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

   1 /*
   2  * Copyright (c) 2006, 2007 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         int fnd = 0;
 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                         fnd = 1;
 257                         break;
 258                 }
 259         }
 260
 261         spin_unlock_irqrestore(&qpt->lock, flags);
 262
 263         if (!fnd)
 264                 return;
 265
 266         free_qpn(qpt, qp->ibqp.qp_num);
 267
 268         wait_event(qp->wait, !atomic_read(&qp->refcount));
 269 }
 270
 271 /**
 272  * ipath_free_all_qps - remove all QPs from the table
 273  * @qpt: the QP table to empty
 274  */
 275 void ipath_free_all_qps(struct ipath_qp_table *qpt)
 276 {
 277         unsigned long flags;
 278         struct ipath_qp *qp, *nqp;
 279         u32 n;
 280
 281         for (n = 0; n < qpt->max; n++) {
 282                 spin_lock_irqsave(&qpt->lock, flags);
 283                 qp = qpt->table[n];
 284                 qpt->table[n] = NULL;
 285                 spin_unlock_irqrestore(&qpt->lock, flags);
 286
 287                 while (qp) {
 288                         nqp = qp->next;
 289                         free_qpn(qpt, qp->ibqp.qp_num);
 290                         if (!atomic_dec_and_test(&qp->refcount) ||
 291                             !ipath_destroy_qp(&qp->ibqp))
 292                                 ipath_dbg("QP memory leak!\n");
 293                         qp = nqp;
 294                 }
 295         }
 296
 297         for (n = 0; n < ARRAY_SIZE(qpt->map); n++) {
 298                 if (qpt->map[n].page)
 299                         free_page((unsigned long)qpt->map[n].page);
 300         }
 301 }
 302
 303 /**
 304  * ipath_lookup_qpn - return the QP with the given QPN
 305  * @qpt: the QP table
 306  * @qpn: the QP number to look up
 307  *
 308  * The caller is responsible for decrementing the QP reference count
 309  * when done.
 310  */
 311 struct ipath_qp *ipath_lookup_qpn(struct ipath_qp_table *qpt, u32 qpn)
 312 {
 313         unsigned long flags;
 314         struct ipath_qp *qp;
 315
 316         spin_lock_irqsave(&qpt->lock, flags);
 317
 318         for (qp = qpt->table[qpn % qpt->max]; qp; qp = qp->next) {
 319                 if (qp->ibqp.qp_num == qpn) {
 320                         atomic_inc(&qp->refcount);
 321                         break;
 322                 }
 323         }
 324
 325         spin_unlock_irqrestore(&qpt->lock, flags);
 326         return qp;
 327 }
 328
 329 /**
 330  * ipath_reset_qp - initialize the QP state to the reset state
 331  * @qp: the QP to reset
 332  */
 333 static void ipath_reset_qp(struct ipath_qp *qp)
 334 {
 335         qp->remote_qpn = 0;
 336         qp->qkey = 0;
 337         qp->qp_access_flags = 0;
 338         qp->s_busy = 0;
 339         qp->s_flags &= IPATH_S_SIGNAL_REQ_WR;
 340         qp->s_hdrwords = 0;
 341         qp->s_wqe = NULL;
 342         qp->s_psn = 0;
 343         qp->r_psn = 0;
 344         qp->r_msn = 0;
 345         if (qp->ibqp.qp_type == IB_QPT_RC) {
 346                 qp->s_state = IB_OPCODE_RC_SEND_LAST;
 347                 qp->r_state = IB_OPCODE_RC_SEND_LAST;
 348         } else {
 349                 qp->s_state = IB_OPCODE_UC_SEND_LAST;
 350                 qp->r_state = IB_OPCODE_UC_SEND_LAST;
 351         }
 352         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 353         qp->r_nak_state = 0;
 354         qp->r_wrid_valid = 0;
 355         qp->s_rnr_timeout = 0;
 356         qp->s_head = 0;
 357         qp->s_tail = 0;
 358         qp->s_cur = 0;
 359         qp->s_last = 0;
 360         qp->s_ssn = 1;
 361         qp->s_lsn = 0;
 362         qp->s_wait_credit = 0;
 363         memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue));
 364         qp->r_head_ack_queue = 0;
 365         qp->s_tail_ack_queue = 0;
 366         qp->s_num_rd_atomic = 0;
 367         if (qp->r_rq.wq) {
 368                 qp->r_rq.wq->head = 0;
 369                 qp->r_rq.wq->tail = 0;
 370         }
 371         qp->r_reuse_sge = 0;
 372 }
 373
 374 /**
 375  * ipath_error_qp - put a QP into an error state
 376  * @qp: the QP to put into an error state
 377  * @err: the receive completion error to signal if a RWQE is active
 378  *
 379  * Flushes both send and receive work queues.
 380  * Returns true if last WQE event should be generated.
 381  * The QP s_lock should be held and interrupts disabled.
 382  */
 383
 384 int ipath_error_qp(struct ipath_qp *qp, enum ib_wc_status err)
 385 {
 386         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 387         struct ib_wc wc;
 388         int ret = 0;
 389
 390         ipath_dbg("QP%d/%d in error state\n",
 391                   qp->ibqp.qp_num, qp->remote_qpn);
 392
 393         spin_lock(&dev->pending_lock);
 394         /* XXX What if its already removed by the timeout code? */
 395         if (!list_empty(&qp->timerwait))
 396                 list_del_init(&qp->timerwait);
 397         if (!list_empty(&qp->piowait))
 398                 list_del_init(&qp->piowait);
 399         spin_unlock(&dev->pending_lock);
 400
 401         wc.vendor_err = 0;
 402         wc.byte_len = 0;
 403         wc.imm_data = 0;
 404         wc.qp = &qp->ibqp;
 405         wc.src_qp = 0;
 406         wc.wc_flags = 0;
 407         wc.pkey_index = 0;
 408         wc.slid = 0;
 409         wc.sl = 0;
 410         wc.dlid_path_bits = 0;
 411         wc.port_num = 0;
 412         if (qp->r_wrid_valid) {
 413                 qp->r_wrid_valid = 0;
 414                 wc.wr_id = qp->r_wr_id;
 415                 wc.opcode = IB_WC_RECV;
 416                 wc.status = err;
 417                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
 418         }
 419         wc.status = IB_WC_WR_FLUSH_ERR;
 420
 421         while (qp->s_last != qp->s_head) {
 422                 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 423
 424                 wc.wr_id = wqe->wr.wr_id;
 425                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 426                 if (++qp->s_last >= qp->s_size)
 427                         qp->s_last = 0;
 428                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
 429         }
 430         qp->s_cur = qp->s_tail = qp->s_head;
 431         qp->s_hdrwords = 0;
 432         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 433
 434         if (qp->r_rq.wq) {
 435                 struct ipath_rwq *wq;
 436                 u32 head;
 437                 u32 tail;
 438
 439                 spin_lock(&qp->r_rq.lock);
 440
 441                 /* sanity check pointers before trusting them */
 442                 wq = qp->r_rq.wq;
 443                 head = wq->head;
 444                 if (head >= qp->r_rq.size)
 445                         head = 0;
 446                 tail = wq->tail;
 447                 if (tail >= qp->r_rq.size)
 448                         tail = 0;
 449                 wc.opcode = IB_WC_RECV;
 450                 while (tail != head) {
 451                         wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
 452                         if (++tail >= qp->r_rq.size)
 453                                 tail = 0;
 454                         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 455                 }
 456                 wq->tail = tail;
 457
 458                 spin_unlock(&qp->r_rq.lock);
 459         } else if (qp->ibqp.event_handler)
 460                 ret = 1;
 461
 462         return ret;
 463 }
 464
 465 /**
 466  * ipath_modify_qp - modify the attributes of a queue pair
 467  * @ibqp: the queue pair who's attributes we're modifying
 468  * @attr: the new attributes
 469  * @attr_mask: the mask of attributes to modify
 470  * @udata: user data for ipathverbs.so
 471  *
 472  * Returns 0 on success, otherwise returns an errno.
 473  */
 474 int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 475                     int attr_mask, struct ib_udata *udata)
 476 {
 477         struct ipath_ibdev *dev = to_idev(ibqp->device);
 478         struct ipath_qp *qp = to_iqp(ibqp);
 479         enum ib_qp_state cur_state, new_state;
 480         unsigned long flags;
 481         int lastwqe = 0;
 482         int ret;
 483
 484         spin_lock_irqsave(&qp->s_lock, flags);
 485
 486         cur_state = attr_mask & IB_QP_CUR_STATE ?
 487                 attr->cur_qp_state : qp->state;
 488         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
 489
 490         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
 491                                 attr_mask))
 492                 goto inval;
 493
 494         if (attr_mask & IB_QP_AV) {
 495                 if (attr->ah_attr.dlid == 0 ||
 496                     attr->ah_attr.dlid >= IPATH_MULTICAST_LID_BASE)
 497                         goto inval;
 498
 499                 if ((attr->ah_attr.ah_flags & IB_AH_GRH) &&
 500                     (attr->ah_attr.grh.sgid_index > 1))
 501                         goto inval;
 502         }
 503
 504         if (attr_mask & IB_QP_PKEY_INDEX)
 505                 if (attr->pkey_index >= ipath_get_npkeys(dev->dd))
 506                         goto inval;
 507
 508         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 509                 if (attr->min_rnr_timer > 31)
 510                         goto inval;
 511
 512         if (attr_mask & IB_QP_PORT)
 513                 if (attr->port_num == 0 ||
 514                     attr->port_num > ibqp->device->phys_port_cnt)
 515                         goto inval;
 516
 517         /*
 518          * Note: the chips support a maximum MTU of 4096, but the driver
 519          * hasn't implemented this feature yet, so don't allow Path MTU
 520          * values greater than 2048.
 521          */
 522         if (attr_mask & IB_QP_PATH_MTU)
 523                 if (attr->path_mtu > IB_MTU_2048)
 524                         goto inval;
 525
 526         if (attr_mask & IB_QP_PATH_MIG_STATE)
 527                 if (attr->path_mig_state != IB_MIG_MIGRATED &&
 528                     attr->path_mig_state != IB_MIG_REARM)
 529                         goto inval;
 530
 531         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 532                 if (attr->max_dest_rd_atomic > IPATH_MAX_RDMA_ATOMIC)
 533                         goto inval;
 534
 535         switch (new_state) {
 536         case IB_QPS_RESET:
 537                 ipath_reset_qp(qp);
 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                 break;
 546
 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
 569         if (attr_mask & IB_QP_PATH_MTU)
 570                 qp->path_mtu = attr->path_mtu;
 571
 572         if (attr_mask & IB_QP_RETRY_CNT)
 573                 qp->s_retry = qp->s_retry_cnt = attr->retry_cnt;
 574
 575         if (attr_mask & IB_QP_RNR_RETRY) {
 576                 qp->s_rnr_retry = attr->rnr_retry;
 577                 if (qp->s_rnr_retry > 7)
 578                         qp->s_rnr_retry = 7;
 579                 qp->s_rnr_retry_cnt = qp->s_rnr_retry;
 580         }
 581
 582         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 583                 qp->r_min_rnr_timer = attr->min_rnr_timer;
 584
 585         if (attr_mask & IB_QP_TIMEOUT)
 586                 qp->timeout = attr->timeout;
 587
 588         if (attr_mask & IB_QP_QKEY)
 589                 qp->qkey = attr->qkey;
 590
 591         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 592                 qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
 593
 594         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
 595                 qp->s_max_rd_atomic = attr->max_rd_atomic;
 596
 597         qp->state = new_state;
 598         spin_unlock_irqrestore(&qp->s_lock, flags);
 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_irqrestore(&qp->s_lock, flags);
 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 = 0;
 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;
 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         struct ib_qp *ret;
 749
 750         if (init_attr->cap.max_send_sge > ib_ipath_max_sges ||
 751             init_attr->cap.max_recv_sge > ib_ipath_max_sges ||
 752             init_attr->cap.max_send_wr > ib_ipath_max_qp_wrs ||
 753             init_attr->cap.max_recv_wr > ib_ipath_max_qp_wrs) {
 754                 ret = ERR_PTR(-ENOMEM);
 755                 goto bail;
 756         }
 757
 758         if (init_attr->cap.max_send_sge +
 759             init_attr->cap.max_recv_sge +
 760             init_attr->cap.max_send_wr +
 761             init_attr->cap.max_recv_wr == 0) {
 762                 ret = ERR_PTR(-EINVAL);
 763                 goto bail;
 764         }
 765
 766         switch (init_attr->qp_type) {
 767         case IB_QPT_UC:
 768         case IB_QPT_RC:
 769         case IB_QPT_UD:
 770         case IB_QPT_SMI:
 771         case IB_QPT_GSI:
 772                 sz = sizeof(struct ipath_sge) *
 773                         init_attr->cap.max_send_sge +
 774                         sizeof(struct ipath_swqe);
 775                 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
 776                 if (swq == NULL) {
 777                         ret = ERR_PTR(-ENOMEM);
 778                         goto bail;
 779                 }
 780                 sz = sizeof(*qp);
 781                 if (init_attr->srq) {
 782                         struct ipath_srq *srq = to_isrq(init_attr->srq);
 783
 784                         sz += sizeof(*qp->r_sg_list) *
 785                                 srq->rq.max_sge;
 786                 } else
 787                         sz += sizeof(*qp->r_sg_list) *
 788                                 init_attr->cap.max_recv_sge;
 789                 qp = kmalloc(sz, GFP_KERNEL);
 790                 if (!qp) {
 791                         ret = ERR_PTR(-ENOMEM);
 792                         goto bail_swq;
 793                 }
 794                 if (init_attr->srq) {
 795                         sz = 0;
 796                         qp->r_rq.size = 0;
 797                         qp->r_rq.max_sge = 0;
 798                         qp->r_rq.wq = NULL;
 799                         init_attr->cap.max_recv_wr = 0;
 800                         init_attr->cap.max_recv_sge = 0;
 801                 } else {
 802                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
 803                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
 804                         sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
 805                                 sizeof(struct ipath_rwqe);
 806                         qp->r_rq.wq = vmalloc_user(sizeof(struct ipath_rwq) +
 807                                               qp->r_rq.size * sz);
 808                         if (!qp->r_rq.wq) {
 809                                 ret = ERR_PTR(-ENOMEM);
 810                                 goto bail_qp;
 811                         }
 812                 }
 813
 814                 /*
 815                  * ib_create_qp() will initialize qp->ibqp
 816                  * except for qp->ibqp.qp_num.
 817                  */
 818                 spin_lock_init(&qp->s_lock);
 819                 spin_lock_init(&qp->r_rq.lock);
 820                 atomic_set(&qp->refcount, 0);
 821                 init_waitqueue_head(&qp->wait);
 822                 tasklet_init(&qp->s_task, ipath_do_send, (unsigned long)qp);
 823                 INIT_LIST_HEAD(&qp->piowait);
 824                 INIT_LIST_HEAD(&qp->timerwait);
 825                 qp->state = IB_QPS_RESET;
 826                 qp->s_wq = swq;
 827                 qp->s_size = init_attr->cap.max_send_wr + 1;
 828                 qp->s_max_sge = init_attr->cap.max_send_sge;
 829                 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
 830                         qp->s_flags = IPATH_S_SIGNAL_REQ_WR;
 831                 else
 832                         qp->s_flags = 0;
 833                 dev = to_idev(ibpd->device);
 834                 err = ipath_alloc_qpn(&dev->qp_table, qp,
 835                                       init_attr->qp_type);
 836                 if (err) {
 837                         ret = ERR_PTR(err);
 838                         vfree(qp->r_rq.wq);
 839                         goto bail_qp;
 840                 }
 841                 qp->ip = NULL;
 842                 ipath_reset_qp(qp);
 843                 break;
 844
 845         default:
 846                 /* Don't support raw QPs */
 847                 ret = ERR_PTR(-ENOSYS);
 848                 goto bail;
 849         }
 850
 851         init_attr->cap.max_inline_data = 0;
 852
 853         /*
 854          * Return the address of the RWQ as the offset to mmap.
 855          * See ipath_mmap() for details.
 856          */
 857         if (udata && udata->outlen >= sizeof(__u64)) {
 858                 int err;
 859
 860                 if (!qp->r_rq.wq) {
 861                         __u64 offset = 0;
 862
 863                         err = ib_copy_to_udata(udata, &offset,
 864                                                sizeof(offset));
 865                         if (err) {
 866                                 ret = ERR_PTR(err);
 867                                 goto bail_ip;
 868                         }
 869                 } else {
 870                         u32 s = sizeof(struct ipath_rwq) +
 871                                 qp->r_rq.size * sz;
 872
 873                         qp->ip =
 874                             ipath_create_mmap_info(dev, s,
 875                                                    ibpd->uobject->context,
 876                                                    qp->r_rq.wq);
 877                         if (!qp->ip) {
 878                                 ret = ERR_PTR(-ENOMEM);
 879                                 goto bail_ip;
 880                         }
 881
 882                         err = ib_copy_to_udata(udata, &(qp->ip->offset),
 883                                                sizeof(qp->ip->offset));
 884                         if (err) {
 885                                 ret = ERR_PTR(err);
 886                                 goto bail_ip;
 887                         }
 888                 }
 889         }
 890
 891         spin_lock(&dev->n_qps_lock);
 892         if (dev->n_qps_allocated == ib_ipath_max_qps) {
 893                 spin_unlock(&dev->n_qps_lock);
 894                 ret = ERR_PTR(-ENOMEM);
 895                 goto bail_ip;
 896         }
 897
 898         dev->n_qps_allocated++;
 899         spin_unlock(&dev->n_qps_lock);
 900
 901         if (qp->ip) {
 902                 spin_lock_irq(&dev->pending_lock);
 903                 list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps);
 904                 spin_unlock_irq(&dev->pending_lock);
 905         }
 906
 907         ret = &qp->ibqp;
 908         goto bail;
 909
 910 bail_ip:
 911         if (qp->ip)
 912                 kref_put(&qp->ip->ref, ipath_release_mmap_info);
 913         else
 914                 vfree(qp->r_rq.wq);
 915         ipath_free_qp(&dev->qp_table, qp);
 916 bail_qp:
 917         kfree(qp);
 918 bail_swq:
 919         vfree(swq);
 920 bail:
 921         return ret;
 922 }
 923
 924 /**
 925  * ipath_destroy_qp - destroy a queue pair
 926  * @ibqp: the queue pair to destroy
 927  *
 928  * Returns 0 on success.
 929  *
 930  * Note that this can be called while the QP is actively sending or
 931  * receiving!
 932  */
 933 int ipath_destroy_qp(struct ib_qp *ibqp)
 934 {
 935         struct ipath_qp *qp = to_iqp(ibqp);
 936         struct ipath_ibdev *dev = to_idev(ibqp->device);
 937         unsigned long flags;
 938
 939         spin_lock_irqsave(&qp->s_lock, flags);
 940         qp->state = IB_QPS_ERR;
 941         spin_unlock_irqrestore(&qp->s_lock, flags);
 942         spin_lock(&dev->n_qps_lock);
 943         dev->n_qps_allocated--;
 944         spin_unlock(&dev->n_qps_lock);
 945
 946         /* Stop the sending tasklet. */
 947         tasklet_kill(&qp->s_task);
 948
 949         /* Make sure the QP isn't on the timeout list. */
 950         spin_lock_irqsave(&dev->pending_lock, flags);
 951         if (!list_empty(&qp->timerwait))
 952                 list_del_init(&qp->timerwait);
 953         if (!list_empty(&qp->piowait))
 954                 list_del_init(&qp->piowait);
 955         spin_unlock_irqrestore(&dev->pending_lock, flags);
 956
 957         /*
 958          * Make sure that the QP is not in the QPN table so receive
 959          * interrupts will discard packets for this QP.  XXX Also remove QP
 960          * from multicast table.
 961          */
 962         if (atomic_read(&qp->refcount) != 0)
 963                 ipath_free_qp(&dev->qp_table, qp);
 964
 965         if (qp->ip)
 966                 kref_put(&qp->ip->ref, ipath_release_mmap_info);
 967         else
 968                 vfree(qp->r_rq.wq);
 969         vfree(qp->s_wq);
 970         kfree(qp);
 971         return 0;
 972 }
 973
 974 /**
 975  * ipath_init_qp_table - initialize the QP table for a device
 976  * @idev: the device who's QP table we're initializing
 977  * @size: the size of the QP table
 978  *
 979  * Returns 0 on success, otherwise returns an errno.
 980  */
 981 int ipath_init_qp_table(struct ipath_ibdev *idev, int size)
 982 {
 983         int i;
 984         int ret;
 985
 986         idev->qp_table.last = 1;        /* QPN 0 and 1 are special. */
 987         idev->qp_table.max = size;
 988         idev->qp_table.nmaps = 1;
 989         idev->qp_table.table = kzalloc(size * sizeof(*idev->qp_table.table),
 990                                        GFP_KERNEL);
 991         if (idev->qp_table.table == NULL) {
 992                 ret = -ENOMEM;
 993                 goto bail;
 994         }
 995
 996         for (i = 0; i < ARRAY_SIZE(idev->qp_table.map); i++) {
 997                 atomic_set(&idev->qp_table.map[i].n_free, BITS_PER_PAGE);
 998                 idev->qp_table.map[i].page = NULL;
 999         }
1000
1001         ret = 0;
1002
1003 bail:
1004         return ret;
1005 }
1006
1007 /**
1008  * ipath_sqerror_qp - put a QP's send queue into an error state
1009  * @qp: QP who's send queue will be put into an error state
1010  * @wc: the WC responsible for putting the QP in this state
1011  *
1012  * Flushes the send work queue.
1013  * The QP s_lock should be held and interrupts disabled.
1014  */
1015
1016 void ipath_sqerror_qp(struct ipath_qp *qp, struct ib_wc *wc)
1017 {
1018         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
1019         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
1020
1021         ipath_dbg("Send queue error on QP%d/%d: err: %d\n",
1022                   qp->ibqp.qp_num, qp->remote_qpn, wc->status);
1023
1024         spin_lock(&dev->pending_lock);
1025         /* XXX What if its already removed by the timeout code? */
1026         if (!list_empty(&qp->timerwait))
1027                 list_del_init(&qp->timerwait);
1028         if (!list_empty(&qp->piowait))
1029                 list_del_init(&qp->piowait);
1030         spin_unlock(&dev->pending_lock);
1031
1032         ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
1033         if (++qp->s_last >= qp->s_size)
1034                 qp->s_last = 0;
1035
1036         wc->status = IB_WC_WR_FLUSH_ERR;
1037
1038         while (qp->s_last != qp->s_head) {
1039                 wqe = get_swqe_ptr(qp, qp->s_last);
1040                 wc->wr_id = wqe->wr.wr_id;
1041                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1042                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
1043                 if (++qp->s_last >= qp->s_size)
1044                         qp->s_last = 0;
1045         }
1046         qp->s_cur = qp->s_tail = qp->s_head;
1047         qp->state = IB_QPS_SQE;
1048 }
1049
1050 /**
1051  * ipath_get_credit - flush the send work queue of a QP
1052  * @qp: the qp who's send work queue to flush
1053  * @aeth: the Acknowledge Extended Transport Header
1054  *
1055  * The QP s_lock should be held.
1056  */
1057 void ipath_get_credit(struct ipath_qp *qp, u32 aeth)
1058 {
1059         u32 credit = (aeth >> IPATH_AETH_CREDIT_SHIFT) & IPATH_AETH_CREDIT_MASK;
1060
1061         /*
1062          * If the credit is invalid, we can send
1063          * as many packets as we like.  Otherwise, we have to
1064          * honor the credit field.
1065          */
1066         if (credit == IPATH_AETH_CREDIT_INVAL)
1067                 qp->s_lsn = (u32) -1;
1068         else if (qp->s_lsn != (u32) -1) {
1069                 /* Compute new LSN (i.e., MSN + credit) */
1070                 credit = (aeth + credit_table[credit]) & IPATH_MSN_MASK;
1071                 if (ipath_cmp24(credit, qp->s_lsn) > 0)
1072                         qp->s_lsn = credit;
1073         }
1074
1075         /* Restart sending if it was blocked due to lack of credits. */
1076         if (qp->s_cur != qp->s_head &&
1077             (qp->s_lsn == (u32) -1 ||
1078              ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn,
1079                          qp->s_lsn + 1) <= 0))
1080                 tasklet_hi_schedule(&qp->s_task);
1081 }