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