[libvirt] Questions about function virPCIDeviceIsBehindSwitchLackingACS in virpci.c

[Alex Williamson]

On Wed, 20 Sep 2017 21:39:55 -0400
Laine Stump <laine at laine.org> wrote:

> On 09/20/2017 09:25 PM, Wuzongyong (Euler Dept) wrote:
> >> -----Original Message-----
> >> From: sendmail [mailto:justsendmailnothingelse at gmail.com] On Behalf Of
> >> Laine Stump
> >> Sent: Thursday, September 21, 2017 8:57 AM
> >> To: libvirt-list at redhat.com
> >> Cc: Wuzongyong (Euler Dept) <cordius.wu at huawei.com>; Wanzongshun
> >> (Vincent) <wanzongshun at huawei.com>; Alex Williamson
> >> <Alex.Williamson at redhat.com>
> >> Subject: Re: [libvirt] Questions about function
> >> virPCIDeviceIsBehindSwitchLackingACS in virpci.c
> >>
> >> On 09/18/2017 09:24 PM, Wuzongyong (Euler Dept) wrote:
> >> Hi,
> >>
> >> In function virPCIDeviceIsBehindSwitchLackingACS, I noticed that(line 8):
> >>
> >> 1    if (virPCIDeviceGetParent(dev, &parent) < 0)
> >> 2        return -1;
> >> 3    if (!parent) {
> >> 4        /* if we have no parent, and this is the root bus, ACS doesn't come
> >> 5         * into play since devices on the root bus can't P2P without going
> >> 6         * through the root IOMMU.
> >> 7         */
> >> 8        if (dev->address.bus == 0) {
> >> 9            return 0;
> >> 10        } else {
> >> 11            virReportError(VIR_ERR_INTERNAL_ERROR,
> >> 12                           _("Failed to find parent device for %s"),
> >> 13                           dev->name);
> >> 14            return -1;
> >> 15        }
> >> 16    }
> >>
> >> Why we just return 0 only if device’s bus is 0?
> >> In my server, I can see a root bus which bus number is greater than 0, see the
> >> results(just a part) after I run lspci -t:
> >>
> >> +-[0000:80]-+-02.0-[81-83]--+-00.0
> >> |           |               \-00.1
> >> |           +-05.0
> >> |           +-05.1
> >> |           +-05.2
> >> |           \-05.4
> >> +-[0000:7f]-+-08.0
> >> |           +-08.2
> >> |           +-08.3
> >> |           + . . .
> >> |           \-1f.2
> >> \-[0000:00]-+-00.0
> >>              +-01.0-[01]----00.0
> >>              +-02.0-[02]--+-00.0
> >>              |            +-00.1
> >>              |            +-00.2
> >>              |            \-00.3
> >>              +-02.2-[03]--
> >>              +-03.0-[04-0b]----00.0-[05-0b]--+-08.0-[06-08]----00.0
> >>              |                               \-10.0-[09-0b]----00.0
> >>              +-05.0
> >>              +-05.1
> >>              +-05.2
> >>              +-05.4
> >>              +-11.0
> >>              +-11.4
> >>              +-16.0
> >>              +-16.1
> >>              +-1a.0
> >>
> >> If I assign the device 0000:81:00.0 to a VM, I get “Failed to find parent
> >> device”.
> >> I think I should get no error with return value 0 just like bus number is 0,
> >> because
> >> bus 80 is the root bus as well in my case.
> >>
> >> In the <<Intel C610 Series Chipset and Intel X99 Chipset Platform Controller  
> >> Hub(PCH)>>  
> >> Datasheet, I found that(Chapter 9.1):
> >>      For some server platforms, it may be desirable to have multiple PCHs in
> >> the system
> >>      Which means some PCH’s may reside on a bus greater than 0.
> >>
> >> So, is this a bug?
> >>
> >> My memory is that if you're using VFIO for device assignment, all that
> >> checking should be performed by VFIO, and libvirt shouldn't be checking for
> >> ACS at all. (Alex, can you confirm or refute this?)

Yes, the libvirt ACS checking was never complete and is easily
bypassed, vfio handles this now.

> >> virPCIDeviceIsBehindSwitchLackingACS() is only called from
> >> virPCIDeviceIsAssignable(), and that function is only called if the device's
> >> stubDriver is set to something other than "vfio-pci" (see step 1 in
> >> virHostdevPreparePCIDevices()). Digging deeper, it looks like the device's
> >> stubDriver is set by virHostdevGetPCIHostDeviceList(), which appears to set
> >> it to vfio-pci if the backend is specified as vfio (i.e. <driver name='vfio'/> in
> >> the libvirt XML. This *should* be the default setting!)
> >>
> >> Since I assume you're not using RHEL6, meaning that you will be using VFIO
> >> by default, not legacy KVM assignment.
> >>
> >> TL;DR I think the bug here is that the ...CheckACS function is being called *at
> >> all*. That code path should be completely obsolete.  
> > Yeah, I'm using the legacy KVM assignment just for check if virPCIDeviceIsBehindSwitchLackingACS
> > do right thing.
> > So, do you mean legacy KVM assignment code path in libvirt is not maintained any more?  
> 
> 
> Yes, that is correct. I'm pretty sure it's also not maintained in the
> kernel anymore either. As a matter of fact I had thought that legacy KVM
> device assignment had been disabled in the kernel for a very long time
> (it's disabled downstream in RHEL7). As far as I know, there is no
> reason at all that anyone should be using legacy KVM device assignment
> unless they're on an old 2.6.x kernel (like RHEL6/CentOS6).I'm fairly
> certain nobody has *intentionally* used/tested legacy KVM device
> assignment with anything other than the extremely old downstream builds
> of libvirt used by RHEL6/CentOS6 in a very long time. Some day when we
> decide that we no longer want to support building new upstream libvirt
> on those old systems, we should remove the code for it.
> 

Legacy KVM device assignment doesn't exist in the kernel anymore, it
was deprecated in 4.2 and was finally removed in 4.12.  It's long past
time to move to vfio.  I also wouldn't classify legacy KVM device
assignment as "supported" in QEMU either, bugs are not likely to get
fixed and before long we'll probably be looking for an excuse to remove
it from that code base as well.  Thanks,

Alex