Re: [PATCH 1/2] s390x/pci: add support for guests that request direct mapping

[David Hildenbrand]

On 09.12.24 22:45, Matthew Rosato wrote:
> On 12/9/24 4:01 PM, David Hildenbrand wrote:
> > On 09.12.24 20:29, Matthew Rosato wrote:
> >
> > Hi,
> >
> > Trying to wrap my head around that ... you mention that "pin the entirety of guest 
> > memory".
> >
> > Do you mean that we will actually end up longterm pinning all guest RAM in the 
> > kernel, similar to what vfio ends up doing?
>
> Yes.  Actually, the usecase here is specifically PCI passthrough via vfio-pci 
> on s390.  Unlike other platforms, the default s390 approach only pins on-demand 
> and doesn't longterm pin all guest RAM, which is nice from a memory footprint 
> perspective but pays a price via all those guest-2 RPCIT instructions.  The 
> goal here is now provide the optional alternative to longterm pin like other 
> platforms.

Okay, thanks for confirming. One more question: who will trigger this 
longterm-pinning? Is it vfio?

(the code flow from your code to the pinning code would be nice)

> > In that case, it would be incompatible with virtio-balloon (and without 
> > modifications with upcoming virtio-mem). Is there already a mechanism in place 
> > to handle that -- a call  to ram_block_discard_disable() -- or even a way to 
> > support coordinated discarding of RAM (e.g., virtio-mem + vfio)?
>
> Good point, should be calling add ram_block_discard_disable(true) when set 
> register + a corresponding (false) during deregister...  Will add for v2.
>
> As for supporting coordinated discard, I was hoping to subsequently look at 
> virtio-mem for this.

As long as discarding is blocked for now, we're good. To support it, the 
RAMDiscardManager would have to be wired up, similar to vfio.

I think the current way of handling it via

+    IOMMUTLBEvent event = {
+        .type = IOMMU_NOTIFIER_MAP,
+        .entry = {
+            .target_as = &address_space_memory,
+            .translated_addr = 0,
+            .perm = IOMMU_RW,
+        },
+    };


Is probably not ideal: it cannot cope with memory holes (which 
virtio-mem would create).

Likely, you'd instead want an address space notifier, and really only 
map the memory region sections you get notified about.

There, you can test for RAMDiscardManager and handle it like vfio does.

--
Cheers,

David / dhildenb