Re: [Qemu-devel] [PATCH v2 0/3] Make KVM/MSI code device-ID-aware

[Michael S. Tsirkin]

On Wed, Oct 14, 2015 at 11:46:02AM +0300, Pavel Fedin wrote:
>  Hello!
> 
> > I'm confused by pci changes.  It's all arm specific stuff leaking
> > out to pci core.
> 
>  But... How else would you implement it? I tried to do this as abstract as 
> possible. Just pass device pointer to functions which
> will use it. Or, would you like to override all these functions for ARM 
> architecture? That would be a huge code duplication. And,
> still, they would need PCI device specifier, which they don't have, and what 
> patch 0003 takes care about.
> 
> > Maybe it was a mistake to put stream id in struct MemTxAttrs: there's
> > still nothing using it.
> 
>  It will be used by ITS emulation too. Take a look at RFC, ITS code injects 
> it into the kernel together with MSI data. Also, i know
> that it has something to do with SMMU too, so, if we ever implement ARM SMMU, 
> we'll likely have to deal with them too.
>  Presence of device ID pretty well emulates what the hardware does - it just 
> sends this ID along with PCI transaction.
>  We will use it as soon as we either:
> a) Get kernel API approved
> b) Get software GICv3 emulation from Shlomo, then i'll be able to rebase my 
> old software-emulated ITS on top of that.
> 
> > How about we stick PCIDevice * (or even DeviceState *) there?
> 
>  "there" = where? In MemTxAttrs? Would it be good to have 64-bit pointer 
> instead of 16-bit ID there? It is passed by value, and
> current version fits into register even on 32-bit host.

OK, so we could rename stream id to requester id. That at least would
make sense at the pci level, and would make it unique per hierarchy.

But I suspect all these optimizations break later when we need to support
multiple root complex devices.


> Kind regards,
> Pavel Fedin
> Expert Engineer
> Samsung Electronics Research center Russia
>