[RFC 0/7] Introduce swiotlb throttling

[Christoph Hellwig]

On Mon, Aug 26, 2024 at 03:27:30PM +0000, Michael Kelley wrote:
> OK, this makes sense to me. The DMA_ATTR_* symbols are currently
> defined as just values that are not part of an enum or any other higher
> level abstraction, and the "attrs" parameter to the dma_* functions is
> just "unsigned long". Are you thinking that the separate namespace is
> based only on the symbolic name (i.e., DMA_MAP_* vs DMA_ATTR_*),
> with the values being disjoint? That seems straightforward to me.

Yes. Although initially I'd just keep ATTR for the allocation and then
maybe do a scripted run to convert it.

> Changing the "attrs" parameter to an enum is a much bigger change ....

I don't think an enum makes much sense as we have bits defined.  A
__bitwise type would be nice, but not required.

> For a transition period we can have both DMA_ATTR_SKIP_CPU_SYNC
> and DMA_MAP_SKIP_CPU_SYNC, and then work to change all
> occurrences of the former to the latter.
> 
> I'll have to look more closely at WEAK_ORDERING and NO_WARN.
> 
> There are also a couple of places where DMA_ATTR_NO_KERNEL_MAPPING
> is used for dma_map_* calls, but those are clearly bogus since that
> attribute is never tested in the map path.

Yeah, these kinds of bogus things is what I'd like to kill..


> > Note that this also in general involves changes to the block drivers
> > to set that flag, which is a bit annoying, but I guess there is not
> > easy way around it without paying the price for the BLK_MQ_F_BLOCKING
> > overhead everywhere.
> 
> Agreed. I assumed there was some cost to BLK_MQ_F_BLOCKING since
> the default is !BLK_MQ_F_BLOCKING, but I don't really know what
> that is. Do you have a short summary, just for my education?

I think the biggest issue is that synchronize_srcu is pretty damn
expensive, but there's also a whole bunch of places that unconditionally
defer to the workqueue.