Re: [PATCH v3 00/14] staging: media: zoran: fusion in one module

[LABBE Corentin]

Le Wed, Nov 03, 2021 at 05:29:46PM +0100, Hans Verkuil a écrit :
> On 03/11/2021 16:57, LABBE Corentin wrote:
> > Le Wed, Nov 03, 2021 at 04:21:02PM +0100, Hans Verkuil a écrit :
> >> Hi Corentin,
> >>
> >> On 26/10/2021 21:34, Corentin Labbe wrote:
> >>> Hello
> >>>
> >>> The main change of this serie is to fusion all zoran related modules in
> >>> one.
> >>> This fixes the load order problem when everything is built-in.
> >>
> >> I've been testing this series, and while the module load/unload is now working,
> >> I'm running into a lot of other v4l2 compliance issues.
> >>
> >> I've fixed various issues in some follow-up patches available in my tree:
> >>
> >> https://git.linuxtv.org/hverkuil/media_tree.git/log/?h=zoran
> >>
> >> At least some of the worst offenders are now resolved. Note that the patch
> >> dropping read/write support relies on this patch:
> >>
> >> https://patchwork.linuxtv.org/project/linux-media/patch/4f89b139-13b7-eee6-9662-996626b778b0@xxxxxxxxx/
> > 
> > Hello
> > 
> > My test branch already included your "zoran: fix various V4L2 compliance errors"
> > I have quickly checked other patch and I am ok with them.
> > I will add and test with them.
> > 
> >>
> >> But there is one really major bug that makes me hesitant to merge this:
> >>
> >> This works:
> >>
> >> v4l2-ctl -v pixelformat=MJPG,width=768,height=576
> >> v4l2-ctl --stream-mmap
> >>
> >> This fails:
> >>
> >> v4l2-ctl -v pixelformat=MJPG,width=768,height=288
> >> v4l2-ctl --stream-mmap
> >>
> >> It's an immediate lock up with nothing to indicate what is wrong.
> >> As soon as the height is 288 or less, this happens.
> >>
> >> Both with my DC30 and DC30D.
> > 
> > Just for curiosity, what is the difference between thoses two ?
> 
> It's the DC30 variant without an adv7175.
> 
> > 
> >>
> >> Do you see the same? Any idea what is going on? I would feel much happier
> >> if this is fixed.
> >>
> >> Note that the same problem is present without this patch series, so it's
> >> been there for some time.
> >>
> > 
> > I will start on digging this problem and add thoses commands to my CI.
> > And I know there are a huge quantity of problem since origins.
> > A simple example is that just setting MJPEG as default input format does not work.
> > 
> > But since it is not related to my serie, can you please merge it.
> 
> Before I do that, I would really like to know a bit more about this issue:
> can you reproduce it? Is it DC30 specific or a general problem with zoran?
> 
> The problem with this hard hang is that it is hard to do regression testing
> with v4l2-compliance, since it will hang as soon as MJPG pixelformat is
> tested.
> 
> I would feel much happier if the hang can be avoided, even if it is just
> with a temporary hack. It will make it much easier going forward.
> 

I found the bug

The null pointer deref was in zoran_reap_stat_com() due to 
buf = zr->inuse[i];
...
buf->vbuf.vb2_buf.timestamp = ktime_get_ns();
with buf = NULL;

It is due to miscalculation of "i".

I will resend my serie with the fix for that.

Regards