drivers/infiniband/hw/ipath/ipath_ruc.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 "ipath_verbs.h"
  35
  36 /*
  37  * Convert the AETH RNR timeout code into the number of milliseconds.
  38  */
  39 const u32 ib_ipath_rnr_table[32] = {
  40         656,                    /* 0 */
  41         1,                      /* 1 */
  42         1,                      /* 2 */
  43         1,                      /* 3 */
  44         1,                      /* 4 */
  45         1,                      /* 5 */
  46         1,                      /* 6 */
  47         1,                      /* 7 */
  48         1,                      /* 8 */
  49         1,                      /* 9 */
  50         1,                      /* A */
  51         1,                      /* B */
  52         1,                      /* C */
  53         1,                      /* D */
  54         2,                      /* E */
  55         2,                      /* F */
  56         3,                      /* 10 */
  57         4,                      /* 11 */
  58         6,                      /* 12 */
  59         8,                      /* 13 */
  60         11,                     /* 14 */
  61         16,                     /* 15 */
  62         21,                     /* 16 */
  63         31,                     /* 17 */
  64         41,                     /* 18 */
  65         62,                     /* 19 */
  66         82,                     /* 1A */
  67         123,                    /* 1B */
  68         164,                    /* 1C */
  69         246,                    /* 1D */
  70         328,                    /* 1E */
  71         492                     /* 1F */
  72 };
  73
  74 /**
  75  * ipath_insert_rnr_queue - put QP on the RNR timeout list for the device
  76  * @qp: the QP
  77  *
  78  * XXX Use a simple list for now.  We might need a priority
  79  * queue if we have lots of QPs waiting for RNR timeouts
  80  * but that should be rare.
  81  */
  82 void ipath_insert_rnr_queue(struct ipath_qp *qp)
  83 {
  84         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
  85         unsigned long flags;
  86
  87         spin_lock_irqsave(&dev->pending_lock, flags);
  88         if (list_empty(&dev->rnrwait))
  89                 list_add(&qp->timerwait, &dev->rnrwait);
  90         else {
  91                 struct list_head *l = &dev->rnrwait;
  92                 struct ipath_qp *nqp = list_entry(l->next, struct ipath_qp,
  93                                                   timerwait);
  94
  95                 while (qp->s_rnr_timeout >= nqp->s_rnr_timeout) {
  96                         qp->s_rnr_timeout -= nqp->s_rnr_timeout;
  97                         l = l->next;
  98                         if (l->next == &dev->rnrwait)
  99                                 break;
 100                         nqp = list_entry(l->next, struct ipath_qp,
 101                                          timerwait);
 102                 }
 103                 list_add(&qp->timerwait, l);
 104         }
 105         spin_unlock_irqrestore(&dev->pending_lock, flags);
 106 }
 107
 108 /**
 109  * ipath_get_rwqe - copy the next RWQE into the QP's RWQE
 110  * @qp: the QP
 111  * @wr_id_only: update wr_id only, not SGEs
 112  *
 113  * Return 0 if no RWQE is available, otherwise return 1.
 114  *
 115  * Called at interrupt level with the QP r_rq.lock held.
 116  */
 117 int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
 118 {
 119         struct ipath_rq *rq;
 120         struct ipath_srq *srq;
 121         struct ipath_rwqe *wqe;
 122         int ret;
 123
 124         if (!qp->ibqp.srq) {
 125                 rq = &qp->r_rq;
 126                 if (unlikely(rq->tail == rq->head)) {
 127                         ret = 0;
 128                         goto bail;
 129                 }
 130                 wqe = get_rwqe_ptr(rq, rq->tail);
 131                 qp->r_wr_id = wqe->wr_id;
 132                 if (!wr_id_only) {
 133                         qp->r_sge.sge = wqe->sg_list[0];
 134                         qp->r_sge.sg_list = wqe->sg_list + 1;
 135                         qp->r_sge.num_sge = wqe->num_sge;
 136                         qp->r_len = wqe->length;
 137                 }
 138                 if (++rq->tail >= rq->size)
 139                         rq->tail = 0;
 140                 ret = 1;
 141                 goto bail;
 142         }
 143
 144         srq = to_isrq(qp->ibqp.srq);
 145         rq = &srq->rq;
 146         spin_lock(&rq->lock);
 147         if (unlikely(rq->tail == rq->head)) {
 148                 spin_unlock(&rq->lock);
 149                 ret = 0;
 150                 goto bail;
 151         }
 152         wqe = get_rwqe_ptr(rq, rq->tail);
 153         qp->r_wr_id = wqe->wr_id;
 154         if (!wr_id_only) {
 155                 qp->r_sge.sge = wqe->sg_list[0];
 156                 qp->r_sge.sg_list = wqe->sg_list + 1;
 157                 qp->r_sge.num_sge = wqe->num_sge;
 158                 qp->r_len = wqe->length;
 159         }
 160         if (++rq->tail >= rq->size)
 161                 rq->tail = 0;
 162         if (srq->ibsrq.event_handler) {
 163                 struct ib_event ev;
 164                 u32 n;
 165
 166                 if (rq->head < rq->tail)
 167                         n = rq->size + rq->head - rq->tail;
 168                 else
 169                         n = rq->head - rq->tail;
 170                 if (n < srq->limit) {
 171                         srq->limit = 0;
 172                         spin_unlock(&rq->lock);
 173                         ev.device = qp->ibqp.device;
 174                         ev.element.srq = qp->ibqp.srq;
 175                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
 176                         srq->ibsrq.event_handler(&ev,
 177                                                  srq->ibsrq.srq_context);
 178                 } else
 179                         spin_unlock(&rq->lock);
 180         } else
 181                 spin_unlock(&rq->lock);
 182         ret = 1;
 183
 184 bail:
 185         return ret;
 186 }
 187
 188 /**
 189  * ipath_ruc_loopback - handle UC and RC lookback requests
 190  * @sqp: the loopback QP
 191  * @wc: the work completion entry
 192  *
 193  * This is called from ipath_do_uc_send() or ipath_do_rc_send() to
 194  * forward a WQE addressed to the same HCA.
 195  * Note that although we are single threaded due to the tasklet, we still
 196  * have to protect against post_send().  We don't have to worry about
 197  * receive interrupts since this is a connected protocol and all packets
 198  * will pass through here.
 199  */
 200 void ipath_ruc_loopback(struct ipath_qp *sqp, struct ib_wc *wc)
 201 {
 202         struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
 203         struct ipath_qp *qp;
 204         struct ipath_swqe *wqe;
 205         struct ipath_sge *sge;
 206         unsigned long flags;
 207         u64 sdata;
 208
 209         qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
 210         if (!qp) {
 211                 dev->n_pkt_drops++;
 212                 return;
 213         }
 214
 215 again:
 216         spin_lock_irqsave(&sqp->s_lock, flags);
 217
 218         if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_SEND_OK)) {
 219                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 220                 goto done;
 221         }
 222
 223         /* Get the next send request. */
 224         if (sqp->s_last == sqp->s_head) {
 225                 /* Send work queue is empty. */
 226                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 227                 goto done;
 228         }
 229
 230         /*
 231          * We can rely on the entry not changing without the s_lock
 232          * being held until we update s_last.
 233          */
 234         wqe = get_swqe_ptr(sqp, sqp->s_last);
 235         spin_unlock_irqrestore(&sqp->s_lock, flags);
 236
 237         wc->wc_flags = 0;
 238         wc->imm_data = 0;
 239
 240         sqp->s_sge.sge = wqe->sg_list[0];
 241         sqp->s_sge.sg_list = wqe->sg_list + 1;
 242         sqp->s_sge.num_sge = wqe->wr.num_sge;
 243         sqp->s_len = wqe->length;
 244         switch (wqe->wr.opcode) {
 245         case IB_WR_SEND_WITH_IMM:
 246                 wc->wc_flags = IB_WC_WITH_IMM;
 247                 wc->imm_data = wqe->wr.imm_data;
 248                 /* FALLTHROUGH */
 249         case IB_WR_SEND:
 250                 spin_lock_irqsave(&qp->r_rq.lock, flags);
 251                 if (!ipath_get_rwqe(qp, 0)) {
 252                 rnr_nak:
 253                         spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 254                         /* Handle RNR NAK */
 255                         if (qp->ibqp.qp_type == IB_QPT_UC)
 256                                 goto send_comp;
 257                         if (sqp->s_rnr_retry == 0) {
 258                                 wc->status = IB_WC_RNR_RETRY_EXC_ERR;
 259                                 goto err;
 260                         }
 261                         if (sqp->s_rnr_retry_cnt < 7)
 262                                 sqp->s_rnr_retry--;
 263                         dev->n_rnr_naks++;
 264                         sqp->s_rnr_timeout =
 265                                 ib_ipath_rnr_table[sqp->s_min_rnr_timer];
 266                         ipath_insert_rnr_queue(sqp);
 267                         goto done;
 268                 }
 269                 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 270                 break;
 271
 272         case IB_WR_RDMA_WRITE_WITH_IMM:
 273                 wc->wc_flags = IB_WC_WITH_IMM;
 274                 wc->imm_data = wqe->wr.imm_data;
 275                 spin_lock_irqsave(&qp->r_rq.lock, flags);
 276                 if (!ipath_get_rwqe(qp, 1))
 277                         goto rnr_nak;
 278                 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 279                 /* FALLTHROUGH */
 280         case IB_WR_RDMA_WRITE:
 281                 if (wqe->length == 0)
 282                         break;
 283                 if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge, wqe->length,
 284                                             wqe->wr.wr.rdma.remote_addr,
 285                                             wqe->wr.wr.rdma.rkey,
 286                                             IB_ACCESS_REMOTE_WRITE))) {
 287                 acc_err:
 288                         wc->status = IB_WC_REM_ACCESS_ERR;
 289                 err:
 290                         wc->wr_id = wqe->wr.wr_id;
 291                         wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 292                         wc->vendor_err = 0;
 293                         wc->byte_len = 0;
 294                         wc->qp_num = sqp->ibqp.qp_num;
 295                         wc->src_qp = sqp->remote_qpn;
 296                         wc->pkey_index = 0;
 297                         wc->slid = sqp->remote_ah_attr.dlid;
 298                         wc->sl = sqp->remote_ah_attr.sl;
 299                         wc->dlid_path_bits = 0;
 300                         wc->port_num = 0;
 301                         ipath_sqerror_qp(sqp, wc);
 302                         goto done;
 303                 }
 304                 break;
 305
 306         case IB_WR_RDMA_READ:
 307                 if (unlikely(!ipath_rkey_ok(dev, &sqp->s_sge, wqe->length,
 308                                             wqe->wr.wr.rdma.remote_addr,
 309                                             wqe->wr.wr.rdma.rkey,
 310                                             IB_ACCESS_REMOTE_READ)))
 311                         goto acc_err;
 312                 if (unlikely(!(qp->qp_access_flags &
 313                                IB_ACCESS_REMOTE_READ)))
 314                         goto acc_err;
 315                 qp->r_sge.sge = wqe->sg_list[0];
 316                 qp->r_sge.sg_list = wqe->sg_list + 1;
 317                 qp->r_sge.num_sge = wqe->wr.num_sge;
 318                 break;
 319
 320         case IB_WR_ATOMIC_CMP_AND_SWP:
 321         case IB_WR_ATOMIC_FETCH_AND_ADD:
 322                 if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge, sizeof(u64),
 323                                             wqe->wr.wr.rdma.remote_addr,
 324                                             wqe->wr.wr.rdma.rkey,
 325                                             IB_ACCESS_REMOTE_ATOMIC)))
 326                         goto acc_err;
 327                 /* Perform atomic OP and save result. */
 328                 sdata = wqe->wr.wr.atomic.swap;
 329                 spin_lock_irqsave(&dev->pending_lock, flags);
 330                 qp->r_atomic_data = *(u64 *) qp->r_sge.sge.vaddr;
 331                 if (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
 332                         *(u64 *) qp->r_sge.sge.vaddr =
 333                                 qp->r_atomic_data + sdata;
 334                 else if (qp->r_atomic_data == wqe->wr.wr.atomic.compare_add)
 335                         *(u64 *) qp->r_sge.sge.vaddr = sdata;
 336                 spin_unlock_irqrestore(&dev->pending_lock, flags);
 337                 *(u64 *) sqp->s_sge.sge.vaddr = qp->r_atomic_data;
 338                 goto send_comp;
 339
 340         default:
 341                 goto done;
 342         }
 343
 344         sge = &sqp->s_sge.sge;
 345         while (sqp->s_len) {
 346                 u32 len = sqp->s_len;
 347
 348                 if (len > sge->length)
 349                         len = sge->length;
 350                 BUG_ON(len == 0);
 351                 ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
 352                 sge->vaddr += len;
 353                 sge->length -= len;
 354                 sge->sge_length -= len;
 355                 if (sge->sge_length == 0) {
 356                         if (--sqp->s_sge.num_sge)
 357                                 *sge = *sqp->s_sge.sg_list++;
 358                 } else if (sge->length == 0 && sge->mr != NULL) {
 359                         if (++sge->n >= IPATH_SEGSZ) {
 360                                 if (++sge->m >= sge->mr->mapsz)
 361                                         break;
 362                                 sge->n = 0;
 363                         }
 364                         sge->vaddr =
 365                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 366                         sge->length =
 367                                 sge->mr->map[sge->m]->segs[sge->n].length;
 368                 }
 369                 sqp->s_len -= len;
 370         }
 371
 372         if (wqe->wr.opcode == IB_WR_RDMA_WRITE ||
 373             wqe->wr.opcode == IB_WR_RDMA_READ)
 374                 goto send_comp;
 375
 376         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
 377                 wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
 378         else
 379                 wc->opcode = IB_WC_RECV;
 380         wc->wr_id = qp->r_wr_id;
 381         wc->status = IB_WC_SUCCESS;
 382         wc->vendor_err = 0;
 383         wc->byte_len = wqe->length;
 384         wc->qp_num = qp->ibqp.qp_num;
 385         wc->src_qp = qp->remote_qpn;
 386         /* XXX do we know which pkey matched? Only needed for GSI. */
 387         wc->pkey_index = 0;
 388         wc->slid = qp->remote_ah_attr.dlid;
 389         wc->sl = qp->remote_ah_attr.sl;
 390         wc->dlid_path_bits = 0;
 391         /* Signal completion event if the solicited bit is set. */
 392         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), wc,
 393                        wqe->wr.send_flags & IB_SEND_SOLICITED);
 394
 395 send_comp:
 396         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
 397
 398         if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &sqp->s_flags) ||
 399             (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
 400                 wc->wr_id = wqe->wr.wr_id;
 401                 wc->status = IB_WC_SUCCESS;
 402                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 403                 wc->vendor_err = 0;
 404                 wc->byte_len = wqe->length;
 405                 wc->qp_num = sqp->ibqp.qp_num;
 406                 wc->src_qp = 0;
 407                 wc->pkey_index = 0;
 408                 wc->slid = 0;
 409                 wc->sl = 0;
 410                 wc->dlid_path_bits = 0;
 411                 wc->port_num = 0;
 412                 ipath_cq_enter(to_icq(sqp->ibqp.send_cq), wc, 0);
 413         }
 414
 415         /* Update s_last now that we are finished with the SWQE */
 416         spin_lock_irqsave(&sqp->s_lock, flags);
 417         if (++sqp->s_last >= sqp->s_size)
 418                 sqp->s_last = 0;
 419         spin_unlock_irqrestore(&sqp->s_lock, flags);
 420         goto again;
 421
 422 done:
 423         if (atomic_dec_and_test(&qp->refcount))
 424                 wake_up(&qp->wait);
 425 }
 426
 427 /**
 428  * ipath_no_bufs_available - tell the layer driver we need buffers
 429  * @qp: the QP that caused the problem
 430  * @dev: the device we ran out of buffers on
 431  *
 432  * Called when we run out of PIO buffers.
 433  */
 434 void ipath_no_bufs_available(struct ipath_qp *qp, struct ipath_ibdev *dev)
 435 {
 436         unsigned long flags;
 437
 438         spin_lock_irqsave(&dev->pending_lock, flags);
 439         if (list_empty(&qp->piowait))
 440                 list_add_tail(&qp->piowait, &dev->piowait);
 441         spin_unlock_irqrestore(&dev->pending_lock, flags);
 442         /*
 443          * Note that as soon as ipath_layer_want_buffer() is called and
 444          * possibly before it returns, ipath_ib_piobufavail()
 445          * could be called.  If we are still in the tasklet function,
 446          * tasklet_hi_schedule() will not call us until the next time
 447          * tasklet_hi_schedule() is called.
 448          * We clear the tasklet flag now since we are committing to return
 449          * from the tasklet function.
 450          */
 451         clear_bit(IPATH_S_BUSY, &qp->s_flags);
 452         tasklet_unlock(&qp->s_task);
 453         ipath_layer_want_buffer(dev->dd);
 454         dev->n_piowait++;
 455 }
 456
 457 /**
 458  * ipath_post_rc_send - post RC and UC sends
 459  * @qp: the QP to post on
 460  * @wr: the work request to send
 461  */
 462 int ipath_post_rc_send(struct ipath_qp *qp, struct ib_send_wr *wr)
 463 {
 464         struct ipath_swqe *wqe;
 465         unsigned long flags;
 466         u32 next;
 467         int i, j;
 468         int acc;
 469         int ret;
 470
 471         /*
 472          * Don't allow RDMA reads or atomic operations on UC or
 473          * undefined operations.
 474          * Make sure buffer is large enough to hold the result for atomics.
 475          */
 476         if (qp->ibqp.qp_type == IB_QPT_UC) {
 477                 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ) {
 478                         ret = -EINVAL;
 479                         goto bail;
 480                 }
 481         } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) {
 482                 ret = -EINVAL;
 483                 goto bail;
 484         } else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
 485                    (wr->num_sge == 0 ||
 486                     wr->sg_list[0].length < sizeof(u64) ||
 487                     wr->sg_list[0].addr & (sizeof(u64) - 1))) {
 488                 ret = -EINVAL;
 489                 goto bail;
 490         }
 491         /* IB spec says that num_sge == 0 is OK. */
 492         if (wr->num_sge > qp->s_max_sge) {
 493                 ret = -ENOMEM;
 494                 goto bail;
 495         }
 496         spin_lock_irqsave(&qp->s_lock, flags);
 497         next = qp->s_head + 1;
 498         if (next >= qp->s_size)
 499                 next = 0;
 500         if (next == qp->s_last) {
 501                 spin_unlock_irqrestore(&qp->s_lock, flags);
 502                 ret = -EINVAL;
 503                 goto bail;
 504         }
 505
 506         wqe = get_swqe_ptr(qp, qp->s_head);
 507         wqe->wr = *wr;
 508         wqe->ssn = qp->s_ssn++;
 509         wqe->sg_list[0].mr = NULL;
 510         wqe->sg_list[0].vaddr = NULL;
 511         wqe->sg_list[0].length = 0;
 512         wqe->sg_list[0].sge_length = 0;
 513         wqe->length = 0;
 514         acc = wr->opcode >= IB_WR_RDMA_READ ? IB_ACCESS_LOCAL_WRITE : 0;
 515         for (i = 0, j = 0; i < wr->num_sge; i++) {
 516                 if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
 517                         spin_unlock_irqrestore(&qp->s_lock, flags);
 518                         ret = -EINVAL;
 519                         goto bail;
 520                 }
 521                 if (wr->sg_list[i].length == 0)
 522                         continue;
 523                 if (!ipath_lkey_ok(&to_idev(qp->ibqp.device)->lk_table,
 524                                    &wqe->sg_list[j], &wr->sg_list[i],
 525                                    acc)) {
 526                         spin_unlock_irqrestore(&qp->s_lock, flags);
 527                         ret = -EINVAL;
 528                         goto bail;
 529                 }
 530                 wqe->length += wr->sg_list[i].length;
 531                 j++;
 532         }
 533         wqe->wr.num_sge = j;
 534         qp->s_head = next;
 535         spin_unlock_irqrestore(&qp->s_lock, flags);
 536
 537         if (qp->ibqp.qp_type == IB_QPT_UC)
 538                 ipath_do_uc_send((unsigned long) qp);
 539         else
 540                 ipath_do_rc_send((unsigned long) qp);
 541
 542         ret = 0;
 543
 544 bail:
 545         return ret;
 546 }