drivers/infiniband/hw/qib/qib_qp.c

   1 /*
   2  * Copyright (c) 2012, 2013 Intel Corporation.  All rights reserved.
   3  * Copyright (c) 2006 - 2012 QLogic Corporation.  * All rights reserved.
   4  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include <linux/err.h>
  36 #include <linux/vmalloc.h>
  37 #include <rdma/rdma_vt.h>
  38 #ifdef CONFIG_DEBUG_FS
  39 #include <linux/seq_file.h>
  40 #endif
  41
  42 #include "qib.h"
  43
  44 /*
  45  * mask field which was present in now deleted qib_qpn_table
  46  * is not present in rvt_qpn_table. Defining the same field
  47  * as qpt_mask here instead of adding the mask field to
  48  * rvt_qpn_table.
  49  */
  50 u16 qpt_mask;
  51
  52 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt,
  53                               struct rvt_qpn_map *map, unsigned off)
  54 {
  55         return (map - qpt->map) * RVT_BITS_PER_PAGE + off;
  56 }
  57
  58 static inline unsigned find_next_offset(struct rvt_qpn_table *qpt,
  59                                         struct rvt_qpn_map *map, unsigned off,
  60                                         unsigned n)
  61 {
  62         if (qpt_mask) {
  63                 off++;
  64                 if (((off & qpt_mask) >> 1) >= n)
  65                         off = (off | qpt_mask) + 2;
  66         } else {
  67                 off = find_next_zero_bit(map->page, RVT_BITS_PER_PAGE, off);
  68         }
  69         return off;
  70 }
  71
  72 /*
  73  * Convert the AETH credit code into the number of credits.
  74  */
  75 static u32 credit_table[31] = {
  76         0,                      /* 0 */
  77         1,                      /* 1 */
  78         2,                      /* 2 */
  79         3,                      /* 3 */
  80         4,                      /* 4 */
  81         6,                      /* 5 */
  82         8,                      /* 6 */
  83         12,                     /* 7 */
  84         16,                     /* 8 */
  85         24,                     /* 9 */
  86         32,                     /* A */
  87         48,                     /* B */
  88         64,                     /* C */
  89         96,                     /* D */
  90         128,                    /* E */
  91         192,                    /* F */
  92         256,                    /* 10 */
  93         384,                    /* 11 */
  94         512,                    /* 12 */
  95         768,                    /* 13 */
  96         1024,                   /* 14 */
  97         1536,                   /* 15 */
  98         2048,                   /* 16 */
  99         3072,                   /* 17 */
 100         4096,                   /* 18 */
 101         6144,                   /* 19 */
 102         8192,                   /* 1A */
 103         12288,                  /* 1B */
 104         16384,                  /* 1C */
 105         24576,                  /* 1D */
 106         32768                   /* 1E */
 107 };
 108
 109 static void get_map_page(struct rvt_qpn_table *qpt, struct rvt_qpn_map *map,
 110                          gfp_t gfp)
 111 {
 112         unsigned long page = get_zeroed_page(gfp);
 113
 114         /*
 115          * Free the page if someone raced with us installing it.
 116          */
 117
 118         spin_lock(&qpt->lock);
 119         if (map->page)
 120                 free_page(page);
 121         else
 122                 map->page = (void *)page;
 123         spin_unlock(&qpt->lock);
 124 }
 125
 126 /*
 127  * Allocate the next available QPN or
 128  * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI.
 129  */
 130 int qib_alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt,
 131                   enum ib_qp_type type, u8 port, gfp_t gfp)
 132 {
 133         u32 i, offset, max_scan, qpn;
 134         struct rvt_qpn_map *map;
 135         u32 ret;
 136         struct qib_ibdev *verbs_dev = container_of(rdi, struct qib_ibdev, rdi);
 137         struct qib_devdata *dd = container_of(verbs_dev, struct qib_devdata,
 138                                               verbs_dev);
 139
 140         if (type == IB_QPT_SMI || type == IB_QPT_GSI) {
 141                 unsigned n;
 142
 143                 ret = type == IB_QPT_GSI;
 144                 n = 1 << (ret + 2 * (port - 1));
 145                 spin_lock(&qpt->lock);
 146                 if (qpt->flags & n)
 147                         ret = -EINVAL;
 148                 else
 149                         qpt->flags |= n;
 150                 spin_unlock(&qpt->lock);
 151                 goto bail;
 152         }
 153
 154         qpn = qpt->last + 2;
 155         if (qpn >= RVT_QPN_MAX)
 156                 qpn = 2;
 157         if (qpt_mask && ((qpn & qpt_mask) >> 1) >= dd->n_krcv_queues)
 158                 qpn = (qpn | qpt_mask) + 2;
 159         offset = qpn & RVT_BITS_PER_PAGE_MASK;
 160         map = &qpt->map[qpn / RVT_BITS_PER_PAGE];
 161         max_scan = qpt->nmaps - !offset;
 162         for (i = 0;;) {
 163                 if (unlikely(!map->page)) {
 164                         get_map_page(qpt, map, gfp);
 165                         if (unlikely(!map->page))
 166                                 break;
 167                 }
 168                 do {
 169                         if (!test_and_set_bit(offset, map->page)) {
 170                                 qpt->last = qpn;
 171                                 ret = qpn;
 172                                 goto bail;
 173                         }
 174                         offset = find_next_offset(qpt, map, offset,
 175                                 dd->n_krcv_queues);
 176                         qpn = mk_qpn(qpt, map, offset);
 177                         /*
 178                          * This test differs from alloc_pidmap().
 179                          * If find_next_offset() does find a zero
 180                          * bit, we don't need to check for QPN
 181                          * wrapping around past our starting QPN.
 182                          * We just need to be sure we don't loop
 183                          * forever.
 184                          */
 185                 } while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX);
 186                 /*
 187                  * In order to keep the number of pages allocated to a
 188                  * minimum, we scan the all existing pages before increasing
 189                  * the size of the bitmap table.
 190                  */
 191                 if (++i > max_scan) {
 192                         if (qpt->nmaps == RVT_QPNMAP_ENTRIES)
 193                                 break;
 194                         map = &qpt->map[qpt->nmaps++];
 195                         offset = 0;
 196                 } else if (map < &qpt->map[qpt->nmaps]) {
 197                         ++map;
 198                         offset = 0;
 199                 } else {
 200                         map = &qpt->map[0];
 201                         offset = 2;
 202                 }
 203                 qpn = mk_qpn(qpt, map, offset);
 204         }
 205
 206         ret = -ENOMEM;
 207
 208 bail:
 209         return ret;
 210 }
 211
 212 /**
 213  * qib_free_all_qps - check for QPs still in use
 214  */
 215 unsigned qib_free_all_qps(struct rvt_dev_info *rdi)
 216 {
 217         struct qib_ibdev *verbs_dev = container_of(rdi, struct qib_ibdev, rdi);
 218         struct qib_devdata *dd = container_of(verbs_dev, struct qib_devdata,
 219                                               verbs_dev);
 220         unsigned n, qp_inuse = 0;
 221
 222         for (n = 0; n < dd->num_pports; n++) {
 223                 struct qib_ibport *ibp = &dd->pport[n].ibport_data;
 224
 225                 rcu_read_lock();
 226                 if (rcu_dereference(ibp->rvp.qp[0]))
 227                         qp_inuse++;
 228                 if (rcu_dereference(ibp->rvp.qp[1]))
 229                         qp_inuse++;
 230                 rcu_read_unlock();
 231         }
 232         return qp_inuse;
 233 }
 234
 235 void qib_notify_qp_reset(struct rvt_qp *qp)
 236 {
 237         struct qib_qp_priv *priv = qp->priv;
 238
 239         atomic_set(&priv->s_dma_busy, 0);
 240 }
 241
 242 void qib_notify_error_qp(struct rvt_qp *qp)
 243 {
 244         struct qib_qp_priv *priv = qp->priv;
 245         struct qib_ibdev *dev = to_idev(qp->ibqp.device);
 246
 247         spin_lock(&dev->rdi.pending_lock);
 248         if (!list_empty(&priv->iowait) && !(qp->s_flags & RVT_S_BUSY)) {
 249                 qp->s_flags &= ~RVT_S_ANY_WAIT_IO;
 250                 list_del_init(&priv->iowait);
 251         }
 252         spin_unlock(&dev->rdi.pending_lock);
 253
 254         if (!(qp->s_flags & RVT_S_BUSY)) {
 255                 qp->s_hdrwords = 0;
 256                 if (qp->s_rdma_mr) {
 257                         rvt_put_mr(qp->s_rdma_mr);
 258                         qp->s_rdma_mr = NULL;
 259                 }
 260                 if (priv->s_tx) {
 261                         qib_put_txreq(priv->s_tx);
 262                         priv->s_tx = NULL;
 263                 }
 264         }
 265 }
 266
 267 static int mtu_to_enum(u32 mtu)
 268 {
 269         int enum_mtu;
 270
 271         switch (mtu) {
 272         case 4096:
 273                 enum_mtu = IB_MTU_4096;
 274                 break;
 275         case 2048:
 276                 enum_mtu = IB_MTU_2048;
 277                 break;
 278         case 1024:
 279                 enum_mtu = IB_MTU_1024;
 280                 break;
 281         case 512:
 282                 enum_mtu = IB_MTU_512;
 283                 break;
 284         case 256:
 285                 enum_mtu = IB_MTU_256;
 286                 break;
 287         default:
 288                 enum_mtu = IB_MTU_2048;
 289         }
 290         return enum_mtu;
 291 }
 292
 293 int qib_get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
 294                            struct ib_qp_attr *attr)
 295 {
 296         int mtu, pmtu, pidx = qp->port_num - 1;
 297         struct qib_ibdev *verbs_dev = container_of(rdi, struct qib_ibdev, rdi);
 298         struct qib_devdata *dd = container_of(verbs_dev, struct qib_devdata,
 299                                               verbs_dev);
 300         mtu = ib_mtu_enum_to_int(attr->path_mtu);
 301         if (mtu == -1)
 302                 return -EINVAL;
 303
 304         if (mtu > dd->pport[pidx].ibmtu)
 305                 pmtu = mtu_to_enum(dd->pport[pidx].ibmtu);
 306         else
 307                 pmtu = attr->path_mtu;
 308         return pmtu;
 309 }
 310
 311 int qib_mtu_to_path_mtu(u32 mtu)
 312 {
 313         return mtu_to_enum(mtu);
 314 }
 315
 316 u32 qib_mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
 317 {
 318         return ib_mtu_enum_to_int(pmtu);
 319 }
 320
 321 /**
 322  * qib_compute_aeth - compute the AETH (syndrome + MSN)
 323  * @qp: the queue pair to compute the AETH for
 324  *
 325  * Returns the AETH.
 326  */
 327 __be32 qib_compute_aeth(struct rvt_qp *qp)
 328 {
 329         u32 aeth = qp->r_msn & QIB_MSN_MASK;
 330
 331         if (qp->ibqp.srq) {
 332                 /*
 333                  * Shared receive queues don't generate credits.
 334                  * Set the credit field to the invalid value.
 335                  */
 336                 aeth |= QIB_AETH_CREDIT_INVAL << QIB_AETH_CREDIT_SHIFT;
 337         } else {
 338                 u32 min, max, x;
 339                 u32 credits;
 340                 struct rvt_rwq *wq = qp->r_rq.wq;
 341                 u32 head;
 342                 u32 tail;
 343
 344                 /* sanity check pointers before trusting them */
 345                 head = wq->head;
 346                 if (head >= qp->r_rq.size)
 347                         head = 0;
 348                 tail = wq->tail;
 349                 if (tail >= qp->r_rq.size)
 350                         tail = 0;
 351                 /*
 352                  * Compute the number of credits available (RWQEs).
 353                  * XXX Not holding the r_rq.lock here so there is a small
 354                  * chance that the pair of reads are not atomic.
 355                  */
 356                 credits = head - tail;
 357                 if ((int)credits < 0)
 358                         credits += qp->r_rq.size;
 359                 /*
 360                  * Binary search the credit table to find the code to
 361                  * use.
 362                  */
 363                 min = 0;
 364                 max = 31;
 365                 for (;;) {
 366                         x = (min + max) / 2;
 367                         if (credit_table[x] == credits)
 368                                 break;
 369                         if (credit_table[x] > credits)
 370                                 max = x;
 371                         else if (min == x)
 372                                 break;
 373                         else
 374                                 min = x;
 375                 }
 376                 aeth |= x << QIB_AETH_CREDIT_SHIFT;
 377         }
 378         return cpu_to_be32(aeth);
 379 }
 380
 381 void *qib_qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp, gfp_t gfp)
 382 {
 383         struct qib_qp_priv *priv;
 384
 385         priv = kzalloc(sizeof(*priv), gfp);
 386         if (!priv)
 387                 return ERR_PTR(-ENOMEM);
 388         priv->owner = qp;
 389
 390         priv->s_hdr = kzalloc(sizeof(*priv->s_hdr), gfp);
 391         if (!priv->s_hdr) {
 392                 kfree(priv);
 393                 return ERR_PTR(-ENOMEM);
 394         }
 395         init_waitqueue_head(&priv->wait_dma);
 396         INIT_WORK(&priv->s_work, _qib_do_send);
 397         INIT_LIST_HEAD(&priv->iowait);
 398
 399         return priv;
 400 }
 401
 402 void qib_qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
 403 {
 404         struct qib_qp_priv *priv = qp->priv;
 405
 406         kfree(priv->s_hdr);
 407         kfree(priv);
 408 }
 409
 410 void qib_stop_send_queue(struct rvt_qp *qp)
 411 {
 412         struct qib_qp_priv *priv = qp->priv;
 413
 414         cancel_work_sync(&priv->s_work);
 415         del_timer_sync(&qp->s_timer);
 416 }
 417
 418 void qib_quiesce_qp(struct rvt_qp *qp)
 419 {
 420         struct qib_qp_priv *priv = qp->priv;
 421
 422         wait_event(priv->wait_dma, !atomic_read(&priv->s_dma_busy));
 423         if (priv->s_tx) {
 424                 qib_put_txreq(priv->s_tx);
 425                 priv->s_tx = NULL;
 426         }
 427 }
 428
 429 void qib_flush_qp_waiters(struct rvt_qp *qp)
 430 {
 431         struct qib_qp_priv *priv = qp->priv;
 432         struct qib_ibdev *dev = to_idev(qp->ibqp.device);
 433
 434         spin_lock(&dev->rdi.pending_lock);
 435         if (!list_empty(&priv->iowait))
 436                 list_del_init(&priv->iowait);
 437         spin_unlock(&dev->rdi.pending_lock);
 438 }
 439
 440 /**
 441  * qib_get_credit - flush the send work queue of a QP
 442  * @qp: the qp who's send work queue to flush
 443  * @aeth: the Acknowledge Extended Transport Header
 444  *
 445  * The QP s_lock should be held.
 446  */
 447 void qib_get_credit(struct rvt_qp *qp, u32 aeth)
 448 {
 449         u32 credit = (aeth >> QIB_AETH_CREDIT_SHIFT) & QIB_AETH_CREDIT_MASK;
 450
 451         /*
 452          * If the credit is invalid, we can send
 453          * as many packets as we like.  Otherwise, we have to
 454          * honor the credit field.
 455          */
 456         if (credit == QIB_AETH_CREDIT_INVAL) {
 457                 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) {
 458                         qp->s_flags |= RVT_S_UNLIMITED_CREDIT;
 459                         if (qp->s_flags & RVT_S_WAIT_SSN_CREDIT) {
 460                                 qp->s_flags &= ~RVT_S_WAIT_SSN_CREDIT;
 461                                 qib_schedule_send(qp);
 462                         }
 463                 }
 464         } else if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) {
 465                 /* Compute new LSN (i.e., MSN + credit) */
 466                 credit = (aeth + credit_table[credit]) & QIB_MSN_MASK;
 467                 if (qib_cmp24(credit, qp->s_lsn) > 0) {
 468                         qp->s_lsn = credit;
 469                         if (qp->s_flags & RVT_S_WAIT_SSN_CREDIT) {
 470                                 qp->s_flags &= ~RVT_S_WAIT_SSN_CREDIT;
 471                                 qib_schedule_send(qp);
 472                         }
 473                 }
 474         }
 475 }
 476
 477 /**
 478  * qib_check_send_wqe - validate wr/wqe
 479  * @qp - The qp
 480  * @wqe - The built wqe
 481  *
 482  * validate wr/wqe.  This is called
 483  * prior to inserting the wqe into
 484  * the ring but after the wqe has been
 485  * setup.
 486  *
 487  * Returns 1 to force direct progress, 0 otherwise, -EINVAL on failure
 488  */
 489 int qib_check_send_wqe(struct rvt_qp *qp,
 490                        struct rvt_swqe *wqe)
 491 {
 492         struct rvt_ah *ah;
 493         int ret = 0;
 494
 495         switch (qp->ibqp.qp_type) {
 496         case IB_QPT_RC:
 497         case IB_QPT_UC:
 498                 if (wqe->length > 0x80000000U)
 499                         return -EINVAL;
 500                 break;
 501         case IB_QPT_SMI:
 502         case IB_QPT_GSI:
 503         case IB_QPT_UD:
 504                 ah = ibah_to_rvtah(wqe->ud_wr.ah);
 505                 if (wqe->length > (1 << ah->log_pmtu))
 506                         return -EINVAL;
 507                 /* progress hint */
 508                 ret = 1;
 509                 break;
 510         default:
 511                 break;
 512         }
 513         return ret;
 514 }
 515
 516 #ifdef CONFIG_DEBUG_FS
 517
 518 struct qib_qp_iter {
 519         struct qib_ibdev *dev;
 520         struct rvt_qp *qp;
 521         int n;
 522 };
 523
 524 struct qib_qp_iter *qib_qp_iter_init(struct qib_ibdev *dev)
 525 {
 526         struct qib_qp_iter *iter;
 527
 528         iter = kzalloc(sizeof(*iter), GFP_KERNEL);
 529         if (!iter)
 530                 return NULL;
 531
 532         iter->dev = dev;
 533         if (qib_qp_iter_next(iter)) {
 534                 kfree(iter);
 535                 return NULL;
 536         }
 537
 538         return iter;
 539 }
 540
 541 int qib_qp_iter_next(struct qib_qp_iter *iter)
 542 {
 543         struct qib_ibdev *dev = iter->dev;
 544         int n = iter->n;
 545         int ret = 1;
 546         struct rvt_qp *pqp = iter->qp;
 547         struct rvt_qp *qp;
 548
 549         for (; n < dev->rdi.qp_dev->qp_table_size; n++) {
 550                 if (pqp)
 551                         qp = rcu_dereference(pqp->next);
 552                 else
 553                         qp = rcu_dereference(dev->rdi.qp_dev->qp_table[n]);
 554                 pqp = qp;
 555                 if (qp) {
 556                         iter->qp = qp;
 557                         iter->n = n;
 558                         return 0;
 559                 }
 560         }
 561         return ret;
 562 }
 563
 564 static const char * const qp_type_str[] = {
 565         "SMI", "GSI", "RC", "UC", "UD",
 566 };
 567
 568 void qib_qp_iter_print(struct seq_file *s, struct qib_qp_iter *iter)
 569 {
 570         struct rvt_swqe *wqe;
 571         struct rvt_qp *qp = iter->qp;
 572         struct qib_qp_priv *priv = qp->priv;
 573
 574         wqe = rvt_get_swqe_ptr(qp, qp->s_last);
 575         seq_printf(s,
 576                    "N %d QP%u %s %u %u %u f=%x %u %u %u %u %u PSN %x %x %x %x %x (%u %u %u %u %u %u) QP%u LID %x\n",
 577                    iter->n,
 578                    qp->ibqp.qp_num,
 579                    qp_type_str[qp->ibqp.qp_type],
 580                    qp->state,
 581                    wqe->wr.opcode,
 582                    qp->s_hdrwords,
 583                    qp->s_flags,
 584                    atomic_read(&priv->s_dma_busy),
 585                    !list_empty(&priv->iowait),
 586                    qp->timeout,
 587                    wqe->ssn,
 588                    qp->s_lsn,
 589                    qp->s_last_psn,
 590                    qp->s_psn, qp->s_next_psn,
 591                    qp->s_sending_psn, qp->s_sending_hpsn,
 592                    qp->s_last, qp->s_acked, qp->s_cur,
 593                    qp->s_tail, qp->s_head, qp->s_size,
 594                    qp->remote_qpn,
 595                    qp->remote_ah_attr.dlid);
 596 }
 597
 598 #endif