Re: [PATCH v1 0/4] Expose blk-mq tag sets through debugfs

[Lei Chen]

Hi Jens,

A gentle ping on this series.

It adds a tag-set-level debugfs view to inspect CPU mappings and
driver tags, with links from request queues to their shared tag set.

I'd appreciate feedback on the overall approach, particularly the
driver tag lifetime changes in patch 2 that support the debugfs view.

Thanks,
Lei

On Fri, Oct 9, 2026 at 11:01 AM Lei Chen <lei.chen@xxxxxxxxxx> wrote:
>
> Block debugfs provides views of request queues and hardware contexts, but
> not the tag set they share. This series adds a tag set view for inspecting
> configuration, CPU mappings and driver tags, with links from request queue
> directories to identify queues using the same tag set.
>
> The series first removes a stale CPU mapping comment and moves driver tag
> allocation and reclamation from individual request queue initialization to
> the tag set level. This gives the debugfs tag files a lifetime that can be
> coordinated with tag allocation and release. It then adds tagset/<id>
> directories under the debugfs root and links each request queue to its tag
> set directory.
>
> Tag reads hold a debugfs active reference only while copying tag fields to
> local variables. Releasing it before copying to userspace lets tag file
> removal complete even if a reader faults on its user buffer and needs I/O
> to a frozen queue.
>
> Based on Linux v7.3-rc2.
>
> Testing on a 192-CPU x86_64 host with KASAN and lockdep enabled included:
>
>   - null_blk queue growth and shrink under concurrent direct I/O and
>     debugfs reads, including 100 cycles between 1 and 192 queues.
>   - Shared tag sets, shared tag bitmaps, three-map configurations, and
>     repeated device teardown and recreation with old file descriptors.
>   - Unmapped tag reclamation with 256 poll queues: tracing confirmed
>     allocation, reclamation, reallocation and reclamation again for
>     indices that became unmapped as the queue count changed.
>   - Task-local allocation fault injection, including hctx allocation
>     failures that rolled the hardware queue count back.
>   - NVMe loop with two namespaces, 50 controller resets and data-checked
>     direct I/O. These resets retained the same hardware queue count.
>   - userfaultfd-controlled stalls in tag reads: null_blk queue resizing
>     and NVMe loop controller deletion completed while copying to the
>     user buffer was blocked.
>
> No KASAN or lockdep reports were observed in these tests. A hardware queue
> count change with NVMe queues already frozen has not been tested.
>
> Lei Chen (4):
>   blk-mq: remove stale comment about non-present CPU mapping
>   blk-mq: manage driver tags at the tag set level
>   block: expose blk-mq tag sets through debugfs
>   block: link request queue debugfs directories to tag sets
>
>  block/blk-mq-debugfs.c | 298 +++++++++++++++++++++++++++++++++++++++++
>  block/blk-mq-debugfs.h |  22 +++
>  block/blk-mq.c         | 143 +++++++++++++++-----
>  include/linux/blk-mq.h |  13 ++
>  4 files changed, 445 insertions(+), 31 deletions(-)
>
> --
> 2.43.0
>