kernel/drivers/infiniband/hw/ipath/ipath_intr.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
   3  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/pci.h>
  35 #include <linux/delay.h>
  36 #include <linux/sched.h>
  37
  38 #include "ipath_kernel.h"
  39 #include "ipath_verbs.h"
  40 #include "ipath_common.h"
  41
  42
  43 /*
  44  * Called when we might have an error that is specific to a particular
  45  * PIO buffer, and may need to cancel that buffer, so it can be re-used.
  46  */
  47 void ipath_disarm_senderrbufs(struct ipath_devdata *dd)
  48 {
  49         u32 piobcnt;
  50         unsigned long sbuf[4];
  51         /*
  52          * it's possible that sendbuffererror could have bits set; might
  53          * have already done this as a result of hardware error handling
  54          */
  55         piobcnt = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
  56         /* read these before writing errorclear */
  57         sbuf[0] = ipath_read_kreg64(
  58                 dd, dd->ipath_kregs->kr_sendbuffererror);
  59         sbuf[1] = ipath_read_kreg64(
  60                 dd, dd->ipath_kregs->kr_sendbuffererror + 1);
  61         if (piobcnt > 128)
  62                 sbuf[2] = ipath_read_kreg64(
  63                         dd, dd->ipath_kregs->kr_sendbuffererror + 2);
  64         if (piobcnt > 192)
  65                 sbuf[3] = ipath_read_kreg64(
  66                         dd, dd->ipath_kregs->kr_sendbuffererror + 3);
  67         else
  68                 sbuf[3] = 0;
  69
  70         if (sbuf[0] || sbuf[1] || (piobcnt > 128 && (sbuf[2] || sbuf[3]))) {
  71                 int i;
  72                 if (ipath_debug & (__IPATH_PKTDBG|__IPATH_DBG) &&
  73                         time_after(dd->ipath_lastcancel, jiffies)) {
  74                         __IPATH_DBG_WHICH(__IPATH_PKTDBG|__IPATH_DBG,
  75                                           "SendbufErrs %lx %lx", sbuf[0],
  76                                           sbuf[1]);
  77                         if (ipath_debug & __IPATH_PKTDBG && piobcnt > 128)
  78                                 printk(" %lx %lx ", sbuf[2], sbuf[3]);
  79                         printk("\n");
  80                 }
  81
  82                 for (i = 0; i < piobcnt; i++)
  83                         if (test_bit(i, sbuf))
  84                                 ipath_disarm_piobufs(dd, i, 1);
  85                 /* ignore armlaunch errs for a bit */
  86                 dd->ipath_lastcancel = jiffies+3;
  87         }
  88 }
  89
  90
  91 /* These are all rcv-related errors which we want to count for stats */
  92 #define E_SUM_PKTERRS \
  93         (INFINIPATH_E_RHDRLEN | INFINIPATH_E_RBADTID | \
  94          INFINIPATH_E_RBADVERSION | INFINIPATH_E_RHDR | \
  95          INFINIPATH_E_RLONGPKTLEN | INFINIPATH_E_RSHORTPKTLEN | \
  96          INFINIPATH_E_RMAXPKTLEN | INFINIPATH_E_RMINPKTLEN | \
  97          INFINIPATH_E_RFORMATERR | INFINIPATH_E_RUNSUPVL | \
  98          INFINIPATH_E_RUNEXPCHAR | INFINIPATH_E_REBP)
  99
 100 /* These are all send-related errors which we want to count for stats */
 101 #define E_SUM_ERRS \
 102         (INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SUNEXPERRPKTNUM | \
 103          INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
 104          INFINIPATH_E_SMAXPKTLEN | INFINIPATH_E_SUNSUPVL | \
 105          INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
 106          INFINIPATH_E_INVALIDADDR)
 107
 108 /*
 109  * this is similar to E_SUM_ERRS, but can't ignore armlaunch, don't ignore
 110  * errors not related to freeze and cancelling buffers.  Can't ignore
 111  * armlaunch because could get more while still cleaning up, and need
 112  * to cancel those as they happen.
 113  */
 114 #define E_SPKT_ERRS_IGNORE \
 115          (INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
 116          INFINIPATH_E_SMAXPKTLEN | INFINIPATH_E_SMINPKTLEN | \
 117          INFINIPATH_E_SPKTLEN)
 118
 119 /*
 120  * these are errors that can occur when the link changes state while
 121  * a packet is being sent or received.  This doesn't cover things
 122  * like EBP or VCRC that can be the result of a sending having the
 123  * link change state, so we receive a "known bad" packet.
 124  */
 125 #define E_SUM_LINK_PKTERRS \
 126         (INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
 127          INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
 128          INFINIPATH_E_RSHORTPKTLEN | INFINIPATH_E_RMINPKTLEN | \
 129          INFINIPATH_E_RUNEXPCHAR)
 130
 131 static u64 handle_e_sum_errs(struct ipath_devdata *dd, ipath_err_t errs)
 132 {
 133         u64 ignore_this_time = 0;
 134
 135         ipath_disarm_senderrbufs(dd);
 136         if ((errs & E_SUM_LINK_PKTERRS) &&
 137             !(dd->ipath_flags & IPATH_LINKACTIVE)) {
 138                 /*
 139                  * This can happen when SMA is trying to bring the link
 140                  * up, but the IB link changes state at the "wrong" time.
 141                  * The IB logic then complains that the packet isn't
 142                  * valid.  We don't want to confuse people, so we just
 143                  * don't print them, except at debug
 144                  */
 145                 ipath_dbg("Ignoring packet errors %llx, because link not "
 146                           "ACTIVE\n", (unsigned long long) errs);
 147                 ignore_this_time = errs & E_SUM_LINK_PKTERRS;
 148         }
 149
 150         return ignore_this_time;
 151 }
 152
 153 /* generic hw error messages... */
 154 #define INFINIPATH_HWE_TXEMEMPARITYERR_MSG(a) \
 155         { \
 156                 .mask = ( INFINIPATH_HWE_TXEMEMPARITYERR_##a <<    \
 157                           INFINIPATH_HWE_TXEMEMPARITYERR_SHIFT ),   \
 158                 .msg = "TXE " #a " Memory Parity"            \
 159         }
 160 #define INFINIPATH_HWE_RXEMEMPARITYERR_MSG(a) \
 161         { \
 162                 .mask = ( INFINIPATH_HWE_RXEMEMPARITYERR_##a <<    \
 163                           INFINIPATH_HWE_RXEMEMPARITYERR_SHIFT ),   \
 164                 .msg = "RXE " #a " Memory Parity"            \
 165         }
 166
 167 static const struct ipath_hwerror_msgs ipath_generic_hwerror_msgs[] = {
 168         INFINIPATH_HWE_MSG(IBCBUSFRSPCPARITYERR, "IPATH2IB Parity"),
 169         INFINIPATH_HWE_MSG(IBCBUSTOSPCPARITYERR, "IB2IPATH Parity"),
 170
 171         INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOBUF),
 172         INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOPBC),
 173         INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOLAUNCHFIFO),
 174
 175         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(RCVBUF),
 176         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(LOOKUPQ),
 177         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(EAGERTID),
 178         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(EXPTID),
 179         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(FLAGBUF),
 180         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(DATAINFO),
 181         INFINIPATH_HWE_RXEMEMPARITYERR_MSG(HDRINFO),
 182 };
 183
 184 /**
 185  * ipath_format_hwmsg - format a single hwerror message
 186  * @msg message buffer
 187  * @msgl length of message buffer
 188  * @hwmsg message to add to message buffer
 189  */
 190 static void ipath_format_hwmsg(char *msg, size_t msgl, const char *hwmsg)
 191 {
 192         strlcat(msg, "[", msgl);
 193         strlcat(msg, hwmsg, msgl);
 194         strlcat(msg, "]", msgl);
 195 }
 196
 197 /**
 198  * ipath_format_hwerrors - format hardware error messages for display
 199  * @hwerrs hardware errors bit vector
 200  * @hwerrmsgs hardware error descriptions
 201  * @nhwerrmsgs number of hwerrmsgs
 202  * @msg message buffer
 203  * @msgl message buffer length
 204  */
 205 void ipath_format_hwerrors(u64 hwerrs,
 206                            const struct ipath_hwerror_msgs *hwerrmsgs,
 207                            size_t nhwerrmsgs,
 208                            char *msg, size_t msgl)
 209 {
 210         int i;
 211         const int glen =
 212             ARRAY_SIZE(ipath_generic_hwerror_msgs);
 213
 214         for (i=0; i<glen; i++) {
 215                 if (hwerrs & ipath_generic_hwerror_msgs[i].mask) {
 216                         ipath_format_hwmsg(msg, msgl,
 217                                            ipath_generic_hwerror_msgs[i].msg);
 218                 }
 219         }
 220
 221         for (i=0; i<nhwerrmsgs; i++) {
 222                 if (hwerrs & hwerrmsgs[i].mask) {
 223                         ipath_format_hwmsg(msg, msgl, hwerrmsgs[i].msg);
 224                 }
 225         }
 226 }
 227
 228 /* return the strings for the most common link states */
 229 static char *ib_linkstate(struct ipath_devdata *dd, u64 ibcs)
 230 {
 231         char *ret;
 232         u32 state;
 233
 234         state = ipath_ib_state(dd, ibcs);
 235         if (state == dd->ib_init)
 236                 ret = "Init";
 237         else if (state == dd->ib_arm)
 238                 ret = "Arm";
 239         else if (state == dd->ib_active)
 240                 ret = "Active";
 241         else
 242                 ret = "Down";
 243         return ret;
 244 }
 245
 246 void signal_ib_event(struct ipath_devdata *dd, enum ib_event_type ev)
 247 {
 248         struct ib_event event;
 249
 250         event.device = &dd->verbs_dev->ibdev;
 251         event.element.port_num = 1;
 252         event.event = ev;
 253         ib_dispatch_event(&event);
 254 }
 255
 256 static void handle_e_ibstatuschanged(struct ipath_devdata *dd,
 257                                      ipath_err_t errs)
 258 {
 259         u32 ltstate, lstate, ibstate, lastlstate;
 260         u32 init = dd->ib_init;
 261         u32 arm = dd->ib_arm;
 262         u32 active = dd->ib_active;
 263         const u64 ibcs = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
 264
 265         lstate = ipath_ib_linkstate(dd, ibcs); /* linkstate */
 266         ibstate = ipath_ib_state(dd, ibcs);
 267         /* linkstate at last interrupt */
 268         lastlstate = ipath_ib_linkstate(dd, dd->ipath_lastibcstat);
 269         ltstate = ipath_ib_linktrstate(dd, ibcs); /* linktrainingtate */
 270
 271         /*
 272          * Since going into a recovery state causes the link state to go
 273          * down and since recovery is transitory, it is better if we "miss"
 274          * ever seeing the link training state go into recovery (i.e.,
 275          * ignore this transition for link state special handling purposes)
 276          * without even updating ipath_lastibcstat.
 277          */
 278         if ((ltstate == INFINIPATH_IBCS_LT_STATE_RECOVERRETRAIN) ||
 279             (ltstate == INFINIPATH_IBCS_LT_STATE_RECOVERWAITRMT) ||
 280             (ltstate == INFINIPATH_IBCS_LT_STATE_RECOVERIDLE))
 281                 goto done;
 282
 283         /*
 284          * if linkstate transitions into INIT from any of the various down
 285          * states, or if it transitions from any of the up (INIT or better)
 286          * states into any of the down states (except link recovery), then
 287          * call the chip-specific code to take appropriate actions.
 288          */
 289         if (lstate >= INFINIPATH_IBCS_L_STATE_INIT &&
 290                 lastlstate == INFINIPATH_IBCS_L_STATE_DOWN) {
 291                 /* transitioned to UP */
 292                 if (dd->ipath_f_ib_updown(dd, 1, ibcs)) {
 293                         /* link came up, so we must no longer be disabled */
 294                         dd->ipath_flags &= ~IPATH_IB_LINK_DISABLED;
 295                         ipath_cdbg(LINKVERB, "LinkUp handled, skipped\n");
 296                         goto skip_ibchange; /* chip-code handled */
 297                 }
 298         } else if ((lastlstate >= INFINIPATH_IBCS_L_STATE_INIT ||
 299                 (dd->ipath_flags & IPATH_IB_FORCE_NOTIFY)) &&
 300                 ltstate <= INFINIPATH_IBCS_LT_STATE_CFGWAITRMT &&
 301                 ltstate != INFINIPATH_IBCS_LT_STATE_LINKUP) {
 302                 int handled;
 303                 handled = dd->ipath_f_ib_updown(dd, 0, ibcs);
 304                 dd->ipath_flags &= ~IPATH_IB_FORCE_NOTIFY;
 305                 if (handled) {
 306                         ipath_cdbg(LINKVERB, "LinkDown handled, skipped\n");
 307                         goto skip_ibchange; /* chip-code handled */
 308                 }
 309         }
 310
 311         /*
 312          * Significant enough to always print and get into logs, if it was
 313          * unexpected.  If it was a requested state change, we'll have
 314          * already cleared the flags, so we won't print this warning
 315          */
 316         if ((ibstate != arm && ibstate != active) &&
 317             (dd->ipath_flags & (IPATH_LINKARMED | IPATH_LINKACTIVE))) {
 318                 dev_info(&dd->pcidev->dev, "Link state changed from %s "
 319                          "to %s\n", (dd->ipath_flags & IPATH_LINKARMED) ?
 320                          "ARM" : "ACTIVE", ib_linkstate(dd, ibcs));
 321         }
 322
 323         if (ltstate == INFINIPATH_IBCS_LT_STATE_POLLACTIVE ||
 324             ltstate == INFINIPATH_IBCS_LT_STATE_POLLQUIET) {
 325                 u32 lastlts;
 326                 lastlts = ipath_ib_linktrstate(dd, dd->ipath_lastibcstat);
 327                 /*
 328                  * Ignore cycling back and forth from Polling.Active to
 329                  * Polling.Quiet while waiting for the other end of the link
 330                  * to come up, except to try and decide if we are connected
 331                  * to a live IB device or not.  We will cycle back and
 332                  * forth between them if no cable is plugged in, the other
 333                  * device is powered off or disabled, etc.
 334                  */
 335                 if (lastlts == INFINIPATH_IBCS_LT_STATE_POLLACTIVE ||
 336                     lastlts == INFINIPATH_IBCS_LT_STATE_POLLQUIET) {
 337                         if (!(dd->ipath_flags & IPATH_IB_AUTONEG_INPROG) &&
 338                              (++dd->ipath_ibpollcnt == 40)) {
 339                                 dd->ipath_flags |= IPATH_NOCABLE;
 340                                 *dd->ipath_statusp |=
 341                                         IPATH_STATUS_IB_NOCABLE;
 342                                 ipath_cdbg(LINKVERB, "Set NOCABLE\n");
 343                         }
 344                         ipath_cdbg(LINKVERB, "POLL change to %s (%x)\n",
 345                                 ipath_ibcstatus_str[ltstate], ibstate);
 346                         goto skip_ibchange;
 347                 }
 348         }
 349
 350         dd->ipath_ibpollcnt = 0; /* not poll*, now */
 351         ipath_stats.sps_iblink++;
 352
 353         if (ibstate != init && dd->ipath_lastlinkrecov && ipath_linkrecovery) {
 354                 u64 linkrecov;
 355                 linkrecov = ipath_snap_cntr(dd,
 356                         dd->ipath_cregs->cr_iblinkerrrecovcnt);
 357                 if (linkrecov != dd->ipath_lastlinkrecov) {
 358                         ipath_dbg("IB linkrecov up %Lx (%s %s) recov %Lu\n",
 359                                 (unsigned long long) ibcs,
 360                                 ib_linkstate(dd, ibcs),
 361                                 ipath_ibcstatus_str[ltstate],
 362                                 (unsigned long long) linkrecov);
 363                         /* and no more until active again */
 364                         dd->ipath_lastlinkrecov = 0;
 365                         ipath_set_linkstate(dd, IPATH_IB_LINKDOWN);
 366                         goto skip_ibchange;
 367                 }
 368         }
 369
 370         if (ibstate == init || ibstate == arm || ibstate == active) {
 371                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_NOCABLE;
 372                 if (ibstate == init || ibstate == arm) {
 373                         *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
 374                         if (dd->ipath_flags & IPATH_LINKACTIVE)
 375                                 signal_ib_event(dd, IB_EVENT_PORT_ERR);
 376                 }
 377                 if (ibstate == arm) {
 378                         dd->ipath_flags |= IPATH_LINKARMED;
 379                         dd->ipath_flags &= ~(IPATH_LINKUNK |
 380                                 IPATH_LINKINIT | IPATH_LINKDOWN |
 381                                 IPATH_LINKACTIVE | IPATH_NOCABLE);
 382                         ipath_hol_down(dd);
 383                 } else  if (ibstate == init) {
 384                         /*
 385                          * set INIT and DOWN.  Down is checked by
 386                          * most of the other code, but INIT is
 387                          * useful to know in a few places.
 388                          */
 389                         dd->ipath_flags |= IPATH_LINKINIT |
 390                                 IPATH_LINKDOWN;
 391                         dd->ipath_flags &= ~(IPATH_LINKUNK |
 392                                 IPATH_LINKARMED | IPATH_LINKACTIVE |
 393                                 IPATH_NOCABLE);
 394                         ipath_hol_down(dd);
 395                 } else {  /* active */
 396                         dd->ipath_lastlinkrecov = ipath_snap_cntr(dd,
 397                                 dd->ipath_cregs->cr_iblinkerrrecovcnt);
 398                         *dd->ipath_statusp |=
 399                                 IPATH_STATUS_IB_READY | IPATH_STATUS_IB_CONF;
 400                         dd->ipath_flags |= IPATH_LINKACTIVE;
 401                         dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
 402                                 | IPATH_LINKDOWN | IPATH_LINKARMED |
 403                                 IPATH_NOCABLE);
 404                         if (dd->ipath_flags & IPATH_HAS_SEND_DMA)
 405                                 ipath_restart_sdma(dd);
 406                         signal_ib_event(dd, IB_EVENT_PORT_ACTIVE);
 407                         /* LED active not handled in chip _f_updown */
 408                         dd->ipath_f_setextled(dd, lstate, ltstate);
 409                         ipath_hol_up(dd);
 410                 }
 411
 412                 /*
 413                  * print after we've already done the work, so as not to
 414                  * delay the state changes and notifications, for debugging
 415                  */
 416                 if (lstate == lastlstate)
 417                         ipath_cdbg(LINKVERB, "Unchanged from last: %s "
 418                                 "(%x)\n", ib_linkstate(dd, ibcs), ibstate);
 419                 else
 420                         ipath_cdbg(VERBOSE, "Unit %u: link up to %s %s (%x)\n",
 421                                   dd->ipath_unit, ib_linkstate(dd, ibcs),
 422                                   ipath_ibcstatus_str[ltstate],  ibstate);
 423         } else { /* down */
 424                 if (dd->ipath_flags & IPATH_LINKACTIVE)
 425                         signal_ib_event(dd, IB_EVENT_PORT_ERR);
 426                 dd->ipath_flags |= IPATH_LINKDOWN;
 427                 dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
 428                                      | IPATH_LINKACTIVE |
 429                                      IPATH_LINKARMED);
 430                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
 431                 dd->ipath_lli_counter = 0;
 432
 433                 if (lastlstate != INFINIPATH_IBCS_L_STATE_DOWN)
 434                         ipath_cdbg(VERBOSE, "Unit %u link state down "
 435                                    "(state 0x%x), from %s\n",
 436                                    dd->ipath_unit, lstate,
 437                                    ib_linkstate(dd, dd->ipath_lastibcstat));
 438                 else
 439                         ipath_cdbg(LINKVERB, "Unit %u link state changed "
 440                                    "to %s (0x%x) from down (%x)\n",
 441                                    dd->ipath_unit,
 442                                    ipath_ibcstatus_str[ltstate],
 443                                    ibstate, lastlstate);
 444         }
 445
 446 skip_ibchange:
 447         dd->ipath_lastibcstat = ibcs;
 448 done:
 449         return;
 450 }
 451
 452 static void handle_supp_msgs(struct ipath_devdata *dd,
 453                              unsigned supp_msgs, char *msg, u32 msgsz)
 454 {
 455         /*
 456          * Print the message unless it's ibc status change only, which
 457          * happens so often we never want to count it.
 458          */
 459         if (dd->ipath_lasterror & ~INFINIPATH_E_IBSTATUSCHANGED) {
 460                 int iserr;
 461                 ipath_err_t mask;
 462                 iserr = ipath_decode_err(dd, msg, msgsz,
 463                                          dd->ipath_lasterror &
 464                                          ~INFINIPATH_E_IBSTATUSCHANGED);
 465
 466                 mask = INFINIPATH_E_RRCVEGRFULL | INFINIPATH_E_RRCVHDRFULL |
 467                         INFINIPATH_E_PKTERRS | INFINIPATH_E_SDMADISABLED;
 468
 469                 /* if we're in debug, then don't mask SDMADISABLED msgs */
 470                 if (ipath_debug & __IPATH_DBG)
 471                         mask &= ~INFINIPATH_E_SDMADISABLED;
 472
 473                 if (dd->ipath_lasterror & ~mask)
 474                         ipath_dev_err(dd, "Suppressed %u messages for "
 475                                       "fast-repeating errors (%s) (%llx)\n",
 476                                       supp_msgs, msg,
 477                                       (unsigned long long)
 478                                       dd->ipath_lasterror);
 479                 else {
 480                         /*
 481                          * rcvegrfull and rcvhdrqfull are "normal", for some
 482                          * types of processes (mostly benchmarks) that send
 483                          * huge numbers of messages, while not processing
 484                          * them. So only complain about these at debug
 485                          * level.
 486                          */
 487                         if (iserr)
 488                                 ipath_dbg("Suppressed %u messages for %s\n",
 489                                           supp_msgs, msg);
 490                         else
 491                                 ipath_cdbg(ERRPKT,
 492                                         "Suppressed %u messages for %s\n",
 493                                           supp_msgs, msg);
 494                 }
 495         }
 496 }
 497
 498 static unsigned handle_frequent_errors(struct ipath_devdata *dd,
 499                                        ipath_err_t errs, char *msg,
 500                                        u32 msgsz, int *noprint)
 501 {
 502         unsigned long nc;
 503         static unsigned long nextmsg_time;
 504         static unsigned nmsgs, supp_msgs;
 505
 506         /*
 507          * Throttle back "fast" messages to no more than 10 per 5 seconds.
 508          * This isn't perfect, but it's a reasonable heuristic. If we get
 509          * more than 10, give a 6x longer delay.
 510          */
 511         nc = jiffies;
 512         if (nmsgs > 10) {
 513                 if (time_before(nc, nextmsg_time)) {
 514                         *noprint = 1;
 515                         if (!supp_msgs++)
 516                                 nextmsg_time = nc + HZ * 3;
 517                 }
 518                 else if (supp_msgs) {
 519                         handle_supp_msgs(dd, supp_msgs, msg, msgsz);
 520                         supp_msgs = 0;
 521                         nmsgs = 0;
 522                 }
 523         }
 524         else if (!nmsgs++ || time_after(nc, nextmsg_time))
 525                 nextmsg_time = nc + HZ / 2;
 526
 527         return supp_msgs;
 528 }
 529
 530 static void handle_sdma_errors(struct ipath_devdata *dd, ipath_err_t errs)
 531 {
 532         unsigned long flags;
 533         int expected;
 534
 535         if (ipath_debug & __IPATH_DBG) {
 536                 char msg[128];
 537                 ipath_decode_err(dd, msg, sizeof msg, errs &
 538                         INFINIPATH_E_SDMAERRS);
 539                 ipath_dbg("errors %lx (%s)\n", (unsigned long)errs, msg);
 540         }
 541         if (ipath_debug & __IPATH_VERBDBG) {
 542                 unsigned long tl, hd, status, lengen;
 543                 tl = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmatail);
 544                 hd = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmahead);
 545                 status = ipath_read_kreg64(dd
 546                         , dd->ipath_kregs->kr_senddmastatus);
 547                 lengen = ipath_read_kreg64(dd,
 548                         dd->ipath_kregs->kr_senddmalengen);
 549                 ipath_cdbg(VERBOSE, "sdma tl 0x%lx hd 0x%lx status 0x%lx "
 550                         "lengen 0x%lx\n", tl, hd, status, lengen);
 551         }
 552
 553         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 554         __set_bit(IPATH_SDMA_DISABLED, &dd->ipath_sdma_status);
 555         expected = test_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status);
 556         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 557         if (!expected)
 558                 ipath_cancel_sends(dd, 1);
 559 }
 560
 561 static void handle_sdma_intr(struct ipath_devdata *dd, u64 istat)
 562 {
 563         unsigned long flags;
 564         int expected;
 565
 566         if ((istat & INFINIPATH_I_SDMAINT) &&
 567             !test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
 568                 ipath_sdma_intr(dd);
 569
 570         if (istat & INFINIPATH_I_SDMADISABLED) {
 571                 expected = test_bit(IPATH_SDMA_ABORTING,
 572                         &dd->ipath_sdma_status);
 573                 ipath_dbg("%s SDmaDisabled intr\n",
 574                         expected ? "expected" : "unexpected");
 575                 spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 576                 __set_bit(IPATH_SDMA_DISABLED, &dd->ipath_sdma_status);
 577                 spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 578                 if (!expected)
 579                         ipath_cancel_sends(dd, 1);
 580                 if (!test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
 581                         tasklet_hi_schedule(&dd->ipath_sdma_abort_task);
 582         }
 583 }
 584
 585 static int handle_hdrq_full(struct ipath_devdata *dd)
 586 {
 587         int chkerrpkts = 0;
 588         u32 hd, tl;
 589         u32 i;
 590
 591         ipath_stats.sps_hdrqfull++;
 592         for (i = 0; i < dd->ipath_cfgports; i++) {
 593                 struct ipath_portdata *pd = dd->ipath_pd[i];
 594
 595                 if (i == 0) {
 596                         /*
 597                          * For kernel receive queues, we just want to know
 598                          * if there are packets in the queue that we can
 599                          * process.
 600                          */
 601                         if (pd->port_head != ipath_get_hdrqtail(pd))
 602                                 chkerrpkts |= 1 << i;
 603                         continue;
 604                 }
 605
 606                 /* Skip if user context is not open */
 607                 if (!pd || !pd->port_cnt)
 608                         continue;
 609
 610                 /* Don't report the same point multiple times. */
 611                 if (dd->ipath_flags & IPATH_NODMA_RTAIL)
 612                         tl = ipath_read_ureg32(dd, ur_rcvhdrtail, i);
 613                 else
 614                         tl = ipath_get_rcvhdrtail(pd);
 615                 if (tl == pd->port_lastrcvhdrqtail)
 616                         continue;
 617
 618                 hd = ipath_read_ureg32(dd, ur_rcvhdrhead, i);
 619                 if (hd == (tl + 1) || (!hd && tl == dd->ipath_hdrqlast)) {
 620                         pd->port_lastrcvhdrqtail = tl;
 621                         pd->port_hdrqfull++;
 622                         /* flush hdrqfull so that poll() sees it */
 623                         wmb();
 624                         wake_up_interruptible(&pd->port_wait);
 625                 }
 626         }
 627
 628         return chkerrpkts;
 629 }
 630
 631 static int handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
 632 {
 633         char msg[128];
 634         u64 ignore_this_time = 0;
 635         u64 iserr = 0;
 636         int chkerrpkts = 0, noprint = 0;
 637         unsigned supp_msgs;
 638         int log_idx;
 639
 640         /*
 641          * don't report errors that are masked, either at init
 642          * (not set in ipath_errormask), or temporarily (set in
 643          * ipath_maskederrs)
 644          */
 645         errs &= dd->ipath_errormask & ~dd->ipath_maskederrs;
 646
 647         supp_msgs = handle_frequent_errors(dd, errs, msg, (u32)sizeof msg,
 648                 &noprint);
 649
 650         /* do these first, they are most important */
 651         if (errs & INFINIPATH_E_HARDWARE) {
 652                 /* reuse same msg buf */
 653                 dd->ipath_f_handle_hwerrors(dd, msg, sizeof msg);
 654         } else {
 655                 u64 mask;
 656                 for (log_idx = 0; log_idx < IPATH_EEP_LOG_CNT; ++log_idx) {
 657                         mask = dd->ipath_eep_st_masks[log_idx].errs_to_log;
 658                         if (errs & mask)
 659                                 ipath_inc_eeprom_err(dd, log_idx, 1);
 660                 }
 661         }
 662
 663         if (errs & INFINIPATH_E_SDMAERRS)
 664                 handle_sdma_errors(dd, errs);
 665
 666         if (!noprint && (errs & ~dd->ipath_e_bitsextant))
 667                 ipath_dev_err(dd, "error interrupt with unknown errors "
 668                               "%llx set\n", (unsigned long long)
 669                               (errs & ~dd->ipath_e_bitsextant));
 670
 671         if (errs & E_SUM_ERRS)
 672                 ignore_this_time = handle_e_sum_errs(dd, errs);
 673         else if ((errs & E_SUM_LINK_PKTERRS) &&
 674             !(dd->ipath_flags & IPATH_LINKACTIVE)) {
 675                 /*
 676                  * This can happen when SMA is trying to bring the link
 677                  * up, but the IB link changes state at the "wrong" time.
 678                  * The IB logic then complains that the packet isn't
 679                  * valid.  We don't want to confuse people, so we just
 680                  * don't print them, except at debug
 681                  */
 682                 ipath_dbg("Ignoring packet errors %llx, because link not "
 683                           "ACTIVE\n", (unsigned long long) errs);
 684                 ignore_this_time = errs & E_SUM_LINK_PKTERRS;
 685         }
 686
 687         if (supp_msgs == 250000) {
 688                 int s_iserr;
 689                 /*
 690                  * It's not entirely reasonable assuming that the errors set
 691                  * in the last clear period are all responsible for the
 692                  * problem, but the alternative is to assume it's the only
 693                  * ones on this particular interrupt, which also isn't great
 694                  */
 695                 dd->ipath_maskederrs |= dd->ipath_lasterror | errs;
 696
 697                 dd->ipath_errormask &= ~dd->ipath_maskederrs;
 698                 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 699                                  dd->ipath_errormask);
 700                 s_iserr = ipath_decode_err(dd, msg, sizeof msg,
 701                                            dd->ipath_maskederrs);
 702
 703                 if (dd->ipath_maskederrs &
 704                     ~(INFINIPATH_E_RRCVEGRFULL |
 705                       INFINIPATH_E_RRCVHDRFULL | INFINIPATH_E_PKTERRS))
 706                         ipath_dev_err(dd, "Temporarily disabling "
 707                             "error(s) %llx reporting; too frequent (%s)\n",
 708                                 (unsigned long long) dd->ipath_maskederrs,
 709                                 msg);
 710                 else {
 711                         /*
 712                          * rcvegrfull and rcvhdrqfull are "normal",
 713                          * for some types of processes (mostly benchmarks)
 714                          * that send huge numbers of messages, while not
 715                          * processing them.  So only complain about
 716                          * these at debug level.
 717                          */
 718                         if (s_iserr)
 719                                 ipath_dbg("Temporarily disabling reporting "
 720                                     "too frequent queue full errors (%s)\n",
 721                                     msg);
 722                         else
 723                                 ipath_cdbg(ERRPKT,
 724                                     "Temporarily disabling reporting too"
 725                                     " frequent packet errors (%s)\n",
 726                                     msg);
 727                 }
 728
 729                 /*
 730                  * Re-enable the masked errors after around 3 minutes.  in
 731                  * ipath_get_faststats().  If we have a series of fast
 732                  * repeating but different errors, the interval will keep
 733                  * stretching out, but that's OK, as that's pretty
 734                  * catastrophic.
 735                  */
 736                 dd->ipath_unmasktime = jiffies + HZ * 180;
 737         }
 738
 739         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, errs);
 740         if (ignore_this_time)
 741                 errs &= ~ignore_this_time;
 742         if (errs & ~dd->ipath_lasterror) {
 743                 errs &= ~dd->ipath_lasterror;
 744                 /* never suppress duplicate hwerrors or ibstatuschange */
 745                 dd->ipath_lasterror |= errs &
 746                         ~(INFINIPATH_E_HARDWARE |
 747                           INFINIPATH_E_IBSTATUSCHANGED);
 748         }
 749
 750         if (errs & INFINIPATH_E_SENDSPECIALTRIGGER) {
 751                 dd->ipath_spectriggerhit++;
 752                 ipath_dbg("%lu special trigger hits\n",
 753                         dd->ipath_spectriggerhit);
 754         }
 755
 756         /* likely due to cancel; so suppress message unless verbose */
 757         if ((errs & (INFINIPATH_E_SPKTLEN | INFINIPATH_E_SPIOARMLAUNCH)) &&
 758                 time_after(dd->ipath_lastcancel, jiffies)) {
 759                 /* armlaunch takes precedence; it often causes both. */
 760                 ipath_cdbg(VERBOSE,
 761                         "Suppressed %s error (%llx) after sendbuf cancel\n",
 762                         (errs &  INFINIPATH_E_SPIOARMLAUNCH) ?
 763                         "armlaunch" : "sendpktlen", (unsigned long long)errs);
 764                 errs &= ~(INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SPKTLEN);
 765         }
 766
 767         if (!errs)
 768                 return 0;
 769
 770         if (!noprint) {
 771                 ipath_err_t mask;
 772                 /*
 773                  * The ones we mask off are handled specially below
 774                  * or above.  Also mask SDMADISABLED by default as it
 775                  * is too chatty.
 776                  */
 777                 mask = INFINIPATH_E_IBSTATUSCHANGED |
 778                         INFINIPATH_E_RRCVEGRFULL | INFINIPATH_E_RRCVHDRFULL |
 779                         INFINIPATH_E_HARDWARE | INFINIPATH_E_SDMADISABLED;
 780
 781                 /* if we're in debug, then don't mask SDMADISABLED msgs */
 782                 if (ipath_debug & __IPATH_DBG)
 783                         mask &= ~INFINIPATH_E_SDMADISABLED;
 784
 785                 ipath_decode_err(dd, msg, sizeof msg, errs & ~mask);
 786         } else
 787                 /* so we don't need if (!noprint) at strlcat's below */
 788                 *msg = 0;
 789
 790         if (errs & E_SUM_PKTERRS) {
 791                 ipath_stats.sps_pkterrs++;
 792                 chkerrpkts = 1;
 793         }
 794         if (errs & E_SUM_ERRS)
 795                 ipath_stats.sps_errs++;
 796
 797         if (errs & (INFINIPATH_E_RICRC | INFINIPATH_E_RVCRC)) {
 798                 ipath_stats.sps_crcerrs++;
 799                 chkerrpkts = 1;
 800         }
 801         iserr = errs & ~(E_SUM_PKTERRS | INFINIPATH_E_PKTERRS);
 802
 803
 804         /*
 805          * We don't want to print these two as they happen, or we can make
 806          * the situation even worse, because it takes so long to print
 807          * messages to serial consoles.  Kernel ports get printed from
 808          * fast_stats, no more than every 5 seconds, user ports get printed
 809          * on close
 810          */
 811         if (errs & INFINIPATH_E_RRCVHDRFULL)
 812                 chkerrpkts |= handle_hdrq_full(dd);
 813         if (errs & INFINIPATH_E_RRCVEGRFULL) {
 814                 struct ipath_portdata *pd = dd->ipath_pd[0];
 815
 816                 /*
 817                  * since this is of less importance and not likely to
 818                  * happen without also getting hdrfull, only count
 819                  * occurrences; don't check each port (or even the kernel
 820                  * vs user)
 821                  */
 822                 ipath_stats.sps_etidfull++;
 823                 if (pd->port_head != ipath_get_hdrqtail(pd))
 824                         chkerrpkts |= 1;
 825         }
 826
 827         /*
 828          * do this before IBSTATUSCHANGED, in case both bits set in a single
 829          * interrupt; we want the STATUSCHANGE to "win", so we do our
 830          * internal copy of state machine correctly
 831          */
 832         if (errs & INFINIPATH_E_RIBLOSTLINK) {
 833                 /*
 834                  * force through block below
 835                  */
 836                 errs |= INFINIPATH_E_IBSTATUSCHANGED;
 837                 ipath_stats.sps_iblink++;
 838                 dd->ipath_flags |= IPATH_LINKDOWN;
 839                 dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
 840                                      | IPATH_LINKARMED | IPATH_LINKACTIVE);
 841                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
 842
 843                 ipath_dbg("Lost link, link now down (%s)\n",
 844                         ipath_ibcstatus_str[ipath_read_kreg64(dd,
 845                         dd->ipath_kregs->kr_ibcstatus) & 0xf]);
 846         }
 847         if (errs & INFINIPATH_E_IBSTATUSCHANGED)
 848                 handle_e_ibstatuschanged(dd, errs);
 849
 850         if (errs & INFINIPATH_E_RESET) {
 851                 if (!noprint)
 852                         ipath_dev_err(dd, "Got reset, requires re-init "
 853                                       "(unload and reload driver)\n");
 854                 dd->ipath_flags &= ~IPATH_INITTED;      /* needs re-init */
 855                 /* mark as having had error */
 856                 *dd->ipath_statusp |= IPATH_STATUS_HWERROR;
 857                 *dd->ipath_statusp &= ~IPATH_STATUS_IB_CONF;
 858         }
 859
 860         if (!noprint && *msg) {
 861                 if (iserr)
 862                         ipath_dev_err(dd, "%s error\n", msg);
 863         }
 864         if (dd->ipath_state_wanted & dd->ipath_flags) {
 865                 ipath_cdbg(VERBOSE, "driver wanted state %x, iflags now %x, "
 866                            "waking\n", dd->ipath_state_wanted,
 867                            dd->ipath_flags);
 868                 wake_up_interruptible(&ipath_state_wait);
 869         }
 870
 871         return chkerrpkts;
 872 }
 873
 874 /*
 875  * try to cleanup as much as possible for anything that might have gone
 876  * wrong while in freeze mode, such as pio buffers being written by user
 877  * processes (causing armlaunch), send errors due to going into freeze mode,
 878  * etc., and try to avoid causing extra interrupts while doing so.
 879  * Forcibly update the in-memory pioavail register copies after cleanup
 880  * because the chip won't do it while in freeze mode (the register values
 881  * themselves are kept correct).
 882  * Make sure that we don't lose any important interrupts by using the chip
 883  * feature that says that writing 0 to a bit in *clear that is set in
 884  * *status will cause an interrupt to be generated again (if allowed by
 885  * the *mask value).
 886  */
 887 void ipath_clear_freeze(struct ipath_devdata *dd)
 888 {
 889         /* disable error interrupts, to avoid confusion */
 890         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask, 0ULL);
 891
 892         /* also disable interrupts; errormask is sometimes overwriten */
 893         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
 894
 895         ipath_cancel_sends(dd, 1);
 896
 897         /* clear the freeze, and be sure chip saw it */
 898         ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
 899                          dd->ipath_control);
 900         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 901
 902         /* force in-memory update now we are out of freeze */
 903         ipath_force_pio_avail_update(dd);
 904
 905         /*
 906          * force new interrupt if any hwerr, error or interrupt bits are
 907          * still set, and clear "safe" send packet errors related to freeze
 908          * and cancelling sends.  Re-enable error interrupts before possible
 909          * force of re-interrupt on pending interrupts.
 910          */
 911         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear, 0ULL);
 912         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear,
 913                 E_SPKT_ERRS_IGNORE);
 914         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 915                 dd->ipath_errormask);
 916         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, -1LL);
 917         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
 918 }
 919
 920
 921 /* this is separate to allow for better optimization of ipath_intr() */
 922
 923 static noinline void ipath_bad_intr(struct ipath_devdata *dd, u32 *unexpectp)
 924 {
 925         /*
 926          * sometimes happen during driver init and unload, don't want
 927          * to process any interrupts at that point
 928          */
 929
 930         /* this is just a bandaid, not a fix, if something goes badly
 931          * wrong */
 932         if (++*unexpectp > 100) {
 933                 if (++*unexpectp > 105) {
 934                         /*
 935                          * ok, we must be taking somebody else's interrupts,
 936                          * due to a messed up mptable and/or PIRQ table, so
 937                          * unregister the interrupt.  We've seen this during
 938                          * linuxbios development work, and it may happen in
 939                          * the future again.
 940                          */
 941                         if (dd->pcidev && dd->ipath_irq) {
 942                                 ipath_dev_err(dd, "Now %u unexpected "
 943                                               "interrupts, unregistering "
 944                                               "interrupt handler\n",
 945                                               *unexpectp);
 946                                 ipath_dbg("free_irq of irq %d\n",
 947                                           dd->ipath_irq);
 948                                 dd->ipath_f_free_irq(dd);
 949                         }
 950                 }
 951                 if (ipath_read_ireg(dd, dd->ipath_kregs->kr_intmask)) {
 952                         ipath_dev_err(dd, "%u unexpected interrupts, "
 953                                       "disabling interrupts completely\n",
 954                                       *unexpectp);
 955                         /*
 956                          * disable all interrupts, something is very wrong
 957                          */
 958                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
 959                                          0ULL);
 960                 }
 961         } else if (*unexpectp > 1)
 962                 ipath_dbg("Interrupt when not ready, should not happen, "
 963                           "ignoring\n");
 964 }
 965
 966 static noinline void ipath_bad_regread(struct ipath_devdata *dd)
 967 {
 968         static int allbits;
 969
 970         /* separate routine, for better optimization of ipath_intr() */
 971
 972         /*
 973          * We print the message and disable interrupts, in hope of
 974          * having a better chance of debugging the problem.
 975          */
 976         ipath_dev_err(dd,
 977                       "Read of interrupt status failed (all bits set)\n");
 978         if (allbits++) {
 979                 /* disable all interrupts, something is very wrong */
 980                 ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
 981                 if (allbits == 2) {
 982                         ipath_dev_err(dd, "Still bad interrupt status, "
 983                                       "unregistering interrupt\n");
 984                         dd->ipath_f_free_irq(dd);
 985                 } else if (allbits > 2) {
 986                         if ((allbits % 10000) == 0)
 987                                 printk(".");
 988                 } else
 989                         ipath_dev_err(dd, "Disabling interrupts, "
 990                                       "multiple errors\n");
 991         }
 992 }
 993
 994 static void handle_layer_pioavail(struct ipath_devdata *dd)
 995 {
 996         unsigned long flags;
 997         int ret;
 998
 999         ret = ipath_ib_piobufavail(dd->verbs_dev);
1000         if (ret > 0)
1001                 goto set;
1002
1003         return;
1004 set:
1005         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
1006         dd->ipath_sendctrl |= INFINIPATH_S_PIOINTBUFAVAIL;
1007         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1008                          dd->ipath_sendctrl);
1009         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1010         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
1011 }
1012
1013 /*
1014  * Handle receive interrupts for user ports; this means a user
1015  * process was waiting for a packet to arrive, and didn't want
1016  * to poll
1017  */
1018 static void handle_urcv(struct ipath_devdata *dd, u64 istat)
1019 {
1020         u64 portr;
1021         int i;
1022         int rcvdint = 0;
1023
1024         /*
1025          * test_and_clear_bit(IPATH_PORT_WAITING_RCV) and
1026          * test_and_clear_bit(IPATH_PORT_WAITING_URG) below
1027          * would both like timely updates of the bits so that
1028          * we don't pass them by unnecessarily.  the rmb()
1029          * here ensures that we see them promptly -- the
1030          * corresponding wmb()'s are in ipath_poll_urgent()
1031          * and ipath_poll_next()...
1032          */
1033         rmb();
1034         portr = ((istat >> dd->ipath_i_rcvavail_shift) &
1035                  dd->ipath_i_rcvavail_mask) |
1036                 ((istat >> dd->ipath_i_rcvurg_shift) &
1037                  dd->ipath_i_rcvurg_mask);
1038         for (i = 1; i < dd->ipath_cfgports; i++) {
1039                 struct ipath_portdata *pd = dd->ipath_pd[i];
1040
1041                 if (portr & (1 << i) && pd && pd->port_cnt) {
1042                         if (test_and_clear_bit(IPATH_PORT_WAITING_RCV,
1043                                                &pd->port_flag)) {
1044                                 clear_bit(i + dd->ipath_r_intravail_shift,
1045                                           &dd->ipath_rcvctrl);
1046                                 wake_up_interruptible(&pd->port_wait);
1047                                 rcvdint = 1;
1048                         } else if (test_and_clear_bit(IPATH_PORT_WAITING_URG,
1049                                                       &pd->port_flag)) {
1050                                 pd->port_urgent++;
1051                                 wake_up_interruptible(&pd->port_wait);
1052                         }
1053                 }
1054         }
1055         if (rcvdint) {
1056                 /* only want to take one interrupt, so turn off the rcv
1057                  * interrupt for all the ports that we set the rcv_waiting
1058                  * (but never for kernel port)
1059                  */
1060                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1061                                  dd->ipath_rcvctrl);
1062         }
1063 }
1064
1065 irqreturn_t ipath_intr(int irq, void *data)
1066 {
1067         struct ipath_devdata *dd = data;
1068         u64 istat, chk0rcv = 0;
1069         ipath_err_t estat = 0;
1070         irqreturn_t ret;
1071         static unsigned unexpected = 0;
1072         u64 kportrbits;
1073
1074         ipath_stats.sps_ints++;
1075
1076         if (dd->ipath_int_counter != (u32) -1)
1077                 dd->ipath_int_counter++;
1078
1079         if (!(dd->ipath_flags & IPATH_PRESENT)) {
1080                 /*
1081                  * This return value is not great, but we do not want the
1082                  * interrupt core code to remove our interrupt handler
1083                  * because we don't appear to be handling an interrupt
1084                  * during a chip reset.
1085                  */
1086                 return IRQ_HANDLED;
1087         }
1088
1089         /*
1090          * this needs to be flags&initted, not statusp, so we keep
1091          * taking interrupts even after link goes down, etc.
1092          * Also, we *must* clear the interrupt at some point, or we won't
1093          * take it again, which can be real bad for errors, etc...
1094          */
1095
1096         if (!(dd->ipath_flags & IPATH_INITTED)) {
1097                 ipath_bad_intr(dd, &unexpected);
1098                 ret = IRQ_NONE;
1099                 goto bail;
1100         }
1101
1102         istat = ipath_read_ireg(dd, dd->ipath_kregs->kr_intstatus);
1103
1104         if (unlikely(!istat)) {
1105                 ipath_stats.sps_nullintr++;
1106                 ret = IRQ_NONE; /* not our interrupt, or already handled */
1107                 goto bail;
1108         }
1109         if (unlikely(istat == -1)) {
1110                 ipath_bad_regread(dd);
1111                 /* don't know if it was our interrupt or not */
1112                 ret = IRQ_NONE;
1113                 goto bail;
1114         }
1115
1116         if (unexpected)
1117                 unexpected = 0;
1118
1119         if (unlikely(istat & ~dd->ipath_i_bitsextant))
1120                 ipath_dev_err(dd,
1121                               "interrupt with unknown interrupts %Lx set\n",
1122                               (unsigned long long)
1123                               istat & ~dd->ipath_i_bitsextant);
1124         else if (istat & ~INFINIPATH_I_ERROR) /* errors do own printing */
1125                 ipath_cdbg(VERBOSE, "intr stat=0x%Lx\n",
1126                         (unsigned long long) istat);
1127
1128         if (istat & INFINIPATH_I_ERROR) {
1129                 ipath_stats.sps_errints++;
1130                 estat = ipath_read_kreg64(dd,
1131                                           dd->ipath_kregs->kr_errorstatus);
1132                 if (!estat)
1133                         dev_info(&dd->pcidev->dev, "error interrupt (%Lx), "
1134                                  "but no error bits set!\n",
1135                                  (unsigned long long) istat);
1136                 else if (estat == -1LL)
1137                         /*
1138                          * should we try clearing all, or hope next read
1139                          * works?
1140                          */
1141                         ipath_dev_err(dd, "Read of error status failed "
1142                                       "(all bits set); ignoring\n");
1143                 else
1144                         chk0rcv |= handle_errors(dd, estat);
1145         }
1146
1147         if (istat & INFINIPATH_I_GPIO) {
1148                 /*
1149                  * GPIO interrupts fall in two broad classes:
1150                  * GPIO_2 indicates (on some HT4xx boards) that a packet
1151                  *        has arrived for Port 0. Checking for this
1152                  *        is controlled by flag IPATH_GPIO_INTR.
1153                  * GPIO_3..5 on IBA6120 Rev2 and IBA6110 Rev4 chips indicate
1154                  *        errors that we need to count. Checking for this
1155                  *        is controlled by flag IPATH_GPIO_ERRINTRS.
1156                  */
1157                 u32 gpiostatus;
1158                 u32 to_clear = 0;
1159
1160                 gpiostatus = ipath_read_kreg32(
1161                         dd, dd->ipath_kregs->kr_gpio_status);
1162                 /* First the error-counter case. */
1163                 if ((gpiostatus & IPATH_GPIO_ERRINTR_MASK) &&
1164                     (dd->ipath_flags & IPATH_GPIO_ERRINTRS)) {
1165                         /* want to clear the bits we see asserted. */
1166                         to_clear |= (gpiostatus & IPATH_GPIO_ERRINTR_MASK);
1167
1168                         /*
1169                          * Count appropriately, clear bits out of our copy,
1170                          * as they have been "handled".
1171                          */
1172                         if (gpiostatus & (1 << IPATH_GPIO_RXUVL_BIT)) {
1173                                 ipath_dbg("FlowCtl on UnsupVL\n");
1174                                 dd->ipath_rxfc_unsupvl_errs++;
1175                         }
1176                         if (gpiostatus & (1 << IPATH_GPIO_OVRUN_BIT)) {
1177                                 ipath_dbg("Overrun Threshold exceeded\n");
1178                                 dd->ipath_overrun_thresh_errs++;
1179                         }
1180                         if (gpiostatus & (1 << IPATH_GPIO_LLI_BIT)) {
1181                                 ipath_dbg("Local Link Integrity error\n");
1182                                 dd->ipath_lli_errs++;
1183                         }
1184                         gpiostatus &= ~IPATH_GPIO_ERRINTR_MASK;
1185                 }
1186                 /* Now the Port0 Receive case */
1187                 if ((gpiostatus & (1 << IPATH_GPIO_PORT0_BIT)) &&
1188                     (dd->ipath_flags & IPATH_GPIO_INTR)) {
1189                         /*
1190                          * GPIO status bit 2 is set, and we expected it.
1191                          * clear it and indicate in p0bits.
1192                          * This probably only happens if a Port0 pkt
1193                          * arrives at _just_ the wrong time, and we
1194                          * handle that by seting chk0rcv;
1195                          */
1196                         to_clear |= (1 << IPATH_GPIO_PORT0_BIT);
1197                         gpiostatus &= ~(1 << IPATH_GPIO_PORT0_BIT);
1198                         chk0rcv = 1;
1199                 }
1200                 if (gpiostatus) {
1201                         /*
1202                          * Some unexpected bits remain. If they could have
1203                          * caused the interrupt, complain and clear.
1204                          * To avoid repetition of this condition, also clear
1205                          * the mask. It is almost certainly due to error.
1206                          */
1207                         const u32 mask = (u32) dd->ipath_gpio_mask;
1208
1209                         if (mask & gpiostatus) {
1210                                 ipath_dbg("Unexpected GPIO IRQ bits %x\n",
1211                                   gpiostatus & mask);
1212                                 to_clear |= (gpiostatus & mask);
1213                                 dd->ipath_gpio_mask &= ~(gpiostatus & mask);
1214                                 ipath_write_kreg(dd,
1215                                         dd->ipath_kregs->kr_gpio_mask,
1216                                         dd->ipath_gpio_mask);
1217                         }
1218                 }
1219                 if (to_clear) {
1220                         ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
1221                                         (u64) to_clear);
1222                 }
1223         }
1224
1225         /*
1226          * Clear the interrupt bits we found set, unless they are receive
1227          * related, in which case we already cleared them above, and don't
1228          * want to clear them again, because we might lose an interrupt.
1229          * Clear it early, so we "know" know the chip will have seen this by
1230          * the time we process the queue, and will re-interrupt if necessary.
1231          * The processor itself won't take the interrupt again until we return.
1232          */
1233         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
1234
1235         /*
1236          * Handle kernel receive queues before checking for pio buffers
1237          * available since receives can overflow; piobuf waiters can afford
1238          * a few extra cycles, since they were waiting anyway, and user's
1239          * waiting for receive are at the bottom.
1240          */
1241         kportrbits = (1ULL << dd->ipath_i_rcvavail_shift) |
1242                 (1ULL << dd->ipath_i_rcvurg_shift);
1243         if (chk0rcv || (istat & kportrbits)) {
1244                 istat &= ~kportrbits;
1245                 ipath_kreceive(dd->ipath_pd[0]);
1246         }
1247
1248         if (istat & ((dd->ipath_i_rcvavail_mask << dd->ipath_i_rcvavail_shift) |
1249                      (dd->ipath_i_rcvurg_mask << dd->ipath_i_rcvurg_shift)))
1250                 handle_urcv(dd, istat);
1251
1252         if (istat & (INFINIPATH_I_SDMAINT | INFINIPATH_I_SDMADISABLED))
1253                 handle_sdma_intr(dd, istat);
1254
1255         if (istat & INFINIPATH_I_SPIOBUFAVAIL) {
1256                 unsigned long flags;
1257
1258                 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
1259                 dd->ipath_sendctrl &= ~INFINIPATH_S_PIOINTBUFAVAIL;
1260                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1261                                  dd->ipath_sendctrl);
1262                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1263                 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
1264
1265                 /* always process; sdma verbs uses PIO for acks and VL15  */
1266                 handle_layer_pioavail(dd);
1267         }
1268
1269         ret = IRQ_HANDLED;
1270
1271 bail:
1272         return ret;
1273 }