Re: [PATCH v7 9/9] block: Avoid deadlocks with bio allocation by stacking drivers

[Vivek Goyal <vgoyal@redhat.com>]

On Thu, Aug 30, 2012 at 06:43:59PM -0700, Kent Overstreet wrote:
> On Thu, Aug 30, 2012 at 06:07:45PM -0400, Vivek Goyal wrote:
> > On Wed, Aug 29, 2012 at 10:13:45AM -0700, Kent Overstreet wrote:
> > 
> > [..]
> > > > Performance aside, punting submission to per device worker in case of deep
> > > > stack usage sounds cleaner solution to me.
> > > 
> > > Agreed, but performance tends to matter in the real world. And either
> > > way the tricky bits are going to be confined to a few functions, so I
> > > don't think it matters that much.
> > > 
> > > If someone wants to code up the workqueue version and test it, they're
> > > more than welcome...
> > 
> > Here is one quick and dirty proof of concept patch. It checks for stack
> > depth and if remaining space is less than 20% of stack size, then it
> > defers the bio submission to per queue worker.
> 
> I can't think of any correctness issues. I see some stuff that could be
> simplified (blk_drain_deferred_bios() is redundant, just make it a
> wrapper around blk_deffered_bio_work()).
> 
> Still skeptical about the performance impact, though - frankly, on some
> of the hardware I've been running bcache on this would be a visible
> performance regression - probably double digit percentages but I'd have
> to benchmark it.  That kind of of hardware/usage is not normal today,
> but I've put a lot of work into performance and I don't want to make
> things worse without good reason.

Would you like to give this patch a quick try and see with bcache on your
hardware how much performance impact do you see. 

Given the fact that submission through worker happens only in case of 
when stack usage is high, that should reduce the impact of the patch
and common use cases should reamin unaffected.

> 
> Have you tested/benchmarked it?

No, I have not. I will run some simple workloads on SSD.

> 
> There's scheduling behaviour, too. We really want the workqueue thread's
> cpu time to be charged to the process that submitted the bio. (We could
> use a mechanism like that in other places, too... not like this is a new
> issue).
> 
> This is going to be a real issue for users that need strong isolation -
> for any driver that uses non negligable cpu (i.e. dm crypt), we're
> breaking that (not that it wasn't broken already, but this makes it
> worse).

There are so many places in kernel where worker threads do work on behalf
of each process. I think this is really a minor concern and I would not
be too worried about it.

What is concerning though really is the greater stack usage due to
recursive nature of make_request() and performance impact of deferral
to a worker thread.

Thanks
Vivek