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 static MALLOC_DEFINE(M_MCA, "MCA", "Machine Check Architecture");
95 static volatile int mca_count; /* Number of records stored. */
96 static int mca_banks; /* Number of per-CPU register banks. */
98 static SYSCTL_NODE(_hw, OID_AUTO, mca, CTLFLAG_RD, NULL,
99 "Machine Check Architecture");
101 static int mca_enabled = 1;
102 SYSCTL_INT(_hw_mca, OID_AUTO, enabled, CTLFLAG_RDTUN, &mca_enabled, 0,
103 "Administrative toggle for machine check support");
105 static int amd10h_L1TP = 1;
106 SYSCTL_INT(_hw_mca, OID_AUTO, amd10h_L1TP, CTLFLAG_RDTUN, &amd10h_L1TP, 0,
107 "Administrative toggle for logging of level one TLB parity (L1TP) errors");
109 static int intel6h_HSD131;
110 SYSCTL_INT(_hw_mca, OID_AUTO, intel6h_HSD131, CTLFLAG_RDTUN, &intel6h_HSD131, 0,
111 "Administrative toggle for logging of spurious corrected errors");
113 int workaround_erratum383;
114 SYSCTL_INT(_hw_mca, OID_AUTO, erratum383, CTLFLAG_RDTUN,
115 &workaround_erratum383, 0,
116 "Is the workaround for Erratum 383 on AMD Family 10h processors enabled?");
118 static STAILQ_HEAD(, mca_internal) mca_freelist;
119 static int mca_freecount;
120 static STAILQ_HEAD(, mca_internal) mca_records;
121 static struct callout mca_timer;
122 static int mca_ticks = 3600; /* Check hourly by default. */
123 static struct taskqueue *mca_tq;
124 static struct task mca_refill_task, mca_scan_task;
125 static struct mtx mca_lock;
128 static struct cmc_state **cmc_state; /* Indexed by cpuid, bank. */
129 static struct amd_et_state **amd_et_state; /* Indexed by cpuid, bank. */
130 static int cmc_throttle = 60; /* Time in seconds to throttle CMCI. */
132 static int amd_elvt = -1;
135 amd_thresholding_supported(void)
137 if (cpu_vendor_id != CPU_VENDOR_AMD &&
138 cpu_vendor_id != CPU_VENDOR_HYGON)
141 * The RASCap register is wholly reserved in families 0x10-0x15 (through model 1F).
143 * It begins to be documented in family 0x15 model 30 and family 0x16,
144 * but neither of these families documents the ScalableMca bit, which
145 * supposedly defines the presence of this feature on family 0x17.
147 if (CPUID_TO_FAMILY(cpu_id) >= 0x10 && CPUID_TO_FAMILY(cpu_id) <= 0x16)
149 if (CPUID_TO_FAMILY(cpu_id) >= 0x17)
150 return ((amd_rascap & AMDRAS_SCALABLE_MCA) != 0);
156 cmci_supported(uint64_t mcg_cap)
159 * MCG_CAP_CMCI_P bit is reserved in AMD documentation. Until
160 * it is defined, do not use it to check for CMCI support.
162 if (cpu_vendor_id != CPU_VENDOR_INTEL)
164 return ((mcg_cap & MCG_CAP_CMCI_P) != 0);
168 sysctl_positive_int(SYSCTL_HANDLER_ARGS)
172 value = *(int *)arg1;
173 error = sysctl_handle_int(oidp, &value, 0, req);
174 if (error || req->newptr == NULL)
178 *(int *)arg1 = value;
183 sysctl_mca_records(SYSCTL_HANDLER_ARGS)
185 int *name = (int *)arg1;
186 u_int namelen = arg2;
187 struct mca_record record;
188 struct mca_internal *rec;
194 if (name[0] < 0 || name[0] >= mca_count)
197 mtx_lock_spin(&mca_lock);
198 if (name[0] >= mca_count) {
199 mtx_unlock_spin(&mca_lock);
203 STAILQ_FOREACH(rec, &mca_records, link) {
210 mtx_unlock_spin(&mca_lock);
211 return (SYSCTL_OUT(req, &record, sizeof(record)));
215 mca_error_ttype(uint16_t mca_error)
218 switch ((mca_error & 0x000c) >> 2) {
230 mca_error_level(uint16_t mca_error)
233 switch (mca_error & 0x0003) {
247 mca_error_request(uint16_t mca_error)
250 switch ((mca_error & 0x00f0) >> 4) {
274 mca_error_mmtype(uint16_t mca_error)
277 switch ((mca_error & 0x70) >> 4) {
293 mca_mute(const struct mca_record *rec)
297 * Skip spurious corrected parity errors generated by Intel Haswell-
298 * and Broadwell-based CPUs (see HSD131, HSM142, HSW131 and BDM48
299 * erratum respectively), unless reporting is enabled.
300 * Note that these errors also have been observed with the D0-stepping
301 * of Haswell, while at least initially the CPU specification updates
302 * suggested only the C0-stepping to be affected. Similarly, Celeron
303 * 2955U with a CPU ID of 0x45 apparently are also concerned with the
304 * same problem, with HSM142 only referring to 0x3c and 0x46.
306 if (cpu_vendor_id == CPU_VENDOR_INTEL &&
307 CPUID_TO_FAMILY(cpu_id) == 0x6 &&
308 (CPUID_TO_MODEL(cpu_id) == 0x3c || /* HSD131, HSM142, HSW131 */
309 CPUID_TO_MODEL(cpu_id) == 0x3d || /* BDM48 */
310 CPUID_TO_MODEL(cpu_id) == 0x45 ||
311 CPUID_TO_MODEL(cpu_id) == 0x46) && /* HSM142 */
313 (rec->mr_status & 0xa0000000ffffffff) == 0x80000000000f0005 &&
320 /* Dump details about a single machine check. */
322 mca_log(const struct mca_record *rec)
329 printf("MCA: Bank %d, Status 0x%016llx\n", rec->mr_bank,
330 (long long)rec->mr_status);
331 printf("MCA: Global Cap 0x%016llx, Status 0x%016llx\n",
332 (long long)rec->mr_mcg_cap, (long long)rec->mr_mcg_status);
333 printf("MCA: Vendor \"%s\", ID 0x%x, APIC ID %d\n", cpu_vendor,
334 rec->mr_cpu_id, rec->mr_apic_id);
335 printf("MCA: CPU %d ", rec->mr_cpu);
336 if (rec->mr_status & MC_STATUS_UC)
340 if (cmci_supported(rec->mr_mcg_cap))
341 printf("(%lld) ", ((long long)rec->mr_status &
342 MC_STATUS_COR_COUNT) >> 38);
344 if (rec->mr_status & MC_STATUS_PCC)
346 if (rec->mr_status & MC_STATUS_OVER)
348 mca_error = rec->mr_status & MC_STATUS_MCA_ERROR;
350 /* Simple error codes. */
355 printf("unclassified error");
358 printf("ucode ROM parity error");
361 printf("external error");
367 printf("internal parity error");
370 printf("internal timer error");
373 if ((mca_error & 0xfc00) == 0x0400) {
374 printf("internal error %x", mca_error & 0x03ff);
378 /* Compound error codes. */
380 /* Memory hierarchy error. */
381 if ((mca_error & 0xeffc) == 0x000c) {
382 printf("%s memory error", mca_error_level(mca_error));
387 if ((mca_error & 0xeff0) == 0x0010) {
388 printf("%sTLB %s error", mca_error_ttype(mca_error),
389 mca_error_level(mca_error));
393 /* Memory controller error. */
394 if ((mca_error & 0xef80) == 0x0080) {
395 printf("%s channel ", mca_error_mmtype(mca_error));
396 if ((mca_error & 0x000f) != 0x000f)
397 printf("%d", mca_error & 0x000f);
400 printf(" memory error");
405 if ((mca_error & 0xef00) == 0x0100) {
406 printf("%sCACHE %s %s error",
407 mca_error_ttype(mca_error),
408 mca_error_level(mca_error),
409 mca_error_request(mca_error));
413 /* Bus and/or Interconnect error. */
414 if ((mca_error & 0xe800) == 0x0800) {
415 printf("BUS%s ", mca_error_level(mca_error));
416 switch ((mca_error & 0x0600) >> 9) {
430 printf(" %s ", mca_error_request(mca_error));
431 switch ((mca_error & 0x000c) >> 2) {
445 if (mca_error & 0x0100)
446 printf(" timed out");
450 printf("unknown error %x", mca_error);
454 if (rec->mr_status & MC_STATUS_ADDRV)
455 printf("MCA: Address 0x%llx\n", (long long)rec->mr_addr);
456 if (rec->mr_status & MC_STATUS_MISCV)
457 printf("MCA: Misc 0x%llx\n", (long long)rec->mr_misc);
461 mca_check_status(int bank, struct mca_record *rec)
466 status = rdmsr(MSR_MC_STATUS(bank));
467 if (!(status & MC_STATUS_VAL))
470 /* Save exception information. */
471 rec->mr_status = status;
474 if (status & MC_STATUS_ADDRV)
475 rec->mr_addr = rdmsr(MSR_MC_ADDR(bank));
477 if (status & MC_STATUS_MISCV)
478 rec->mr_misc = rdmsr(MSR_MC_MISC(bank));
479 rec->mr_tsc = rdtsc();
480 rec->mr_apic_id = PCPU_GET(apic_id);
481 rec->mr_mcg_cap = rdmsr(MSR_MCG_CAP);
482 rec->mr_mcg_status = rdmsr(MSR_MCG_STATUS);
483 rec->mr_cpu_id = cpu_id;
484 rec->mr_cpu_vendor_id = cpu_vendor_id;
485 rec->mr_cpu = PCPU_GET(cpuid);
488 * Clear machine check. Don't do this for uncorrectable
489 * errors so that the BIOS can see them.
491 if (!(rec->mr_status & (MC_STATUS_PCC | MC_STATUS_UC))) {
492 wrmsr(MSR_MC_STATUS(bank), 0);
499 mca_fill_freelist(void)
501 struct mca_internal *rec;
505 * Ensure we have at least one record for each bank and one
508 desired = imax(mp_ncpus, mca_banks);
509 mtx_lock_spin(&mca_lock);
510 while (mca_freecount < desired) {
511 mtx_unlock_spin(&mca_lock);
512 rec = malloc(sizeof(*rec), M_MCA, M_WAITOK);
513 mtx_lock_spin(&mca_lock);
514 STAILQ_INSERT_TAIL(&mca_freelist, rec, link);
517 mtx_unlock_spin(&mca_lock);
521 mca_refill(void *context, int pending)
528 mca_record_entry(enum scan_mode mode, const struct mca_record *record)
530 struct mca_internal *rec;
532 if (mode == POLLED) {
533 rec = malloc(sizeof(*rec), M_MCA, M_WAITOK);
534 mtx_lock_spin(&mca_lock);
536 mtx_lock_spin(&mca_lock);
537 rec = STAILQ_FIRST(&mca_freelist);
539 printf("MCA: Unable to allocate space for an event.\n");
541 mtx_unlock_spin(&mca_lock);
544 STAILQ_REMOVE_HEAD(&mca_freelist, link);
550 STAILQ_INSERT_TAIL(&mca_records, rec, link);
552 mtx_unlock_spin(&mca_lock);
553 if (mode == CMCI && !cold)
554 taskqueue_enqueue(mca_tq, &mca_refill_task);
559 * Update the interrupt threshold for a CMCI. The strategy is to use
560 * a low trigger that interrupts as soon as the first event occurs.
561 * However, if a steady stream of events arrive, the threshold is
562 * increased until the interrupts are throttled to once every
563 * cmc_throttle seconds or the periodic scan. If a periodic scan
564 * finds that the threshold is too high, it is lowered.
567 update_threshold(enum scan_mode mode, int valid, int last_intr, int count,
568 int cur_threshold, int max_threshold)
573 delta = (u_int)(time_uptime - last_intr);
574 limit = cur_threshold;
577 * If an interrupt was received less than cmc_throttle seconds
578 * since the previous interrupt and the count from the current
579 * event is greater than or equal to the current threshold,
580 * double the threshold up to the max.
582 if (mode == CMCI && valid) {
583 if (delta < cmc_throttle && count >= limit &&
584 limit < max_threshold) {
585 limit = min(limit << 1, max_threshold);
591 * When the banks are polled, check to see if the threshold
597 /* If a CMCI occured recently, do nothing for now. */
598 if (delta < cmc_throttle)
602 * Compute a new limit based on the average rate of events per
603 * cmc_throttle seconds since the last interrupt.
606 limit = count * cmc_throttle / delta;
609 else if (limit > max_threshold)
610 limit = max_threshold;
618 cmci_update(enum scan_mode mode, int bank, int valid, struct mca_record *rec)
620 struct cmc_state *cc;
622 int cur_threshold, new_threshold;
625 /* Fetch the current limit for this bank. */
626 cc = &cmc_state[PCPU_GET(cpuid)][bank];
627 ctl = rdmsr(MSR_MC_CTL2(bank));
628 count = (rec->mr_status & MC_STATUS_COR_COUNT) >> 38;
629 cur_threshold = ctl & MC_CTL2_THRESHOLD;
631 new_threshold = update_threshold(mode, valid, cc->last_intr, count,
632 cur_threshold, cc->max_threshold);
634 if (mode == CMCI && valid)
635 cc->last_intr = time_uptime;
636 if (new_threshold != cur_threshold) {
637 ctl &= ~MC_CTL2_THRESHOLD;
638 ctl |= new_threshold;
639 wrmsr(MSR_MC_CTL2(bank), ctl);
644 amd_thresholding_update(enum scan_mode mode, int bank, int valid)
646 struct amd_et_state *cc;
651 cc = &amd_et_state[PCPU_GET(cpuid)][bank];
652 misc = rdmsr(MSR_MC_MISC(bank));
653 count = (misc & MC_MISC_AMD_CNT_MASK) >> MC_MISC_AMD_CNT_SHIFT;
654 count = count - (MC_MISC_AMD_CNT_MAX - cc->cur_threshold);
656 new_threshold = update_threshold(mode, valid, cc->last_intr, count,
657 cc->cur_threshold, MC_MISC_AMD_CNT_MAX);
659 cc->cur_threshold = new_threshold;
660 misc &= ~MC_MISC_AMD_CNT_MASK;
661 misc |= (uint64_t)(MC_MISC_AMD_CNT_MAX - cc->cur_threshold)
662 << MC_MISC_AMD_CNT_SHIFT;
663 misc &= ~MC_MISC_AMD_OVERFLOW;
664 wrmsr(MSR_MC_MISC(bank), misc);
665 if (mode == CMCI && valid)
666 cc->last_intr = time_uptime;
671 * This scans all the machine check banks of the current CPU to see if
672 * there are any machine checks. Any non-recoverable errors are
673 * reported immediately via mca_log(). The current thread must be
674 * pinned when this is called. The 'mode' parameter indicates if we
675 * are being called from the MC exception handler, the CMCI handler,
676 * or the periodic poller. In the MC exception case this function
677 * returns true if the system is restartable. Otherwise, it returns a
678 * count of the number of valid MC records found.
681 mca_scan(enum scan_mode mode, int *recoverablep)
683 struct mca_record rec;
684 uint64_t mcg_cap, ucmask;
685 int count, i, recoverable, valid;
689 ucmask = MC_STATUS_UC | MC_STATUS_PCC;
691 /* When handling a MCE#, treat the OVER flag as non-restartable. */
693 ucmask |= MC_STATUS_OVER;
694 mcg_cap = rdmsr(MSR_MCG_CAP);
695 for (i = 0; i < (mcg_cap & MCG_CAP_COUNT); i++) {
698 * For a CMCI, only check banks this CPU is
701 if (mode == CMCI && !(PCPU_GET(cmci_mask) & 1 << i))
705 valid = mca_check_status(i, &rec);
708 if (rec.mr_status & ucmask) {
710 mtx_lock_spin(&mca_lock);
712 mtx_unlock_spin(&mca_lock);
714 mca_record_entry(mode, &rec);
719 * If this is a bank this CPU monitors via CMCI,
720 * update the threshold.
722 if (PCPU_GET(cmci_mask) & 1 << i) {
723 if (cmc_state != NULL)
724 cmci_update(mode, i, valid, &rec);
726 amd_thresholding_update(mode, i, valid);
732 if (recoverablep != NULL)
733 *recoverablep = recoverable;
738 * Scan the machine check banks on all CPUs by binding to each CPU in
739 * turn. If any of the CPUs contained new machine check records, log
740 * them to the console.
743 mca_scan_cpus(void *context, int pending)
745 struct mca_internal *mca;
756 count += mca_scan(POLLED, NULL);
762 mtx_lock_spin(&mca_lock);
763 STAILQ_FOREACH(mca, &mca_records, link) {
769 mtx_unlock_spin(&mca_lock);
774 mca_periodic_scan(void *arg)
777 taskqueue_enqueue(mca_tq, &mca_scan_task);
778 callout_reset(&mca_timer, mca_ticks * hz, mca_periodic_scan, NULL);
782 sysctl_mca_scan(SYSCTL_HANDLER_ARGS)
787 error = sysctl_handle_int(oidp, &i, 0, req);
791 taskqueue_enqueue(mca_tq, &mca_scan_task);
796 mca_createtq(void *dummy)
801 mca_tq = taskqueue_create_fast("mca", M_WAITOK,
802 taskqueue_thread_enqueue, &mca_tq);
803 taskqueue_start_threads(&mca_tq, 1, PI_SWI(SWI_TQ), "mca taskq");
805 /* CMCIs during boot may have claimed items from the freelist. */
808 SYSINIT(mca_createtq, SI_SUB_CONFIGURE, SI_ORDER_ANY, mca_createtq, NULL);
811 mca_startup(void *dummy)
817 callout_reset(&mca_timer, mca_ticks * hz, mca_periodic_scan, NULL);
819 #ifdef EARLY_AP_STARTUP
820 SYSINIT(mca_startup, SI_SUB_KICK_SCHEDULER, SI_ORDER_ANY, mca_startup, NULL);
822 SYSINIT(mca_startup, SI_SUB_SMP, SI_ORDER_ANY, mca_startup, NULL);
831 cmc_state = malloc((mp_maxid + 1) * sizeof(struct cmc_state *), M_MCA,
833 for (i = 0; i <= mp_maxid; i++)
834 cmc_state[i] = malloc(sizeof(struct cmc_state) * mca_banks,
835 M_MCA, M_WAITOK | M_ZERO);
836 SYSCTL_ADD_PROC(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
837 "cmc_throttle", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
838 &cmc_throttle, 0, sysctl_positive_int, "I",
839 "Interval in seconds to throttle corrected MC interrupts");
843 amd_thresholding_setup(void)
847 amd_et_state = malloc((mp_maxid + 1) * sizeof(struct amd_et_state *),
849 for (i = 0; i <= mp_maxid; i++)
850 amd_et_state[i] = malloc(sizeof(struct amd_et_state) *
851 mca_banks, M_MCA, M_WAITOK | M_ZERO);
852 SYSCTL_ADD_PROC(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
853 "cmc_throttle", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
854 &cmc_throttle, 0, sysctl_positive_int, "I",
855 "Interval in seconds to throttle corrected MC interrupts");
860 mca_setup(uint64_t mcg_cap)
864 * On AMD Family 10h processors, unless logging of level one TLB
865 * parity (L1TP) errors is disabled, enable the recommended workaround
868 if (cpu_vendor_id == CPU_VENDOR_AMD &&
869 CPUID_TO_FAMILY(cpu_id) == 0x10 && amd10h_L1TP)
870 workaround_erratum383 = 1;
872 mca_banks = mcg_cap & MCG_CAP_COUNT;
873 mtx_init(&mca_lock, "mca", NULL, MTX_SPIN);
874 STAILQ_INIT(&mca_records);
875 TASK_INIT(&mca_scan_task, 0, mca_scan_cpus, NULL);
876 callout_init(&mca_timer, 1);
877 STAILQ_INIT(&mca_freelist);
878 TASK_INIT(&mca_refill_task, 0, mca_refill, NULL);
880 SYSCTL_ADD_INT(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
881 "count", CTLFLAG_RD, (int *)(uintptr_t)&mca_count, 0,
883 SYSCTL_ADD_PROC(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
884 "interval", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, &mca_ticks,
885 0, sysctl_positive_int, "I",
886 "Periodic interval in seconds to scan for machine checks");
887 SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
888 "records", CTLFLAG_RD, sysctl_mca_records, "Machine check records");
889 SYSCTL_ADD_PROC(NULL, SYSCTL_STATIC_CHILDREN(_hw_mca), OID_AUTO,
890 "force_scan", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
891 sysctl_mca_scan, "I", "Force an immediate scan for machine checks");
893 if (cmci_supported(mcg_cap))
895 else if (amd_thresholding_supported())
896 amd_thresholding_setup();
902 * See if we should monitor CMCI for this bank. If CMCI_EN is already
903 * set in MC_CTL2, then another CPU is responsible for this bank, so
904 * ignore it. If CMCI_EN returns zero after being set, then this bank
905 * does not support CMCI_EN. If this CPU sets CMCI_EN, then it should
906 * now monitor this bank.
911 struct cmc_state *cc;
914 KASSERT(i < mca_banks, ("CPU %d has more MC banks", PCPU_GET(cpuid)));
917 * It is possible for some APs to report CMCI support even if the BSP
918 * does not, apparently due to a BIOS bug.
920 if (cmc_state == NULL) {
923 "AP %d (%d,%d) reports CMCI support but the BSP does not\n",
924 PCPU_GET(cpuid), PCPU_GET(apic_id),
930 ctl = rdmsr(MSR_MC_CTL2(i));
931 if (ctl & MC_CTL2_CMCI_EN)
932 /* Already monitored by another CPU. */
935 /* Set the threshold to one event for now. */
936 ctl &= ~MC_CTL2_THRESHOLD;
937 ctl |= MC_CTL2_CMCI_EN | 1;
938 wrmsr(MSR_MC_CTL2(i), ctl);
939 ctl = rdmsr(MSR_MC_CTL2(i));
940 if (!(ctl & MC_CTL2_CMCI_EN))
941 /* This bank does not support CMCI. */
944 cc = &cmc_state[PCPU_GET(cpuid)][i];
946 /* Determine maximum threshold. */
947 ctl &= ~MC_CTL2_THRESHOLD;
949 wrmsr(MSR_MC_CTL2(i), ctl);
950 ctl = rdmsr(MSR_MC_CTL2(i));
951 cc->max_threshold = ctl & MC_CTL2_THRESHOLD;
953 /* Start off with a threshold of 1. */
954 ctl &= ~MC_CTL2_THRESHOLD;
956 wrmsr(MSR_MC_CTL2(i), ctl);
958 /* Mark this bank as monitored. */
959 PCPU_SET(cmci_mask, PCPU_GET(cmci_mask) | 1 << i);
963 * For resume, reset the threshold for any banks we monitor back to
964 * one and throw away the timestamp of the last interrupt.
969 struct cmc_state *cc;
972 KASSERT(i < mca_banks, ("CPU %d has more MC banks", PCPU_GET(cpuid)));
974 /* See cmci_monitor(). */
975 if (cmc_state == NULL)
978 /* Ignore banks not monitored by this CPU. */
979 if (!(PCPU_GET(cmci_mask) & 1 << i))
982 cc = &cmc_state[PCPU_GET(cpuid)][i];
984 ctl = rdmsr(MSR_MC_CTL2(i));
985 ctl &= ~MC_CTL2_THRESHOLD;
986 ctl |= MC_CTL2_CMCI_EN | 1;
987 wrmsr(MSR_MC_CTL2(i), ctl);
991 * Apply an AMD ET configuration to the corresponding MSR.
994 amd_thresholding_start(struct amd_et_state *cc, int bank)
998 KASSERT(amd_elvt >= 0, ("ELVT offset is not set"));
1000 misc = rdmsr(MSR_MC_MISC(bank));
1002 misc &= ~MC_MISC_AMD_INT_MASK;
1003 misc |= MC_MISC_AMD_INT_LVT;
1005 misc &= ~MC_MISC_AMD_LVT_MASK;
1006 misc |= (uint64_t)amd_elvt << MC_MISC_AMD_LVT_SHIFT;
1008 misc &= ~MC_MISC_AMD_CNT_MASK;
1009 misc |= (uint64_t)(MC_MISC_AMD_CNT_MAX - cc->cur_threshold)
1010 << MC_MISC_AMD_CNT_SHIFT;
1012 misc &= ~MC_MISC_AMD_OVERFLOW;
1013 misc |= MC_MISC_AMD_CNTEN;
1015 wrmsr(MSR_MC_MISC(bank), misc);
1019 amd_thresholding_monitor(int i)
1021 struct amd_et_state *cc;
1025 * Kludge: On 10h, banks after 4 are not thresholding but also may have
1026 * bogus Valid bits. Skip them. This is definitely fixed in 15h, but
1027 * I have not investigated whether it is fixed in earlier models.
1029 if (CPUID_TO_FAMILY(cpu_id) < 0x15 && i >= 5)
1032 /* The counter must be valid and present. */
1033 misc = rdmsr(MSR_MC_MISC(i));
1034 if ((misc & (MC_MISC_AMD_VAL | MC_MISC_AMD_CNTP)) !=
1035 (MC_MISC_AMD_VAL | MC_MISC_AMD_CNTP))
1038 /* The register should not be locked. */
1039 if ((misc & MC_MISC_AMD_LOCK) != 0) {
1041 printf("%s: 0x%jx: Bank %d: locked\n", __func__,
1042 (uintmax_t)misc, i);
1047 * If counter is enabled then either the firmware or another CPU
1048 * has already claimed it.
1050 if ((misc & MC_MISC_AMD_CNTEN) != 0) {
1052 printf("%s: 0x%jx: Bank %d: already enabled\n",
1053 __func__, (uintmax_t)misc, i);
1058 * Configure an Extended Interrupt LVT register for reporting
1059 * counter overflows if that feature is supported and the first
1060 * extended register is available.
1062 amd_elvt = lapic_enable_mca_elvt();
1064 printf("%s: Bank %d: lapic enable mca elvt failed: %d\n",
1065 __func__, i, amd_elvt);
1069 /* Re-use Intel CMC support infrastructure. */
1071 printf("%s: Starting AMD thresholding on bank %d\n", __func__,
1074 cc = &amd_et_state[PCPU_GET(cpuid)][i];
1075 cc->cur_threshold = 1;
1076 amd_thresholding_start(cc, i);
1078 /* Mark this bank as monitored. */
1079 PCPU_SET(cmci_mask, PCPU_GET(cmci_mask) | 1 << i);
1083 amd_thresholding_resume(int i)
1085 struct amd_et_state *cc;
1087 KASSERT(i < mca_banks, ("CPU %d has more MC banks", PCPU_GET(cpuid)));
1089 /* Ignore banks not monitored by this CPU. */
1090 if (!(PCPU_GET(cmci_mask) & 1 << i))
1093 cc = &amd_et_state[PCPU_GET(cpuid)][i];
1095 cc->cur_threshold = 1;
1096 amd_thresholding_start(cc, i);
1101 * Initializes per-CPU machine check registers and enables corrected
1102 * machine check interrupts.
1109 int i, skip, family;
1111 family = CPUID_TO_FAMILY(cpu_id);
1113 /* MCE is required. */
1114 if (!mca_enabled || !(cpu_feature & CPUID_MCE))
1117 if (cpu_feature & CPUID_MCA) {
1119 PCPU_SET(cmci_mask, 0);
1121 mcg_cap = rdmsr(MSR_MCG_CAP);
1122 if (mcg_cap & MCG_CAP_CTL_P)
1123 /* Enable MCA features. */
1124 wrmsr(MSR_MCG_CTL, MCG_CTL_ENABLE);
1125 if (IS_BSP() && boot)
1129 * Disable logging of level one TLB parity (L1TP) errors by
1130 * the data cache as an alternative workaround for AMD Family
1131 * 10h Erratum 383. Unlike the recommended workaround, there
1132 * is no performance penalty to this workaround. However,
1133 * L1TP errors will go unreported.
1135 if (cpu_vendor_id == CPU_VENDOR_AMD && family == 0x10 &&
1137 mask = rdmsr(MSR_MC0_CTL_MASK);
1138 if ((mask & (1UL << 5)) == 0)
1139 wrmsr(MSR_MC0_CTL_MASK, mask | (1UL << 5));
1143 * The cmci_monitor() must not be executed
1144 * simultaneously by several CPUs.
1147 mtx_lock_spin(&mca_lock);
1149 for (i = 0; i < (mcg_cap & MCG_CAP_COUNT); i++) {
1150 /* By default enable logging of all errors. */
1151 ctl = 0xffffffffffffffffUL;
1154 if (cpu_vendor_id == CPU_VENDOR_INTEL) {
1156 * For P6 models before Nehalem MC0_CTL is
1157 * always enabled and reserved.
1159 if (i == 0 && family == 0x6
1160 && CPUID_TO_MODEL(cpu_id) < 0x1a)
1162 } else if (cpu_vendor_id == CPU_VENDOR_AMD) {
1163 /* BKDG for Family 10h: unset GartTblWkEn. */
1164 if (i == MC_AMDNB_BANK && family >= 0xf)
1165 ctl &= ~(1UL << 10);
1169 wrmsr(MSR_MC_CTL(i), ctl);
1172 if (cmci_supported(mcg_cap)) {
1177 } else if (amd_thresholding_supported()) {
1179 amd_thresholding_monitor(i);
1181 amd_thresholding_resume(i);
1185 /* Clear all errors. */
1186 wrmsr(MSR_MC_STATUS(i), 0);
1189 mtx_unlock_spin(&mca_lock);
1192 if (!amd_thresholding_supported() &&
1193 PCPU_GET(cmci_mask) != 0 && boot)
1198 load_cr4(rcr4() | CR4_MCE);
1201 /* Must be executed on each CPU during boot. */
1209 /* Must be executed on each CPU during resume. */
1218 * The machine check registers for the BSP cannot be initialized until
1219 * the local APIC is initialized. This happens at SI_SUB_CPU,
1223 mca_init_bsp(void *arg __unused)
1228 SYSINIT(mca_init_bsp, SI_SUB_CPU, SI_ORDER_ANY, mca_init_bsp, NULL);
1230 /* Called when a machine check exception fires. */
1234 uint64_t mcg_status;
1235 int recoverable, count;
1237 if (!(cpu_feature & CPUID_MCA)) {
1239 * Just print the values of the old Pentium registers
1242 printf("MC Type: 0x%jx Address: 0x%jx\n",
1243 (uintmax_t)rdmsr(MSR_P5_MC_TYPE),
1244 (uintmax_t)rdmsr(MSR_P5_MC_ADDR));
1245 panic("Machine check");
1248 /* Scan the banks and check for any non-recoverable errors. */
1249 count = mca_scan(MCE, &recoverable);
1250 mcg_status = rdmsr(MSR_MCG_STATUS);
1251 if (!(mcg_status & MCG_STATUS_RIPV))
1256 * Only panic if the error was detected local to this CPU.
1257 * Some errors will assert a machine check on all CPUs, but
1258 * only certain CPUs will find a valid bank to log.
1263 panic("Unrecoverable machine check exception");
1267 wrmsr(MSR_MCG_STATUS, mcg_status & ~MCG_STATUS_MCIP);
1271 /* Called for a CMCI (correctable machine check interrupt). */
1275 struct mca_internal *mca;
1279 * Serialize MCA bank scanning to prevent collisions from
1282 count = mca_scan(CMCI, NULL);
1284 /* If we found anything, log them to the console. */
1286 mtx_lock_spin(&mca_lock);
1287 STAILQ_FOREACH(mca, &mca_records, link) {
1293 mtx_unlock_spin(&mca_lock);