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

[Ingo Molnar <mingo@elte.hu>]

* Joerg Roedel <joro@8bytes.org> wrote:

> On Fri, Nov 21, 2008 at 06:43:48PM +0100, Ingo Molnar wrote:
> > 
> > * Joerg Roedel <joerg.roedel@amd.com> wrote:
> > 
> > > +static struct list_head dma_entry_hash[HASH_SIZE];
> > > +
> > > +/* A slab cache to allocate dma_map_entries fast */
> > > +static struct kmem_cache *dma_entry_cache;
> > > +
> > > +/* lock to protect the data structures */
> > > +static DEFINE_SPINLOCK(dma_lock);
> > 
> > some more generic comments about the data structure: it's main purpose 
> > is to provide a mapping based on (dev,addr). There's little if any 
> > cross-entry interaction - same-address+same-dev DMA is checked.
> > 
> > 1)
> > 
> > the hash:
> > 
> > + 	return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
> > 
> > should mix in entry->dev as well - that way we get not just per 
> > address but per device hash space separation as well.
> > 
> > 2)
> > 
> > HASH_FN_SHIFT is 1MB chunks right now - that's probably fine in 
> > practice albeit perhaps a bit too small. There's seldom any coherency 
> > between the physical addresses of DMA - we rarely have any real 
> > (performance-relevant) physical co-location of DMA addresses beyond 4K 
> > granularity. So using 1MB chunking here will discard a good deal of 
> > random low bits we should be hashing on.
> > 
> > 3)
> > 
> > And the most scalable locking would be per hash bucket locking - no 
> > global lock is needed. The bucket hash heads should probably be 
> > cacheline sized - so we'd get one lock per bucket.
> 
> Hmm, I just had the idea of saving this data in struct device. How 
> about that? The locking should scale too and we can extend it 
> easier. For example it simplifys a per-device disable function for 
> the checking. Or another future feature might be leak tracing.

that will help with spreading the hash across devices, but brings in 
lifetime issues: you must be absolutely sure all DMA has drained at 
the point a device is deinitialized.

Dunno ... i think it's still better to have a central hash for all DMA 
ops that is aware of per device details.

The moment we spread this out to struct device we've lost the ability 
to _potentially_ do inter-device checking. (for example to make sure 
no other device is DMA-ing on the same address - which is a possible 
bug pattern as well although right now Linux does not really avoid it 
actively)

Hm?

Btw., also have a look at lib/debugobjects.c: i think we should also 
consider extending that facility and then just hook it up to the DMA 
ops. The DMA checking is kind of a similar "op lifetime" scenario - 
with a few extras to extend lib/debugobjects.c with. Note how it 
already has pooling, a good hash, etc.

	Ingo