SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_qp.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. 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 static u32 alloc_qpn(struct ipath_qp_table *qpt)
  85 {
  86         u32 i, offset, max_scan, qpn;
  87         struct qpn_map *map;
  88         u32 ret;
  89
  90         qpn = qpt->last + 1;
  91         if (qpn >= QPN_MAX)
  92                 qpn = 2;
  93         offset = qpn & BITS_PER_PAGE_MASK;
  94         map = &qpt->map[qpn / BITS_PER_PAGE];
  95         max_scan = qpt->nmaps - !offset;
  96         for (i = 0;;) {
  97                 if (unlikely(!map->page)) {
  98                         unsigned long page = get_zeroed_page(GFP_KERNEL);
  99                         unsigned long flags;
 100
 101                         /*
 102                          * Free the page if someone raced with us
 103                          * installing it:
 104                          */
 105                         spin_lock_irqsave(&qpt->lock, flags);
 106                         if (map->page)
 107                                 free_page(page);
 108                         else
 109                                 map->page = (void *)page;
 110                         spin_unlock_irqrestore(&qpt->lock, flags);
 111                         if (unlikely(!map->page))
 112                                 break;
 113                 }
 114                 if (likely(atomic_read(&map->n_free))) {
 115                         do {
 116                                 if (!test_and_set_bit(offset, map->page)) {
 117                                         atomic_dec(&map->n_free);
 118                                         qpt->last = qpn;
 119                                         ret = qpn;
 120                                         goto bail;
 121                                 }
 122                                 offset = find_next_offset(map, offset);
 123                                 qpn = mk_qpn(qpt, map, offset);
 124                                 /*
 125                                  * This test differs from alloc_pidmap().
 126                                  * If find_next_offset() does find a zero
 127                                  * bit, we don't need to check for QPN
 128                                  * wrapping around past our starting QPN.
 129                                  * We just need to be sure we don't loop
 130                                  * forever.
 131                                  */
 132                         } while (offset < BITS_PER_PAGE && qpn < QPN_MAX);
 133                 }
 134                 /*
 135                  * In order to keep the number of pages allocated to a
 136                  * minimum, we scan the all existing pages before increasing
 137                  * the size of the bitmap table.
 138                  */
 139                 if (++i > max_scan) {
 140                         if (qpt->nmaps == QPNMAP_ENTRIES)
 141                                 break;
 142                         map = &qpt->map[qpt->nmaps++];
 143                         offset = 0;
 144                 } else if (map < &qpt->map[qpt->nmaps]) {
 145                         ++map;
 146                         offset = 0;
 147                 } else {
 148                         map = &qpt->map[0];
 149                         offset = 2;
 150                 }
 151                 qpn = mk_qpn(qpt, map, offset);
 152         }
 153
 154         ret = 0;
 155
 156 bail:
 157         return ret;
 158 }
 159
 160 static void free_qpn(struct ipath_qp_table *qpt, u32 qpn)
 161 {
 162         struct qpn_map *map;
 163
 164         map = qpt->map + qpn / BITS_PER_PAGE;
 165         if (map->page)
 166                 clear_bit(qpn & BITS_PER_PAGE_MASK, map->page);
 167         atomic_inc(&map->n_free);
 168 }
 169
 170 /**
 171  * ipath_alloc_qpn - allocate a QP number
 172  * @qpt: the QP table
 173  * @qp: the QP
 174  * @type: the QP type (IB_QPT_SMI and IB_QPT_GSI are special)
 175  *
 176  * Allocate the next available QPN and put the QP into the hash table.
 177  * The hash table holds a reference to the QP.
 178  */
 179 static int ipath_alloc_qpn(struct ipath_qp_table *qpt, struct ipath_qp *qp,
 180                            enum ib_qp_type type)
 181 {
 182         unsigned long flags;
 183         u32 qpn;
 184         int ret;
 185
 186         if (type == IB_QPT_SMI)
 187                 qpn = 0;
 188         else if (type == IB_QPT_GSI)
 189                 qpn = 1;
 190         else {
 191                 /* Allocate the next available QPN */
 192                 qpn = alloc_qpn(qpt);
 193                 if (qpn == 0) {
 194                         ret = -ENOMEM;
 195                         goto bail;
 196                 }
 197         }
 198         qp->ibqp.qp_num = qpn;
 199
 200         /* Add the QP to the hash table. */
 201         spin_lock_irqsave(&qpt->lock, flags);
 202
 203         qpn %= qpt->max;
 204         qp->next = qpt->table[qpn];
 205         qpt->table[qpn] = qp;
 206         atomic_inc(&qp->refcount);
 207
 208         spin_unlock_irqrestore(&qpt->lock, flags);
 209         ret = 0;
 210
 211 bail:
 212         return ret;
 213 }
 214
 215 /**
 216  * ipath_free_qp - remove a QP from the QP table
 217  * @qpt: the QP table
 218  * @qp: the QP to remove
 219  *
 220  * Remove the QP from the table so it can't be found asynchronously by
 221  * the receive interrupt routine.
 222  */
 223 static void ipath_free_qp(struct ipath_qp_table *qpt, struct ipath_qp *qp)
 224 {
 225         struct ipath_qp *q, **qpp;
 226         unsigned long flags;
 227         int fnd = 0;
 228
 229         spin_lock_irqsave(&qpt->lock, flags);
 230
 231         /* Remove QP from the hash table. */
 232         qpp = &qpt->table[qp->ibqp.qp_num % qpt->max];
 233         for (; (q = *qpp) != NULL; qpp = &q->next) {
 234                 if (q == qp) {
 235                         *qpp = qp->next;
 236                         qp->next = NULL;
 237                         atomic_dec(&qp->refcount);
 238                         fnd = 1;
 239                         break;
 240                 }
 241         }
 242
 243         spin_unlock_irqrestore(&qpt->lock, flags);
 244
 245         if (!fnd)
 246                 return;
 247
 248         /* If QPN is not reserved, mark QPN free in the bitmap. */
 249         if (qp->ibqp.qp_num > 1)
 250                 free_qpn(qpt, qp->ibqp.qp_num);
 251
 252         wait_event(qp->wait, !atomic_read(&qp->refcount));
 253 }
 254
 255 /**
 256  * ipath_free_all_qps - remove all QPs from the table
 257  * @qpt: the QP table to empty
 258  */
 259 void ipath_free_all_qps(struct ipath_qp_table *qpt)
 260 {
 261         unsigned long flags;
 262         struct ipath_qp *qp, *nqp;
 263         u32 n;
 264
 265         for (n = 0; n < qpt->max; n++) {
 266                 spin_lock_irqsave(&qpt->lock, flags);
 267                 qp = qpt->table[n];
 268                 qpt->table[n] = NULL;
 269                 spin_unlock_irqrestore(&qpt->lock, flags);
 270
 271                 while (qp) {
 272                         nqp = qp->next;
 273                         if (qp->ibqp.qp_num > 1)
 274                                 free_qpn(qpt, qp->ibqp.qp_num);
 275                         if (!atomic_dec_and_test(&qp->refcount) ||
 276                             !ipath_destroy_qp(&qp->ibqp))
 277                                 _VERBS_INFO("QP memory leak!\n");
 278                         qp = nqp;
 279                 }
 280         }
 281
 282         for (n = 0; n < ARRAY_SIZE(qpt->map); n++) {
 283                 if (qpt->map[n].page)
 284                         free_page((unsigned long)qpt->map[n].page);
 285         }
 286 }
 287
 288 /**
 289  * ipath_lookup_qpn - return the QP with the given QPN
 290  * @qpt: the QP table
 291  * @qpn: the QP number to look up
 292  *
 293  * The caller is responsible for decrementing the QP reference count
 294  * when done.
 295  */
 296 struct ipath_qp *ipath_lookup_qpn(struct ipath_qp_table *qpt, u32 qpn)
 297 {
 298         unsigned long flags;
 299         struct ipath_qp *qp;
 300
 301         spin_lock_irqsave(&qpt->lock, flags);
 302
 303         for (qp = qpt->table[qpn % qpt->max]; qp; qp = qp->next) {
 304                 if (qp->ibqp.qp_num == qpn) {
 305                         atomic_inc(&qp->refcount);
 306                         break;
 307                 }
 308         }
 309
 310         spin_unlock_irqrestore(&qpt->lock, flags);
 311         return qp;
 312 }
 313
 314 /**
 315  * ipath_reset_qp - initialize the QP state to the reset state
 316  * @qp: the QP to reset
 317  */
 318 static void ipath_reset_qp(struct ipath_qp *qp)
 319 {
 320         qp->remote_qpn = 0;
 321         qp->qkey = 0;
 322         qp->qp_access_flags = 0;
 323         clear_bit(IPATH_S_BUSY, &qp->s_flags);
 324         qp->s_hdrwords = 0;
 325         qp->s_psn = 0;
 326         qp->r_psn = 0;
 327         qp->r_msn = 0;
 328         if (qp->ibqp.qp_type == IB_QPT_RC) {
 329                 qp->s_state = IB_OPCODE_RC_SEND_LAST;
 330                 qp->r_state = IB_OPCODE_RC_SEND_LAST;
 331         } else {
 332                 qp->s_state = IB_OPCODE_UC_SEND_LAST;
 333                 qp->r_state = IB_OPCODE_UC_SEND_LAST;
 334         }
 335         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 336         qp->r_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 337         qp->r_nak_state = 0;
 338         qp->s_rnr_timeout = 0;
 339         qp->s_head = 0;
 340         qp->s_tail = 0;
 341         qp->s_cur = 0;
 342         qp->s_last = 0;
 343         qp->s_ssn = 1;
 344         qp->s_lsn = 0;
 345         if (qp->r_rq.wq) {
 346                 qp->r_rq.wq->head = 0;
 347                 qp->r_rq.wq->tail = 0;
 348         }
 349         qp->r_reuse_sge = 0;
 350 }
 351
 352 /**
 353  * ipath_error_qp - put a QP into an error state
 354  * @qp: the QP to put into an error state
 355  *
 356  * Flushes both send and receive work queues.
 357  * QP s_lock should be held and interrupts disabled.
 358  */
 359
 360 void ipath_error_qp(struct ipath_qp *qp)
 361 {
 362         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 363         struct ib_wc wc;
 364
 365         _VERBS_INFO("QP%d/%d in error state\n",
 366                     qp->ibqp.qp_num, qp->remote_qpn);
 367
 368         spin_lock(&dev->pending_lock);
 369         /* XXX What if its already removed by the timeout code? */
 370         if (!list_empty(&qp->timerwait))
 371                 list_del_init(&qp->timerwait);
 372         if (!list_empty(&qp->piowait))
 373                 list_del_init(&qp->piowait);
 374         spin_unlock(&dev->pending_lock);
 375
 376         wc.status = IB_WC_WR_FLUSH_ERR;
 377         wc.vendor_err = 0;
 378         wc.byte_len = 0;
 379         wc.imm_data = 0;
 380         wc.qp_num = qp->ibqp.qp_num;
 381         wc.src_qp = 0;
 382         wc.wc_flags = 0;
 383         wc.pkey_index = 0;
 384         wc.slid = 0;
 385         wc.sl = 0;
 386         wc.dlid_path_bits = 0;
 387         wc.port_num = 0;
 388
 389         while (qp->s_last != qp->s_head) {
 390                 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 391
 392                 wc.wr_id = wqe->wr.wr_id;
 393                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 394                 if (++qp->s_last >= qp->s_size)
 395                         qp->s_last = 0;
 396                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
 397         }
 398         qp->s_cur = qp->s_tail = qp->s_head;
 399         qp->s_hdrwords = 0;
 400         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 401
 402         if (qp->r_rq.wq) {
 403                 struct ipath_rwq *wq;
 404                 u32 head;
 405                 u32 tail;
 406
 407                 spin_lock(&qp->r_rq.lock);
 408
 409                 /* sanity check pointers before trusting them */
 410                 wq = qp->r_rq.wq;
 411                 head = wq->head;
 412                 if (head >= qp->r_rq.size)
 413                         head = 0;
 414                 tail = wq->tail;
 415                 if (tail >= qp->r_rq.size)
 416                         tail = 0;
 417                 wc.opcode = IB_WC_RECV;
 418                 while (tail != head) {
 419                         wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
 420                         if (++tail >= qp->r_rq.size)
 421                                 tail = 0;
 422                         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 423                 }
 424                 wq->tail = tail;
 425
 426                 spin_unlock(&qp->r_rq.lock);
 427         }
 428 }
 429
 430 /**
 431  * ipath_modify_qp - modify the attributes of a queue pair
 432  * @ibqp: the queue pair who's attributes we're modifying
 433  * @attr: the new attributes
 434  * @attr_mask: the mask of attributes to modify
 435  * @udata: user data for ipathverbs.so
 436  *
 437  * Returns 0 on success, otherwise returns an errno.
 438  */
 439 int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 440                     int attr_mask, struct ib_udata *udata)
 441 {
 442         struct ipath_ibdev *dev = to_idev(ibqp->device);
 443         struct ipath_qp *qp = to_iqp(ibqp);
 444         enum ib_qp_state cur_state, new_state;
 445         unsigned long flags;
 446         int ret;
 447
 448         spin_lock_irqsave(&qp->s_lock, flags);
 449
 450         cur_state = attr_mask & IB_QP_CUR_STATE ?
 451                 attr->cur_qp_state : qp->state;
 452         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
 453
 454         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
 455                                 attr_mask))
 456                 goto inval;
 457
 458         if (attr_mask & IB_QP_AV)
 459                 if (attr->ah_attr.dlid == 0 ||
 460                     attr->ah_attr.dlid >= IPATH_MULTICAST_LID_BASE)
 461                         goto inval;
 462
 463         if (attr_mask & IB_QP_PKEY_INDEX)
 464                 if (attr->pkey_index >= ipath_get_npkeys(dev->dd))
 465                         goto inval;
 466
 467         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 468                 if (attr->min_rnr_timer > 31)
 469                         goto inval;
 470
 471         switch (new_state) {
 472         case IB_QPS_RESET:
 473                 ipath_reset_qp(qp);
 474                 break;
 475
 476         case IB_QPS_ERR:
 477                 ipath_error_qp(qp);
 478                 break;
 479
 480         default:
 481                 break;
 482
 483         }
 484
 485         if (attr_mask & IB_QP_PKEY_INDEX)
 486                 qp->s_pkey_index = attr->pkey_index;
 487
 488         if (attr_mask & IB_QP_DEST_QPN)
 489                 qp->remote_qpn = attr->dest_qp_num;
 490
 491         if (attr_mask & IB_QP_SQ_PSN) {
 492                 qp->s_next_psn = attr->sq_psn;
 493                 qp->s_last_psn = qp->s_next_psn - 1;
 494         }
 495
 496         if (attr_mask & IB_QP_RQ_PSN)
 497                 qp->r_psn = attr->rq_psn;
 498
 499         if (attr_mask & IB_QP_ACCESS_FLAGS)
 500                 qp->qp_access_flags = attr->qp_access_flags;
 501
 502         if (attr_mask & IB_QP_AV)
 503                 qp->remote_ah_attr = attr->ah_attr;
 504
 505         if (attr_mask & IB_QP_PATH_MTU)
 506                 qp->path_mtu = attr->path_mtu;
 507
 508         if (attr_mask & IB_QP_RETRY_CNT)
 509                 qp->s_retry = qp->s_retry_cnt = attr->retry_cnt;
 510
 511         if (attr_mask & IB_QP_RNR_RETRY) {
 512                 qp->s_rnr_retry = attr->rnr_retry;
 513                 if (qp->s_rnr_retry > 7)
 514                         qp->s_rnr_retry = 7;
 515                 qp->s_rnr_retry_cnt = qp->s_rnr_retry;
 516         }
 517
 518         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 519                 qp->r_min_rnr_timer = attr->min_rnr_timer;
 520
 521         if (attr_mask & IB_QP_QKEY)
 522                 qp->qkey = attr->qkey;
 523
 524         qp->state = new_state;
 525         spin_unlock_irqrestore(&qp->s_lock, flags);
 526
 527         ret = 0;
 528         goto bail;
 529
 530 inval:
 531         spin_unlock_irqrestore(&qp->s_lock, flags);
 532         ret = -EINVAL;
 533
 534 bail:
 535         return ret;
 536 }
 537
 538 int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 539                    int attr_mask, struct ib_qp_init_attr *init_attr)
 540 {
 541         struct ipath_qp *qp = to_iqp(ibqp);
 542
 543         attr->qp_state = qp->state;
 544         attr->cur_qp_state = attr->qp_state;
 545         attr->path_mtu = qp->path_mtu;
 546         attr->path_mig_state = 0;
 547         attr->qkey = qp->qkey;
 548         attr->rq_psn = qp->r_psn;
 549         attr->sq_psn = qp->s_next_psn;
 550         attr->dest_qp_num = qp->remote_qpn;
 551         attr->qp_access_flags = qp->qp_access_flags;
 552         attr->cap.max_send_wr = qp->s_size - 1;
 553         attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
 554         attr->cap.max_send_sge = qp->s_max_sge;
 555         attr->cap.max_recv_sge = qp->r_rq.max_sge;
 556         attr->cap.max_inline_data = 0;
 557         attr->ah_attr = qp->remote_ah_attr;
 558         memset(&attr->alt_ah_attr, 0, sizeof(attr->alt_ah_attr));
 559         attr->pkey_index = qp->s_pkey_index;
 560         attr->alt_pkey_index = 0;
 561         attr->en_sqd_async_notify = 0;
 562         attr->sq_draining = 0;
 563         attr->max_rd_atomic = 1;
 564         attr->max_dest_rd_atomic = 1;
 565         attr->min_rnr_timer = qp->r_min_rnr_timer;
 566         attr->port_num = 1;
 567         attr->timeout = 0;
 568         attr->retry_cnt = qp->s_retry_cnt;
 569         attr->rnr_retry = qp->s_rnr_retry;
 570         attr->alt_port_num = 0;
 571         attr->alt_timeout = 0;
 572
 573         init_attr->event_handler = qp->ibqp.event_handler;
 574         init_attr->qp_context = qp->ibqp.qp_context;
 575         init_attr->send_cq = qp->ibqp.send_cq;
 576         init_attr->recv_cq = qp->ibqp.recv_cq;
 577         init_attr->srq = qp->ibqp.srq;
 578         init_attr->cap = attr->cap;
 579         init_attr->sq_sig_type =
 580                 (qp->s_flags & (1 << IPATH_S_SIGNAL_REQ_WR))
 581                 ? IB_SIGNAL_REQ_WR : 0;
 582         init_attr->qp_type = qp->ibqp.qp_type;
 583         init_attr->port_num = 1;
 584         return 0;
 585 }
 586
 587 /**
 588  * ipath_compute_aeth - compute the AETH (syndrome + MSN)
 589  * @qp: the queue pair to compute the AETH for
 590  *
 591  * Returns the AETH.
 592  */
 593 __be32 ipath_compute_aeth(struct ipath_qp *qp)
 594 {
 595         u32 aeth = qp->r_msn & IPATH_MSN_MASK;
 596
 597         if (qp->ibqp.srq) {
 598                 /*
 599                  * Shared receive queues don't generate credits.
 600                  * Set the credit field to the invalid value.
 601                  */
 602                 aeth |= IPATH_AETH_CREDIT_INVAL << IPATH_AETH_CREDIT_SHIFT;
 603         } else {
 604                 u32 min, max, x;
 605                 u32 credits;
 606                 struct ipath_rwq *wq = qp->r_rq.wq;
 607                 u32 head;
 608                 u32 tail;
 609
 610                 /* sanity check pointers before trusting them */
 611                 head = wq->head;
 612                 if (head >= qp->r_rq.size)
 613                         head = 0;
 614                 tail = wq->tail;
 615                 if (tail >= qp->r_rq.size)
 616                         tail = 0;
 617                 /*
 618                  * Compute the number of credits available (RWQEs).
 619                  * XXX Not holding the r_rq.lock here so there is a small
 620                  * chance that the pair of reads are not atomic.
 621                  */
 622                 credits = head - tail;
 623                 if ((int)credits < 0)
 624                         credits += qp->r_rq.size;
 625                 /*
 626                  * Binary search the credit table to find the code to
 627                  * use.
 628                  */
 629                 min = 0;
 630                 max = 31;
 631                 for (;;) {
 632                         x = (min + max) / 2;
 633                         if (credit_table[x] == credits)
 634                                 break;
 635                         if (credit_table[x] > credits)
 636                                 max = x;
 637                         else if (min == x)
 638                                 break;
 639                         else
 640                                 min = x;
 641                 }
 642                 aeth |= x << IPATH_AETH_CREDIT_SHIFT;
 643         }
 644         return cpu_to_be32(aeth);
 645 }
 646
 647 /**
 648  * set_verbs_flags - set the verbs layer flags
 649  * @dd: the infinipath device
 650  * @flags: the flags to set
 651  */
 652 static int set_verbs_flags(struct ipath_devdata *dd, unsigned flags)
 653 {
 654         struct ipath_devdata *ss;
 655         unsigned long lflags;
 656
 657         spin_lock_irqsave(&ipath_devs_lock, lflags);
 658
 659         list_for_each_entry(ss, &ipath_dev_list, ipath_list) {
 660                 if (!(ss->ipath_flags & IPATH_INITTED))
 661                         continue;
 662                 if ((flags & IPATH_VERBS_KERNEL_SMA) &&
 663                     !(*ss->ipath_statusp & IPATH_STATUS_SMA))
 664                         *ss->ipath_statusp |= IPATH_STATUS_OIB_SMA;
 665                 else
 666                         *ss->ipath_statusp &= ~IPATH_STATUS_OIB_SMA;
 667         }
 668
 669         spin_unlock_irqrestore(&ipath_devs_lock, lflags);
 670
 671         return 0;
 672 }
 673
 674 /**
 675  * ipath_create_qp - create a queue pair for a device
 676  * @ibpd: the protection domain who's device we create the queue pair for
 677  * @init_attr: the attributes of the queue pair
 678  * @udata: unused by InfiniPath
 679  *
 680  * Returns the queue pair on success, otherwise returns an errno.
 681  *
 682  * Called by the ib_create_qp() core verbs function.
 683  */
 684 struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
 685                               struct ib_qp_init_attr *init_attr,
 686                               struct ib_udata *udata)
 687 {
 688         struct ipath_qp *qp;
 689         int err;
 690         struct ipath_swqe *swq = NULL;
 691         struct ipath_ibdev *dev;
 692         size_t sz;
 693         struct ib_qp *ret;
 694
 695         if (init_attr->cap.max_send_sge > ib_ipath_max_sges ||
 696             init_attr->cap.max_recv_sge > ib_ipath_max_sges ||
 697             init_attr->cap.max_send_wr > ib_ipath_max_qp_wrs ||
 698             init_attr->cap.max_recv_wr > ib_ipath_max_qp_wrs) {
 699                 ret = ERR_PTR(-ENOMEM);
 700                 goto bail;
 701         }
 702
 703         if (init_attr->cap.max_send_sge +
 704             init_attr->cap.max_recv_sge +
 705             init_attr->cap.max_send_wr +
 706             init_attr->cap.max_recv_wr == 0) {
 707                 ret = ERR_PTR(-EINVAL);
 708                 goto bail;
 709         }
 710
 711         switch (init_attr->qp_type) {
 712         case IB_QPT_UC:
 713         case IB_QPT_RC:
 714                 sz = sizeof(struct ipath_sge) *
 715                         init_attr->cap.max_send_sge +
 716                         sizeof(struct ipath_swqe);
 717                 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
 718                 if (swq == NULL) {
 719                         ret = ERR_PTR(-ENOMEM);
 720                         goto bail;
 721                 }
 722                 /* FALLTHROUGH */
 723         case IB_QPT_UD:
 724         case IB_QPT_SMI:
 725         case IB_QPT_GSI:
 726                 sz = sizeof(*qp);
 727                 if (init_attr->srq) {
 728                         struct ipath_srq *srq = to_isrq(init_attr->srq);
 729
 730                         sz += sizeof(*qp->r_sg_list) *
 731                                 srq->rq.max_sge;
 732                 } else
 733                         sz += sizeof(*qp->r_sg_list) *
 734                                 init_attr->cap.max_recv_sge;
 735                 qp = kmalloc(sz, GFP_KERNEL);
 736                 if (!qp) {
 737                         ret = ERR_PTR(-ENOMEM);
 738                         goto bail_swq;
 739                 }
 740                 if (init_attr->srq) {
 741                         sz = 0;
 742                         qp->r_rq.size = 0;
 743                         qp->r_rq.max_sge = 0;
 744                         qp->r_rq.wq = NULL;
 745                         init_attr->cap.max_recv_wr = 0;
 746                         init_attr->cap.max_recv_sge = 0;
 747                 } else {
 748                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
 749                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
 750                         sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
 751                                 sizeof(struct ipath_rwqe);
 752                         qp->r_rq.wq = vmalloc_user(sizeof(struct ipath_rwq) +
 753                                               qp->r_rq.size * sz);
 754                         if (!qp->r_rq.wq) {
 755                                 ret = ERR_PTR(-ENOMEM);
 756                                 goto bail_qp;
 757                         }
 758                 }
 759
 760                 /*
 761                  * ib_create_qp() will initialize qp->ibqp
 762                  * except for qp->ibqp.qp_num.
 763                  */
 764                 spin_lock_init(&qp->s_lock);
 765                 spin_lock_init(&qp->r_rq.lock);
 766                 atomic_set(&qp->refcount, 0);
 767                 init_waitqueue_head(&qp->wait);
 768                 tasklet_init(&qp->s_task, ipath_do_ruc_send,
 769                              (unsigned long)qp);
 770                 INIT_LIST_HEAD(&qp->piowait);
 771                 INIT_LIST_HEAD(&qp->timerwait);
 772                 qp->state = IB_QPS_RESET;
 773                 qp->s_wq = swq;
 774                 qp->s_size = init_attr->cap.max_send_wr + 1;
 775                 qp->s_max_sge = init_attr->cap.max_send_sge;
 776                 qp->s_flags = init_attr->sq_sig_type == IB_SIGNAL_REQ_WR ?
 777                         1 << IPATH_S_SIGNAL_REQ_WR : 0;
 778                 dev = to_idev(ibpd->device);
 779                 err = ipath_alloc_qpn(&dev->qp_table, qp,
 780                                       init_attr->qp_type);
 781                 if (err) {
 782                         ret = ERR_PTR(err);
 783                         goto bail_rwq;
 784                 }
 785                 qp->ip = NULL;
 786                 ipath_reset_qp(qp);
 787
 788                 /* Tell the core driver that the kernel SMA is present. */
 789                 if (init_attr->qp_type == IB_QPT_SMI)
 790                         set_verbs_flags(dev->dd, IPATH_VERBS_KERNEL_SMA);
 791                 break;
 792
 793         default:
 794                 /* Don't support raw QPs */
 795                 ret = ERR_PTR(-ENOSYS);
 796                 goto bail;
 797         }
 798
 799         init_attr->cap.max_inline_data = 0;
 800
 801         /*
 802          * Return the address of the RWQ as the offset to mmap.
 803          * See ipath_mmap() for details.
 804          */
 805         if (udata && udata->outlen >= sizeof(__u64)) {
 806                 struct ipath_mmap_info *ip;
 807                 __u64 offset = (__u64) qp->r_rq.wq;
 808                 int err;
 809
 810                 err = ib_copy_to_udata(udata, &offset, sizeof(offset));
 811                 if (err) {
 812                         ret = ERR_PTR(err);
 813                         goto bail_rwq;
 814                 }
 815
 816                 if (qp->r_rq.wq) {
 817                         /* Allocate info for ipath_mmap(). */
 818                         ip = kmalloc(sizeof(*ip), GFP_KERNEL);
 819                         if (!ip) {
 820                                 ret = ERR_PTR(-ENOMEM);
 821                                 goto bail_rwq;
 822                         }
 823                         qp->ip = ip;
 824                         ip->context = ibpd->uobject->context;
 825                         ip->obj = qp->r_rq.wq;
 826                         kref_init(&ip->ref);
 827                         ip->mmap_cnt = 0;
 828                         ip->size = PAGE_ALIGN(sizeof(struct ipath_rwq) +
 829                                               qp->r_rq.size * sz);
 830                         spin_lock_irq(&dev->pending_lock);
 831                         ip->next = dev->pending_mmaps;
 832                         dev->pending_mmaps = ip;
 833                         spin_unlock_irq(&dev->pending_lock);
 834                 }
 835         }
 836
 837         ret = &qp->ibqp;
 838         goto bail;
 839
 840 bail_rwq:
 841         vfree(qp->r_rq.wq);
 842 bail_qp:
 843         kfree(qp);
 844 bail_swq:
 845         vfree(swq);
 846 bail:
 847         return ret;
 848 }
 849
 850 /**
 851  * ipath_destroy_qp - destroy a queue pair
 852  * @ibqp: the queue pair to destroy
 853  *
 854  * Returns 0 on success.
 855  *
 856  * Note that this can be called while the QP is actively sending or
 857  * receiving!
 858  */
 859 int ipath_destroy_qp(struct ib_qp *ibqp)
 860 {
 861         struct ipath_qp *qp = to_iqp(ibqp);
 862         struct ipath_ibdev *dev = to_idev(ibqp->device);
 863         unsigned long flags;
 864
 865         /* Tell the core driver that the kernel SMA is gone. */
 866         if (qp->ibqp.qp_type == IB_QPT_SMI)
 867                 set_verbs_flags(dev->dd, 0);
 868
 869         spin_lock_irqsave(&qp->s_lock, flags);
 870         qp->state = IB_QPS_ERR;
 871         spin_unlock_irqrestore(&qp->s_lock, flags);
 872
 873         /* Stop the sending tasklet. */
 874         tasklet_kill(&qp->s_task);
 875
 876         /* Make sure the QP isn't on the timeout list. */
 877         spin_lock_irqsave(&dev->pending_lock, flags);
 878         if (!list_empty(&qp->timerwait))
 879                 list_del_init(&qp->timerwait);
 880         if (!list_empty(&qp->piowait))
 881                 list_del_init(&qp->piowait);
 882         spin_unlock_irqrestore(&dev->pending_lock, flags);
 883
 884         /*
 885          * Make sure that the QP is not in the QPN table so receive
 886          * interrupts will discard packets for this QP.  XXX Also remove QP
 887          * from multicast table.
 888          */
 889         if (atomic_read(&qp->refcount) != 0)
 890                 ipath_free_qp(&dev->qp_table, qp);
 891
 892         if (qp->ip)
 893                 kref_put(&qp->ip->ref, ipath_release_mmap_info);
 894         else
 895                 vfree(qp->r_rq.wq);
 896         vfree(qp->s_wq);
 897         kfree(qp);
 898         return 0;
 899 }
 900
 901 /**
 902  * ipath_init_qp_table - initialize the QP table for a device
 903  * @idev: the device who's QP table we're initializing
 904  * @size: the size of the QP table
 905  *
 906  * Returns 0 on success, otherwise returns an errno.
 907  */
 908 int ipath_init_qp_table(struct ipath_ibdev *idev, int size)
 909 {
 910         int i;
 911         int ret;
 912
 913         idev->qp_table.last = 1;        /* QPN 0 and 1 are special. */
 914         idev->qp_table.max = size;
 915         idev->qp_table.nmaps = 1;
 916         idev->qp_table.table = kzalloc(size * sizeof(*idev->qp_table.table),
 917                                        GFP_KERNEL);
 918         if (idev->qp_table.table == NULL) {
 919                 ret = -ENOMEM;
 920                 goto bail;
 921         }
 922
 923         for (i = 0; i < ARRAY_SIZE(idev->qp_table.map); i++) {
 924                 atomic_set(&idev->qp_table.map[i].n_free, BITS_PER_PAGE);
 925                 idev->qp_table.map[i].page = NULL;
 926         }
 927
 928         ret = 0;
 929
 930 bail:
 931         return ret;
 932 }
 933
 934 /**
 935  * ipath_sqerror_qp - put a QP's send queue into an error state
 936  * @qp: QP who's send queue will be put into an error state
 937  * @wc: the WC responsible for putting the QP in this state
 938  *
 939  * Flushes the send work queue.
 940  * The QP s_lock should be held.
 941  */
 942
 943 void ipath_sqerror_qp(struct ipath_qp *qp, struct ib_wc *wc)
 944 {
 945         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 946         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 947
 948         _VERBS_INFO("Send queue error on QP%d/%d: err: %d\n",
 949                     qp->ibqp.qp_num, qp->remote_qpn, wc->status);
 950
 951         spin_lock(&dev->pending_lock);
 952         /* XXX What if its already removed by the timeout code? */
 953         if (!list_empty(&qp->timerwait))
 954                 list_del_init(&qp->timerwait);
 955         if (!list_empty(&qp->piowait))
 956                 list_del_init(&qp->piowait);
 957         spin_unlock(&dev->pending_lock);
 958
 959         ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
 960         if (++qp->s_last >= qp->s_size)
 961                 qp->s_last = 0;
 962
 963         wc->status = IB_WC_WR_FLUSH_ERR;
 964
 965         while (qp->s_last != qp->s_head) {
 966                 wc->wr_id = wqe->wr.wr_id;
 967                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 968                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
 969                 if (++qp->s_last >= qp->s_size)
 970                         qp->s_last = 0;
 971                 wqe = get_swqe_ptr(qp, qp->s_last);
 972         }
 973         qp->s_cur = qp->s_tail = qp->s_head;
 974         qp->state = IB_QPS_SQE;
 975 }
 976
 977 /**
 978  * ipath_get_credit - flush the send work queue of a QP
 979  * @qp: the qp who's send work queue to flush
 980  * @aeth: the Acknowledge Extended Transport Header
 981  *
 982  * The QP s_lock should be held.
 983  */
 984 void ipath_get_credit(struct ipath_qp *qp, u32 aeth)
 985 {
 986         u32 credit = (aeth >> IPATH_AETH_CREDIT_SHIFT) & IPATH_AETH_CREDIT_MASK;
 987
 988         /*
 989          * If the credit is invalid, we can send
 990          * as many packets as we like.  Otherwise, we have to
 991          * honor the credit field.
 992          */
 993         if (credit == IPATH_AETH_CREDIT_INVAL)
 994                 qp->s_lsn = (u32) -1;
 995         else if (qp->s_lsn != (u32) -1) {
 996                 /* Compute new LSN (i.e., MSN + credit) */
 997                 credit = (aeth + credit_table[credit]) & IPATH_MSN_MASK;
 998                 if (ipath_cmp24(credit, qp->s_lsn) > 0)
 999                         qp->s_lsn = credit;
1000         }
1001
1002         /* Restart sending if it was blocked due to lack of credits. */
1003         if (qp->s_cur != qp->s_head &&
1004             (qp->s_lsn == (u32) -1 ||
1005              ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn,
1006                          qp->s_lsn + 1) <= 0))
1007                 tasklet_hi_schedule(&qp->s_task);
1008 }