Re: [Bug #13371] s2disk hangs with kernel 2.6.29 and later, SATA, Gigabyte EG45M-DS2H - culprit found

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Tuesday 26 May 2009, gapeters@worldnet.att.net wrote:
> On Tue, May 26, 2009 at 01:03:00AM +0200, Richard Atterer wrote:
> > Hello,
> > 
> > "Jerry" contacted me off-list with information about the root of the 
> > problem - many thanks indeed!!!
> > 
> > It turns out that the e100 driver is responsible. (The motherboard actually 
> > has a r8169 chip, but I added an old Intel PCI ethernet card I own.)
> > 
> > By rmmoding e100 before s2disk, the problem goes away. Jerry mentions that 
> > he once saw a message about e100 failing to load its microcode - I never 
> > saw such a message.
> > 
> > The most recent e100 commit ac7c992 by Thadeu Lima de Souza Cascardo (Apr 
> > 28) made some changes to the behaviour in the case of shutdown vs. suspend, 
> > maybe the problem is there? (But note that during bisecting, I tested 
> > kernels older than that commit, and not all of them worked.) Before that, 
> > f902283 (Mar 3 2008) mentions suspend in the log message.
> > 
> > Cheers,
> > 
> >   Richard
> > 
> > -- 
> 
> I think I saw the firmware loader failure during resume. If you leave the
> system alone for ~2 minutes it will resume, but the e100 won't work &
> any thing "touching" it will go into an un-interruptable sleep
> (ifconfig, cat'ing some of its /sys files).
> 
> IIRC started with 2.6.29, with the change to net drivers to use the
> firmware loader. During the hibernate, for some reason, the e100 is
> suspended then resumed, at which point it requests its firmware.
> Userspace is frozen so the firmware load times out (the 2 minute timeout
> which is why I guessed you were having the same problem).
> 
> During resume, the e100 is apparently resumed *before* userspace is
> un-frozen, so again it times out loading its firmware. At this point
> it's unusable.

This theory makes sense.

Generally firmware loading is not going to work from the driver's ->suspend()
and ->resume() callbacks.  For this reason the firmware should be loaded
into RAM before suspend, for example from a PM notifier.

Thanks,
Rafael