Re: [PATCH -mm 1/2] scsi: remove dma_is_consistent usage in 53c700

[James Bottomley]

On Mon, 2010-06-28 at 12:37 +0900, FUJITA Tomonori wrote:
> On Sun, 27 Jun 2010 10:08:48 -0500
> James Bottomley <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx> wrote:
> 
> > On Sun, 2010-06-27 at 19:10 +0900, FUJITA Tomonori wrote:
> > > 53c700 is the only user of dma_is_consistent():
> > > 
> > > BUG_ON(!dma_is_consistent(hostdata->dev, pScript) && L1_CACHE_BYTES < dma_get_cache_alignment());
> > > 
> > > The above code tries to see if the system can allocate coherent memory
> > > or not. It's for some old systems that can't allocate coherent memory
> > > at all (e.g some parisc systems).
> > 
> > Actually, that's not the right explanation.  The BUG_ON is because of an
> > efficiency in the driver ... it's nothing to do with the architecture.
> > 
> > The driver uses a set of mailboxes, but for efficiency's sake, it packs
> > them into a single coherent area and separates the different usages by a
> > L1 cache stride).  On architectures capable of manufacturing coherent
> > memory, this is a nice speed up in the DMA infrastructure.  However, for
> > incoherent architectures, it's fatal if the dma coherence stride is
> > greater than the L1 cache size, because now we'll get data corruption
> > due to cacheline interference.  That's what the BUG_ON is checking for.
> 
> Sorry, I should have looked the details of the driver.
> 
> You are talking about the following tricks, right?
> 
> #define	MSG_ARRAY_SIZE	8
> #define	MSGOUT_OFFSET	(L1_CACHE_ALIGN(sizeof(SCRIPT)))
> 	__u8	*msgout;
> #define MSGIN_OFFSET	(MSGOUT_OFFSET + L1_CACHE_ALIGN(MSG_ARRAY_SIZE))
> 	__u8	*msgin;
> #define STATUS_OFFSET	(MSGIN_OFFSET + L1_CACHE_ALIGN(MSG_ARRAY_SIZE))
> 	__u8	*status;
> #define SLOTS_OFFSET	(STATUS_OFFSET + L1_CACHE_ALIGN(MSG_ARRAY_SIZE))
> 	struct NCR_700_command_slot	*slots;
> #define	TOTAL_MEM_SIZE	(SLOTS_OFFSET + L1_CACHE_ALIGN(sizeof(struct NCR_700_command_slot) * NCR_700_COMMAND_SLOTS_PER_HOST))

Yes, that's it.  The mailboxes themselves are pretty small, and the
minimum coherent allocation is usually a page, so we'd waste orders of
magnitude more coherent memory than we actually need without this trick
(and on a lot of platforms, coherent memory is scarce).

> > > I think that we can safely remove the above usage:
> > > 
> > > - such old systems haven't triger the above checking for long.
> > > 
> > > - the above condition is important for systems that can't allocate
> > >   coherent memory if these systems do DMA. So probably it would be
> > >   better to have such checking in arch's DMA initialization code
> > >   instead of a driver.
> > 
> > Well, we can't check in the architecture because it's a driver specific
> > thing ... I suppose making it a rule that dma_get_cache_alignment()
> > *must* be <= L1_CACHE_BYTES fixes it ... we seem to have no architecture
> > violating that, so just add it to the documentation, and the check can
> > go.
> 
> Seems that on some architectures (arm and mips at least),
> dma_get_cache_alignment() could greater than L1_CACHE_BYTES. But they
> simply return the possible maximum size of cache size like:
> 
> static inline int dma_get_cache_alignment(void)
> {
> 	/* XXX Largest on any MIPS */
> 	return 128;
> }
> 
> So practically, we should be safe. I guess that we can simply convert
> them to return L1_CACHE_BYTES.

As long as that's architecturally true, yes.   I mean I can't imagine
any architecture that had a dma alignment requirement that was greater
than its L1 cache width ... but I've been surprised be for making
"Obviously this can't happen ..." type statements where MIPS is
concerned.

> Some PARISC and mips are only the fully non-coherent architectures
> that we support now?

I think there might be some ARM SoC systems as well ... there were some
strange ones that had tight limits on the addresses the other SoC
components could DMA to which made it very difficult to make consistent
areas.

>  We can remove the above checking if
> dma_get_cache_alignment() is <= L1_CACHE_BYTES on PARISC and mips?

I don't think we need to check, just document that
dma_get_cache_alignment cannot be greater than the L1 cache stride.

James


--
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/