RE: [linux-usb-devel] [4/4] 2.6.23-rc3: known regressions

[Linus Torvalds <torvalds@linux-foundation.org>]

On Wed, 22 Aug 2007, Stuart_Hayes@Dell.com wrote:
> 
> If you were running 2.6.26-rc3, that's quite possibly because you didn't
> have the follow-up patch that fixed my original patch... it wasn't in
> 2.6.26-rc3

Well, I was running "current git", and it's never been there. So not just 
a -rc3 issue.

> That could definitely cause mouse lock-ups.  Sorry, that should have
> occurred to me yesterday when you mentioned the problem your kids were
> seeing, but it didn't for some reason.

Btw, could it have caused the USB stack to be *really* confused? Some of 
those mouse lockups ended up also locking the machine hard (ie no ping, no 
nothing), and I'm a bit worried that there was something else going on 
too..

That said, if you can actually re-create the MMF problems, could you 
please try the patch that Arjan suggested? Ie add a

	/*
	 * Some broadcom chips are buggy and can't take more than 5 usec as DMA
	 * latency; inform the rest of kernel of this.
	 */
	if (weird_broadcom_chip())
		set_acceptable_latency("ehci", 5);

to the USB driver, and then add something like

	static inline int cpufreq_acceptable_latency(struct cpufreq_policy *policy)
	{
		unsigned long latency;

		/* Policy latency in usec */
		latency = policy->cpuinfo.transition_latency / 1000;

		if (latency > system_latency_constraint())
			return -EINVAL;

		return 0;
	}

adn then add calls to this from both the "__cpufreq_set_policy()" function 
and the "__cpufreq_driver_target()" one too..

That should disable cpufreq with that broken chip, which is perhaps a big 
draconian, but it's certainly better than having the USB layer know about 
cpufreq internals directly.

In the longer run, I think we can move the "system_latency_constraint()" 
checking from the policy registration into each CPU frequency driver, so 
that it could be more dynamically decide about "can we do it right _now_" 
rather than globally saying "we can't do it with this hardware".

				Linus