drivers/infiniband/hw/ipath/ipath_intr.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008 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/pci.h>
  35 #include <linux/delay.h>
  36
  37 #include "ipath_kernel.h"
  38 #include "ipath_verbs.h"
  39 #include "ipath_common.h"
  40
  41 /*
  42  * clear (write) a pio buffer, to clear a parity error.   This routine
  43  * should only be called when in freeze mode, and the buffer should be
  44  * canceled afterwards.
  45  */
  46 static void ipath_clrpiobuf(struct ipath_devdata *dd, u32 pnum)
  47 {
  48         u32 __iomem *pbuf;
  49         u32 dwcnt; /* dword count to write */
  50         if (pnum < dd->ipath_piobcnt2k) {
  51                 pbuf = (u32 __iomem *) (dd->ipath_pio2kbase + pnum *
  52                         dd->ipath_palign);
  53                 dwcnt = dd->ipath_piosize2k >> 2;
  54         }
  55         else {
  56                 pbuf = (u32 __iomem *) (dd->ipath_pio4kbase +
  57                         (pnum - dd->ipath_piobcnt2k) * dd->ipath_4kalign);
  58                 dwcnt = dd->ipath_piosize4k >> 2;
  59         }
  60         dev_info(&dd->pcidev->dev,
  61                 "Rewrite PIO buffer %u, to recover from parity error\n",
  62                 pnum);
  63
  64         /* no flush required, since already in freeze */
  65         writel(dwcnt + 1, pbuf);
  66         while (--dwcnt)
  67                 writel(0, pbuf++);
  68 }
  69
  70 /*
  71  * Called when we might have an error that is specific to a particular
  72  * PIO buffer, and may need to cancel that buffer, so it can be re-used.
  73  * If rewrite is true, and bits are set in the sendbufferror registers,
  74  * we'll write to the buffer, for error recovery on parity errors.
  75  */
  76 void ipath_disarm_senderrbufs(struct ipath_devdata *dd, int rewrite)
  77 {
  78         u32 piobcnt;
  79         unsigned long sbuf[4];
  80         /*
  81          * it's possible that sendbuffererror could have bits set; might
  82          * have already done this as a result of hardware error handling
  83          */
  84         piobcnt = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
  85         /* read these before writing errorclear */
  86         sbuf[0] = ipath_read_kreg64(
  87                 dd, dd->ipath_kregs->kr_sendbuffererror);
  88         sbuf[1] = ipath_read_kreg64(
  89                 dd, dd->ipath_kregs->kr_sendbuffererror + 1);
  90         if (piobcnt > 128)
  91                 sbuf[2] = ipath_read_kreg64(
  92                         dd, dd->ipath_kregs->kr_sendbuffererror + 2);
  93         if (piobcnt > 192)
  94                 sbuf[3] = ipath_read_kreg64(
  95                         dd, dd->ipath_kregs->kr_sendbuffererror + 3);
  96         else
  97                 sbuf[3] = 0;
  98
  99         if (sbuf[0] || sbuf[1] || (piobcnt > 128 && (sbuf[2] || sbuf[3]))) {
 100                 int i;
 101                 if (ipath_debug & (__IPATH_PKTDBG|__IPATH_DBG) &&
 102                         dd->ipath_lastcancel > jiffies) {
 103                         __IPATH_DBG_WHICH(__IPATH_PKTDBG|__IPATH_DBG,
 104                                           "SendbufErrs %lx %lx", sbuf[0],
 105                                           sbuf[1]);
 106                         if (ipath_debug & __IPATH_PKTDBG && piobcnt > 128)
 107                                 printk(" %lx %lx ", sbuf[2], sbuf[3]);
 108                         printk("\n");
 109                 }
 110
 111                 for (i = 0; i < piobcnt; i++)
 112                         if (test_bit(i, sbuf)) {
 113                                 if (rewrite)
 114                                         ipath_clrpiobuf(dd, i);
 115                                 ipath_disarm_piobufs(dd, i, 1);
 116                         }
 117                 /* ignore armlaunch errs for a bit */
 118                 dd->ipath_lastcancel = jiffies+3;
 119         }
 120 }
 121
 122
 123 /* These are all rcv-related errors which we want to count for stats */
 124 #define E_SUM_PKTERRS \
 125         (INFINIPATH_E_RHDRLEN | INFINIPATH_E_RBADTID | \
 126          INFINIPATH_E_RBADVERSION | INFINIPATH_E_RHDR | \
 127          INFINIPATH_E_RLONGPKTLEN | INFINIPATH_E_RSHORTPKTLEN | \
 128          INFINIPATH_E_RMAXPKTLEN | INFINIPATH_E_RMINPKTLEN | \
 129          INFINIPATH_E_RFORMATERR | INFINIPATH_E_RUNSUPVL | \
 130          INFINIPATH_E_RUNEXPCHAR | INFINIPATH_E_REBP)
 131
 132 /* These are all send-related errors which we want to count for stats */
 133 #define E_SUM_ERRS \
 134         (INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SUNEXPERRPKTNUM | \
 135          INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
 136          INFINIPATH_E_SMAXPKTLEN | INFINIPATH_E_SUNSUPVL | \
 137          INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
 138          INFINIPATH_E_INVALIDADDR)
 139
 140 /*
 141  * this is similar to E_SUM_ERRS, but can't ignore armlaunch, don't ignore
 142  * errors not related to freeze and cancelling buffers.  Can't ignore
 143  * armlaunch because could get more while still cleaning up, and need
 144  * to cancel those as they happen.
 145  */
 146 #define E_SPKT_ERRS_IGNORE \
 147          (INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
 148          INFINIPATH_E_SMAXPKTLEN | INFINIPATH_E_SMINPKTLEN | \
 149          INFINIPATH_E_SPKTLEN)
 150
 151 /*
 152  * these are errors that can occur when the link changes state while
 153  * a packet is being sent or received.  This doesn't cover things
 154  * like EBP or VCRC that can be the result of a sending having the
 155  * link change state, so we receive a "known bad" packet.
 156  */
 157 #define E_SUM_LINK_PKTERRS \
 158         (INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
 159          INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
 160          INFINIPATH_E_RSHORTPKTLEN | INFINIPATH_E_RMINPKTLEN | \
 161          INFINIPATH_E_RUNEXPCHAR)
 162
 163 static u64 handle_e_sum_errs(struct ipath_devdata *dd, ipath_err_t errs)
 164 {
 165         u64 ignore_this_time = 0;
 166
 167         ipath_disarm_senderrbufs(dd, 0);
 168         if ((errs & E_SUM_LINK_PKTERRS) &&
 169             !(dd->ipath_flags & IPATH_LINKACTIVE)) {
 170                 /*
 171                  * This can happen when SMA is trying to bring the link
 172                  * up, but the IB link changes state at the "wrong" time.
 173                  * The IB logic then complains that the packet isn't
 174                  * valid.  We don't want to confuse people, so we just
 175                  * don't print them, except at debug
 176                  */
 177                 ipath_dbg("Ignoring packet errors %llx, because link not "
 178                           "ACTIVE\n", (unsigned long long) errs);
 179                 ignore_this_time = errs & E_SUM_LINK_PKTERRS;
 180         }
 181
 182         return ignore_this_time;
 183 }
 184
 185 /* generic hw error messages... */
 186 #define INFINIPATH_HWE_TXEMEMPARITYERR_MSG(a) \
 187         { \
 188                 .mask = ( INFINIPATH_HWE_TXEMEMPARITYERR_##a <<    \
 189                           INFINIPATH_HWE_TXEMEMPARITYERR_SHIFT ),   \
 190                 .msg = "TXE " #a " Memory Parity"            \
 191         }
 192 #define INFINIPATH_HWE_RXEMEMPARITYERR_MSG(a) \
 193         { \
 194                 .mask = ( INFINIPATH_HWE_RXEMEMPARITYERR_##a <<    \
 195                           INFINIPATH_HWE_RXEMEMPARITYERR_SHIFT ),   \
 196                 .msg = "RXE " #a " Memory Parity"            \
 197         }
 198
 199 static const struct ipath_hwerror_msgs ipath_generic_hwerror_msgs[] = {
 200         INFINIPATH_HWE_MSG(IBCBUSFRSPCPARITYERR, "IPATH2IB Parity"),
 201         INFINIPATH_HWE_MSG(IBCBUSTOSPCPARITYERR, "IB2IPATH Parity"),
 202
 203         INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOBUF),
 204         INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOPBC),
 205         INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOLAUNCHFIFO),
 206
 207         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(RCVBUF),
 208         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(LOOKUPQ),
 209         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(EAGERTID),
 210         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(EXPTID),
 211         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(FLAGBUF),
 212         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(DATAINFO),
 213         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(HDRINFO),
 214 };
 215
 216 /**
 217  * ipath_format_hwmsg - format a single hwerror message
 218  * @msg message buffer
 219  * @msgl length of message buffer
 220  * @hwmsg message to add to message buffer
 221  */
 222 static void ipath_format_hwmsg(char *msg, size_t msgl, const char *hwmsg)
 223 {
 224         strlcat(msg, "[", msgl);
 225         strlcat(msg, hwmsg, msgl);
 226         strlcat(msg, "]", msgl);
 227 }
 228
 229 /**
 230  * ipath_format_hwerrors - format hardware error messages for display
 231  * @hwerrs hardware errors bit vector
 232  * @hwerrmsgs hardware error descriptions
 233  * @nhwerrmsgs number of hwerrmsgs
 234  * @msg message buffer
 235  * @msgl message buffer length
 236  */
 237 void ipath_format_hwerrors(u64 hwerrs,
 238                            const struct ipath_hwerror_msgs *hwerrmsgs,
 239                            size_t nhwerrmsgs,
 240                            char *msg, size_t msgl)
 241 {
 242         int i;
 243         const int glen =
 244             sizeof(ipath_generic_hwerror_msgs) /
 245             sizeof(ipath_generic_hwerror_msgs[0]);
 246
 247         for (i=0; i<glen; i++) {
 248                 if (hwerrs & ipath_generic_hwerror_msgs[i].mask) {
 249                         ipath_format_hwmsg(msg, msgl,
 250                                            ipath_generic_hwerror_msgs[i].msg);
 251                 }
 252         }
 253
 254         for (i=0; i<nhwerrmsgs; i++) {
 255                 if (hwerrs & hwerrmsgs[i].mask) {
 256                         ipath_format_hwmsg(msg, msgl, hwerrmsgs[i].msg);
 257                 }
 258         }
 259 }
 260
 261 /* return the strings for the most common link states */
 262 static char *ib_linkstate(struct ipath_devdata *dd, u64 ibcs)
 263 {
 264         char *ret;
 265         u32 state;
 266
 267         state = ipath_ib_state(dd, ibcs);
 268         if (state == dd->ib_init)
 269                 ret = "Init";
 270         else if (state == dd->ib_arm)
 271                 ret = "Arm";
 272         else if (state == dd->ib_active)
 273                 ret = "Active";
 274         else
 275                 ret = "Down";
 276         return ret;
 277 }
 278
 279 void signal_ib_event(struct ipath_devdata *dd, enum ib_event_type ev)
 280 {
 281         struct ib_event event;
 282
 283         event.device = &dd->verbs_dev->ibdev;
 284         event.element.port_num = 1;
 285         event.event = ev;
 286         ib_dispatch_event(&event);
 287 }
 288
 289 static void handle_e_ibstatuschanged(struct ipath_devdata *dd,
 290                                      ipath_err_t errs)
 291 {
 292         u32 ltstate, lstate, ibstate, lastlstate;
 293         u32 init = dd->ib_init;
 294         u32 arm = dd->ib_arm;
 295         u32 active = dd->ib_active;
 296         const u64 ibcs = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
 297
 298         lstate = ipath_ib_linkstate(dd, ibcs); /* linkstate */
 299         ibstate = ipath_ib_state(dd, ibcs);
 300         /* linkstate at last interrupt */
 301         lastlstate = ipath_ib_linkstate(dd, dd->ipath_lastibcstat);
 302         ltstate = ipath_ib_linktrstate(dd, ibcs); /* linktrainingtate */
 303
 304         /*
 305          * Since going into a recovery state causes the link state to go
 306          * down and since recovery is transitory, it is better if we "miss"
 307          * ever seeing the link training state go into recovery (i.e.,
 308          * ignore this transition for link state special handling purposes)
 309          * without even updating ipath_lastibcstat.
 310          */
 311         if ((ltstate == INFINIPATH_IBCS_LT_STATE_RECOVERRETRAIN) ||
 312             (ltstate == INFINIPATH_IBCS_LT_STATE_RECOVERWAITRMT) ||
 313             (ltstate == INFINIPATH_IBCS_LT_STATE_RECOVERIDLE))
 314                 goto done;
 315
 316         /*
 317          * if linkstate transitions into INIT from any of the various down
 318          * states, or if it transitions from any of the up (INIT or better)
 319          * states into any of the down states (except link recovery), then
 320          * call the chip-specific code to take appropriate actions.
 321          */
 322         if (lstate >= INFINIPATH_IBCS_L_STATE_INIT &&
 323                 lastlstate == INFINIPATH_IBCS_L_STATE_DOWN) {
 324                 /* transitioned to UP */
 325                 if (dd->ipath_f_ib_updown(dd, 1, ibcs)) {
 326                         /* link came up, so we must no longer be disabled */
 327                         dd->ipath_flags &= ~IPATH_IB_LINK_DISABLED;
 328                         ipath_cdbg(LINKVERB, "LinkUp handled, skipped\n");
 329                         goto skip_ibchange; /* chip-code handled */
 330                 }
 331         } else if ((lastlstate >= INFINIPATH_IBCS_L_STATE_INIT ||
 332                 (dd->ipath_flags & IPATH_IB_FORCE_NOTIFY)) &&
 333                 ltstate <= INFINIPATH_IBCS_LT_STATE_CFGWAITRMT &&
 334                 ltstate != INFINIPATH_IBCS_LT_STATE_LINKUP) {
 335                 int handled;
 336                 handled = dd->ipath_f_ib_updown(dd, 0, ibcs);
 337                 dd->ipath_flags &= ~IPATH_IB_FORCE_NOTIFY;
 338                 if (handled) {
 339                         ipath_cdbg(LINKVERB, "LinkDown handled, skipped\n");
 340                         goto skip_ibchange; /* chip-code handled */
 341                 }
 342         }
 343
 344         /*
 345          * Significant enough to always print and get into logs, if it was
 346          * unexpected.  If it was a requested state change, we'll have
 347          * already cleared the flags, so we won't print this warning
 348          */
 349         if ((ibstate != arm && ibstate != active) &&
 350             (dd->ipath_flags & (IPATH_LINKARMED | IPATH_LINKACTIVE))) {
 351                 dev_info(&dd->pcidev->dev, "Link state changed from %s "
 352                          "to %s\n", (dd->ipath_flags & IPATH_LINKARMED) ?
 353                          "ARM" : "ACTIVE", ib_linkstate(dd, ibcs));
 354         }
 355
 356         if (ltstate == INFINIPATH_IBCS_LT_STATE_POLLACTIVE ||
 357             ltstate == INFINIPATH_IBCS_LT_STATE_POLLQUIET) {
 358                 u32 lastlts;
 359                 lastlts = ipath_ib_linktrstate(dd, dd->ipath_lastibcstat);
 360                 /*
 361                  * Ignore cycling back and forth from Polling.Active to
 362                  * Polling.Quiet while waiting for the other end of the link
 363                  * to come up, except to try and decide if we are connected
 364                  * to a live IB device or not.  We will cycle back and
 365                  * forth between them if no cable is plugged in, the other
 366                  * device is powered off or disabled, etc.
 367                  */
 368                 if (lastlts == INFINIPATH_IBCS_LT_STATE_POLLACTIVE ||
 369                     lastlts == INFINIPATH_IBCS_LT_STATE_POLLQUIET) {
 370                         if (!(dd->ipath_flags & IPATH_IB_AUTONEG_INPROG) &&
 371                              (++dd->ipath_ibpollcnt == 40)) {
 372                                 dd->ipath_flags |= IPATH_NOCABLE;
 373                                 *dd->ipath_statusp |=
 374                                         IPATH_STATUS_IB_NOCABLE;
 375                                 ipath_cdbg(LINKVERB, "Set NOCABLE\n");
 376                         }
 377                         ipath_cdbg(LINKVERB, "POLL change to %s (%x)\n",
 378                                 ipath_ibcstatus_str[ltstate], ibstate);
 379                         goto skip_ibchange;
 380                 }
 381         }
 382
 383         dd->ipath_ibpollcnt = 0; /* not poll*, now */
 384         ipath_stats.sps_iblink++;
 385
 386         if (ibstate != init && dd->ipath_lastlinkrecov && ipath_linkrecovery) {
 387                 u64 linkrecov;
 388                 linkrecov = ipath_snap_cntr(dd,
 389                         dd->ipath_cregs->cr_iblinkerrrecovcnt);
 390                 if (linkrecov != dd->ipath_lastlinkrecov) {
 391                         ipath_dbg("IB linkrecov up %Lx (%s %s) recov %Lu\n",
 392                                 ibcs, ib_linkstate(dd, ibcs),
 393                                 ipath_ibcstatus_str[ltstate],
 394                                 linkrecov);
 395                         /* and no more until active again */
 396                         dd->ipath_lastlinkrecov = 0;
 397                         ipath_set_linkstate(dd, IPATH_IB_LINKDOWN);
 398                         goto skip_ibchange;
 399                 }
 400         }
 401
 402         if (ibstate == init || ibstate == arm || ibstate == active) {
 403                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_NOCABLE;
 404                 if (ibstate == init || ibstate == arm) {
 405                         *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
 406                         if (dd->ipath_flags & IPATH_LINKACTIVE)
 407                                 signal_ib_event(dd, IB_EVENT_PORT_ERR);
 408                 }
 409                 if (ibstate == arm) {
 410                         dd->ipath_flags |= IPATH_LINKARMED;
 411                         dd->ipath_flags &= ~(IPATH_LINKUNK |
 412                                 IPATH_LINKINIT | IPATH_LINKDOWN |
 413                                 IPATH_LINKACTIVE | IPATH_NOCABLE);
 414                         ipath_hol_down(dd);
 415                 } else  if (ibstate == init) {
 416                         /*
 417                          * set INIT and DOWN.  Down is checked by
 418                          * most of the other code, but INIT is
 419                          * useful to know in a few places.
 420                          */
 421                         dd->ipath_flags |= IPATH_LINKINIT |
 422                                 IPATH_LINKDOWN;
 423                         dd->ipath_flags &= ~(IPATH_LINKUNK |
 424                                 IPATH_LINKARMED | IPATH_LINKACTIVE |
 425                                 IPATH_NOCABLE);
 426                         ipath_hol_down(dd);
 427                 } else {  /* active */
 428                         dd->ipath_lastlinkrecov = ipath_snap_cntr(dd,
 429                                 dd->ipath_cregs->cr_iblinkerrrecovcnt);
 430                         *dd->ipath_statusp |=
 431                                 IPATH_STATUS_IB_READY | IPATH_STATUS_IB_CONF;
 432                         dd->ipath_flags |= IPATH_LINKACTIVE;
 433                         dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
 434                                 | IPATH_LINKDOWN | IPATH_LINKARMED |
 435                                 IPATH_NOCABLE);
 436                         if (dd->ipath_flags & IPATH_HAS_SEND_DMA)
 437                                 ipath_restart_sdma(dd);
 438                         signal_ib_event(dd, IB_EVENT_PORT_ACTIVE);
 439                         /* LED active not handled in chip _f_updown */
 440                         dd->ipath_f_setextled(dd, lstate, ltstate);
 441                         ipath_hol_up(dd);
 442                 }
 443
 444                 /*
 445                  * print after we've already done the work, so as not to
 446                  * delay the state changes and notifications, for debugging
 447                  */
 448                 if (lstate == lastlstate)
 449                         ipath_cdbg(LINKVERB, "Unchanged from last: %s "
 450                                 "(%x)\n", ib_linkstate(dd, ibcs), ibstate);
 451                 else
 452                         ipath_cdbg(VERBOSE, "Unit %u: link up to %s %s (%x)\n",
 453                                   dd->ipath_unit, ib_linkstate(dd, ibcs),
 454                                   ipath_ibcstatus_str[ltstate],  ibstate);
 455         } else { /* down */
 456                 if (dd->ipath_flags & IPATH_LINKACTIVE)
 457                         signal_ib_event(dd, IB_EVENT_PORT_ERR);
 458                 dd->ipath_flags |= IPATH_LINKDOWN;
 459                 dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
 460                                      | IPATH_LINKACTIVE |
 461                                      IPATH_LINKARMED);
 462                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
 463                 dd->ipath_lli_counter = 0;
 464
 465                 if (lastlstate != INFINIPATH_IBCS_L_STATE_DOWN)
 466                         ipath_cdbg(VERBOSE, "Unit %u link state down "
 467                                    "(state 0x%x), from %s\n",
 468                                    dd->ipath_unit, lstate,
 469                                    ib_linkstate(dd, dd->ipath_lastibcstat));
 470                 else
 471                         ipath_cdbg(LINKVERB, "Unit %u link state changed "
 472                                    "to %s (0x%x) from down (%x)\n",
 473                                    dd->ipath_unit,
 474                                    ipath_ibcstatus_str[ltstate],
 475                                    ibstate, lastlstate);
 476         }
 477
 478 skip_ibchange:
 479         dd->ipath_lastibcstat = ibcs;
 480 done:
 481         return;
 482 }
 483
 484 static void handle_supp_msgs(struct ipath_devdata *dd,
 485                              unsigned supp_msgs, char *msg, u32 msgsz)
 486 {
 487         /*
 488          * Print the message unless it's ibc status change only, which
 489          * happens so often we never want to count it.
 490          */
 491         if (dd->ipath_lasterror & ~INFINIPATH_E_IBSTATUSCHANGED) {
 492                 int iserr;
 493                 ipath_err_t mask;
 494                 iserr = ipath_decode_err(dd, msg, msgsz,
 495                                          dd->ipath_lasterror &
 496                                          ~INFINIPATH_E_IBSTATUSCHANGED);
 497
 498                 mask = INFINIPATH_E_RRCVEGRFULL | INFINIPATH_E_RRCVHDRFULL |
 499                         INFINIPATH_E_PKTERRS | INFINIPATH_E_SDMADISABLED;
 500
 501                 /* if we're in debug, then don't mask SDMADISABLED msgs */
 502                 if (ipath_debug & __IPATH_DBG)
 503                         mask &= ~INFINIPATH_E_SDMADISABLED;
 504
 505                 if (dd->ipath_lasterror & ~mask)
 506                         ipath_dev_err(dd, "Suppressed %u messages for "
 507                                       "fast-repeating errors (%s) (%llx)\n",
 508                                       supp_msgs, msg,
 509                                       (unsigned long long)
 510                                       dd->ipath_lasterror);
 511                 else {
 512                         /*
 513                          * rcvegrfull and rcvhdrqfull are "normal", for some
 514                          * types of processes (mostly benchmarks) that send
 515                          * huge numbers of messages, while not processing
 516                          * them. So only complain about these at debug
 517                          * level.
 518                          */
 519                         if (iserr)
 520                                 ipath_dbg("Suppressed %u messages for %s\n",
 521                                           supp_msgs, msg);
 522                         else
 523                                 ipath_cdbg(ERRPKT,
 524                                         "Suppressed %u messages for %s\n",
 525                                           supp_msgs, msg);
 526                 }
 527         }
 528 }
 529
 530 static unsigned handle_frequent_errors(struct ipath_devdata *dd,
 531                                        ipath_err_t errs, char *msg,
 532                                        u32 msgsz, int *noprint)
 533 {
 534         unsigned long nc;
 535         static unsigned long nextmsg_time;
 536         static unsigned nmsgs, supp_msgs;
 537
 538         /*
 539          * Throttle back "fast" messages to no more than 10 per 5 seconds.
 540          * This isn't perfect, but it's a reasonable heuristic. If we get
 541          * more than 10, give a 6x longer delay.
 542          */
 543         nc = jiffies;
 544         if (nmsgs > 10) {
 545                 if (time_before(nc, nextmsg_time)) {
 546                         *noprint = 1;
 547                         if (!supp_msgs++)
 548                                 nextmsg_time = nc + HZ * 3;
 549                 }
 550                 else if (supp_msgs) {
 551                         handle_supp_msgs(dd, supp_msgs, msg, msgsz);
 552                         supp_msgs = 0;
 553                         nmsgs = 0;
 554                 }
 555         }
 556         else if (!nmsgs++ || time_after(nc, nextmsg_time))
 557                 nextmsg_time = nc + HZ / 2;
 558
 559         return supp_msgs;
 560 }
 561
 562 static void handle_sdma_errors(struct ipath_devdata *dd, ipath_err_t errs)
 563 {
 564         unsigned long flags;
 565         int expected;
 566
 567         if (ipath_debug & __IPATH_DBG) {
 568                 char msg[128];
 569                 ipath_decode_err(dd, msg, sizeof msg, errs &
 570                         INFINIPATH_E_SDMAERRS);
 571                 ipath_dbg("errors %lx (%s)\n", (unsigned long)errs, msg);
 572         }
 573         if (ipath_debug & __IPATH_VERBDBG) {
 574                 unsigned long tl, hd, status, lengen;
 575                 tl = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmatail);
 576                 hd = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmahead);
 577                 status = ipath_read_kreg64(dd
 578                         , dd->ipath_kregs->kr_senddmastatus);
 579                 lengen = ipath_read_kreg64(dd,
 580                         dd->ipath_kregs->kr_senddmalengen);
 581                 ipath_cdbg(VERBOSE, "sdma tl 0x%lx hd 0x%lx status 0x%lx "
 582                         "lengen 0x%lx\n", tl, hd, status, lengen);
 583         }
 584
 585         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 586         __set_bit(IPATH_SDMA_DISABLED, &dd->ipath_sdma_status);
 587         expected = test_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status);
 588         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 589         if (!expected)
 590                 ipath_cancel_sends(dd, 1);
 591 }
 592
 593 static void handle_sdma_intr(struct ipath_devdata *dd, u64 istat)
 594 {
 595         unsigned long flags;
 596         int expected;
 597
 598         if ((istat & INFINIPATH_I_SDMAINT) &&
 599             !test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
 600                 ipath_sdma_intr(dd);
 601
 602         if (istat & INFINIPATH_I_SDMADISABLED) {
 603                 expected = test_bit(IPATH_SDMA_ABORTING,
 604                         &dd->ipath_sdma_status);
 605                 ipath_dbg("%s SDmaDisabled intr\n",
 606                         expected ? "expected" : "unexpected");
 607                 spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 608                 __set_bit(IPATH_SDMA_DISABLED, &dd->ipath_sdma_status);
 609                 spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 610                 if (!expected)
 611                         ipath_cancel_sends(dd, 1);
 612                 if (!test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
 613                         tasklet_hi_schedule(&dd->ipath_sdma_abort_task);
 614         }
 615 }
 616
 617 static int handle_hdrq_full(struct ipath_devdata *dd)
 618 {
 619         int chkerrpkts = 0;
 620         u32 hd, tl;
 621         u32 i;
 622
 623         ipath_stats.sps_hdrqfull++;
 624         for (i = 0; i < dd->ipath_cfgports; i++) {
 625                 struct ipath_portdata *pd = dd->ipath_pd[i];
 626
 627                 if (i == 0) {
 628                         /*
 629                          * For kernel receive queues, we just want to know
 630                          * if there are packets in the queue that we can
 631                          * process.
 632                          */
 633                         if (pd->port_head != ipath_get_hdrqtail(pd))
 634                                 chkerrpkts |= 1 << i;
 635                         continue;
 636                 }
 637
 638                 /* Skip if user context is not open */
 639                 if (!pd || !pd->port_cnt)
 640                         continue;
 641
 642                 /* Don't report the same point multiple times. */
 643                 if (dd->ipath_flags & IPATH_NODMA_RTAIL)
 644                         tl = ipath_read_ureg32(dd, ur_rcvhdrtail, i);
 645                 else
 646                         tl = ipath_get_rcvhdrtail(pd);
 647                 if (tl == pd->port_lastrcvhdrqtail)
 648                         continue;
 649
 650                 hd = ipath_read_ureg32(dd, ur_rcvhdrhead, i);
 651                 if (hd == (tl + 1) || (!hd && tl == dd->ipath_hdrqlast)) {
 652                         pd->port_lastrcvhdrqtail = tl;
 653                         pd->port_hdrqfull++;
 654                         /* flush hdrqfull so that poll() sees it */
 655                         wmb();
 656                         wake_up_interruptible(&pd->port_wait);
 657                 }
 658         }
 659
 660         return chkerrpkts;
 661 }
 662
 663 static int handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
 664 {
 665         char msg[128];
 666         u64 ignore_this_time = 0;
 667         u64 iserr = 0;
 668         int chkerrpkts = 0, noprint = 0;
 669         unsigned supp_msgs;
 670         int log_idx;
 671
 672         /*
 673          * don't report errors that are masked, either at init
 674          * (not set in ipath_errormask), or temporarily (set in
 675          * ipath_maskederrs)
 676          */
 677         errs &= dd->ipath_errormask & ~dd->ipath_maskederrs;
 678
 679         supp_msgs = handle_frequent_errors(dd, errs, msg, (u32)sizeof msg,
 680                 &noprint);
 681
 682         /* do these first, they are most important */
 683         if (errs & INFINIPATH_E_HARDWARE) {
 684                 /* reuse same msg buf */
 685                 dd->ipath_f_handle_hwerrors(dd, msg, sizeof msg);
 686         } else {
 687                 u64 mask;
 688                 for (log_idx = 0; log_idx < IPATH_EEP_LOG_CNT; ++log_idx) {
 689                         mask = dd->ipath_eep_st_masks[log_idx].errs_to_log;
 690                         if (errs & mask)
 691                                 ipath_inc_eeprom_err(dd, log_idx, 1);
 692                 }
 693         }
 694
 695         if (errs & INFINIPATH_E_SDMAERRS)
 696                 handle_sdma_errors(dd, errs);
 697
 698         if (!noprint && (errs & ~dd->ipath_e_bitsextant))
 699                 ipath_dev_err(dd, "error interrupt with unknown errors "
 700                               "%llx set\n", (unsigned long long)
 701                               (errs & ~dd->ipath_e_bitsextant));
 702
 703         if (errs & E_SUM_ERRS)
 704                 ignore_this_time = handle_e_sum_errs(dd, errs);
 705         else if ((errs & E_SUM_LINK_PKTERRS) &&
 706             !(dd->ipath_flags & IPATH_LINKACTIVE)) {
 707                 /*
 708                  * This can happen when SMA is trying to bring the link
 709                  * up, but the IB link changes state at the "wrong" time.
 710                  * The IB logic then complains that the packet isn't
 711                  * valid.  We don't want to confuse people, so we just
 712                  * don't print them, except at debug
 713                  */
 714                 ipath_dbg("Ignoring packet errors %llx, because link not "
 715                           "ACTIVE\n", (unsigned long long) errs);
 716                 ignore_this_time = errs & E_SUM_LINK_PKTERRS;
 717         }
 718
 719         if (supp_msgs == 250000) {
 720                 int s_iserr;
 721                 /*
 722                  * It's not entirely reasonable assuming that the errors set
 723                  * in the last clear period are all responsible for the
 724                  * problem, but the alternative is to assume it's the only
 725                  * ones on this particular interrupt, which also isn't great
 726                  */
 727                 dd->ipath_maskederrs |= dd->ipath_lasterror | errs;
 728
 729                 dd->ipath_errormask &= ~dd->ipath_maskederrs;
 730                 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 731                                  dd->ipath_errormask);
 732                 s_iserr = ipath_decode_err(dd, msg, sizeof msg,
 733                                            dd->ipath_maskederrs);
 734
 735                 if (dd->ipath_maskederrs &
 736                     ~(INFINIPATH_E_RRCVEGRFULL |
 737                       INFINIPATH_E_RRCVHDRFULL | INFINIPATH_E_PKTERRS))
 738                         ipath_dev_err(dd, "Temporarily disabling "
 739                             "error(s) %llx reporting; too frequent (%s)\n",
 740                                 (unsigned long long) dd->ipath_maskederrs,
 741                                 msg);
 742                 else {
 743                         /*
 744                          * rcvegrfull and rcvhdrqfull are "normal",
 745                          * for some types of processes (mostly benchmarks)
 746                          * that send huge numbers of messages, while not
 747                          * processing them.  So only complain about
 748                          * these at debug level.
 749                          */
 750                         if (s_iserr)
 751                                 ipath_dbg("Temporarily disabling reporting "
 752                                     "too frequent queue full errors (%s)\n",
 753                                     msg);
 754                         else
 755                                 ipath_cdbg(ERRPKT,
 756                                     "Temporarily disabling reporting too"
 757                                     " frequent packet errors (%s)\n",
 758                                     msg);
 759                 }
 760
 761                 /*
 762                  * Re-enable the masked errors after around 3 minutes.  in
 763                  * ipath_get_faststats().  If we have a series of fast
 764                  * repeating but different errors, the interval will keep
 765                  * stretching out, but that's OK, as that's pretty
 766                  * catastrophic.
 767                  */
 768                 dd->ipath_unmasktime = jiffies + HZ * 180;
 769         }
 770
 771         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, errs);
 772         if (ignore_this_time)
 773                 errs &= ~ignore_this_time;
 774         if (errs & ~dd->ipath_lasterror) {
 775                 errs &= ~dd->ipath_lasterror;
 776                 /* never suppress duplicate hwerrors or ibstatuschange */
 777                 dd->ipath_lasterror |= errs &
 778                         ~(INFINIPATH_E_HARDWARE |
 779                           INFINIPATH_E_IBSTATUSCHANGED);
 780         }
 781
 782         if (errs & INFINIPATH_E_SENDSPECIALTRIGGER) {
 783                 dd->ipath_spectriggerhit++;
 784                 ipath_dbg("%lu special trigger hits\n",
 785                         dd->ipath_spectriggerhit);
 786         }
 787
 788         /* likely due to cancel; so suppress message unless verbose */
 789         if ((errs & (INFINIPATH_E_SPKTLEN | INFINIPATH_E_SPIOARMLAUNCH)) &&
 790                 dd->ipath_lastcancel > jiffies) {
 791                 /* armlaunch takes precedence; it often causes both. */
 792                 ipath_cdbg(VERBOSE,
 793                         "Suppressed %s error (%llx) after sendbuf cancel\n",
 794                         (errs &  INFINIPATH_E_SPIOARMLAUNCH) ?
 795                         "armlaunch" : "sendpktlen", (unsigned long long)errs);
 796                 errs &= ~(INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SPKTLEN);
 797         }
 798
 799         if (!errs)
 800                 return 0;
 801
 802         if (!noprint) {
 803                 ipath_err_t mask;
 804                 /*
 805                  * The ones we mask off are handled specially below
 806                  * or above.  Also mask SDMADISABLED by default as it
 807                  * is too chatty.
 808                  */
 809                 mask = INFINIPATH_E_IBSTATUSCHANGED |
 810                         INFINIPATH_E_RRCVEGRFULL | INFINIPATH_E_RRCVHDRFULL |
 811                         INFINIPATH_E_HARDWARE | INFINIPATH_E_SDMADISABLED;
 812
 813                 /* if we're in debug, then don't mask SDMADISABLED msgs */
 814                 if (ipath_debug & __IPATH_DBG)
 815                         mask &= ~INFINIPATH_E_SDMADISABLED;
 816
 817                 ipath_decode_err(dd, msg, sizeof msg, errs & ~mask);
 818         } else
 819                 /* so we don't need if (!noprint) at strlcat's below */
 820                 *msg = 0;
 821
 822         if (errs & E_SUM_PKTERRS) {
 823                 ipath_stats.sps_pkterrs++;
 824                 chkerrpkts = 1;
 825         }
 826         if (errs & E_SUM_ERRS)
 827                 ipath_stats.sps_errs++;
 828
 829         if (errs & (INFINIPATH_E_RICRC | INFINIPATH_E_RVCRC)) {
 830                 ipath_stats.sps_crcerrs++;
 831                 chkerrpkts = 1;
 832         }
 833         iserr = errs & ~(E_SUM_PKTERRS | INFINIPATH_E_PKTERRS);
 834
 835
 836         /*
 837          * We don't want to print these two as they happen, or we can make
 838          * the situation even worse, because it takes so long to print
 839          * messages to serial consoles.  Kernel ports get printed from
 840          * fast_stats, no more than every 5 seconds, user ports get printed
 841          * on close
 842          */
 843         if (errs & INFINIPATH_E_RRCVHDRFULL)
 844                 chkerrpkts |= handle_hdrq_full(dd);
 845         if (errs & INFINIPATH_E_RRCVEGRFULL) {
 846                 struct ipath_portdata *pd = dd->ipath_pd[0];
 847
 848                 /*
 849                  * since this is of less importance and not likely to
 850                  * happen without also getting hdrfull, only count
 851                  * occurrences; don't check each port (or even the kernel
 852                  * vs user)
 853                  */
 854                 ipath_stats.sps_etidfull++;
 855                 if (pd->port_head != ipath_get_hdrqtail(pd))
 856                         chkerrpkts |= 1;
 857         }
 858
 859         /*
 860          * do this before IBSTATUSCHANGED, in case both bits set in a single
 861          * interrupt; we want the STATUSCHANGE to "win", so we do our
 862          * internal copy of state machine correctly
 863          */
 864         if (errs & INFINIPATH_E_RIBLOSTLINK) {
 865                 /*
 866                  * force through block below
 867                  */
 868                 errs |= INFINIPATH_E_IBSTATUSCHANGED;
 869                 ipath_stats.sps_iblink++;
 870                 dd->ipath_flags |= IPATH_LINKDOWN;
 871                 dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
 872                                      | IPATH_LINKARMED | IPATH_LINKACTIVE);
 873                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
 874
 875                 ipath_dbg("Lost link, link now down (%s)\n",
 876                         ipath_ibcstatus_str[ipath_read_kreg64(dd,
 877                         dd->ipath_kregs->kr_ibcstatus) & 0xf]);
 878         }
 879         if (errs & INFINIPATH_E_IBSTATUSCHANGED)
 880                 handle_e_ibstatuschanged(dd, errs);
 881
 882         if (errs & INFINIPATH_E_RESET) {
 883                 if (!noprint)
 884                         ipath_dev_err(dd, "Got reset, requires re-init "
 885                                       "(unload and reload driver)\n");
 886                 dd->ipath_flags &= ~IPATH_INITTED;      /* needs re-init */
 887                 /* mark as having had error */
 888                 *dd->ipath_statusp |= IPATH_STATUS_HWERROR;
 889                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_CONF;
 890         }
 891
 892         if (!noprint && *msg) {
 893                 if (iserr)
 894                         ipath_dev_err(dd, "%s error\n", msg);
 895         }
 896         if (dd->ipath_state_wanted & dd->ipath_flags) {
 897                 ipath_cdbg(VERBOSE, "driver wanted state %x, iflags now %x, "
 898                            "waking\n", dd->ipath_state_wanted,
 899                            dd->ipath_flags);
 900                 wake_up_interruptible(&ipath_state_wait);
 901         }
 902
 903         return chkerrpkts;
 904 }
 905
 906 /*
 907  * try to cleanup as much as possible for anything that might have gone
 908  * wrong while in freeze mode, such as pio buffers being written by user
 909  * processes (causing armlaunch), send errors due to going into freeze mode,
 910  * etc., and try to avoid causing extra interrupts while doing so.
 911  * Forcibly update the in-memory pioavail register copies after cleanup
 912  * because the chip won't do it for anything changing while in freeze mode
 913  * (we don't want to wait for the next pio buffer state change).
 914  * Make sure that we don't lose any important interrupts by using the chip
 915  * feature that says that writing 0 to a bit in *clear that is set in
 916  * *status will cause an interrupt to be generated again (if allowed by
 917  * the *mask value).
 918  */
 919 void ipath_clear_freeze(struct ipath_devdata *dd)
 920 {
 921         int i, im;
 922         u64 val;
 923
 924         /* disable error interrupts, to avoid confusion */
 925         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask, 0ULL);
 926
 927         /* also disable interrupts; errormask is sometimes overwriten */
 928         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
 929
 930         /*
 931          * clear all sends, because they have may been
 932          * completed by usercode while in freeze mode, and
 933          * therefore would not be sent, and eventually
 934          * might cause the process to run out of bufs
 935          */
 936         ipath_cancel_sends(dd, 0);
 937         ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
 938                          dd->ipath_control);
 939
 940         /* ensure pio avail updates continue */
 941         ipath_force_pio_avail_update(dd);
 942
 943         /*
 944          * We just enabled pioavailupdate, so dma copy is almost certainly
 945          * not yet right, so read the registers directly.  Similar to init
 946          */
 947         for (i = 0; i < dd->ipath_pioavregs; i++) {
 948                 /* deal with 6110 chip bug */
 949                 im = (i > 3 && (dd->ipath_flags & IPATH_SWAP_PIOBUFS)) ?
 950                         i ^ 1 : i;
 951                 val = ipath_read_kreg64(dd, (0x1000 / sizeof(u64)) + im);
 952                 dd->ipath_pioavailregs_dma[i] = cpu_to_le64(val);
 953                 dd->ipath_pioavailshadow[i] = val |
 954                         (~dd->ipath_pioavailkernel[i] <<
 955                         INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT);
 956         }
 957
 958         /*
 959          * force new interrupt if any hwerr, error or interrupt bits are
 960          * still set, and clear "safe" send packet errors related to freeze
 961          * and cancelling sends.  Re-enable error interrupts before possible
 962          * force of re-interrupt on pending interrupts.
 963          */
 964         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear, 0ULL);
 965         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear,
 966                 E_SPKT_ERRS_IGNORE);
 967         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 968                 dd->ipath_errormask);
 969         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, -1LL);
 970         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
 971 }
 972
 973
 974 /* this is separate to allow for better optimization of ipath_intr() */
 975
 976 static noinline void ipath_bad_intr(struct ipath_devdata *dd, u32 *unexpectp)
 977 {
 978         /*
 979          * sometimes happen during driver init and unload, don't want
 980          * to process any interrupts at that point
 981          */
 982
 983         /* this is just a bandaid, not a fix, if something goes badly
 984          * wrong */
 985         if (++*unexpectp > 100) {
 986                 if (++*unexpectp > 105) {
 987                         /*
 988                          * ok, we must be taking somebody else's interrupts,
 989                          * due to a messed up mptable and/or PIRQ table, so
 990                          * unregister the interrupt.  We've seen this during
 991                          * linuxbios development work, and it may happen in
 992                          * the future again.
 993                          */
 994                         if (dd->pcidev && dd->ipath_irq) {
 995                                 ipath_dev_err(dd, "Now %u unexpected "
 996                                               "interrupts, unregistering "
 997                                               "interrupt handler\n",
 998                                               *unexpectp);
 999                                 ipath_dbg("free_irq of irq %d\n",
1000                                           dd->ipath_irq);
1001                                 dd->ipath_f_free_irq(dd);
1002                         }
1003                 }
1004                 if (ipath_read_ireg(dd, dd->ipath_kregs->kr_intmask)) {
1005                         ipath_dev_err(dd, "%u unexpected interrupts, "
1006                                       "disabling interrupts completely\n",
1007                                       *unexpectp);
1008                         /*
1009                          * disable all interrupts, something is very wrong
1010                          */
1011                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
1012                                          0ULL);
1013                 }
1014         } else if (*unexpectp > 1)
1015                 ipath_dbg("Interrupt when not ready, should not happen, "
1016                           "ignoring\n");
1017 }
1018
1019 static noinline void ipath_bad_regread(struct ipath_devdata *dd)
1020 {
1021         static int allbits;
1022
1023         /* separate routine, for better optimization of ipath_intr() */
1024
1025         /*
1026          * We print the message and disable interrupts, in hope of
1027          * having a better chance of debugging the problem.
1028          */
1029         ipath_dev_err(dd,
1030                       "Read of interrupt status failed (all bits set)\n");
1031         if (allbits++) {
1032                 /* disable all interrupts, something is very wrong */
1033                 ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
1034                 if (allbits == 2) {
1035                         ipath_dev_err(dd, "Still bad interrupt status, "
1036                                       "unregistering interrupt\n");
1037                         dd->ipath_f_free_irq(dd);
1038                 } else if (allbits > 2) {
1039                         if ((allbits % 10000) == 0)
1040                                 printk(".");
1041                 } else
1042                         ipath_dev_err(dd, "Disabling interrupts, "
1043                                       "multiple errors\n");
1044         }
1045 }
1046
1047 static void handle_layer_pioavail(struct ipath_devdata *dd)
1048 {
1049         unsigned long flags;
1050         int ret;
1051
1052         ret = ipath_ib_piobufavail(dd->verbs_dev);
1053         if (ret > 0)
1054                 goto set;
1055
1056         return;
1057 set:
1058         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
1059         dd->ipath_sendctrl |= INFINIPATH_S_PIOINTBUFAVAIL;
1060         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1061                          dd->ipath_sendctrl);
1062         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1063         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
1064 }
1065
1066 /*
1067  * Handle receive interrupts for user ports; this means a user
1068  * process was waiting for a packet to arrive, and didn't want
1069  * to poll
1070  */
1071 static void handle_urcv(struct ipath_devdata *dd, u64 istat)
1072 {
1073         u64 portr;
1074         int i;
1075         int rcvdint = 0;
1076
1077         /*
1078          * test_and_clear_bit(IPATH_PORT_WAITING_RCV) and
1079          * test_and_clear_bit(IPATH_PORT_WAITING_URG) below
1080          * would both like timely updates of the bits so that
1081          * we don't pass them by unnecessarily.  the rmb()
1082          * here ensures that we see them promptly -- the
1083          * corresponding wmb()'s are in ipath_poll_urgent()
1084          * and ipath_poll_next()...
1085          */
1086         rmb();
1087         portr = ((istat >> dd->ipath_i_rcvavail_shift) &
1088                  dd->ipath_i_rcvavail_mask) |
1089                 ((istat >> dd->ipath_i_rcvurg_shift) &
1090                  dd->ipath_i_rcvurg_mask);
1091         for (i = 1; i < dd->ipath_cfgports; i++) {
1092                 struct ipath_portdata *pd = dd->ipath_pd[i];
1093
1094                 if (portr & (1 << i) && pd && pd->port_cnt) {
1095                         if (test_and_clear_bit(IPATH_PORT_WAITING_RCV,
1096                                                &pd->port_flag)) {
1097                                 clear_bit(i + dd->ipath_r_intravail_shift,
1098                                           &dd->ipath_rcvctrl);
1099                                 wake_up_interruptible(&pd->port_wait);
1100                                 rcvdint = 1;
1101                         } else if (test_and_clear_bit(IPATH_PORT_WAITING_URG,
1102                                                       &pd->port_flag)) {
1103                                 pd->port_urgent++;
1104                                 wake_up_interruptible(&pd->port_wait);
1105                         }
1106                 }
1107         }
1108         if (rcvdint) {
1109                 /* only want to take one interrupt, so turn off the rcv
1110                  * interrupt for all the ports that we set the rcv_waiting
1111                  * (but never for kernel port)
1112                  */
1113                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1114                                  dd->ipath_rcvctrl);
1115         }
1116 }
1117
1118 irqreturn_t ipath_intr(int irq, void *data)
1119 {
1120         struct ipath_devdata *dd = data;
1121         u64 istat, chk0rcv = 0;
1122         ipath_err_t estat = 0;
1123         irqreturn_t ret;
1124         static unsigned unexpected = 0;
1125         u64 kportrbits;
1126
1127         ipath_stats.sps_ints++;
1128
1129         if (dd->ipath_int_counter != (u32) -1)
1130                 dd->ipath_int_counter++;
1131
1132         if (!(dd->ipath_flags & IPATH_PRESENT)) {
1133                 /*
1134                  * This return value is not great, but we do not want the
1135                  * interrupt core code to remove our interrupt handler
1136                  * because we don't appear to be handling an interrupt
1137                  * during a chip reset.
1138                  */
1139                 return IRQ_HANDLED;
1140         }
1141
1142         /*
1143          * this needs to be flags&initted, not statusp, so we keep
1144          * taking interrupts even after link goes down, etc.
1145          * Also, we *must* clear the interrupt at some point, or we won't
1146          * take it again, which can be real bad for errors, etc...
1147          */
1148
1149         if (!(dd->ipath_flags & IPATH_INITTED)) {
1150                 ipath_bad_intr(dd, &unexpected);
1151                 ret = IRQ_NONE;
1152                 goto bail;
1153         }
1154
1155         istat = ipath_read_ireg(dd, dd->ipath_kregs->kr_intstatus);
1156
1157         if (unlikely(!istat)) {
1158                 ipath_stats.sps_nullintr++;
1159                 ret = IRQ_NONE; /* not our interrupt, or already handled */
1160                 goto bail;
1161         }
1162         if (unlikely(istat == -1)) {
1163                 ipath_bad_regread(dd);
1164                 /* don't know if it was our interrupt or not */
1165                 ret = IRQ_NONE;
1166                 goto bail;
1167         }
1168
1169         if (unexpected)
1170                 unexpected = 0;
1171
1172         if (unlikely(istat & ~dd->ipath_i_bitsextant))
1173                 ipath_dev_err(dd,
1174                               "interrupt with unknown interrupts %Lx set\n",
1175                               istat & ~dd->ipath_i_bitsextant);
1176         else if (istat & ~INFINIPATH_I_ERROR) /* errors do own printing */
1177                 ipath_cdbg(VERBOSE, "intr stat=0x%Lx\n", istat);
1178
1179         if (istat & INFINIPATH_I_ERROR) {
1180                 ipath_stats.sps_errints++;
1181                 estat = ipath_read_kreg64(dd,
1182                                           dd->ipath_kregs->kr_errorstatus);
1183                 if (!estat)
1184                         dev_info(&dd->pcidev->dev, "error interrupt (%Lx), "
1185                                  "but no error bits set!\n", istat);
1186                 else if (estat == -1LL)
1187                         /*
1188                          * should we try clearing all, or hope next read
1189                          * works?
1190                          */
1191                         ipath_dev_err(dd, "Read of error status failed "
1192                                       "(all bits set); ignoring\n");
1193                 else
1194                         chk0rcv |= handle_errors(dd, estat);
1195         }
1196
1197         if (istat & INFINIPATH_I_GPIO) {
1198                 /*
1199                  * GPIO interrupts fall in two broad classes:
1200                  * GPIO_2 indicates (on some HT4xx boards) that a packet
1201                  *        has arrived for Port 0. Checking for this
1202                  *        is controlled by flag IPATH_GPIO_INTR.
1203                  * GPIO_3..5 on IBA6120 Rev2 and IBA6110 Rev4 chips indicate
1204                  *        errors that we need to count. Checking for this
1205                  *        is controlled by flag IPATH_GPIO_ERRINTRS.
1206                  */
1207                 u32 gpiostatus;
1208                 u32 to_clear = 0;
1209
1210                 gpiostatus = ipath_read_kreg32(
1211                         dd, dd->ipath_kregs->kr_gpio_status);
1212                 /* First the error-counter case. */
1213                 if ((gpiostatus & IPATH_GPIO_ERRINTR_MASK) &&
1214                     (dd->ipath_flags & IPATH_GPIO_ERRINTRS)) {
1215                         /* want to clear the bits we see asserted. */
1216                         to_clear |= (gpiostatus & IPATH_GPIO_ERRINTR_MASK);
1217
1218                         /*
1219                          * Count appropriately, clear bits out of our copy,
1220                          * as they have been "handled".
1221                          */
1222                         if (gpiostatus & (1 << IPATH_GPIO_RXUVL_BIT)) {
1223                                 ipath_dbg("FlowCtl on UnsupVL\n");
1224                                 dd->ipath_rxfc_unsupvl_errs++;
1225                         }
1226                         if (gpiostatus & (1 << IPATH_GPIO_OVRUN_BIT)) {
1227                                 ipath_dbg("Overrun Threshold exceeded\n");
1228                                 dd->ipath_overrun_thresh_errs++;
1229                         }
1230                         if (gpiostatus & (1 << IPATH_GPIO_LLI_BIT)) {
1231                                 ipath_dbg("Local Link Integrity error\n");
1232                                 dd->ipath_lli_errs++;
1233                         }
1234                         gpiostatus &= ~IPATH_GPIO_ERRINTR_MASK;
1235                 }
1236                 /* Now the Port0 Receive case */
1237                 if ((gpiostatus & (1 << IPATH_GPIO_PORT0_BIT)) &&
1238                     (dd->ipath_flags & IPATH_GPIO_INTR)) {
1239                         /*
1240                          * GPIO status bit 2 is set, and we expected it.
1241                          * clear it and indicate in p0bits.
1242                          * This probably only happens if a Port0 pkt
1243                          * arrives at _just_ the wrong time, and we
1244                          * handle that by seting chk0rcv;
1245                          */
1246                         to_clear |= (1 << IPATH_GPIO_PORT0_BIT);
1247                         gpiostatus &= ~(1 << IPATH_GPIO_PORT0_BIT);
1248                         chk0rcv = 1;
1249                 }
1250                 if (gpiostatus) {
1251                         /*
1252                          * Some unexpected bits remain. If they could have
1253                          * caused the interrupt, complain and clear.
1254                          * To avoid repetition of this condition, also clear
1255                          * the mask. It is almost certainly due to error.
1256                          */
1257                         const u32 mask = (u32) dd->ipath_gpio_mask;
1258
1259                         if (mask & gpiostatus) {
1260                                 ipath_dbg("Unexpected GPIO IRQ bits %x\n",
1261                                   gpiostatus & mask);
1262                                 to_clear |= (gpiostatus & mask);
1263                                 dd->ipath_gpio_mask &= ~(gpiostatus & mask);
1264                                 ipath_write_kreg(dd,
1265                                         dd->ipath_kregs->kr_gpio_mask,
1266                                         dd->ipath_gpio_mask);
1267                         }
1268                 }
1269                 if (to_clear) {
1270                         ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
1271                                         (u64) to_clear);
1272                 }
1273         }
1274
1275         /*
1276          * Clear the interrupt bits we found set, unless they are receive
1277          * related, in which case we already cleared them above, and don't
1278          * want to clear them again, because we might lose an interrupt.
1279          * Clear it early, so we "know" know the chip will have seen this by
1280          * the time we process the queue, and will re-interrupt if necessary.
1281          * The processor itself won't take the interrupt again until we return.
1282          */
1283         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
1284
1285         /*
1286          * Handle kernel receive queues before checking for pio buffers
1287          * available since receives can overflow; piobuf waiters can afford
1288          * a few extra cycles, since they were waiting anyway, and user's
1289          * waiting for receive are at the bottom.
1290          */
1291         kportrbits = (1ULL << dd->ipath_i_rcvavail_shift) |
1292                 (1ULL << dd->ipath_i_rcvurg_shift);
1293         if (chk0rcv || (istat & kportrbits)) {
1294                 istat &= ~kportrbits;
1295                 ipath_kreceive(dd->ipath_pd[0]);
1296         }
1297
1298         if (istat & ((dd->ipath_i_rcvavail_mask << dd->ipath_i_rcvavail_shift) |
1299                      (dd->ipath_i_rcvurg_mask << dd->ipath_i_rcvurg_shift)))
1300                 handle_urcv(dd, istat);
1301
1302         if (istat & (INFINIPATH_I_SDMAINT | INFINIPATH_I_SDMADISABLED))
1303                 handle_sdma_intr(dd, istat);
1304
1305         if (istat & INFINIPATH_I_SPIOBUFAVAIL) {
1306                 unsigned long flags;
1307
1308                 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
1309                 dd->ipath_sendctrl &= ~INFINIPATH_S_PIOINTBUFAVAIL;
1310                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1311                                  dd->ipath_sendctrl);
1312                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1313                 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
1314
1315                 if (!(dd->ipath_flags & IPATH_HAS_SEND_DMA))
1316                         handle_layer_pioavail(dd);
1317                 else
1318                         ipath_dbg("unexpected BUFAVAIL intr\n");
1319         }
1320
1321         ret = IRQ_HANDLED;
1322
1323 bail:
1324         return ret;
1325 }