6.1-STABLE hangs, ddb shows 'acpi_timer_read'?

[Robert Watson]

On Fri, 20 Oct 2006, Karl Pielorz wrote:

>> On Fri, Oct 20, 2006 at 11:56:33AM +0100, Karl Pielorz wrote:
>>> 
>>> I have a crash dump from it - which I've saved (I'm moderately familiar 
>>> with working with dumps, but this one is split into two?)
>>> 
>>> If anyone has any pointers, or can do some hand holding to get more info 
>>> from the dump, or what to do next time it happens?
>> 
>> See 
>> http://www.freebsd.org/doc/en_US.ISO8859-1/books/developers-handbook/kern 
>> eldebug-deadlocks.html
>
> Thanks for the link! Anyone know much is all that lot (such as 
> INVARIANTS/WITNESS) etc. likely to slow the machine down? - A few percent? 
> More?
>
> I'm just a little hesitant to put it all in, and end up with a machine 
> that's 80% slower :(

Depends a lot on your workload.  WITNESS used to really, really slow things 
down for kernel lock intensive workloads (VFS especially); now it just really 
slows things down.  INVARIANTS overhead is generally measurable, but for most 
workloads it is likely <20%.  The place INVARIANTS hits you is when the kernel 
is allocating and freeing lots of memory, during which INVARIANTS will be 
scrubbing and checking for use after free, etc.  It's worth running with 
WITNESS when debugging deadlocks if possible, because it is, essentially, a 
deadlock debugging tool.  A moderate number of people run with INVARIANTS in 
production, but you generally don't want to do that with WITNESS.

Robert N M Watson
Computer Laboratory
University of Cambridge