[rescue] Happy New Year! RIP, Sun/Solaris...

[Joshua Boyd]

On Wed, Jan 05, 2011 at 03:02:17PM +0000, Peter Corlett wrote:

> One possible reason why USB and NFS interact badly is that the kernel stack
> is being exhaused due to there being too many layers. But I'm not a kernel
> hacker and not minded to investigate it. There's probably a compile-time
> option for the Linux kernel to crank the stack size up a bit.

I do not think that kernel stack size is the issue, unless your error
messages specifically say that it is exhausted.

What I see in USB and cifs is that some unexpected event happens, or an
expected event doesn't happen, or there is garbage in a response, and
the whole thing throughs up its hands and quits with no way to restart
except a reboot.

When I have trouble with either USB or CIFS, the rest of the machine is
fully usable, so I can potentially put off a reboot for awhile, but I
must do it eventually if I want to use either service again.

THAT makes me with that USB and CIFS were userland services.  Of course,
that would potentially cause a performance hit.  As it stands, I believe
you would have both multiple data copies and multiple context switches.
In theory, something could probably be done to make going between two
userland processes like that zero copy, but I don't know how you would
get around the context switches.

OTOH, maybe I don't need to care about context switches on a quad core
machine.