2 * Copyright (c) 2006, 2007 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
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:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
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.
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
34 #include <linux/pci.h>
35 #include <linux/delay.h>
37 #include "ipath_kernel.h"
38 #include "ipath_verbs.h"
39 #include "ipath_common.h"
42 * clear (write) a pio buffer, to clear a parity error. This routine
43 * should only be called when in freeze mode, and the buffer should be
44 * canceled afterwards.
46 static void ipath_clrpiobuf(struct ipath_devdata *dd, u32 pnum)
49 u32 dwcnt; /* dword count to write */
50 if (pnum < dd->ipath_piobcnt2k) {
51 pbuf = (u32 __iomem *) (dd->ipath_pio2kbase + pnum *
53 dwcnt = dd->ipath_piosize2k >> 2;
56 pbuf = (u32 __iomem *) (dd->ipath_pio4kbase +
57 (pnum - dd->ipath_piobcnt2k) * dd->ipath_4kalign);
58 dwcnt = dd->ipath_piosize4k >> 2;
60 dev_info(&dd->pcidev->dev,
61 "Rewrite PIO buffer %u, to recover from parity error\n",
64 /* no flush required, since already in freeze */
65 writel(dwcnt + 1, pbuf);
71 * Called when we might have an error that is specific to a particular
72 * PIO buffer, and may need to cancel that buffer, so it can be re-used.
73 * If rewrite is true, and bits are set in the sendbufferror registers,
74 * we'll write to the buffer, for error recovery on parity errors.
76 static void ipath_disarm_senderrbufs(struct ipath_devdata *dd, int rewrite)
79 unsigned long sbuf[4];
81 * it's possible that sendbuffererror could have bits set; might
82 * have already done this as a result of hardware error handling
84 piobcnt = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
85 /* read these before writing errorclear */
86 sbuf[0] = ipath_read_kreg64(
87 dd, dd->ipath_kregs->kr_sendbuffererror);
88 sbuf[1] = ipath_read_kreg64(
89 dd, dd->ipath_kregs->kr_sendbuffererror + 1);
91 sbuf[2] = ipath_read_kreg64(
92 dd, dd->ipath_kregs->kr_sendbuffererror + 2);
93 sbuf[3] = ipath_read_kreg64(
94 dd, dd->ipath_kregs->kr_sendbuffererror + 3);
97 if (sbuf[0] || sbuf[1] || (piobcnt > 128 && (sbuf[2] || sbuf[3]))) {
99 if (ipath_debug & (__IPATH_PKTDBG|__IPATH_DBG) &&
100 dd->ipath_lastcancel > jiffies) {
101 __IPATH_DBG_WHICH(__IPATH_PKTDBG|__IPATH_DBG,
102 "SendbufErrs %lx %lx", sbuf[0],
104 if (ipath_debug & __IPATH_PKTDBG && piobcnt > 128)
105 printk(" %lx %lx ", sbuf[2], sbuf[3]);
109 for (i = 0; i < piobcnt; i++)
110 if (test_bit(i, sbuf)) {
112 ipath_clrpiobuf(dd, i);
113 ipath_disarm_piobufs(dd, i, 1);
115 /* ignore armlaunch errs for a bit */
116 dd->ipath_lastcancel = jiffies+3;
121 /* These are all rcv-related errors which we want to count for stats */
122 #define E_SUM_PKTERRS \
123 (INFINIPATH_E_RHDRLEN | INFINIPATH_E_RBADTID | \
124 INFINIPATH_E_RBADVERSION | INFINIPATH_E_RHDR | \
125 INFINIPATH_E_RLONGPKTLEN | INFINIPATH_E_RSHORTPKTLEN | \
126 INFINIPATH_E_RMAXPKTLEN | INFINIPATH_E_RMINPKTLEN | \
127 INFINIPATH_E_RFORMATERR | INFINIPATH_E_RUNSUPVL | \
128 INFINIPATH_E_RUNEXPCHAR | INFINIPATH_E_REBP)
130 /* These are all send-related errors which we want to count for stats */
132 (INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SUNEXPERRPKTNUM | \
133 INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
134 INFINIPATH_E_SMAXPKTLEN | INFINIPATH_E_SUNSUPVL | \
135 INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
136 INFINIPATH_E_INVALIDADDR)
139 * this is similar to E_SUM_ERRS, but can't ignore armlaunch, don't ignore
140 * errors not related to freeze and cancelling buffers. Can't ignore
141 * armlaunch because could get more while still cleaning up, and need
142 * to cancel those as they happen.
144 #define E_SPKT_ERRS_IGNORE \
145 (INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
146 INFINIPATH_E_SMAXPKTLEN | INFINIPATH_E_SMINPKTLEN | \
147 INFINIPATH_E_SPKTLEN)
150 * these are errors that can occur when the link changes state while
151 * a packet is being sent or received. This doesn't cover things
152 * like EBP or VCRC that can be the result of a sending having the
153 * link change state, so we receive a "known bad" packet.
155 #define E_SUM_LINK_PKTERRS \
156 (INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
157 INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
158 INFINIPATH_E_RSHORTPKTLEN | INFINIPATH_E_RMINPKTLEN | \
159 INFINIPATH_E_RUNEXPCHAR)
161 static u64 handle_e_sum_errs(struct ipath_devdata *dd, ipath_err_t errs)
163 u64 ignore_this_time = 0;
165 ipath_disarm_senderrbufs(dd, 0);
166 if ((errs & E_SUM_LINK_PKTERRS) &&
167 !(dd->ipath_flags & IPATH_LINKACTIVE)) {
169 * This can happen when SMA is trying to bring the link
170 * up, but the IB link changes state at the "wrong" time.
171 * The IB logic then complains that the packet isn't
172 * valid. We don't want to confuse people, so we just
173 * don't print them, except at debug
175 ipath_dbg("Ignoring packet errors %llx, because link not "
176 "ACTIVE\n", (unsigned long long) errs);
177 ignore_this_time = errs & E_SUM_LINK_PKTERRS;
180 return ignore_this_time;
183 /* generic hw error messages... */
184 #define INFINIPATH_HWE_TXEMEMPARITYERR_MSG(a) \
186 .mask = ( INFINIPATH_HWE_TXEMEMPARITYERR_##a << \
187 INFINIPATH_HWE_TXEMEMPARITYERR_SHIFT ), \
188 .msg = "TXE " #a " Memory Parity" \
190 #define INFINIPATH_HWE_RXEMEMPARITYERR_MSG(a) \
192 .mask = ( INFINIPATH_HWE_RXEMEMPARITYERR_##a << \
193 INFINIPATH_HWE_RXEMEMPARITYERR_SHIFT ), \
194 .msg = "RXE " #a " Memory Parity" \
197 static const struct ipath_hwerror_msgs ipath_generic_hwerror_msgs[] = {
198 INFINIPATH_HWE_MSG(IBCBUSFRSPCPARITYERR, "IPATH2IB Parity"),
199 INFINIPATH_HWE_MSG(IBCBUSTOSPCPARITYERR, "IB2IPATH Parity"),
201 INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOBUF),
202 INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOPBC),
203 INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOLAUNCHFIFO),
205 INFINIPATH_HWE_RXEMEMPARITYERR_MSG(RCVBUF),
206 INFINIPATH_HWE_RXEMEMPARITYERR_MSG(LOOKUPQ),
207 INFINIPATH_HWE_RXEMEMPARITYERR_MSG(EAGERTID),
208 INFINIPATH_HWE_RXEMEMPARITYERR_MSG(EXPTID),
209 INFINIPATH_HWE_RXEMEMPARITYERR_MSG(FLAGBUF),
210 INFINIPATH_HWE_RXEMEMPARITYERR_MSG(DATAINFO),
211 INFINIPATH_HWE_RXEMEMPARITYERR_MSG(HDRINFO),
215 * ipath_format_hwmsg - format a single hwerror message
216 * @msg message buffer
217 * @msgl length of message buffer
218 * @hwmsg message to add to message buffer
220 static void ipath_format_hwmsg(char *msg, size_t msgl, const char *hwmsg)
222 strlcat(msg, "[", msgl);
223 strlcat(msg, hwmsg, msgl);
224 strlcat(msg, "]", msgl);
228 * ipath_format_hwerrors - format hardware error messages for display
229 * @hwerrs hardware errors bit vector
230 * @hwerrmsgs hardware error descriptions
231 * @nhwerrmsgs number of hwerrmsgs
232 * @msg message buffer
233 * @msgl message buffer length
235 void ipath_format_hwerrors(u64 hwerrs,
236 const struct ipath_hwerror_msgs *hwerrmsgs,
238 char *msg, size_t msgl)
242 sizeof(ipath_generic_hwerror_msgs) /
243 sizeof(ipath_generic_hwerror_msgs[0]);
245 for (i=0; i<glen; i++) {
246 if (hwerrs & ipath_generic_hwerror_msgs[i].mask) {
247 ipath_format_hwmsg(msg, msgl,
248 ipath_generic_hwerror_msgs[i].msg);
252 for (i=0; i<nhwerrmsgs; i++) {
253 if (hwerrs & hwerrmsgs[i].mask) {
254 ipath_format_hwmsg(msg, msgl, hwerrmsgs[i].msg);
259 /* return the strings for the most common link states */
260 static char *ib_linkstate(struct ipath_devdata *dd, u64 ibcs)
265 state = ipath_ib_state(dd, ibcs);
266 if (state == dd->ib_init)
268 else if (state == dd->ib_arm)
270 else if (state == dd->ib_active)
277 void signal_ib_event(struct ipath_devdata *dd, enum ib_event_type ev)
279 struct ib_event event;
281 event.device = &dd->verbs_dev->ibdev;
282 event.element.port_num = 1;
284 ib_dispatch_event(&event);
287 static void handle_e_ibstatuschanged(struct ipath_devdata *dd,
290 u32 ltstate, lstate, ibstate, lastlstate;
291 u32 init = dd->ib_init;
292 u32 arm = dd->ib_arm;
293 u32 active = dd->ib_active;
294 const u64 ibcs = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
296 lstate = ipath_ib_linkstate(dd, ibcs); /* linkstate */
297 ibstate = ipath_ib_state(dd, ibcs);
298 /* linkstate at last interrupt */
299 lastlstate = ipath_ib_linkstate(dd, dd->ipath_lastibcstat);
300 ltstate = ipath_ib_linktrstate(dd, ibcs); /* linktrainingtate */
303 * Since going into a recovery state causes the link state to go
304 * down and since recovery is transitory, it is better if we "miss"
305 * ever seeing the link training state go into recovery (i.e.,
306 * ignore this transition for link state special handling purposes)
307 * without even updating ipath_lastibcstat.
309 if ((ltstate == INFINIPATH_IBCS_LT_STATE_RECOVERRETRAIN) ||
310 (ltstate == INFINIPATH_IBCS_LT_STATE_RECOVERWAITRMT) ||
311 (ltstate == INFINIPATH_IBCS_LT_STATE_RECOVERIDLE))
315 * if linkstate transitions into INIT from any of the various down
316 * states, or if it transitions from any of the up (INIT or better)
317 * states into any of the down states (except link recovery), then
318 * call the chip-specific code to take appropriate actions.
320 if (lstate >= INFINIPATH_IBCS_L_STATE_INIT &&
321 lastlstate == INFINIPATH_IBCS_L_STATE_DOWN) {
322 /* transitioned to UP */
323 if (dd->ipath_f_ib_updown(dd, 1, ibcs)) {
324 /* link came up, so we must no longer be disabled */
325 dd->ipath_flags &= ~IPATH_IB_LINK_DISABLED;
326 ipath_cdbg(LINKVERB, "LinkUp handled, skipped\n");
327 goto skip_ibchange; /* chip-code handled */
329 } else if ((lastlstate >= INFINIPATH_IBCS_L_STATE_INIT ||
330 (dd->ipath_flags & IPATH_IB_FORCE_NOTIFY)) &&
331 ltstate <= INFINIPATH_IBCS_LT_STATE_CFGWAITRMT &&
332 ltstate != INFINIPATH_IBCS_LT_STATE_LINKUP) {
334 handled = dd->ipath_f_ib_updown(dd, 0, ibcs);
335 dd->ipath_flags &= ~IPATH_IB_FORCE_NOTIFY;
337 ipath_cdbg(LINKVERB, "LinkDown handled, skipped\n");
338 goto skip_ibchange; /* chip-code handled */
343 * Significant enough to always print and get into logs, if it was
344 * unexpected. If it was a requested state change, we'll have
345 * already cleared the flags, so we won't print this warning
347 if ((ibstate != arm && ibstate != active) &&
348 (dd->ipath_flags & (IPATH_LINKARMED | IPATH_LINKACTIVE))) {
349 dev_info(&dd->pcidev->dev, "Link state changed from %s "
350 "to %s\n", (dd->ipath_flags & IPATH_LINKARMED) ?
351 "ARM" : "ACTIVE", ib_linkstate(dd, ibcs));
354 if (ltstate == INFINIPATH_IBCS_LT_STATE_POLLACTIVE ||
355 ltstate == INFINIPATH_IBCS_LT_STATE_POLLQUIET) {
357 lastlts = ipath_ib_linktrstate(dd, dd->ipath_lastibcstat);
359 * Ignore cycling back and forth from Polling.Active to
360 * Polling.Quiet while waiting for the other end of the link
361 * to come up, except to try and decide if we are connected
362 * to a live IB device or not. We will cycle back and
363 * forth between them if no cable is plugged in, the other
364 * device is powered off or disabled, etc.
366 if (lastlts == INFINIPATH_IBCS_LT_STATE_POLLACTIVE ||
367 lastlts == INFINIPATH_IBCS_LT_STATE_POLLQUIET) {
368 if (++dd->ipath_ibpollcnt == 40) {
369 dd->ipath_flags |= IPATH_NOCABLE;
370 *dd->ipath_statusp |=
371 IPATH_STATUS_IB_NOCABLE;
372 ipath_cdbg(LINKVERB, "Set NOCABLE\n");
374 ipath_cdbg(LINKVERB, "POLL change to %s (%x)\n",
375 ipath_ibcstatus_str[ltstate], ibstate);
380 dd->ipath_ibpollcnt = 0; /* not poll*, now */
381 ipath_stats.sps_iblink++;
383 if (ibstate != init && dd->ipath_lastlinkrecov && ipath_linkrecovery) {
385 linkrecov = ipath_snap_cntr(dd,
386 dd->ipath_cregs->cr_iblinkerrrecovcnt);
387 if (linkrecov != dd->ipath_lastlinkrecov) {
388 ipath_dbg("IB linkrecov up %Lx (%s %s) recov %Lu\n",
389 ibcs, ib_linkstate(dd, ibcs),
390 ipath_ibcstatus_str[ltstate],
392 /* and no more until active again */
393 dd->ipath_lastlinkrecov = 0;
394 ipath_set_linkstate(dd, IPATH_IB_LINKDOWN);
399 if (ibstate == init || ibstate == arm || ibstate == active) {
400 *dd->ipath_statusp &= ~IPATH_STATUS_IB_NOCABLE;
401 if (ibstate == init || ibstate == arm) {
402 *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
403 if (dd->ipath_flags & IPATH_LINKACTIVE)
404 signal_ib_event(dd, IB_EVENT_PORT_ERR);
406 if (ibstate == arm) {
407 dd->ipath_flags |= IPATH_LINKARMED;
408 dd->ipath_flags &= ~(IPATH_LINKUNK |
409 IPATH_LINKINIT | IPATH_LINKDOWN |
410 IPATH_LINKACTIVE | IPATH_NOCABLE);
412 } else if (ibstate == init) {
414 * set INIT and DOWN. Down is checked by
415 * most of the other code, but INIT is
416 * useful to know in a few places.
418 dd->ipath_flags |= IPATH_LINKINIT |
420 dd->ipath_flags &= ~(IPATH_LINKUNK |
421 IPATH_LINKARMED | IPATH_LINKACTIVE |
424 } else { /* active */
425 dd->ipath_lastlinkrecov = ipath_snap_cntr(dd,
426 dd->ipath_cregs->cr_iblinkerrrecovcnt);
427 *dd->ipath_statusp |=
428 IPATH_STATUS_IB_READY | IPATH_STATUS_IB_CONF;
429 dd->ipath_flags |= IPATH_LINKACTIVE;
430 dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
431 | IPATH_LINKDOWN | IPATH_LINKARMED |
433 signal_ib_event(dd, IB_EVENT_PORT_ACTIVE);
434 /* LED active not handled in chip _f_updown */
435 dd->ipath_f_setextled(dd, lstate, ltstate);
440 * print after we've already done the work, so as not to
441 * delay the state changes and notifications, for debugging
443 if (lstate == lastlstate)
444 ipath_cdbg(LINKVERB, "Unchanged from last: %s "
445 "(%x)\n", ib_linkstate(dd, ibcs), ibstate);
447 ipath_cdbg(VERBOSE, "Unit %u: link up to %s %s (%x)\n",
448 dd->ipath_unit, ib_linkstate(dd, ibcs),
449 ipath_ibcstatus_str[ltstate], ibstate);
451 if (dd->ipath_flags & IPATH_LINKACTIVE)
452 signal_ib_event(dd, IB_EVENT_PORT_ERR);
453 dd->ipath_flags |= IPATH_LINKDOWN;
454 dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
457 *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
458 dd->ipath_lli_counter = 0;
460 if (lastlstate != INFINIPATH_IBCS_L_STATE_DOWN)
461 ipath_cdbg(VERBOSE, "Unit %u link state down "
462 "(state 0x%x), from %s\n",
463 dd->ipath_unit, lstate,
464 ib_linkstate(dd, dd->ipath_lastibcstat));
466 ipath_cdbg(LINKVERB, "Unit %u link state changed "
467 "to %s (0x%x) from down (%x)\n",
469 ipath_ibcstatus_str[ltstate],
470 ibstate, lastlstate);
474 dd->ipath_lastibcstat = ibcs;
479 static void handle_supp_msgs(struct ipath_devdata *dd,
480 unsigned supp_msgs, char *msg, int msgsz)
483 * Print the message unless it's ibc status change only, which
484 * happens so often we never want to count it.
486 if (dd->ipath_lasterror & ~INFINIPATH_E_IBSTATUSCHANGED) {
488 iserr = ipath_decode_err(msg, msgsz,
489 dd->ipath_lasterror &
490 ~INFINIPATH_E_IBSTATUSCHANGED);
491 if (dd->ipath_lasterror &
492 ~(INFINIPATH_E_RRCVEGRFULL |
493 INFINIPATH_E_RRCVHDRFULL | INFINIPATH_E_PKTERRS))
494 ipath_dev_err(dd, "Suppressed %u messages for "
495 "fast-repeating errors (%s) (%llx)\n",
498 dd->ipath_lasterror);
501 * rcvegrfull and rcvhdrqfull are "normal", for some
502 * types of processes (mostly benchmarks) that send
503 * huge numbers of messages, while not processing
504 * them. So only complain about these at debug
508 ipath_dbg("Suppressed %u messages for %s\n",
512 "Suppressed %u messages for %s\n",
518 static unsigned handle_frequent_errors(struct ipath_devdata *dd,
519 ipath_err_t errs, char *msg,
520 int msgsz, int *noprint)
523 static unsigned long nextmsg_time;
524 static unsigned nmsgs, supp_msgs;
527 * Throttle back "fast" messages to no more than 10 per 5 seconds.
528 * This isn't perfect, but it's a reasonable heuristic. If we get
529 * more than 10, give a 6x longer delay.
533 if (time_before(nc, nextmsg_time)) {
536 nextmsg_time = nc + HZ * 3;
538 else if (supp_msgs) {
539 handle_supp_msgs(dd, supp_msgs, msg, msgsz);
544 else if (!nmsgs++ || time_after(nc, nextmsg_time))
545 nextmsg_time = nc + HZ / 2;
550 static int handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
553 u64 ignore_this_time = 0;
555 int chkerrpkts = 0, noprint = 0;
559 supp_msgs = handle_frequent_errors(dd, errs, msg, sizeof msg, &noprint);
561 /* don't report errors that are masked */
562 errs &= ~dd->ipath_maskederrs;
564 /* do these first, they are most important */
565 if (errs & INFINIPATH_E_HARDWARE) {
566 /* reuse same msg buf */
567 dd->ipath_f_handle_hwerrors(dd, msg, sizeof msg);
570 for (log_idx = 0; log_idx < IPATH_EEP_LOG_CNT; ++log_idx) {
571 mask = dd->ipath_eep_st_masks[log_idx].errs_to_log;
573 ipath_inc_eeprom_err(dd, log_idx, 1);
577 if (!noprint && (errs & ~dd->ipath_e_bitsextant))
578 ipath_dev_err(dd, "error interrupt with unknown errors "
579 "%llx set\n", (unsigned long long)
580 (errs & ~dd->ipath_e_bitsextant));
582 if (errs & E_SUM_ERRS)
583 ignore_this_time = handle_e_sum_errs(dd, errs);
584 else if ((errs & E_SUM_LINK_PKTERRS) &&
585 !(dd->ipath_flags & IPATH_LINKACTIVE)) {
587 * This can happen when SMA is trying to bring the link
588 * up, but the IB link changes state at the "wrong" time.
589 * The IB logic then complains that the packet isn't
590 * valid. We don't want to confuse people, so we just
591 * don't print them, except at debug
593 ipath_dbg("Ignoring packet errors %llx, because link not "
594 "ACTIVE\n", (unsigned long long) errs);
595 ignore_this_time = errs & E_SUM_LINK_PKTERRS;
598 if (supp_msgs == 250000) {
601 * It's not entirely reasonable assuming that the errors set
602 * in the last clear period are all responsible for the
603 * problem, but the alternative is to assume it's the only
604 * ones on this particular interrupt, which also isn't great
606 dd->ipath_maskederrs |= dd->ipath_lasterror | errs;
608 dd->ipath_errormask &= ~dd->ipath_maskederrs;
609 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
610 dd->ipath_errormask);
611 s_iserr = ipath_decode_err(msg, sizeof msg,
612 dd->ipath_maskederrs);
614 if (dd->ipath_maskederrs &
615 ~(INFINIPATH_E_RRCVEGRFULL |
616 INFINIPATH_E_RRCVHDRFULL | INFINIPATH_E_PKTERRS))
617 ipath_dev_err(dd, "Temporarily disabling "
618 "error(s) %llx reporting; too frequent (%s)\n",
619 (unsigned long long) dd->ipath_maskederrs,
623 * rcvegrfull and rcvhdrqfull are "normal",
624 * for some types of processes (mostly benchmarks)
625 * that send huge numbers of messages, while not
626 * processing them. So only complain about
627 * these at debug level.
630 ipath_dbg("Temporarily disabling reporting "
631 "too frequent queue full errors (%s)\n",
635 "Temporarily disabling reporting too"
636 " frequent packet errors (%s)\n",
641 * Re-enable the masked errors after around 3 minutes. in
642 * ipath_get_faststats(). If we have a series of fast
643 * repeating but different errors, the interval will keep
644 * stretching out, but that's OK, as that's pretty
647 dd->ipath_unmasktime = jiffies + HZ * 180;
650 ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, errs);
651 if (ignore_this_time)
652 errs &= ~ignore_this_time;
653 if (errs & ~dd->ipath_lasterror) {
654 errs &= ~dd->ipath_lasterror;
655 /* never suppress duplicate hwerrors or ibstatuschange */
656 dd->ipath_lasterror |= errs &
657 ~(INFINIPATH_E_HARDWARE |
658 INFINIPATH_E_IBSTATUSCHANGED);
661 /* likely due to cancel, so suppress */
662 if ((errs & (INFINIPATH_E_SPKTLEN | INFINIPATH_E_SPIOARMLAUNCH)) &&
663 dd->ipath_lastcancel > jiffies) {
664 ipath_dbg("Suppressed armlaunch/spktlen after error send cancel\n");
665 errs &= ~(INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SPKTLEN);
673 * the ones we mask off are handled specially below or above
675 ipath_decode_err(msg, sizeof msg,
676 errs & ~(INFINIPATH_E_IBSTATUSCHANGED |
677 INFINIPATH_E_RRCVEGRFULL |
678 INFINIPATH_E_RRCVHDRFULL |
679 INFINIPATH_E_HARDWARE));
681 /* so we don't need if (!noprint) at strlcat's below */
684 if (errs & E_SUM_PKTERRS) {
685 ipath_stats.sps_pkterrs++;
688 if (errs & E_SUM_ERRS)
689 ipath_stats.sps_errs++;
691 if (errs & (INFINIPATH_E_RICRC | INFINIPATH_E_RVCRC)) {
692 ipath_stats.sps_crcerrs++;
695 iserr = errs & ~(E_SUM_PKTERRS | INFINIPATH_E_PKTERRS);
699 * We don't want to print these two as they happen, or we can make
700 * the situation even worse, because it takes so long to print
701 * messages to serial consoles. Kernel ports get printed from
702 * fast_stats, no more than every 5 seconds, user ports get printed
705 if (errs & INFINIPATH_E_RRCVHDRFULL) {
707 ipath_stats.sps_hdrqfull++;
708 for (i = 0; i < dd->ipath_cfgports; i++) {
709 struct ipath_portdata *pd = dd->ipath_pd[i];
712 tl = ipath_get_hdrqtail(pd);
713 } else if (pd && pd->port_cnt &&
714 pd->port_rcvhdrtail_kvaddr) {
716 * don't report same point multiple times,
719 tl = *(u64 *) pd->port_rcvhdrtail_kvaddr;
720 if (tl == pd->port_lastrcvhdrqtail)
722 hd = ipath_read_ureg32(dd, ur_rcvhdrhead,
726 if (hd == (tl + 1) ||
727 (!hd && tl == dd->ipath_hdrqlast)) {
730 pd->port_lastrcvhdrqtail = tl;
732 /* flush hdrqfull so that poll() sees it */
734 wake_up_interruptible(&pd->port_wait);
738 if (errs & INFINIPATH_E_RRCVEGRFULL) {
739 struct ipath_portdata *pd = dd->ipath_pd[0];
742 * since this is of less importance and not likely to
743 * happen without also getting hdrfull, only count
744 * occurrences; don't check each port (or even the kernel
747 ipath_stats.sps_etidfull++;
748 if (pd->port_head != ipath_get_hdrqtail(pd))
753 * do this before IBSTATUSCHANGED, in case both bits set in a single
754 * interrupt; we want the STATUSCHANGE to "win", so we do our
755 * internal copy of state machine correctly
757 if (errs & INFINIPATH_E_RIBLOSTLINK) {
759 * force through block below
761 errs |= INFINIPATH_E_IBSTATUSCHANGED;
762 ipath_stats.sps_iblink++;
763 dd->ipath_flags |= IPATH_LINKDOWN;
764 dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
765 | IPATH_LINKARMED | IPATH_LINKACTIVE);
766 *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
768 ipath_dbg("Lost link, link now down (%s)\n",
769 ipath_ibcstatus_str[ipath_read_kreg64(dd,
770 dd->ipath_kregs->kr_ibcstatus) & 0xf]);
772 if (errs & INFINIPATH_E_IBSTATUSCHANGED)
773 handle_e_ibstatuschanged(dd, errs);
775 if (errs & INFINIPATH_E_RESET) {
777 ipath_dev_err(dd, "Got reset, requires re-init "
778 "(unload and reload driver)\n");
779 dd->ipath_flags &= ~IPATH_INITTED; /* needs re-init */
780 /* mark as having had error */
781 *dd->ipath_statusp |= IPATH_STATUS_HWERROR;
782 *dd->ipath_statusp &= ~IPATH_STATUS_IB_CONF;
785 if (!noprint && *msg) {
787 ipath_dev_err(dd, "%s error\n", msg);
789 dev_info(&dd->pcidev->dev, "%s packet problems\n",
792 if (dd->ipath_state_wanted & dd->ipath_flags) {
793 ipath_cdbg(VERBOSE, "driver wanted state %x, iflags now %x, "
794 "waking\n", dd->ipath_state_wanted,
796 wake_up_interruptible(&ipath_state_wait);
803 * try to cleanup as much as possible for anything that might have gone
804 * wrong while in freeze mode, such as pio buffers being written by user
805 * processes (causing armlaunch), send errors due to going into freeze mode,
806 * etc., and try to avoid causing extra interrupts while doing so.
807 * Forcibly update the in-memory pioavail register copies after cleanup
808 * because the chip won't do it for anything changing while in freeze mode
809 * (we don't want to wait for the next pio buffer state change).
810 * Make sure that we don't lose any important interrupts by using the chip
811 * feature that says that writing 0 to a bit in *clear that is set in
812 * *status will cause an interrupt to be generated again (if allowed by
815 void ipath_clear_freeze(struct ipath_devdata *dd)
820 /* disable error interrupts, to avoid confusion */
821 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask, 0ULL);
823 /* also disable interrupts; errormask is sometimes overwriten */
824 ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
827 * clear all sends, because they have may been
828 * completed by usercode while in freeze mode, and
829 * therefore would not be sent, and eventually
830 * might cause the process to run out of bufs
832 ipath_cancel_sends(dd, 0);
833 ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
836 /* ensure pio avail updates continue */
837 ipath_force_pio_avail_update(dd);
840 * We just enabled pioavailupdate, so dma copy is almost certainly
841 * not yet right, so read the registers directly. Similar to init
843 for (i = 0; i < dd->ipath_pioavregs; i++) {
844 /* deal with 6110 chip bug */
845 im = (i > 3 && (dd->ipath_flags & IPATH_SWAP_PIOBUFS)) ?
847 val = ipath_read_kreg64(dd, (0x1000 / sizeof(u64)) + im);
848 dd->ipath_pioavailregs_dma[i] = cpu_to_le64(val);
849 dd->ipath_pioavailshadow[i] = val |
850 (~dd->ipath_pioavailkernel[i] <<
851 INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT);
855 * force new interrupt if any hwerr, error or interrupt bits are
856 * still set, and clear "safe" send packet errors related to freeze
857 * and cancelling sends. Re-enable error interrupts before possible
858 * force of re-interrupt on pending interrupts.
860 ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear, 0ULL);
861 ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear,
863 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
864 dd->ipath_errormask);
865 ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, -1LL);
866 ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
870 /* this is separate to allow for better optimization of ipath_intr() */
872 static noinline void ipath_bad_intr(struct ipath_devdata *dd, u32 *unexpectp)
875 * sometimes happen during driver init and unload, don't want
876 * to process any interrupts at that point
879 /* this is just a bandaid, not a fix, if something goes badly
881 if (++*unexpectp > 100) {
882 if (++*unexpectp > 105) {
884 * ok, we must be taking somebody else's interrupts,
885 * due to a messed up mptable and/or PIRQ table, so
886 * unregister the interrupt. We've seen this during
887 * linuxbios development work, and it may happen in
890 if (dd->pcidev && dd->ipath_irq) {
891 ipath_dev_err(dd, "Now %u unexpected "
892 "interrupts, unregistering "
893 "interrupt handler\n",
895 ipath_dbg("free_irq of irq %d\n",
897 dd->ipath_f_free_irq(dd);
900 if (ipath_read_ireg(dd, dd->ipath_kregs->kr_intmask)) {
901 ipath_dev_err(dd, "%u unexpected interrupts, "
902 "disabling interrupts completely\n",
905 * disable all interrupts, something is very wrong
907 ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
910 } else if (*unexpectp > 1)
911 ipath_dbg("Interrupt when not ready, should not happen, "
915 static noinline void ipath_bad_regread(struct ipath_devdata *dd)
919 /* separate routine, for better optimization of ipath_intr() */
922 * We print the message and disable interrupts, in hope of
923 * having a better chance of debugging the problem.
926 "Read of interrupt status failed (all bits set)\n");
928 /* disable all interrupts, something is very wrong */
929 ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
931 ipath_dev_err(dd, "Still bad interrupt status, "
932 "unregistering interrupt\n");
933 dd->ipath_f_free_irq(dd);
934 } else if (allbits > 2) {
935 if ((allbits % 10000) == 0)
938 ipath_dev_err(dd, "Disabling interrupts, "
939 "multiple errors\n");
943 static void handle_layer_pioavail(struct ipath_devdata *dd)
948 ret = ipath_ib_piobufavail(dd->verbs_dev);
954 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
955 dd->ipath_sendctrl |= INFINIPATH_S_PIOINTBUFAVAIL;
956 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
958 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
959 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
963 * Handle receive interrupts for user ports; this means a user
964 * process was waiting for a packet to arrive, and didn't want
967 static void handle_urcv(struct ipath_devdata *dd, u64 istat)
974 * test_and_clear_bit(IPATH_PORT_WAITING_RCV) and
975 * test_and_clear_bit(IPATH_PORT_WAITING_URG) below
976 * would both like timely updates of the bits so that
977 * we don't pass them by unnecessarily. the rmb()
978 * here ensures that we see them promptly -- the
979 * corresponding wmb()'s are in ipath_poll_urgent()
980 * and ipath_poll_next()...
983 portr = ((istat >> dd->ipath_i_rcvavail_shift) &
984 dd->ipath_i_rcvavail_mask) |
985 ((istat >> dd->ipath_i_rcvurg_shift) &
986 dd->ipath_i_rcvurg_mask);
987 for (i = 1; i < dd->ipath_cfgports; i++) {
988 struct ipath_portdata *pd = dd->ipath_pd[i];
990 if (portr & (1 << i) && pd && pd->port_cnt) {
991 if (test_and_clear_bit(IPATH_PORT_WAITING_RCV,
993 clear_bit(i + dd->ipath_r_intravail_shift,
995 wake_up_interruptible(&pd->port_wait);
997 } else if (test_and_clear_bit(IPATH_PORT_WAITING_URG,
1000 wake_up_interruptible(&pd->port_wait);
1005 /* only want to take one interrupt, so turn off the rcv
1006 * interrupt for all the ports that we set the rcv_waiting
1007 * (but never for kernel port)
1009 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1014 irqreturn_t ipath_intr(int irq, void *data)
1016 struct ipath_devdata *dd = data;
1017 u32 istat, chk0rcv = 0;
1018 ipath_err_t estat = 0;
1020 static unsigned unexpected = 0;
1023 ipath_stats.sps_ints++;
1025 if (dd->ipath_int_counter != (u32) -1)
1026 dd->ipath_int_counter++;
1028 if (!(dd->ipath_flags & IPATH_PRESENT)) {
1030 * This return value is not great, but we do not want the
1031 * interrupt core code to remove our interrupt handler
1032 * because we don't appear to be handling an interrupt
1033 * during a chip reset.
1039 * this needs to be flags&initted, not statusp, so we keep
1040 * taking interrupts even after link goes down, etc.
1041 * Also, we *must* clear the interrupt at some point, or we won't
1042 * take it again, which can be real bad for errors, etc...
1045 if (!(dd->ipath_flags & IPATH_INITTED)) {
1046 ipath_bad_intr(dd, &unexpected);
1051 istat = ipath_read_ireg(dd, dd->ipath_kregs->kr_intstatus);
1053 if (unlikely(!istat)) {
1054 ipath_stats.sps_nullintr++;
1055 ret = IRQ_NONE; /* not our interrupt, or already handled */
1058 if (unlikely(istat == -1)) {
1059 ipath_bad_regread(dd);
1060 /* don't know if it was our interrupt or not */
1068 if (unlikely(istat & ~dd->ipath_i_bitsextant))
1070 "interrupt with unknown interrupts %x set\n",
1071 istat & (u32) ~ dd->ipath_i_bitsextant);
1073 ipath_cdbg(VERBOSE, "intr stat=0x%x\n", istat);
1075 if (unlikely(istat & INFINIPATH_I_ERROR)) {
1076 ipath_stats.sps_errints++;
1077 estat = ipath_read_kreg64(dd,
1078 dd->ipath_kregs->kr_errorstatus);
1080 dev_info(&dd->pcidev->dev, "error interrupt (%x), "
1081 "but no error bits set!\n", istat);
1082 else if (estat == -1LL)
1084 * should we try clearing all, or hope next read
1087 ipath_dev_err(dd, "Read of error status failed "
1088 "(all bits set); ignoring\n");
1090 chk0rcv |= handle_errors(dd, estat);
1093 if (istat & INFINIPATH_I_GPIO) {
1095 * GPIO interrupts fall in two broad classes:
1096 * GPIO_2 indicates (on some HT4xx boards) that a packet
1097 * has arrived for Port 0. Checking for this
1098 * is controlled by flag IPATH_GPIO_INTR.
1099 * GPIO_3..5 on IBA6120 Rev2 and IBA6110 Rev4 chips indicate
1100 * errors that we need to count. Checking for this
1101 * is controlled by flag IPATH_GPIO_ERRINTRS.
1106 gpiostatus = ipath_read_kreg32(
1107 dd, dd->ipath_kregs->kr_gpio_status);
1108 /* First the error-counter case. */
1109 if ((gpiostatus & IPATH_GPIO_ERRINTR_MASK) &&
1110 (dd->ipath_flags & IPATH_GPIO_ERRINTRS)) {
1111 /* want to clear the bits we see asserted. */
1112 to_clear |= (gpiostatus & IPATH_GPIO_ERRINTR_MASK);
1115 * Count appropriately, clear bits out of our copy,
1116 * as they have been "handled".
1118 if (gpiostatus & (1 << IPATH_GPIO_RXUVL_BIT)) {
1119 ipath_dbg("FlowCtl on UnsupVL\n");
1120 dd->ipath_rxfc_unsupvl_errs++;
1122 if (gpiostatus & (1 << IPATH_GPIO_OVRUN_BIT)) {
1123 ipath_dbg("Overrun Threshold exceeded\n");
1124 dd->ipath_overrun_thresh_errs++;
1126 if (gpiostatus & (1 << IPATH_GPIO_LLI_BIT)) {
1127 ipath_dbg("Local Link Integrity error\n");
1128 dd->ipath_lli_errs++;
1130 gpiostatus &= ~IPATH_GPIO_ERRINTR_MASK;
1132 /* Now the Port0 Receive case */
1133 if ((gpiostatus & (1 << IPATH_GPIO_PORT0_BIT)) &&
1134 (dd->ipath_flags & IPATH_GPIO_INTR)) {
1136 * GPIO status bit 2 is set, and we expected it.
1137 * clear it and indicate in p0bits.
1138 * This probably only happens if a Port0 pkt
1139 * arrives at _just_ the wrong time, and we
1140 * handle that by seting chk0rcv;
1142 to_clear |= (1 << IPATH_GPIO_PORT0_BIT);
1143 gpiostatus &= ~(1 << IPATH_GPIO_PORT0_BIT);
1148 * Some unexpected bits remain. If they could have
1149 * caused the interrupt, complain and clear.
1150 * To avoid repetition of this condition, also clear
1151 * the mask. It is almost certainly due to error.
1153 const u32 mask = (u32) dd->ipath_gpio_mask;
1155 if (mask & gpiostatus) {
1156 ipath_dbg("Unexpected GPIO IRQ bits %x\n",
1158 to_clear |= (gpiostatus & mask);
1159 dd->ipath_gpio_mask &= ~(gpiostatus & mask);
1160 ipath_write_kreg(dd,
1161 dd->ipath_kregs->kr_gpio_mask,
1162 dd->ipath_gpio_mask);
1166 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
1172 * Clear the interrupt bits we found set, unless they are receive
1173 * related, in which case we already cleared them above, and don't
1174 * want to clear them again, because we might lose an interrupt.
1175 * Clear it early, so we "know" know the chip will have seen this by
1176 * the time we process the queue, and will re-interrupt if necessary.
1177 * The processor itself won't take the interrupt again until we return.
1179 ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
1182 * Handle kernel receive queues before checking for pio buffers
1183 * available since receives can overflow; piobuf waiters can afford
1184 * a few extra cycles, since they were waiting anyway, and user's
1185 * waiting for receive are at the bottom.
1187 kportrbits = (1ULL << dd->ipath_i_rcvavail_shift) |
1188 (1ULL << dd->ipath_i_rcvurg_shift);
1189 if (chk0rcv || (istat & kportrbits)) {
1190 istat &= ~kportrbits;
1191 ipath_kreceive(dd->ipath_pd[0]);
1194 if (istat & ((dd->ipath_i_rcvavail_mask << dd->ipath_i_rcvavail_shift) |
1195 (dd->ipath_i_rcvurg_mask << dd->ipath_i_rcvurg_shift)))
1196 handle_urcv(dd, istat);
1198 if (istat & INFINIPATH_I_SPIOBUFAVAIL) {
1199 unsigned long flags;
1201 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
1202 dd->ipath_sendctrl &= ~INFINIPATH_S_PIOINTBUFAVAIL;
1203 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1204 dd->ipath_sendctrl);
1205 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1206 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
1208 handle_layer_pioavail(dd);