On Tue, Jan 21, 2020 at 12:00:57PM +0800, Jason Wang wrote:

On 2020/1/21 上午1:49, Jason Gunthorpe wrote:

quoted

On Mon, Jan 20, 2020 at 04:43:53PM +0800, Jason Wang wrote:

quoted

This is similar to the design of platform IOMMU part of vhost-vdpa. We
decide to send diffs to platform IOMMU there. If it's ok to do that in
driver, we can replace set_map with incremental API like map()/unmap().

Then driver need to maintain rbtree itself.

I think we really need to see two modes, one where there is a fixed
translation without dynamic vIOMMU driven changes and one that
supports vIOMMU.

I think in this case, you meant the method proposed by Shahaf that sends
diffs of "fixed translation" to device?

It would be kind of tricky to deal with the following case for example:

old map [4G, 16G) new map [4G, 8G)

If we do

1) flush [4G, 16G)
2) add [4G, 8G)

There could be a window between 1) and 2).

It requires the IOMMU that can do

1) remove [8G, 16G)
2) flush [8G, 16G)
3) change [4G, 8G)

....

Basically what I had in mind is something like qemu memory api

0. begin
1. remove [8G, 16G)
2. add [4G, 8G)
3. commit

Anyway, I'm fine with a one-shot API for now, we can
improve it later.

quoted

There are different optimization goals in the drivers for these two
configurations.

quoted

quoted

If the first one, then I think memory hotplug is a heavy flow
regardless. Do you think the extra cycles for the tree traverse
will be visible in any way?

I think if the driver can pause the DMA during the time for setting up new
mapping, it should be fine.

This is very tricky for any driver if the mapping change hits the
virtio rings. :(

Even a IOMMU using driver is going to have problems with that..

Jason

Or I wonder whether ATS/PRI can help here. E.g during I/O page fault,
driver/device can wait for the new mapping to be set and then replay the
DMA.

Thanks

-- 
MST