At Thu, 9 Aug 2012 14:49:04 +0200, Thierry Reding wrote:
On Thu, Aug 09, 2012 at 02:32:38PM +0200, Takashi Iwai wrote:
At Thu, 9 Aug 2012 12:34:30 +0200, Thierry Reding wrote:
On Thu, Aug 09, 2012 at 10:21:15AM +0200, Takashi Iwai wrote:
At Thu, 9 Aug 2012 10:07:13 +0200, Thierry Reding wrote:
On Thu, Aug 09, 2012 at 09:42:48AM +0200, Takashi Iwai wrote:
At Thu, 9 Aug 2012 09:36:42 +0200, Thierry Reding wrote: > > On Thu, Aug 09, 2012 at 09:31:30AM +0200, Takashi Iwai wrote: > > At Thu, 9 Aug 2012 09:08:13 +0200, > > Thierry Reding wrote: > > > > > > On Thu, Aug 09, 2012 at 08:57:13AM +0200, Takashi Iwai wrote: > > > > At Thu, 9 Aug 2012 08:45:23 +0200, > > > > Thierry Reding wrote: > > > > > > > > > > Recent changes to the firmware loading helpers cause drivers to stall > > > > > when firmware is loaded during the module_init() call. The snd-hda-intel > > > > > module requests firmware if the patch= parameter is used to load a patch > > > > > file. This patch works around the problem by deferring the probe in such > > > > > cases, which will cause the module to load successfully and the driver > > > > > binding to the device outside the module_init() call. > > > > > > > > Is the "recent" change meant 3.6 kernel, or in linux-next? > > > > > > > > In anyway, I don't understand why such a change was allowed. Most > > > > drivers do call request_firmware() at the device probing time. > > > > If this really has to be resolved in the driver side, it must be a bug > > > > in the firmware loader core code. > > > > > > A good explanation of the problem and subsequent discussion can be found > > > here: > > > > > > http://article.gmane.org/gmane.linux.drivers.video-input-infrastructure/4997... > > > > Yeah, but it doesn't justify this ugly module option. > > It's a simple bug. Papering over it with this option doesn't fix > > anything. > > It's not an option, all it does is defer probing if and only if the > patch parameter was specified to make sure the firmware load won't > stall. I realize that this may not be an optimal solution, but at least > it fixes the problem with no fallout.
Ah sorry, I misread the patch.
Then it shouldn't be checked at that point. Since 3.5 kernel, the probing code was already split for vga_switcheroo support.
Yes, I saw that. But unless you actually use vga_switcheroo, the second stage, azx_probe_continue(), will still be called from azx_probe() and therefore ultimately from module_init().
Yeah, but this could be easily delayed. The split was already done, so the next step would be to return after the first half at probe, then call the second half later.
Before coming up with this patch I actually did play around a bit with using the asynchronous firmware load functions but it turned out to be rather difficult to do so I opted for the easy way. The biggest problem I faced was that since patch loading needs to be done very early on, a lot of the initialization would need to be done after .probe() and many things could still fail, so cleaning up after errors would become increasingly difficult.
async probe is also on my TODO list, but it's deferred ;)
The point you added is the second stage.
I don't understand this sentence.
I meant that your patch added the check at the second-half probing function (azx_probe_contine()). That is, it could be already the point triggered by vga_switcheroo handler, not via module_init any longer.
So, after rethinking what you suggested, I wrote a quick patch below. Could you check whether this works?
It oopses, though I can't quite tell where. I need to test some more later to see where it goes wrong.
Yeah, I tested it here and noticed, too. As mentioned, the behavior of -EPROBE_DEFER is somehow flaky. For example, when it's used for modules, the deferred probe will be never triggered (unless a new device is bound).
Yes, the idea is that probing is only retried after other drivers have bound to other devices because otherwise nothing about any missing resources can have changed. This however would indicate that deferred probing is not the right solution here, after all we're not waiting for another resource to become available. Something like delayed work might be better suited.
Well, asynchronous firmware load is actually the right solution, delayed work just might be a better workaround. =)
For completeness I should say that I've been using deferred probing with modules quite successfully on another platform, so it is not a general problem. Rather as you said, it is only triggered if another module is loaded after the deferral.
Yes. But if you reload just snd-hda-intel module, it'll be stuck.
Considering the problem again, it's currently an issue only for the built-in sound driver, right? AFAIK, request_firmware() works fine for modules. If so, a simple "fix" to avoid the unexpected behavior is to make CONFIG_SND_HDA_PATCH_LOADER depending on CONFIG_SND_HDA=m. It'd be simple enough for merging 3.6 kernel.
I've actually seen this problem with snd-hda-intel built as a module as well, so I don't think this kind of temporary fix will do.
Hmm, for modules, request_firmware() seems working fine on my systems, no matter whether it succeeds or not. Could you double-check? If the user-helper isn't ready at module loading time, of course, it stalls, just like the kernel driver. But it should be available there.
I almost finished writing the patch to use request_firmware_nowait() version, but I'm afraid it's too intrusive for 3.6. If disabling the patch loader for the built-in driver is OK, I'd queue the request_firmware_nowait() patches to 3.7 queue.
I'm in no hurry, and the patch that I carry works for me. If there are no other reasons to pull a corresponding fix into 3.6 I can certainly wait for 3.7. The particular setup that I have requires other patches that may not go into 3.6 anyway.
OK.
Takashi