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 "ips_common.h"
  38 #include "ipath_layer.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                 if (!noprint) {
 266                         if (((dd->ipath_lastibcstat >>
 267                               INFINIPATH_IBCS_LINKSTATE_SHIFT) &
 268                              INFINIPATH_IBCS_LINKSTATE_MASK)
 269                             == INFINIPATH_IBCS_L_STATE_ACTIVE)
 270                                 /* if from up to down be more vocal */
 271                                 ipath_cdbg(SMA,
 272                                            "Unit %u link now down (%s)\n",
 273                                            dd->ipath_unit,
 274                                            ipath_ibcstatus_str[ltstate]);
 275                         else
 276                                 ipath_cdbg(VERBOSE, "Unit %u link is "
 277                                            "down (%s)\n", dd->ipath_unit,
 278                                            ipath_ibcstatus_str[ltstate]);
 279                 }
 280
 281                 dd->ipath_f_setextled(dd, lstate, ltstate);
 282         } else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_ACTIVE) {
 283                 dd->ipath_flags |= IPATH_LINKACTIVE;
 284                 dd->ipath_flags &=
 285                         ~(IPATH_LINKUNK | IPATH_LINKINIT | IPATH_LINKDOWN |
 286                           IPATH_LINKARMED | IPATH_NOCABLE);
 287                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_NOCABLE;
 288                 *dd->ipath_statusp |=
 289                         IPATH_STATUS_IB_READY | IPATH_STATUS_IB_CONF;
 290                 dd->ipath_f_setextled(dd, lstate, ltstate);
 291
 292                 __ipath_layer_intr(dd, IPATH_LAYER_INT_IF_UP);
 293         } else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_INIT) {
 294                 /*
 295                  * set INIT and DOWN.  Down is checked by most of the other
 296                  * code, but INIT is useful to know in a few places.
 297                  */
 298                 dd->ipath_flags |= IPATH_LINKINIT | IPATH_LINKDOWN;
 299                 dd->ipath_flags &=
 300                         ~(IPATH_LINKUNK | IPATH_LINKACTIVE | IPATH_LINKARMED
 301                           | IPATH_NOCABLE);
 302                 *dd->ipath_statusp &= ~(IPATH_STATUS_IB_NOCABLE
 303                                         | IPATH_STATUS_IB_READY);
 304                 dd->ipath_f_setextled(dd, lstate, ltstate);
 305         } else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_ARM) {
 306                 dd->ipath_flags |= IPATH_LINKARMED;
 307                 dd->ipath_flags &=
 308                         ~(IPATH_LINKUNK | IPATH_LINKDOWN | IPATH_LINKINIT |
 309                           IPATH_LINKACTIVE | IPATH_NOCABLE);
 310                 *dd->ipath_statusp &= ~(IPATH_STATUS_IB_NOCABLE
 311                                         | IPATH_STATUS_IB_READY);
 312                 dd->ipath_f_setextled(dd, lstate, ltstate);
 313         } else {
 314                 if (!noprint)
 315                         ipath_dbg("IBstatuschange unit %u: %s (%x)\n",
 316                                   dd->ipath_unit,
 317                                   ipath_ibcstatus_str[ltstate], ltstate);
 318         }
 319 skip_ibchange:
 320         dd->ipath_lastibcstat = val;
 321 }
 322
 323 static void handle_supp_msgs(struct ipath_devdata *dd,
 324                              unsigned supp_msgs, char msg[512])
 325 {
 326         /*
 327          * Print the message unless it's ibc status change only, which
 328          * happens so often we never want to count it.
 329          */
 330         if (dd->ipath_lasterror & ~INFINIPATH_E_IBSTATUSCHANGED) {
 331                 ipath_decode_err(msg, sizeof msg, dd->ipath_lasterror &
 332                                  ~INFINIPATH_E_IBSTATUSCHANGED);
 333                 if (dd->ipath_lasterror &
 334                     ~(INFINIPATH_E_RRCVEGRFULL | INFINIPATH_E_RRCVHDRFULL))
 335                         ipath_dev_err(dd, "Suppressed %u messages for "
 336                                       "fast-repeating errors (%s) (%llx)\n",
 337                                       supp_msgs, msg,
 338                                       (unsigned long long)
 339                                       dd->ipath_lasterror);
 340                 else {
 341                         /*
 342                          * rcvegrfull and rcvhdrqfull are "normal", for some
 343                          * types of processes (mostly benchmarks) that send
 344                          * huge numbers of messages, while not processing
 345                          * them. So only complain about these at debug
 346                          * level.
 347                          */
 348                         ipath_dbg("Suppressed %u messages for %s\n",
 349                                   supp_msgs, msg);
 350                 }
 351         }
 352 }
 353
 354 static unsigned handle_frequent_errors(struct ipath_devdata *dd,
 355                                        ipath_err_t errs, char msg[512],
 356                                        int *noprint)
 357 {
 358         unsigned long nc;
 359         static unsigned long nextmsg_time;
 360         static unsigned nmsgs, supp_msgs;
 361
 362         /*
 363          * Throttle back "fast" messages to no more than 10 per 5 seconds.
 364          * This isn't perfect, but it's a reasonable heuristic. If we get
 365          * more than 10, give a 6x longer delay.
 366          */
 367         nc = jiffies;
 368         if (nmsgs > 10) {
 369                 if (time_before(nc, nextmsg_time)) {
 370                         *noprint = 1;
 371                         if (!supp_msgs++)
 372                                 nextmsg_time = nc + HZ * 3;
 373                 }
 374                 else if (supp_msgs) {
 375                         handle_supp_msgs(dd, supp_msgs, msg);
 376                         supp_msgs = 0;
 377                         nmsgs = 0;
 378                 }
 379         }
 380         else if (!nmsgs++ || time_after(nc, nextmsg_time))
 381                 nextmsg_time = nc + HZ / 2;
 382
 383         return supp_msgs;
 384 }
 385
 386 static int handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
 387 {
 388         char msg[512];
 389         u64 ignore_this_time = 0;
 390         int i;
 391         int chkerrpkts = 0, noprint = 0;
 392         unsigned supp_msgs;
 393
 394         supp_msgs = handle_frequent_errors(dd, errs, msg, &noprint);
 395
 396         /*
 397          * don't report errors that are masked (includes those always
 398          * ignored)
 399          */
 400         errs &= ~dd->ipath_maskederrs;
 401
 402         /* do these first, they are most important */
 403         if (errs & INFINIPATH_E_HARDWARE) {
 404                 /* reuse same msg buf */
 405                 dd->ipath_f_handle_hwerrors(dd, msg, sizeof msg);
 406         }
 407
 408         if (!noprint && (errs & ~infinipath_e_bitsextant))
 409                 ipath_dev_err(dd, "error interrupt with unknown errors "
 410                               "%llx set\n", (unsigned long long)
 411                               (errs & ~infinipath_e_bitsextant));
 412
 413         if (errs & E_SUM_ERRS)
 414                 ignore_this_time = handle_e_sum_errs(dd, errs);
 415         else if ((errs & E_SUM_LINK_PKTERRS) &&
 416             !(dd->ipath_flags & IPATH_LINKACTIVE)) {
 417                 /*
 418                  * This can happen when SMA is trying to bring the link
 419                  * up, but the IB link changes state at the "wrong" time.
 420                  * The IB logic then complains that the packet isn't
 421                  * valid.  We don't want to confuse people, so we just
 422                  * don't print them, except at debug
 423                  */
 424                 ipath_dbg("Ignoring packet errors %llx, because link not "
 425                           "ACTIVE\n", (unsigned long long) errs);
 426                 ignore_this_time = errs & E_SUM_LINK_PKTERRS;
 427         }
 428
 429         if (supp_msgs == 250000) {
 430                 /*
 431                  * It's not entirely reasonable assuming that the errors set
 432                  * in the last clear period are all responsible for the
 433                  * problem, but the alternative is to assume it's the only
 434                  * ones on this particular interrupt, which also isn't great
 435                  */
 436                 dd->ipath_maskederrs |= dd->ipath_lasterror | errs;
 437                 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 438                                  ~dd->ipath_maskederrs);
 439                 ipath_decode_err(msg, sizeof msg,
 440                                  (dd->ipath_maskederrs & ~dd->
 441                                   ipath_ignorederrs));
 442
 443                 if ((dd->ipath_maskederrs & ~dd->ipath_ignorederrs) &
 444                     ~(INFINIPATH_E_RRCVEGRFULL | INFINIPATH_E_RRCVHDRFULL))
 445                         ipath_dev_err(dd, "Disabling error(s) %llx because "
 446                                       "occurring too frequently (%s)\n",
 447                                       (unsigned long long)
 448                                       (dd->ipath_maskederrs &
 449                                        ~dd->ipath_ignorederrs), msg);
 450                 else {
 451                         /*
 452                          * rcvegrfull and rcvhdrqfull are "normal",
 453                          * for some types of processes (mostly benchmarks)
 454                          * that send huge numbers of messages, while not
 455                          * processing them.  So only complain about
 456                          * these at debug level.
 457                          */
 458                         ipath_dbg("Disabling frequent queue full errors "
 459                                   "(%s)\n", msg);
 460                 }
 461
 462                 /*
 463                  * Re-enable the masked errors after around 3 minutes.  in
 464                  * ipath_get_faststats().  If we have a series of fast
 465                  * repeating but different errors, the interval will keep
 466                  * stretching out, but that's OK, as that's pretty
 467                  * catastrophic.
 468                  */
 469                 dd->ipath_unmasktime = jiffies + HZ * 180;
 470         }
 471
 472         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, errs);
 473         if (ignore_this_time)
 474                 errs &= ~ignore_this_time;
 475         if (errs & ~dd->ipath_lasterror) {
 476                 errs &= ~dd->ipath_lasterror;
 477                 /* never suppress duplicate hwerrors or ibstatuschange */
 478                 dd->ipath_lasterror |= errs &
 479                         ~(INFINIPATH_E_HARDWARE |
 480                           INFINIPATH_E_IBSTATUSCHANGED);
 481         }
 482         if (!errs)
 483                 return 0;
 484
 485         if (!noprint)
 486                 /*
 487                  * the ones we mask off are handled specially below or above
 488                  */
 489                 ipath_decode_err(msg, sizeof msg,
 490                                  errs & ~(INFINIPATH_E_IBSTATUSCHANGED |
 491                                           INFINIPATH_E_RRCVEGRFULL |
 492                                           INFINIPATH_E_RRCVHDRFULL |
 493                                           INFINIPATH_E_HARDWARE));
 494         else
 495                 /* so we don't need if (!noprint) at strlcat's below */
 496                 *msg = 0;
 497
 498         if (errs & E_SUM_PKTERRS) {
 499                 ipath_stats.sps_pkterrs++;
 500                 chkerrpkts = 1;
 501         }
 502         if (errs & E_SUM_ERRS)
 503                 ipath_stats.sps_errs++;
 504
 505         if (errs & (INFINIPATH_E_RICRC | INFINIPATH_E_RVCRC)) {
 506                 ipath_stats.sps_crcerrs++;
 507                 chkerrpkts = 1;
 508         }
 509
 510         /*
 511          * We don't want to print these two as they happen, or we can make
 512          * the situation even worse, because it takes so long to print
 513          * messages to serial consoles.  Kernel ports get printed from
 514          * fast_stats, no more than every 5 seconds, user ports get printed
 515          * on close
 516          */
 517         if (errs & INFINIPATH_E_RRCVHDRFULL) {
 518                 int any;
 519                 u32 hd, tl;
 520                 ipath_stats.sps_hdrqfull++;
 521                 for (any = i = 0; i < dd->ipath_cfgports; i++) {
 522                         struct ipath_portdata *pd = dd->ipath_pd[i];
 523                         if (i == 0) {
 524                                 hd = dd->ipath_port0head;
 525                                 tl = (u32) le64_to_cpu(
 526                                         *dd->ipath_hdrqtailptr);
 527                         } else if (pd && pd->port_cnt &&
 528                                    pd->port_rcvhdrtail_kvaddr) {
 529                                 /*
 530                                  * don't report same point multiple times,
 531                                  * except kernel
 532                                  */
 533                                 tl = (u32) * pd->port_rcvhdrtail_kvaddr;
 534                                 if (tl == dd->ipath_lastrcvhdrqtails[i])
 535                                         continue;
 536                                 hd = ipath_read_ureg32(dd, ur_rcvhdrhead,
 537                                                        i);
 538                         } else
 539                                 continue;
 540                         if (hd == (tl + 1) ||
 541                             (!hd && tl == dd->ipath_hdrqlast)) {
 542                                 if (i == 0)
 543                                         chkerrpkts = 1;
 544                                 dd->ipath_lastrcvhdrqtails[i] = tl;
 545                                 pd->port_hdrqfull++;
 546                         }
 547                 }
 548         }
 549         if (errs & INFINIPATH_E_RRCVEGRFULL) {
 550                 /*
 551                  * since this is of less importance and not likely to
 552                  * happen without also getting hdrfull, only count
 553                  * occurrences; don't check each port (or even the kernel
 554                  * vs user)
 555                  */
 556                 ipath_stats.sps_etidfull++;
 557                 if (dd->ipath_port0head !=
 558                     (u32) le64_to_cpu(*dd->ipath_hdrqtailptr))
 559                         chkerrpkts = 1;
 560         }
 561
 562         /*
 563          * do this before IBSTATUSCHANGED, in case both bits set in a single
 564          * interrupt; we want the STATUSCHANGE to "win", so we do our
 565          * internal copy of state machine correctly
 566          */
 567         if (errs & INFINIPATH_E_RIBLOSTLINK) {
 568                 /*
 569                  * force through block below
 570                  */
 571                 errs |= INFINIPATH_E_IBSTATUSCHANGED;
 572                 ipath_stats.sps_iblink++;
 573                 dd->ipath_flags |= IPATH_LINKDOWN;
 574                 dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
 575                                      | IPATH_LINKARMED | IPATH_LINKACTIVE);
 576                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
 577                 if (!noprint) {
 578                         u64 st = ipath_read_kreg64(
 579                                 dd, dd->ipath_kregs->kr_ibcstatus);
 580
 581                         ipath_dbg("Lost link, link now down (%s)\n",
 582                                   ipath_ibcstatus_str[st & 0xf]);
 583                 }
 584         }
 585         if (errs & INFINIPATH_E_IBSTATUSCHANGED)
 586                 handle_e_ibstatuschanged(dd, errs, noprint);
 587
 588         if (errs & INFINIPATH_E_RESET) {
 589                 if (!noprint)
 590                         ipath_dev_err(dd, "Got reset, requires re-init "
 591                                       "(unload and reload driver)\n");
 592                 dd->ipath_flags &= ~IPATH_INITTED;      /* needs re-init */
 593                 /* mark as having had error */
 594                 *dd->ipath_statusp |= IPATH_STATUS_HWERROR;
 595                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_CONF;
 596         }
 597
 598         if (!noprint && *msg)
 599                 ipath_dev_err(dd, "%s error\n", msg);
 600         if (dd->ipath_sma_state_wanted & dd->ipath_flags) {
 601                 ipath_cdbg(VERBOSE, "sma wanted state %x, iflags now %x, "
 602                            "waking\n", dd->ipath_sma_state_wanted,
 603                            dd->ipath_flags);
 604                 wake_up_interruptible(&ipath_sma_state_wait);
 605         }
 606
 607         return chkerrpkts;
 608 }
 609
 610 /* this is separate to allow for better optimization of ipath_intr() */
 611
 612 static void ipath_bad_intr(struct ipath_devdata *dd, u32 * unexpectp)
 613 {
 614         /*
 615          * sometimes happen during driver init and unload, don't want
 616          * to process any interrupts at that point
 617          */
 618
 619         /* this is just a bandaid, not a fix, if something goes badly
 620          * wrong */
 621         if (++*unexpectp > 100) {
 622                 if (++*unexpectp > 105) {
 623                         /*
 624                          * ok, we must be taking somebody else's interrupts,
 625                          * due to a messed up mptable and/or PIRQ table, so
 626                          * unregister the interrupt.  We've seen this during
 627                          * linuxbios development work, and it may happen in
 628                          * the future again.
 629                          */
 630                         if (dd->pcidev && dd->pcidev->irq) {
 631                                 ipath_dev_err(dd, "Now %u unexpected "
 632                                               "interrupts, unregistering "
 633                                               "interrupt handler\n",
 634                                               *unexpectp);
 635                                 ipath_dbg("free_irq of irq %x\n",
 636                                           dd->pcidev->irq);
 637                                 free_irq(dd->pcidev->irq, dd);
 638                         }
 639                 }
 640                 if (ipath_read_kreg32(dd, dd->ipath_kregs->kr_intmask)) {
 641                         ipath_dev_err(dd, "%u unexpected interrupts, "
 642                                       "disabling interrupts completely\n",
 643                                       *unexpectp);
 644                         /*
 645                          * disable all interrupts, something is very wrong
 646                          */
 647                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
 648                                          0ULL);
 649                 }
 650         } else if (*unexpectp > 1)
 651                 ipath_dbg("Interrupt when not ready, should not happen, "
 652                           "ignoring\n");
 653 }
 654
 655 static void ipath_bad_regread(struct ipath_devdata *dd)
 656 {
 657         static int allbits;
 658
 659         /* separate routine, for better optimization of ipath_intr() */
 660
 661         /*
 662          * We print the message and disable interrupts, in hope of
 663          * having a better chance of debugging the problem.
 664          */
 665         ipath_dev_err(dd,
 666                       "Read of interrupt status failed (all bits set)\n");
 667         if (allbits++) {
 668                 /* disable all interrupts, something is very wrong */
 669                 ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
 670                 if (allbits == 2) {
 671                         ipath_dev_err(dd, "Still bad interrupt status, "
 672                                       "unregistering interrupt\n");
 673                         free_irq(dd->pcidev->irq, dd);
 674                 } else if (allbits > 2) {
 675                         if ((allbits % 10000) == 0)
 676                                 printk(".");
 677                 } else
 678                         ipath_dev_err(dd, "Disabling interrupts, "
 679                                       "multiple errors\n");
 680         }
 681 }
 682
 683 static void handle_port_pioavail(struct ipath_devdata *dd)
 684 {
 685         u32 i;
 686         /*
 687          * start from port 1, since for now port 0  is never using
 688          * wait_event for PIO
 689          */
 690         for (i = 1; dd->ipath_portpiowait && i < dd->ipath_cfgports; i++) {
 691                 struct ipath_portdata *pd = dd->ipath_pd[i];
 692
 693                 if (pd && pd->port_cnt &&
 694                     dd->ipath_portpiowait & (1U << i)) {
 695                         clear_bit(i, &dd->ipath_portpiowait);
 696                         if (test_bit(IPATH_PORT_WAITING_PIO,
 697                                      &pd->port_flag)) {
 698                                 clear_bit(IPATH_PORT_WAITING_PIO,
 699                                           &pd->port_flag);
 700                                 wake_up_interruptible(&pd->port_wait);
 701                         }
 702                 }
 703         }
 704 }
 705
 706 static void handle_layer_pioavail(struct ipath_devdata *dd)
 707 {
 708         int ret;
 709
 710         ret = __ipath_layer_intr(dd, IPATH_LAYER_INT_SEND_CONTINUE);
 711         if (ret > 0)
 712                 goto set;
 713
 714         ret = __ipath_verbs_piobufavail(dd);
 715         if (ret > 0)
 716                 goto set;
 717
 718         return;
 719 set:
 720         set_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl);
 721         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 722                          dd->ipath_sendctrl);
 723 }
 724
 725 /*
 726  * Handle receive interrupts for user ports; this means a user
 727  * process was waiting for a packet to arrive, and didn't want
 728  * to poll
 729  */
 730 static void handle_urcv(struct ipath_devdata *dd, u32 istat)
 731 {
 732         u64 portr;
 733         int i;
 734         int rcvdint = 0;
 735
 736         portr = ((istat >> INFINIPATH_I_RCVAVAIL_SHIFT) &
 737                  infinipath_i_rcvavail_mask)
 738                 | ((istat >> INFINIPATH_I_RCVURG_SHIFT) &
 739                    infinipath_i_rcvurg_mask);
 740         for (i = 1; i < dd->ipath_cfgports; i++) {
 741                 struct ipath_portdata *pd = dd->ipath_pd[i];
 742                 if (portr & (1 << i) && pd && pd->port_cnt &&
 743                         test_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag)) {
 744                         int rcbit;
 745                         clear_bit(IPATH_PORT_WAITING_RCV,
 746                                   &pd->port_flag);
 747                         rcbit = i + INFINIPATH_R_INTRAVAIL_SHIFT;
 748                         clear_bit(1UL << rcbit, &dd->ipath_rcvctrl);
 749                         wake_up_interruptible(&pd->port_wait);
 750                         rcvdint = 1;
 751                 }
 752         }
 753         if (rcvdint) {
 754                 /* only want to take one interrupt, so turn off the rcv
 755                  * interrupt for all the ports that we did the wakeup on
 756                  * (but never for kernel port)
 757                  */
 758                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 759                                  dd->ipath_rcvctrl);
 760         }
 761 }
 762
 763 irqreturn_t ipath_intr(int irq, void *data, struct pt_regs *regs)
 764 {
 765         struct ipath_devdata *dd = data;
 766         u32 istat, chk0rcv = 0;
 767         ipath_err_t estat = 0;
 768         irqreturn_t ret;
 769         u32 oldhead, curtail;
 770         static unsigned unexpected = 0;
 771         static const u32 port0rbits = (1U<<INFINIPATH_I_RCVAVAIL_SHIFT) |
 772                  (1U<<INFINIPATH_I_RCVURG_SHIFT);
 773
 774         ipath_stats.sps_ints++;
 775
 776         if (!(dd->ipath_flags & IPATH_PRESENT)) {
 777                 /*
 778                  * This return value is not great, but we do not want the
 779                  * interrupt core code to remove our interrupt handler
 780                  * because we don't appear to be handling an interrupt
 781                  * during a chip reset.
 782                  */
 783                 return IRQ_HANDLED;
 784         }
 785
 786         /*
 787          * this needs to be flags&initted, not statusp, so we keep
 788          * taking interrupts even after link goes down, etc.
 789          * Also, we *must* clear the interrupt at some point, or we won't
 790          * take it again, which can be real bad for errors, etc...
 791          */
 792
 793         if (!(dd->ipath_flags & IPATH_INITTED)) {
 794                 ipath_bad_intr(dd, &unexpected);
 795                 ret = IRQ_NONE;
 796                 goto bail;
 797         }
 798
 799         /*
 800          * We try to avoid reading the interrupt status register, since
 801          * that's a PIO read, and stalls the processor for up to about
 802          * ~0.25 usec. The idea is that if we processed a port0 packet,
 803          * we blindly clear the  port 0 receive interrupt bits, and nothing
 804          * else, then return.  If other interrupts are pending, the chip
 805          * will re-interrupt us as soon as we write the intclear register.
 806          * We then won't process any more kernel packets (if not the 2nd
 807          * time, then the 3rd or 4th) and we'll then handle the other
 808          * interrupts.   We clear the interrupts first so that we don't
 809          * lose intr for later packets that arrive while we are processing.
 810          */
 811         oldhead = dd->ipath_port0head;
 812         curtail = (u32)le64_to_cpu(*dd->ipath_hdrqtailptr);
 813         if (oldhead != curtail) {
 814                 if (dd->ipath_flags & IPATH_GPIO_INTR) {
 815                         ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
 816                                          (u64) (1 << 2));
 817                         istat = port0rbits | INFINIPATH_I_GPIO;
 818                 }
 819                 else
 820                         istat = port0rbits;
 821                 ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
 822                 ipath_kreceive(dd);
 823                 if (oldhead != dd->ipath_port0head) {
 824                         ipath_stats.sps_fastrcvint++;
 825                         goto done;
 826                 }
 827         }
 828
 829         istat = ipath_read_kreg32(dd, dd->ipath_kregs->kr_intstatus);
 830
 831         if (unlikely(!istat)) {
 832                 ipath_stats.sps_nullintr++;
 833                 ret = IRQ_NONE; /* not our interrupt, or already handled */
 834                 goto bail;
 835         }
 836         if (unlikely(istat == -1)) {
 837                 ipath_bad_regread(dd);
 838                 /* don't know if it was our interrupt or not */
 839                 ret = IRQ_NONE;
 840                 goto bail;
 841         }
 842
 843         if (unexpected)
 844                 unexpected = 0;
 845
 846         if (unlikely(istat & ~infinipath_i_bitsextant))
 847                 ipath_dev_err(dd,
 848                               "interrupt with unknown interrupts %x set\n",
 849                               istat & (u32) ~ infinipath_i_bitsextant);
 850         else
 851                 ipath_cdbg(VERBOSE, "intr stat=0x%x\n", istat);
 852
 853         if (unlikely(istat & INFINIPATH_I_ERROR)) {
 854                 ipath_stats.sps_errints++;
 855                 estat = ipath_read_kreg64(dd,
 856                                           dd->ipath_kregs->kr_errorstatus);
 857                 if (!estat)
 858                         dev_info(&dd->pcidev->dev, "error interrupt (%x), "
 859                                  "but no error bits set!\n", istat);
 860                 else if (estat == -1LL)
 861                         /*
 862                          * should we try clearing all, or hope next read
 863                          * works?
 864                          */
 865                         ipath_dev_err(dd, "Read of error status failed "
 866                                       "(all bits set); ignoring\n");
 867                 else
 868                         if (handle_errors(dd, estat))
 869                                 /* force calling ipath_kreceive() */
 870                                 chk0rcv = 1;
 871         }
 872
 873         if (istat & INFINIPATH_I_GPIO) {
 874                 /*
 875                  * Packets are available in the port 0 rcv queue.
 876                  * Eventually this needs to be generalized to check
 877                  * IPATH_GPIO_INTR, and the specific GPIO bit, if
 878                  * GPIO interrupts are used for anything else.
 879                  */
 880                 if (unlikely(!(dd->ipath_flags & IPATH_GPIO_INTR))) {
 881                         u32 gpiostatus;
 882                         gpiostatus = ipath_read_kreg32(
 883                                 dd, dd->ipath_kregs->kr_gpio_status);
 884                         ipath_dbg("Unexpected GPIO interrupt bits %x\n",
 885                                   gpiostatus);
 886                         ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
 887                                          gpiostatus);
 888                 }
 889                 else {
 890                         /* Clear GPIO status bit 2 */
 891                         ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
 892                                         (u64) (1 << 2));
 893                         chk0rcv = 1;
 894                 }
 895         }
 896         chk0rcv |= istat & port0rbits;
 897
 898         /*
 899          * Clear the interrupt bits we found set, unless they are receive
 900          * related, in which case we already cleared them above, and don't
 901          * want to clear them again, because we might lose an interrupt.
 902          * Clear it early, so we "know" know the chip will have seen this by
 903          * the time we process the queue, and will re-interrupt if necessary.
 904          * The processor itself won't take the interrupt again until we return.
 905          */
 906         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
 907
 908         /*
 909          * handle port0 receive  before checking for pio buffers available,
 910          * since receives can overflow; piobuf waiters can afford a few
 911          * extra cycles, since they were waiting anyway, and user's waiting
 912          * for receive are at the bottom.
 913          */
 914         if (chk0rcv) {
 915                 ipath_kreceive(dd);
 916                 istat &= ~port0rbits;
 917         }
 918
 919         if (istat & ((infinipath_i_rcvavail_mask <<
 920                       INFINIPATH_I_RCVAVAIL_SHIFT)
 921                      | (infinipath_i_rcvurg_mask <<
 922                         INFINIPATH_I_RCVURG_SHIFT)))
 923                 handle_urcv(dd, istat);
 924
 925         if (istat & INFINIPATH_I_SPIOBUFAVAIL) {
 926                 clear_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl);
 927                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 928                                  dd->ipath_sendctrl);
 929
 930                 if (dd->ipath_portpiowait)
 931                         handle_port_pioavail(dd);
 932
 933                 handle_layer_pioavail(dd);
 934         }
 935
 936 done:
 937         ret = IRQ_HANDLED;
 938
 939 bail:
 940         return ret;
 941 }