Re: [Qemu-devel] buffer alignment for block backends

[Anthony Liguori]

Stefano Stabellini wrote:
> Anthony Liguori wrote:
>
>   
> > > If you do not want to do that, another possible solution is to create a
> > > new function called "qemu_blockalign" that would be implemented as
> > > qemu_memalign(512, size);
> > >       
> > This is fine, but this is purely an optimization, it cannot be relied 
> > upon in the general case.
> >     
>
>
>
> well, there aren't many places that allocate buffers for the block
> backends, I can count only the following for dma operations:
>
> - block.c:bdrv_aio_rw_vector
>   

This bounces a scatter/gather list into a single linear buffer since not 
all backends handle scatter/gather lists today.

> and this one for other ide read and write operations:
>
> - ide.c:ide_init2
>   

This buffer is only used when not doing DMA.  When doing DMA, we are 
able to do zero-copy IO so the alignment of the request depends on how 
the guest aligned the request.  I suspect you'll find a lot of guests 
that, in practice, do not align requests at 4k boundaries.  I don't know 
what the requirements are for IDE but I would be surprised if it was 4k.

> I think it would be important at least for dma operations.
>   

We have a lot of places with explicit memalign's because the block raw 
backend code degrades into synchronous IO when performing non-aligned IO 
with cache=off.  I've never liked this much personally.

That's not saying that I think we shouldn't try to align DMA buffers 
when we're allocating them in QEMU.  A block level function to do this 
would be pretty nice in fact.

> > >  so we don't have to write 512 bytes everywhere
> > > but only in one place, thus making life easier to people like me that
> > > have to change the value for a special case.
> > >
> > > Thanks in advance for your sympathy :)
> > >   
> > >       
> > Why does your backend requirement page alignment and who's notion of page?
> >
> >     
>
> my backend (block-vbd) needs page aligned buffers because blkfront needs
> page aligned buffers. I could allocate a new page aligned buffer every
> time and the memcpy but it is just a waste.
>   

You'll need to check the alignment of the request and bounce it if 
necessary.  In the case that you have zero-copy requests coming from the 
guest that aren't page aligned, someone has to bounce the thing to make 
it page aligned.

So I presume you're implementing blkfront in userspace?  Does minios 
provide a userspace interface for grant tables that looks similar to the 
interfaces on Linux?  Were you planning on submitting this for inclusion 
in upstream QEMU?

I think it's a reasonable thing to include provided it's relatively 
self-contained.

Regards,

Anthony Liguori