Re: 2.6.1-rc1-tiny2

[Jens Axboe <axboe@suse.de>]

On Wed, Jan 07 2004, Matt Mackall wrote:
> On Wed, Jan 07, 2004 at 03:06:40PM +0100, Jens Axboe wrote:
> > On Mon, Jan 05 2004, Matt Mackall wrote:
> > > This is the fourth release of the -tiny kernel tree. The aim of this
> > > tree is to collect patches that reduce kernel disk and memory
> > > footprint as well as tools for working on small systems. Target users
> > > are things like embedded systems, small or legacy desktop folks, and
> > > handhelds.
> > > 
> > > Latest release includes:
> > >  - various compile fixes for last release
> > >  - actually include Andi Kleen's bloat-o-meter this time
> > >  - optional mempool removal
> > 
> > Your CONFIG_MEMPOOL is completely broken as you are no longer giving the
> > same guarentees (you have no reserve at all). Might as well change it to
> > CONFIG_DEADLOCK instead.
> 
> It's equivalent to a pool size of zero, yes, so deadlock odds are
> significantly higher with some usage scenarios. I'll add a big fat
> warning.

Precisely. In most scenarios it makes deadlocks possible, where it was
safe before (more below).

> On the other hand, the existence of pre-allocated mempools can greatly
> increase the likelihood of starvation, oom, and deadlock on the rest
> of the system, especially as it becomes a greater percentage of the
> total free memory on a small system. In other words, I had to cut this
> corner to make running in 2M work with my config. When I merge
> CONFIG_BLOCK, it'll be more generally useful.

It needs to be carefulled tuned, definitely.

> For the sake of our other readers, I'll point out that mempool doesn't
> intrinisically reduce deadlock odds to zero unless we have a hard
> limit on requests in flight that's strictly less than pool size.

That's not true, depends entirely on usage. It's not a magic wand. And
you don't need a hard limit, you only need progress guarentee. Typically
just a single pre-allocated object can make you 100% deadlock free, if
stacking is not involved. So for most cases, I think it would be much
better if you just hard wired min_nr to 1, that would move you from 90%
to 99% safe :-)

-- 
Jens Axboe