Re: [RFC PATCH v7 6/9] media: tegra: Add Tegra210 Video input driver

From: Sowjanya Komatineni
Date: Wed Apr 15 2020 - 19:08:29 EST


With minor change of not using vi reference after host1x_client_unregister and freeing vi during v4l2 device release works.

For csi, we can use devm_kzalloc for now untill we decide later if we want to expose async subdev nodes during sensor support.

Will have this fix in v8 with a comment in vi_remove to make sure not to use vi reference after host1x_client_unregister.

Will test more and will release v8 with above fix to allow direct host1x client driver unbind.

Thanks

sowjanya


On 4/15/20 12:51 PM, Sowjanya Komatineni wrote:

On 4/15/20 12:21 PM, Dmitry Osipenko wrote:
External email: Use caution opening links or attachments


15.04.2020 21:53, Sowjanya Komatineni ÐÐÑÐÑ:
...
Have you tried to test this driver under KASAN? I suspect that
you just
masked the problem, instead of fixing it.
Tested with kmemleak scan and did not see any memory leaks
You should get use-after-free and not memleak.
I don't see use-after-free bugs during the testing.

But as mentioned when direct vi/csi client driver unbind happens while video device node is kept opened, vi driver remove will free vi structure memory but actual video device memory which is part of channels remains but list head gets lost when vi structure is freed.

So, when device node is released and executes release callback as list head is lost it can't free allocated channels which is not good.

This happens only with direct host1x client vi/csi driver unbind.

Need to find better place to free host1x client driver data structure to allow direct client driver unbind->bind.