[PATCHv4 0/8] nvme: export additional diagnostic counters via sysfs

[Nilay Shroff]

Hi Keith,

On 5/17/26 12:06 AM, Nilay Shroff wrote:
> Hi,
> 
> The NVMe driver encounters various events and conditions during normal
> operation that are either not tracked today or not exposed to userspace
> via sysfs. Lack of visibility into these events can make it difficult to
> diagnose subtle issues related to controller behavior, multipath
> stability, and I/O reliability.
> 
> This patchset adds several diagnostic counters that provide improved
> observability into NVMe behavior. These counters are intended to help
> users understand events such as transient path unavailability,
> controller retries/reconnect/reset, failovers, and I/O failures. They
> can also be consumed by monitoring tools such as nvme-top.
> 
> Specifically, this series proposes to export the following counters via
> sysfs:
>    - Command retry count
>    - Multipath failover count
>    - Command error count
>    - I/O requeue count
>    - I/O failure count
>    - Controller reset event counts
>    - Controller reconnect counts
> 
> The first patch in the series adds a new diag attribute group under per-path,
> ns-head and ctrl sysfs directories so that all diagnostics counters could be
> grouped together under diag sub-directory. The subsequent patches in the series
> adds diagnostics counters listed above.
> 
> Please note that this patchset doesn't make any functional change but
> rather export relevant counters to user space via sysfs.
> 
> As usual, feedback/comments/suggestions are welcome!
> 
> Changes from v3:
>    - To be consistent in naming, all counters are suffixed with _count
>      (Keith Busch)
>    - The first patch in the series creates new attribute group named
>      diag and all counters are now grouped under this new sysfs
>      attribute group (Keith Busch)
>    - Counters are defined as atomic_long_t instead of size_t (Keith Busch)
>    - Removed RB and TB tags due to above changes
> Link to v3: https://lore.kernel.org/all/20260220175024.292898-1-nilay@linux.ibm.com/

As discussed during LSFMM, I have incorporated all of your feedback listed in the
changelog above in current patchset, except for integrating command_error_count into
the generic gendisk statistics.

While working through the implementation, I realized that although some namespace/
path-level error statistics could conceptually fit within generic block-layer accounting,
controller-wide error statistics do not naturally map onto a gendisk object. This is
because NVMe controllers themselves do not have an associated block device representation.

As a result, controller-scoped telemetry cannot be cleanly integrated into generic gendisk
statistics. Therefore, this patchset exports these counters through NVMe-specific sysfs
attributes instead. Let me know if this approach looks reasonable.

Thanks,
--Nilay