.\"
.\" Copyright (c) 2026 The FreeBSD Project
.\"
.\" SPDX-License-Identifier: BSD-2-Clause
.\"
.Dd January 14, 2026
.Dt MCA 4 amd64
.Os
.Sh NAME
.Nm mca
.Nd Machine Check Architecture
.Sh DESCRIPTION
The
.Nm
subsystem provides support for the x86 Machine Check Architecture.
The CPU uses this architecture to report various hardware problems,
ranging from correctible errors to uncorrectible fatal errors.
The
.Nm
subsystem processes the errors reported by the CPU and logs them
according to the user-supplied configuration.
.Pp
The
.Nm
subsystem is automatically compiled into every x86 kernel.
.Sh LOGGING
By default, every message is logged to four locations:
.Bl -bullet
.It
The console
.It
The system log (using the
.Dv LOG_KERN
.Xr syslog 3
facility)
.It
An in-kernel cache of event records
.It
A statistics array
.El
.Pp
An administrator can disable console logging of non-fatal errors using the
.Va hw.mca.uselog
.Xr sysctl 8
or tunable setting.
Fatal errors are always logged to the console.
.Pp
The in-kernel cache of event records can be accessed from userspace using the
.Va hw.mca.records
.Xr sysctl 8
node.
By default, the in-kernel cache of event records grows without bound and
contains records for all
.Nm
events received since system boot.
The cache can be limited (in which case it turns into a ring buffer) or
disabled altogether using the
.Va hw.mca.maxcount
.Xr sysctl 8
or tunable setting.
.Pp
The statistics array can be retrieved using the
.Va hw.mca.stats
.Xr sysctl 8
node.
.Sh SYSCTL VARIABLES
The subsystem has a number of configuration options to control the
way events are processed and logged.
These options are configured via
.Xr sysctl 8
variables or tunables.
These settings control things such as log destination, event limiting,
and sampling.
These settings directly impact both the completeness of logs and the performance
impact of processing
.Nm
events.
A system administrator may want to review these and modify them to obtain
the appropriate behavior in their environment.
.Bl -tag -width indent
.It Va hw.mca.enabled
(Only settable as a tunable.)
.Pp
If set to 0, the CPU is not configured to send
.Nm
messages to the kernel.
.Pp
Default is 1 (enabled).
.It Va hw.mca.log_corrected
(Settable as a tunable or sysctl.)
.Pp
If enabled, corrected messages are logged to console, syslog, the in-kernel
event cache, and the statistics array (subject to separate configuration
controlling those facilities).
If disabled, corrected messages are only logged to the in-kernel event
cache (subject to separate configuration controlling that facility).
.Pp
Default is 1 (enabled).
.It Va hw.mca.intel6h_HSD131
(Only settable as a tunable.)
.Pp
This setting enables a workaround for benign corrected parity errors which
may be reported by certain Intel desktop Haswell CPUs.
(The name "HSD131" comes from the name of the Intel erratum report about this
issue.)
.Pp
Default is 0 (disabled).
.It Va hw.mca.amd10h_L1TP
(Only settable as a tunable.)
.Pp
Enable logging of level one TLB parity errors on certain AMD CPUs.
This option has no impact on other CPUs.
.Pp
Default is 1 (enabled).
.It Va hw.mca.erratum383
(Only settable as a tunable.)
.Pp
This setting enables a workaround for Erratum 383 on AMD Family 10h CPUs.
The erratum changes the way pages are promoted to or demoted from being
super-pages.
On affected AMD CPUs, either
.Va hw.mca.amd10h_L1TP
or
.Va hw.mca.erratum383
must be on.
If both are off, the system will dynamically enable
.Va hw.mca.erratum383
at boot time.
.Pp
Default is 0 (disabled).
.It Va hw.mca.uselog
(Settable as a tunable or sysctl.)
.Pp
If enabled, the system will send messages about non-fatal
.Nm
events to syslog and not to the console.
If disabled, the system will send messages about non-fatal
.Nm
events to both syslog and the console.
Fatal events are always logged to the console.
.Pp
Default is false (disabled).
.It Va hw.mca.stats
(Read-only sysctl.)
.Pp
This returns an array of
.Va MCA_T_COUNT
uint64_t values.
The
.Vt "enum mca_stat_types"
definition in
.In x86/mca.h
provides the value of
.Va MCA_T_COUNT
and the index values for various event types.
.It Va hw.mca.log_interval
(Settable as a tunable or sysctl.)
.Pp
This sets the minimum time (in seconds) between logging correctible errors.
The rate limit is only applied after the system records a reasonable
number of errors of the same type.
The goal is to reduce the impact of the system seeing and attempting to log
a burst of similar errors, which can be expensive (especially when printed
to the console).
If this setting is 0, no rate limit is applied.
.Pp
Default is 0 (no rate limit).
.It Va hw.mca.cmc_throttle
(Settable as a tunable or sysctl. Only available if
.Xr apic 4
support is enabled and the hardware supports CMC interrupt throttling.)
.Pp
This sets the maximum time (in seconds) to throttle CMC interrupts.
In normal operation, the system attempts to receive CMC interrupts as soon as
an event occurs.
However, if a high rate of events occurs in a short time, the system will
begin throttling the CMC interrupts.
While the events continue to occur at a high rate, the system will gradually
increase the throttling interval until it reaches the
.Va hw.mca.cmc_throttle
setting.
.Pp
Default is 60 seconds.
.It Va hw.mca.count
(Read-only sysctl.)
.Pp
This returns the current number of
.Nm
records in the in-kernel cache.
This can be used to determine how many records are available to read with the
.Va hw.mca.records
.Xr sysctl 8
interface.
It can also be used to monitor the amount of memory used by the in-kernel
record cache.
.It Va hw.mca.maxcount
(Settable as a tunable or sysctl.)
.Pp
This setting controls the size and behavior of the in-kernel cache of
.Nm
records.
If the setting is -1, the in-kernel cache grows without bounds and contains a
complete record events since boot or the last time the
.Va hw.mca.maxcount
setting was changed.
If the setting is 0, the in-kernel cache is disabled.
If the setting is a positive integer, the in-kernel cache functions as a ring
buffer with the number of entries defined by the setting.
.Pp
Default is -1 (unlimited).
.It Va hw.mca.interval
(Settable as a tunable or sysctl.)
.Pp
This setting controls how often (in seconds) the kernel should proactively
scan for new
.Nm
events.
In many circumstances, the CPU will send an interrupt to signal new events.
However, there are cases where the periodic scan for events will discover
new events.
.Pp
Default is 300 seconds.
.It Va hw.mca.force_scan
(Settable only as a sysctl.)
.Pp
Setting this to any non-zero value will force an immediate scan for
.Nm
events.
Setting this to zero has no effect.
This is functionally a write-only setting.
The current value is always 0, even when a scan is running.
.Pp
Default is 0 (no scan requested).
.It Va hw.mca.records.n
(Read-only sysctl.)
.Pp
This is used to copy
.Nm
records from the in-kernel cache to a user space process.
The
.Va n
value in the
.Xr sysctl 8
node is an integer index into the in-kernel cache.
(Or, put differently, the "new" value describes how many records the kernel
should skip to find the desired record.)
The return value is a
.Vt "struct mca_record" ,
which is defined in
.In x86/mca.h .
.El
.Sh LOADER TUNABLES
Tunables can be set at the
.Xr loader 8
prompt before booting the kernel or stored in
.Pa /boot/loader.conf .
The tunables all have corresponding
.Xr sysctl 8
entries.
The tunables are listed above in the
.Sx SYSCTL VARIABLES
section.
.Sh COMPATIBILITY
.Nm
is only available on x86 systems.
.Sh SEE ALSO
.Xr loader.conf 5 ,
.Xr sysctl 8 ,
.Xr syslogd 8
.Sh AUTHORS
The
.Nm
subsystem was originally written by
.An John Baldwin Aq Mt jhb@FreeBSD.org .
.Pp
This manual page was written by
.An Jonathan Looney Aq Mt jtl@FreeBSD.org .