SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_verbs.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. All rights reserved.
   3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <rdma/ib_mad.h>
  35 #include <rdma/ib_user_verbs.h>
  36 #include <linux/io.h>
  37 #include <linux/utsname.h>
  38
  39 #include "ipath_kernel.h"
  40 #include "ipath_verbs.h"
  41 #include "ipath_common.h"
  42
  43 static unsigned int ib_ipath_qp_table_size = 251;
  44 module_param_named(qp_table_size, ib_ipath_qp_table_size, uint, S_IRUGO);
  45 MODULE_PARM_DESC(qp_table_size, "QP table size");
  46
  47 unsigned int ib_ipath_lkey_table_size = 12;
  48 module_param_named(lkey_table_size, ib_ipath_lkey_table_size, uint,
  49                    S_IRUGO);
  50 MODULE_PARM_DESC(lkey_table_size,
  51                  "LKEY table size in bits (2^n, 1 <= n <= 23)");
  52
  53 static unsigned int ib_ipath_max_pds = 0xFFFF;
  54 module_param_named(max_pds, ib_ipath_max_pds, uint, S_IWUSR | S_IRUGO);
  55 MODULE_PARM_DESC(max_pds,
  56                  "Maximum number of protection domains to support");
  57
  58 static unsigned int ib_ipath_max_ahs = 0xFFFF;
  59 module_param_named(max_ahs, ib_ipath_max_ahs, uint, S_IWUSR | S_IRUGO);
  60 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
  61
  62 unsigned int ib_ipath_max_cqes = 0x2FFFF;
  63 module_param_named(max_cqes, ib_ipath_max_cqes, uint, S_IWUSR | S_IRUGO);
  64 MODULE_PARM_DESC(max_cqes,
  65                  "Maximum number of completion queue entries to support");
  66
  67 unsigned int ib_ipath_max_cqs = 0x1FFFF;
  68 module_param_named(max_cqs, ib_ipath_max_cqs, uint, S_IWUSR | S_IRUGO);
  69 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
  70
  71 unsigned int ib_ipath_max_qp_wrs = 0x3FFF;
  72 module_param_named(max_qp_wrs, ib_ipath_max_qp_wrs, uint,
  73                    S_IWUSR | S_IRUGO);
  74 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
  75
  76 unsigned int ib_ipath_max_qps = 16384;
  77 module_param_named(max_qps, ib_ipath_max_qps, uint, S_IWUSR | S_IRUGO);
  78 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
  79
  80 unsigned int ib_ipath_max_sges = 0x60;
  81 module_param_named(max_sges, ib_ipath_max_sges, uint, S_IWUSR | S_IRUGO);
  82 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
  83
  84 unsigned int ib_ipath_max_mcast_grps = 16384;
  85 module_param_named(max_mcast_grps, ib_ipath_max_mcast_grps, uint,
  86                    S_IWUSR | S_IRUGO);
  87 MODULE_PARM_DESC(max_mcast_grps,
  88                  "Maximum number of multicast groups to support");
  89
  90 unsigned int ib_ipath_max_mcast_qp_attached = 16;
  91 module_param_named(max_mcast_qp_attached, ib_ipath_max_mcast_qp_attached,
  92                    uint, S_IWUSR | S_IRUGO);
  93 MODULE_PARM_DESC(max_mcast_qp_attached,
  94                  "Maximum number of attached QPs to support");
  95
  96 unsigned int ib_ipath_max_srqs = 1024;
  97 module_param_named(max_srqs, ib_ipath_max_srqs, uint, S_IWUSR | S_IRUGO);
  98 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
  99
 100 unsigned int ib_ipath_max_srq_sges = 128;
 101 module_param_named(max_srq_sges, ib_ipath_max_srq_sges,
 102                    uint, S_IWUSR | S_IRUGO);
 103 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
 104
 105 unsigned int ib_ipath_max_srq_wrs = 0x1FFFF;
 106 module_param_named(max_srq_wrs, ib_ipath_max_srq_wrs,
 107                    uint, S_IWUSR | S_IRUGO);
 108 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
 109
 110 const int ib_ipath_state_ops[IB_QPS_ERR + 1] = {
 111         [IB_QPS_RESET] = 0,
 112         [IB_QPS_INIT] = IPATH_POST_RECV_OK,
 113         [IB_QPS_RTR] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK,
 114         [IB_QPS_RTS] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
 115             IPATH_POST_SEND_OK | IPATH_PROCESS_SEND_OK,
 116         [IB_QPS_SQD] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
 117             IPATH_POST_SEND_OK,
 118         [IB_QPS_SQE] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK,
 119         [IB_QPS_ERR] = 0,
 120 };
 121
 122 struct ipath_ucontext {
 123         struct ib_ucontext ibucontext;
 124 };
 125
 126 static inline struct ipath_ucontext *to_iucontext(struct ib_ucontext
 127                                                   *ibucontext)
 128 {
 129         return container_of(ibucontext, struct ipath_ucontext, ibucontext);
 130 }
 131
 132 /*
 133  * Translate ib_wr_opcode into ib_wc_opcode.
 134  */
 135 const enum ib_wc_opcode ib_ipath_wc_opcode[] = {
 136         [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
 137         [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
 138         [IB_WR_SEND] = IB_WC_SEND,
 139         [IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
 140         [IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
 141         [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
 142         [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD
 143 };
 144
 145 /*
 146  * System image GUID.
 147  */
 148 static __be64 sys_image_guid;
 149
 150 /**
 151  * ipath_copy_sge - copy data to SGE memory
 152  * @ss: the SGE state
 153  * @data: the data to copy
 154  * @length: the length of the data
 155  */
 156 void ipath_copy_sge(struct ipath_sge_state *ss, void *data, u32 length)
 157 {
 158         struct ipath_sge *sge = &ss->sge;
 159
 160         while (length) {
 161                 u32 len = sge->length;
 162
 163                 BUG_ON(len == 0);
 164                 if (len > length)
 165                         len = length;
 166                 memcpy(sge->vaddr, data, len);
 167                 sge->vaddr += len;
 168                 sge->length -= len;
 169                 sge->sge_length -= len;
 170                 if (sge->sge_length == 0) {
 171                         if (--ss->num_sge)
 172                                 *sge = *ss->sg_list++;
 173                 } else if (sge->length == 0 && sge->mr != NULL) {
 174                         if (++sge->n >= IPATH_SEGSZ) {
 175                                 if (++sge->m >= sge->mr->mapsz)
 176                                         break;
 177                                 sge->n = 0;
 178                         }
 179                         sge->vaddr =
 180                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 181                         sge->length =
 182                                 sge->mr->map[sge->m]->segs[sge->n].length;
 183                 }
 184                 data += len;
 185                 length -= len;
 186         }
 187 }
 188
 189 /**
 190  * ipath_skip_sge - skip over SGE memory - XXX almost dup of prev func
 191  * @ss: the SGE state
 192  * @length: the number of bytes to skip
 193  */
 194 void ipath_skip_sge(struct ipath_sge_state *ss, u32 length)
 195 {
 196         struct ipath_sge *sge = &ss->sge;
 197
 198         while (length) {
 199                 u32 len = sge->length;
 200
 201                 BUG_ON(len == 0);
 202                 if (len > length)
 203                         len = length;
 204                 sge->vaddr += len;
 205                 sge->length -= len;
 206                 sge->sge_length -= len;
 207                 if (sge->sge_length == 0) {
 208                         if (--ss->num_sge)
 209                                 *sge = *ss->sg_list++;
 210                 } else if (sge->length == 0 && sge->mr != NULL) {
 211                         if (++sge->n >= IPATH_SEGSZ) {
 212                                 if (++sge->m >= sge->mr->mapsz)
 213                                         break;
 214                                 sge->n = 0;
 215                         }
 216                         sge->vaddr =
 217                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 218                         sge->length =
 219                                 sge->mr->map[sge->m]->segs[sge->n].length;
 220                 }
 221                 length -= len;
 222         }
 223 }
 224
 225 /**
 226  * ipath_post_send - post a send on a QP
 227  * @ibqp: the QP to post the send on
 228  * @wr: the list of work requests to post
 229  * @bad_wr: the first bad WR is put here
 230  *
 231  * This may be called from interrupt context.
 232  */
 233 static int ipath_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 234                            struct ib_send_wr **bad_wr)
 235 {
 236         struct ipath_qp *qp = to_iqp(ibqp);
 237         int err = 0;
 238
 239         /* Check that state is OK to post send. */
 240         if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_SEND_OK)) {
 241                 *bad_wr = wr;
 242                 err = -EINVAL;
 243                 goto bail;
 244         }
 245
 246         for (; wr; wr = wr->next) {
 247                 switch (qp->ibqp.qp_type) {
 248                 case IB_QPT_UC:
 249                 case IB_QPT_RC:
 250                         err = ipath_post_ruc_send(qp, wr);
 251                         break;
 252
 253                 case IB_QPT_SMI:
 254                 case IB_QPT_GSI:
 255                 case IB_QPT_UD:
 256                         err = ipath_post_ud_send(qp, wr);
 257                         break;
 258
 259                 default:
 260                         err = -EINVAL;
 261                 }
 262                 if (err) {
 263                         *bad_wr = wr;
 264                         break;
 265                 }
 266         }
 267
 268 bail:
 269         return err;
 270 }
 271
 272 /**
 273  * ipath_post_receive - post a receive on a QP
 274  * @ibqp: the QP to post the receive on
 275  * @wr: the WR to post
 276  * @bad_wr: the first bad WR is put here
 277  *
 278  * This may be called from interrupt context.
 279  */
 280 static int ipath_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 281                               struct ib_recv_wr **bad_wr)
 282 {
 283         struct ipath_qp *qp = to_iqp(ibqp);
 284         struct ipath_rwq *wq = qp->r_rq.wq;
 285         unsigned long flags;
 286         int ret;
 287
 288         /* Check that state is OK to post receive. */
 289         if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_RECV_OK) || !wq) {
 290                 *bad_wr = wr;
 291                 ret = -EINVAL;
 292                 goto bail;
 293         }
 294
 295         for (; wr; wr = wr->next) {
 296                 struct ipath_rwqe *wqe;
 297                 u32 next;
 298                 int i;
 299
 300                 if ((unsigned) wr->num_sge > qp->r_rq.max_sge) {
 301                         *bad_wr = wr;
 302                         ret = -ENOMEM;
 303                         goto bail;
 304                 }
 305
 306                 spin_lock_irqsave(&qp->r_rq.lock, flags);
 307                 next = wq->head + 1;
 308                 if (next >= qp->r_rq.size)
 309                         next = 0;
 310                 if (next == wq->tail) {
 311                         spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 312                         *bad_wr = wr;
 313                         ret = -ENOMEM;
 314                         goto bail;
 315                 }
 316
 317                 wqe = get_rwqe_ptr(&qp->r_rq, wq->head);
 318                 wqe->wr_id = wr->wr_id;
 319                 wqe->num_sge = wr->num_sge;
 320                 for (i = 0; i < wr->num_sge; i++)
 321                         wqe->sg_list[i] = wr->sg_list[i];
 322                 wq->head = next;
 323                 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 324         }
 325         ret = 0;
 326
 327 bail:
 328         return ret;
 329 }
 330
 331 /**
 332  * ipath_qp_rcv - processing an incoming packet on a QP
 333  * @dev: the device the packet came on
 334  * @hdr: the packet header
 335  * @has_grh: true if the packet has a GRH
 336  * @data: the packet data
 337  * @tlen: the packet length
 338  * @qp: the QP the packet came on
 339  *
 340  * This is called from ipath_ib_rcv() to process an incoming packet
 341  * for the given QP.
 342  * Called at interrupt level.
 343  */
 344 static void ipath_qp_rcv(struct ipath_ibdev *dev,
 345                          struct ipath_ib_header *hdr, int has_grh,
 346                          void *data, u32 tlen, struct ipath_qp *qp)
 347 {
 348         /* Check for valid receive state. */
 349         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
 350                 dev->n_pkt_drops++;
 351                 return;
 352         }
 353
 354         switch (qp->ibqp.qp_type) {
 355         case IB_QPT_SMI:
 356         case IB_QPT_GSI:
 357         case IB_QPT_UD:
 358                 ipath_ud_rcv(dev, hdr, has_grh, data, tlen, qp);
 359                 break;
 360
 361         case IB_QPT_RC:
 362                 ipath_rc_rcv(dev, hdr, has_grh, data, tlen, qp);
 363                 break;
 364
 365         case IB_QPT_UC:
 366                 ipath_uc_rcv(dev, hdr, has_grh, data, tlen, qp);
 367                 break;
 368
 369         default:
 370                 break;
 371         }
 372 }
 373
 374 /**
 375  * ipath_ib_rcv - process an incoming packet
 376  * @arg: the device pointer
 377  * @rhdr: the header of the packet
 378  * @data: the packet data
 379  * @tlen: the packet length
 380  *
 381  * This is called from ipath_kreceive() to process an incoming packet at
 382  * interrupt level. Tlen is the length of the header + data + CRC in bytes.
 383  */
 384 void ipath_ib_rcv(struct ipath_ibdev *dev, void *rhdr, void *data,
 385                   u32 tlen)
 386 {
 387         struct ipath_ib_header *hdr = rhdr;
 388         struct ipath_other_headers *ohdr;
 389         struct ipath_qp *qp;
 390         u32 qp_num;
 391         int lnh;
 392         u8 opcode;
 393         u16 lid;
 394
 395         if (unlikely(dev == NULL))
 396                 goto bail;
 397
 398         if (unlikely(tlen < 24)) {      /* LRH+BTH+CRC */
 399                 dev->rcv_errors++;
 400                 goto bail;
 401         }
 402
 403         /* Check for a valid destination LID (see ch. 7.11.1). */
 404         lid = be16_to_cpu(hdr->lrh[1]);
 405         if (lid < IPATH_MULTICAST_LID_BASE) {
 406                 lid &= ~((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
 407                 if (unlikely(lid != dev->dd->ipath_lid)) {
 408                         dev->rcv_errors++;
 409                         goto bail;
 410                 }
 411         }
 412
 413         /* Check for GRH */
 414         lnh = be16_to_cpu(hdr->lrh[0]) & 3;
 415         if (lnh == IPATH_LRH_BTH)
 416                 ohdr = &hdr->u.oth;
 417         else if (lnh == IPATH_LRH_GRH)
 418                 ohdr = &hdr->u.l.oth;
 419         else {
 420                 dev->rcv_errors++;
 421                 goto bail;
 422         }
 423
 424         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
 425         dev->opstats[opcode].n_bytes += tlen;
 426         dev->opstats[opcode].n_packets++;
 427
 428         /* Get the destination QP number. */
 429         qp_num = be32_to_cpu(ohdr->bth[1]) & IPATH_QPN_MASK;
 430         if (qp_num == IPATH_MULTICAST_QPN) {
 431                 struct ipath_mcast *mcast;
 432                 struct ipath_mcast_qp *p;
 433
 434                 mcast = ipath_mcast_find(&hdr->u.l.grh.dgid);
 435                 if (mcast == NULL) {
 436                         dev->n_pkt_drops++;
 437                         goto bail;
 438                 }
 439                 dev->n_multicast_rcv++;
 440                 list_for_each_entry_rcu(p, &mcast->qp_list, list)
 441                         ipath_qp_rcv(dev, hdr, lnh == IPATH_LRH_GRH, data,
 442                                      tlen, p->qp);
 443                 /*
 444                  * Notify ipath_multicast_detach() if it is waiting for us
 445                  * to finish.
 446                  */
 447                 if (atomic_dec_return(&mcast->refcount) <= 1)
 448                         wake_up(&mcast->wait);
 449         } else {
 450                 qp = ipath_lookup_qpn(&dev->qp_table, qp_num);
 451                 if (qp) {
 452                         dev->n_unicast_rcv++;
 453                         ipath_qp_rcv(dev, hdr, lnh == IPATH_LRH_GRH, data,
 454                                      tlen, qp);
 455                         /*
 456                          * Notify ipath_destroy_qp() if it is waiting
 457                          * for us to finish.
 458                          */
 459                         if (atomic_dec_and_test(&qp->refcount))
 460                                 wake_up(&qp->wait);
 461                 } else
 462                         dev->n_pkt_drops++;
 463         }
 464
 465 bail:;
 466 }
 467
 468 /**
 469  * ipath_ib_timer - verbs timer
 470  * @arg: the device pointer
 471  *
 472  * This is called from ipath_do_rcv_timer() at interrupt level to check for
 473  * QPs which need retransmits and to collect performance numbers.
 474  */
 475 void ipath_ib_timer(struct ipath_ibdev *dev)
 476 {
 477         struct ipath_qp *resend = NULL;
 478         struct list_head *last;
 479         struct ipath_qp *qp;
 480         unsigned long flags;
 481
 482         if (dev == NULL)
 483                 return;
 484
 485         spin_lock_irqsave(&dev->pending_lock, flags);
 486         /* Start filling the next pending queue. */
 487         if (++dev->pending_index >= ARRAY_SIZE(dev->pending))
 488                 dev->pending_index = 0;
 489         /* Save any requests still in the new queue, they have timed out. */
 490         last = &dev->pending[dev->pending_index];
 491         while (!list_empty(last)) {
 492                 qp = list_entry(last->next, struct ipath_qp, timerwait);
 493                 list_del_init(&qp->timerwait);
 494                 qp->timer_next = resend;
 495                 resend = qp;
 496                 atomic_inc(&qp->refcount);
 497         }
 498         last = &dev->rnrwait;
 499         if (!list_empty(last)) {
 500                 qp = list_entry(last->next, struct ipath_qp, timerwait);
 501                 if (--qp->s_rnr_timeout == 0) {
 502                         do {
 503                                 list_del_init(&qp->timerwait);
 504                                 tasklet_hi_schedule(&qp->s_task);
 505                                 if (list_empty(last))
 506                                         break;
 507                                 qp = list_entry(last->next, struct ipath_qp,
 508                                                 timerwait);
 509                         } while (qp->s_rnr_timeout == 0);
 510                 }
 511         }
 512         /*
 513          * We should only be in the started state if pma_sample_start != 0
 514          */
 515         if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_STARTED &&
 516             --dev->pma_sample_start == 0) {
 517                 dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_RUNNING;
 518                 ipath_snapshot_counters(dev->dd, &dev->ipath_sword,
 519                                         &dev->ipath_rword,
 520                                         &dev->ipath_spkts,
 521                                         &dev->ipath_rpkts,
 522                                         &dev->ipath_xmit_wait);
 523         }
 524         if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_RUNNING) {
 525                 if (dev->pma_sample_interval == 0) {
 526                         u64 ta, tb, tc, td, te;
 527
 528                         dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_DONE;
 529                         ipath_snapshot_counters(dev->dd, &ta, &tb,
 530                                                 &tc, &td, &te);
 531
 532                         dev->ipath_sword = ta - dev->ipath_sword;
 533                         dev->ipath_rword = tb - dev->ipath_rword;
 534                         dev->ipath_spkts = tc - dev->ipath_spkts;
 535                         dev->ipath_rpkts = td - dev->ipath_rpkts;
 536                         dev->ipath_xmit_wait = te - dev->ipath_xmit_wait;
 537                 }
 538                 else
 539                         dev->pma_sample_interval--;
 540         }
 541         spin_unlock_irqrestore(&dev->pending_lock, flags);
 542
 543         /* XXX What if timer fires again while this is running? */
 544         for (qp = resend; qp != NULL; qp = qp->timer_next) {
 545                 struct ib_wc wc;
 546
 547                 spin_lock_irqsave(&qp->s_lock, flags);
 548                 if (qp->s_last != qp->s_tail && qp->state == IB_QPS_RTS) {
 549                         dev->n_timeouts++;
 550                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
 551                 }
 552                 spin_unlock_irqrestore(&qp->s_lock, flags);
 553
 554                 /* Notify ipath_destroy_qp() if it is waiting. */
 555                 if (atomic_dec_and_test(&qp->refcount))
 556                         wake_up(&qp->wait);
 557         }
 558 }
 559
 560 static void update_sge(struct ipath_sge_state *ss, u32 length)
 561 {
 562         struct ipath_sge *sge = &ss->sge;
 563
 564         sge->vaddr += length;
 565         sge->length -= length;
 566         sge->sge_length -= length;
 567         if (sge->sge_length == 0) {
 568                 if (--ss->num_sge)
 569                         *sge = *ss->sg_list++;
 570         } else if (sge->length == 0 && sge->mr != NULL) {
 571                 if (++sge->n >= IPATH_SEGSZ) {
 572                         if (++sge->m >= sge->mr->mapsz)
 573                                 return;
 574                         sge->n = 0;
 575                 }
 576                 sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
 577                 sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
 578         }
 579 }
 580
 581 #ifdef __LITTLE_ENDIAN
 582 static inline u32 get_upper_bits(u32 data, u32 shift)
 583 {
 584         return data >> shift;
 585 }
 586
 587 static inline u32 set_upper_bits(u32 data, u32 shift)
 588 {
 589         return data << shift;
 590 }
 591
 592 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
 593 {
 594         data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
 595         data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
 596         return data;
 597 }
 598 #else
 599 static inline u32 get_upper_bits(u32 data, u32 shift)
 600 {
 601         return data << shift;
 602 }
 603
 604 static inline u32 set_upper_bits(u32 data, u32 shift)
 605 {
 606         return data >> shift;
 607 }
 608
 609 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
 610 {
 611         data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
 612         data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
 613         return data;
 614 }
 615 #endif
 616
 617 static void copy_io(u32 __iomem *piobuf, struct ipath_sge_state *ss,
 618                     u32 length)
 619 {
 620         u32 extra = 0;
 621         u32 data = 0;
 622         u32 last;
 623
 624         while (1) {
 625                 u32 len = ss->sge.length;
 626                 u32 off;
 627
 628                 BUG_ON(len == 0);
 629                 if (len > length)
 630                         len = length;
 631                 if (len > ss->sge.sge_length)
 632                         len = ss->sge.sge_length;
 633                 /* If the source address is not aligned, try to align it. */
 634                 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
 635                 if (off) {
 636                         u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
 637                                             ~(sizeof(u32) - 1));
 638                         u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
 639                         u32 y;
 640
 641                         y = sizeof(u32) - off;
 642                         if (len > y)
 643                                 len = y;
 644                         if (len + extra >= sizeof(u32)) {
 645                                 data |= set_upper_bits(v, extra *
 646                                                        BITS_PER_BYTE);
 647                                 len = sizeof(u32) - extra;
 648                                 if (len == length) {
 649                                         last = data;
 650                                         break;
 651                                 }
 652                                 __raw_writel(data, piobuf);
 653                                 piobuf++;
 654                                 extra = 0;
 655                                 data = 0;
 656                         } else {
 657                                 /* Clear unused upper bytes */
 658                                 data |= clear_upper_bytes(v, len, extra);
 659                                 if (len == length) {
 660                                         last = data;
 661                                         break;
 662                                 }
 663                                 extra += len;
 664                         }
 665                 } else if (extra) {
 666                         /* Source address is aligned. */
 667                         u32 *addr = (u32 *) ss->sge.vaddr;
 668                         int shift = extra * BITS_PER_BYTE;
 669                         int ushift = 32 - shift;
 670                         u32 l = len;
 671
 672                         while (l >= sizeof(u32)) {
 673                                 u32 v = *addr;
 674
 675                                 data |= set_upper_bits(v, shift);
 676                                 __raw_writel(data, piobuf);
 677                                 data = get_upper_bits(v, ushift);
 678                                 piobuf++;
 679                                 addr++;
 680                                 l -= sizeof(u32);
 681                         }
 682                         /*
 683                          * We still have 'extra' number of bytes leftover.
 684                          */
 685                         if (l) {
 686                                 u32 v = *addr;
 687
 688                                 if (l + extra >= sizeof(u32)) {
 689                                         data |= set_upper_bits(v, shift);
 690                                         len -= l + extra - sizeof(u32);
 691                                         if (len == length) {
 692                                                 last = data;
 693                                                 break;
 694                                         }
 695                                         __raw_writel(data, piobuf);
 696                                         piobuf++;
 697                                         extra = 0;
 698                                         data = 0;
 699                                 } else {
 700                                         /* Clear unused upper bytes */
 701                                         data |= clear_upper_bytes(v, l,
 702                                                                   extra);
 703                                         if (len == length) {
 704                                                 last = data;
 705                                                 break;
 706                                         }
 707                                         extra += l;
 708                                 }
 709                         } else if (len == length) {
 710                                 last = data;
 711                                 break;
 712                         }
 713                 } else if (len == length) {
 714                         u32 w;
 715
 716                         /*
 717                          * Need to round up for the last dword in the
 718                          * packet.
 719                          */
 720                         w = (len + 3) >> 2;
 721                         __iowrite32_copy(piobuf, ss->sge.vaddr, w - 1);
 722                         piobuf += w - 1;
 723                         last = ((u32 *) ss->sge.vaddr)[w - 1];
 724                         break;
 725                 } else {
 726                         u32 w = len >> 2;
 727
 728                         __iowrite32_copy(piobuf, ss->sge.vaddr, w);
 729                         piobuf += w;
 730
 731                         extra = len & (sizeof(u32) - 1);
 732                         if (extra) {
 733                                 u32 v = ((u32 *) ss->sge.vaddr)[w];
 734
 735                                 /* Clear unused upper bytes */
 736                                 data = clear_upper_bytes(v, extra, 0);
 737                         }
 738                 }
 739                 update_sge(ss, len);
 740                 length -= len;
 741         }
 742         /* Update address before sending packet. */
 743         update_sge(ss, length);
 744         /* must flush early everything before trigger word */
 745         ipath_flush_wc();
 746         __raw_writel(last, piobuf);
 747         /* be sure trigger word is written */
 748         ipath_flush_wc();
 749 }
 750
 751 /**
 752  * ipath_verbs_send - send a packet
 753  * @dd: the infinipath device
 754  * @hdrwords: the number of words in the header
 755  * @hdr: the packet header
 756  * @len: the length of the packet in bytes
 757  * @ss: the SGE to send
 758  */
 759 int ipath_verbs_send(struct ipath_devdata *dd, u32 hdrwords,
 760                      u32 *hdr, u32 len, struct ipath_sge_state *ss)
 761 {
 762         u32 __iomem *piobuf;
 763         u32 plen;
 764         int ret;
 765
 766         /* +1 is for the qword padding of pbc */
 767         plen = hdrwords + ((len + 3) >> 2) + 1;
 768         if (unlikely((plen << 2) > dd->ipath_ibmaxlen)) {
 769                 ipath_dbg("packet len 0x%x too long, failing\n", plen);
 770                 ret = -EINVAL;
 771                 goto bail;
 772         }
 773
 774         /* Get a PIO buffer to use. */
 775         piobuf = ipath_getpiobuf(dd, NULL);
 776         if (unlikely(piobuf == NULL)) {
 777                 ret = -EBUSY;
 778                 goto bail;
 779         }
 780
 781         /*
 782          * Write len to control qword, no flags.
 783          * We have to flush after the PBC for correctness on some cpus
 784          * or WC buffer can be written out of order.
 785          */
 786         writeq(plen, piobuf);
 787         ipath_flush_wc();
 788         piobuf += 2;
 789         if (len == 0) {
 790                 /*
 791                  * If there is just the header portion, must flush before
 792                  * writing last word of header for correctness, and after
 793                  * the last header word (trigger word).
 794                  */
 795                 __iowrite32_copy(piobuf, hdr, hdrwords - 1);
 796                 ipath_flush_wc();
 797                 __raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
 798                 ipath_flush_wc();
 799                 ret = 0;
 800                 goto bail;
 801         }
 802
 803         __iowrite32_copy(piobuf, hdr, hdrwords);
 804         piobuf += hdrwords;
 805
 806         /* The common case is aligned and contained in one segment. */
 807         if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
 808                    !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
 809                 u32 w;
 810                 u32 *addr = (u32 *) ss->sge.vaddr;
 811
 812                 /* Update address before sending packet. */
 813                 update_sge(ss, len);
 814                 /* Need to round up for the last dword in the packet. */
 815                 w = (len + 3) >> 2;
 816                 __iowrite32_copy(piobuf, addr, w - 1);
 817                 /* must flush early everything before trigger word */
 818                 ipath_flush_wc();
 819                 __raw_writel(addr[w - 1], piobuf + w - 1);
 820                 /* be sure trigger word is written */
 821                 ipath_flush_wc();
 822                 ret = 0;
 823                 goto bail;
 824         }
 825         copy_io(piobuf, ss, len);
 826         ret = 0;
 827
 828 bail:
 829         return ret;
 830 }
 831
 832 int ipath_snapshot_counters(struct ipath_devdata *dd, u64 *swords,
 833                             u64 *rwords, u64 *spkts, u64 *rpkts,
 834                             u64 *xmit_wait)
 835 {
 836         int ret;
 837
 838         if (!(dd->ipath_flags & IPATH_INITTED)) {
 839                 /* no hardware, freeze, etc. */
 840                 ipath_dbg("unit %u not usable\n", dd->ipath_unit);
 841                 ret = -EINVAL;
 842                 goto bail;
 843         }
 844         *swords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt);
 845         *rwords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
 846         *spkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
 847         *rpkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
 848         *xmit_wait = ipath_snap_cntr(dd, dd->ipath_cregs->cr_sendstallcnt);
 849
 850         ret = 0;
 851
 852 bail:
 853         return ret;
 854 }
 855
 856 /**
 857  * ipath_get_counters - get various chip counters
 858  * @dd: the infinipath device
 859  * @cntrs: counters are placed here
 860  *
 861  * Return the counters needed by recv_pma_get_portcounters().
 862  */
 863 int ipath_get_counters(struct ipath_devdata *dd,
 864                        struct ipath_verbs_counters *cntrs)
 865 {
 866         int ret;
 867
 868         if (!(dd->ipath_flags & IPATH_INITTED)) {
 869                 /* no hardware, freeze, etc. */
 870                 ipath_dbg("unit %u not usable\n", dd->ipath_unit);
 871                 ret = -EINVAL;
 872                 goto bail;
 873         }
 874         cntrs->symbol_error_counter =
 875                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_ibsymbolerrcnt);
 876         cntrs->link_error_recovery_counter =
 877                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_iblinkerrrecovcnt);
 878         /*
 879          * The link downed counter counts when the other side downs the
 880          * connection.  We add in the number of times we downed the link
 881          * due to local link integrity errors to compensate.
 882          */
 883         cntrs->link_downed_counter =
 884                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_iblinkdowncnt);
 885         cntrs->port_rcv_errors =
 886                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_rxdroppktcnt) +
 887                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_rcvovflcnt) +
 888                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_portovflcnt) +
 889                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_err_rlencnt) +
 890                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_invalidrlencnt) +
 891                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_erricrccnt) +
 892                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_errvcrccnt) +
 893                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_errlpcrccnt) +
 894                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_badformatcnt);
 895         cntrs->port_rcv_remphys_errors =
 896                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_rcvebpcnt);
 897         cntrs->port_xmit_discards =
 898                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_unsupvlcnt);
 899         cntrs->port_xmit_data =
 900                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt);
 901         cntrs->port_rcv_data =
 902                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
 903         cntrs->port_xmit_packets =
 904                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
 905         cntrs->port_rcv_packets =
 906                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
 907         cntrs->local_link_integrity_errors = dd->ipath_lli_errors;
 908         cntrs->excessive_buffer_overrun_errors = 0; /* XXX */
 909
 910         ret = 0;
 911
 912 bail:
 913         return ret;
 914 }
 915
 916 /**
 917  * ipath_ib_piobufavail - callback when a PIO buffer is available
 918  * @arg: the device pointer
 919  *
 920  * This is called from ipath_intr() at interrupt level when a PIO buffer is
 921  * available after ipath_verbs_send() returned an error that no buffers were
 922  * available.  Return 1 if we consumed all the PIO buffers and we still have
 923  * QPs waiting for buffers (for now, just do a tasklet_hi_schedule and
 924  * return zero).
 925  */
 926 int ipath_ib_piobufavail(struct ipath_ibdev *dev)
 927 {
 928         struct ipath_qp *qp;
 929         unsigned long flags;
 930
 931         if (dev == NULL)
 932                 goto bail;
 933
 934         spin_lock_irqsave(&dev->pending_lock, flags);
 935         while (!list_empty(&dev->piowait)) {
 936                 qp = list_entry(dev->piowait.next, struct ipath_qp,
 937                                 piowait);
 938                 list_del_init(&qp->piowait);
 939                 tasklet_hi_schedule(&qp->s_task);
 940         }
 941         spin_unlock_irqrestore(&dev->pending_lock, flags);
 942
 943 bail:
 944         return 0;
 945 }
 946
 947 static int ipath_query_device(struct ib_device *ibdev,
 948                               struct ib_device_attr *props)
 949 {
 950         struct ipath_ibdev *dev = to_idev(ibdev);
 951
 952         memset(props, 0, sizeof(*props));
 953
 954         props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
 955                 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
 956                 IB_DEVICE_SYS_IMAGE_GUID;
 957         props->page_size_cap = PAGE_SIZE;
 958         props->vendor_id = dev->dd->ipath_vendorid;
 959         props->vendor_part_id = dev->dd->ipath_deviceid;
 960         props->hw_ver = dev->dd->ipath_pcirev;
 961
 962         props->sys_image_guid = dev->sys_image_guid;
 963
 964         props->max_mr_size = ~0ull;
 965         props->max_qp = ib_ipath_max_qps;
 966         props->max_qp_wr = ib_ipath_max_qp_wrs;
 967         props->max_sge = ib_ipath_max_sges;
 968         props->max_cq = ib_ipath_max_cqs;
 969         props->max_ah = ib_ipath_max_ahs;
 970         props->max_cqe = ib_ipath_max_cqes;
 971         props->max_mr = dev->lk_table.max;
 972         props->max_pd = ib_ipath_max_pds;
 973         props->max_qp_rd_atom = 1;
 974         props->max_qp_init_rd_atom = 1;
 975         /* props->max_res_rd_atom */
 976         props->max_srq = ib_ipath_max_srqs;
 977         props->max_srq_wr = ib_ipath_max_srq_wrs;
 978         props->max_srq_sge = ib_ipath_max_srq_sges;
 979         /* props->local_ca_ack_delay */
 980         props->atomic_cap = IB_ATOMIC_HCA;
 981         props->max_pkeys = ipath_get_npkeys(dev->dd);
 982         props->max_mcast_grp = ib_ipath_max_mcast_grps;
 983         props->max_mcast_qp_attach = ib_ipath_max_mcast_qp_attached;
 984         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
 985                 props->max_mcast_grp;
 986
 987         return 0;
 988 }
 989
 990 const u8 ipath_cvt_physportstate[16] = {
 991         [INFINIPATH_IBCS_LT_STATE_DISABLED] = 3,
 992         [INFINIPATH_IBCS_LT_STATE_LINKUP] = 5,
 993         [INFINIPATH_IBCS_LT_STATE_POLLACTIVE] = 2,
 994         [INFINIPATH_IBCS_LT_STATE_POLLQUIET] = 2,
 995         [INFINIPATH_IBCS_LT_STATE_SLEEPDELAY] = 1,
 996         [INFINIPATH_IBCS_LT_STATE_SLEEPQUIET] = 1,
 997         [INFINIPATH_IBCS_LT_STATE_CFGDEBOUNCE] = 4,
 998         [INFINIPATH_IBCS_LT_STATE_CFGRCVFCFG] = 4,
 999         [INFINIPATH_IBCS_LT_STATE_CFGWAITRMT] = 4,
1000         [INFINIPATH_IBCS_LT_STATE_CFGIDLE] = 4,
1001         [INFINIPATH_IBCS_LT_STATE_RECOVERRETRAIN] = 6,
1002         [INFINIPATH_IBCS_LT_STATE_RECOVERWAITRMT] = 6,
1003         [INFINIPATH_IBCS_LT_STATE_RECOVERIDLE] = 6,
1004 };
1005
1006 u32 ipath_get_cr_errpkey(struct ipath_devdata *dd)
1007 {
1008         return ipath_read_creg32(dd, dd->ipath_cregs->cr_errpkey);
1009 }
1010
1011 static int ipath_query_port(struct ib_device *ibdev,
1012                             u8 port, struct ib_port_attr *props)
1013 {
1014         struct ipath_ibdev *dev = to_idev(ibdev);
1015         enum ib_mtu mtu;
1016         u16 lid = dev->dd->ipath_lid;
1017         u64 ibcstat;
1018
1019         memset(props, 0, sizeof(*props));
1020         props->lid = lid ? lid : __constant_be16_to_cpu(IB_LID_PERMISSIVE);
1021         props->lmc = dev->mkeyprot_resv_lmc & 7;
1022         props->sm_lid = dev->sm_lid;
1023         props->sm_sl = dev->sm_sl;
1024         ibcstat = dev->dd->ipath_lastibcstat;
1025         props->state = ((ibcstat >> 4) & 0x3) + 1;
1026         /* See phys_state_show() */
1027         props->phys_state = ipath_cvt_physportstate[
1028                 dev->dd->ipath_lastibcstat & 0xf];
1029         props->port_cap_flags = dev->port_cap_flags;
1030         props->gid_tbl_len = 1;
1031         props->max_msg_sz = 0x80000000;
1032         props->pkey_tbl_len = ipath_get_npkeys(dev->dd);
1033         props->bad_pkey_cntr = ipath_get_cr_errpkey(dev->dd) -
1034                 dev->z_pkey_violations;
1035         props->qkey_viol_cntr = dev->qkey_violations;
1036         props->active_width = IB_WIDTH_4X;
1037         /* See rate_show() */
1038         props->active_speed = 1;        /* Regular 10Mbs speed. */
1039         props->max_vl_num = 1;          /* VLCap = VL0 */
1040         props->init_type_reply = 0;
1041
1042         props->max_mtu = IB_MTU_4096;
1043         switch (dev->dd->ipath_ibmtu) {
1044         case 4096:
1045                 mtu = IB_MTU_4096;
1046                 break;
1047         case 2048:
1048                 mtu = IB_MTU_2048;
1049                 break;
1050         case 1024:
1051                 mtu = IB_MTU_1024;
1052                 break;
1053         case 512:
1054                 mtu = IB_MTU_512;
1055                 break;
1056         case 256:
1057                 mtu = IB_MTU_256;
1058                 break;
1059         default:
1060                 mtu = IB_MTU_2048;
1061         }
1062         props->active_mtu = mtu;
1063         props->subnet_timeout = dev->subnet_timeout;
1064
1065         return 0;
1066 }
1067
1068 static int ipath_modify_device(struct ib_device *device,
1069                                int device_modify_mask,
1070                                struct ib_device_modify *device_modify)
1071 {
1072         int ret;
1073
1074         if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1075                                    IB_DEVICE_MODIFY_NODE_DESC)) {
1076                 ret = -EOPNOTSUPP;
1077                 goto bail;
1078         }
1079
1080         if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC)
1081                 memcpy(device->node_desc, device_modify->node_desc, 64);
1082
1083         if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
1084                 to_idev(device)->sys_image_guid =
1085                         cpu_to_be64(device_modify->sys_image_guid);
1086
1087         ret = 0;
1088
1089 bail:
1090         return ret;
1091 }
1092
1093 static int ipath_modify_port(struct ib_device *ibdev,
1094                              u8 port, int port_modify_mask,
1095                              struct ib_port_modify *props)
1096 {
1097         struct ipath_ibdev *dev = to_idev(ibdev);
1098
1099         dev->port_cap_flags |= props->set_port_cap_mask;
1100         dev->port_cap_flags &= ~props->clr_port_cap_mask;
1101         if (port_modify_mask & IB_PORT_SHUTDOWN)
1102                 ipath_set_linkstate(dev->dd, IPATH_IB_LINKDOWN);
1103         if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR)
1104                 dev->qkey_violations = 0;
1105         return 0;
1106 }
1107
1108 static int ipath_query_gid(struct ib_device *ibdev, u8 port,
1109                            int index, union ib_gid *gid)
1110 {
1111         struct ipath_ibdev *dev = to_idev(ibdev);
1112         int ret;
1113
1114         if (index >= 1) {
1115                 ret = -EINVAL;
1116                 goto bail;
1117         }
1118         gid->global.subnet_prefix = dev->gid_prefix;
1119         gid->global.interface_id = dev->dd->ipath_guid;
1120
1121         ret = 0;
1122
1123 bail:
1124         return ret;
1125 }
1126
1127 static struct ib_pd *ipath_alloc_pd(struct ib_device *ibdev,
1128                                     struct ib_ucontext *context,
1129                                     struct ib_udata *udata)
1130 {
1131         struct ipath_ibdev *dev = to_idev(ibdev);
1132         struct ipath_pd *pd;
1133         struct ib_pd *ret;
1134
1135         /*
1136          * This is actually totally arbitrary.  Some correctness tests
1137          * assume there's a maximum number of PDs that can be allocated.
1138          * We don't actually have this limit, but we fail the test if
1139          * we allow allocations of more than we report for this value.
1140          */
1141
1142         pd = kmalloc(sizeof *pd, GFP_KERNEL);
1143         if (!pd) {
1144                 ret = ERR_PTR(-ENOMEM);
1145                 goto bail;
1146         }
1147
1148         spin_lock(&dev->n_pds_lock);
1149         if (dev->n_pds_allocated == ib_ipath_max_pds) {
1150                 spin_unlock(&dev->n_pds_lock);
1151                 kfree(pd);
1152                 ret = ERR_PTR(-ENOMEM);
1153                 goto bail;
1154         }
1155
1156         dev->n_pds_allocated++;
1157         spin_unlock(&dev->n_pds_lock);
1158
1159         /* ib_alloc_pd() will initialize pd->ibpd. */
1160         pd->user = udata != NULL;
1161
1162         ret = &pd->ibpd;
1163
1164 bail:
1165         return ret;
1166 }
1167
1168 static int ipath_dealloc_pd(struct ib_pd *ibpd)
1169 {
1170         struct ipath_pd *pd = to_ipd(ibpd);
1171         struct ipath_ibdev *dev = to_idev(ibpd->device);
1172
1173         spin_lock(&dev->n_pds_lock);
1174         dev->n_pds_allocated--;
1175         spin_unlock(&dev->n_pds_lock);
1176
1177         kfree(pd);
1178
1179         return 0;
1180 }
1181
1182 /**
1183  * ipath_create_ah - create an address handle
1184  * @pd: the protection domain
1185  * @ah_attr: the attributes of the AH
1186  *
1187  * This may be called from interrupt context.
1188  */
1189 static struct ib_ah *ipath_create_ah(struct ib_pd *pd,
1190                                      struct ib_ah_attr *ah_attr)
1191 {
1192         struct ipath_ah *ah;
1193         struct ib_ah *ret;
1194         struct ipath_ibdev *dev = to_idev(pd->device);
1195
1196         /* A multicast address requires a GRH (see ch. 8.4.1). */
1197         if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
1198             ah_attr->dlid != IPATH_PERMISSIVE_LID &&
1199             !(ah_attr->ah_flags & IB_AH_GRH)) {
1200                 ret = ERR_PTR(-EINVAL);
1201                 goto bail;
1202         }
1203
1204         if (ah_attr->dlid == 0) {
1205                 ret = ERR_PTR(-EINVAL);
1206                 goto bail;
1207         }
1208
1209         if (ah_attr->port_num < 1 ||
1210             ah_attr->port_num > pd->device->phys_port_cnt) {
1211                 ret = ERR_PTR(-EINVAL);
1212                 goto bail;
1213         }
1214
1215         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
1216         if (!ah) {
1217                 ret = ERR_PTR(-ENOMEM);
1218                 goto bail;
1219         }
1220
1221         spin_lock(&dev->n_ahs_lock);
1222         if (dev->n_ahs_allocated == ib_ipath_max_ahs) {
1223                 spin_unlock(&dev->n_ahs_lock);
1224                 kfree(ah);
1225                 ret = ERR_PTR(-ENOMEM);
1226                 goto bail;
1227         }
1228
1229         dev->n_ahs_allocated++;
1230         spin_unlock(&dev->n_ahs_lock);
1231
1232         /* ib_create_ah() will initialize ah->ibah. */
1233         ah->attr = *ah_attr;
1234
1235         ret = &ah->ibah;
1236
1237 bail:
1238         return ret;
1239 }
1240
1241 /**
1242  * ipath_destroy_ah - destroy an address handle
1243  * @ibah: the AH to destroy
1244  *
1245  * This may be called from interrupt context.
1246  */
1247 static int ipath_destroy_ah(struct ib_ah *ibah)
1248 {
1249         struct ipath_ibdev *dev = to_idev(ibah->device);
1250         struct ipath_ah *ah = to_iah(ibah);
1251
1252         spin_lock(&dev->n_ahs_lock);
1253         dev->n_ahs_allocated--;
1254         spin_unlock(&dev->n_ahs_lock);
1255
1256         kfree(ah);
1257
1258         return 0;
1259 }
1260
1261 static int ipath_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1262 {
1263         struct ipath_ah *ah = to_iah(ibah);
1264
1265         *ah_attr = ah->attr;
1266
1267         return 0;
1268 }
1269
1270 /**
1271  * ipath_get_npkeys - return the size of the PKEY table for port 0
1272  * @dd: the infinipath device
1273  */
1274 unsigned ipath_get_npkeys(struct ipath_devdata *dd)
1275 {
1276         return ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys);
1277 }
1278
1279 /**
1280  * ipath_get_pkey - return the indexed PKEY from the port 0 PKEY table
1281  * @dd: the infinipath device
1282  * @index: the PKEY index
1283  */
1284 unsigned ipath_get_pkey(struct ipath_devdata *dd, unsigned index)
1285 {
1286         unsigned ret;
1287
1288         if (index >= ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys))
1289                 ret = 0;
1290         else
1291                 ret = dd->ipath_pd[0]->port_pkeys[index];
1292
1293         return ret;
1294 }
1295
1296 static int ipath_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1297                             u16 *pkey)
1298 {
1299         struct ipath_ibdev *dev = to_idev(ibdev);
1300         int ret;
1301
1302         if (index >= ipath_get_npkeys(dev->dd)) {
1303                 ret = -EINVAL;
1304                 goto bail;
1305         }
1306
1307         *pkey = ipath_get_pkey(dev->dd, index);
1308         ret = 0;
1309
1310 bail:
1311         return ret;
1312 }
1313
1314 /**
1315  * ipath_alloc_ucontext - allocate a ucontest
1316  * @ibdev: the infiniband device
1317  * @udata: not used by the InfiniPath driver
1318  */
1319
1320 static struct ib_ucontext *ipath_alloc_ucontext(struct ib_device *ibdev,
1321                                                 struct ib_udata *udata)
1322 {
1323         struct ipath_ucontext *context;
1324         struct ib_ucontext *ret;
1325
1326         context = kmalloc(sizeof *context, GFP_KERNEL);
1327         if (!context) {
1328                 ret = ERR_PTR(-ENOMEM);
1329                 goto bail;
1330         }
1331
1332         ret = &context->ibucontext;
1333
1334 bail:
1335         return ret;
1336 }
1337
1338 static int ipath_dealloc_ucontext(struct ib_ucontext *context)
1339 {
1340         kfree(to_iucontext(context));
1341         return 0;
1342 }
1343
1344 static int ipath_verbs_register_sysfs(struct ib_device *dev);
1345
1346 static void __verbs_timer(unsigned long arg)
1347 {
1348         struct ipath_devdata *dd = (struct ipath_devdata *) arg;
1349
1350         /*
1351          * If port 0 receive packet interrupts are not available, or
1352          * can be missed, poll the receive queue
1353          */
1354         if (dd->ipath_flags & IPATH_POLL_RX_INTR)
1355                 ipath_kreceive(dd);
1356
1357         /* Handle verbs layer timeouts. */
1358         ipath_ib_timer(dd->verbs_dev);
1359
1360         mod_timer(&dd->verbs_timer, jiffies + 1);
1361 }
1362
1363 static int enable_timer(struct ipath_devdata *dd)
1364 {
1365         /*
1366          * Early chips had a design flaw where the chip and kernel idea
1367          * of the tail register don't always agree, and therefore we won't
1368          * get an interrupt on the next packet received.
1369          * If the board supports per packet receive interrupts, use it.
1370          * Otherwise, the timer function periodically checks for packets
1371          * to cover this case.
1372          * Either way, the timer is needed for verbs layer related
1373          * processing.
1374          */
1375         if (dd->ipath_flags & IPATH_GPIO_INTR) {
1376                 ipath_write_kreg(dd, dd->ipath_kregs->kr_debugportselect,
1377                                  0x2074076542310ULL);
1378                 /* Enable GPIO bit 2 interrupt */
1379                 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1380                                  (u64) (1 << 2));
1381         }
1382
1383         init_timer(&dd->verbs_timer);
1384         dd->verbs_timer.function = __verbs_timer;
1385         dd->verbs_timer.data = (unsigned long)dd;
1386         dd->verbs_timer.expires = jiffies + 1;
1387         add_timer(&dd->verbs_timer);
1388
1389         return 0;
1390 }
1391
1392 static int disable_timer(struct ipath_devdata *dd)
1393 {
1394         /* Disable GPIO bit 2 interrupt */
1395         if (dd->ipath_flags & IPATH_GPIO_INTR)
1396                 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask, 0);
1397
1398         del_timer_sync(&dd->verbs_timer);
1399
1400         return 0;
1401 }
1402
1403 /**
1404  * ipath_register_ib_device - register our device with the infiniband core
1405  * @dd: the device data structure
1406  * Return the allocated ipath_ibdev pointer or NULL on error.
1407  */
1408 int ipath_register_ib_device(struct ipath_devdata *dd)
1409 {
1410         struct ipath_verbs_counters cntrs;
1411         struct ipath_ibdev *idev;
1412         struct ib_device *dev;
1413         int ret;
1414
1415         idev = (struct ipath_ibdev *)ib_alloc_device(sizeof *idev);
1416         if (idev == NULL) {
1417                 ret = -ENOMEM;
1418                 goto bail;
1419         }
1420
1421         dev = &idev->ibdev;
1422
1423         /* Only need to initialize non-zero fields. */
1424         spin_lock_init(&idev->n_pds_lock);
1425         spin_lock_init(&idev->n_ahs_lock);
1426         spin_lock_init(&idev->n_cqs_lock);
1427         spin_lock_init(&idev->n_qps_lock);
1428         spin_lock_init(&idev->n_srqs_lock);
1429         spin_lock_init(&idev->n_mcast_grps_lock);
1430
1431         spin_lock_init(&idev->qp_table.lock);
1432         spin_lock_init(&idev->lk_table.lock);
1433         idev->sm_lid = __constant_be16_to_cpu(IB_LID_PERMISSIVE);
1434         /* Set the prefix to the default value (see ch. 4.1.1) */
1435         idev->gid_prefix = __constant_cpu_to_be64(0xfe80000000000000ULL);
1436
1437         ret = ipath_init_qp_table(idev, ib_ipath_qp_table_size);
1438         if (ret)
1439                 goto err_qp;
1440
1441         /*
1442          * The top ib_ipath_lkey_table_size bits are used to index the
1443          * table.  The lower 8 bits can be owned by the user (copied from
1444          * the LKEY).  The remaining bits act as a generation number or tag.
1445          */
1446         idev->lk_table.max = 1 << ib_ipath_lkey_table_size;
1447         idev->lk_table.table = kzalloc(idev->lk_table.max *
1448                                        sizeof(*idev->lk_table.table),
1449                                        GFP_KERNEL);
1450         if (idev->lk_table.table == NULL) {
1451                 ret = -ENOMEM;
1452                 goto err_lk;
1453         }
1454         spin_lock_init(&idev->pending_lock);
1455         INIT_LIST_HEAD(&idev->pending[0]);
1456         INIT_LIST_HEAD(&idev->pending[1]);
1457         INIT_LIST_HEAD(&idev->pending[2]);
1458         INIT_LIST_HEAD(&idev->piowait);
1459         INIT_LIST_HEAD(&idev->rnrwait);
1460         idev->pending_index = 0;
1461         idev->port_cap_flags =
1462                 IB_PORT_SYS_IMAGE_GUID_SUP | IB_PORT_CLIENT_REG_SUP;
1463         idev->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1464         idev->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1465         idev->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1466         idev->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1467         idev->pma_counter_select[5] = IB_PMA_PORT_XMIT_WAIT;
1468         idev->link_width_enabled = 3;   /* 1x or 4x */
1469
1470         /* Snapshot current HW counters to "clear" them. */
1471         ipath_get_counters(dd, &cntrs);
1472         idev->z_symbol_error_counter = cntrs.symbol_error_counter;
1473         idev->z_link_error_recovery_counter =
1474                 cntrs.link_error_recovery_counter;
1475         idev->z_link_downed_counter = cntrs.link_downed_counter;
1476         idev->z_port_rcv_errors = cntrs.port_rcv_errors;
1477         idev->z_port_rcv_remphys_errors =
1478                 cntrs.port_rcv_remphys_errors;
1479         idev->z_port_xmit_discards = cntrs.port_xmit_discards;
1480         idev->z_port_xmit_data = cntrs.port_xmit_data;
1481         idev->z_port_rcv_data = cntrs.port_rcv_data;
1482         idev->z_port_xmit_packets = cntrs.port_xmit_packets;
1483         idev->z_port_rcv_packets = cntrs.port_rcv_packets;
1484         idev->z_local_link_integrity_errors =
1485                 cntrs.local_link_integrity_errors;
1486         idev->z_excessive_buffer_overrun_errors =
1487                 cntrs.excessive_buffer_overrun_errors;
1488
1489         /*
1490          * The system image GUID is supposed to be the same for all
1491          * IB HCAs in a single system but since there can be other
1492          * device types in the system, we can't be sure this is unique.
1493          */
1494         if (!sys_image_guid)
1495                 sys_image_guid = dd->ipath_guid;
1496         idev->sys_image_guid = sys_image_guid;
1497         idev->ib_unit = dd->ipath_unit;
1498         idev->dd = dd;
1499
1500         strlcpy(dev->name, "ipath%d", IB_DEVICE_NAME_MAX);
1501         dev->owner = THIS_MODULE;
1502         dev->node_guid = dd->ipath_guid;
1503         dev->uverbs_abi_ver = IPATH_UVERBS_ABI_VERSION;
1504         dev->uverbs_cmd_mask =
1505                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
1506                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
1507                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
1508                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
1509                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
1510                 (1ull << IB_USER_VERBS_CMD_CREATE_AH)           |
1511                 (1ull << IB_USER_VERBS_CMD_DESTROY_AH)          |
1512                 (1ull << IB_USER_VERBS_CMD_QUERY_AH)            |
1513                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
1514                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
1515                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1516                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
1517                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
1518                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
1519                 (1ull << IB_USER_VERBS_CMD_POLL_CQ)             |
1520                 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)       |
1521                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
1522                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
1523                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
1524                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
1525                 (1ull << IB_USER_VERBS_CMD_POST_SEND)           |
1526                 (1ull << IB_USER_VERBS_CMD_POST_RECV)           |
1527                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
1528                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
1529                 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
1530                 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
1531                 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
1532                 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
1533                 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
1534         dev->node_type = IB_NODE_CA;
1535         dev->phys_port_cnt = 1;
1536         dev->dma_device = &dd->pcidev->dev;
1537         dev->class_dev.dev = dev->dma_device;
1538         dev->query_device = ipath_query_device;
1539         dev->modify_device = ipath_modify_device;
1540         dev->query_port = ipath_query_port;
1541         dev->modify_port = ipath_modify_port;
1542         dev->query_pkey = ipath_query_pkey;
1543         dev->query_gid = ipath_query_gid;
1544         dev->alloc_ucontext = ipath_alloc_ucontext;
1545         dev->dealloc_ucontext = ipath_dealloc_ucontext;
1546         dev->alloc_pd = ipath_alloc_pd;
1547         dev->dealloc_pd = ipath_dealloc_pd;
1548         dev->create_ah = ipath_create_ah;
1549         dev->destroy_ah = ipath_destroy_ah;
1550         dev->query_ah = ipath_query_ah;
1551         dev->create_srq = ipath_create_srq;
1552         dev->modify_srq = ipath_modify_srq;
1553         dev->query_srq = ipath_query_srq;
1554         dev->destroy_srq = ipath_destroy_srq;
1555         dev->create_qp = ipath_create_qp;
1556         dev->modify_qp = ipath_modify_qp;
1557         dev->query_qp = ipath_query_qp;
1558         dev->destroy_qp = ipath_destroy_qp;
1559         dev->post_send = ipath_post_send;
1560         dev->post_recv = ipath_post_receive;
1561         dev->post_srq_recv = ipath_post_srq_receive;
1562         dev->create_cq = ipath_create_cq;
1563         dev->destroy_cq = ipath_destroy_cq;
1564         dev->resize_cq = ipath_resize_cq;
1565         dev->poll_cq = ipath_poll_cq;
1566         dev->req_notify_cq = ipath_req_notify_cq;
1567         dev->get_dma_mr = ipath_get_dma_mr;
1568         dev->reg_phys_mr = ipath_reg_phys_mr;
1569         dev->reg_user_mr = ipath_reg_user_mr;
1570         dev->dereg_mr = ipath_dereg_mr;
1571         dev->alloc_fmr = ipath_alloc_fmr;
1572         dev->map_phys_fmr = ipath_map_phys_fmr;
1573         dev->unmap_fmr = ipath_unmap_fmr;
1574         dev->dealloc_fmr = ipath_dealloc_fmr;
1575         dev->attach_mcast = ipath_multicast_attach;
1576         dev->detach_mcast = ipath_multicast_detach;
1577         dev->process_mad = ipath_process_mad;
1578         dev->mmap = ipath_mmap;
1579
1580         snprintf(dev->node_desc, sizeof(dev->node_desc),
1581                  IPATH_IDSTR " %s", system_utsname.nodename);
1582
1583         ret = ib_register_device(dev);
1584         if (ret)
1585                 goto err_reg;
1586
1587         if (ipath_verbs_register_sysfs(dev))
1588                 goto err_class;
1589
1590         enable_timer(dd);
1591
1592         goto bail;
1593
1594 err_class:
1595         ib_unregister_device(dev);
1596 err_reg:
1597         kfree(idev->lk_table.table);
1598 err_lk:
1599         kfree(idev->qp_table.table);
1600 err_qp:
1601         ib_dealloc_device(dev);
1602         ipath_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1603         idev = NULL;
1604
1605 bail:
1606         dd->verbs_dev = idev;
1607         return ret;
1608 }
1609
1610 void ipath_unregister_ib_device(struct ipath_ibdev *dev)
1611 {
1612         struct ib_device *ibdev = &dev->ibdev;
1613
1614         disable_timer(dev->dd);
1615
1616         ib_unregister_device(ibdev);
1617
1618         if (!list_empty(&dev->pending[0]) ||
1619             !list_empty(&dev->pending[1]) ||
1620             !list_empty(&dev->pending[2]))
1621                 ipath_dev_err(dev->dd, "pending list not empty!\n");
1622         if (!list_empty(&dev->piowait))
1623                 ipath_dev_err(dev->dd, "piowait list not empty!\n");
1624         if (!list_empty(&dev->rnrwait))
1625                 ipath_dev_err(dev->dd, "rnrwait list not empty!\n");
1626         if (!ipath_mcast_tree_empty())
1627                 ipath_dev_err(dev->dd, "multicast table memory leak!\n");
1628         /*
1629          * Note that ipath_unregister_ib_device() can be called before all
1630          * the QPs are destroyed!
1631          */
1632         ipath_free_all_qps(&dev->qp_table);
1633         kfree(dev->qp_table.table);
1634         kfree(dev->lk_table.table);
1635         ib_dealloc_device(ibdev);
1636 }
1637
1638 static ssize_t show_rev(struct class_device *cdev, char *buf)
1639 {
1640         struct ipath_ibdev *dev =
1641                 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1642
1643         return sprintf(buf, "%x\n", dev->dd->ipath_pcirev);
1644 }
1645
1646 static ssize_t show_hca(struct class_device *cdev, char *buf)
1647 {
1648         struct ipath_ibdev *dev =
1649                 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1650         int ret;
1651
1652         ret = dev->dd->ipath_f_get_boardname(dev->dd, buf, 128);
1653         if (ret < 0)
1654                 goto bail;
1655         strcat(buf, "\n");
1656         ret = strlen(buf);
1657
1658 bail:
1659         return ret;
1660 }
1661
1662 static ssize_t show_stats(struct class_device *cdev, char *buf)
1663 {
1664         struct ipath_ibdev *dev =
1665                 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1666         int i;
1667         int len;
1668
1669         len = sprintf(buf,
1670                       "RC resends  %d\n"
1671                       "RC no QACK  %d\n"
1672                       "RC ACKs     %d\n"
1673                       "RC SEQ NAKs %d\n"
1674                       "RC RDMA seq %d\n"
1675                       "RC RNR NAKs %d\n"
1676                       "RC OTH NAKs %d\n"
1677                       "RC timeouts %d\n"
1678                       "RC RDMA dup %d\n"
1679                       "piobuf wait %d\n"
1680                       "no piobuf   %d\n"
1681                       "PKT drops   %d\n"
1682                       "WQE errs    %d\n",
1683                       dev->n_rc_resends, dev->n_rc_qacks, dev->n_rc_acks,
1684                       dev->n_seq_naks, dev->n_rdma_seq, dev->n_rnr_naks,
1685                       dev->n_other_naks, dev->n_timeouts,
1686                       dev->n_rdma_dup_busy, dev->n_piowait,
1687                       dev->n_no_piobuf, dev->n_pkt_drops, dev->n_wqe_errs);
1688         for (i = 0; i < ARRAY_SIZE(dev->opstats); i++) {
1689                 const struct ipath_opcode_stats *si = &dev->opstats[i];
1690
1691                 if (!si->n_packets && !si->n_bytes)
1692                         continue;
1693                 len += sprintf(buf + len, "%02x %llu/%llu\n", i,
1694                                (unsigned long long) si->n_packets,
1695                                (unsigned long long) si->n_bytes);
1696         }
1697         return len;
1698 }
1699
1700 static CLASS_DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
1701 static CLASS_DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
1702 static CLASS_DEVICE_ATTR(board_id, S_IRUGO, show_hca, NULL);
1703 static CLASS_DEVICE_ATTR(stats, S_IRUGO, show_stats, NULL);
1704
1705 static struct class_device_attribute *ipath_class_attributes[] = {
1706         &class_device_attr_hw_rev,
1707         &class_device_attr_hca_type,
1708         &class_device_attr_board_id,
1709         &class_device_attr_stats
1710 };
1711
1712 static int ipath_verbs_register_sysfs(struct ib_device *dev)
1713 {
1714         int i;
1715         int ret;
1716
1717         for (i = 0; i < ARRAY_SIZE(ipath_class_attributes); ++i)
1718                 if (class_device_create_file(&dev->class_dev,
1719                                              ipath_class_attributes[i])) {
1720                         ret = 1;
1721                         goto bail;
1722                 }
1723
1724         ret = 0;
1725
1726 bail:
1727         return ret;
1728 }