Re: [RFC][PATCH 2/9] deadlock prevention core
From: Peter Zijlstra
Date: Wed Aug 09 2006 - 02:58:47 EST
On Tue, 2006-08-08 at 22:44 -0700, Daniel Phillips wrote:
> David Miller wrote:
> >From: Daniel Phillips <phillips@xxxxxxxxxx>
> >>David Miller wrote:
> >>>I think the new atomic operation that will seemingly occur on every
> >>>device SKB free is unacceptable.
> >>Alternate suggestion?
> > Sorry, I have none. But you're unlikely to get your changes
> > considered seriously unless you can avoid any new overhead your patch
> > has which is of this level.
> We just skip anything new unless the socket is actively carrying block
> IO traffic, in which case we pay a miniscule price to avoid severe
> performance artifacts or in the worst case, deadlock. So in this design
> the new atomic operation does not occur on every device SKP free.
> All atomic ops sit behind the cheap test:
> (dev->flags & IFF_MEMALLOC)
> or if any have escaped that is just an oversight. Peter?
That should be so indeed. Except on the allocation path ofcourse, there
it only occurs when the first allocation fails.
> > We're busy trying to make these data structures smaller, and eliminate
> > atomic operations, as much as possible. Therefore anything which adds
> > new datastructure elements and new atomic operations will be met with
> > fierce resistence unless it results an equal or greater shrink of
> > datastructures elsewhere or removes atomic operations elsewhere in
> > the critical path.
> Right now we have a problem because our network stack cannot support
> block IO reliably. Without that, Linux is no enterprise storage
Indeed, surely not all wanted new features come with zero cost. If its a
hard condition that all new features remove data and operations progress
is going to be challenging.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/