On Thu, 29 Jan 2026 05:34:10 -0800 Breno Leitao <[email protected]> wrote:
> Add a sysfs file at /sys/kernel/vmcore_stats and expose hardware error
> recovery statistics that are already tracked by the kernel. This allows
> userspace monitoring tools to track recovered hardware errors without
> requiring kernel crashes.
>
> This is useful to track recoverable hardware errors in a time series,
> even if the host doesn't crash.
>
> Create a generic vmcore_stats sysfs, and add a section for
> hwerr_recovery that shows the counts per subsystem and timestamps:
>
> - cpu: CPU-related errors (MCE, ARM processor errors)
> - memory: Memory-related errors
> - pci: PCI/PCIe AER non-fatal errors
> - cxl: CXL errors
> - other: Other hardware errors
>
> Example output:
> hwerr_recovery:
> cpu: 0 (0)
> memory: 2 (1738148257)
> pci: 1 (1738147000)
> cxl: 0 (0)
> other: 0 (0)
sysfs rules (which are widely ignored) say "one value per file".
As a compromise the above could be squished into a single line. Harder
for humans to read, but it sounds like that isn't the expected use case.
Or screw sysfs rules ;)
> The value in parentheses is the timestamp (seconds since epoch) of the
> last error of that type, or 0 if no errors have occurred.
>
> These statistics provide visibility into the health of the system's
> hardware and can be used by system administrators to proactively detect
> failing components before they cause system crashes.
>
> ...
>
> +/* sysfs interface for hardware error recovery statistics */
> +static ssize_t vmcore_stats_show(struct kobject *kobj,
> + struct kobj_attribute *attr, char *buf)
> +{
> + return sysfs_emit(buf,
> + "Recovered hardware errors:\n"
> + " cpu: %d (%lld)\n"
> + " memory: %d (%lld)\n"
> + " pci: %d (%lld)\n"
> + " cxl: %d (%lld)\n"
> + " other: %d (%lld)\n",
> + atomic_read(&hwerr_data[HWERR_RECOV_CPU].count),
> + (long
> long)READ_ONCE(hwerr_data[HWERR_RECOV_CPU].timestamp),
> + atomic_read(&hwerr_data[HWERR_RECOV_MEMORY].count),
> + (long
> long)READ_ONCE(hwerr_data[HWERR_RECOV_MEMORY].timestamp),
vsprintf has `%ptT' for time64_t. Is it usable here?
> + atomic_read(&hwerr_data[HWERR_RECOV_PCI].count),
> + (long
> long)READ_ONCE(hwerr_data[HWERR_RECOV_PCI].timestamp),
> + atomic_read(&hwerr_data[HWERR_RECOV_CXL].count),
> + (long
> long)READ_ONCE(hwerr_data[HWERR_RECOV_CXL].timestamp),
> + atomic_read(&hwerr_data[HWERR_RECOV_OTHERS].count),
> + (long
> long)READ_ONCE(hwerr_data[HWERR_RECOV_OTHERS].timestamp));
> +}
> +
>
> ...
>
