drivers/infiniband/hw/qib/qib_qp.c

   1 /*
   2  * Copyright (c) 2012 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 <linux/jhash.h>
  38
  39 #include "qib.h"
  40
  41 #define BITS_PER_PAGE           (PAGE_SIZE*BITS_PER_BYTE)
  42 #define BITS_PER_PAGE_MASK      (BITS_PER_PAGE-1)
  43
  44 static inline unsigned mk_qpn(struct qib_qpn_table *qpt,
  45                               struct qpn_map *map, unsigned off)
  46 {
  47         return (map - qpt->map) * BITS_PER_PAGE + off;
  48 }
  49
  50 static inline unsigned find_next_offset(struct qib_qpn_table *qpt,
  51                                         struct qpn_map *map, unsigned off,
  52                                         unsigned n)
  53 {
  54         if (qpt->mask) {
  55                 off++;
  56                 if (((off & qpt->mask) >> 1) >= n)
  57                         off = (off | qpt->mask) + 2;
  58         } else
  59                 off = find_next_zero_bit(map->page, BITS_PER_PAGE, off);
  60         return off;
  61 }
  62
  63 /*
  64  * Convert the AETH credit code into the number of credits.
  65  */
  66 static u32 credit_table[31] = {
  67         0,                      /* 0 */
  68         1,                      /* 1 */
  69         2,                      /* 2 */
  70         3,                      /* 3 */
  71         4,                      /* 4 */
  72         6,                      /* 5 */
  73         8,                      /* 6 */
  74         12,                     /* 7 */
  75         16,                     /* 8 */
  76         24,                     /* 9 */
  77         32,                     /* A */
  78         48,                     /* B */
  79         64,                     /* C */
  80         96,                     /* D */
  81         128,                    /* E */
  82         192,                    /* F */
  83         256,                    /* 10 */
  84         384,                    /* 11 */
  85         512,                    /* 12 */
  86         768,                    /* 13 */
  87         1024,                   /* 14 */
  88         1536,                   /* 15 */
  89         2048,                   /* 16 */
  90         3072,                   /* 17 */
  91         4096,                   /* 18 */
  92         6144,                   /* 19 */
  93         8192,                   /* 1A */
  94         12288,                  /* 1B */
  95         16384,                  /* 1C */
  96         24576,                  /* 1D */
  97         32768                   /* 1E */
  98 };
  99
 100 static void get_map_page(struct qib_qpn_table *qpt, struct qpn_map *map)
 101 {
 102         unsigned long page = get_zeroed_page(GFP_KERNEL);
 103
 104         /*
 105          * Free the page if someone raced with us installing it.
 106          */
 107
 108         spin_lock(&qpt->lock);
 109         if (map->page)
 110                 free_page(page);
 111         else
 112                 map->page = (void *)page;
 113         spin_unlock(&qpt->lock);
 114 }
 115
 116 /*
 117  * Allocate the next available QPN or
 118  * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI.
 119  */
 120 static int alloc_qpn(struct qib_devdata *dd, struct qib_qpn_table *qpt,
 121                      enum ib_qp_type type, u8 port)
 122 {
 123         u32 i, offset, max_scan, qpn;
 124         struct qpn_map *map;
 125         u32 ret;
 126
 127         if (type == IB_QPT_SMI || type == IB_QPT_GSI) {
 128                 unsigned n;
 129
 130                 ret = type == IB_QPT_GSI;
 131                 n = 1 << (ret + 2 * (port - 1));
 132                 spin_lock(&qpt->lock);
 133                 if (qpt->flags & n)
 134                         ret = -EINVAL;
 135                 else
 136                         qpt->flags |= n;
 137                 spin_unlock(&qpt->lock);
 138                 goto bail;
 139         }
 140
 141         qpn = qpt->last + 2;
 142         if (qpn >= QPN_MAX)
 143                 qpn = 2;
 144         if (qpt->mask && ((qpn & qpt->mask) >> 1) >= dd->n_krcv_queues)
 145                 qpn = (qpn | qpt->mask) + 2;
 146         offset = qpn & BITS_PER_PAGE_MASK;
 147         map = &qpt->map[qpn / BITS_PER_PAGE];
 148         max_scan = qpt->nmaps - !offset;
 149         for (i = 0;;) {
 150                 if (unlikely(!map->page)) {
 151                         get_map_page(qpt, map);
 152                         if (unlikely(!map->page))
 153                                 break;
 154                 }
 155                 do {
 156                         if (!test_and_set_bit(offset, map->page)) {
 157                                 qpt->last = qpn;
 158                                 ret = qpn;
 159                                 goto bail;
 160                         }
 161                         offset = find_next_offset(qpt, map, offset,
 162                                 dd->n_krcv_queues);
 163                         qpn = mk_qpn(qpt, map, offset);
 164                         /*
 165                          * This test differs from alloc_pidmap().
 166                          * If find_next_offset() does find a zero
 167                          * bit, we don't need to check for QPN
 168                          * wrapping around past our starting QPN.
 169                          * We just need to be sure we don't loop
 170                          * forever.
 171                          */
 172                 } while (offset < BITS_PER_PAGE && qpn < QPN_MAX);
 173                 /*
 174                  * In order to keep the number of pages allocated to a
 175                  * minimum, we scan the all existing pages before increasing
 176                  * the size of the bitmap table.
 177                  */
 178                 if (++i > max_scan) {
 179                         if (qpt->nmaps == QPNMAP_ENTRIES)
 180                                 break;
 181                         map = &qpt->map[qpt->nmaps++];
 182                         offset = 0;
 183                 } else if (map < &qpt->map[qpt->nmaps]) {
 184                         ++map;
 185                         offset = 0;
 186                 } else {
 187                         map = &qpt->map[0];
 188                         offset = 2;
 189                 }
 190                 qpn = mk_qpn(qpt, map, offset);
 191         }
 192
 193         ret = -ENOMEM;
 194
 195 bail:
 196         return ret;
 197 }
 198
 199 static void free_qpn(struct qib_qpn_table *qpt, u32 qpn)
 200 {
 201         struct qpn_map *map;
 202
 203         map = qpt->map + qpn / BITS_PER_PAGE;
 204         if (map->page)
 205                 clear_bit(qpn & BITS_PER_PAGE_MASK, map->page);
 206 }
 207
 208 static inline unsigned qpn_hash(struct qib_ibdev *dev, u32 qpn)
 209 {
 210         return jhash_1word(qpn, dev->qp_rnd) &
 211                 (dev->qp_table_size - 1);
 212 }
 213
 214
 215 /*
 216  * Put the QP into the hash table.
 217  * The hash table holds a reference to the QP.
 218  */
 219 static void insert_qp(struct qib_ibdev *dev, struct qib_qp *qp)
 220 {
 221         struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 222         unsigned long flags;
 223         unsigned n = qpn_hash(dev, qp->ibqp.qp_num);
 224
 225         spin_lock_irqsave(&dev->qpt_lock, flags);
 226         atomic_inc(&qp->refcount);
 227
 228         if (qp->ibqp.qp_num == 0)
 229                 rcu_assign_pointer(ibp->qp0, qp);
 230         else if (qp->ibqp.qp_num == 1)
 231                 rcu_assign_pointer(ibp->qp1, qp);
 232         else {
 233                 qp->next = dev->qp_table[n];
 234                 rcu_assign_pointer(dev->qp_table[n], qp);
 235         }
 236
 237         spin_unlock_irqrestore(&dev->qpt_lock, flags);
 238         synchronize_rcu();
 239 }
 240
 241 /*
 242  * Remove the QP from the table so it can't be found asynchronously by
 243  * the receive interrupt routine.
 244  */
 245 static void remove_qp(struct qib_ibdev *dev, struct qib_qp *qp)
 246 {
 247         struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 248         unsigned n = qpn_hash(dev, qp->ibqp.qp_num);
 249         unsigned long flags;
 250
 251         spin_lock_irqsave(&dev->qpt_lock, flags);
 252
 253         if (rcu_dereference_protected(ibp->qp0,
 254                         lockdep_is_held(&dev->qpt_lock)) == qp) {
 255                 atomic_dec(&qp->refcount);
 256                 rcu_assign_pointer(ibp->qp0, NULL);
 257         } else if (rcu_dereference_protected(ibp->qp1,
 258                         lockdep_is_held(&dev->qpt_lock)) == qp) {
 259                 atomic_dec(&qp->refcount);
 260                 rcu_assign_pointer(ibp->qp1, NULL);
 261         } else {
 262                 struct qib_qp *q;
 263                 struct qib_qp __rcu **qpp;
 264
 265                 qpp = &dev->qp_table[n];
 266                 for (; (q = rcu_dereference_protected(*qpp,
 267                                 lockdep_is_held(&dev->qpt_lock))) != NULL;
 268                                 qpp = &q->next)
 269                         if (q == qp) {
 270                                 atomic_dec(&qp->refcount);
 271                                 *qpp = qp->next;
 272                                 rcu_assign_pointer(qp->next, NULL);
 273                                 break;
 274                         }
 275         }
 276
 277         spin_unlock_irqrestore(&dev->qpt_lock, flags);
 278         synchronize_rcu();
 279 }
 280
 281 /**
 282  * qib_free_all_qps - check for QPs still in use
 283  * @qpt: the QP table to empty
 284  *
 285  * There should not be any QPs still in use.
 286  * Free memory for table.
 287  */
 288 unsigned qib_free_all_qps(struct qib_devdata *dd)
 289 {
 290         struct qib_ibdev *dev = &dd->verbs_dev;
 291         unsigned long flags;
 292         struct qib_qp *qp;
 293         unsigned n, qp_inuse = 0;
 294
 295         for (n = 0; n < dd->num_pports; n++) {
 296                 struct qib_ibport *ibp = &dd->pport[n].ibport_data;
 297
 298                 if (!qib_mcast_tree_empty(ibp))
 299                         qp_inuse++;
 300                 rcu_read_lock();
 301                 if (rcu_dereference(ibp->qp0))
 302                         qp_inuse++;
 303                 if (rcu_dereference(ibp->qp1))
 304                         qp_inuse++;
 305                 rcu_read_unlock();
 306         }
 307
 308         spin_lock_irqsave(&dev->qpt_lock, flags);
 309         for (n = 0; n < dev->qp_table_size; n++) {
 310                 qp = rcu_dereference_protected(dev->qp_table[n],
 311                         lockdep_is_held(&dev->qpt_lock));
 312                 rcu_assign_pointer(dev->qp_table[n], NULL);
 313
 314                 for (; qp; qp = rcu_dereference_protected(qp->next,
 315                                         lockdep_is_held(&dev->qpt_lock)))
 316                         qp_inuse++;
 317         }
 318         spin_unlock_irqrestore(&dev->qpt_lock, flags);
 319         synchronize_rcu();
 320
 321         return qp_inuse;
 322 }
 323
 324 /**
 325  * qib_lookup_qpn - return the QP with the given QPN
 326  * @qpt: the QP table
 327  * @qpn: the QP number to look up
 328  *
 329  * The caller is responsible for decrementing the QP reference count
 330  * when done.
 331  */
 332 struct qib_qp *qib_lookup_qpn(struct qib_ibport *ibp, u32 qpn)
 333 {
 334         struct qib_qp *qp = NULL;
 335
 336         if (unlikely(qpn <= 1)) {
 337                 rcu_read_lock();
 338                 if (qpn == 0)
 339                         qp = rcu_dereference(ibp->qp0);
 340                 else
 341                         qp = rcu_dereference(ibp->qp1);
 342         } else {
 343                 struct qib_ibdev *dev = &ppd_from_ibp(ibp)->dd->verbs_dev;
 344                 unsigned n = qpn_hash(dev, qpn);
 345
 346                 rcu_read_lock();
 347                 for (qp = rcu_dereference(dev->qp_table[n]); qp;
 348                         qp = rcu_dereference(qp->next))
 349                         if (qp->ibqp.qp_num == qpn)
 350                                 break;
 351         }
 352         if (qp)
 353                 if (unlikely(!atomic_inc_not_zero(&qp->refcount)))
 354                         qp = NULL;
 355
 356         rcu_read_unlock();
 357         return qp;
 358 }
 359
 360 /**
 361  * qib_reset_qp - initialize the QP state to the reset state
 362  * @qp: the QP to reset
 363  * @type: the QP type
 364  */
 365 static void qib_reset_qp(struct qib_qp *qp, enum ib_qp_type type)
 366 {
 367         qp->remote_qpn = 0;
 368         qp->qkey = 0;
 369         qp->qp_access_flags = 0;
 370         atomic_set(&qp->s_dma_busy, 0);
 371         qp->s_flags &= QIB_S_SIGNAL_REQ_WR;
 372         qp->s_hdrwords = 0;
 373         qp->s_wqe = NULL;
 374         qp->s_draining = 0;
 375         qp->s_next_psn = 0;
 376         qp->s_last_psn = 0;
 377         qp->s_sending_psn = 0;
 378         qp->s_sending_hpsn = 0;
 379         qp->s_psn = 0;
 380         qp->r_psn = 0;
 381         qp->r_msn = 0;
 382         if (type == IB_QPT_RC) {
 383                 qp->s_state = IB_OPCODE_RC_SEND_LAST;
 384                 qp->r_state = IB_OPCODE_RC_SEND_LAST;
 385         } else {
 386                 qp->s_state = IB_OPCODE_UC_SEND_LAST;
 387                 qp->r_state = IB_OPCODE_UC_SEND_LAST;
 388         }
 389         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 390         qp->r_nak_state = 0;
 391         qp->r_aflags = 0;
 392         qp->r_flags = 0;
 393         qp->s_head = 0;
 394         qp->s_tail = 0;
 395         qp->s_cur = 0;
 396         qp->s_acked = 0;
 397         qp->s_last = 0;
 398         qp->s_ssn = 1;
 399         qp->s_lsn = 0;
 400         qp->s_mig_state = IB_MIG_MIGRATED;
 401         memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue));
 402         qp->r_head_ack_queue = 0;
 403         qp->s_tail_ack_queue = 0;
 404         qp->s_num_rd_atomic = 0;
 405         if (qp->r_rq.wq) {
 406                 qp->r_rq.wq->head = 0;
 407                 qp->r_rq.wq->tail = 0;
 408         }
 409         qp->r_sge.num_sge = 0;
 410 }
 411
 412 static void clear_mr_refs(struct qib_qp *qp, int clr_sends)
 413 {
 414         unsigned n;
 415
 416         if (test_and_clear_bit(QIB_R_REWIND_SGE, &qp->r_aflags))
 417                 qib_put_ss(&qp->s_rdma_read_sge);
 418
 419         qib_put_ss(&qp->r_sge);
 420
 421         if (clr_sends) {
 422                 while (qp->s_last != qp->s_head) {
 423                         struct qib_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 424                         unsigned i;
 425
 426                         for (i = 0; i < wqe->wr.num_sge; i++) {
 427                                 struct qib_sge *sge = &wqe->sg_list[i];
 428
 429                                 qib_put_mr(sge->mr);
 430                         }
 431                         if (qp->ibqp.qp_type == IB_QPT_UD ||
 432                             qp->ibqp.qp_type == IB_QPT_SMI ||
 433                             qp->ibqp.qp_type == IB_QPT_GSI)
 434                                 atomic_dec(&to_iah(wqe->wr.wr.ud.ah)->refcount);
 435                         if (++qp->s_last >= qp->s_size)
 436                                 qp->s_last = 0;
 437                 }
 438                 if (qp->s_rdma_mr) {
 439                         qib_put_mr(qp->s_rdma_mr);
 440                         qp->s_rdma_mr = NULL;
 441                 }
 442         }
 443
 444         if (qp->ibqp.qp_type != IB_QPT_RC)
 445                 return;
 446
 447         for (n = 0; n < ARRAY_SIZE(qp->s_ack_queue); n++) {
 448                 struct qib_ack_entry *e = &qp->s_ack_queue[n];
 449
 450                 if (e->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST &&
 451                     e->rdma_sge.mr) {
 452                         qib_put_mr(e->rdma_sge.mr);
 453                         e->rdma_sge.mr = NULL;
 454                 }
 455         }
 456 }
 457
 458 /**
 459  * qib_error_qp - put a QP into the error state
 460  * @qp: the QP to put into the error state
 461  * @err: the receive completion error to signal if a RWQE is active
 462  *
 463  * Flushes both send and receive work queues.
 464  * Returns true if last WQE event should be generated.
 465  * The QP r_lock and s_lock should be held and interrupts disabled.
 466  * If we are already in error state, just return.
 467  */
 468 int qib_error_qp(struct qib_qp *qp, enum ib_wc_status err)
 469 {
 470         struct qib_ibdev *dev = to_idev(qp->ibqp.device);
 471         struct ib_wc wc;
 472         int ret = 0;
 473
 474         if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
 475                 goto bail;
 476
 477         qp->state = IB_QPS_ERR;
 478
 479         if (qp->s_flags & (QIB_S_TIMER | QIB_S_WAIT_RNR)) {
 480                 qp->s_flags &= ~(QIB_S_TIMER | QIB_S_WAIT_RNR);
 481                 del_timer(&qp->s_timer);
 482         }
 483
 484         if (qp->s_flags & QIB_S_ANY_WAIT_SEND)
 485                 qp->s_flags &= ~QIB_S_ANY_WAIT_SEND;
 486
 487         spin_lock(&dev->pending_lock);
 488         if (!list_empty(&qp->iowait) && !(qp->s_flags & QIB_S_BUSY)) {
 489                 qp->s_flags &= ~QIB_S_ANY_WAIT_IO;
 490                 list_del_init(&qp->iowait);
 491         }
 492         spin_unlock(&dev->pending_lock);
 493
 494         if (!(qp->s_flags & QIB_S_BUSY)) {
 495                 qp->s_hdrwords = 0;
 496                 if (qp->s_rdma_mr) {
 497                         qib_put_mr(qp->s_rdma_mr);
 498                         qp->s_rdma_mr = NULL;
 499                 }
 500                 if (qp->s_tx) {
 501                         qib_put_txreq(qp->s_tx);
 502                         qp->s_tx = NULL;
 503                 }
 504         }
 505
 506         /* Schedule the sending tasklet to drain the send work queue. */
 507         if (qp->s_last != qp->s_head)
 508                 qib_schedule_send(qp);
 509
 510         clear_mr_refs(qp, 0);
 511
 512         memset(&wc, 0, sizeof(wc));
 513         wc.qp = &qp->ibqp;
 514         wc.opcode = IB_WC_RECV;
 515
 516         if (test_and_clear_bit(QIB_R_WRID_VALID, &qp->r_aflags)) {
 517                 wc.wr_id = qp->r_wr_id;
 518                 wc.status = err;
 519                 qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 520         }
 521         wc.status = IB_WC_WR_FLUSH_ERR;
 522
 523         if (qp->r_rq.wq) {
 524                 struct qib_rwq *wq;
 525                 u32 head;
 526                 u32 tail;
 527
 528                 spin_lock(&qp->r_rq.lock);
 529
 530                 /* sanity check pointers before trusting them */
 531                 wq = qp->r_rq.wq;
 532                 head = wq->head;
 533                 if (head >= qp->r_rq.size)
 534                         head = 0;
 535                 tail = wq->tail;
 536                 if (tail >= qp->r_rq.size)
 537                         tail = 0;
 538                 while (tail != head) {
 539                         wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
 540                         if (++tail >= qp->r_rq.size)
 541                                 tail = 0;
 542                         qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 543                 }
 544                 wq->tail = tail;
 545
 546                 spin_unlock(&qp->r_rq.lock);
 547         } else if (qp->ibqp.event_handler)
 548                 ret = 1;
 549
 550 bail:
 551         return ret;
 552 }
 553
 554 /**
 555  * qib_modify_qp - modify the attributes of a queue pair
 556  * @ibqp: the queue pair who's attributes we're modifying
 557  * @attr: the new attributes
 558  * @attr_mask: the mask of attributes to modify
 559  * @udata: user data for libibverbs.so
 560  *
 561  * Returns 0 on success, otherwise returns an errno.
 562  */
 563 int qib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 564                   int attr_mask, struct ib_udata *udata)
 565 {
 566         struct qib_ibdev *dev = to_idev(ibqp->device);
 567         struct qib_qp *qp = to_iqp(ibqp);
 568         enum ib_qp_state cur_state, new_state;
 569         struct ib_event ev;
 570         int lastwqe = 0;
 571         int mig = 0;
 572         int ret;
 573         u32 pmtu = 0; /* for gcc warning only */
 574
 575         spin_lock_irq(&qp->r_lock);
 576         spin_lock(&qp->s_lock);
 577
 578         cur_state = attr_mask & IB_QP_CUR_STATE ?
 579                 attr->cur_qp_state : qp->state;
 580         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
 581
 582         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
 583                                 attr_mask))
 584                 goto inval;
 585
 586         if (attr_mask & IB_QP_AV) {
 587                 if (attr->ah_attr.dlid >= QIB_MULTICAST_LID_BASE)
 588                         goto inval;
 589                 if (qib_check_ah(qp->ibqp.device, &attr->ah_attr))
 590                         goto inval;
 591         }
 592
 593         if (attr_mask & IB_QP_ALT_PATH) {
 594                 if (attr->alt_ah_attr.dlid >= QIB_MULTICAST_LID_BASE)
 595                         goto inval;
 596                 if (qib_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
 597                         goto inval;
 598                 if (attr->alt_pkey_index >= qib_get_npkeys(dd_from_dev(dev)))
 599                         goto inval;
 600         }
 601
 602         if (attr_mask & IB_QP_PKEY_INDEX)
 603                 if (attr->pkey_index >= qib_get_npkeys(dd_from_dev(dev)))
 604                         goto inval;
 605
 606         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 607                 if (attr->min_rnr_timer > 31)
 608                         goto inval;
 609
 610         if (attr_mask & IB_QP_PORT)
 611                 if (qp->ibqp.qp_type == IB_QPT_SMI ||
 612                     qp->ibqp.qp_type == IB_QPT_GSI ||
 613                     attr->port_num == 0 ||
 614                     attr->port_num > ibqp->device->phys_port_cnt)
 615                         goto inval;
 616
 617         if (attr_mask & IB_QP_DEST_QPN)
 618                 if (attr->dest_qp_num > QIB_QPN_MASK)
 619                         goto inval;
 620
 621         if (attr_mask & IB_QP_RETRY_CNT)
 622                 if (attr->retry_cnt > 7)
 623                         goto inval;
 624
 625         if (attr_mask & IB_QP_RNR_RETRY)
 626                 if (attr->rnr_retry > 7)
 627                         goto inval;
 628
 629         /*
 630          * Don't allow invalid path_mtu values.  OK to set greater
 631          * than the active mtu (or even the max_cap, if we have tuned
 632          * that to a small mtu.  We'll set qp->path_mtu
 633          * to the lesser of requested attribute mtu and active,
 634          * for packetizing messages.
 635          * Note that the QP port has to be set in INIT and MTU in RTR.
 636          */
 637         if (attr_mask & IB_QP_PATH_MTU) {
 638                 struct qib_devdata *dd = dd_from_dev(dev);
 639                 int mtu, pidx = qp->port_num - 1;
 640
 641                 mtu = ib_mtu_enum_to_int(attr->path_mtu);
 642                 if (mtu == -1)
 643                         goto inval;
 644                 if (mtu > dd->pport[pidx].ibmtu) {
 645                         switch (dd->pport[pidx].ibmtu) {
 646                         case 4096:
 647                                 pmtu = IB_MTU_4096;
 648                                 break;
 649                         case 2048:
 650                                 pmtu = IB_MTU_2048;
 651                                 break;
 652                         case 1024:
 653                                 pmtu = IB_MTU_1024;
 654                                 break;
 655                         case 512:
 656                                 pmtu = IB_MTU_512;
 657                                 break;
 658                         case 256:
 659                                 pmtu = IB_MTU_256;
 660                                 break;
 661                         default:
 662                                 pmtu = IB_MTU_2048;
 663                         }
 664                 } else
 665                         pmtu = attr->path_mtu;
 666         }
 667
 668         if (attr_mask & IB_QP_PATH_MIG_STATE) {
 669                 if (attr->path_mig_state == IB_MIG_REARM) {
 670                         if (qp->s_mig_state == IB_MIG_ARMED)
 671                                 goto inval;
 672                         if (new_state != IB_QPS_RTS)
 673                                 goto inval;
 674                 } else if (attr->path_mig_state == IB_MIG_MIGRATED) {
 675                         if (qp->s_mig_state == IB_MIG_REARM)
 676                                 goto inval;
 677                         if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
 678                                 goto inval;
 679                         if (qp->s_mig_state == IB_MIG_ARMED)
 680                                 mig = 1;
 681                 } else
 682                         goto inval;
 683         }
 684
 685         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 686                 if (attr->max_dest_rd_atomic > QIB_MAX_RDMA_ATOMIC)
 687                         goto inval;
 688
 689         switch (new_state) {
 690         case IB_QPS_RESET:
 691                 if (qp->state != IB_QPS_RESET) {
 692                         qp->state = IB_QPS_RESET;
 693                         spin_lock(&dev->pending_lock);
 694                         if (!list_empty(&qp->iowait))
 695                                 list_del_init(&qp->iowait);
 696                         spin_unlock(&dev->pending_lock);
 697                         qp->s_flags &= ~(QIB_S_TIMER | QIB_S_ANY_WAIT);
 698                         spin_unlock(&qp->s_lock);
 699                         spin_unlock_irq(&qp->r_lock);
 700                         /* Stop the sending work queue and retry timer */
 701                         cancel_work_sync(&qp->s_work);
 702                         del_timer_sync(&qp->s_timer);
 703                         wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
 704                         if (qp->s_tx) {
 705                                 qib_put_txreq(qp->s_tx);
 706                                 qp->s_tx = NULL;
 707                         }
 708                         remove_qp(dev, qp);
 709                         wait_event(qp->wait, !atomic_read(&qp->refcount));
 710                         spin_lock_irq(&qp->r_lock);
 711                         spin_lock(&qp->s_lock);
 712                         clear_mr_refs(qp, 1);
 713                         qib_reset_qp(qp, ibqp->qp_type);
 714                 }
 715                 break;
 716
 717         case IB_QPS_RTR:
 718                 /* Allow event to retrigger if QP set to RTR more than once */
 719                 qp->r_flags &= ~QIB_R_COMM_EST;
 720                 qp->state = new_state;
 721                 break;
 722
 723         case IB_QPS_SQD:
 724                 qp->s_draining = qp->s_last != qp->s_cur;
 725                 qp->state = new_state;
 726                 break;
 727
 728         case IB_QPS_SQE:
 729                 if (qp->ibqp.qp_type == IB_QPT_RC)
 730                         goto inval;
 731                 qp->state = new_state;
 732                 break;
 733
 734         case IB_QPS_ERR:
 735                 lastwqe = qib_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 736                 break;
 737
 738         default:
 739                 qp->state = new_state;
 740                 break;
 741         }
 742
 743         if (attr_mask & IB_QP_PKEY_INDEX)
 744                 qp->s_pkey_index = attr->pkey_index;
 745
 746         if (attr_mask & IB_QP_PORT)
 747                 qp->port_num = attr->port_num;
 748
 749         if (attr_mask & IB_QP_DEST_QPN)
 750                 qp->remote_qpn = attr->dest_qp_num;
 751
 752         if (attr_mask & IB_QP_SQ_PSN) {
 753                 qp->s_next_psn = attr->sq_psn & QIB_PSN_MASK;
 754                 qp->s_psn = qp->s_next_psn;
 755                 qp->s_sending_psn = qp->s_next_psn;
 756                 qp->s_last_psn = qp->s_next_psn - 1;
 757                 qp->s_sending_hpsn = qp->s_last_psn;
 758         }
 759
 760         if (attr_mask & IB_QP_RQ_PSN)
 761                 qp->r_psn = attr->rq_psn & QIB_PSN_MASK;
 762
 763         if (attr_mask & IB_QP_ACCESS_FLAGS)
 764                 qp->qp_access_flags = attr->qp_access_flags;
 765
 766         if (attr_mask & IB_QP_AV) {
 767                 qp->remote_ah_attr = attr->ah_attr;
 768                 qp->s_srate = attr->ah_attr.static_rate;
 769         }
 770
 771         if (attr_mask & IB_QP_ALT_PATH) {
 772                 qp->alt_ah_attr = attr->alt_ah_attr;
 773                 qp->s_alt_pkey_index = attr->alt_pkey_index;
 774         }
 775
 776         if (attr_mask & IB_QP_PATH_MIG_STATE) {
 777                 qp->s_mig_state = attr->path_mig_state;
 778                 if (mig) {
 779                         qp->remote_ah_attr = qp->alt_ah_attr;
 780                         qp->port_num = qp->alt_ah_attr.port_num;
 781                         qp->s_pkey_index = qp->s_alt_pkey_index;
 782                 }
 783         }
 784
 785         if (attr_mask & IB_QP_PATH_MTU) {
 786                 qp->path_mtu = pmtu;
 787                 qp->pmtu = ib_mtu_enum_to_int(pmtu);
 788         }
 789
 790         if (attr_mask & IB_QP_RETRY_CNT) {
 791                 qp->s_retry_cnt = attr->retry_cnt;
 792                 qp->s_retry = attr->retry_cnt;
 793         }
 794
 795         if (attr_mask & IB_QP_RNR_RETRY) {
 796                 qp->s_rnr_retry_cnt = attr->rnr_retry;
 797                 qp->s_rnr_retry = attr->rnr_retry;
 798         }
 799
 800         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 801                 qp->r_min_rnr_timer = attr->min_rnr_timer;
 802
 803         if (attr_mask & IB_QP_TIMEOUT) {
 804                 qp->timeout = attr->timeout;
 805                 qp->timeout_jiffies =
 806                         usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
 807                                 1000UL);
 808         }
 809
 810         if (attr_mask & IB_QP_QKEY)
 811                 qp->qkey = attr->qkey;
 812
 813         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 814                 qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
 815
 816         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
 817                 qp->s_max_rd_atomic = attr->max_rd_atomic;
 818
 819         spin_unlock(&qp->s_lock);
 820         spin_unlock_irq(&qp->r_lock);
 821
 822         if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
 823                 insert_qp(dev, qp);
 824
 825         if (lastwqe) {
 826                 ev.device = qp->ibqp.device;
 827                 ev.element.qp = &qp->ibqp;
 828                 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 829                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 830         }
 831         if (mig) {
 832                 ev.device = qp->ibqp.device;
 833                 ev.element.qp = &qp->ibqp;
 834                 ev.event = IB_EVENT_PATH_MIG;
 835                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 836         }
 837         ret = 0;
 838         goto bail;
 839
 840 inval:
 841         spin_unlock(&qp->s_lock);
 842         spin_unlock_irq(&qp->r_lock);
 843         ret = -EINVAL;
 844
 845 bail:
 846         return ret;
 847 }
 848
 849 int qib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 850                  int attr_mask, struct ib_qp_init_attr *init_attr)
 851 {
 852         struct qib_qp *qp = to_iqp(ibqp);
 853
 854         attr->qp_state = qp->state;
 855         attr->cur_qp_state = attr->qp_state;
 856         attr->path_mtu = qp->path_mtu;
 857         attr->path_mig_state = qp->s_mig_state;
 858         attr->qkey = qp->qkey;
 859         attr->rq_psn = qp->r_psn & QIB_PSN_MASK;
 860         attr->sq_psn = qp->s_next_psn & QIB_PSN_MASK;
 861         attr->dest_qp_num = qp->remote_qpn;
 862         attr->qp_access_flags = qp->qp_access_flags;
 863         attr->cap.max_send_wr = qp->s_size - 1;
 864         attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
 865         attr->cap.max_send_sge = qp->s_max_sge;
 866         attr->cap.max_recv_sge = qp->r_rq.max_sge;
 867         attr->cap.max_inline_data = 0;
 868         attr->ah_attr = qp->remote_ah_attr;
 869         attr->alt_ah_attr = qp->alt_ah_attr;
 870         attr->pkey_index = qp->s_pkey_index;
 871         attr->alt_pkey_index = qp->s_alt_pkey_index;
 872         attr->en_sqd_async_notify = 0;
 873         attr->sq_draining = qp->s_draining;
 874         attr->max_rd_atomic = qp->s_max_rd_atomic;
 875         attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
 876         attr->min_rnr_timer = qp->r_min_rnr_timer;
 877         attr->port_num = qp->port_num;
 878         attr->timeout = qp->timeout;
 879         attr->retry_cnt = qp->s_retry_cnt;
 880         attr->rnr_retry = qp->s_rnr_retry_cnt;
 881         attr->alt_port_num = qp->alt_ah_attr.port_num;
 882         attr->alt_timeout = qp->alt_timeout;
 883
 884         init_attr->event_handler = qp->ibqp.event_handler;
 885         init_attr->qp_context = qp->ibqp.qp_context;
 886         init_attr->send_cq = qp->ibqp.send_cq;
 887         init_attr->recv_cq = qp->ibqp.recv_cq;
 888         init_attr->srq = qp->ibqp.srq;
 889         init_attr->cap = attr->cap;
 890         if (qp->s_flags & QIB_S_SIGNAL_REQ_WR)
 891                 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 892         else
 893                 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
 894         init_attr->qp_type = qp->ibqp.qp_type;
 895         init_attr->port_num = qp->port_num;
 896         return 0;
 897 }
 898
 899 /**
 900  * qib_compute_aeth - compute the AETH (syndrome + MSN)
 901  * @qp: the queue pair to compute the AETH for
 902  *
 903  * Returns the AETH.
 904  */
 905 __be32 qib_compute_aeth(struct qib_qp *qp)
 906 {
 907         u32 aeth = qp->r_msn & QIB_MSN_MASK;
 908
 909         if (qp->ibqp.srq) {
 910                 /*
 911                  * Shared receive queues don't generate credits.
 912                  * Set the credit field to the invalid value.
 913                  */
 914                 aeth |= QIB_AETH_CREDIT_INVAL << QIB_AETH_CREDIT_SHIFT;
 915         } else {
 916                 u32 min, max, x;
 917                 u32 credits;
 918                 struct qib_rwq *wq = qp->r_rq.wq;
 919                 u32 head;
 920                 u32 tail;
 921
 922                 /* sanity check pointers before trusting them */
 923                 head = wq->head;
 924                 if (head >= qp->r_rq.size)
 925                         head = 0;
 926                 tail = wq->tail;
 927                 if (tail >= qp->r_rq.size)
 928                         tail = 0;
 929                 /*
 930                  * Compute the number of credits available (RWQEs).
 931                  * XXX Not holding the r_rq.lock here so there is a small
 932                  * chance that the pair of reads are not atomic.
 933                  */
 934                 credits = head - tail;
 935                 if ((int)credits < 0)
 936                         credits += qp->r_rq.size;
 937                 /*
 938                  * Binary search the credit table to find the code to
 939                  * use.
 940                  */
 941                 min = 0;
 942                 max = 31;
 943                 for (;;) {
 944                         x = (min + max) / 2;
 945                         if (credit_table[x] == credits)
 946                                 break;
 947                         if (credit_table[x] > credits)
 948                                 max = x;
 949                         else if (min == x)
 950                                 break;
 951                         else
 952                                 min = x;
 953                 }
 954                 aeth |= x << QIB_AETH_CREDIT_SHIFT;
 955         }
 956         return cpu_to_be32(aeth);
 957 }
 958
 959 /**
 960  * qib_create_qp - create a queue pair for a device
 961  * @ibpd: the protection domain who's device we create the queue pair for
 962  * @init_attr: the attributes of the queue pair
 963  * @udata: user data for libibverbs.so
 964  *
 965  * Returns the queue pair on success, otherwise returns an errno.
 966  *
 967  * Called by the ib_create_qp() core verbs function.
 968  */
 969 struct ib_qp *qib_create_qp(struct ib_pd *ibpd,
 970                             struct ib_qp_init_attr *init_attr,
 971                             struct ib_udata *udata)
 972 {
 973         struct qib_qp *qp;
 974         int err;
 975         struct qib_swqe *swq = NULL;
 976         struct qib_ibdev *dev;
 977         struct qib_devdata *dd;
 978         size_t sz;
 979         size_t sg_list_sz;
 980         struct ib_qp *ret;
 981
 982         if (init_attr->cap.max_send_sge > ib_qib_max_sges ||
 983             init_attr->cap.max_send_wr > ib_qib_max_qp_wrs) {
 984                 ret = ERR_PTR(-EINVAL);
 985                 goto bail;
 986         }
 987
 988         /* Check receive queue parameters if no SRQ is specified. */
 989         if (!init_attr->srq) {
 990                 if (init_attr->cap.max_recv_sge > ib_qib_max_sges ||
 991                     init_attr->cap.max_recv_wr > ib_qib_max_qp_wrs) {
 992                         ret = ERR_PTR(-EINVAL);
 993                         goto bail;
 994                 }
 995                 if (init_attr->cap.max_send_sge +
 996                     init_attr->cap.max_send_wr +
 997                     init_attr->cap.max_recv_sge +
 998                     init_attr->cap.max_recv_wr == 0) {
 999                         ret = ERR_PTR(-EINVAL);
1000                         goto bail;
1001                 }
1002         }
1003
1004         switch (init_attr->qp_type) {
1005         case IB_QPT_SMI:
1006         case IB_QPT_GSI:
1007                 if (init_attr->port_num == 0 ||
1008                     init_attr->port_num > ibpd->device->phys_port_cnt) {
1009                         ret = ERR_PTR(-EINVAL);
1010                         goto bail;
1011                 }
1012         case IB_QPT_UC:
1013         case IB_QPT_RC:
1014         case IB_QPT_UD:
1015                 sz = sizeof(struct qib_sge) *
1016                         init_attr->cap.max_send_sge +
1017                         sizeof(struct qib_swqe);
1018                 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
1019                 if (swq == NULL) {
1020                         ret = ERR_PTR(-ENOMEM);
1021                         goto bail;
1022                 }
1023                 sz = sizeof(*qp);
1024                 sg_list_sz = 0;
1025                 if (init_attr->srq) {
1026                         struct qib_srq *srq = to_isrq(init_attr->srq);
1027
1028                         if (srq->rq.max_sge > 1)
1029                                 sg_list_sz = sizeof(*qp->r_sg_list) *
1030                                         (srq->rq.max_sge - 1);
1031                 } else if (init_attr->cap.max_recv_sge > 1)
1032                         sg_list_sz = sizeof(*qp->r_sg_list) *
1033                                 (init_attr->cap.max_recv_sge - 1);
1034                 qp = kzalloc(sz + sg_list_sz, GFP_KERNEL);
1035                 if (!qp) {
1036                         ret = ERR_PTR(-ENOMEM);
1037                         goto bail_swq;
1038                 }
1039                 RCU_INIT_POINTER(qp->next, NULL);
1040                 qp->s_hdr = kzalloc(sizeof(*qp->s_hdr), GFP_KERNEL);
1041                 if (!qp->s_hdr) {
1042                         ret = ERR_PTR(-ENOMEM);
1043                         goto bail_qp;
1044                 }
1045                 qp->timeout_jiffies =
1046                         usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
1047                                 1000UL);
1048                 if (init_attr->srq)
1049                         sz = 0;
1050                 else {
1051                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
1052                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
1053                         sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
1054                                 sizeof(struct qib_rwqe);
1055                         qp->r_rq.wq = vmalloc_user(sizeof(struct qib_rwq) +
1056                                                    qp->r_rq.size * sz);
1057                         if (!qp->r_rq.wq) {
1058                                 ret = ERR_PTR(-ENOMEM);
1059                                 goto bail_qp;
1060                         }
1061                 }
1062
1063                 /*
1064                  * ib_create_qp() will initialize qp->ibqp
1065                  * except for qp->ibqp.qp_num.
1066                  */
1067                 spin_lock_init(&qp->r_lock);
1068                 spin_lock_init(&qp->s_lock);
1069                 spin_lock_init(&qp->r_rq.lock);
1070                 atomic_set(&qp->refcount, 0);
1071                 init_waitqueue_head(&qp->wait);
1072                 init_waitqueue_head(&qp->wait_dma);
1073                 init_timer(&qp->s_timer);
1074                 qp->s_timer.data = (unsigned long)qp;
1075                 INIT_WORK(&qp->s_work, qib_do_send);
1076                 INIT_LIST_HEAD(&qp->iowait);
1077                 INIT_LIST_HEAD(&qp->rspwait);
1078                 qp->state = IB_QPS_RESET;
1079                 qp->s_wq = swq;
1080                 qp->s_size = init_attr->cap.max_send_wr + 1;
1081                 qp->s_max_sge = init_attr->cap.max_send_sge;
1082                 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
1083                         qp->s_flags = QIB_S_SIGNAL_REQ_WR;
1084                 dev = to_idev(ibpd->device);
1085                 dd = dd_from_dev(dev);
1086                 err = alloc_qpn(dd, &dev->qpn_table, init_attr->qp_type,
1087                                 init_attr->port_num);
1088                 if (err < 0) {
1089                         ret = ERR_PTR(err);
1090                         vfree(qp->r_rq.wq);
1091                         goto bail_qp;
1092                 }
1093                 qp->ibqp.qp_num = err;
1094                 qp->port_num = init_attr->port_num;
1095                 qib_reset_qp(qp, init_attr->qp_type);
1096                 break;
1097
1098         default:
1099                 /* Don't support raw QPs */
1100                 ret = ERR_PTR(-ENOSYS);
1101                 goto bail;
1102         }
1103
1104         init_attr->cap.max_inline_data = 0;
1105
1106         /*
1107          * Return the address of the RWQ as the offset to mmap.
1108          * See qib_mmap() for details.
1109          */
1110         if (udata && udata->outlen >= sizeof(__u64)) {
1111                 if (!qp->r_rq.wq) {
1112                         __u64 offset = 0;
1113
1114                         err = ib_copy_to_udata(udata, &offset,
1115                                                sizeof(offset));
1116                         if (err) {
1117                                 ret = ERR_PTR(err);
1118                                 goto bail_ip;
1119                         }
1120                 } else {
1121                         u32 s = sizeof(struct qib_rwq) + qp->r_rq.size * sz;
1122
1123                         qp->ip = qib_create_mmap_info(dev, s,
1124                                                       ibpd->uobject->context,
1125                                                       qp->r_rq.wq);
1126                         if (!qp->ip) {
1127                                 ret = ERR_PTR(-ENOMEM);
1128                                 goto bail_ip;
1129                         }
1130
1131                         err = ib_copy_to_udata(udata, &(qp->ip->offset),
1132                                                sizeof(qp->ip->offset));
1133                         if (err) {
1134                                 ret = ERR_PTR(err);
1135                                 goto bail_ip;
1136                         }
1137                 }
1138         }
1139
1140         spin_lock(&dev->n_qps_lock);
1141         if (dev->n_qps_allocated == ib_qib_max_qps) {
1142                 spin_unlock(&dev->n_qps_lock);
1143                 ret = ERR_PTR(-ENOMEM);
1144                 goto bail_ip;
1145         }
1146
1147         dev->n_qps_allocated++;
1148         spin_unlock(&dev->n_qps_lock);
1149
1150         if (qp->ip) {
1151                 spin_lock_irq(&dev->pending_lock);
1152                 list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps);
1153                 spin_unlock_irq(&dev->pending_lock);
1154         }
1155
1156         ret = &qp->ibqp;
1157         goto bail;
1158
1159 bail_ip:
1160         if (qp->ip)
1161                 kref_put(&qp->ip->ref, qib_release_mmap_info);
1162         else
1163                 vfree(qp->r_rq.wq);
1164         free_qpn(&dev->qpn_table, qp->ibqp.qp_num);
1165 bail_qp:
1166         kfree(qp->s_hdr);
1167         kfree(qp);
1168 bail_swq:
1169         vfree(swq);
1170 bail:
1171         return ret;
1172 }
1173
1174 /**
1175  * qib_destroy_qp - destroy a queue pair
1176  * @ibqp: the queue pair to destroy
1177  *
1178  * Returns 0 on success.
1179  *
1180  * Note that this can be called while the QP is actively sending or
1181  * receiving!
1182  */
1183 int qib_destroy_qp(struct ib_qp *ibqp)
1184 {
1185         struct qib_qp *qp = to_iqp(ibqp);
1186         struct qib_ibdev *dev = to_idev(ibqp->device);
1187
1188         /* Make sure HW and driver activity is stopped. */
1189         spin_lock_irq(&qp->s_lock);
1190         if (qp->state != IB_QPS_RESET) {
1191                 qp->state = IB_QPS_RESET;
1192                 spin_lock(&dev->pending_lock);
1193                 if (!list_empty(&qp->iowait))
1194                         list_del_init(&qp->iowait);
1195                 spin_unlock(&dev->pending_lock);
1196                 qp->s_flags &= ~(QIB_S_TIMER | QIB_S_ANY_WAIT);
1197                 spin_unlock_irq(&qp->s_lock);
1198                 cancel_work_sync(&qp->s_work);
1199                 del_timer_sync(&qp->s_timer);
1200                 wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
1201                 if (qp->s_tx) {
1202                         qib_put_txreq(qp->s_tx);
1203                         qp->s_tx = NULL;
1204                 }
1205                 remove_qp(dev, qp);
1206                 wait_event(qp->wait, !atomic_read(&qp->refcount));
1207                 clear_mr_refs(qp, 1);
1208         } else
1209                 spin_unlock_irq(&qp->s_lock);
1210
1211         /* all user's cleaned up, mark it available */
1212         free_qpn(&dev->qpn_table, qp->ibqp.qp_num);
1213         spin_lock(&dev->n_qps_lock);
1214         dev->n_qps_allocated--;
1215         spin_unlock(&dev->n_qps_lock);
1216
1217         if (qp->ip)
1218                 kref_put(&qp->ip->ref, qib_release_mmap_info);
1219         else
1220                 vfree(qp->r_rq.wq);
1221         vfree(qp->s_wq);
1222         kfree(qp->s_hdr);
1223         kfree(qp);
1224         return 0;
1225 }
1226
1227 /**
1228  * qib_init_qpn_table - initialize the QP number table for a device
1229  * @qpt: the QPN table
1230  */
1231 void qib_init_qpn_table(struct qib_devdata *dd, struct qib_qpn_table *qpt)
1232 {
1233         spin_lock_init(&qpt->lock);
1234         qpt->last = 1;          /* start with QPN 2 */
1235         qpt->nmaps = 1;
1236         qpt->mask = dd->qpn_mask;
1237 }
1238
1239 /**
1240  * qib_free_qpn_table - free the QP number table for a device
1241  * @qpt: the QPN table
1242  */
1243 void qib_free_qpn_table(struct qib_qpn_table *qpt)
1244 {
1245         int i;
1246
1247         for (i = 0; i < ARRAY_SIZE(qpt->map); i++)
1248                 if (qpt->map[i].page)
1249                         free_page((unsigned long) qpt->map[i].page);
1250 }
1251
1252 /**
1253  * qib_get_credit - flush the send work queue of a QP
1254  * @qp: the qp who's send work queue to flush
1255  * @aeth: the Acknowledge Extended Transport Header
1256  *
1257  * The QP s_lock should be held.
1258  */
1259 void qib_get_credit(struct qib_qp *qp, u32 aeth)
1260 {
1261         u32 credit = (aeth >> QIB_AETH_CREDIT_SHIFT) & QIB_AETH_CREDIT_MASK;
1262
1263         /*
1264          * If the credit is invalid, we can send
1265          * as many packets as we like.  Otherwise, we have to
1266          * honor the credit field.
1267          */
1268         if (credit == QIB_AETH_CREDIT_INVAL) {
1269                 if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT)) {
1270                         qp->s_flags |= QIB_S_UNLIMITED_CREDIT;
1271                         if (qp->s_flags & QIB_S_WAIT_SSN_CREDIT) {
1272                                 qp->s_flags &= ~QIB_S_WAIT_SSN_CREDIT;
1273                                 qib_schedule_send(qp);
1274                         }
1275                 }
1276         } else if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT)) {
1277                 /* Compute new LSN (i.e., MSN + credit) */
1278                 credit = (aeth + credit_table[credit]) & QIB_MSN_MASK;
1279                 if (qib_cmp24(credit, qp->s_lsn) > 0) {
1280                         qp->s_lsn = credit;
1281                         if (qp->s_flags & QIB_S_WAIT_SSN_CREDIT) {
1282                                 qp->s_flags &= ~QIB_S_WAIT_SSN_CREDIT;
1283                                 qib_schedule_send(qp);
1284                         }
1285                 }
1286         }
1287 }