Re: [PATCH v2] slab,slub: ignore __GFP_WAIT if we're booting or suspending

[Nick Piggin]

On Sun, Jun 21, 2009 at 07:31:05PM +1000, Benjamin Herrenschmidt wrote:
> 
> > > Right, that might be something to look into, though we haven't yet
> > > applied the technique for suspend & resume. My main issue with it at the
> > > moment is how do I synchronize with allocations that are already
> > > sleeping when changing the gfp flag mask without bloating the normal
> > 
> > Well, but the problem already exists, no? If someone is already
> > sleeping due to __GFP_WAIT, he'll probably sleep till the resume.
> 
> Yes. In fact, without the masking, a driver that hasn't been suspended
> yet could well start sleeping in GFP_KERNEL after the disk driver has
> suspended. It may do so while holding a mutex or similar, which might
> deadlock its own suspend() callback. It's not something that drivers can
> trivially address by having a pre-suspend hook, and avoid allocations,
> since allocations may be done by subsystems on behalf of the driver or
> such. It's a can of worms, which is why I believe the only sane approach
> is to stop allocators from doing IOs once we start suspend.

Maybe so. Masking off __GFP_WAIT up in slab and page allocator
isn't really needed though (or necessarily a good idea to throw
out that information far from where it is used).

Checking for suspend active and avoiding writeout from reclaim
for example might be a better idea.

 
> So yes, just applying the mask would help, but wouldn't completely fix
> it unless we also find a way to synchronize.

You could potentially use srcu or something like that in page
reclaim in order to have a way to be able to kick everyone
out. page reclaim entry/exit from the page allocator isn't such
a fastpath though, so even a simple mutex or something may be
possible.

--
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/