Re: [PATCH v6 2/2] accel/rocket: Fix iommu_group leak and unsafe IRQ register access

[Igor Paunovic <royalnet026@gmail.com>]

Hi Zhao,

I ran into the iommu_group leak independently today - Sashiko flagged
it again on my rocket clks fix [1] - and before writing anything I
went through the lists and found your series, so let me try to help
here rather than duplicate it.

I read v6 2/2 and the review history, and I think the patch got stuck
because it bundles two separate things:

 (a) the iommu_group leak in rocket_job_handle_irq(). Your fix for
     this is exactly right: reusing the cached core->iommu_group
     matches what rocket_job_run() and rocket_reset() already do, and
     rocket_core_init() sets up the cache before the IRQ handler can
     run and only puts it in rocket_core_fini(). That part is a clean
     one-liner with Fixes: 0810d5ad88a1.

 (b) the runtime-PM guards around the IRQ register accesses - which is
     where each revision drew new review comments.

Would you consider respinning just (a) as a standalone v7? I would be
glad to test it on RK3588 (Orange Pi 5 Plus) and reply with Tested-by:
I have a benchmark here that exercises exactly that line ~150 times
per second with bit-exact output verification. (b) can then be
discussed separately without holding the simple fix hostage.

For what it is worth regarding the AI review noise: the leak is real
but bounded - it happens once per job (the intermediate-task IRQs
return before that line), and the group kobject kref saturates with a
one-time WARN rather than overflowing - and the one-line fix cannot
regress the NULL-group case either, since it passes the same value the
current code already passes. So (a) stands on its own merits.

[1] https://lore.kernel.org/linux-rockchip/20260729092939.118779-1-royalnet026@gmail.com/

Thanks,
Igor Paunovic