Le 26/08/2025 à 14:41, Jason Gunthorpe a écrit :

On Tue, Aug 26, 2025 at 11:52:37AM +0200, Benjamin Gaignard wrote:

quoted

Le 25/08/2025 à 20:31, Jason Gunthorpe a écrit :

quoted

On Mon, Aug 25, 2025 at 01:50:16PM -0400, Nicolas Dufresne wrote:

quoted

Jason, the point is that the iommu and the VPU are not separate devices, which
comes with side effects. On RKVDec side, the iommu configuration get resets
whenever a decoding error leads to a VPU "self reset". I can't remember who from
the iommu subsystem suggested that, but the empty domain method was agreed to be

IDK, that seems really goofy too me an defiantly needs to be
extensively documented this is restoring the default with some lore
link of the original suggestion.

quoted

the least invasive way to workaround that issue. I believe Detlev tried multiple
time to add APIs for that before the discussion lead to this path.

You mean this:

https://lore.kernel.org/linux-iommu/20250318152049.14781-1-detlev.casanova@collabora.com/ (local)

Which came back with the same remark I would give:

   Please have some kind of proper reset notifier mechanism - in fact
   with runtime PM could you not already invoke a suspend/resume cycle
   via the device links?

when doing parallel decode suspend/resume are not invoked.

It was a proposal for an error recovery path.

quoted

quoted

Or another reasonable option:

    Or at worst just export a public interface for the other driver to
    invoke rk_iommu_resume() directly.

Sigh.

An other solution which is working is to call iommu_flush_iotlb_all()
before decoding each frame.

That was already proposed and shot down, it makes no sense at all use
to use flushing to reset the registers because the HW weirdly lost
them, and flushing should never happen outside mapping contexts.

If the HW is really resetting the iommu registers after every frame
that is really just painfully broken, and makes me wonder if it really
should be an iommu subsystem driver at all if it is so tightly coupled
to the computing HW. :\

Like we wouldn't try to put a GPU MMU into the iommu subsystem though
they do fairly similar things.

I will add vsi_iommu_restore_ctx() function where all locks will be
managed and call vsi_iommu_enable(). Decoder driver will call this function
before each frame.
I think it similar to some omap iommu code I have found.

Regards,
Benjamin

Jason