Re: [PATCH 03/10] x86: add initialization code for DMA-API debugging

[Ingo Molnar <mingo@elte.hu>]

* Joerg Roedel <joerg.roedel@amd.com> wrote:

> On Fri, Nov 21, 2008 at 05:56:28PM +0100, Ingo Molnar wrote:
> > > +	return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
> > > +}
> > > +
> > > +static struct dma_debug_entry *dma_entry_alloc(void)
> > > +{
> > > +	gfp_t gfp = GFP_KERNEL | __GFP_ZERO;
> > > +
> > > +	if (in_atomic())
> > > +		gfp |= GFP_ATOMIC;
> > > +
> > > +	return kmem_cache_alloc(dma_entry_cache, gfp);
> > 
> > hm. that slab allocation in the middle of DMA mapping ops makes me a 
> > bit nervous. the DMA mapping ops are generally rather atomic.
> > 
> > and in_atomic() check there is a bug on !PREEMPT kernels, so it wont 
> > fly.
> 
> I am not sure I understand this correctly. You say the check for 
> in_atomic() will break on !PREEMPT kernels?

Correct. There is no check to be able to tell whether we can schedule 
or not. I.e. on !PREEMPT your patches will crash and burn.

> > We dont have a gfp flag passed in as all the DMA mapping APIs 
> > really expect all allocations having been done in advance already.
> 
> Hmm, I can change the code to pre-allocate a certain (configurable?) 
> number of these entries and disble the checking if we run out of it.

yeah, that's a good approach too - that's what lockdep does. Perhaps 
make the max entries nr a Kconfig entity - so it can be tuned up/down 
depending on what type of iommu scheme is enabled.

	Ingo