From: Igor Mammedov <imammedo@redhat.com>
To: Peter Xu <peterx@redhat.com>
Cc: Peter Maydell <peter.maydell@linaro.org>,
"Daniel P . Berrange" <berrange@redhat.com>,
Eduardo Habkost <ehabkost@redhat.com>,
"Michael S . Tsirkin" <mst@redhat.com>,
Jason Wang <jasowang@redhat.com>,
David Hildenbrand <david@redhat.com>,
qemu-devel@nongnu.org, Markus Armbruster <armbru@redhat.com>,
Shannon Zhao <shannon.zhaosl@gmail.com>,
Alex Williamson <alex.williamson@redhat.com>,
Paolo Bonzini <pbonzini@redhat.com>,
Eric Auger <eric.auger@redhat.com>,
David Gibson <david@gibson.dropbear.id.au>
Subject: Re: [PATCH 8/8] x86-iommu: Fail early if vIOMMU specified after vfio-pci
Date: Tue, 26 Oct 2021 17:11:39 +0200 [thread overview]
Message-ID: <20211026171139.30bf0c80@redhat.com> (raw)
In-Reply-To: <YXIehUnQUZ/Odgw7@xz-m1.local>
On Fri, 22 Oct 2021 10:14:29 +0800
Peter Xu <peterx@redhat.com> wrote:
> Hi, Alex,
>
> On Thu, Oct 21, 2021 at 04:30:39PM -0600, Alex Williamson wrote:
> > On Thu, 21 Oct 2021 18:42:59 +0800
> > Peter Xu <peterx@redhat.com> wrote:
> >
> > > Scan the pci bus to make sure there's no vfio-pci device attached before vIOMMU
> > > is realized.
> >
> > Sorry, I'm not onboard with this solution at all.
> >
> > It would be really useful though if this commit log or a code comment
> > described exactly the incompatibility for which vfio-pci devices are
> > being called out here. Otherwise I see this as a bit of magic voodoo
> > that gets lost in lore and copied elsewhere and we're constantly trying
> > to figure out specific incompatibilities when vfio-pci devices are
> > trying really hard to be "just another device".
>
> Sure, I can enrich the commit message.
>
> >
> > I infer from the link of the previous alternate solution that this is
> > to do with the fact that vfio devices attach a memory listener to the
> > device address space.
>
> IMHO it's not about the memory listeners, I think that' after vfio detected
> some vIOMMU memory regions already, which must be based on an vIOMMU address
> space being available. I think the problem is that when realize() vfio-pci we
> fetch the dma address space specifically for getting the vfio group, while that
> could happen too early, even before vIOMMU is created.
>
> > Interestingly that previous cover letter also discusses how vdpa devices
> > might have a similar issue, which makes it confusing again that we're calling
> > out vfio-pci devices by name rather than for a behavior.
>
> Yes I'll need to see whether this approach will be accepted first. I think
> similar thing could help VDPA but it's not required there because VDPA has
> already worked around using pci_device_iommu_address_space(). So potentially
> the only one to "fix" is the vfio-pci device using along with vIOMMU, when the
> device ordering is specified in the wrong order. I'll leave the VDPA problem
> to Jason to see whether he prefers keeping current code, or switch to a simpler
> one. That should be after this one.
>
> >
> > If the behavior here is that vfio-pci devices attach a listener to the
> > device address space, then that provides a couple possible options. We
> > could look for devices that have recorded an interest in their address
> > space, such as by setting a flag on PCIDevice when someone calls
> > pci_device_iommu_address_space(), where we could walk all devices using
> > the code in this series to find a device with such a flag.
>
> Right, we can set a flag for all the pci devices that needs to consolidate
> pci_device_iommu_address_space() result, however then it'll be vfio-pci only so
> far. Btw, I actually proposed similar things two months ago, and I think Igor
> showed concern on that flag being vague on meaning:
(1)
> https://lore.kernel.org/qemu-devel/20210906104915.7dd5c934@redhat.com/
>
> > > Does it need to be a pre_plug hook? I thought we might just need a flag in the
> > > pci device classes showing that it should be after vIOMMUs, then in vIOMMU
> > > realize functions we walk pci bus to make sure no such device exist?
> > >
> > > We could have a base vIOMMU class, then that could be in the realize() of the
> > > common class.
> >
> > We basically don't know if device needs IOMMU or not and can work
> > with/without it just fine. In this case I'd think about IOMMU as board
> > feature that morphs PCI buses (some of them) (address space, bus numers, ...).
> > So I don't perceive any iommu flag as a device property at all.
> >
> > As for realize vs pre_plug, the later is the part of abstract realize
> > (see: device_set_realized) and is already used by some PCI infrastructure:
> > ex: pcie_cap_slot_pre_plug_cb/spapr_pci_pre_plug
>
> I still think that flag will work, that flag should only shows "whether this
> device needs to be specified earlier than vIOMMU", but I can get the point from
> Igor that it's at least confusing on what does the flag mean.
> Meanwhile I
> don't think that flag will be required, as this is not the first time we name a
> special device in the code, e.g. pc_machine_device_pre_plug_cb().
> intel_iommu.c has it too upon vfio-pci already on making sure caching-mode=on
> in vtd_machine_done_notify_one().
I pointed to specifically to _pre_plug() handler and not as
implemented here in realize().
Reasoning behind it is that some_device_realize() should not poke
into other devices, while pc_machine_device_pre_plug_cb() is
part of machine code can/may legitimately access its child devices and verify/
configure them. (Hence I'd drop my suggested-by with current impl.)
> If Igor is okay with adding such a flag for PCIDevice class, I can do that in
> the new version. I don't have a strong opinion on this.
Also, I've suggested to use pre_plug only as the last resort in case
vfio-pci can't be made independent of the order (see [1] for reset time
suggestion).
So why 'reset' approach didn't work out?
(this need to be cover letter/commit message as a reason why
we are resorting to a hack)
>
> Thanks,
>
next prev parent reply other threads:[~2021-10-26 15:30 UTC|newest]
Thread overview: 44+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-10-21 10:42 [PATCH 0/8] pci/iommu: Fail early if vfio-pci detected before vIOMMU Peter Xu
2021-10-21 10:42 ` [PATCH 1/8] pci: Define pci_bus_dev_fn type Peter Xu
2021-10-21 10:53 ` David Hildenbrand
2021-10-21 11:15 ` Eric Auger
2021-10-22 2:16 ` Peter Xu
2021-10-21 11:36 ` Philippe Mathieu-Daudé
2021-10-21 10:42 ` [PATCH 2/8] pci: Export pci_for_each_device_under_bus*() Peter Xu
2021-10-21 10:54 ` David Hildenbrand
2021-10-21 11:32 ` Eric Auger
2021-10-21 11:39 ` Philippe Mathieu-Daudé
2021-10-21 10:42 ` [PATCH 3/8] pci: Use pci_for_each_device_under_bus*() Peter Xu
2021-10-21 10:56 ` David Hildenbrand
2021-10-21 11:34 ` Eric Auger
2021-10-22 2:19 ` Peter Xu
2021-10-21 10:42 ` [PATCH 4/8] pci: Define pci_bus_fn/pci_bus_ret_fn type Peter Xu
2021-10-21 10:57 ` David Hildenbrand
2021-10-21 11:37 ` Eric Auger
2021-10-21 11:44 ` Philippe Mathieu-Daudé
2021-10-21 12:54 ` Philippe Mathieu-Daudé
2021-10-22 2:24 ` Peter Xu
2021-10-21 10:42 ` [PATCH 5/8] pci: Add pci_for_each_root_bus() Peter Xu
2021-10-21 11:00 ` David Hildenbrand
2021-10-21 12:22 ` Eric Auger
2021-10-25 13:16 ` Michael S. Tsirkin
2021-10-28 2:56 ` Peter Xu
2021-10-21 10:42 ` [PATCH 6/8] pci: Use pci_for_each_root_bus() in current code Peter Xu
2021-10-21 11:06 ` David Hildenbrand
2021-10-21 12:28 ` Eric Auger
2021-10-21 10:42 ` [PATCH 7/8] pci: Add pci_for_each_device_all() Peter Xu
2021-10-21 10:54 ` Michael S. Tsirkin
2021-10-22 2:33 ` Peter Xu
2021-10-22 8:43 ` Michael S. Tsirkin
2021-10-25 12:57 ` Peter Xu
2021-10-25 13:13 ` Michael S. Tsirkin
2021-10-21 10:42 ` [PATCH 8/8] x86-iommu: Fail early if vIOMMU specified after vfio-pci Peter Xu
2021-10-21 10:49 ` Michael S. Tsirkin
2021-10-21 12:38 ` Eric Auger
2021-10-22 2:37 ` Peter Xu
2021-10-21 22:30 ` Alex Williamson
2021-10-22 2:14 ` Peter Xu
2021-10-26 15:11 ` Igor Mammedov [this message]
2021-10-26 15:38 ` Alex Williamson
2021-10-27 8:30 ` Peter Xu
2021-10-28 2:30 ` Peter Xu
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20211026171139.30bf0c80@redhat.com \
--to=imammedo@redhat.com \
--cc=alex.williamson@redhat.com \
--cc=armbru@redhat.com \
--cc=berrange@redhat.com \
--cc=david@gibson.dropbear.id.au \
--cc=david@redhat.com \
--cc=ehabkost@redhat.com \
--cc=eric.auger@redhat.com \
--cc=jasowang@redhat.com \
--cc=mst@redhat.com \
--cc=pbonzini@redhat.com \
--cc=peter.maydell@linaro.org \
--cc=peterx@redhat.com \
--cc=qemu-devel@nongnu.org \
--cc=shannon.zhaosl@gmail.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).