
1718
1719         dev->n_pds_allocated++;
1720         spin_unlock(&dev->n_pds_lock);
1721
1722         /* ib_alloc_pd() will initialize pd->ibpd. */
1723         pd->user = udata != NULL;
1724
1725         ret = &pd->ibpd;
1726
1727 bail:
1728         return ret;
1729 }
1730
1731 static int ipath_dealloc_pd(struct ib_pd *ibpd)
1732 {
1733         struct ipath_pd *pd = to_ipd(ibpd);
1734         struct ipath_ibdev *dev = to_idev(ibpd->device);
1735
1736         spin_lock(&dev->n_pds_lock);
1737         dev->n_pds_allocated--;
1738         spin_unlock(&dev->n_pds_lock);
1739
1740         kfree(pd);
1741
1742         return 0;
1743 }
1744
1745 /**
1746  * ipath_create_ah - create an address handle
1747  * @pd: the protection domain
1748  * @ah_attr: the attributes of the AH
1749  *
1750  * This may be called from interrupt context.
1751  */
1752 static struct ib_ah *ipath_create_ah(struct ib_pd *pd,
1753                                      struct ib_ah_attr *ah_attr)
1754 {
1755         struct ipath_ah *ah;
1756         struct ib_ah *ret;
1757         struct ipath_ibdev *dev = to_idev(pd->device);
1758         unsigned long flags;
1759
1760         /* A multicast address requires a GRH (see ch. 8.4.1). */
1761         if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
1762             ah_attr->dlid != IPATH_PERMISSIVE_LID &&
1763             !(ah_attr->ah_flags & IB_AH_GRH)) {
1764                 ret = ERR_PTR(-EINVAL);
1765                 goto bail;
1766         }
1767
1768         if (ah_attr->dlid == 0) {
1769                 ret = ERR_PTR(-EINVAL);
1770                 goto bail;
1771         }
1772
1773         if (ah_attr->port_num < 1 ||
1774             ah_attr->port_num > pd->device->phys_port_cnt) {
1775                 ret = ERR_PTR(-EINVAL);
1776                 goto bail;
1777         }
1778
1779         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
1780         if (!ah) {
1781                 ret = ERR_PTR(-ENOMEM);
1782                 goto bail;
1783         }
1784
1785         spin_lock_irqsave(&dev->n_ahs_lock, flags);
1786         if (dev->n_ahs_allocated == ib_ipath_max_ahs) {
1787                 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1788                 kfree(ah);
1789                 ret = ERR_PTR(-ENOMEM);
1790                 goto bail;
1791         }
1792
1793         dev->n_ahs_allocated++;
1794         spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1795
1796         /* ib_create_ah() will initialize ah->ibah. */
1797         ah->attr = *ah_attr;
1798         ah->attr.static_rate = ipath_ib_rate_to_mult(ah_attr->static_rate);
1799
1800         ret = &ah->ibah;
1801
1802 bail:
1803         return ret;
1804 }
1805
1806 /**
1807  * ipath_destroy_ah - destroy an address handle
1808  * @ibah: the AH to destroy
1809  *
1810  * This may be called from interrupt context.
1811  */
1812 static int ipath_destroy_ah(struct ib_ah *ibah)
1813 {
1814         struct ipath_ibdev *dev = to_idev(ibah->device);
1815         struct ipath_ah *ah = to_iah(ibah);
1816         unsigned long flags;
1817
1818         spin_lock_irqsave(&dev->n_ahs_lock, flags);
1819         dev->n_ahs_allocated--;
1820         spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1821
1822         kfree(ah);
1823
1824         return 0;
1825 }
1826
1827 static int ipath_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1828 {
1829         struct ipath_ah *ah = to_iah(ibah);
1830
1831         *ah_attr = ah->attr;
1832         ah_attr->static_rate = ipath_mult_to_ib_rate(ah->attr.static_rate);
1833
1834         return 0;
1835 }
1836
1837 /**
1838  * ipath_get_npkeys - return the size of the PKEY table for port 0
1839  * @dd: the infinipath device
1840  */
1841 unsigned ipath_get_npkeys(struct ipath_devdata *dd)
1842 {
1843         return ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys);
1844 }
1845
1846 /**
1847  * ipath_get_pkey - return the indexed PKEY from the port 0 PKEY table
1848  * @dd: the infinipath device
1849  * @index: the PKEY index
1850  */
1851 unsigned ipath_get_pkey(struct ipath_devdata *dd, unsigned index)
1852 {
1853         unsigned ret;
1854
1855         if (index >= ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys))
1856                 ret = 0;
1857         else
1858                 ret = dd->ipath_pd[0]->port_pkeys[index];
1859
1860         return ret;
1861 }
1862
1863 static int ipath_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1864                             u16 *pkey)
1865 {
1866         struct ipath_ibdev *dev = to_idev(ibdev);
1867         int ret;
1868
1869         if (index >= ipath_get_npkeys(dev->dd)) {
1870                 ret = -EINVAL;
1871                 goto bail;
1872         }
1873
1874         *pkey = ipath_get_pkey(dev->dd, index);
1875         ret = 0;
1876
1877 bail:
1878         return ret;
1879 }
1880
1881 /**
1882  * ipath_alloc_ucontext - allocate a ucontest
1883  * @ibdev: the infiniband device
1884  * @udata: not used by the InfiniPath driver
1885  */
1886
1887 static struct ib_ucontext *ipath_alloc_ucontext(struct ib_device *ibdev,
1888                                                 struct ib_udata *udata)
1889 {
1890         struct ipath_ucontext *context;
1891         struct ib_ucontext *ret;
1892
1893         context = kmalloc(sizeof *context, GFP_KERNEL);
1894         if (!context) {
1895                 ret = ERR_PTR(-ENOMEM);
1896                 goto bail;
1897         }
1898
1899         ret = &context->ibucontext;
1900
1901 bail:
1902         return ret;
1903 }
1904
1905 static int ipath_dealloc_ucontext(struct ib_ucontext *context)
1906 {
1907         kfree(to_iucontext(context));
1908         return 0;
1909 }
1910
1911 static int ipath_verbs_register_sysfs(struct ib_device *dev);
1912
1913 static void __verbs_timer(unsigned long arg)
1914 {
1915         struct ipath_devdata *dd = (struct ipath_devdata *) arg;
1916
1917         /* Handle verbs layer timeouts. */
1918         ipath_ib_timer(dd->verbs_dev);
1919
1920         mod_timer(&dd->verbs_timer, jiffies + 1);
1921 }
1922
1923 static int enable_timer(struct ipath_devdata *dd)
1924 {
1925         /*
1926          * Early chips had a design flaw where the chip and kernel idea
1927          * of the tail register don't always agree, and therefore we won't
1928          * get an interrupt on the next packet received.
1929          * If the board supports per packet receive interrupts, use it.
1930          * Otherwise, the timer function periodically checks for packets
1931          * to cover this case.
1932          * Either way, the timer is needed for verbs layer related
1933          * processing.
1934          */
1935         if (dd->ipath_flags & IPATH_GPIO_INTR) {
1936                 ipath_write_kreg(dd, dd->ipath_kregs->kr_debugportselect,
1937                                  0x2074076542310ULL);
1938                 /* Enable GPIO bit 2 interrupt */
1939                 dd->ipath_gpio_mask |= (u64) (1 << IPATH_GPIO_PORT0_BIT);
1940                 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1941                                  dd->ipath_gpio_mask);
1942         }
1943
1944         init_timer(&dd->verbs_timer);
1945         dd->verbs_timer.function = __verbs_timer;
1946         dd->verbs_timer.data = (unsigned long)dd;
1947         dd->verbs_timer.expires = jiffies + 1;
1948         add_timer(&dd->verbs_timer);
1949
1950         return 0;
1951 }
1952
1953 static int disable_timer(struct ipath_devdata *dd)
1954 {
1955         /* Disable GPIO bit 2 interrupt */
1956         if (dd->ipath_flags & IPATH_GPIO_INTR) {
1957                 /* Disable GPIO bit 2 interrupt */
1958                 dd->ipath_gpio_mask &= ~((u64) (1 << IPATH_GPIO_PORT0_BIT));
1959                 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1960                                  dd->ipath_gpio_mask);
1961                 /*
1962                  * We might want to undo changes to debugportselect,
1963                  * but how?
1964                  */
1965         }
1966
1967         del_timer_sync(&dd->verbs_timer);
1968
1969         return 0;
1970 }
1971
1972 /**
1973  * ipath_register_ib_device - register our device with the infiniband core
1974  * @dd: the device data structure
1975  * Return the allocated ipath_ibdev pointer or NULL on error.
1976  */
1977 int ipath_register_ib_device(struct ipath_devdata *dd)
1978 {
1979         struct ipath_verbs_counters cntrs;
1980         struct ipath_ibdev *idev;
1981         struct ib_device *dev;
1982         struct ipath_verbs_txreq *tx;
1983         unsigned i;
1984         int ret;
1985
1986         idev = (struct ipath_ibdev *)ib_alloc_device(sizeof *idev);
1987         if (idev == NULL) {
1988                 ret = -ENOMEM;
1989                 goto bail;
1990         }
1991
1992         dev = &idev->ibdev;
1993
1994         if (dd->ipath_sdma_descq_cnt) {
1995                 tx = kmalloc(dd->ipath_sdma_descq_cnt * sizeof *tx,
1996                              GFP_KERNEL);
1997                 if (tx == NULL) {
1998                         ret = -ENOMEM;
1999                         goto err_tx;
2000                 }
2001         } else
2002                 tx = NULL;
2003         idev->txreq_bufs = tx;
2004
2005         /* Only need to initialize non-zero fields. */
2006         spin_lock_init(&idev->n_pds_lock);
2007         spin_lock_init(&idev->n_ahs_lock);
2008         spin_lock_init(&idev->n_cqs_lock);
2009         spin_lock_init(&idev->n_qps_lock);
2010         spin_lock_init(&idev->n_srqs_lock);
2011         spin_lock_init(&idev->n_mcast_grps_lock);
2012
2013         spin_lock_init(&idev->qp_table.lock);
2014         spin_lock_init(&idev->lk_table.lock);
2015         idev->sm_lid = __constant_be16_to_cpu(IB_LID_PERMISSIVE);
2016         /* Set the prefix to the default value (see ch. 4.1.1) */
2017         idev->gid_prefix = __constant_cpu_to_be64(0xfe80000000000000ULL);
2018
2019         ret = ipath_init_qp_table(idev, ib_ipath_qp_table_size);
2020         if (ret)
2021                 goto err_qp;
2022
2023         /*
2024          * The top ib_ipath_lkey_table_size bits are used to index the
2025          * table.  The lower 8 bits can be owned by the user (copied from
2026          * the LKEY).  The remaining bits act as a generation number or tag.
2027          */
2028         idev->lk_table.max = 1 << ib_ipath_lkey_table_size;
2029         idev->lk_table.table = kzalloc(idev->lk_table.max *
2030                                        sizeof(*idev->lk_table.table),
2031                                        GFP_KERNEL);
2032         if (idev->lk_table.table == NULL) {
2033                 ret = -ENOMEM;
2034                 goto err_lk;
2035         }
2036         INIT_LIST_HEAD(&idev->pending_mmaps);
2037         spin_lock_init(&idev->pending_lock);
2038         idev->mmap_offset = PAGE_SIZE;
2039         spin_lock_init(&idev->mmap_offset_lock);
2040         INIT_LIST_HEAD(&idev->pending[0]);
2041         INIT_LIST_HEAD(&idev->pending[1]);
2042         INIT_LIST_HEAD(&idev->pending[2]);
2043         INIT_LIST_HEAD(&idev->piowait);
2044         INIT_LIST_HEAD(&idev->rnrwait);
2045         INIT_LIST_HEAD(&idev->txreq_free);
2046         idev->pending_index = 0;
2047         idev->port_cap_flags =
2048                 IB_PORT_SYS_IMAGE_GUID_SUP | IB_PORT_CLIENT_REG_SUP;
2049         if (dd->ipath_flags & IPATH_HAS_LINK_LATENCY)
2050                 idev->port_cap_flags |= IB_PORT_LINK_LATENCY_SUP;
2051         idev->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
2052         idev->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
2053         idev->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
2054         idev->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
2055         idev->pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
2056
2057         /* Snapshot current HW counters to "clear" them. */
2058         ipath_get_counters(dd, &cntrs);
2059         idev->z_symbol_error_counter = cntrs.symbol_error_counter;
2060         idev->z_link_error_recovery_counter =
2061                 cntrs.link_error_recovery_counter;
2062         idev->z_link_downed_counter = cntrs.link_downed_counter;
2063         idev->z_port_rcv_errors = cntrs.port_rcv_errors;
2064         idev->z_port_rcv_remphys_errors =
2065                 cntrs.port_rcv_remphys_errors;
2066         idev->z_port_xmit_discards = cntrs.port_xmit_discards;
2067         idev->z_port_xmit_data = cntrs.port_xmit_data;
2068         idev->z_port_rcv_data = cntrs.port_rcv_data;
2069         idev->z_port_xmit_packets = cntrs.port_xmit_packets;
2070         idev->z_port_rcv_packets = cntrs.port_rcv_packets;
2071         idev->z_local_link_integrity_errors =
2072                 cntrs.local_link_integrity_errors;
2073         idev->z_excessive_buffer_overrun_errors =
2074                 cntrs.excessive_buffer_overrun_errors;
2075         idev->z_vl15_dropped = cntrs.vl15_dropped;
2076
2077         for (i = 0; i < dd->ipath_sdma_descq_cnt; i++, tx++)
2078                 list_add(&tx->txreq.list, &idev->txreq_free);
2079
2080         /*
2081          * The system image GUID is supposed to be the same for all
2082          * IB HCAs in a single system but since there can be other
2083          * device types in the system, we can't be sure this is unique.
2084          */
2085         if (!sys_image_guid)
2086                 sys_image_guid = dd->ipath_guid;
2087         idev->sys_image_guid = sys_image_guid;
2088         idev->ib_unit = dd->ipath_unit;
2089         idev->dd = dd;
2090
2091         strlcpy(dev->name, "ipath%d", IB_DEVICE_NAME_MAX);
2092         dev->owner = THIS_MODULE;
2093         dev->node_guid = dd->ipath_guid;
2094         dev->uverbs_abi_ver = IPATH_UVERBS_ABI_VERSION;
2095         dev->uverbs_cmd_mask =
2096                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
2097                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
2098                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
2099                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
2100                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
2101                 (1ull << IB_USER_VERBS_CMD_CREATE_AH)           |
2102                 (1ull << IB_USER_VERBS_CMD_DESTROY_AH)          |
2103                 (1ull << IB_USER_VERBS_CMD_QUERY_AH)            |
2104                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
2105                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
2106                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
2107                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
2108                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
2109                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
2110                 (1ull << IB_USER_VERBS_CMD_POLL_CQ)             |
2111                 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)       |
2112                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
2113                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
2114                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
2115                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
2116                 (1ull << IB_USER_VERBS_CMD_POST_SEND)           |
2117                 (1ull << IB_USER_VERBS_CMD_POST_RECV)           |
2118                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
2119                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
2120                 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
2121                 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
2122                 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
2123                 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
2124                 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
2125         dev->node_type = RDMA_NODE_IB_CA;
2126         dev->phys_port_cnt = 1;
2127         dev->num_comp_vectors = 1;
2128         dev->dma_device = &dd->pcidev->dev;
2129         dev->query_device = ipath_query_device;
2130         dev->modify_device = ipath_modify_device;
2131         dev->query_port = ipath_query_port;
2132         dev->modify_port = ipath_modify_port;
2133         dev->query_pkey = ipath_query_pkey;
2134         dev->query_gid = ipath_query_gid;
2135         dev->alloc_ucontext = ipath_alloc_ucontext;
2136         dev->dealloc_ucontext = ipath_dealloc_ucontext;
2137         dev->alloc_pd = ipath_alloc_pd;
2138         dev->dealloc_pd = ipath_dealloc_pd;
2139         dev->create_ah = ipath_create_ah;
2140         dev->destroy_ah = ipath_destroy_ah;
2141         dev->query_ah = ipath_query_ah;
2142         dev->create_srq = ipath_create_srq;
2143         dev->modify_srq = ipath_modify_srq;
2144         dev->query_srq = ipath_query_srq;
2145         dev->destroy_srq = ipath_destroy_srq;
2146         dev->create_qp = ipath_create_qp;
2147         dev->modify_qp = ipath_modify_qp;
2148         dev->query_qp = ipath_query_qp;
2149         dev->destroy_qp = ipath_destroy_qp;
2150         dev->post_send = ipath_post_send;
2151         dev->post_recv = ipath_post_receive;
2152         dev->post_srq_recv = ipath_post_srq_receive;
2153         dev->create_cq = ipath_create_cq;
2154         dev->destroy_cq = ipath_destroy_cq;
2155         dev->resize_cq = ipath_resize_cq;
2156         dev->poll_cq = ipath_poll_cq;
2157         dev->req_notify_cq = ipath_req_notify_cq;
2158         dev->get_dma_mr = ipath_get_dma_mr;
2159         dev->reg_phys_mr = ipath_reg_phys_mr;
2160         dev->reg_user_mr = ipath_reg_user_mr;
2161         dev->dereg_mr = ipath_dereg_mr;
2162         dev->alloc_fmr = ipath_alloc_fmr;
2163         dev->map_phys_fmr = ipath_map_phys_fmr;
2164         dev->unmap_fmr = ipath_unmap_fmr;
2165         dev->dealloc_fmr = ipath_dealloc_fmr;
2166         dev->attach_mcast = ipath_multicast_attach;
2167         dev->detach_mcast = ipath_multicast_detach;
2168         dev->process_mad = ipath_process_mad;
2169         dev->mmap = ipath_mmap;
2170         dev->dma_ops = &ipath_dma_mapping_ops;
2171
2172         snprintf(dev->node_desc, sizeof(dev->node_desc),
2173                  IPATH_IDSTR " %s", init_utsname()->nodename);
2174
2175         ret = ib_register_device(dev);
2176         if (ret)
2177                 goto err_reg;
2178
2179         if (ipath_verbs_register_sysfs(dev))
2180                 goto err_class;
2181
2182         enable_timer(dd);
2183
2184         goto bail;
2185
2186 err_class:
2187         ib_unregister_device(dev);
2188 err_reg:
2189         kfree(idev->lk_table.table);
2190 err_lk:
2191         kfree(idev->qp_table.table);
2192 err_qp:
2193         kfree(idev->txreq_bufs);
2194 err_tx:
2195         ib_dealloc_device(dev);
2196         ipath_dev_err(dd, "cannot register verbs: %d!\n", -ret);
2197         idev = NULL;
2198
2199 bail:
2200         dd->verbs_dev = idev;
2201         return ret;
2202 }
2203
2204 void ipath_unregister_ib_device(struct ipath_ibdev *dev)
2205 {
2206         struct ib_device *ibdev = &dev->ibdev;
2207         u32 qps_inuse;
2208
2209         ib_unregister_device(ibdev);
2210
2211         disable_timer(dev->dd);
2212
2213         if (!list_empty(&dev->pending[0]) ||
2214             !list_empty(&dev->pending[1]) ||
2215             !list_empty(&dev->pending[2]))
2216                 ipath_dev_err(dev->dd, "pending list not empty!\n");
2217         if (!list_empty(&dev->piowait))
2218                 ipath_dev_err(dev->dd, "piowait list not empty!\n");
2219         if (!list_empty(&dev->rnrwait))
2220                 ipath_dev_err(dev->dd, "rnrwait list not empty!\n");
2221         if (!ipath_mcast_tree_empty())
2222                 ipath_dev_err(dev->dd, "multicast table memory leak!\n");
2223         /*
2224          * Note that ipath_unregister_ib_device() can be called before all
2225          * the QPs are destroyed!
2226          */
2227         qps_inuse = ipath_free_all_qps(&dev->qp_table);
2228         if (qps_inuse)
2229                 ipath_dev_err(dev->dd, "QP memory leak! %u still in use\n",
2230                         qps_inuse);
2231         kfree(dev->qp_table.table);
2232         kfree(dev->lk_table.table);
2233         kfree(dev->txreq_bufs);
2234         ib_dealloc_device(ibdev);
2235 }
2236
2237 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2238                         char *buf)
2239 {
2240         struct ipath_ibdev *dev =
2241                 container_of(device, struct ipath_ibdev, ibdev.dev);
2242
2243         return sprintf(buf, "%x\n", dev->dd->ipath_pcirev);
2244 }
2245
2246 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2247                         char *buf)
2248 {
2249         struct ipath_ibdev *dev =
2250                 container_of(device, struct ipath_ibdev, ibdev.dev);
2251         int ret;
2252
2253         ret = dev->dd->ipath_f_get_boardname(dev->dd, buf, 128);
2254         if (ret < 0)
2255                 goto bail;
2256         strcat(buf, "\n");
2257         ret = strlen(buf);
2258
2259 bail:
2260         return ret;
2261 }
2262
2263 static ssize_t show_stats(struct device *device, struct device_attribute *attr,
2264                           char *buf)
2265 {
2266         struct ipath_ibdev *dev =
2267                 container_of(device, struct ipath_ibdev, ibdev.dev);
2268         int i;
2269         int len;
2270
2271         len = sprintf(buf,
2272                       "RC resends  %d\n"
2273                       "RC no QACK  %d\n"
2274                       "RC ACKs     %d\n"
2275                       "RC SEQ NAKs %d\n"
2276                       "RC RDMA seq %d\n"
2277                       "RC RNR NAKs %d\n"
2278                       "RC OTH NAKs %d\n"
2279                       "RC timeouts %d\n"
2280                       "RC RDMA dup %d\n"
2281                       "piobuf wait %d\n"
2282                       "unaligned   %d\n"
2283                       "PKT drops   %d\n"
2284                       "WQE errs    %d\n",
2285                       dev->n_rc_resends, dev->n_rc_qacks, dev->n_rc_acks,
2286                       dev->n_seq_naks, dev->n_rdma_seq, dev->n_rnr_naks,
2287                       dev->n_other_naks, dev->n_timeouts,
2288                       dev->n_rdma_dup_busy, dev->n_piowait, dev->n_unaligned,
2289                       dev->n_pkt_drops, dev->n_wqe_errs);
2290         for (i = 0; i < ARRAY_SIZE(dev->opstats); i++) {
2291                 const struct ipath_opcode_stats *si = &dev->opstats[i];
2292
2293                 if (!si->n_packets && !si->n_bytes)
2294                         continue;
2295                 len += sprintf(buf + len, "%02x %llu/%llu\n", i,
2296                                (unsigned long long) si->n_packets,
2297                                (unsigned long long) si->n_bytes);
2298         }
2299         return len;
2300 }
2301
2302 static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
2303 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
2304 static DEVICE_ATTR(board_id, S_IRUGO, show_hca, NULL);
2305 static DEVICE_ATTR(stats, S_IRUGO, show_stats, NULL);
2306
2307 static struct device_attribute *ipath_class_attributes[] = {
2308         &dev_attr_hw_rev,
2309         &dev_attr_hca_type,
2310         &dev_attr_board_id,
2311         &dev_attr_stats
2312 };
2313
2314 static int ipath_verbs_register_sysfs(struct ib_device *dev)
2315 {
2316         int i;
2317         int ret;
2318
2319         for (i = 0; i < ARRAY_SIZE(ipath_class_attributes); ++i)
2320                 if (device_create_file(&dev->dev,
2321                                        ipath_class_attributes[i])) {
2322                         ret = 1;
2323                         goto bail;
2324                 }
2325
2326         ret = 0;
2327
2328 bail:
2329         return ret;
2330 }