Hi Tomas!

Am 20.10.2012 um 12:06 schrieb Tomas Pospisek <t...@sourcepole.ch>:

> can you see a kernel stacktrace anywhere? Are you in a GUI when the system 
> freezes, i.e. running some desktop? In that case you probably won't be able 
> to see the stacktrace, but you could try to CTRL-ALT-Fx to some of your 
> consoles and maybe be lucky. I have not investigated this any further but you 
> can find a lot of ressources on the internet that try to explain how to get a 
> kerneldump when your system freezes. Maybe one could proceed from there with 
> the help of kernel people.
> 
> One guess what causes this bug is, that the system runs into memory pressure 
> - i.e. something similar like [1] - that is the system runs into a situation 
> where it's lacking memory, it wants to swap stuff out, in order to swap out 
> it needs to allocate memory, or swap something back in - like f.ex. noflushd 
> -> deadlock.

In the spindown mode used by noflushd, spinup should be handled by the 
disk/controller itself. The kernel should only notice the same delay as during 
any other disk spinup.

> I feel that the noflushd package has its value. Getting further with this bug 
> might prevent noflushd being dropped from the next release.


The problem at hand is most likely caused by a bug in the kernel, maybe even 
one that's only present in squeeze, not in the wheezy kernel. And while 
noflushd might still have its merits, it's a fact that it does weird things to 
some kernel components and therefore keeps triggering corner-case bugs that 
turn out to be extremely hard to debug. Frankly, I'm not sure whether expecting 
kernel developers to spend significant time hunting down noflushd fallout is 
fair, especially as other interfaces like laptop-mode seem to cover most 
people's needs, and noflushd has really grown into a niche application. Feel 
free to convince me otherwise, but at the moment, I'm pretty much in favour of 
dropping noflushd from Debian altogether.

Regards,

Daniel.


--
To UNSUBSCRIBE, email to debian-bugs-dist-requ...@lists.debian.org
with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org

Reply via email to