Re: [PATCH v2 00/12] New paravirtual PCI front-end for Hyper-V VMs

[Marc Zyngier]

On 14/09/15 18:59, Jake Oshins wrote:
>> -----Original Message-----
>> From: Marc Zyngier [mailto:marc.zyngier@xxxxxxx]
>> Sent: Monday, September 14, 2015 8:01 AM
>> To: Jake Oshins <jakeo@xxxxxxxxxxxxx>; gregkh@xxxxxxxxxxxxxxxxxxx; KY
>> Srinivasan <kys@xxxxxxxxxxxxx>; linux-kernel@xxxxxxxxxxxxxxx;
>> devel@xxxxxxxxxxxxxxxxxxxxxx; olaf@xxxxxxxxx; apw@xxxxxxxxxxxxx;
>> vkuznets@xxxxxxxxxx; linux-pci@xxxxxxxxxxxxxxx; bhelgaas@xxxxxxxxxx;
>> tglx@xxxxxxxxxxxxx; Jiang Liu <jiang.liu@xxxxxxxxxxxxxxx>
>> Subject: Re: [PATCH v2 00/12] New paravirtual PCI front-end for Hyper-V
>> VMs
>>
>> Hi Jake,
>>
>> In the future, please CC me on anything that touches irqdomains, along
>> with Jiang Liu as we both co-maintain this piece of code.
>>
> 
> Absolutely.  Sorry for that omission.
> 
>> On 11/09/15 01:00, jakeo@xxxxxxxxxxxxx wrote:
>>> From: Jake Oshins <jakeo@xxxxxxxxxxxxx>
>>>
>>> The patch series updates the one sent about a month ago in three ways.  It
>>> integrated with other IRQ domain work done in linux-next in that time, it
>>> distributes interrupts to multiple virtual processors in the guest VM, and it
>>> incorporates feedback from Thomas Gleixner and others.
>>>
>>> These patches change the IRQ domain code so that an IRQ domain can
>> match on both
>>> bus type and on the PCI domain.  The IRQ domain match code is modified
>> so that
>>> IRQ domains can have a "rank," allowing for a default one which matches
>> every
>>> x86 PC and more specific ones that replace the default.
>>
>> I'm not really fond of this approach. We already have a way to match an
>> IRQ domain, and that's the device node. It looks to me that you're going
>> through a lot of pain inventing a new infrastructure to avoid divorcing
>> the two. If you could lookup your PCI IRQ domain directly based some
>> (non-DT) identifier, and then possibly fallback to the default one,
>> would that help?
>>
>> If so, here's the deal: I have been working on a patch series that
>> addresses the above for unrelated reasons (ACPI support on arm64). It
>> has been posted twice already:
>>
>> http://lists.infradead.org/pipermail/linux-arm-kernel/2015-July/358768.html
>>
>> and the latest version is there:
>>
>> https://git.kernel.org/cgit/linux/kernel/git/maz/arm-
>> platforms.git/log/?h=irq/gsi-irq-domain-v3
>>
>> I have the feeling that you could replace a lot of your patches with
>> this infrastructure.
>>
>> Thoughts?
>>
>> 	M.
>> --
> 
> First, thank you so much for reviewing this.  I've read the patch
> series above, but I'm sure that I might have misinterpreted it.  It
> seems to merge the DT and ACPI GSI infrastructure, which I think is a
> great idea.  I'm not sure, however, that it would, as it stands,
> provide what I need here.  Please do tell me if I'm wrong.
> 
> The series above allows you to supply different IRQ domains for
> separate parts of the ACPI GSI space, which is fine for IRQs which
> are actually defined by ACPI.  Message-signaled interrupts (MSI),
> however, aren't defined by ACPI.  ACPI only talks about the routing
> of interrupts with pins and traces (or ones which have equivalent
> mechanisms like the INTx# protocol in PCI Express.)
> 
> What the older DT layer code allowed was for the PCI driver to look
> up an IRQ domain by walking up the device tree looking for a node
> that claimed to be an IRQ domain.  The match() function on the IRQ
> domain allowed it to say that it supported interrupts on PCI buses.
> 
> What's not clear to me is how I would create an IRQ domain that
> matches not on ACPI GSI ranges (because ACPI doesn't talk about MSI)
> and not just on generic PCI buses.  I need to be able to ask for an
> IRQ domain "from my parent" which doesn't really exist without the OF
> device tree or "for a specific PCI bus domain."  That second one is
> what I was trying to enable.
> 
> Is there a way to do that with the infrastructure that you're
> introducing?

The ACPI/GSI stuff is a red herring, and is completely unrelated to the
problem you're trying to solve. What I think is of interest to you is
contained in the first three patches.

In your 4th patch, you have the following code:

+	pci_domain = pci_domain_nr(bus);
+	d = irq_find_matching_host(NULL, DOMAIN_BUS_PCI_MSI, &pci_domain);

which really feels like you're trying to create a namespace that is
parallel to the one defined by the device_node parameter. What I'm
trying to do is to be able to replace the device_node by something more
generic (at the moment, you can either pass a device_node or some token
that the irqdomain subsystem generates for you - see patch #7 for an
example).

You could pass this token to pci_msi_create_irq_domain (which obviously
needs some repainting not to take a device_node), store it in your bus
structure, and perform the lookup based on this value. Or store the
actual domain there, whatever.

What I want to do is really to make this device_node pointer for systems
that do not have a DT node to pass there, which is exactly your case (by
the look of it, the bus number is your identifier of choice, but I
suspect a pointer to an internal structure would be better suited).

	M.
-- 
Jazz is not dead. It just smells funny...
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/