drivers/infiniband/hw/qib/qib_verbs.c

   1 /*
   2  * Copyright (c) 2012, 2013 Intel Corporation.  All rights reserved.
   3  * Copyright (c) 2006 - 2012 QLogic Corporation. All rights reserved.
   4  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include <rdma/ib_mad.h>
  36 #include <rdma/ib_user_verbs.h>
  37 #include <linux/io.h>
  38 #include <linux/module.h>
  39 #include <linux/utsname.h>
  40 #include <linux/rculist.h>
  41 #include <linux/mm.h>
  42 #include <linux/random.h>
  43 #include <linux/vmalloc.h>
  44 #include <rdma/rdma_vt.h>
  45
  46 #include "qib.h"
  47 #include "qib_common.h"
  48
  49 static unsigned int ib_qib_qp_table_size = 256;
  50 module_param_named(qp_table_size, ib_qib_qp_table_size, uint, S_IRUGO);
  51 MODULE_PARM_DESC(qp_table_size, "QP table size");
  52
  53 static unsigned int qib_lkey_table_size = 16;
  54 module_param_named(lkey_table_size, qib_lkey_table_size, uint,
  55                    S_IRUGO);
  56 MODULE_PARM_DESC(lkey_table_size,
  57                  "LKEY table size in bits (2^n, 1 <= n <= 23)");
  58
  59 static unsigned int ib_qib_max_pds = 0xFFFF;
  60 module_param_named(max_pds, ib_qib_max_pds, uint, S_IRUGO);
  61 MODULE_PARM_DESC(max_pds,
  62                  "Maximum number of protection domains to support");
  63
  64 static unsigned int ib_qib_max_ahs = 0xFFFF;
  65 module_param_named(max_ahs, ib_qib_max_ahs, uint, S_IRUGO);
  66 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
  67
  68 unsigned int ib_qib_max_cqes = 0x2FFFF;
  69 module_param_named(max_cqes, ib_qib_max_cqes, uint, S_IRUGO);
  70 MODULE_PARM_DESC(max_cqes,
  71                  "Maximum number of completion queue entries to support");
  72
  73 unsigned int ib_qib_max_cqs = 0x1FFFF;
  74 module_param_named(max_cqs, ib_qib_max_cqs, uint, S_IRUGO);
  75 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
  76
  77 unsigned int ib_qib_max_qp_wrs = 0x3FFF;
  78 module_param_named(max_qp_wrs, ib_qib_max_qp_wrs, uint, S_IRUGO);
  79 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
  80
  81 unsigned int ib_qib_max_qps = 16384;
  82 module_param_named(max_qps, ib_qib_max_qps, uint, S_IRUGO);
  83 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
  84
  85 unsigned int ib_qib_max_sges = 0x60;
  86 module_param_named(max_sges, ib_qib_max_sges, uint, S_IRUGO);
  87 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
  88
  89 unsigned int ib_qib_max_mcast_grps = 16384;
  90 module_param_named(max_mcast_grps, ib_qib_max_mcast_grps, uint, S_IRUGO);
  91 MODULE_PARM_DESC(max_mcast_grps,
  92                  "Maximum number of multicast groups to support");
  93
  94 unsigned int ib_qib_max_mcast_qp_attached = 16;
  95 module_param_named(max_mcast_qp_attached, ib_qib_max_mcast_qp_attached,
  96                    uint, S_IRUGO);
  97 MODULE_PARM_DESC(max_mcast_qp_attached,
  98                  "Maximum number of attached QPs to support");
  99
 100 unsigned int ib_qib_max_srqs = 1024;
 101 module_param_named(max_srqs, ib_qib_max_srqs, uint, S_IRUGO);
 102 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
 103
 104 unsigned int ib_qib_max_srq_sges = 128;
 105 module_param_named(max_srq_sges, ib_qib_max_srq_sges, uint, S_IRUGO);
 106 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
 107
 108 unsigned int ib_qib_max_srq_wrs = 0x1FFFF;
 109 module_param_named(max_srq_wrs, ib_qib_max_srq_wrs, uint, S_IRUGO);
 110 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
 111
 112 static unsigned int ib_qib_disable_sma;
 113 module_param_named(disable_sma, ib_qib_disable_sma, uint, S_IWUSR | S_IRUGO);
 114 MODULE_PARM_DESC(disable_sma, "Disable the SMA");
 115
 116 /*
 117  * Translate ib_wr_opcode into ib_wc_opcode.
 118  */
 119 const enum ib_wc_opcode ib_qib_wc_opcode[] = {
 120         [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
 121         [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
 122         [IB_WR_SEND] = IB_WC_SEND,
 123         [IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
 124         [IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
 125         [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
 126         [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD
 127 };
 128
 129 /*
 130  * System image GUID.
 131  */
 132 __be64 ib_qib_sys_image_guid;
 133
 134 /**
 135  * qib_copy_sge - copy data to SGE memory
 136  * @ss: the SGE state
 137  * @data: the data to copy
 138  * @length: the length of the data
 139  */
 140 void qib_copy_sge(struct rvt_sge_state *ss, void *data, u32 length, int release)
 141 {
 142         struct rvt_sge *sge = &ss->sge;
 143
 144         while (length) {
 145                 u32 len = rvt_get_sge_length(sge, length);
 146
 147                 WARN_ON_ONCE(len == 0);
 148                 memcpy(sge->vaddr, data, len);
 149                 rvt_update_sge(ss, len, release);
 150                 data += len;
 151                 length -= len;
 152         }
 153 }
 154
 155 /*
 156  * Count the number of DMA descriptors needed to send length bytes of data.
 157  * Don't modify the qib_sge_state to get the count.
 158  * Return zero if any of the segments is not aligned.
 159  */
 160 static u32 qib_count_sge(struct rvt_sge_state *ss, u32 length)
 161 {
 162         struct rvt_sge *sg_list = ss->sg_list;
 163         struct rvt_sge sge = ss->sge;
 164         u8 num_sge = ss->num_sge;
 165         u32 ndesc = 1;  /* count the header */
 166
 167         while (length) {
 168                 u32 len = sge.length;
 169
 170                 if (len > length)
 171                         len = length;
 172                 if (len > sge.sge_length)
 173                         len = sge.sge_length;
 174                 BUG_ON(len == 0);
 175                 if (((long) sge.vaddr & (sizeof(u32) - 1)) ||
 176                     (len != length && (len & (sizeof(u32) - 1)))) {
 177                         ndesc = 0;
 178                         break;
 179                 }
 180                 ndesc++;
 181                 sge.vaddr += len;
 182                 sge.length -= len;
 183                 sge.sge_length -= len;
 184                 if (sge.sge_length == 0) {
 185                         if (--num_sge)
 186                                 sge = *sg_list++;
 187                 } else if (sge.length == 0 && sge.mr->lkey) {
 188                         if (++sge.n >= RVT_SEGSZ) {
 189                                 if (++sge.m >= sge.mr->mapsz)
 190                                         break;
 191                                 sge.n = 0;
 192                         }
 193                         sge.vaddr =
 194                                 sge.mr->map[sge.m]->segs[sge.n].vaddr;
 195                         sge.length =
 196                                 sge.mr->map[sge.m]->segs[sge.n].length;
 197                 }
 198                 length -= len;
 199         }
 200         return ndesc;
 201 }
 202
 203 /*
 204  * Copy from the SGEs to the data buffer.
 205  */
 206 static void qib_copy_from_sge(void *data, struct rvt_sge_state *ss, u32 length)
 207 {
 208         struct rvt_sge *sge = &ss->sge;
 209
 210         while (length) {
 211                 u32 len = sge->length;
 212
 213                 if (len > length)
 214                         len = length;
 215                 if (len > sge->sge_length)
 216                         len = sge->sge_length;
 217                 BUG_ON(len == 0);
 218                 memcpy(data, sge->vaddr, len);
 219                 sge->vaddr += len;
 220                 sge->length -= len;
 221                 sge->sge_length -= len;
 222                 if (sge->sge_length == 0) {
 223                         if (--ss->num_sge)
 224                                 *sge = *ss->sg_list++;
 225                 } else if (sge->length == 0 && sge->mr->lkey) {
 226                         if (++sge->n >= RVT_SEGSZ) {
 227                                 if (++sge->m >= sge->mr->mapsz)
 228                                         break;
 229                                 sge->n = 0;
 230                         }
 231                         sge->vaddr =
 232                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 233                         sge->length =
 234                                 sge->mr->map[sge->m]->segs[sge->n].length;
 235                 }
 236                 data += len;
 237                 length -= len;
 238         }
 239 }
 240
 241 /**
 242  * qib_qp_rcv - processing an incoming packet on a QP
 243  * @rcd: the context pointer
 244  * @hdr: the packet header
 245  * @has_grh: true if the packet has a GRH
 246  * @data: the packet data
 247  * @tlen: the packet length
 248  * @qp: the QP the packet came on
 249  *
 250  * This is called from qib_ib_rcv() to process an incoming packet
 251  * for the given QP.
 252  * Called at interrupt level.
 253  */
 254 static void qib_qp_rcv(struct qib_ctxtdata *rcd, struct ib_header *hdr,
 255                        int has_grh, void *data, u32 tlen, struct rvt_qp *qp)
 256 {
 257         struct qib_ibport *ibp = &rcd->ppd->ibport_data;
 258
 259         spin_lock(&qp->r_lock);
 260
 261         /* Check for valid receive state. */
 262         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
 263                 ibp->rvp.n_pkt_drops++;
 264                 goto unlock;
 265         }
 266
 267         switch (qp->ibqp.qp_type) {
 268         case IB_QPT_SMI:
 269         case IB_QPT_GSI:
 270                 if (ib_qib_disable_sma)
 271                         break;
 272                 /* FALLTHROUGH */
 273         case IB_QPT_UD:
 274                 qib_ud_rcv(ibp, hdr, has_grh, data, tlen, qp);
 275                 break;
 276
 277         case IB_QPT_RC:
 278                 qib_rc_rcv(rcd, hdr, has_grh, data, tlen, qp);
 279                 break;
 280
 281         case IB_QPT_UC:
 282                 qib_uc_rcv(ibp, hdr, has_grh, data, tlen, qp);
 283                 break;
 284
 285         default:
 286                 break;
 287         }
 288
 289 unlock:
 290         spin_unlock(&qp->r_lock);
 291 }
 292
 293 /**
 294  * qib_ib_rcv - process an incoming packet
 295  * @rcd: the context pointer
 296  * @rhdr: the header of the packet
 297  * @data: the packet payload
 298  * @tlen: the packet length
 299  *
 300  * This is called from qib_kreceive() to process an incoming packet at
 301  * interrupt level. Tlen is the length of the header + data + CRC in bytes.
 302  */
 303 void qib_ib_rcv(struct qib_ctxtdata *rcd, void *rhdr, void *data, u32 tlen)
 304 {
 305         struct qib_pportdata *ppd = rcd->ppd;
 306         struct qib_ibport *ibp = &ppd->ibport_data;
 307         struct ib_header *hdr = rhdr;
 308         struct qib_devdata *dd = ppd->dd;
 309         struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
 310         struct ib_other_headers *ohdr;
 311         struct rvt_qp *qp;
 312         u32 qp_num;
 313         int lnh;
 314         u8 opcode;
 315         u16 lid;
 316
 317         /* 24 == LRH+BTH+CRC */
 318         if (unlikely(tlen < 24))
 319                 goto drop;
 320
 321         /* Check for a valid destination LID (see ch. 7.11.1). */
 322         lid = be16_to_cpu(hdr->lrh[1]);
 323         if (lid < be16_to_cpu(IB_MULTICAST_LID_BASE)) {
 324                 lid &= ~((1 << ppd->lmc) - 1);
 325                 if (unlikely(lid != ppd->lid))
 326                         goto drop;
 327         }
 328
 329         /* Check for GRH */
 330         lnh = be16_to_cpu(hdr->lrh[0]) & 3;
 331         if (lnh == QIB_LRH_BTH)
 332                 ohdr = &hdr->u.oth;
 333         else if (lnh == QIB_LRH_GRH) {
 334                 u32 vtf;
 335
 336                 ohdr = &hdr->u.l.oth;
 337                 if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
 338                         goto drop;
 339                 vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
 340                 if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
 341                         goto drop;
 342         } else
 343                 goto drop;
 344
 345         opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0x7f;
 346 #ifdef CONFIG_DEBUG_FS
 347         rcd->opstats->stats[opcode].n_bytes += tlen;
 348         rcd->opstats->stats[opcode].n_packets++;
 349 #endif
 350
 351         /* Get the destination QP number. */
 352         qp_num = be32_to_cpu(ohdr->bth[1]) & RVT_QPN_MASK;
 353         if (qp_num == QIB_MULTICAST_QPN) {
 354                 struct rvt_mcast *mcast;
 355                 struct rvt_mcast_qp *p;
 356
 357                 if (lnh != QIB_LRH_GRH)
 358                         goto drop;
 359                 mcast = rvt_mcast_find(&ibp->rvp, &hdr->u.l.grh.dgid, lid);
 360                 if (mcast == NULL)
 361                         goto drop;
 362                 this_cpu_inc(ibp->pmastats->n_multicast_rcv);
 363                 list_for_each_entry_rcu(p, &mcast->qp_list, list)
 364                         qib_qp_rcv(rcd, hdr, 1, data, tlen, p->qp);
 365                 /*
 366                  * Notify rvt_multicast_detach() if it is waiting for us
 367                  * to finish.
 368                  */
 369                 if (atomic_dec_return(&mcast->refcount) <= 1)
 370                         wake_up(&mcast->wait);
 371         } else {
 372                 rcu_read_lock();
 373                 qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
 374                 if (!qp) {
 375                         rcu_read_unlock();
 376                         goto drop;
 377                 }
 378                 this_cpu_inc(ibp->pmastats->n_unicast_rcv);
 379                 qib_qp_rcv(rcd, hdr, lnh == QIB_LRH_GRH, data, tlen, qp);
 380                 rcu_read_unlock();
 381         }
 382         return;
 383
 384 drop:
 385         ibp->rvp.n_pkt_drops++;
 386 }
 387
 388 /*
 389  * This is called from a timer to check for QPs
 390  * which need kernel memory in order to send a packet.
 391  */
 392 static void mem_timer(unsigned long data)
 393 {
 394         struct qib_ibdev *dev = (struct qib_ibdev *) data;
 395         struct list_head *list = &dev->memwait;
 396         struct rvt_qp *qp = NULL;
 397         struct qib_qp_priv *priv = NULL;
 398         unsigned long flags;
 399
 400         spin_lock_irqsave(&dev->rdi.pending_lock, flags);
 401         if (!list_empty(list)) {
 402                 priv = list_entry(list->next, struct qib_qp_priv, iowait);
 403                 qp = priv->owner;
 404                 list_del_init(&priv->iowait);
 405                 rvt_get_qp(qp);
 406                 if (!list_empty(list))
 407                         mod_timer(&dev->mem_timer, jiffies + 1);
 408         }
 409         spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
 410
 411         if (qp) {
 412                 spin_lock_irqsave(&qp->s_lock, flags);
 413                 if (qp->s_flags & RVT_S_WAIT_KMEM) {
 414                         qp->s_flags &= ~RVT_S_WAIT_KMEM;
 415                         qib_schedule_send(qp);
 416                 }
 417                 spin_unlock_irqrestore(&qp->s_lock, flags);
 418                 rvt_put_qp(qp);
 419         }
 420 }
 421
 422 #ifdef __LITTLE_ENDIAN
 423 static inline u32 get_upper_bits(u32 data, u32 shift)
 424 {
 425         return data >> shift;
 426 }
 427
 428 static inline u32 set_upper_bits(u32 data, u32 shift)
 429 {
 430         return data << shift;
 431 }
 432
 433 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
 434 {
 435         data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
 436         data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
 437         return data;
 438 }
 439 #else
 440 static inline u32 get_upper_bits(u32 data, u32 shift)
 441 {
 442         return data << shift;
 443 }
 444
 445 static inline u32 set_upper_bits(u32 data, u32 shift)
 446 {
 447         return data >> shift;
 448 }
 449
 450 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
 451 {
 452         data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
 453         data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
 454         return data;
 455 }
 456 #endif
 457
 458 static void copy_io(u32 __iomem *piobuf, struct rvt_sge_state *ss,
 459                     u32 length, unsigned flush_wc)
 460 {
 461         u32 extra = 0;
 462         u32 data = 0;
 463         u32 last;
 464
 465         while (1) {
 466                 u32 len = ss->sge.length;
 467                 u32 off;
 468
 469                 if (len > length)
 470                         len = length;
 471                 if (len > ss->sge.sge_length)
 472                         len = ss->sge.sge_length;
 473                 BUG_ON(len == 0);
 474                 /* If the source address is not aligned, try to align it. */
 475                 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
 476                 if (off) {
 477                         u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
 478                                             ~(sizeof(u32) - 1));
 479                         u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
 480                         u32 y;
 481
 482                         y = sizeof(u32) - off;
 483                         if (len > y)
 484                                 len = y;
 485                         if (len + extra >= sizeof(u32)) {
 486                                 data |= set_upper_bits(v, extra *
 487                                                        BITS_PER_BYTE);
 488                                 len = sizeof(u32) - extra;
 489                                 if (len == length) {
 490                                         last = data;
 491                                         break;
 492                                 }
 493                                 __raw_writel(data, piobuf);
 494                                 piobuf++;
 495                                 extra = 0;
 496                                 data = 0;
 497                         } else {
 498                                 /* Clear unused upper bytes */
 499                                 data |= clear_upper_bytes(v, len, extra);
 500                                 if (len == length) {
 501                                         last = data;
 502                                         break;
 503                                 }
 504                                 extra += len;
 505                         }
 506                 } else if (extra) {
 507                         /* Source address is aligned. */
 508                         u32 *addr = (u32 *) ss->sge.vaddr;
 509                         int shift = extra * BITS_PER_BYTE;
 510                         int ushift = 32 - shift;
 511                         u32 l = len;
 512
 513                         while (l >= sizeof(u32)) {
 514                                 u32 v = *addr;
 515
 516                                 data |= set_upper_bits(v, shift);
 517                                 __raw_writel(data, piobuf);
 518                                 data = get_upper_bits(v, ushift);
 519                                 piobuf++;
 520                                 addr++;
 521                                 l -= sizeof(u32);
 522                         }
 523                         /*
 524                          * We still have 'extra' number of bytes leftover.
 525                          */
 526                         if (l) {
 527                                 u32 v = *addr;
 528
 529                                 if (l + extra >= sizeof(u32)) {
 530                                         data |= set_upper_bits(v, shift);
 531                                         len -= l + extra - sizeof(u32);
 532                                         if (len == length) {
 533                                                 last = data;
 534                                                 break;
 535                                         }
 536                                         __raw_writel(data, piobuf);
 537                                         piobuf++;
 538                                         extra = 0;
 539                                         data = 0;
 540                                 } else {
 541                                         /* Clear unused upper bytes */
 542                                         data |= clear_upper_bytes(v, l, extra);
 543                                         if (len == length) {
 544                                                 last = data;
 545                                                 break;
 546                                         }
 547                                         extra += l;
 548                                 }
 549                         } else if (len == length) {
 550                                 last = data;
 551                                 break;
 552                         }
 553                 } else if (len == length) {
 554                         u32 w;
 555
 556                         /*
 557                          * Need to round up for the last dword in the
 558                          * packet.
 559                          */
 560                         w = (len + 3) >> 2;
 561                         qib_pio_copy(piobuf, ss->sge.vaddr, w - 1);
 562                         piobuf += w - 1;
 563                         last = ((u32 *) ss->sge.vaddr)[w - 1];
 564                         break;
 565                 } else {
 566                         u32 w = len >> 2;
 567
 568                         qib_pio_copy(piobuf, ss->sge.vaddr, w);
 569                         piobuf += w;
 570
 571                         extra = len & (sizeof(u32) - 1);
 572                         if (extra) {
 573                                 u32 v = ((u32 *) ss->sge.vaddr)[w];
 574
 575                                 /* Clear unused upper bytes */
 576                                 data = clear_upper_bytes(v, extra, 0);
 577                         }
 578                 }
 579                 rvt_update_sge(ss, len, false);
 580                 length -= len;
 581         }
 582         /* Update address before sending packet. */
 583         rvt_update_sge(ss, length, false);
 584         if (flush_wc) {
 585                 /* must flush early everything before trigger word */
 586                 qib_flush_wc();
 587                 __raw_writel(last, piobuf);
 588                 /* be sure trigger word is written */
 589                 qib_flush_wc();
 590         } else
 591                 __raw_writel(last, piobuf);
 592 }
 593
 594 static noinline struct qib_verbs_txreq *__get_txreq(struct qib_ibdev *dev,
 595                                            struct rvt_qp *qp)
 596 {
 597         struct qib_qp_priv *priv = qp->priv;
 598         struct qib_verbs_txreq *tx;
 599         unsigned long flags;
 600
 601         spin_lock_irqsave(&qp->s_lock, flags);
 602         spin_lock(&dev->rdi.pending_lock);
 603
 604         if (!list_empty(&dev->txreq_free)) {
 605                 struct list_head *l = dev->txreq_free.next;
 606
 607                 list_del(l);
 608                 spin_unlock(&dev->rdi.pending_lock);
 609                 spin_unlock_irqrestore(&qp->s_lock, flags);
 610                 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
 611         } else {
 612                 if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK &&
 613                     list_empty(&priv->iowait)) {
 614                         dev->n_txwait++;
 615                         qp->s_flags |= RVT_S_WAIT_TX;
 616                         list_add_tail(&priv->iowait, &dev->txwait);
 617                 }
 618                 qp->s_flags &= ~RVT_S_BUSY;
 619                 spin_unlock(&dev->rdi.pending_lock);
 620                 spin_unlock_irqrestore(&qp->s_lock, flags);
 621                 tx = ERR_PTR(-EBUSY);
 622         }
 623         return tx;
 624 }
 625
 626 static inline struct qib_verbs_txreq *get_txreq(struct qib_ibdev *dev,
 627                                          struct rvt_qp *qp)
 628 {
 629         struct qib_verbs_txreq *tx;
 630         unsigned long flags;
 631
 632         spin_lock_irqsave(&dev->rdi.pending_lock, flags);
 633         /* assume the list non empty */
 634         if (likely(!list_empty(&dev->txreq_free))) {
 635                 struct list_head *l = dev->txreq_free.next;
 636
 637                 list_del(l);
 638                 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
 639                 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
 640         } else {
 641                 /* call slow path to get the extra lock */
 642                 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
 643                 tx =  __get_txreq(dev, qp);
 644         }
 645         return tx;
 646 }
 647
 648 void qib_put_txreq(struct qib_verbs_txreq *tx)
 649 {
 650         struct qib_ibdev *dev;
 651         struct rvt_qp *qp;
 652         struct qib_qp_priv *priv;
 653         unsigned long flags;
 654
 655         qp = tx->qp;
 656         dev = to_idev(qp->ibqp.device);
 657
 658         if (tx->mr) {
 659                 rvt_put_mr(tx->mr);
 660                 tx->mr = NULL;
 661         }
 662         if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF) {
 663                 tx->txreq.flags &= ~QIB_SDMA_TXREQ_F_FREEBUF;
 664                 dma_unmap_single(&dd_from_dev(dev)->pcidev->dev,
 665                                  tx->txreq.addr, tx->hdr_dwords << 2,
 666                                  DMA_TO_DEVICE);
 667                 kfree(tx->align_buf);
 668         }
 669
 670         spin_lock_irqsave(&dev->rdi.pending_lock, flags);
 671
 672         /* Put struct back on free list */
 673         list_add(&tx->txreq.list, &dev->txreq_free);
 674
 675         if (!list_empty(&dev->txwait)) {
 676                 /* Wake up first QP wanting a free struct */
 677                 priv = list_entry(dev->txwait.next, struct qib_qp_priv,
 678                                   iowait);
 679                 qp = priv->owner;
 680                 list_del_init(&priv->iowait);
 681                 rvt_get_qp(qp);
 682                 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
 683
 684                 spin_lock_irqsave(&qp->s_lock, flags);
 685                 if (qp->s_flags & RVT_S_WAIT_TX) {
 686                         qp->s_flags &= ~RVT_S_WAIT_TX;
 687                         qib_schedule_send(qp);
 688                 }
 689                 spin_unlock_irqrestore(&qp->s_lock, flags);
 690
 691                 rvt_put_qp(qp);
 692         } else
 693                 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
 694 }
 695
 696 /*
 697  * This is called when there are send DMA descriptors that might be
 698  * available.
 699  *
 700  * This is called with ppd->sdma_lock held.
 701  */
 702 void qib_verbs_sdma_desc_avail(struct qib_pportdata *ppd, unsigned avail)
 703 {
 704         struct rvt_qp *qp, *nqp;
 705         struct qib_qp_priv *qpp, *nqpp;
 706         struct rvt_qp *qps[20];
 707         struct qib_ibdev *dev;
 708         unsigned i, n;
 709
 710         n = 0;
 711         dev = &ppd->dd->verbs_dev;
 712         spin_lock(&dev->rdi.pending_lock);
 713
 714         /* Search wait list for first QP wanting DMA descriptors. */
 715         list_for_each_entry_safe(qpp, nqpp, &dev->dmawait, iowait) {
 716                 qp = qpp->owner;
 717                 nqp = nqpp->owner;
 718                 if (qp->port_num != ppd->port)
 719                         continue;
 720                 if (n == ARRAY_SIZE(qps))
 721                         break;
 722                 if (qpp->s_tx->txreq.sg_count > avail)
 723                         break;
 724                 avail -= qpp->s_tx->txreq.sg_count;
 725                 list_del_init(&qpp->iowait);
 726                 rvt_get_qp(qp);
 727                 qps[n++] = qp;
 728         }
 729
 730         spin_unlock(&dev->rdi.pending_lock);
 731
 732         for (i = 0; i < n; i++) {
 733                 qp = qps[i];
 734                 spin_lock(&qp->s_lock);
 735                 if (qp->s_flags & RVT_S_WAIT_DMA_DESC) {
 736                         qp->s_flags &= ~RVT_S_WAIT_DMA_DESC;
 737                         qib_schedule_send(qp);
 738                 }
 739                 spin_unlock(&qp->s_lock);
 740                 rvt_put_qp(qp);
 741         }
 742 }
 743
 744 /*
 745  * This is called with ppd->sdma_lock held.
 746  */
 747 static void sdma_complete(struct qib_sdma_txreq *cookie, int status)
 748 {
 749         struct qib_verbs_txreq *tx =
 750                 container_of(cookie, struct qib_verbs_txreq, txreq);
 751         struct rvt_qp *qp = tx->qp;
 752         struct qib_qp_priv *priv = qp->priv;
 753
 754         spin_lock(&qp->s_lock);
 755         if (tx->wqe)
 756                 qib_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
 757         else if (qp->ibqp.qp_type == IB_QPT_RC) {
 758                 struct ib_header *hdr;
 759
 760                 if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF)
 761                         hdr = &tx->align_buf->hdr;
 762                 else {
 763                         struct qib_ibdev *dev = to_idev(qp->ibqp.device);
 764
 765                         hdr = &dev->pio_hdrs[tx->hdr_inx].hdr;
 766                 }
 767                 qib_rc_send_complete(qp, hdr);
 768         }
 769         if (atomic_dec_and_test(&priv->s_dma_busy)) {
 770                 if (qp->state == IB_QPS_RESET)
 771                         wake_up(&priv->wait_dma);
 772                 else if (qp->s_flags & RVT_S_WAIT_DMA) {
 773                         qp->s_flags &= ~RVT_S_WAIT_DMA;
 774                         qib_schedule_send(qp);
 775                 }
 776         }
 777         spin_unlock(&qp->s_lock);
 778
 779         qib_put_txreq(tx);
 780 }
 781
 782 static int wait_kmem(struct qib_ibdev *dev, struct rvt_qp *qp)
 783 {
 784         struct qib_qp_priv *priv = qp->priv;
 785         unsigned long flags;
 786         int ret = 0;
 787
 788         spin_lock_irqsave(&qp->s_lock, flags);
 789         if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
 790                 spin_lock(&dev->rdi.pending_lock);
 791                 if (list_empty(&priv->iowait)) {
 792                         if (list_empty(&dev->memwait))
 793                                 mod_timer(&dev->mem_timer, jiffies + 1);
 794                         qp->s_flags |= RVT_S_WAIT_KMEM;
 795                         list_add_tail(&priv->iowait, &dev->memwait);
 796                 }
 797                 spin_unlock(&dev->rdi.pending_lock);
 798                 qp->s_flags &= ~RVT_S_BUSY;
 799                 ret = -EBUSY;
 800         }
 801         spin_unlock_irqrestore(&qp->s_lock, flags);
 802
 803         return ret;
 804 }
 805
 806 static int qib_verbs_send_dma(struct rvt_qp *qp, struct ib_header *hdr,
 807                               u32 hdrwords, struct rvt_sge_state *ss, u32 len,
 808                               u32 plen, u32 dwords)
 809 {
 810         struct qib_qp_priv *priv = qp->priv;
 811         struct qib_ibdev *dev = to_idev(qp->ibqp.device);
 812         struct qib_devdata *dd = dd_from_dev(dev);
 813         struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 814         struct qib_pportdata *ppd = ppd_from_ibp(ibp);
 815         struct qib_verbs_txreq *tx;
 816         struct qib_pio_header *phdr;
 817         u32 control;
 818         u32 ndesc;
 819         int ret;
 820
 821         tx = priv->s_tx;
 822         if (tx) {
 823                 priv->s_tx = NULL;
 824                 /* resend previously constructed packet */
 825                 ret = qib_sdma_verbs_send(ppd, tx->ss, tx->dwords, tx);
 826                 goto bail;
 827         }
 828
 829         tx = get_txreq(dev, qp);
 830         if (IS_ERR(tx))
 831                 goto bail_tx;
 832
 833         control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
 834                                        be16_to_cpu(hdr->lrh[0]) >> 12);
 835         tx->qp = qp;
 836         tx->wqe = qp->s_wqe;
 837         tx->mr = qp->s_rdma_mr;
 838         if (qp->s_rdma_mr)
 839                 qp->s_rdma_mr = NULL;
 840         tx->txreq.callback = sdma_complete;
 841         if (dd->flags & QIB_HAS_SDMA_TIMEOUT)
 842                 tx->txreq.flags = QIB_SDMA_TXREQ_F_HEADTOHOST;
 843         else
 844                 tx->txreq.flags = QIB_SDMA_TXREQ_F_INTREQ;
 845         if (plen + 1 > dd->piosize2kmax_dwords)
 846                 tx->txreq.flags |= QIB_SDMA_TXREQ_F_USELARGEBUF;
 847
 848         if (len) {
 849                 /*
 850                  * Don't try to DMA if it takes more descriptors than
 851                  * the queue holds.
 852                  */
 853                 ndesc = qib_count_sge(ss, len);
 854                 if (ndesc >= ppd->sdma_descq_cnt)
 855                         ndesc = 0;
 856         } else
 857                 ndesc = 1;
 858         if (ndesc) {
 859                 phdr = &dev->pio_hdrs[tx->hdr_inx];
 860                 phdr->pbc[0] = cpu_to_le32(plen);
 861                 phdr->pbc[1] = cpu_to_le32(control);
 862                 memcpy(&phdr->hdr, hdr, hdrwords << 2);
 863                 tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEDESC;
 864                 tx->txreq.sg_count = ndesc;
 865                 tx->txreq.addr = dev->pio_hdrs_phys +
 866                         tx->hdr_inx * sizeof(struct qib_pio_header);
 867                 tx->hdr_dwords = hdrwords + 2; /* add PBC length */
 868                 ret = qib_sdma_verbs_send(ppd, ss, dwords, tx);
 869                 goto bail;
 870         }
 871
 872         /* Allocate a buffer and copy the header and payload to it. */
 873         tx->hdr_dwords = plen + 1;
 874         phdr = kmalloc(tx->hdr_dwords << 2, GFP_ATOMIC);
 875         if (!phdr)
 876                 goto err_tx;
 877         phdr->pbc[0] = cpu_to_le32(plen);
 878         phdr->pbc[1] = cpu_to_le32(control);
 879         memcpy(&phdr->hdr, hdr, hdrwords << 2);
 880         qib_copy_from_sge((u32 *) &phdr->hdr + hdrwords, ss, len);
 881
 882         tx->txreq.addr = dma_map_single(&dd->pcidev->dev, phdr,
 883                                         tx->hdr_dwords << 2, DMA_TO_DEVICE);
 884         if (dma_mapping_error(&dd->pcidev->dev, tx->txreq.addr))
 885                 goto map_err;
 886         tx->align_buf = phdr;
 887         tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEBUF;
 888         tx->txreq.sg_count = 1;
 889         ret = qib_sdma_verbs_send(ppd, NULL, 0, tx);
 890         goto unaligned;
 891
 892 map_err:
 893         kfree(phdr);
 894 err_tx:
 895         qib_put_txreq(tx);
 896         ret = wait_kmem(dev, qp);
 897 unaligned:
 898         ibp->rvp.n_unaligned++;
 899 bail:
 900         return ret;
 901 bail_tx:
 902         ret = PTR_ERR(tx);
 903         goto bail;
 904 }
 905
 906 /*
 907  * If we are now in the error state, return zero to flush the
 908  * send work request.
 909  */
 910 static int no_bufs_available(struct rvt_qp *qp)
 911 {
 912         struct qib_qp_priv *priv = qp->priv;
 913         struct qib_ibdev *dev = to_idev(qp->ibqp.device);
 914         struct qib_devdata *dd;
 915         unsigned long flags;
 916         int ret = 0;
 917
 918         /*
 919          * Note that as soon as want_buffer() is called and
 920          * possibly before it returns, qib_ib_piobufavail()
 921          * could be called. Therefore, put QP on the I/O wait list before
 922          * enabling the PIO avail interrupt.
 923          */
 924         spin_lock_irqsave(&qp->s_lock, flags);
 925         if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
 926                 spin_lock(&dev->rdi.pending_lock);
 927                 if (list_empty(&priv->iowait)) {
 928                         dev->n_piowait++;
 929                         qp->s_flags |= RVT_S_WAIT_PIO;
 930                         list_add_tail(&priv->iowait, &dev->piowait);
 931                         dd = dd_from_dev(dev);
 932                         dd->f_wantpiobuf_intr(dd, 1);
 933                 }
 934                 spin_unlock(&dev->rdi.pending_lock);
 935                 qp->s_flags &= ~RVT_S_BUSY;
 936                 ret = -EBUSY;
 937         }
 938         spin_unlock_irqrestore(&qp->s_lock, flags);
 939         return ret;
 940 }
 941
 942 static int qib_verbs_send_pio(struct rvt_qp *qp, struct ib_header *ibhdr,
 943                               u32 hdrwords, struct rvt_sge_state *ss, u32 len,
 944                               u32 plen, u32 dwords)
 945 {
 946         struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
 947         struct qib_pportdata *ppd = dd->pport + qp->port_num - 1;
 948         u32 *hdr = (u32 *) ibhdr;
 949         u32 __iomem *piobuf_orig;
 950         u32 __iomem *piobuf;
 951         u64 pbc;
 952         unsigned long flags;
 953         unsigned flush_wc;
 954         u32 control;
 955         u32 pbufn;
 956
 957         control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
 958                 be16_to_cpu(ibhdr->lrh[0]) >> 12);
 959         pbc = ((u64) control << 32) | plen;
 960         piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn);
 961         if (unlikely(piobuf == NULL))
 962                 return no_bufs_available(qp);
 963
 964         /*
 965          * Write the pbc.
 966          * We have to flush after the PBC for correctness on some cpus
 967          * or WC buffer can be written out of order.
 968          */
 969         writeq(pbc, piobuf);
 970         piobuf_orig = piobuf;
 971         piobuf += 2;
 972
 973         flush_wc = dd->flags & QIB_PIO_FLUSH_WC;
 974         if (len == 0) {
 975                 /*
 976                  * If there is just the header portion, must flush before
 977                  * writing last word of header for correctness, and after
 978                  * the last header word (trigger word).
 979                  */
 980                 if (flush_wc) {
 981                         qib_flush_wc();
 982                         qib_pio_copy(piobuf, hdr, hdrwords - 1);
 983                         qib_flush_wc();
 984                         __raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
 985                         qib_flush_wc();
 986                 } else
 987                         qib_pio_copy(piobuf, hdr, hdrwords);
 988                 goto done;
 989         }
 990
 991         if (flush_wc)
 992                 qib_flush_wc();
 993         qib_pio_copy(piobuf, hdr, hdrwords);
 994         piobuf += hdrwords;
 995
 996         /* The common case is aligned and contained in one segment. */
 997         if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
 998                    !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
 999                 u32 *addr = (u32 *) ss->sge.vaddr;
1000
1001                 /* Update address before sending packet. */
1002                 rvt_update_sge(ss, len, false);
1003                 if (flush_wc) {
1004                         qib_pio_copy(piobuf, addr, dwords - 1);
1005                         /* must flush early everything before trigger word */
1006                         qib_flush_wc();
1007                         __raw_writel(addr[dwords - 1], piobuf + dwords - 1);
1008                         /* be sure trigger word is written */
1009                         qib_flush_wc();
1010                 } else
1011                         qib_pio_copy(piobuf, addr, dwords);
1012                 goto done;
1013         }
1014         copy_io(piobuf, ss, len, flush_wc);
1015 done:
1016         if (dd->flags & QIB_USE_SPCL_TRIG) {
1017                 u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023;
1018
1019                 qib_flush_wc();
1020                 __raw_writel(0xaebecede, piobuf_orig + spcl_off);
1021         }
1022         qib_sendbuf_done(dd, pbufn);
1023         if (qp->s_rdma_mr) {
1024                 rvt_put_mr(qp->s_rdma_mr);
1025                 qp->s_rdma_mr = NULL;
1026         }
1027         if (qp->s_wqe) {
1028                 spin_lock_irqsave(&qp->s_lock, flags);
1029                 qib_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
1030                 spin_unlock_irqrestore(&qp->s_lock, flags);
1031         } else if (qp->ibqp.qp_type == IB_QPT_RC) {
1032                 spin_lock_irqsave(&qp->s_lock, flags);
1033                 qib_rc_send_complete(qp, ibhdr);
1034                 spin_unlock_irqrestore(&qp->s_lock, flags);
1035         }
1036         return 0;
1037 }
1038
1039 /**
1040  * qib_verbs_send - send a packet
1041  * @qp: the QP to send on
1042  * @hdr: the packet header
1043  * @hdrwords: the number of 32-bit words in the header
1044  * @ss: the SGE to send
1045  * @len: the length of the packet in bytes
1046  *
1047  * Return zero if packet is sent or queued OK.
1048  * Return non-zero and clear qp->s_flags RVT_S_BUSY otherwise.
1049  */
1050 int qib_verbs_send(struct rvt_qp *qp, struct ib_header *hdr,
1051                    u32 hdrwords, struct rvt_sge_state *ss, u32 len)
1052 {
1053         struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1054         u32 plen;
1055         int ret;
1056         u32 dwords = (len + 3) >> 2;
1057
1058         /*
1059          * Calculate the send buffer trigger address.
1060          * The +1 counts for the pbc control dword following the pbc length.
1061          */
1062         plen = hdrwords + dwords + 1;
1063
1064         /*
1065          * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1066          * can defer SDMA restart until link goes ACTIVE without
1067          * worrying about just how we got there.
1068          */
1069         if (qp->ibqp.qp_type == IB_QPT_SMI ||
1070             !(dd->flags & QIB_HAS_SEND_DMA))
1071                 ret = qib_verbs_send_pio(qp, hdr, hdrwords, ss, len,
1072                                          plen, dwords);
1073         else
1074                 ret = qib_verbs_send_dma(qp, hdr, hdrwords, ss, len,
1075                                          plen, dwords);
1076
1077         return ret;
1078 }
1079
1080 int qib_snapshot_counters(struct qib_pportdata *ppd, u64 *swords,
1081                           u64 *rwords, u64 *spkts, u64 *rpkts,
1082                           u64 *xmit_wait)
1083 {
1084         int ret;
1085         struct qib_devdata *dd = ppd->dd;
1086
1087         if (!(dd->flags & QIB_PRESENT)) {
1088                 /* no hardware, freeze, etc. */
1089                 ret = -EINVAL;
1090                 goto bail;
1091         }
1092         *swords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDSEND);
1093         *rwords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDRCV);
1094         *spkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTSEND);
1095         *rpkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTRCV);
1096         *xmit_wait = dd->f_portcntr(ppd, QIBPORTCNTR_SENDSTALL);
1097
1098         ret = 0;
1099
1100 bail:
1101         return ret;
1102 }
1103
1104 /**
1105  * qib_get_counters - get various chip counters
1106  * @dd: the qlogic_ib device
1107  * @cntrs: counters are placed here
1108  *
1109  * Return the counters needed by recv_pma_get_portcounters().
1110  */
1111 int qib_get_counters(struct qib_pportdata *ppd,
1112                      struct qib_verbs_counters *cntrs)
1113 {
1114         int ret;
1115
1116         if (!(ppd->dd->flags & QIB_PRESENT)) {
1117                 /* no hardware, freeze, etc. */
1118                 ret = -EINVAL;
1119                 goto bail;
1120         }
1121         cntrs->symbol_error_counter =
1122                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBSYMBOLERR);
1123         cntrs->link_error_recovery_counter =
1124                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKERRRECOV);
1125         /*
1126          * The link downed counter counts when the other side downs the
1127          * connection.  We add in the number of times we downed the link
1128          * due to local link integrity errors to compensate.
1129          */
1130         cntrs->link_downed_counter =
1131                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKDOWN);
1132         cntrs->port_rcv_errors =
1133                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXDROPPKT) +
1134                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVOVFL) +
1135                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERR_RLEN) +
1136                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_INVALIDRLEN) +
1137                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLINK) +
1138                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRICRC) +
1139                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRVCRC) +
1140                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLPCRC) +
1141                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_BADFORMAT);
1142         cntrs->port_rcv_errors +=
1143                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXLOCALPHYERR);
1144         cntrs->port_rcv_errors +=
1145                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXVLERR);
1146         cntrs->port_rcv_remphys_errors =
1147                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVEBP);
1148         cntrs->port_xmit_discards =
1149                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_UNSUPVL);
1150         cntrs->port_xmit_data = ppd->dd->f_portcntr(ppd,
1151                         QIBPORTCNTR_WORDSEND);
1152         cntrs->port_rcv_data = ppd->dd->f_portcntr(ppd,
1153                         QIBPORTCNTR_WORDRCV);
1154         cntrs->port_xmit_packets = ppd->dd->f_portcntr(ppd,
1155                         QIBPORTCNTR_PKTSEND);
1156         cntrs->port_rcv_packets = ppd->dd->f_portcntr(ppd,
1157                         QIBPORTCNTR_PKTRCV);
1158         cntrs->local_link_integrity_errors =
1159                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_LLI);
1160         cntrs->excessive_buffer_overrun_errors =
1161                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_EXCESSBUFOVFL);
1162         cntrs->vl15_dropped =
1163                 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_VL15PKTDROP);
1164
1165         ret = 0;
1166
1167 bail:
1168         return ret;
1169 }
1170
1171 /**
1172  * qib_ib_piobufavail - callback when a PIO buffer is available
1173  * @dd: the device pointer
1174  *
1175  * This is called from qib_intr() at interrupt level when a PIO buffer is
1176  * available after qib_verbs_send() returned an error that no buffers were
1177  * available. Disable the interrupt if there are no more QPs waiting.
1178  */
1179 void qib_ib_piobufavail(struct qib_devdata *dd)
1180 {
1181         struct qib_ibdev *dev = &dd->verbs_dev;
1182         struct list_head *list;
1183         struct rvt_qp *qps[5];
1184         struct rvt_qp *qp;
1185         unsigned long flags;
1186         unsigned i, n;
1187         struct qib_qp_priv *priv;
1188
1189         list = &dev->piowait;
1190         n = 0;
1191
1192         /*
1193          * Note: checking that the piowait list is empty and clearing
1194          * the buffer available interrupt needs to be atomic or we
1195          * could end up with QPs on the wait list with the interrupt
1196          * disabled.
1197          */
1198         spin_lock_irqsave(&dev->rdi.pending_lock, flags);
1199         while (!list_empty(list)) {
1200                 if (n == ARRAY_SIZE(qps))
1201                         goto full;
1202                 priv = list_entry(list->next, struct qib_qp_priv, iowait);
1203                 qp = priv->owner;
1204                 list_del_init(&priv->iowait);
1205                 rvt_get_qp(qp);
1206                 qps[n++] = qp;
1207         }
1208         dd->f_wantpiobuf_intr(dd, 0);
1209 full:
1210         spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
1211
1212         for (i = 0; i < n; i++) {
1213                 qp = qps[i];
1214
1215                 spin_lock_irqsave(&qp->s_lock, flags);
1216                 if (qp->s_flags & RVT_S_WAIT_PIO) {
1217                         qp->s_flags &= ~RVT_S_WAIT_PIO;
1218                         qib_schedule_send(qp);
1219                 }
1220                 spin_unlock_irqrestore(&qp->s_lock, flags);
1221
1222                 /* Notify qib_destroy_qp() if it is waiting. */
1223                 rvt_put_qp(qp);
1224         }
1225 }
1226
1227 static int qib_query_port(struct rvt_dev_info *rdi, u8 port_num,
1228                           struct ib_port_attr *props)
1229 {
1230         struct qib_ibdev *ibdev = container_of(rdi, struct qib_ibdev, rdi);
1231         struct qib_devdata *dd = dd_from_dev(ibdev);
1232         struct qib_pportdata *ppd = &dd->pport[port_num - 1];
1233         enum ib_mtu mtu;
1234         u16 lid = ppd->lid;
1235
1236         /* props being zeroed by the caller, avoid zeroing it here */
1237         props->lid = lid ? lid : be16_to_cpu(IB_LID_PERMISSIVE);
1238         props->lmc = ppd->lmc;
1239         props->state = dd->f_iblink_state(ppd->lastibcstat);
1240         props->phys_state = dd->f_ibphys_portstate(ppd->lastibcstat);
1241         props->gid_tbl_len = QIB_GUIDS_PER_PORT;
1242         props->active_width = ppd->link_width_active;
1243         /* See rate_show() */
1244         props->active_speed = ppd->link_speed_active;
1245         props->max_vl_num = qib_num_vls(ppd->vls_supported);
1246
1247         props->max_mtu = qib_ibmtu ? qib_ibmtu : IB_MTU_4096;
1248         switch (ppd->ibmtu) {
1249         case 4096:
1250                 mtu = IB_MTU_4096;
1251                 break;
1252         case 2048:
1253                 mtu = IB_MTU_2048;
1254                 break;
1255         case 1024:
1256                 mtu = IB_MTU_1024;
1257                 break;
1258         case 512:
1259                 mtu = IB_MTU_512;
1260                 break;
1261         case 256:
1262                 mtu = IB_MTU_256;
1263                 break;
1264         default:
1265                 mtu = IB_MTU_2048;
1266         }
1267         props->active_mtu = mtu;
1268
1269         return 0;
1270 }
1271
1272 static int qib_modify_device(struct ib_device *device,
1273                              int device_modify_mask,
1274                              struct ib_device_modify *device_modify)
1275 {
1276         struct qib_devdata *dd = dd_from_ibdev(device);
1277         unsigned i;
1278         int ret;
1279
1280         if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1281                                    IB_DEVICE_MODIFY_NODE_DESC)) {
1282                 ret = -EOPNOTSUPP;
1283                 goto bail;
1284         }
1285
1286         if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1287                 memcpy(device->node_desc, device_modify->node_desc,
1288                        IB_DEVICE_NODE_DESC_MAX);
1289                 for (i = 0; i < dd->num_pports; i++) {
1290                         struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1291
1292                         qib_node_desc_chg(ibp);
1293                 }
1294         }
1295
1296         if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1297                 ib_qib_sys_image_guid =
1298                         cpu_to_be64(device_modify->sys_image_guid);
1299                 for (i = 0; i < dd->num_pports; i++) {
1300                         struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1301
1302                         qib_sys_guid_chg(ibp);
1303                 }
1304         }
1305
1306         ret = 0;
1307
1308 bail:
1309         return ret;
1310 }
1311
1312 static int qib_shut_down_port(struct rvt_dev_info *rdi, u8 port_num)
1313 {
1314         struct qib_ibdev *ibdev = container_of(rdi, struct qib_ibdev, rdi);
1315         struct qib_devdata *dd = dd_from_dev(ibdev);
1316         struct qib_pportdata *ppd = &dd->pport[port_num - 1];
1317
1318         qib_set_linkstate(ppd, QIB_IB_LINKDOWN);
1319
1320         return 0;
1321 }
1322
1323 static int qib_get_guid_be(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
1324                            int guid_index, __be64 *guid)
1325 {
1326         struct qib_ibport *ibp = container_of(rvp, struct qib_ibport, rvp);
1327         struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1328
1329         if (guid_index == 0)
1330                 *guid = ppd->guid;
1331         else if (guid_index < QIB_GUIDS_PER_PORT)
1332                 *guid = ibp->guids[guid_index - 1];
1333         else
1334                 return -EINVAL;
1335
1336         return 0;
1337 }
1338
1339 int qib_check_ah(struct ib_device *ibdev, struct rdma_ah_attr *ah_attr)
1340 {
1341         if (rdma_ah_get_sl(ah_attr) > 15)
1342                 return -EINVAL;
1343
1344         return 0;
1345 }
1346
1347 static void qib_notify_new_ah(struct ib_device *ibdev,
1348                               struct rdma_ah_attr *ah_attr,
1349                               struct rvt_ah *ah)
1350 {
1351         struct qib_ibport *ibp;
1352         struct qib_pportdata *ppd;
1353
1354         /*
1355          * Do not trust reading anything from rvt_ah at this point as it is not
1356          * done being setup. We can however modify things which we need to set.
1357          */
1358
1359         ibp = to_iport(ibdev, rdma_ah_get_port_num(ah_attr));
1360         ppd = ppd_from_ibp(ibp);
1361         ah->vl = ibp->sl_to_vl[rdma_ah_get_sl(&ah->attr)];
1362         ah->log_pmtu = ilog2(ppd->ibmtu);
1363 }
1364
1365 struct ib_ah *qib_create_qp0_ah(struct qib_ibport *ibp, u16 dlid)
1366 {
1367         struct rdma_ah_attr attr;
1368         struct ib_ah *ah = ERR_PTR(-EINVAL);
1369         struct rvt_qp *qp0;
1370         struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1371         struct qib_devdata *dd = dd_from_ppd(ppd);
1372         u8 port_num = ppd->port;
1373
1374         memset(&attr, 0, sizeof(attr));
1375         attr.type = rdma_ah_find_type(&dd->verbs_dev.rdi.ibdev, port_num);
1376         rdma_ah_set_dlid(&attr, dlid);
1377         rdma_ah_set_port_num(&attr, port_num);
1378         rcu_read_lock();
1379         qp0 = rcu_dereference(ibp->rvp.qp[0]);
1380         if (qp0)
1381                 ah = rdma_create_ah(qp0->ibqp.pd, &attr);
1382         rcu_read_unlock();
1383         return ah;
1384 }
1385
1386 /**
1387  * qib_get_npkeys - return the size of the PKEY table for context 0
1388  * @dd: the qlogic_ib device
1389  */
1390 unsigned qib_get_npkeys(struct qib_devdata *dd)
1391 {
1392         return ARRAY_SIZE(dd->rcd[0]->pkeys);
1393 }
1394
1395 /*
1396  * Return the indexed PKEY from the port PKEY table.
1397  * No need to validate rcd[ctxt]; the port is setup if we are here.
1398  */
1399 unsigned qib_get_pkey(struct qib_ibport *ibp, unsigned index)
1400 {
1401         struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1402         struct qib_devdata *dd = ppd->dd;
1403         unsigned ctxt = ppd->hw_pidx;
1404         unsigned ret;
1405
1406         /* dd->rcd null if mini_init or some init failures */
1407         if (!dd->rcd || index >= ARRAY_SIZE(dd->rcd[ctxt]->pkeys))
1408                 ret = 0;
1409         else
1410                 ret = dd->rcd[ctxt]->pkeys[index];
1411
1412         return ret;
1413 }
1414
1415 static void init_ibport(struct qib_pportdata *ppd)
1416 {
1417         struct qib_verbs_counters cntrs;
1418         struct qib_ibport *ibp = &ppd->ibport_data;
1419
1420         spin_lock_init(&ibp->rvp.lock);
1421         /* Set the prefix to the default value (see ch. 4.1.1) */
1422         ibp->rvp.gid_prefix = IB_DEFAULT_GID_PREFIX;
1423         ibp->rvp.sm_lid = be16_to_cpu(IB_LID_PERMISSIVE);
1424         ibp->rvp.port_cap_flags = IB_PORT_SYS_IMAGE_GUID_SUP |
1425                 IB_PORT_CLIENT_REG_SUP | IB_PORT_SL_MAP_SUP |
1426                 IB_PORT_TRAP_SUP | IB_PORT_AUTO_MIGR_SUP |
1427                 IB_PORT_DR_NOTICE_SUP | IB_PORT_CAP_MASK_NOTICE_SUP |
1428                 IB_PORT_OTHER_LOCAL_CHANGES_SUP;
1429         if (ppd->dd->flags & QIB_HAS_LINK_LATENCY)
1430                 ibp->rvp.port_cap_flags |= IB_PORT_LINK_LATENCY_SUP;
1431         ibp->rvp.pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1432         ibp->rvp.pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1433         ibp->rvp.pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1434         ibp->rvp.pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1435         ibp->rvp.pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1436
1437         /* Snapshot current HW counters to "clear" them. */
1438         qib_get_counters(ppd, &cntrs);
1439         ibp->z_symbol_error_counter = cntrs.symbol_error_counter;
1440         ibp->z_link_error_recovery_counter =
1441                 cntrs.link_error_recovery_counter;
1442         ibp->z_link_downed_counter = cntrs.link_downed_counter;
1443         ibp->z_port_rcv_errors = cntrs.port_rcv_errors;
1444         ibp->z_port_rcv_remphys_errors = cntrs.port_rcv_remphys_errors;
1445         ibp->z_port_xmit_discards = cntrs.port_xmit_discards;
1446         ibp->z_port_xmit_data = cntrs.port_xmit_data;
1447         ibp->z_port_rcv_data = cntrs.port_rcv_data;
1448         ibp->z_port_xmit_packets = cntrs.port_xmit_packets;
1449         ibp->z_port_rcv_packets = cntrs.port_rcv_packets;
1450         ibp->z_local_link_integrity_errors =
1451                 cntrs.local_link_integrity_errors;
1452         ibp->z_excessive_buffer_overrun_errors =
1453                 cntrs.excessive_buffer_overrun_errors;
1454         ibp->z_vl15_dropped = cntrs.vl15_dropped;
1455         RCU_INIT_POINTER(ibp->rvp.qp[0], NULL);
1456         RCU_INIT_POINTER(ibp->rvp.qp[1], NULL);
1457 }
1458
1459 /**
1460  * qib_fill_device_attr - Fill in rvt dev info device attributes.
1461  * @dd: the device data structure
1462  */
1463 static void qib_fill_device_attr(struct qib_devdata *dd)
1464 {
1465         struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
1466
1467         memset(&rdi->dparms.props, 0, sizeof(rdi->dparms.props));
1468
1469         rdi->dparms.props.max_pd = ib_qib_max_pds;
1470         rdi->dparms.props.max_ah = ib_qib_max_ahs;
1471         rdi->dparms.props.device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1472                 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1473                 IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1474                 IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE;
1475         rdi->dparms.props.page_size_cap = PAGE_SIZE;
1476         rdi->dparms.props.vendor_id =
1477                 QIB_SRC_OUI_1 << 16 | QIB_SRC_OUI_2 << 8 | QIB_SRC_OUI_3;
1478         rdi->dparms.props.vendor_part_id = dd->deviceid;
1479         rdi->dparms.props.hw_ver = dd->minrev;
1480         rdi->dparms.props.sys_image_guid = ib_qib_sys_image_guid;
1481         rdi->dparms.props.max_mr_size = ~0ULL;
1482         rdi->dparms.props.max_qp = ib_qib_max_qps;
1483         rdi->dparms.props.max_qp_wr = ib_qib_max_qp_wrs;
1484         rdi->dparms.props.max_sge = ib_qib_max_sges;
1485         rdi->dparms.props.max_sge_rd = ib_qib_max_sges;
1486         rdi->dparms.props.max_cq = ib_qib_max_cqs;
1487         rdi->dparms.props.max_cqe = ib_qib_max_cqes;
1488         rdi->dparms.props.max_ah = ib_qib_max_ahs;
1489         rdi->dparms.props.max_mr = rdi->lkey_table.max;
1490         rdi->dparms.props.max_fmr = rdi->lkey_table.max;
1491         rdi->dparms.props.max_map_per_fmr = 32767;
1492         rdi->dparms.props.max_qp_rd_atom = QIB_MAX_RDMA_ATOMIC;
1493         rdi->dparms.props.max_qp_init_rd_atom = 255;
1494         rdi->dparms.props.max_srq = ib_qib_max_srqs;
1495         rdi->dparms.props.max_srq_wr = ib_qib_max_srq_wrs;
1496         rdi->dparms.props.max_srq_sge = ib_qib_max_srq_sges;
1497         rdi->dparms.props.atomic_cap = IB_ATOMIC_GLOB;
1498         rdi->dparms.props.max_pkeys = qib_get_npkeys(dd);
1499         rdi->dparms.props.max_mcast_grp = ib_qib_max_mcast_grps;
1500         rdi->dparms.props.max_mcast_qp_attach = ib_qib_max_mcast_qp_attached;
1501         rdi->dparms.props.max_total_mcast_qp_attach =
1502                                         rdi->dparms.props.max_mcast_qp_attach *
1503                                         rdi->dparms.props.max_mcast_grp;
1504         /* post send table */
1505         dd->verbs_dev.rdi.post_parms = qib_post_parms;
1506 }
1507
1508 /**
1509  * qib_register_ib_device - register our device with the infiniband core
1510  * @dd: the device data structure
1511  * Return the allocated qib_ibdev pointer or NULL on error.
1512  */
1513 int qib_register_ib_device(struct qib_devdata *dd)
1514 {
1515         struct qib_ibdev *dev = &dd->verbs_dev;
1516         struct ib_device *ibdev = &dev->rdi.ibdev;
1517         struct qib_pportdata *ppd = dd->pport;
1518         unsigned i, ctxt;
1519         int ret;
1520
1521         get_random_bytes(&dev->qp_rnd, sizeof(dev->qp_rnd));
1522         for (i = 0; i < dd->num_pports; i++)
1523                 init_ibport(ppd + i);
1524
1525         /* Only need to initialize non-zero fields. */
1526         setup_timer(&dev->mem_timer, mem_timer, (unsigned long)dev);
1527
1528         INIT_LIST_HEAD(&dev->piowait);
1529         INIT_LIST_HEAD(&dev->dmawait);
1530         INIT_LIST_HEAD(&dev->txwait);
1531         INIT_LIST_HEAD(&dev->memwait);
1532         INIT_LIST_HEAD(&dev->txreq_free);
1533
1534         if (ppd->sdma_descq_cnt) {
1535                 dev->pio_hdrs = dma_alloc_coherent(&dd->pcidev->dev,
1536                                                 ppd->sdma_descq_cnt *
1537                                                 sizeof(struct qib_pio_header),
1538                                                 &dev->pio_hdrs_phys,
1539                                                 GFP_KERNEL);
1540                 if (!dev->pio_hdrs) {
1541                         ret = -ENOMEM;
1542                         goto err_hdrs;
1543                 }
1544         }
1545
1546         for (i = 0; i < ppd->sdma_descq_cnt; i++) {
1547                 struct qib_verbs_txreq *tx;
1548
1549                 tx = kzalloc(sizeof(*tx), GFP_KERNEL);
1550                 if (!tx) {
1551                         ret = -ENOMEM;
1552                         goto err_tx;
1553                 }
1554                 tx->hdr_inx = i;
1555                 list_add(&tx->txreq.list, &dev->txreq_free);
1556         }
1557
1558         /*
1559          * The system image GUID is supposed to be the same for all
1560          * IB HCAs in a single system but since there can be other
1561          * device types in the system, we can't be sure this is unique.
1562          */
1563         if (!ib_qib_sys_image_guid)
1564                 ib_qib_sys_image_guid = ppd->guid;
1565
1566         strlcpy(ibdev->name, "qib%d", IB_DEVICE_NAME_MAX);
1567         ibdev->owner = THIS_MODULE;
1568         ibdev->node_guid = ppd->guid;
1569         ibdev->phys_port_cnt = dd->num_pports;
1570         ibdev->dev.parent = &dd->pcidev->dev;
1571         ibdev->modify_device = qib_modify_device;
1572         ibdev->process_mad = qib_process_mad;
1573
1574         snprintf(ibdev->node_desc, sizeof(ibdev->node_desc),
1575                  "Intel Infiniband HCA %s", init_utsname()->nodename);
1576
1577         /*
1578          * Fill in rvt info object.
1579          */
1580         dd->verbs_dev.rdi.driver_f.port_callback = qib_create_port_files;
1581         dd->verbs_dev.rdi.driver_f.get_card_name = qib_get_card_name;
1582         dd->verbs_dev.rdi.driver_f.get_pci_dev = qib_get_pci_dev;
1583         dd->verbs_dev.rdi.driver_f.check_ah = qib_check_ah;
1584         dd->verbs_dev.rdi.driver_f.check_send_wqe = qib_check_send_wqe;
1585         dd->verbs_dev.rdi.driver_f.notify_new_ah = qib_notify_new_ah;
1586         dd->verbs_dev.rdi.driver_f.alloc_qpn = qib_alloc_qpn;
1587         dd->verbs_dev.rdi.driver_f.qp_priv_alloc = qib_qp_priv_alloc;
1588         dd->verbs_dev.rdi.driver_f.qp_priv_free = qib_qp_priv_free;
1589         dd->verbs_dev.rdi.driver_f.free_all_qps = qib_free_all_qps;
1590         dd->verbs_dev.rdi.driver_f.notify_qp_reset = qib_notify_qp_reset;
1591         dd->verbs_dev.rdi.driver_f.do_send = qib_do_send;
1592         dd->verbs_dev.rdi.driver_f.schedule_send = qib_schedule_send;
1593         dd->verbs_dev.rdi.driver_f.quiesce_qp = qib_quiesce_qp;
1594         dd->verbs_dev.rdi.driver_f.stop_send_queue = qib_stop_send_queue;
1595         dd->verbs_dev.rdi.driver_f.flush_qp_waiters = qib_flush_qp_waiters;
1596         dd->verbs_dev.rdi.driver_f.notify_error_qp = qib_notify_error_qp;
1597         dd->verbs_dev.rdi.driver_f.notify_restart_rc = qib_restart_rc;
1598         dd->verbs_dev.rdi.driver_f.mtu_to_path_mtu = qib_mtu_to_path_mtu;
1599         dd->verbs_dev.rdi.driver_f.mtu_from_qp = qib_mtu_from_qp;
1600         dd->verbs_dev.rdi.driver_f.get_pmtu_from_attr = qib_get_pmtu_from_attr;
1601         dd->verbs_dev.rdi.driver_f.schedule_send_no_lock = _qib_schedule_send;
1602         dd->verbs_dev.rdi.driver_f.query_port_state = qib_query_port;
1603         dd->verbs_dev.rdi.driver_f.shut_down_port = qib_shut_down_port;
1604         dd->verbs_dev.rdi.driver_f.cap_mask_chg = qib_cap_mask_chg;
1605         dd->verbs_dev.rdi.driver_f.notify_create_mad_agent =
1606                                                 qib_notify_create_mad_agent;
1607         dd->verbs_dev.rdi.driver_f.notify_free_mad_agent =
1608                                                 qib_notify_free_mad_agent;
1609
1610         dd->verbs_dev.rdi.dparms.max_rdma_atomic = QIB_MAX_RDMA_ATOMIC;
1611         dd->verbs_dev.rdi.driver_f.get_guid_be = qib_get_guid_be;
1612         dd->verbs_dev.rdi.dparms.lkey_table_size = qib_lkey_table_size;
1613         dd->verbs_dev.rdi.dparms.qp_table_size = ib_qib_qp_table_size;
1614         dd->verbs_dev.rdi.dparms.qpn_start = 1;
1615         dd->verbs_dev.rdi.dparms.qpn_res_start = QIB_KD_QP;
1616         dd->verbs_dev.rdi.dparms.qpn_res_end = QIB_KD_QP; /* Reserve one QP */
1617         dd->verbs_dev.rdi.dparms.qpn_inc = 1;
1618         dd->verbs_dev.rdi.dparms.qos_shift = 1;
1619         dd->verbs_dev.rdi.dparms.psn_mask = QIB_PSN_MASK;
1620         dd->verbs_dev.rdi.dparms.psn_shift = QIB_PSN_SHIFT;
1621         dd->verbs_dev.rdi.dparms.psn_modify_mask = QIB_PSN_MASK;
1622         dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
1623         dd->verbs_dev.rdi.dparms.npkeys = qib_get_npkeys(dd);
1624         dd->verbs_dev.rdi.dparms.node = dd->assigned_node_id;
1625         dd->verbs_dev.rdi.dparms.core_cap_flags = RDMA_CORE_PORT_IBA_IB;
1626         dd->verbs_dev.rdi.dparms.max_mad_size = IB_MGMT_MAD_SIZE;
1627
1628         snprintf(dd->verbs_dev.rdi.dparms.cq_name,
1629                  sizeof(dd->verbs_dev.rdi.dparms.cq_name),
1630                  "qib_cq%d", dd->unit);
1631
1632         qib_fill_device_attr(dd);
1633
1634         ppd = dd->pport;
1635         for (i = 0; i < dd->num_pports; i++, ppd++) {
1636                 ctxt = ppd->hw_pidx;
1637                 rvt_init_port(&dd->verbs_dev.rdi,
1638                               &ppd->ibport_data.rvp,
1639                               i,
1640                               dd->rcd[ctxt]->pkeys);
1641         }
1642
1643         ret = rvt_register_device(&dd->verbs_dev.rdi);
1644         if (ret)
1645                 goto err_tx;
1646
1647         ret = qib_verbs_register_sysfs(dd);
1648         if (ret)
1649                 goto err_class;
1650
1651         return ret;
1652
1653 err_class:
1654         rvt_unregister_device(&dd->verbs_dev.rdi);
1655 err_tx:
1656         while (!list_empty(&dev->txreq_free)) {
1657                 struct list_head *l = dev->txreq_free.next;
1658                 struct qib_verbs_txreq *tx;
1659
1660                 list_del(l);
1661                 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
1662                 kfree(tx);
1663         }
1664         if (ppd->sdma_descq_cnt)
1665                 dma_free_coherent(&dd->pcidev->dev,
1666                                   ppd->sdma_descq_cnt *
1667                                         sizeof(struct qib_pio_header),
1668                                   dev->pio_hdrs, dev->pio_hdrs_phys);
1669 err_hdrs:
1670         qib_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1671         return ret;
1672 }
1673
1674 void qib_unregister_ib_device(struct qib_devdata *dd)
1675 {
1676         struct qib_ibdev *dev = &dd->verbs_dev;
1677
1678         qib_verbs_unregister_sysfs(dd);
1679
1680         rvt_unregister_device(&dd->verbs_dev.rdi);
1681
1682         if (!list_empty(&dev->piowait))
1683                 qib_dev_err(dd, "piowait list not empty!\n");
1684         if (!list_empty(&dev->dmawait))
1685                 qib_dev_err(dd, "dmawait list not empty!\n");
1686         if (!list_empty(&dev->txwait))
1687                 qib_dev_err(dd, "txwait list not empty!\n");
1688         if (!list_empty(&dev->memwait))
1689                 qib_dev_err(dd, "memwait list not empty!\n");
1690
1691         del_timer_sync(&dev->mem_timer);
1692         while (!list_empty(&dev->txreq_free)) {
1693                 struct list_head *l = dev->txreq_free.next;
1694                 struct qib_verbs_txreq *tx;
1695
1696                 list_del(l);
1697                 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
1698                 kfree(tx);
1699         }
1700         if (dd->pport->sdma_descq_cnt)
1701                 dma_free_coherent(&dd->pcidev->dev,
1702                                   dd->pport->sdma_descq_cnt *
1703                                         sizeof(struct qib_pio_header),
1704                                   dev->pio_hdrs, dev->pio_hdrs_phys);
1705 }
1706
1707 /**
1708  * _qib_schedule_send - schedule progress
1709  * @qp - the qp
1710  *
1711  * This schedules progress w/o regard to the s_flags.
1712  *
1713  * It is only used in post send, which doesn't hold
1714  * the s_lock.
1715  */
1716 void _qib_schedule_send(struct rvt_qp *qp)
1717 {
1718         struct qib_ibport *ibp =
1719                 to_iport(qp->ibqp.device, qp->port_num);
1720         struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1721         struct qib_qp_priv *priv = qp->priv;
1722
1723         queue_work(ppd->qib_wq, &priv->s_work);
1724 }
1725
1726 /**
1727  * qib_schedule_send - schedule progress
1728  * @qp - the qp
1729  *
1730  * This schedules qp progress.  The s_lock
1731  * should be held.
1732  */
1733 void qib_schedule_send(struct rvt_qp *qp)
1734 {
1735         if (qib_send_ok(qp))
1736                 _qib_schedule_send(qp);
1737 }