Re: [RFC][PATCH] PM: Force GFP_NOIO during suspend/resume (was: Re: [linux-pm] Memory allocations in .suspend became very unreliable)

[Rafael J. Wysocki]

On Wednesday 20 January 2010, Oliver Neukum wrote:
> Am Mittwoch, 20. Januar 2010 00:17:51 schrieb Benjamin Herrenschmidt:
> > On Tue, 2010-01-19 at 10:04 +0100, Bastien ROUCARIES wrote:
> > > Instead of masking bit could we only check if incompatible flags are
> > > used during suspend, and warm deeply. Call stack will be therefore
> > > identified, and we could have some metrics about such problem.
> > > 
> > > It will be a debug option like lockdep but pretty low cost.
> > 
> > I still believe it would just be a giant can of worms to require every
> > call site of memory allocators to "know" whether suspend has been
> > started or not.... Along the same reasons why we added that stuff for
> > boot time allocs.
> 
> But we have the freezer. So generally we don't require that knowledge.
> We can expect no normal IO to happen.
> The question is in the suspend paths. We never may use anything
> but GFP_NOIO (and GFP_ATOMIC) in the suspend() path. We can
> take care of that requirement in the allocator only if the whole system
> is suspended. As soon as a driver does runtime power management,
> it is on its own.

If you start new kernel threads using the async framework, for example,
GFP_KERNEL allocations are going to be used.

As I said before, IMnshO , duplicating every piece of code that allocates
memory and can be run during suspend/resume as well as in other circumstances
doesn't make sense.

Rafael
--
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/