ipath_intr.c revision 78d1e02fac0595a8aa8a5064d1bd0c0ea55b22b0
1/*
2 * Copyright (c) 2006, 2007 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
36#include "ipath_kernel.h"
37#include "ipath_verbs.h"
38#include "ipath_common.h"
39
40/*
41 * clear (write) a pio buffer, to clear a parity error.   This routine
42 * should only be called when in freeze mode, and the buffer should be
43 * canceled afterwards.
44 */
45static void ipath_clrpiobuf(struct ipath_devdata *dd, u32 pnum)
46{
47	u32 __iomem *pbuf;
48	u32 dwcnt; /* dword count to write */
49	if (pnum < dd->ipath_piobcnt2k) {
50		pbuf = (u32 __iomem *) (dd->ipath_pio2kbase + pnum *
51			dd->ipath_palign);
52		dwcnt = dd->ipath_piosize2k >> 2;
53	}
54	else {
55		pbuf = (u32 __iomem *) (dd->ipath_pio4kbase +
56			(pnum - dd->ipath_piobcnt2k) * dd->ipath_4kalign);
57		dwcnt = dd->ipath_piosize4k >> 2;
58	}
59	dev_info(&dd->pcidev->dev,
60		"Rewrite PIO buffer %u, to recover from parity error\n",
61		pnum);
62	*pbuf = dwcnt+1; /* no flush required, since already in freeze */
63	while(--dwcnt)
64		*pbuf++ = 0;
65}
66
67/*
68 * Called when we might have an error that is specific to a particular
69 * PIO buffer, and may need to cancel that buffer, so it can be re-used.
70 * If rewrite is true, and bits are set in the sendbufferror registers,
71 * we'll write to the buffer, for error recovery on parity errors.
72 */
73static void ipath_disarm_senderrbufs(struct ipath_devdata *dd, int rewrite)
74{
75	u32 piobcnt;
76	unsigned long sbuf[4];
77	/*
78	 * it's possible that sendbuffererror could have bits set; might
79	 * have already done this as a result of hardware error handling
80	 */
81	piobcnt = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
82	/* read these before writing errorclear */
83	sbuf[0] = ipath_read_kreg64(
84		dd, dd->ipath_kregs->kr_sendbuffererror);
85	sbuf[1] = ipath_read_kreg64(
86		dd, dd->ipath_kregs->kr_sendbuffererror + 1);
87	if (piobcnt > 128) {
88		sbuf[2] = ipath_read_kreg64(
89			dd, dd->ipath_kregs->kr_sendbuffererror + 2);
90		sbuf[3] = ipath_read_kreg64(
91			dd, dd->ipath_kregs->kr_sendbuffererror + 3);
92	}
93
94	if (sbuf[0] || sbuf[1] || (piobcnt > 128 && (sbuf[2] || sbuf[3]))) {
95		int i;
96		if (ipath_debug & (__IPATH_PKTDBG|__IPATH_DBG) &&
97			dd->ipath_lastcancel > jiffies) {
98			__IPATH_DBG_WHICH(__IPATH_PKTDBG|__IPATH_DBG,
99					  "SendbufErrs %lx %lx", sbuf[0],
100					  sbuf[1]);
101			if (ipath_debug & __IPATH_PKTDBG && piobcnt > 128)
102				printk(" %lx %lx ", sbuf[2], sbuf[3]);
103			printk("\n");
104		}
105
106		for (i = 0; i < piobcnt; i++)
107			if (test_bit(i, sbuf)) {
108				if (rewrite)
109					ipath_clrpiobuf(dd, i);
110				ipath_disarm_piobufs(dd, i, 1);
111			}
112		/* ignore armlaunch errs for a bit */
113		dd->ipath_lastcancel = jiffies+3;
114	}
115}
116
117
118/* These are all rcv-related errors which we want to count for stats */
119#define E_SUM_PKTERRS \
120	(INFINIPATH_E_RHDRLEN | INFINIPATH_E_RBADTID | \
121	 INFINIPATH_E_RBADVERSION | INFINIPATH_E_RHDR | \
122	 INFINIPATH_E_RLONGPKTLEN | INFINIPATH_E_RSHORTPKTLEN | \
123	 INFINIPATH_E_RMAXPKTLEN | INFINIPATH_E_RMINPKTLEN | \
124	 INFINIPATH_E_RFORMATERR | INFINIPATH_E_RUNSUPVL | \
125	 INFINIPATH_E_RUNEXPCHAR | INFINIPATH_E_REBP)
126
127/* These are all send-related errors which we want to count for stats */
128#define E_SUM_ERRS \
129	(INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SUNEXPERRPKTNUM | \
130	 INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
131	 INFINIPATH_E_SMAXPKTLEN | INFINIPATH_E_SUNSUPVL | \
132	 INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
133	 INFINIPATH_E_INVALIDADDR)
134
135/*
136 * this is similar to E_SUM_ERRS, but can't ignore armlaunch, don't ignore
137 * errors not related to freeze and cancelling buffers.  Can't ignore
138 * armlaunch because could get more while still cleaning up, and need
139 * to cancel those as they happen.
140 */
141#define E_SPKT_ERRS_IGNORE \
142	 (INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
143	 INFINIPATH_E_SMAXPKTLEN | INFINIPATH_E_SMINPKTLEN | \
144	 INFINIPATH_E_SPKTLEN)
145
146/*
147 * these are errors that can occur when the link changes state while
148 * a packet is being sent or received.  This doesn't cover things
149 * like EBP or VCRC that can be the result of a sending having the
150 * link change state, so we receive a "known bad" packet.
151 */
152#define E_SUM_LINK_PKTERRS \
153	(INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
154	 INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
155	 INFINIPATH_E_RSHORTPKTLEN | INFINIPATH_E_RMINPKTLEN | \
156	 INFINIPATH_E_RUNEXPCHAR)
157
158static u64 handle_e_sum_errs(struct ipath_devdata *dd, ipath_err_t errs)
159{
160	u64 ignore_this_time = 0;
161
162	ipath_disarm_senderrbufs(dd, 0);
163	if ((errs & E_SUM_LINK_PKTERRS) &&
164	    !(dd->ipath_flags & IPATH_LINKACTIVE)) {
165		/*
166		 * This can happen when SMA is trying to bring the link
167		 * up, but the IB link changes state at the "wrong" time.
168		 * The IB logic then complains that the packet isn't
169		 * valid.  We don't want to confuse people, so we just
170		 * don't print them, except at debug
171		 */
172		ipath_dbg("Ignoring packet errors %llx, because link not "
173			  "ACTIVE\n", (unsigned long long) errs);
174		ignore_this_time = errs & E_SUM_LINK_PKTERRS;
175	}
176
177	return ignore_this_time;
178}
179
180/* generic hw error messages... */
181#define INFINIPATH_HWE_TXEMEMPARITYERR_MSG(a) \
182	{ \
183		.mask = ( INFINIPATH_HWE_TXEMEMPARITYERR_##a <<    \
184			  INFINIPATH_HWE_TXEMEMPARITYERR_SHIFT ),   \
185		.msg = "TXE " #a " Memory Parity"	     \
186	}
187#define INFINIPATH_HWE_RXEMEMPARITYERR_MSG(a) \
188	{ \
189		.mask = ( INFINIPATH_HWE_RXEMEMPARITYERR_##a <<    \
190			  INFINIPATH_HWE_RXEMEMPARITYERR_SHIFT ),   \
191		.msg = "RXE " #a " Memory Parity"	     \
192	}
193
194static const struct ipath_hwerror_msgs ipath_generic_hwerror_msgs[] = {
195	INFINIPATH_HWE_MSG(IBCBUSFRSPCPARITYERR, "IPATH2IB Parity"),
196	INFINIPATH_HWE_MSG(IBCBUSTOSPCPARITYERR, "IB2IPATH Parity"),
197
198	INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOBUF),
199	INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOPBC),
200	INFINIPATH_HWE_TXEMEMPARITYERR_MSG(PIOLAUNCHFIFO),
201
202	INFINIPATH_HWE_RXEMEMPARITYERR_MSG(RCVBUF),
203	INFINIPATH_HWE_RXEMEMPARITYERR_MSG(LOOKUPQ),
204	INFINIPATH_HWE_RXEMEMPARITYERR_MSG(EAGERTID),
205	INFINIPATH_HWE_RXEMEMPARITYERR_MSG(EXPTID),
206	INFINIPATH_HWE_RXEMEMPARITYERR_MSG(FLAGBUF),
207	INFINIPATH_HWE_RXEMEMPARITYERR_MSG(DATAINFO),
208	INFINIPATH_HWE_RXEMEMPARITYERR_MSG(HDRINFO),
209};
210
211/**
212 * ipath_format_hwmsg - format a single hwerror message
213 * @msg message buffer
214 * @msgl length of message buffer
215 * @hwmsg message to add to message buffer
216 */
217static void ipath_format_hwmsg(char *msg, size_t msgl, const char *hwmsg)
218{
219	strlcat(msg, "[", msgl);
220	strlcat(msg, hwmsg, msgl);
221	strlcat(msg, "]", msgl);
222}
223
224/**
225 * ipath_format_hwerrors - format hardware error messages for display
226 * @hwerrs hardware errors bit vector
227 * @hwerrmsgs hardware error descriptions
228 * @nhwerrmsgs number of hwerrmsgs
229 * @msg message buffer
230 * @msgl message buffer length
231 */
232void ipath_format_hwerrors(u64 hwerrs,
233			   const struct ipath_hwerror_msgs *hwerrmsgs,
234			   size_t nhwerrmsgs,
235			   char *msg, size_t msgl)
236{
237	int i;
238	const int glen =
239	    sizeof(ipath_generic_hwerror_msgs) /
240	    sizeof(ipath_generic_hwerror_msgs[0]);
241
242	for (i=0; i<glen; i++) {
243		if (hwerrs & ipath_generic_hwerror_msgs[i].mask) {
244			ipath_format_hwmsg(msg, msgl,
245					   ipath_generic_hwerror_msgs[i].msg);
246		}
247	}
248
249	for (i=0; i<nhwerrmsgs; i++) {
250		if (hwerrs & hwerrmsgs[i].mask) {
251			ipath_format_hwmsg(msg, msgl, hwerrmsgs[i].msg);
252		}
253	}
254}
255
256/* return the strings for the most common link states */
257static char *ib_linkstate(u32 linkstate)
258{
259	char *ret;
260
261	switch (linkstate) {
262	case IPATH_IBSTATE_INIT:
263		ret = "Init";
264		break;
265	case IPATH_IBSTATE_ARM:
266		ret = "Arm";
267		break;
268	case IPATH_IBSTATE_ACTIVE:
269		ret = "Active";
270		break;
271	default:
272		ret = "Down";
273	}
274
275	return ret;
276}
277
278static void handle_e_ibstatuschanged(struct ipath_devdata *dd,
279				     ipath_err_t errs, int noprint)
280{
281	u64 val;
282	u32 ltstate, lstate;
283
284	/*
285	 * even if diags are enabled, we want to notice LINKINIT, etc.
286	 * We just don't want to change the LED state, or
287	 * dd->ipath_kregs->kr_ibcctrl
288	 */
289	val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
290	lstate = val & IPATH_IBSTATE_MASK;
291
292	/*
293	 * this is confusing enough when it happens that I want to always put it
294	 * on the console and in the logs.  If it was a requested state change,
295	 * we'll have already cleared the flags, so we won't print this warning
296	 */
297	if ((lstate != IPATH_IBSTATE_ARM && lstate != IPATH_IBSTATE_ACTIVE)
298		&& (dd->ipath_flags & (IPATH_LINKARMED | IPATH_LINKACTIVE))) {
299		dev_info(&dd->pcidev->dev, "Link state changed from %s to %s\n",
300				 (dd->ipath_flags & IPATH_LINKARMED) ? "ARM" : "ACTIVE",
301				 ib_linkstate(lstate));
302		/*
303		 * Flush all queued sends when link went to DOWN or INIT,
304		 * to be sure that they don't block SMA and other MAD packets
305		 */
306		ipath_cancel_sends(dd, 1);
307	}
308	else if (lstate == IPATH_IBSTATE_INIT || lstate == IPATH_IBSTATE_ARM ||
309	    lstate == IPATH_IBSTATE_ACTIVE) {
310		/*
311		 * only print at SMA if there is a change, debug if not
312		 * (sometimes we want to know that, usually not).
313		 */
314		if (lstate == ((unsigned) dd->ipath_lastibcstat
315			       & IPATH_IBSTATE_MASK)) {
316			ipath_dbg("Status change intr but no change (%s)\n",
317				  ib_linkstate(lstate));
318		}
319		else
320			ipath_cdbg(VERBOSE, "Unit %u link state %s, last "
321				   "was %s\n", dd->ipath_unit,
322				   ib_linkstate(lstate),
323				   ib_linkstate((unsigned)
324						dd->ipath_lastibcstat
325						& IPATH_IBSTATE_MASK));
326	}
327	else {
328		lstate = dd->ipath_lastibcstat & IPATH_IBSTATE_MASK;
329		if (lstate == IPATH_IBSTATE_INIT ||
330		    lstate == IPATH_IBSTATE_ARM ||
331		    lstate == IPATH_IBSTATE_ACTIVE)
332			ipath_cdbg(VERBOSE, "Unit %u link state down"
333				   " (state 0x%x), from %s\n",
334				   dd->ipath_unit,
335				   (u32)val & IPATH_IBSTATE_MASK,
336				   ib_linkstate(lstate));
337		else
338			ipath_cdbg(VERBOSE, "Unit %u link state changed "
339				   "to 0x%x from down (%x)\n",
340				   dd->ipath_unit, (u32) val, lstate);
341	}
342	ltstate = (val >> INFINIPATH_IBCS_LINKTRAININGSTATE_SHIFT) &
343		INFINIPATH_IBCS_LINKTRAININGSTATE_MASK;
344	lstate = (val >> INFINIPATH_IBCS_LINKSTATE_SHIFT) &
345		INFINIPATH_IBCS_LINKSTATE_MASK;
346
347	if (ltstate == INFINIPATH_IBCS_LT_STATE_POLLACTIVE ||
348	    ltstate == INFINIPATH_IBCS_LT_STATE_POLLQUIET) {
349		u32 last_ltstate;
350
351		/*
352		 * Ignore cycling back and forth from Polling.Active
353		 * to Polling.Quiet while waiting for the other end of
354		 * the link to come up. We will cycle back and forth
355		 * between them if no cable is plugged in,
356		 * the other device is powered off or disabled, etc.
357		 */
358		last_ltstate = (dd->ipath_lastibcstat >>
359				INFINIPATH_IBCS_LINKTRAININGSTATE_SHIFT)
360			& INFINIPATH_IBCS_LINKTRAININGSTATE_MASK;
361		if (last_ltstate == INFINIPATH_IBCS_LT_STATE_POLLACTIVE
362		    || last_ltstate ==
363		    INFINIPATH_IBCS_LT_STATE_POLLQUIET) {
364			if (dd->ipath_ibpollcnt > 40) {
365				dd->ipath_flags |= IPATH_NOCABLE;
366				*dd->ipath_statusp |=
367					IPATH_STATUS_IB_NOCABLE;
368			} else
369				dd->ipath_ibpollcnt++;
370			goto skip_ibchange;
371		}
372	}
373	dd->ipath_ibpollcnt = 0;	/* some state other than 2 or 3 */
374	ipath_stats.sps_iblink++;
375	if (ltstate != INFINIPATH_IBCS_LT_STATE_LINKUP) {
376		dd->ipath_flags |= IPATH_LINKDOWN;
377		dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
378				     | IPATH_LINKACTIVE |
379				     IPATH_LINKARMED);
380		*dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
381		dd->ipath_lli_counter = 0;
382		if (!noprint) {
383			if (((dd->ipath_lastibcstat >>
384			      INFINIPATH_IBCS_LINKSTATE_SHIFT) &
385			     INFINIPATH_IBCS_LINKSTATE_MASK)
386			    == INFINIPATH_IBCS_L_STATE_ACTIVE)
387				/* if from up to down be more vocal */
388				ipath_cdbg(VERBOSE,
389					   "Unit %u link now down (%s)\n",
390					   dd->ipath_unit,
391					   ipath_ibcstatus_str[ltstate]);
392			else
393				ipath_cdbg(VERBOSE, "Unit %u link is "
394					   "down (%s)\n", dd->ipath_unit,
395					   ipath_ibcstatus_str[ltstate]);
396		}
397
398		dd->ipath_f_setextled(dd, lstate, ltstate);
399	} else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_ACTIVE) {
400		dd->ipath_flags |= IPATH_LINKACTIVE;
401		dd->ipath_flags &=
402			~(IPATH_LINKUNK | IPATH_LINKINIT | IPATH_LINKDOWN |
403			  IPATH_LINKARMED | IPATH_NOCABLE);
404		*dd->ipath_statusp &= ~IPATH_STATUS_IB_NOCABLE;
405		*dd->ipath_statusp |=
406			IPATH_STATUS_IB_READY | IPATH_STATUS_IB_CONF;
407		dd->ipath_f_setextled(dd, lstate, ltstate);
408	} else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_INIT) {
409		/*
410		 * set INIT and DOWN.  Down is checked by most of the other
411		 * code, but INIT is useful to know in a few places.
412		 */
413		dd->ipath_flags |= IPATH_LINKINIT | IPATH_LINKDOWN;
414		dd->ipath_flags &=
415			~(IPATH_LINKUNK | IPATH_LINKACTIVE | IPATH_LINKARMED
416			  | IPATH_NOCABLE);
417		*dd->ipath_statusp &= ~(IPATH_STATUS_IB_NOCABLE
418					| IPATH_STATUS_IB_READY);
419		dd->ipath_f_setextled(dd, lstate, ltstate);
420	} else if ((val & IPATH_IBSTATE_MASK) == IPATH_IBSTATE_ARM) {
421		dd->ipath_flags |= IPATH_LINKARMED;
422		dd->ipath_flags &=
423			~(IPATH_LINKUNK | IPATH_LINKDOWN | IPATH_LINKINIT |
424			  IPATH_LINKACTIVE | IPATH_NOCABLE);
425		*dd->ipath_statusp &= ~(IPATH_STATUS_IB_NOCABLE
426					| IPATH_STATUS_IB_READY);
427		dd->ipath_f_setextled(dd, lstate, ltstate);
428	} else {
429		if (!noprint)
430			ipath_dbg("IBstatuschange unit %u: %s (%x)\n",
431				  dd->ipath_unit,
432				  ipath_ibcstatus_str[ltstate], ltstate);
433	}
434skip_ibchange:
435	dd->ipath_lastibcstat = val;
436}
437
438static void handle_supp_msgs(struct ipath_devdata *dd,
439			     unsigned supp_msgs, char msg[512])
440{
441	/*
442	 * Print the message unless it's ibc status change only, which
443	 * happens so often we never want to count it.
444	 */
445	if (dd->ipath_lasterror & ~INFINIPATH_E_IBSTATUSCHANGED) {
446		int iserr;
447		iserr = ipath_decode_err(msg, sizeof msg,
448				dd->ipath_lasterror &
449				~INFINIPATH_E_IBSTATUSCHANGED);
450		if (dd->ipath_lasterror &
451			~(INFINIPATH_E_RRCVEGRFULL |
452			INFINIPATH_E_RRCVHDRFULL | INFINIPATH_E_PKTERRS))
453			ipath_dev_err(dd, "Suppressed %u messages for "
454				      "fast-repeating errors (%s) (%llx)\n",
455				      supp_msgs, msg,
456				      (unsigned long long)
457				      dd->ipath_lasterror);
458		else {
459			/*
460			 * rcvegrfull and rcvhdrqfull are "normal", for some
461			 * types of processes (mostly benchmarks) that send
462			 * huge numbers of messages, while not processing
463			 * them. So only complain about these at debug
464			 * level.
465			 */
466			if (iserr)
467				ipath_dbg("Suppressed %u messages for %s\n",
468					  supp_msgs, msg);
469			else
470				ipath_cdbg(ERRPKT,
471					"Suppressed %u messages for %s\n",
472					  supp_msgs, msg);
473		}
474	}
475}
476
477static unsigned handle_frequent_errors(struct ipath_devdata *dd,
478				       ipath_err_t errs, char msg[512],
479				       int *noprint)
480{
481	unsigned long nc;
482	static unsigned long nextmsg_time;
483	static unsigned nmsgs, supp_msgs;
484
485	/*
486	 * Throttle back "fast" messages to no more than 10 per 5 seconds.
487	 * This isn't perfect, but it's a reasonable heuristic. If we get
488	 * more than 10, give a 6x longer delay.
489	 */
490	nc = jiffies;
491	if (nmsgs > 10) {
492		if (time_before(nc, nextmsg_time)) {
493			*noprint = 1;
494			if (!supp_msgs++)
495				nextmsg_time = nc + HZ * 3;
496		}
497		else if (supp_msgs) {
498			handle_supp_msgs(dd, supp_msgs, msg);
499			supp_msgs = 0;
500			nmsgs = 0;
501		}
502	}
503	else if (!nmsgs++ || time_after(nc, nextmsg_time))
504		nextmsg_time = nc + HZ / 2;
505
506	return supp_msgs;
507}
508
509static int handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
510{
511	char msg[512];
512	u64 ignore_this_time = 0;
513	int i, iserr = 0;
514	int chkerrpkts = 0, noprint = 0;
515	unsigned supp_msgs;
516	int log_idx;
517
518	supp_msgs = handle_frequent_errors(dd, errs, msg, &noprint);
519
520	/* don't report errors that are masked */
521	errs &= ~dd->ipath_maskederrs;
522
523	/* do these first, they are most important */
524	if (errs & INFINIPATH_E_HARDWARE) {
525		/* reuse same msg buf */
526		dd->ipath_f_handle_hwerrors(dd, msg, sizeof msg);
527	} else {
528		u64 mask;
529		for (log_idx = 0; log_idx < IPATH_EEP_LOG_CNT; ++log_idx) {
530			mask = dd->ipath_eep_st_masks[log_idx].errs_to_log;
531			if (errs & mask)
532				ipath_inc_eeprom_err(dd, log_idx, 1);
533		}
534	}
535
536	if (!noprint && (errs & ~dd->ipath_e_bitsextant))
537		ipath_dev_err(dd, "error interrupt with unknown errors "
538			      "%llx set\n", (unsigned long long)
539			      (errs & ~dd->ipath_e_bitsextant));
540
541	if (errs & E_SUM_ERRS)
542		ignore_this_time = handle_e_sum_errs(dd, errs);
543	else if ((errs & E_SUM_LINK_PKTERRS) &&
544	    !(dd->ipath_flags & IPATH_LINKACTIVE)) {
545		/*
546		 * This can happen when SMA is trying to bring the link
547		 * up, but the IB link changes state at the "wrong" time.
548		 * The IB logic then complains that the packet isn't
549		 * valid.  We don't want to confuse people, so we just
550		 * don't print them, except at debug
551		 */
552		ipath_dbg("Ignoring packet errors %llx, because link not "
553			  "ACTIVE\n", (unsigned long long) errs);
554		ignore_this_time = errs & E_SUM_LINK_PKTERRS;
555	}
556
557	if (supp_msgs == 250000) {
558		int s_iserr;
559		/*
560		 * It's not entirely reasonable assuming that the errors set
561		 * in the last clear period are all responsible for the
562		 * problem, but the alternative is to assume it's the only
563		 * ones on this particular interrupt, which also isn't great
564		 */
565		dd->ipath_maskederrs |= dd->ipath_lasterror | errs;
566		dd->ipath_errormask &= ~dd->ipath_maskederrs;
567		ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
568			dd->ipath_errormask);
569		s_iserr = ipath_decode_err(msg, sizeof msg,
570			dd->ipath_maskederrs);
571
572		if (dd->ipath_maskederrs &
573			~(INFINIPATH_E_RRCVEGRFULL |
574			INFINIPATH_E_RRCVHDRFULL | INFINIPATH_E_PKTERRS))
575			ipath_dev_err(dd, "Temporarily disabling "
576			    "error(s) %llx reporting; too frequent (%s)\n",
577				(unsigned long long)dd->ipath_maskederrs,
578				msg);
579		else {
580			/*
581			 * rcvegrfull and rcvhdrqfull are "normal",
582			 * for some types of processes (mostly benchmarks)
583			 * that send huge numbers of messages, while not
584			 * processing them.  So only complain about
585			 * these at debug level.
586			 */
587			if (s_iserr)
588				ipath_dbg("Temporarily disabling reporting "
589				    "too frequent queue full errors (%s)\n",
590				    msg);
591			else
592				ipath_cdbg(ERRPKT,
593				    "Temporarily disabling reporting too"
594				    " frequent packet errors (%s)\n",
595				    msg);
596		}
597
598		/*
599		 * Re-enable the masked errors after around 3 minutes.  in
600		 * ipath_get_faststats().  If we have a series of fast
601		 * repeating but different errors, the interval will keep
602		 * stretching out, but that's OK, as that's pretty
603		 * catastrophic.
604		 */
605		dd->ipath_unmasktime = jiffies + HZ * 180;
606	}
607
608	ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, errs);
609	if (ignore_this_time)
610		errs &= ~ignore_this_time;
611	if (errs & ~dd->ipath_lasterror) {
612		errs &= ~dd->ipath_lasterror;
613		/* never suppress duplicate hwerrors or ibstatuschange */
614		dd->ipath_lasterror |= errs &
615			~(INFINIPATH_E_HARDWARE |
616			  INFINIPATH_E_IBSTATUSCHANGED);
617	}
618
619	/* likely due to cancel, so suppress */
620	if ((errs & (INFINIPATH_E_SPKTLEN | INFINIPATH_E_SPIOARMLAUNCH)) &&
621		dd->ipath_lastcancel > jiffies) {
622		ipath_dbg("Suppressed armlaunch/spktlen after error send cancel\n");
623		errs &= ~(INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SPKTLEN);
624	}
625
626	if (!errs)
627		return 0;
628
629	if (!noprint)
630		/*
631		 * the ones we mask off are handled specially below or above
632		 */
633		ipath_decode_err(msg, sizeof msg,
634				 errs & ~(INFINIPATH_E_IBSTATUSCHANGED |
635					  INFINIPATH_E_RRCVEGRFULL |
636					  INFINIPATH_E_RRCVHDRFULL |
637					  INFINIPATH_E_HARDWARE));
638	else
639		/* so we don't need if (!noprint) at strlcat's below */
640		*msg = 0;
641
642	if (errs & E_SUM_PKTERRS) {
643		ipath_stats.sps_pkterrs++;
644		chkerrpkts = 1;
645	}
646	if (errs & E_SUM_ERRS)
647		ipath_stats.sps_errs++;
648
649	if (errs & (INFINIPATH_E_RICRC | INFINIPATH_E_RVCRC)) {
650		ipath_stats.sps_crcerrs++;
651		chkerrpkts = 1;
652	}
653	iserr = errs & ~(E_SUM_PKTERRS | INFINIPATH_E_PKTERRS);
654
655
656	/*
657	 * We don't want to print these two as they happen, or we can make
658	 * the situation even worse, because it takes so long to print
659	 * messages to serial consoles.  Kernel ports get printed from
660	 * fast_stats, no more than every 5 seconds, user ports get printed
661	 * on close
662	 */
663	if (errs & INFINIPATH_E_RRCVHDRFULL) {
664		u32 hd, tl;
665		ipath_stats.sps_hdrqfull++;
666		for (i = 0; i < dd->ipath_cfgports; i++) {
667			struct ipath_portdata *pd = dd->ipath_pd[i];
668			if (i == 0) {
669				hd = dd->ipath_port0head;
670				tl = (u32) le64_to_cpu(
671					*dd->ipath_hdrqtailptr);
672			} else if (pd && pd->port_cnt &&
673				   pd->port_rcvhdrtail_kvaddr) {
674				/*
675				 * don't report same point multiple times,
676				 * except kernel
677				 */
678				tl = *(u64 *) pd->port_rcvhdrtail_kvaddr;
679				if (tl == dd->ipath_lastrcvhdrqtails[i])
680					continue;
681				hd = ipath_read_ureg32(dd, ur_rcvhdrhead,
682						       i);
683			} else
684				continue;
685			if (hd == (tl + 1) ||
686			    (!hd && tl == dd->ipath_hdrqlast)) {
687				if (i == 0)
688					chkerrpkts = 1;
689				dd->ipath_lastrcvhdrqtails[i] = tl;
690				pd->port_hdrqfull++;
691				if (test_bit(IPATH_PORT_WAITING_OVERFLOW,
692					     &pd->port_flag)) {
693					clear_bit(
694					  IPATH_PORT_WAITING_OVERFLOW,
695					  &pd->port_flag);
696					set_bit(
697					  IPATH_PORT_WAITING_OVERFLOW,
698					  &pd->int_flag);
699					wake_up_interruptible(
700					  &pd->port_wait);
701				}
702			}
703		}
704	}
705	if (errs & INFINIPATH_E_RRCVEGRFULL) {
706		/*
707		 * since this is of less importance and not likely to
708		 * happen without also getting hdrfull, only count
709		 * occurrences; don't check each port (or even the kernel
710		 * vs user)
711		 */
712		ipath_stats.sps_etidfull++;
713		if (dd->ipath_port0head !=
714		    (u32) le64_to_cpu(*dd->ipath_hdrqtailptr))
715			chkerrpkts = 1;
716	}
717
718	/*
719	 * do this before IBSTATUSCHANGED, in case both bits set in a single
720	 * interrupt; we want the STATUSCHANGE to "win", so we do our
721	 * internal copy of state machine correctly
722	 */
723	if (errs & INFINIPATH_E_RIBLOSTLINK) {
724		/*
725		 * force through block below
726		 */
727		errs |= INFINIPATH_E_IBSTATUSCHANGED;
728		ipath_stats.sps_iblink++;
729		dd->ipath_flags |= IPATH_LINKDOWN;
730		dd->ipath_flags &= ~(IPATH_LINKUNK | IPATH_LINKINIT
731				     | IPATH_LINKARMED | IPATH_LINKACTIVE);
732		*dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
733		if (!noprint) {
734			u64 st = ipath_read_kreg64(
735				dd, dd->ipath_kregs->kr_ibcstatus);
736
737			ipath_dbg("Lost link, link now down (%s)\n",
738				  ipath_ibcstatus_str[st & 0xf]);
739		}
740	}
741	if (errs & INFINIPATH_E_IBSTATUSCHANGED)
742		handle_e_ibstatuschanged(dd, errs, noprint);
743
744	if (errs & INFINIPATH_E_RESET) {
745		if (!noprint)
746			ipath_dev_err(dd, "Got reset, requires re-init "
747				      "(unload and reload driver)\n");
748		dd->ipath_flags &= ~IPATH_INITTED;	/* needs re-init */
749		/* mark as having had error */
750		*dd->ipath_statusp |= IPATH_STATUS_HWERROR;
751		*dd->ipath_statusp &= ~IPATH_STATUS_IB_CONF;
752	}
753
754	if (!noprint && *msg) {
755		if (iserr)
756			ipath_dev_err(dd, "%s error\n", msg);
757		else
758			dev_info(&dd->pcidev->dev, "%s packet problems\n",
759				msg);
760	}
761	if (dd->ipath_state_wanted & dd->ipath_flags) {
762		ipath_cdbg(VERBOSE, "driver wanted state %x, iflags now %x, "
763			   "waking\n", dd->ipath_state_wanted,
764			   dd->ipath_flags);
765		wake_up_interruptible(&ipath_state_wait);
766	}
767
768	return chkerrpkts;
769}
770
771
772/*
773 * try to cleanup as much as possible for anything that might have gone
774 * wrong while in freeze mode, such as pio buffers being written by user
775 * processes (causing armlaunch), send errors due to going into freeze mode,
776 * etc., and try to avoid causing extra interrupts while doing so.
777 * Forcibly update the in-memory pioavail register copies after cleanup
778 * because the chip won't do it for anything changing while in freeze mode
779 * (we don't want to wait for the next pio buffer state change).
780 * Make sure that we don't lose any important interrupts by using the chip
781 * feature that says that writing 0 to a bit in *clear that is set in
782 * *status will cause an interrupt to be generated again (if allowed by
783 * the *mask value).
784 */
785void ipath_clear_freeze(struct ipath_devdata *dd)
786{
787	int i, im;
788	__le64 val;
789
790	/* disable error interrupts, to avoid confusion */
791	ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask, 0ULL);
792
793	/* also disable interrupts; errormask is sometimes overwriten */
794	ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
795
796	/*
797	 * clear all sends, because they have may been
798	 * completed by usercode while in freeze mode, and
799	 * therefore would not be sent, and eventually
800	 * might cause the process to run out of bufs
801	 */
802	ipath_cancel_sends(dd, 0);
803	ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
804			 dd->ipath_control);
805
806	/* ensure pio avail updates continue */
807	ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
808		 dd->ipath_sendctrl & ~INFINIPATH_S_PIOBUFAVAILUPD);
809	ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
810	ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
811		 dd->ipath_sendctrl);
812
813	/*
814	 * We just enabled pioavailupdate, so dma copy is almost certainly
815	 * not yet right, so read the registers directly.  Similar to init
816	 */
817	for (i = 0; i < dd->ipath_pioavregs; i++) {
818		/* deal with 6110 chip bug */
819		im = i > 3 ? ((i&1) ? i-1 : i+1) : i;
820		val = ipath_read_kreg64(dd, (0x1000/sizeof(u64))+im);
821		dd->ipath_pioavailregs_dma[i] = dd->ipath_pioavailshadow[i]
822			= le64_to_cpu(val);
823	}
824
825	/*
826	 * force new interrupt if any hwerr, error or interrupt bits are
827	 * still set, and clear "safe" send packet errors related to freeze
828	 * and cancelling sends.  Re-enable error interrupts before possible
829	 * force of re-interrupt on pending interrupts.
830	 */
831	ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear, 0ULL);
832	ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear,
833		E_SPKT_ERRS_IGNORE);
834	ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
835		dd->ipath_errormask);
836	ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, -1LL);
837	ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
838}
839
840
841/* this is separate to allow for better optimization of ipath_intr() */
842
843static void ipath_bad_intr(struct ipath_devdata *dd, u32 * unexpectp)
844{
845	/*
846	 * sometimes happen during driver init and unload, don't want
847	 * to process any interrupts at that point
848	 */
849
850	/* this is just a bandaid, not a fix, if something goes badly
851	 * wrong */
852	if (++*unexpectp > 100) {
853		if (++*unexpectp > 105) {
854			/*
855			 * ok, we must be taking somebody else's interrupts,
856			 * due to a messed up mptable and/or PIRQ table, so
857			 * unregister the interrupt.  We've seen this during
858			 * linuxbios development work, and it may happen in
859			 * the future again.
860			 */
861			if (dd->pcidev && dd->ipath_irq) {
862				ipath_dev_err(dd, "Now %u unexpected "
863					      "interrupts, unregistering "
864					      "interrupt handler\n",
865					      *unexpectp);
866				ipath_dbg("free_irq of irq %d\n",
867					  dd->ipath_irq);
868				dd->ipath_f_free_irq(dd);
869			}
870		}
871		if (ipath_read_kreg32(dd, dd->ipath_kregs->kr_intmask)) {
872			ipath_dev_err(dd, "%u unexpected interrupts, "
873				      "disabling interrupts completely\n",
874				      *unexpectp);
875			/*
876			 * disable all interrupts, something is very wrong
877			 */
878			ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
879					 0ULL);
880		}
881	} else if (*unexpectp > 1)
882		ipath_dbg("Interrupt when not ready, should not happen, "
883			  "ignoring\n");
884}
885
886static void ipath_bad_regread(struct ipath_devdata *dd)
887{
888	static int allbits;
889
890	/* separate routine, for better optimization of ipath_intr() */
891
892	/*
893	 * We print the message and disable interrupts, in hope of
894	 * having a better chance of debugging the problem.
895	 */
896	ipath_dev_err(dd,
897		      "Read of interrupt status failed (all bits set)\n");
898	if (allbits++) {
899		/* disable all interrupts, something is very wrong */
900		ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
901		if (allbits == 2) {
902			ipath_dev_err(dd, "Still bad interrupt status, "
903				      "unregistering interrupt\n");
904			dd->ipath_f_free_irq(dd);
905		} else if (allbits > 2) {
906			if ((allbits % 10000) == 0)
907				printk(".");
908		} else
909			ipath_dev_err(dd, "Disabling interrupts, "
910				      "multiple errors\n");
911	}
912}
913
914static void handle_port_pioavail(struct ipath_devdata *dd)
915{
916	u32 i;
917	/*
918	 * start from port 1, since for now port 0  is never using
919	 * wait_event for PIO
920	 */
921	for (i = 1; dd->ipath_portpiowait && i < dd->ipath_cfgports; i++) {
922		struct ipath_portdata *pd = dd->ipath_pd[i];
923
924		if (pd && pd->port_cnt &&
925		    dd->ipath_portpiowait & (1U << i)) {
926			clear_bit(i, &dd->ipath_portpiowait);
927			if (test_bit(IPATH_PORT_WAITING_PIO,
928				     &pd->port_flag)) {
929				clear_bit(IPATH_PORT_WAITING_PIO,
930					  &pd->port_flag);
931				wake_up_interruptible(&pd->port_wait);
932			}
933		}
934	}
935}
936
937static void handle_layer_pioavail(struct ipath_devdata *dd)
938{
939	int ret;
940
941	ret = ipath_ib_piobufavail(dd->verbs_dev);
942	if (ret > 0)
943		goto set;
944
945	return;
946set:
947	set_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl);
948	ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
949			 dd->ipath_sendctrl);
950}
951
952/*
953 * Handle receive interrupts for user ports; this means a user
954 * process was waiting for a packet to arrive, and didn't want
955 * to poll
956 */
957static void handle_urcv(struct ipath_devdata *dd, u32 istat)
958{
959	u64 portr;
960	int i;
961	int rcvdint = 0;
962
963	portr = ((istat >> INFINIPATH_I_RCVAVAIL_SHIFT) &
964		 dd->ipath_i_rcvavail_mask)
965		| ((istat >> INFINIPATH_I_RCVURG_SHIFT) &
966		   dd->ipath_i_rcvurg_mask);
967	for (i = 1; i < dd->ipath_cfgports; i++) {
968		struct ipath_portdata *pd = dd->ipath_pd[i];
969		if (portr & (1 << i) && pd && pd->port_cnt) {
970			if (test_bit(IPATH_PORT_WAITING_RCV,
971				     &pd->port_flag)) {
972				clear_bit(IPATH_PORT_WAITING_RCV,
973					  &pd->port_flag);
974				set_bit(IPATH_PORT_WAITING_RCV,
975					&pd->int_flag);
976				clear_bit(i + INFINIPATH_R_INTRAVAIL_SHIFT,
977					  &dd->ipath_rcvctrl);
978				wake_up_interruptible(&pd->port_wait);
979				rcvdint = 1;
980			} else if (test_bit(IPATH_PORT_WAITING_URG,
981					    &pd->port_flag)) {
982				clear_bit(IPATH_PORT_WAITING_URG,
983					  &pd->port_flag);
984				set_bit(IPATH_PORT_WAITING_URG,
985					&pd->int_flag);
986				wake_up_interruptible(&pd->port_wait);
987			}
988		}
989	}
990	if (rcvdint) {
991		/* only want to take one interrupt, so turn off the rcv
992		 * interrupt for all the ports that we did the wakeup on
993		 * (but never for kernel port)
994		 */
995		ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
996				 dd->ipath_rcvctrl);
997	}
998}
999
1000irqreturn_t ipath_intr(int irq, void *data)
1001{
1002	struct ipath_devdata *dd = data;
1003	u32 istat, chk0rcv = 0;
1004	ipath_err_t estat = 0;
1005	irqreturn_t ret;
1006	static unsigned unexpected = 0;
1007	static const u32 port0rbits = (1U<<INFINIPATH_I_RCVAVAIL_SHIFT) |
1008		 (1U<<INFINIPATH_I_RCVURG_SHIFT);
1009
1010	ipath_stats.sps_ints++;
1011
1012	if (dd->ipath_int_counter != (u32) -1)
1013		dd->ipath_int_counter++;
1014
1015	if (!(dd->ipath_flags & IPATH_PRESENT)) {
1016		/*
1017		 * This return value is not great, but we do not want the
1018		 * interrupt core code to remove our interrupt handler
1019		 * because we don't appear to be handling an interrupt
1020		 * during a chip reset.
1021		 */
1022		return IRQ_HANDLED;
1023	}
1024
1025	/*
1026	 * this needs to be flags&initted, not statusp, so we keep
1027	 * taking interrupts even after link goes down, etc.
1028	 * Also, we *must* clear the interrupt at some point, or we won't
1029	 * take it again, which can be real bad for errors, etc...
1030	 */
1031
1032	if (!(dd->ipath_flags & IPATH_INITTED)) {
1033		ipath_bad_intr(dd, &unexpected);
1034		ret = IRQ_NONE;
1035		goto bail;
1036	}
1037
1038	istat = ipath_read_kreg32(dd, dd->ipath_kregs->kr_intstatus);
1039
1040	if (unlikely(!istat)) {
1041		ipath_stats.sps_nullintr++;
1042		ret = IRQ_NONE; /* not our interrupt, or already handled */
1043		goto bail;
1044	}
1045	if (unlikely(istat == -1)) {
1046		ipath_bad_regread(dd);
1047		/* don't know if it was our interrupt or not */
1048		ret = IRQ_NONE;
1049		goto bail;
1050	}
1051
1052	if (unexpected)
1053		unexpected = 0;
1054
1055	if (unlikely(istat & ~dd->ipath_i_bitsextant))
1056		ipath_dev_err(dd,
1057			      "interrupt with unknown interrupts %x set\n",
1058			      istat & (u32) ~ dd->ipath_i_bitsextant);
1059	else
1060		ipath_cdbg(VERBOSE, "intr stat=0x%x\n", istat);
1061
1062	if (unlikely(istat & INFINIPATH_I_ERROR)) {
1063		ipath_stats.sps_errints++;
1064		estat = ipath_read_kreg64(dd,
1065					  dd->ipath_kregs->kr_errorstatus);
1066		if (!estat)
1067			dev_info(&dd->pcidev->dev, "error interrupt (%x), "
1068				 "but no error bits set!\n", istat);
1069		else if (estat == -1LL)
1070			/*
1071			 * should we try clearing all, or hope next read
1072			 * works?
1073			 */
1074			ipath_dev_err(dd, "Read of error status failed "
1075				      "(all bits set); ignoring\n");
1076		else
1077			if (handle_errors(dd, estat))
1078				/* force calling ipath_kreceive() */
1079				chk0rcv = 1;
1080	}
1081
1082	if (istat & INFINIPATH_I_GPIO) {
1083		/*
1084		 * GPIO interrupts fall in two broad classes:
1085		 * GPIO_2 indicates (on some HT4xx boards) that a packet
1086		 *        has arrived for Port 0. Checking for this
1087		 *        is controlled by flag IPATH_GPIO_INTR.
1088		 * GPIO_3..5 on IBA6120 Rev2 chips indicate errors
1089		 *        that we need to count. Checking for this
1090		 *        is controlled by flag IPATH_GPIO_ERRINTRS.
1091		 */
1092		u32 gpiostatus;
1093		u32 to_clear = 0;
1094
1095		gpiostatus = ipath_read_kreg32(
1096			dd, dd->ipath_kregs->kr_gpio_status);
1097		/* First the error-counter case.
1098		 */
1099		if ((gpiostatus & IPATH_GPIO_ERRINTR_MASK) &&
1100		    (dd->ipath_flags & IPATH_GPIO_ERRINTRS)) {
1101			/* want to clear the bits we see asserted. */
1102			to_clear |= (gpiostatus & IPATH_GPIO_ERRINTR_MASK);
1103
1104			/*
1105			 * Count appropriately, clear bits out of our copy,
1106			 * as they have been "handled".
1107			 */
1108			if (gpiostatus & (1 << IPATH_GPIO_RXUVL_BIT)) {
1109				ipath_dbg("FlowCtl on UnsupVL\n");
1110				dd->ipath_rxfc_unsupvl_errs++;
1111			}
1112			if (gpiostatus & (1 << IPATH_GPIO_OVRUN_BIT)) {
1113				ipath_dbg("Overrun Threshold exceeded\n");
1114				dd->ipath_overrun_thresh_errs++;
1115			}
1116			if (gpiostatus & (1 << IPATH_GPIO_LLI_BIT)) {
1117				ipath_dbg("Local Link Integrity error\n");
1118				dd->ipath_lli_errs++;
1119			}
1120			gpiostatus &= ~IPATH_GPIO_ERRINTR_MASK;
1121		}
1122		/* Now the Port0 Receive case */
1123		if ((gpiostatus & (1 << IPATH_GPIO_PORT0_BIT)) &&
1124		    (dd->ipath_flags & IPATH_GPIO_INTR)) {
1125			/*
1126			 * GPIO status bit 2 is set, and we expected it.
1127			 * clear it and indicate in p0bits.
1128			 * This probably only happens if a Port0 pkt
1129			 * arrives at _just_ the wrong time, and we
1130			 * handle that by seting chk0rcv;
1131			 */
1132			to_clear |= (1 << IPATH_GPIO_PORT0_BIT);
1133			gpiostatus &= ~(1 << IPATH_GPIO_PORT0_BIT);
1134			chk0rcv = 1;
1135		}
1136		if (gpiostatus) {
1137			/*
1138			 * Some unexpected bits remain. If they could have
1139			 * caused the interrupt, complain and clear.
1140			 * MEA: this is almost certainly non-ideal.
1141			 * we should look into auto-disable of unexpected
1142			 * GPIO interrupts, possibly on a "three strikes"
1143			 * basis.
1144			 */
1145			const u32 mask = (u32) dd->ipath_gpio_mask;
1146
1147			if (mask & gpiostatus) {
1148				ipath_dbg("Unexpected GPIO IRQ bits %x\n",
1149				  gpiostatus & mask);
1150				to_clear |= (gpiostatus & mask);
1151			}
1152		}
1153		if (to_clear) {
1154			ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
1155					(u64) to_clear);
1156		}
1157	}
1158	chk0rcv |= istat & port0rbits;
1159
1160	/*
1161	 * Clear the interrupt bits we found set, unless they are receive
1162	 * related, in which case we already cleared them above, and don't
1163	 * want to clear them again, because we might lose an interrupt.
1164	 * Clear it early, so we "know" know the chip will have seen this by
1165	 * the time we process the queue, and will re-interrupt if necessary.
1166	 * The processor itself won't take the interrupt again until we return.
1167	 */
1168	ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
1169
1170	/*
1171	 * handle port0 receive  before checking for pio buffers available,
1172	 * since receives can overflow; piobuf waiters can afford a few
1173	 * extra cycles, since they were waiting anyway, and user's waiting
1174	 * for receive are at the bottom.
1175	 */
1176	if (chk0rcv) {
1177		ipath_kreceive(dd);
1178		istat &= ~port0rbits;
1179	}
1180
1181	if (istat & ((dd->ipath_i_rcvavail_mask <<
1182		      INFINIPATH_I_RCVAVAIL_SHIFT)
1183		     | (dd->ipath_i_rcvurg_mask <<
1184			INFINIPATH_I_RCVURG_SHIFT)))
1185		handle_urcv(dd, istat);
1186
1187	if (istat & INFINIPATH_I_SPIOBUFAVAIL) {
1188		clear_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl);
1189		ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1190				 dd->ipath_sendctrl);
1191
1192		if (dd->ipath_portpiowait)
1193			handle_port_pioavail(dd);
1194
1195		handle_layer_pioavail(dd);
1196	}
1197
1198	ret = IRQ_HANDLED;
1199
1200bail:
1201	return ret;
1202}
1203