drivers/infiniband/hw/qib/qib_driver.c

   1 /*
   2  * Copyright (c) 2013 Intel Corporation. All rights reserved.
   3  * Copyright (c) 2006, 2007, 2008, 2009 QLogic Corporation. All rights reserved.
   4  * Copyright (c) 2003, 2004, 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 <linux/spinlock.h>
  36 #include <linux/pci.h>
  37 #include <linux/io.h>
  38 #include <linux/delay.h>
  39 #include <linux/netdevice.h>
  40 #include <linux/vmalloc.h>
  41 #include <linux/module.h>
  42 #include <linux/prefetch.h>
  43
  44 #include "qib.h"
  45
  46 /*
  47  * The size has to be longer than this string, so we can append
  48  * board/chip information to it in the init code.
  49  */
  50 const char ib_qib_version[] = QIB_DRIVER_VERSION "\n";
  51
  52 DEFINE_SPINLOCK(qib_devs_lock);
  53 LIST_HEAD(qib_dev_list);
  54 DEFINE_MUTEX(qib_mutex);        /* general driver use */
  55
  56 unsigned qib_ibmtu;
  57 module_param_named(ibmtu, qib_ibmtu, uint, S_IRUGO);
  58 MODULE_PARM_DESC(ibmtu, "Set max IB MTU (0=2KB, 1=256, 2=512, ... 5=4096");
  59
  60 unsigned qib_compat_ddr_negotiate = 1;
  61 module_param_named(compat_ddr_negotiate, qib_compat_ddr_negotiate, uint,
  62                    S_IWUSR | S_IRUGO);
  63 MODULE_PARM_DESC(compat_ddr_negotiate,
  64                  "Attempt pre-IBTA 1.2 DDR speed negotiation");
  65
  66 MODULE_LICENSE("Dual BSD/GPL");
  67 MODULE_AUTHOR("Intel <ibsupport@intel.com>");
  68 MODULE_DESCRIPTION("Intel IB driver");
  69
  70 /*
  71  * QIB_PIO_MAXIBHDR is the max IB header size allowed for in our
  72  * PIO send buffers.  This is well beyond anything currently
  73  * defined in the InfiniBand spec.
  74  */
  75 #define QIB_PIO_MAXIBHDR 128
  76
  77 /*
  78  * QIB_MAX_PKT_RCV is the max # if packets processed per receive interrupt.
  79  */
  80 #define QIB_MAX_PKT_RECV 64
  81
  82 struct qlogic_ib_stats qib_stats;
  83
  84 const char *qib_get_unit_name(int unit)
  85 {
  86         static char iname[16];
  87
  88         snprintf(iname, sizeof(iname), "infinipath%u", unit);
  89         return iname;
  90 }
  91
  92 const char *qib_get_card_name(struct rvt_dev_info *rdi)
  93 {
  94         struct qib_ibdev *ibdev = container_of(rdi, struct qib_ibdev, rdi);
  95         struct qib_devdata *dd = container_of(ibdev,
  96                                               struct qib_devdata, verbs_dev);
  97         return qib_get_unit_name(dd->unit);
  98 }
  99
 100 struct pci_dev *qib_get_pci_dev(struct rvt_dev_info *rdi)
 101 {
 102         struct qib_ibdev *ibdev = container_of(rdi, struct qib_ibdev, rdi);
 103         struct qib_devdata *dd = container_of(ibdev,
 104                                               struct qib_devdata, verbs_dev);
 105         return dd->pcidev;
 106 }
 107
 108 /*
 109  * Return count of units with at least one port ACTIVE.
 110  */
 111 int qib_count_active_units(void)
 112 {
 113         struct qib_devdata *dd;
 114         struct qib_pportdata *ppd;
 115         unsigned long flags;
 116         int pidx, nunits_active = 0;
 117
 118         spin_lock_irqsave(&qib_devs_lock, flags);
 119         list_for_each_entry(dd, &qib_dev_list, list) {
 120                 if (!(dd->flags & QIB_PRESENT) || !dd->kregbase)
 121                         continue;
 122                 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 123                         ppd = dd->pport + pidx;
 124                         if (ppd->lid && (ppd->lflags & (QIBL_LINKINIT |
 125                                          QIBL_LINKARMED | QIBL_LINKACTIVE))) {
 126                                 nunits_active++;
 127                                 break;
 128                         }
 129                 }
 130         }
 131         spin_unlock_irqrestore(&qib_devs_lock, flags);
 132         return nunits_active;
 133 }
 134
 135 /*
 136  * Return count of all units, optionally return in arguments
 137  * the number of usable (present) units, and the number of
 138  * ports that are up.
 139  */
 140 int qib_count_units(int *npresentp, int *nupp)
 141 {
 142         int nunits = 0, npresent = 0, nup = 0;
 143         struct qib_devdata *dd;
 144         unsigned long flags;
 145         int pidx;
 146         struct qib_pportdata *ppd;
 147
 148         spin_lock_irqsave(&qib_devs_lock, flags);
 149
 150         list_for_each_entry(dd, &qib_dev_list, list) {
 151                 nunits++;
 152                 if ((dd->flags & QIB_PRESENT) && dd->kregbase)
 153                         npresent++;
 154                 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 155                         ppd = dd->pport + pidx;
 156                         if (ppd->lid && (ppd->lflags & (QIBL_LINKINIT |
 157                                          QIBL_LINKARMED | QIBL_LINKACTIVE)))
 158                                 nup++;
 159                 }
 160         }
 161
 162         spin_unlock_irqrestore(&qib_devs_lock, flags);
 163
 164         if (npresentp)
 165                 *npresentp = npresent;
 166         if (nupp)
 167                 *nupp = nup;
 168
 169         return nunits;
 170 }
 171
 172 /**
 173  * qib_wait_linkstate - wait for an IB link state change to occur
 174  * @dd: the qlogic_ib device
 175  * @state: the state to wait for
 176  * @msecs: the number of milliseconds to wait
 177  *
 178  * wait up to msecs milliseconds for IB link state change to occur for
 179  * now, take the easy polling route.  Currently used only by
 180  * qib_set_linkstate.  Returns 0 if state reached, otherwise
 181  * -ETIMEDOUT state can have multiple states set, for any of several
 182  * transitions.
 183  */
 184 int qib_wait_linkstate(struct qib_pportdata *ppd, u32 state, int msecs)
 185 {
 186         int ret;
 187         unsigned long flags;
 188
 189         spin_lock_irqsave(&ppd->lflags_lock, flags);
 190         if (ppd->state_wanted) {
 191                 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 192                 ret = -EBUSY;
 193                 goto bail;
 194         }
 195         ppd->state_wanted = state;
 196         spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 197         wait_event_interruptible_timeout(ppd->state_wait,
 198                                          (ppd->lflags & state),
 199                                          msecs_to_jiffies(msecs));
 200         spin_lock_irqsave(&ppd->lflags_lock, flags);
 201         ppd->state_wanted = 0;
 202         spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 203
 204         if (!(ppd->lflags & state))
 205                 ret = -ETIMEDOUT;
 206         else
 207                 ret = 0;
 208 bail:
 209         return ret;
 210 }
 211
 212 int qib_set_linkstate(struct qib_pportdata *ppd, u8 newstate)
 213 {
 214         u32 lstate;
 215         int ret;
 216         struct qib_devdata *dd = ppd->dd;
 217         unsigned long flags;
 218
 219         switch (newstate) {
 220         case QIB_IB_LINKDOWN_ONLY:
 221                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 222                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_NOP);
 223                 /* don't wait */
 224                 ret = 0;
 225                 goto bail;
 226
 227         case QIB_IB_LINKDOWN:
 228                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 229                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_POLL);
 230                 /* don't wait */
 231                 ret = 0;
 232                 goto bail;
 233
 234         case QIB_IB_LINKDOWN_SLEEP:
 235                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 236                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_SLEEP);
 237                 /* don't wait */
 238                 ret = 0;
 239                 goto bail;
 240
 241         case QIB_IB_LINKDOWN_DISABLE:
 242                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 243                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_DISABLE);
 244                 /* don't wait */
 245                 ret = 0;
 246                 goto bail;
 247
 248         case QIB_IB_LINKARM:
 249                 if (ppd->lflags & QIBL_LINKARMED) {
 250                         ret = 0;
 251                         goto bail;
 252                 }
 253                 if (!(ppd->lflags & (QIBL_LINKINIT | QIBL_LINKACTIVE))) {
 254                         ret = -EINVAL;
 255                         goto bail;
 256                 }
 257                 /*
 258                  * Since the port can be ACTIVE when we ask for ARMED,
 259                  * clear QIBL_LINKV so we can wait for a transition.
 260                  * If the link isn't ARMED, then something else happened
 261                  * and there is no point waiting for ARMED.
 262                  */
 263                 spin_lock_irqsave(&ppd->lflags_lock, flags);
 264                 ppd->lflags &= ~QIBL_LINKV;
 265                 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 266                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 267                                  IB_LINKCMD_ARMED | IB_LINKINITCMD_NOP);
 268                 lstate = QIBL_LINKV;
 269                 break;
 270
 271         case QIB_IB_LINKACTIVE:
 272                 if (ppd->lflags & QIBL_LINKACTIVE) {
 273                         ret = 0;
 274                         goto bail;
 275                 }
 276                 if (!(ppd->lflags & QIBL_LINKARMED)) {
 277                         ret = -EINVAL;
 278                         goto bail;
 279                 }
 280                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 281                                  IB_LINKCMD_ACTIVE | IB_LINKINITCMD_NOP);
 282                 lstate = QIBL_LINKACTIVE;
 283                 break;
 284
 285         default:
 286                 ret = -EINVAL;
 287                 goto bail;
 288         }
 289         ret = qib_wait_linkstate(ppd, lstate, 10);
 290
 291 bail:
 292         return ret;
 293 }
 294
 295 /*
 296  * Get address of eager buffer from it's index (allocated in chunks, not
 297  * contiguous).
 298  */
 299 static inline void *qib_get_egrbuf(const struct qib_ctxtdata *rcd, u32 etail)
 300 {
 301         const u32 chunk = etail >> rcd->rcvegrbufs_perchunk_shift;
 302         const u32 idx =  etail & ((u32)rcd->rcvegrbufs_perchunk - 1);
 303
 304         return rcd->rcvegrbuf[chunk] + (idx << rcd->dd->rcvegrbufsize_shift);
 305 }
 306
 307 /*
 308  * Returns 1 if error was a CRC, else 0.
 309  * Needed for some chip's synthesized error counters.
 310  */
 311 static u32 qib_rcv_hdrerr(struct qib_ctxtdata *rcd, struct qib_pportdata *ppd,
 312                           u32 ctxt, u32 eflags, u32 l, u32 etail,
 313                           __le32 *rhf_addr, struct qib_message_header *rhdr)
 314 {
 315         u32 ret = 0;
 316
 317         if (eflags & (QLOGIC_IB_RHF_H_ICRCERR | QLOGIC_IB_RHF_H_VCRCERR))
 318                 ret = 1;
 319         else if (eflags == QLOGIC_IB_RHF_H_TIDERR) {
 320                 /* For TIDERR and RC QPs premptively schedule a NAK */
 321                 struct ib_header *hdr = (struct ib_header *)rhdr;
 322                 struct ib_other_headers *ohdr = NULL;
 323                 struct qib_ibport *ibp = &ppd->ibport_data;
 324                 struct qib_devdata *dd = ppd->dd;
 325                 struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
 326                 struct rvt_qp *qp = NULL;
 327                 u32 tlen = qib_hdrget_length_in_bytes(rhf_addr);
 328                 u16 lid  = be16_to_cpu(hdr->lrh[1]);
 329                 int lnh = be16_to_cpu(hdr->lrh[0]) & 3;
 330                 u32 qp_num;
 331                 u32 opcode;
 332                 u32 psn;
 333                 int diff;
 334
 335                 /* Sanity check packet */
 336                 if (tlen < 24)
 337                         goto drop;
 338
 339                 if (lid < be16_to_cpu(IB_MULTICAST_LID_BASE)) {
 340                         lid &= ~((1 << ppd->lmc) - 1);
 341                         if (unlikely(lid != ppd->lid))
 342                                 goto drop;
 343                 }
 344
 345                 /* Check for GRH */
 346                 if (lnh == QIB_LRH_BTH)
 347                         ohdr = &hdr->u.oth;
 348                 else if (lnh == QIB_LRH_GRH) {
 349                         u32 vtf;
 350
 351                         ohdr = &hdr->u.l.oth;
 352                         if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
 353                                 goto drop;
 354                         vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
 355                         if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
 356                                 goto drop;
 357                 } else
 358                         goto drop;
 359
 360                 /* Get opcode and PSN from packet */
 361                 opcode = be32_to_cpu(ohdr->bth[0]);
 362                 opcode >>= 24;
 363                 psn = be32_to_cpu(ohdr->bth[2]);
 364
 365                 /* Get the destination QP number. */
 366                 qp_num = be32_to_cpu(ohdr->bth[1]) & RVT_QPN_MASK;
 367                 if (qp_num != QIB_MULTICAST_QPN) {
 368                         int ruc_res;
 369
 370                         rcu_read_lock();
 371                         qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
 372                         if (!qp) {
 373                                 rcu_read_unlock();
 374                                 goto drop;
 375                         }
 376
 377                         /*
 378                          * Handle only RC QPs - for other QP types drop error
 379                          * packet.
 380                          */
 381                         spin_lock(&qp->r_lock);
 382
 383                         /* Check for valid receive state. */
 384                         if (!(ib_rvt_state_ops[qp->state] &
 385                               RVT_PROCESS_RECV_OK)) {
 386                                 ibp->rvp.n_pkt_drops++;
 387                                 goto unlock;
 388                         }
 389
 390                         switch (qp->ibqp.qp_type) {
 391                         case IB_QPT_RC:
 392                                 ruc_res =
 393                                         qib_ruc_check_hdr(
 394                                                 ibp, hdr,
 395                                                 lnh == QIB_LRH_GRH,
 396                                                 qp,
 397                                                 be32_to_cpu(ohdr->bth[0]));
 398                                 if (ruc_res)
 399                                         goto unlock;
 400
 401                                 /* Only deal with RDMA Writes for now */
 402                                 if (opcode <
 403                                     IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) {
 404                                         diff = qib_cmp24(psn, qp->r_psn);
 405                                         if (!qp->r_nak_state && diff >= 0) {
 406                                                 ibp->rvp.n_rc_seqnak++;
 407                                                 qp->r_nak_state =
 408                                                         IB_NAK_PSN_ERROR;
 409                                                 /* Use the expected PSN. */
 410                                                 qp->r_ack_psn = qp->r_psn;
 411                                                 /*
 412                                                  * Wait to send the sequence
 413                                                  * NAK until all packets
 414                                                  * in the receive queue have
 415                                                  * been processed.
 416                                                  * Otherwise, we end up
 417                                                  * propagating congestion.
 418                                                  */
 419                                                 if (list_empty(&qp->rspwait)) {
 420                                                         qp->r_flags |=
 421                                                                 RVT_R_RSP_NAK;
 422                                                         rvt_get_qp(qp);
 423                                                         list_add_tail(
 424                                                          &qp->rspwait,
 425                                                          &rcd->qp_wait_list);
 426                                                 }
 427                                         } /* Out of sequence NAK */
 428                                 } /* QP Request NAKs */
 429                                 break;
 430                         case IB_QPT_SMI:
 431                         case IB_QPT_GSI:
 432                         case IB_QPT_UD:
 433                         case IB_QPT_UC:
 434                         default:
 435                                 /* For now don't handle any other QP types */
 436                                 break;
 437                         }
 438
 439 unlock:
 440                         spin_unlock(&qp->r_lock);
 441                         rcu_read_unlock();
 442                 } /* Unicast QP */
 443         } /* Valid packet with TIDErr */
 444
 445 drop:
 446         return ret;
 447 }
 448
 449 /*
 450  * qib_kreceive - receive a packet
 451  * @rcd: the qlogic_ib context
 452  * @llic: gets count of good packets needed to clear lli,
 453  *          (used with chips that need need to track crcs for lli)
 454  *
 455  * called from interrupt handler for errors or receive interrupt
 456  * Returns number of CRC error packets, needed by some chips for
 457  * local link integrity tracking.   crcs are adjusted down by following
 458  * good packets, if any, and count of good packets is also tracked.
 459  */
 460 u32 qib_kreceive(struct qib_ctxtdata *rcd, u32 *llic, u32 *npkts)
 461 {
 462         struct qib_devdata *dd = rcd->dd;
 463         struct qib_pportdata *ppd = rcd->ppd;
 464         __le32 *rhf_addr;
 465         void *ebuf;
 466         const u32 rsize = dd->rcvhdrentsize;        /* words */
 467         const u32 maxcnt = dd->rcvhdrcnt * rsize;   /* words */
 468         u32 etail = -1, l, hdrqtail;
 469         struct qib_message_header *hdr;
 470         u32 eflags, etype, tlen, i = 0, updegr = 0, crcs = 0;
 471         int last;
 472         u64 lval;
 473         struct rvt_qp *qp, *nqp;
 474
 475         l = rcd->head;
 476         rhf_addr = (__le32 *) rcd->rcvhdrq + l + dd->rhf_offset;
 477         if (dd->flags & QIB_NODMA_RTAIL) {
 478                 u32 seq = qib_hdrget_seq(rhf_addr);
 479
 480                 if (seq != rcd->seq_cnt)
 481                         goto bail;
 482                 hdrqtail = 0;
 483         } else {
 484                 hdrqtail = qib_get_rcvhdrtail(rcd);
 485                 if (l == hdrqtail)
 486                         goto bail;
 487                 smp_rmb();  /* prevent speculative reads of dma'ed hdrq */
 488         }
 489
 490         for (last = 0, i = 1; !last; i += !last) {
 491                 hdr = dd->f_get_msgheader(dd, rhf_addr);
 492                 eflags = qib_hdrget_err_flags(rhf_addr);
 493                 etype = qib_hdrget_rcv_type(rhf_addr);
 494                 /* total length */
 495                 tlen = qib_hdrget_length_in_bytes(rhf_addr);
 496                 ebuf = NULL;
 497                 if ((dd->flags & QIB_NODMA_RTAIL) ?
 498                     qib_hdrget_use_egr_buf(rhf_addr) :
 499                     (etype != RCVHQ_RCV_TYPE_EXPECTED)) {
 500                         etail = qib_hdrget_index(rhf_addr);
 501                         updegr = 1;
 502                         if (tlen > sizeof(*hdr) ||
 503                             etype >= RCVHQ_RCV_TYPE_NON_KD) {
 504                                 ebuf = qib_get_egrbuf(rcd, etail);
 505                                 prefetch_range(ebuf, tlen - sizeof(*hdr));
 506                         }
 507                 }
 508                 if (!eflags) {
 509                         u16 lrh_len = be16_to_cpu(hdr->lrh[2]) << 2;
 510
 511                         if (lrh_len != tlen) {
 512                                 qib_stats.sps_lenerrs++;
 513                                 goto move_along;
 514                         }
 515                 }
 516                 if (etype == RCVHQ_RCV_TYPE_NON_KD && !eflags &&
 517                     ebuf == NULL &&
 518                     tlen > (dd->rcvhdrentsize - 2 + 1 -
 519                                 qib_hdrget_offset(rhf_addr)) << 2) {
 520                         goto move_along;
 521                 }
 522
 523                 /*
 524                  * Both tiderr and qibhdrerr are set for all plain IB
 525                  * packets; only qibhdrerr should be set.
 526                  */
 527                 if (unlikely(eflags))
 528                         crcs += qib_rcv_hdrerr(rcd, ppd, rcd->ctxt, eflags, l,
 529                                                etail, rhf_addr, hdr);
 530                 else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
 531                         qib_ib_rcv(rcd, hdr, ebuf, tlen);
 532                         if (crcs)
 533                                 crcs--;
 534                         else if (llic && *llic)
 535                                 --*llic;
 536                 }
 537 move_along:
 538                 l += rsize;
 539                 if (l >= maxcnt)
 540                         l = 0;
 541                 if (i == QIB_MAX_PKT_RECV)
 542                         last = 1;
 543
 544                 rhf_addr = (__le32 *) rcd->rcvhdrq + l + dd->rhf_offset;
 545                 if (dd->flags & QIB_NODMA_RTAIL) {
 546                         u32 seq = qib_hdrget_seq(rhf_addr);
 547
 548                         if (++rcd->seq_cnt > 13)
 549                                 rcd->seq_cnt = 1;
 550                         if (seq != rcd->seq_cnt)
 551                                 last = 1;
 552                 } else if (l == hdrqtail)
 553                         last = 1;
 554                 /*
 555                  * Update head regs etc., every 16 packets, if not last pkt,
 556                  * to help prevent rcvhdrq overflows, when many packets
 557                  * are processed and queue is nearly full.
 558                  * Don't request an interrupt for intermediate updates.
 559                  */
 560                 lval = l;
 561                 if (!last && !(i & 0xf)) {
 562                         dd->f_update_usrhead(rcd, lval, updegr, etail, i);
 563                         updegr = 0;
 564                 }
 565         }
 566
 567         rcd->head = l;
 568
 569         /*
 570          * Iterate over all QPs waiting to respond.
 571          * The list won't change since the IRQ is only run on one CPU.
 572          */
 573         list_for_each_entry_safe(qp, nqp, &rcd->qp_wait_list, rspwait) {
 574                 list_del_init(&qp->rspwait);
 575                 if (qp->r_flags & RVT_R_RSP_NAK) {
 576                         qp->r_flags &= ~RVT_R_RSP_NAK;
 577                         qib_send_rc_ack(qp);
 578                 }
 579                 if (qp->r_flags & RVT_R_RSP_SEND) {
 580                         unsigned long flags;
 581
 582                         qp->r_flags &= ~RVT_R_RSP_SEND;
 583                         spin_lock_irqsave(&qp->s_lock, flags);
 584                         if (ib_rvt_state_ops[qp->state] &
 585                                         RVT_PROCESS_OR_FLUSH_SEND)
 586                                 qib_schedule_send(qp);
 587                         spin_unlock_irqrestore(&qp->s_lock, flags);
 588                 }
 589                 rvt_put_qp(qp);
 590         }
 591
 592 bail:
 593         /* Report number of packets consumed */
 594         if (npkts)
 595                 *npkts = i;
 596
 597         /*
 598          * Always write head at end, and setup rcv interrupt, even
 599          * if no packets were processed.
 600          */
 601         lval = (u64)rcd->head | dd->rhdrhead_intr_off;
 602         dd->f_update_usrhead(rcd, lval, updegr, etail, i);
 603         return crcs;
 604 }
 605
 606 /**
 607  * qib_set_mtu - set the MTU
 608  * @ppd: the perport data
 609  * @arg: the new MTU
 610  *
 611  * We can handle "any" incoming size, the issue here is whether we
 612  * need to restrict our outgoing size.   For now, we don't do any
 613  * sanity checking on this, and we don't deal with what happens to
 614  * programs that are already running when the size changes.
 615  * NOTE: changing the MTU will usually cause the IBC to go back to
 616  * link INIT state...
 617  */
 618 int qib_set_mtu(struct qib_pportdata *ppd, u16 arg)
 619 {
 620         u32 piosize;
 621         int ret, chk;
 622
 623         if (arg != 256 && arg != 512 && arg != 1024 && arg != 2048 &&
 624             arg != 4096) {
 625                 ret = -EINVAL;
 626                 goto bail;
 627         }
 628         chk = ib_mtu_enum_to_int(qib_ibmtu);
 629         if (chk > 0 && arg > chk) {
 630                 ret = -EINVAL;
 631                 goto bail;
 632         }
 633
 634         piosize = ppd->ibmaxlen;
 635         ppd->ibmtu = arg;
 636
 637         if (arg >= (piosize - QIB_PIO_MAXIBHDR)) {
 638                 /* Only if it's not the initial value (or reset to it) */
 639                 if (piosize != ppd->init_ibmaxlen) {
 640                         if (arg > piosize && arg <= ppd->init_ibmaxlen)
 641                                 piosize = ppd->init_ibmaxlen - 2 * sizeof(u32);
 642                         ppd->ibmaxlen = piosize;
 643                 }
 644         } else if ((arg + QIB_PIO_MAXIBHDR) != ppd->ibmaxlen) {
 645                 piosize = arg + QIB_PIO_MAXIBHDR - 2 * sizeof(u32);
 646                 ppd->ibmaxlen = piosize;
 647         }
 648
 649         ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_MTU, 0);
 650
 651         ret = 0;
 652
 653 bail:
 654         return ret;
 655 }
 656
 657 int qib_set_lid(struct qib_pportdata *ppd, u32 lid, u8 lmc)
 658 {
 659         struct qib_devdata *dd = ppd->dd;
 660
 661         ppd->lid = lid;
 662         ppd->lmc = lmc;
 663
 664         dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LIDLMC,
 665                          lid | (~((1U << lmc) - 1)) << 16);
 666
 667         qib_devinfo(dd->pcidev, "IB%u:%u got a lid: 0x%x\n",
 668                     dd->unit, ppd->port, lid);
 669
 670         return 0;
 671 }
 672
 673 /*
 674  * Following deal with the "obviously simple" task of overriding the state
 675  * of the LEDS, which normally indicate link physical and logical status.
 676  * The complications arise in dealing with different hardware mappings
 677  * and the board-dependent routine being called from interrupts.
 678  * and then there's the requirement to _flash_ them.
 679  */
 680 #define LED_OVER_FREQ_SHIFT 8
 681 #define LED_OVER_FREQ_MASK (0xFF<<LED_OVER_FREQ_SHIFT)
 682 /* Below is "non-zero" to force override, but both actual LEDs are off */
 683 #define LED_OVER_BOTH_OFF (8)
 684
 685 static void qib_run_led_override(unsigned long opaque)
 686 {
 687         struct qib_pportdata *ppd = (struct qib_pportdata *)opaque;
 688         struct qib_devdata *dd = ppd->dd;
 689         int timeoff;
 690         int ph_idx;
 691
 692         if (!(dd->flags & QIB_INITTED))
 693                 return;
 694
 695         ph_idx = ppd->led_override_phase++ & 1;
 696         ppd->led_override = ppd->led_override_vals[ph_idx];
 697         timeoff = ppd->led_override_timeoff;
 698
 699         dd->f_setextled(ppd, 1);
 700         /*
 701          * don't re-fire the timer if user asked for it to be off; we let
 702          * it fire one more time after they turn it off to simplify
 703          */
 704         if (ppd->led_override_vals[0] || ppd->led_override_vals[1])
 705                 mod_timer(&ppd->led_override_timer, jiffies + timeoff);
 706 }
 707
 708 void qib_set_led_override(struct qib_pportdata *ppd, unsigned int val)
 709 {
 710         struct qib_devdata *dd = ppd->dd;
 711         int timeoff, freq;
 712
 713         if (!(dd->flags & QIB_INITTED))
 714                 return;
 715
 716         /* First check if we are blinking. If not, use 1HZ polling */
 717         timeoff = HZ;
 718         freq = (val & LED_OVER_FREQ_MASK) >> LED_OVER_FREQ_SHIFT;
 719
 720         if (freq) {
 721                 /* For blink, set each phase from one nybble of val */
 722                 ppd->led_override_vals[0] = val & 0xF;
 723                 ppd->led_override_vals[1] = (val >> 4) & 0xF;
 724                 timeoff = (HZ << 4)/freq;
 725         } else {
 726                 /* Non-blink set both phases the same. */
 727                 ppd->led_override_vals[0] = val & 0xF;
 728                 ppd->led_override_vals[1] = val & 0xF;
 729         }
 730         ppd->led_override_timeoff = timeoff;
 731
 732         /*
 733          * If the timer has not already been started, do so. Use a "quick"
 734          * timeout so the function will be called soon, to look at our request.
 735          */
 736         if (atomic_inc_return(&ppd->led_override_timer_active) == 1) {
 737                 /* Need to start timer */
 738                 init_timer(&ppd->led_override_timer);
 739                 ppd->led_override_timer.function = qib_run_led_override;
 740                 ppd->led_override_timer.data = (unsigned long) ppd;
 741                 ppd->led_override_timer.expires = jiffies + 1;
 742                 add_timer(&ppd->led_override_timer);
 743         } else {
 744                 if (ppd->led_override_vals[0] || ppd->led_override_vals[1])
 745                         mod_timer(&ppd->led_override_timer, jiffies + 1);
 746                 atomic_dec(&ppd->led_override_timer_active);
 747         }
 748 }
 749
 750 /**
 751  * qib_reset_device - reset the chip if possible
 752  * @unit: the device to reset
 753  *
 754  * Whether or not reset is successful, we attempt to re-initialize the chip
 755  * (that is, much like a driver unload/reload).  We clear the INITTED flag
 756  * so that the various entry points will fail until we reinitialize.  For
 757  * now, we only allow this if no user contexts are open that use chip resources
 758  */
 759 int qib_reset_device(int unit)
 760 {
 761         int ret, i;
 762         struct qib_devdata *dd = qib_lookup(unit);
 763         struct qib_pportdata *ppd;
 764         unsigned long flags;
 765         int pidx;
 766
 767         if (!dd) {
 768                 ret = -ENODEV;
 769                 goto bail;
 770         }
 771
 772         qib_devinfo(dd->pcidev, "Reset on unit %u requested\n", unit);
 773
 774         if (!dd->kregbase || !(dd->flags & QIB_PRESENT)) {
 775                 qib_devinfo(dd->pcidev,
 776                         "Invalid unit number %u or not initialized or not present\n",
 777                         unit);
 778                 ret = -ENXIO;
 779                 goto bail;
 780         }
 781
 782         spin_lock_irqsave(&dd->uctxt_lock, flags);
 783         if (dd->rcd)
 784                 for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
 785                         if (!dd->rcd[i] || !dd->rcd[i]->cnt)
 786                                 continue;
 787                         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 788                         ret = -EBUSY;
 789                         goto bail;
 790                 }
 791         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 792
 793         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 794                 ppd = dd->pport + pidx;
 795                 if (atomic_read(&ppd->led_override_timer_active)) {
 796                         /* Need to stop LED timer, _then_ shut off LEDs */
 797                         del_timer_sync(&ppd->led_override_timer);
 798                         atomic_set(&ppd->led_override_timer_active, 0);
 799                 }
 800
 801                 /* Shut off LEDs after we are sure timer is not running */
 802                 ppd->led_override = LED_OVER_BOTH_OFF;
 803                 dd->f_setextled(ppd, 0);
 804                 if (dd->flags & QIB_HAS_SEND_DMA)
 805                         qib_teardown_sdma(ppd);
 806         }
 807
 808         ret = dd->f_reset(dd);
 809         if (ret == 1)
 810                 ret = qib_init(dd, 1);
 811         else
 812                 ret = -EAGAIN;
 813         if (ret)
 814                 qib_dev_err(dd,
 815                         "Reinitialize unit %u after reset failed with %d\n",
 816                         unit, ret);
 817         else
 818                 qib_devinfo(dd->pcidev,
 819                         "Reinitialized unit %u after resetting\n",
 820                         unit);
 821
 822 bail:
 823         return ret;
 824 }