On 04/04/2017 02:59 PM, Eduardo Habkost wrote:
On Tue, Apr 04, 2017 at 09:02:28AM +0200, Alexander Graf wrote:

On 04.04.17 08:58, Thomas Huth wrote:
On 04.04.2017 08:53, Alexander Graf wrote:

On 03.04.17 23:00, Eduardo Habkost wrote:
On Mon, Apr 03, 2017 at 10:15:44PM +0200, Alexander Graf wrote:

On 03.04.17 22:10, Eduardo Habkost wrote:
On Mon, Apr 03, 2017 at 08:49:16PM +0100, Peter Maydell wrote:
On 1 April 2017 at 01:46, Eduardo Habkost <ehabk...@redhat.com> wrote:
commit 33cd52b5d7b9adfd009e95f07e6c64dd88ae2a31 unset
cannot_instantiate_with_device_add_yet in TYPE_SYSBUS, making
all kinds of untested devices available to -device and
device_add.

The problem with that is: setting has_dynamic_sysbus on a
machine-type lets it accept all the 288 sysbus device types we
have in QEMU, and most of them were never meant to be used with
-device. That's a lot of untested code.

Fortunately today we have just a few has_dynamic_sysbus=1
machines: virt, pc-q35-*, ppce500, and spapr.

virt, ppce500, and spapr have extra checks to ensure just a few
device types can be instantiated:

* virt supports only TYPE_VFIO_CALXEDA_XGMAC, TYPE_VFIO_AMD_XGBE.
* ppce500 supports only TYPE_ETSEC_COMMON.
* spapr supports only TYPE_SPAPR_PCI_HOST_BRIDGE.

q35 has no code to block unsupported sysbus devices, however, and
accepts all device types. Fortunately, only the following 20
device types are compiled into the qemu-system-x86_64 and
qemu-system-i386 binaries:

* allwinner-ahci
* amd-iommu
* cfi.pflash01
* esp
* fw_cfg_io
* fw_cfg_mem
* generic-sdhci
* hpet
* intel-iommu
* ioapic
* isabus-bridge
* kvmclock
* kvm-ioapic
* kvmvapic
* SUNW,fdtwo
* sysbus-ahci
* sysbus-fdc
* sysbus-ohci
* unimplemented-device
* virtio-mmio

Instead of requiring each machine-type with has_dynamic_sysbus=1
to implement its own mechanism to block unsupported devices, we
can use the user_creatable flag to ensure we won't let the user
plug anything that will never work.
How does this work? Which devices can be dynamically
plugged is machine dependent. You can't dynamically-plug
an intel-iommu on the ARM virt board, and you can't
dynamically-plug the vfio-calxeda-xgmac on the spapr
board, and so on. So I don't see how we can just have
a flag on the device itself that controls whether
it can be dynamically plugged.

So I'm definitely coming around to the opinion that
it's just a bug in the q35 board that it doesn't have
any device whitelisting, and we should fix that.
OK, let's assume q35 must implement a whitelist:

To build that whitelist, we need to be able to know what should
be in the whitelist, or not. And nobody knew for sure what was
user-creatable in q35 by accident, and what was really supposed
to be user-creatable in q35. See the "q35 and sysbus devices"
thread I started ~2 weeks ago.

Building a q35 whitelist will be much easier if make
sys-bus-devices non-user-creatable by default.
So why are they user creatable in the first place?

We used to have boards that were dynamic sysbus aware (ppce500, virt)
that
allowed dynamic creation and every other board did not. I don't
remember the
exact mechanism behind it though.

When did that behavior change? It sounds like a regression somewhere.
See patch description:

commit 33cd52b5d7b9adfd009e95f07e6c64dd88ae2a31 unset
cannot_instantiate_with_device_add_yet in TYPE_SYSBUS,
Note that the commit above is not a regression[1] (because q35
didn't have has_dynamic_sysbus=1 yet), but having sysbus devices
have cannot_instantiate_with_device_add_yet=false/user_creatable=true
by default makes it harder to build the whitelist for q35 (or
other machines that will have has_dynamic_sysbus=1 in the
future).
I seem to miss the bigger picture.

Why would anyone set has_dynamic_sysbus=1 in a board file without an
explicit whitelist? The whitelist is *not* device specific. It's board
specific, because your board needs to know how to wire up a device and
how to expose the fact that it exists to the OS.

So the real bug is that someone set has_dynamic_sysbus=1 in q35 without
implementing all of the dynamic sysbus logic, no?
According to commit bf8d492405feaee2c1685b3b9d5e03228ed3e47f this was
just introduced for allowing the "intel-iommu" device, so I guess this
is the device that we want to have in a whitelist there?
If you want a dynamic intel-iommu, you also need to have dynamic ACPI
generation to create DMAR tables the OS can use to drive the IOMMU, no? At
last at that point, someone should have realized that a "whitelist" is
mandatory.

But yes, please just only add a whitelist for intel-iommu to the q35 board.
That is the real bug.
Look, I don't disagree that we need a whitelist on q35. But I
don't know why you assume it is as simple as adding intel-iommu.

So why has intel-iommu worked in the first place? Who wired up its memory regions? Who wired up fault IRQs? Who added the DMAR table to ACPI when it was created?

We *don't know* what should be on q35 whitelist until we review
each device that is accepted by q35 today, and make sure it
really is not supposed to be supported on -device.

If in doubt, not a single sysbus device should have ever worked without explicit code around it. Really :).

Making has_dynamic_sysbus/user_creatable consistent on sysbus
devices helps on that. It is not necessary nor sufficient to fix
q35, that's true, but it helps *a lot*.

See, after I start this series, I already found two exceptions to
the "just add intel-iommu" rule:

1) amd-iommu
2) xen-backend

I'm not sure yet if we have others, until people review the other
"Remove user_creatable from <device>" patches in this
series.

Same question as above there. How do they get mapped? How does the OS learn they exist?


The basic idea behind dynamic sysbus is that you move the responsibility of
sysbus spawnability checks from the sysbus layer to the board file. If the
board file ignores to do those checks, it's at fault.
I think I will do this:

I will submit v2 of this thread pretending it is just going to
fix the "info qdm" regression introduced by commit
bf8d492405feaee2c1685b3b9d5e03228ed3e47f, and remove any mention
of the q35 bug from the series and patch description.

I hopt this will make us stop diverging the discussion to "you
should add a whitelist to q35 first", and stop ignoring that:
1) cannot_instantiate_with_device_add_yet is being set
    incorrectly on the sysbus device classes and that's a bad
    idea;
2) cannot_instantiate_with_device_add_yet is an awful name.


We need to make sure that the board has control over which devices are spawnable. I fail to see how this series helps or achieves that, but I'm happy to learn.



Alex



Reply via email to