Message-ID: <alpine.DEB.2.22.394.2210171323450.2351079@ubuntu-linux-20-04-desktop>
Date:   Mon, 17 Oct 2022 13:39:35 -0700 (PDT)
From:   Stefano Stabellini <sstabellini@...nel.org>
To:     Oleksandr Tyshchenko <Oleksandr_Tyshchenko@...m.com>
cc:     Stefano Stabellini <sstabellini@...nel.org>,
        Oleksandr Tyshchenko <olekstysh@...il.com>,
        "xen-devel@...ts.xenproject.org" <xen-devel@...ts.xenproject.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        Juergen Gross <jgross@...e.com>
Subject: Re: [PATCH] xen/virtio: Handle PCI devices which Host controller is
 described in DT

On Sat, 15 Oct 2022, Oleksandr Tyshchenko wrote:
> On 13.10.22 03:33, Stefano Stabellini wrote:
> 
> Hello Stefano
> 
> > On Wed, 12 Oct 2022, Oleksandr Tyshchenko wrote:
> >>> On Thu, 6 Oct 2022, Oleksandr Tyshchenko wrote:
> >>>> From: Oleksandr Tyshchenko <oleksandr_tyshchenko@...m.com>
> >>>>
> >>>> Use the same "xen-grant-dma" device concept (based on generic IOMMU
> >>>> device-tree bindings) for the PCI devices behind device-tree based
> >>>> PCI Host controller.
> >>>>
> >>>> Signed-off-by: Oleksandr Tyshchenko <oleksandr_tyshchenko@...m.com>
> >>>> ---
> >>>> Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices
> >>>> on Arm at some point in the future. The Xen toolstack side is not published yet.
> >>>> Here, for PCI devices we use the same way to pass backend domid to the guest as for
> >>>> platform devices.
> >>>>
> >>>> Depends on Juergen's series:
> >>>> https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/20221006071500.15689-1-jgross@suse.com/__;!!GF_29dbcQIUBPA!waOk2Goc7qlhNo5-csRObryil_GzMF_e61EJR501oJ08cH2dnJulsZXWlelBDTBqa63TVoUcWQTB5NecJ1p4xFNgh2_EuA$  [lore[.]kernel[.]org]
> >>>> ---
> >>>>    drivers/xen/grant-dma-ops.c | 51 +++++++++++++++++++++++++++++--------
> >>>>    1 file changed, 41 insertions(+), 10 deletions(-)
> >>>>
> >>>> diff --git a/drivers/xen/grant-dma-ops.c b/drivers/xen/grant-dma-ops.c
> >>>> index ff9be3aff87e..79d13122ec08 100644
> >>>> --- a/drivers/xen/grant-dma-ops.c
> >>>> +++ b/drivers/xen/grant-dma-ops.c
> >>>> @@ -10,6 +10,7 @@
> >>>>    #include <linux/module.h>
> >>>>    #include <linux/dma-map-ops.h>
> >>>>    #include <linux/of.h>
> >>>> +#include <linux/pci.h>
> >>>>    #include <linux/pfn.h>
> >>>>    #include <linux/xarray.h>
> >>>>    #include <linux/virtio_anchor.h>
> >>>> @@ -273,12 +274,28 @@ static const struct dma_map_ops xen_grant_dma_ops = {
> >>>>    	.dma_supported = xen_grant_dma_supported,
> >>>>    };
> >>>>    
> >>>> -static bool xen_is_dt_grant_dma_device(struct device *dev)
> >>>> +static struct device_node *xen_dt_get_node(struct device *dev)
> >>>> +{
> >>>> +	if (dev_is_pci(dev)) {
> >>>> +		struct pci_dev *pdev = to_pci_dev(dev);
> >>>> +		struct pci_bus *bus = pdev->bus;
> >>>> +
> >>>> +		/* Walk up to the root bus to look for PCI Host controller */
> >>>> +		while (!pci_is_root_bus(bus))
> >>>> +			bus = bus->parent;
> >>>> +
> >>>> +		return of_node_get(bus->bridge->parent->of_node);
> >>>> +	}
> >>> Is it possible to have multiple virtio devices under a single virtio-pci
> >>> root complex?
> >> yes
> >>
> >>
> >>> What if virtio-net has the backend in dom0 and
> >>> virtio-block has the backend in dom1?
> >>>
> >>> Or each virtio PCI device shows up under a different PCI root complex?
> >>
> >> Good questions. To be honest, it is not 100% clear to me yet how it is
> >> supposed to be. But let's guess...
> >>
> >> I think that having a PCI Host bridge per virtio-pci device is overkill.
> >>
> >> So, I see two options here:
> >> 1. We provide PCI Host bridge per backends domain to the guest, so each
> >> PCI Host bridge covers only virtio-pci devices whose backends are
> >> running within *the same* domain.
> >> With that we would be able to use property at PCI Host bridge level.
> >>
> >> 2. We provide only a single PCI Host bridge to the guest, so that single
> >> PCI Host bridge covers all virtio-pci devices assigned to this guest.
> >> No matter where the corresponding backends are running (the
> >> virtio-devices under that PCI Host bridge can have the backends in
> >> different domains).
> >> With that we wouldn’t be able to use property at PCI Host bridge level.
> >> And we need a more flexible option(s) to be able distinguish between
> >> virtio-pci devices.
> >>
> >> Taking into account that for virtio-pci on Arm we need to emulate a
> >> specific PCI Host bridge in Xen to intercept the guest PCI config space
> >> accesses
> >> (detect what PCI device is targeted) and forward them to the appropriate
> >> backend (IOREQ Server),
> >> it feels to me that we likely need to go with the second option here
> >> (one PCI host bridge per guest), I may mistake,
> >> but I don’t think that we want to emulate several PCI Host bridges for a
> >> single guest (more code, more resources, etc).
> >>
> >>
> >>
> >>> If we can have multiple virtio PCI devices under a single PCI root
> >>> complex, then I think it would be better to check for a per-device
> >>> property, rather than a single property at the PCI root complex level.
> >> Completely agree.
> >>
> >>
> >>> The first thing that comes to mind is to describe each PCI device under
> >>> the root complex in device tree. Although it is uncommon (usually only
> >>> the PCI root complex is described in device tree), it is possible to
> >>> also describe in device tree all the individual PCI devices under the
> >>> root complex.
> >>>
> >>> Given that the domU device tree is generated by Xen and/or the Xen
> >>> toolstack, it would be easy to arrange for it to happen.
> >> Technically yes. If we decide to provide only a single PCI Host bridge
> >> to the guest, we will have have to deal with the virtio-pci devices with
> >> various backend_domid,
> >> so we can consider using more flexible property
> >> “iommu-map”/”iommu-map-mask” specially introduced for such purposes:
> >> https://urldefense.com/v3/__https://www.kernel.org/doc/Documentation/devicetree/bindings/pci/pci-iommu.txt__;!!GF_29dbcQIUBPA!17Dk_s8c_5QCQDmLM1FYp695LuUMSTrUV6HPfRA9BgmVz9TRRDZBOuzsdZw_c6G0ogky1P11gql9CHAOshukWmv7zWS0dQ$  [kernel[.]org]
> >> I think, with that we would be able to describe, i.e that virtio-pci
> >> device A connects to stream_id (backend_domid) X and virtio-pci device B
> >> to stream_id Y,
> >> and virtio-pci device C to nothing (so is not required to use grants),
> >> unless I missed something.
> >>
> >> I have looked at it and I don’t see at the moment why the idea wouldn’t
> >> work, but I haven’t experimented with that yet in such context.
> > I think it would work too
> 
> 
> I have experimented with that, it works. And I have already created a patch.
> 
> https://lore.kernel.org/xen-devel/20221015153409.918775-1-olekstysh@gmail.com/
> 
> 
> What I was thinking is although generic PCI-IOMMU bindings 
> ("xen-grant-dma") wouldn't likely be suitable for *future* hotplug support,
> it would allow us to have the working solution on Arm (with a minimal 
> changes, only drivers/xen/grant-dma-ops.c is touched)
> at least for PCI devices which are known at the domain creation time. Of 
> course, this needs the proper support in the toolstack.

Yeah, it is hard to argue against this, as we don't have a good
alternative :-)


> >>> That would solve the issue as far as I can tell, but I worry it might
> >>> not be a good idea because if we rely on the per-device device tree node
> >>> to be present then it becomes harder to implement virtio hotplug
> >>> (Virtio hotplug is important to add dom0less support.)
> >>>
> >>> Let's say that we create a dom0less domU with an emulated PCI root
> >>> complex without any devices under it, then after Dom0 is fully booted,
> >>> we add a virtio-net emulated device. How do we tell the guest what is
> >>> the backend domain id?
> >>>
> >>> Device tree and other firmware tables are not relevant anymore.
> >>>
> >>> We could reuse a PCI config space register to expose the backend id.
> >>> However this solution requires a backend change (QEMU) to expose the
> >>> backend id via an emulated register for each emulated device.
> >>>
> >>> To avoid having to introduce a special config space register in all
> >>> emulated PCI devices (virtio-net, virtio-block, etc) I wonder if we
> >>> could add a special PCI config space register at the emulated PCI Root
> >>> Complex level.
> >>>
> >>> Basically the workflow would be as follow:
> >>>
> >>> - Linux recognizes the PCI Root Complex as a Xen PCI Root Complex
> >>> - Linux writes to special PCI config space register of the Xen PCI Root
> >>>     Complex the PCI device id (basically the BDF)
> >>> - The Xen PCI Root Complex emulated by Xen answers by writing back to
> >>>     the same location the backend id (domid of the backend)
> >>> - Linux reads back the same PCI config space register of the Xen PCI
> >>>     Root Complex and learn the relevant domid
> >>>
> >>> What do you think?
> >>
> >> I think the idea sounds indeed interesting and would probably work, but
> >> would require guest modifications other than just in drivers/xen (and
> >> likely the specification changes as well).
> >> Which ideally of course should be avoided.
> >> Also I was thinking it would be nice not to diverge much between
> >> communicating the backend_domid for platform and PCI devices on Arm with
> >> device tree.
> >>
> >> If we managed to re-use generic IOMMU device-tree bindings for
> >> virtio-mmio, we would likely be able to re-use PCI-IOMMU device-tree
> >> bindings for virtio-pci,
> >> at least for boot PCI devices (which are known at the domain creation time).
> >> The more, the bindings is already present:
> >> https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/devicetree/bindings/iommu/xen,grant-dma.yaml__;!!GF_29dbcQIUBPA!17Dk_s8c_5QCQDmLM1FYp695LuUMSTrUV6HPfRA9BgmVz9TRRDZBOuzsdZw_c6G0ogky1P11gql9CHAOshukWmvpfF8GmA$  [git[.]kernel[.]org]
> > I think using a special config space register in the root complex would
> > not be terrible in terms of guest changes because it is easy to
> > introduce a new root complex driver in Linux and other OSes. The root
> > complex would still be ECAM compatible so the regular ECAM driver would
> > still work. A new driver would only be necessary if you want to be able
> > to access the special config space register.
> 
> 
> This needs an additional investigation.
> 
> 
> >
> > That said, for sure the fewer changes to the guest the better and I like
> > the existing xen,grant-dma binding.
> >
> >
> >> Regarding the hotplug devices, yes it is a valid use-case which should
> >> be also supported with virtio-pci, I assume the Xenstore could be
> >> re-used for that purpose if it would be available.
> >> The Xenstore is available with the usual Dom0 and toolstack, is it
> >> available with dom0less?
> > Xenstore is available for dom0less if we have a dom0 running with
> > xenstored as one of the dom0less domains. We currently rely on it for
> > Xen PV drivers with dom0less.  After dom0 is fully booted, we use "xl
> > network-attach" to create a vif interface dynamically in the domU.
> >
> > That is why I was thinking of using virtio hotplug to solve the same
> > problem with virtio, I was imagining that after dom0 is fully booted we
> > would do "xl virtio-attach" and create a new virtio interface in the
> > domU. But I cannot see an easy way to make virtio hotplug work together
> > with the xen,grant-dma bindings. I think it would be better if we find a
> > way to make it work without xenstore (because xenstore would be a
> > safety-certification dependency).
> 
> I got your concern regarding the usage of xenstore in general.
> Also I got that hotplug is the only way to get virtual devices (either 
> PV or virtio) working for the dom0less system, is my understanding correct?

In a dom0less system domains boot in parallel. The backend is typically
in a larger and slower domain to boot (Linux). So to solve the problem
with Xen PV drivers, we hotplug Xen PV devices after dom0 is booted ("xl
network-attach" for instance).

With virtio, we could either do the same (use virtio hotplug) or find a
way to tell the frontend to delay initialization. The point is that we
don't want to frontend to try to access backend resources before the
backend is up and running.


> The virtio hotplug could *probably* work together with the 
> "xen,grant-dma" bindings with some prerequisites (backend domid should 
> be known in advance),
> but I am not quite as I am not too familiar with dom0less details. But 
> anyway, I will try to describe it...
> 
> As far as I know the dom0less DomUs are configured from the device-tree. 
> So the configuration is known and prepared beforehand.
> I may guess that we know in what DomU1 we are going to run the virtio 
> backends and what PCI devices we are going to assign to DomU2. So when 
> Xen generating device-tree
> for DomU2 it could create proper iommu-map for the PCI Host bridge node. 
> Although for this to work we would need an ability to configure domain 
> ID (d->domain_id) via device-tree
> and likely reserve some range of domain IDs (to not cross with 
> max_init_domid). But, it wouldn't be 100% hotplug then.

That's fine it doesn't have to be 100% hotplug. In reality, this is a
static configuration so we know all the information beforehand (which VM
is the backend, which is the frontend, which devices are
shared/emulated). The only issue is that we need a way to tell VM2 to
wait for the backend in VM1 to come online. But the device tree could
contain all information from the start.
 

> > Maybe we need to think outside the box and find another solution that
> > doesn't rely on hotplug.
> >
> > For instance, let's say that we expose the virtio devices in device tree
> > in a dom0less configuration too but with status = "disabled". When dom0
> > (or backend domain) is up and running it can signal that it is ready.
> > Maybe if we had a special Xen-specific PCI Root Complex driver in the
> > guest, it could wait for the Xen signal and then continue PCI probing at
> > that point honoring xen,grant-dma bindings if present in device tree
> > even if the devices had status = "disabled" initially.
> >
> > It looks like that would require many guest changes unfortunately.
> 
> 
> It looks like yes, also you mentioned "it can signal that it is ready",
> the question is by what means (xenstore would be a god fit here...)?

Maybe xenstore, yes. The problem is that we want something that works
with minimal drivers changes, and the problem is that if we present the
virtio devices in device tree from boot, the drivers will try to probe
them immediately. We need a way to delay that.


> And I haven't seen that virtio-pci devices are described in device-tree 
> somewhere, only generic PCI host bridge node
> is described. The virtio-pci devices will be detected the same way as 
> usual PCI devices during boot. Unless I missed something.

Yes exactly, and that is the problem. How do we make those driver "wait"
before probing.


> Regarding the virtio-mmio (platform) devices, yes, we could expose them 
> with status "disabled", and they won't get probed by default.
> To be honest, I have experimented with that, when I was thinking of 
> possible hotplug for virtio-mmio devices (I know, this sounds uncommon 
> and strange).
> I used Linux feature (CONFIG_OF_DYNAMIC, overlays) to update the 
> device-tree on running guest, so the toolstack initially inserts 
> virtio-mmio device nodes for non-boot devices
> with status "disabled", and at the runtime, once we receive an event for 
> example, we change the status to "ok" and the corresponding virtio-mmio 
> device gets probed.
> But again, it is not a 100% hotplug, as we need to pre-allocate memory 
> range and interrupt in advance (when generating guest device tree).

Actually this is really cool! Does it work? It doesn't matter to me if
the virtio devices are pci or mmio as long as we can solve the "wait"
problem. So this could be a good solution.


> > As an alternative I wonder, given that Xen emulates the PCI root
> > complex, if we can reuse one of the PCI link up/down delays for this
> > instead, like "pcie_wait_for_link". It looks like the wait time is in
> > millisec while we would need potentially several seconds here but it
> > might be possible?
> 
> I am not sure that I understand this alternative idea.

The PCI subsystem has already a concept of wait times. Just have a look
at pcie_wait_for_link under drivers/pci. The question was whether we can
find a way to reuse one of the existing wait times to deal with our
"wait" problem.

 
> >
> > Other ideas?
> 
> Another (crazy?) idea is to reuse CONFIG_XEN_VIRTIO_FORCE_GRANT for 
> dom0less system (I mean without "xen,grant-dma" bindings at all).
> If virtio backends are always going to run in Dom0 when we have it up 
> and running, then it should work as domid == 0 is reserved for Dom0.
> If there is a need to run virtio backends in other *backend* domain (for 
> the domain ID to be always known we could reserve an ID for it, so it 
> would be a const value),
> we could probably introduce something configurable like 
> CONFIG_XEN_VIRTIO_FORCE_GRANT_BE_DOMID with 0 by default (or cmd line 
> option).

The problem in a dom0less system is not much how to tell which is the
backend domid, because that is known in advance and could be added to
device tree at boot somehow. The issue is how to ask the frontend to
"wait" and then how to tell the frontend to "proceed" after the backend
comes online.

Powered by blists - more mailing lists