SAFE public projects git trees. - safe/jmp/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  * The QP s_lock should be held and interrupts disabled.
 381  */
 382
 383 void ipath_error_qp(struct ipath_qp *qp, enum ib_wc_status err)
 384 {
 385         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 386         struct ib_wc wc;
 387
 388         ipath_dbg("QP%d/%d in error state\n",
 389                   qp->ibqp.qp_num, qp->remote_qpn);
 390
 391         spin_lock(&dev->pending_lock);
 392         /* XXX What if its already removed by the timeout code? */
 393         if (!list_empty(&qp->timerwait))
 394                 list_del_init(&qp->timerwait);
 395         if (!list_empty(&qp->piowait))
 396                 list_del_init(&qp->piowait);
 397         spin_unlock(&dev->pending_lock);
 398
 399         wc.vendor_err = 0;
 400         wc.byte_len = 0;
 401         wc.imm_data = 0;
 402         wc.qp = &qp->ibqp;
 403         wc.src_qp = 0;
 404         wc.wc_flags = 0;
 405         wc.pkey_index = 0;
 406         wc.slid = 0;
 407         wc.sl = 0;
 408         wc.dlid_path_bits = 0;
 409         wc.port_num = 0;
 410         if (qp->r_wrid_valid) {
 411                 qp->r_wrid_valid = 0;
 412                 wc.wr_id = qp->r_wr_id;
 413                 wc.opcode = IB_WC_RECV;
 414                 wc.status = err;
 415                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
 416         }
 417         wc.status = IB_WC_WR_FLUSH_ERR;
 418
 419         while (qp->s_last != qp->s_head) {
 420                 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 421
 422                 wc.wr_id = wqe->wr.wr_id;
 423                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 424                 if (++qp->s_last >= qp->s_size)
 425                         qp->s_last = 0;
 426                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
 427         }
 428         qp->s_cur = qp->s_tail = qp->s_head;
 429         qp->s_hdrwords = 0;
 430         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 431
 432         if (qp->r_rq.wq) {
 433                 struct ipath_rwq *wq;
 434                 u32 head;
 435                 u32 tail;
 436
 437                 spin_lock(&qp->r_rq.lock);
 438
 439                 /* sanity check pointers before trusting them */
 440                 wq = qp->r_rq.wq;
 441                 head = wq->head;
 442                 if (head >= qp->r_rq.size)
 443                         head = 0;
 444                 tail = wq->tail;
 445                 if (tail >= qp->r_rq.size)
 446                         tail = 0;
 447                 wc.opcode = IB_WC_RECV;
 448                 while (tail != head) {
 449                         wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
 450                         if (++tail >= qp->r_rq.size)
 451                                 tail = 0;
 452                         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 453                 }
 454                 wq->tail = tail;
 455
 456                 spin_unlock(&qp->r_rq.lock);
 457         }
 458 }
 459
 460 /**
 461  * ipath_modify_qp - modify the attributes of a queue pair
 462  * @ibqp: the queue pair who's attributes we're modifying
 463  * @attr: the new attributes
 464  * @attr_mask: the mask of attributes to modify
 465  * @udata: user data for ipathverbs.so
 466  *
 467  * Returns 0 on success, otherwise returns an errno.
 468  */
 469 int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 470                     int attr_mask, struct ib_udata *udata)
 471 {
 472         struct ipath_ibdev *dev = to_idev(ibqp->device);
 473         struct ipath_qp *qp = to_iqp(ibqp);
 474         enum ib_qp_state cur_state, new_state;
 475         unsigned long flags;
 476         int ret;
 477
 478         spin_lock_irqsave(&qp->s_lock, flags);
 479
 480         cur_state = attr_mask & IB_QP_CUR_STATE ?
 481                 attr->cur_qp_state : qp->state;
 482         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
 483
 484         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
 485                                 attr_mask))
 486                 goto inval;
 487
 488         if (attr_mask & IB_QP_AV) {
 489                 if (attr->ah_attr.dlid == 0 ||
 490                     attr->ah_attr.dlid >= IPATH_MULTICAST_LID_BASE)
 491                         goto inval;
 492
 493                 if ((attr->ah_attr.ah_flags & IB_AH_GRH) &&
 494                     (attr->ah_attr.grh.sgid_index > 1))
 495                         goto inval;
 496         }
 497
 498         if (attr_mask & IB_QP_PKEY_INDEX)
 499                 if (attr->pkey_index >= ipath_get_npkeys(dev->dd))
 500                         goto inval;
 501
 502         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 503                 if (attr->min_rnr_timer > 31)
 504                         goto inval;
 505
 506         if (attr_mask & IB_QP_PORT)
 507                 if (attr->port_num == 0 ||
 508                     attr->port_num > ibqp->device->phys_port_cnt)
 509                         goto inval;
 510
 511         /*
 512          * Note: the chips support a maximum MTU of 4096, but the driver
 513          * hasn't implemented this feature yet, so don't allow Path MTU
 514          * values greater than 2048.
 515          */
 516         if (attr_mask & IB_QP_PATH_MTU)
 517                 if (attr->path_mtu > IB_MTU_2048)
 518                         goto inval;
 519
 520         if (attr_mask & IB_QP_PATH_MIG_STATE)
 521                 if (attr->path_mig_state != IB_MIG_MIGRATED &&
 522                     attr->path_mig_state != IB_MIG_REARM)
 523                         goto inval;
 524
 525         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 526                 if (attr->max_dest_rd_atomic > IPATH_MAX_RDMA_ATOMIC)
 527                         goto inval;
 528
 529         switch (new_state) {
 530         case IB_QPS_RESET:
 531                 ipath_reset_qp(qp);
 532                 break;
 533
 534         case IB_QPS_ERR:
 535                 ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 536                 break;
 537
 538         default:
 539                 break;
 540
 541         }
 542
 543         if (attr_mask & IB_QP_PKEY_INDEX)
 544                 qp->s_pkey_index = attr->pkey_index;
 545
 546         if (attr_mask & IB_QP_DEST_QPN)
 547                 qp->remote_qpn = attr->dest_qp_num;
 548
 549         if (attr_mask & IB_QP_SQ_PSN) {
 550                 qp->s_psn = qp->s_next_psn = attr->sq_psn;
 551                 qp->s_last_psn = qp->s_next_psn - 1;
 552         }
 553
 554         if (attr_mask & IB_QP_RQ_PSN)
 555                 qp->r_psn = attr->rq_psn;
 556
 557         if (attr_mask & IB_QP_ACCESS_FLAGS)
 558                 qp->qp_access_flags = attr->qp_access_flags;
 559
 560         if (attr_mask & IB_QP_AV)
 561                 qp->remote_ah_attr = attr->ah_attr;
 562
 563         if (attr_mask & IB_QP_PATH_MTU)
 564                 qp->path_mtu = attr->path_mtu;
 565
 566         if (attr_mask & IB_QP_RETRY_CNT)
 567                 qp->s_retry = qp->s_retry_cnt = attr->retry_cnt;
 568
 569         if (attr_mask & IB_QP_RNR_RETRY) {
 570                 qp->s_rnr_retry = attr->rnr_retry;
 571                 if (qp->s_rnr_retry > 7)
 572                         qp->s_rnr_retry = 7;
 573                 qp->s_rnr_retry_cnt = qp->s_rnr_retry;
 574         }
 575
 576         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 577                 qp->r_min_rnr_timer = attr->min_rnr_timer;
 578
 579         if (attr_mask & IB_QP_TIMEOUT)
 580                 qp->timeout = attr->timeout;
 581
 582         if (attr_mask & IB_QP_QKEY)
 583                 qp->qkey = attr->qkey;
 584
 585         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 586                 qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
 587
 588         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
 589                 qp->s_max_rd_atomic = attr->max_rd_atomic;
 590
 591         qp->state = new_state;
 592         spin_unlock_irqrestore(&qp->s_lock, flags);
 593
 594         ret = 0;
 595         goto bail;
 596
 597 inval:
 598         spin_unlock_irqrestore(&qp->s_lock, flags);
 599         ret = -EINVAL;
 600
 601 bail:
 602         return ret;
 603 }
 604
 605 int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 606                    int attr_mask, struct ib_qp_init_attr *init_attr)
 607 {
 608         struct ipath_qp *qp = to_iqp(ibqp);
 609
 610         attr->qp_state = qp->state;
 611         attr->cur_qp_state = attr->qp_state;
 612         attr->path_mtu = qp->path_mtu;
 613         attr->path_mig_state = 0;
 614         attr->qkey = qp->qkey;
 615         attr->rq_psn = qp->r_psn;
 616         attr->sq_psn = qp->s_next_psn;
 617         attr->dest_qp_num = qp->remote_qpn;
 618         attr->qp_access_flags = qp->qp_access_flags;
 619         attr->cap.max_send_wr = qp->s_size - 1;
 620         attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
 621         attr->cap.max_send_sge = qp->s_max_sge;
 622         attr->cap.max_recv_sge = qp->r_rq.max_sge;
 623         attr->cap.max_inline_data = 0;
 624         attr->ah_attr = qp->remote_ah_attr;
 625         memset(&attr->alt_ah_attr, 0, sizeof(attr->alt_ah_attr));
 626         attr->pkey_index = qp->s_pkey_index;
 627         attr->alt_pkey_index = 0;
 628         attr->en_sqd_async_notify = 0;
 629         attr->sq_draining = 0;
 630         attr->max_rd_atomic = qp->s_max_rd_atomic;
 631         attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
 632         attr->min_rnr_timer = qp->r_min_rnr_timer;
 633         attr->port_num = 1;
 634         attr->timeout = qp->timeout;
 635         attr->retry_cnt = qp->s_retry_cnt;
 636         attr->rnr_retry = qp->s_rnr_retry;
 637         attr->alt_port_num = 0;
 638         attr->alt_timeout = 0;
 639
 640         init_attr->event_handler = qp->ibqp.event_handler;
 641         init_attr->qp_context = qp->ibqp.qp_context;
 642         init_attr->send_cq = qp->ibqp.send_cq;
 643         init_attr->recv_cq = qp->ibqp.recv_cq;
 644         init_attr->srq = qp->ibqp.srq;
 645         init_attr->cap = attr->cap;
 646         if (qp->s_flags & IPATH_S_SIGNAL_REQ_WR)
 647                 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 648         else
 649                 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
 650         init_attr->qp_type = qp->ibqp.qp_type;
 651         init_attr->port_num = 1;
 652         return 0;
 653 }
 654
 655 /**
 656  * ipath_compute_aeth - compute the AETH (syndrome + MSN)
 657  * @qp: the queue pair to compute the AETH for
 658  *
 659  * Returns the AETH.
 660  */
 661 __be32 ipath_compute_aeth(struct ipath_qp *qp)
 662 {
 663         u32 aeth = qp->r_msn & IPATH_MSN_MASK;
 664
 665         if (qp->ibqp.srq) {
 666                 /*
 667                  * Shared receive queues don't generate credits.
 668                  * Set the credit field to the invalid value.
 669                  */
 670                 aeth |= IPATH_AETH_CREDIT_INVAL << IPATH_AETH_CREDIT_SHIFT;
 671         } else {
 672                 u32 min, max, x;
 673                 u32 credits;
 674                 struct ipath_rwq *wq = qp->r_rq.wq;
 675                 u32 head;
 676                 u32 tail;
 677
 678                 /* sanity check pointers before trusting them */
 679                 head = wq->head;
 680                 if (head >= qp->r_rq.size)
 681                         head = 0;
 682                 tail = wq->tail;
 683                 if (tail >= qp->r_rq.size)
 684                         tail = 0;
 685                 /*
 686                  * Compute the number of credits available (RWQEs).
 687                  * XXX Not holding the r_rq.lock here so there is a small
 688                  * chance that the pair of reads are not atomic.
 689                  */
 690                 credits = head - tail;
 691                 if ((int)credits < 0)
 692                         credits += qp->r_rq.size;
 693                 /*
 694                  * Binary search the credit table to find the code to
 695                  * use.
 696                  */
 697                 min = 0;
 698                 max = 31;
 699                 for (;;) {
 700                         x = (min + max) / 2;
 701                         if (credit_table[x] == credits)
 702                                 break;
 703                         if (credit_table[x] > credits)
 704                                 max = x;
 705                         else if (min == x)
 706                                 break;
 707                         else
 708                                 min = x;
 709                 }
 710                 aeth |= x << IPATH_AETH_CREDIT_SHIFT;
 711         }
 712         return cpu_to_be32(aeth);
 713 }
 714
 715 /**
 716  * ipath_create_qp - create a queue pair for a device
 717  * @ibpd: the protection domain who's device we create the queue pair for
 718  * @init_attr: the attributes of the queue pair
 719  * @udata: unused by InfiniPath
 720  *
 721  * Returns the queue pair on success, otherwise returns an errno.
 722  *
 723  * Called by the ib_create_qp() core verbs function.
 724  */
 725 struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
 726                               struct ib_qp_init_attr *init_attr,
 727                               struct ib_udata *udata)
 728 {
 729         struct ipath_qp *qp;
 730         int err;
 731         struct ipath_swqe *swq = NULL;
 732         struct ipath_ibdev *dev;
 733         size_t sz;
 734         struct ib_qp *ret;
 735
 736         if (init_attr->cap.max_send_sge > ib_ipath_max_sges ||
 737             init_attr->cap.max_recv_sge > ib_ipath_max_sges ||
 738             init_attr->cap.max_send_wr > ib_ipath_max_qp_wrs ||
 739             init_attr->cap.max_recv_wr > ib_ipath_max_qp_wrs) {
 740                 ret = ERR_PTR(-ENOMEM);
 741                 goto bail;
 742         }
 743
 744         if (init_attr->cap.max_send_sge +
 745             init_attr->cap.max_recv_sge +
 746             init_attr->cap.max_send_wr +
 747             init_attr->cap.max_recv_wr == 0) {
 748                 ret = ERR_PTR(-EINVAL);
 749                 goto bail;
 750         }
 751
 752         switch (init_attr->qp_type) {
 753         case IB_QPT_UC:
 754         case IB_QPT_RC:
 755         case IB_QPT_UD:
 756         case IB_QPT_SMI:
 757         case IB_QPT_GSI:
 758                 sz = sizeof(struct ipath_sge) *
 759                         init_attr->cap.max_send_sge +
 760                         sizeof(struct ipath_swqe);
 761                 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
 762                 if (swq == NULL) {
 763                         ret = ERR_PTR(-ENOMEM);
 764                         goto bail;
 765                 }
 766                 sz = sizeof(*qp);
 767                 if (init_attr->srq) {
 768                         struct ipath_srq *srq = to_isrq(init_attr->srq);
 769
 770                         sz += sizeof(*qp->r_sg_list) *
 771                                 srq->rq.max_sge;
 772                 } else
 773                         sz += sizeof(*qp->r_sg_list) *
 774                                 init_attr->cap.max_recv_sge;
 775                 qp = kmalloc(sz, GFP_KERNEL);
 776                 if (!qp) {
 777                         ret = ERR_PTR(-ENOMEM);
 778                         goto bail_swq;
 779                 }
 780                 if (init_attr->srq) {
 781                         sz = 0;
 782                         qp->r_rq.size = 0;
 783                         qp->r_rq.max_sge = 0;
 784                         qp->r_rq.wq = NULL;
 785                         init_attr->cap.max_recv_wr = 0;
 786                         init_attr->cap.max_recv_sge = 0;
 787                 } else {
 788                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
 789                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
 790                         sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
 791                                 sizeof(struct ipath_rwqe);
 792                         qp->r_rq.wq = vmalloc_user(sizeof(struct ipath_rwq) +
 793                                               qp->r_rq.size * sz);
 794                         if (!qp->r_rq.wq) {
 795                                 ret = ERR_PTR(-ENOMEM);
 796                                 goto bail_qp;
 797                         }
 798                 }
 799
 800                 /*
 801                  * ib_create_qp() will initialize qp->ibqp
 802                  * except for qp->ibqp.qp_num.
 803                  */
 804                 spin_lock_init(&qp->s_lock);
 805                 spin_lock_init(&qp->r_rq.lock);
 806                 atomic_set(&qp->refcount, 0);
 807                 init_waitqueue_head(&qp->wait);
 808                 tasklet_init(&qp->s_task, ipath_do_send, (unsigned long)qp);
 809                 INIT_LIST_HEAD(&qp->piowait);
 810                 INIT_LIST_HEAD(&qp->timerwait);
 811                 qp->state = IB_QPS_RESET;
 812                 qp->s_wq = swq;
 813                 qp->s_size = init_attr->cap.max_send_wr + 1;
 814                 qp->s_max_sge = init_attr->cap.max_send_sge;
 815                 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
 816                         qp->s_flags = IPATH_S_SIGNAL_REQ_WR;
 817                 else
 818                         qp->s_flags = 0;
 819                 dev = to_idev(ibpd->device);
 820                 err = ipath_alloc_qpn(&dev->qp_table, qp,
 821                                       init_attr->qp_type);
 822                 if (err) {
 823                         ret = ERR_PTR(err);
 824                         goto bail_rwq;
 825                 }
 826                 qp->ip = NULL;
 827                 ipath_reset_qp(qp);
 828                 break;
 829
 830         default:
 831                 /* Don't support raw QPs */
 832                 ret = ERR_PTR(-ENOSYS);
 833                 goto bail;
 834         }
 835
 836         init_attr->cap.max_inline_data = 0;
 837
 838         /*
 839          * Return the address of the RWQ as the offset to mmap.
 840          * See ipath_mmap() for details.
 841          */
 842         if (udata && udata->outlen >= sizeof(__u64)) {
 843                 int err;
 844
 845                 if (!qp->r_rq.wq) {
 846                         __u64 offset = 0;
 847
 848                         err = ib_copy_to_udata(udata, &offset,
 849                                                sizeof(offset));
 850                         if (err) {
 851                                 ret = ERR_PTR(err);
 852                                 goto bail_rwq;
 853                         }
 854                 } else {
 855                         u32 s = sizeof(struct ipath_rwq) +
 856                                 qp->r_rq.size * sz;
 857
 858                         qp->ip =
 859                             ipath_create_mmap_info(dev, s,
 860                                                    ibpd->uobject->context,
 861                                                    qp->r_rq.wq);
 862                         if (!qp->ip) {
 863                                 ret = ERR_PTR(-ENOMEM);
 864                                 goto bail_rwq;
 865                         }
 866
 867                         err = ib_copy_to_udata(udata, &(qp->ip->offset),
 868                                                sizeof(qp->ip->offset));
 869                         if (err) {
 870                                 ret = ERR_PTR(err);
 871                                 goto bail_ip;
 872                         }
 873                 }
 874         }
 875
 876         spin_lock(&dev->n_qps_lock);
 877         if (dev->n_qps_allocated == ib_ipath_max_qps) {
 878                 spin_unlock(&dev->n_qps_lock);
 879                 ret = ERR_PTR(-ENOMEM);
 880                 goto bail_ip;
 881         }
 882
 883         dev->n_qps_allocated++;
 884         spin_unlock(&dev->n_qps_lock);
 885
 886         if (qp->ip) {
 887                 spin_lock_irq(&dev->pending_lock);
 888                 list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps);
 889                 spin_unlock_irq(&dev->pending_lock);
 890         }
 891
 892         ret = &qp->ibqp;
 893         goto bail;
 894
 895 bail_ip:
 896         kfree(qp->ip);
 897 bail_rwq:
 898         vfree(qp->r_rq.wq);
 899 bail_qp:
 900         kfree(qp);
 901 bail_swq:
 902         vfree(swq);
 903 bail:
 904         return ret;
 905 }
 906
 907 /**
 908  * ipath_destroy_qp - destroy a queue pair
 909  * @ibqp: the queue pair to destroy
 910  *
 911  * Returns 0 on success.
 912  *
 913  * Note that this can be called while the QP is actively sending or
 914  * receiving!
 915  */
 916 int ipath_destroy_qp(struct ib_qp *ibqp)
 917 {
 918         struct ipath_qp *qp = to_iqp(ibqp);
 919         struct ipath_ibdev *dev = to_idev(ibqp->device);
 920         unsigned long flags;
 921
 922         spin_lock_irqsave(&qp->s_lock, flags);
 923         qp->state = IB_QPS_ERR;
 924         spin_unlock_irqrestore(&qp->s_lock, flags);
 925         spin_lock(&dev->n_qps_lock);
 926         dev->n_qps_allocated--;
 927         spin_unlock(&dev->n_qps_lock);
 928
 929         /* Stop the sending tasklet. */
 930         tasklet_kill(&qp->s_task);
 931
 932         /* Make sure the QP isn't on the timeout list. */
 933         spin_lock_irqsave(&dev->pending_lock, flags);
 934         if (!list_empty(&qp->timerwait))
 935                 list_del_init(&qp->timerwait);
 936         if (!list_empty(&qp->piowait))
 937                 list_del_init(&qp->piowait);
 938         spin_unlock_irqrestore(&dev->pending_lock, flags);
 939
 940         /*
 941          * Make sure that the QP is not in the QPN table so receive
 942          * interrupts will discard packets for this QP.  XXX Also remove QP
 943          * from multicast table.
 944          */
 945         if (atomic_read(&qp->refcount) != 0)
 946                 ipath_free_qp(&dev->qp_table, qp);
 947
 948         if (qp->ip)
 949                 kref_put(&qp->ip->ref, ipath_release_mmap_info);
 950         else
 951                 vfree(qp->r_rq.wq);
 952         vfree(qp->s_wq);
 953         kfree(qp);
 954         return 0;
 955 }
 956
 957 /**
 958  * ipath_init_qp_table - initialize the QP table for a device
 959  * @idev: the device who's QP table we're initializing
 960  * @size: the size of the QP table
 961  *
 962  * Returns 0 on success, otherwise returns an errno.
 963  */
 964 int ipath_init_qp_table(struct ipath_ibdev *idev, int size)
 965 {
 966         int i;
 967         int ret;
 968
 969         idev->qp_table.last = 1;        /* QPN 0 and 1 are special. */
 970         idev->qp_table.max = size;
 971         idev->qp_table.nmaps = 1;
 972         idev->qp_table.table = kzalloc(size * sizeof(*idev->qp_table.table),
 973                                        GFP_KERNEL);
 974         if (idev->qp_table.table == NULL) {
 975                 ret = -ENOMEM;
 976                 goto bail;
 977         }
 978
 979         for (i = 0; i < ARRAY_SIZE(idev->qp_table.map); i++) {
 980                 atomic_set(&idev->qp_table.map[i].n_free, BITS_PER_PAGE);
 981                 idev->qp_table.map[i].page = NULL;
 982         }
 983
 984         ret = 0;
 985
 986 bail:
 987         return ret;
 988 }
 989
 990 /**
 991  * ipath_sqerror_qp - put a QP's send queue into an error state
 992  * @qp: QP who's send queue will be put into an error state
 993  * @wc: the WC responsible for putting the QP in this state
 994  *
 995  * Flushes the send work queue.
 996  * The QP s_lock should be held and interrupts disabled.
 997  */
 998
 999 void ipath_sqerror_qp(struct ipath_qp *qp, struct ib_wc *wc)
1000 {
1001         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
1002         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
1003
1004         ipath_dbg("Send queue error on QP%d/%d: err: %d\n",
1005                   qp->ibqp.qp_num, qp->remote_qpn, wc->status);
1006
1007         spin_lock(&dev->pending_lock);
1008         /* XXX What if its already removed by the timeout code? */
1009         if (!list_empty(&qp->timerwait))
1010                 list_del_init(&qp->timerwait);
1011         if (!list_empty(&qp->piowait))
1012                 list_del_init(&qp->piowait);
1013         spin_unlock(&dev->pending_lock);
1014
1015         ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
1016         if (++qp->s_last >= qp->s_size)
1017                 qp->s_last = 0;
1018
1019         wc->status = IB_WC_WR_FLUSH_ERR;
1020
1021         while (qp->s_last != qp->s_head) {
1022                 wqe = get_swqe_ptr(qp, qp->s_last);
1023                 wc->wr_id = wqe->wr.wr_id;
1024                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1025                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
1026                 if (++qp->s_last >= qp->s_size)
1027                         qp->s_last = 0;
1028         }
1029         qp->s_cur = qp->s_tail = qp->s_head;
1030         qp->state = IB_QPS_SQE;
1031 }
1032
1033 /**
1034  * ipath_get_credit - flush the send work queue of a QP
1035  * @qp: the qp who's send work queue to flush
1036  * @aeth: the Acknowledge Extended Transport Header
1037  *
1038  * The QP s_lock should be held.
1039  */
1040 void ipath_get_credit(struct ipath_qp *qp, u32 aeth)
1041 {
1042         u32 credit = (aeth >> IPATH_AETH_CREDIT_SHIFT) & IPATH_AETH_CREDIT_MASK;
1043
1044         /*
1045          * If the credit is invalid, we can send
1046          * as many packets as we like.  Otherwise, we have to
1047          * honor the credit field.
1048          */
1049         if (credit == IPATH_AETH_CREDIT_INVAL)
1050                 qp->s_lsn = (u32) -1;
1051         else if (qp->s_lsn != (u32) -1) {
1052                 /* Compute new LSN (i.e., MSN + credit) */
1053                 credit = (aeth + credit_table[credit]) & IPATH_MSN_MASK;
1054                 if (ipath_cmp24(credit, qp->s_lsn) > 0)
1055                         qp->s_lsn = credit;
1056         }
1057
1058         /* Restart sending if it was blocked due to lack of credits. */
1059         if (qp->s_cur != qp->s_head &&
1060             (qp->s_lsn == (u32) -1 ||
1061              ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn,
1062                          qp->s_lsn + 1) <= 0))
1063                 tasklet_hi_schedule(&qp->s_task);
1064 }