git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_intr.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. 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_layer.h"
  38 #include "ipath_common.h"
  39
  40 /* These are all rcv-related errors which we want to count for stats */
  41 #define E_SUM_PKTERRS \
  42         (INFINIPATH_E_RHDRLEN | INFINIPATH_E_RBADTID | \
  43          INFINIPATH_E_RBADVERSION | INFINIPATH_E_RHDR | \
  44          INFINIPATH_E_RLONGPKTLEN | INFINIPATH_E_RSHORTPKTLEN | \
  45          INFINIPATH_E_RMAXPKTLEN | INFINIPATH_E_RMINPKTLEN | \
  46          INFINIPATH_E_RFORMATERR | INFINIPATH_E_RUNSUPVL | \
  47          INFINIPATH_E_RUNEXPCHAR | INFINIPATH_E_REBP)
  48
  49 /* These are all send-related errors which we want to count for stats */
  50 #define E_SUM_ERRS \
  51         (INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SUNEXPERRPKTNUM | \
  52          INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
  53          INFINIPATH_E_SMAXPKTLEN | INFINIPATH_E_SUNSUPVL | \
  54          INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
  55          INFINIPATH_E_INVALIDADDR)
  56
  57 /*
  58  * these are errors that can occur when the link changes state while
  59  * a packet is being sent or received.  This doesn't cover things
  60  * like EBP or VCRC that can be the result of a sending having the
  61  * link change state, so we receive a "known bad" packet.
  62  */
  63 #define E_SUM_LINK_PKTERRS \
  64         (INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
  65          INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
  66          INFINIPATH_E_RSHORTPKTLEN | INFINIPATH_E_RMINPKTLEN | \
  67          INFINIPATH_E_RUNEXPCHAR)
  68
  69 static u64 handle_e_sum_errs(struct ipath_devdata *dd, ipath_err_t errs)
  70 {
  71         unsigned long sbuf[4];
  72         u64 ignore_this_time = 0;
  73         u32 piobcnt;
  74
  75         /* if possible that sendbuffererror could be valid */
  76         piobcnt = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
  77         /* read these before writing errorclear */
  78         sbuf[0] = ipath_read_kreg64(
  79                 dd, dd->ipath_kregs->kr_sendbuffererror);
  80         sbuf[1] = ipath_read_kreg64(
  81                 dd, dd->ipath_kregs->kr_sendbuffererror + 1);
  82         if (piobcnt > 128) {
  83                 sbuf[2] = ipath_read_kreg64(
  84                         dd, dd->ipath_kregs->kr_sendbuffererror + 2);
  85                 sbuf[3] = ipath_read_kreg64(
  86                         dd, dd->ipath_kregs->kr_sendbuffererror + 3);
  87         }
  88
  89         if (sbuf[0] || sbuf[1] || (piobcnt > 128 && (sbuf[2] || sbuf[3]))) {
  90                 int i;
  91
  92                 ipath_cdbg(PKT, "SendbufErrs %lx %lx ", sbuf[0], sbuf[1]);
  93                 if (ipath_debug & __IPATH_PKTDBG && piobcnt > 128)
  94                         printk("%lx %lx ", sbuf[2], sbuf[3]);
  95                 for (i = 0; i < piobcnt; i++) {
  96                         if (test_bit(i, sbuf)) {
  97                                 u32 __iomem *piobuf;
  98                                 if (i < dd->ipath_piobcnt2k)
  99                                         piobuf = (u32 __iomem *)
 100                                                 (dd->ipath_pio2kbase +
 101                                                  i * dd->ipath_palign);
 102                                 else
 103                                         piobuf = (u32 __iomem *)
 104                                                 (dd->ipath_pio4kbase +
 105                                                  (i - dd->ipath_piobcnt2k) *
 106                                                  dd->ipath_4kalign);
 107
 108                                 ipath_cdbg(PKT,
 109                                            "PIObuf[%u] @%p pbc is %x; ",
 110                                            i, piobuf, readl(piobuf));
 111
 112                                 ipath_disarm_piobufs(dd, i, 1);
 113                         }
 114                 }
 115                 if (ipath_debug & __IPATH_PKTDBG)
 116                         printk("\n");
 117         }
 118         if ((errs & E_SUM_LINK_PKTERRS) &&
 119             !(dd->ipath_flags & IPATH_LINKACTIVE)) {
 120                 /*
 121                  * This can happen when SMA is trying to bring the link
 122                  * up, but the IB link changes state at the "wrong" time.
 123                  * The IB logic then complains that the packet isn't
 124                  * valid.  We don't want to confuse people, so we just
 125                  * don't print them, except at debug
 126                  */
 127                 ipath_dbg("Ignoring packet errors %llx, because link not "
 128                           "ACTIVE\n", (unsigned long long) errs);
 129                 ignore_this_time = errs & E_SUM_LINK_PKTERRS;
 130         }
 131
 132         return ignore_this_time;
 133 }
 134
 135 /* return the strings for the most common link states */
 136 static char *ib_linkstate(u32 linkstate)
 137 {
 138         char *ret;
 139
 140         switch (linkstate) {
 141         case IPATH_IBSTATE_INIT:
 142                 ret = "Init";
 143                 break;
 144         case IPATH_IBSTATE_ARM:
 145                 ret = "Arm";
 146                 break;
 147         case IPATH_IBSTATE_ACTIVE:
 148                 ret = "Active";
 149                 break;
 150         default:
 151                 ret = "Down";
 152         }
 153
 154         return ret;
 155 }
 156
 157 static void handle_e_ibstatuschanged(struct ipath_devdata *dd,
 158                                      ipath_err_t errs, int noprint)
 159 {
 160         u64 val;
 161         u32 ltstate, lstate;
 162
 163         /*
 164          * even if diags are enabled, we want to notice LINKINIT, etc.
 165          * We just don't want to change the LED state, or
 166          * dd->ipath_kregs->kr_ibcctrl
 167          */
 168         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
 169         lstate = val & IPATH_IBSTATE_MASK;
 170
 171         /*
 172          * this is confusing enough when it happens that I want to always put it
 173          * on the console and in the logs.  If it was a requested state change,
 174          * we'll have already cleared the flags, so we won't print this warning
 175          */
 176         if ((lstate != IPATH_IBSTATE_ARM && lstate != IPATH_IBSTATE_ACTIVE)
 177                 && (dd->ipath_flags & (IPATH_LINKARMED | IPATH_LINKACTIVE))) {
 178                 dev_info(&dd->pcidev->dev, "Link state changed from %s to %s\n",
 179                                  (dd->ipath_flags & IPATH_LINKARMED) ? "ARM" : "ACTIVE",
 180                                  ib_linkstate(lstate));
 181                 /*
 182                  * Flush all queued sends when link went to DOWN or INIT,
 183                  * to be sure that they don't block SMA and other MAD packets
 184                  */
 185                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 186                                  INFINIPATH_S_ABORT);
 187                 ipath_disarm_piobufs(dd, dd->ipath_lastport_piobuf,
 188                                                         (unsigned)(dd->ipath_piobcnt2k +
 189                                         dd->ipath_piobcnt4k) -
 190                                         dd->ipath_lastport_piobuf);
 191         }
 192         else if (lstate == IPATH_IBSTATE_INIT || lstate == IPATH_IBSTATE_ARM ||
 193             lstate == IPATH_IBSTATE_ACTIVE) {
 194                 /*
 195                  * only print at SMA if there is a change, debug if not
 196                  * (sometimes we want to know that, usually not).
 197                  */
 198                 if (lstate == ((unsigned) dd->ipath_lastibcstat
 199                                & IPATH_IBSTATE_MASK)) {
 200                         ipath_dbg("Status change intr but no change (%s)\n",
 201                                   ib_linkstate(lstate));
 202                 }
 203                 else
 204                         ipath_cdbg(SMA, "Unit %u link state %s, last "
 205                                    "was %s\n", dd->ipath_unit,
 206                                    ib_linkstate(lstate),
 207                                    ib_linkstate((unsigned)
 208                                                 dd->ipath_lastibcstat
 209                                                 & IPATH_IBSTATE_MASK));
 210         }
 211         else {
 212                 lstate = dd->ipath_lastibcstat & IPATH_IBSTATE_MASK;
 213                 if (lstate == IPATH_IBSTATE_INIT ||
 214                     lstate == IPATH_IBSTATE_ARM ||
 215                     lstate == IPATH_IBSTATE_ACTIVE)
 216                         ipath_cdbg(SMA, "Unit %u link state down"
 217                                    " (state 0x%x), from %s\n",
 218                                    dd->ipath_unit,
 219                                    (u32)val & IPATH_IBSTATE_MASK,
 220                                    ib_linkstate(lstate));
 221                 else
 222                         ipath_cdbg(VERBOSE, "Unit %u link state changed "
 223                                    "to 0x%x from down (%x)\n",
 224                                    dd->ipath_unit, (u32) val, lstate);
 225         }
 226         ltstate = (val >> INFINIPATH_IBCS_LINKTRAININGSTATE_SHIFT) &
 227                 INFINIPATH_IBCS_LINKTRAININGSTATE_MASK;
 228         lstate = (val >> INFINIPATH_IBCS_LINKSTATE_SHIFT) &
 229                 INFINIPATH_IBCS_LINKSTATE_MASK;
 230
 231         if (ltstate == INFINIPATH_IBCS_LT_STATE_POLLACTIVE ||
 232             ltstate == INFINIPATH_IBCS_LT_STATE_POLLQUIET) {
 233                 u32 last_ltstate;
 234
 235                 /*
 236                  * Ignore cycling back and forth from Polling.Active
 237                  * to Polling.Quiet while waiting for the other end of
 238                  * the link to come up. We will cycle back and forth
 239                  * between them if no cable is plugged in,
 240                  * the other device is powered off or disabled, etc.
 241                  */
 242                 last_ltstate = (dd->ipath_lastibcstat >>
 243                                 INFINIPATH_IBCS_LINKTRAININGSTATE_SHIFT)
 244                         & INFINIPATH_IBCS_LINKTRAININGSTATE_MASK;
 245                 if (last_ltstate == INFINIPATH_IBCS_LT_STATE_POLLACTIVE
 246                     || last_ltstate ==
 247                     INFINIPATH_IBCS_LT_STATE_POLLQUIET) {
 248                         if (dd->ipath_ibpollcnt > 40) {
 249                                 dd->ipath_flags |= IPATH_NOCABLE;
 250                                 *dd->ipath_statusp |=
 251                                         IPATH_STATUS_IB_NOCABLE;
 252                         } else
 253                                 dd->ipath_ibpollcnt++;
 254                         goto skip_ibchange;
 255                 }
 256         }
 257         dd->ipath_ibpollcnt = 0;        /* some state other than 2 or 3 */
 258         ipath_stats.sps_iblink++;
 259         if (ltstate != INFINIPATH_IBCS_LT_STATE_LINKUP) {
 260                 dd->ipath_flags |= IPATH_LINKDOWN;
 261                 dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
 262                                      | IPATH_LINKACTIVE |
 263                                      IPATH_LINKARMED);
 264                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
 265                 dd->ipath_lli_counter = 0;
 266                 if (!noprint) {
 267                         if (((dd->ipath_lastibcstat >>
 268                               INFINIPATH_IBCS_LINKSTATE_SHIFT) &
 269                              INFINIPATH_IBCS_LINKSTATE_MASK)
 270                             == INFINIPATH_IBCS_L_STATE_ACTIVE)
 271                                 /* if from up to down be more vocal */
 272                                 ipath_cdbg(SMA,
 273                                            "Unit %u link now down (%s)\n",
 274                                            dd->ipath_unit,
 275                                            ipath_ibcstatus_str[ltstate]);
 276                         else
 277                                 ipath_cdbg(VERBOSE, "Unit %u link is "
 278                                            "down (%s)\n", dd->ipath_unit,
 279                                            ipath_ibcstatus_str[ltstate]);
 280                 }
 281
 282                 dd->ipath_f_setextled(dd, lstate, ltstate);
 283         } else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_ACTIVE) {
 284                 dd->ipath_flags |= IPATH_LINKACTIVE;
 285                 dd->ipath_flags &=
 286                         ~(IPATH_LINKUNK | IPATH_LINKINIT | IPATH_LINKDOWN |
 287                           IPATH_LINKARMED | IPATH_NOCABLE);
 288                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_NOCABLE;
 289                 *dd->ipath_statusp |=
 290                         IPATH_STATUS_IB_READY | IPATH_STATUS_IB_CONF;
 291                 dd->ipath_f_setextled(dd, lstate, ltstate);
 292
 293                 __ipath_layer_intr(dd, IPATH_LAYER_INT_IF_UP);
 294         } else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_INIT) {
 295                 /*
 296                  * set INIT and DOWN.  Down is checked by most of the other
 297                  * code, but INIT is useful to know in a few places.
 298                  */
 299                 dd->ipath_flags |= IPATH_LINKINIT | IPATH_LINKDOWN;
 300                 dd->ipath_flags &=
 301                         ~(IPATH_LINKUNK | IPATH_LINKACTIVE | IPATH_LINKARMED
 302                           | IPATH_NOCABLE);
 303                 *dd->ipath_statusp &= ~(IPATH_STATUS_IB_NOCABLE
 304                                         | IPATH_STATUS_IB_READY);
 305                 dd->ipath_f_setextled(dd, lstate, ltstate);
 306         } else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_ARM) {
 307                 dd->ipath_flags |= IPATH_LINKARMED;
 308                 dd->ipath_flags &=
 309                         ~(IPATH_LINKUNK | IPATH_LINKDOWN | IPATH_LINKINIT |
 310                           IPATH_LINKACTIVE | IPATH_NOCABLE);
 311                 *dd->ipath_statusp &= ~(IPATH_STATUS_IB_NOCABLE
 312                                         | IPATH_STATUS_IB_READY);
 313                 dd->ipath_f_setextled(dd, lstate, ltstate);
 314         } else {
 315                 if (!noprint)
 316                         ipath_dbg("IBstatuschange unit %u: %s (%x)\n",
 317                                   dd->ipath_unit,
 318                                   ipath_ibcstatus_str[ltstate], ltstate);
 319         }
 320 skip_ibchange:
 321         dd->ipath_lastibcstat = val;
 322 }
 323
 324 static void handle_supp_msgs(struct ipath_devdata *dd,
 325                              unsigned supp_msgs, char msg[512])
 326 {
 327         /*
 328          * Print the message unless it's ibc status change only, which
 329          * happens so often we never want to count it.
 330          */
 331         if (dd->ipath_lasterror & ~INFINIPATH_E_IBSTATUSCHANGED) {
 332                 ipath_decode_err(msg, sizeof msg, dd->ipath_lasterror &
 333                                  ~INFINIPATH_E_IBSTATUSCHANGED);
 334                 if (dd->ipath_lasterror &
 335                     ~(INFINIPATH_E_RRCVEGRFULL | INFINIPATH_E_RRCVHDRFULL))
 336                         ipath_dev_err(dd, "Suppressed %u messages for "
 337                                       "fast-repeating errors (%s) (%llx)\n",
 338                                       supp_msgs, msg,
 339                                       (unsigned long long)
 340                                       dd->ipath_lasterror);
 341                 else {
 342                         /*
 343                          * rcvegrfull and rcvhdrqfull are "normal", for some
 344                          * types of processes (mostly benchmarks) that send
 345                          * huge numbers of messages, while not processing
 346                          * them. So only complain about these at debug
 347                          * level.
 348                          */
 349                         ipath_dbg("Suppressed %u messages for %s\n",
 350                                   supp_msgs, msg);
 351                 }
 352         }
 353 }
 354
 355 static unsigned handle_frequent_errors(struct ipath_devdata *dd,
 356                                        ipath_err_t errs, char msg[512],
 357                                        int *noprint)
 358 {
 359         unsigned long nc;
 360         static unsigned long nextmsg_time;
 361         static unsigned nmsgs, supp_msgs;
 362
 363         /*
 364          * Throttle back "fast" messages to no more than 10 per 5 seconds.
 365          * This isn't perfect, but it's a reasonable heuristic. If we get
 366          * more than 10, give a 6x longer delay.
 367          */
 368         nc = jiffies;
 369         if (nmsgs > 10) {
 370                 if (time_before(nc, nextmsg_time)) {
 371                         *noprint = 1;
 372                         if (!supp_msgs++)
 373                                 nextmsg_time = nc + HZ * 3;
 374                 }
 375                 else if (supp_msgs) {
 376                         handle_supp_msgs(dd, supp_msgs, msg);
 377                         supp_msgs = 0;
 378                         nmsgs = 0;
 379                 }
 380         }
 381         else if (!nmsgs++ || time_after(nc, nextmsg_time))
 382                 nextmsg_time = nc + HZ / 2;
 383
 384         return supp_msgs;
 385 }
 386
 387 static int handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
 388 {
 389         char msg[512];
 390         u64 ignore_this_time = 0;
 391         int i;
 392         int chkerrpkts = 0, noprint = 0;
 393         unsigned supp_msgs;
 394
 395         supp_msgs = handle_frequent_errors(dd, errs, msg, &noprint);
 396
 397         /*
 398          * don't report errors that are masked (includes those always
 399          * ignored)
 400          */
 401         errs &= ~dd->ipath_maskederrs;
 402
 403         /* do these first, they are most important */
 404         if (errs & INFINIPATH_E_HARDWARE) {
 405                 /* reuse same msg buf */
 406                 dd->ipath_f_handle_hwerrors(dd, msg, sizeof msg);
 407         }
 408
 409         if (!noprint && (errs & ~infinipath_e_bitsextant))
 410                 ipath_dev_err(dd, "error interrupt with unknown errors "
 411                               "%llx set\n", (unsigned long long)
 412                               (errs & ~infinipath_e_bitsextant));
 413
 414         if (errs & E_SUM_ERRS)
 415                 ignore_this_time = handle_e_sum_errs(dd, errs);
 416         else if ((errs & E_SUM_LINK_PKTERRS) &&
 417             !(dd->ipath_flags & IPATH_LINKACTIVE)) {
 418                 /*
 419                  * This can happen when SMA is trying to bring the link
 420                  * up, but the IB link changes state at the "wrong" time.
 421                  * The IB logic then complains that the packet isn't
 422                  * valid.  We don't want to confuse people, so we just
 423                  * don't print them, except at debug
 424                  */
 425                 ipath_dbg("Ignoring packet errors %llx, because link not "
 426                           "ACTIVE\n", (unsigned long long) errs);
 427                 ignore_this_time = errs & E_SUM_LINK_PKTERRS;
 428         }
 429
 430         if (supp_msgs == 250000) {
 431                 /*
 432                  * It's not entirely reasonable assuming that the errors set
 433                  * in the last clear period are all responsible for the
 434                  * problem, but the alternative is to assume it's the only
 435                  * ones on this particular interrupt, which also isn't great
 436                  */
 437                 dd->ipath_maskederrs |= dd->ipath_lasterror | errs;
 438                 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 439                                  ~dd->ipath_maskederrs);
 440                 ipath_decode_err(msg, sizeof msg,
 441                                  (dd->ipath_maskederrs & ~dd->
 442                                   ipath_ignorederrs));
 443
 444                 if ((dd->ipath_maskederrs & ~dd->ipath_ignorederrs) &
 445                     ~(INFINIPATH_E_RRCVEGRFULL | INFINIPATH_E_RRCVHDRFULL))
 446                         ipath_dev_err(dd, "Disabling error(s) %llx because "
 447                                       "occurring too frequently (%s)\n",
 448                                       (unsigned long long)
 449                                       (dd->ipath_maskederrs &
 450                                        ~dd->ipath_ignorederrs), msg);
 451                 else {
 452                         /*
 453                          * rcvegrfull and rcvhdrqfull are "normal",
 454                          * for some types of processes (mostly benchmarks)
 455                          * that send huge numbers of messages, while not
 456                          * processing them.  So only complain about
 457                          * these at debug level.
 458                          */
 459                         ipath_dbg("Disabling frequent queue full errors "
 460                                   "(%s)\n", msg);
 461                 }
 462
 463                 /*
 464                  * Re-enable the masked errors after around 3 minutes.  in
 465                  * ipath_get_faststats().  If we have a series of fast
 466                  * repeating but different errors, the interval will keep
 467                  * stretching out, but that's OK, as that's pretty
 468                  * catastrophic.
 469                  */
 470                 dd->ipath_unmasktime = jiffies + HZ * 180;
 471         }
 472
 473         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, errs);
 474         if (ignore_this_time)
 475                 errs &= ~ignore_this_time;
 476         if (errs & ~dd->ipath_lasterror) {
 477                 errs &= ~dd->ipath_lasterror;
 478                 /* never suppress duplicate hwerrors or ibstatuschange */
 479                 dd->ipath_lasterror |= errs &
 480                         ~(INFINIPATH_E_HARDWARE |
 481                           INFINIPATH_E_IBSTATUSCHANGED);
 482         }
 483         if (!errs)
 484                 return 0;
 485
 486         if (!noprint)
 487                 /*
 488                  * the ones we mask off are handled specially below or above
 489                  */
 490                 ipath_decode_err(msg, sizeof msg,
 491                                  errs & ~(INFINIPATH_E_IBSTATUSCHANGED |
 492                                           INFINIPATH_E_RRCVEGRFULL |
 493                                           INFINIPATH_E_RRCVHDRFULL |
 494                                           INFINIPATH_E_HARDWARE));
 495         else
 496                 /* so we don't need if (!noprint) at strlcat's below */
 497                 *msg = 0;
 498
 499         if (errs & E_SUM_PKTERRS) {
 500                 ipath_stats.sps_pkterrs++;
 501                 chkerrpkts = 1;
 502         }
 503         if (errs & E_SUM_ERRS)
 504                 ipath_stats.sps_errs++;
 505
 506         if (errs & (INFINIPATH_E_RICRC | INFINIPATH_E_RVCRC)) {
 507                 ipath_stats.sps_crcerrs++;
 508                 chkerrpkts = 1;
 509         }
 510
 511         /*
 512          * We don't want to print these two as they happen, or we can make
 513          * the situation even worse, because it takes so long to print
 514          * messages to serial consoles.  Kernel ports get printed from
 515          * fast_stats, no more than every 5 seconds, user ports get printed
 516          * on close
 517          */
 518         if (errs & INFINIPATH_E_RRCVHDRFULL) {
 519                 int any;
 520                 u32 hd, tl;
 521                 ipath_stats.sps_hdrqfull++;
 522                 for (any = i = 0; i < dd->ipath_cfgports; i++) {
 523                         struct ipath_portdata *pd = dd->ipath_pd[i];
 524                         if (i == 0) {
 525                                 hd = dd->ipath_port0head;
 526                                 tl = (u32) le64_to_cpu(
 527                                         *dd->ipath_hdrqtailptr);
 528                         } else if (pd && pd->port_cnt &&
 529                                    pd->port_rcvhdrtail_kvaddr) {
 530                                 /*
 531                                  * don't report same point multiple times,
 532                                  * except kernel
 533                                  */
 534                                 tl = (u32) * pd->port_rcvhdrtail_kvaddr;
 535                                 if (tl == dd->ipath_lastrcvhdrqtails[i])
 536                                         continue;
 537                                 hd = ipath_read_ureg32(dd, ur_rcvhdrhead,
 538                                                        i);
 539                         } else
 540                                 continue;
 541                         if (hd == (tl + 1) ||
 542                             (!hd && tl == dd->ipath_hdrqlast)) {
 543                                 if (i == 0)
 544                                         chkerrpkts = 1;
 545                                 dd->ipath_lastrcvhdrqtails[i] = tl;
 546                                 pd->port_hdrqfull++;
 547                         }
 548                 }
 549         }
 550         if (errs & INFINIPATH_E_RRCVEGRFULL) {
 551                 /*
 552                  * since this is of less importance and not likely to
 553                  * happen without also getting hdrfull, only count
 554                  * occurrences; don't check each port (or even the kernel
 555                  * vs user)
 556                  */
 557                 ipath_stats.sps_etidfull++;
 558                 if (dd->ipath_port0head !=
 559                     (u32) le64_to_cpu(*dd->ipath_hdrqtailptr))
 560                         chkerrpkts = 1;
 561         }
 562
 563         /*
 564          * do this before IBSTATUSCHANGED, in case both bits set in a single
 565          * interrupt; we want the STATUSCHANGE to "win", so we do our
 566          * internal copy of state machine correctly
 567          */
 568         if (errs & INFINIPATH_E_RIBLOSTLINK) {
 569                 /*
 570                  * force through block below
 571                  */
 572                 errs |= INFINIPATH_E_IBSTATUSCHANGED;
 573                 ipath_stats.sps_iblink++;
 574                 dd->ipath_flags |= IPATH_LINKDOWN;
 575                 dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
 576                                      | IPATH_LINKARMED | IPATH_LINKACTIVE);
 577                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
 578                 if (!noprint) {
 579                         u64 st = ipath_read_kreg64(
 580                                 dd, dd->ipath_kregs->kr_ibcstatus);
 581
 582                         ipath_dbg("Lost link, link now down (%s)\n",
 583                                   ipath_ibcstatus_str[st & 0xf]);
 584                 }
 585         }
 586         if (errs & INFINIPATH_E_IBSTATUSCHANGED)
 587                 handle_e_ibstatuschanged(dd, errs, noprint);
 588
 589         if (errs & INFINIPATH_E_RESET) {
 590                 if (!noprint)
 591                         ipath_dev_err(dd, "Got reset, requires re-init "
 592                                       "(unload and reload driver)\n");
 593                 dd->ipath_flags &= ~IPATH_INITTED;      /* needs re-init */
 594                 /* mark as having had error */
 595                 *dd->ipath_statusp |= IPATH_STATUS_HWERROR;
 596                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_CONF;
 597         }
 598
 599         if (!noprint && *msg)
 600                 ipath_dev_err(dd, "%s error\n", msg);
 601         if (dd->ipath_sma_state_wanted & dd->ipath_flags) {
 602                 ipath_cdbg(VERBOSE, "sma wanted state %x, iflags now %x, "
 603                            "waking\n", dd->ipath_sma_state_wanted,
 604                            dd->ipath_flags);
 605                 wake_up_interruptible(&ipath_sma_state_wait);
 606         }
 607
 608         return chkerrpkts;
 609 }
 610
 611 /* this is separate to allow for better optimization of ipath_intr() */
 612
 613 static void ipath_bad_intr(struct ipath_devdata *dd, u32 * unexpectp)
 614 {
 615         /*
 616          * sometimes happen during driver init and unload, don't want
 617          * to process any interrupts at that point
 618          */
 619
 620         /* this is just a bandaid, not a fix, if something goes badly
 621          * wrong */
 622         if (++*unexpectp > 100) {
 623                 if (++*unexpectp > 105) {
 624                         /*
 625                          * ok, we must be taking somebody else's interrupts,
 626                          * due to a messed up mptable and/or PIRQ table, so
 627                          * unregister the interrupt.  We've seen this during
 628                          * linuxbios development work, and it may happen in
 629                          * the future again.
 630                          */
 631                         if (dd->pcidev && dd->pcidev->irq) {
 632                                 ipath_dev_err(dd, "Now %u unexpected "
 633                                               "interrupts, unregistering "
 634                                               "interrupt handler\n",
 635                                               *unexpectp);
 636                                 ipath_dbg("free_irq of irq %x\n",
 637                                           dd->pcidev->irq);
 638                                 free_irq(dd->pcidev->irq, dd);
 639                         }
 640                 }
 641                 if (ipath_read_kreg32(dd, dd->ipath_kregs->kr_intmask)) {
 642                         ipath_dev_err(dd, "%u unexpected interrupts, "
 643                                       "disabling interrupts completely\n",
 644                                       *unexpectp);
 645                         /*
 646                          * disable all interrupts, something is very wrong
 647                          */
 648                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
 649                                          0ULL);
 650                 }
 651         } else if (*unexpectp > 1)
 652                 ipath_dbg("Interrupt when not ready, should not happen, "
 653                           "ignoring\n");
 654 }
 655
 656 static void ipath_bad_regread(struct ipath_devdata *dd)
 657 {
 658         static int allbits;
 659
 660         /* separate routine, for better optimization of ipath_intr() */
 661
 662         /*
 663          * We print the message and disable interrupts, in hope of
 664          * having a better chance of debugging the problem.
 665          */
 666         ipath_dev_err(dd,
 667                       "Read of interrupt status failed (all bits set)\n");
 668         if (allbits++) {
 669                 /* disable all interrupts, something is very wrong */
 670                 ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
 671                 if (allbits == 2) {
 672                         ipath_dev_err(dd, "Still bad interrupt status, "
 673                                       "unregistering interrupt\n");
 674                         free_irq(dd->pcidev->irq, dd);
 675                 } else if (allbits > 2) {
 676                         if ((allbits % 10000) == 0)
 677                                 printk(".");
 678                 } else
 679                         ipath_dev_err(dd, "Disabling interrupts, "
 680                                       "multiple errors\n");
 681         }
 682 }
 683
 684 static void handle_port_pioavail(struct ipath_devdata *dd)
 685 {
 686         u32 i;
 687         /*
 688          * start from port 1, since for now port 0  is never using
 689          * wait_event for PIO
 690          */
 691         for (i = 1; dd->ipath_portpiowait && i < dd->ipath_cfgports; i++) {
 692                 struct ipath_portdata *pd = dd->ipath_pd[i];
 693
 694                 if (pd && pd->port_cnt &&
 695                     dd->ipath_portpiowait & (1U << i)) {
 696                         clear_bit(i, &dd->ipath_portpiowait);
 697                         if (test_bit(IPATH_PORT_WAITING_PIO,
 698                                      &pd->port_flag)) {
 699                                 clear_bit(IPATH_PORT_WAITING_PIO,
 700                                           &pd->port_flag);
 701                                 wake_up_interruptible(&pd->port_wait);
 702                         }
 703                 }
 704         }
 705 }
 706
 707 static void handle_layer_pioavail(struct ipath_devdata *dd)
 708 {
 709         int ret;
 710
 711         ret = __ipath_layer_intr(dd, IPATH_LAYER_INT_SEND_CONTINUE);
 712         if (ret > 0)
 713                 goto set;
 714
 715         ret = __ipath_verbs_piobufavail(dd);
 716         if (ret > 0)
 717                 goto set;
 718
 719         return;
 720 set:
 721         set_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl);
 722         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 723                          dd->ipath_sendctrl);
 724 }
 725
 726 /*
 727  * Handle receive interrupts for user ports; this means a user
 728  * process was waiting for a packet to arrive, and didn't want
 729  * to poll
 730  */
 731 static void handle_urcv(struct ipath_devdata *dd, u32 istat)
 732 {
 733         u64 portr;
 734         int i;
 735         int rcvdint = 0;
 736
 737         portr = ((istat >> INFINIPATH_I_RCVAVAIL_SHIFT) &
 738                  infinipath_i_rcvavail_mask)
 739                 | ((istat >> INFINIPATH_I_RCVURG_SHIFT) &
 740                    infinipath_i_rcvurg_mask);
 741         for (i = 1; i < dd->ipath_cfgports; i++) {
 742                 struct ipath_portdata *pd = dd->ipath_pd[i];
 743                 if (portr & (1 << i) && pd && pd->port_cnt &&
 744                         test_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag)) {
 745                         int rcbit;
 746                         clear_bit(IPATH_PORT_WAITING_RCV,
 747                                   &pd->port_flag);
 748                         rcbit = i + INFINIPATH_R_INTRAVAIL_SHIFT;
 749                         clear_bit(1UL << rcbit, &dd->ipath_rcvctrl);
 750                         wake_up_interruptible(&pd->port_wait);
 751                         rcvdint = 1;
 752                 }
 753         }
 754         if (rcvdint) {
 755                 /* only want to take one interrupt, so turn off the rcv
 756                  * interrupt for all the ports that we did the wakeup on
 757                  * (but never for kernel port)
 758                  */
 759                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 760                                  dd->ipath_rcvctrl);
 761         }
 762 }
 763
 764 irqreturn_t ipath_intr(int irq, void *data, struct pt_regs *regs)
 765 {
 766         struct ipath_devdata *dd = data;
 767         u32 istat, chk0rcv = 0;
 768         ipath_err_t estat = 0;
 769         irqreturn_t ret;
 770         u32 oldhead, curtail;
 771         static unsigned unexpected = 0;
 772         static const u32 port0rbits = (1U<<INFINIPATH_I_RCVAVAIL_SHIFT) |
 773                  (1U<<INFINIPATH_I_RCVURG_SHIFT);
 774
 775         ipath_stats.sps_ints++;
 776
 777         if (!(dd->ipath_flags & IPATH_PRESENT)) {
 778                 /*
 779                  * This return value is not great, but we do not want the
 780                  * interrupt core code to remove our interrupt handler
 781                  * because we don't appear to be handling an interrupt
 782                  * during a chip reset.
 783                  */
 784                 return IRQ_HANDLED;
 785         }
 786
 787         /*
 788          * this needs to be flags&initted, not statusp, so we keep
 789          * taking interrupts even after link goes down, etc.
 790          * Also, we *must* clear the interrupt at some point, or we won't
 791          * take it again, which can be real bad for errors, etc...
 792          */
 793
 794         if (!(dd->ipath_flags & IPATH_INITTED)) {
 795                 ipath_bad_intr(dd, &unexpected);
 796                 ret = IRQ_NONE;
 797                 goto bail;
 798         }
 799
 800         /*
 801          * We try to avoid reading the interrupt status register, since
 802          * that's a PIO read, and stalls the processor for up to about
 803          * ~0.25 usec. The idea is that if we processed a port0 packet,
 804          * we blindly clear the  port 0 receive interrupt bits, and nothing
 805          * else, then return.  If other interrupts are pending, the chip
 806          * will re-interrupt us as soon as we write the intclear register.
 807          * We then won't process any more kernel packets (if not the 2nd
 808          * time, then the 3rd or 4th) and we'll then handle the other
 809          * interrupts.   We clear the interrupts first so that we don't
 810          * lose intr for later packets that arrive while we are processing.
 811          */
 812         oldhead = dd->ipath_port0head;
 813         curtail = (u32)le64_to_cpu(*dd->ipath_hdrqtailptr);
 814         if (oldhead != curtail) {
 815                 if (dd->ipath_flags & IPATH_GPIO_INTR) {
 816                         ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
 817                                          (u64) (1 << 2));
 818                         istat = port0rbits | INFINIPATH_I_GPIO;
 819                 }
 820                 else
 821                         istat = port0rbits;
 822                 ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
 823                 ipath_kreceive(dd);
 824                 if (oldhead != dd->ipath_port0head) {
 825                         ipath_stats.sps_fastrcvint++;
 826                         goto done;
 827                 }
 828         }
 829
 830         istat = ipath_read_kreg32(dd, dd->ipath_kregs->kr_intstatus);
 831
 832         if (unlikely(!istat)) {
 833                 ipath_stats.sps_nullintr++;
 834                 ret = IRQ_NONE; /* not our interrupt, or already handled */
 835                 goto bail;
 836         }
 837         if (unlikely(istat == -1)) {
 838                 ipath_bad_regread(dd);
 839                 /* don't know if it was our interrupt or not */
 840                 ret = IRQ_NONE;
 841                 goto bail;
 842         }
 843
 844         if (unexpected)
 845                 unexpected = 0;
 846
 847         if (unlikely(istat & ~infinipath_i_bitsextant))
 848                 ipath_dev_err(dd,
 849                               "interrupt with unknown interrupts %x set\n",
 850                               istat & (u32) ~ infinipath_i_bitsextant);
 851         else
 852                 ipath_cdbg(VERBOSE, "intr stat=0x%x\n", istat);
 853
 854         if (unlikely(istat & INFINIPATH_I_ERROR)) {
 855                 ipath_stats.sps_errints++;
 856                 estat = ipath_read_kreg64(dd,
 857                                           dd->ipath_kregs->kr_errorstatus);
 858                 if (!estat)
 859                         dev_info(&dd->pcidev->dev, "error interrupt (%x), "
 860                                  "but no error bits set!\n", istat);
 861                 else if (estat == -1LL)
 862                         /*
 863                          * should we try clearing all, or hope next read
 864                          * works?
 865                          */
 866                         ipath_dev_err(dd, "Read of error status failed "
 867                                       "(all bits set); ignoring\n");
 868                 else
 869                         if (handle_errors(dd, estat))
 870                                 /* force calling ipath_kreceive() */
 871                                 chk0rcv = 1;
 872         }
 873
 874         if (istat & INFINIPATH_I_GPIO) {
 875                 /*
 876                  * Packets are available in the port 0 rcv queue.
 877                  * Eventually this needs to be generalized to check
 878                  * IPATH_GPIO_INTR, and the specific GPIO bit, if
 879                  * GPIO interrupts are used for anything else.
 880                  */
 881                 if (unlikely(!(dd->ipath_flags & IPATH_GPIO_INTR))) {
 882                         u32 gpiostatus;
 883                         gpiostatus = ipath_read_kreg32(
 884                                 dd, dd->ipath_kregs->kr_gpio_status);
 885                         ipath_dbg("Unexpected GPIO interrupt bits %x\n",
 886                                   gpiostatus);
 887                         ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
 888                                          gpiostatus);
 889                 }
 890                 else {
 891                         /* Clear GPIO status bit 2 */
 892                         ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
 893                                         (u64) (1 << 2));
 894                         chk0rcv = 1;
 895                 }
 896         }
 897         chk0rcv |= istat & port0rbits;
 898
 899         /*
 900          * Clear the interrupt bits we found set, unless they are receive
 901          * related, in which case we already cleared them above, and don't
 902          * want to clear them again, because we might lose an interrupt.
 903          * Clear it early, so we "know" know the chip will have seen this by
 904          * the time we process the queue, and will re-interrupt if necessary.
 905          * The processor itself won't take the interrupt again until we return.
 906          */
 907         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
 908
 909         /*
 910          * handle port0 receive  before checking for pio buffers available,
 911          * since receives can overflow; piobuf waiters can afford a few
 912          * extra cycles, since they were waiting anyway, and user's waiting
 913          * for receive are at the bottom.
 914          */
 915         if (chk0rcv) {
 916                 ipath_kreceive(dd);
 917                 istat &= ~port0rbits;
 918         }
 919
 920         if (istat & ((infinipath_i_rcvavail_mask <<
 921                       INFINIPATH_I_RCVAVAIL_SHIFT)
 922                      | (infinipath_i_rcvurg_mask <<
 923                         INFINIPATH_I_RCVURG_SHIFT)))
 924                 handle_urcv(dd, istat);
 925
 926         if (istat & INFINIPATH_I_SPIOBUFAVAIL) {
 927                 clear_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl);
 928                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 929                                  dd->ipath_sendctrl);
 930
 931                 if (dd->ipath_portpiowait)
 932                         handle_port_pioavail(dd);
 933
 934                 handle_layer_pioavail(dd);
 935         }
 936
 937 done:
 938         ret = IRQ_HANDLED;
 939
 940 bail:
 941         return ret;
 942 }