git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_qp.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
   3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/err.h>
  35 #include <linux/vmalloc.h>
  36
  37 #include "ipath_verbs.h"
  38 #include "ipath_kernel.h"
  39
  40 #define BITS_PER_PAGE           (PAGE_SIZE*BITS_PER_BYTE)
  41 #define BITS_PER_PAGE_MASK      (BITS_PER_PAGE-1)
  42 #define mk_qpn(qpt, map, off)   (((map) - (qpt)->map) * BITS_PER_PAGE + \
  43                                  (off))
  44 #define find_next_offset(map, off) find_next_zero_bit((map)->page, \
  45                                                       BITS_PER_PAGE, off)
  46
  47 /*
  48  * Convert the AETH credit code into the number of credits.
  49  */
  50 static u32 credit_table[31] = {
  51         0,                      /* 0 */
  52         1,                      /* 1 */
  53         2,                      /* 2 */
  54         3,                      /* 3 */
  55         4,                      /* 4 */
  56         6,                      /* 5 */
  57         8,                      /* 6 */
  58         12,                     /* 7 */
  59         16,                     /* 8 */
  60         24,                     /* 9 */
  61         32,                     /* A */
  62         48,                     /* B */
  63         64,                     /* C */
  64         96,                     /* D */
  65         128,                    /* E */
  66         192,                    /* F */
  67         256,                    /* 10 */
  68         384,                    /* 11 */
  69         512,                    /* 12 */
  70         768,                    /* 13 */
  71         1024,                   /* 14 */
  72         1536,                   /* 15 */
  73         2048,                   /* 16 */
  74         3072,                   /* 17 */
  75         4096,                   /* 18 */
  76         6144,                   /* 19 */
  77         8192,                   /* 1A */
  78         12288,                  /* 1B */
  79         16384,                  /* 1C */
  80         24576,                  /* 1D */
  81         32768                   /* 1E */
  82 };
  83
  84
  85 static void get_map_page(struct ipath_qp_table *qpt, struct qpn_map *map)
  86 {
  87         unsigned long page = get_zeroed_page(GFP_KERNEL);
  88         unsigned long flags;
  89
  90         /*
  91          * Free the page if someone raced with us installing it.
  92          */
  93
  94         spin_lock_irqsave(&qpt->lock, flags);
  95         if (map->page)
  96                 free_page(page);
  97         else
  98                 map->page = (void *)page;
  99         spin_unlock_irqrestore(&qpt->lock, flags);
 100 }
 101
 102
 103 static int alloc_qpn(struct ipath_qp_table *qpt, enum ib_qp_type type)
 104 {
 105         u32 i, offset, max_scan, qpn;
 106         struct qpn_map *map;
 107         u32 ret = -1;
 108
 109         if (type == IB_QPT_SMI)
 110                 ret = 0;
 111         else if (type == IB_QPT_GSI)
 112                 ret = 1;
 113
 114         if (ret != -1) {
 115                 map = &qpt->map[0];
 116                 if (unlikely(!map->page)) {
 117                         get_map_page(qpt, map);
 118                         if (unlikely(!map->page)) {
 119                                 ret = -ENOMEM;
 120                                 goto bail;
 121                         }
 122                 }
 123                 if (!test_and_set_bit(ret, map->page))
 124                         atomic_dec(&map->n_free);
 125                 else
 126                         ret = -EBUSY;
 127                 goto bail;
 128         }
 129
 130         qpn = qpt->last + 1;
 131         if (qpn >= QPN_MAX)
 132                 qpn = 2;
 133         offset = qpn & BITS_PER_PAGE_MASK;
 134         map = &qpt->map[qpn / BITS_PER_PAGE];
 135         max_scan = qpt->nmaps - !offset;
 136         for (i = 0;;) {
 137                 if (unlikely(!map->page)) {
 138                         get_map_page(qpt, map);
 139                         if (unlikely(!map->page))
 140                                 break;
 141                 }
 142                 if (likely(atomic_read(&map->n_free))) {
 143                         do {
 144                                 if (!test_and_set_bit(offset, map->page)) {
 145                                         atomic_dec(&map->n_free);
 146                                         qpt->last = qpn;
 147                                         ret = qpn;
 148                                         goto bail;
 149                                 }
 150                                 offset = find_next_offset(map, offset);
 151                                 qpn = mk_qpn(qpt, map, offset);
 152                                 /*
 153                                  * This test differs from alloc_pidmap().
 154                                  * If find_next_offset() does find a zero
 155                                  * bit, we don't need to check for QPN
 156                                  * wrapping around past our starting QPN.
 157                                  * We just need to be sure we don't loop
 158                                  * forever.
 159                                  */
 160                         } while (offset < BITS_PER_PAGE && qpn < QPN_MAX);
 161                 }
 162                 /*
 163                  * In order to keep the number of pages allocated to a
 164                  * minimum, we scan the all existing pages before increasing
 165                  * the size of the bitmap table.
 166                  */
 167                 if (++i > max_scan) {
 168                         if (qpt->nmaps == QPNMAP_ENTRIES)
 169                                 break;
 170                         map = &qpt->map[qpt->nmaps++];
 171                         offset = 0;
 172                 } else if (map < &qpt->map[qpt->nmaps]) {
 173                         ++map;
 174                         offset = 0;
 175                 } else {
 176                         map = &qpt->map[0];
 177                         offset = 2;
 178                 }
 179                 qpn = mk_qpn(qpt, map, offset);
 180         }
 181
 182         ret = -ENOMEM;
 183
 184 bail:
 185         return ret;
 186 }
 187
 188 static void free_qpn(struct ipath_qp_table *qpt, u32 qpn)
 189 {
 190         struct qpn_map *map;
 191
 192         map = qpt->map + qpn / BITS_PER_PAGE;
 193         if (map->page)
 194                 clear_bit(qpn & BITS_PER_PAGE_MASK, map->page);
 195         atomic_inc(&map->n_free);
 196 }
 197
 198 /**
 199  * ipath_alloc_qpn - allocate a QP number
 200  * @qpt: the QP table
 201  * @qp: the QP
 202  * @type: the QP type (IB_QPT_SMI and IB_QPT_GSI are special)
 203  *
 204  * Allocate the next available QPN and put the QP into the hash table.
 205  * The hash table holds a reference to the QP.
 206  */
 207 static int ipath_alloc_qpn(struct ipath_qp_table *qpt, struct ipath_qp *qp,
 208                            enum ib_qp_type type)
 209 {
 210         unsigned long flags;
 211         int ret;
 212
 213         ret = alloc_qpn(qpt, type);
 214         if (ret < 0)
 215                 goto bail;
 216         qp->ibqp.qp_num = ret;
 217
 218         /* Add the QP to the hash table. */
 219         spin_lock_irqsave(&qpt->lock, flags);
 220
 221         ret %= qpt->max;
 222         qp->next = qpt->table[ret];
 223         qpt->table[ret] = qp;
 224         atomic_inc(&qp->refcount);
 225
 226         spin_unlock_irqrestore(&qpt->lock, flags);
 227         ret = 0;
 228
 229 bail:
 230         return ret;
 231 }
 232
 233 /**
 234  * ipath_free_qp - remove a QP from the QP table
 235  * @qpt: the QP table
 236  * @qp: the QP to remove
 237  *
 238  * Remove the QP from the table so it can't be found asynchronously by
 239  * the receive interrupt routine.
 240  */
 241 static void ipath_free_qp(struct ipath_qp_table *qpt, struct ipath_qp *qp)
 242 {
 243         struct ipath_qp *q, **qpp;
 244         unsigned long flags;
 245
 246         spin_lock_irqsave(&qpt->lock, flags);
 247
 248         /* Remove QP from the hash table. */
 249         qpp = &qpt->table[qp->ibqp.qp_num % qpt->max];
 250         for (; (q = *qpp) != NULL; qpp = &q->next) {
 251                 if (q == qp) {
 252                         *qpp = qp->next;
 253                         qp->next = NULL;
 254                         atomic_dec(&qp->refcount);
 255                         break;
 256                 }
 257         }
 258
 259         spin_unlock_irqrestore(&qpt->lock, flags);
 260 }
 261
 262 /**
 263  * ipath_free_all_qps - check for QPs still in use
 264  * @qpt: the QP table to empty
 265  *
 266  * There should not be any QPs still in use.
 267  * Free memory for table.
 268  */
 269 unsigned ipath_free_all_qps(struct ipath_qp_table *qpt)
 270 {
 271         unsigned long flags;
 272         struct ipath_qp *qp;
 273         u32 n, qp_inuse = 0;
 274
 275         spin_lock_irqsave(&qpt->lock, flags);
 276         for (n = 0; n < qpt->max; n++) {
 277                 qp = qpt->table[n];
 278                 qpt->table[n] = NULL;
 279
 280                 for (; qp; qp = qp->next)
 281                         qp_inuse++;
 282         }
 283         spin_unlock_irqrestore(&qpt->lock, flags);
 284
 285         for (n = 0; n < ARRAY_SIZE(qpt->map); n++)
 286                 if (qpt->map[n].page)
 287                         free_page((unsigned long) qpt->map[n].page);
 288         return qp_inuse;
 289 }
 290
 291 /**
 292  * ipath_lookup_qpn - return the QP with the given QPN
 293  * @qpt: the QP table
 294  * @qpn: the QP number to look up
 295  *
 296  * The caller is responsible for decrementing the QP reference count
 297  * when done.
 298  */
 299 struct ipath_qp *ipath_lookup_qpn(struct ipath_qp_table *qpt, u32 qpn)
 300 {
 301         unsigned long flags;
 302         struct ipath_qp *qp;
 303
 304         spin_lock_irqsave(&qpt->lock, flags);
 305
 306         for (qp = qpt->table[qpn % qpt->max]; qp; qp = qp->next) {
 307                 if (qp->ibqp.qp_num == qpn) {
 308                         atomic_inc(&qp->refcount);
 309                         break;
 310                 }
 311         }
 312
 313         spin_unlock_irqrestore(&qpt->lock, flags);
 314         return qp;
 315 }
 316
 317 /**
 318  * ipath_reset_qp - initialize the QP state to the reset state
 319  * @qp: the QP to reset
 320  * @type: the QP type
 321  */
 322 static void ipath_reset_qp(struct ipath_qp *qp, enum ib_qp_type type)
 323 {
 324         qp->remote_qpn = 0;
 325         qp->qkey = 0;
 326         qp->qp_access_flags = 0;
 327         atomic_set(&qp->s_dma_busy, 0);
 328         qp->s_flags &= IPATH_S_SIGNAL_REQ_WR;
 329         qp->s_hdrwords = 0;
 330         qp->s_wqe = NULL;
 331         qp->s_pkt_delay = 0;
 332         qp->s_draining = 0;
 333         qp->s_psn = 0;
 334         qp->r_psn = 0;
 335         qp->r_msn = 0;
 336         if (type == IB_QPT_RC) {
 337                 qp->s_state = IB_OPCODE_RC_SEND_LAST;
 338                 qp->r_state = IB_OPCODE_RC_SEND_LAST;
 339         } else {
 340                 qp->s_state = IB_OPCODE_UC_SEND_LAST;
 341                 qp->r_state = IB_OPCODE_UC_SEND_LAST;
 342         }
 343         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 344         qp->r_nak_state = 0;
 345         qp->r_aflags = 0;
 346         qp->r_flags = 0;
 347         qp->s_rnr_timeout = 0;
 348         qp->s_head = 0;
 349         qp->s_tail = 0;
 350         qp->s_cur = 0;
 351         qp->s_last = 0;
 352         qp->s_ssn = 1;
 353         qp->s_lsn = 0;
 354         memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue));
 355         qp->r_head_ack_queue = 0;
 356         qp->s_tail_ack_queue = 0;
 357         qp->s_num_rd_atomic = 0;
 358         if (qp->r_rq.wq) {
 359                 qp->r_rq.wq->head = 0;
 360                 qp->r_rq.wq->tail = 0;
 361         }
 362 }
 363
 364 /**
 365  * ipath_error_qp - put a QP into the error state
 366  * @qp: the QP to put into the error state
 367  * @err: the receive completion error to signal if a RWQE is active
 368  *
 369  * Flushes both send and receive work queues.
 370  * Returns true if last WQE event should be generated.
 371  * The QP s_lock should be held and interrupts disabled.
 372  * If we are already in error state, just return.
 373  */
 374
 375 int ipath_error_qp(struct ipath_qp *qp, enum ib_wc_status err)
 376 {
 377         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 378         struct ib_wc wc;
 379         int ret = 0;
 380
 381         if (qp->state == IB_QPS_ERR)
 382                 goto bail;
 383
 384         qp->state = IB_QPS_ERR;
 385
 386         spin_lock(&dev->pending_lock);
 387         if (!list_empty(&qp->timerwait))
 388                 list_del_init(&qp->timerwait);
 389         if (!list_empty(&qp->piowait))
 390                 list_del_init(&qp->piowait);
 391         spin_unlock(&dev->pending_lock);
 392
 393         /* Schedule the sending tasklet to drain the send work queue. */
 394         if (qp->s_last != qp->s_head)
 395                 ipath_schedule_send(qp);
 396
 397         memset(&wc, 0, sizeof(wc));
 398         wc.qp = &qp->ibqp;
 399         wc.opcode = IB_WC_RECV;
 400
 401         if (test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags)) {
 402                 wc.wr_id = qp->r_wr_id;
 403                 wc.status = err;
 404                 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 405         }
 406         wc.status = IB_WC_WR_FLUSH_ERR;
 407
 408         if (qp->r_rq.wq) {
 409                 struct ipath_rwq *wq;
 410                 u32 head;
 411                 u32 tail;
 412
 413                 spin_lock(&qp->r_rq.lock);
 414
 415                 /* sanity check pointers before trusting them */
 416                 wq = qp->r_rq.wq;
 417                 head = wq->head;
 418                 if (head >= qp->r_rq.size)
 419                         head = 0;
 420                 tail = wq->tail;
 421                 if (tail >= qp->r_rq.size)
 422                         tail = 0;
 423                 while (tail != head) {
 424                         wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
 425                         if (++tail >= qp->r_rq.size)
 426                                 tail = 0;
 427                         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 428                 }
 429                 wq->tail = tail;
 430
 431                 spin_unlock(&qp->r_rq.lock);
 432         } else if (qp->ibqp.event_handler)
 433                 ret = 1;
 434
 435 bail:
 436         return ret;
 437 }
 438
 439 /**
 440  * ipath_modify_qp - modify the attributes of a queue pair
 441  * @ibqp: the queue pair who's attributes we're modifying
 442  * @attr: the new attributes
 443  * @attr_mask: the mask of attributes to modify
 444  * @udata: user data for ipathverbs.so
 445  *
 446  * Returns 0 on success, otherwise returns an errno.
 447  */
 448 int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 449                     int attr_mask, struct ib_udata *udata)
 450 {
 451         struct ipath_ibdev *dev = to_idev(ibqp->device);
 452         struct ipath_qp *qp = to_iqp(ibqp);
 453         enum ib_qp_state cur_state, new_state;
 454         int lastwqe = 0;
 455         int ret;
 456
 457         spin_lock_irq(&qp->s_lock);
 458
 459         cur_state = attr_mask & IB_QP_CUR_STATE ?
 460                 attr->cur_qp_state : qp->state;
 461         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
 462
 463         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
 464                                 attr_mask))
 465                 goto inval;
 466
 467         if (attr_mask & IB_QP_AV) {
 468                 if (attr->ah_attr.dlid == 0 ||
 469                     attr->ah_attr.dlid >= IPATH_MULTICAST_LID_BASE)
 470                         goto inval;
 471
 472                 if ((attr->ah_attr.ah_flags & IB_AH_GRH) &&
 473                     (attr->ah_attr.grh.sgid_index > 1))
 474                         goto inval;
 475         }
 476
 477         if (attr_mask & IB_QP_PKEY_INDEX)
 478                 if (attr->pkey_index >= ipath_get_npkeys(dev->dd))
 479                         goto inval;
 480
 481         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 482                 if (attr->min_rnr_timer > 31)
 483                         goto inval;
 484
 485         if (attr_mask & IB_QP_PORT)
 486                 if (attr->port_num == 0 ||
 487                     attr->port_num > ibqp->device->phys_port_cnt)
 488                         goto inval;
 489
 490         /*
 491          * don't allow invalid Path MTU values or greater than 2048
 492          * unless we are configured for a 4KB MTU
 493          */
 494         if ((attr_mask & IB_QP_PATH_MTU) &&
 495                 (ib_mtu_enum_to_int(attr->path_mtu) == -1 ||
 496                 (attr->path_mtu > IB_MTU_2048 && !ipath_mtu4096)))
 497                 goto inval;
 498
 499         if (attr_mask & IB_QP_PATH_MIG_STATE)
 500                 if (attr->path_mig_state != IB_MIG_MIGRATED &&
 501                     attr->path_mig_state != IB_MIG_REARM)
 502                         goto inval;
 503
 504         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 505                 if (attr->max_dest_rd_atomic > IPATH_MAX_RDMA_ATOMIC)
 506                         goto inval;
 507
 508         switch (new_state) {
 509         case IB_QPS_RESET:
 510                 if (qp->state != IB_QPS_RESET) {
 511                         qp->state = IB_QPS_RESET;
 512                         spin_lock(&dev->pending_lock);
 513                         if (!list_empty(&qp->timerwait))
 514                                 list_del_init(&qp->timerwait);
 515                         if (!list_empty(&qp->piowait))
 516                                 list_del_init(&qp->piowait);
 517                         spin_unlock(&dev->pending_lock);
 518                         qp->s_flags &= ~IPATH_S_ANY_WAIT;
 519                         spin_unlock_irq(&qp->s_lock);
 520                         /* Stop the sending tasklet */
 521                         tasklet_kill(&qp->s_task);
 522                         wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
 523                         spin_lock_irq(&qp->s_lock);
 524                 }
 525                 ipath_reset_qp(qp, ibqp->qp_type);
 526                 break;
 527
 528         case IB_QPS_SQD:
 529                 qp->s_draining = qp->s_last != qp->s_cur;
 530                 qp->state = new_state;
 531                 break;
 532
 533         case IB_QPS_SQE:
 534                 if (qp->ibqp.qp_type == IB_QPT_RC)
 535                         goto inval;
 536                 qp->state = new_state;
 537                 break;
 538
 539         case IB_QPS_ERR:
 540                 lastwqe = ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 541                 break;
 542
 543         default:
 544                 qp->state = new_state;
 545                 break;
 546         }
 547
 548         if (attr_mask & IB_QP_PKEY_INDEX)
 549                 qp->s_pkey_index = attr->pkey_index;
 550
 551         if (attr_mask & IB_QP_DEST_QPN)
 552                 qp->remote_qpn = attr->dest_qp_num;
 553
 554         if (attr_mask & IB_QP_SQ_PSN) {
 555                 qp->s_psn = qp->s_next_psn = attr->sq_psn;
 556                 qp->s_last_psn = qp->s_next_psn - 1;
 557         }
 558
 559         if (attr_mask & IB_QP_RQ_PSN)
 560                 qp->r_psn = attr->rq_psn;
 561
 562         if (attr_mask & IB_QP_ACCESS_FLAGS)
 563                 qp->qp_access_flags = attr->qp_access_flags;
 564
 565         if (attr_mask & IB_QP_AV) {
 566                 qp->remote_ah_attr = attr->ah_attr;
 567                 qp->s_dmult = ipath_ib_rate_to_mult(attr->ah_attr.static_rate);
 568         }
 569
 570         if (attr_mask & IB_QP_PATH_MTU)
 571                 qp->path_mtu = attr->path_mtu;
 572
 573         if (attr_mask & IB_QP_RETRY_CNT)
 574                 qp->s_retry = qp->s_retry_cnt = attr->retry_cnt;
 575
 576         if (attr_mask & IB_QP_RNR_RETRY) {
 577                 qp->s_rnr_retry = attr->rnr_retry;
 578                 if (qp->s_rnr_retry > 7)
 579                         qp->s_rnr_retry = 7;
 580                 qp->s_rnr_retry_cnt = qp->s_rnr_retry;
 581         }
 582
 583         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 584                 qp->r_min_rnr_timer = attr->min_rnr_timer;
 585
 586         if (attr_mask & IB_QP_TIMEOUT)
 587                 qp->timeout = attr->timeout;
 588
 589         if (attr_mask & IB_QP_QKEY)
 590                 qp->qkey = attr->qkey;
 591
 592         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 593                 qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
 594
 595         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
 596                 qp->s_max_rd_atomic = attr->max_rd_atomic;
 597
 598         spin_unlock_irq(&qp->s_lock);
 599
 600         if (lastwqe) {
 601                 struct ib_event ev;
 602
 603                 ev.device = qp->ibqp.device;
 604                 ev.element.qp = &qp->ibqp;
 605                 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 606                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 607         }
 608         ret = 0;
 609         goto bail;
 610
 611 inval:
 612         spin_unlock_irq(&qp->s_lock);
 613         ret = -EINVAL;
 614
 615 bail:
 616         return ret;
 617 }
 618
 619 int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 620                    int attr_mask, struct ib_qp_init_attr *init_attr)
 621 {
 622         struct ipath_qp *qp = to_iqp(ibqp);
 623
 624         attr->qp_state = qp->state;
 625         attr->cur_qp_state = attr->qp_state;
 626         attr->path_mtu = qp->path_mtu;
 627         attr->path_mig_state = 0;
 628         attr->qkey = qp->qkey;
 629         attr->rq_psn = qp->r_psn;
 630         attr->sq_psn = qp->s_next_psn;
 631         attr->dest_qp_num = qp->remote_qpn;
 632         attr->qp_access_flags = qp->qp_access_flags;
 633         attr->cap.max_send_wr = qp->s_size - 1;
 634         attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
 635         attr->cap.max_send_sge = qp->s_max_sge;
 636         attr->cap.max_recv_sge = qp->r_rq.max_sge;
 637         attr->cap.max_inline_data = 0;
 638         attr->ah_attr = qp->remote_ah_attr;
 639         memset(&attr->alt_ah_attr, 0, sizeof(attr->alt_ah_attr));
 640         attr->pkey_index = qp->s_pkey_index;
 641         attr->alt_pkey_index = 0;
 642         attr->en_sqd_async_notify = 0;
 643         attr->sq_draining = qp->s_draining;
 644         attr->max_rd_atomic = qp->s_max_rd_atomic;
 645         attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
 646         attr->min_rnr_timer = qp->r_min_rnr_timer;
 647         attr->port_num = 1;
 648         attr->timeout = qp->timeout;
 649         attr->retry_cnt = qp->s_retry_cnt;
 650         attr->rnr_retry = qp->s_rnr_retry_cnt;
 651         attr->alt_port_num = 0;
 652         attr->alt_timeout = 0;
 653
 654         init_attr->event_handler = qp->ibqp.event_handler;
 655         init_attr->qp_context = qp->ibqp.qp_context;
 656         init_attr->send_cq = qp->ibqp.send_cq;
 657         init_attr->recv_cq = qp->ibqp.recv_cq;
 658         init_attr->srq = qp->ibqp.srq;
 659         init_attr->cap = attr->cap;
 660         if (qp->s_flags & IPATH_S_SIGNAL_REQ_WR)
 661                 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 662         else
 663                 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
 664         init_attr->qp_type = qp->ibqp.qp_type;
 665         init_attr->port_num = 1;
 666         return 0;
 667 }
 668
 669 /**
 670  * ipath_compute_aeth - compute the AETH (syndrome + MSN)
 671  * @qp: the queue pair to compute the AETH for
 672  *
 673  * Returns the AETH.
 674  */
 675 __be32 ipath_compute_aeth(struct ipath_qp *qp)
 676 {
 677         u32 aeth = qp->r_msn & IPATH_MSN_MASK;
 678
 679         if (qp->ibqp.srq) {
 680                 /*
 681                  * Shared receive queues don't generate credits.
 682                  * Set the credit field to the invalid value.
 683                  */
 684                 aeth |= IPATH_AETH_CREDIT_INVAL << IPATH_AETH_CREDIT_SHIFT;
 685         } else {
 686                 u32 min, max, x;
 687                 u32 credits;
 688                 struct ipath_rwq *wq = qp->r_rq.wq;
 689                 u32 head;
 690                 u32 tail;
 691
 692                 /* sanity check pointers before trusting them */
 693                 head = wq->head;
 694                 if (head >= qp->r_rq.size)
 695                         head = 0;
 696                 tail = wq->tail;
 697                 if (tail >= qp->r_rq.size)
 698                         tail = 0;
 699                 /*
 700                  * Compute the number of credits available (RWQEs).
 701                  * XXX Not holding the r_rq.lock here so there is a small
 702                  * chance that the pair of reads are not atomic.
 703                  */
 704                 credits = head - tail;
 705                 if ((int)credits < 0)
 706                         credits += qp->r_rq.size;
 707                 /*
 708                  * Binary search the credit table to find the code to
 709                  * use.
 710                  */
 711                 min = 0;
 712                 max = 31;
 713                 for (;;) {
 714                         x = (min + max) / 2;
 715                         if (credit_table[x] == credits)
 716                                 break;
 717                         if (credit_table[x] > credits)
 718                                 max = x;
 719                         else if (min == x)
 720                                 break;
 721                         else
 722                                 min = x;
 723                 }
 724                 aeth |= x << IPATH_AETH_CREDIT_SHIFT;
 725         }
 726         return cpu_to_be32(aeth);
 727 }
 728
 729 /**
 730  * ipath_create_qp - create a queue pair for a device
 731  * @ibpd: the protection domain who's device we create the queue pair for
 732  * @init_attr: the attributes of the queue pair
 733  * @udata: unused by InfiniPath
 734  *
 735  * Returns the queue pair on success, otherwise returns an errno.
 736  *
 737  * Called by the ib_create_qp() core verbs function.
 738  */
 739 struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
 740                               struct ib_qp_init_attr *init_attr,
 741                               struct ib_udata *udata)
 742 {
 743         struct ipath_qp *qp;
 744         int err;
 745         struct ipath_swqe *swq = NULL;
 746         struct ipath_ibdev *dev;
 747         size_t sz;
 748         size_t sg_list_sz;
 749         struct ib_qp *ret;
 750
 751         if (init_attr->create_flags) {
 752                 ret = ERR_PTR(-EINVAL);
 753                 goto bail;
 754         }
 755
 756         if (init_attr->cap.max_send_sge > ib_ipath_max_sges ||
 757             init_attr->cap.max_send_wr > ib_ipath_max_qp_wrs) {
 758                 ret = ERR_PTR(-EINVAL);
 759                 goto bail;
 760         }
 761
 762         /* Check receive queue parameters if no SRQ is specified. */
 763         if (!init_attr->srq) {
 764                 if (init_attr->cap.max_recv_sge > ib_ipath_max_sges ||
 765                     init_attr->cap.max_recv_wr > ib_ipath_max_qp_wrs) {
 766                         ret = ERR_PTR(-EINVAL);
 767                         goto bail;
 768                 }
 769                 if (init_attr->cap.max_send_sge +
 770                     init_attr->cap.max_send_wr +
 771                     init_attr->cap.max_recv_sge +
 772                     init_attr->cap.max_recv_wr == 0) {
 773                         ret = ERR_PTR(-EINVAL);
 774                         goto bail;
 775                 }
 776         }
 777
 778         switch (init_attr->qp_type) {
 779         case IB_QPT_UC:
 780         case IB_QPT_RC:
 781         case IB_QPT_UD:
 782         case IB_QPT_SMI:
 783         case IB_QPT_GSI:
 784                 sz = sizeof(struct ipath_sge) *
 785                         init_attr->cap.max_send_sge +
 786                         sizeof(struct ipath_swqe);
 787                 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
 788                 if (swq == NULL) {
 789                         ret = ERR_PTR(-ENOMEM);
 790                         goto bail;
 791                 }
 792                 sz = sizeof(*qp);
 793                 sg_list_sz = 0;
 794                 if (init_attr->srq) {
 795                         struct ipath_srq *srq = to_isrq(init_attr->srq);
 796
 797                         if (srq->rq.max_sge > 1)
 798                                 sg_list_sz = sizeof(*qp->r_sg_list) *
 799                                         (srq->rq.max_sge - 1);
 800                 } else if (init_attr->cap.max_recv_sge > 1)
 801                         sg_list_sz = sizeof(*qp->r_sg_list) *
 802                                 (init_attr->cap.max_recv_sge - 1);
 803                 qp = kmalloc(sz + sg_list_sz, GFP_KERNEL);
 804                 if (!qp) {
 805                         ret = ERR_PTR(-ENOMEM);
 806                         goto bail_swq;
 807                 }
 808                 if (sg_list_sz && (init_attr->qp_type == IB_QPT_UD ||
 809                     init_attr->qp_type == IB_QPT_SMI ||
 810                     init_attr->qp_type == IB_QPT_GSI)) {
 811                         qp->r_ud_sg_list = kmalloc(sg_list_sz, GFP_KERNEL);
 812                         if (!qp->r_ud_sg_list) {
 813                                 ret = ERR_PTR(-ENOMEM);
 814                                 goto bail_qp;
 815                         }
 816                 } else
 817                         qp->r_ud_sg_list = NULL;
 818                 if (init_attr->srq) {
 819                         sz = 0;
 820                         qp->r_rq.size = 0;
 821                         qp->r_rq.max_sge = 0;
 822                         qp->r_rq.wq = NULL;
 823                         init_attr->cap.max_recv_wr = 0;
 824                         init_attr->cap.max_recv_sge = 0;
 825                 } else {
 826                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
 827                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
 828                         sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
 829                                 sizeof(struct ipath_rwqe);
 830                         qp->r_rq.wq = vmalloc_user(sizeof(struct ipath_rwq) +
 831                                               qp->r_rq.size * sz);
 832                         if (!qp->r_rq.wq) {
 833                                 ret = ERR_PTR(-ENOMEM);
 834                                 goto bail_sg_list;
 835                         }
 836                 }
 837
 838                 /*
 839                  * ib_create_qp() will initialize qp->ibqp
 840                  * except for qp->ibqp.qp_num.
 841                  */
 842                 spin_lock_init(&qp->s_lock);
 843                 spin_lock_init(&qp->r_rq.lock);
 844                 atomic_set(&qp->refcount, 0);
 845                 init_waitqueue_head(&qp->wait);
 846                 init_waitqueue_head(&qp->wait_dma);
 847                 tasklet_init(&qp->s_task, ipath_do_send, (unsigned long)qp);
 848                 INIT_LIST_HEAD(&qp->piowait);
 849                 INIT_LIST_HEAD(&qp->timerwait);
 850                 qp->state = IB_QPS_RESET;
 851                 qp->s_wq = swq;
 852                 qp->s_size = init_attr->cap.max_send_wr + 1;
 853                 qp->s_max_sge = init_attr->cap.max_send_sge;
 854                 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
 855                         qp->s_flags = IPATH_S_SIGNAL_REQ_WR;
 856                 else
 857                         qp->s_flags = 0;
 858                 dev = to_idev(ibpd->device);
 859                 err = ipath_alloc_qpn(&dev->qp_table, qp,
 860                                       init_attr->qp_type);
 861                 if (err) {
 862                         ret = ERR_PTR(err);
 863                         vfree(qp->r_rq.wq);
 864                         goto bail_sg_list;
 865                 }
 866                 qp->ip = NULL;
 867                 qp->s_tx = NULL;
 868                 ipath_reset_qp(qp, init_attr->qp_type);
 869                 break;
 870
 871         default:
 872                 /* Don't support raw QPs */
 873                 ret = ERR_PTR(-ENOSYS);
 874                 goto bail;
 875         }
 876
 877         init_attr->cap.max_inline_data = 0;
 878
 879         /*
 880          * Return the address of the RWQ as the offset to mmap.
 881          * See ipath_mmap() for details.
 882          */
 883         if (udata && udata->outlen >= sizeof(__u64)) {
 884                 if (!qp->r_rq.wq) {
 885                         __u64 offset = 0;
 886
 887                         err = ib_copy_to_udata(udata, &offset,
 888                                                sizeof(offset));
 889                         if (err) {
 890                                 ret = ERR_PTR(err);
 891                                 goto bail_ip;
 892                         }
 893                 } else {
 894                         u32 s = sizeof(struct ipath_rwq) +
 895                                 qp->r_rq.size * sz;
 896
 897                         qp->ip =
 898                             ipath_create_mmap_info(dev, s,
 899                                                    ibpd->uobject->context,
 900                                                    qp->r_rq.wq);
 901                         if (!qp->ip) {
 902                                 ret = ERR_PTR(-ENOMEM);
 903                                 goto bail_ip;
 904                         }
 905
 906                         err = ib_copy_to_udata(udata, &(qp->ip->offset),
 907                                                sizeof(qp->ip->offset));
 908                         if (err) {
 909                                 ret = ERR_PTR(err);
 910                                 goto bail_ip;
 911                         }
 912                 }
 913         }
 914
 915         spin_lock(&dev->n_qps_lock);
 916         if (dev->n_qps_allocated == ib_ipath_max_qps) {
 917                 spin_unlock(&dev->n_qps_lock);
 918                 ret = ERR_PTR(-ENOMEM);
 919                 goto bail_ip;
 920         }
 921
 922         dev->n_qps_allocated++;
 923         spin_unlock(&dev->n_qps_lock);
 924
 925         if (qp->ip) {
 926                 spin_lock_irq(&dev->pending_lock);
 927                 list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps);
 928                 spin_unlock_irq(&dev->pending_lock);
 929         }
 930
 931         ret = &qp->ibqp;
 932         goto bail;
 933
 934 bail_ip:
 935         if (qp->ip)
 936                 kref_put(&qp->ip->ref, ipath_release_mmap_info);
 937         else
 938                 vfree(qp->r_rq.wq);
 939         ipath_free_qp(&dev->qp_table, qp);
 940         free_qpn(&dev->qp_table, qp->ibqp.qp_num);
 941 bail_sg_list:
 942         kfree(qp->r_ud_sg_list);
 943 bail_qp:
 944         kfree(qp);
 945 bail_swq:
 946         vfree(swq);
 947 bail:
 948         return ret;
 949 }
 950
 951 /**
 952  * ipath_destroy_qp - destroy a queue pair
 953  * @ibqp: the queue pair to destroy
 954  *
 955  * Returns 0 on success.
 956  *
 957  * Note that this can be called while the QP is actively sending or
 958  * receiving!
 959  */
 960 int ipath_destroy_qp(struct ib_qp *ibqp)
 961 {
 962         struct ipath_qp *qp = to_iqp(ibqp);
 963         struct ipath_ibdev *dev = to_idev(ibqp->device);
 964
 965         /* Make sure HW and driver activity is stopped. */
 966         spin_lock_irq(&qp->s_lock);
 967         if (qp->state != IB_QPS_RESET) {
 968                 qp->state = IB_QPS_RESET;
 969                 spin_lock(&dev->pending_lock);
 970                 if (!list_empty(&qp->timerwait))
 971                         list_del_init(&qp->timerwait);
 972                 if (!list_empty(&qp->piowait))
 973                         list_del_init(&qp->piowait);
 974                 spin_unlock(&dev->pending_lock);
 975                 qp->s_flags &= ~IPATH_S_ANY_WAIT;
 976                 spin_unlock_irq(&qp->s_lock);
 977                 /* Stop the sending tasklet */
 978                 tasklet_kill(&qp->s_task);
 979                 wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
 980         } else
 981                 spin_unlock_irq(&qp->s_lock);
 982
 983         ipath_free_qp(&dev->qp_table, qp);
 984
 985         if (qp->s_tx) {
 986                 atomic_dec(&qp->refcount);
 987                 if (qp->s_tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEBUF)
 988                         kfree(qp->s_tx->txreq.map_addr);
 989                 spin_lock_irq(&dev->pending_lock);
 990                 list_add(&qp->s_tx->txreq.list, &dev->txreq_free);
 991                 spin_unlock_irq(&dev->pending_lock);
 992                 qp->s_tx = NULL;
 993         }
 994
 995         wait_event(qp->wait, !atomic_read(&qp->refcount));
 996
 997         /* all user's cleaned up, mark it available */
 998         free_qpn(&dev->qp_table, qp->ibqp.qp_num);
 999         spin_lock(&dev->n_qps_lock);
1000         dev->n_qps_allocated--;
1001         spin_unlock(&dev->n_qps_lock);
1002
1003         if (qp->ip)
1004                 kref_put(&qp->ip->ref, ipath_release_mmap_info);
1005         else
1006                 vfree(qp->r_rq.wq);
1007         kfree(qp->r_ud_sg_list);
1008         vfree(qp->s_wq);
1009         kfree(qp);
1010         return 0;
1011 }
1012
1013 /**
1014  * ipath_init_qp_table - initialize the QP table for a device
1015  * @idev: the device who's QP table we're initializing
1016  * @size: the size of the QP table
1017  *
1018  * Returns 0 on success, otherwise returns an errno.
1019  */
1020 int ipath_init_qp_table(struct ipath_ibdev *idev, int size)
1021 {
1022         int i;
1023         int ret;
1024
1025         idev->qp_table.last = 1;        /* QPN 0 and 1 are special. */
1026         idev->qp_table.max = size;
1027         idev->qp_table.nmaps = 1;
1028         idev->qp_table.table = kzalloc(size * sizeof(*idev->qp_table.table),
1029                                        GFP_KERNEL);
1030         if (idev->qp_table.table == NULL) {
1031                 ret = -ENOMEM;
1032                 goto bail;
1033         }
1034
1035         for (i = 0; i < ARRAY_SIZE(idev->qp_table.map); i++) {
1036                 atomic_set(&idev->qp_table.map[i].n_free, BITS_PER_PAGE);
1037                 idev->qp_table.map[i].page = NULL;
1038         }
1039
1040         ret = 0;
1041
1042 bail:
1043         return ret;
1044 }
1045
1046 /**
1047  * ipath_get_credit - flush the send work queue of a QP
1048  * @qp: the qp who's send work queue to flush
1049  * @aeth: the Acknowledge Extended Transport Header
1050  *
1051  * The QP s_lock should be held.
1052  */
1053 void ipath_get_credit(struct ipath_qp *qp, u32 aeth)
1054 {
1055         u32 credit = (aeth >> IPATH_AETH_CREDIT_SHIFT) & IPATH_AETH_CREDIT_MASK;
1056
1057         /*
1058          * If the credit is invalid, we can send
1059          * as many packets as we like.  Otherwise, we have to
1060          * honor the credit field.
1061          */
1062         if (credit == IPATH_AETH_CREDIT_INVAL)
1063                 qp->s_lsn = (u32) -1;
1064         else if (qp->s_lsn != (u32) -1) {
1065                 /* Compute new LSN (i.e., MSN + credit) */
1066                 credit = (aeth + credit_table[credit]) & IPATH_MSN_MASK;
1067                 if (ipath_cmp24(credit, qp->s_lsn) > 0)
1068                         qp->s_lsn = credit;
1069         }
1070
1071         /* Restart sending if it was blocked due to lack of credits. */
1072         if ((qp->s_flags & IPATH_S_WAIT_SSN_CREDIT) &&
1073             qp->s_cur != qp->s_head &&
1074             (qp->s_lsn == (u32) -1 ||
1075              ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn,
1076                          qp->s_lsn + 1) <= 0))
1077                 ipath_schedule_send(qp);
1078 }