SAFE public projects git trees. - safe/jmp/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/sched.h>
  36 #include <linux/vmalloc.h>
  37
  38 #include "ipath_verbs.h"
  39 #include "ipath_kernel.h"
  40
  41 #define BITS_PER_PAGE           (PAGE_SIZE*BITS_PER_BYTE)
  42 #define BITS_PER_PAGE_MASK      (BITS_PER_PAGE-1)
  43 #define mk_qpn(qpt, map, off)   (((map) - (qpt)->map) * BITS_PER_PAGE + \
  44                                  (off))
  45 #define find_next_offset(map, off) find_next_zero_bit((map)->page, \
  46                                                       BITS_PER_PAGE, off)
  47
  48 /*
  49  * Convert the AETH credit code into the number of credits.
  50  */
  51 static u32 credit_table[31] = {
  52         0,                      /* 0 */
  53         1,                      /* 1 */
  54         2,                      /* 2 */
  55         3,                      /* 3 */
  56         4,                      /* 4 */
  57         6,                      /* 5 */
  58         8,                      /* 6 */
  59         12,                     /* 7 */
  60         16,                     /* 8 */
  61         24,                     /* 9 */
  62         32,                     /* A */
  63         48,                     /* B */
  64         64,                     /* C */
  65         96,                     /* D */
  66         128,                    /* E */
  67         192,                    /* F */
  68         256,                    /* 10 */
  69         384,                    /* 11 */
  70         512,                    /* 12 */
  71         768,                    /* 13 */
  72         1024,                   /* 14 */
  73         1536,                   /* 15 */
  74         2048,                   /* 16 */
  75         3072,                   /* 17 */
  76         4096,                   /* 18 */
  77         6144,                   /* 19 */
  78         8192,                   /* 1A */
  79         12288,                  /* 1B */
  80         16384,                  /* 1C */
  81         24576,                  /* 1D */
  82         32768                   /* 1E */
  83 };
  84
  85
  86 static void get_map_page(struct ipath_qp_table *qpt, struct qpn_map *map)
  87 {
  88         unsigned long page = get_zeroed_page(GFP_KERNEL);
  89         unsigned long flags;
  90
  91         /*
  92          * Free the page if someone raced with us installing it.
  93          */
  94
  95         spin_lock_irqsave(&qpt->lock, flags);
  96         if (map->page)
  97                 free_page(page);
  98         else
  99                 map->page = (void *)page;
 100         spin_unlock_irqrestore(&qpt->lock, flags);
 101 }
 102
 103
 104 static int alloc_qpn(struct ipath_qp_table *qpt, enum ib_qp_type type)
 105 {
 106         u32 i, offset, max_scan, qpn;
 107         struct qpn_map *map;
 108         u32 ret = -1;
 109
 110         if (type == IB_QPT_SMI)
 111                 ret = 0;
 112         else if (type == IB_QPT_GSI)
 113                 ret = 1;
 114
 115         if (ret != -1) {
 116                 map = &qpt->map[0];
 117                 if (unlikely(!map->page)) {
 118                         get_map_page(qpt, map);
 119                         if (unlikely(!map->page)) {
 120                                 ret = -ENOMEM;
 121                                 goto bail;
 122                         }
 123                 }
 124                 if (!test_and_set_bit(ret, map->page))
 125                         atomic_dec(&map->n_free);
 126                 else
 127                         ret = -EBUSY;
 128                 goto bail;
 129         }
 130
 131         qpn = qpt->last + 1;
 132         if (qpn >= QPN_MAX)
 133                 qpn = 2;
 134         offset = qpn & BITS_PER_PAGE_MASK;
 135         map = &qpt->map[qpn / BITS_PER_PAGE];
 136         max_scan = qpt->nmaps - !offset;
 137         for (i = 0;;) {
 138                 if (unlikely(!map->page)) {
 139                         get_map_page(qpt, map);
 140                         if (unlikely(!map->page))
 141                                 break;
 142                 }
 143                 if (likely(atomic_read(&map->n_free))) {
 144                         do {
 145                                 if (!test_and_set_bit(offset, map->page)) {
 146                                         atomic_dec(&map->n_free);
 147                                         qpt->last = qpn;
 148                                         ret = qpn;
 149                                         goto bail;
 150                                 }
 151                                 offset = find_next_offset(map, offset);
 152                                 qpn = mk_qpn(qpt, map, offset);
 153                                 /*
 154                                  * This test differs from alloc_pidmap().
 155                                  * If find_next_offset() does find a zero
 156                                  * bit, we don't need to check for QPN
 157                                  * wrapping around past our starting QPN.
 158                                  * We just need to be sure we don't loop
 159                                  * forever.
 160                                  */
 161                         } while (offset < BITS_PER_PAGE && qpn < QPN_MAX);
 162                 }
 163                 /*
 164                  * In order to keep the number of pages allocated to a
 165                  * minimum, we scan the all existing pages before increasing
 166                  * the size of the bitmap table.
 167                  */
 168                 if (++i > max_scan) {
 169                         if (qpt->nmaps == QPNMAP_ENTRIES)
 170                                 break;
 171                         map = &qpt->map[qpt->nmaps++];
 172                         offset = 0;
 173                 } else if (map < &qpt->map[qpt->nmaps]) {
 174                         ++map;
 175                         offset = 0;
 176                 } else {
 177                         map = &qpt->map[0];
 178                         offset = 2;
 179                 }
 180                 qpn = mk_qpn(qpt, map, offset);
 181         }
 182
 183         ret = -ENOMEM;
 184
 185 bail:
 186         return ret;
 187 }
 188
 189 static void free_qpn(struct ipath_qp_table *qpt, u32 qpn)
 190 {
 191         struct qpn_map *map;
 192
 193         map = qpt->map + qpn / BITS_PER_PAGE;
 194         if (map->page)
 195                 clear_bit(qpn & BITS_PER_PAGE_MASK, map->page);
 196         atomic_inc(&map->n_free);
 197 }
 198
 199 /**
 200  * ipath_alloc_qpn - allocate a QP number
 201  * @qpt: the QP table
 202  * @qp: the QP
 203  * @type: the QP type (IB_QPT_SMI and IB_QPT_GSI are special)
 204  *
 205  * Allocate the next available QPN and put the QP into the hash table.
 206  * The hash table holds a reference to the QP.
 207  */
 208 static int ipath_alloc_qpn(struct ipath_qp_table *qpt, struct ipath_qp *qp,
 209                            enum ib_qp_type type)
 210 {
 211         unsigned long flags;
 212         int ret;
 213
 214         ret = alloc_qpn(qpt, type);
 215         if (ret < 0)
 216                 goto bail;
 217         qp->ibqp.qp_num = ret;
 218
 219         /* Add the QP to the hash table. */
 220         spin_lock_irqsave(&qpt->lock, flags);
 221
 222         ret %= qpt->max;
 223         qp->next = qpt->table[ret];
 224         qpt->table[ret] = qp;
 225         atomic_inc(&qp->refcount);
 226
 227         spin_unlock_irqrestore(&qpt->lock, flags);
 228         ret = 0;
 229
 230 bail:
 231         return ret;
 232 }
 233
 234 /**
 235  * ipath_free_qp - remove a QP from the QP table
 236  * @qpt: the QP table
 237  * @qp: the QP to remove
 238  *
 239  * Remove the QP from the table so it can't be found asynchronously by
 240  * the receive interrupt routine.
 241  */
 242 static void ipath_free_qp(struct ipath_qp_table *qpt, struct ipath_qp *qp)
 243 {
 244         struct ipath_qp *q, **qpp;
 245         unsigned long flags;
 246
 247         spin_lock_irqsave(&qpt->lock, flags);
 248
 249         /* Remove QP from the hash table. */
 250         qpp = &qpt->table[qp->ibqp.qp_num % qpt->max];
 251         for (; (q = *qpp) != NULL; qpp = &q->next) {
 252                 if (q == qp) {
 253                         *qpp = qp->next;
 254                         qp->next = NULL;
 255                         atomic_dec(&qp->refcount);
 256                         break;
 257                 }
 258         }
 259
 260         spin_unlock_irqrestore(&qpt->lock, flags);
 261 }
 262
 263 /**
 264  * ipath_free_all_qps - check for QPs still in use
 265  * @qpt: the QP table to empty
 266  *
 267  * There should not be any QPs still in use.
 268  * Free memory for table.
 269  */
 270 unsigned ipath_free_all_qps(struct ipath_qp_table *qpt)
 271 {
 272         unsigned long flags;
 273         struct ipath_qp *qp;
 274         u32 n, qp_inuse = 0;
 275
 276         spin_lock_irqsave(&qpt->lock, flags);
 277         for (n = 0; n < qpt->max; n++) {
 278                 qp = qpt->table[n];
 279                 qpt->table[n] = NULL;
 280
 281                 for (; qp; qp = qp->next)
 282                         qp_inuse++;
 283         }
 284         spin_unlock_irqrestore(&qpt->lock, flags);
 285
 286         for (n = 0; n < ARRAY_SIZE(qpt->map); n++)
 287                 if (qpt->map[n].page)
 288                         free_page((unsigned long) qpt->map[n].page);
 289         return qp_inuse;
 290 }
 291
 292 /**
 293  * ipath_lookup_qpn - return the QP with the given QPN
 294  * @qpt: the QP table
 295  * @qpn: the QP number to look up
 296  *
 297  * The caller is responsible for decrementing the QP reference count
 298  * when done.
 299  */
 300 struct ipath_qp *ipath_lookup_qpn(struct ipath_qp_table *qpt, u32 qpn)
 301 {
 302         unsigned long flags;
 303         struct ipath_qp *qp;
 304
 305         spin_lock_irqsave(&qpt->lock, flags);
 306
 307         for (qp = qpt->table[qpn % qpt->max]; qp; qp = qp->next) {
 308                 if (qp->ibqp.qp_num == qpn) {
 309                         atomic_inc(&qp->refcount);
 310                         break;
 311                 }
 312         }
 313
 314         spin_unlock_irqrestore(&qpt->lock, flags);
 315         return qp;
 316 }
 317
 318 /**
 319  * ipath_reset_qp - initialize the QP state to the reset state
 320  * @qp: the QP to reset
 321  * @type: the QP type
 322  */
 323 static void ipath_reset_qp(struct ipath_qp *qp, enum ib_qp_type type)
 324 {
 325         qp->remote_qpn = 0;
 326         qp->qkey = 0;
 327         qp->qp_access_flags = 0;
 328         atomic_set(&qp->s_dma_busy, 0);
 329         qp->s_flags &= IPATH_S_SIGNAL_REQ_WR;
 330         qp->s_hdrwords = 0;
 331         qp->s_wqe = NULL;
 332         qp->s_pkt_delay = 0;
 333         qp->s_draining = 0;
 334         qp->s_psn = 0;
 335         qp->r_psn = 0;
 336         qp->r_msn = 0;
 337         if (type == IB_QPT_RC) {
 338                 qp->s_state = IB_OPCODE_RC_SEND_LAST;
 339                 qp->r_state = IB_OPCODE_RC_SEND_LAST;
 340         } else {
 341                 qp->s_state = IB_OPCODE_UC_SEND_LAST;
 342                 qp->r_state = IB_OPCODE_UC_SEND_LAST;
 343         }
 344         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 345         qp->r_nak_state = 0;
 346         qp->r_aflags = 0;
 347         qp->r_flags = 0;
 348         qp->s_rnr_timeout = 0;
 349         qp->s_head = 0;
 350         qp->s_tail = 0;
 351         qp->s_cur = 0;
 352         qp->s_last = 0;
 353         qp->s_ssn = 1;
 354         qp->s_lsn = 0;
 355         memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue));
 356         qp->r_head_ack_queue = 0;
 357         qp->s_tail_ack_queue = 0;
 358         qp->s_num_rd_atomic = 0;
 359         if (qp->r_rq.wq) {
 360                 qp->r_rq.wq->head = 0;
 361                 qp->r_rq.wq->tail = 0;
 362         }
 363 }
 364
 365 /**
 366  * ipath_error_qp - put a QP into the error state
 367  * @qp: the QP to put into the error state
 368  * @err: the receive completion error to signal if a RWQE is active
 369  *
 370  * Flushes both send and receive work queues.
 371  * Returns true if last WQE event should be generated.
 372  * The QP s_lock should be held and interrupts disabled.
 373  * If we are already in error state, just return.
 374  */
 375
 376 int ipath_error_qp(struct ipath_qp *qp, enum ib_wc_status err)
 377 {
 378         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 379         struct ib_wc wc;
 380         int ret = 0;
 381
 382         if (qp->state == IB_QPS_ERR)
 383                 goto bail;
 384
 385         qp->state = IB_QPS_ERR;
 386
 387         spin_lock(&dev->pending_lock);
 388         if (!list_empty(&qp->timerwait))
 389                 list_del_init(&qp->timerwait);
 390         if (!list_empty(&qp->piowait))
 391                 list_del_init(&qp->piowait);
 392         spin_unlock(&dev->pending_lock);
 393
 394         /* Schedule the sending tasklet to drain the send work queue. */
 395         if (qp->s_last != qp->s_head)
 396                 ipath_schedule_send(qp);
 397
 398         memset(&wc, 0, sizeof(wc));
 399         wc.qp = &qp->ibqp;
 400         wc.opcode = IB_WC_RECV;
 401
 402         if (test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags)) {
 403                 wc.wr_id = qp->r_wr_id;
 404                 wc.status = err;
 405                 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 406         }
 407         wc.status = IB_WC_WR_FLUSH_ERR;
 408
 409         if (qp->r_rq.wq) {
 410                 struct ipath_rwq *wq;
 411                 u32 head;
 412                 u32 tail;
 413
 414                 spin_lock(&qp->r_rq.lock);
 415
 416                 /* sanity check pointers before trusting them */
 417                 wq = qp->r_rq.wq;
 418                 head = wq->head;
 419                 if (head >= qp->r_rq.size)
 420                         head = 0;
 421                 tail = wq->tail;
 422                 if (tail >= qp->r_rq.size)
 423                         tail = 0;
 424                 while (tail != head) {
 425                         wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
 426                         if (++tail >= qp->r_rq.size)
 427                                 tail = 0;
 428                         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 429                 }
 430                 wq->tail = tail;
 431
 432                 spin_unlock(&qp->r_rq.lock);
 433         } else if (qp->ibqp.event_handler)
 434                 ret = 1;
 435
 436 bail:
 437         return ret;
 438 }
 439
 440 /**
 441  * ipath_modify_qp - modify the attributes of a queue pair
 442  * @ibqp: the queue pair who's attributes we're modifying
 443  * @attr: the new attributes
 444  * @attr_mask: the mask of attributes to modify
 445  * @udata: user data for ipathverbs.so
 446  *
 447  * Returns 0 on success, otherwise returns an errno.
 448  */
 449 int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 450                     int attr_mask, struct ib_udata *udata)
 451 {
 452         struct ipath_ibdev *dev = to_idev(ibqp->device);
 453         struct ipath_qp *qp = to_iqp(ibqp);
 454         enum ib_qp_state cur_state, new_state;
 455         int lastwqe = 0;
 456         int ret;
 457
 458         spin_lock_irq(&qp->s_lock);
 459
 460         cur_state = attr_mask & IB_QP_CUR_STATE ?
 461                 attr->cur_qp_state : qp->state;
 462         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
 463
 464         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
 465                                 attr_mask))
 466                 goto inval;
 467
 468         if (attr_mask & IB_QP_AV) {
 469                 if (attr->ah_attr.dlid == 0 ||
 470                     attr->ah_attr.dlid >= IPATH_MULTICAST_LID_BASE)
 471                         goto inval;
 472
 473                 if ((attr->ah_attr.ah_flags & IB_AH_GRH) &&
 474                     (attr->ah_attr.grh.sgid_index > 1))
 475                         goto inval;
 476         }
 477
 478         if (attr_mask & IB_QP_PKEY_INDEX)
 479                 if (attr->pkey_index >= ipath_get_npkeys(dev->dd))
 480                         goto inval;
 481
 482         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 483                 if (attr->min_rnr_timer > 31)
 484                         goto inval;
 485
 486         if (attr_mask & IB_QP_PORT)
 487                 if (attr->port_num == 0 ||
 488                     attr->port_num > ibqp->device->phys_port_cnt)
 489                         goto inval;
 490
 491         /*
 492          * don't allow invalid Path MTU values or greater than 2048
 493          * unless we are configured for a 4KB MTU
 494          */
 495         if ((attr_mask & IB_QP_PATH_MTU) &&
 496                 (ib_mtu_enum_to_int(attr->path_mtu) == -1 ||
 497                 (attr->path_mtu > IB_MTU_2048 && !ipath_mtu4096)))
 498                 goto inval;
 499
 500         if (attr_mask & IB_QP_PATH_MIG_STATE)
 501                 if (attr->path_mig_state != IB_MIG_MIGRATED &&
 502                     attr->path_mig_state != IB_MIG_REARM)
 503                         goto inval;
 504
 505         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 506                 if (attr->max_dest_rd_atomic > IPATH_MAX_RDMA_ATOMIC)
 507                         goto inval;
 508
 509         switch (new_state) {
 510         case IB_QPS_RESET:
 511                 if (qp->state != IB_QPS_RESET) {
 512                         qp->state = IB_QPS_RESET;
 513                         spin_lock(&dev->pending_lock);
 514                         if (!list_empty(&qp->timerwait))
 515                                 list_del_init(&qp->timerwait);
 516                         if (!list_empty(&qp->piowait))
 517                                 list_del_init(&qp->piowait);
 518                         spin_unlock(&dev->pending_lock);
 519                         qp->s_flags &= ~IPATH_S_ANY_WAIT;
 520                         spin_unlock_irq(&qp->s_lock);
 521                         /* Stop the sending tasklet */
 522                         tasklet_kill(&qp->s_task);
 523                         wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
 524                         spin_lock_irq(&qp->s_lock);
 525                 }
 526                 ipath_reset_qp(qp, ibqp->qp_type);
 527                 break;
 528
 529         case IB_QPS_SQD:
 530                 qp->s_draining = qp->s_last != qp->s_cur;
 531                 qp->state = new_state;
 532                 break;
 533
 534         case IB_QPS_SQE:
 535                 if (qp->ibqp.qp_type == IB_QPT_RC)
 536                         goto inval;
 537                 qp->state = new_state;
 538                 break;
 539
 540         case IB_QPS_ERR:
 541                 lastwqe = ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 542                 break;
 543
 544         default:
 545                 qp->state = new_state;
 546                 break;
 547         }
 548
 549         if (attr_mask & IB_QP_PKEY_INDEX)
 550                 qp->s_pkey_index = attr->pkey_index;
 551
 552         if (attr_mask & IB_QP_DEST_QPN)
 553                 qp->remote_qpn = attr->dest_qp_num;
 554
 555         if (attr_mask & IB_QP_SQ_PSN) {
 556                 qp->s_psn = qp->s_next_psn = attr->sq_psn;
 557                 qp->s_last_psn = qp->s_next_psn - 1;
 558         }
 559
 560         if (attr_mask & IB_QP_RQ_PSN)
 561                 qp->r_psn = attr->rq_psn;
 562
 563         if (attr_mask & IB_QP_ACCESS_FLAGS)
 564                 qp->qp_access_flags = attr->qp_access_flags;
 565
 566         if (attr_mask & IB_QP_AV) {
 567                 qp->remote_ah_attr = attr->ah_attr;
 568                 qp->s_dmult = ipath_ib_rate_to_mult(attr->ah_attr.static_rate);
 569         }
 570
 571         if (attr_mask & IB_QP_PATH_MTU)
 572                 qp->path_mtu = attr->path_mtu;
 573
 574         if (attr_mask & IB_QP_RETRY_CNT)
 575                 qp->s_retry = qp->s_retry_cnt = attr->retry_cnt;
 576
 577         if (attr_mask & IB_QP_RNR_RETRY) {
 578                 qp->s_rnr_retry = attr->rnr_retry;
 579                 if (qp->s_rnr_retry > 7)
 580                         qp->s_rnr_retry = 7;
 581                 qp->s_rnr_retry_cnt = qp->s_rnr_retry;
 582         }
 583
 584         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 585                 qp->r_min_rnr_timer = attr->min_rnr_timer;
 586
 587         if (attr_mask & IB_QP_TIMEOUT)
 588                 qp->timeout = attr->timeout;
 589
 590         if (attr_mask & IB_QP_QKEY)
 591                 qp->qkey = attr->qkey;
 592
 593         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 594                 qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
 595
 596         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
 597                 qp->s_max_rd_atomic = attr->max_rd_atomic;
 598
 599         spin_unlock_irq(&qp->s_lock);
 600
 601         if (lastwqe) {
 602                 struct ib_event ev;
 603
 604                 ev.device = qp->ibqp.device;
 605                 ev.element.qp = &qp->ibqp;
 606                 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 607                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 608         }
 609         ret = 0;
 610         goto bail;
 611
 612 inval:
 613         spin_unlock_irq(&qp->s_lock);
 614         ret = -EINVAL;
 615
 616 bail:
 617         return ret;
 618 }
 619
 620 int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 621                    int attr_mask, struct ib_qp_init_attr *init_attr)
 622 {
 623         struct ipath_qp *qp = to_iqp(ibqp);
 624
 625         attr->qp_state = qp->state;
 626         attr->cur_qp_state = attr->qp_state;
 627         attr->path_mtu = qp->path_mtu;
 628         attr->path_mig_state = 0;
 629         attr->qkey = qp->qkey;
 630         attr->rq_psn = qp->r_psn;
 631         attr->sq_psn = qp->s_next_psn;
 632         attr->dest_qp_num = qp->remote_qpn;
 633         attr->qp_access_flags = qp->qp_access_flags;
 634         attr->cap.max_send_wr = qp->s_size - 1;
 635         attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
 636         attr->cap.max_send_sge = qp->s_max_sge;
 637         attr->cap.max_recv_sge = qp->r_rq.max_sge;
 638         attr->cap.max_inline_data = 0;
 639         attr->ah_attr = qp->remote_ah_attr;
 640         memset(&attr->alt_ah_attr, 0, sizeof(attr->alt_ah_attr));
 641         attr->pkey_index = qp->s_pkey_index;
 642         attr->alt_pkey_index = 0;
 643         attr->en_sqd_async_notify = 0;
 644         attr->sq_draining = qp->s_draining;
 645         attr->max_rd_atomic = qp->s_max_rd_atomic;
 646         attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
 647         attr->min_rnr_timer = qp->r_min_rnr_timer;
 648         attr->port_num = 1;
 649         attr->timeout = qp->timeout;
 650         attr->retry_cnt = qp->s_retry_cnt;
 651         attr->rnr_retry = qp->s_rnr_retry_cnt;
 652         attr->alt_port_num = 0;
 653         attr->alt_timeout = 0;
 654
 655         init_attr->event_handler = qp->ibqp.event_handler;
 656         init_attr->qp_context = qp->ibqp.qp_context;
 657         init_attr->send_cq = qp->ibqp.send_cq;
 658         init_attr->recv_cq = qp->ibqp.recv_cq;
 659         init_attr->srq = qp->ibqp.srq;
 660         init_attr->cap = attr->cap;
 661         if (qp->s_flags & IPATH_S_SIGNAL_REQ_WR)
 662                 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 663         else
 664                 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
 665         init_attr->qp_type = qp->ibqp.qp_type;
 666         init_attr->port_num = 1;
 667         return 0;
 668 }
 669
 670 /**
 671  * ipath_compute_aeth - compute the AETH (syndrome + MSN)
 672  * @qp: the queue pair to compute the AETH for
 673  *
 674  * Returns the AETH.
 675  */
 676 __be32 ipath_compute_aeth(struct ipath_qp *qp)
 677 {
 678         u32 aeth = qp->r_msn & IPATH_MSN_MASK;
 679
 680         if (qp->ibqp.srq) {
 681                 /*
 682                  * Shared receive queues don't generate credits.
 683                  * Set the credit field to the invalid value.
 684                  */
 685                 aeth |= IPATH_AETH_CREDIT_INVAL << IPATH_AETH_CREDIT_SHIFT;
 686         } else {
 687                 u32 min, max, x;
 688                 u32 credits;
 689                 struct ipath_rwq *wq = qp->r_rq.wq;
 690                 u32 head;
 691                 u32 tail;
 692
 693                 /* sanity check pointers before trusting them */
 694                 head = wq->head;
 695                 if (head >= qp->r_rq.size)
 696                         head = 0;
 697                 tail = wq->tail;
 698                 if (tail >= qp->r_rq.size)
 699                         tail = 0;
 700                 /*
 701                  * Compute the number of credits available (RWQEs).
 702                  * XXX Not holding the r_rq.lock here so there is a small
 703                  * chance that the pair of reads are not atomic.
 704                  */
 705                 credits = head - tail;
 706                 if ((int)credits < 0)
 707                         credits += qp->r_rq.size;
 708                 /*
 709                  * Binary search the credit table to find the code to
 710                  * use.
 711                  */
 712                 min = 0;
 713                 max = 31;
 714                 for (;;) {
 715                         x = (min + max) / 2;
 716                         if (credit_table[x] == credits)
 717                                 break;
 718                         if (credit_table[x] > credits)
 719                                 max = x;
 720                         else if (min == x)
 721                                 break;
 722                         else
 723                                 min = x;
 724                 }
 725                 aeth |= x << IPATH_AETH_CREDIT_SHIFT;
 726         }
 727         return cpu_to_be32(aeth);
 728 }
 729
 730 /**
 731  * ipath_create_qp - create a queue pair for a device
 732  * @ibpd: the protection domain who's device we create the queue pair for
 733  * @init_attr: the attributes of the queue pair
 734  * @udata: unused by InfiniPath
 735  *
 736  * Returns the queue pair on success, otherwise returns an errno.
 737  *
 738  * Called by the ib_create_qp() core verbs function.
 739  */
 740 struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
 741                               struct ib_qp_init_attr *init_attr,
 742                               struct ib_udata *udata)
 743 {
 744         struct ipath_qp *qp;
 745         int err;
 746         struct ipath_swqe *swq = NULL;
 747         struct ipath_ibdev *dev;
 748         size_t sz;
 749         size_t sg_list_sz;
 750         struct ib_qp *ret;
 751
 752         if (init_attr->create_flags) {
 753                 ret = ERR_PTR(-EINVAL);
 754                 goto bail;
 755         }
 756
 757         if (init_attr->cap.max_send_sge > ib_ipath_max_sges ||
 758             init_attr->cap.max_send_wr > ib_ipath_max_qp_wrs) {
 759                 ret = ERR_PTR(-EINVAL);
 760                 goto bail;
 761         }
 762
 763         /* Check receive queue parameters if no SRQ is specified. */
 764         if (!init_attr->srq) {
 765                 if (init_attr->cap.max_recv_sge > ib_ipath_max_sges ||
 766                     init_attr->cap.max_recv_wr > ib_ipath_max_qp_wrs) {
 767                         ret = ERR_PTR(-EINVAL);
 768                         goto bail;
 769                 }
 770                 if (init_attr->cap.max_send_sge +
 771                     init_attr->cap.max_send_wr +
 772                     init_attr->cap.max_recv_sge +
 773                     init_attr->cap.max_recv_wr == 0) {
 774                         ret = ERR_PTR(-EINVAL);
 775                         goto bail;
 776                 }
 777         }
 778
 779         switch (init_attr->qp_type) {
 780         case IB_QPT_UC:
 781         case IB_QPT_RC:
 782         case IB_QPT_UD:
 783         case IB_QPT_SMI:
 784         case IB_QPT_GSI:
 785                 sz = sizeof(struct ipath_sge) *
 786                         init_attr->cap.max_send_sge +
 787                         sizeof(struct ipath_swqe);
 788                 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
 789                 if (swq == NULL) {
 790                         ret = ERR_PTR(-ENOMEM);
 791                         goto bail;
 792                 }
 793                 sz = sizeof(*qp);
 794                 sg_list_sz = 0;
 795                 if (init_attr->srq) {
 796                         struct ipath_srq *srq = to_isrq(init_attr->srq);
 797
 798                         if (srq->rq.max_sge > 1)
 799                                 sg_list_sz = sizeof(*qp->r_sg_list) *
 800                                         (srq->rq.max_sge - 1);
 801                 } else if (init_attr->cap.max_recv_sge > 1)
 802                         sg_list_sz = sizeof(*qp->r_sg_list) *
 803                                 (init_attr->cap.max_recv_sge - 1);
 804                 qp = kmalloc(sz + sg_list_sz, GFP_KERNEL);
 805                 if (!qp) {
 806                         ret = ERR_PTR(-ENOMEM);
 807                         goto bail_swq;
 808                 }
 809                 if (sg_list_sz && (init_attr->qp_type == IB_QPT_UD ||
 810                     init_attr->qp_type == IB_QPT_SMI ||
 811                     init_attr->qp_type == IB_QPT_GSI)) {
 812                         qp->r_ud_sg_list = kmalloc(sg_list_sz, GFP_KERNEL);
 813                         if (!qp->r_ud_sg_list) {
 814                                 ret = ERR_PTR(-ENOMEM);
 815                                 goto bail_qp;
 816                         }
 817                 } else
 818                         qp->r_ud_sg_list = NULL;
 819                 if (init_attr->srq) {
 820                         sz = 0;
 821                         qp->r_rq.size = 0;
 822                         qp->r_rq.max_sge = 0;
 823                         qp->r_rq.wq = NULL;
 824                         init_attr->cap.max_recv_wr = 0;
 825                         init_attr->cap.max_recv_sge = 0;
 826                 } else {
 827                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
 828                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
 829                         sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
 830                                 sizeof(struct ipath_rwqe);
 831                         qp->r_rq.wq = vmalloc_user(sizeof(struct ipath_rwq) +
 832                                               qp->r_rq.size * sz);
 833                         if (!qp->r_rq.wq) {
 834                                 ret = ERR_PTR(-ENOMEM);
 835                                 goto bail_sg_list;
 836                         }
 837                 }
 838
 839                 /*
 840                  * ib_create_qp() will initialize qp->ibqp
 841                  * except for qp->ibqp.qp_num.
 842                  */
 843                 spin_lock_init(&qp->s_lock);
 844                 spin_lock_init(&qp->r_rq.lock);
 845                 atomic_set(&qp->refcount, 0);
 846                 init_waitqueue_head(&qp->wait);
 847                 init_waitqueue_head(&qp->wait_dma);
 848                 tasklet_init(&qp->s_task, ipath_do_send, (unsigned long)qp);
 849                 INIT_LIST_HEAD(&qp->piowait);
 850                 INIT_LIST_HEAD(&qp->timerwait);
 851                 qp->state = IB_QPS_RESET;
 852                 qp->s_wq = swq;
 853                 qp->s_size = init_attr->cap.max_send_wr + 1;
 854                 qp->s_max_sge = init_attr->cap.max_send_sge;
 855                 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
 856                         qp->s_flags = IPATH_S_SIGNAL_REQ_WR;
 857                 else
 858                         qp->s_flags = 0;
 859                 dev = to_idev(ibpd->device);
 860                 err = ipath_alloc_qpn(&dev->qp_table, qp,
 861                                       init_attr->qp_type);
 862                 if (err) {
 863                         ret = ERR_PTR(err);
 864                         vfree(qp->r_rq.wq);
 865                         goto bail_sg_list;
 866                 }
 867                 qp->ip = NULL;
 868                 qp->s_tx = NULL;
 869                 ipath_reset_qp(qp, init_attr->qp_type);
 870                 break;
 871
 872         default:
 873                 /* Don't support raw QPs */
 874                 ret = ERR_PTR(-ENOSYS);
 875                 goto bail;
 876         }
 877
 878         init_attr->cap.max_inline_data = 0;
 879
 880         /*
 881          * Return the address of the RWQ as the offset to mmap.
 882          * See ipath_mmap() for details.
 883          */
 884         if (udata && udata->outlen >= sizeof(__u64)) {
 885                 if (!qp->r_rq.wq) {
 886                         __u64 offset = 0;
 887
 888                         err = ib_copy_to_udata(udata, &offset,
 889                                                sizeof(offset));
 890                         if (err) {
 891                                 ret = ERR_PTR(err);
 892                                 goto bail_ip;
 893                         }
 894                 } else {
 895                         u32 s = sizeof(struct ipath_rwq) +
 896                                 qp->r_rq.size * sz;
 897
 898                         qp->ip =
 899                             ipath_create_mmap_info(dev, s,
 900                                                    ibpd->uobject->context,
 901                                                    qp->r_rq.wq);
 902                         if (!qp->ip) {
 903                                 ret = ERR_PTR(-ENOMEM);
 904                                 goto bail_ip;
 905                         }
 906
 907                         err = ib_copy_to_udata(udata, &(qp->ip->offset),
 908                                                sizeof(qp->ip->offset));
 909                         if (err) {
 910                                 ret = ERR_PTR(err);
 911                                 goto bail_ip;
 912                         }
 913                 }
 914         }
 915
 916         spin_lock(&dev->n_qps_lock);
 917         if (dev->n_qps_allocated == ib_ipath_max_qps) {
 918                 spin_unlock(&dev->n_qps_lock);
 919                 ret = ERR_PTR(-ENOMEM);
 920                 goto bail_ip;
 921         }
 922
 923         dev->n_qps_allocated++;
 924         spin_unlock(&dev->n_qps_lock);
 925
 926         if (qp->ip) {
 927                 spin_lock_irq(&dev->pending_lock);
 928                 list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps);
 929                 spin_unlock_irq(&dev->pending_lock);
 930         }
 931
 932         ret = &qp->ibqp;
 933         goto bail;
 934
 935 bail_ip:
 936         if (qp->ip)
 937                 kref_put(&qp->ip->ref, ipath_release_mmap_info);
 938         else
 939                 vfree(qp->r_rq.wq);
 940         ipath_free_qp(&dev->qp_table, qp);
 941         free_qpn(&dev->qp_table, qp->ibqp.qp_num);
 942 bail_sg_list:
 943         kfree(qp->r_ud_sg_list);
 944 bail_qp:
 945         kfree(qp);
 946 bail_swq:
 947         vfree(swq);
 948 bail:
 949         return ret;
 950 }
 951
 952 /**
 953  * ipath_destroy_qp - destroy a queue pair
 954  * @ibqp: the queue pair to destroy
 955  *
 956  * Returns 0 on success.
 957  *
 958  * Note that this can be called while the QP is actively sending or
 959  * receiving!
 960  */
 961 int ipath_destroy_qp(struct ib_qp *ibqp)
 962 {
 963         struct ipath_qp *qp = to_iqp(ibqp);
 964         struct ipath_ibdev *dev = to_idev(ibqp->device);
 965
 966         /* Make sure HW and driver activity is stopped. */
 967         spin_lock_irq(&qp->s_lock);
 968         if (qp->state != IB_QPS_RESET) {
 969                 qp->state = IB_QPS_RESET;
 970                 spin_lock(&dev->pending_lock);
 971                 if (!list_empty(&qp->timerwait))
 972                         list_del_init(&qp->timerwait);
 973                 if (!list_empty(&qp->piowait))
 974                         list_del_init(&qp->piowait);
 975                 spin_unlock(&dev->pending_lock);
 976                 qp->s_flags &= ~IPATH_S_ANY_WAIT;
 977                 spin_unlock_irq(&qp->s_lock);
 978                 /* Stop the sending tasklet */
 979                 tasklet_kill(&qp->s_task);
 980                 wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
 981         } else
 982                 spin_unlock_irq(&qp->s_lock);
 983
 984         ipath_free_qp(&dev->qp_table, qp);
 985
 986         if (qp->s_tx) {
 987                 atomic_dec(&qp->refcount);
 988                 if (qp->s_tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEBUF)
 989                         kfree(qp->s_tx->txreq.map_addr);
 990                 spin_lock_irq(&dev->pending_lock);
 991                 list_add(&qp->s_tx->txreq.list, &dev->txreq_free);
 992                 spin_unlock_irq(&dev->pending_lock);
 993                 qp->s_tx = NULL;
 994         }
 995
 996         wait_event(qp->wait, !atomic_read(&qp->refcount));
 997
 998         /* all user's cleaned up, mark it available */
 999         free_qpn(&dev->qp_table, qp->ibqp.qp_num);
1000         spin_lock(&dev->n_qps_lock);
1001         dev->n_qps_allocated--;
1002         spin_unlock(&dev->n_qps_lock);
1003
1004         if (qp->ip)
1005                 kref_put(&qp->ip->ref, ipath_release_mmap_info);
1006         else
1007                 vfree(qp->r_rq.wq);
1008         kfree(qp->r_ud_sg_list);
1009         vfree(qp->s_wq);
1010         kfree(qp);
1011         return 0;
1012 }
1013
1014 /**
1015  * ipath_init_qp_table - initialize the QP table for a device
1016  * @idev: the device who's QP table we're initializing
1017  * @size: the size of the QP table
1018  *
1019  * Returns 0 on success, otherwise returns an errno.
1020  */
1021 int ipath_init_qp_table(struct ipath_ibdev *idev, int size)
1022 {
1023         int i;
1024         int ret;
1025
1026         idev->qp_table.last = 1;        /* QPN 0 and 1 are special. */
1027         idev->qp_table.max = size;
1028         idev->qp_table.nmaps = 1;
1029         idev->qp_table.table = kzalloc(size * sizeof(*idev->qp_table.table),
1030                                        GFP_KERNEL);
1031         if (idev->qp_table.table == NULL) {
1032                 ret = -ENOMEM;
1033                 goto bail;
1034         }
1035
1036         for (i = 0; i < ARRAY_SIZE(idev->qp_table.map); i++) {
1037                 atomic_set(&idev->qp_table.map[i].n_free, BITS_PER_PAGE);
1038                 idev->qp_table.map[i].page = NULL;
1039         }
1040
1041         ret = 0;
1042
1043 bail:
1044         return ret;
1045 }
1046
1047 /**
1048  * ipath_get_credit - flush the send work queue of a QP
1049  * @qp: the qp who's send work queue to flush
1050  * @aeth: the Acknowledge Extended Transport Header
1051  *
1052  * The QP s_lock should be held.
1053  */
1054 void ipath_get_credit(struct ipath_qp *qp, u32 aeth)
1055 {
1056         u32 credit = (aeth >> IPATH_AETH_CREDIT_SHIFT) & IPATH_AETH_CREDIT_MASK;
1057
1058         /*
1059          * If the credit is invalid, we can send
1060          * as many packets as we like.  Otherwise, we have to
1061          * honor the credit field.
1062          */
1063         if (credit == IPATH_AETH_CREDIT_INVAL)
1064                 qp->s_lsn = (u32) -1;
1065         else if (qp->s_lsn != (u32) -1) {
1066                 /* Compute new LSN (i.e., MSN + credit) */
1067                 credit = (aeth + credit_table[credit]) & IPATH_MSN_MASK;
1068                 if (ipath_cmp24(credit, qp->s_lsn) > 0)
1069                         qp->s_lsn = credit;
1070         }
1071
1072         /* Restart sending if it was blocked due to lack of credits. */
1073         if ((qp->s_flags & IPATH_S_WAIT_SSN_CREDIT) &&
1074             qp->s_cur != qp->s_head &&
1075             (qp->s_lsn == (u32) -1 ||
1076              ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn,
1077                          qp->s_lsn + 1) <= 0))
1078                 ipath_schedule_send(qp);
1079 }