Some odd behaviour of vmstat and vmtotal...

[Sam Leffler]

Alfred Perlstein wrote:
> * gnn at freebsd.org <gnn at freebsd.org> [080430 18:13] wrote:
>   
>> Howdy,
>>
>> In deploying 7.0 at work we were finding a persistent problem when
>> running "vmstat 1" on systems.  The problem shows up as a 10ms "pause"
>> in processing, usually packet stamping and forwarding by a user level
>> process.  This is due to the fact that vmstat calls the vmtotal()
>> routine which in turn does a whole lot of locking.  The vmtotal call
>> locks and walks the VM object list twice in a mark and sweep
>> operation.
>>
>> So, the question is, "What is the right way to fix this?"  I could
>> remove the locking since the O_ACTIVE bit is not used by any other
>> routine besides vmtotal, but I'm not too happy about that.
>>
>> The relevant code can be found in src/sys/vm/vm_meter.c in vmtotal().
>>
>> Thoughts?
>>     
>
> Can you _really_ remove the locking?  I don't think that would be safe
> if the list is being manipulated, but I haven't checked.
>
> Typically one uses a sentinal to get around such problems.  Basically
> insert a dummy object that no one should touch other than you, then
> drop the lock periodically and insert it into the list, then resume.
>
>   

net80211 uses a scan generation # to walk it's node lists.  No sentinel.

    Sam