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  * @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_psn = 0;
 344         qp->r_psn = 0;
 345         qp->r_msn = 0;
 346         if (type == IB_QPT_RC) {
 347                 qp->s_state = IB_OPCODE_RC_SEND_LAST;
 348                 qp->r_state = IB_OPCODE_RC_SEND_LAST;
 349         } else {
 350                 qp->s_state = IB_OPCODE_UC_SEND_LAST;
 351                 qp->r_state = IB_OPCODE_UC_SEND_LAST;
 352         }
 353         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 354         qp->r_nak_state = 0;
 355         qp->r_wrid_valid = 0;
 356         qp->s_rnr_timeout = 0;
 357         qp->s_head = 0;
 358         qp->s_tail = 0;
 359         qp->s_cur = 0;
 360         qp->s_last = 0;
 361         qp->s_ssn = 1;
 362         qp->s_lsn = 0;
 363         qp->s_wait_credit = 0;
 364         memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue));
 365         qp->r_head_ack_queue = 0;
 366         qp->s_tail_ack_queue = 0;
 367         qp->s_num_rd_atomic = 0;
 368         if (qp->r_rq.wq) {
 369                 qp->r_rq.wq->head = 0;
 370                 qp->r_rq.wq->tail = 0;
 371         }
 372         qp->r_reuse_sge = 0;
 373 }
 374
 375 /**
 376  * ipath_error_qp - put a QP into an error state
 377  * @qp: the QP to put into an error state
 378  * @err: the receive completion error to signal if a RWQE is active
 379  *
 380  * Flushes both send and receive work queues.
 381  * Returns true if last WQE event should be generated.
 382  * The QP s_lock should be held and interrupts disabled.
 383  */
 384
 385 int ipath_error_qp(struct ipath_qp *qp, enum ib_wc_status err)
 386 {
 387         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 388         struct ib_wc wc;
 389         int ret = 0;
 390
 391         ipath_dbg("QP%d/%d in error state (%d)\n",
 392                   qp->ibqp.qp_num, qp->remote_qpn, err);
 393
 394         spin_lock(&dev->pending_lock);
 395         /* XXX What if its already removed by the timeout code? */
 396         if (!list_empty(&qp->timerwait))
 397                 list_del_init(&qp->timerwait);
 398         if (!list_empty(&qp->piowait))
 399                 list_del_init(&qp->piowait);
 400         spin_unlock(&dev->pending_lock);
 401
 402         wc.vendor_err = 0;
 403         wc.byte_len = 0;
 404         wc.imm_data = 0;
 405         wc.qp = &qp->ibqp;
 406         wc.src_qp = 0;
 407         wc.wc_flags = 0;
 408         wc.pkey_index = 0;
 409         wc.slid = 0;
 410         wc.sl = 0;
 411         wc.dlid_path_bits = 0;
 412         wc.port_num = 0;
 413         if (qp->r_wrid_valid) {
 414                 qp->r_wrid_valid = 0;
 415                 wc.wr_id = qp->r_wr_id;
 416                 wc.opcode = IB_WC_RECV;
 417                 wc.status = err;
 418                 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 419         }
 420         wc.status = IB_WC_WR_FLUSH_ERR;
 421
 422         while (qp->s_last != qp->s_head) {
 423                 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 424
 425                 wc.wr_id = wqe->wr.wr_id;
 426                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 427                 if (++qp->s_last >= qp->s_size)
 428                         qp->s_last = 0;
 429                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
 430         }
 431         qp->s_cur = qp->s_tail = qp->s_head;
 432         qp->s_hdrwords = 0;
 433         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 434
 435         if (qp->r_rq.wq) {
 436                 struct ipath_rwq *wq;
 437                 u32 head;
 438                 u32 tail;
 439
 440                 spin_lock(&qp->r_rq.lock);
 441
 442                 /* sanity check pointers before trusting them */
 443                 wq = qp->r_rq.wq;
 444                 head = wq->head;
 445                 if (head >= qp->r_rq.size)
 446                         head = 0;
 447                 tail = wq->tail;
 448                 if (tail >= qp->r_rq.size)
 449                         tail = 0;
 450                 wc.opcode = IB_WC_RECV;
 451                 while (tail != head) {
 452                         wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
 453                         if (++tail >= qp->r_rq.size)
 454                                 tail = 0;
 455                         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 456                 }
 457                 wq->tail = tail;
 458
 459                 spin_unlock(&qp->r_rq.lock);
 460         } else if (qp->ibqp.event_handler)
 461                 ret = 1;
 462
 463         return ret;
 464 }
 465
 466 /**
 467  * ipath_modify_qp - modify the attributes of a queue pair
 468  * @ibqp: the queue pair who's attributes we're modifying
 469  * @attr: the new attributes
 470  * @attr_mask: the mask of attributes to modify
 471  * @udata: user data for ipathverbs.so
 472  *
 473  * Returns 0 on success, otherwise returns an errno.
 474  */
 475 int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 476                     int attr_mask, struct ib_udata *udata)
 477 {
 478         struct ipath_ibdev *dev = to_idev(ibqp->device);
 479         struct ipath_qp *qp = to_iqp(ibqp);
 480         enum ib_qp_state cur_state, new_state;
 481         unsigned long flags;
 482         int lastwqe = 0;
 483         int ret;
 484
 485         spin_lock_irqsave(&qp->s_lock, flags);
 486
 487         cur_state = attr_mask & IB_QP_CUR_STATE ?
 488                 attr->cur_qp_state : qp->state;
 489         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
 490
 491         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
 492                                 attr_mask))
 493                 goto inval;
 494
 495         if (attr_mask & IB_QP_AV) {
 496                 if (attr->ah_attr.dlid == 0 ||
 497                     attr->ah_attr.dlid >= IPATH_MULTICAST_LID_BASE)
 498                         goto inval;
 499
 500                 if ((attr->ah_attr.ah_flags & IB_AH_GRH) &&
 501                     (attr->ah_attr.grh.sgid_index > 1))
 502                         goto inval;
 503         }
 504
 505         if (attr_mask & IB_QP_PKEY_INDEX)
 506                 if (attr->pkey_index >= ipath_get_npkeys(dev->dd))
 507                         goto inval;
 508
 509         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 510                 if (attr->min_rnr_timer > 31)
 511                         goto inval;
 512
 513         if (attr_mask & IB_QP_PORT)
 514                 if (attr->port_num == 0 ||
 515                     attr->port_num > ibqp->device->phys_port_cnt)
 516                         goto inval;
 517
 518         /*
 519          * Note: the chips support a maximum MTU of 4096, but the driver
 520          * hasn't implemented this feature yet, so don't allow Path MTU
 521          * values greater than 2048.
 522          */
 523         if (attr_mask & IB_QP_PATH_MTU)
 524                 if (attr->path_mtu > IB_MTU_2048)
 525                         goto inval;
 526
 527         if (attr_mask & IB_QP_PATH_MIG_STATE)
 528                 if (attr->path_mig_state != IB_MIG_MIGRATED &&
 529                     attr->path_mig_state != IB_MIG_REARM)
 530                         goto inval;
 531
 532         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 533                 if (attr->max_dest_rd_atomic > IPATH_MAX_RDMA_ATOMIC)
 534                         goto inval;
 535
 536         switch (new_state) {
 537         case IB_QPS_RESET:
 538                 ipath_reset_qp(qp, ibqp->qp_type);
 539                 break;
 540
 541         case IB_QPS_ERR:
 542                 lastwqe = ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 543                 break;
 544
 545         default:
 546                 break;
 547
 548         }
 549
 550         if (attr_mask & IB_QP_PKEY_INDEX)
 551                 qp->s_pkey_index = attr->pkey_index;
 552
 553         if (attr_mask & IB_QP_DEST_QPN)
 554                 qp->remote_qpn = attr->dest_qp_num;
 555
 556         if (attr_mask & IB_QP_SQ_PSN) {
 557                 qp->s_psn = qp->s_next_psn = attr->sq_psn;
 558                 qp->s_last_psn = qp->s_next_psn - 1;
 559         }
 560
 561         if (attr_mask & IB_QP_RQ_PSN)
 562                 qp->r_psn = attr->rq_psn;
 563
 564         if (attr_mask & IB_QP_ACCESS_FLAGS)
 565                 qp->qp_access_flags = attr->qp_access_flags;
 566
 567         if (attr_mask & IB_QP_AV)
 568                 qp->remote_ah_attr = attr->ah_attr;
 569
 570         if (attr_mask & IB_QP_PATH_MTU)
 571                 qp->path_mtu = attr->path_mtu;
 572
 573         if (attr_mask & IB_QP_RETRY_CNT)
 574                 qp->s_retry = qp->s_retry_cnt = attr->retry_cnt;
 575
 576         if (attr_mask & IB_QP_RNR_RETRY) {
 577                 qp->s_rnr_retry = attr->rnr_retry;
 578                 if (qp->s_rnr_retry > 7)
 579                         qp->s_rnr_retry = 7;
 580                 qp->s_rnr_retry_cnt = qp->s_rnr_retry;
 581         }
 582
 583         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 584                 qp->r_min_rnr_timer = attr->min_rnr_timer;
 585
 586         if (attr_mask & IB_QP_TIMEOUT)
 587                 qp->timeout = attr->timeout;
 588
 589         if (attr_mask & IB_QP_QKEY)
 590                 qp->qkey = attr->qkey;
 591
 592         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 593                 qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
 594
 595         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
 596                 qp->s_max_rd_atomic = attr->max_rd_atomic;
 597
 598         qp->state = new_state;
 599         spin_unlock_irqrestore(&qp->s_lock, flags);
 600
 601         if (lastwqe) {
 602                 struct ib_event ev;
 603
 604                 ev.device = qp->ibqp.device;
 605                 ev.element.qp = &qp->ibqp;
 606                 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 607                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 608         }
 609         ret = 0;
 610         goto bail;
 611
 612 inval:
 613         spin_unlock_irqrestore(&qp->s_lock, flags);
 614         ret = -EINVAL;
 615
 616 bail:
 617         return ret;
 618 }
 619
 620 int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 621                    int attr_mask, struct ib_qp_init_attr *init_attr)
 622 {
 623         struct ipath_qp *qp = to_iqp(ibqp);
 624
 625         attr->qp_state = qp->state;
 626         attr->cur_qp_state = attr->qp_state;
 627         attr->path_mtu = qp->path_mtu;
 628         attr->path_mig_state = 0;
 629         attr->qkey = qp->qkey;
 630         attr->rq_psn = qp->r_psn;
 631         attr->sq_psn = qp->s_next_psn;
 632         attr->dest_qp_num = qp->remote_qpn;
 633         attr->qp_access_flags = qp->qp_access_flags;
 634         attr->cap.max_send_wr = qp->s_size - 1;
 635         attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
 636         attr->cap.max_send_sge = qp->s_max_sge;
 637         attr->cap.max_recv_sge = qp->r_rq.max_sge;
 638         attr->cap.max_inline_data = 0;
 639         attr->ah_attr = qp->remote_ah_attr;
 640         memset(&attr->alt_ah_attr, 0, sizeof(attr->alt_ah_attr));
 641         attr->pkey_index = qp->s_pkey_index;
 642         attr->alt_pkey_index = 0;
 643         attr->en_sqd_async_notify = 0;
 644         attr->sq_draining = 0;
 645         attr->max_rd_atomic = qp->s_max_rd_atomic;
 646         attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
 647         attr->min_rnr_timer = qp->r_min_rnr_timer;
 648         attr->port_num = 1;
 649         attr->timeout = qp->timeout;
 650         attr->retry_cnt = qp->s_retry_cnt;
 651         attr->rnr_retry = qp->s_rnr_retry_cnt;
 652         attr->alt_port_num = 0;
 653         attr->alt_timeout = 0;
 654
 655         init_attr->event_handler = qp->ibqp.event_handler;
 656         init_attr->qp_context = qp->ibqp.qp_context;
 657         init_attr->send_cq = qp->ibqp.send_cq;
 658         init_attr->recv_cq = qp->ibqp.recv_cq;
 659         init_attr->srq = qp->ibqp.srq;
 660         init_attr->cap = attr->cap;
 661         if (qp->s_flags & IPATH_S_SIGNAL_REQ_WR)
 662                 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 663         else
 664                 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
 665         init_attr->qp_type = qp->ibqp.qp_type;
 666         init_attr->port_num = 1;
 667         return 0;
 668 }
 669
 670 /**
 671  * ipath_compute_aeth - compute the AETH (syndrome + MSN)
 672  * @qp: the queue pair to compute the AETH for
 673  *
 674  * Returns the AETH.
 675  */
 676 __be32 ipath_compute_aeth(struct ipath_qp *qp)
 677 {
 678         u32 aeth = qp->r_msn & IPATH_MSN_MASK;
 679
 680         if (qp->ibqp.srq) {
 681                 /*
 682                  * Shared receive queues don't generate credits.
 683                  * Set the credit field to the invalid value.
 684                  */
 685                 aeth |= IPATH_AETH_CREDIT_INVAL << IPATH_AETH_CREDIT_SHIFT;
 686         } else {
 687                 u32 min, max, x;
 688                 u32 credits;
 689                 struct ipath_rwq *wq = qp->r_rq.wq;
 690                 u32 head;
 691                 u32 tail;
 692
 693                 /* sanity check pointers before trusting them */
 694                 head = wq->head;
 695                 if (head >= qp->r_rq.size)
 696                         head = 0;
 697                 tail = wq->tail;
 698                 if (tail >= qp->r_rq.size)
 699                         tail = 0;
 700                 /*
 701                  * Compute the number of credits available (RWQEs).
 702                  * XXX Not holding the r_rq.lock here so there is a small
 703                  * chance that the pair of reads are not atomic.
 704                  */
 705                 credits = head - tail;
 706                 if ((int)credits < 0)
 707                         credits += qp->r_rq.size;
 708                 /*
 709                  * Binary search the credit table to find the code to
 710                  * use.
 711                  */
 712                 min = 0;
 713                 max = 31;
 714                 for (;;) {
 715                         x = (min + max) / 2;
 716                         if (credit_table[x] == credits)
 717                                 break;
 718                         if (credit_table[x] > credits)
 719                                 max = x;
 720                         else if (min == x)
 721                                 break;
 722                         else
 723                                 min = x;
 724                 }
 725                 aeth |= x << IPATH_AETH_CREDIT_SHIFT;
 726         }
 727         return cpu_to_be32(aeth);
 728 }
 729
 730 /**
 731  * ipath_create_qp - create a queue pair for a device
 732  * @ibpd: the protection domain who's device we create the queue pair for
 733  * @init_attr: the attributes of the queue pair
 734  * @udata: unused by InfiniPath
 735  *
 736  * Returns the queue pair on success, otherwise returns an errno.
 737  *
 738  * Called by the ib_create_qp() core verbs function.
 739  */
 740 struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
 741                               struct ib_qp_init_attr *init_attr,
 742                               struct ib_udata *udata)
 743 {
 744         struct ipath_qp *qp;
 745         int err;
 746         struct ipath_swqe *swq = NULL;
 747         struct ipath_ibdev *dev;
 748         size_t sz;
 749         struct ib_qp *ret;
 750
 751         if (init_attr->cap.max_send_sge > ib_ipath_max_sges ||
 752             init_attr->cap.max_recv_sge > ib_ipath_max_sges ||
 753             init_attr->cap.max_send_wr > ib_ipath_max_qp_wrs ||
 754             init_attr->cap.max_recv_wr > ib_ipath_max_qp_wrs) {
 755                 ret = ERR_PTR(-ENOMEM);
 756                 goto bail;
 757         }
 758
 759         if (init_attr->cap.max_send_sge +
 760             init_attr->cap.max_recv_sge +
 761             init_attr->cap.max_send_wr +
 762             init_attr->cap.max_recv_wr == 0) {
 763                 ret = ERR_PTR(-EINVAL);
 764                 goto bail;
 765         }
 766
 767         switch (init_attr->qp_type) {
 768         case IB_QPT_UC:
 769         case IB_QPT_RC:
 770         case IB_QPT_UD:
 771         case IB_QPT_SMI:
 772         case IB_QPT_GSI:
 773                 sz = sizeof(struct ipath_sge) *
 774                         init_attr->cap.max_send_sge +
 775                         sizeof(struct ipath_swqe);
 776                 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
 777                 if (swq == NULL) {
 778                         ret = ERR_PTR(-ENOMEM);
 779                         goto bail;
 780                 }
 781                 sz = sizeof(*qp);
 782                 if (init_attr->srq) {
 783                         struct ipath_srq *srq = to_isrq(init_attr->srq);
 784
 785                         sz += sizeof(*qp->r_sg_list) *
 786                                 srq->rq.max_sge;
 787                 } else
 788                         sz += sizeof(*qp->r_sg_list) *
 789                                 init_attr->cap.max_recv_sge;
 790                 qp = kmalloc(sz, GFP_KERNEL);
 791                 if (!qp) {
 792                         ret = ERR_PTR(-ENOMEM);
 793                         goto bail_swq;
 794                 }
 795                 if (init_attr->srq) {
 796                         sz = 0;
 797                         qp->r_rq.size = 0;
 798                         qp->r_rq.max_sge = 0;
 799                         qp->r_rq.wq = NULL;
 800                         init_attr->cap.max_recv_wr = 0;
 801                         init_attr->cap.max_recv_sge = 0;
 802                 } else {
 803                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
 804                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
 805                         sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
 806                                 sizeof(struct ipath_rwqe);
 807                         qp->r_rq.wq = vmalloc_user(sizeof(struct ipath_rwq) +
 808                                               qp->r_rq.size * sz);
 809                         if (!qp->r_rq.wq) {
 810                                 ret = ERR_PTR(-ENOMEM);
 811                                 goto bail_qp;
 812                         }
 813                 }
 814
 815                 /*
 816                  * ib_create_qp() will initialize qp->ibqp
 817                  * except for qp->ibqp.qp_num.
 818                  */
 819                 spin_lock_init(&qp->s_lock);
 820                 spin_lock_init(&qp->r_rq.lock);
 821                 atomic_set(&qp->refcount, 0);
 822                 init_waitqueue_head(&qp->wait);
 823                 tasklet_init(&qp->s_task, ipath_do_send, (unsigned long)qp);
 824                 INIT_LIST_HEAD(&qp->piowait);
 825                 INIT_LIST_HEAD(&qp->timerwait);
 826                 qp->state = IB_QPS_RESET;
 827                 qp->s_wq = swq;
 828                 qp->s_size = init_attr->cap.max_send_wr + 1;
 829                 qp->s_max_sge = init_attr->cap.max_send_sge;
 830                 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
 831                         qp->s_flags = IPATH_S_SIGNAL_REQ_WR;
 832                 else
 833                         qp->s_flags = 0;
 834                 dev = to_idev(ibpd->device);
 835                 err = ipath_alloc_qpn(&dev->qp_table, qp,
 836                                       init_attr->qp_type);
 837                 if (err) {
 838                         ret = ERR_PTR(err);
 839                         vfree(qp->r_rq.wq);
 840                         goto bail_qp;
 841                 }
 842                 qp->ip = NULL;
 843                 ipath_reset_qp(qp, init_attr->qp_type);
 844                 break;
 845
 846         default:
 847                 /* Don't support raw QPs */
 848                 ret = ERR_PTR(-ENOSYS);
 849                 goto bail;
 850         }
 851
 852         init_attr->cap.max_inline_data = 0;
 853
 854         /*
 855          * Return the address of the RWQ as the offset to mmap.
 856          * See ipath_mmap() for details.
 857          */
 858         if (udata && udata->outlen >= sizeof(__u64)) {
 859                 if (!qp->r_rq.wq) {
 860                         __u64 offset = 0;
 861
 862                         err = ib_copy_to_udata(udata, &offset,
 863                                                sizeof(offset));
 864                         if (err) {
 865                                 ret = ERR_PTR(err);
 866                                 goto bail_ip;
 867                         }
 868                 } else {
 869                         u32 s = sizeof(struct ipath_rwq) +
 870                                 qp->r_rq.size * sz;
 871
 872                         qp->ip =
 873                             ipath_create_mmap_info(dev, s,
 874                                                    ibpd->uobject->context,
 875                                                    qp->r_rq.wq);
 876                         if (!qp->ip) {
 877                                 ret = ERR_PTR(-ENOMEM);
 878                                 goto bail_ip;
 879                         }
 880
 881                         err = ib_copy_to_udata(udata, &(qp->ip->offset),
 882                                                sizeof(qp->ip->offset));
 883                         if (err) {
 884                                 ret = ERR_PTR(err);
 885                                 goto bail_ip;
 886                         }
 887                 }
 888         }
 889
 890         spin_lock(&dev->n_qps_lock);
 891         if (dev->n_qps_allocated == ib_ipath_max_qps) {
 892                 spin_unlock(&dev->n_qps_lock);
 893                 ret = ERR_PTR(-ENOMEM);
 894                 goto bail_ip;
 895         }
 896
 897         dev->n_qps_allocated++;
 898         spin_unlock(&dev->n_qps_lock);
 899
 900         if (qp->ip) {
 901                 spin_lock_irq(&dev->pending_lock);
 902                 list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps);
 903                 spin_unlock_irq(&dev->pending_lock);
 904         }
 905
 906         ret = &qp->ibqp;
 907         goto bail;
 908
 909 bail_ip:
 910         if (qp->ip)
 911                 kref_put(&qp->ip->ref, ipath_release_mmap_info);
 912         else
 913                 vfree(qp->r_rq.wq);
 914         ipath_free_qp(&dev->qp_table, qp);
 915 bail_qp:
 916         kfree(qp);
 917 bail_swq:
 918         vfree(swq);
 919 bail:
 920         return ret;
 921 }
 922
 923 /**
 924  * ipath_destroy_qp - destroy a queue pair
 925  * @ibqp: the queue pair to destroy
 926  *
 927  * Returns 0 on success.
 928  *
 929  * Note that this can be called while the QP is actively sending or
 930  * receiving!
 931  */
 932 int ipath_destroy_qp(struct ib_qp *ibqp)
 933 {
 934         struct ipath_qp *qp = to_iqp(ibqp);
 935         struct ipath_ibdev *dev = to_idev(ibqp->device);
 936         unsigned long flags;
 937
 938         spin_lock_irqsave(&qp->s_lock, flags);
 939         qp->state = IB_QPS_ERR;
 940         spin_unlock_irqrestore(&qp->s_lock, flags);
 941         spin_lock(&dev->n_qps_lock);
 942         dev->n_qps_allocated--;
 943         spin_unlock(&dev->n_qps_lock);
 944
 945         /* Stop the sending tasklet. */
 946         tasklet_kill(&qp->s_task);
 947
 948         /* Make sure the QP isn't on the timeout list. */
 949         spin_lock_irqsave(&dev->pending_lock, flags);
 950         if (!list_empty(&qp->timerwait))
 951                 list_del_init(&qp->timerwait);
 952         if (!list_empty(&qp->piowait))
 953                 list_del_init(&qp->piowait);
 954         spin_unlock_irqrestore(&dev->pending_lock, flags);
 955
 956         /*
 957          * Make sure that the QP is not in the QPN table so receive
 958          * interrupts will discard packets for this QP.  XXX Also remove QP
 959          * from multicast table.
 960          */
 961         if (atomic_read(&qp->refcount) != 0)
 962                 ipath_free_qp(&dev->qp_table, qp);
 963
 964         if (qp->ip)
 965                 kref_put(&qp->ip->ref, ipath_release_mmap_info);
 966         else
 967                 vfree(qp->r_rq.wq);
 968         vfree(qp->s_wq);
 969         kfree(qp);
 970         return 0;
 971 }
 972
 973 /**
 974  * ipath_init_qp_table - initialize the QP table for a device
 975  * @idev: the device who's QP table we're initializing
 976  * @size: the size of the QP table
 977  *
 978  * Returns 0 on success, otherwise returns an errno.
 979  */
 980 int ipath_init_qp_table(struct ipath_ibdev *idev, int size)
 981 {
 982         int i;
 983         int ret;
 984
 985         idev->qp_table.last = 1;        /* QPN 0 and 1 are special. */
 986         idev->qp_table.max = size;
 987         idev->qp_table.nmaps = 1;
 988         idev->qp_table.table = kzalloc(size * sizeof(*idev->qp_table.table),
 989                                        GFP_KERNEL);
 990         if (idev->qp_table.table == NULL) {
 991                 ret = -ENOMEM;
 992                 goto bail;
 993         }
 994
 995         for (i = 0; i < ARRAY_SIZE(idev->qp_table.map); i++) {
 996                 atomic_set(&idev->qp_table.map[i].n_free, BITS_PER_PAGE);
 997                 idev->qp_table.map[i].page = NULL;
 998         }
 999
1000         ret = 0;
1001
1002 bail:
1003         return ret;
1004 }
1005
1006 /**
1007  * ipath_sqerror_qp - put a QP's send queue into an error state
1008  * @qp: QP who's send queue will be put into an error state
1009  * @wc: the WC responsible for putting the QP in this state
1010  *
1011  * Flushes the send work queue.
1012  * The QP s_lock should be held and interrupts disabled.
1013  */
1014
1015 void ipath_sqerror_qp(struct ipath_qp *qp, struct ib_wc *wc)
1016 {
1017         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
1018         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
1019
1020         ipath_dbg("Send queue error on QP%d/%d: err: %d\n",
1021                   qp->ibqp.qp_num, qp->remote_qpn, wc->status);
1022
1023         spin_lock(&dev->pending_lock);
1024         /* XXX What if its already removed by the timeout code? */
1025         if (!list_empty(&qp->timerwait))
1026                 list_del_init(&qp->timerwait);
1027         if (!list_empty(&qp->piowait))
1028                 list_del_init(&qp->piowait);
1029         spin_unlock(&dev->pending_lock);
1030
1031         ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
1032         if (++qp->s_last >= qp->s_size)
1033                 qp->s_last = 0;
1034
1035         wc->status = IB_WC_WR_FLUSH_ERR;
1036
1037         while (qp->s_last != qp->s_head) {
1038                 wqe = get_swqe_ptr(qp, qp->s_last);
1039                 wc->wr_id = wqe->wr.wr_id;
1040                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1041                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
1042                 if (++qp->s_last >= qp->s_size)
1043                         qp->s_last = 0;
1044         }
1045         qp->s_cur = qp->s_tail = qp->s_head;
1046         qp->state = IB_QPS_SQE;
1047 }
1048
1049 /**
1050  * ipath_get_credit - flush the send work queue of a QP
1051  * @qp: the qp who's send work queue to flush
1052  * @aeth: the Acknowledge Extended Transport Header
1053  *
1054  * The QP s_lock should be held.
1055  */
1056 void ipath_get_credit(struct ipath_qp *qp, u32 aeth)
1057 {
1058         u32 credit = (aeth >> IPATH_AETH_CREDIT_SHIFT) & IPATH_AETH_CREDIT_MASK;
1059
1060         /*
1061          * If the credit is invalid, we can send
1062          * as many packets as we like.  Otherwise, we have to
1063          * honor the credit field.
1064          */
1065         if (credit == IPATH_AETH_CREDIT_INVAL)
1066                 qp->s_lsn = (u32) -1;
1067         else if (qp->s_lsn != (u32) -1) {
1068                 /* Compute new LSN (i.e., MSN + credit) */
1069                 credit = (aeth + credit_table[credit]) & IPATH_MSN_MASK;
1070                 if (ipath_cmp24(credit, qp->s_lsn) > 0)
1071                         qp->s_lsn = credit;
1072         }
1073
1074         /* Restart sending if it was blocked due to lack of credits. */
1075         if (qp->s_cur != qp->s_head &&
1076             (qp->s_lsn == (u32) -1 ||
1077              ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn,
1078                          qp->s_lsn + 1) <= 0))
1079                 tasklet_hi_schedule(&qp->s_task);
1080 }