Again about pbuf_mtx

[Alexander Motin]

On 05.09.2013 15:40, Alexander Motin wrote:
> Some may remember that not so long ago I complained about high lock
> congestion on pbuf_mtx. At that time switching the mutex to padalign
> reduced the problem. But now after improving scalability in CAM and GEOM
> and doing more then half million IOPS on 32-core system I again heavily
> hit that problem -- hwpmc shows about 30% of CPU time spent on that
> mutex spinning and another 30% of time spent on attempt of threads to go
> to sleep on that mutex and getting more collisions there.
>
> Trying to mitigate that I've made a patch
> (http://people.freebsd.org/~mav/pcpu_pbuf.patch) to split single queue
> of pbufs into several. That definitely cost some amount of KVA and
> memory, but on my tests it fixes problem redically, removing any
> measurable congestion there. The patch is not complete and don't even
> boot on i386 now, but I would like to hear opinions about the approach,
> or may be some better propositions.

On kib@ proposition I've tried to reimplement that patch using vmem(9). 
Code indeed looks much better (at least looked before workarounds):
http://people.freebsd.org/~mav/pbuf_vmem.patch
and it works fast, but I have found number of problems:
  - now we have only 256 (or even less) pbufs and UMA used by vmem for 
quick caches tend to allocate up to 256 items per CPU and never release 
them back. I've partially workarounded that by passing fake MAGIC_SIZE 
value to vmem and down to UMA as size to make initial bucket sizes 
smaller, but that is a hack and not always sufficient since size may 
grow on congestion and again never shrink back.
  - UMA panics with "uma_zalloc: Bucket pointer mangled." if I am giving 
vmem zero as valid pointer. I've workarounded that by adding an offset 
to the value, but I think that assertion in UMA should be removed if we 
are going to use it for abstract values now.

> Another patch I've made
> (http://people.freebsd.org/~mav/si_threadcount.patch) removes lock
> acquisition from dev_relthread() by using atomics for reference
> counting. That fixes another congestion I see. This patch looks fine to
> me and the only congestion I see after that is on HBA driver locks, but
> may be I am missing something?

-- 
Alexander Motin