Re: [RFC v1 3/5] xen/arm: introduce SCMI-SMC mediator driver

[Julien Grall <julien@xxxxxxx>]

Hi,

On 06/01/2022 15:43, Oleksii Moisieiev wrote:
> On Thu, Jan 06, 2022 at 02:02:10PM +0000, Julien Grall wrote:
> > On 06/01/2022 13:53, Oleksii Moisieiev wrote:
> > > Hi Julien,
> > Hi,
> >
> > > On Mon, Jan 03, 2022 at 01:14:17PM +0000, Julien Grall wrote:
> > > > Hi,
> > > >
> > > > On 24/12/2021 17:02, Oleksii Moisieiev wrote:
> > > > > On Fri, Dec 24, 2021 at 03:42:42PM +0100, Julien Grall wrote:
> > > > > > On 20/12/2021 16:41, Oleksii Moisieiev wrote:
> > > > > > > >       2) What are the expected memory attribute for the regions?
> > > > > > > xen uses iommu_permit_access to pass agent page to the guest. So guest can 
> > > > > > > access the page directly.
> > > > > > I think you misunderstood my comment. Memory can be mapped with various type
> > > > > > (e.g. Device, Memory) and attribute (cacheable, non-cacheable...). What will
> > > > > > the firmware expect? What will the guest OS usually?
> > > > > >
> > > > > > The reason I am asking is the attributes have to matched to avoid any
> > > > > > coherency issues. At the moment, if you use XEN_DOMCTL_memory_mapping, Xen
> > > > > > will configure the stage-2 to use Device nGnRnE. As the result, the result
> > > > > > memory access will be Device nGnRnE which is very strict.
> > > > > Let me share with you the configuration example:
> > > > > scmi expects memory to be configured in the device-tree:
> > > > >
> > > > > cpu_scp_shm: scp-shmem@0xXXXXXXX {
> > > > >         compatible = "arm,scmi-shmem";
> > > > >         reg = <0x0 0xXXXXXX 0x0 0x1000>;
> > > > > };
> > > > >
> > > > > where XXXXXX address I allocate in alloc_magic_pages function.
> > > > The goal of alloc_magic_pages() is to allocate RAM. However, what you want
> > > > is a guest physical address and then map a part of the shared page.
> > > Do you mean that I can't use alloc_magic_pages to allocate shared
> > > memory region for SCMI?
> > Correct. alloc_magic_pages() will allocate a RAM page and then assign to the
> > guest. From your description, this is not what you want because you will
> > call XEN_DOMCTL_memory_mapping (and therefore replace the mapping).
> Ok thanks, I will refactor this part in v2.
>
> > > > I can see two options here:
> > > >     1) Hardcode the SCMI region in the memory map
> > > >     2) Create a new region in the memory map that can be used for reserving
> > > > memory for mapping.
> > > Could you please explain what do you mean under the "new region in the
> > > memory map"?
> > I mean reserving some guest physical address that could be used for map host
> > physical address (e.g. SCMI region, GIC CPU interface...).
> >
> > So rather than hardcoding the address, we have something more flexible.
> Ok, I will fix that in v2.
Just for avoidance of doubt. I was clarify option 2 and not requesting 
to implement. That said, if you want to implement option 2 I would be 
happy to review it.
> > > > We still have plenty of space in the guest memory map. So the former is
> > > > probably going to be fine for now.
> > > >
> > > > > Then I get paddr of the scmi channel for this domain and use
> > > > > XEN_DOMCTL_memory_mapping to map scmi channel address to gfn.
> > > > >    > Hope I wass able to answer your question.
> > > > What you provided me is how the guest OS will locate the shared memory. This
> > > > still doesn't tell me which memory attribute will be used to map the page in
> > > > Stage-1 (guest page-tables).
> > > >
> > > > To find that out, you want to look at the driver and how the mapping is
> > > > done. The Linux driver (drivers/firmware/arm_scmi) is using devm_ioremap()
> > > > (see smc_chan_setup()).
> > > >
> > > > Under the hood, the function devm_ioremap() is using PROT_DEVICE_nGnRE
> > > > (arm64) which is one of the most restrictive memory attribute.
> > > >
> > > > This means the firmware should be able to deal with the most restrictive
> > > > attribute and therefore using XEN_DOMCTL_memory_mapping to map the shared
> > > > page in stage-2 should be fine.
> > > I'm using vmap call to map channel memory (see smc_create_channel()).
> > > vmap call sets PAGE_HYPERVISOR flag which sets MT_NORMAL (0x7) flag.
> > You want to use ioremap().
> I've used ioremap originally, but changed it to vmap because ioremap
> doesn't support memcpy.
> What if I use __vmap with MT_DEVICE_nGnRE flag?
That's not going to help. Our implementation of memcpy() is using 
unaligned access (which is forbidden on Device memory).
You will need something similar to memcpy_toio() in Linux. I don't think 
we have one today in Xen, so I would suggest to import the 
implementation from Linux.
> > > Considering that protocol is synchronous and only one agent per channel is
> > > expected - this works fine for now.
> > > But I agree that the same memory attributes should be used in xen and
> > > kernel. I fill fix that in v2.
> > I am a bit confused. Are you mapping the full shared memory area in Xen? If
> > yes, why do you need to map the memory that is going to be shared with a
> > domain?
> Xen should have an access to all agent channels because it should send
> SCMI_BASE_DISCOVER_AGENT to each channel and receive agent_id during
> scmi_probe call.
Hmmm... Just to confirm, this will only happen during Xen boot? IOW, Xen 
will never write to the channel when a domain is running?
If yes, then I think it would be best to unmap the channel once they are 
used. This would prevent all sort of issues (e.g. Xen mistakenly written 
in them).
Cheers,

--
Julien Grall