drivers/infiniband/hw/qib/qib_driver.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008, 2009 QLogic Corporation. All rights reserved.
   3  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/spinlock.h>
  35 #include <linux/pci.h>
  36 #include <linux/io.h>
  37 #include <linux/delay.h>
  38 #include <linux/netdevice.h>
  39 #include <linux/vmalloc.h>
  40 #include <linux/module.h>
  41
  42 #include "qib.h"
  43
  44 /*
  45  * The size has to be longer than this string, so we can append
  46  * board/chip information to it in the init code.
  47  */
  48 const char ib_qib_version[] = QIB_IDSTR "\n";
  49
  50 DEFINE_SPINLOCK(qib_devs_lock);
  51 LIST_HEAD(qib_dev_list);
  52 DEFINE_MUTEX(qib_mutex);        /* general driver use */
  53
  54 unsigned qib_ibmtu;
  55 module_param_named(ibmtu, qib_ibmtu, uint, S_IRUGO);
  56 MODULE_PARM_DESC(ibmtu, "Set max IB MTU (0=2KB, 1=256, 2=512, ... 5=4096");
  57
  58 unsigned qib_compat_ddr_negotiate = 1;
  59 module_param_named(compat_ddr_negotiate, qib_compat_ddr_negotiate, uint,
  60                    S_IWUSR | S_IRUGO);
  61 MODULE_PARM_DESC(compat_ddr_negotiate,
  62                  "Attempt pre-IBTA 1.2 DDR speed negotiation");
  63
  64 MODULE_LICENSE("Dual BSD/GPL");
  65 MODULE_AUTHOR("QLogic <support@qlogic.com>");
  66 MODULE_DESCRIPTION("QLogic IB driver");
  67
  68 /*
  69  * QIB_PIO_MAXIBHDR is the max IB header size allowed for in our
  70  * PIO send buffers.  This is well beyond anything currently
  71  * defined in the InfiniBand spec.
  72  */
  73 #define QIB_PIO_MAXIBHDR 128
  74
  75 /*
  76  * QIB_MAX_PKT_RCV is the max # if packets processed per receive interrupt.
  77  */
  78 #define QIB_MAX_PKT_RECV 64
  79
  80 struct qlogic_ib_stats qib_stats;
  81
  82 const char *qib_get_unit_name(int unit)
  83 {
  84         static char iname[16];
  85
  86         snprintf(iname, sizeof iname, "infinipath%u", unit);
  87         return iname;
  88 }
  89
  90 /*
  91  * Return count of units with at least one port ACTIVE.
  92  */
  93 int qib_count_active_units(void)
  94 {
  95         struct qib_devdata *dd;
  96         struct qib_pportdata *ppd;
  97         unsigned long flags;
  98         int pidx, nunits_active = 0;
  99
 100         spin_lock_irqsave(&qib_devs_lock, flags);
 101         list_for_each_entry(dd, &qib_dev_list, list) {
 102                 if (!(dd->flags & QIB_PRESENT) || !dd->kregbase)
 103                         continue;
 104                 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 105                         ppd = dd->pport + pidx;
 106                         if (ppd->lid && (ppd->lflags & (QIBL_LINKINIT |
 107                                          QIBL_LINKARMED | QIBL_LINKACTIVE))) {
 108                                 nunits_active++;
 109                                 break;
 110                         }
 111                 }
 112         }
 113         spin_unlock_irqrestore(&qib_devs_lock, flags);
 114         return nunits_active;
 115 }
 116
 117 /*
 118  * Return count of all units, optionally return in arguments
 119  * the number of usable (present) units, and the number of
 120  * ports that are up.
 121  */
 122 int qib_count_units(int *npresentp, int *nupp)
 123 {
 124         int nunits = 0, npresent = 0, nup = 0;
 125         struct qib_devdata *dd;
 126         unsigned long flags;
 127         int pidx;
 128         struct qib_pportdata *ppd;
 129
 130         spin_lock_irqsave(&qib_devs_lock, flags);
 131
 132         list_for_each_entry(dd, &qib_dev_list, list) {
 133                 nunits++;
 134                 if ((dd->flags & QIB_PRESENT) && dd->kregbase)
 135                         npresent++;
 136                 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 137                         ppd = dd->pport + pidx;
 138                         if (ppd->lid && (ppd->lflags & (QIBL_LINKINIT |
 139                                          QIBL_LINKARMED | QIBL_LINKACTIVE)))
 140                                 nup++;
 141                 }
 142         }
 143
 144         spin_unlock_irqrestore(&qib_devs_lock, flags);
 145
 146         if (npresentp)
 147                 *npresentp = npresent;
 148         if (nupp)
 149                 *nupp = nup;
 150
 151         return nunits;
 152 }
 153
 154 /**
 155  * qib_wait_linkstate - wait for an IB link state change to occur
 156  * @dd: the qlogic_ib device
 157  * @state: the state to wait for
 158  * @msecs: the number of milliseconds to wait
 159  *
 160  * wait up to msecs milliseconds for IB link state change to occur for
 161  * now, take the easy polling route.  Currently used only by
 162  * qib_set_linkstate.  Returns 0 if state reached, otherwise
 163  * -ETIMEDOUT state can have multiple states set, for any of several
 164  * transitions.
 165  */
 166 int qib_wait_linkstate(struct qib_pportdata *ppd, u32 state, int msecs)
 167 {
 168         int ret;
 169         unsigned long flags;
 170
 171         spin_lock_irqsave(&ppd->lflags_lock, flags);
 172         if (ppd->state_wanted) {
 173                 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 174                 ret = -EBUSY;
 175                 goto bail;
 176         }
 177         ppd->state_wanted = state;
 178         spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 179         wait_event_interruptible_timeout(ppd->state_wait,
 180                                          (ppd->lflags & state),
 181                                          msecs_to_jiffies(msecs));
 182         spin_lock_irqsave(&ppd->lflags_lock, flags);
 183         ppd->state_wanted = 0;
 184         spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 185
 186         if (!(ppd->lflags & state))
 187                 ret = -ETIMEDOUT;
 188         else
 189                 ret = 0;
 190 bail:
 191         return ret;
 192 }
 193
 194 int qib_set_linkstate(struct qib_pportdata *ppd, u8 newstate)
 195 {
 196         u32 lstate;
 197         int ret;
 198         struct qib_devdata *dd = ppd->dd;
 199         unsigned long flags;
 200
 201         switch (newstate) {
 202         case QIB_IB_LINKDOWN_ONLY:
 203                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 204                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_NOP);
 205                 /* don't wait */
 206                 ret = 0;
 207                 goto bail;
 208
 209         case QIB_IB_LINKDOWN:
 210                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 211                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_POLL);
 212                 /* don't wait */
 213                 ret = 0;
 214                 goto bail;
 215
 216         case QIB_IB_LINKDOWN_SLEEP:
 217                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 218                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_SLEEP);
 219                 /* don't wait */
 220                 ret = 0;
 221                 goto bail;
 222
 223         case QIB_IB_LINKDOWN_DISABLE:
 224                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 225                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_DISABLE);
 226                 /* don't wait */
 227                 ret = 0;
 228                 goto bail;
 229
 230         case QIB_IB_LINKARM:
 231                 if (ppd->lflags & QIBL_LINKARMED) {
 232                         ret = 0;
 233                         goto bail;
 234                 }
 235                 if (!(ppd->lflags & (QIBL_LINKINIT | QIBL_LINKACTIVE))) {
 236                         ret = -EINVAL;
 237                         goto bail;
 238                 }
 239                 /*
 240                  * Since the port can be ACTIVE when we ask for ARMED,
 241                  * clear QIBL_LINKV so we can wait for a transition.
 242                  * If the link isn't ARMED, then something else happened
 243                  * and there is no point waiting for ARMED.
 244                  */
 245                 spin_lock_irqsave(&ppd->lflags_lock, flags);
 246                 ppd->lflags &= ~QIBL_LINKV;
 247                 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 248                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 249                                  IB_LINKCMD_ARMED | IB_LINKINITCMD_NOP);
 250                 lstate = QIBL_LINKV;
 251                 break;
 252
 253         case QIB_IB_LINKACTIVE:
 254                 if (ppd->lflags & QIBL_LINKACTIVE) {
 255                         ret = 0;
 256                         goto bail;
 257                 }
 258                 if (!(ppd->lflags & QIBL_LINKARMED)) {
 259                         ret = -EINVAL;
 260                         goto bail;
 261                 }
 262                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 263                                  IB_LINKCMD_ACTIVE | IB_LINKINITCMD_NOP);
 264                 lstate = QIBL_LINKACTIVE;
 265                 break;
 266
 267         default:
 268                 ret = -EINVAL;
 269                 goto bail;
 270         }
 271         ret = qib_wait_linkstate(ppd, lstate, 10);
 272
 273 bail:
 274         return ret;
 275 }
 276
 277 /*
 278  * Get address of eager buffer from it's index (allocated in chunks, not
 279  * contiguous).
 280  */
 281 static inline void *qib_get_egrbuf(const struct qib_ctxtdata *rcd, u32 etail)
 282 {
 283         const u32 chunk = etail >> rcd->rcvegrbufs_perchunk_shift;
 284         const u32 idx =  etail & ((u32)rcd->rcvegrbufs_perchunk - 1);
 285
 286         return rcd->rcvegrbuf[chunk] + (idx << rcd->dd->rcvegrbufsize_shift);
 287 }
 288
 289 /*
 290  * Returns 1 if error was a CRC, else 0.
 291  * Needed for some chip's synthesized error counters.
 292  */
 293 static u32 qib_rcv_hdrerr(struct qib_ctxtdata *rcd, struct qib_pportdata *ppd,
 294                           u32 ctxt, u32 eflags, u32 l, u32 etail,
 295                           __le32 *rhf_addr, struct qib_message_header *rhdr)
 296 {
 297         u32 ret = 0;
 298
 299         if (eflags & (QLOGIC_IB_RHF_H_ICRCERR | QLOGIC_IB_RHF_H_VCRCERR))
 300                 ret = 1;
 301         else if (eflags == QLOGIC_IB_RHF_H_TIDERR) {
 302                 /* For TIDERR and RC QPs premptively schedule a NAK */
 303                 struct qib_ib_header *hdr = (struct qib_ib_header *) rhdr;
 304                 struct qib_other_headers *ohdr = NULL;
 305                 struct qib_ibport *ibp = &ppd->ibport_data;
 306                 struct qib_qp *qp = NULL;
 307                 u32 tlen = qib_hdrget_length_in_bytes(rhf_addr);
 308                 u16 lid  = be16_to_cpu(hdr->lrh[1]);
 309                 int lnh = be16_to_cpu(hdr->lrh[0]) & 3;
 310                 u32 qp_num;
 311                 u32 opcode;
 312                 u32 psn;
 313                 int diff;
 314
 315                 /* Sanity check packet */
 316                 if (tlen < 24)
 317                         goto drop;
 318
 319                 if (lid < QIB_MULTICAST_LID_BASE) {
 320                         lid &= ~((1 << ppd->lmc) - 1);
 321                         if (unlikely(lid != ppd->lid))
 322                                 goto drop;
 323                 }
 324
 325                 /* Check for GRH */
 326                 if (lnh == QIB_LRH_BTH)
 327                         ohdr = &hdr->u.oth;
 328                 else if (lnh == QIB_LRH_GRH) {
 329                         u32 vtf;
 330
 331                         ohdr = &hdr->u.l.oth;
 332                         if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
 333                                 goto drop;
 334                         vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
 335                         if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
 336                                 goto drop;
 337                 } else
 338                         goto drop;
 339
 340                 /* Get opcode and PSN from packet */
 341                 opcode = be32_to_cpu(ohdr->bth[0]);
 342                 opcode >>= 24;
 343                 psn = be32_to_cpu(ohdr->bth[2]);
 344
 345                 /* Get the destination QP number. */
 346                 qp_num = be32_to_cpu(ohdr->bth[1]) & QIB_QPN_MASK;
 347                 if (qp_num != QIB_MULTICAST_QPN) {
 348                         int ruc_res;
 349                         qp = qib_lookup_qpn(ibp, qp_num);
 350                         if (!qp)
 351                                 goto drop;
 352
 353                         /*
 354                          * Handle only RC QPs - for other QP types drop error
 355                          * packet.
 356                          */
 357                         spin_lock(&qp->r_lock);
 358
 359                         /* Check for valid receive state. */
 360                         if (!(ib_qib_state_ops[qp->state] &
 361                               QIB_PROCESS_RECV_OK)) {
 362                                 ibp->n_pkt_drops++;
 363                                 goto unlock;
 364                         }
 365
 366                         switch (qp->ibqp.qp_type) {
 367                         case IB_QPT_RC:
 368                                 ruc_res =
 369                                         qib_ruc_check_hdr(
 370                                                 ibp, hdr,
 371                                                 lnh == QIB_LRH_GRH,
 372                                                 qp,
 373                                                 be32_to_cpu(ohdr->bth[0]));
 374                                 if (ruc_res)
 375                                         goto unlock;
 376
 377                                 /* Only deal with RDMA Writes for now */
 378                                 if (opcode <
 379                                     IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) {
 380                                         diff = qib_cmp24(psn, qp->r_psn);
 381                                         if (!qp->r_nak_state && diff >= 0) {
 382                                                 ibp->n_rc_seqnak++;
 383                                                 qp->r_nak_state =
 384                                                         IB_NAK_PSN_ERROR;
 385                                                 /* Use the expected PSN. */
 386                                                 qp->r_ack_psn = qp->r_psn;
 387                                                 /*
 388                                                  * Wait to send the sequence
 389                                                  * NAK until all packets
 390                                                  * in the receive queue have
 391                                                  * been processed.
 392                                                  * Otherwise, we end up
 393                                                  * propagating congestion.
 394                                                  */
 395                                                 if (list_empty(&qp->rspwait)) {
 396                                                         qp->r_flags |=
 397                                                                 QIB_R_RSP_NAK;
 398                                                         atomic_inc(
 399                                                                 &qp->refcount);
 400                                                         list_add_tail(
 401                                                          &qp->rspwait,
 402                                                          &rcd->qp_wait_list);
 403                                                 }
 404                                         } /* Out of sequence NAK */
 405                                 } /* QP Request NAKs */
 406                                 break;
 407                         case IB_QPT_SMI:
 408                         case IB_QPT_GSI:
 409                         case IB_QPT_UD:
 410                         case IB_QPT_UC:
 411                         default:
 412                                 /* For now don't handle any other QP types */
 413                                 break;
 414                         }
 415
 416 unlock:
 417                         spin_unlock(&qp->r_lock);
 418                         /*
 419                          * Notify qib_destroy_qp() if it is waiting
 420                          * for us to finish.
 421                          */
 422                         if (atomic_dec_and_test(&qp->refcount))
 423                                 wake_up(&qp->wait);
 424                 } /* Unicast QP */
 425         } /* Valid packet with TIDErr */
 426
 427 drop:
 428         return ret;
 429 }
 430
 431 /*
 432  * qib_kreceive - receive a packet
 433  * @rcd: the qlogic_ib context
 434  * @llic: gets count of good packets needed to clear lli,
 435  *          (used with chips that need need to track crcs for lli)
 436  *
 437  * called from interrupt handler for errors or receive interrupt
 438  * Returns number of CRC error packets, needed by some chips for
 439  * local link integrity tracking.   crcs are adjusted down by following
 440  * good packets, if any, and count of good packets is also tracked.
 441  */
 442 u32 qib_kreceive(struct qib_ctxtdata *rcd, u32 *llic, u32 *npkts)
 443 {
 444         struct qib_devdata *dd = rcd->dd;
 445         struct qib_pportdata *ppd = rcd->ppd;
 446         __le32 *rhf_addr;
 447         void *ebuf;
 448         const u32 rsize = dd->rcvhdrentsize;        /* words */
 449         const u32 maxcnt = dd->rcvhdrcnt * rsize;   /* words */
 450         u32 etail = -1, l, hdrqtail;
 451         struct qib_message_header *hdr;
 452         u32 eflags, etype, tlen, i = 0, updegr = 0, crcs = 0;
 453         int last;
 454         u64 lval;
 455         struct qib_qp *qp, *nqp;
 456
 457         l = rcd->head;
 458         rhf_addr = (__le32 *) rcd->rcvhdrq + l + dd->rhf_offset;
 459         if (dd->flags & QIB_NODMA_RTAIL) {
 460                 u32 seq = qib_hdrget_seq(rhf_addr);
 461                 if (seq != rcd->seq_cnt)
 462                         goto bail;
 463                 hdrqtail = 0;
 464         } else {
 465                 hdrqtail = qib_get_rcvhdrtail(rcd);
 466                 if (l == hdrqtail)
 467                         goto bail;
 468                 smp_rmb();  /* prevent speculative reads of dma'ed hdrq */
 469         }
 470
 471         for (last = 0, i = 1; !last; i += !last) {
 472                 hdr = dd->f_get_msgheader(dd, rhf_addr);
 473                 eflags = qib_hdrget_err_flags(rhf_addr);
 474                 etype = qib_hdrget_rcv_type(rhf_addr);
 475                 /* total length */
 476                 tlen = qib_hdrget_length_in_bytes(rhf_addr);
 477                 ebuf = NULL;
 478                 if ((dd->flags & QIB_NODMA_RTAIL) ?
 479                     qib_hdrget_use_egr_buf(rhf_addr) :
 480                     (etype != RCVHQ_RCV_TYPE_EXPECTED)) {
 481                         etail = qib_hdrget_index(rhf_addr);
 482                         updegr = 1;
 483                         if (tlen > sizeof(*hdr) ||
 484                             etype >= RCVHQ_RCV_TYPE_NON_KD)
 485                                 ebuf = qib_get_egrbuf(rcd, etail);
 486                 }
 487                 if (!eflags) {
 488                         u16 lrh_len = be16_to_cpu(hdr->lrh[2]) << 2;
 489
 490                         if (lrh_len != tlen) {
 491                                 qib_stats.sps_lenerrs++;
 492                                 goto move_along;
 493                         }
 494                 }
 495                 if (etype == RCVHQ_RCV_TYPE_NON_KD && !eflags &&
 496                     ebuf == NULL &&
 497                     tlen > (dd->rcvhdrentsize - 2 + 1 -
 498                                 qib_hdrget_offset(rhf_addr)) << 2) {
 499                         goto move_along;
 500                 }
 501
 502                 /*
 503                  * Both tiderr and qibhdrerr are set for all plain IB
 504                  * packets; only qibhdrerr should be set.
 505                  */
 506                 if (unlikely(eflags))
 507                         crcs += qib_rcv_hdrerr(rcd, ppd, rcd->ctxt, eflags, l,
 508                                                etail, rhf_addr, hdr);
 509                 else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
 510                         qib_ib_rcv(rcd, hdr, ebuf, tlen);
 511                         if (crcs)
 512                                 crcs--;
 513                         else if (llic && *llic)
 514                                 --*llic;
 515                 }
 516 move_along:
 517                 l += rsize;
 518                 if (l >= maxcnt)
 519                         l = 0;
 520                 if (i == QIB_MAX_PKT_RECV)
 521                         last = 1;
 522
 523                 rhf_addr = (__le32 *) rcd->rcvhdrq + l + dd->rhf_offset;
 524                 if (dd->flags & QIB_NODMA_RTAIL) {
 525                         u32 seq = qib_hdrget_seq(rhf_addr);
 526
 527                         if (++rcd->seq_cnt > 13)
 528                                 rcd->seq_cnt = 1;
 529                         if (seq != rcd->seq_cnt)
 530                                 last = 1;
 531                 } else if (l == hdrqtail)
 532                         last = 1;
 533                 /*
 534                  * Update head regs etc., every 16 packets, if not last pkt,
 535                  * to help prevent rcvhdrq overflows, when many packets
 536                  * are processed and queue is nearly full.
 537                  * Don't request an interrupt for intermediate updates.
 538                  */
 539                 lval = l;
 540                 if (!last && !(i & 0xf)) {
 541                         dd->f_update_usrhead(rcd, lval, updegr, etail, i);
 542                         updegr = 0;
 543                 }
 544         }
 545         /*
 546          * Notify qib_destroy_qp() if it is waiting
 547          * for lookaside_qp to finish.
 548          */
 549         if (rcd->lookaside_qp) {
 550                 if (atomic_dec_and_test(&rcd->lookaside_qp->refcount))
 551                         wake_up(&rcd->lookaside_qp->wait);
 552                 rcd->lookaside_qp = NULL;
 553         }
 554
 555         rcd->head = l;
 556         rcd->pkt_count += i;
 557
 558         /*
 559          * Iterate over all QPs waiting to respond.
 560          * The list won't change since the IRQ is only run on one CPU.
 561          */
 562         list_for_each_entry_safe(qp, nqp, &rcd->qp_wait_list, rspwait) {
 563                 list_del_init(&qp->rspwait);
 564                 if (qp->r_flags & QIB_R_RSP_NAK) {
 565                         qp->r_flags &= ~QIB_R_RSP_NAK;
 566                         qib_send_rc_ack(qp);
 567                 }
 568                 if (qp->r_flags & QIB_R_RSP_SEND) {
 569                         unsigned long flags;
 570
 571                         qp->r_flags &= ~QIB_R_RSP_SEND;
 572                         spin_lock_irqsave(&qp->s_lock, flags);
 573                         if (ib_qib_state_ops[qp->state] &
 574                                         QIB_PROCESS_OR_FLUSH_SEND)
 575                                 qib_schedule_send(qp);
 576                         spin_unlock_irqrestore(&qp->s_lock, flags);
 577                 }
 578                 if (atomic_dec_and_test(&qp->refcount))
 579                         wake_up(&qp->wait);
 580         }
 581
 582 bail:
 583         /* Report number of packets consumed */
 584         if (npkts)
 585                 *npkts = i;
 586
 587         /*
 588          * Always write head at end, and setup rcv interrupt, even
 589          * if no packets were processed.
 590          */
 591         lval = (u64)rcd->head | dd->rhdrhead_intr_off;
 592         dd->f_update_usrhead(rcd, lval, updegr, etail, i);
 593         return crcs;
 594 }
 595
 596 /**
 597  * qib_set_mtu - set the MTU
 598  * @ppd: the perport data
 599  * @arg: the new MTU
 600  *
 601  * We can handle "any" incoming size, the issue here is whether we
 602  * need to restrict our outgoing size.   For now, we don't do any
 603  * sanity checking on this, and we don't deal with what happens to
 604  * programs that are already running when the size changes.
 605  * NOTE: changing the MTU will usually cause the IBC to go back to
 606  * link INIT state...
 607  */
 608 int qib_set_mtu(struct qib_pportdata *ppd, u16 arg)
 609 {
 610         u32 piosize;
 611         int ret, chk;
 612
 613         if (arg != 256 && arg != 512 && arg != 1024 && arg != 2048 &&
 614             arg != 4096) {
 615                 ret = -EINVAL;
 616                 goto bail;
 617         }
 618         chk = ib_mtu_enum_to_int(qib_ibmtu);
 619         if (chk > 0 && arg > chk) {
 620                 ret = -EINVAL;
 621                 goto bail;
 622         }
 623
 624         piosize = ppd->ibmaxlen;
 625         ppd->ibmtu = arg;
 626
 627         if (arg >= (piosize - QIB_PIO_MAXIBHDR)) {
 628                 /* Only if it's not the initial value (or reset to it) */
 629                 if (piosize != ppd->init_ibmaxlen) {
 630                         if (arg > piosize && arg <= ppd->init_ibmaxlen)
 631                                 piosize = ppd->init_ibmaxlen - 2 * sizeof(u32);
 632                         ppd->ibmaxlen = piosize;
 633                 }
 634         } else if ((arg + QIB_PIO_MAXIBHDR) != ppd->ibmaxlen) {
 635                 piosize = arg + QIB_PIO_MAXIBHDR - 2 * sizeof(u32);
 636                 ppd->ibmaxlen = piosize;
 637         }
 638
 639         ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_MTU, 0);
 640
 641         ret = 0;
 642
 643 bail:
 644         return ret;
 645 }
 646
 647 int qib_set_lid(struct qib_pportdata *ppd, u32 lid, u8 lmc)
 648 {
 649         struct qib_devdata *dd = ppd->dd;
 650         ppd->lid = lid;
 651         ppd->lmc = lmc;
 652
 653         dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LIDLMC,
 654                          lid | (~((1U << lmc) - 1)) << 16);
 655
 656         qib_devinfo(dd->pcidev, "IB%u:%u got a lid: 0x%x\n",
 657                     dd->unit, ppd->port, lid);
 658
 659         return 0;
 660 }
 661
 662 /*
 663  * Following deal with the "obviously simple" task of overriding the state
 664  * of the LEDS, which normally indicate link physical and logical status.
 665  * The complications arise in dealing with different hardware mappings
 666  * and the board-dependent routine being called from interrupts.
 667  * and then there's the requirement to _flash_ them.
 668  */
 669 #define LED_OVER_FREQ_SHIFT 8
 670 #define LED_OVER_FREQ_MASK (0xFF<<LED_OVER_FREQ_SHIFT)
 671 /* Below is "non-zero" to force override, but both actual LEDs are off */
 672 #define LED_OVER_BOTH_OFF (8)
 673
 674 static void qib_run_led_override(unsigned long opaque)
 675 {
 676         struct qib_pportdata *ppd = (struct qib_pportdata *)opaque;
 677         struct qib_devdata *dd = ppd->dd;
 678         int timeoff;
 679         int ph_idx;
 680
 681         if (!(dd->flags & QIB_INITTED))
 682                 return;
 683
 684         ph_idx = ppd->led_override_phase++ & 1;
 685         ppd->led_override = ppd->led_override_vals[ph_idx];
 686         timeoff = ppd->led_override_timeoff;
 687
 688         dd->f_setextled(ppd, 1);
 689         /*
 690          * don't re-fire the timer if user asked for it to be off; we let
 691          * it fire one more time after they turn it off to simplify
 692          */
 693         if (ppd->led_override_vals[0] || ppd->led_override_vals[1])
 694                 mod_timer(&ppd->led_override_timer, jiffies + timeoff);
 695 }
 696
 697 void qib_set_led_override(struct qib_pportdata *ppd, unsigned int val)
 698 {
 699         struct qib_devdata *dd = ppd->dd;
 700         int timeoff, freq;
 701
 702         if (!(dd->flags & QIB_INITTED))
 703                 return;
 704
 705         /* First check if we are blinking. If not, use 1HZ polling */
 706         timeoff = HZ;
 707         freq = (val & LED_OVER_FREQ_MASK) >> LED_OVER_FREQ_SHIFT;
 708
 709         if (freq) {
 710                 /* For blink, set each phase from one nybble of val */
 711                 ppd->led_override_vals[0] = val & 0xF;
 712                 ppd->led_override_vals[1] = (val >> 4) & 0xF;
 713                 timeoff = (HZ << 4)/freq;
 714         } else {
 715                 /* Non-blink set both phases the same. */
 716                 ppd->led_override_vals[0] = val & 0xF;
 717                 ppd->led_override_vals[1] = val & 0xF;
 718         }
 719         ppd->led_override_timeoff = timeoff;
 720
 721         /*
 722          * If the timer has not already been started, do so. Use a "quick"
 723          * timeout so the function will be called soon, to look at our request.
 724          */
 725         if (atomic_inc_return(&ppd->led_override_timer_active) == 1) {
 726                 /* Need to start timer */
 727                 init_timer(&ppd->led_override_timer);
 728                 ppd->led_override_timer.function = qib_run_led_override;
 729                 ppd->led_override_timer.data = (unsigned long) ppd;
 730                 ppd->led_override_timer.expires = jiffies + 1;
 731                 add_timer(&ppd->led_override_timer);
 732         } else {
 733                 if (ppd->led_override_vals[0] || ppd->led_override_vals[1])
 734                         mod_timer(&ppd->led_override_timer, jiffies + 1);
 735                 atomic_dec(&ppd->led_override_timer_active);
 736         }
 737 }
 738
 739 /**
 740  * qib_reset_device - reset the chip if possible
 741  * @unit: the device to reset
 742  *
 743  * Whether or not reset is successful, we attempt to re-initialize the chip
 744  * (that is, much like a driver unload/reload).  We clear the INITTED flag
 745  * so that the various entry points will fail until we reinitialize.  For
 746  * now, we only allow this if no user contexts are open that use chip resources
 747  */
 748 int qib_reset_device(int unit)
 749 {
 750         int ret, i;
 751         struct qib_devdata *dd = qib_lookup(unit);
 752         struct qib_pportdata *ppd;
 753         unsigned long flags;
 754         int pidx;
 755
 756         if (!dd) {
 757                 ret = -ENODEV;
 758                 goto bail;
 759         }
 760
 761         qib_devinfo(dd->pcidev, "Reset on unit %u requested\n", unit);
 762
 763         if (!dd->kregbase || !(dd->flags & QIB_PRESENT)) {
 764                 qib_devinfo(dd->pcidev, "Invalid unit number %u or "
 765                             "not initialized or not present\n", unit);
 766                 ret = -ENXIO;
 767                 goto bail;
 768         }
 769
 770         spin_lock_irqsave(&dd->uctxt_lock, flags);
 771         if (dd->rcd)
 772                 for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
 773                         if (!dd->rcd[i] || !dd->rcd[i]->cnt)
 774                                 continue;
 775                         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 776                         ret = -EBUSY;
 777                         goto bail;
 778                 }
 779         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 780
 781         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 782                 ppd = dd->pport + pidx;
 783                 if (atomic_read(&ppd->led_override_timer_active)) {
 784                         /* Need to stop LED timer, _then_ shut off LEDs */
 785                         del_timer_sync(&ppd->led_override_timer);
 786                         atomic_set(&ppd->led_override_timer_active, 0);
 787                 }
 788
 789                 /* Shut off LEDs after we are sure timer is not running */
 790                 ppd->led_override = LED_OVER_BOTH_OFF;
 791                 dd->f_setextled(ppd, 0);
 792                 if (dd->flags & QIB_HAS_SEND_DMA)
 793                         qib_teardown_sdma(ppd);
 794         }
 795
 796         ret = dd->f_reset(dd);
 797         if (ret == 1)
 798                 ret = qib_init(dd, 1);
 799         else
 800                 ret = -EAGAIN;
 801         if (ret)
 802                 qib_dev_err(dd, "Reinitialize unit %u after "
 803                             "reset failed with %d\n", unit, ret);
 804         else
 805                 qib_devinfo(dd->pcidev, "Reinitialized unit %u after "
 806                             "resetting\n", unit);
 807
 808 bail:
 809         return ret;
 810 }