Re: [PATCH] blk-mq-dma: always initialize dma state

[Keith Busch <kbusch@kernel.org>]

On Wed, Dec 10, 2025 at 07:54:07AM +0100, Christoph Hellwig wrote:
> On Tue, Dec 09, 2025 at 10:49:15PM -0800, Keith Busch wrote:
> > -	if (blk_can_dma_map_iova(req, dma_dev) &&
> > -	    dma_iova_try_alloc(dma_dev, state, vec.paddr, total_len))
> > +	if (!blk_can_dma_map_iova(req, dma_dev))
> > +		memset(state, 0, sizeof(*state));
> > +	else if (dma_iova_try_alloc(dma_dev, state, vec.paddr, total_len))
> >  		return blk_rq_dma_map_iova(req, dma_dev, state, iter, &vec);
> 
> What about just doing the memset unconditionally?  It's just two
> 64-bit fields so no real overhead, and it gives us a clean slate that
> avoid introducing other bugs later on.

I didn't do that is because dma_iova_try_alloc() also does that, so it'd
be two repeated and unnecessary memset's on the same address. That feels
undesirable no matter how small the struct is. I could remove the memset
in dma_iova_try_alloc instead and make the caller initialize it. There
are only two existing users of the API, so not a big deal to change it.