Re: [PATCH v2 -mm 0/2] x86: per-device dma_mapping_ops

[Amit Shah <amit.shah@qumranet.com>]

On Monday 26 May 2008 11:41:52 FUJITA Tomonori wrote:
> On Mon, 26 May 2008 09:39:20 +0530
>
> Amit Shah <amit.shah@qumranet.com> wrote:
> > On Sunday 25 May 2008 12:50:11 Muli Ben-Yehuda wrote:
> > > On Thu, May 22, 2008 at 04:13:02PM +0530, Amit Shah wrote:
> > > > OK; this sounds helpful. the hook can make a hypercall and confirm
> > > > with the host kernel if the device in question is an assigned
> > > > physical device. If yes, we replace the dma_ops. Though, the
> > > > original intent of having stackable ops is that we might want to go
> > > > through the swiotlb in the guest even for an assigned device if the
> > > > guest dma addresses are not in the addressable range of the guest
> > > > chipset.
> > > >
> > > > > created (it works with hot plugging). It enables IOMMUs to set up
> > > > > an appropriate dma_mapping_ops per device.
> > > >
> > > > From what we've discussed so far, it looks like stackable dma ops
> > > > will definitely be needed. Does this patchset provide something that
> > > > stacking won't?
> > >
> > > Yes---this patchset let's you have a per-device dma-ops, whereas with
> > > stackable you only get global dma-ops. I think it's clear we need
> > > both, and I think per-device dma-ops are the first thing that's
> > > needed. Stacking can then be introduced on a per-device basis.
> >
> > When we would want stacking, we'll want it globally and not per-device,
> > isn't it? Or at least for devices on a particular bus.
> >
> > When an IOMMU driver registers itself, it should tell which devices it's
> > interested in (each device behind a bus or by enumerating each device it
> > cares for). This should take care of all the scenarios and we won't have
> > the need for per-device dma_ops.
>
> Well, without per-device dma_ops, IOMMUs could live. But it's pretty
> hacky. Every time a dma operation is called, IOMMUs need to figure out
> how a device should be handled.

What if this information could be hidden behind (a slightly complicated) 
get_dma_ops()? Also, each of the operations in dma_ops will see if there's 
something else down the stack that might be interested in the current device.

My contention is that we are going to need stackable ops, and a full-fledged 
stackable implementation is going to solve this problem as well. However, 
this current implementation of per-device dma_ops looks like a really simple 
and non-intrusive solution to one problem, that of getting rid of some 
overheads in the IOMMU code.

> If IOMMUs can set dma_ops for the device when a new device is created,
> IOMMUs don't care anything any more. That's much clean. That's What
> the POWER architecture does.
>
> > For something like pvdma, we can walk through the list of pci devices and
> > make a hypercall for each of them to get this information and have the
> > pvdma version of dma_ops registered for that device. This sounds like
> > it's per-device dma_ops, but it's not -- internally, the dma operations
> > walk through each of the IOMMUs registered and call them in sequence.
>
> As Muli poinsted out, For pvdma, you need stacking per-device
> dma_ops. With per-device dma_ops, you don't need hack like adding
> is_pv_device hook in dma_ops. You can set your dma_ops to only pci
> devices that you are interested.

The hack was added only because there's no stackable dma api we have now. 
Sure, per-device dma_ops is going to solve this problem and I like it. I'm 
only saying we're also going to need stacking ops and in effect, per-device 
dma_ops would just be replaced by them once we get the complete solution.