Re: A workaround for request_firmware() stuck in module_init

[Takashi Iwai <tiwai@suse.de>]

At Tue, 4 Sep 2012 23:52:15 +0800,
Ming Lei wrote:
> 
> On Tue, Sep 4, 2012 at 9:06 PM, Takashi Iwai <tiwai@suse.de> wrote:
> > Hi,
> >
> > as I've got recently a few bug reports regarding the stuck with
> > request_firmware() in module_init of some sound drivers, I started
> > looking at the issue.  Strangely, the problem doesn't happen on
> > openSUSE 12.2 although it has the same udev version with libkmod as
> > Fedora.  So I installed Fedora 17, and indeed I could see a problem
> > there.
> 
> It should be a bug in udev, as discussed in the below link:
> 
>           http://marc.info/?t=134552745100002&r=1&w=2

Yeah, if udev has a fix, I'm fine.  I'll proactively ignore these bug
reports.  But did it really happen...?

> > Obviously a solution would be to rewrite the driver code to use
> > request_firmware_nowait() instead.  But it'd need a lot of code
> > shuffling, and most of such drivers are old stuff I don't want to do a
> > serious surgery.
> >
> > So I tried an easier workaround by using the deferred probing.
> > An experimental patch is below.  As you can see, from the driver side,
> > it's simple: just add two lines at the head of each probe function.
> 
> In fact, the defer probe should only be applied in situations which
> driver is built in kernel and request_firmware is called in .probe().

Is it?  I thought the deferred probe is basically not for this problem
but rather for the dependency problem with other modules.  That's the
reason it's triggered only upon the successful binding.

And IMO, the deferred probe for the built-in kernel is also silly.
Did anyone really make it working for built-in kernel driver and
external firmware files?  (For the resume, it's of course a different
issue.  And I guess it's been solved by your fw cache patch, right?)

> > Do you think this kind of hack is OK?  If not, any better (IOW easier)
> > solution?
> 
> Looks your solution is a bit complicated, and I have wrote a similar
> patch to address the problem, but Linus thought that it may hide the
> problem of drivers.
> 
>         http://marc.info/?t=134278790800004&r=1&w=2
> 
> IMO, driver core needn't to be changed, and the defer probe can be
> supported just by changes in request_firmware() and its caller.

Ideally, yes.  But it won't work in some drivers like sound drivers,
that have its own device number counting changed at each probe call.
For such drivers, the deferred probe check must be done before
entering the normal probe procedure, and returning -EPROBE_DEFER would
be too late (or need more complex fallbacks).


thanks,

Takashi