drivers/infiniband/hw/ipath/ipath_intr.c

   1 /*
   2  * Copyright (c) 2006, 2007 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
  36 #include "ipath_kernel.h"
  37 #include "ipath_verbs.h"
  38 #include "ipath_common.h"
  39
  40 /*
  41  * clear (write) a pio buffer, to clear a parity error.   This routine
  42  * should only be called when in freeze mode, and the buffer should be
  43  * canceled afterwards.
  44  */
  45 static void ipath_clrpiobuf(struct ipath_devdata *dd, u32 pnum)
  46 {
  47         u32 __iomem *pbuf;
  48         u32 dwcnt; /* dword count to write */
  49         if (pnum < dd->ipath_piobcnt2k) {
  50                 pbuf = (u32 __iomem *) (dd->ipath_pio2kbase + pnum *
  51                         dd->ipath_palign);
  52                 dwcnt = dd->ipath_piosize2k >> 2;
  53         }
  54         else {
  55                 pbuf = (u32 __iomem *) (dd->ipath_pio4kbase +
  56                         (pnum - dd->ipath_piobcnt2k) * dd->ipath_4kalign);
  57                 dwcnt = dd->ipath_piosize4k >> 2;
  58         }
  59         dev_info(&dd->pcidev->dev,
  60                 "Rewrite PIO buffer %u, to recover from parity error\n",
  61                 pnum);
  62
  63         /* no flush required, since already in freeze */
  64         writel(dwcnt + 1, pbuf);
  65         while (--dwcnt)
  66                 writel(0, pbuf++);
  67 }
  68
  69 /*
  70  * Called when we might have an error that is specific to a particular
  71  * PIO buffer, and may need to cancel that buffer, so it can be re-used.
  72  * If rewrite is true, and bits are set in the sendbufferror registers,
  73  * we'll write to the buffer, for error recovery on parity errors.
  74  */
  75 static void ipath_disarm_senderrbufs(struct ipath_devdata *dd, int rewrite)
  76 {
  77         u32 piobcnt;
  78         unsigned long sbuf[4];
  79         /*
  80          * it's possible that sendbuffererror could have bits set; might
  81          * have already done this as a result of hardware error handling
  82          */
  83         piobcnt = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
  84         /* read these before writing errorclear */
  85         sbuf[0] = ipath_read_kreg64(
  86                 dd, dd->ipath_kregs->kr_sendbuffererror);
  87         sbuf[1] = ipath_read_kreg64(
  88                 dd, dd->ipath_kregs->kr_sendbuffererror + 1);
  89         if (piobcnt > 128) {
  90                 sbuf[2] = ipath_read_kreg64(
  91                         dd, dd->ipath_kregs->kr_sendbuffererror + 2);
  92                 sbuf[3] = ipath_read_kreg64(
  93                         dd, dd->ipath_kregs->kr_sendbuffererror + 3);
  94         }
  95
  96         if (sbuf[0] || sbuf[1] || (piobcnt > 128 && (sbuf[2] || sbuf[3]))) {
  97                 int i;
  98                 if (ipath_debug & (__IPATH_PKTDBG|__IPATH_DBG) &&
  99                         dd->ipath_lastcancel > jiffies) {
 100                         __IPATH_DBG_WHICH(__IPATH_PKTDBG|__IPATH_DBG,
 101                                           "SendbufErrs %lx %lx", sbuf[0],
 102                                           sbuf[1]);
 103                         if (ipath_debug & __IPATH_PKTDBG && piobcnt > 128)
 104                                 printk(" %lx %lx ", sbuf[2], sbuf[3]);
 105                         printk("\n");
 106                 }
 107
 108                 for (i = 0; i < piobcnt; i++)
 109                         if (test_bit(i, sbuf)) {
 110                                 if (rewrite)
 111                                         ipath_clrpiobuf(dd, i);
 112                                 ipath_disarm_piobufs(dd, i, 1);
 113                         }
 114                 /* ignore armlaunch errs for a bit */
 115                 dd->ipath_lastcancel = jiffies+3;
 116         }
 117 }
 118
 119
 120 /* These are all rcv-related errors which we want to count for stats */
 121 #define E_SUM_PKTERRS \
 122         (INFINIPATH_E_RHDRLEN | INFINIPATH_E_RBADTID | \
 123          INFINIPATH_E_RBADVERSION | INFINIPATH_E_RHDR | \
 124          INFINIPATH_E_RLONGPKTLEN | INFINIPATH_E_RSHORTPKTLEN | \
 125          INFINIPATH_E_RMAXPKTLEN | INFINIPATH_E_RMINPKTLEN | \
 126          INFINIPATH_E_RFORMATERR | INFINIPATH_E_RUNSUPVL | \
 127          INFINIPATH_E_RUNEXPCHAR | INFINIPATH_E_REBP)
 128
 129 /* These are all send-related errors which we want to count for stats */
 130 #define E_SUM_ERRS \
 131         (INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SUNEXPERRPKTNUM | \
 132          INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
 133          INFINIPATH_E_SMAXPKTLEN | INFINIPATH_E_SUNSUPVL | \
 134          INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
 135          INFINIPATH_E_INVALIDADDR)
 136
 137 /*
 138  * this is similar to E_SUM_ERRS, but can't ignore armlaunch, don't ignore
 139  * errors not related to freeze and cancelling buffers.  Can't ignore
 140  * armlaunch because could get more while still cleaning up, and need
 141  * to cancel those as they happen.
 142  */
 143 #define E_SPKT_ERRS_IGNORE \
 144          (INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
 145          INFINIPATH_E_SMAXPKTLEN | INFINIPATH_E_SMINPKTLEN | \
 146          INFINIPATH_E_SPKTLEN)
 147
 148 /*
 149  * these are errors that can occur when the link changes state while
 150  * a packet is being sent or received.  This doesn't cover things
 151  * like EBP or VCRC that can be the result of a sending having the
 152  * link change state, so we receive a "known bad" packet.
 153  */
 154 #define E_SUM_LINK_PKTERRS \
 155         (INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
 156          INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
 157          INFINIPATH_E_RSHORTPKTLEN | INFINIPATH_E_RMINPKTLEN | \
 158          INFINIPATH_E_RUNEXPCHAR)
 159
 160 static u64 handle_e_sum_errs(struct ipath_devdata *dd, ipath_err_t errs)
 161 {
 162         u64 ignore_this_time = 0;
 163
 164         ipath_disarm_senderrbufs(dd, 0);
 165         if ((errs & E_SUM_LINK_PKTERRS) &&
 166             !(dd->ipath_flags & IPATH_LINKACTIVE)) {
 167                 /*
 168                  * This can happen when SMA is trying to bring the link
 169                  * up, but the IB link changes state at the "wrong" time.
 170                  * The IB logic then complains that the packet isn't
 171                  * valid.  We don't want to confuse people, so we just
 172                  * don't print them, except at debug
 173                  */
 174                 ipath_dbg("Ignoring packet errors %llx, because link not "
 175                           "ACTIVE\n", (unsigned long long) errs);
 176                 ignore_this_time = errs & E_SUM_LINK_PKTERRS;
 177         }
 178
 179         return ignore_this_time;
 180 }
 181
 182 /* generic hw error messages... */
 183 #define INFINIPATH_HWE_TXEMEMPARITYERR_MSG(a) \
 184         { \
 185                 .mask = ( INFINIPATH_HWE_TXEMEMPARITYERR_##a <<    \
 186                           INFINIPATH_HWE_TXEMEMPARITYERR_SHIFT ),   \
 187                 .msg = "TXE " #a " Memory Parity"            \
 188         }
 189 #define INFINIPATH_HWE_RXEMEMPARITYERR_MSG(a) \
 190         { \
 191                 .mask = ( INFINIPATH_HWE_RXEMEMPARITYERR_##a <<    \
 192                           INFINIPATH_HWE_RXEMEMPARITYERR_SHIFT ),   \
 193                 .msg = "RXE " #a " Memory Parity"            \
 194         }
 195
 196 static const struct ipath_hwerror_msgs ipath_generic_hwerror_msgs[] = {
 197         INFINIPATH_HWE_MSG(IBCBUSFRSPCPARITYERR, "IPATH2IB Parity"),
 198         INFINIPATH_HWE_MSG(IBCBUSTOSPCPARITYERR, "IB2IPATH Parity"),
 199
 200         INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOBUF),
 201         INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOPBC),
 202         INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOLAUNCHFIFO),
 203
 204         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(RCVBUF),
 205         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(LOOKUPQ),
 206         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(EAGERTID),
 207         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(EXPTID),
 208         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(FLAGBUF),
 209         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(DATAINFO),
 210         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(HDRINFO),
 211 };
 212
 213 /**
 214  * ipath_format_hwmsg - format a single hwerror message
 215  * @msg message buffer
 216  * @msgl length of message buffer
 217  * @hwmsg message to add to message buffer
 218  */
 219 static void ipath_format_hwmsg(char *msg, size_t msgl, const char *hwmsg)
 220 {
 221         strlcat(msg, "[", msgl);
 222         strlcat(msg, hwmsg, msgl);
 223         strlcat(msg, "]", msgl);
 224 }
 225
 226 /**
 227  * ipath_format_hwerrors - format hardware error messages for display
 228  * @hwerrs hardware errors bit vector
 229  * @hwerrmsgs hardware error descriptions
 230  * @nhwerrmsgs number of hwerrmsgs
 231  * @msg message buffer
 232  * @msgl message buffer length
 233  */
 234 void ipath_format_hwerrors(u64 hwerrs,
 235                            const struct ipath_hwerror_msgs *hwerrmsgs,
 236                            size_t nhwerrmsgs,
 237                            char *msg, size_t msgl)
 238 {
 239         int i;
 240         const int glen =
 241             sizeof(ipath_generic_hwerror_msgs) /
 242             sizeof(ipath_generic_hwerror_msgs[0]);
 243
 244         for (i=0; i<glen; i++) {
 245                 if (hwerrs & ipath_generic_hwerror_msgs[i].mask) {
 246                         ipath_format_hwmsg(msg, msgl,
 247                                            ipath_generic_hwerror_msgs[i].msg);
 248                 }
 249         }
 250
 251         for (i=0; i<nhwerrmsgs; i++) {
 252                 if (hwerrs & hwerrmsgs[i].mask) {
 253                         ipath_format_hwmsg(msg, msgl, hwerrmsgs[i].msg);
 254                 }
 255         }
 256 }
 257
 258 /* return the strings for the most common link states */
 259 static char *ib_linkstate(u32 linkstate)
 260 {
 261         char *ret;
 262
 263         switch (linkstate) {
 264         case IPATH_IBSTATE_INIT:
 265                 ret = "Init";
 266                 break;
 267         case IPATH_IBSTATE_ARM:
 268                 ret = "Arm";
 269                 break;
 270         case IPATH_IBSTATE_ACTIVE:
 271                 ret = "Active";
 272                 break;
 273         default:
 274                 ret = "Down";
 275         }
 276
 277         return ret;
 278 }
 279
 280 void signal_ib_event(struct ipath_devdata *dd, enum ib_event_type ev)
 281 {
 282         struct ib_event event;
 283
 284         event.device = &dd->verbs_dev->ibdev;
 285         event.element.port_num = 1;
 286         event.event = ev;
 287         ib_dispatch_event(&event);
 288 }
 289
 290 static void handle_e_ibstatuschanged(struct ipath_devdata *dd,
 291                                      ipath_err_t errs, int noprint)
 292 {
 293         u64 val;
 294         u32 ltstate, lstate;
 295
 296         /*
 297          * even if diags are enabled, we want to notice LINKINIT, etc.
 298          * We just don't want to change the LED state, or
 299          * dd->ipath_kregs->kr_ibcctrl
 300          */
 301         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
 302         lstate = val & IPATH_IBSTATE_MASK;
 303
 304         /*
 305          * this is confusing enough when it happens that I want to always put it
 306          * on the console and in the logs.  If it was a requested state change,
 307          * we'll have already cleared the flags, so we won't print this warning
 308          */
 309         if ((lstate != IPATH_IBSTATE_ARM && lstate != IPATH_IBSTATE_ACTIVE)
 310                 && (dd->ipath_flags & (IPATH_LINKARMED | IPATH_LINKACTIVE))) {
 311                 dev_info(&dd->pcidev->dev, "Link state changed from %s to %s\n",
 312                                  (dd->ipath_flags & IPATH_LINKARMED) ? "ARM" : "ACTIVE",
 313                                  ib_linkstate(lstate));
 314                 /*
 315                  * Flush all queued sends when link went to DOWN or INIT,
 316                  * to be sure that they don't block SMA and other MAD packets
 317                  */
 318                 ipath_cancel_sends(dd, 1);
 319         }
 320         else if (lstate == IPATH_IBSTATE_INIT || lstate == IPATH_IBSTATE_ARM ||
 321             lstate == IPATH_IBSTATE_ACTIVE) {
 322                 /*
 323                  * only print at SMA if there is a change, debug if not
 324                  * (sometimes we want to know that, usually not).
 325                  */
 326                 if (lstate == ((unsigned) dd->ipath_lastibcstat
 327                                & IPATH_IBSTATE_MASK)) {
 328                         ipath_dbg("Status change intr but no change (%s)\n",
 329                                   ib_linkstate(lstate));
 330                 }
 331                 else
 332                         ipath_cdbg(VERBOSE, "Unit %u link state %s, last "
 333                                    "was %s\n", dd->ipath_unit,
 334                                    ib_linkstate(lstate),
 335                                    ib_linkstate((unsigned)
 336                                                 dd->ipath_lastibcstat
 337                                                 & IPATH_IBSTATE_MASK));
 338         }
 339         else {
 340                 lstate = dd->ipath_lastibcstat & IPATH_IBSTATE_MASK;
 341                 if (lstate == IPATH_IBSTATE_INIT ||
 342                     lstate == IPATH_IBSTATE_ARM ||
 343                     lstate == IPATH_IBSTATE_ACTIVE)
 344                         ipath_cdbg(VERBOSE, "Unit %u link state down"
 345                                    " (state 0x%x), from %s\n",
 346                                    dd->ipath_unit,
 347                                    (u32)val & IPATH_IBSTATE_MASK,
 348                                    ib_linkstate(lstate));
 349                 else
 350                         ipath_cdbg(VERBOSE, "Unit %u link state changed "
 351                                    "to 0x%x from down (%x)\n",
 352                                    dd->ipath_unit, (u32) val, lstate);
 353         }
 354         ltstate = (val >> INFINIPATH_IBCS_LINKTRAININGSTATE_SHIFT) &
 355                 INFINIPATH_IBCS_LINKTRAININGSTATE_MASK;
 356         lstate = (val >> INFINIPATH_IBCS_LINKSTATE_SHIFT) &
 357                 INFINIPATH_IBCS_LINKSTATE_MASK;
 358
 359         if (ltstate == INFINIPATH_IBCS_LT_STATE_POLLACTIVE ||
 360             ltstate == INFINIPATH_IBCS_LT_STATE_POLLQUIET) {
 361                 u32 last_ltstate;
 362
 363                 /*
 364                  * Ignore cycling back and forth from Polling.Active
 365                  * to Polling.Quiet while waiting for the other end of
 366                  * the link to come up. We will cycle back and forth
 367                  * between them if no cable is plugged in,
 368                  * the other device is powered off or disabled, etc.
 369                  */
 370                 last_ltstate = (dd->ipath_lastibcstat >>
 371                                 INFINIPATH_IBCS_LINKTRAININGSTATE_SHIFT)
 372                         & INFINIPATH_IBCS_LINKTRAININGSTATE_MASK;
 373                 if (last_ltstate == INFINIPATH_IBCS_LT_STATE_POLLACTIVE
 374                     || last_ltstate ==
 375                     INFINIPATH_IBCS_LT_STATE_POLLQUIET) {
 376                         if (dd->ipath_ibpollcnt > 40) {
 377                                 dd->ipath_flags |= IPATH_NOCABLE;
 378                                 *dd->ipath_statusp |=
 379                                         IPATH_STATUS_IB_NOCABLE;
 380                         } else
 381                                 dd->ipath_ibpollcnt++;
 382                         goto skip_ibchange;
 383                 }
 384         }
 385         dd->ipath_ibpollcnt = 0;        /* some state other than 2 or 3 */
 386         ipath_stats.sps_iblink++;
 387         if (ltstate != INFINIPATH_IBCS_LT_STATE_LINKUP) {
 388                 if (dd->ipath_flags & IPATH_LINKACTIVE)
 389                         signal_ib_event(dd, IB_EVENT_PORT_ERR);
 390                 dd->ipath_flags |= IPATH_LINKDOWN;
 391                 dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
 392                                      | IPATH_LINKACTIVE |
 393                                      IPATH_LINKARMED);
 394                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
 395                 dd->ipath_lli_counter = 0;
 396                 if (!noprint) {
 397                         if (((dd->ipath_lastibcstat >>
 398                               INFINIPATH_IBCS_LINKSTATE_SHIFT) &
 399                              INFINIPATH_IBCS_LINKSTATE_MASK)
 400                             == INFINIPATH_IBCS_L_STATE_ACTIVE)
 401                                 /* if from up to down be more vocal */
 402                                 ipath_cdbg(VERBOSE,
 403                                            "Unit %u link now down (%s)\n",
 404                                            dd->ipath_unit,
 405                                            ipath_ibcstatus_str[ltstate]);
 406                         else
 407                                 ipath_cdbg(VERBOSE, "Unit %u link is "
 408                                            "down (%s)\n", dd->ipath_unit,
 409                                            ipath_ibcstatus_str[ltstate]);
 410                 }
 411
 412                 dd->ipath_f_setextled(dd, lstate, ltstate);
 413         } else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_ACTIVE) {
 414                 dd->ipath_flags |= IPATH_LINKACTIVE;
 415                 dd->ipath_flags &=
 416                         ~(IPATH_LINKUNK | IPATH_LINKINIT | IPATH_LINKDOWN |
 417                           IPATH_LINKARMED | IPATH_NOCABLE);
 418                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_NOCABLE;
 419                 *dd->ipath_statusp |=
 420                         IPATH_STATUS_IB_READY | IPATH_STATUS_IB_CONF;
 421                 dd->ipath_f_setextled(dd, lstate, ltstate);
 422                 signal_ib_event(dd, IB_EVENT_PORT_ACTIVE);
 423         } else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_INIT) {
 424                 if (dd->ipath_flags & IPATH_LINKACTIVE)
 425                         signal_ib_event(dd, IB_EVENT_PORT_ERR);
 426                 /*
 427                  * set INIT and DOWN.  Down is checked by most of the other
 428                  * code, but INIT is useful to know in a few places.
 429                  */
 430                 dd->ipath_flags |= IPATH_LINKINIT | IPATH_LINKDOWN;
 431                 dd->ipath_flags &=
 432                         ~(IPATH_LINKUNK | IPATH_LINKACTIVE | IPATH_LINKARMED
 433                           | IPATH_NOCABLE);
 434                 *dd->ipath_statusp &= ~(IPATH_STATUS_IB_NOCABLE
 435                                         | IPATH_STATUS_IB_READY);
 436                 dd->ipath_f_setextled(dd, lstate, ltstate);
 437         } else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_ARM) {
 438                 if (dd->ipath_flags & IPATH_LINKACTIVE)
 439                         signal_ib_event(dd, IB_EVENT_PORT_ERR);
 440                 dd->ipath_flags |= IPATH_LINKARMED;
 441                 dd->ipath_flags &=
 442                         ~(IPATH_LINKUNK | IPATH_LINKDOWN | IPATH_LINKINIT |
 443                           IPATH_LINKACTIVE | IPATH_NOCABLE);
 444                 *dd->ipath_statusp &= ~(IPATH_STATUS_IB_NOCABLE
 445                                         | IPATH_STATUS_IB_READY);
 446                 dd->ipath_f_setextled(dd, lstate, ltstate);
 447         } else {
 448                 if (!noprint)
 449                         ipath_dbg("IBstatuschange unit %u: %s (%x)\n",
 450                                   dd->ipath_unit,
 451                                   ipath_ibcstatus_str[ltstate], ltstate);
 452         }
 453 skip_ibchange:
 454         dd->ipath_lastibcstat = val;
 455 }
 456
 457 static void handle_supp_msgs(struct ipath_devdata *dd,
 458                              unsigned supp_msgs, char *msg, int msgsz)
 459 {
 460         /*
 461          * Print the message unless it's ibc status change only, which
 462          * happens so often we never want to count it.
 463          */
 464         if (dd->ipath_lasterror & ~INFINIPATH_E_IBSTATUSCHANGED) {
 465                 int iserr;
 466                 iserr = ipath_decode_err(msg, msgsz,
 467                                          dd->ipath_lasterror &
 468                                          ~INFINIPATH_E_IBSTATUSCHANGED);
 469                 if (dd->ipath_lasterror &
 470                         ~(INFINIPATH_E_RRCVEGRFULL |
 471                         INFINIPATH_E_RRCVHDRFULL | INFINIPATH_E_PKTERRS))
 472                         ipath_dev_err(dd, "Suppressed %u messages for "
 473                                       "fast-repeating errors (%s) (%llx)\n",
 474                                       supp_msgs, msg,
 475                                       (unsigned long long)
 476                                       dd->ipath_lasterror);
 477                 else {
 478                         /*
 479                          * rcvegrfull and rcvhdrqfull are "normal", for some
 480                          * types of processes (mostly benchmarks) that send
 481                          * huge numbers of messages, while not processing
 482                          * them. So only complain about these at debug
 483                          * level.
 484                          */
 485                         if (iserr)
 486                                 ipath_dbg("Suppressed %u messages for %s\n",
 487                                           supp_msgs, msg);
 488                         else
 489                                 ipath_cdbg(ERRPKT,
 490                                         "Suppressed %u messages for %s\n",
 491                                           supp_msgs, msg);
 492                 }
 493         }
 494 }
 495
 496 static unsigned handle_frequent_errors(struct ipath_devdata *dd,
 497                                        ipath_err_t errs, char *msg,
 498                                        int msgsz, int *noprint)
 499 {
 500         unsigned long nc;
 501         static unsigned long nextmsg_time;
 502         static unsigned nmsgs, supp_msgs;
 503
 504         /*
 505          * Throttle back "fast" messages to no more than 10 per 5 seconds.
 506          * This isn't perfect, but it's a reasonable heuristic. If we get
 507          * more than 10, give a 6x longer delay.
 508          */
 509         nc = jiffies;
 510         if (nmsgs > 10) {
 511                 if (time_before(nc, nextmsg_time)) {
 512                         *noprint = 1;
 513                         if (!supp_msgs++)
 514                                 nextmsg_time = nc + HZ * 3;
 515                 }
 516                 else if (supp_msgs) {
 517                         handle_supp_msgs(dd, supp_msgs, msg, msgsz);
 518                         supp_msgs = 0;
 519                         nmsgs = 0;
 520                 }
 521         }
 522         else if (!nmsgs++ || time_after(nc, nextmsg_time))
 523                 nextmsg_time = nc + HZ / 2;
 524
 525         return supp_msgs;
 526 }
 527
 528 static int handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
 529 {
 530         char msg[128];
 531         u64 ignore_this_time = 0;
 532         int i, iserr = 0;
 533         int chkerrpkts = 0, noprint = 0;
 534         unsigned supp_msgs;
 535         int log_idx;
 536
 537         supp_msgs = handle_frequent_errors(dd, errs, msg, sizeof msg, &noprint);
 538
 539         /* don't report errors that are masked */
 540         errs &= ~dd->ipath_maskederrs;
 541
 542         /* do these first, they are most important */
 543         if (errs & INFINIPATH_E_HARDWARE) {
 544                 /* reuse same msg buf */
 545                 dd->ipath_f_handle_hwerrors(dd, msg, sizeof msg);
 546         } else {
 547                 u64 mask;
 548                 for (log_idx = 0; log_idx < IPATH_EEP_LOG_CNT; ++log_idx) {
 549                         mask = dd->ipath_eep_st_masks[log_idx].errs_to_log;
 550                         if (errs & mask)
 551                                 ipath_inc_eeprom_err(dd, log_idx, 1);
 552                 }
 553         }
 554
 555         if (!noprint && (errs & ~dd->ipath_e_bitsextant))
 556                 ipath_dev_err(dd, "error interrupt with unknown errors "
 557                               "%llx set\n", (unsigned long long)
 558                               (errs & ~dd->ipath_e_bitsextant));
 559
 560         if (errs & E_SUM_ERRS)
 561                 ignore_this_time = handle_e_sum_errs(dd, errs);
 562         else if ((errs & E_SUM_LINK_PKTERRS) &&
 563             !(dd->ipath_flags & IPATH_LINKACTIVE)) {
 564                 /*
 565                  * This can happen when SMA is trying to bring the link
 566                  * up, but the IB link changes state at the "wrong" time.
 567                  * The IB logic then complains that the packet isn't
 568                  * valid.  We don't want to confuse people, so we just
 569                  * don't print them, except at debug
 570                  */
 571                 ipath_dbg("Ignoring packet errors %llx, because link not "
 572                           "ACTIVE\n", (unsigned long long) errs);
 573                 ignore_this_time = errs & E_SUM_LINK_PKTERRS;
 574         }
 575
 576         if (supp_msgs == 250000) {
 577                 int s_iserr;
 578                 /*
 579                  * It's not entirely reasonable assuming that the errors set
 580                  * in the last clear period are all responsible for the
 581                  * problem, but the alternative is to assume it's the only
 582                  * ones on this particular interrupt, which also isn't great
 583                  */
 584                 dd->ipath_maskederrs |= dd->ipath_lasterror | errs;
 585                 dd->ipath_errormask &= ~dd->ipath_maskederrs;
 586                 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 587                         dd->ipath_errormask);
 588                 s_iserr = ipath_decode_err(msg, sizeof msg,
 589                         dd->ipath_maskederrs);
 590
 591                 if (dd->ipath_maskederrs &
 592                         ~(INFINIPATH_E_RRCVEGRFULL |
 593                         INFINIPATH_E_RRCVHDRFULL | INFINIPATH_E_PKTERRS))
 594                         ipath_dev_err(dd, "Temporarily disabling "
 595                             "error(s) %llx reporting; too frequent (%s)\n",
 596                                 (unsigned long long)dd->ipath_maskederrs,
 597                                 msg);
 598                 else {
 599                         /*
 600                          * rcvegrfull and rcvhdrqfull are "normal",
 601                          * for some types of processes (mostly benchmarks)
 602                          * that send huge numbers of messages, while not
 603                          * processing them.  So only complain about
 604                          * these at debug level.
 605                          */
 606                         if (s_iserr)
 607                                 ipath_dbg("Temporarily disabling reporting "
 608                                     "too frequent queue full errors (%s)\n",
 609                                     msg);
 610                         else
 611                                 ipath_cdbg(ERRPKT,
 612                                     "Temporarily disabling reporting too"
 613                                     " frequent packet errors (%s)\n",
 614                                     msg);
 615                 }
 616
 617                 /*
 618                  * Re-enable the masked errors after around 3 minutes.  in
 619                  * ipath_get_faststats().  If we have a series of fast
 620                  * repeating but different errors, the interval will keep
 621                  * stretching out, but that's OK, as that's pretty
 622                  * catastrophic.
 623                  */
 624                 dd->ipath_unmasktime = jiffies + HZ * 180;
 625         }
 626
 627         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, errs);
 628         if (ignore_this_time)
 629                 errs &= ~ignore_this_time;
 630         if (errs & ~dd->ipath_lasterror) {
 631                 errs &= ~dd->ipath_lasterror;
 632                 /* never suppress duplicate hwerrors or ibstatuschange */
 633                 dd->ipath_lasterror |= errs &
 634                         ~(INFINIPATH_E_HARDWARE |
 635                           INFINIPATH_E_IBSTATUSCHANGED);
 636         }
 637
 638         /* likely due to cancel, so suppress */
 639         if ((errs & (INFINIPATH_E_SPKTLEN | INFINIPATH_E_SPIOARMLAUNCH)) &&
 640                 dd->ipath_lastcancel > jiffies) {
 641                 ipath_dbg("Suppressed armlaunch/spktlen after error send cancel\n");
 642                 errs &= ~(INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SPKTLEN);
 643         }
 644
 645         if (!errs)
 646                 return 0;
 647
 648         if (!noprint)
 649                 /*
 650                  * the ones we mask off are handled specially below or above
 651                  */
 652                 ipath_decode_err(msg, sizeof msg,
 653                                  errs & ~(INFINIPATH_E_IBSTATUSCHANGED |
 654                                           INFINIPATH_E_RRCVEGRFULL |
 655                                           INFINIPATH_E_RRCVHDRFULL |
 656                                           INFINIPATH_E_HARDWARE));
 657         else
 658                 /* so we don't need if (!noprint) at strlcat's below */
 659                 *msg = 0;
 660
 661         if (errs & E_SUM_PKTERRS) {
 662                 ipath_stats.sps_pkterrs++;
 663                 chkerrpkts = 1;
 664         }
 665         if (errs & E_SUM_ERRS)
 666                 ipath_stats.sps_errs++;
 667
 668         if (errs & (INFINIPATH_E_RICRC | INFINIPATH_E_RVCRC)) {
 669                 ipath_stats.sps_crcerrs++;
 670                 chkerrpkts = 1;
 671         }
 672         iserr = errs & ~(E_SUM_PKTERRS | INFINIPATH_E_PKTERRS);
 673
 674
 675         /*
 676          * We don't want to print these two as they happen, or we can make
 677          * the situation even worse, because it takes so long to print
 678          * messages to serial consoles.  Kernel ports get printed from
 679          * fast_stats, no more than every 5 seconds, user ports get printed
 680          * on close
 681          */
 682         if (errs & INFINIPATH_E_RRCVHDRFULL) {
 683                 u32 hd, tl;
 684                 ipath_stats.sps_hdrqfull++;
 685                 for (i = 0; i < dd->ipath_cfgports; i++) {
 686                         struct ipath_portdata *pd = dd->ipath_pd[i];
 687                         if (i == 0) {
 688                                 hd = pd->port_head;
 689                                 tl = (u32) le64_to_cpu(
 690                                         *dd->ipath_hdrqtailptr);
 691                         } else if (pd && pd->port_cnt &&
 692                                    pd->port_rcvhdrtail_kvaddr) {
 693                                 /*
 694                                  * don't report same point multiple times,
 695                                  * except kernel
 696                                  */
 697                                 tl = *(u64 *) pd->port_rcvhdrtail_kvaddr;
 698                                 if (tl == pd->port_lastrcvhdrqtail)
 699                                         continue;
 700                                 hd = ipath_read_ureg32(dd, ur_rcvhdrhead,
 701                                                        i);
 702                         } else
 703                                 continue;
 704                         if (hd == (tl + 1) ||
 705                             (!hd && tl == dd->ipath_hdrqlast)) {
 706                                 if (i == 0)
 707                                         chkerrpkts = 1;
 708                                 pd->port_lastrcvhdrqtail = tl;
 709                                 pd->port_hdrqfull++;
 710                                 /* flush hdrqfull so that poll() sees it */
 711                                 wmb();
 712                                 wake_up_interruptible(&pd->port_wait);
 713                         }
 714                 }
 715         }
 716         if (errs & INFINIPATH_E_RRCVEGRFULL) {
 717                 struct ipath_portdata *pd = dd->ipath_pd[0];
 718
 719                 /*
 720                  * since this is of less importance and not likely to
 721                  * happen without also getting hdrfull, only count
 722                  * occurrences; don't check each port (or even the kernel
 723                  * vs user)
 724                  */
 725                 ipath_stats.sps_etidfull++;
 726                 if (pd->port_head !=
 727                     (u32) le64_to_cpu(*dd->ipath_hdrqtailptr))
 728                         chkerrpkts = 1;
 729         }
 730
 731         /*
 732          * do this before IBSTATUSCHANGED, in case both bits set in a single
 733          * interrupt; we want the STATUSCHANGE to "win", so we do our
 734          * internal copy of state machine correctly
 735          */
 736         if (errs & INFINIPATH_E_RIBLOSTLINK) {
 737                 /*
 738                  * force through block below
 739                  */
 740                 errs |= INFINIPATH_E_IBSTATUSCHANGED;
 741                 ipath_stats.sps_iblink++;
 742                 dd->ipath_flags |= IPATH_LINKDOWN;
 743                 dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
 744                                      | IPATH_LINKARMED | IPATH_LINKACTIVE);
 745                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
 746                 if (!noprint) {
 747                         u64 st = ipath_read_kreg64(
 748                                 dd, dd->ipath_kregs->kr_ibcstatus);
 749
 750                         ipath_dbg("Lost link, link now down (%s)\n",
 751                                   ipath_ibcstatus_str[st & 0xf]);
 752                 }
 753         }
 754         if (errs & INFINIPATH_E_IBSTATUSCHANGED)
 755                 handle_e_ibstatuschanged(dd, errs, noprint);
 756
 757         if (errs & INFINIPATH_E_RESET) {
 758                 if (!noprint)
 759                         ipath_dev_err(dd, "Got reset, requires re-init "
 760                                       "(unload and reload driver)\n");
 761                 dd->ipath_flags &= ~IPATH_INITTED;      /* needs re-init */
 762                 /* mark as having had error */
 763                 *dd->ipath_statusp |= IPATH_STATUS_HWERROR;
 764                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_CONF;
 765         }
 766
 767         if (!noprint && *msg) {
 768                 if (iserr)
 769                         ipath_dev_err(dd, "%s error\n", msg);
 770                 else
 771                         dev_info(&dd->pcidev->dev, "%s packet problems\n",
 772                                 msg);
 773         }
 774         if (dd->ipath_state_wanted & dd->ipath_flags) {
 775                 ipath_cdbg(VERBOSE, "driver wanted state %x, iflags now %x, "
 776                            "waking\n", dd->ipath_state_wanted,
 777                            dd->ipath_flags);
 778                 wake_up_interruptible(&ipath_state_wait);
 779         }
 780
 781         return chkerrpkts;
 782 }
 783
 784
 785 /*
 786  * try to cleanup as much as possible for anything that might have gone
 787  * wrong while in freeze mode, such as pio buffers being written by user
 788  * processes (causing armlaunch), send errors due to going into freeze mode,
 789  * etc., and try to avoid causing extra interrupts while doing so.
 790  * Forcibly update the in-memory pioavail register copies after cleanup
 791  * because the chip won't do it for anything changing while in freeze mode
 792  * (we don't want to wait for the next pio buffer state change).
 793  * Make sure that we don't lose any important interrupts by using the chip
 794  * feature that says that writing 0 to a bit in *clear that is set in
 795  * *status will cause an interrupt to be generated again (if allowed by
 796  * the *mask value).
 797  */
 798 void ipath_clear_freeze(struct ipath_devdata *dd)
 799 {
 800         int i, im;
 801         __le64 val;
 802         unsigned long flags;
 803
 804         /* disable error interrupts, to avoid confusion */
 805         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask, 0ULL);
 806
 807         /* also disable interrupts; errormask is sometimes overwriten */
 808         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
 809
 810         /*
 811          * clear all sends, because they have may been
 812          * completed by usercode while in freeze mode, and
 813          * therefore would not be sent, and eventually
 814          * might cause the process to run out of bufs
 815          */
 816         ipath_cancel_sends(dd, 0);
 817         ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
 818                          dd->ipath_control);
 819
 820         /* ensure pio avail updates continue */
 821         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
 822         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 823                  dd->ipath_sendctrl & ~INFINIPATH_S_PIOBUFAVAILUPD);
 824         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 825         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 826                          dd->ipath_sendctrl);
 827         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 828         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
 829
 830         /*
 831          * We just enabled pioavailupdate, so dma copy is almost certainly
 832          * not yet right, so read the registers directly.  Similar to init
 833          */
 834         for (i = 0; i < dd->ipath_pioavregs; i++) {
 835                 /* deal with 6110 chip bug */
 836                 im = i > 3 ? i ^ 1 : i;
 837                 val = ipath_read_kreg64(dd, (0x1000 / sizeof(u64)) + im);
 838                 dd->ipath_pioavailregs_dma[i] = dd->ipath_pioavailshadow[i]
 839                         = le64_to_cpu(val);
 840         }
 841
 842         /*
 843          * force new interrupt if any hwerr, error or interrupt bits are
 844          * still set, and clear "safe" send packet errors related to freeze
 845          * and cancelling sends.  Re-enable error interrupts before possible
 846          * force of re-interrupt on pending interrupts.
 847          */
 848         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear, 0ULL);
 849         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear,
 850                 E_SPKT_ERRS_IGNORE);
 851         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 852                 dd->ipath_errormask);
 853         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, -1LL);
 854         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
 855 }
 856
 857
 858 /* this is separate to allow for better optimization of ipath_intr() */
 859
 860 static noinline void ipath_bad_intr(struct ipath_devdata *dd, u32 *unexpectp)
 861 {
 862         /*
 863          * sometimes happen during driver init and unload, don't want
 864          * to process any interrupts at that point
 865          */
 866
 867         /* this is just a bandaid, not a fix, if something goes badly
 868          * wrong */
 869         if (++*unexpectp > 100) {
 870                 if (++*unexpectp > 105) {
 871                         /*
 872                          * ok, we must be taking somebody else's interrupts,
 873                          * due to a messed up mptable and/or PIRQ table, so
 874                          * unregister the interrupt.  We've seen this during
 875                          * linuxbios development work, and it may happen in
 876                          * the future again.
 877                          */
 878                         if (dd->pcidev && dd->ipath_irq) {
 879                                 ipath_dev_err(dd, "Now %u unexpected "
 880                                               "interrupts, unregistering "
 881                                               "interrupt handler\n",
 882                                               *unexpectp);
 883                                 ipath_dbg("free_irq of irq %d\n",
 884                                           dd->ipath_irq);
 885                                 dd->ipath_f_free_irq(dd);
 886                         }
 887                 }
 888                 if (ipath_read_ireg(dd, dd->ipath_kregs->kr_intmask)) {
 889                         ipath_dev_err(dd, "%u unexpected interrupts, "
 890                                       "disabling interrupts completely\n",
 891                                       *unexpectp);
 892                         /*
 893                          * disable all interrupts, something is very wrong
 894                          */
 895                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
 896                                          0ULL);
 897                 }
 898         } else if (*unexpectp > 1)
 899                 ipath_dbg("Interrupt when not ready, should not happen, "
 900                           "ignoring\n");
 901 }
 902
 903 static noinline void ipath_bad_regread(struct ipath_devdata *dd)
 904 {
 905         static int allbits;
 906
 907         /* separate routine, for better optimization of ipath_intr() */
 908
 909         /*
 910          * We print the message and disable interrupts, in hope of
 911          * having a better chance of debugging the problem.
 912          */
 913         ipath_dev_err(dd,
 914                       "Read of interrupt status failed (all bits set)\n");
 915         if (allbits++) {
 916                 /* disable all interrupts, something is very wrong */
 917                 ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
 918                 if (allbits == 2) {
 919                         ipath_dev_err(dd, "Still bad interrupt status, "
 920                                       "unregistering interrupt\n");
 921                         dd->ipath_f_free_irq(dd);
 922                 } else if (allbits > 2) {
 923                         if ((allbits % 10000) == 0)
 924                                 printk(".");
 925                 } else
 926                         ipath_dev_err(dd, "Disabling interrupts, "
 927                                       "multiple errors\n");
 928         }
 929 }
 930
 931 static void handle_layer_pioavail(struct ipath_devdata *dd)
 932 {
 933         unsigned long flags;
 934         int ret;
 935
 936         ret = ipath_ib_piobufavail(dd->verbs_dev);
 937         if (ret > 0)
 938                 goto set;
 939
 940         return;
 941 set:
 942         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
 943         dd->ipath_sendctrl |= INFINIPATH_S_PIOINTBUFAVAIL;
 944         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 945                          dd->ipath_sendctrl);
 946         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 947         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
 948 }
 949
 950 /*
 951  * Handle receive interrupts for user ports; this means a user
 952  * process was waiting for a packet to arrive, and didn't want
 953  * to poll
 954  */
 955 static void handle_urcv(struct ipath_devdata *dd, u32 istat)
 956 {
 957         u64 portr;
 958         int i;
 959         int rcvdint = 0;
 960
 961         /*
 962          * test_and_clear_bit(IPATH_PORT_WAITING_RCV) and
 963          * test_and_clear_bit(IPATH_PORT_WAITING_URG) below
 964          * would both like timely updates of the bits so that
 965          * we don't pass them by unnecessarily.  the rmb()
 966          * here ensures that we see them promptly -- the
 967          * corresponding wmb()'s are in ipath_poll_urgent()
 968          * and ipath_poll_next()...
 969          */
 970         rmb();
 971         portr = ((istat >> INFINIPATH_I_RCVAVAIL_SHIFT) &
 972                  dd->ipath_i_rcvavail_mask)
 973                 | ((istat >> INFINIPATH_I_RCVURG_SHIFT) &
 974                    dd->ipath_i_rcvurg_mask);
 975         for (i = 1; i < dd->ipath_cfgports; i++) {
 976                 struct ipath_portdata *pd = dd->ipath_pd[i];
 977                 if (portr & (1 << i) && pd && pd->port_cnt) {
 978                         if (test_and_clear_bit(IPATH_PORT_WAITING_RCV,
 979                                                &pd->port_flag)) {
 980                                 clear_bit(i + dd->ipath_r_intravail_shift,
 981                                           &dd->ipath_rcvctrl);
 982                                 wake_up_interruptible(&pd->port_wait);
 983                                 rcvdint = 1;
 984                         } else if (test_and_clear_bit(IPATH_PORT_WAITING_URG,
 985                                                       &pd->port_flag)) {
 986                                 pd->port_urgent++;
 987                                 wake_up_interruptible(&pd->port_wait);
 988                         }
 989                 }
 990         }
 991         if (rcvdint) {
 992                 /* only want to take one interrupt, so turn off the rcv
 993                  * interrupt for all the ports that we did the wakeup on
 994                  * (but never for kernel port)
 995                  */
 996                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 997                                  dd->ipath_rcvctrl);
 998         }
 999 }
1000
1001 irqreturn_t ipath_intr(int irq, void *data)
1002 {
1003         struct ipath_devdata *dd = data;
1004         u32 istat, chk0rcv = 0;
1005         ipath_err_t estat = 0;
1006         irqreturn_t ret;
1007         static unsigned unexpected = 0;
1008         static const u32 port0rbits = (1U<<INFINIPATH_I_RCVAVAIL_SHIFT) |
1009                  (1U<<INFINIPATH_I_RCVURG_SHIFT);
1010
1011         ipath_stats.sps_ints++;
1012
1013         if (dd->ipath_int_counter != (u32) -1)
1014                 dd->ipath_int_counter++;
1015
1016         if (!(dd->ipath_flags & IPATH_PRESENT)) {
1017                 /*
1018                  * This return value is not great, but we do not want the
1019                  * interrupt core code to remove our interrupt handler
1020                  * because we don't appear to be handling an interrupt
1021                  * during a chip reset.
1022                  */
1023                 return IRQ_HANDLED;
1024         }
1025
1026         /*
1027          * this needs to be flags&initted, not statusp, so we keep
1028          * taking interrupts even after link goes down, etc.
1029          * Also, we *must* clear the interrupt at some point, or we won't
1030          * take it again, which can be real bad for errors, etc...
1031          */
1032
1033         if (!(dd->ipath_flags & IPATH_INITTED)) {
1034                 ipath_bad_intr(dd, &unexpected);
1035                 ret = IRQ_NONE;
1036                 goto bail;
1037         }
1038
1039         istat = ipath_read_ireg(dd, dd->ipath_kregs->kr_intstatus);
1040
1041         if (unlikely(!istat)) {
1042                 ipath_stats.sps_nullintr++;
1043                 ret = IRQ_NONE; /* not our interrupt, or already handled */
1044                 goto bail;
1045         }
1046         if (unlikely(istat == -1)) {
1047                 ipath_bad_regread(dd);
1048                 /* don't know if it was our interrupt or not */
1049                 ret = IRQ_NONE;
1050                 goto bail;
1051         }
1052
1053         if (unexpected)
1054                 unexpected = 0;
1055
1056         if (unlikely(istat & ~dd->ipath_i_bitsextant))
1057                 ipath_dev_err(dd,
1058                               "interrupt with unknown interrupts %x set\n",
1059                               istat & (u32) ~ dd->ipath_i_bitsextant);
1060         else
1061                 ipath_cdbg(VERBOSE, "intr stat=0x%x\n", istat);
1062
1063         if (unlikely(istat & INFINIPATH_I_ERROR)) {
1064                 ipath_stats.sps_errints++;
1065                 estat = ipath_read_kreg64(dd,
1066                                           dd->ipath_kregs->kr_errorstatus);
1067                 if (!estat)
1068                         dev_info(&dd->pcidev->dev, "error interrupt (%x), "
1069                                  "but no error bits set!\n", istat);
1070                 else if (estat == -1LL)
1071                         /*
1072                          * should we try clearing all, or hope next read
1073                          * works?
1074                          */
1075                         ipath_dev_err(dd, "Read of error status failed "
1076                                       "(all bits set); ignoring\n");
1077                 else
1078                         if (handle_errors(dd, estat))
1079                                 /* force calling ipath_kreceive() */
1080                                 chk0rcv = 1;
1081         }
1082
1083         if (istat & INFINIPATH_I_GPIO) {
1084                 /*
1085                  * GPIO interrupts fall in two broad classes:
1086                  * GPIO_2 indicates (on some HT4xx boards) that a packet
1087                  *        has arrived for Port 0. Checking for this
1088                  *        is controlled by flag IPATH_GPIO_INTR.
1089                  * GPIO_3..5 on IBA6120 Rev2 and IBA6110 Rev4 chips indicate
1090                  *        errors that we need to count. Checking for this
1091                  *        is controlled by flag IPATH_GPIO_ERRINTRS.
1092                  */
1093                 u32 gpiostatus;
1094                 u32 to_clear = 0;
1095
1096                 gpiostatus = ipath_read_kreg32(
1097                         dd, dd->ipath_kregs->kr_gpio_status);
1098                 /* First the error-counter case.
1099                  */
1100                 if ((gpiostatus & IPATH_GPIO_ERRINTR_MASK) &&
1101                     (dd->ipath_flags & IPATH_GPIO_ERRINTRS)) {
1102                         /* want to clear the bits we see asserted. */
1103                         to_clear |= (gpiostatus & IPATH_GPIO_ERRINTR_MASK);
1104
1105                         /*
1106                          * Count appropriately, clear bits out of our copy,
1107                          * as they have been "handled".
1108                          */
1109                         if (gpiostatus & (1 << IPATH_GPIO_RXUVL_BIT)) {
1110                                 ipath_dbg("FlowCtl on UnsupVL\n");
1111                                 dd->ipath_rxfc_unsupvl_errs++;
1112                         }
1113                         if (gpiostatus & (1 << IPATH_GPIO_OVRUN_BIT)) {
1114                                 ipath_dbg("Overrun Threshold exceeded\n");
1115                                 dd->ipath_overrun_thresh_errs++;
1116                         }
1117                         if (gpiostatus & (1 << IPATH_GPIO_LLI_BIT)) {
1118                                 ipath_dbg("Local Link Integrity error\n");
1119                                 dd->ipath_lli_errs++;
1120                         }
1121                         gpiostatus &= ~IPATH_GPIO_ERRINTR_MASK;
1122                 }
1123                 /* Now the Port0 Receive case */
1124                 if ((gpiostatus & (1 << IPATH_GPIO_PORT0_BIT)) &&
1125                     (dd->ipath_flags & IPATH_GPIO_INTR)) {
1126                         /*
1127                          * GPIO status bit 2 is set, and we expected it.
1128                          * clear it and indicate in p0bits.
1129                          * This probably only happens if a Port0 pkt
1130                          * arrives at _just_ the wrong time, and we
1131                          * handle that by seting chk0rcv;
1132                          */
1133                         to_clear |= (1 << IPATH_GPIO_PORT0_BIT);
1134                         gpiostatus &= ~(1 << IPATH_GPIO_PORT0_BIT);
1135                         chk0rcv = 1;
1136                 }
1137                 if (gpiostatus) {
1138                         /*
1139                          * Some unexpected bits remain. If they could have
1140                          * caused the interrupt, complain and clear.
1141                          * To avoid repetition of this condition, also clear
1142                          * the mask. It is almost certainly due to error.
1143                          */
1144                         const u32 mask = (u32) dd->ipath_gpio_mask;
1145
1146                         if (mask & gpiostatus) {
1147                                 ipath_dbg("Unexpected GPIO IRQ bits %x\n",
1148                                   gpiostatus & mask);
1149                                 to_clear |= (gpiostatus & mask);
1150                                 dd->ipath_gpio_mask &= ~(gpiostatus & mask);
1151                                 ipath_write_kreg(dd,
1152                                         dd->ipath_kregs->kr_gpio_mask,
1153                                         dd->ipath_gpio_mask);
1154                         }
1155                 }
1156                 if (to_clear) {
1157                         ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
1158                                         (u64) to_clear);
1159                 }
1160         }
1161         chk0rcv |= istat & port0rbits;
1162
1163         /*
1164          * Clear the interrupt bits we found set, unless they are receive
1165          * related, in which case we already cleared them above, and don't
1166          * want to clear them again, because we might lose an interrupt.
1167          * Clear it early, so we "know" know the chip will have seen this by
1168          * the time we process the queue, and will re-interrupt if necessary.
1169          * The processor itself won't take the interrupt again until we return.
1170          */
1171         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
1172
1173         /*
1174          * handle port0 receive  before checking for pio buffers available,
1175          * since receives can overflow; piobuf waiters can afford a few
1176          * extra cycles, since they were waiting anyway, and user's waiting
1177          * for receive are at the bottom.
1178          */
1179         if (chk0rcv) {
1180                 ipath_kreceive(dd->ipath_pd[0]);
1181                 istat &= ~port0rbits;
1182         }
1183
1184         if (istat & ((dd->ipath_i_rcvavail_mask <<
1185                       INFINIPATH_I_RCVAVAIL_SHIFT)
1186                      | (dd->ipath_i_rcvurg_mask <<
1187                         INFINIPATH_I_RCVURG_SHIFT)))
1188                 handle_urcv(dd, istat);
1189
1190         if (istat & INFINIPATH_I_SPIOBUFAVAIL) {
1191                 unsigned long flags;
1192
1193                 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
1194                 dd->ipath_sendctrl &= ~INFINIPATH_S_PIOINTBUFAVAIL;
1195                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1196                                  dd->ipath_sendctrl);
1197                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1198                 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
1199
1200                 handle_layer_pioavail(dd);
1201         }
1202
1203         ret = IRQ_HANDLED;
1204
1205 bail:
1206         return ret;
1207 }