2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2009 Hudson River Trading LLC
5 * Written by: John H. Baldwin <jhb@FreeBSD.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * Support for x86 machine check architecture.
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
43 #include <sys/param.h>
45 #include <sys/interrupt.h>
46 #include <sys/kernel.h>
48 #include <sys/malloc.h>
49 #include <sys/mutex.h>
51 #include <sys/sched.h>
53 #include <sys/sysctl.h>
54 #include <sys/systm.h>
55 #include <sys/taskqueue.h>
56 #include <machine/intr_machdep.h>
57 #include <x86/apicvar.h>
58 #include <machine/cpu.h>
59 #include <machine/cputypes.h>
61 #include <machine/md_var.h>
62 #include <machine/specialreg.h>
64 /* Modes for mca_scan() */
73 * State maintained for each monitored MCx bank to control the
74 * corrected machine check interrupt threshold.
88 struct mca_record rec;
90 STAILQ_ENTRY(mca_internal) link;
93 struct mca_enumerator_ops {
94 unsigned int (*ctl)(int);
95 unsigned int (*status)(int);
96 unsigned int (*addr)(int);
97 unsigned int (*misc)(int);
100 static MALLOC_DEFINE(M_MCA, "MCA", "Machine Check Architecture");
102 static volatile int mca_count; /* Number of records stored. */
103 static int mca_banks; /* Number of per-CPU register banks. */
105 static SYSCTL_NODE(_hw, OID_AUTO, mca, CTLFLAG_RD, NULL,
106 "Machine Check Architecture");
108 static int mca_enabled = 1;
109 SYSCTL_INT(_hw_mca, OID_AUTO, enabled, CTLFLAG_RDTUN, &mca_enabled, 0,
110 "Administrative toggle for machine check support");
112 static int amd10h_L1TP = 1;
113 SYSCTL_INT(_hw_mca, OID_AUTO, amd10h_L1TP, CTLFLAG_RDTUN, &amd10h_L1TP, 0,
114 "Administrative toggle for logging of level one TLB parity (L1TP) errors");
116 static int intel6h_HSD131;
117 SYSCTL_INT(_hw_mca, OID_AUTO, intel6h_HSD131, CTLFLAG_RDTUN, &intel6h_HSD131, 0,
118 "Administrative toggle for logging of spurious corrected errors");
120 int workaround_erratum383;
121 SYSCTL_INT(_hw_mca, OID_AUTO, erratum383, CTLFLAG_RDTUN,
122 &workaround_erratum383, 0,
123 "Is the workaround for Erratum 383 on AMD Family 10h processors enabled?");
125 static STAILQ_HEAD(, mca_internal) mca_freelist;
126 static int mca_freecount;
127 static STAILQ_HEAD(, mca_internal) mca_records;
128 static struct callout mca_timer;
129 static int mca_ticks = 3600; /* Check hourly by default. */
130 static struct taskqueue *mca_tq;
131 static struct task mca_refill_task, mca_scan_task;
132 static struct mtx mca_lock;
135 mca_ia32_ctl_reg(int bank)
137 return (MSR_MC_CTL(bank));
141 mca_ia32_status_reg(int bank)
143 return (MSR_MC_STATUS(bank));
147 mca_ia32_addr_reg(int bank)
149 return (MSR_MC_ADDR(bank));
153 mca_ia32_misc_reg(int bank)
155 return (MSR_MC_MISC(bank));
159 mca_smca_ctl_reg(int bank)
161 return (MSR_SMCA_MC_CTL(bank));
165 mca_smca_status_reg(int bank)
167 return (MSR_SMCA_MC_STATUS(bank));
171 mca_smca_addr_reg(int bank)
173 return (MSR_SMCA_MC_ADDR(bank));
177 mca_smca_misc_reg(int bank)
179 return (MSR_SMCA_MC_MISC(bank));
182 static struct mca_enumerator_ops mca_msr_ops = {
183 .ctl = mca_ia32_ctl_reg,
184 .status = mca_ia32_status_reg,
185 .addr = mca_ia32_addr_reg,
186 .misc = mca_ia32_misc_reg
190 static struct cmc_state **cmc_state; /* Indexed by cpuid, bank. */
191 static struct amd_et_state **amd_et_state; /* Indexed by cpuid, bank. */
192 static int cmc_throttle = 60; /* Time in seconds to throttle CMCI. */
194 static int amd_elvt = -1;
197 amd_thresholding_supported(void)
199 if (cpu_vendor_id != CPU_VENDOR_AMD)
202 * The RASCap register is wholly reserved in families 0x10-0x15 (through model 1F).
204 * It begins to be documented in family 0x15 model 30 and family 0x16,
205 * but neither of these families documents the ScalableMca bit, which
206 * supposedly defines the presence of this feature on family 0x17.
208 if (CPUID_TO_FAMILY(cpu_id) >= 0x10 && CPUID_TO_FAMILY(cpu_id) <= 0x16)
210 if (CPUID_TO_FAMILY(cpu_id) >= 0x17)
211 return ((amd_rascap & AMDRAS_SCALABLE_MCA) != 0);
217 cmci_supported(uint64_t mcg_cap)
220 * MCG_CAP_CMCI_P bit is reserved in AMD documentation. Until
221 * it is defined, do not use it to check for CMCI support.
223 if (cpu_vendor_id != CPU_VENDOR_INTEL)
225 return ((mcg_cap & MCG_CAP_CMCI_P) != 0);
229 sysctl_positive_int(SYSCTL_HANDLER_ARGS)
233 value = *(int *)arg1;
234 error = sysctl_handle_int(oidp, &value, 0, req);
235 if (error || req->newptr == NULL)
239 *(int *)arg1 = value;
244 sysctl_mca_records(SYSCTL_HANDLER_ARGS)
246 int *name = (int *)arg1;
247 u_int namelen = arg2;
248 struct mca_record record;
249 struct mca_internal *rec;
255 if (name[0] < 0 || name[0] >= mca_count)
258 mtx_lock_spin(&mca_lock);
259 if (name[0] >= mca_count) {
260 mtx_unlock_spin(&mca_lock);
264 STAILQ_FOREACH(rec, &mca_records, link) {
271 mtx_unlock_spin(&mca_lock);
272 return (SYSCTL_OUT(req, &record, sizeof(record)));
276 mca_error_ttype(uint16_t mca_error)
279 switch ((mca_error & 0x000c) >> 2) {
291 mca_error_level(uint16_t mca_error)
294 switch (mca_error & 0x0003) {
308 mca_error_request(uint16_t mca_error)
311 switch ((mca_error & 0x00f0) >> 4) {
335 mca_error_mmtype(uint16_t mca_error)
338 switch ((mca_error & 0x70) >> 4) {
354 mca_mute(const struct mca_record *rec)
358 * Skip spurious corrected parity errors generated by Intel Haswell-
359 * and Broadwell-based CPUs (see HSD131, HSM142, HSW131 and BDM48
360 * erratum respectively), unless reporting is enabled.
361 * Note that these errors also have been observed with the D0-stepping
362 * of Haswell, while at least initially the CPU specification updates
363 * suggested only the C0-stepping to be affected. Similarly, Celeron
364 * 2955U with a CPU ID of 0x45 apparently are also concerned with the
365 * same problem, with HSM142 only referring to 0x3c and 0x46.
367 if (cpu_vendor_id == CPU_VENDOR_INTEL &&
368 CPUID_TO_FAMILY(cpu_id) == 0x6 &&
369 (CPUID_TO_MODEL(cpu_id) == 0x3c || /* HSD131, HSM142, HSW131 */
370 CPUID_TO_MODEL(cpu_id) == 0x3d || /* BDM48 */
371 CPUID_TO_MODEL(cpu_id) == 0x45 ||
372 CPUID_TO_MODEL(cpu_id) == 0x46) && /* HSM142 */
374 (rec->mr_status & 0xa0000000ffffffff) == 0x80000000000f0005 &&
381 /* Dump details about a single machine check. */
383 mca_log(const struct mca_record *rec)
390 printf("MCA: Bank %d, Status 0x%016llx\n", rec->mr_bank,
391 (long long)rec->mr_status);
392 printf("MCA: Global Cap 0x%016llx, Status 0x%016llx\n",
393 (long long)rec->mr_mcg_cap, (long long)rec->mr_mcg_status);
394 printf("MCA: Vendor \"%s\", ID 0x%x, APIC ID %d\n", cpu_vendor,
395 rec->mr_cpu_id, rec->mr_apic_id);
396 printf("MCA: CPU %d ", rec->mr_cpu);
397 if (rec->mr_status & MC_STATUS_UC)
401 if (cmci_supported(rec->mr_mcg_cap))
402 printf("(%lld) ", ((long long)rec->mr_status &
403 MC_STATUS_COR_COUNT) >> 38);
405 if (rec->mr_status & MC_STATUS_PCC)
407 if (rec->mr_status & MC_STATUS_OVER)
409 mca_error = rec->mr_status & MC_STATUS_MCA_ERROR;
411 /* Simple error codes. */
416 printf("unclassified error");
419 printf("ucode ROM parity error");
422 printf("external error");
428 printf("internal parity error");
431 printf("internal timer error");
434 if ((mca_error & 0xfc00) == 0x0400) {
435 printf("internal error %x", mca_error & 0x03ff);
439 /* Compound error codes. */
441 /* Memory hierarchy error. */
442 if ((mca_error & 0xeffc) == 0x000c) {
443 printf("%s memory error", mca_error_level(mca_error));
448 if ((mca_error & 0xeff0) == 0x0010) {
449 printf("%sTLB %s error", mca_error_ttype(mca_error),
450 mca_error_level(mca_error));
454 /* Memory controller error. */
455 if ((mca_error & 0xef80) == 0x0080) {
456 printf("%s channel ", mca_error_mmtype(mca_error));
457 if ((mca_error & 0x000f) != 0x000f)
458 printf("%d", mca_error & 0x000f);
461 printf(" memory error");
466 if ((mca_error & 0xef00) == 0x0100) {
467 printf("%sCACHE %s %s error",
468 mca_error_ttype(mca_error),
469 mca_error_level(mca_error),
470 mca_error_request(mca_error));
474 /* Bus and/or Interconnect error. */
475 if ((mca_error & 0xe800) == 0x0800) {
476 printf("BUS%s ", mca_error_level(mca_error));
477 switch ((mca_error & 0x0600) >> 9) {
491 printf(" %s ", mca_error_request(mca_error));
492 switch ((mca_error & 0x000c) >> 2) {
506 if (mca_error & 0x0100)
507 printf(" timed out");
511 printf("unknown error %x", mca_error);
515 if (rec->mr_status & MC_STATUS_ADDRV)
516 printf("MCA: Address 0x%llx\n", (long long)rec->mr_addr);
517 if (rec->mr_status & MC_STATUS_MISCV)
518 printf("MCA: Misc 0x%llx\n", (long long)rec->mr_misc);
522 mca_check_status(int bank, struct mca_record *rec)
527 status = rdmsr(mca_msr_ops.status(bank));
528 if (!(status & MC_STATUS_VAL))
531 /* Save exception information. */
532 rec->mr_status = status;
535 if (status & MC_STATUS_ADDRV)
536 rec->mr_addr = rdmsr(mca_msr_ops.addr(bank));
538 if (status & MC_STATUS_MISCV)
539 rec->mr_misc = rdmsr(mca_msr_ops.misc(bank));
540 rec->mr_tsc = rdtsc();
541 rec->mr_apic_id = PCPU_GET(apic_id);
542 rec->mr_mcg_cap = rdmsr(MSR_MCG_CAP);
543 rec->mr_mcg_status = rdmsr(MSR_MCG_STATUS);
544 rec->mr_cpu_id = cpu_id;
545 rec->mr_cpu_vendor_id = cpu_vendor_id;
546 rec->mr_cpu = PCPU_GET(cpuid);
549 * Clear machine check. Don't do this for uncorrectable
550 * errors so that the BIOS can see them.
552 if (!(rec->mr_status & (MC_STATUS_PCC | MC_STATUS_UC))) {
553 wrmsr(mca_msr_ops.status(bank), 0);
560 mca_fill_freelist(void)
562 struct mca_internal *rec;
566 * Ensure we have at least one record for each bank and one
569 desired = imax(mp_ncpus, mca_banks);
570 mtx_lock_spin(&mca_lock);
571 while (mca_freecount < desired) {
572 mtx_unlock_spin(&mca_lock);
573 rec = malloc(sizeof(*rec), M_MCA, M_WAITOK);
574 mtx_lock_spin(&mca_lock);
575 STAILQ_INSERT_TAIL(&mca_freelist, rec, link);
578 mtx_unlock_spin(&mca_lock);
582 mca_refill(void *context, int pending)
589 mca_record_entry(enum scan_mode mode, const struct mca_record *record)
591 struct mca_internal *rec;
593 if (mode == POLLED) {
594 rec = malloc(sizeof(*rec), M_MCA, M_WAITOK);
595 mtx_lock_spin(&mca_lock);
597 mtx_lock_spin(&mca_lock);
598 rec = STAILQ_FIRST(&mca_freelist);
600 printf("MCA: Unable to allocate space for an event.\n");
602 mtx_unlock_spin(&mca_lock);
605 STAILQ_REMOVE_HEAD(&mca_freelist, link);
611 STAILQ_INSERT_TAIL(&mca_records, rec, link);
613 mtx_unlock_spin(&mca_lock);
614 if (mode == CMCI && !cold)
615 taskqueue_enqueue(mca_tq, &mca_refill_task);
620 * Update the interrupt threshold for a CMCI. The strategy is to use
621 * a low trigger that interrupts as soon as the first event occurs.
622 * However, if a steady stream of events arrive, the threshold is
623 * increased until the interrupts are throttled to once every
624 * cmc_throttle seconds or the periodic scan. If a periodic scan
625 * finds that the threshold is too high, it is lowered.
628 update_threshold(enum scan_mode mode, int valid, int last_intr, int count,
629 int cur_threshold, int max_threshold)
634 delta = (u_int)(time_uptime - last_intr);
635 limit = cur_threshold;
638 * If an interrupt was received less than cmc_throttle seconds
639 * since the previous interrupt and the count from the current
640 * event is greater than or equal to the current threshold,
641 * double the threshold up to the max.
643 if (mode == CMCI && valid) {
644 if (delta < cmc_throttle && count >= limit &&
645 limit < max_threshold) {
646 limit = min(limit << 1, max_threshold);
652 * When the banks are polled, check to see if the threshold
658 /* If a CMCI occured recently, do nothing for now. */
659 if (delta < cmc_throttle)
663 * Compute a new limit based on the average rate of events per
664 * cmc_throttle seconds since the last interrupt.
667 limit = count * cmc_throttle / delta;
670 else if (limit > max_threshold)
671 limit = max_threshold;
679 cmci_update(enum scan_mode mode, int bank, int valid, struct mca_record *rec)
681 struct cmc_state *cc;
683 int cur_threshold, new_threshold;
686 /* Fetch the current limit for this bank. */
687 cc = &cmc_state[PCPU_GET(cpuid)][bank];
688 ctl = rdmsr(MSR_MC_CTL2(bank));
689 count = (rec->mr_status & MC_STATUS_COR_COUNT) >> 38;
690 cur_threshold = ctl & MC_CTL2_THRESHOLD;
692 new_threshold = update_threshold(mode, valid, cc->last_intr, count,
693 cur_threshold, cc->max_threshold);
695 if (mode == CMCI && valid)
696 cc->last_intr = time_uptime;
697 if (new_threshold != cur_threshold) {
698 ctl &= ~MC_CTL2_THRESHOLD;
699 ctl |= new_threshold;
700 wrmsr(MSR_MC_CTL2(bank), ctl);
705 amd_thresholding_update(enum scan_mode mode, int bank, int valid)
707 struct amd_et_state *cc;
712 cc = &amd_et_state[PCPU_GET(cpuid)][bank];
713 misc = rdmsr(mca_msr_ops.misc(bank));
714 count = (misc & MC_MISC_AMD_CNT_MASK) >> MC_MISC_AMD_CNT_SHIFT;
715 count = count - (MC_MISC_AMD_CNT_MAX - cc->cur_threshold);
717 new_threshold = update_threshold(mode, valid, cc->last_intr, count,
718 cc->cur_threshold, MC_MISC_AMD_CNT_MAX);
720 cc->cur_threshold = new_threshold;
721 misc &= ~MC_MISC_AMD_CNT_MASK;
722 misc |= (uint64_t)(MC_MISC_AMD_CNT_MAX - cc->cur_threshold)
723 << MC_MISC_AMD_CNT_SHIFT;
724 misc &= ~MC_MISC_AMD_OVERFLOW;
725 wrmsr(mca_msr_ops.misc(bank), misc);
726 if (mode == CMCI && valid)
727 cc->last_intr = time_uptime;
732 * This scans all the machine check banks of the current CPU to see if
733 * there are any machine checks. Any non-recoverable errors are
734 * reported immediately via mca_log(). The current thread must be
735 * pinned when this is called. The 'mode' parameter indicates if we
736 * are being called from the MC exception handler, the CMCI handler,
737 * or the periodic poller. In the MC exception case this function
738 * returns true if the system is restartable. Otherwise, it returns a
739 * count of the number of valid MC records found.
742 mca_scan(enum scan_mode mode, int *recoverablep)
744 struct mca_record rec;
745 uint64_t mcg_cap, ucmask;
746 int count, i, recoverable, valid;
750 ucmask = MC_STATUS_UC | MC_STATUS_PCC;
752 /* When handling a MCE#, treat the OVER flag as non-restartable. */
754 ucmask |= MC_STATUS_OVER;
755 mcg_cap = rdmsr(MSR_MCG_CAP);
756 for (i = 0; i < (mcg_cap & MCG_CAP_COUNT); i++) {
759 * For a CMCI, only check banks this CPU is
762 if (mode == CMCI && !(PCPU_GET(cmci_mask) & 1 << i))
766 valid = mca_check_status(i, &rec);
769 if (rec.mr_status & ucmask) {
771 mtx_lock_spin(&mca_lock);
773 mtx_unlock_spin(&mca_lock);
775 mca_record_entry(mode, &rec);
780 * If this is a bank this CPU monitors via CMCI,
781 * update the threshold.
783 if (PCPU_GET(cmci_mask) & 1 << i) {
784 if (cmc_state != NULL)
785 cmci_update(mode, i, valid, &rec);
787 amd_thresholding_update(mode, i, valid);
793 if (recoverablep != NULL)
794 *recoverablep = recoverable;
799 * Scan the machine check banks on all CPUs by binding to each CPU in
800 * turn. If any of the CPUs contained new machine check records, log
801 * them to the console.
804 mca_scan_cpus(void *context, int pending)
806 struct mca_internal *mca;
817 count += mca_scan(POLLED, NULL);
823 mtx_lock_spin(&mca_lock);
824 STAILQ_FOREACH(mca, &mca_records, link) {
830 mtx_unlock_spin(&mca_lock);
835 mca_periodic_scan(void *arg)
838 taskqueue_enqueue(mca_tq, &mca_scan_task);
839 callout_reset(&mca_timer, mca_ticks * hz, mca_periodic_scan, NULL);
843 sysctl_mca_scan(SYSCTL_HANDLER_ARGS)
848 error = sysctl_handle_int(oidp, &i, 0, req);
852 taskqueue_enqueue(mca_tq, &mca_scan_task);
857 mca_createtq(void *dummy)
862 mca_tq = taskqueue_create_fast("mca", M_WAITOK,
863 taskqueue_thread_enqueue, &mca_tq);
864 taskqueue_start_threads(&mca_tq, 1, PI_SWI(SWI_TQ), "mca taskq");
866 /* CMCIs during boot may have claimed items from the freelist. */
869 SYSINIT(mca_createtq, SI_SUB_CONFIGURE, SI_ORDER_ANY, mca_createtq, NULL);
872 mca_startup(void *dummy)
878 callout_reset(&mca_timer, mca_ticks * hz, mca_periodic_scan, NULL);
880 #ifdef EARLY_AP_STARTUP
881 SYSINIT(mca_startup, SI_SUB_KICK_SCHEDULER, SI_ORDER_ANY, mca_startup, NULL);
883 SYSINIT(mca_startup, SI_SUB_SMP, SI_ORDER_ANY, mca_startup, NULL);
892 cmc_state = malloc((mp_maxid + 1) * sizeof(struct cmc_state *), M_MCA,
894 for (i = 0; i <= mp_maxid; i++)
895 cmc_state[i] = malloc(sizeof(struct cmc_state) * mca_banks,
896 M_MCA, M_WAITOK | M_ZERO);
897 SYSCTL_ADD_PROC(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
898 "cmc_throttle", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
899 &cmc_throttle, 0, sysctl_positive_int, "I",
900 "Interval in seconds to throttle corrected MC interrupts");
904 amd_thresholding_setup(void)
908 amd_et_state = malloc((mp_maxid + 1) * sizeof(struct amd_et_state *),
910 for (i = 0; i <= mp_maxid; i++)
911 amd_et_state[i] = malloc(sizeof(struct amd_et_state) *
912 mca_banks, M_MCA, M_WAITOK | M_ZERO);
913 SYSCTL_ADD_PROC(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
914 "cmc_throttle", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
915 &cmc_throttle, 0, sysctl_positive_int, "I",
916 "Interval in seconds to throttle corrected MC interrupts");
921 mca_setup(uint64_t mcg_cap)
925 * On AMD Family 10h processors, unless logging of level one TLB
926 * parity (L1TP) errors is disabled, enable the recommended workaround
929 if (cpu_vendor_id == CPU_VENDOR_AMD &&
930 CPUID_TO_FAMILY(cpu_id) == 0x10 && amd10h_L1TP)
931 workaround_erratum383 = 1;
933 mca_banks = mcg_cap & MCG_CAP_COUNT;
934 mtx_init(&mca_lock, "mca", NULL, MTX_SPIN);
935 STAILQ_INIT(&mca_records);
936 TASK_INIT(&mca_scan_task, 0, mca_scan_cpus, NULL);
937 callout_init(&mca_timer, 1);
938 STAILQ_INIT(&mca_freelist);
939 TASK_INIT(&mca_refill_task, 0, mca_refill, NULL);
941 SYSCTL_ADD_INT(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
942 "count", CTLFLAG_RD, (int *)(uintptr_t)&mca_count, 0,
944 SYSCTL_ADD_PROC(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
945 "interval", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, &mca_ticks,
946 0, sysctl_positive_int, "I",
947 "Periodic interval in seconds to scan for machine checks");
948 SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
949 "records", CTLFLAG_RD, sysctl_mca_records, "Machine check records");
950 SYSCTL_ADD_PROC(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
951 "force_scan", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
952 sysctl_mca_scan, "I", "Force an immediate scan for machine checks");
954 if (cmci_supported(mcg_cap))
956 else if (amd_thresholding_supported())
957 amd_thresholding_setup();
963 * See if we should monitor CMCI for this bank. If CMCI_EN is already
964 * set in MC_CTL2, then another CPU is responsible for this bank, so
965 * ignore it. If CMCI_EN returns zero after being set, then this bank
966 * does not support CMCI_EN. If this CPU sets CMCI_EN, then it should
967 * now monitor this bank.
972 struct cmc_state *cc;
975 KASSERT(i < mca_banks, ("CPU %d has more MC banks", PCPU_GET(cpuid)));
977 ctl = rdmsr(MSR_MC_CTL2(i));
978 if (ctl & MC_CTL2_CMCI_EN)
979 /* Already monitored by another CPU. */
982 /* Set the threshold to one event for now. */
983 ctl &= ~MC_CTL2_THRESHOLD;
984 ctl |= MC_CTL2_CMCI_EN | 1;
985 wrmsr(MSR_MC_CTL2(i), ctl);
986 ctl = rdmsr(MSR_MC_CTL2(i));
987 if (!(ctl & MC_CTL2_CMCI_EN))
988 /* This bank does not support CMCI. */
991 cc = &cmc_state[PCPU_GET(cpuid)][i];
993 /* Determine maximum threshold. */
994 ctl &= ~MC_CTL2_THRESHOLD;
996 wrmsr(MSR_MC_CTL2(i), ctl);
997 ctl = rdmsr(MSR_MC_CTL2(i));
998 cc->max_threshold = ctl & MC_CTL2_THRESHOLD;
1000 /* Start off with a threshold of 1. */
1001 ctl &= ~MC_CTL2_THRESHOLD;
1003 wrmsr(MSR_MC_CTL2(i), ctl);
1005 /* Mark this bank as monitored. */
1006 PCPU_SET(cmci_mask, PCPU_GET(cmci_mask) | 1 << i);
1010 * For resume, reset the threshold for any banks we monitor back to
1011 * one and throw away the timestamp of the last interrupt.
1016 struct cmc_state *cc;
1019 KASSERT(i < mca_banks, ("CPU %d has more MC banks", PCPU_GET(cpuid)));
1021 /* Ignore banks not monitored by this CPU. */
1022 if (!(PCPU_GET(cmci_mask) & 1 << i))
1025 cc = &cmc_state[PCPU_GET(cpuid)][i];
1027 ctl = rdmsr(MSR_MC_CTL2(i));
1028 ctl &= ~MC_CTL2_THRESHOLD;
1029 ctl |= MC_CTL2_CMCI_EN | 1;
1030 wrmsr(MSR_MC_CTL2(i), ctl);
1034 * Apply an AMD ET configuration to the corresponding MSR.
1037 amd_thresholding_start(struct amd_et_state *cc, int bank)
1041 KASSERT(amd_elvt >= 0, ("ELVT offset is not set"));
1043 misc = rdmsr(mca_msr_ops.misc(bank));
1045 misc &= ~MC_MISC_AMD_INT_MASK;
1046 misc |= MC_MISC_AMD_INT_LVT;
1048 misc &= ~MC_MISC_AMD_LVT_MASK;
1049 misc |= (uint64_t)amd_elvt << MC_MISC_AMD_LVT_SHIFT;
1051 misc &= ~MC_MISC_AMD_CNT_MASK;
1052 misc |= (uint64_t)(MC_MISC_AMD_CNT_MAX - cc->cur_threshold)
1053 << MC_MISC_AMD_CNT_SHIFT;
1055 misc &= ~MC_MISC_AMD_OVERFLOW;
1056 misc |= MC_MISC_AMD_CNTEN;
1058 wrmsr(mca_msr_ops.misc(bank), misc);
1062 amd_thresholding_monitor(int i)
1064 struct amd_et_state *cc;
1068 * Kludge: On 10h, banks after 4 are not thresholding but also may have
1069 * bogus Valid bits. Skip them. This is definitely fixed in 15h, but
1070 * I have not investigated whether it is fixed in earlier models.
1072 if (CPUID_TO_FAMILY(cpu_id) < 0x15 && i >= 5)
1075 /* The counter must be valid and present. */
1076 misc = rdmsr(mca_msr_ops.misc(i));
1077 if ((misc & (MC_MISC_AMD_VAL | MC_MISC_AMD_CNTP)) !=
1078 (MC_MISC_AMD_VAL | MC_MISC_AMD_CNTP))
1081 /* The register should not be locked. */
1082 if ((misc & MC_MISC_AMD_LOCK) != 0) {
1084 printf("%s: 0x%jx: Bank %d: locked\n", __func__,
1085 (uintmax_t)misc, i);
1090 * If counter is enabled then either the firmware or another CPU
1091 * has already claimed it.
1093 if ((misc & MC_MISC_AMD_CNTEN) != 0) {
1095 printf("%s: 0x%jx: Bank %d: already enabled\n",
1096 __func__, (uintmax_t)misc, i);
1101 * Configure an Extended Interrupt LVT register for reporting
1102 * counter overflows if that feature is supported and the first
1103 * extended register is available.
1105 amd_elvt = lapic_enable_mca_elvt();
1107 printf("%s: Bank %d: lapic enable mca elvt failed: %d\n",
1108 __func__, i, amd_elvt);
1112 /* Re-use Intel CMC support infrastructure. */
1114 printf("%s: Starting AMD thresholding on bank %d\n", __func__,
1117 cc = &amd_et_state[PCPU_GET(cpuid)][i];
1118 cc->cur_threshold = 1;
1119 amd_thresholding_start(cc, i);
1121 /* Mark this bank as monitored. */
1122 PCPU_SET(cmci_mask, PCPU_GET(cmci_mask) | 1 << i);
1126 amd_thresholding_resume(int i)
1128 struct amd_et_state *cc;
1130 KASSERT(i < mca_banks, ("CPU %d has more MC banks", PCPU_GET(cpuid)));
1132 /* Ignore banks not monitored by this CPU. */
1133 if (!(PCPU_GET(cmci_mask) & 1 << i))
1136 cc = &amd_et_state[PCPU_GET(cpuid)][i];
1138 cc->cur_threshold = 1;
1139 amd_thresholding_start(cc, i);
1144 * Initializes per-CPU machine check registers and enables corrected
1145 * machine check interrupts.
1152 int i, skip, family;
1154 family = CPUID_TO_FAMILY(cpu_id);
1156 /* MCE is required. */
1157 if (!mca_enabled || !(cpu_feature & CPUID_MCE))
1160 if (cpu_feature & CPUID_MCA) {
1162 PCPU_SET(cmci_mask, 0);
1164 mcg_cap = rdmsr(MSR_MCG_CAP);
1165 if (mcg_cap & MCG_CAP_CTL_P)
1166 /* Enable MCA features. */
1167 wrmsr(MSR_MCG_CTL, MCG_CTL_ENABLE);
1168 if (IS_BSP() && boot)
1172 * Disable logging of level one TLB parity (L1TP) errors by
1173 * the data cache as an alternative workaround for AMD Family
1174 * 10h Erratum 383. Unlike the recommended workaround, there
1175 * is no performance penalty to this workaround. However,
1176 * L1TP errors will go unreported.
1178 if (cpu_vendor_id == CPU_VENDOR_AMD && family == 0x10 &&
1180 mask = rdmsr(MSR_MC0_CTL_MASK);
1181 if ((mask & (1UL << 5)) == 0)
1182 wrmsr(MSR_MC0_CTL_MASK, mask | (1UL << 5));
1184 if (amd_rascap & AMDRAS_SCALABLE_MCA) {
1185 mca_msr_ops.ctl = mca_smca_ctl_reg;
1186 mca_msr_ops.status = mca_smca_status_reg;
1187 mca_msr_ops.addr = mca_smca_addr_reg;
1188 mca_msr_ops.misc = mca_smca_misc_reg;
1192 * The cmci_monitor() must not be executed
1193 * simultaneously by several CPUs.
1196 mtx_lock_spin(&mca_lock);
1198 for (i = 0; i < (mcg_cap & MCG_CAP_COUNT); i++) {
1199 /* By default enable logging of all errors. */
1200 ctl = 0xffffffffffffffffUL;
1203 if (cpu_vendor_id == CPU_VENDOR_INTEL) {
1205 * For P6 models before Nehalem MC0_CTL is
1206 * always enabled and reserved.
1208 if (i == 0 && family == 0x6
1209 && CPUID_TO_MODEL(cpu_id) < 0x1a)
1211 } else if (cpu_vendor_id == CPU_VENDOR_AMD) {
1212 /* BKDG for Family 10h: unset GartTblWkEn. */
1213 if (i == MC_AMDNB_BANK && family >= 0xf &&
1215 ctl &= ~(1UL << 10);
1219 wrmsr(mca_msr_ops.ctl(i), ctl);
1222 if (cmci_supported(mcg_cap)) {
1227 } else if (amd_thresholding_supported()) {
1229 amd_thresholding_monitor(i);
1231 amd_thresholding_resume(i);
1235 /* Clear all errors. */
1236 wrmsr(mca_msr_ops.status(i), 0);
1239 mtx_unlock_spin(&mca_lock);
1242 if (!amd_thresholding_supported() &&
1243 PCPU_GET(cmci_mask) != 0 && boot)
1248 load_cr4(rcr4() | CR4_MCE);
1251 /* Must be executed on each CPU during boot. */
1259 /* Must be executed on each CPU during resume. */
1268 * The machine check registers for the BSP cannot be initialized until
1269 * the local APIC is initialized. This happens at SI_SUB_CPU,
1273 mca_init_bsp(void *arg __unused)
1278 SYSINIT(mca_init_bsp, SI_SUB_CPU, SI_ORDER_ANY, mca_init_bsp, NULL);
1280 /* Called when a machine check exception fires. */
1284 uint64_t mcg_status;
1285 int recoverable, count;
1287 if (!(cpu_feature & CPUID_MCA)) {
1289 * Just print the values of the old Pentium registers
1292 printf("MC Type: 0x%jx Address: 0x%jx\n",
1293 (uintmax_t)rdmsr(MSR_P5_MC_TYPE),
1294 (uintmax_t)rdmsr(MSR_P5_MC_ADDR));
1295 panic("Machine check");
1298 /* Scan the banks and check for any non-recoverable errors. */
1299 count = mca_scan(MCE, &recoverable);
1300 mcg_status = rdmsr(MSR_MCG_STATUS);
1301 if (!(mcg_status & MCG_STATUS_RIPV))
1306 * Only panic if the error was detected local to this CPU.
1307 * Some errors will assert a machine check on all CPUs, but
1308 * only certain CPUs will find a valid bank to log.
1313 panic("Unrecoverable machine check exception");
1317 wrmsr(MSR_MCG_STATUS, mcg_status & ~MCG_STATUS_MCIP);
1321 /* Called for a CMCI (correctable machine check interrupt). */
1325 struct mca_internal *mca;
1329 * Serialize MCA bank scanning to prevent collisions from
1332 count = mca_scan(CMCI, NULL);
1334 /* If we found anything, log them to the console. */
1336 mtx_lock_spin(&mca_lock);
1337 STAILQ_FOREACH(mca, &mca_records, link) {
1343 mtx_unlock_spin(&mca_lock);