[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH v2 0/2] ARM: ACPI: IORT: Hide SMMU from hardware domain's IORT table



Hello Julien,

On 10/4/2017 12:12 AM, Julien Grall wrote:
Hello,

On 25/09/17 05:22, Manish Jaggi wrote:
On 9/22/2017 7:42 PM, Andre Przywara wrote:
Hi Manish,

On 11/09/17 22:33, mjaggi@xxxxxxxxxxxxxxxxxx wrote:
From: Manish Jaggi <mjaggi@xxxxxxxxxx>

The set is divided into two patches. First one calculates the size of IORT
while second one writes the IORT table itself.
It would be good if you could give a quick introduction *why* this set
is needed here (and introduce IORT to the casual reader).
In general some more high-level documentation on your functions would be good, as it took me quite some time to understand what each function does.
ok, will add more documentation.
So my understanding is:
phase 1:
- go over each entry in each RC node
Rather than each entry (which could be a large number) I am taking the complete range and checking it with the same logic. If the ID range is a subset or a super-set of id range in smmu, new id range is created.

So if pci_rc node has an id map {p_input-base,p_output-base,p_out_ref, p_count} and it an output reference to smmu node with id-map {s_input-base, s_output-base,s_out_ref, s_count}, based on the the the s_count and s_input/p_output the new id-map is created with {p_input, s_output, s_out_ref, adjusted_count}

update_id_mapping function does that.

So I am following the same logic. We can chat over IRC / I can give a code walk-through ...

-   if that points to an SMMU node, go over each outgoing ITS entry and
find overlaps with this RC entry
-     for each overlap create a new entry in a list with this RC
pointing to the ITS directly

phase 2, creating the new IORT
- go over each RC node
-   if that points to an ITS, copy through IORT entries
-   if that points to an SMMU, replace with the remapped entries
- go over each ITS node
-   copy through IORT entries
Thats exactly what this patch does.
What are you comments on the current patch approach to hide smmu nodes.
I have answered to your comments, see below.
IMHO we can reuse most of the fixup code here.

So I believe this would do the trick and you end up with an efficient
representation of the IORT without SMMUs - at least for RC nodes.

After some brainstorming with Julien we found two problems:
1) This only covers RC nodes, but not "named components" (platform
devices), which we will need. That should be fixable by removing the
hardcoded IORT node types in the code and treating NC nodes like RC nodes.
Yes, so first we can take this as a base, once this is ok, I can add support for named components.
2) Eventually we will need *virtual* deviceID support, for DomUs. Now we
I am a bit surprised that you answered to the e-mail but didn't provide any opinion on 2).
Apologies for that.
could start introducing that already, also doing some virtual mapping
for Dom0. The ITS code would then translate each virtual device ID that
Dom0 requests into a hardware device ID.
I agree that this means a lot more work, but we will need it anyway.

I am a bit surprised that you answered to the e-mail but didn't provide any opinion on 2).
Apologies for that. Sorry to surprise you twice :)

IMHO It was a bit obvious for DomU and I was waiting to hear what other would say on this.
as (2) below.
Moreover we need to discuss IORT generation for DomU
- could be done by xl tools
or xen should do it.

Also this is the part of PCI passthrough flow so that also might change few things.

But from pov of dom0 smmu hiding, it is a different flow and is coupled with PCI PT.



I think 1) can be solved using this series as a base. I have quite some
comments ready for the patches, shall we follow this route.

2) obviously would change the game completely. We need to sit down and
design this properly. Probably this means that Xen parses the IORT and
builds internal representations of the mappings,
Can you please add more detail on the internal representations of the mappings. IIUC the information is already there in ACPI tables, would it not add extra overhead of abstractions to maintain. Enumeration of PCI devices would generate a pci list which would be anyways separate.
which are consulted as
needed when passing through devices. The guest's (that would include
Dom0) IORT would then be generated completely from scratch.

I have a different opinion here, dom0 IORT would is most cases be very close to host IORT sans smmu nodes and few platform devices. And which platform devices to hide would probably depend on the xen command line, For instance for dom0 we would copy ITS information while for domU it would have to be generated, so scratch would be more for domU. We could have a common code for creating IORT structure but it would be a bit complex code with lot of abstractions and callbacks, so I suggest that keeping code simpler would be better.
I would like to hear your opinion on this. I will try to discuss the
feasibility of 2) with people at Connect. It would be good if we could
decide whether this is the way to go or we should use a solution based
on this series.

Andre, Stefano and I had a chat during connect about how we want to see IORT support in Xen. In the near future, IORT will be used for different components accross the hypervisor (ITS, SMMU...) and as a way to
communicate the topology to the guests.
Are we thinking about virtual SMMU here as well?

The IORT for the hardware domain is just a specific case as it is based pre-existing information. But because of removing nodes (e.g SMMU nodes and probably the PMU nodes), it is basically a full re-write.

So I would consider of full separate the logic of generating the IORT table from the host IORT table. By that I mean not browsing the host IORT when generating the host.

by "the host" you mean dom0 IORT  ?
This has two benefits:
1) The code generation could be re-used for generating the guest IORT later on.
See my comment above
    2) See 2) from Andre
3) We could decide in a finer grain which devices (e.g platform device) Dom0 can see.
ok,

So, IHMO, we should take a different approach from this series and extending the scope of it. Rather than focusing on only IORT for the hardware, I would be better to see IORT as a whole. I.e how IORT will interact with the hypervisor?

For instance, likely you would need to parse the IORT in quite a few places in Xen. It would be better to get IORT parsed only once and store the information in Xen like data-structures.

I am thinking of reusing much of ACPI tables for that and introducing less abstractions.
This require more work than this current scope of this series. But I think it would helful for future work such as PCI passthrough support.

Any opinions?

for DomUs it is tied to PCI PT, so both design should evolve together.
Cheers,



_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel

 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.