Hi Dmitry,
On Sat, Feb 20, 2021 at 08:16:22AM +0300, Dmitry Osipenko wrote:
> 19.02.2021 01:07, Nicolin Chen пишет:
> > Commit 25938c73cd79 ("iommu/tegra-smmu: Rework tegra_smmu_probe_device()")
> > removed certain hack in the tegra_smmu_probe() by relying on IOMMU core to
> > of_xlate SMMU's SID per device, so as to get rid of tegra_smmu_find() and
> > tegra_smmu_configure() that are typically done in the IOMMU core also.
> >
> > This approach works for both existing devices that have DT nodes and other
> > devices (like PCI device) that don't exist in DT, on Tegra210 and Tegra3
> > upon testing. However, Page Fault errors are reported on tegra124-Nyan:
> >
> > tegra-mc 70019000.memory-controller: display0a: read @0xfe056b40:
> > EMEM address decode error (SMMU translation error [--S])
> > tegra-mc 70019000.memory-controller: display0a: read @0xfe056b40:
> > Page fault (SMMU translation error [--S])
> >
> > After debugging, I found that the mentioned commit changed some function
> > callback sequence of tegra-smmu's, resulting in enabling SMMU for display
> > client before display driver gets initialized. I couldn't reproduce exact
> > same issue on Tegra210 as Tegra124 (arm-32) differs at arch-level code.
>
> Hello Nicolin,
>
> Could you please explain in a more details what exactly makes the
> difference for the callback sequence?
Here is a log with 5.11.0-rc6:
https://lava.collabora.co.uk/scheduler/job/3187849
[dump_stack was added in some tegra-smmu functions]
And here is a corresponding log with reverting the original commit:
https://lava.collabora.co.uk/scheduler/job/3187851
Here is a log with 5.11.0-rc7-next-20210210:
https://lava.collabora.co.uk/scheduler/job/3210245
And here is a corresponding log with reverting the original commit:
https://lava.collabora.co.uk/scheduler/job/3210596
Both failing logs show that mc errors started right after client DC
got enabled by ->attach_dev() callback that in the passing logs was
not called until Host1x driver init. And note that two failing logs
show that ->attach_dev() could be called from two different sources,
of_dma_configure_id() or arch_setup_dma_ops().
The reason why ->attach_dev() gets called is probably related to the
following reasons (sorry, can't be 100% sure as I don't have Tegra124
or other 32bit Tegra board to test):
1) With the commit reverted, all clients are probed in "arch" stage,
which is even prior to iommu core initialization -- including it
setting default domain type. This probably messed up the type of
allocating domains against the default domain type. Also internal
group is somehow affected. So some condition check in iommu core
failed and then it bypassed ->attach_dev callback in really_probe
stage, until Host1x driver does attach_dev again.
2) 32bit ARM has arch_setup_dma_ops() does an additional set of iommu
domain allocation + attach_dev(), after of_dma_configure_id() did
once. This isn't reproducible for me on Tegra210.
As debugging online isn't very efficient, and given that Thierry has
been working on the linear mapping of framebuffer carveout, I choose
to partially revert as a quick fix.
Thanks
