drivers/infiniband/hw/ipath/ipath_qp.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. All rights reserved.
   3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/err.h>
  35 #include <linux/vmalloc.h>
  36
  37 #include "ipath_verbs.h"
  38 #include "ips_common.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 #define TRANS_INVALID   0
  48 #define TRANS_ANY2RST   1
  49 #define TRANS_RST2INIT  2
  50 #define TRANS_INIT2INIT 3
  51 #define TRANS_INIT2RTR  4
  52 #define TRANS_RTR2RTS   5
  53 #define TRANS_RTS2RTS   6
  54 #define TRANS_SQERR2RTS 7
  55 #define TRANS_ANY2ERR   8
  56 #define TRANS_RTS2SQD   9  /* XXX Wait for expected ACKs & signal event */
  57 #define TRANS_SQD2SQD   10 /* error if not drained & parameter change */
  58 #define TRANS_SQD2RTS   11 /* error if not drained */
  59
  60 /*
  61  * Convert the AETH credit code into the number of credits.
  62  */
  63 static u32 credit_table[31] = {
  64         0,                      /* 0 */
  65         1,                      /* 1 */
  66         2,                      /* 2 */
  67         3,                      /* 3 */
  68         4,                      /* 4 */
  69         6,                      /* 5 */
  70         8,                      /* 6 */
  71         12,                     /* 7 */
  72         16,                     /* 8 */
  73         24,                     /* 9 */
  74         32,                     /* A */
  75         48,                     /* B */
  76         64,                     /* C */
  77         96,                     /* D */
  78         128,                    /* E */
  79         192,                    /* F */
  80         256,                    /* 10 */
  81         384,                    /* 11 */
  82         512,                    /* 12 */
  83         768,                    /* 13 */
  84         1024,                   /* 14 */
  85         1536,                   /* 15 */
  86         2048,                   /* 16 */
  87         3072,                   /* 17 */
  88         4096,                   /* 18 */
  89         6144,                   /* 19 */
  90         8192,                   /* 1A */
  91         12288,                  /* 1B */
  92         16384,                  /* 1C */
  93         24576,                  /* 1D */
  94         32768                   /* 1E */
  95 };
  96
  97 static u32 alloc_qpn(struct ipath_qp_table *qpt)
  98 {
  99         u32 i, offset, max_scan, qpn;
 100         struct qpn_map *map;
 101         u32 ret;
 102
 103         qpn = qpt->last + 1;
 104         if (qpn >= QPN_MAX)
 105                 qpn = 2;
 106         offset = qpn & BITS_PER_PAGE_MASK;
 107         map = &qpt->map[qpn / BITS_PER_PAGE];
 108         max_scan = qpt->nmaps - !offset;
 109         for (i = 0;;) {
 110                 if (unlikely(!map->page)) {
 111                         unsigned long page = get_zeroed_page(GFP_KERNEL);
 112                         unsigned long flags;
 113
 114                         /*
 115                          * Free the page if someone raced with us
 116                          * installing it:
 117                          */
 118                         spin_lock_irqsave(&qpt->lock, flags);
 119                         if (map->page)
 120                                 free_page(page);
 121                         else
 122                                 map->page = (void *)page;
 123                         spin_unlock_irqrestore(&qpt->lock, flags);
 124                         if (unlikely(!map->page))
 125                                 break;
 126                 }
 127                 if (likely(atomic_read(&map->n_free))) {
 128                         do {
 129                                 if (!test_and_set_bit(offset, map->page)) {
 130                                         atomic_dec(&map->n_free);
 131                                         qpt->last = qpn;
 132                                         ret = qpn;
 133                                         goto bail;
 134                                 }
 135                                 offset = find_next_offset(map, offset);
 136                                 qpn = mk_qpn(qpt, map, offset);
 137                                 /*
 138                                  * This test differs from alloc_pidmap().
 139                                  * If find_next_offset() does find a zero
 140                                  * bit, we don't need to check for QPN
 141                                  * wrapping around past our starting QPN.
 142                                  * We just need to be sure we don't loop
 143                                  * forever.
 144                                  */
 145                         } while (offset < BITS_PER_PAGE && qpn < QPN_MAX);
 146                 }
 147                 /*
 148                  * In order to keep the number of pages allocated to a
 149                  * minimum, we scan the all existing pages before increasing
 150                  * the size of the bitmap table.
 151                  */
 152                 if (++i > max_scan) {
 153                         if (qpt->nmaps == QPNMAP_ENTRIES)
 154                                 break;
 155                         map = &qpt->map[qpt->nmaps++];
 156                         offset = 0;
 157                 } else if (map < &qpt->map[qpt->nmaps]) {
 158                         ++map;
 159                         offset = 0;
 160                 } else {
 161                         map = &qpt->map[0];
 162                         offset = 2;
 163                 }
 164                 qpn = mk_qpn(qpt, map, offset);
 165         }
 166
 167         ret = 0;
 168
 169 bail:
 170         return ret;
 171 }
 172
 173 static void free_qpn(struct ipath_qp_table *qpt, u32 qpn)
 174 {
 175         struct qpn_map *map;
 176
 177         map = qpt->map + qpn / BITS_PER_PAGE;
 178         if (map->page)
 179                 clear_bit(qpn & BITS_PER_PAGE_MASK, map->page);
 180         atomic_inc(&map->n_free);
 181 }
 182
 183 /**
 184  * ipath_alloc_qpn - allocate a QP number
 185  * @qpt: the QP table
 186  * @qp: the QP
 187  * @type: the QP type (IB_QPT_SMI and IB_QPT_GSI are special)
 188  *
 189  * Allocate the next available QPN and put the QP into the hash table.
 190  * The hash table holds a reference to the QP.
 191  */
 192 static int ipath_alloc_qpn(struct ipath_qp_table *qpt, struct ipath_qp *qp,
 193                            enum ib_qp_type type)
 194 {
 195         unsigned long flags;
 196         u32 qpn;
 197         int ret;
 198
 199         if (type == IB_QPT_SMI)
 200                 qpn = 0;
 201         else if (type == IB_QPT_GSI)
 202                 qpn = 1;
 203         else {
 204                 /* Allocate the next available QPN */
 205                 qpn = alloc_qpn(qpt);
 206                 if (qpn == 0) {
 207                         ret = -ENOMEM;
 208                         goto bail;
 209                 }
 210         }
 211         qp->ibqp.qp_num = qpn;
 212
 213         /* Add the QP to the hash table. */
 214         spin_lock_irqsave(&qpt->lock, flags);
 215
 216         qpn %= qpt->max;
 217         qp->next = qpt->table[qpn];
 218         qpt->table[qpn] = qp;
 219         atomic_inc(&qp->refcount);
 220
 221         spin_unlock_irqrestore(&qpt->lock, flags);
 222         ret = 0;
 223
 224 bail:
 225         return ret;
 226 }
 227
 228 /**
 229  * ipath_free_qp - remove a QP from the QP table
 230  * @qpt: the QP table
 231  * @qp: the QP to remove
 232  *
 233  * Remove the QP from the table so it can't be found asynchronously by
 234  * the receive interrupt routine.
 235  */
 236 static void ipath_free_qp(struct ipath_qp_table *qpt, struct ipath_qp *qp)
 237 {
 238         struct ipath_qp *q, **qpp;
 239         unsigned long flags;
 240         int fnd = 0;
 241
 242         spin_lock_irqsave(&qpt->lock, flags);
 243
 244         /* Remove QP from the hash table. */
 245         qpp = &qpt->table[qp->ibqp.qp_num % qpt->max];
 246         for (; (q = *qpp) != NULL; qpp = &q->next) {
 247                 if (q == qp) {
 248                         *qpp = qp->next;
 249                         qp->next = NULL;
 250                         atomic_dec(&qp->refcount);
 251                         fnd = 1;
 252                         break;
 253                 }
 254         }
 255
 256         spin_unlock_irqrestore(&qpt->lock, flags);
 257
 258         if (!fnd)
 259                 return;
 260
 261         /* If QPN is not reserved, mark QPN free in the bitmap. */
 262         if (qp->ibqp.qp_num > 1)
 263                 free_qpn(qpt, qp->ibqp.qp_num);
 264
 265         wait_event(qp->wait, !atomic_read(&qp->refcount));
 266 }
 267
 268 /**
 269  * ipath_free_all_qps - remove all QPs from the table
 270  * @qpt: the QP table to empty
 271  */
 272 void ipath_free_all_qps(struct ipath_qp_table *qpt)
 273 {
 274         unsigned long flags;
 275         struct ipath_qp *qp, *nqp;
 276         u32 n;
 277
 278         for (n = 0; n < qpt->max; n++) {
 279                 spin_lock_irqsave(&qpt->lock, flags);
 280                 qp = qpt->table[n];
 281                 qpt->table[n] = NULL;
 282                 spin_unlock_irqrestore(&qpt->lock, flags);
 283
 284                 while (qp) {
 285                         nqp = qp->next;
 286                         if (qp->ibqp.qp_num > 1)
 287                                 free_qpn(qpt, qp->ibqp.qp_num);
 288                         if (!atomic_dec_and_test(&qp->refcount) ||
 289                             !ipath_destroy_qp(&qp->ibqp))
 290                                 _VERBS_INFO("QP memory leak!\n");
 291                         qp = nqp;
 292                 }
 293         }
 294
 295         for (n = 0; n < ARRAY_SIZE(qpt->map); n++) {
 296                 if (qpt->map[n].page)
 297                         free_page((unsigned long)qpt->map[n].page);
 298         }
 299 }
 300
 301 /**
 302  * ipath_lookup_qpn - return the QP with the given QPN
 303  * @qpt: the QP table
 304  * @qpn: the QP number to look up
 305  *
 306  * The caller is responsible for decrementing the QP reference count
 307  * when done.
 308  */
 309 struct ipath_qp *ipath_lookup_qpn(struct ipath_qp_table *qpt, u32 qpn)
 310 {
 311         unsigned long flags;
 312         struct ipath_qp *qp;
 313
 314         spin_lock_irqsave(&qpt->lock, flags);
 315
 316         for (qp = qpt->table[qpn % qpt->max]; qp; qp = qp->next) {
 317                 if (qp->ibqp.qp_num == qpn) {
 318                         atomic_inc(&qp->refcount);
 319                         break;
 320                 }
 321         }
 322
 323         spin_unlock_irqrestore(&qpt->lock, flags);
 324         return qp;
 325 }
 326
 327 /**
 328  * ipath_reset_qp - initialize the QP state to the reset state
 329  * @qp: the QP to reset
 330  */
 331 static void ipath_reset_qp(struct ipath_qp *qp)
 332 {
 333         qp->remote_qpn = 0;
 334         qp->qkey = 0;
 335         qp->qp_access_flags = 0;
 336         clear_bit(IPATH_S_BUSY, &qp->s_flags);
 337         qp->s_hdrwords = 0;
 338         qp->s_psn = 0;
 339         qp->r_psn = 0;
 340         qp->r_msn = 0;
 341         if (qp->ibqp.qp_type == IB_QPT_RC) {
 342                 qp->s_state = IB_OPCODE_RC_SEND_LAST;
 343                 qp->r_state = IB_OPCODE_RC_SEND_LAST;
 344         } else {
 345                 qp->s_state = IB_OPCODE_UC_SEND_LAST;
 346                 qp->r_state = IB_OPCODE_UC_SEND_LAST;
 347         }
 348         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 349         qp->r_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 350         qp->r_nak_state = 0;
 351         qp->s_rnr_timeout = 0;
 352         qp->s_head = 0;
 353         qp->s_tail = 0;
 354         qp->s_cur = 0;
 355         qp->s_last = 0;
 356         qp->s_ssn = 1;
 357         qp->s_lsn = 0;
 358         qp->r_rq.head = 0;
 359         qp->r_rq.tail = 0;
 360         qp->r_reuse_sge = 0;
 361 }
 362
 363 /**
 364  * ipath_error_qp - put a QP into an error state
 365  * @qp: the QP to put into an error state
 366  *
 367  * Flushes both send and receive work queues.
 368  * QP s_lock should be held and interrupts disabled.
 369  */
 370
 371 void ipath_error_qp(struct ipath_qp *qp)
 372 {
 373         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 374         struct ib_wc wc;
 375
 376         _VERBS_INFO("QP%d/%d in error state\n",
 377                     qp->ibqp.qp_num, qp->remote_qpn);
 378
 379         spin_lock(&dev->pending_lock);
 380         /* XXX What if its already removed by the timeout code? */
 381         if (!list_empty(&qp->timerwait))
 382                 list_del_init(&qp->timerwait);
 383         if (!list_empty(&qp->piowait))
 384                 list_del_init(&qp->piowait);
 385         spin_unlock(&dev->pending_lock);
 386
 387         wc.status = IB_WC_WR_FLUSH_ERR;
 388         wc.vendor_err = 0;
 389         wc.byte_len = 0;
 390         wc.imm_data = 0;
 391         wc.qp_num = qp->ibqp.qp_num;
 392         wc.src_qp = 0;
 393         wc.wc_flags = 0;
 394         wc.pkey_index = 0;
 395         wc.slid = 0;
 396         wc.sl = 0;
 397         wc.dlid_path_bits = 0;
 398         wc.port_num = 0;
 399
 400         while (qp->s_last != qp->s_head) {
 401                 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 402
 403                 wc.wr_id = wqe->wr.wr_id;
 404                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 405                 if (++qp->s_last >= qp->s_size)
 406                         qp->s_last = 0;
 407                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
 408         }
 409         qp->s_cur = qp->s_tail = qp->s_head;
 410         qp->s_hdrwords = 0;
 411         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 412
 413         wc.opcode = IB_WC_RECV;
 414         spin_lock(&qp->r_rq.lock);
 415         while (qp->r_rq.tail != qp->r_rq.head) {
 416                 wc.wr_id = get_rwqe_ptr(&qp->r_rq, qp->r_rq.tail)->wr_id;
 417                 if (++qp->r_rq.tail >= qp->r_rq.size)
 418                         qp->r_rq.tail = 0;
 419                 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 420         }
 421         spin_unlock(&qp->r_rq.lock);
 422 }
 423
 424 /**
 425  * ipath_modify_qp - modify the attributes of a queue pair
 426  * @ibqp: the queue pair who's attributes we're modifying
 427  * @attr: the new attributes
 428  * @attr_mask: the mask of attributes to modify
 429  *
 430  * Returns 0 on success, otherwise returns an errno.
 431  */
 432 int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 433                     int attr_mask)
 434 {
 435         struct ipath_ibdev *dev = to_idev(ibqp->device);
 436         struct ipath_qp *qp = to_iqp(ibqp);
 437         enum ib_qp_state cur_state, new_state;
 438         unsigned long flags;
 439         int ret;
 440
 441         spin_lock_irqsave(&qp->s_lock, flags);
 442
 443         cur_state = attr_mask & IB_QP_CUR_STATE ?
 444                 attr->cur_qp_state : qp->state;
 445         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
 446
 447         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
 448                                 attr_mask))
 449                 goto inval;
 450
 451         if (attr_mask & IB_QP_AV)
 452                 if (attr->ah_attr.dlid == 0 ||
 453                     attr->ah_attr.dlid >= IPS_MULTICAST_LID_BASE)
 454                         goto inval;
 455
 456         if (attr_mask & IB_QP_PKEY_INDEX)
 457                 if (attr->pkey_index >= ipath_layer_get_npkeys(dev->dd))
 458                         goto inval;
 459
 460         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 461                 if (attr->min_rnr_timer > 31)
 462                         goto inval;
 463
 464         switch (new_state) {
 465         case IB_QPS_RESET:
 466                 ipath_reset_qp(qp);
 467                 break;
 468
 469         case IB_QPS_ERR:
 470                 ipath_error_qp(qp);
 471                 break;
 472
 473         default:
 474                 break;
 475
 476         }
 477
 478         if (attr_mask & IB_QP_PKEY_INDEX)
 479                 qp->s_pkey_index = attr->pkey_index;
 480
 481         if (attr_mask & IB_QP_DEST_QPN)
 482                 qp->remote_qpn = attr->dest_qp_num;
 483
 484         if (attr_mask & IB_QP_SQ_PSN) {
 485                 qp->s_next_psn = attr->sq_psn;
 486                 qp->s_last_psn = qp->s_next_psn - 1;
 487         }
 488
 489         if (attr_mask & IB_QP_RQ_PSN)
 490                 qp->r_psn = attr->rq_psn;
 491
 492         if (attr_mask & IB_QP_ACCESS_FLAGS)
 493                 qp->qp_access_flags = attr->qp_access_flags;
 494
 495         if (attr_mask & IB_QP_AV)
 496                 qp->remote_ah_attr = attr->ah_attr;
 497
 498         if (attr_mask & IB_QP_PATH_MTU)
 499                 qp->path_mtu = attr->path_mtu;
 500
 501         if (attr_mask & IB_QP_RETRY_CNT)
 502                 qp->s_retry = qp->s_retry_cnt = attr->retry_cnt;
 503
 504         if (attr_mask & IB_QP_RNR_RETRY) {
 505                 qp->s_rnr_retry = attr->rnr_retry;
 506                 if (qp->s_rnr_retry > 7)
 507                         qp->s_rnr_retry = 7;
 508                 qp->s_rnr_retry_cnt = qp->s_rnr_retry;
 509         }
 510
 511         if (attr_mask & IB_QP_MIN_RNR_TIMER)
 512                 qp->r_min_rnr_timer = attr->min_rnr_timer;
 513
 514         if (attr_mask & IB_QP_QKEY)
 515                 qp->qkey = attr->qkey;
 516
 517         qp->state = new_state;
 518         spin_unlock_irqrestore(&qp->s_lock, flags);
 519
 520         ret = 0;
 521         goto bail;
 522
 523 inval:
 524         spin_unlock_irqrestore(&qp->s_lock, flags);
 525         ret = -EINVAL;
 526
 527 bail:
 528         return ret;
 529 }
 530
 531 int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 532                    int attr_mask, struct ib_qp_init_attr *init_attr)
 533 {
 534         struct ipath_qp *qp = to_iqp(ibqp);
 535
 536         attr->qp_state = qp->state;
 537         attr->cur_qp_state = attr->qp_state;
 538         attr->path_mtu = qp->path_mtu;
 539         attr->path_mig_state = 0;
 540         attr->qkey = qp->qkey;
 541         attr->rq_psn = qp->r_psn;
 542         attr->sq_psn = qp->s_next_psn;
 543         attr->dest_qp_num = qp->remote_qpn;
 544         attr->qp_access_flags = qp->qp_access_flags;
 545         attr->cap.max_send_wr = qp->s_size - 1;
 546         attr->cap.max_recv_wr = qp->r_rq.size - 1;
 547         attr->cap.max_send_sge = qp->s_max_sge;
 548         attr->cap.max_recv_sge = qp->r_rq.max_sge;
 549         attr->cap.max_inline_data = 0;
 550         attr->ah_attr = qp->remote_ah_attr;
 551         memset(&attr->alt_ah_attr, 0, sizeof(attr->alt_ah_attr));
 552         attr->pkey_index = qp->s_pkey_index;
 553         attr->alt_pkey_index = 0;
 554         attr->en_sqd_async_notify = 0;
 555         attr->sq_draining = 0;
 556         attr->max_rd_atomic = 1;
 557         attr->max_dest_rd_atomic = 1;
 558         attr->min_rnr_timer = qp->r_min_rnr_timer;
 559         attr->port_num = 1;
 560         attr->timeout = 0;
 561         attr->retry_cnt = qp->s_retry_cnt;
 562         attr->rnr_retry = qp->s_rnr_retry;
 563         attr->alt_port_num = 0;
 564         attr->alt_timeout = 0;
 565
 566         init_attr->event_handler = qp->ibqp.event_handler;
 567         init_attr->qp_context = qp->ibqp.qp_context;
 568         init_attr->send_cq = qp->ibqp.send_cq;
 569         init_attr->recv_cq = qp->ibqp.recv_cq;
 570         init_attr->srq = qp->ibqp.srq;
 571         init_attr->cap = attr->cap;
 572         init_attr->sq_sig_type =
 573                 (qp->s_flags & (1 << IPATH_S_SIGNAL_REQ_WR))
 574                 ? IB_SIGNAL_REQ_WR : 0;
 575         init_attr->qp_type = qp->ibqp.qp_type;
 576         init_attr->port_num = 1;
 577         return 0;
 578 }
 579
 580 /**
 581  * ipath_compute_aeth - compute the AETH (syndrome + MSN)
 582  * @qp: the queue pair to compute the AETH for
 583  *
 584  * Returns the AETH.
 585  */
 586 __be32 ipath_compute_aeth(struct ipath_qp *qp)
 587 {
 588         u32 aeth = qp->r_msn & IPS_MSN_MASK;
 589
 590         if (qp->ibqp.srq) {
 591                 /*
 592                  * Shared receive queues don't generate credits.
 593                  * Set the credit field to the invalid value.
 594                  */
 595                 aeth |= IPS_AETH_CREDIT_INVAL << IPS_AETH_CREDIT_SHIFT;
 596         } else {
 597                 u32 min, max, x;
 598                 u32 credits;
 599
 600                 /*
 601                  * Compute the number of credits available (RWQEs).
 602                  * XXX Not holding the r_rq.lock here so there is a small
 603                  * chance that the pair of reads are not atomic.
 604                  */
 605                 credits = qp->r_rq.head - qp->r_rq.tail;
 606                 if ((int)credits < 0)
 607                         credits += qp->r_rq.size;
 608                 /*
 609                  * Binary search the credit table to find the code to
 610                  * use.
 611                  */
 612                 min = 0;
 613                 max = 31;
 614                 for (;;) {
 615                         x = (min + max) / 2;
 616                         if (credit_table[x] == credits)
 617                                 break;
 618                         if (credit_table[x] > credits)
 619                                 max = x;
 620                         else if (min == x)
 621                                 break;
 622                         else
 623                                 min = x;
 624                 }
 625                 aeth |= x << IPS_AETH_CREDIT_SHIFT;
 626         }
 627         return cpu_to_be32(aeth);
 628 }
 629
 630 /**
 631  * ipath_create_qp - create a queue pair for a device
 632  * @ibpd: the protection domain who's device we create the queue pair for
 633  * @init_attr: the attributes of the queue pair
 634  * @udata: unused by InfiniPath
 635  *
 636  * Returns the queue pair on success, otherwise returns an errno.
 637  *
 638  * Called by the ib_create_qp() core verbs function.
 639  */
 640 struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
 641                               struct ib_qp_init_attr *init_attr,
 642                               struct ib_udata *udata)
 643 {
 644         struct ipath_qp *qp;
 645         int err;
 646         struct ipath_swqe *swq = NULL;
 647         struct ipath_ibdev *dev;
 648         size_t sz;
 649         struct ib_qp *ret;
 650
 651         if (init_attr->cap.max_send_sge > ib_ipath_max_sges ||
 652             init_attr->cap.max_recv_sge > ib_ipath_max_sges ||
 653             init_attr->cap.max_send_wr > ib_ipath_max_qp_wrs ||
 654             init_attr->cap.max_recv_wr > ib_ipath_max_qp_wrs) {
 655                 ret = ERR_PTR(-ENOMEM);
 656                 goto bail;
 657         }
 658
 659         if (init_attr->cap.max_send_sge +
 660             init_attr->cap.max_recv_sge +
 661             init_attr->cap.max_send_wr +
 662             init_attr->cap.max_recv_wr == 0) {
 663                 ret = ERR_PTR(-EINVAL);
 664                 goto bail;
 665         }
 666
 667         switch (init_attr->qp_type) {
 668         case IB_QPT_UC:
 669         case IB_QPT_RC:
 670                 sz = sizeof(struct ipath_sge) *
 671                         init_attr->cap.max_send_sge +
 672                         sizeof(struct ipath_swqe);
 673                 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
 674                 if (swq == NULL) {
 675                         ret = ERR_PTR(-ENOMEM);
 676                         goto bail;
 677                 }
 678                 /* FALLTHROUGH */
 679         case IB_QPT_UD:
 680         case IB_QPT_SMI:
 681         case IB_QPT_GSI:
 682                 qp = kmalloc(sizeof(*qp), GFP_KERNEL);
 683                 if (!qp) {
 684                         vfree(swq);
 685                         ret = ERR_PTR(-ENOMEM);
 686                         goto bail;
 687                 }
 688                 if (init_attr->srq) {
 689                         qp->r_rq.size = 0;
 690                         qp->r_rq.max_sge = 0;
 691                         qp->r_rq.wq = NULL;
 692                 } else {
 693                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
 694                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
 695                         sz = (sizeof(struct ipath_sge) * qp->r_rq.max_sge) +
 696                                 sizeof(struct ipath_rwqe);
 697                         qp->r_rq.wq = vmalloc(qp->r_rq.size * sz);
 698                         if (!qp->r_rq.wq) {
 699                                 kfree(qp);
 700                                 vfree(swq);
 701                                 ret = ERR_PTR(-ENOMEM);
 702                                 goto bail;
 703                         }
 704                 }
 705
 706                 /*
 707                  * ib_create_qp() will initialize qp->ibqp
 708                  * except for qp->ibqp.qp_num.
 709                  */
 710                 spin_lock_init(&qp->s_lock);
 711                 spin_lock_init(&qp->r_rq.lock);
 712                 atomic_set(&qp->refcount, 0);
 713                 init_waitqueue_head(&qp->wait);
 714                 tasklet_init(&qp->s_task, ipath_do_ruc_send,
 715                              (unsigned long)qp);
 716                 INIT_LIST_HEAD(&qp->piowait);
 717                 INIT_LIST_HEAD(&qp->timerwait);
 718                 qp->state = IB_QPS_RESET;
 719                 qp->s_wq = swq;
 720                 qp->s_size = init_attr->cap.max_send_wr + 1;
 721                 qp->s_max_sge = init_attr->cap.max_send_sge;
 722                 qp->s_flags = init_attr->sq_sig_type == IB_SIGNAL_REQ_WR ?
 723                         1 << IPATH_S_SIGNAL_REQ_WR : 0;
 724                 dev = to_idev(ibpd->device);
 725                 err = ipath_alloc_qpn(&dev->qp_table, qp,
 726                                       init_attr->qp_type);
 727                 if (err) {
 728                         vfree(swq);
 729                         vfree(qp->r_rq.wq);
 730                         kfree(qp);
 731                         ret = ERR_PTR(err);
 732                         goto bail;
 733                 }
 734                 ipath_reset_qp(qp);
 735
 736                 /* Tell the core driver that the kernel SMA is present. */
 737                 if (init_attr->qp_type == IB_QPT_SMI)
 738                         ipath_layer_set_verbs_flags(dev->dd,
 739                                                     IPATH_VERBS_KERNEL_SMA);
 740                 break;
 741
 742         default:
 743                 /* Don't support raw QPs */
 744                 ret = ERR_PTR(-ENOSYS);
 745                 goto bail;
 746         }
 747
 748         init_attr->cap.max_inline_data = 0;
 749
 750         ret = &qp->ibqp;
 751
 752 bail:
 753         return ret;
 754 }
 755
 756 /**
 757  * ipath_destroy_qp - destroy a queue pair
 758  * @ibqp: the queue pair to destroy
 759  *
 760  * Returns 0 on success.
 761  *
 762  * Note that this can be called while the QP is actively sending or
 763  * receiving!
 764  */
 765 int ipath_destroy_qp(struct ib_qp *ibqp)
 766 {
 767         struct ipath_qp *qp = to_iqp(ibqp);
 768         struct ipath_ibdev *dev = to_idev(ibqp->device);
 769         unsigned long flags;
 770
 771         /* Tell the core driver that the kernel SMA is gone. */
 772         if (qp->ibqp.qp_type == IB_QPT_SMI)
 773                 ipath_layer_set_verbs_flags(dev->dd, 0);
 774
 775         spin_lock_irqsave(&qp->r_rq.lock, flags);
 776         spin_lock(&qp->s_lock);
 777         qp->state = IB_QPS_ERR;
 778         spin_unlock(&qp->s_lock);
 779         spin_unlock_irqrestore(&qp->r_rq.lock, flags);
 780
 781         /* Stop the sending tasklet. */
 782         tasklet_kill(&qp->s_task);
 783
 784         /* Make sure the QP isn't on the timeout list. */
 785         spin_lock_irqsave(&dev->pending_lock, flags);
 786         if (!list_empty(&qp->timerwait))
 787                 list_del_init(&qp->timerwait);
 788         if (!list_empty(&qp->piowait))
 789                 list_del_init(&qp->piowait);
 790         spin_unlock_irqrestore(&dev->pending_lock, flags);
 791
 792         /*
 793          * Make sure that the QP is not in the QPN table so receive
 794          * interrupts will discard packets for this QP.  XXX Also remove QP
 795          * from multicast table.
 796          */
 797         if (atomic_read(&qp->refcount) != 0)
 798                 ipath_free_qp(&dev->qp_table, qp);
 799
 800         vfree(qp->s_wq);
 801         vfree(qp->r_rq.wq);
 802         kfree(qp);
 803         return 0;
 804 }
 805
 806 /**
 807  * ipath_init_qp_table - initialize the QP table for a device
 808  * @idev: the device who's QP table we're initializing
 809  * @size: the size of the QP table
 810  *
 811  * Returns 0 on success, otherwise returns an errno.
 812  */
 813 int ipath_init_qp_table(struct ipath_ibdev *idev, int size)
 814 {
 815         int i;
 816         int ret;
 817
 818         idev->qp_table.last = 1;        /* QPN 0 and 1 are special. */
 819         idev->qp_table.max = size;
 820         idev->qp_table.nmaps = 1;
 821         idev->qp_table.table = kzalloc(size * sizeof(*idev->qp_table.table),
 822                                        GFP_KERNEL);
 823         if (idev->qp_table.table == NULL) {
 824                 ret = -ENOMEM;
 825                 goto bail;
 826         }
 827
 828         for (i = 0; i < ARRAY_SIZE(idev->qp_table.map); i++) {
 829                 atomic_set(&idev->qp_table.map[i].n_free, BITS_PER_PAGE);
 830                 idev->qp_table.map[i].page = NULL;
 831         }
 832
 833         ret = 0;
 834
 835 bail:
 836         return ret;
 837 }
 838
 839 /**
 840  * ipath_sqerror_qp - put a QP's send queue into an error state
 841  * @qp: QP who's send queue will be put into an error state
 842  * @wc: the WC responsible for putting the QP in this state
 843  *
 844  * Flushes the send work queue.
 845  * The QP s_lock should be held.
 846  */
 847
 848 void ipath_sqerror_qp(struct ipath_qp *qp, struct ib_wc *wc)
 849 {
 850         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 851         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 852
 853         _VERBS_INFO("Send queue error on QP%d/%d: err: %d\n",
 854                     qp->ibqp.qp_num, qp->remote_qpn, wc->status);
 855
 856         spin_lock(&dev->pending_lock);
 857         /* XXX What if its already removed by the timeout code? */
 858         if (!list_empty(&qp->timerwait))
 859                 list_del_init(&qp->timerwait);
 860         if (!list_empty(&qp->piowait))
 861                 list_del_init(&qp->piowait);
 862         spin_unlock(&dev->pending_lock);
 863
 864         ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
 865         if (++qp->s_last >= qp->s_size)
 866                 qp->s_last = 0;
 867
 868         wc->status = IB_WC_WR_FLUSH_ERR;
 869
 870         while (qp->s_last != qp->s_head) {
 871                 wc->wr_id = wqe->wr.wr_id;
 872                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 873                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
 874                 if (++qp->s_last >= qp->s_size)
 875                         qp->s_last = 0;
 876                 wqe = get_swqe_ptr(qp, qp->s_last);
 877         }
 878         qp->s_cur = qp->s_tail = qp->s_head;
 879         qp->state = IB_QPS_SQE;
 880 }
 881
 882 /**
 883  * ipath_get_credit - flush the send work queue of a QP
 884  * @qp: the qp who's send work queue to flush
 885  * @aeth: the Acknowledge Extended Transport Header
 886  *
 887  * The QP s_lock should be held.
 888  */
 889 void ipath_get_credit(struct ipath_qp *qp, u32 aeth)
 890 {
 891         u32 credit = (aeth >> IPS_AETH_CREDIT_SHIFT) & IPS_AETH_CREDIT_MASK;
 892
 893         /*
 894          * If the credit is invalid, we can send
 895          * as many packets as we like.  Otherwise, we have to
 896          * honor the credit field.
 897          */
 898         if (credit == IPS_AETH_CREDIT_INVAL)
 899                 qp->s_lsn = (u32) -1;
 900         else if (qp->s_lsn != (u32) -1) {
 901                 /* Compute new LSN (i.e., MSN + credit) */
 902                 credit = (aeth + credit_table[credit]) & IPS_MSN_MASK;
 903                 if (ipath_cmp24(credit, qp->s_lsn) > 0)
 904                         qp->s_lsn = credit;
 905         }
 906
 907         /* Restart sending if it was blocked due to lack of credits. */
 908         if (qp->s_cur != qp->s_head &&
 909             (qp->s_lsn == (u32) -1 ||
 910              ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn,
 911                          qp->s_lsn + 1) <= 0))
 912                 tasklet_hi_schedule(&qp->s_task);
 913 }