FreeBSD/Linux Kernel Cross Reference
sys/x86/x86/mca.c
1 /*-
2 * Copyright (c) 2009 Hudson River Trading LLC
3 * Written by: John H. Baldwin <jhb@FreeBSD.org>
4 * All rights reserved.
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
8 * are met:
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 *
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25 * SUCH DAMAGE.
26 */
27
28 /*
29 * Support for x86 machine check architecture.
30 */
31
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34
35 #ifdef __amd64__
36 #define DEV_APIC
37 #else
38 #include "opt_apic.h"
39 #endif
40
41 #include <sys/param.h>
42 #include <sys/bus.h>
43 #include <sys/interrupt.h>
44 #include <sys/kernel.h>
45 #include <sys/lock.h>
46 #include <sys/malloc.h>
47 #include <sys/mutex.h>
48 #include <sys/proc.h>
49 #include <sys/sched.h>
50 #include <sys/smp.h>
51 #include <sys/sysctl.h>
52 #include <sys/systm.h>
53 #include <sys/taskqueue.h>
54 #include <machine/intr_machdep.h>
55 #include <machine/apicvar.h>
56 #include <machine/cpu.h>
57 #include <machine/cputypes.h>
58 #include <machine/mca.h>
59 #include <machine/md_var.h>
60 #include <machine/specialreg.h>
61
62 /* Modes for mca_scan() */
63 enum scan_mode {
64 POLLED,
65 MCE,
66 CMCI,
67 };
68
69 #ifdef DEV_APIC
70 /*
71 * State maintained for each monitored MCx bank to control the
72 * corrected machine check interrupt threshold.
73 */
74 struct cmc_state {
75 int max_threshold;
76 int last_intr;
77 };
78 #endif
79
80 struct mca_internal {
81 struct mca_record rec;
82 int logged;
83 STAILQ_ENTRY(mca_internal) link;
84 };
85
86 static MALLOC_DEFINE(M_MCA, "MCA", "Machine Check Architecture");
87
88 static volatile int mca_count; /* Number of records stored. */
89 static int mca_banks; /* Number of per-CPU register banks. */
90
91 SYSCTL_NODE(_hw, OID_AUTO, mca, CTLFLAG_RD, NULL, "Machine Check Architecture");
92
93 static int mca_enabled = 1;
94 TUNABLE_INT("hw.mca.enabled", &mca_enabled);
95 SYSCTL_INT(_hw_mca, OID_AUTO, enabled, CTLFLAG_RDTUN, &mca_enabled, 0,
96 "Administrative toggle for machine check support");
97
98 static int amd10h_L1TP = 1;
99 TUNABLE_INT("hw.mca.amd10h_L1TP", &amd10h_L1TP);
100 SYSCTL_INT(_hw_mca, OID_AUTO, amd10h_L1TP, CTLFLAG_RDTUN, &amd10h_L1TP, 0,
101 "Administrative toggle for logging of level one TLB parity (L1TP) errors");
102
103 static int intel6h_HSD131;
104 TUNABLE_INT("hw.mca.intel6h_hsd131", &intel6h_HSD131);
105 SYSCTL_INT(_hw_mca, OID_AUTO, intel6h_HSD131, CTLFLAG_RDTUN, &intel6h_HSD131, 0,
106 "Administrative toggle for logging of spurious corrected errors");
107
108 int workaround_erratum383;
109 SYSCTL_INT(_hw_mca, OID_AUTO, erratum383, CTLFLAG_RD, &workaround_erratum383, 0,
110 "Is the workaround for Erratum 383 on AMD Family 10h processors enabled?");
111
112 static STAILQ_HEAD(, mca_internal) mca_freelist;
113 static int mca_freecount;
114 static STAILQ_HEAD(, mca_internal) mca_records;
115 static struct callout mca_timer;
116 static int mca_ticks = 3600; /* Check hourly by default. */
117 static struct taskqueue *mca_tq;
118 static struct task mca_refill_task, mca_scan_task;
119 static struct mtx mca_lock;
120
121 #ifdef DEV_APIC
122 static struct cmc_state **cmc_state; /* Indexed by cpuid, bank */
123 static int cmc_throttle = 60; /* Time in seconds to throttle CMCI. */
124 #endif
125
126 static int
127 sysctl_positive_int(SYSCTL_HANDLER_ARGS)
128 {
129 int error, value;
130
131 value = *(int *)arg1;
132 error = sysctl_handle_int(oidp, &value, 0, req);
133 if (error || req->newptr == NULL)
134 return (error);
135 if (value <= 0)
136 return (EINVAL);
137 *(int *)arg1 = value;
138 return (0);
139 }
140
141 static int
142 sysctl_mca_records(SYSCTL_HANDLER_ARGS)
143 {
144 int *name = (int *)arg1;
145 u_int namelen = arg2;
146 struct mca_record record;
147 struct mca_internal *rec;
148 int i;
149
150 if (namelen != 1)
151 return (EINVAL);
152
153 if (name[0] < 0 || name[0] >= mca_count)
154 return (EINVAL);
155
156 mtx_lock_spin(&mca_lock);
157 if (name[0] >= mca_count) {
158 mtx_unlock_spin(&mca_lock);
159 return (EINVAL);
160 }
161 i = 0;
162 STAILQ_FOREACH(rec, &mca_records, link) {
163 if (i == name[0]) {
164 record = rec->rec;
165 break;
166 }
167 i++;
168 }
169 mtx_unlock_spin(&mca_lock);
170 return (SYSCTL_OUT(req, &record, sizeof(record)));
171 }
172
173 static const char *
174 mca_error_ttype(uint16_t mca_error)
175 {
176
177 switch ((mca_error & 0x000c) >> 2) {
178 case 0:
179 return ("I");
180 case 1:
181 return ("D");
182 case 2:
183 return ("G");
184 }
185 return ("?");
186 }
187
188 static const char *
189 mca_error_level(uint16_t mca_error)
190 {
191
192 switch (mca_error & 0x0003) {
193 case 0:
194 return ("L0");
195 case 1:
196 return ("L1");
197 case 2:
198 return ("L2");
199 case 3:
200 return ("LG");
201 }
202 return ("L?");
203 }
204
205 static const char *
206 mca_error_request(uint16_t mca_error)
207 {
208
209 switch ((mca_error & 0x00f0) >> 4) {
210 case 0x0:
211 return ("ERR");
212 case 0x1:
213 return ("RD");
214 case 0x2:
215 return ("WR");
216 case 0x3:
217 return ("DRD");
218 case 0x4:
219 return ("DWR");
220 case 0x5:
221 return ("IRD");
222 case 0x6:
223 return ("PREFETCH");
224 case 0x7:
225 return ("EVICT");
226 case 0x8:
227 return ("SNOOP");
228 }
229 return ("???");
230 }
231
232 static const char *
233 mca_error_mmtype(uint16_t mca_error)
234 {
235
236 switch ((mca_error & 0x70) >> 4) {
237 case 0x0:
238 return ("GEN");
239 case 0x1:
240 return ("RD");
241 case 0x2:
242 return ("WR");
243 case 0x3:
244 return ("AC");
245 case 0x4:
246 return ("MS");
247 }
248 return ("???");
249 }
250
251 static int __nonnull(1)
252 mca_mute(const struct mca_record *rec)
253 {
254
255 /*
256 * Skip spurious corrected parity errors generated by Intel Haswell-
257 * and Broadwell-based CPUs (see HSD131, HSM142, HSW131 and BDM48
258 * erratum respectively), unless reporting is enabled.
259 * Note that these errors also have been observed with the D0-stepping
260 * of Haswell, while at least initially the CPU specification updates
261 * suggested only the C0-stepping to be affected. Similarly, Celeron
262 * 2955U with a CPU ID of 0x45 apparently are also concerned with the
263 * same problem, with HSM142 only referring to 0x3c and 0x46.
264 */
265 if (cpu_vendor_id == CPU_VENDOR_INTEL &&
266 CPUID_TO_FAMILY(cpu_id) == 0x6 &&
267 (CPUID_TO_MODEL(cpu_id) == 0x3c || /* HSD131, HSM142, HSW131 */
268 CPUID_TO_MODEL(cpu_id) == 0x3d || /* BDM48 */
269 CPUID_TO_MODEL(cpu_id) == 0x45 ||
270 CPUID_TO_MODEL(cpu_id) == 0x46) && /* HSM142 */
271 rec->mr_bank == 0 &&
272 (rec->mr_status & 0xa0000000ffffffff) == 0x80000000000f0005 &&
273 !intel6h_HSD131)
274 return (1);
275
276 return (0);
277 }
278
279 /* Dump details about a single machine check. */
280 static void __nonnull(1)
281 mca_log(const struct mca_record *rec)
282 {
283 uint16_t mca_error;
284
285 if (mca_mute(rec))
286 return;
287
288 printf("MCA: Bank %d, Status 0x%016llx\n", rec->mr_bank,
289 (long long)rec->mr_status);
290 printf("MCA: Global Cap 0x%016llx, Status 0x%016llx\n",
291 (long long)rec->mr_mcg_cap, (long long)rec->mr_mcg_status);
292 printf("MCA: Vendor \"%s\", ID 0x%x, APIC ID %d\n", cpu_vendor,
293 rec->mr_cpu_id, rec->mr_apic_id);
294 printf("MCA: CPU %d ", rec->mr_cpu);
295 if (rec->mr_status & MC_STATUS_UC)
296 printf("UNCOR ");
297 else {
298 printf("COR ");
299 if (rec->mr_mcg_cap & MCG_CAP_CMCI_P)
300 printf("(%lld) ", ((long long)rec->mr_status &
301 MC_STATUS_COR_COUNT) >> 38);
302 }
303 if (rec->mr_status & MC_STATUS_PCC)
304 printf("PCC ");
305 if (rec->mr_status & MC_STATUS_OVER)
306 printf("OVER ");
307 mca_error = rec->mr_status & MC_STATUS_MCA_ERROR;
308 switch (mca_error) {
309 /* Simple error codes. */
310 case 0x0000:
311 printf("no error");
312 break;
313 case 0x0001:
314 printf("unclassified error");
315 break;
316 case 0x0002:
317 printf("ucode ROM parity error");
318 break;
319 case 0x0003:
320 printf("external error");
321 break;
322 case 0x0004:
323 printf("FRC error");
324 break;
325 case 0x0005:
326 printf("internal parity error");
327 break;
328 case 0x0400:
329 printf("internal timer error");
330 break;
331 default:
332 if ((mca_error & 0xfc00) == 0x0400) {
333 printf("internal error %x", mca_error & 0x03ff);
334 break;
335 }
336
337 /* Compound error codes. */
338
339 /* Memory hierarchy error. */
340 if ((mca_error & 0xeffc) == 0x000c) {
341 printf("%s memory error", mca_error_level(mca_error));
342 break;
343 }
344
345 /* TLB error. */
346 if ((mca_error & 0xeff0) == 0x0010) {
347 printf("%sTLB %s error", mca_error_ttype(mca_error),
348 mca_error_level(mca_error));
349 break;
350 }
351
352 /* Memory controller error. */
353 if ((mca_error & 0xef80) == 0x0080) {
354 printf("%s channel ", mca_error_mmtype(mca_error));
355 if ((mca_error & 0x000f) != 0x000f)
356 printf("%d", mca_error & 0x000f);
357 else
358 printf("??");
359 printf(" memory error");
360 break;
361 }
362
363 /* Cache error. */
364 if ((mca_error & 0xef00) == 0x0100) {
365 printf("%sCACHE %s %s error",
366 mca_error_ttype(mca_error),
367 mca_error_level(mca_error),
368 mca_error_request(mca_error));
369 break;
370 }
371
372 /* Bus and/or Interconnect error. */
373 if ((mca_error & 0xe800) == 0x0800) {
374 printf("BUS%s ", mca_error_level(mca_error));
375 switch ((mca_error & 0x0600) >> 9) {
376 case 0:
377 printf("Source");
378 break;
379 case 1:
380 printf("Responder");
381 break;
382 case 2:
383 printf("Observer");
384 break;
385 default:
386 printf("???");
387 break;
388 }
389 printf(" %s ", mca_error_request(mca_error));
390 switch ((mca_error & 0x000c) >> 2) {
391 case 0:
392 printf("Memory");
393 break;
394 case 2:
395 printf("I/O");
396 break;
397 case 3:
398 printf("Other");
399 break;
400 default:
401 printf("???");
402 break;
403 }
404 if (mca_error & 0x0100)
405 printf(" timed out");
406 break;
407 }
408
409 printf("unknown error %x", mca_error);
410 break;
411 }
412 printf("\n");
413 if (rec->mr_status & MC_STATUS_ADDRV)
414 printf("MCA: Address 0x%llx\n", (long long)rec->mr_addr);
415 if (rec->mr_status & MC_STATUS_MISCV)
416 printf("MCA: Misc 0x%llx\n", (long long)rec->mr_misc);
417 }
418
419 static int __nonnull(2)
420 mca_check_status(int bank, struct mca_record *rec)
421 {
422 uint64_t status;
423 u_int p[4];
424
425 status = rdmsr(MSR_MC_STATUS(bank));
426 if (!(status & MC_STATUS_VAL))
427 return (0);
428
429 /* Save exception information. */
430 rec->mr_status = status;
431 rec->mr_bank = bank;
432 rec->mr_addr = 0;
433 if (status & MC_STATUS_ADDRV)
434 rec->mr_addr = rdmsr(MSR_MC_ADDR(bank));
435 rec->mr_misc = 0;
436 if (status & MC_STATUS_MISCV)
437 rec->mr_misc = rdmsr(MSR_MC_MISC(bank));
438 rec->mr_tsc = rdtsc();
439 rec->mr_apic_id = PCPU_GET(apic_id);
440 rec->mr_mcg_cap = rdmsr(MSR_MCG_CAP);
441 rec->mr_mcg_status = rdmsr(MSR_MCG_STATUS);
442 rec->mr_cpu_id = cpu_id;
443 rec->mr_cpu_vendor_id = cpu_vendor_id;
444 rec->mr_cpu = PCPU_GET(cpuid);
445
446 /*
447 * Clear machine check. Don't do this for uncorrectable
448 * errors so that the BIOS can see them.
449 */
450 if (!(rec->mr_status & (MC_STATUS_PCC | MC_STATUS_UC))) {
451 wrmsr(MSR_MC_STATUS(bank), 0);
452 do_cpuid(0, p);
453 }
454 return (1);
455 }
456
457 static void
458 mca_fill_freelist(void)
459 {
460 struct mca_internal *rec;
461 int desired;
462
463 /*
464 * Ensure we have at least one record for each bank and one
465 * record per CPU.
466 */
467 desired = imax(mp_ncpus, mca_banks);
468 mtx_lock_spin(&mca_lock);
469 while (mca_freecount < desired) {
470 mtx_unlock_spin(&mca_lock);
471 rec = malloc(sizeof(*rec), M_MCA, M_WAITOK);
472 mtx_lock_spin(&mca_lock);
473 STAILQ_INSERT_TAIL(&mca_freelist, rec, link);
474 mca_freecount++;
475 }
476 mtx_unlock_spin(&mca_lock);
477 }
478
479 static void
480 mca_refill(void *context, int pending)
481 {
482
483 mca_fill_freelist();
484 }
485
486 static void __nonnull(2)
487 mca_record_entry(enum scan_mode mode, const struct mca_record *record)
488 {
489 struct mca_internal *rec;
490
491 if (mode == POLLED) {
492 rec = malloc(sizeof(*rec), M_MCA, M_WAITOK);
493 mtx_lock_spin(&mca_lock);
494 } else {
495 mtx_lock_spin(&mca_lock);
496 rec = STAILQ_FIRST(&mca_freelist);
497 if (rec == NULL) {
498 printf("MCA: Unable to allocate space for an event.\n");
499 mca_log(record);
500 mtx_unlock_spin(&mca_lock);
501 return;
502 }
503 STAILQ_REMOVE_HEAD(&mca_freelist, link);
504 mca_freecount--;
505 }
506
507 rec->rec = *record;
508 rec->logged = 0;
509 STAILQ_INSERT_TAIL(&mca_records, rec, link);
510 mca_count++;
511 mtx_unlock_spin(&mca_lock);
512 if (mode == CMCI)
513 taskqueue_enqueue_fast(mca_tq, &mca_refill_task);
514 }
515
516 #ifdef DEV_APIC
517 /*
518 * Update the interrupt threshold for a CMCI. The strategy is to use
519 * a low trigger that interrupts as soon as the first event occurs.
520 * However, if a steady stream of events arrive, the threshold is
521 * increased until the interrupts are throttled to once every
522 * cmc_throttle seconds or the periodic scan. If a periodic scan
523 * finds that the threshold is too high, it is lowered.
524 */
525 static void
526 cmci_update(enum scan_mode mode, int bank, int valid, struct mca_record *rec)
527 {
528 struct cmc_state *cc;
529 uint64_t ctl;
530 u_int delta;
531 int count, limit;
532
533 /* Fetch the current limit for this bank. */
534 cc = &cmc_state[PCPU_GET(cpuid)][bank];
535 ctl = rdmsr(MSR_MC_CTL2(bank));
536 count = (rec->mr_status & MC_STATUS_COR_COUNT) >> 38;
537 delta = (u_int)(ticks - cc->last_intr);
538
539 /*
540 * If an interrupt was received less than cmc_throttle seconds
541 * since the previous interrupt and the count from the current
542 * event is greater than or equal to the current threshold,
543 * double the threshold up to the max.
544 */
545 if (mode == CMCI && valid) {
546 limit = ctl & MC_CTL2_THRESHOLD;
547 if (delta < cmc_throttle && count >= limit &&
548 limit < cc->max_threshold) {
549 limit = min(limit << 1, cc->max_threshold);
550 ctl &= ~MC_CTL2_THRESHOLD;
551 ctl |= limit;
552 wrmsr(MSR_MC_CTL2(bank), limit);
553 }
554 cc->last_intr = ticks;
555 return;
556 }
557
558 /*
559 * When the banks are polled, check to see if the threshold
560 * should be lowered.
561 */
562 if (mode != POLLED)
563 return;
564
565 /* If a CMCI occured recently, do nothing for now. */
566 if (delta < cmc_throttle)
567 return;
568
569 /*
570 * Compute a new limit based on the average rate of events per
571 * cmc_throttle seconds since the last interrupt.
572 */
573 if (valid) {
574 count = (rec->mr_status & MC_STATUS_COR_COUNT) >> 38;
575 limit = count * cmc_throttle / delta;
576 if (limit <= 0)
577 limit = 1;
578 else if (limit > cc->max_threshold)
579 limit = cc->max_threshold;
580 } else
581 limit = 1;
582 if ((ctl & MC_CTL2_THRESHOLD) != limit) {
583 ctl &= ~MC_CTL2_THRESHOLD;
584 ctl |= limit;
585 wrmsr(MSR_MC_CTL2(bank), limit);
586 }
587 }
588 #endif
589
590 /*
591 * This scans all the machine check banks of the current CPU to see if
592 * there are any machine checks. Any non-recoverable errors are
593 * reported immediately via mca_log(). The current thread must be
594 * pinned when this is called. The 'mode' parameter indicates if we
595 * are being called from the MC exception handler, the CMCI handler,
596 * or the periodic poller. In the MC exception case this function
597 * returns true if the system is restartable. Otherwise, it returns a
598 * count of the number of valid MC records found.
599 */
600 static int
601 mca_scan(enum scan_mode mode)
602 {
603 struct mca_record rec;
604 uint64_t mcg_cap, ucmask;
605 int count, i, recoverable, valid;
606
607 count = 0;
608 recoverable = 1;
609 ucmask = MC_STATUS_UC | MC_STATUS_PCC;
610
611 /* When handling a MCE#, treat the OVER flag as non-restartable. */
612 if (mode == MCE)
613 ucmask |= MC_STATUS_OVER;
614 mcg_cap = rdmsr(MSR_MCG_CAP);
615 for (i = 0; i < (mcg_cap & MCG_CAP_COUNT); i++) {
616 #ifdef DEV_APIC
617 /*
618 * For a CMCI, only check banks this CPU is
619 * responsible for.
620 */
621 if (mode == CMCI && !(PCPU_GET(cmci_mask) & 1 << i))
622 continue;
623 #endif
624
625 valid = mca_check_status(i, &rec);
626 if (valid) {
627 count++;
628 if (rec.mr_status & ucmask) {
629 recoverable = 0;
630 mtx_lock_spin(&mca_lock);
631 mca_log(&rec);
632 mtx_unlock_spin(&mca_lock);
633 }
634 mca_record_entry(mode, &rec);
635 }
636
637 #ifdef DEV_APIC
638 /*
639 * If this is a bank this CPU monitors via CMCI,
640 * update the threshold.
641 */
642 if (PCPU_GET(cmci_mask) & 1 << i)
643 cmci_update(mode, i, valid, &rec);
644 #endif
645 }
646 if (mode == POLLED)
647 mca_fill_freelist();
648 return (mode == MCE ? recoverable : count);
649 }
650
651 /*
652 * Scan the machine check banks on all CPUs by binding to each CPU in
653 * turn. If any of the CPUs contained new machine check records, log
654 * them to the console.
655 */
656 static void
657 mca_scan_cpus(void *context, int pending)
658 {
659 struct mca_internal *mca;
660 struct thread *td;
661 int count, cpu;
662
663 mca_fill_freelist();
664 td = curthread;
665 count = 0;
666 thread_lock(td);
667 CPU_FOREACH(cpu) {
668 sched_bind(td, cpu);
669 thread_unlock(td);
670 count += mca_scan(POLLED);
671 thread_lock(td);
672 sched_unbind(td);
673 }
674 thread_unlock(td);
675 if (count != 0) {
676 mtx_lock_spin(&mca_lock);
677 STAILQ_FOREACH(mca, &mca_records, link) {
678 if (!mca->logged) {
679 mca->logged = 1;
680 mca_log(&mca->rec);
681 }
682 }
683 mtx_unlock_spin(&mca_lock);
684 }
685 }
686
687 static void
688 mca_periodic_scan(void *arg)
689 {
690
691 taskqueue_enqueue_fast(mca_tq, &mca_scan_task);
692 callout_reset(&mca_timer, mca_ticks * hz, mca_periodic_scan, NULL);
693 }
694
695 static int
696 sysctl_mca_scan(SYSCTL_HANDLER_ARGS)
697 {
698 int error, i;
699
700 i = 0;
701 error = sysctl_handle_int(oidp, &i, 0, req);
702 if (error)
703 return (error);
704 if (i)
705 taskqueue_enqueue_fast(mca_tq, &mca_scan_task);
706 return (0);
707 }
708
709 static void
710 mca_createtq(void *dummy)
711 {
712 if (mca_banks <= 0)
713 return;
714
715 mca_tq = taskqueue_create_fast("mca", M_WAITOK,
716 taskqueue_thread_enqueue, &mca_tq);
717 taskqueue_start_threads(&mca_tq, 1, PI_SWI(SWI_TQ), "mca taskq");
718 }
719 SYSINIT(mca_createtq, SI_SUB_CONFIGURE, SI_ORDER_ANY, mca_createtq, NULL);
720
721 static void
722 mca_startup(void *dummy)
723 {
724
725 if (mca_banks <= 0)
726 return;
727
728 callout_reset(&mca_timer, mca_ticks * hz, mca_periodic_scan, NULL);
729 }
730 SYSINIT(mca_startup, SI_SUB_SMP, SI_ORDER_ANY, mca_startup, NULL);
731
732 #ifdef DEV_APIC
733 static void
734 cmci_setup(void)
735 {
736 int i;
737
738 cmc_state = malloc((mp_maxid + 1) * sizeof(struct cmc_state *), M_MCA,
739 M_WAITOK);
740 for (i = 0; i <= mp_maxid; i++)
741 cmc_state[i] = malloc(sizeof(struct cmc_state) * mca_banks,
742 M_MCA, M_WAITOK | M_ZERO);
743 SYSCTL_ADD_PROC(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
744 "cmc_throttle", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
745 &cmc_throttle, 0, sysctl_positive_int, "I",
746 "Interval in seconds to throttle corrected MC interrupts");
747 }
748 #endif
749
750 static void
751 mca_setup(uint64_t mcg_cap)
752 {
753
754 /*
755 * On AMD Family 10h processors, unless logging of level one TLB
756 * parity (L1TP) errors is disabled, enable the recommended workaround
757 * for Erratum 383.
758 */
759 if (cpu_vendor_id == CPU_VENDOR_AMD &&
760 CPUID_TO_FAMILY(cpu_id) == 0x10 && amd10h_L1TP)
761 workaround_erratum383 = 1;
762
763 mca_banks = mcg_cap & MCG_CAP_COUNT;
764 mtx_init(&mca_lock, "mca", NULL, MTX_SPIN);
765 STAILQ_INIT(&mca_records);
766 TASK_INIT(&mca_scan_task, 0, mca_scan_cpus, NULL);
767 callout_init(&mca_timer, CALLOUT_MPSAFE);
768 STAILQ_INIT(&mca_freelist);
769 TASK_INIT(&mca_refill_task, 0, mca_refill, NULL);
770 mca_fill_freelist();
771 SYSCTL_ADD_INT(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
772 "count", CTLFLAG_RD, (int *)(uintptr_t)&mca_count, 0,
773 "Record count");
774 SYSCTL_ADD_PROC(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
775 "interval", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, &mca_ticks,
776 0, sysctl_positive_int, "I",
777 "Periodic interval in seconds to scan for machine checks");
778 SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
779 "records", CTLFLAG_RD, sysctl_mca_records, "Machine check records");
780 SYSCTL_ADD_PROC(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
781 "force_scan", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
782 sysctl_mca_scan, "I", "Force an immediate scan for machine checks");
783 #ifdef DEV_APIC
784 if (mcg_cap & MCG_CAP_CMCI_P)
785 cmci_setup();
786 #endif
787 }
788
789 #ifdef DEV_APIC
790 /*
791 * See if we should monitor CMCI for this bank. If CMCI_EN is already
792 * set in MC_CTL2, then another CPU is responsible for this bank, so
793 * ignore it. If CMCI_EN returns zero after being set, then this bank
794 * does not support CMCI_EN. If this CPU sets CMCI_EN, then it should
795 * now monitor this bank.
796 */
797 static void
798 cmci_monitor(int i)
799 {
800 struct cmc_state *cc;
801 uint64_t ctl;
802
803 KASSERT(i < mca_banks, ("CPU %d has more MC banks", PCPU_GET(cpuid)));
804
805 ctl = rdmsr(MSR_MC_CTL2(i));
806 if (ctl & MC_CTL2_CMCI_EN)
807 /* Already monitored by another CPU. */
808 return;
809
810 /* Set the threshold to one event for now. */
811 ctl &= ~MC_CTL2_THRESHOLD;
812 ctl |= MC_CTL2_CMCI_EN | 1;
813 wrmsr(MSR_MC_CTL2(i), ctl);
814 ctl = rdmsr(MSR_MC_CTL2(i));
815 if (!(ctl & MC_CTL2_CMCI_EN))
816 /* This bank does not support CMCI. */
817 return;
818
819 cc = &cmc_state[PCPU_GET(cpuid)][i];
820
821 /* Determine maximum threshold. */
822 ctl &= ~MC_CTL2_THRESHOLD;
823 ctl |= 0x7fff;
824 wrmsr(MSR_MC_CTL2(i), ctl);
825 ctl = rdmsr(MSR_MC_CTL2(i));
826 cc->max_threshold = ctl & MC_CTL2_THRESHOLD;
827
828 /* Start off with a threshold of 1. */
829 ctl &= ~MC_CTL2_THRESHOLD;
830 ctl |= 1;
831 wrmsr(MSR_MC_CTL2(i), ctl);
832
833 /* Mark this bank as monitored. */
834 PCPU_SET(cmci_mask, PCPU_GET(cmci_mask) | 1 << i);
835 }
836
837 /*
838 * For resume, reset the threshold for any banks we monitor back to
839 * one and throw away the timestamp of the last interrupt.
840 */
841 static void
842 cmci_resume(int i)
843 {
844 struct cmc_state *cc;
845 uint64_t ctl;
846
847 KASSERT(i < mca_banks, ("CPU %d has more MC banks", PCPU_GET(cpuid)));
848
849 /* Ignore banks not monitored by this CPU. */
850 if (!(PCPU_GET(cmci_mask) & 1 << i))
851 return;
852
853 cc = &cmc_state[PCPU_GET(cpuid)][i];
854 cc->last_intr = -ticks;
855 ctl = rdmsr(MSR_MC_CTL2(i));
856 ctl &= ~MC_CTL2_THRESHOLD;
857 ctl |= MC_CTL2_CMCI_EN | 1;
858 wrmsr(MSR_MC_CTL2(i), ctl);
859 }
860 #endif
861
862 /*
863 * Initializes per-CPU machine check registers and enables corrected
864 * machine check interrupts.
865 */
866 static void
867 _mca_init(int boot)
868 {
869 uint64_t mcg_cap;
870 uint64_t ctl, mask;
871 int i, skip;
872
873 /* MCE is required. */
874 if (!mca_enabled || !(cpu_feature & CPUID_MCE))
875 return;
876
877 if (cpu_feature & CPUID_MCA) {
878 if (boot)
879 PCPU_SET(cmci_mask, 0);
880
881 mcg_cap = rdmsr(MSR_MCG_CAP);
882 if (mcg_cap & MCG_CAP_CTL_P)
883 /* Enable MCA features. */
884 wrmsr(MSR_MCG_CTL, MCG_CTL_ENABLE);
885 if (PCPU_GET(cpuid) == 0 && boot)
886 mca_setup(mcg_cap);
887
888 /*
889 * Disable logging of level one TLB parity (L1TP) errors by
890 * the data cache as an alternative workaround for AMD Family
891 * 10h Erratum 383. Unlike the recommended workaround, there
892 * is no performance penalty to this workaround. However,
893 * L1TP errors will go unreported.
894 */
895 if (cpu_vendor_id == CPU_VENDOR_AMD &&
896 CPUID_TO_FAMILY(cpu_id) == 0x10 && !amd10h_L1TP) {
897 mask = rdmsr(MSR_MC0_CTL_MASK);
898 if ((mask & (1UL << 5)) == 0)
899 wrmsr(MSR_MC0_CTL_MASK, mask | (1UL << 5));
900 }
901 for (i = 0; i < (mcg_cap & MCG_CAP_COUNT); i++) {
902 /* By default enable logging of all errors. */
903 ctl = 0xffffffffffffffffUL;
904 skip = 0;
905
906 if (cpu_vendor_id == CPU_VENDOR_INTEL) {
907 /*
908 * For P6 models before Nehalem MC0_CTL is
909 * always enabled and reserved.
910 */
911 if (i == 0 && CPUID_TO_FAMILY(cpu_id) == 0x6
912 && CPUID_TO_MODEL(cpu_id) < 0x1a)
913 skip = 1;
914 } else if (cpu_vendor_id == CPU_VENDOR_AMD) {
915 /* BKDG for Family 10h: unset GartTblWkEn. */
916 if (i == 4 && CPUID_TO_FAMILY(cpu_id) >= 0xf)
917 ctl &= ~(1UL << 10);
918 }
919
920 if (!skip)
921 wrmsr(MSR_MC_CTL(i), ctl);
922
923 #ifdef DEV_APIC
924 if (mcg_cap & MCG_CAP_CMCI_P) {
925 if (boot)
926 cmci_monitor(i);
927 else
928 cmci_resume(i);
929 }
930 #endif
931
932 /* Clear all errors. */
933 wrmsr(MSR_MC_STATUS(i), 0);
934 }
935
936 #ifdef DEV_APIC
937 if (PCPU_GET(cmci_mask) != 0 && boot)
938 lapic_enable_cmc();
939 #endif
940 }
941
942 load_cr4(rcr4() | CR4_MCE);
943 }
944
945 /* Must be executed on each CPU during boot. */
946 void
947 mca_init(void)
948 {
949
950 _mca_init(1);
951 }
952
953 /* Must be executed on each CPU during resume. */
954 void
955 mca_resume(void)
956 {
957
958 _mca_init(0);
959 }
960
961 /*
962 * The machine check registers for the BSP cannot be initialized until
963 * the local APIC is initialized. This happens at SI_SUB_CPU,
964 * SI_ORDER_SECOND.
965 */
966 static void
967 mca_init_bsp(void *arg __unused)
968 {
969
970 mca_init();
971 }
972 SYSINIT(mca_init_bsp, SI_SUB_CPU, SI_ORDER_ANY, mca_init_bsp, NULL);
973
974 /* Called when a machine check exception fires. */
975 void
976 mca_intr(void)
977 {
978 uint64_t mcg_status;
979 int old_count, recoverable;
980
981 if (!(cpu_feature & CPUID_MCA)) {
982 /*
983 * Just print the values of the old Pentium registers
984 * and panic.
985 */
986 printf("MC Type: 0x%jx Address: 0x%jx\n",
987 (uintmax_t)rdmsr(MSR_P5_MC_TYPE),
988 (uintmax_t)rdmsr(MSR_P5_MC_ADDR));
989 panic("Machine check");
990 }
991
992 /* Scan the banks and check for any non-recoverable errors. */
993 old_count = mca_count;
994 recoverable = mca_scan(MCE);
995 mcg_status = rdmsr(MSR_MCG_STATUS);
996 if (!(mcg_status & MCG_STATUS_RIPV))
997 recoverable = 0;
998
999 if (!recoverable) {
1000 /*
1001 * Wait for at least one error to be logged before
1002 * panic'ing. Some errors will assert a machine check
1003 * on all CPUs, but only certain CPUs will find a valid
1004 * bank to log.
1005 */
1006 while (mca_count == old_count)
1007 cpu_spinwait();
1008
1009 panic("Unrecoverable machine check exception");
1010 }
1011
1012 /* Clear MCIP. */
1013 wrmsr(MSR_MCG_STATUS, mcg_status & ~MCG_STATUS_MCIP);
1014 }
1015
1016 #ifdef DEV_APIC
1017 /* Called for a CMCI (correctable machine check interrupt). */
1018 void
1019 cmc_intr(void)
1020 {
1021 struct mca_internal *mca;
1022 int count;
1023
1024 /*
1025 * Serialize MCA bank scanning to prevent collisions from
1026 * sibling threads.
1027 */
1028 count = mca_scan(CMCI);
1029
1030 /* If we found anything, log them to the console. */
1031 if (count != 0) {
1032 mtx_lock_spin(&mca_lock);
1033 STAILQ_FOREACH(mca, &mca_records, link) {
1034 if (!mca->logged) {
1035 mca->logged = 1;
1036 mca_log(&mca->rec);
1037 }
1038 }
1039 mtx_unlock_spin(&mca_lock);
1040 }
1041 }
1042 #endif
Cache object: dc646f168011ae31929a7ff70ffd309d
|