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 #define BITS_PER_PAGE           (PAGE_SIZE*BITS_PER_BYTE)
  46 #define BITS_PER_PAGE_MASK      (BITS_PER_PAGE-1)
  47
  48 static inline unsigned mk_qpn(struct qib_qpn_table *qpt,
  49                               struct qpn_map *map, unsigned off)
  50 {
  51         return (map - qpt->map) * BITS_PER_PAGE + off;
  52 }
  53
  54 static inline unsigned find_next_offset(struct qib_qpn_table *qpt,
  55                                         struct qpn_map *map, unsigned off,
  56                                         unsigned n)
  57 {
  58         if (qpt->mask) {
  59                 off++;
  60                 if (((off & qpt->mask) >> 1) >= n)
  61                         off = (off | qpt->mask) + 2;
  62         } else
  63                 off = find_next_zero_bit(map->page, BITS_PER_PAGE, off);
  64         return off;
  65 }
  66
  67 /*
  68  * Convert the AETH credit code into the number of credits.
  69  */
  70 static u32 credit_table[31] = {
  71         0,                      /* 0 */
  72         1,                      /* 1 */
  73         2,                      /* 2 */
  74         3,                      /* 3 */
  75         4,                      /* 4 */
  76         6,                      /* 5 */
  77         8,                      /* 6 */
  78         12,                     /* 7 */
  79         16,                     /* 8 */
  80         24,                     /* 9 */
  81         32,                     /* A */
  82         48,                     /* B */
  83         64,                     /* C */
  84         96,                     /* D */
  85         128,                    /* E */
  86         192,                    /* F */
  87         256,                    /* 10 */
  88         384,                    /* 11 */
  89         512,                    /* 12 */
  90         768,                    /* 13 */
  91         1024,                   /* 14 */
  92         1536,                   /* 15 */
  93         2048,                   /* 16 */
  94         3072,                   /* 17 */
  95         4096,                   /* 18 */
  96         6144,                   /* 19 */
  97         8192,                   /* 1A */
  98         12288,                  /* 1B */
  99         16384,                  /* 1C */
 100         24576,                  /* 1D */
 101         32768                   /* 1E */
 102 };
 103
 104 static void get_map_page(struct qib_qpn_table *qpt, struct qpn_map *map,
 105                          gfp_t gfp)
 106 {
 107         unsigned long page = get_zeroed_page(gfp);
 108
 109         /*
 110          * Free the page if someone raced with us installing it.
 111          */
 112
 113         spin_lock(&qpt->lock);
 114         if (map->page)
 115                 free_page(page);
 116         else
 117                 map->page = (void *)page;
 118         spin_unlock(&qpt->lock);
 119 }
 120
 121 /*
 122  * Allocate the next available QPN or
 123  * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI.
 124  */
 125 static int alloc_qpn(struct qib_devdata *dd, struct qib_qpn_table *qpt,
 126                      enum ib_qp_type type, u8 port, gfp_t gfp)
 127 {
 128         u32 i, offset, max_scan, qpn;
 129         struct qpn_map *map;
 130         u32 ret;
 131
 132         if (type == IB_QPT_SMI || type == IB_QPT_GSI) {
 133                 unsigned n;
 134
 135                 ret = type == IB_QPT_GSI;
 136                 n = 1 << (ret + 2 * (port - 1));
 137                 spin_lock(&qpt->lock);
 138                 if (qpt->flags & n)
 139                         ret = -EINVAL;
 140                 else
 141                         qpt->flags |= n;
 142                 spin_unlock(&qpt->lock);
 143                 goto bail;
 144         }
 145
 146         qpn = qpt->last + 2;
 147         if (qpn >= QPN_MAX)
 148                 qpn = 2;
 149         if (qpt->mask && ((qpn & qpt->mask) >> 1) >= dd->n_krcv_queues)
 150                 qpn = (qpn | qpt->mask) + 2;
 151         offset = qpn & BITS_PER_PAGE_MASK;
 152         map = &qpt->map[qpn / BITS_PER_PAGE];
 153         max_scan = qpt->nmaps - !offset;
 154         for (i = 0;;) {
 155                 if (unlikely(!map->page)) {
 156                         get_map_page(qpt, map, gfp);
 157                         if (unlikely(!map->page))
 158                                 break;
 159                 }
 160                 do {
 161                         if (!test_and_set_bit(offset, map->page)) {
 162                                 qpt->last = qpn;
 163                                 ret = qpn;
 164                                 goto bail;
 165                         }
 166                         offset = find_next_offset(qpt, map, offset,
 167                                 dd->n_krcv_queues);
 168                         qpn = mk_qpn(qpt, map, offset);
 169                         /*
 170                          * This test differs from alloc_pidmap().
 171                          * If find_next_offset() does find a zero
 172                          * bit, we don't need to check for QPN
 173                          * wrapping around past our starting QPN.
 174                          * We just need to be sure we don't loop
 175                          * forever.
 176                          */
 177                 } while (offset < BITS_PER_PAGE && qpn < QPN_MAX);
 178                 /*
 179                  * In order to keep the number of pages allocated to a
 180                  * minimum, we scan the all existing pages before increasing
 181                  * the size of the bitmap table.
 182                  */
 183                 if (++i > max_scan) {
 184                         if (qpt->nmaps == QPNMAP_ENTRIES)
 185                                 break;
 186                         map = &qpt->map[qpt->nmaps++];
 187                         offset = 0;
 188                 } else if (map < &qpt->map[qpt->nmaps]) {
 189                         ++map;
 190                         offset = 0;
 191                 } else {
 192                         map = &qpt->map[0];
 193                         offset = 2;
 194                 }
 195                 qpn = mk_qpn(qpt, map, offset);
 196         }
 197
 198         ret = -ENOMEM;
 199
 200 bail:
 201         return ret;
 202 }
 203
 204 static void free_qpn(struct qib_qpn_table *qpt, u32 qpn)
 205 {
 206         struct qpn_map *map;
 207
 208         map = qpt->map + qpn / BITS_PER_PAGE;
 209         if (map->page)
 210                 clear_bit(qpn & BITS_PER_PAGE_MASK, map->page);
 211 }
 212
 213 static inline unsigned qpn_hash(struct qib_ibdev *dev, u32 qpn)
 214 {
 215         return jhash_1word(qpn, dev->qp_rnd) &
 216                 (dev->qp_table_size - 1);
 217 }
 218
 219
 220 /*
 221  * Put the QP into the hash table.
 222  * The hash table holds a reference to the QP.
 223  */
 224 static void insert_qp(struct qib_ibdev *dev, struct qib_qp *qp)
 225 {
 226         struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 227         unsigned long flags;
 228         unsigned n = qpn_hash(dev, qp->ibqp.qp_num);
 229
 230         atomic_inc(&qp->refcount);
 231         spin_lock_irqsave(&dev->qpt_lock, flags);
 232
 233         if (qp->ibqp.qp_num == 0)
 234                 rcu_assign_pointer(ibp->qp0, qp);
 235         else if (qp->ibqp.qp_num == 1)
 236                 rcu_assign_pointer(ibp->qp1, qp);
 237         else {
 238                 qp->next = dev->qp_table[n];
 239                 rcu_assign_pointer(dev->qp_table[n], qp);
 240         }
 241
 242         spin_unlock_irqrestore(&dev->qpt_lock, flags);
 243 }
 244
 245 /*
 246  * Remove the QP from the table so it can't be found asynchronously by
 247  * the receive interrupt routine.
 248  */
 249 static void remove_qp(struct qib_ibdev *dev, struct qib_qp *qp)
 250 {
 251         struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 252         unsigned n = qpn_hash(dev, qp->ibqp.qp_num);
 253         unsigned long flags;
 254         int removed = 1;
 255
 256         spin_lock_irqsave(&dev->qpt_lock, flags);
 257
 258         if (rcu_dereference_protected(ibp->qp0,
 259                         lockdep_is_held(&dev->qpt_lock)) == qp) {
 260                 RCU_INIT_POINTER(ibp->qp0, NULL);
 261         } else if (rcu_dereference_protected(ibp->qp1,
 262                         lockdep_is_held(&dev->qpt_lock)) == qp) {
 263                 RCU_INIT_POINTER(ibp->qp1, NULL);
 264         } else {
 265                 struct qib_qp *q;
 266                 struct qib_qp __rcu **qpp;
 267
 268                 removed = 0;
 269                 qpp = &dev->qp_table[n];
 270                 for (; (q = rcu_dereference_protected(*qpp,
 271                                 lockdep_is_held(&dev->qpt_lock))) != NULL;
 272                                 qpp = &q->next)
 273                         if (q == qp) {
 274                                 RCU_INIT_POINTER(*qpp,
 275                                         rcu_dereference_protected(qp->next,
 276                                          lockdep_is_held(&dev->qpt_lock)));
 277                                 removed = 1;
 278                                 break;
 279                         }
 280         }
 281
 282         spin_unlock_irqrestore(&dev->qpt_lock, flags);
 283         if (removed) {
 284                 synchronize_rcu();
 285                 atomic_dec(&qp->refcount);
 286         }
 287 }
 288
 289 /**
 290  * qib_free_all_qps - check for QPs still in use
 291  * @qpt: the QP table to empty
 292  *
 293  * There should not be any QPs still in use.
 294  * Free memory for table.
 295  */
 296 unsigned qib_free_all_qps(struct qib_devdata *dd)
 297 {
 298         struct qib_ibdev *dev = &dd->verbs_dev;
 299         unsigned long flags;
 300         struct qib_qp *qp;
 301         unsigned n, qp_inuse = 0;
 302
 303         for (n = 0; n < dd->num_pports; n++) {
 304                 struct qib_ibport *ibp = &dd->pport[n].ibport_data;
 305
 306                 if (!qib_mcast_tree_empty(ibp))
 307                         qp_inuse++;
 308                 rcu_read_lock();
 309                 if (rcu_dereference(ibp->qp0))
 310                         qp_inuse++;
 311                 if (rcu_dereference(ibp->qp1))
 312                         qp_inuse++;
 313                 rcu_read_unlock();
 314         }
 315
 316         spin_lock_irqsave(&dev->qpt_lock, flags);
 317         for (n = 0; n < dev->qp_table_size; n++) {
 318                 qp = rcu_dereference_protected(dev->qp_table[n],
 319                         lockdep_is_held(&dev->qpt_lock));
 320                 RCU_INIT_POINTER(dev->qp_table[n], NULL);
 321
 322                 for (; qp; qp = rcu_dereference_protected(qp->next,
 323                                         lockdep_is_held(&dev->qpt_lock)))
 324                         qp_inuse++;
 325         }
 326         spin_unlock_irqrestore(&dev->qpt_lock, flags);
 327         synchronize_rcu();
 328
 329         return qp_inuse;
 330 }
 331
 332 /**
 333  * qib_lookup_qpn - return the QP with the given QPN
 334  * @qpt: the QP table
 335  * @qpn: the QP number to look up
 336  *
 337  * The caller is responsible for decrementing the QP reference count
 338  * when done.
 339  */
 340 struct qib_qp *qib_lookup_qpn(struct qib_ibport *ibp, u32 qpn)
 341 {
 342         struct qib_qp *qp = NULL;
 343
 344         rcu_read_lock();
 345         if (unlikely(qpn <= 1)) {
 346                 if (qpn == 0)
 347                         qp = rcu_dereference(ibp->qp0);
 348                 else
 349                         qp = rcu_dereference(ibp->qp1);
 350                 if (qp)
 351                         atomic_inc(&qp->refcount);
 352         } else {
 353                 struct qib_ibdev *dev = &ppd_from_ibp(ibp)->dd->verbs_dev;
 354                 unsigned n = qpn_hash(dev, qpn);
 355
 356                 for (qp = rcu_dereference(dev->qp_table[n]); qp;
 357                         qp = rcu_dereference(qp->next))
 358                         if (qp->ibqp.qp_num == qpn) {
 359                                 atomic_inc(&qp->refcount);
 360                                 break;
 361                         }
 362         }
 363         rcu_read_unlock();
 364         return qp;
 365 }
 366
 367 /**
 368  * qib_reset_qp - initialize the QP state to the reset state
 369  * @qp: the QP to reset
 370  * @type: the QP type
 371  */
 372 static void qib_reset_qp(struct qib_qp *qp, enum ib_qp_type type)
 373 {
 374         struct qib_qp_priv *priv = qp->priv;
 375         qp->remote_qpn = 0;
 376         qp->qkey = 0;
 377         qp->qp_access_flags = 0;
 378         atomic_set(&priv->s_dma_busy, 0);
 379         qp->s_flags &= QIB_S_SIGNAL_REQ_WR;
 380         qp->s_hdrwords = 0;
 381         qp->s_wqe = NULL;
 382         qp->s_draining = 0;
 383         qp->s_next_psn = 0;
 384         qp->s_last_psn = 0;
 385         qp->s_sending_psn = 0;
 386         qp->s_sending_hpsn = 0;
 387         qp->s_psn = 0;
 388         qp->r_psn = 0;
 389         qp->r_msn = 0;
 390         if (type == IB_QPT_RC) {
 391                 qp->s_state = IB_OPCODE_RC_SEND_LAST;
 392                 qp->r_state = IB_OPCODE_RC_SEND_LAST;
 393         } else {
 394                 qp->s_state = IB_OPCODE_UC_SEND_LAST;
 395                 qp->r_state = IB_OPCODE_UC_SEND_LAST;
 396         }
 397         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 398         qp->r_nak_state = 0;
 399         qp->r_aflags = 0;
 400         qp->r_flags = 0;
 401         qp->s_head = 0;
 402         qp->s_tail = 0;
 403         qp->s_cur = 0;
 404         qp->s_acked = 0;
 405         qp->s_last = 0;
 406         qp->s_ssn = 1;
 407         qp->s_lsn = 0;
 408         qp->s_mig_state = IB_MIG_MIGRATED;
 409         memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue));
 410         qp->r_head_ack_queue = 0;
 411         qp->s_tail_ack_queue = 0;
 412         qp->s_num_rd_atomic = 0;
 413         if (qp->r_rq.wq) {
 414                 qp->r_rq.wq->head = 0;
 415                 qp->r_rq.wq->tail = 0;
 416         }
 417         qp->r_sge.num_sge = 0;
 418 }
 419
 420 static void clear_mr_refs(struct qib_qp *qp, int clr_sends)
 421 {
 422         unsigned n;
 423
 424         if (test_and_clear_bit(QIB_R_REWIND_SGE, &qp->r_aflags))
 425                 qib_put_ss(&qp->s_rdma_read_sge);
 426
 427         qib_put_ss(&qp->r_sge);
 428
 429         if (clr_sends) {
 430                 while (qp->s_last != qp->s_head) {
 431                         struct qib_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 432                         unsigned i;
 433
 434                         for (i = 0; i < wqe->wr.num_sge; i++) {
 435                                 struct qib_sge *sge = &wqe->sg_list[i];
 436
 437                                 qib_put_mr(sge->mr);
 438                         }
 439                         if (qp->ibqp.qp_type == IB_QPT_UD ||
 440                             qp->ibqp.qp_type == IB_QPT_SMI ||
 441                             qp->ibqp.qp_type == IB_QPT_GSI)
 442                                 atomic_dec(&to_iah(wqe->ud_wr.ah)->refcount);
 443                         if (++qp->s_last >= qp->s_size)
 444                                 qp->s_last = 0;
 445                 }
 446                 if (qp->s_rdma_mr) {
 447                         qib_put_mr(qp->s_rdma_mr);
 448                         qp->s_rdma_mr = NULL;
 449                 }
 450         }
 451
 452         if (qp->ibqp.qp_type != IB_QPT_RC)
 453                 return;
 454
 455         for (n = 0; n < ARRAY_SIZE(qp->s_ack_queue); n++) {
 456                 struct qib_ack_entry *e = &qp->s_ack_queue[n];
 457
 458                 if (e->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST &&
 459                     e->rdma_sge.mr) {
 460                         qib_put_mr(e->rdma_sge.mr);
 461                         e->rdma_sge.mr = NULL;
 462                 }
 463         }
 464 }
 465
 466 /**
 467  * qib_error_qp - put a QP into the error state
 468  * @qp: the QP to put into the error state
 469  * @err: the receive completion error to signal if a RWQE is active
 470  *
 471  * Flushes both send and receive work queues.
 472  * Returns true if last WQE event should be generated.
 473  * The QP r_lock and s_lock should be held and interrupts disabled.
 474  * If we are already in error state, just return.
 475  */
 476 int qib_error_qp(struct qib_qp *qp, enum ib_wc_status err)
 477 {
 478         struct qib_qp_priv *priv = qp->priv;
 479         struct qib_ibdev *dev = to_idev(qp->ibqp.device);
 480         struct ib_wc wc;
 481         int ret = 0;
 482
 483         if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
 484                 goto bail;
 485
 486         qp->state = IB_QPS_ERR;
 487
 488         if (qp->s_flags & (QIB_S_TIMER | QIB_S_WAIT_RNR)) {
 489                 qp->s_flags &= ~(QIB_S_TIMER | QIB_S_WAIT_RNR);
 490                 del_timer(&qp->s_timer);
 491         }
 492
 493         if (qp->s_flags & QIB_S_ANY_WAIT_SEND)
 494                 qp->s_flags &= ~QIB_S_ANY_WAIT_SEND;
 495
 496         spin_lock(&dev->pending_lock);
 497         if (!list_empty(&priv->iowait) && !(qp->s_flags & QIB_S_BUSY)) {
 498                 qp->s_flags &= ~QIB_S_ANY_WAIT_IO;
 499                 list_del_init(&priv->iowait);
 500         }
 501         spin_unlock(&dev->pending_lock);
 502
 503         if (!(qp->s_flags & QIB_S_BUSY)) {
 504                 qp->s_hdrwords = 0;
 505                 if (qp->s_rdma_mr) {
 506                         qib_put_mr(qp->s_rdma_mr);
 507                         qp->s_rdma_mr = NULL;
 508                 }
 509                 if (priv->s_tx) {
 510                         qib_put_txreq(priv->s_tx);
 511                         priv->s_tx = NULL;
 512                 }
 513         }
 514
 515         /* Schedule the sending tasklet to drain the send work queue. */
 516         if (qp->s_last != qp->s_head)
 517                 qib_schedule_send(qp);
 518
 519         clear_mr_refs(qp, 0);
 520
 521         memset(&wc, 0, sizeof(wc));
 522         wc.qp = &qp->ibqp;
 523         wc.opcode = IB_WC_RECV;
 524
 525         if (test_and_clear_bit(QIB_R_WRID_VALID, &qp->r_aflags)) {
 526                 wc.wr_id = qp->r_wr_id;
 527                 wc.status = err;
 528                 qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 529         }
 530         wc.status = IB_WC_WR_FLUSH_ERR;
 531
 532         if (qp->r_rq.wq) {
 533                 struct qib_rwq *wq;
 534                 u32 head;
 535                 u32 tail;
 536
 537                 spin_lock(&qp->r_rq.lock);
 538
 539                 /* sanity check pointers before trusting them */
 540                 wq = qp->r_rq.wq;
 541                 head = wq->head;
 542                 if (head >= qp->r_rq.size)
 543                         head = 0;
 544                 tail = wq->tail;
 545                 if (tail >= qp->r_rq.size)
 546                         tail = 0;
 547                 while (tail != head) {
 548                         wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
 549                         if (++tail >= qp->r_rq.size)
 550                                 tail = 0;
 551                         qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 552                 }
 553                 wq->tail = tail;
 554
 555                 spin_unlock(&qp->r_rq.lock);
 556         } else if (qp->ibqp.event_handler)
 557                 ret = 1;
 558
 559 bail:
 560         return ret;
 561 }
 562
 563 /**
 564  * qib_modify_qp - modify the attributes of a queue pair
 565  * @ibqp: the queue pair who's attributes we're modifying
 566  * @attr: the new attributes
 567  * @attr_mask: the mask of attributes to modify
 568  * @udata: user data for libibverbs.so
 569  *
 570  * Returns 0 on success, otherwise returns an errno.
 571  */
 572 int qib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 573                   int attr_mask, struct ib_udata *udata)
 574 {
 575         struct qib_ibdev *dev = to_idev(ibqp->device);
 576         struct qib_qp *qp = to_iqp(ibqp);
 577         struct qib_qp_priv *priv = qp->priv;
 578         enum ib_qp_state cur_state, new_state;
 579         struct ib_event ev;
 580         int lastwqe = 0;
 581         int mig = 0;
 582         int ret;
 583         u32 pmtu = 0; /* for gcc warning only */
 584
 585         spin_lock_irq(&qp->r_lock);
 586         spin_lock(&qp->s_lock);
 587
 588         cur_state = attr_mask & IB_QP_CUR_STATE ?
 589                 attr->cur_qp_state : qp->state;
 590         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
 591
 592         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
 593                                 attr_mask, IB_LINK_LAYER_UNSPECIFIED))
 594                 goto inval;
 595
 596         if (attr_mask & IB_QP_AV) {
 597                 if (attr->ah_attr.dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE))
 598                         goto inval;
 599                 if (qib_check_ah(qp->ibqp.device, &attr->ah_attr))
 600                         goto inval;
 601         }
 602
 603         if (attr_mask & IB_QP_ALT_PATH) {
 604                 if (attr->alt_ah_attr.dlid >=
 605                     be16_to_cpu(IB_MULTICAST_LID_BASE))
 606                         goto inval;
 607                 if (qib_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
 608                         goto inval;
 609                 if (attr->alt_pkey_index >= qib_get_npkeys(dd_from_dev(dev)))
 610                         goto inval;
 611         }
 612
 613         if (attr_mask & IB_QP_PKEY_INDEX)
 614                 if (attr->pkey_index >= qib_get_npkeys(dd_from_dev(dev)))
 615                         goto inval;
 616
 617         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 618                 if (attr->min_rnr_timer > 31)
 619                         goto inval;
 620
 621         if (attr_mask & IB_QP_PORT)
 622                 if (qp->ibqp.qp_type == IB_QPT_SMI ||
 623                     qp->ibqp.qp_type == IB_QPT_GSI ||
 624                     attr->port_num == 0 ||
 625                     attr->port_num > ibqp->device->phys_port_cnt)
 626                         goto inval;
 627
 628         if (attr_mask & IB_QP_DEST_QPN)
 629                 if (attr->dest_qp_num > QIB_QPN_MASK)
 630                         goto inval;
 631
 632         if (attr_mask & IB_QP_RETRY_CNT)
 633                 if (attr->retry_cnt > 7)
 634                         goto inval;
 635
 636         if (attr_mask & IB_QP_RNR_RETRY)
 637                 if (attr->rnr_retry > 7)
 638                         goto inval;
 639
 640         /*
 641          * Don't allow invalid path_mtu values.  OK to set greater
 642          * than the active mtu (or even the max_cap, if we have tuned
 643          * that to a small mtu.  We'll set qp->path_mtu
 644          * to the lesser of requested attribute mtu and active,
 645          * for packetizing messages.
 646          * Note that the QP port has to be set in INIT and MTU in RTR.
 647          */
 648         if (attr_mask & IB_QP_PATH_MTU) {
 649                 struct qib_devdata *dd = dd_from_dev(dev);
 650                 int mtu, pidx = qp->port_num - 1;
 651
 652                 mtu = ib_mtu_enum_to_int(attr->path_mtu);
 653                 if (mtu == -1)
 654                         goto inval;
 655                 if (mtu > dd->pport[pidx].ibmtu) {
 656                         switch (dd->pport[pidx].ibmtu) {
 657                         case 4096:
 658                                 pmtu = IB_MTU_4096;
 659                                 break;
 660                         case 2048:
 661                                 pmtu = IB_MTU_2048;
 662                                 break;
 663                         case 1024:
 664                                 pmtu = IB_MTU_1024;
 665                                 break;
 666                         case 512:
 667                                 pmtu = IB_MTU_512;
 668                                 break;
 669                         case 256:
 670                                 pmtu = IB_MTU_256;
 671                                 break;
 672                         default:
 673                                 pmtu = IB_MTU_2048;
 674                         }
 675                 } else
 676                         pmtu = attr->path_mtu;
 677         }
 678
 679         if (attr_mask & IB_QP_PATH_MIG_STATE) {
 680                 if (attr->path_mig_state == IB_MIG_REARM) {
 681                         if (qp->s_mig_state == IB_MIG_ARMED)
 682                                 goto inval;
 683                         if (new_state != IB_QPS_RTS)
 684                                 goto inval;
 685                 } else if (attr->path_mig_state == IB_MIG_MIGRATED) {
 686                         if (qp->s_mig_state == IB_MIG_REARM)
 687                                 goto inval;
 688                         if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
 689                                 goto inval;
 690                         if (qp->s_mig_state == IB_MIG_ARMED)
 691                                 mig = 1;
 692                 } else
 693                         goto inval;
 694         }
 695
 696         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 697                 if (attr->max_dest_rd_atomic > QIB_MAX_RDMA_ATOMIC)
 698                         goto inval;
 699
 700         switch (new_state) {
 701         case IB_QPS_RESET:
 702                 if (qp->state != IB_QPS_RESET) {
 703                         qp->state = IB_QPS_RESET;
 704                         spin_lock(&dev->pending_lock);
 705                         if (!list_empty(&priv->iowait))
 706                                 list_del_init(&priv->iowait);
 707                         spin_unlock(&dev->pending_lock);
 708                         qp->s_flags &= ~(QIB_S_TIMER | QIB_S_ANY_WAIT);
 709                         spin_unlock(&qp->s_lock);
 710                         spin_unlock_irq(&qp->r_lock);
 711                         /* Stop the sending work queue and retry timer */
 712                         cancel_work_sync(&priv->s_work);
 713                         del_timer_sync(&qp->s_timer);
 714                         wait_event(priv->wait_dma,
 715                                    !atomic_read(&priv->s_dma_busy));
 716                         if (priv->s_tx) {
 717                                 qib_put_txreq(priv->s_tx);
 718                                 priv->s_tx = NULL;
 719                         }
 720                         remove_qp(dev, qp);
 721                         wait_event(qp->wait, !atomic_read(&qp->refcount));
 722                         spin_lock_irq(&qp->r_lock);
 723                         spin_lock(&qp->s_lock);
 724                         clear_mr_refs(qp, 1);
 725                         qib_reset_qp(qp, ibqp->qp_type);
 726                 }
 727                 break;
 728
 729         case IB_QPS_RTR:
 730                 /* Allow event to retrigger if QP set to RTR more than once */
 731                 qp->r_flags &= ~QIB_R_COMM_EST;
 732                 qp->state = new_state;
 733                 break;
 734
 735         case IB_QPS_SQD:
 736                 qp->s_draining = qp->s_last != qp->s_cur;
 737                 qp->state = new_state;
 738                 break;
 739
 740         case IB_QPS_SQE:
 741                 if (qp->ibqp.qp_type == IB_QPT_RC)
 742                         goto inval;
 743                 qp->state = new_state;
 744                 break;
 745
 746         case IB_QPS_ERR:
 747                 lastwqe = qib_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 748                 break;
 749
 750         default:
 751                 qp->state = new_state;
 752                 break;
 753         }
 754
 755         if (attr_mask & IB_QP_PKEY_INDEX)
 756                 qp->s_pkey_index = attr->pkey_index;
 757
 758         if (attr_mask & IB_QP_PORT)
 759                 qp->port_num = attr->port_num;
 760
 761         if (attr_mask & IB_QP_DEST_QPN)
 762                 qp->remote_qpn = attr->dest_qp_num;
 763
 764         if (attr_mask & IB_QP_SQ_PSN) {
 765                 qp->s_next_psn = attr->sq_psn & QIB_PSN_MASK;
 766                 qp->s_psn = qp->s_next_psn;
 767                 qp->s_sending_psn = qp->s_next_psn;
 768                 qp->s_last_psn = qp->s_next_psn - 1;
 769                 qp->s_sending_hpsn = qp->s_last_psn;
 770         }
 771
 772         if (attr_mask & IB_QP_RQ_PSN)
 773                 qp->r_psn = attr->rq_psn & QIB_PSN_MASK;
 774
 775         if (attr_mask & IB_QP_ACCESS_FLAGS)
 776                 qp->qp_access_flags = attr->qp_access_flags;
 777
 778         if (attr_mask & IB_QP_AV) {
 779                 qp->remote_ah_attr = attr->ah_attr;
 780                 qp->s_srate = attr->ah_attr.static_rate;
 781         }
 782
 783         if (attr_mask & IB_QP_ALT_PATH) {
 784                 qp->alt_ah_attr = attr->alt_ah_attr;
 785                 qp->s_alt_pkey_index = attr->alt_pkey_index;
 786         }
 787
 788         if (attr_mask & IB_QP_PATH_MIG_STATE) {
 789                 qp->s_mig_state = attr->path_mig_state;
 790                 if (mig) {
 791                         qp->remote_ah_attr = qp->alt_ah_attr;
 792                         qp->port_num = qp->alt_ah_attr.port_num;
 793                         qp->s_pkey_index = qp->s_alt_pkey_index;
 794                 }
 795         }
 796
 797         if (attr_mask & IB_QP_PATH_MTU) {
 798                 qp->path_mtu = pmtu;
 799                 qp->pmtu = ib_mtu_enum_to_int(pmtu);
 800         }
 801
 802         if (attr_mask & IB_QP_RETRY_CNT) {
 803                 qp->s_retry_cnt = attr->retry_cnt;
 804                 qp->s_retry = attr->retry_cnt;
 805         }
 806
 807         if (attr_mask & IB_QP_RNR_RETRY) {
 808                 qp->s_rnr_retry_cnt = attr->rnr_retry;
 809                 qp->s_rnr_retry = attr->rnr_retry;
 810         }
 811
 812         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 813                 qp->r_min_rnr_timer = attr->min_rnr_timer;
 814
 815         if (attr_mask & IB_QP_TIMEOUT) {
 816                 qp->timeout = attr->timeout;
 817                 qp->timeout_jiffies =
 818                         usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
 819                                 1000UL);
 820         }
 821
 822         if (attr_mask & IB_QP_QKEY)
 823                 qp->qkey = attr->qkey;
 824
 825         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 826                 qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
 827
 828         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
 829                 qp->s_max_rd_atomic = attr->max_rd_atomic;
 830
 831         spin_unlock(&qp->s_lock);
 832         spin_unlock_irq(&qp->r_lock);
 833
 834         if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
 835                 insert_qp(dev, qp);
 836
 837         if (lastwqe) {
 838                 ev.device = qp->ibqp.device;
 839                 ev.element.qp = &qp->ibqp;
 840                 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 841                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 842         }
 843         if (mig) {
 844                 ev.device = qp->ibqp.device;
 845                 ev.element.qp = &qp->ibqp;
 846                 ev.event = IB_EVENT_PATH_MIG;
 847                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 848         }
 849         ret = 0;
 850         goto bail;
 851
 852 inval:
 853         spin_unlock(&qp->s_lock);
 854         spin_unlock_irq(&qp->r_lock);
 855         ret = -EINVAL;
 856
 857 bail:
 858         return ret;
 859 }
 860
 861 int qib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 862                  int attr_mask, struct ib_qp_init_attr *init_attr)
 863 {
 864         struct qib_qp *qp = to_iqp(ibqp);
 865
 866         attr->qp_state = qp->state;
 867         attr->cur_qp_state = attr->qp_state;
 868         attr->path_mtu = qp->path_mtu;
 869         attr->path_mig_state = qp->s_mig_state;
 870         attr->qkey = qp->qkey;
 871         attr->rq_psn = qp->r_psn & QIB_PSN_MASK;
 872         attr->sq_psn = qp->s_next_psn & QIB_PSN_MASK;
 873         attr->dest_qp_num = qp->remote_qpn;
 874         attr->qp_access_flags = qp->qp_access_flags;
 875         attr->cap.max_send_wr = qp->s_size - 1;
 876         attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
 877         attr->cap.max_send_sge = qp->s_max_sge;
 878         attr->cap.max_recv_sge = qp->r_rq.max_sge;
 879         attr->cap.max_inline_data = 0;
 880         attr->ah_attr = qp->remote_ah_attr;
 881         attr->alt_ah_attr = qp->alt_ah_attr;
 882         attr->pkey_index = qp->s_pkey_index;
 883         attr->alt_pkey_index = qp->s_alt_pkey_index;
 884         attr->en_sqd_async_notify = 0;
 885         attr->sq_draining = qp->s_draining;
 886         attr->max_rd_atomic = qp->s_max_rd_atomic;
 887         attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
 888         attr->min_rnr_timer = qp->r_min_rnr_timer;
 889         attr->port_num = qp->port_num;
 890         attr->timeout = qp->timeout;
 891         attr->retry_cnt = qp->s_retry_cnt;
 892         attr->rnr_retry = qp->s_rnr_retry_cnt;
 893         attr->alt_port_num = qp->alt_ah_attr.port_num;
 894         attr->alt_timeout = qp->alt_timeout;
 895
 896         init_attr->event_handler = qp->ibqp.event_handler;
 897         init_attr->qp_context = qp->ibqp.qp_context;
 898         init_attr->send_cq = qp->ibqp.send_cq;
 899         init_attr->recv_cq = qp->ibqp.recv_cq;
 900         init_attr->srq = qp->ibqp.srq;
 901         init_attr->cap = attr->cap;
 902         if (qp->s_flags & QIB_S_SIGNAL_REQ_WR)
 903                 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 904         else
 905                 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
 906         init_attr->qp_type = qp->ibqp.qp_type;
 907         init_attr->port_num = qp->port_num;
 908         return 0;
 909 }
 910
 911 /**
 912  * qib_compute_aeth - compute the AETH (syndrome + MSN)
 913  * @qp: the queue pair to compute the AETH for
 914  *
 915  * Returns the AETH.
 916  */
 917 __be32 qib_compute_aeth(struct qib_qp *qp)
 918 {
 919         u32 aeth = qp->r_msn & QIB_MSN_MASK;
 920
 921         if (qp->ibqp.srq) {
 922                 /*
 923                  * Shared receive queues don't generate credits.
 924                  * Set the credit field to the invalid value.
 925                  */
 926                 aeth |= QIB_AETH_CREDIT_INVAL << QIB_AETH_CREDIT_SHIFT;
 927         } else {
 928                 u32 min, max, x;
 929                 u32 credits;
 930                 struct qib_rwq *wq = qp->r_rq.wq;
 931                 u32 head;
 932                 u32 tail;
 933
 934                 /* sanity check pointers before trusting them */
 935                 head = wq->head;
 936                 if (head >= qp->r_rq.size)
 937                         head = 0;
 938                 tail = wq->tail;
 939                 if (tail >= qp->r_rq.size)
 940                         tail = 0;
 941                 /*
 942                  * Compute the number of credits available (RWQEs).
 943                  * XXX Not holding the r_rq.lock here so there is a small
 944                  * chance that the pair of reads are not atomic.
 945                  */
 946                 credits = head - tail;
 947                 if ((int)credits < 0)
 948                         credits += qp->r_rq.size;
 949                 /*
 950                  * Binary search the credit table to find the code to
 951                  * use.
 952                  */
 953                 min = 0;
 954                 max = 31;
 955                 for (;;) {
 956                         x = (min + max) / 2;
 957                         if (credit_table[x] == credits)
 958                                 break;
 959                         if (credit_table[x] > credits)
 960                                 max = x;
 961                         else if (min == x)
 962                                 break;
 963                         else
 964                                 min = x;
 965                 }
 966                 aeth |= x << QIB_AETH_CREDIT_SHIFT;
 967         }
 968         return cpu_to_be32(aeth);
 969 }
 970
 971 /**
 972  * qib_create_qp - create a queue pair for a device
 973  * @ibpd: the protection domain who's device we create the queue pair for
 974  * @init_attr: the attributes of the queue pair
 975  * @udata: user data for libibverbs.so
 976  *
 977  * Returns the queue pair on success, otherwise returns an errno.
 978  *
 979  * Called by the ib_create_qp() core verbs function.
 980  */
 981 struct ib_qp *qib_create_qp(struct ib_pd *ibpd,
 982                             struct ib_qp_init_attr *init_attr,
 983                             struct ib_udata *udata)
 984 {
 985         struct qib_qp *qp;
 986         int err;
 987         struct qib_swqe *swq = NULL;
 988         struct qib_ibdev *dev;
 989         struct qib_devdata *dd;
 990         size_t sz;
 991         size_t sg_list_sz;
 992         struct ib_qp *ret;
 993         gfp_t gfp;
 994         struct qib_qp_priv *priv;
 995
 996         if (init_attr->cap.max_send_sge > ib_qib_max_sges ||
 997             init_attr->cap.max_send_wr > ib_qib_max_qp_wrs ||
 998             init_attr->create_flags & ~(IB_QP_CREATE_USE_GFP_NOIO))
 999                 return ERR_PTR(-EINVAL);
1000
1001         /* GFP_NOIO is applicable in RC QPs only */
1002         if (init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO &&
1003             init_attr->qp_type != IB_QPT_RC)
1004                 return ERR_PTR(-EINVAL);
1005
1006         gfp = init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO ?
1007                         GFP_NOIO : GFP_KERNEL;
1008
1009         /* Check receive queue parameters if no SRQ is specified. */
1010         if (!init_attr->srq) {
1011                 if (init_attr->cap.max_recv_sge > ib_qib_max_sges ||
1012                     init_attr->cap.max_recv_wr > ib_qib_max_qp_wrs) {
1013                         ret = ERR_PTR(-EINVAL);
1014                         goto bail;
1015                 }
1016                 if (init_attr->cap.max_send_sge +
1017                     init_attr->cap.max_send_wr +
1018                     init_attr->cap.max_recv_sge +
1019                     init_attr->cap.max_recv_wr == 0) {
1020                         ret = ERR_PTR(-EINVAL);
1021                         goto bail;
1022                 }
1023         }
1024
1025         switch (init_attr->qp_type) {
1026         case IB_QPT_SMI:
1027         case IB_QPT_GSI:
1028                 if (init_attr->port_num == 0 ||
1029                     init_attr->port_num > ibpd->device->phys_port_cnt) {
1030                         ret = ERR_PTR(-EINVAL);
1031                         goto bail;
1032                 }
1033         case IB_QPT_UC:
1034         case IB_QPT_RC:
1035         case IB_QPT_UD:
1036                 sz = sizeof(struct qib_sge) *
1037                         init_attr->cap.max_send_sge +
1038                         sizeof(struct qib_swqe);
1039                 swq = __vmalloc((init_attr->cap.max_send_wr + 1) * sz,
1040                                 gfp, PAGE_KERNEL);
1041                 if (swq == NULL) {
1042                         ret = ERR_PTR(-ENOMEM);
1043                         goto bail;
1044                 }
1045                 sz = sizeof(*qp);
1046                 sg_list_sz = 0;
1047                 if (init_attr->srq) {
1048                         struct qib_srq *srq = to_isrq(init_attr->srq);
1049
1050                         if (srq->rq.max_sge > 1)
1051                                 sg_list_sz = sizeof(*qp->r_sg_list) *
1052                                         (srq->rq.max_sge - 1);
1053                 } else if (init_attr->cap.max_recv_sge > 1)
1054                         sg_list_sz = sizeof(*qp->r_sg_list) *
1055                                 (init_attr->cap.max_recv_sge - 1);
1056                 qp = kzalloc(sz + sg_list_sz, gfp);
1057                 if (!qp) {
1058                         ret = ERR_PTR(-ENOMEM);
1059                         goto bail_swq;
1060                 }
1061                 RCU_INIT_POINTER(qp->next, NULL);
1062                 priv = kzalloc(sizeof(*priv), gfp);
1063                 if (!priv) {
1064                         ret = ERR_PTR(-ENOMEM);
1065                         goto bail_qp_hdr;
1066                 }
1067                 priv->owner = qp;
1068                 priv->s_hdr = kzalloc(sizeof(*priv->s_hdr), gfp);
1069                 if (!priv->s_hdr) {
1070                         ret = ERR_PTR(-ENOMEM);
1071                         goto bail_qp;
1072                 }
1073                 qp->priv = priv;
1074                 qp->timeout_jiffies =
1075                         usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
1076                                 1000UL);
1077                 if (init_attr->srq)
1078                         sz = 0;
1079                 else {
1080                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
1081                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
1082                         sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
1083                                 sizeof(struct qib_rwqe);
1084                         if (gfp != GFP_NOIO)
1085                                 qp->r_rq.wq = vmalloc_user(
1086                                                 sizeof(struct qib_rwq) +
1087                                                 qp->r_rq.size * sz);
1088                         else
1089                                 qp->r_rq.wq = __vmalloc(
1090                                                 sizeof(struct qib_rwq) +
1091                                                 qp->r_rq.size * sz,
1092                                                 gfp, PAGE_KERNEL);
1093
1094                         if (!qp->r_rq.wq) {
1095                                 ret = ERR_PTR(-ENOMEM);
1096                                 goto bail_qp;
1097                         }
1098                 }
1099
1100                 /*
1101                  * ib_create_qp() will initialize qp->ibqp
1102                  * except for qp->ibqp.qp_num.
1103                  */
1104                 spin_lock_init(&qp->r_lock);
1105                 spin_lock_init(&qp->s_lock);
1106                 spin_lock_init(&qp->r_rq.lock);
1107                 atomic_set(&qp->refcount, 0);
1108                 init_waitqueue_head(&qp->wait);
1109                 init_waitqueue_head(&priv->wait_dma);
1110                 init_timer(&qp->s_timer);
1111                 qp->s_timer.data = (unsigned long)qp;
1112                 INIT_WORK(&priv->s_work, qib_do_send);
1113                 INIT_LIST_HEAD(&priv->iowait);
1114                 INIT_LIST_HEAD(&qp->rspwait);
1115                 qp->state = IB_QPS_RESET;
1116                 qp->s_wq = swq;
1117                 qp->s_size = init_attr->cap.max_send_wr + 1;
1118                 qp->s_max_sge = init_attr->cap.max_send_sge;
1119                 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
1120                         qp->s_flags = QIB_S_SIGNAL_REQ_WR;
1121                 dev = to_idev(ibpd->device);
1122                 dd = dd_from_dev(dev);
1123                 err = alloc_qpn(dd, &dev->qpn_table, init_attr->qp_type,
1124                                 init_attr->port_num, gfp);
1125                 if (err < 0) {
1126                         ret = ERR_PTR(err);
1127                         vfree(qp->r_rq.wq);
1128                         goto bail_qp;
1129                 }
1130                 qp->ibqp.qp_num = err;
1131                 qp->port_num = init_attr->port_num;
1132                 qib_reset_qp(qp, init_attr->qp_type);
1133                 break;
1134
1135         default:
1136                 /* Don't support raw QPs */
1137                 ret = ERR_PTR(-ENOSYS);
1138                 goto bail;
1139         }
1140
1141         init_attr->cap.max_inline_data = 0;
1142
1143         /*
1144          * Return the address of the RWQ as the offset to mmap.
1145          * See qib_mmap() for details.
1146          */
1147         if (udata && udata->outlen >= sizeof(__u64)) {
1148                 if (!qp->r_rq.wq) {
1149                         __u64 offset = 0;
1150
1151                         err = ib_copy_to_udata(udata, &offset,
1152                                                sizeof(offset));
1153                         if (err) {
1154                                 ret = ERR_PTR(err);
1155                                 goto bail_ip;
1156                         }
1157                 } else {
1158                         u32 s = sizeof(struct qib_rwq) + qp->r_rq.size * sz;
1159
1160                         qp->ip = qib_create_mmap_info(dev, s,
1161                                                       ibpd->uobject->context,
1162                                                       qp->r_rq.wq);
1163                         if (!qp->ip) {
1164                                 ret = ERR_PTR(-ENOMEM);
1165                                 goto bail_ip;
1166                         }
1167
1168                         err = ib_copy_to_udata(udata, &(qp->ip->offset),
1169                                                sizeof(qp->ip->offset));
1170                         if (err) {
1171                                 ret = ERR_PTR(err);
1172                                 goto bail_ip;
1173                         }
1174                 }
1175         }
1176
1177         spin_lock(&dev->n_qps_lock);
1178         if (dev->n_qps_allocated == ib_qib_max_qps) {
1179                 spin_unlock(&dev->n_qps_lock);
1180                 ret = ERR_PTR(-ENOMEM);
1181                 goto bail_ip;
1182         }
1183
1184         dev->n_qps_allocated++;
1185         spin_unlock(&dev->n_qps_lock);
1186
1187         if (qp->ip) {
1188                 spin_lock_irq(&dev->pending_lock);
1189                 list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps);
1190                 spin_unlock_irq(&dev->pending_lock);
1191         }
1192
1193         ret = &qp->ibqp;
1194         goto bail;
1195
1196 bail_ip:
1197         if (qp->ip)
1198                 kref_put(&qp->ip->ref, qib_release_mmap_info);
1199         else
1200                 vfree(qp->r_rq.wq);
1201         free_qpn(&dev->qpn_table, qp->ibqp.qp_num);
1202 bail_qp:
1203         kfree(priv->s_hdr);
1204         kfree(priv);
1205 bail_qp_hdr:
1206         kfree(qp);
1207 bail_swq:
1208         vfree(swq);
1209 bail:
1210         return ret;
1211 }
1212
1213 /**
1214  * qib_destroy_qp - destroy a queue pair
1215  * @ibqp: the queue pair to destroy
1216  *
1217  * Returns 0 on success.
1218  *
1219  * Note that this can be called while the QP is actively sending or
1220  * receiving!
1221  */
1222 int qib_destroy_qp(struct ib_qp *ibqp)
1223 {
1224         struct qib_qp *qp = to_iqp(ibqp);
1225         struct qib_ibdev *dev = to_idev(ibqp->device);
1226         struct qib_qp_priv *priv = qp->priv;
1227
1228         /* Make sure HW and driver activity is stopped. */
1229         spin_lock_irq(&qp->s_lock);
1230         if (qp->state != IB_QPS_RESET) {
1231                 qp->state = IB_QPS_RESET;
1232                 spin_lock(&dev->pending_lock);
1233                 if (!list_empty(&priv->iowait))
1234                         list_del_init(&priv->iowait);
1235                 spin_unlock(&dev->pending_lock);
1236                 qp->s_flags &= ~(QIB_S_TIMER | QIB_S_ANY_WAIT);
1237                 spin_unlock_irq(&qp->s_lock);
1238                 cancel_work_sync(&priv->s_work);
1239                 del_timer_sync(&qp->s_timer);
1240                 wait_event(priv->wait_dma, !atomic_read(&priv->s_dma_busy));
1241                 if (priv->s_tx) {
1242                         qib_put_txreq(priv->s_tx);
1243                         priv->s_tx = NULL;
1244                 }
1245                 remove_qp(dev, qp);
1246                 wait_event(qp->wait, !atomic_read(&qp->refcount));
1247                 clear_mr_refs(qp, 1);
1248         } else
1249                 spin_unlock_irq(&qp->s_lock);
1250
1251         /* all user's cleaned up, mark it available */
1252         free_qpn(&dev->qpn_table, qp->ibqp.qp_num);
1253         spin_lock(&dev->n_qps_lock);
1254         dev->n_qps_allocated--;
1255         spin_unlock(&dev->n_qps_lock);
1256
1257         if (qp->ip)
1258                 kref_put(&qp->ip->ref, qib_release_mmap_info);
1259         else
1260                 vfree(qp->r_rq.wq);
1261         vfree(qp->s_wq);
1262         kfree(priv->s_hdr);
1263         kfree(priv);
1264         kfree(qp);
1265         return 0;
1266 }
1267
1268 /**
1269  * qib_init_qpn_table - initialize the QP number table for a device
1270  * @qpt: the QPN table
1271  */
1272 void qib_init_qpn_table(struct qib_devdata *dd, struct qib_qpn_table *qpt)
1273 {
1274         spin_lock_init(&qpt->lock);
1275         qpt->last = 1;          /* start with QPN 2 */
1276         qpt->nmaps = 1;
1277         qpt->mask = dd->qpn_mask;
1278 }
1279
1280 /**
1281  * qib_free_qpn_table - free the QP number table for a device
1282  * @qpt: the QPN table
1283  */
1284 void qib_free_qpn_table(struct qib_qpn_table *qpt)
1285 {
1286         int i;
1287
1288         for (i = 0; i < ARRAY_SIZE(qpt->map); i++)
1289                 if (qpt->map[i].page)
1290                         free_page((unsigned long) qpt->map[i].page);
1291 }
1292
1293 /**
1294  * qib_get_credit - flush the send work queue of a QP
1295  * @qp: the qp who's send work queue to flush
1296  * @aeth: the Acknowledge Extended Transport Header
1297  *
1298  * The QP s_lock should be held.
1299  */
1300 void qib_get_credit(struct qib_qp *qp, u32 aeth)
1301 {
1302         u32 credit = (aeth >> QIB_AETH_CREDIT_SHIFT) & QIB_AETH_CREDIT_MASK;
1303
1304         /*
1305          * If the credit is invalid, we can send
1306          * as many packets as we like.  Otherwise, we have to
1307          * honor the credit field.
1308          */
1309         if (credit == QIB_AETH_CREDIT_INVAL) {
1310                 if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT)) {
1311                         qp->s_flags |= QIB_S_UNLIMITED_CREDIT;
1312                         if (qp->s_flags & QIB_S_WAIT_SSN_CREDIT) {
1313                                 qp->s_flags &= ~QIB_S_WAIT_SSN_CREDIT;
1314                                 qib_schedule_send(qp);
1315                         }
1316                 }
1317         } else if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT)) {
1318                 /* Compute new LSN (i.e., MSN + credit) */
1319                 credit = (aeth + credit_table[credit]) & QIB_MSN_MASK;
1320                 if (qib_cmp24(credit, qp->s_lsn) > 0) {
1321                         qp->s_lsn = credit;
1322                         if (qp->s_flags & QIB_S_WAIT_SSN_CREDIT) {
1323                                 qp->s_flags &= ~QIB_S_WAIT_SSN_CREDIT;
1324                                 qib_schedule_send(qp);
1325                         }
1326                 }
1327         }
1328 }
1329
1330 #ifdef CONFIG_DEBUG_FS
1331
1332 struct qib_qp_iter {
1333         struct qib_ibdev *dev;
1334         struct qib_qp *qp;
1335         int n;
1336 };
1337
1338 struct qib_qp_iter *qib_qp_iter_init(struct qib_ibdev *dev)
1339 {
1340         struct qib_qp_iter *iter;
1341
1342         iter = kzalloc(sizeof(*iter), GFP_KERNEL);
1343         if (!iter)
1344                 return NULL;
1345
1346         iter->dev = dev;
1347         if (qib_qp_iter_next(iter)) {
1348                 kfree(iter);
1349                 return NULL;
1350         }
1351
1352         return iter;
1353 }
1354
1355 int qib_qp_iter_next(struct qib_qp_iter *iter)
1356 {
1357         struct qib_ibdev *dev = iter->dev;
1358         int n = iter->n;
1359         int ret = 1;
1360         struct qib_qp *pqp = iter->qp;
1361         struct qib_qp *qp;
1362
1363         for (; n < dev->qp_table_size; n++) {
1364                 if (pqp)
1365                         qp = rcu_dereference(pqp->next);
1366                 else
1367                         qp = rcu_dereference(dev->qp_table[n]);
1368                 pqp = qp;
1369                 if (qp) {
1370                         iter->qp = qp;
1371                         iter->n = n;
1372                         return 0;
1373                 }
1374         }
1375         return ret;
1376 }
1377
1378 static const char * const qp_type_str[] = {
1379         "SMI", "GSI", "RC", "UC", "UD",
1380 };
1381
1382 void qib_qp_iter_print(struct seq_file *s, struct qib_qp_iter *iter)
1383 {
1384         struct qib_swqe *wqe;
1385         struct qib_qp *qp = iter->qp;
1386         struct qib_qp_priv *priv = qp->priv;
1387
1388         wqe = get_swqe_ptr(qp, qp->s_last);
1389         seq_printf(s,
1390                    "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",
1391                    iter->n,
1392                    qp->ibqp.qp_num,
1393                    qp_type_str[qp->ibqp.qp_type],
1394                    qp->state,
1395                    wqe->wr.opcode,
1396                    qp->s_hdrwords,
1397                    qp->s_flags,
1398                    atomic_read(&priv->s_dma_busy),
1399                    !list_empty(&priv->iowait),
1400                    qp->timeout,
1401                    wqe->ssn,
1402                    qp->s_lsn,
1403                    qp->s_last_psn,
1404                    qp->s_psn, qp->s_next_psn,
1405                    qp->s_sending_psn, qp->s_sending_hpsn,
1406                    qp->s_last, qp->s_acked, qp->s_cur,
1407                    qp->s_tail, qp->s_head, qp->s_size,
1408                    qp->remote_qpn,
1409                    qp->remote_ah_attr.dlid);
1410 }
1411
1412 #endif