2 * Copyright (c) 2006 QLogic, Inc. 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>
36 #include "ipath_kernel.h"
37 #include "ipath_verbs.h"
38 #include "ipath_common.h"
41 * Called when we might have an error that is specific to a particular
42 * PIO buffer, and may need to cancel that buffer, so it can be re-used.
44 void ipath_disarm_senderrbufs(struct ipath_devdata *dd)
47 unsigned long sbuf[4];
49 * it's possible that sendbuffererror could have bits set; might
50 * have already done this as a result of hardware error handling
52 piobcnt = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
53 /* read these before writing errorclear */
54 sbuf[0] = ipath_read_kreg64(
55 dd, dd->ipath_kregs->kr_sendbuffererror);
56 sbuf[1] = ipath_read_kreg64(
57 dd, dd->ipath_kregs->kr_sendbuffererror + 1);
59 sbuf[2] = ipath_read_kreg64(
60 dd, dd->ipath_kregs->kr_sendbuffererror + 2);
61 sbuf[3] = ipath_read_kreg64(
62 dd, dd->ipath_kregs->kr_sendbuffererror + 3);
65 if (sbuf[0] || sbuf[1] || (piobcnt > 128 && (sbuf[2] || sbuf[3]))) {
67 if (ipath_debug & (__IPATH_PKTDBG|__IPATH_DBG)) {
68 __IPATH_DBG_WHICH(__IPATH_PKTDBG|__IPATH_DBG,
69 "SendbufErrs %lx %lx", sbuf[0],
71 if (ipath_debug & __IPATH_PKTDBG && piobcnt > 128)
72 printk(" %lx %lx ", sbuf[2], sbuf[3]);
76 for (i = 0; i < piobcnt; i++)
77 if (test_bit(i, sbuf))
78 ipath_disarm_piobufs(dd, i, 1);
79 dd->ipath_lastcancel = jiffies+3; /* no armlaunch for a bit */
84 /* These are all rcv-related errors which we want to count for stats */
85 #define E_SUM_PKTERRS \
86 (INFINIPATH_E_RHDRLEN | INFINIPATH_E_RBADTID | \
87 INFINIPATH_E_RBADVERSION | INFINIPATH_E_RHDR | \
88 INFINIPATH_E_RLONGPKTLEN | INFINIPATH_E_RSHORTPKTLEN | \
89 INFINIPATH_E_RMAXPKTLEN | INFINIPATH_E_RMINPKTLEN | \
90 INFINIPATH_E_RFORMATERR | INFINIPATH_E_RUNSUPVL | \
91 INFINIPATH_E_RUNEXPCHAR | INFINIPATH_E_REBP)
93 /* These are all send-related errors which we want to count for stats */
95 (INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SUNEXPERRPKTNUM | \
96 INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
97 INFINIPATH_E_SMAXPKTLEN | INFINIPATH_E_SUNSUPVL | \
98 INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
99 INFINIPATH_E_INVALIDADDR)
102 * these are errors that can occur when the link changes state while
103 * a packet is being sent or received. This doesn't cover things
104 * like EBP or VCRC that can be the result of a sending having the
105 * link change state, so we receive a "known bad" packet.
107 #define E_SUM_LINK_PKTERRS \
108 (INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
109 INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
110 INFINIPATH_E_RSHORTPKTLEN | INFINIPATH_E_RMINPKTLEN | \
111 INFINIPATH_E_RUNEXPCHAR)
113 static u64 handle_e_sum_errs(struct ipath_devdata *dd, ipath_err_t errs)
115 u64 ignore_this_time = 0;
117 ipath_disarm_senderrbufs(dd);
118 if ((errs & E_SUM_LINK_PKTERRS) &&
119 !(dd->ipath_flags & IPATH_LINKACTIVE)) {
121 * This can happen when SMA is trying to bring the link
122 * up, but the IB link changes state at the "wrong" time.
123 * The IB logic then complains that the packet isn't
124 * valid. We don't want to confuse people, so we just
125 * don't print them, except at debug
127 ipath_dbg("Ignoring packet errors %llx, because link not "
128 "ACTIVE\n", (unsigned long long) errs);
129 ignore_this_time = errs & E_SUM_LINK_PKTERRS;
132 return ignore_this_time;
135 /* generic hw error messages... */
136 #define INFINIPATH_HWE_TXEMEMPARITYERR_MSG(a) \
138 .mask = ( INFINIPATH_HWE_TXEMEMPARITYERR_##a << \
139 INFINIPATH_HWE_TXEMEMPARITYERR_SHIFT ), \
140 .msg = "TXE " #a " Memory Parity" \
142 #define INFINIPATH_HWE_RXEMEMPARITYERR_MSG(a) \
144 .mask = ( INFINIPATH_HWE_RXEMEMPARITYERR_##a << \
145 INFINIPATH_HWE_RXEMEMPARITYERR_SHIFT ), \
146 .msg = "RXE " #a " Memory Parity" \
149 static const struct ipath_hwerror_msgs ipath_generic_hwerror_msgs[] = {
150 INFINIPATH_HWE_MSG(IBCBUSFRSPCPARITYERR, "IPATH2IB Parity"),
151 INFINIPATH_HWE_MSG(IBCBUSTOSPCPARITYERR, "IB2IPATH Parity"),
153 INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOBUF),
154 INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOPBC),
155 INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOLAUNCHFIFO),
157 INFINIPATH_HWE_RXEMEMPARITYERR_MSG(RCVBUF),
158 INFINIPATH_HWE_RXEMEMPARITYERR_MSG(LOOKUPQ),
159 INFINIPATH_HWE_RXEMEMPARITYERR_MSG(EAGERTID),
160 INFINIPATH_HWE_RXEMEMPARITYERR_MSG(EXPTID),
161 INFINIPATH_HWE_RXEMEMPARITYERR_MSG(FLAGBUF),
162 INFINIPATH_HWE_RXEMEMPARITYERR_MSG(DATAINFO),
163 INFINIPATH_HWE_RXEMEMPARITYERR_MSG(HDRINFO),
167 * ipath_format_hwmsg - format a single hwerror message
168 * @msg message buffer
169 * @msgl length of message buffer
170 * @hwmsg message to add to message buffer
172 static void ipath_format_hwmsg(char *msg, size_t msgl, const char *hwmsg)
174 strlcat(msg, "[", msgl);
175 strlcat(msg, hwmsg, msgl);
176 strlcat(msg, "]", msgl);
180 * ipath_format_hwerrors - format hardware error messages for display
181 * @hwerrs hardware errors bit vector
182 * @hwerrmsgs hardware error descriptions
183 * @nhwerrmsgs number of hwerrmsgs
184 * @msg message buffer
185 * @msgl message buffer length
187 void ipath_format_hwerrors(u64 hwerrs,
188 const struct ipath_hwerror_msgs *hwerrmsgs,
190 char *msg, size_t msgl)
194 sizeof(ipath_generic_hwerror_msgs) /
195 sizeof(ipath_generic_hwerror_msgs[0]);
197 for (i=0; i<glen; i++) {
198 if (hwerrs & ipath_generic_hwerror_msgs[i].mask) {
199 ipath_format_hwmsg(msg, msgl,
200 ipath_generic_hwerror_msgs[i].msg);
204 for (i=0; i<nhwerrmsgs; i++) {
205 if (hwerrs & hwerrmsgs[i].mask) {
206 ipath_format_hwmsg(msg, msgl, hwerrmsgs[i].msg);
211 /* return the strings for the most common link states */
212 static char *ib_linkstate(u32 linkstate)
217 case IPATH_IBSTATE_INIT:
220 case IPATH_IBSTATE_ARM:
223 case IPATH_IBSTATE_ACTIVE:
233 static void handle_e_ibstatuschanged(struct ipath_devdata *dd,
234 ipath_err_t errs, int noprint)
240 * even if diags are enabled, we want to notice LINKINIT, etc.
241 * We just don't want to change the LED state, or
242 * dd->ipath_kregs->kr_ibcctrl
244 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
245 lstate = val & IPATH_IBSTATE_MASK;
248 * this is confusing enough when it happens that I want to always put it
249 * on the console and in the logs. If it was a requested state change,
250 * we'll have already cleared the flags, so we won't print this warning
252 if ((lstate != IPATH_IBSTATE_ARM && lstate != IPATH_IBSTATE_ACTIVE)
253 && (dd->ipath_flags & (IPATH_LINKARMED | IPATH_LINKACTIVE))) {
254 dev_info(&dd->pcidev->dev, "Link state changed from %s to %s\n",
255 (dd->ipath_flags & IPATH_LINKARMED) ? "ARM" : "ACTIVE",
256 ib_linkstate(lstate));
258 * Flush all queued sends when link went to DOWN or INIT,
259 * to be sure that they don't block SMA and other MAD packets
261 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
263 ipath_disarm_piobufs(dd, dd->ipath_lastport_piobuf,
264 (unsigned)(dd->ipath_piobcnt2k +
265 dd->ipath_piobcnt4k) -
266 dd->ipath_lastport_piobuf);
268 else if (lstate == IPATH_IBSTATE_INIT || lstate == IPATH_IBSTATE_ARM ||
269 lstate == IPATH_IBSTATE_ACTIVE) {
271 * only print at SMA if there is a change, debug if not
272 * (sometimes we want to know that, usually not).
274 if (lstate == ((unsigned) dd->ipath_lastibcstat
275 & IPATH_IBSTATE_MASK)) {
276 ipath_dbg("Status change intr but no change (%s)\n",
277 ib_linkstate(lstate));
280 ipath_cdbg(VERBOSE, "Unit %u link state %s, last "
281 "was %s\n", dd->ipath_unit,
282 ib_linkstate(lstate),
283 ib_linkstate((unsigned)
284 dd->ipath_lastibcstat
285 & IPATH_IBSTATE_MASK));
288 lstate = dd->ipath_lastibcstat & IPATH_IBSTATE_MASK;
289 if (lstate == IPATH_IBSTATE_INIT ||
290 lstate == IPATH_IBSTATE_ARM ||
291 lstate == IPATH_IBSTATE_ACTIVE)
292 ipath_cdbg(VERBOSE, "Unit %u link state down"
293 " (state 0x%x), from %s\n",
295 (u32)val & IPATH_IBSTATE_MASK,
296 ib_linkstate(lstate));
298 ipath_cdbg(VERBOSE, "Unit %u link state changed "
299 "to 0x%x from down (%x)\n",
300 dd->ipath_unit, (u32) val, lstate);
302 ltstate = (val >> INFINIPATH_IBCS_LINKTRAININGSTATE_SHIFT) &
303 INFINIPATH_IBCS_LINKTRAININGSTATE_MASK;
304 lstate = (val >> INFINIPATH_IBCS_LINKSTATE_SHIFT) &
305 INFINIPATH_IBCS_LINKSTATE_MASK;
307 if (ltstate == INFINIPATH_IBCS_LT_STATE_POLLACTIVE ||
308 ltstate == INFINIPATH_IBCS_LT_STATE_POLLQUIET) {
312 * Ignore cycling back and forth from Polling.Active
313 * to Polling.Quiet while waiting for the other end of
314 * the link to come up. We will cycle back and forth
315 * between them if no cable is plugged in,
316 * the other device is powered off or disabled, etc.
318 last_ltstate = (dd->ipath_lastibcstat >>
319 INFINIPATH_IBCS_LINKTRAININGSTATE_SHIFT)
320 & INFINIPATH_IBCS_LINKTRAININGSTATE_MASK;
321 if (last_ltstate == INFINIPATH_IBCS_LT_STATE_POLLACTIVE
323 INFINIPATH_IBCS_LT_STATE_POLLQUIET) {
324 if (dd->ipath_ibpollcnt > 40) {
325 dd->ipath_flags |= IPATH_NOCABLE;
326 *dd->ipath_statusp |=
327 IPATH_STATUS_IB_NOCABLE;
329 dd->ipath_ibpollcnt++;
333 dd->ipath_ibpollcnt = 0; /* some state other than 2 or 3 */
334 ipath_stats.sps_iblink++;
335 if (ltstate != INFINIPATH_IBCS_LT_STATE_LINKUP) {
336 dd->ipath_flags |= IPATH_LINKDOWN;
337 dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
340 *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
341 dd->ipath_lli_counter = 0;
343 if (((dd->ipath_lastibcstat >>
344 INFINIPATH_IBCS_LINKSTATE_SHIFT) &
345 INFINIPATH_IBCS_LINKSTATE_MASK)
346 == INFINIPATH_IBCS_L_STATE_ACTIVE)
347 /* if from up to down be more vocal */
349 "Unit %u link now down (%s)\n",
351 ipath_ibcstatus_str[ltstate]);
353 ipath_cdbg(VERBOSE, "Unit %u link is "
354 "down (%s)\n", dd->ipath_unit,
355 ipath_ibcstatus_str[ltstate]);
358 dd->ipath_f_setextled(dd, lstate, ltstate);
359 } else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_ACTIVE) {
360 dd->ipath_flags |= IPATH_LINKACTIVE;
362 ~(IPATH_LINKUNK | IPATH_LINKINIT | IPATH_LINKDOWN |
363 IPATH_LINKARMED | IPATH_NOCABLE);
364 *dd->ipath_statusp &= ~IPATH_STATUS_IB_NOCABLE;
365 *dd->ipath_statusp |=
366 IPATH_STATUS_IB_READY | IPATH_STATUS_IB_CONF;
367 dd->ipath_f_setextled(dd, lstate, ltstate);
368 } else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_INIT) {
370 * set INIT and DOWN. Down is checked by most of the other
371 * code, but INIT is useful to know in a few places.
373 dd->ipath_flags |= IPATH_LINKINIT | IPATH_LINKDOWN;
375 ~(IPATH_LINKUNK | IPATH_LINKACTIVE | IPATH_LINKARMED
377 *dd->ipath_statusp &= ~(IPATH_STATUS_IB_NOCABLE
378 | IPATH_STATUS_IB_READY);
379 dd->ipath_f_setextled(dd, lstate, ltstate);
380 } else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_ARM) {
381 dd->ipath_flags |= IPATH_LINKARMED;
383 ~(IPATH_LINKUNK | IPATH_LINKDOWN | IPATH_LINKINIT |
384 IPATH_LINKACTIVE | IPATH_NOCABLE);
385 *dd->ipath_statusp &= ~(IPATH_STATUS_IB_NOCABLE
386 | IPATH_STATUS_IB_READY);
387 dd->ipath_f_setextled(dd, lstate, ltstate);
390 ipath_dbg("IBstatuschange unit %u: %s (%x)\n",
392 ipath_ibcstatus_str[ltstate], ltstate);
395 dd->ipath_lastibcstat = val;
398 static void handle_supp_msgs(struct ipath_devdata *dd,
399 unsigned supp_msgs, char msg[512])
402 * Print the message unless it's ibc status change only, which
403 * happens so often we never want to count it.
405 if (dd->ipath_lasterror & ~INFINIPATH_E_IBSTATUSCHANGED) {
407 iserr = ipath_decode_err(msg, sizeof msg,
408 dd->ipath_lasterror &
409 ~INFINIPATH_E_IBSTATUSCHANGED);
410 if (dd->ipath_lasterror &
411 ~(INFINIPATH_E_RRCVEGRFULL |
412 INFINIPATH_E_RRCVHDRFULL | INFINIPATH_E_PKTERRS))
413 ipath_dev_err(dd, "Suppressed %u messages for "
414 "fast-repeating errors (%s) (%llx)\n",
417 dd->ipath_lasterror);
420 * rcvegrfull and rcvhdrqfull are "normal", for some
421 * types of processes (mostly benchmarks) that send
422 * huge numbers of messages, while not processing
423 * them. So only complain about these at debug
427 ipath_dbg("Suppressed %u messages for %s\n",
431 "Suppressed %u messages for %s\n",
437 static unsigned handle_frequent_errors(struct ipath_devdata *dd,
438 ipath_err_t errs, char msg[512],
442 static unsigned long nextmsg_time;
443 static unsigned nmsgs, supp_msgs;
446 * Throttle back "fast" messages to no more than 10 per 5 seconds.
447 * This isn't perfect, but it's a reasonable heuristic. If we get
448 * more than 10, give a 6x longer delay.
452 if (time_before(nc, nextmsg_time)) {
455 nextmsg_time = nc + HZ * 3;
457 else if (supp_msgs) {
458 handle_supp_msgs(dd, supp_msgs, msg);
463 else if (!nmsgs++ || time_after(nc, nextmsg_time))
464 nextmsg_time = nc + HZ / 2;
469 static int handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
472 u64 ignore_this_time = 0;
474 int chkerrpkts = 0, noprint = 0;
477 supp_msgs = handle_frequent_errors(dd, errs, msg, &noprint);
480 * don't report errors that are masked (includes those always
483 errs &= ~dd->ipath_maskederrs;
485 /* do these first, they are most important */
486 if (errs & INFINIPATH_E_HARDWARE) {
487 /* reuse same msg buf */
488 dd->ipath_f_handle_hwerrors(dd, msg, sizeof msg);
491 if (!noprint && (errs & ~dd->ipath_e_bitsextant))
492 ipath_dev_err(dd, "error interrupt with unknown errors "
493 "%llx set\n", (unsigned long long)
494 (errs & ~dd->ipath_e_bitsextant));
496 if (errs & E_SUM_ERRS)
497 ignore_this_time = handle_e_sum_errs(dd, errs);
498 else if ((errs & E_SUM_LINK_PKTERRS) &&
499 !(dd->ipath_flags & IPATH_LINKACTIVE)) {
501 * This can happen when SMA is trying to bring the link
502 * up, but the IB link changes state at the "wrong" time.
503 * The IB logic then complains that the packet isn't
504 * valid. We don't want to confuse people, so we just
505 * don't print them, except at debug
507 ipath_dbg("Ignoring packet errors %llx, because link not "
508 "ACTIVE\n", (unsigned long long) errs);
509 ignore_this_time = errs & E_SUM_LINK_PKTERRS;
512 if (supp_msgs == 250000) {
515 * It's not entirely reasonable assuming that the errors set
516 * in the last clear period are all responsible for the
517 * problem, but the alternative is to assume it's the only
518 * ones on this particular interrupt, which also isn't great
520 dd->ipath_maskederrs |= dd->ipath_lasterror | errs;
521 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
522 ~dd->ipath_maskederrs);
523 s_iserr = ipath_decode_err(msg, sizeof msg,
524 (dd->ipath_maskederrs & ~dd->
527 if ((dd->ipath_maskederrs & ~dd->ipath_ignorederrs) &
528 ~(INFINIPATH_E_RRCVEGRFULL |
529 INFINIPATH_E_RRCVHDRFULL | INFINIPATH_E_PKTERRS))
530 ipath_dev_err(dd, "Temporarily disabling "
531 "error(s) %llx reporting; too frequent (%s)\n",
532 (unsigned long long) (dd->ipath_maskederrs &
533 ~dd->ipath_ignorederrs), msg);
536 * rcvegrfull and rcvhdrqfull are "normal",
537 * for some types of processes (mostly benchmarks)
538 * that send huge numbers of messages, while not
539 * processing them. So only complain about
540 * these at debug level.
543 ipath_dbg("Temporarily disabling reporting "
544 "too frequent queue full errors (%s)\n",
548 "Temporarily disabling reporting too"
549 " frequent packet errors (%s)\n",
554 * Re-enable the masked errors after around 3 minutes. in
555 * ipath_get_faststats(). If we have a series of fast
556 * repeating but different errors, the interval will keep
557 * stretching out, but that's OK, as that's pretty
560 dd->ipath_unmasktime = jiffies + HZ * 180;
563 ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, errs);
564 if (ignore_this_time)
565 errs &= ~ignore_this_time;
566 if (errs & ~dd->ipath_lasterror) {
567 errs &= ~dd->ipath_lasterror;
568 /* never suppress duplicate hwerrors or ibstatuschange */
569 dd->ipath_lasterror |= errs &
570 ~(INFINIPATH_E_HARDWARE |
571 INFINIPATH_E_IBSTATUSCHANGED);
574 /* likely due to cancel, so suppress */
575 if ((errs & (INFINIPATH_E_SPKTLEN | INFINIPATH_E_SPIOARMLAUNCH)) &&
576 dd->ipath_lastcancel > jiffies) {
577 ipath_dbg("Suppressed armlaunch/spktlen after error send cancel\n");
578 errs &= ~(INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SPKTLEN);
586 * the ones we mask off are handled specially below or above
588 ipath_decode_err(msg, sizeof msg,
589 errs & ~(INFINIPATH_E_IBSTATUSCHANGED |
590 INFINIPATH_E_RRCVEGRFULL |
591 INFINIPATH_E_RRCVHDRFULL |
592 INFINIPATH_E_HARDWARE));
594 /* so we don't need if (!noprint) at strlcat's below */
597 if (errs & E_SUM_PKTERRS) {
598 ipath_stats.sps_pkterrs++;
601 if (errs & E_SUM_ERRS)
602 ipath_stats.sps_errs++;
604 if (errs & (INFINIPATH_E_RICRC | INFINIPATH_E_RVCRC)) {
605 ipath_stats.sps_crcerrs++;
608 iserr = errs & ~(E_SUM_PKTERRS | INFINIPATH_E_PKTERRS);
612 * We don't want to print these two as they happen, or we can make
613 * the situation even worse, because it takes so long to print
614 * messages to serial consoles. Kernel ports get printed from
615 * fast_stats, no more than every 5 seconds, user ports get printed
618 if (errs & INFINIPATH_E_RRCVHDRFULL) {
620 ipath_stats.sps_hdrqfull++;
621 for (i = 0; i < dd->ipath_cfgports; i++) {
622 struct ipath_portdata *pd = dd->ipath_pd[i];
624 hd = dd->ipath_port0head;
625 tl = (u32) le64_to_cpu(
626 *dd->ipath_hdrqtailptr);
627 } else if (pd && pd->port_cnt &&
628 pd->port_rcvhdrtail_kvaddr) {
630 * don't report same point multiple times,
633 tl = *(u64 *) pd->port_rcvhdrtail_kvaddr;
634 if (tl == dd->ipath_lastrcvhdrqtails[i])
636 hd = ipath_read_ureg32(dd, ur_rcvhdrhead,
640 if (hd == (tl + 1) ||
641 (!hd && tl == dd->ipath_hdrqlast)) {
644 dd->ipath_lastrcvhdrqtails[i] = tl;
649 if (errs & INFINIPATH_E_RRCVEGRFULL) {
651 * since this is of less importance and not likely to
652 * happen without also getting hdrfull, only count
653 * occurrences; don't check each port (or even the kernel
656 ipath_stats.sps_etidfull++;
657 if (dd->ipath_port0head !=
658 (u32) le64_to_cpu(*dd->ipath_hdrqtailptr))
663 * do this before IBSTATUSCHANGED, in case both bits set in a single
664 * interrupt; we want the STATUSCHANGE to "win", so we do our
665 * internal copy of state machine correctly
667 if (errs & INFINIPATH_E_RIBLOSTLINK) {
669 * force through block below
671 errs |= INFINIPATH_E_IBSTATUSCHANGED;
672 ipath_stats.sps_iblink++;
673 dd->ipath_flags |= IPATH_LINKDOWN;
674 dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
675 | IPATH_LINKARMED | IPATH_LINKACTIVE);
676 *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
678 u64 st = ipath_read_kreg64(
679 dd, dd->ipath_kregs->kr_ibcstatus);
681 ipath_dbg("Lost link, link now down (%s)\n",
682 ipath_ibcstatus_str[st & 0xf]);
685 if (errs & INFINIPATH_E_IBSTATUSCHANGED)
686 handle_e_ibstatuschanged(dd, errs, noprint);
688 if (errs & INFINIPATH_E_RESET) {
690 ipath_dev_err(dd, "Got reset, requires re-init "
691 "(unload and reload driver)\n");
692 dd->ipath_flags &= ~IPATH_INITTED; /* needs re-init */
693 /* mark as having had error */
694 *dd->ipath_statusp |= IPATH_STATUS_HWERROR;
695 *dd->ipath_statusp &= ~IPATH_STATUS_IB_CONF;
698 if (!noprint && *msg) {
700 ipath_dev_err(dd, "%s error\n", msg);
702 dev_info(&dd->pcidev->dev, "%s packet problems\n",
705 if (dd->ipath_state_wanted & dd->ipath_flags) {
706 ipath_cdbg(VERBOSE, "driver wanted state %x, iflags now %x, "
707 "waking\n", dd->ipath_state_wanted,
709 wake_up_interruptible(&ipath_state_wait);
715 /* this is separate to allow for better optimization of ipath_intr() */
717 static void ipath_bad_intr(struct ipath_devdata *dd, u32 * unexpectp)
720 * sometimes happen during driver init and unload, don't want
721 * to process any interrupts at that point
724 /* this is just a bandaid, not a fix, if something goes badly
726 if (++*unexpectp > 100) {
727 if (++*unexpectp > 105) {
729 * ok, we must be taking somebody else's interrupts,
730 * due to a messed up mptable and/or PIRQ table, so
731 * unregister the interrupt. We've seen this during
732 * linuxbios development work, and it may happen in
735 if (dd->pcidev && dd->ipath_irq) {
736 ipath_dev_err(dd, "Now %u unexpected "
737 "interrupts, unregistering "
738 "interrupt handler\n",
740 ipath_dbg("free_irq of irq %d\n",
742 dd->ipath_f_free_irq(dd);
745 if (ipath_read_kreg32(dd, dd->ipath_kregs->kr_intmask)) {
746 ipath_dev_err(dd, "%u unexpected interrupts, "
747 "disabling interrupts completely\n",
750 * disable all interrupts, something is very wrong
752 ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
755 } else if (*unexpectp > 1)
756 ipath_dbg("Interrupt when not ready, should not happen, "
760 static void ipath_bad_regread(struct ipath_devdata *dd)
764 /* separate routine, for better optimization of ipath_intr() */
767 * We print the message and disable interrupts, in hope of
768 * having a better chance of debugging the problem.
771 "Read of interrupt status failed (all bits set)\n");
773 /* disable all interrupts, something is very wrong */
774 ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
776 ipath_dev_err(dd, "Still bad interrupt status, "
777 "unregistering interrupt\n");
778 dd->ipath_f_free_irq(dd);
779 } else if (allbits > 2) {
780 if ((allbits % 10000) == 0)
783 ipath_dev_err(dd, "Disabling interrupts, "
784 "multiple errors\n");
788 static void handle_port_pioavail(struct ipath_devdata *dd)
792 * start from port 1, since for now port 0 is never using
795 for (i = 1; dd->ipath_portpiowait && i < dd->ipath_cfgports; i++) {
796 struct ipath_portdata *pd = dd->ipath_pd[i];
798 if (pd && pd->port_cnt &&
799 dd->ipath_portpiowait & (1U << i)) {
800 clear_bit(i, &dd->ipath_portpiowait);
801 if (test_bit(IPATH_PORT_WAITING_PIO,
803 clear_bit(IPATH_PORT_WAITING_PIO,
805 wake_up_interruptible(&pd->port_wait);
811 static void handle_layer_pioavail(struct ipath_devdata *dd)
815 ret = ipath_ib_piobufavail(dd->verbs_dev);
821 set_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl);
822 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
827 * Handle receive interrupts for user ports; this means a user
828 * process was waiting for a packet to arrive, and didn't want
831 static void handle_urcv(struct ipath_devdata *dd, u32 istat)
837 portr = ((istat >> INFINIPATH_I_RCVAVAIL_SHIFT) &
838 dd->ipath_i_rcvavail_mask)
839 | ((istat >> INFINIPATH_I_RCVURG_SHIFT) &
840 dd->ipath_i_rcvurg_mask);
841 for (i = 1; i < dd->ipath_cfgports; i++) {
842 struct ipath_portdata *pd = dd->ipath_pd[i];
843 if (portr & (1 << i) && pd && pd->port_cnt &&
844 test_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag)) {
845 clear_bit(IPATH_PORT_WAITING_RCV,
847 clear_bit(i + INFINIPATH_R_INTRAVAIL_SHIFT,
849 wake_up_interruptible(&pd->port_wait);
854 /* only want to take one interrupt, so turn off the rcv
855 * interrupt for all the ports that we did the wakeup on
856 * (but never for kernel port)
858 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
863 irqreturn_t ipath_intr(int irq, void *data)
865 struct ipath_devdata *dd = data;
866 u32 istat, chk0rcv = 0;
867 ipath_err_t estat = 0;
869 u32 oldhead, curtail;
870 static unsigned unexpected = 0;
871 static const u32 port0rbits = (1U<<INFINIPATH_I_RCVAVAIL_SHIFT) |
872 (1U<<INFINIPATH_I_RCVURG_SHIFT);
874 ipath_stats.sps_ints++;
876 if (!(dd->ipath_flags & IPATH_PRESENT)) {
878 * This return value is not great, but we do not want the
879 * interrupt core code to remove our interrupt handler
880 * because we don't appear to be handling an interrupt
881 * during a chip reset.
887 * this needs to be flags&initted, not statusp, so we keep
888 * taking interrupts even after link goes down, etc.
889 * Also, we *must* clear the interrupt at some point, or we won't
890 * take it again, which can be real bad for errors, etc...
893 if (!(dd->ipath_flags & IPATH_INITTED)) {
894 ipath_bad_intr(dd, &unexpected);
900 * We try to avoid reading the interrupt status register, since
901 * that's a PIO read, and stalls the processor for up to about
902 * ~0.25 usec. The idea is that if we processed a port0 packet,
903 * we blindly clear the port 0 receive interrupt bits, and nothing
904 * else, then return. If other interrupts are pending, the chip
905 * will re-interrupt us as soon as we write the intclear register.
906 * We then won't process any more kernel packets (if not the 2nd
907 * time, then the 3rd or 4th) and we'll then handle the other
908 * interrupts. We clear the interrupts first so that we don't
909 * lose intr for later packets that arrive while we are processing.
911 oldhead = dd->ipath_port0head;
912 curtail = (u32)le64_to_cpu(*dd->ipath_hdrqtailptr);
913 if (oldhead != curtail) {
914 if (dd->ipath_flags & IPATH_GPIO_INTR) {
915 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
916 (u64) (1 << IPATH_GPIO_PORT0_BIT));
917 istat = port0rbits | INFINIPATH_I_GPIO;
921 ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
923 if (oldhead != dd->ipath_port0head) {
924 ipath_stats.sps_fastrcvint++;
929 istat = ipath_read_kreg32(dd, dd->ipath_kregs->kr_intstatus);
931 if (unlikely(!istat)) {
932 ipath_stats.sps_nullintr++;
933 ret = IRQ_NONE; /* not our interrupt, or already handled */
936 if (unlikely(istat == -1)) {
937 ipath_bad_regread(dd);
938 /* don't know if it was our interrupt or not */
946 if (unlikely(istat & ~dd->ipath_i_bitsextant))
948 "interrupt with unknown interrupts %x set\n",
949 istat & (u32) ~ dd->ipath_i_bitsextant);
951 ipath_cdbg(VERBOSE, "intr stat=0x%x\n", istat);
953 if (unlikely(istat & INFINIPATH_I_ERROR)) {
954 ipath_stats.sps_errints++;
955 estat = ipath_read_kreg64(dd,
956 dd->ipath_kregs->kr_errorstatus);
958 dev_info(&dd->pcidev->dev, "error interrupt (%x), "
959 "but no error bits set!\n", istat);
960 else if (estat == -1LL)
962 * should we try clearing all, or hope next read
965 ipath_dev_err(dd, "Read of error status failed "
966 "(all bits set); ignoring\n");
968 if (handle_errors(dd, estat))
969 /* force calling ipath_kreceive() */
973 if (istat & INFINIPATH_I_GPIO) {
975 * GPIO interrupts fall in two broad classes:
976 * GPIO_2 indicates (on some HT4xx boards) that a packet
977 * has arrived for Port 0. Checking for this
978 * is controlled by flag IPATH_GPIO_INTR.
979 * GPIO_3..5 on IBA6120 Rev2 chips indicate errors
980 * that we need to count. Checking for this
981 * is controlled by flag IPATH_GPIO_ERRINTRS.
986 gpiostatus = ipath_read_kreg32(
987 dd, dd->ipath_kregs->kr_gpio_status);
988 /* First the error-counter case.
990 if ((gpiostatus & IPATH_GPIO_ERRINTR_MASK) &&
991 (dd->ipath_flags & IPATH_GPIO_ERRINTRS)) {
992 /* want to clear the bits we see asserted. */
993 to_clear |= (gpiostatus & IPATH_GPIO_ERRINTR_MASK);
996 * Count appropriately, clear bits out of our copy,
997 * as they have been "handled".
999 if (gpiostatus & (1 << IPATH_GPIO_RXUVL_BIT)) {
1000 ipath_dbg("FlowCtl on UnsupVL\n");
1001 dd->ipath_rxfc_unsupvl_errs++;
1003 if (gpiostatus & (1 << IPATH_GPIO_OVRUN_BIT)) {
1004 ipath_dbg("Overrun Threshold exceeded\n");
1005 dd->ipath_overrun_thresh_errs++;
1007 if (gpiostatus & (1 << IPATH_GPIO_LLI_BIT)) {
1008 ipath_dbg("Local Link Integrity error\n");
1009 dd->ipath_lli_errs++;
1011 gpiostatus &= ~IPATH_GPIO_ERRINTR_MASK;
1013 /* Now the Port0 Receive case */
1014 if ((gpiostatus & (1 << IPATH_GPIO_PORT0_BIT)) &&
1015 (dd->ipath_flags & IPATH_GPIO_INTR)) {
1017 * GPIO status bit 2 is set, and we expected it.
1018 * clear it and indicate in p0bits.
1019 * This probably only happens if a Port0 pkt
1020 * arrives at _just_ the wrong time, and we
1021 * handle that by seting chk0rcv;
1023 to_clear |= (1 << IPATH_GPIO_PORT0_BIT);
1024 gpiostatus &= ~(1 << IPATH_GPIO_PORT0_BIT);
1027 if (unlikely(gpiostatus)) {
1029 * Some unexpected bits remain. If they could have
1030 * caused the interrupt, complain and clear.
1031 * MEA: this is almost certainly non-ideal.
1032 * we should look into auto-disable of unexpected
1033 * GPIO interrupts, possibly on a "three strikes"
1037 mask = ipath_read_kreg32(
1038 dd, dd->ipath_kregs->kr_gpio_mask);
1039 if (mask & gpiostatus) {
1040 ipath_dbg("Unexpected GPIO IRQ bits %x\n",
1042 to_clear |= (gpiostatus & mask);
1046 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
1050 chk0rcv |= istat & port0rbits;
1053 * Clear the interrupt bits we found set, unless they are receive
1054 * related, in which case we already cleared them above, and don't
1055 * want to clear them again, because we might lose an interrupt.
1056 * Clear it early, so we "know" know the chip will have seen this by
1057 * the time we process the queue, and will re-interrupt if necessary.
1058 * The processor itself won't take the interrupt again until we return.
1060 ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
1063 * handle port0 receive before checking for pio buffers available,
1064 * since receives can overflow; piobuf waiters can afford a few
1065 * extra cycles, since they were waiting anyway, and user's waiting
1066 * for receive are at the bottom.
1070 istat &= ~port0rbits;
1073 if (istat & ((dd->ipath_i_rcvavail_mask <<
1074 INFINIPATH_I_RCVAVAIL_SHIFT)
1075 | (dd->ipath_i_rcvurg_mask <<
1076 INFINIPATH_I_RCVURG_SHIFT)))
1077 handle_urcv(dd, istat);
1079 if (istat & INFINIPATH_I_SPIOBUFAVAIL) {
1080 clear_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl);
1081 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1082 dd->ipath_sendctrl);
1084 if (dd->ipath_portpiowait)
1085 handle_port_pioavail(dd);
1087 handle_layer_pioavail(dd);