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

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