drivers/staging/rdma/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/slab.h>
  38 #include <linux/module.h>
  39 #include <linux/utsname.h>
  40 #include <linux/rculist.h>
  41
  42 #include "ipath_kernel.h"
  43 #include "ipath_verbs.h"
  44 #include "ipath_common.h"
  45
  46 static unsigned int ib_ipath_qp_table_size = 251;
  47 module_param_named(qp_table_size, ib_ipath_qp_table_size, uint, S_IRUGO);
  48 MODULE_PARM_DESC(qp_table_size, "QP table size");
  49
  50 unsigned int ib_ipath_lkey_table_size = 12;
  51 module_param_named(lkey_table_size, ib_ipath_lkey_table_size, uint,
  52                    S_IRUGO);
  53 MODULE_PARM_DESC(lkey_table_size,
  54                  "LKEY table size in bits (2^n, 1 <= n <= 23)");
  55
  56 static unsigned int ib_ipath_max_pds = 0xFFFF;
  57 module_param_named(max_pds, ib_ipath_max_pds, uint, S_IWUSR | S_IRUGO);
  58 MODULE_PARM_DESC(max_pds,
  59                  "Maximum number of protection domains to support");
  60
  61 static unsigned int ib_ipath_max_ahs = 0xFFFF;
  62 module_param_named(max_ahs, ib_ipath_max_ahs, uint, S_IWUSR | S_IRUGO);
  63 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
  64
  65 unsigned int ib_ipath_max_cqes = 0x2FFFF;
  66 module_param_named(max_cqes, ib_ipath_max_cqes, uint, S_IWUSR | S_IRUGO);
  67 MODULE_PARM_DESC(max_cqes,
  68                  "Maximum number of completion queue entries to support");
  69
  70 unsigned int ib_ipath_max_cqs = 0x1FFFF;
  71 module_param_named(max_cqs, ib_ipath_max_cqs, uint, S_IWUSR | S_IRUGO);
  72 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
  73
  74 unsigned int ib_ipath_max_qp_wrs = 0x3FFF;
  75 module_param_named(max_qp_wrs, ib_ipath_max_qp_wrs, uint,
  76                    S_IWUSR | S_IRUGO);
  77 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
  78
  79 unsigned int ib_ipath_max_qps = 16384;
  80 module_param_named(max_qps, ib_ipath_max_qps, uint, S_IWUSR | S_IRUGO);
  81 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
  82
  83 unsigned int ib_ipath_max_sges = 0x60;
  84 module_param_named(max_sges, ib_ipath_max_sges, uint, S_IWUSR | S_IRUGO);
  85 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
  86
  87 unsigned int ib_ipath_max_mcast_grps = 16384;
  88 module_param_named(max_mcast_grps, ib_ipath_max_mcast_grps, uint,
  89                    S_IWUSR | S_IRUGO);
  90 MODULE_PARM_DESC(max_mcast_grps,
  91                  "Maximum number of multicast groups to support");
  92
  93 unsigned int ib_ipath_max_mcast_qp_attached = 16;
  94 module_param_named(max_mcast_qp_attached, ib_ipath_max_mcast_qp_attached,
  95                    uint, S_IWUSR | S_IRUGO);
  96 MODULE_PARM_DESC(max_mcast_qp_attached,
  97                  "Maximum number of attached QPs to support");
  98
  99 unsigned int ib_ipath_max_srqs = 1024;
 100 module_param_named(max_srqs, ib_ipath_max_srqs, uint, S_IWUSR | S_IRUGO);
 101 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
 102
 103 unsigned int ib_ipath_max_srq_sges = 128;
 104 module_param_named(max_srq_sges, ib_ipath_max_srq_sges,
 105                    uint, S_IWUSR | S_IRUGO);
 106 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
 107
 108 unsigned int ib_ipath_max_srq_wrs = 0x1FFFF;
 109 module_param_named(max_srq_wrs, ib_ipath_max_srq_wrs,
 110                    uint, S_IWUSR | S_IRUGO);
 111 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
 112
 113 static unsigned int ib_ipath_disable_sma;
 114 module_param_named(disable_sma, ib_ipath_disable_sma, uint, S_IWUSR | S_IRUGO);
 115 MODULE_PARM_DESC(disable_sma, "Disable the SMA");
 116
 117 /*
 118  * Note that it is OK to post send work requests in the SQE and ERR
 119  * states; ipath_do_send() will process them and generate error
 120  * completions as per IB 1.2 C10-96.
 121  */
 122 const int ib_ipath_state_ops[IB_QPS_ERR + 1] = {
 123         [IB_QPS_RESET] = 0,
 124         [IB_QPS_INIT] = IPATH_POST_RECV_OK,
 125         [IB_QPS_RTR] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK,
 126         [IB_QPS_RTS] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
 127             IPATH_POST_SEND_OK | IPATH_PROCESS_SEND_OK |
 128             IPATH_PROCESS_NEXT_SEND_OK,
 129         [IB_QPS_SQD] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
 130             IPATH_POST_SEND_OK | IPATH_PROCESS_SEND_OK,
 131         [IB_QPS_SQE] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
 132             IPATH_POST_SEND_OK | IPATH_FLUSH_SEND,
 133         [IB_QPS_ERR] = IPATH_POST_RECV_OK | IPATH_FLUSH_RECV |
 134             IPATH_POST_SEND_OK | IPATH_FLUSH_SEND,
 135 };
 136
 137 struct ipath_ucontext {
 138         struct ib_ucontext ibucontext;
 139 };
 140
 141 static inline struct ipath_ucontext *to_iucontext(struct ib_ucontext
 142                                                   *ibucontext)
 143 {
 144         return container_of(ibucontext, struct ipath_ucontext, ibucontext);
 145 }
 146
 147 /*
 148  * Translate ib_wr_opcode into ib_wc_opcode.
 149  */
 150 const enum ib_wc_opcode ib_ipath_wc_opcode[] = {
 151         [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
 152         [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
 153         [IB_WR_SEND] = IB_WC_SEND,
 154         [IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
 155         [IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
 156         [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
 157         [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD
 158 };
 159
 160 /*
 161  * System image GUID.
 162  */
 163 static __be64 sys_image_guid;
 164
 165 /**
 166  * ipath_copy_sge - copy data to SGE memory
 167  * @ss: the SGE state
 168  * @data: the data to copy
 169  * @length: the length of the data
 170  */
 171 void ipath_copy_sge(struct ipath_sge_state *ss, void *data, u32 length)
 172 {
 173         struct ipath_sge *sge = &ss->sge;
 174
 175         while (length) {
 176                 u32 len = sge->length;
 177
 178                 if (len > length)
 179                         len = length;
 180                 if (len > sge->sge_length)
 181                         len = sge->sge_length;
 182                 BUG_ON(len == 0);
 183                 memcpy(sge->vaddr, data, len);
 184                 sge->vaddr += len;
 185                 sge->length -= len;
 186                 sge->sge_length -= len;
 187                 if (sge->sge_length == 0) {
 188                         if (--ss->num_sge)
 189                                 *sge = *ss->sg_list++;
 190                 } else if (sge->length == 0 && sge->mr != NULL) {
 191                         if (++sge->n >= IPATH_SEGSZ) {
 192                                 if (++sge->m >= sge->mr->mapsz)
 193                                         break;
 194                                 sge->n = 0;
 195                         }
 196                         sge->vaddr =
 197                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 198                         sge->length =
 199                                 sge->mr->map[sge->m]->segs[sge->n].length;
 200                 }
 201                 data += len;
 202                 length -= len;
 203         }
 204 }
 205
 206 /**
 207  * ipath_skip_sge - skip over SGE memory - XXX almost dup of prev func
 208  * @ss: the SGE state
 209  * @length: the number of bytes to skip
 210  */
 211 void ipath_skip_sge(struct ipath_sge_state *ss, u32 length)
 212 {
 213         struct ipath_sge *sge = &ss->sge;
 214
 215         while (length) {
 216                 u32 len = sge->length;
 217
 218                 if (len > length)
 219                         len = length;
 220                 if (len > sge->sge_length)
 221                         len = sge->sge_length;
 222                 BUG_ON(len == 0);
 223                 sge->vaddr += len;
 224                 sge->length -= len;
 225                 sge->sge_length -= len;
 226                 if (sge->sge_length == 0) {
 227                         if (--ss->num_sge)
 228                                 *sge = *ss->sg_list++;
 229                 } else if (sge->length == 0 && sge->mr != NULL) {
 230                         if (++sge->n >= IPATH_SEGSZ) {
 231                                 if (++sge->m >= sge->mr->mapsz)
 232                                         break;
 233                                 sge->n = 0;
 234                         }
 235                         sge->vaddr =
 236                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 237                         sge->length =
 238                                 sge->mr->map[sge->m]->segs[sge->n].length;
 239                 }
 240                 length -= len;
 241         }
 242 }
 243
 244 /*
 245  * Count the number of DMA descriptors needed to send length bytes of data.
 246  * Don't modify the ipath_sge_state to get the count.
 247  * Return zero if any of the segments is not aligned.
 248  */
 249 static u32 ipath_count_sge(struct ipath_sge_state *ss, u32 length)
 250 {
 251         struct ipath_sge *sg_list = ss->sg_list;
 252         struct ipath_sge sge = ss->sge;
 253         u8 num_sge = ss->num_sge;
 254         u32 ndesc = 1;  /* count the header */
 255
 256         while (length) {
 257                 u32 len = sge.length;
 258
 259                 if (len > length)
 260                         len = length;
 261                 if (len > sge.sge_length)
 262                         len = sge.sge_length;
 263                 BUG_ON(len == 0);
 264                 if (((long) sge.vaddr & (sizeof(u32) - 1)) ||
 265                     (len != length && (len & (sizeof(u32) - 1)))) {
 266                         ndesc = 0;
 267                         break;
 268                 }
 269                 ndesc++;
 270                 sge.vaddr += len;
 271                 sge.length -= len;
 272                 sge.sge_length -= len;
 273                 if (sge.sge_length == 0) {
 274                         if (--num_sge)
 275                                 sge = *sg_list++;
 276                 } else if (sge.length == 0 && sge.mr != NULL) {
 277                         if (++sge.n >= IPATH_SEGSZ) {
 278                                 if (++sge.m >= sge.mr->mapsz)
 279                                         break;
 280                                 sge.n = 0;
 281                         }
 282                         sge.vaddr =
 283                                 sge.mr->map[sge.m]->segs[sge.n].vaddr;
 284                         sge.length =
 285                                 sge.mr->map[sge.m]->segs[sge.n].length;
 286                 }
 287                 length -= len;
 288         }
 289         return ndesc;
 290 }
 291
 292 /*
 293  * Copy from the SGEs to the data buffer.
 294  */
 295 static void ipath_copy_from_sge(void *data, struct ipath_sge_state *ss,
 296                                 u32 length)
 297 {
 298         struct ipath_sge *sge = &ss->sge;
 299
 300         while (length) {
 301                 u32 len = sge->length;
 302
 303                 if (len > length)
 304                         len = length;
 305                 if (len > sge->sge_length)
 306                         len = sge->sge_length;
 307                 BUG_ON(len == 0);
 308                 memcpy(data, sge->vaddr, len);
 309                 sge->vaddr += len;
 310                 sge->length -= len;
 311                 sge->sge_length -= len;
 312                 if (sge->sge_length == 0) {
 313                         if (--ss->num_sge)
 314                                 *sge = *ss->sg_list++;
 315                 } else if (sge->length == 0 && sge->mr != NULL) {
 316                         if (++sge->n >= IPATH_SEGSZ) {
 317                                 if (++sge->m >= sge->mr->mapsz)
 318                                         break;
 319                                 sge->n = 0;
 320                         }
 321                         sge->vaddr =
 322                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 323                         sge->length =
 324                                 sge->mr->map[sge->m]->segs[sge->n].length;
 325                 }
 326                 data += len;
 327                 length -= len;
 328         }
 329 }
 330
 331 /**
 332  * ipath_post_one_send - post one RC, UC, or UD send work request
 333  * @qp: the QP to post on
 334  * @wr: the work request to send
 335  */
 336 static int ipath_post_one_send(struct ipath_qp *qp, struct ib_send_wr *wr)
 337 {
 338         struct ipath_swqe *wqe;
 339         u32 next;
 340         int i;
 341         int j;
 342         int acc;
 343         int ret;
 344         unsigned long flags;
 345         struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
 346
 347         spin_lock_irqsave(&qp->s_lock, flags);
 348
 349         if (qp->ibqp.qp_type != IB_QPT_SMI &&
 350             !(dd->ipath_flags & IPATH_LINKACTIVE)) {
 351                 ret = -ENETDOWN;
 352                 goto bail;
 353         }
 354
 355         /* Check that state is OK to post send. */
 356         if (unlikely(!(ib_ipath_state_ops[qp->state] & IPATH_POST_SEND_OK)))
 357                 goto bail_inval;
 358
 359         /* IB spec says that num_sge == 0 is OK. */
 360         if (wr->num_sge > qp->s_max_sge)
 361                 goto bail_inval;
 362
 363         /*
 364          * Don't allow RDMA reads or atomic operations on UC or
 365          * undefined operations.
 366          * Make sure buffer is large enough to hold the result for atomics.
 367          */
 368         if (qp->ibqp.qp_type == IB_QPT_UC) {
 369                 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ)
 370                         goto bail_inval;
 371         } else if (qp->ibqp.qp_type == IB_QPT_UD) {
 372                 /* Check UD opcode */
 373                 if (wr->opcode != IB_WR_SEND &&
 374                     wr->opcode != IB_WR_SEND_WITH_IMM)
 375                         goto bail_inval;
 376                 /* Check UD destination address PD */
 377                 if (qp->ibqp.pd != ud_wr(wr)->ah->pd)
 378                         goto bail_inval;
 379         } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD)
 380                 goto bail_inval;
 381         else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
 382                    (wr->num_sge == 0 ||
 383                     wr->sg_list[0].length < sizeof(u64) ||
 384                     wr->sg_list[0].addr & (sizeof(u64) - 1)))
 385                 goto bail_inval;
 386         else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic)
 387                 goto bail_inval;
 388
 389         next = qp->s_head + 1;
 390         if (next >= qp->s_size)
 391                 next = 0;
 392         if (next == qp->s_last) {
 393                 ret = -ENOMEM;
 394                 goto bail;
 395         }
 396
 397         wqe = get_swqe_ptr(qp, qp->s_head);
 398
 399         if (qp->ibqp.qp_type != IB_QPT_UC &&
 400             qp->ibqp.qp_type != IB_QPT_RC)
 401                 memcpy(&wqe->ud_wr, ud_wr(wr), sizeof(wqe->ud_wr));
 402         else if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
 403                  wr->opcode == IB_WR_RDMA_WRITE ||
 404                  wr->opcode == IB_WR_RDMA_READ)
 405                 memcpy(&wqe->rdma_wr, rdma_wr(wr), sizeof(wqe->rdma_wr));
 406         else if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 407                  wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
 408                 memcpy(&wqe->atomic_wr, atomic_wr(wr), sizeof(wqe->atomic_wr));
 409         else
 410                 memcpy(&wqe->wr, wr, sizeof(wqe->wr));
 411
 412         wqe->length = 0;
 413         if (wr->num_sge) {
 414                 acc = wr->opcode >= IB_WR_RDMA_READ ?
 415                         IB_ACCESS_LOCAL_WRITE : 0;
 416                 for (i = 0, j = 0; i < wr->num_sge; i++) {
 417                         u32 length = wr->sg_list[i].length;
 418                         int ok;
 419
 420                         if (length == 0)
 421                                 continue;
 422                         ok = ipath_lkey_ok(qp, &wqe->sg_list[j],
 423                                            &wr->sg_list[i], acc);
 424                         if (!ok)
 425                                 goto bail_inval;
 426                         wqe->length += length;
 427                         j++;
 428                 }
 429                 wqe->wr.num_sge = j;
 430         }
 431         if (qp->ibqp.qp_type == IB_QPT_UC ||
 432             qp->ibqp.qp_type == IB_QPT_RC) {
 433                 if (wqe->length > 0x80000000U)
 434                         goto bail_inval;
 435         } else if (wqe->length > to_idev(qp->ibqp.device)->dd->ipath_ibmtu)
 436                 goto bail_inval;
 437         wqe->ssn = qp->s_ssn++;
 438         qp->s_head = next;
 439
 440         ret = 0;
 441         goto bail;
 442
 443 bail_inval:
 444         ret = -EINVAL;
 445 bail:
 446         spin_unlock_irqrestore(&qp->s_lock, flags);
 447         return ret;
 448 }
 449
 450 /**
 451  * ipath_post_send - post a send on a QP
 452  * @ibqp: the QP to post the send on
 453  * @wr: the list of work requests to post
 454  * @bad_wr: the first bad WR is put here
 455  *
 456  * This may be called from interrupt context.
 457  */
 458 static int ipath_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 459                            struct ib_send_wr **bad_wr)
 460 {
 461         struct ipath_qp *qp = to_iqp(ibqp);
 462         int err = 0;
 463
 464         for (; wr; wr = wr->next) {
 465                 err = ipath_post_one_send(qp, wr);
 466                 if (err) {
 467                         *bad_wr = wr;
 468                         goto bail;
 469                 }
 470         }
 471
 472         /* Try to do the send work in the caller's context. */
 473         ipath_do_send((unsigned long) qp);
 474
 475 bail:
 476         return err;
 477 }
 478
 479 /**
 480  * ipath_post_receive - post a receive on a QP
 481  * @ibqp: the QP to post the receive on
 482  * @wr: the WR to post
 483  * @bad_wr: the first bad WR is put here
 484  *
 485  * This may be called from interrupt context.
 486  */
 487 static int ipath_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 488                               struct ib_recv_wr **bad_wr)
 489 {
 490         struct ipath_qp *qp = to_iqp(ibqp);
 491         struct ipath_rwq *wq = qp->r_rq.wq;
 492         unsigned long flags;
 493         int ret;
 494
 495         /* Check that state is OK to post receive. */
 496         if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_RECV_OK) || !wq) {
 497                 *bad_wr = wr;
 498                 ret = -EINVAL;
 499                 goto bail;
 500         }
 501
 502         for (; wr; wr = wr->next) {
 503                 struct ipath_rwqe *wqe;
 504                 u32 next;
 505                 int i;
 506
 507                 if ((unsigned) wr->num_sge > qp->r_rq.max_sge) {
 508                         *bad_wr = wr;
 509                         ret = -EINVAL;
 510                         goto bail;
 511                 }
 512
 513                 spin_lock_irqsave(&qp->r_rq.lock, flags);
 514                 next = wq->head + 1;
 515                 if (next >= qp->r_rq.size)
 516                         next = 0;
 517                 if (next == wq->tail) {
 518                         spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 519                         *bad_wr = wr;
 520                         ret = -ENOMEM;
 521                         goto bail;
 522                 }
 523
 524                 wqe = get_rwqe_ptr(&qp->r_rq, wq->head);
 525                 wqe->wr_id = wr->wr_id;
 526                 wqe->num_sge = wr->num_sge;
 527                 for (i = 0; i < wr->num_sge; i++)
 528                         wqe->sg_list[i] = wr->sg_list[i];
 529                 /* Make sure queue entry is written before the head index. */
 530                 smp_wmb();
 531                 wq->head = next;
 532                 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 533         }
 534         ret = 0;
 535
 536 bail:
 537         return ret;
 538 }
 539
 540 /**
 541  * ipath_qp_rcv - processing an incoming packet on a QP
 542  * @dev: the device the packet came on
 543  * @hdr: the packet header
 544  * @has_grh: true if the packet has a GRH
 545  * @data: the packet data
 546  * @tlen: the packet length
 547  * @qp: the QP the packet came on
 548  *
 549  * This is called from ipath_ib_rcv() to process an incoming packet
 550  * for the given QP.
 551  * Called at interrupt level.
 552  */
 553 static void ipath_qp_rcv(struct ipath_ibdev *dev,
 554                          struct ipath_ib_header *hdr, int has_grh,
 555                          void *data, u32 tlen, struct ipath_qp *qp)
 556 {
 557         /* Check for valid receive state. */
 558         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
 559                 dev->n_pkt_drops++;
 560                 return;
 561         }
 562
 563         switch (qp->ibqp.qp_type) {
 564         case IB_QPT_SMI:
 565         case IB_QPT_GSI:
 566                 if (ib_ipath_disable_sma)
 567                         break;
 568                 /* FALLTHROUGH */
 569         case IB_QPT_UD:
 570                 ipath_ud_rcv(dev, hdr, has_grh, data, tlen, qp);
 571                 break;
 572
 573         case IB_QPT_RC:
 574                 ipath_rc_rcv(dev, hdr, has_grh, data, tlen, qp);
 575                 break;
 576
 577         case IB_QPT_UC:
 578                 ipath_uc_rcv(dev, hdr, has_grh, data, tlen, qp);
 579                 break;
 580
 581         default:
 582                 break;
 583         }
 584 }
 585
 586 /**
 587  * ipath_ib_rcv - process an incoming packet
 588  * @arg: the device pointer
 589  * @rhdr: the header of the packet
 590  * @data: the packet data
 591  * @tlen: the packet length
 592  *
 593  * This is called from ipath_kreceive() to process an incoming packet at
 594  * interrupt level. Tlen is the length of the header + data + CRC in bytes.
 595  */
 596 void ipath_ib_rcv(struct ipath_ibdev *dev, void *rhdr, void *data,
 597                   u32 tlen)
 598 {
 599         struct ipath_ib_header *hdr = rhdr;
 600         struct ipath_other_headers *ohdr;
 601         struct ipath_qp *qp;
 602         u32 qp_num;
 603         int lnh;
 604         u8 opcode;
 605         u16 lid;
 606
 607         if (unlikely(dev == NULL))
 608                 goto bail;
 609
 610         if (unlikely(tlen < 24)) {      /* LRH+BTH+CRC */
 611                 dev->rcv_errors++;
 612                 goto bail;
 613         }
 614
 615         /* Check for a valid destination LID (see ch. 7.11.1). */
 616         lid = be16_to_cpu(hdr->lrh[1]);
 617         if (lid < IPATH_MULTICAST_LID_BASE) {
 618                 lid &= ~((1 << dev->dd->ipath_lmc) - 1);
 619                 if (unlikely(lid != dev->dd->ipath_lid)) {
 620                         dev->rcv_errors++;
 621                         goto bail;
 622                 }
 623         }
 624
 625         /* Check for GRH */
 626         lnh = be16_to_cpu(hdr->lrh[0]) & 3;
 627         if (lnh == IPATH_LRH_BTH)
 628                 ohdr = &hdr->u.oth;
 629         else if (lnh == IPATH_LRH_GRH)
 630                 ohdr = &hdr->u.l.oth;
 631         else {
 632                 dev->rcv_errors++;
 633                 goto bail;
 634         }
 635
 636         opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0x7f;
 637         dev->opstats[opcode].n_bytes += tlen;
 638         dev->opstats[opcode].n_packets++;
 639
 640         /* Get the destination QP number. */
 641         qp_num = be32_to_cpu(ohdr->bth[1]) & IPATH_QPN_MASK;
 642         if (qp_num == IPATH_MULTICAST_QPN) {
 643                 struct ipath_mcast *mcast;
 644                 struct ipath_mcast_qp *p;
 645
 646                 if (lnh != IPATH_LRH_GRH) {
 647                         dev->n_pkt_drops++;
 648                         goto bail;
 649                 }
 650                 mcast = ipath_mcast_find(&hdr->u.l.grh.dgid);
 651                 if (mcast == NULL) {
 652                         dev->n_pkt_drops++;
 653                         goto bail;
 654                 }
 655                 dev->n_multicast_rcv++;
 656                 list_for_each_entry_rcu(p, &mcast->qp_list, list)
 657                         ipath_qp_rcv(dev, hdr, 1, data, tlen, p->qp);
 658                 /*
 659                  * Notify ipath_multicast_detach() if it is waiting for us
 660                  * to finish.
 661                  */
 662                 if (atomic_dec_return(&mcast->refcount) <= 1)
 663                         wake_up(&mcast->wait);
 664         } else {
 665                 qp = ipath_lookup_qpn(&dev->qp_table, qp_num);
 666                 if (qp) {
 667                         dev->n_unicast_rcv++;
 668                         ipath_qp_rcv(dev, hdr, lnh == IPATH_LRH_GRH, data,
 669                                      tlen, qp);
 670                         /*
 671                          * Notify ipath_destroy_qp() if it is waiting
 672                          * for us to finish.
 673                          */
 674                         if (atomic_dec_and_test(&qp->refcount))
 675                                 wake_up(&qp->wait);
 676                 } else
 677                         dev->n_pkt_drops++;
 678         }
 679
 680 bail:;
 681 }
 682
 683 /**
 684  * ipath_ib_timer - verbs timer
 685  * @arg: the device pointer
 686  *
 687  * This is called from ipath_do_rcv_timer() at interrupt level to check for
 688  * QPs which need retransmits and to collect performance numbers.
 689  */
 690 static void ipath_ib_timer(struct ipath_ibdev *dev)
 691 {
 692         struct ipath_qp *resend = NULL;
 693         struct ipath_qp *rnr = NULL;
 694         struct list_head *last;
 695         struct ipath_qp *qp;
 696         unsigned long flags;
 697
 698         if (dev == NULL)
 699                 return;
 700
 701         spin_lock_irqsave(&dev->pending_lock, flags);
 702         /* Start filling the next pending queue. */
 703         if (++dev->pending_index >= ARRAY_SIZE(dev->pending))
 704                 dev->pending_index = 0;
 705         /* Save any requests still in the new queue, they have timed out. */
 706         last = &dev->pending[dev->pending_index];
 707         while (!list_empty(last)) {
 708                 qp = list_entry(last->next, struct ipath_qp, timerwait);
 709                 list_del_init(&qp->timerwait);
 710                 qp->timer_next = resend;
 711                 resend = qp;
 712                 atomic_inc(&qp->refcount);
 713         }
 714         last = &dev->rnrwait;
 715         if (!list_empty(last)) {
 716                 qp = list_entry(last->next, struct ipath_qp, timerwait);
 717                 if (--qp->s_rnr_timeout == 0) {
 718                         do {
 719                                 list_del_init(&qp->timerwait);
 720                                 qp->timer_next = rnr;
 721                                 rnr = qp;
 722                                 atomic_inc(&qp->refcount);
 723                                 if (list_empty(last))
 724                                         break;
 725                                 qp = list_entry(last->next, struct ipath_qp,
 726                                                 timerwait);
 727                         } while (qp->s_rnr_timeout == 0);
 728                 }
 729         }
 730         /*
 731          * We should only be in the started state if pma_sample_start != 0
 732          */
 733         if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_STARTED &&
 734             --dev->pma_sample_start == 0) {
 735                 dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_RUNNING;
 736                 ipath_snapshot_counters(dev->dd, &dev->ipath_sword,
 737                                         &dev->ipath_rword,
 738                                         &dev->ipath_spkts,
 739                                         &dev->ipath_rpkts,
 740                                         &dev->ipath_xmit_wait);
 741         }
 742         if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_RUNNING) {
 743                 if (dev->pma_sample_interval == 0) {
 744                         u64 ta, tb, tc, td, te;
 745
 746                         dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_DONE;
 747                         ipath_snapshot_counters(dev->dd, &ta, &tb,
 748                                                 &tc, &td, &te);
 749
 750                         dev->ipath_sword = ta - dev->ipath_sword;
 751                         dev->ipath_rword = tb - dev->ipath_rword;
 752                         dev->ipath_spkts = tc - dev->ipath_spkts;
 753                         dev->ipath_rpkts = td - dev->ipath_rpkts;
 754                         dev->ipath_xmit_wait = te - dev->ipath_xmit_wait;
 755                 } else {
 756                         dev->pma_sample_interval--;
 757                 }
 758         }
 759         spin_unlock_irqrestore(&dev->pending_lock, flags);
 760
 761         /* XXX What if timer fires again while this is running? */
 762         while (resend != NULL) {
 763                 qp = resend;
 764                 resend = qp->timer_next;
 765
 766                 spin_lock_irqsave(&qp->s_lock, flags);
 767                 if (qp->s_last != qp->s_tail &&
 768                     ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) {
 769                         dev->n_timeouts++;
 770                         ipath_restart_rc(qp, qp->s_last_psn + 1);
 771                 }
 772                 spin_unlock_irqrestore(&qp->s_lock, flags);
 773
 774                 /* Notify ipath_destroy_qp() if it is waiting. */
 775                 if (atomic_dec_and_test(&qp->refcount))
 776                         wake_up(&qp->wait);
 777         }
 778         while (rnr != NULL) {
 779                 qp = rnr;
 780                 rnr = qp->timer_next;
 781
 782                 spin_lock_irqsave(&qp->s_lock, flags);
 783                 if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)
 784                         ipath_schedule_send(qp);
 785                 spin_unlock_irqrestore(&qp->s_lock, flags);
 786
 787                 /* Notify ipath_destroy_qp() if it is waiting. */
 788                 if (atomic_dec_and_test(&qp->refcount))
 789                         wake_up(&qp->wait);
 790         }
 791 }
 792
 793 static void update_sge(struct ipath_sge_state *ss, u32 length)
 794 {
 795         struct ipath_sge *sge = &ss->sge;
 796
 797         sge->vaddr += length;
 798         sge->length -= length;
 799         sge->sge_length -= length;
 800         if (sge->sge_length == 0) {
 801                 if (--ss->num_sge)
 802                         *sge = *ss->sg_list++;
 803         } else if (sge->length == 0 && sge->mr != NULL) {
 804                 if (++sge->n >= IPATH_SEGSZ) {
 805                         if (++sge->m >= sge->mr->mapsz)
 806                                 return;
 807                         sge->n = 0;
 808                 }
 809                 sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
 810                 sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
 811         }
 812 }
 813
 814 #ifdef __LITTLE_ENDIAN
 815 static inline u32 get_upper_bits(u32 data, u32 shift)
 816 {
 817         return data >> shift;
 818 }
 819
 820 static inline u32 set_upper_bits(u32 data, u32 shift)
 821 {
 822         return data << shift;
 823 }
 824
 825 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
 826 {
 827         data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
 828         data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
 829         return data;
 830 }
 831 #else
 832 static inline u32 get_upper_bits(u32 data, u32 shift)
 833 {
 834         return data << shift;
 835 }
 836
 837 static inline u32 set_upper_bits(u32 data, u32 shift)
 838 {
 839         return data >> shift;
 840 }
 841
 842 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
 843 {
 844         data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
 845         data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
 846         return data;
 847 }
 848 #endif
 849
 850 static void copy_io(u32 __iomem *piobuf, struct ipath_sge_state *ss,
 851                     u32 length, unsigned flush_wc)
 852 {
 853         u32 extra = 0;
 854         u32 data = 0;
 855         u32 last;
 856
 857         while (1) {
 858                 u32 len = ss->sge.length;
 859                 u32 off;
 860
 861                 if (len > length)
 862                         len = length;
 863                 if (len > ss->sge.sge_length)
 864                         len = ss->sge.sge_length;
 865                 BUG_ON(len == 0);
 866                 /* If the source address is not aligned, try to align it. */
 867                 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
 868                 if (off) {
 869                         u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
 870                                             ~(sizeof(u32) - 1));
 871                         u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
 872                         u32 y;
 873
 874                         y = sizeof(u32) - off;
 875                         if (len > y)
 876                                 len = y;
 877                         if (len + extra >= sizeof(u32)) {
 878                                 data |= set_upper_bits(v, extra *
 879                                                        BITS_PER_BYTE);
 880                                 len = sizeof(u32) - extra;
 881                                 if (len == length) {
 882                                         last = data;
 883                                         break;
 884                                 }
 885                                 __raw_writel(data, piobuf);
 886                                 piobuf++;
 887                                 extra = 0;
 888                                 data = 0;
 889                         } else {
 890                                 /* Clear unused upper bytes */
 891                                 data |= clear_upper_bytes(v, len, extra);
 892                                 if (len == length) {
 893                                         last = data;
 894                                         break;
 895                                 }
 896                                 extra += len;
 897                         }
 898                 } else if (extra) {
 899                         /* Source address is aligned. */
 900                         u32 *addr = (u32 *) ss->sge.vaddr;
 901                         int shift = extra * BITS_PER_BYTE;
 902                         int ushift = 32 - shift;
 903                         u32 l = len;
 904
 905                         while (l >= sizeof(u32)) {
 906                                 u32 v = *addr;
 907
 908                                 data |= set_upper_bits(v, shift);
 909                                 __raw_writel(data, piobuf);
 910                                 data = get_upper_bits(v, ushift);
 911                                 piobuf++;
 912                                 addr++;
 913                                 l -= sizeof(u32);
 914                         }
 915                         /*
 916                          * We still have 'extra' number of bytes leftover.
 917                          */
 918                         if (l) {
 919                                 u32 v = *addr;
 920
 921                                 if (l + extra >= sizeof(u32)) {
 922                                         data |= set_upper_bits(v, shift);
 923                                         len -= l + extra - sizeof(u32);
 924                                         if (len == length) {
 925                                                 last = data;
 926                                                 break;
 927                                         }
 928                                         __raw_writel(data, piobuf);
 929                                         piobuf++;
 930                                         extra = 0;
 931                                         data = 0;
 932                                 } else {
 933                                         /* Clear unused upper bytes */
 934                                         data |= clear_upper_bytes(v, l,
 935                                                                   extra);
 936                                         if (len == length) {
 937                                                 last = data;
 938                                                 break;
 939                                         }
 940                                         extra += l;
 941                                 }
 942                         } else if (len == length) {
 943                                 last = data;
 944                                 break;
 945                         }
 946                 } else if (len == length) {
 947                         u32 w;
 948
 949                         /*
 950                          * Need to round up for the last dword in the
 951                          * packet.
 952                          */
 953                         w = (len + 3) >> 2;
 954                         __iowrite32_copy(piobuf, ss->sge.vaddr, w - 1);
 955                         piobuf += w - 1;
 956                         last = ((u32 *) ss->sge.vaddr)[w - 1];
 957                         break;
 958                 } else {
 959                         u32 w = len >> 2;
 960
 961                         __iowrite32_copy(piobuf, ss->sge.vaddr, w);
 962                         piobuf += w;
 963
 964                         extra = len & (sizeof(u32) - 1);
 965                         if (extra) {
 966                                 u32 v = ((u32 *) ss->sge.vaddr)[w];
 967
 968                                 /* Clear unused upper bytes */
 969                                 data = clear_upper_bytes(v, extra, 0);
 970                         }
 971                 }
 972                 update_sge(ss, len);
 973                 length -= len;
 974         }
 975         /* Update address before sending packet. */
 976         update_sge(ss, length);
 977         if (flush_wc) {
 978                 /* must flush early everything before trigger word */
 979                 ipath_flush_wc();
 980                 __raw_writel(last, piobuf);
 981                 /* be sure trigger word is written */
 982                 ipath_flush_wc();
 983         } else
 984                 __raw_writel(last, piobuf);
 985 }
 986
 987 /*
 988  * Convert IB rate to delay multiplier.
 989  */
 990 unsigned ipath_ib_rate_to_mult(enum ib_rate rate)
 991 {
 992         switch (rate) {
 993         case IB_RATE_2_5_GBPS: return 8;
 994         case IB_RATE_5_GBPS:   return 4;
 995         case IB_RATE_10_GBPS:  return 2;
 996         case IB_RATE_20_GBPS:  return 1;
 997         default:               return 0;
 998         }
 999 }
1000
1001 /*
1002  * Convert delay multiplier to IB rate
1003  */
1004 static enum ib_rate ipath_mult_to_ib_rate(unsigned mult)
1005 {
1006         switch (mult) {
1007         case 8:  return IB_RATE_2_5_GBPS;
1008         case 4:  return IB_RATE_5_GBPS;
1009         case 2:  return IB_RATE_10_GBPS;
1010         case 1:  return IB_RATE_20_GBPS;
1011         default: return IB_RATE_PORT_CURRENT;
1012         }
1013 }
1014
1015 static inline struct ipath_verbs_txreq *get_txreq(struct ipath_ibdev *dev)
1016 {
1017         struct ipath_verbs_txreq *tx = NULL;
1018         unsigned long flags;
1019
1020         spin_lock_irqsave(&dev->pending_lock, flags);
1021         if (!list_empty(&dev->txreq_free)) {
1022                 struct list_head *l = dev->txreq_free.next;
1023
1024                 list_del(l);
1025                 tx = list_entry(l, struct ipath_verbs_txreq, txreq.list);
1026         }
1027         spin_unlock_irqrestore(&dev->pending_lock, flags);
1028         return tx;
1029 }
1030
1031 static inline void put_txreq(struct ipath_ibdev *dev,
1032                              struct ipath_verbs_txreq *tx)
1033 {
1034         unsigned long flags;
1035
1036         spin_lock_irqsave(&dev->pending_lock, flags);
1037         list_add(&tx->txreq.list, &dev->txreq_free);
1038         spin_unlock_irqrestore(&dev->pending_lock, flags);
1039 }
1040
1041 static void sdma_complete(void *cookie, int status)
1042 {
1043         struct ipath_verbs_txreq *tx = cookie;
1044         struct ipath_qp *qp = tx->qp;
1045         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
1046         unsigned long flags;
1047         enum ib_wc_status ibs = status == IPATH_SDMA_TXREQ_S_OK ?
1048                 IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR;
1049
1050         if (atomic_dec_and_test(&qp->s_dma_busy)) {
1051                 spin_lock_irqsave(&qp->s_lock, flags);
1052                 if (tx->wqe)
1053                         ipath_send_complete(qp, tx->wqe, ibs);
1054                 if ((ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND &&
1055                      qp->s_last != qp->s_head) ||
1056                     (qp->s_flags & IPATH_S_WAIT_DMA))
1057                         ipath_schedule_send(qp);
1058                 spin_unlock_irqrestore(&qp->s_lock, flags);
1059                 wake_up(&qp->wait_dma);
1060         } else if (tx->wqe) {
1061                 spin_lock_irqsave(&qp->s_lock, flags);
1062                 ipath_send_complete(qp, tx->wqe, ibs);
1063                 spin_unlock_irqrestore(&qp->s_lock, flags);
1064         }
1065
1066         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEBUF)
1067                 kfree(tx->txreq.map_addr);
1068         put_txreq(dev, tx);
1069
1070         if (atomic_dec_and_test(&qp->refcount))
1071                 wake_up(&qp->wait);
1072 }
1073
1074 static void decrement_dma_busy(struct ipath_qp *qp)
1075 {
1076         unsigned long flags;
1077
1078         if (atomic_dec_and_test(&qp->s_dma_busy)) {
1079                 spin_lock_irqsave(&qp->s_lock, flags);
1080                 if ((ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND &&
1081                      qp->s_last != qp->s_head) ||
1082                     (qp->s_flags & IPATH_S_WAIT_DMA))
1083                         ipath_schedule_send(qp);
1084                 spin_unlock_irqrestore(&qp->s_lock, flags);
1085                 wake_up(&qp->wait_dma);
1086         }
1087 }
1088
1089 /*
1090  * Compute the number of clock cycles of delay before sending the next packet.
1091  * The multipliers reflect the number of clocks for the fastest rate so
1092  * one tick at 4xDDR is 8 ticks at 1xSDR.
1093  * If the destination port will take longer to receive a packet than
1094  * the outgoing link can send it, we need to delay sending the next packet
1095  * by the difference in time it takes the receiver to receive and the sender
1096  * to send this packet.
1097  * Note that this delay is always correct for UC and RC but not always
1098  * optimal for UD. For UD, the destination HCA can be different for each
1099  * packet, in which case, we could send packets to a different destination
1100  * while "waiting" for the delay. The overhead for doing this without
1101  * HW support is more than just paying the cost of delaying some packets
1102  * unnecessarily.
1103  */
1104 static inline unsigned ipath_pkt_delay(u32 plen, u8 snd_mult, u8 rcv_mult)
1105 {
1106         return (rcv_mult > snd_mult) ?
1107                 (plen * (rcv_mult - snd_mult) + 1) >> 1 : 0;
1108 }
1109
1110 static int ipath_verbs_send_dma(struct ipath_qp *qp,
1111                                 struct ipath_ib_header *hdr, u32 hdrwords,
1112                                 struct ipath_sge_state *ss, u32 len,
1113                                 u32 plen, u32 dwords)
1114 {
1115         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
1116         struct ipath_devdata *dd = dev->dd;
1117         struct ipath_verbs_txreq *tx;
1118         u32 *piobuf;
1119         u32 control;
1120         u32 ndesc;
1121         int ret;
1122
1123         tx = qp->s_tx;
1124         if (tx) {
1125                 qp->s_tx = NULL;
1126                 /* resend previously constructed packet */
1127                 atomic_inc(&qp->s_dma_busy);
1128                 ret = ipath_sdma_verbs_send(dd, tx->ss, tx->len, tx);
1129                 if (ret) {
1130                         qp->s_tx = tx;
1131                         decrement_dma_busy(qp);
1132                 }
1133                 goto bail;
1134         }
1135
1136         tx = get_txreq(dev);
1137         if (!tx) {
1138                 ret = -EBUSY;
1139                 goto bail;
1140         }
1141
1142         /*
1143          * Get the saved delay count we computed for the previous packet
1144          * and save the delay count for this packet to be used next time
1145          * we get here.
1146          */
1147         control = qp->s_pkt_delay;
1148         qp->s_pkt_delay = ipath_pkt_delay(plen, dd->delay_mult, qp->s_dmult);
1149
1150         tx->qp = qp;
1151         atomic_inc(&qp->refcount);
1152         tx->wqe = qp->s_wqe;
1153         tx->txreq.callback = sdma_complete;
1154         tx->txreq.callback_cookie = tx;
1155         tx->txreq.flags = IPATH_SDMA_TXREQ_F_HEADTOHOST |
1156                 IPATH_SDMA_TXREQ_F_INTREQ | IPATH_SDMA_TXREQ_F_FREEDESC;
1157         if (plen + 1 >= IPATH_SMALLBUF_DWORDS)
1158                 tx->txreq.flags |= IPATH_SDMA_TXREQ_F_USELARGEBUF;
1159
1160         /* VL15 packets bypass credit check */
1161         if ((be16_to_cpu(hdr->lrh[0]) >> 12) == 15) {
1162                 control |= 1ULL << 31;
1163                 tx->txreq.flags |= IPATH_SDMA_TXREQ_F_VL15;
1164         }
1165
1166         if (len) {
1167                 /*
1168                  * Don't try to DMA if it takes more descriptors than
1169                  * the queue holds.
1170                  */
1171                 ndesc = ipath_count_sge(ss, len);
1172                 if (ndesc >= dd->ipath_sdma_descq_cnt)
1173                         ndesc = 0;
1174         } else
1175                 ndesc = 1;
1176         if (ndesc) {
1177                 tx->hdr.pbc[0] = cpu_to_le32(plen);
1178                 tx->hdr.pbc[1] = cpu_to_le32(control);
1179                 memcpy(&tx->hdr.hdr, hdr, hdrwords << 2);
1180                 tx->txreq.sg_count = ndesc;
1181                 tx->map_len = (hdrwords + 2) << 2;
1182                 tx->txreq.map_addr = &tx->hdr;
1183                 atomic_inc(&qp->s_dma_busy);
1184                 ret = ipath_sdma_verbs_send(dd, ss, dwords, tx);
1185                 if (ret) {
1186                         /* save ss and length in dwords */
1187                         tx->ss = ss;
1188                         tx->len = dwords;
1189                         qp->s_tx = tx;
1190                         decrement_dma_busy(qp);
1191                 }
1192                 goto bail;
1193         }
1194
1195         /* Allocate a buffer and copy the header and payload to it. */
1196         tx->map_len = (plen + 1) << 2;
1197         piobuf = kmalloc(tx->map_len, GFP_ATOMIC);
1198         if (unlikely(piobuf == NULL)) {
1199                 ret = -EBUSY;
1200                 goto err_tx;
1201         }
1202         tx->txreq.map_addr = piobuf;
1203         tx->txreq.flags |= IPATH_SDMA_TXREQ_F_FREEBUF;
1204         tx->txreq.sg_count = 1;
1205
1206         *piobuf++ = (__force u32) cpu_to_le32(plen);
1207         *piobuf++ = (__force u32) cpu_to_le32(control);
1208         memcpy(piobuf, hdr, hdrwords << 2);
1209         ipath_copy_from_sge(piobuf + hdrwords, ss, len);
1210
1211         atomic_inc(&qp->s_dma_busy);
1212         ret = ipath_sdma_verbs_send(dd, NULL, 0, tx);
1213         /*
1214          * If we couldn't queue the DMA request, save the info
1215          * and try again later rather than destroying the
1216          * buffer and undoing the side effects of the copy.
1217          */
1218         if (ret) {
1219                 tx->ss = NULL;
1220                 tx->len = 0;
1221                 qp->s_tx = tx;
1222                 decrement_dma_busy(qp);
1223         }
1224         dev->n_unaligned++;
1225         goto bail;
1226
1227 err_tx:
1228         if (atomic_dec_and_test(&qp->refcount))
1229                 wake_up(&qp->wait);
1230         put_txreq(dev, tx);
1231 bail:
1232         return ret;
1233 }
1234
1235 static int ipath_verbs_send_pio(struct ipath_qp *qp,
1236                                 struct ipath_ib_header *ibhdr, u32 hdrwords,
1237                                 struct ipath_sge_state *ss, u32 len,
1238                                 u32 plen, u32 dwords)
1239 {
1240         struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
1241         u32 *hdr = (u32 *) ibhdr;
1242         u32 __iomem *piobuf;
1243         unsigned flush_wc;
1244         u32 control;
1245         int ret;
1246         unsigned long flags;
1247
1248         piobuf = ipath_getpiobuf(dd, plen, NULL);
1249         if (unlikely(piobuf == NULL)) {
1250                 ret = -EBUSY;
1251                 goto bail;
1252         }
1253
1254         /*
1255          * Get the saved delay count we computed for the previous packet
1256          * and save the delay count for this packet to be used next time
1257          * we get here.
1258          */
1259         control = qp->s_pkt_delay;
1260         qp->s_pkt_delay = ipath_pkt_delay(plen, dd->delay_mult, qp->s_dmult);
1261
1262         /* VL15 packets bypass credit check */
1263         if ((be16_to_cpu(ibhdr->lrh[0]) >> 12) == 15)
1264                 control |= 1ULL << 31;
1265
1266         /*
1267          * Write the length to the control qword plus any needed flags.
1268          * We have to flush after the PBC for correctness on some cpus
1269          * or WC buffer can be written out of order.
1270          */
1271         writeq(((u64) control << 32) | plen, piobuf);
1272         piobuf += 2;
1273
1274         flush_wc = dd->ipath_flags & IPATH_PIO_FLUSH_WC;
1275         if (len == 0) {
1276                 /*
1277                  * If there is just the header portion, must flush before
1278                  * writing last word of header for correctness, and after
1279                  * the last header word (trigger word).
1280                  */
1281                 if (flush_wc) {
1282                         ipath_flush_wc();
1283                         __iowrite32_copy(piobuf, hdr, hdrwords - 1);
1284                         ipath_flush_wc();
1285                         __raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
1286                         ipath_flush_wc();
1287                 } else
1288                         __iowrite32_copy(piobuf, hdr, hdrwords);
1289                 goto done;
1290         }
1291
1292         if (flush_wc)
1293                 ipath_flush_wc();
1294         __iowrite32_copy(piobuf, hdr, hdrwords);
1295         piobuf += hdrwords;
1296
1297         /* The common case is aligned and contained in one segment. */
1298         if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
1299                    !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
1300                 u32 *addr = (u32 *) ss->sge.vaddr;
1301
1302                 /* Update address before sending packet. */
1303                 update_sge(ss, len);
1304                 if (flush_wc) {
1305                         __iowrite32_copy(piobuf, addr, dwords - 1);
1306                         /* must flush early everything before trigger word */
1307                         ipath_flush_wc();
1308                         __raw_writel(addr[dwords - 1], piobuf + dwords - 1);
1309                         /* be sure trigger word is written */
1310                         ipath_flush_wc();
1311                 } else
1312                         __iowrite32_copy(piobuf, addr, dwords);
1313                 goto done;
1314         }
1315         copy_io(piobuf, ss, len, flush_wc);
1316 done:
1317         if (qp->s_wqe) {
1318                 spin_lock_irqsave(&qp->s_lock, flags);
1319                 ipath_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
1320                 spin_unlock_irqrestore(&qp->s_lock, flags);
1321         }
1322         ret = 0;
1323 bail:
1324         return ret;
1325 }
1326
1327 /**
1328  * ipath_verbs_send - send a packet
1329  * @qp: the QP to send on
1330  * @hdr: the packet header
1331  * @hdrwords: the number of 32-bit words in the header
1332  * @ss: the SGE to send
1333  * @len: the length of the packet in bytes
1334  */
1335 int ipath_verbs_send(struct ipath_qp *qp, struct ipath_ib_header *hdr,
1336                      u32 hdrwords, struct ipath_sge_state *ss, u32 len)
1337 {
1338         struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
1339         u32 plen;
1340         int ret;
1341         u32 dwords = (len + 3) >> 2;
1342
1343         /*
1344          * Calculate the send buffer trigger address.
1345          * The +1 counts for the pbc control dword following the pbc length.
1346          */
1347         plen = hdrwords + dwords + 1;
1348
1349         /*
1350          * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1351          * can defer SDMA restart until link goes ACTIVE without
1352          * worrying about just how we got there.
1353          */
1354         if (qp->ibqp.qp_type == IB_QPT_SMI ||
1355             !(dd->ipath_flags & IPATH_HAS_SEND_DMA))
1356                 ret = ipath_verbs_send_pio(qp, hdr, hdrwords, ss, len,
1357                                            plen, dwords);
1358         else
1359                 ret = ipath_verbs_send_dma(qp, hdr, hdrwords, ss, len,
1360                                            plen, dwords);
1361
1362         return ret;
1363 }
1364
1365 int ipath_snapshot_counters(struct ipath_devdata *dd, u64 *swords,
1366                             u64 *rwords, u64 *spkts, u64 *rpkts,
1367                             u64 *xmit_wait)
1368 {
1369         int ret;
1370
1371         if (!(dd->ipath_flags & IPATH_INITTED)) {
1372                 /* no hardware, freeze, etc. */
1373                 ret = -EINVAL;
1374                 goto bail;
1375         }
1376         *swords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt);
1377         *rwords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
1378         *spkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
1379         *rpkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
1380         *xmit_wait = ipath_snap_cntr(dd, dd->ipath_cregs->cr_sendstallcnt);
1381
1382         ret = 0;
1383
1384 bail:
1385         return ret;
1386 }
1387
1388 /**
1389  * ipath_get_counters - get various chip counters
1390  * @dd: the infinipath device
1391  * @cntrs: counters are placed here
1392  *
1393  * Return the counters needed by recv_pma_get_portcounters().
1394  */
1395 int ipath_get_counters(struct ipath_devdata *dd,
1396                        struct ipath_verbs_counters *cntrs)
1397 {
1398         struct ipath_cregs const *crp = dd->ipath_cregs;
1399         int ret;
1400
1401         if (!(dd->ipath_flags & IPATH_INITTED)) {
1402                 /* no hardware, freeze, etc. */
1403                 ret = -EINVAL;
1404                 goto bail;
1405         }
1406         cntrs->symbol_error_counter =
1407                 ipath_snap_cntr(dd, crp->cr_ibsymbolerrcnt);
1408         cntrs->link_error_recovery_counter =
1409                 ipath_snap_cntr(dd, crp->cr_iblinkerrrecovcnt);
1410         /*
1411          * The link downed counter counts when the other side downs the
1412          * connection.  We add in the number of times we downed the link
1413          * due to local link integrity errors to compensate.
1414          */
1415         cntrs->link_downed_counter =
1416                 ipath_snap_cntr(dd, crp->cr_iblinkdowncnt);
1417         cntrs->port_rcv_errors =
1418                 ipath_snap_cntr(dd, crp->cr_rxdroppktcnt) +
1419                 ipath_snap_cntr(dd, crp->cr_rcvovflcnt) +
1420                 ipath_snap_cntr(dd, crp->cr_portovflcnt) +
1421                 ipath_snap_cntr(dd, crp->cr_err_rlencnt) +
1422                 ipath_snap_cntr(dd, crp->cr_invalidrlencnt) +
1423                 ipath_snap_cntr(dd, crp->cr_errlinkcnt) +
1424                 ipath_snap_cntr(dd, crp->cr_erricrccnt) +
1425                 ipath_snap_cntr(dd, crp->cr_errvcrccnt) +
1426                 ipath_snap_cntr(dd, crp->cr_errlpcrccnt) +
1427                 ipath_snap_cntr(dd, crp->cr_badformatcnt) +
1428                 dd->ipath_rxfc_unsupvl_errs;
1429         if (crp->cr_rxotherlocalphyerrcnt)
1430                 cntrs->port_rcv_errors +=
1431                         ipath_snap_cntr(dd, crp->cr_rxotherlocalphyerrcnt);
1432         if (crp->cr_rxvlerrcnt)
1433                 cntrs->port_rcv_errors +=
1434                         ipath_snap_cntr(dd, crp->cr_rxvlerrcnt);
1435         cntrs->port_rcv_remphys_errors =
1436                 ipath_snap_cntr(dd, crp->cr_rcvebpcnt);
1437         cntrs->port_xmit_discards = ipath_snap_cntr(dd, crp->cr_unsupvlcnt);
1438         cntrs->port_xmit_data = ipath_snap_cntr(dd, crp->cr_wordsendcnt);
1439         cntrs->port_rcv_data = ipath_snap_cntr(dd, crp->cr_wordrcvcnt);
1440         cntrs->port_xmit_packets = ipath_snap_cntr(dd, crp->cr_pktsendcnt);
1441         cntrs->port_rcv_packets = ipath_snap_cntr(dd, crp->cr_pktrcvcnt);
1442         cntrs->local_link_integrity_errors =
1443                 crp->cr_locallinkintegrityerrcnt ?
1444                 ipath_snap_cntr(dd, crp->cr_locallinkintegrityerrcnt) :
1445                 ((dd->ipath_flags & IPATH_GPIO_ERRINTRS) ?
1446                  dd->ipath_lli_errs : dd->ipath_lli_errors);
1447         cntrs->excessive_buffer_overrun_errors =
1448                 crp->cr_excessbufferovflcnt ?
1449                 ipath_snap_cntr(dd, crp->cr_excessbufferovflcnt) :
1450                 dd->ipath_overrun_thresh_errs;
1451         cntrs->vl15_dropped = crp->cr_vl15droppedpktcnt ?
1452                 ipath_snap_cntr(dd, crp->cr_vl15droppedpktcnt) : 0;
1453
1454         ret = 0;
1455
1456 bail:
1457         return ret;
1458 }
1459
1460 /**
1461  * ipath_ib_piobufavail - callback when a PIO buffer is available
1462  * @arg: the device pointer
1463  *
1464  * This is called from ipath_intr() at interrupt level when a PIO buffer is
1465  * available after ipath_verbs_send() returned an error that no buffers were
1466  * available.  Return 1 if we consumed all the PIO buffers and we still have
1467  * QPs waiting for buffers (for now, just restart the send tasklet and
1468  * return zero).
1469  */
1470 int ipath_ib_piobufavail(struct ipath_ibdev *dev)
1471 {
1472         struct list_head *list;
1473         struct ipath_qp *qplist;
1474         struct ipath_qp *qp;
1475         unsigned long flags;
1476
1477         if (dev == NULL)
1478                 goto bail;
1479
1480         list = &dev->piowait;
1481         qplist = NULL;
1482
1483         spin_lock_irqsave(&dev->pending_lock, flags);
1484         while (!list_empty(list)) {
1485                 qp = list_entry(list->next, struct ipath_qp, piowait);
1486                 list_del_init(&qp->piowait);
1487                 qp->pio_next = qplist;
1488                 qplist = qp;
1489                 atomic_inc(&qp->refcount);
1490         }
1491         spin_unlock_irqrestore(&dev->pending_lock, flags);
1492
1493         while (qplist != NULL) {
1494                 qp = qplist;
1495                 qplist = qp->pio_next;
1496
1497                 spin_lock_irqsave(&qp->s_lock, flags);
1498                 if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)
1499                         ipath_schedule_send(qp);
1500                 spin_unlock_irqrestore(&qp->s_lock, flags);
1501
1502                 /* Notify ipath_destroy_qp() if it is waiting. */
1503                 if (atomic_dec_and_test(&qp->refcount))
1504                         wake_up(&qp->wait);
1505         }
1506
1507 bail:
1508         return 0;
1509 }
1510
1511 static int ipath_query_device(struct ib_device *ibdev, struct ib_device_attr *props,
1512                               struct ib_udata *uhw)
1513 {
1514         struct ipath_ibdev *dev = to_idev(ibdev);
1515
1516         if (uhw->inlen || uhw->outlen)
1517                 return -EINVAL;
1518
1519         memset(props, 0, sizeof(*props));
1520
1521         props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1522                 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1523                 IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1524                 IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE;
1525         props->page_size_cap = PAGE_SIZE;
1526         props->vendor_id =
1527                 IPATH_SRC_OUI_1 << 16 | IPATH_SRC_OUI_2 << 8 | IPATH_SRC_OUI_3;
1528         props->vendor_part_id = dev->dd->ipath_deviceid;
1529         props->hw_ver = dev->dd->ipath_pcirev;
1530
1531         props->sys_image_guid = dev->sys_image_guid;
1532
1533         props->max_mr_size = ~0ull;
1534         props->max_qp = ib_ipath_max_qps;
1535         props->max_qp_wr = ib_ipath_max_qp_wrs;
1536         props->max_sge = ib_ipath_max_sges;
1537         props->max_sge_rd = ib_ipath_max_sges;
1538         props->max_cq = ib_ipath_max_cqs;
1539         props->max_ah = ib_ipath_max_ahs;
1540         props->max_cqe = ib_ipath_max_cqes;
1541         props->max_mr = dev->lk_table.max;
1542         props->max_fmr = dev->lk_table.max;
1543         props->max_map_per_fmr = 32767;
1544         props->max_pd = ib_ipath_max_pds;
1545         props->max_qp_rd_atom = IPATH_MAX_RDMA_ATOMIC;
1546         props->max_qp_init_rd_atom = 255;
1547         /* props->max_res_rd_atom */
1548         props->max_srq = ib_ipath_max_srqs;
1549         props->max_srq_wr = ib_ipath_max_srq_wrs;
1550         props->max_srq_sge = ib_ipath_max_srq_sges;
1551         /* props->local_ca_ack_delay */
1552         props->atomic_cap = IB_ATOMIC_GLOB;
1553         props->max_pkeys = ipath_get_npkeys(dev->dd);
1554         props->max_mcast_grp = ib_ipath_max_mcast_grps;
1555         props->max_mcast_qp_attach = ib_ipath_max_mcast_qp_attached;
1556         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1557                 props->max_mcast_grp;
1558
1559         return 0;
1560 }
1561
1562 const u8 ipath_cvt_physportstate[32] = {
1563         [INFINIPATH_IBCS_LT_STATE_DISABLED] = IB_PHYSPORTSTATE_DISABLED,
1564         [INFINIPATH_IBCS_LT_STATE_LINKUP] = IB_PHYSPORTSTATE_LINKUP,
1565         [INFINIPATH_IBCS_LT_STATE_POLLACTIVE] = IB_PHYSPORTSTATE_POLL,
1566         [INFINIPATH_IBCS_LT_STATE_POLLQUIET] = IB_PHYSPORTSTATE_POLL,
1567         [INFINIPATH_IBCS_LT_STATE_SLEEPDELAY] = IB_PHYSPORTSTATE_SLEEP,
1568         [INFINIPATH_IBCS_LT_STATE_SLEEPQUIET] = IB_PHYSPORTSTATE_SLEEP,
1569         [INFINIPATH_IBCS_LT_STATE_CFGDEBOUNCE] =
1570                 IB_PHYSPORTSTATE_CFG_TRAIN,
1571         [INFINIPATH_IBCS_LT_STATE_CFGRCVFCFG] =
1572                 IB_PHYSPORTSTATE_CFG_TRAIN,
1573         [INFINIPATH_IBCS_LT_STATE_CFGWAITRMT] =
1574                 IB_PHYSPORTSTATE_CFG_TRAIN,
1575         [INFINIPATH_IBCS_LT_STATE_CFGIDLE] = IB_PHYSPORTSTATE_CFG_TRAIN,
1576         [INFINIPATH_IBCS_LT_STATE_RECOVERRETRAIN] =
1577                 IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1578         [INFINIPATH_IBCS_LT_STATE_RECOVERWAITRMT] =
1579                 IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1580         [INFINIPATH_IBCS_LT_STATE_RECOVERIDLE] =
1581                 IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1582         [0x10] = IB_PHYSPORTSTATE_CFG_TRAIN,
1583         [0x11] = IB_PHYSPORTSTATE_CFG_TRAIN,
1584         [0x12] = IB_PHYSPORTSTATE_CFG_TRAIN,
1585         [0x13] = IB_PHYSPORTSTATE_CFG_TRAIN,
1586         [0x14] = IB_PHYSPORTSTATE_CFG_TRAIN,
1587         [0x15] = IB_PHYSPORTSTATE_CFG_TRAIN,
1588         [0x16] = IB_PHYSPORTSTATE_CFG_TRAIN,
1589         [0x17] = IB_PHYSPORTSTATE_CFG_TRAIN
1590 };
1591
1592 u32 ipath_get_cr_errpkey(struct ipath_devdata *dd)
1593 {
1594         return ipath_read_creg32(dd, dd->ipath_cregs->cr_errpkey);
1595 }
1596
1597 static int ipath_query_port(struct ib_device *ibdev,
1598                             u8 port, struct ib_port_attr *props)
1599 {
1600         struct ipath_ibdev *dev = to_idev(ibdev);
1601         struct ipath_devdata *dd = dev->dd;
1602         enum ib_mtu mtu;
1603         u16 lid = dd->ipath_lid;
1604         u64 ibcstat;
1605
1606         memset(props, 0, sizeof(*props));
1607         props->lid = lid ? lid : be16_to_cpu(IB_LID_PERMISSIVE);
1608         props->lmc = dd->ipath_lmc;
1609         props->sm_lid = dev->sm_lid;
1610         props->sm_sl = dev->sm_sl;
1611         ibcstat = dd->ipath_lastibcstat;
1612         /* map LinkState to IB portinfo values.  */
1613         props->state = ipath_ib_linkstate(dd, ibcstat) + 1;
1614
1615         /* See phys_state_show() */
1616         props->phys_state = /* MEA: assumes shift == 0 */
1617                 ipath_cvt_physportstate[dd->ipath_lastibcstat &
1618                 dd->ibcs_lts_mask];
1619         props->port_cap_flags = dev->port_cap_flags;
1620         props->gid_tbl_len = 1;
1621         props->max_msg_sz = 0x80000000;
1622         props->pkey_tbl_len = ipath_get_npkeys(dd);
1623         props->bad_pkey_cntr = ipath_get_cr_errpkey(dd) -
1624                 dev->z_pkey_violations;
1625         props->qkey_viol_cntr = dev->qkey_violations;
1626         props->active_width = dd->ipath_link_width_active;
1627         /* See rate_show() */
1628         props->active_speed = dd->ipath_link_speed_active;
1629         props->max_vl_num = 1;          /* VLCap = VL0 */
1630         props->init_type_reply = 0;
1631
1632         props->max_mtu = ipath_mtu4096 ? IB_MTU_4096 : IB_MTU_2048;
1633         switch (dd->ipath_ibmtu) {
1634         case 4096:
1635                 mtu = IB_MTU_4096;
1636                 break;
1637         case 2048:
1638                 mtu = IB_MTU_2048;
1639                 break;
1640         case 1024:
1641                 mtu = IB_MTU_1024;
1642                 break;
1643         case 512:
1644                 mtu = IB_MTU_512;
1645                 break;
1646         case 256:
1647                 mtu = IB_MTU_256;
1648                 break;
1649         default:
1650                 mtu = IB_MTU_2048;
1651         }
1652         props->active_mtu = mtu;
1653         props->subnet_timeout = dev->subnet_timeout;
1654
1655         return 0;
1656 }
1657
1658 static int ipath_modify_device(struct ib_device *device,
1659                                int device_modify_mask,
1660                                struct ib_device_modify *device_modify)
1661 {
1662         int ret;
1663
1664         if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1665                                    IB_DEVICE_MODIFY_NODE_DESC)) {
1666                 ret = -EOPNOTSUPP;
1667                 goto bail;
1668         }
1669
1670         if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC)
1671                 memcpy(device->node_desc, device_modify->node_desc, 64);
1672
1673         if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
1674                 to_idev(device)->sys_image_guid =
1675                         cpu_to_be64(device_modify->sys_image_guid);
1676
1677         ret = 0;
1678
1679 bail:
1680         return ret;
1681 }
1682
1683 static int ipath_modify_port(struct ib_device *ibdev,
1684                              u8 port, int port_modify_mask,
1685                              struct ib_port_modify *props)
1686 {
1687         struct ipath_ibdev *dev = to_idev(ibdev);
1688
1689         dev->port_cap_flags |= props->set_port_cap_mask;
1690         dev->port_cap_flags &= ~props->clr_port_cap_mask;
1691         if (port_modify_mask & IB_PORT_SHUTDOWN)
1692                 ipath_set_linkstate(dev->dd, IPATH_IB_LINKDOWN);
1693         if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR)
1694                 dev->qkey_violations = 0;
1695         return 0;
1696 }
1697
1698 static int ipath_query_gid(struct ib_device *ibdev, u8 port,
1699                            int index, union ib_gid *gid)
1700 {
1701         struct ipath_ibdev *dev = to_idev(ibdev);
1702         int ret;
1703
1704         if (index >= 1) {
1705                 ret = -EINVAL;
1706                 goto bail;
1707         }
1708         gid->global.subnet_prefix = dev->gid_prefix;
1709         gid->global.interface_id = dev->dd->ipath_guid;
1710
1711         ret = 0;
1712
1713 bail:
1714         return ret;
1715 }
1716
1717 static struct ib_pd *ipath_alloc_pd(struct ib_device *ibdev,
1718                                     struct ib_ucontext *context,
1719                                     struct ib_udata *udata)
1720 {
1721         struct ipath_ibdev *dev = to_idev(ibdev);
1722         struct ipath_pd *pd;
1723         struct ib_pd *ret;
1724
1725         /*
1726          * This is actually totally arbitrary.  Some correctness tests
1727          * assume there's a maximum number of PDs that can be allocated.
1728          * We don't actually have this limit, but we fail the test if
1729          * we allow allocations of more than we report for this value.
1730          */
1731
1732         pd = kmalloc(sizeof *pd, GFP_KERNEL);
1733         if (!pd) {
1734                 ret = ERR_PTR(-ENOMEM);
1735                 goto bail;
1736         }
1737
1738         spin_lock(&dev->n_pds_lock);
1739         if (dev->n_pds_allocated == ib_ipath_max_pds) {
1740                 spin_unlock(&dev->n_pds_lock);
1741                 kfree(pd);
1742                 ret = ERR_PTR(-ENOMEM);
1743                 goto bail;
1744         }
1745
1746         dev->n_pds_allocated++;
1747         spin_unlock(&dev->n_pds_lock);
1748
1749         /* ib_alloc_pd() will initialize pd->ibpd. */
1750         pd->user = udata != NULL;
1751
1752         ret = &pd->ibpd;
1753
1754 bail:
1755         return ret;
1756 }
1757
1758 static int ipath_dealloc_pd(struct ib_pd *ibpd)
1759 {
1760         struct ipath_pd *pd = to_ipd(ibpd);
1761         struct ipath_ibdev *dev = to_idev(ibpd->device);
1762
1763         spin_lock(&dev->n_pds_lock);
1764         dev->n_pds_allocated--;
1765         spin_unlock(&dev->n_pds_lock);
1766
1767         kfree(pd);
1768
1769         return 0;
1770 }
1771
1772 /**
1773  * ipath_create_ah - create an address handle
1774  * @pd: the protection domain
1775  * @ah_attr: the attributes of the AH
1776  *
1777  * This may be called from interrupt context.
1778  */
1779 static struct ib_ah *ipath_create_ah(struct ib_pd *pd,
1780                                      struct ib_ah_attr *ah_attr)
1781 {
1782         struct ipath_ah *ah;
1783         struct ib_ah *ret;
1784         struct ipath_ibdev *dev = to_idev(pd->device);
1785         unsigned long flags;
1786
1787         /* A multicast address requires a GRH (see ch. 8.4.1). */
1788         if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
1789             ah_attr->dlid != IPATH_PERMISSIVE_LID &&
1790             !(ah_attr->ah_flags & IB_AH_GRH)) {
1791                 ret = ERR_PTR(-EINVAL);
1792                 goto bail;
1793         }
1794
1795         if (ah_attr->dlid == 0) {
1796                 ret = ERR_PTR(-EINVAL);
1797                 goto bail;
1798         }
1799
1800         if (ah_attr->port_num < 1 ||
1801             ah_attr->port_num > pd->device->phys_port_cnt) {
1802                 ret = ERR_PTR(-EINVAL);
1803                 goto bail;
1804         }
1805
1806         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
1807         if (!ah) {
1808                 ret = ERR_PTR(-ENOMEM);
1809                 goto bail;
1810         }
1811
1812         spin_lock_irqsave(&dev->n_ahs_lock, flags);
1813         if (dev->n_ahs_allocated == ib_ipath_max_ahs) {
1814                 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1815                 kfree(ah);
1816                 ret = ERR_PTR(-ENOMEM);
1817                 goto bail;
1818         }
1819
1820         dev->n_ahs_allocated++;
1821         spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1822
1823         /* ib_create_ah() will initialize ah->ibah. */
1824         ah->attr = *ah_attr;
1825         ah->attr.static_rate = ipath_ib_rate_to_mult(ah_attr->static_rate);
1826
1827         ret = &ah->ibah;
1828
1829 bail:
1830         return ret;
1831 }
1832
1833 /**
1834  * ipath_destroy_ah - destroy an address handle
1835  * @ibah: the AH to destroy
1836  *
1837  * This may be called from interrupt context.
1838  */
1839 static int ipath_destroy_ah(struct ib_ah *ibah)
1840 {
1841         struct ipath_ibdev *dev = to_idev(ibah->device);
1842         struct ipath_ah *ah = to_iah(ibah);
1843         unsigned long flags;
1844
1845         spin_lock_irqsave(&dev->n_ahs_lock, flags);
1846         dev->n_ahs_allocated--;
1847         spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1848
1849         kfree(ah);
1850
1851         return 0;
1852 }
1853
1854 static int ipath_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1855 {
1856         struct ipath_ah *ah = to_iah(ibah);
1857
1858         *ah_attr = ah->attr;
1859         ah_attr->static_rate = ipath_mult_to_ib_rate(ah->attr.static_rate);
1860
1861         return 0;
1862 }
1863
1864 /**
1865  * ipath_get_npkeys - return the size of the PKEY table for port 0
1866  * @dd: the infinipath device
1867  */
1868 unsigned ipath_get_npkeys(struct ipath_devdata *dd)
1869 {
1870         return ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys);
1871 }
1872
1873 /**
1874  * ipath_get_pkey - return the indexed PKEY from the port PKEY table
1875  * @dd: the infinipath device
1876  * @index: the PKEY index
1877  */
1878 unsigned ipath_get_pkey(struct ipath_devdata *dd, unsigned index)
1879 {
1880         unsigned ret;
1881
1882         /* always a kernel port, no locking needed */
1883         if (index >= ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys))
1884                 ret = 0;
1885         else
1886                 ret = dd->ipath_pd[0]->port_pkeys[index];
1887
1888         return ret;
1889 }
1890
1891 static int ipath_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1892                             u16 *pkey)
1893 {
1894         struct ipath_ibdev *dev = to_idev(ibdev);
1895         int ret;
1896
1897         if (index >= ipath_get_npkeys(dev->dd)) {
1898                 ret = -EINVAL;
1899                 goto bail;
1900         }
1901
1902         *pkey = ipath_get_pkey(dev->dd, index);
1903         ret = 0;
1904
1905 bail:
1906         return ret;
1907 }
1908
1909 /**
1910  * ipath_alloc_ucontext - allocate a ucontest
1911  * @ibdev: the infiniband device
1912  * @udata: not used by the InfiniPath driver
1913  */
1914
1915 static struct ib_ucontext *ipath_alloc_ucontext(struct ib_device *ibdev,
1916                                                 struct ib_udata *udata)
1917 {
1918         struct ipath_ucontext *context;
1919         struct ib_ucontext *ret;
1920
1921         context = kmalloc(sizeof *context, GFP_KERNEL);
1922         if (!context) {
1923                 ret = ERR_PTR(-ENOMEM);
1924                 goto bail;
1925         }
1926
1927         ret = &context->ibucontext;
1928
1929 bail:
1930         return ret;
1931 }
1932
1933 static int ipath_dealloc_ucontext(struct ib_ucontext *context)
1934 {
1935         kfree(to_iucontext(context));
1936         return 0;
1937 }
1938
1939 static int ipath_verbs_register_sysfs(struct ib_device *dev);
1940
1941 static void __verbs_timer(unsigned long arg)
1942 {
1943         struct ipath_devdata *dd = (struct ipath_devdata *) arg;
1944
1945         /* Handle verbs layer timeouts. */
1946         ipath_ib_timer(dd->verbs_dev);
1947
1948         mod_timer(&dd->verbs_timer, jiffies + 1);
1949 }
1950
1951 static int enable_timer(struct ipath_devdata *dd)
1952 {
1953         /*
1954          * Early chips had a design flaw where the chip and kernel idea
1955          * of the tail register don't always agree, and therefore we won't
1956          * get an interrupt on the next packet received.
1957          * If the board supports per packet receive interrupts, use it.
1958          * Otherwise, the timer function periodically checks for packets
1959          * to cover this case.
1960          * Either way, the timer is needed for verbs layer related
1961          * processing.
1962          */
1963         if (dd->ipath_flags & IPATH_GPIO_INTR) {
1964                 ipath_write_kreg(dd, dd->ipath_kregs->kr_debugportselect,
1965                                  0x2074076542310ULL);
1966                 /* Enable GPIO bit 2 interrupt */
1967                 dd->ipath_gpio_mask |= (u64) (1 << IPATH_GPIO_PORT0_BIT);
1968                 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1969                                  dd->ipath_gpio_mask);
1970         }
1971
1972         setup_timer(&dd->verbs_timer, __verbs_timer, (unsigned long)dd);
1973
1974         dd->verbs_timer.expires = jiffies + 1;
1975         add_timer(&dd->verbs_timer);
1976
1977         return 0;
1978 }
1979
1980 static int disable_timer(struct ipath_devdata *dd)
1981 {
1982         /* Disable GPIO bit 2 interrupt */
1983         if (dd->ipath_flags & IPATH_GPIO_INTR) {
1984                 /* Disable GPIO bit 2 interrupt */
1985                 dd->ipath_gpio_mask &= ~((u64) (1 << IPATH_GPIO_PORT0_BIT));
1986                 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1987                                  dd->ipath_gpio_mask);
1988                 /*
1989                  * We might want to undo changes to debugportselect,
1990                  * but how?
1991                  */
1992         }
1993
1994         del_timer_sync(&dd->verbs_timer);
1995
1996         return 0;
1997 }
1998
1999 static int ipath_port_immutable(struct ib_device *ibdev, u8 port_num,
2000                                 struct ib_port_immutable *immutable)
2001 {
2002         struct ib_port_attr attr;
2003         int err;
2004
2005         err = ipath_query_port(ibdev, port_num, &attr);
2006         if (err)
2007                 return err;
2008
2009         immutable->pkey_tbl_len = attr.pkey_tbl_len;
2010         immutable->gid_tbl_len = attr.gid_tbl_len;
2011         immutable->core_cap_flags = RDMA_CORE_PORT_IBA_IB;
2012         immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2013
2014         return 0;
2015 }
2016
2017 /**
2018  * ipath_register_ib_device - register our device with the infiniband core
2019  * @dd: the device data structure
2020  * Return the allocated ipath_ibdev pointer or NULL on error.
2021  */
2022 int ipath_register_ib_device(struct ipath_devdata *dd)
2023 {
2024         struct ipath_verbs_counters cntrs;
2025         struct ipath_ibdev *idev;
2026         struct ib_device *dev;
2027         struct ipath_verbs_txreq *tx;
2028         unsigned i;
2029         int ret;
2030
2031         idev = (struct ipath_ibdev *)ib_alloc_device(sizeof *idev);
2032         if (idev == NULL) {
2033                 ret = -ENOMEM;
2034                 goto bail;
2035         }
2036
2037         dev = &idev->ibdev;
2038
2039         if (dd->ipath_sdma_descq_cnt) {
2040                 tx = kmalloc_array(dd->ipath_sdma_descq_cnt, sizeof *tx,
2041                                    GFP_KERNEL);
2042                 if (tx == NULL) {
2043                         ret = -ENOMEM;
2044                         goto err_tx;
2045                 }
2046         } else
2047                 tx = NULL;
2048         idev->txreq_bufs = tx;
2049
2050         /* Only need to initialize non-zero fields. */
2051         spin_lock_init(&idev->n_pds_lock);
2052         spin_lock_init(&idev->n_ahs_lock);
2053         spin_lock_init(&idev->n_cqs_lock);
2054         spin_lock_init(&idev->n_qps_lock);
2055         spin_lock_init(&idev->n_srqs_lock);
2056         spin_lock_init(&idev->n_mcast_grps_lock);
2057
2058         spin_lock_init(&idev->qp_table.lock);
2059         spin_lock_init(&idev->lk_table.lock);
2060         idev->sm_lid = be16_to_cpu(IB_LID_PERMISSIVE);
2061         /* Set the prefix to the default value (see ch. 4.1.1) */
2062         idev->gid_prefix = cpu_to_be64(0xfe80000000000000ULL);
2063
2064         ret = ipath_init_qp_table(idev, ib_ipath_qp_table_size);
2065         if (ret)
2066                 goto err_qp;
2067
2068         /*
2069          * The top ib_ipath_lkey_table_size bits are used to index the
2070          * table.  The lower 8 bits can be owned by the user (copied from
2071          * the LKEY).  The remaining bits act as a generation number or tag.
2072          */
2073         idev->lk_table.max = 1 << ib_ipath_lkey_table_size;
2074         idev->lk_table.table = kcalloc(idev->lk_table.max,
2075                                        sizeof(*idev->lk_table.table),
2076                                        GFP_KERNEL);
2077         if (idev->lk_table.table == NULL) {
2078                 ret = -ENOMEM;
2079                 goto err_lk;
2080         }
2081         INIT_LIST_HEAD(&idev->pending_mmaps);
2082         spin_lock_init(&idev->pending_lock);
2083         idev->mmap_offset = PAGE_SIZE;
2084         spin_lock_init(&idev->mmap_offset_lock);
2085         INIT_LIST_HEAD(&idev->pending[0]);
2086         INIT_LIST_HEAD(&idev->pending[1]);
2087         INIT_LIST_HEAD(&idev->pending[2]);
2088         INIT_LIST_HEAD(&idev->piowait);
2089         INIT_LIST_HEAD(&idev->rnrwait);
2090         INIT_LIST_HEAD(&idev->txreq_free);
2091         idev->pending_index = 0;
2092         idev->port_cap_flags =
2093                 IB_PORT_SYS_IMAGE_GUID_SUP | IB_PORT_CLIENT_REG_SUP;
2094         if (dd->ipath_flags & IPATH_HAS_LINK_LATENCY)
2095                 idev->port_cap_flags |= IB_PORT_LINK_LATENCY_SUP;
2096         idev->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
2097         idev->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
2098         idev->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
2099         idev->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
2100         idev->pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
2101
2102         /* Snapshot current HW counters to "clear" them. */
2103         ipath_get_counters(dd, &cntrs);
2104         idev->z_symbol_error_counter = cntrs.symbol_error_counter;
2105         idev->z_link_error_recovery_counter =
2106                 cntrs.link_error_recovery_counter;
2107         idev->z_link_downed_counter = cntrs.link_downed_counter;
2108         idev->z_port_rcv_errors = cntrs.port_rcv_errors;
2109         idev->z_port_rcv_remphys_errors =
2110                 cntrs.port_rcv_remphys_errors;
2111         idev->z_port_xmit_discards = cntrs.port_xmit_discards;
2112         idev->z_port_xmit_data = cntrs.port_xmit_data;
2113         idev->z_port_rcv_data = cntrs.port_rcv_data;
2114         idev->z_port_xmit_packets = cntrs.port_xmit_packets;
2115         idev->z_port_rcv_packets = cntrs.port_rcv_packets;
2116         idev->z_local_link_integrity_errors =
2117                 cntrs.local_link_integrity_errors;
2118         idev->z_excessive_buffer_overrun_errors =
2119                 cntrs.excessive_buffer_overrun_errors;
2120         idev->z_vl15_dropped = cntrs.vl15_dropped;
2121
2122         for (i = 0; i < dd->ipath_sdma_descq_cnt; i++, tx++)
2123                 list_add(&tx->txreq.list, &idev->txreq_free);
2124
2125         /*
2126          * The system image GUID is supposed to be the same for all
2127          * IB HCAs in a single system but since there can be other
2128          * device types in the system, we can't be sure this is unique.
2129          */
2130         if (!sys_image_guid)
2131                 sys_image_guid = dd->ipath_guid;
2132         idev->sys_image_guid = sys_image_guid;
2133         idev->ib_unit = dd->ipath_unit;
2134         idev->dd = dd;
2135
2136         strlcpy(dev->name, "ipath%d", IB_DEVICE_NAME_MAX);
2137         dev->owner = THIS_MODULE;
2138         dev->node_guid = dd->ipath_guid;
2139         dev->uverbs_abi_ver = IPATH_UVERBS_ABI_VERSION;
2140         dev->uverbs_cmd_mask =
2141                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
2142                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
2143                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
2144                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
2145                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
2146                 (1ull << IB_USER_VERBS_CMD_CREATE_AH)           |
2147                 (1ull << IB_USER_VERBS_CMD_DESTROY_AH)          |
2148                 (1ull << IB_USER_VERBS_CMD_QUERY_AH)            |
2149                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
2150                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
2151                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
2152                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
2153                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
2154                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
2155                 (1ull << IB_USER_VERBS_CMD_POLL_CQ)             |
2156                 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)       |
2157                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
2158                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
2159                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
2160                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
2161                 (1ull << IB_USER_VERBS_CMD_POST_SEND)           |
2162                 (1ull << IB_USER_VERBS_CMD_POST_RECV)           |
2163                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
2164                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
2165                 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
2166                 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
2167                 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
2168                 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
2169                 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
2170         dev->node_type = RDMA_NODE_IB_CA;
2171         dev->phys_port_cnt = 1;
2172         dev->num_comp_vectors = 1;
2173         dev->dma_device = &dd->pcidev->dev;
2174         dev->query_device = ipath_query_device;
2175         dev->modify_device = ipath_modify_device;
2176         dev->query_port = ipath_query_port;
2177         dev->modify_port = ipath_modify_port;
2178         dev->query_pkey = ipath_query_pkey;
2179         dev->query_gid = ipath_query_gid;
2180         dev->alloc_ucontext = ipath_alloc_ucontext;
2181         dev->dealloc_ucontext = ipath_dealloc_ucontext;
2182         dev->alloc_pd = ipath_alloc_pd;
2183         dev->dealloc_pd = ipath_dealloc_pd;
2184         dev->create_ah = ipath_create_ah;
2185         dev->destroy_ah = ipath_destroy_ah;
2186         dev->query_ah = ipath_query_ah;
2187         dev->create_srq = ipath_create_srq;
2188         dev->modify_srq = ipath_modify_srq;
2189         dev->query_srq = ipath_query_srq;
2190         dev->destroy_srq = ipath_destroy_srq;
2191         dev->create_qp = ipath_create_qp;
2192         dev->modify_qp = ipath_modify_qp;
2193         dev->query_qp = ipath_query_qp;
2194         dev->destroy_qp = ipath_destroy_qp;
2195         dev->post_send = ipath_post_send;
2196         dev->post_recv = ipath_post_receive;
2197         dev->post_srq_recv = ipath_post_srq_receive;
2198         dev->create_cq = ipath_create_cq;
2199         dev->destroy_cq = ipath_destroy_cq;
2200         dev->resize_cq = ipath_resize_cq;
2201         dev->poll_cq = ipath_poll_cq;
2202         dev->req_notify_cq = ipath_req_notify_cq;
2203         dev->get_dma_mr = ipath_get_dma_mr;
2204         dev->reg_phys_mr = ipath_reg_phys_mr;
2205         dev->reg_user_mr = ipath_reg_user_mr;
2206         dev->dereg_mr = ipath_dereg_mr;
2207         dev->alloc_fmr = ipath_alloc_fmr;
2208         dev->map_phys_fmr = ipath_map_phys_fmr;
2209         dev->unmap_fmr = ipath_unmap_fmr;
2210         dev->dealloc_fmr = ipath_dealloc_fmr;
2211         dev->attach_mcast = ipath_multicast_attach;
2212         dev->detach_mcast = ipath_multicast_detach;
2213         dev->process_mad = ipath_process_mad;
2214         dev->mmap = ipath_mmap;
2215         dev->dma_ops = &ipath_dma_mapping_ops;
2216         dev->get_port_immutable = ipath_port_immutable;
2217
2218         snprintf(dev->node_desc, sizeof(dev->node_desc),
2219                  IPATH_IDSTR " %s", init_utsname()->nodename);
2220
2221         ret = ib_register_device(dev, NULL);
2222         if (ret)
2223                 goto err_reg;
2224
2225         ret = ipath_verbs_register_sysfs(dev);
2226         if (ret)
2227                 goto err_class;
2228
2229         enable_timer(dd);
2230
2231         goto bail;
2232
2233 err_class:
2234         ib_unregister_device(dev);
2235 err_reg:
2236         kfree(idev->lk_table.table);
2237 err_lk:
2238         kfree(idev->qp_table.table);
2239 err_qp:
2240         kfree(idev->txreq_bufs);
2241 err_tx:
2242         ib_dealloc_device(dev);
2243         ipath_dev_err(dd, "cannot register verbs: %d!\n", -ret);
2244         idev = NULL;
2245
2246 bail:
2247         dd->verbs_dev = idev;
2248         return ret;
2249 }
2250
2251 void ipath_unregister_ib_device(struct ipath_ibdev *dev)
2252 {
2253         struct ib_device *ibdev = &dev->ibdev;
2254         u32 qps_inuse;
2255
2256         ib_unregister_device(ibdev);
2257
2258         disable_timer(dev->dd);
2259
2260         if (!list_empty(&dev->pending[0]) ||
2261             !list_empty(&dev->pending[1]) ||
2262             !list_empty(&dev->pending[2]))
2263                 ipath_dev_err(dev->dd, "pending list not empty!\n");
2264         if (!list_empty(&dev->piowait))
2265                 ipath_dev_err(dev->dd, "piowait list not empty!\n");
2266         if (!list_empty(&dev->rnrwait))
2267                 ipath_dev_err(dev->dd, "rnrwait list not empty!\n");
2268         if (!ipath_mcast_tree_empty())
2269                 ipath_dev_err(dev->dd, "multicast table memory leak!\n");
2270         /*
2271          * Note that ipath_unregister_ib_device() can be called before all
2272          * the QPs are destroyed!
2273          */
2274         qps_inuse = ipath_free_all_qps(&dev->qp_table);
2275         if (qps_inuse)
2276                 ipath_dev_err(dev->dd, "QP memory leak! %u still in use\n",
2277                         qps_inuse);
2278         kfree(dev->qp_table.table);
2279         kfree(dev->lk_table.table);
2280         kfree(dev->txreq_bufs);
2281         ib_dealloc_device(ibdev);
2282 }
2283
2284 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2285                         char *buf)
2286 {
2287         struct ipath_ibdev *dev =
2288                 container_of(device, struct ipath_ibdev, ibdev.dev);
2289
2290         return sprintf(buf, "%x\n", dev->dd->ipath_pcirev);
2291 }
2292
2293 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2294                         char *buf)
2295 {
2296         struct ipath_ibdev *dev =
2297                 container_of(device, struct ipath_ibdev, ibdev.dev);
2298         int ret;
2299
2300         ret = dev->dd->ipath_f_get_boardname(dev->dd, buf, 128);
2301         if (ret < 0)
2302                 goto bail;
2303         strcat(buf, "\n");
2304         ret = strlen(buf);
2305
2306 bail:
2307         return ret;
2308 }
2309
2310 static ssize_t show_stats(struct device *device, struct device_attribute *attr,
2311                           char *buf)
2312 {
2313         struct ipath_ibdev *dev =
2314                 container_of(device, struct ipath_ibdev, ibdev.dev);
2315         int i;
2316         int len;
2317
2318         len = sprintf(buf,
2319                       "RC resends  %d\n"
2320                       "RC no QACK  %d\n"
2321                       "RC ACKs     %d\n"
2322                       "RC SEQ NAKs %d\n"
2323                       "RC RDMA seq %d\n"
2324                       "RC RNR NAKs %d\n"
2325                       "RC OTH NAKs %d\n"
2326                       "RC timeouts %d\n"
2327                       "RC RDMA dup %d\n"
2328                       "piobuf wait %d\n"
2329                       "unaligned   %d\n"
2330                       "PKT drops   %d\n"
2331                       "WQE errs    %d\n",
2332                       dev->n_rc_resends, dev->n_rc_qacks, dev->n_rc_acks,
2333                       dev->n_seq_naks, dev->n_rdma_seq, dev->n_rnr_naks,
2334                       dev->n_other_naks, dev->n_timeouts,
2335                       dev->n_rdma_dup_busy, dev->n_piowait, dev->n_unaligned,
2336                       dev->n_pkt_drops, dev->n_wqe_errs);
2337         for (i = 0; i < ARRAY_SIZE(dev->opstats); i++) {
2338                 const struct ipath_opcode_stats *si = &dev->opstats[i];
2339
2340                 if (!si->n_packets && !si->n_bytes)
2341                         continue;
2342                 len += sprintf(buf + len, "%02x %llu/%llu\n", i,
2343                                (unsigned long long) si->n_packets,
2344                                (unsigned long long) si->n_bytes);
2345         }
2346         return len;
2347 }
2348
2349 static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
2350 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
2351 static DEVICE_ATTR(board_id, S_IRUGO, show_hca, NULL);
2352 static DEVICE_ATTR(stats, S_IRUGO, show_stats, NULL);
2353
2354 static struct device_attribute *ipath_class_attributes[] = {
2355         &dev_attr_hw_rev,
2356         &dev_attr_hca_type,
2357         &dev_attr_board_id,
2358         &dev_attr_stats
2359 };
2360
2361 static int ipath_verbs_register_sysfs(struct ib_device *dev)
2362 {
2363         int i;
2364         int ret;
2365
2366         for (i = 0; i < ARRAY_SIZE(ipath_class_attributes); ++i) {
2367                 ret = device_create_file(&dev->dev,
2368                                        ipath_class_attributes[i]);
2369                 if (ret)
2370                         goto bail;
2371         }
2372         return 0;
2373 bail:
2374         for (i = 0; i < ARRAY_SIZE(ipath_class_attributes); ++i)
2375                 device_remove_file(&dev->dev, ipath_class_attributes[i]);
2376         return ret;
2377 }