[Qemu-devel] Re: [PATCH 4/5] Make MMIO address page aligned in guest.

[Gleb Natapov]

On Mon, Oct 12, 2009 at 11:43:35AM +0200, Michael S. Tsirkin wrote:
> > > Try hot-plugging the device instead of have it present on boot.
> > > Patching BIOS won't help then, will it?  So my question is, if we need
> > > to handle this in qemu, is it worth it to do it in kvm as well?
> > > 
> > It depend how linux assign mmio address to hot pluggable devices. How
> > can you be sure a device driver continue working if you'll misrepresent
> > BAR size BTW?
> 
> Yes, this adds yet another way for device to discover it's running in a
> VM, so this might break some drivers.  If we see many of these in
> practice, we can try adding a PCI-to-PCI bridge with some dummy devices
> behind it to the picture, to increase the chances to get a dedicated
> memory page.
> 
Go ahead. But why all this churn instead of asking linux to align pci
resources and do the same in a bios.

> > > > > > As it stands this
> > > > > > patch is in kvm's bios and is required for assigned devices to work
> > > > > > for some devices, so moving to seabios without this patch will 
> > > > > > introduce
> > > > > > a regression.
> > > > > 
> > > > > I have a question here: if kvm maps a full physical page
> > > > > into guest memory, while device only uses part of the page,
> > > > > won't that mean that guest is granted access outside the
> > > > > device, which it should not have?
> > > > And how is real HW different? It maps a full physical page into OS
> > > > memory even if BAR is smaller then page and grants OS access to
> > > > unassigned mmio region. Access unassigned mmio region shouldn't cause
> > > > any trouble, doesn't it?
> > > 
> > > Unassigned - typically no, but there can be another device there, or a RAM
> > > page.  It is different on real hardware where OS has access to all RAM 
> > > and all
> > > devices, anyway.
> > > 
> > > Here's an example from my laptop:
> > > 
> > > 00:03.0 Communication controller: Intel Corporation Mobile 4 Series 
> > > Chipset MEI Controller (rev 07)
> > >         Subsystem: Lenovo Device 20e6
> > >         Control: I/O- Mem+ BusMaster+ SpecCycle- MemWINV- VGASnoop- 
> > > ParErr- Stepping- SERR- FastB2B- DisINTx-
> > >         Status: Cap+ 66MHz- UDF- FastB2B- ParErr- DEVSEL=fast >TAbort- 
> > > <TAbort- <MAbort- >SERR- <PERR- INTx+
> > >         Latency: 0
> > >         Interrupt: pin A routed to IRQ 11
> > >         Region 0: Memory at fc226800 (64-bit, non-prefetchable) [size=16]
> > >         Capabilities: <access denied>
> > > 
> > > ...
> > > 
> > > 00:1f.2 SATA controller: Intel Corporation ICH9M/M-E SATA AHCI Controller 
> > > (rev 03) (prog-if 01 [AHCI 1.0])
> > >         Subsystem: Lenovo Device 20f8
> > >         Control: I/O+ Mem+ BusMaster+ SpecCycle- MemWINV- VGASnoop- 
> > > ParErr- Stepping- SERR- FastB2B- DisINTx+
> > >         Status: Cap+ 66MHz+ UDF- FastB2B+ ParErr- DEVSEL=medium >TAbort- 
> > > <TAbort- <MAbort- >SERR- <PERR- INTx-
> > >         Latency: 0
> > >         Interrupt: pin B routed to IRQ 28
> > >         Region 0: I/O ports at 1c48 [size=8]
> > >         Region 1: I/O ports at 183c [size=4]
> > >         Region 2: I/O ports at 1c40 [size=8]
> > >         Region 3: I/O ports at 1838 [size=4]
> > >         Region 4: I/O ports at 1c20 [size=32]
> > >         Region 5: Memory at fc226000 (32-bit, non-prefetchable) [size=2K]
> > >         Capabilities: <access denied>
> > >         Kernel driver in use: ahci
> > > 
> > > In this setup, if you assign a page at address fc226000, for SATA,
> > > I think that guest will be able to control Communication controller as 
> > > well.
> > Who configures BARs for assigned device guest or host?
> 
> Host.
Nice, device assignment is broken in one more way.

> 
> > If host you can't safely passthrough one of those devices.
> 
> Why not?
> 
For the reason you stated. If you map fc226000-fc227000 to a guest it
can control device at fc226800.

> > But passthrough is not secure anyway since guest can DMA all over host
> > memory.
> 
> 
> That's why we only enable it with I/O mmu, right?
> 
We do? Not sure about it. It is absolutely required if you want
security of course, but not everyone care.

> > > 
> > > > > Maybe the solution is to disable bypass for sub-page BARs and to
> > > > > handle them in qemu, where we don't have alignment restrictions?
> > > > > 
> > > > Making fast path go through qemu for assigned devices? May be remove 
> > > > this pass through crap from kvm to save us all from this misery then? 
> > > 
> > > Another option is for KVM to check these scenarious and deny assignment if
> > > there's such an overlap.
> > One more constrain for device assignment. Simple real life scenarios
> > don't work for our users as it is. Adding more constrains will not help.
> 
> For linux host, you can force resource alignment using a kernel
> parameter. What do you suggest? Ignore this issue?
> 
I suggest forcing resource alignment in a host using a kernel parameter if you
care about security.

--
                        Gleb.