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 #include <rdma/rdma_vt.h>
  39 #ifdef CONFIG_DEBUG_FS
  40 #include <linux/seq_file.h>
  41 #endif
  42
  43 #include "qib.h"
  44
  45 /*
  46  * mask field which was present in now deleted qib_qpn_table
  47  * is not present in rvt_qpn_table. Defining the same field
  48  * as qpt_mask here instead of adding the mask field to
  49  * rvt_qpn_table.
  50  */
  51 static u16 qpt_mask;
  52
  53 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt,
  54                               struct rvt_qpn_map *map, unsigned off)
  55 {
  56         return (map - qpt->map) * RVT_BITS_PER_PAGE + off;
  57 }
  58
  59 static inline unsigned find_next_offset(struct rvt_qpn_table *qpt,
  60                                         struct rvt_qpn_map *map, unsigned off,
  61                                         unsigned n)
  62 {
  63         if (qpt_mask) {
  64                 off++;
  65                 if (((off & qpt_mask) >> 1) >= n)
  66                         off = (off | qpt_mask) + 2;
  67         } else {
  68                 off = find_next_zero_bit(map->page, RVT_BITS_PER_PAGE, off);
  69         }
  70         return off;
  71 }
  72
  73 /*
  74  * Convert the AETH credit code into the number of credits.
  75  */
  76 static u32 credit_table[31] = {
  77         0,                      /* 0 */
  78         1,                      /* 1 */
  79         2,                      /* 2 */
  80         3,                      /* 3 */
  81         4,                      /* 4 */
  82         6,                      /* 5 */
  83         8,                      /* 6 */
  84         12,                     /* 7 */
  85         16,                     /* 8 */
  86         24,                     /* 9 */
  87         32,                     /* A */
  88         48,                     /* B */
  89         64,                     /* C */
  90         96,                     /* D */
  91         128,                    /* E */
  92         192,                    /* F */
  93         256,                    /* 10 */
  94         384,                    /* 11 */
  95         512,                    /* 12 */
  96         768,                    /* 13 */
  97         1024,                   /* 14 */
  98         1536,                   /* 15 */
  99         2048,                   /* 16 */
 100         3072,                   /* 17 */
 101         4096,                   /* 18 */
 102         6144,                   /* 19 */
 103         8192,                   /* 1A */
 104         12288,                  /* 1B */
 105         16384,                  /* 1C */
 106         24576,                  /* 1D */
 107         32768                   /* 1E */
 108 };
 109
 110 static void get_map_page(struct rvt_qpn_table *qpt, struct rvt_qpn_map *map,
 111                          gfp_t gfp)
 112 {
 113         unsigned long page = get_zeroed_page(gfp);
 114
 115         /*
 116          * Free the page if someone raced with us installing it.
 117          */
 118
 119         spin_lock(&qpt->lock);
 120         if (map->page)
 121                 free_page(page);
 122         else
 123                 map->page = (void *)page;
 124         spin_unlock(&qpt->lock);
 125 }
 126
 127 /*
 128  * Allocate the next available QPN or
 129  * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI.
 130  */
 131 static int alloc_qpn(struct qib_devdata *dd, struct rvt_qpn_table *qpt,
 132                      enum ib_qp_type type, u8 port, gfp_t gfp)
 133 {
 134         u32 i, offset, max_scan, qpn;
 135         struct rvt_qpn_map *map;
 136         u32 ret;
 137
 138         if (type == IB_QPT_SMI || type == IB_QPT_GSI) {
 139                 unsigned n;
 140
 141                 ret = type == IB_QPT_GSI;
 142                 n = 1 << (ret + 2 * (port - 1));
 143                 spin_lock(&qpt->lock);
 144                 if (qpt->flags & n)
 145                         ret = -EINVAL;
 146                 else
 147                         qpt->flags |= n;
 148                 spin_unlock(&qpt->lock);
 149                 goto bail;
 150         }
 151
 152         qpn = qpt->last + 2;
 153         if (qpn >= RVT_QPN_MAX)
 154                 qpn = 2;
 155         if (qpt_mask && ((qpn & qpt_mask) >> 1) >= dd->n_krcv_queues)
 156                 qpn = (qpn | qpt_mask) + 2;
 157         offset = qpn & RVT_BITS_PER_PAGE_MASK;
 158         map = &qpt->map[qpn / RVT_BITS_PER_PAGE];
 159         max_scan = qpt->nmaps - !offset;
 160         for (i = 0;;) {
 161                 if (unlikely(!map->page)) {
 162                         get_map_page(qpt, map, gfp);
 163                         if (unlikely(!map->page))
 164                                 break;
 165                 }
 166                 do {
 167                         if (!test_and_set_bit(offset, map->page)) {
 168                                 qpt->last = qpn;
 169                                 ret = qpn;
 170                                 goto bail;
 171                         }
 172                         offset = find_next_offset(qpt, map, offset,
 173                                 dd->n_krcv_queues);
 174                         qpn = mk_qpn(qpt, map, offset);
 175                         /*
 176                          * This test differs from alloc_pidmap().
 177                          * If find_next_offset() does find a zero
 178                          * bit, we don't need to check for QPN
 179                          * wrapping around past our starting QPN.
 180                          * We just need to be sure we don't loop
 181                          * forever.
 182                          */
 183                 } while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX);
 184                 /*
 185                  * In order to keep the number of pages allocated to a
 186                  * minimum, we scan the all existing pages before increasing
 187                  * the size of the bitmap table.
 188                  */
 189                 if (++i > max_scan) {
 190                         if (qpt->nmaps == RVT_QPNMAP_ENTRIES)
 191                                 break;
 192                         map = &qpt->map[qpt->nmaps++];
 193                         offset = 0;
 194                 } else if (map < &qpt->map[qpt->nmaps]) {
 195                         ++map;
 196                         offset = 0;
 197                 } else {
 198                         map = &qpt->map[0];
 199                         offset = 2;
 200                 }
 201                 qpn = mk_qpn(qpt, map, offset);
 202         }
 203
 204         ret = -ENOMEM;
 205
 206 bail:
 207         return ret;
 208 }
 209
 210 static void free_qpn(struct rvt_qpn_table *qpt, u32 qpn)
 211 {
 212         struct rvt_qpn_map *map;
 213
 214         map = qpt->map + qpn / RVT_BITS_PER_PAGE;
 215         if (map->page)
 216                 clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page);
 217 }
 218
 219 static inline unsigned qpn_hash(struct qib_ibdev *dev, u32 qpn)
 220 {
 221         return jhash_1word(qpn, dev->qp_rnd) &
 222                 (dev->rdi.qp_dev->qp_table_size - 1);
 223 }
 224
 225
 226 /*
 227  * Put the QP into the hash table.
 228  * The hash table holds a reference to the QP.
 229  */
 230 static void insert_qp(struct qib_ibdev *dev, struct rvt_qp *qp)
 231 {
 232         struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 233         unsigned long flags;
 234         unsigned n = qpn_hash(dev, qp->ibqp.qp_num);
 235
 236         atomic_inc(&qp->refcount);
 237         spin_lock_irqsave(&dev->rdi.qp_dev->qpt_lock, flags);
 238
 239         if (qp->ibqp.qp_num == 0)
 240                 rcu_assign_pointer(ibp->rvp.qp[0], qp);
 241         else if (qp->ibqp.qp_num == 1)
 242                 rcu_assign_pointer(ibp->rvp.qp[1], qp);
 243         else {
 244                 qp->next = dev->rdi.qp_dev->qp_table[n];
 245                 rcu_assign_pointer(dev->rdi.qp_dev->qp_table[n], qp);
 246         }
 247
 248         spin_unlock_irqrestore(&dev->rdi.qp_dev->qpt_lock, flags);
 249 }
 250
 251 /*
 252  * Remove the QP from the table so it can't be found asynchronously by
 253  * the receive interrupt routine.
 254  */
 255 static void remove_qp(struct qib_ibdev *dev, struct rvt_qp *qp)
 256 {
 257         struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 258         unsigned n = qpn_hash(dev, qp->ibqp.qp_num);
 259         unsigned long flags;
 260         int removed = 1;
 261         spinlock_t *qpt_lock_ptr; /* Pointer to make checkpatch happy */
 262
 263         spin_lock_irqsave(&dev->rdi.qp_dev->qpt_lock, flags);
 264
 265         qpt_lock_ptr = &dev->rdi.qp_dev->qpt_lock;
 266         if (rcu_dereference_protected(ibp->rvp.qp[0],
 267                                       lockdep_is_held(qpt_lock_ptr)) == qp) {
 268                 RCU_INIT_POINTER(ibp->rvp.qp[0], NULL);
 269         } else if (rcu_dereference_protected(ibp->rvp.qp[1],
 270                         lockdep_is_held(&dev->rdi.qp_dev->qpt_lock)) == qp) {
 271                 RCU_INIT_POINTER(ibp->rvp.qp[1], NULL);
 272         } else {
 273                 struct rvt_qp *q;
 274                 struct rvt_qp __rcu **qpp;
 275
 276                 removed = 0;
 277                 qpp = &dev->rdi.qp_dev->qp_table[n];
 278                 for (; (q = rcu_dereference_protected(*qpp,
 279                                 lockdep_is_held(qpt_lock_ptr))) != NULL;
 280                                 qpp = &q->next)
 281                         if (q == qp) {
 282                                 RCU_INIT_POINTER(*qpp,
 283                                         rcu_dereference_protected(qp->next,
 284                                          lockdep_is_held(qpt_lock_ptr)));
 285                                 removed = 1;
 286                                 break;
 287                         }
 288         }
 289
 290         spin_unlock_irqrestore(&dev->rdi.qp_dev->qpt_lock, flags);
 291         if (removed) {
 292                 synchronize_rcu();
 293                 atomic_dec(&qp->refcount);
 294         }
 295 }
 296
 297 /**
 298  * qib_free_all_qps - check for QPs still in use
 299  * @qpt: the QP table to empty
 300  *
 301  * There should not be any QPs still in use.
 302  * Free memory for table.
 303  */
 304 unsigned qib_free_all_qps(struct qib_devdata *dd)
 305 {
 306         struct qib_ibdev *dev = &dd->verbs_dev;
 307         unsigned long flags;
 308         struct rvt_qp *qp;
 309         unsigned n, qp_inuse = 0;
 310         spinlock_t *qpt_lock_ptr; /* Pointer to make checkpatch happy */
 311
 312         for (n = 0; n < dd->num_pports; n++) {
 313                 struct qib_ibport *ibp = &dd->pport[n].ibport_data;
 314
 315                 if (!qib_mcast_tree_empty(ibp))
 316                         qp_inuse++;
 317                 rcu_read_lock();
 318                 if (rcu_dereference(ibp->rvp.qp[0]))
 319                         qp_inuse++;
 320                 if (rcu_dereference(ibp->rvp.qp[1]))
 321                         qp_inuse++;
 322                 rcu_read_unlock();
 323         }
 324
 325         spin_lock_irqsave(&dev->rdi.qp_dev->qpt_lock, flags);
 326         qpt_lock_ptr = &dev->rdi.qp_dev->qpt_lock;
 327         for (n = 0; n < dev->rdi.qp_dev->qp_table_size; n++) {
 328                 qp = rcu_dereference_protected(dev->rdi.qp_dev->qp_table[n],
 329                                                lockdep_is_held(qpt_lock_ptr));
 330                 RCU_INIT_POINTER(dev->rdi.qp_dev->qp_table[n], NULL);
 331
 332                 for (; qp; qp = rcu_dereference_protected(qp->next,
 333                                         lockdep_is_held(qpt_lock_ptr)))
 334                         qp_inuse++;
 335         }
 336         spin_unlock_irqrestore(&dev->rdi.qp_dev->qpt_lock, flags);
 337         synchronize_rcu();
 338
 339         return qp_inuse;
 340 }
 341
 342 /**
 343  * qib_lookup_qpn - return the QP with the given QPN
 344  * @qpt: the QP table
 345  * @qpn: the QP number to look up
 346  *
 347  * The caller is responsible for decrementing the QP reference count
 348  * when done.
 349  */
 350 struct rvt_qp *qib_lookup_qpn(struct qib_ibport *ibp, u32 qpn)
 351 {
 352         struct rvt_qp *qp = NULL;
 353
 354         rcu_read_lock();
 355         if (unlikely(qpn <= 1)) {
 356                 if (qpn == 0)
 357                         qp = rcu_dereference(ibp->rvp.qp[0]);
 358                 else
 359                         qp = rcu_dereference(ibp->rvp.qp[1]);
 360                 if (qp)
 361                         atomic_inc(&qp->refcount);
 362         } else {
 363                 struct qib_ibdev *dev = &ppd_from_ibp(ibp)->dd->verbs_dev;
 364                 unsigned n = qpn_hash(dev, qpn);
 365
 366                 for (qp = rcu_dereference(dev->rdi.qp_dev->qp_table[n]); qp;
 367                         qp = rcu_dereference(qp->next))
 368                         if (qp->ibqp.qp_num == qpn) {
 369                                 atomic_inc(&qp->refcount);
 370                                 break;
 371                         }
 372         }
 373         rcu_read_unlock();
 374         return qp;
 375 }
 376
 377 /**
 378  * qib_reset_qp - initialize the QP state to the reset state
 379  * @qp: the QP to reset
 380  * @type: the QP type
 381  */
 382 static void qib_reset_qp(struct rvt_qp *qp, enum ib_qp_type type)
 383 {
 384         struct qib_qp_priv *priv = qp->priv;
 385         qp->remote_qpn = 0;
 386         qp->qkey = 0;
 387         qp->qp_access_flags = 0;
 388         atomic_set(&priv->s_dma_busy, 0);
 389         qp->s_flags &= QIB_S_SIGNAL_REQ_WR;
 390         qp->s_hdrwords = 0;
 391         qp->s_wqe = NULL;
 392         qp->s_draining = 0;
 393         qp->s_next_psn = 0;
 394         qp->s_last_psn = 0;
 395         qp->s_sending_psn = 0;
 396         qp->s_sending_hpsn = 0;
 397         qp->s_psn = 0;
 398         qp->r_psn = 0;
 399         qp->r_msn = 0;
 400         if (type == IB_QPT_RC) {
 401                 qp->s_state = IB_OPCODE_RC_SEND_LAST;
 402                 qp->r_state = IB_OPCODE_RC_SEND_LAST;
 403         } else {
 404                 qp->s_state = IB_OPCODE_UC_SEND_LAST;
 405                 qp->r_state = IB_OPCODE_UC_SEND_LAST;
 406         }
 407         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 408         qp->r_nak_state = 0;
 409         qp->r_aflags = 0;
 410         qp->r_flags = 0;
 411         qp->s_head = 0;
 412         qp->s_tail = 0;
 413         qp->s_cur = 0;
 414         qp->s_acked = 0;
 415         qp->s_last = 0;
 416         qp->s_ssn = 1;
 417         qp->s_lsn = 0;
 418         qp->s_mig_state = IB_MIG_MIGRATED;
 419         memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue));
 420         qp->r_head_ack_queue = 0;
 421         qp->s_tail_ack_queue = 0;
 422         qp->s_num_rd_atomic = 0;
 423         if (qp->r_rq.wq) {
 424                 qp->r_rq.wq->head = 0;
 425                 qp->r_rq.wq->tail = 0;
 426         }
 427         qp->r_sge.num_sge = 0;
 428 }
 429
 430 static void clear_mr_refs(struct rvt_qp *qp, int clr_sends)
 431 {
 432         unsigned n;
 433
 434         if (test_and_clear_bit(QIB_R_REWIND_SGE, &qp->r_aflags))
 435                 qib_put_ss(&qp->s_rdma_read_sge);
 436
 437         qib_put_ss(&qp->r_sge);
 438
 439         if (clr_sends) {
 440                 while (qp->s_last != qp->s_head) {
 441                         struct rvt_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 442                         unsigned i;
 443
 444                         for (i = 0; i < wqe->wr.num_sge; i++) {
 445                                 struct rvt_sge *sge = &wqe->sg_list[i];
 446
 447                                 rvt_put_mr(sge->mr);
 448                         }
 449                         if (qp->ibqp.qp_type == IB_QPT_UD ||
 450                             qp->ibqp.qp_type == IB_QPT_SMI ||
 451                             qp->ibqp.qp_type == IB_QPT_GSI)
 452                                 atomic_dec(
 453                                  &ibah_to_rvtah(wqe->ud_wr.ah)->refcount);
 454                         if (++qp->s_last >= qp->s_size)
 455                                 qp->s_last = 0;
 456                 }
 457                 if (qp->s_rdma_mr) {
 458                         rvt_put_mr(qp->s_rdma_mr);
 459                         qp->s_rdma_mr = NULL;
 460                 }
 461         }
 462
 463         if (qp->ibqp.qp_type != IB_QPT_RC)
 464                 return;
 465
 466         for (n = 0; n < ARRAY_SIZE(qp->s_ack_queue); n++) {
 467                 struct rvt_ack_entry *e = &qp->s_ack_queue[n];
 468
 469                 if (e->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST &&
 470                     e->rdma_sge.mr) {
 471                         rvt_put_mr(e->rdma_sge.mr);
 472                         e->rdma_sge.mr = NULL;
 473                 }
 474         }
 475 }
 476
 477 /**
 478  * qib_error_qp - put a QP into the error state
 479  * @qp: the QP to put into the error state
 480  * @err: the receive completion error to signal if a RWQE is active
 481  *
 482  * Flushes both send and receive work queues.
 483  * Returns true if last WQE event should be generated.
 484  * The QP r_lock and s_lock should be held and interrupts disabled.
 485  * If we are already in error state, just return.
 486  */
 487 int qib_error_qp(struct rvt_qp *qp, enum ib_wc_status err)
 488 {
 489         struct qib_qp_priv *priv = qp->priv;
 490         struct qib_ibdev *dev = to_idev(qp->ibqp.device);
 491         struct ib_wc wc;
 492         int ret = 0;
 493
 494         if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
 495                 goto bail;
 496
 497         qp->state = IB_QPS_ERR;
 498
 499         if (qp->s_flags & (QIB_S_TIMER | QIB_S_WAIT_RNR)) {
 500                 qp->s_flags &= ~(QIB_S_TIMER | QIB_S_WAIT_RNR);
 501                 del_timer(&qp->s_timer);
 502         }
 503
 504         if (qp->s_flags & QIB_S_ANY_WAIT_SEND)
 505                 qp->s_flags &= ~QIB_S_ANY_WAIT_SEND;
 506
 507         spin_lock(&dev->rdi.pending_lock);
 508         if (!list_empty(&priv->iowait) && !(qp->s_flags & QIB_S_BUSY)) {
 509                 qp->s_flags &= ~QIB_S_ANY_WAIT_IO;
 510                 list_del_init(&priv->iowait);
 511         }
 512         spin_unlock(&dev->rdi.pending_lock);
 513
 514         if (!(qp->s_flags & QIB_S_BUSY)) {
 515                 qp->s_hdrwords = 0;
 516                 if (qp->s_rdma_mr) {
 517                         rvt_put_mr(qp->s_rdma_mr);
 518                         qp->s_rdma_mr = NULL;
 519                 }
 520                 if (priv->s_tx) {
 521                         qib_put_txreq(priv->s_tx);
 522                         priv->s_tx = NULL;
 523                 }
 524         }
 525
 526         /* Schedule the sending tasklet to drain the send work queue. */
 527         if (qp->s_last != qp->s_head)
 528                 qib_schedule_send(qp);
 529
 530         clear_mr_refs(qp, 0);
 531
 532         memset(&wc, 0, sizeof(wc));
 533         wc.qp = &qp->ibqp;
 534         wc.opcode = IB_WC_RECV;
 535
 536         if (test_and_clear_bit(QIB_R_WRID_VALID, &qp->r_aflags)) {
 537                 wc.wr_id = qp->r_wr_id;
 538                 wc.status = err;
 539                 qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 540         }
 541         wc.status = IB_WC_WR_FLUSH_ERR;
 542
 543         if (qp->r_rq.wq) {
 544                 struct rvt_rwq *wq;
 545                 u32 head;
 546                 u32 tail;
 547
 548                 spin_lock(&qp->r_rq.lock);
 549
 550                 /* sanity check pointers before trusting them */
 551                 wq = qp->r_rq.wq;
 552                 head = wq->head;
 553                 if (head >= qp->r_rq.size)
 554                         head = 0;
 555                 tail = wq->tail;
 556                 if (tail >= qp->r_rq.size)
 557                         tail = 0;
 558                 while (tail != head) {
 559                         wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
 560                         if (++tail >= qp->r_rq.size)
 561                                 tail = 0;
 562                         qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 563                 }
 564                 wq->tail = tail;
 565
 566                 spin_unlock(&qp->r_rq.lock);
 567         } else if (qp->ibqp.event_handler)
 568                 ret = 1;
 569
 570 bail:
 571         return ret;
 572 }
 573
 574 /**
 575  * qib_modify_qp - modify the attributes of a queue pair
 576  * @ibqp: the queue pair who's attributes we're modifying
 577  * @attr: the new attributes
 578  * @attr_mask: the mask of attributes to modify
 579  * @udata: user data for libibverbs.so
 580  *
 581  * Returns 0 on success, otherwise returns an errno.
 582  */
 583 int qib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 584                   int attr_mask, struct ib_udata *udata)
 585 {
 586         struct qib_ibdev *dev = to_idev(ibqp->device);
 587         struct rvt_qp *qp = to_iqp(ibqp);
 588         struct qib_qp_priv *priv = qp->priv;
 589         enum ib_qp_state cur_state, new_state;
 590         struct ib_event ev;
 591         int lastwqe = 0;
 592         int mig = 0;
 593         int ret;
 594         u32 pmtu = 0; /* for gcc warning only */
 595
 596         spin_lock_irq(&qp->r_lock);
 597         spin_lock(&qp->s_lock);
 598
 599         cur_state = attr_mask & IB_QP_CUR_STATE ?
 600                 attr->cur_qp_state : qp->state;
 601         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
 602
 603         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
 604                                 attr_mask, IB_LINK_LAYER_UNSPECIFIED))
 605                 goto inval;
 606
 607         if (attr_mask & IB_QP_AV) {
 608                 if (attr->ah_attr.dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE))
 609                         goto inval;
 610                 if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr))
 611                         goto inval;
 612         }
 613
 614         if (attr_mask & IB_QP_ALT_PATH) {
 615                 if (attr->alt_ah_attr.dlid >=
 616                     be16_to_cpu(IB_MULTICAST_LID_BASE))
 617                         goto inval;
 618                 if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
 619                         goto inval;
 620                 if (attr->alt_pkey_index >= qib_get_npkeys(dd_from_dev(dev)))
 621                         goto inval;
 622         }
 623
 624         if (attr_mask & IB_QP_PKEY_INDEX)
 625                 if (attr->pkey_index >= qib_get_npkeys(dd_from_dev(dev)))
 626                         goto inval;
 627
 628         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 629                 if (attr->min_rnr_timer > 31)
 630                         goto inval;
 631
 632         if (attr_mask & IB_QP_PORT)
 633                 if (qp->ibqp.qp_type == IB_QPT_SMI ||
 634                     qp->ibqp.qp_type == IB_QPT_GSI ||
 635                     attr->port_num == 0 ||
 636                     attr->port_num > ibqp->device->phys_port_cnt)
 637                         goto inval;
 638
 639         if (attr_mask & IB_QP_DEST_QPN)
 640                 if (attr->dest_qp_num > QIB_QPN_MASK)
 641                         goto inval;
 642
 643         if (attr_mask & IB_QP_RETRY_CNT)
 644                 if (attr->retry_cnt > 7)
 645                         goto inval;
 646
 647         if (attr_mask & IB_QP_RNR_RETRY)
 648                 if (attr->rnr_retry > 7)
 649                         goto inval;
 650
 651         /*
 652          * Don't allow invalid path_mtu values.  OK to set greater
 653          * than the active mtu (or even the max_cap, if we have tuned
 654          * that to a small mtu.  We'll set qp->path_mtu
 655          * to the lesser of requested attribute mtu and active,
 656          * for packetizing messages.
 657          * Note that the QP port has to be set in INIT and MTU in RTR.
 658          */
 659         if (attr_mask & IB_QP_PATH_MTU) {
 660                 struct qib_devdata *dd = dd_from_dev(dev);
 661                 int mtu, pidx = qp->port_num - 1;
 662
 663                 mtu = ib_mtu_enum_to_int(attr->path_mtu);
 664                 if (mtu == -1)
 665                         goto inval;
 666                 if (mtu > dd->pport[pidx].ibmtu) {
 667                         switch (dd->pport[pidx].ibmtu) {
 668                         case 4096:
 669                                 pmtu = IB_MTU_4096;
 670                                 break;
 671                         case 2048:
 672                                 pmtu = IB_MTU_2048;
 673                                 break;
 674                         case 1024:
 675                                 pmtu = IB_MTU_1024;
 676                                 break;
 677                         case 512:
 678                                 pmtu = IB_MTU_512;
 679                                 break;
 680                         case 256:
 681                                 pmtu = IB_MTU_256;
 682                                 break;
 683                         default:
 684                                 pmtu = IB_MTU_2048;
 685                         }
 686                 } else
 687                         pmtu = attr->path_mtu;
 688         }
 689
 690         if (attr_mask & IB_QP_PATH_MIG_STATE) {
 691                 if (attr->path_mig_state == IB_MIG_REARM) {
 692                         if (qp->s_mig_state == IB_MIG_ARMED)
 693                                 goto inval;
 694                         if (new_state != IB_QPS_RTS)
 695                                 goto inval;
 696                 } else if (attr->path_mig_state == IB_MIG_MIGRATED) {
 697                         if (qp->s_mig_state == IB_MIG_REARM)
 698                                 goto inval;
 699                         if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
 700                                 goto inval;
 701                         if (qp->s_mig_state == IB_MIG_ARMED)
 702                                 mig = 1;
 703                 } else
 704                         goto inval;
 705         }
 706
 707         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 708                 if (attr->max_dest_rd_atomic > QIB_MAX_RDMA_ATOMIC)
 709                         goto inval;
 710
 711         switch (new_state) {
 712         case IB_QPS_RESET:
 713                 if (qp->state != IB_QPS_RESET) {
 714                         qp->state = IB_QPS_RESET;
 715                         spin_lock(&dev->rdi.pending_lock);
 716                         if (!list_empty(&priv->iowait))
 717                                 list_del_init(&priv->iowait);
 718                         spin_unlock(&dev->rdi.pending_lock);
 719                         qp->s_flags &= ~(QIB_S_TIMER | QIB_S_ANY_WAIT);
 720                         spin_unlock(&qp->s_lock);
 721                         spin_unlock_irq(&qp->r_lock);
 722                         /* Stop the sending work queue and retry timer */
 723                         cancel_work_sync(&priv->s_work);
 724                         del_timer_sync(&qp->s_timer);
 725                         wait_event(priv->wait_dma,
 726                                    !atomic_read(&priv->s_dma_busy));
 727                         if (priv->s_tx) {
 728                                 qib_put_txreq(priv->s_tx);
 729                                 priv->s_tx = NULL;
 730                         }
 731                         remove_qp(dev, qp);
 732                         wait_event(qp->wait, !atomic_read(&qp->refcount));
 733                         spin_lock_irq(&qp->r_lock);
 734                         spin_lock(&qp->s_lock);
 735                         clear_mr_refs(qp, 1);
 736                         qib_reset_qp(qp, ibqp->qp_type);
 737                 }
 738                 break;
 739
 740         case IB_QPS_RTR:
 741                 /* Allow event to retrigger if QP set to RTR more than once */
 742                 qp->r_flags &= ~QIB_R_COMM_EST;
 743                 qp->state = new_state;
 744                 break;
 745
 746         case IB_QPS_SQD:
 747                 qp->s_draining = qp->s_last != qp->s_cur;
 748                 qp->state = new_state;
 749                 break;
 750
 751         case IB_QPS_SQE:
 752                 if (qp->ibqp.qp_type == IB_QPT_RC)
 753                         goto inval;
 754                 qp->state = new_state;
 755                 break;
 756
 757         case IB_QPS_ERR:
 758                 lastwqe = qib_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 759                 break;
 760
 761         default:
 762                 qp->state = new_state;
 763                 break;
 764         }
 765
 766         if (attr_mask & IB_QP_PKEY_INDEX)
 767                 qp->s_pkey_index = attr->pkey_index;
 768
 769         if (attr_mask & IB_QP_PORT)
 770                 qp->port_num = attr->port_num;
 771
 772         if (attr_mask & IB_QP_DEST_QPN)
 773                 qp->remote_qpn = attr->dest_qp_num;
 774
 775         if (attr_mask & IB_QP_SQ_PSN) {
 776                 qp->s_next_psn = attr->sq_psn & QIB_PSN_MASK;
 777                 qp->s_psn = qp->s_next_psn;
 778                 qp->s_sending_psn = qp->s_next_psn;
 779                 qp->s_last_psn = qp->s_next_psn - 1;
 780                 qp->s_sending_hpsn = qp->s_last_psn;
 781         }
 782
 783         if (attr_mask & IB_QP_RQ_PSN)
 784                 qp->r_psn = attr->rq_psn & QIB_PSN_MASK;
 785
 786         if (attr_mask & IB_QP_ACCESS_FLAGS)
 787                 qp->qp_access_flags = attr->qp_access_flags;
 788
 789         if (attr_mask & IB_QP_AV) {
 790                 qp->remote_ah_attr = attr->ah_attr;
 791                 qp->s_srate = attr->ah_attr.static_rate;
 792         }
 793
 794         if (attr_mask & IB_QP_ALT_PATH) {
 795                 qp->alt_ah_attr = attr->alt_ah_attr;
 796                 qp->s_alt_pkey_index = attr->alt_pkey_index;
 797         }
 798
 799         if (attr_mask & IB_QP_PATH_MIG_STATE) {
 800                 qp->s_mig_state = attr->path_mig_state;
 801                 if (mig) {
 802                         qp->remote_ah_attr = qp->alt_ah_attr;
 803                         qp->port_num = qp->alt_ah_attr.port_num;
 804                         qp->s_pkey_index = qp->s_alt_pkey_index;
 805                 }
 806         }
 807
 808         if (attr_mask & IB_QP_PATH_MTU) {
 809                 qp->path_mtu = pmtu;
 810                 qp->pmtu = ib_mtu_enum_to_int(pmtu);
 811         }
 812
 813         if (attr_mask & IB_QP_RETRY_CNT) {
 814                 qp->s_retry_cnt = attr->retry_cnt;
 815                 qp->s_retry = attr->retry_cnt;
 816         }
 817
 818         if (attr_mask & IB_QP_RNR_RETRY) {
 819                 qp->s_rnr_retry_cnt = attr->rnr_retry;
 820                 qp->s_rnr_retry = attr->rnr_retry;
 821         }
 822
 823         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 824                 qp->r_min_rnr_timer = attr->min_rnr_timer;
 825
 826         if (attr_mask & IB_QP_TIMEOUT) {
 827                 qp->timeout = attr->timeout;
 828                 qp->timeout_jiffies =
 829                         usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
 830                                 1000UL);
 831         }
 832
 833         if (attr_mask & IB_QP_QKEY)
 834                 qp->qkey = attr->qkey;
 835
 836         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 837                 qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
 838
 839         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
 840                 qp->s_max_rd_atomic = attr->max_rd_atomic;
 841
 842         spin_unlock(&qp->s_lock);
 843         spin_unlock_irq(&qp->r_lock);
 844
 845         if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
 846                 insert_qp(dev, qp);
 847
 848         if (lastwqe) {
 849                 ev.device = qp->ibqp.device;
 850                 ev.element.qp = &qp->ibqp;
 851                 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 852                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 853         }
 854         if (mig) {
 855                 ev.device = qp->ibqp.device;
 856                 ev.element.qp = &qp->ibqp;
 857                 ev.event = IB_EVENT_PATH_MIG;
 858                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 859         }
 860         ret = 0;
 861         goto bail;
 862
 863 inval:
 864         spin_unlock(&qp->s_lock);
 865         spin_unlock_irq(&qp->r_lock);
 866         ret = -EINVAL;
 867
 868 bail:
 869         return ret;
 870 }
 871
 872 int qib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 873                  int attr_mask, struct ib_qp_init_attr *init_attr)
 874 {
 875         struct rvt_qp *qp = to_iqp(ibqp);
 876
 877         attr->qp_state = qp->state;
 878         attr->cur_qp_state = attr->qp_state;
 879         attr->path_mtu = qp->path_mtu;
 880         attr->path_mig_state = qp->s_mig_state;
 881         attr->qkey = qp->qkey;
 882         attr->rq_psn = qp->r_psn & QIB_PSN_MASK;
 883         attr->sq_psn = qp->s_next_psn & QIB_PSN_MASK;
 884         attr->dest_qp_num = qp->remote_qpn;
 885         attr->qp_access_flags = qp->qp_access_flags;
 886         attr->cap.max_send_wr = qp->s_size - 1;
 887         attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
 888         attr->cap.max_send_sge = qp->s_max_sge;
 889         attr->cap.max_recv_sge = qp->r_rq.max_sge;
 890         attr->cap.max_inline_data = 0;
 891         attr->ah_attr = qp->remote_ah_attr;
 892         attr->alt_ah_attr = qp->alt_ah_attr;
 893         attr->pkey_index = qp->s_pkey_index;
 894         attr->alt_pkey_index = qp->s_alt_pkey_index;
 895         attr->en_sqd_async_notify = 0;
 896         attr->sq_draining = qp->s_draining;
 897         attr->max_rd_atomic = qp->s_max_rd_atomic;
 898         attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
 899         attr->min_rnr_timer = qp->r_min_rnr_timer;
 900         attr->port_num = qp->port_num;
 901         attr->timeout = qp->timeout;
 902         attr->retry_cnt = qp->s_retry_cnt;
 903         attr->rnr_retry = qp->s_rnr_retry_cnt;
 904         attr->alt_port_num = qp->alt_ah_attr.port_num;
 905         attr->alt_timeout = qp->alt_timeout;
 906
 907         init_attr->event_handler = qp->ibqp.event_handler;
 908         init_attr->qp_context = qp->ibqp.qp_context;
 909         init_attr->send_cq = qp->ibqp.send_cq;
 910         init_attr->recv_cq = qp->ibqp.recv_cq;
 911         init_attr->srq = qp->ibqp.srq;
 912         init_attr->cap = attr->cap;
 913         if (qp->s_flags & QIB_S_SIGNAL_REQ_WR)
 914                 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 915         else
 916                 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
 917         init_attr->qp_type = qp->ibqp.qp_type;
 918         init_attr->port_num = qp->port_num;
 919         return 0;
 920 }
 921
 922 /**
 923  * qib_compute_aeth - compute the AETH (syndrome + MSN)
 924  * @qp: the queue pair to compute the AETH for
 925  *
 926  * Returns the AETH.
 927  */
 928 __be32 qib_compute_aeth(struct rvt_qp *qp)
 929 {
 930         u32 aeth = qp->r_msn & QIB_MSN_MASK;
 931
 932         if (qp->ibqp.srq) {
 933                 /*
 934                  * Shared receive queues don't generate credits.
 935                  * Set the credit field to the invalid value.
 936                  */
 937                 aeth |= QIB_AETH_CREDIT_INVAL << QIB_AETH_CREDIT_SHIFT;
 938         } else {
 939                 u32 min, max, x;
 940                 u32 credits;
 941                 struct rvt_rwq *wq = qp->r_rq.wq;
 942                 u32 head;
 943                 u32 tail;
 944
 945                 /* sanity check pointers before trusting them */
 946                 head = wq->head;
 947                 if (head >= qp->r_rq.size)
 948                         head = 0;
 949                 tail = wq->tail;
 950                 if (tail >= qp->r_rq.size)
 951                         tail = 0;
 952                 /*
 953                  * Compute the number of credits available (RWQEs).
 954                  * XXX Not holding the r_rq.lock here so there is a small
 955                  * chance that the pair of reads are not atomic.
 956                  */
 957                 credits = head - tail;
 958                 if ((int)credits < 0)
 959                         credits += qp->r_rq.size;
 960                 /*
 961                  * Binary search the credit table to find the code to
 962                  * use.
 963                  */
 964                 min = 0;
 965                 max = 31;
 966                 for (;;) {
 967                         x = (min + max) / 2;
 968                         if (credit_table[x] == credits)
 969                                 break;
 970                         if (credit_table[x] > credits)
 971                                 max = x;
 972                         else if (min == x)
 973                                 break;
 974                         else
 975                                 min = x;
 976                 }
 977                 aeth |= x << QIB_AETH_CREDIT_SHIFT;
 978         }
 979         return cpu_to_be32(aeth);
 980 }
 981
 982 /**
 983  * qib_create_qp - create a queue pair for a device
 984  * @ibpd: the protection domain who's device we create the queue pair for
 985  * @init_attr: the attributes of the queue pair
 986  * @udata: user data for libibverbs.so
 987  *
 988  * Returns the queue pair on success, otherwise returns an errno.
 989  *
 990  * Called by the ib_create_qp() core verbs function.
 991  */
 992 struct ib_qp *qib_create_qp(struct ib_pd *ibpd,
 993                             struct ib_qp_init_attr *init_attr,
 994                             struct ib_udata *udata)
 995 {
 996         struct rvt_qp *qp;
 997         int err;
 998         struct rvt_swqe *swq = NULL;
 999         struct qib_ibdev *dev;
1000         struct qib_devdata *dd;
1001         size_t sz;
1002         size_t sg_list_sz;
1003         struct ib_qp *ret;
1004         gfp_t gfp;
1005         struct qib_qp_priv *priv;
1006
1007         if (init_attr->cap.max_send_sge > ib_qib_max_sges ||
1008             init_attr->cap.max_send_wr > ib_qib_max_qp_wrs ||
1009             init_attr->create_flags & ~(IB_QP_CREATE_USE_GFP_NOIO))
1010                 return ERR_PTR(-EINVAL);
1011
1012         /* GFP_NOIO is applicable in RC QPs only */
1013         if (init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO &&
1014             init_attr->qp_type != IB_QPT_RC)
1015                 return ERR_PTR(-EINVAL);
1016
1017         gfp = init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO ?
1018                         GFP_NOIO : GFP_KERNEL;
1019
1020         /* Check receive queue parameters if no SRQ is specified. */
1021         if (!init_attr->srq) {
1022                 if (init_attr->cap.max_recv_sge > ib_qib_max_sges ||
1023                     init_attr->cap.max_recv_wr > ib_qib_max_qp_wrs) {
1024                         ret = ERR_PTR(-EINVAL);
1025                         goto bail;
1026                 }
1027                 if (init_attr->cap.max_send_sge +
1028                     init_attr->cap.max_send_wr +
1029                     init_attr->cap.max_recv_sge +
1030                     init_attr->cap.max_recv_wr == 0) {
1031                         ret = ERR_PTR(-EINVAL);
1032                         goto bail;
1033                 }
1034         }
1035
1036         switch (init_attr->qp_type) {
1037         case IB_QPT_SMI:
1038         case IB_QPT_GSI:
1039                 if (init_attr->port_num == 0 ||
1040                     init_attr->port_num > ibpd->device->phys_port_cnt) {
1041                         ret = ERR_PTR(-EINVAL);
1042                         goto bail;
1043                 }
1044         case IB_QPT_UC:
1045         case IB_QPT_RC:
1046         case IB_QPT_UD:
1047                 sz = sizeof(struct rvt_sge) *
1048                         init_attr->cap.max_send_sge +
1049                         sizeof(struct rvt_swqe);
1050                 swq = __vmalloc((init_attr->cap.max_send_wr + 1) * sz,
1051                                 gfp, PAGE_KERNEL);
1052                 if (swq == NULL) {
1053                         ret = ERR_PTR(-ENOMEM);
1054                         goto bail;
1055                 }
1056                 sz = sizeof(*qp);
1057                 sg_list_sz = 0;
1058                 if (init_attr->srq) {
1059                         struct rvt_srq *srq = ibsrq_to_rvtsrq(init_attr->srq);
1060
1061                         if (srq->rq.max_sge > 1)
1062                                 sg_list_sz = sizeof(*qp->r_sg_list) *
1063                                         (srq->rq.max_sge - 1);
1064                 } else if (init_attr->cap.max_recv_sge > 1)
1065                         sg_list_sz = sizeof(*qp->r_sg_list) *
1066                                 (init_attr->cap.max_recv_sge - 1);
1067                 qp = kzalloc(sz + sg_list_sz, gfp);
1068                 if (!qp) {
1069                         ret = ERR_PTR(-ENOMEM);
1070                         goto bail_swq;
1071                 }
1072                 RCU_INIT_POINTER(qp->next, NULL);
1073                 priv = kzalloc(sizeof(*priv), gfp);
1074                 if (!priv) {
1075                         ret = ERR_PTR(-ENOMEM);
1076                         goto bail_qp_hdr;
1077                 }
1078                 priv->owner = qp;
1079                 priv->s_hdr = kzalloc(sizeof(*priv->s_hdr), gfp);
1080                 if (!priv->s_hdr) {
1081                         ret = ERR_PTR(-ENOMEM);
1082                         goto bail_qp;
1083                 }
1084                 qp->priv = priv;
1085                 qp->timeout_jiffies =
1086                         usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
1087                                 1000UL);
1088                 if (init_attr->srq)
1089                         sz = 0;
1090                 else {
1091                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
1092                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
1093                         sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
1094                                 sizeof(struct rvt_rwqe);
1095                         if (gfp != GFP_NOIO)
1096                                 qp->r_rq.wq = vmalloc_user(
1097                                                 sizeof(struct rvt_rwq) +
1098                                                 qp->r_rq.size * sz);
1099                         else
1100                                 qp->r_rq.wq = __vmalloc(
1101                                                 sizeof(struct rvt_rwq) +
1102                                                 qp->r_rq.size * sz,
1103                                                 gfp, PAGE_KERNEL);
1104
1105                         if (!qp->r_rq.wq) {
1106                                 ret = ERR_PTR(-ENOMEM);
1107                                 goto bail_qp;
1108                         }
1109                 }
1110
1111                 /*
1112                  * ib_create_qp() will initialize qp->ibqp
1113                  * except for qp->ibqp.qp_num.
1114                  */
1115                 spin_lock_init(&qp->r_lock);
1116                 spin_lock_init(&qp->s_lock);
1117                 spin_lock_init(&qp->r_rq.lock);
1118                 atomic_set(&qp->refcount, 0);
1119                 init_waitqueue_head(&qp->wait);
1120                 init_waitqueue_head(&priv->wait_dma);
1121                 init_timer(&qp->s_timer);
1122                 qp->s_timer.data = (unsigned long)qp;
1123                 INIT_WORK(&priv->s_work, qib_do_send);
1124                 INIT_LIST_HEAD(&priv->iowait);
1125                 INIT_LIST_HEAD(&qp->rspwait);
1126                 qp->state = IB_QPS_RESET;
1127                 qp->s_wq = swq;
1128                 qp->s_size = init_attr->cap.max_send_wr + 1;
1129                 qp->s_max_sge = init_attr->cap.max_send_sge;
1130                 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
1131                         qp->s_flags = QIB_S_SIGNAL_REQ_WR;
1132                 dev = to_idev(ibpd->device);
1133                 dd = dd_from_dev(dev);
1134                 err = alloc_qpn(dd, &dev->rdi.qp_dev->qpn_table,
1135                                 init_attr->qp_type, init_attr->port_num, gfp);
1136                 if (err < 0) {
1137                         ret = ERR_PTR(err);
1138                         vfree(qp->r_rq.wq);
1139                         goto bail_qp;
1140                 }
1141                 qp->ibqp.qp_num = err;
1142                 qp->port_num = init_attr->port_num;
1143                 qib_reset_qp(qp, init_attr->qp_type);
1144                 break;
1145
1146         default:
1147                 /* Don't support raw QPs */
1148                 ret = ERR_PTR(-ENOSYS);
1149                 goto bail;
1150         }
1151
1152         init_attr->cap.max_inline_data = 0;
1153
1154         /*
1155          * Return the address of the RWQ as the offset to mmap.
1156          * See qib_mmap() for details.
1157          */
1158         if (udata && udata->outlen >= sizeof(__u64)) {
1159                 if (!qp->r_rq.wq) {
1160                         __u64 offset = 0;
1161
1162                         err = ib_copy_to_udata(udata, &offset,
1163                                                sizeof(offset));
1164                         if (err) {
1165                                 ret = ERR_PTR(err);
1166                                 goto bail_ip;
1167                         }
1168                 } else {
1169                         u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz;
1170
1171                         qp->ip = rvt_create_mmap_info(&dev->rdi, s,
1172                                                       ibpd->uobject->context,
1173                                                       qp->r_rq.wq);
1174                         if (!qp->ip) {
1175                                 ret = ERR_PTR(-ENOMEM);
1176                                 goto bail_ip;
1177                         }
1178
1179                         err = ib_copy_to_udata(udata, &(qp->ip->offset),
1180                                                sizeof(qp->ip->offset));
1181                         if (err) {
1182                                 ret = ERR_PTR(err);
1183                                 goto bail_ip;
1184                         }
1185                 }
1186         }
1187
1188         spin_lock(&dev->n_qps_lock);
1189         if (dev->n_qps_allocated == ib_qib_max_qps) {
1190                 spin_unlock(&dev->n_qps_lock);
1191                 ret = ERR_PTR(-ENOMEM);
1192                 goto bail_ip;
1193         }
1194
1195         dev->n_qps_allocated++;
1196         spin_unlock(&dev->n_qps_lock);
1197
1198         if (qp->ip) {
1199                 spin_lock_irq(&dev->rdi.pending_lock);
1200                 list_add(&qp->ip->pending_mmaps, &dev->rdi.pending_mmaps);
1201                 spin_unlock_irq(&dev->rdi.pending_lock);
1202         }
1203
1204         ret = &qp->ibqp;
1205         goto bail;
1206
1207 bail_ip:
1208         if (qp->ip)
1209                 kref_put(&qp->ip->ref, rvt_release_mmap_info);
1210         else
1211                 vfree(qp->r_rq.wq);
1212         free_qpn(&dev->rdi.qp_dev->qpn_table, qp->ibqp.qp_num);
1213 bail_qp:
1214         kfree(priv->s_hdr);
1215         kfree(priv);
1216 bail_qp_hdr:
1217         kfree(qp);
1218 bail_swq:
1219         vfree(swq);
1220 bail:
1221         return ret;
1222 }
1223
1224 /**
1225  * qib_destroy_qp - destroy a queue pair
1226  * @ibqp: the queue pair to destroy
1227  *
1228  * Returns 0 on success.
1229  *
1230  * Note that this can be called while the QP is actively sending or
1231  * receiving!
1232  */
1233 int qib_destroy_qp(struct ib_qp *ibqp)
1234 {
1235         struct rvt_qp *qp = to_iqp(ibqp);
1236         struct qib_ibdev *dev = to_idev(ibqp->device);
1237         struct qib_qp_priv *priv = qp->priv;
1238
1239         /* Make sure HW and driver activity is stopped. */
1240         spin_lock_irq(&qp->s_lock);
1241         if (qp->state != IB_QPS_RESET) {
1242                 qp->state = IB_QPS_RESET;
1243                 spin_lock(&dev->rdi.pending_lock);
1244                 if (!list_empty(&priv->iowait))
1245                         list_del_init(&priv->iowait);
1246                 spin_unlock(&dev->rdi.pending_lock);
1247                 qp->s_flags &= ~(QIB_S_TIMER | QIB_S_ANY_WAIT);
1248                 spin_unlock_irq(&qp->s_lock);
1249                 cancel_work_sync(&priv->s_work);
1250                 del_timer_sync(&qp->s_timer);
1251                 wait_event(priv->wait_dma, !atomic_read(&priv->s_dma_busy));
1252                 if (priv->s_tx) {
1253                         qib_put_txreq(priv->s_tx);
1254                         priv->s_tx = NULL;
1255                 }
1256                 remove_qp(dev, qp);
1257                 wait_event(qp->wait, !atomic_read(&qp->refcount));
1258                 clear_mr_refs(qp, 1);
1259         } else
1260                 spin_unlock_irq(&qp->s_lock);
1261
1262         /* all user's cleaned up, mark it available */
1263         free_qpn(&dev->rdi.qp_dev->qpn_table, qp->ibqp.qp_num);
1264         spin_lock(&dev->n_qps_lock);
1265         dev->n_qps_allocated--;
1266         spin_unlock(&dev->n_qps_lock);
1267
1268         if (qp->ip)
1269                 kref_put(&qp->ip->ref, rvt_release_mmap_info);
1270         else
1271                 vfree(qp->r_rq.wq);
1272         vfree(qp->s_wq);
1273         kfree(priv->s_hdr);
1274         kfree(priv);
1275         kfree(qp);
1276         return 0;
1277 }
1278
1279 /**
1280  * qib_init_qpn_table - initialize the QP number table for a device
1281  * @qpt: the QPN table
1282  */
1283 void qib_init_qpn_table(struct qib_devdata *dd, struct rvt_qpn_table *qpt)
1284 {
1285         spin_lock_init(&qpt->lock);
1286         qpt->last = 1;          /* start with QPN 2 */
1287         qpt->nmaps = 1;
1288         qpt_mask = dd->qpn_mask;
1289 }
1290
1291 /**
1292  * qib_free_qpn_table - free the QP number table for a device
1293  * @qpt: the QPN table
1294  */
1295 void qib_free_qpn_table(struct rvt_qpn_table *qpt)
1296 {
1297         int i;
1298
1299         for (i = 0; i < ARRAY_SIZE(qpt->map); i++)
1300                 if (qpt->map[i].page)
1301                         free_page((unsigned long) qpt->map[i].page);
1302 }
1303
1304 /**
1305  * qib_get_credit - flush the send work queue of a QP
1306  * @qp: the qp who's send work queue to flush
1307  * @aeth: the Acknowledge Extended Transport Header
1308  *
1309  * The QP s_lock should be held.
1310  */
1311 void qib_get_credit(struct rvt_qp *qp, u32 aeth)
1312 {
1313         u32 credit = (aeth >> QIB_AETH_CREDIT_SHIFT) & QIB_AETH_CREDIT_MASK;
1314
1315         /*
1316          * If the credit is invalid, we can send
1317          * as many packets as we like.  Otherwise, we have to
1318          * honor the credit field.
1319          */
1320         if (credit == QIB_AETH_CREDIT_INVAL) {
1321                 if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT)) {
1322                         qp->s_flags |= QIB_S_UNLIMITED_CREDIT;
1323                         if (qp->s_flags & QIB_S_WAIT_SSN_CREDIT) {
1324                                 qp->s_flags &= ~QIB_S_WAIT_SSN_CREDIT;
1325                                 qib_schedule_send(qp);
1326                         }
1327                 }
1328         } else if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT)) {
1329                 /* Compute new LSN (i.e., MSN + credit) */
1330                 credit = (aeth + credit_table[credit]) & QIB_MSN_MASK;
1331                 if (qib_cmp24(credit, qp->s_lsn) > 0) {
1332                         qp->s_lsn = credit;
1333                         if (qp->s_flags & QIB_S_WAIT_SSN_CREDIT) {
1334                                 qp->s_flags &= ~QIB_S_WAIT_SSN_CREDIT;
1335                                 qib_schedule_send(qp);
1336                         }
1337                 }
1338         }
1339 }
1340
1341 #ifdef CONFIG_DEBUG_FS
1342
1343 struct qib_qp_iter {
1344         struct qib_ibdev *dev;
1345         struct rvt_qp *qp;
1346         int n;
1347 };
1348
1349 struct qib_qp_iter *qib_qp_iter_init(struct qib_ibdev *dev)
1350 {
1351         struct qib_qp_iter *iter;
1352
1353         iter = kzalloc(sizeof(*iter), GFP_KERNEL);
1354         if (!iter)
1355                 return NULL;
1356
1357         iter->dev = dev;
1358         if (qib_qp_iter_next(iter)) {
1359                 kfree(iter);
1360                 return NULL;
1361         }
1362
1363         return iter;
1364 }
1365
1366 int qib_qp_iter_next(struct qib_qp_iter *iter)
1367 {
1368         struct qib_ibdev *dev = iter->dev;
1369         int n = iter->n;
1370         int ret = 1;
1371         struct rvt_qp *pqp = iter->qp;
1372         struct rvt_qp *qp;
1373
1374         for (; n < dev->rdi.qp_dev->qp_table_size; n++) {
1375                 if (pqp)
1376                         qp = rcu_dereference(pqp->next);
1377                 else
1378                         qp = rcu_dereference(dev->rdi.qp_dev->qp_table[n]);
1379                 pqp = qp;
1380                 if (qp) {
1381                         iter->qp = qp;
1382                         iter->n = n;
1383                         return 0;
1384                 }
1385         }
1386         return ret;
1387 }
1388
1389 static const char * const qp_type_str[] = {
1390         "SMI", "GSI", "RC", "UC", "UD",
1391 };
1392
1393 void qib_qp_iter_print(struct seq_file *s, struct qib_qp_iter *iter)
1394 {
1395         struct rvt_swqe *wqe;
1396         struct rvt_qp *qp = iter->qp;
1397         struct qib_qp_priv *priv = qp->priv;
1398
1399         wqe = get_swqe_ptr(qp, qp->s_last);
1400         seq_printf(s,
1401                    "N %d QP%u %s %u %u %u f=%x %u %u %u %u %u PSN %x %x %x %x %x (%u %u %u %u %u %u) QP%u LID %x\n",
1402                    iter->n,
1403                    qp->ibqp.qp_num,
1404                    qp_type_str[qp->ibqp.qp_type],
1405                    qp->state,
1406                    wqe->wr.opcode,
1407                    qp->s_hdrwords,
1408                    qp->s_flags,
1409                    atomic_read(&priv->s_dma_busy),
1410                    !list_empty(&priv->iowait),
1411                    qp->timeout,
1412                    wqe->ssn,
1413                    qp->s_lsn,
1414                    qp->s_last_psn,
1415                    qp->s_psn, qp->s_next_psn,
1416                    qp->s_sending_psn, qp->s_sending_hpsn,
1417                    qp->s_last, qp->s_acked, qp->s_cur,
1418                    qp->s_tail, qp->s_head, qp->s_size,
1419                    qp->remote_qpn,
1420                    qp->remote_ah_attr.dlid);
1421 }
1422
1423 #endif