[PATCH v2 00/10] irqchip/gic-v5: Fix misc IRS and ITS host driver issues

Fuad Tabba fuad.tabba at linux.dev
Mon Nov 16 05:28:56 PST 2026


Hi Lorenzo,

On Mon, 16 Nov 2026 12:14:50 +0000, Lorenzo Pieralisi
<lpieralisi at kernel.org> wrote:
[...]
> > invalidate have the same shape. The question is, can an IRS miss the
> > poll, or is a timeout here only ever a dead IRS?
>
> Thank you for your patches Fuad.
>
> The latter, though freeing memory on timeout wasn't a good idea given
> that there is no way to cancel the respective, pending, mapping.

Thanks, that answers it. If you'd like a patch that doesn't free the
memory on a timeout, let me know and I'll send one.

[...]
> > Notes:
> >     Setting L1 on every invalidation turns each registration and
> >     teardown into a range invalidate, both rare paths, and keeps the
> >     function simple. Setting it only when registration allocates the
> >     level 2 table would miss a registration that fails after the
> >     allocation, which leaves the level 1 DTE valid.
>
> I can't follow. I agree that we have to invalidate the L1 DTE on
> L2 allocation but that's a one-off, I don't understand what issues
> you are pointing out in the Notes above, if you can point out please
> the code paths it refers to that'd help.

gicv5_its_device_register() can fail after
gicv5_its_devtab_get_dte_ref() has made the L1 DTE valid, and before
the gicv5_its_device_cache_inv() at the end. That happens if the
EventID bits check fails, or the ITT allocation does. The L1 DTE is
then already valid for the next device registered under it, so that
registration doesn't allocate a level 2 table.

The patch Notes were about setting ITS_INV_DEVICER.L1 in that
gicv5_its_device_cache_inv() call only when the registration allocated
the table. In that case the failed registration never reaches that
call and the next one reaches it with L1 clear, so this L1 DTE is
never invalidated. Invalidating in gicv5_its_alloc_l2_devtab(), right
after the L1 DTE write, is the one-off you describe. I'll rework the
patch to do that instead.

Cheers,
/fuad



More information about the linux-arm-kernel mailing list