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 <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         atomic_inc(&qp->refcount);
 226         spin_lock_irqsave(&dev->qpt_lock, flags);
 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 }
 239
 240 /*
 241  * Remove the QP from the table so it can't be found asynchronously by
 242  * the receive interrupt routine.
 243  */
 244 static void remove_qp(struct qib_ibdev *dev, struct qib_qp *qp)
 245 {
 246         struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 247         unsigned n = qpn_hash(dev, qp->ibqp.qp_num);
 248         unsigned long flags;
 249         int removed = 1;
 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                 rcu_assign_pointer(ibp->qp0, NULL);
 256         } else if (rcu_dereference_protected(ibp->qp1,
 257                         lockdep_is_held(&dev->qpt_lock)) == qp) {
 258                 rcu_assign_pointer(ibp->qp1, NULL);
 259         } else {
 260                 struct qib_qp *q;
 261                 struct qib_qp __rcu **qpp;
 262
 263                 removed = 0;
 264                 qpp = &dev->qp_table[n];
 265                 for (; (q = rcu_dereference_protected(*qpp,
 266                                 lockdep_is_held(&dev->qpt_lock))) != NULL;
 267                                 qpp = &q->next)
 268                         if (q == qp) {
 269                                 rcu_assign_pointer(*qpp,
 270                                         rcu_dereference_protected(qp->next,
 271                                          lockdep_is_held(&dev->qpt_lock)));
 272                                 removed = 1;
 273                                 break;
 274                         }
 275         }
 276
 277         spin_unlock_irqrestore(&dev->qpt_lock, flags);
 278         if (removed) {
 279                 synchronize_rcu();
 280                 atomic_dec(&qp->refcount);
 281         }
 282 }
 283
 284 /**
 285  * qib_free_all_qps - check for QPs still in use
 286  * @qpt: the QP table to empty
 287  *
 288  * There should not be any QPs still in use.
 289  * Free memory for table.
 290  */
 291 unsigned qib_free_all_qps(struct qib_devdata *dd)
 292 {
 293         struct qib_ibdev *dev = &dd->verbs_dev;
 294         unsigned long flags;
 295         struct qib_qp *qp;
 296         unsigned n, qp_inuse = 0;
 297
 298         for (n = 0; n < dd->num_pports; n++) {
 299                 struct qib_ibport *ibp = &dd->pport[n].ibport_data;
 300
 301                 if (!qib_mcast_tree_empty(ibp))
 302                         qp_inuse++;
 303                 rcu_read_lock();
 304                 if (rcu_dereference(ibp->qp0))
 305                         qp_inuse++;
 306                 if (rcu_dereference(ibp->qp1))
 307                         qp_inuse++;
 308                 rcu_read_unlock();
 309         }
 310
 311         spin_lock_irqsave(&dev->qpt_lock, flags);
 312         for (n = 0; n < dev->qp_table_size; n++) {
 313                 qp = rcu_dereference_protected(dev->qp_table[n],
 314                         lockdep_is_held(&dev->qpt_lock));
 315                 rcu_assign_pointer(dev->qp_table[n], NULL);
 316
 317                 for (; qp; qp = rcu_dereference_protected(qp->next,
 318                                         lockdep_is_held(&dev->qpt_lock)))
 319                         qp_inuse++;
 320         }
 321         spin_unlock_irqrestore(&dev->qpt_lock, flags);
 322         synchronize_rcu();
 323
 324         return qp_inuse;
 325 }
 326
 327 /**
 328  * qib_lookup_qpn - return the QP with the given QPN
 329  * @qpt: the QP table
 330  * @qpn: the QP number to look up
 331  *
 332  * The caller is responsible for decrementing the QP reference count
 333  * when done.
 334  */
 335 struct qib_qp *qib_lookup_qpn(struct qib_ibport *ibp, u32 qpn)
 336 {
 337         struct qib_qp *qp = NULL;
 338
 339         rcu_read_lock();
 340         if (unlikely(qpn <= 1)) {
 341                 if (qpn == 0)
 342                         qp = rcu_dereference(ibp->qp0);
 343                 else
 344                         qp = rcu_dereference(ibp->qp1);
 345                 if (qp)
 346                         atomic_inc(&qp->refcount);
 347         } else {
 348                 struct qib_ibdev *dev = &ppd_from_ibp(ibp)->dd->verbs_dev;
 349                 unsigned n = qpn_hash(dev, qpn);
 350
 351                 for (qp = rcu_dereference(dev->qp_table[n]); qp;
 352                         qp = rcu_dereference(qp->next))
 353                         if (qp->ibqp.qp_num == qpn) {
 354                                 atomic_inc(&qp->refcount);
 355                                 break;
 356                         }
 357         }
 358         rcu_read_unlock();
 359         return qp;
 360 }
 361
 362 /**
 363  * qib_reset_qp - initialize the QP state to the reset state
 364  * @qp: the QP to reset
 365  * @type: the QP type
 366  */
 367 static void qib_reset_qp(struct qib_qp *qp, enum ib_qp_type type)
 368 {
 369         qp->remote_qpn = 0;
 370         qp->qkey = 0;
 371         qp->qp_access_flags = 0;
 372         atomic_set(&qp->s_dma_busy, 0);
 373         qp->s_flags &= QIB_S_SIGNAL_REQ_WR;
 374         qp->s_hdrwords = 0;
 375         qp->s_wqe = NULL;
 376         qp->s_draining = 0;
 377         qp->s_next_psn = 0;
 378         qp->s_last_psn = 0;
 379         qp->s_sending_psn = 0;
 380         qp->s_sending_hpsn = 0;
 381         qp->s_psn = 0;
 382         qp->r_psn = 0;
 383         qp->r_msn = 0;
 384         if (type == IB_QPT_RC) {
 385                 qp->s_state = IB_OPCODE_RC_SEND_LAST;
 386                 qp->r_state = IB_OPCODE_RC_SEND_LAST;
 387         } else {
 388                 qp->s_state = IB_OPCODE_UC_SEND_LAST;
 389                 qp->r_state = IB_OPCODE_UC_SEND_LAST;
 390         }
 391         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 392         qp->r_nak_state = 0;
 393         qp->r_aflags = 0;
 394         qp->r_flags = 0;
 395         qp->s_head = 0;
 396         qp->s_tail = 0;
 397         qp->s_cur = 0;
 398         qp->s_acked = 0;
 399         qp->s_last = 0;
 400         qp->s_ssn = 1;
 401         qp->s_lsn = 0;
 402         qp->s_mig_state = IB_MIG_MIGRATED;
 403         memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue));
 404         qp->r_head_ack_queue = 0;
 405         qp->s_tail_ack_queue = 0;
 406         qp->s_num_rd_atomic = 0;
 407         if (qp->r_rq.wq) {
 408                 qp->r_rq.wq->head = 0;
 409                 qp->r_rq.wq->tail = 0;
 410         }
 411         qp->r_sge.num_sge = 0;
 412 }
 413
 414 static void clear_mr_refs(struct qib_qp *qp, int clr_sends)
 415 {
 416         unsigned n;
 417
 418         if (test_and_clear_bit(QIB_R_REWIND_SGE, &qp->r_aflags))
 419                 qib_put_ss(&qp->s_rdma_read_sge);
 420
 421         qib_put_ss(&qp->r_sge);
 422
 423         if (clr_sends) {
 424                 while (qp->s_last != qp->s_head) {
 425                         struct qib_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 426                         unsigned i;
 427
 428                         for (i = 0; i < wqe->wr.num_sge; i++) {
 429                                 struct qib_sge *sge = &wqe->sg_list[i];
 430
 431                                 qib_put_mr(sge->mr);
 432                         }
 433                         if (qp->ibqp.qp_type == IB_QPT_UD ||
 434                             qp->ibqp.qp_type == IB_QPT_SMI ||
 435                             qp->ibqp.qp_type == IB_QPT_GSI)
 436                                 atomic_dec(&to_iah(wqe->wr.wr.ud.ah)->refcount);
 437                         if (++qp->s_last >= qp->s_size)
 438                                 qp->s_last = 0;
 439                 }
 440                 if (qp->s_rdma_mr) {
 441                         qib_put_mr(qp->s_rdma_mr);
 442                         qp->s_rdma_mr = NULL;
 443                 }
 444         }
 445
 446         if (qp->ibqp.qp_type != IB_QPT_RC)
 447                 return;
 448
 449         for (n = 0; n < ARRAY_SIZE(qp->s_ack_queue); n++) {
 450                 struct qib_ack_entry *e = &qp->s_ack_queue[n];
 451
 452                 if (e->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST &&
 453                     e->rdma_sge.mr) {
 454                         qib_put_mr(e->rdma_sge.mr);
 455                         e->rdma_sge.mr = NULL;
 456                 }
 457         }
 458 }
 459
 460 /**
 461  * qib_error_qp - put a QP into the error state
 462  * @qp: the QP to put into the error state
 463  * @err: the receive completion error to signal if a RWQE is active
 464  *
 465  * Flushes both send and receive work queues.
 466  * Returns true if last WQE event should be generated.
 467  * The QP r_lock and s_lock should be held and interrupts disabled.
 468  * If we are already in error state, just return.
 469  */
 470 int qib_error_qp(struct qib_qp *qp, enum ib_wc_status err)
 471 {
 472         struct qib_ibdev *dev = to_idev(qp->ibqp.device);
 473         struct ib_wc wc;
 474         int ret = 0;
 475
 476         if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
 477                 goto bail;
 478
 479         qp->state = IB_QPS_ERR;
 480
 481         if (qp->s_flags & (QIB_S_TIMER | QIB_S_WAIT_RNR)) {
 482                 qp->s_flags &= ~(QIB_S_TIMER | QIB_S_WAIT_RNR);
 483                 del_timer(&qp->s_timer);
 484         }
 485
 486         if (qp->s_flags & QIB_S_ANY_WAIT_SEND)
 487                 qp->s_flags &= ~QIB_S_ANY_WAIT_SEND;
 488
 489         spin_lock(&dev->pending_lock);
 490         if (!list_empty(&qp->iowait) && !(qp->s_flags & QIB_S_BUSY)) {
 491                 qp->s_flags &= ~QIB_S_ANY_WAIT_IO;
 492                 list_del_init(&qp->iowait);
 493         }
 494         spin_unlock(&dev->pending_lock);
 495
 496         if (!(qp->s_flags & QIB_S_BUSY)) {
 497                 qp->s_hdrwords = 0;
 498                 if (qp->s_rdma_mr) {
 499                         qib_put_mr(qp->s_rdma_mr);
 500                         qp->s_rdma_mr = NULL;
 501                 }
 502                 if (qp->s_tx) {
 503                         qib_put_txreq(qp->s_tx);
 504                         qp->s_tx = NULL;
 505                 }
 506         }
 507
 508         /* Schedule the sending tasklet to drain the send work queue. */
 509         if (qp->s_last != qp->s_head)
 510                 qib_schedule_send(qp);
 511
 512         clear_mr_refs(qp, 0);
 513
 514         memset(&wc, 0, sizeof(wc));
 515         wc.qp = &qp->ibqp;
 516         wc.opcode = IB_WC_RECV;
 517
 518         if (test_and_clear_bit(QIB_R_WRID_VALID, &qp->r_aflags)) {
 519                 wc.wr_id = qp->r_wr_id;
 520                 wc.status = err;
 521                 qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 522         }
 523         wc.status = IB_WC_WR_FLUSH_ERR;
 524
 525         if (qp->r_rq.wq) {
 526                 struct qib_rwq *wq;
 527                 u32 head;
 528                 u32 tail;
 529
 530                 spin_lock(&qp->r_rq.lock);
 531
 532                 /* sanity check pointers before trusting them */
 533                 wq = qp->r_rq.wq;
 534                 head = wq->head;
 535                 if (head >= qp->r_rq.size)
 536                         head = 0;
 537                 tail = wq->tail;
 538                 if (tail >= qp->r_rq.size)
 539                         tail = 0;
 540                 while (tail != head) {
 541                         wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
 542                         if (++tail >= qp->r_rq.size)
 543                                 tail = 0;
 544                         qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 545                 }
 546                 wq->tail = tail;
 547
 548                 spin_unlock(&qp->r_rq.lock);
 549         } else if (qp->ibqp.event_handler)
 550                 ret = 1;
 551
 552 bail:
 553         return ret;
 554 }
 555
 556 /**
 557  * qib_modify_qp - modify the attributes of a queue pair
 558  * @ibqp: the queue pair who's attributes we're modifying
 559  * @attr: the new attributes
 560  * @attr_mask: the mask of attributes to modify
 561  * @udata: user data for libibverbs.so
 562  *
 563  * Returns 0 on success, otherwise returns an errno.
 564  */
 565 int qib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 566                   int attr_mask, struct ib_udata *udata)
 567 {
 568         struct qib_ibdev *dev = to_idev(ibqp->device);
 569         struct qib_qp *qp = to_iqp(ibqp);
 570         enum ib_qp_state cur_state, new_state;
 571         struct ib_event ev;
 572         int lastwqe = 0;
 573         int mig = 0;
 574         int ret;
 575         u32 pmtu = 0; /* for gcc warning only */
 576
 577         spin_lock_irq(&qp->r_lock);
 578         spin_lock(&qp->s_lock);
 579
 580         cur_state = attr_mask & IB_QP_CUR_STATE ?
 581                 attr->cur_qp_state : qp->state;
 582         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
 583
 584         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
 585                                 attr_mask))
 586                 goto inval;
 587
 588         if (attr_mask & IB_QP_AV) {
 589                 if (attr->ah_attr.dlid >= QIB_MULTICAST_LID_BASE)
 590                         goto inval;
 591                 if (qib_check_ah(qp->ibqp.device, &attr->ah_attr))
 592                         goto inval;
 593         }
 594
 595         if (attr_mask & IB_QP_ALT_PATH) {
 596                 if (attr->alt_ah_attr.dlid >= QIB_MULTICAST_LID_BASE)
 597                         goto inval;
 598                 if (qib_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
 599                         goto inval;
 600                 if (attr->alt_pkey_index >= qib_get_npkeys(dd_from_dev(dev)))
 601                         goto inval;
 602         }
 603
 604         if (attr_mask & IB_QP_PKEY_INDEX)
 605                 if (attr->pkey_index >= qib_get_npkeys(dd_from_dev(dev)))
 606                         goto inval;
 607
 608         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 609                 if (attr->min_rnr_timer > 31)
 610                         goto inval;
 611
 612         if (attr_mask & IB_QP_PORT)
 613                 if (qp->ibqp.qp_type == IB_QPT_SMI ||
 614                     qp->ibqp.qp_type == IB_QPT_GSI ||
 615                     attr->port_num == 0 ||
 616                     attr->port_num > ibqp->device->phys_port_cnt)
 617                         goto inval;
 618
 619         if (attr_mask & IB_QP_DEST_QPN)
 620                 if (attr->dest_qp_num > QIB_QPN_MASK)
 621                         goto inval;
 622
 623         if (attr_mask & IB_QP_RETRY_CNT)
 624                 if (attr->retry_cnt > 7)
 625                         goto inval;
 626
 627         if (attr_mask & IB_QP_RNR_RETRY)
 628                 if (attr->rnr_retry > 7)
 629                         goto inval;
 630
 631         /*
 632          * Don't allow invalid path_mtu values.  OK to set greater
 633          * than the active mtu (or even the max_cap, if we have tuned
 634          * that to a small mtu.  We'll set qp->path_mtu
 635          * to the lesser of requested attribute mtu and active,
 636          * for packetizing messages.
 637          * Note that the QP port has to be set in INIT and MTU in RTR.
 638          */
 639         if (attr_mask & IB_QP_PATH_MTU) {
 640                 struct qib_devdata *dd = dd_from_dev(dev);
 641                 int mtu, pidx = qp->port_num - 1;
 642
 643                 mtu = ib_mtu_enum_to_int(attr->path_mtu);
 644                 if (mtu == -1)
 645                         goto inval;
 646                 if (mtu > dd->pport[pidx].ibmtu) {
 647                         switch (dd->pport[pidx].ibmtu) {
 648                         case 4096:
 649                                 pmtu = IB_MTU_4096;
 650                                 break;
 651                         case 2048:
 652                                 pmtu = IB_MTU_2048;
 653                                 break;
 654                         case 1024:
 655                                 pmtu = IB_MTU_1024;
 656                                 break;
 657                         case 512:
 658                                 pmtu = IB_MTU_512;
 659                                 break;
 660                         case 256:
 661                                 pmtu = IB_MTU_256;
 662                                 break;
 663                         default:
 664                                 pmtu = IB_MTU_2048;
 665                         }
 666                 } else
 667                         pmtu = attr->path_mtu;
 668         }
 669
 670         if (attr_mask & IB_QP_PATH_MIG_STATE) {
 671                 if (attr->path_mig_state == IB_MIG_REARM) {
 672                         if (qp->s_mig_state == IB_MIG_ARMED)
 673                                 goto inval;
 674                         if (new_state != IB_QPS_RTS)
 675                                 goto inval;
 676                 } else if (attr->path_mig_state == IB_MIG_MIGRATED) {
 677                         if (qp->s_mig_state == IB_MIG_REARM)
 678                                 goto inval;
 679                         if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
 680                                 goto inval;
 681                         if (qp->s_mig_state == IB_MIG_ARMED)
 682                                 mig = 1;
 683                 } else
 684                         goto inval;
 685         }
 686
 687         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 688                 if (attr->max_dest_rd_atomic > QIB_MAX_RDMA_ATOMIC)
 689                         goto inval;
 690
 691         switch (new_state) {
 692         case IB_QPS_RESET:
 693                 if (qp->state != IB_QPS_RESET) {
 694                         qp->state = IB_QPS_RESET;
 695                         spin_lock(&dev->pending_lock);
 696                         if (!list_empty(&qp->iowait))
 697                                 list_del_init(&qp->iowait);
 698                         spin_unlock(&dev->pending_lock);
 699                         qp->s_flags &= ~(QIB_S_TIMER | QIB_S_ANY_WAIT);
 700                         spin_unlock(&qp->s_lock);
 701                         spin_unlock_irq(&qp->r_lock);
 702                         /* Stop the sending work queue and retry timer */
 703                         cancel_work_sync(&qp->s_work);
 704                         del_timer_sync(&qp->s_timer);
 705                         wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
 706                         if (qp->s_tx) {
 707                                 qib_put_txreq(qp->s_tx);
 708                                 qp->s_tx = NULL;
 709                         }
 710                         remove_qp(dev, qp);
 711                         wait_event(qp->wait, !atomic_read(&qp->refcount));
 712                         spin_lock_irq(&qp->r_lock);
 713                         spin_lock(&qp->s_lock);
 714                         clear_mr_refs(qp, 1);
 715                         qib_reset_qp(qp, ibqp->qp_type);
 716                 }
 717                 break;
 718
 719         case IB_QPS_RTR:
 720                 /* Allow event to retrigger if QP set to RTR more than once */
 721                 qp->r_flags &= ~QIB_R_COMM_EST;
 722                 qp->state = new_state;
 723                 break;
 724
 725         case IB_QPS_SQD:
 726                 qp->s_draining = qp->s_last != qp->s_cur;
 727                 qp->state = new_state;
 728                 break;
 729
 730         case IB_QPS_SQE:
 731                 if (qp->ibqp.qp_type == IB_QPT_RC)
 732                         goto inval;
 733                 qp->state = new_state;
 734                 break;
 735
 736         case IB_QPS_ERR:
 737                 lastwqe = qib_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 738                 break;
 739
 740         default:
 741                 qp->state = new_state;
 742                 break;
 743         }
 744
 745         if (attr_mask & IB_QP_PKEY_INDEX)
 746                 qp->s_pkey_index = attr->pkey_index;
 747
 748         if (attr_mask & IB_QP_PORT)
 749                 qp->port_num = attr->port_num;
 750
 751         if (attr_mask & IB_QP_DEST_QPN)
 752                 qp->remote_qpn = attr->dest_qp_num;
 753
 754         if (attr_mask & IB_QP_SQ_PSN) {
 755                 qp->s_next_psn = attr->sq_psn & QIB_PSN_MASK;
 756                 qp->s_psn = qp->s_next_psn;
 757                 qp->s_sending_psn = qp->s_next_psn;
 758                 qp->s_last_psn = qp->s_next_psn - 1;
 759                 qp->s_sending_hpsn = qp->s_last_psn;
 760         }
 761
 762         if (attr_mask & IB_QP_RQ_PSN)
 763                 qp->r_psn = attr->rq_psn & QIB_PSN_MASK;
 764
 765         if (attr_mask & IB_QP_ACCESS_FLAGS)
 766                 qp->qp_access_flags = attr->qp_access_flags;
 767
 768         if (attr_mask & IB_QP_AV) {
 769                 qp->remote_ah_attr = attr->ah_attr;
 770                 qp->s_srate = attr->ah_attr.static_rate;
 771         }
 772
 773         if (attr_mask & IB_QP_ALT_PATH) {
 774                 qp->alt_ah_attr = attr->alt_ah_attr;
 775                 qp->s_alt_pkey_index = attr->alt_pkey_index;
 776         }
 777
 778         if (attr_mask & IB_QP_PATH_MIG_STATE) {
 779                 qp->s_mig_state = attr->path_mig_state;
 780                 if (mig) {
 781                         qp->remote_ah_attr = qp->alt_ah_attr;
 782                         qp->port_num = qp->alt_ah_attr.port_num;
 783                         qp->s_pkey_index = qp->s_alt_pkey_index;
 784                 }
 785         }
 786
 787         if (attr_mask & IB_QP_PATH_MTU) {
 788                 qp->path_mtu = pmtu;
 789                 qp->pmtu = ib_mtu_enum_to_int(pmtu);
 790         }
 791
 792         if (attr_mask & IB_QP_RETRY_CNT) {
 793                 qp->s_retry_cnt = attr->retry_cnt;
 794                 qp->s_retry = attr->retry_cnt;
 795         }
 796
 797         if (attr_mask & IB_QP_RNR_RETRY) {
 798                 qp->s_rnr_retry_cnt = attr->rnr_retry;
 799                 qp->s_rnr_retry = attr->rnr_retry;
 800         }
 801
 802         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 803                 qp->r_min_rnr_timer = attr->min_rnr_timer;
 804
 805         if (attr_mask & IB_QP_TIMEOUT) {
 806                 qp->timeout = attr->timeout;
 807                 qp->timeout_jiffies =
 808                         usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
 809                                 1000UL);
 810         }
 811
 812         if (attr_mask & IB_QP_QKEY)
 813                 qp->qkey = attr->qkey;
 814
 815         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 816                 qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
 817
 818         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
 819                 qp->s_max_rd_atomic = attr->max_rd_atomic;
 820
 821         spin_unlock(&qp->s_lock);
 822         spin_unlock_irq(&qp->r_lock);
 823
 824         if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
 825                 insert_qp(dev, qp);
 826
 827         if (lastwqe) {
 828                 ev.device = qp->ibqp.device;
 829                 ev.element.qp = &qp->ibqp;
 830                 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 831                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 832         }
 833         if (mig) {
 834                 ev.device = qp->ibqp.device;
 835                 ev.element.qp = &qp->ibqp;
 836                 ev.event = IB_EVENT_PATH_MIG;
 837                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 838         }
 839         ret = 0;
 840         goto bail;
 841
 842 inval:
 843         spin_unlock(&qp->s_lock);
 844         spin_unlock_irq(&qp->r_lock);
 845         ret = -EINVAL;
 846
 847 bail:
 848         return ret;
 849 }
 850
 851 int qib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 852                  int attr_mask, struct ib_qp_init_attr *init_attr)
 853 {
 854         struct qib_qp *qp = to_iqp(ibqp);
 855
 856         attr->qp_state = qp->state;
 857         attr->cur_qp_state = attr->qp_state;
 858         attr->path_mtu = qp->path_mtu;
 859         attr->path_mig_state = qp->s_mig_state;
 860         attr->qkey = qp->qkey;
 861         attr->rq_psn = qp->r_psn & QIB_PSN_MASK;
 862         attr->sq_psn = qp->s_next_psn & QIB_PSN_MASK;
 863         attr->dest_qp_num = qp->remote_qpn;
 864         attr->qp_access_flags = qp->qp_access_flags;
 865         attr->cap.max_send_wr = qp->s_size - 1;
 866         attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
 867         attr->cap.max_send_sge = qp->s_max_sge;
 868         attr->cap.max_recv_sge = qp->r_rq.max_sge;
 869         attr->cap.max_inline_data = 0;
 870         attr->ah_attr = qp->remote_ah_attr;
 871         attr->alt_ah_attr = qp->alt_ah_attr;
 872         attr->pkey_index = qp->s_pkey_index;
 873         attr->alt_pkey_index = qp->s_alt_pkey_index;
 874         attr->en_sqd_async_notify = 0;
 875         attr->sq_draining = qp->s_draining;
 876         attr->max_rd_atomic = qp->s_max_rd_atomic;
 877         attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
 878         attr->min_rnr_timer = qp->r_min_rnr_timer;
 879         attr->port_num = qp->port_num;
 880         attr->timeout = qp->timeout;
 881         attr->retry_cnt = qp->s_retry_cnt;
 882         attr->rnr_retry = qp->s_rnr_retry_cnt;
 883         attr->alt_port_num = qp->alt_ah_attr.port_num;
 884         attr->alt_timeout = qp->alt_timeout;
 885
 886         init_attr->event_handler = qp->ibqp.event_handler;
 887         init_attr->qp_context = qp->ibqp.qp_context;
 888         init_attr->send_cq = qp->ibqp.send_cq;
 889         init_attr->recv_cq = qp->ibqp.recv_cq;
 890         init_attr->srq = qp->ibqp.srq;
 891         init_attr->cap = attr->cap;
 892         if (qp->s_flags & QIB_S_SIGNAL_REQ_WR)
 893                 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 894         else
 895                 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
 896         init_attr->qp_type = qp->ibqp.qp_type;
 897         init_attr->port_num = qp->port_num;
 898         return 0;
 899 }
 900
 901 /**
 902  * qib_compute_aeth - compute the AETH (syndrome + MSN)
 903  * @qp: the queue pair to compute the AETH for
 904  *
 905  * Returns the AETH.
 906  */
 907 __be32 qib_compute_aeth(struct qib_qp *qp)
 908 {
 909         u32 aeth = qp->r_msn & QIB_MSN_MASK;
 910
 911         if (qp->ibqp.srq) {
 912                 /*
 913                  * Shared receive queues don't generate credits.
 914                  * Set the credit field to the invalid value.
 915                  */
 916                 aeth |= QIB_AETH_CREDIT_INVAL << QIB_AETH_CREDIT_SHIFT;
 917         } else {
 918                 u32 min, max, x;
 919                 u32 credits;
 920                 struct qib_rwq *wq = qp->r_rq.wq;
 921                 u32 head;
 922                 u32 tail;
 923
 924                 /* sanity check pointers before trusting them */
 925                 head = wq->head;
 926                 if (head >= qp->r_rq.size)
 927                         head = 0;
 928                 tail = wq->tail;
 929                 if (tail >= qp->r_rq.size)
 930                         tail = 0;
 931                 /*
 932                  * Compute the number of credits available (RWQEs).
 933                  * XXX Not holding the r_rq.lock here so there is a small
 934                  * chance that the pair of reads are not atomic.
 935                  */
 936                 credits = head - tail;
 937                 if ((int)credits < 0)
 938                         credits += qp->r_rq.size;
 939                 /*
 940                  * Binary search the credit table to find the code to
 941                  * use.
 942                  */
 943                 min = 0;
 944                 max = 31;
 945                 for (;;) {
 946                         x = (min + max) / 2;
 947                         if (credit_table[x] == credits)
 948                                 break;
 949                         if (credit_table[x] > credits)
 950                                 max = x;
 951                         else if (min == x)
 952                                 break;
 953                         else
 954                                 min = x;
 955                 }
 956                 aeth |= x << QIB_AETH_CREDIT_SHIFT;
 957         }
 958         return cpu_to_be32(aeth);
 959 }
 960
 961 /**
 962  * qib_create_qp - create a queue pair for a device
 963  * @ibpd: the protection domain who's device we create the queue pair for
 964  * @init_attr: the attributes of the queue pair
 965  * @udata: user data for libibverbs.so
 966  *
 967  * Returns the queue pair on success, otherwise returns an errno.
 968  *
 969  * Called by the ib_create_qp() core verbs function.
 970  */
 971 struct ib_qp *qib_create_qp(struct ib_pd *ibpd,
 972                             struct ib_qp_init_attr *init_attr,
 973                             struct ib_udata *udata)
 974 {
 975         struct qib_qp *qp;
 976         int err;
 977         struct qib_swqe *swq = NULL;
 978         struct qib_ibdev *dev;
 979         struct qib_devdata *dd;
 980         size_t sz;
 981         size_t sg_list_sz;
 982         struct ib_qp *ret;
 983
 984         if (init_attr->cap.max_send_sge > ib_qib_max_sges ||
 985             init_attr->cap.max_send_wr > ib_qib_max_qp_wrs) {
 986                 ret = ERR_PTR(-EINVAL);
 987                 goto bail;
 988         }
 989
 990         /* Check receive queue parameters if no SRQ is specified. */
 991         if (!init_attr->srq) {
 992                 if (init_attr->cap.max_recv_sge > ib_qib_max_sges ||
 993                     init_attr->cap.max_recv_wr > ib_qib_max_qp_wrs) {
 994                         ret = ERR_PTR(-EINVAL);
 995                         goto bail;
 996                 }
 997                 if (init_attr->cap.max_send_sge +
 998                     init_attr->cap.max_send_wr +
 999                     init_attr->cap.max_recv_sge +
1000                     init_attr->cap.max_recv_wr == 0) {
1001                         ret = ERR_PTR(-EINVAL);
1002                         goto bail;
1003                 }
1004         }
1005
1006         switch (init_attr->qp_type) {
1007         case IB_QPT_SMI:
1008         case IB_QPT_GSI:
1009                 if (init_attr->port_num == 0 ||
1010                     init_attr->port_num > ibpd->device->phys_port_cnt) {
1011                         ret = ERR_PTR(-EINVAL);
1012                         goto bail;
1013                 }
1014         case IB_QPT_UC:
1015         case IB_QPT_RC:
1016         case IB_QPT_UD:
1017                 sz = sizeof(struct qib_sge) *
1018                         init_attr->cap.max_send_sge +
1019                         sizeof(struct qib_swqe);
1020                 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
1021                 if (swq == NULL) {
1022                         ret = ERR_PTR(-ENOMEM);
1023                         goto bail;
1024                 }
1025                 sz = sizeof(*qp);
1026                 sg_list_sz = 0;
1027                 if (init_attr->srq) {
1028                         struct qib_srq *srq = to_isrq(init_attr->srq);
1029
1030                         if (srq->rq.max_sge > 1)
1031                                 sg_list_sz = sizeof(*qp->r_sg_list) *
1032                                         (srq->rq.max_sge - 1);
1033                 } else if (init_attr->cap.max_recv_sge > 1)
1034                         sg_list_sz = sizeof(*qp->r_sg_list) *
1035                                 (init_attr->cap.max_recv_sge - 1);
1036                 qp = kzalloc(sz + sg_list_sz, GFP_KERNEL);
1037                 if (!qp) {
1038                         ret = ERR_PTR(-ENOMEM);
1039                         goto bail_swq;
1040                 }
1041                 RCU_INIT_POINTER(qp->next, NULL);
1042                 qp->s_hdr = kzalloc(sizeof(*qp->s_hdr), GFP_KERNEL);
1043                 if (!qp->s_hdr) {
1044                         ret = ERR_PTR(-ENOMEM);
1045                         goto bail_qp;
1046                 }
1047                 qp->timeout_jiffies =
1048                         usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
1049                                 1000UL);
1050                 if (init_attr->srq)
1051                         sz = 0;
1052                 else {
1053                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
1054                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
1055                         sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
1056                                 sizeof(struct qib_rwqe);
1057                         qp->r_rq.wq = vmalloc_user(sizeof(struct qib_rwq) +
1058                                                    qp->r_rq.size * sz);
1059                         if (!qp->r_rq.wq) {
1060                                 ret = ERR_PTR(-ENOMEM);
1061                                 goto bail_qp;
1062                         }
1063                 }
1064
1065                 /*
1066                  * ib_create_qp() will initialize qp->ibqp
1067                  * except for qp->ibqp.qp_num.
1068                  */
1069                 spin_lock_init(&qp->r_lock);
1070                 spin_lock_init(&qp->s_lock);
1071                 spin_lock_init(&qp->r_rq.lock);
1072                 atomic_set(&qp->refcount, 0);
1073                 init_waitqueue_head(&qp->wait);
1074                 init_waitqueue_head(&qp->wait_dma);
1075                 init_timer(&qp->s_timer);
1076                 qp->s_timer.data = (unsigned long)qp;
1077                 INIT_WORK(&qp->s_work, qib_do_send);
1078                 INIT_LIST_HEAD(&qp->iowait);
1079                 INIT_LIST_HEAD(&qp->rspwait);
1080                 qp->state = IB_QPS_RESET;
1081                 qp->s_wq = swq;
1082                 qp->s_size = init_attr->cap.max_send_wr + 1;
1083                 qp->s_max_sge = init_attr->cap.max_send_sge;
1084                 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
1085                         qp->s_flags = QIB_S_SIGNAL_REQ_WR;
1086                 dev = to_idev(ibpd->device);
1087                 dd = dd_from_dev(dev);
1088                 err = alloc_qpn(dd, &dev->qpn_table, init_attr->qp_type,
1089                                 init_attr->port_num);
1090                 if (err < 0) {
1091                         ret = ERR_PTR(err);
1092                         vfree(qp->r_rq.wq);
1093                         goto bail_qp;
1094                 }
1095                 qp->ibqp.qp_num = err;
1096                 qp->port_num = init_attr->port_num;
1097                 qib_reset_qp(qp, init_attr->qp_type);
1098                 break;
1099
1100         default:
1101                 /* Don't support raw QPs */
1102                 ret = ERR_PTR(-ENOSYS);
1103                 goto bail;
1104         }
1105
1106         init_attr->cap.max_inline_data = 0;
1107
1108         /*
1109          * Return the address of the RWQ as the offset to mmap.
1110          * See qib_mmap() for details.
1111          */
1112         if (udata && udata->outlen >= sizeof(__u64)) {
1113                 if (!qp->r_rq.wq) {
1114                         __u64 offset = 0;
1115
1116                         err = ib_copy_to_udata(udata, &offset,
1117                                                sizeof(offset));
1118                         if (err) {
1119                                 ret = ERR_PTR(err);
1120                                 goto bail_ip;
1121                         }
1122                 } else {
1123                         u32 s = sizeof(struct qib_rwq) + qp->r_rq.size * sz;
1124
1125                         qp->ip = qib_create_mmap_info(dev, s,
1126                                                       ibpd->uobject->context,
1127                                                       qp->r_rq.wq);
1128                         if (!qp->ip) {
1129                                 ret = ERR_PTR(-ENOMEM);
1130                                 goto bail_ip;
1131                         }
1132
1133                         err = ib_copy_to_udata(udata, &(qp->ip->offset),
1134                                                sizeof(qp->ip->offset));
1135                         if (err) {
1136                                 ret = ERR_PTR(err);
1137                                 goto bail_ip;
1138                         }
1139                 }
1140         }
1141
1142         spin_lock(&dev->n_qps_lock);
1143         if (dev->n_qps_allocated == ib_qib_max_qps) {
1144                 spin_unlock(&dev->n_qps_lock);
1145                 ret = ERR_PTR(-ENOMEM);
1146                 goto bail_ip;
1147         }
1148
1149         dev->n_qps_allocated++;
1150         spin_unlock(&dev->n_qps_lock);
1151
1152         if (qp->ip) {
1153                 spin_lock_irq(&dev->pending_lock);
1154                 list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps);
1155                 spin_unlock_irq(&dev->pending_lock);
1156         }
1157
1158         ret = &qp->ibqp;
1159         goto bail;
1160
1161 bail_ip:
1162         if (qp->ip)
1163                 kref_put(&qp->ip->ref, qib_release_mmap_info);
1164         else
1165                 vfree(qp->r_rq.wq);
1166         free_qpn(&dev->qpn_table, qp->ibqp.qp_num);
1167 bail_qp:
1168         kfree(qp->s_hdr);
1169         kfree(qp);
1170 bail_swq:
1171         vfree(swq);
1172 bail:
1173         return ret;
1174 }
1175
1176 /**
1177  * qib_destroy_qp - destroy a queue pair
1178  * @ibqp: the queue pair to destroy
1179  *
1180  * Returns 0 on success.
1181  *
1182  * Note that this can be called while the QP is actively sending or
1183  * receiving!
1184  */
1185 int qib_destroy_qp(struct ib_qp *ibqp)
1186 {
1187         struct qib_qp *qp = to_iqp(ibqp);
1188         struct qib_ibdev *dev = to_idev(ibqp->device);
1189
1190         /* Make sure HW and driver activity is stopped. */
1191         spin_lock_irq(&qp->s_lock);
1192         if (qp->state != IB_QPS_RESET) {
1193                 qp->state = IB_QPS_RESET;
1194                 spin_lock(&dev->pending_lock);
1195                 if (!list_empty(&qp->iowait))
1196                         list_del_init(&qp->iowait);
1197                 spin_unlock(&dev->pending_lock);
1198                 qp->s_flags &= ~(QIB_S_TIMER | QIB_S_ANY_WAIT);
1199                 spin_unlock_irq(&qp->s_lock);
1200                 cancel_work_sync(&qp->s_work);
1201                 del_timer_sync(&qp->s_timer);
1202                 wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
1203                 if (qp->s_tx) {
1204                         qib_put_txreq(qp->s_tx);
1205                         qp->s_tx = NULL;
1206                 }
1207                 remove_qp(dev, qp);
1208                 wait_event(qp->wait, !atomic_read(&qp->refcount));
1209                 clear_mr_refs(qp, 1);
1210         } else
1211                 spin_unlock_irq(&qp->s_lock);
1212
1213         /* all user's cleaned up, mark it available */
1214         free_qpn(&dev->qpn_table, qp->ibqp.qp_num);
1215         spin_lock(&dev->n_qps_lock);
1216         dev->n_qps_allocated--;
1217         spin_unlock(&dev->n_qps_lock);
1218
1219         if (qp->ip)
1220                 kref_put(&qp->ip->ref, qib_release_mmap_info);
1221         else
1222                 vfree(qp->r_rq.wq);
1223         vfree(qp->s_wq);
1224         kfree(qp->s_hdr);
1225         kfree(qp);
1226         return 0;
1227 }
1228
1229 /**
1230  * qib_init_qpn_table - initialize the QP number table for a device
1231  * @qpt: the QPN table
1232  */
1233 void qib_init_qpn_table(struct qib_devdata *dd, struct qib_qpn_table *qpt)
1234 {
1235         spin_lock_init(&qpt->lock);
1236         qpt->last = 1;          /* start with QPN 2 */
1237         qpt->nmaps = 1;
1238         qpt->mask = dd->qpn_mask;
1239 }
1240
1241 /**
1242  * qib_free_qpn_table - free the QP number table for a device
1243  * @qpt: the QPN table
1244  */
1245 void qib_free_qpn_table(struct qib_qpn_table *qpt)
1246 {
1247         int i;
1248
1249         for (i = 0; i < ARRAY_SIZE(qpt->map); i++)
1250                 if (qpt->map[i].page)
1251                         free_page((unsigned long) qpt->map[i].page);
1252 }
1253
1254 /**
1255  * qib_get_credit - flush the send work queue of a QP
1256  * @qp: the qp who's send work queue to flush
1257  * @aeth: the Acknowledge Extended Transport Header
1258  *
1259  * The QP s_lock should be held.
1260  */
1261 void qib_get_credit(struct qib_qp *qp, u32 aeth)
1262 {
1263         u32 credit = (aeth >> QIB_AETH_CREDIT_SHIFT) & QIB_AETH_CREDIT_MASK;
1264
1265         /*
1266          * If the credit is invalid, we can send
1267          * as many packets as we like.  Otherwise, we have to
1268          * honor the credit field.
1269          */
1270         if (credit == QIB_AETH_CREDIT_INVAL) {
1271                 if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT)) {
1272                         qp->s_flags |= QIB_S_UNLIMITED_CREDIT;
1273                         if (qp->s_flags & QIB_S_WAIT_SSN_CREDIT) {
1274                                 qp->s_flags &= ~QIB_S_WAIT_SSN_CREDIT;
1275                                 qib_schedule_send(qp);
1276                         }
1277                 }
1278         } else if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT)) {
1279                 /* Compute new LSN (i.e., MSN + credit) */
1280                 credit = (aeth + credit_table[credit]) & QIB_MSN_MASK;
1281                 if (qib_cmp24(credit, qp->s_lsn) > 0) {
1282                         qp->s_lsn = credit;
1283                         if (qp->s_flags & QIB_S_WAIT_SSN_CREDIT) {
1284                                 qp->s_flags &= ~QIB_S_WAIT_SSN_CREDIT;
1285                                 qib_schedule_send(qp);
1286                         }
1287                 }
1288         }
1289 }