Re: [Qemu-devel] [PATCH] intel_iommu: do address space switching when reset

[Peter Xu]

On Thu, Sep 06, 2018 at 12:41:36PM -0600, Alex Williamson wrote:
> On Thu, 6 Sep 2018 14:53:12 +0800
> Peter Xu <address@hidden> wrote:
> 
> > On Wed, Sep 05, 2018 at 08:55:50AM -0600, Alex Williamson wrote:
> > > On Wed,  5 Sep 2018 19:31:58 +0800
> > > Peter Xu <address@hidden> wrote:
> > >   
> > > > We will drop all the mappings when system reset, however we'll still
> > > > keep the existing memory layouts.  That'll be problematic since if IOMMU
> > > > is enabled in the guest and then reboot the guest, SeaBIOS will try to
> > > > drive a device that with no page mapped there.  What we need to do is to
> > > > rebuild the GPA->HPA mapping when system resets, hence ease SeaBIOS.
> > > > 
> > > > Without this patch, a guest that boots on an assigned NVMe device might
> > > > fail to find the boot device after a system reboot/reset and we'll be
> > > > able to observe SeaBIOS errors if turned on:
> > > > 
> > > >   WARNING - Timeout at nvme_wait:144!
> > > > 
> > > > With the patch applied, the guest will be able to find the NVMe drive
> > > > and bootstrap there even after multiple reboots or system resets.
> > > > 
> > > > Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1625173
> > > > CC: QEMU Stable <address@hidden>
> > > > Tested-by: Cong Li <address@hidden>
> > > > Signed-off-by: Peter Xu <address@hidden>
> > > > ---
> > > >  hw/i386/intel_iommu.c | 8 ++++++++
> > > >  1 file changed, 8 insertions(+)
> > > > 
> > > > diff --git a/hw/i386/intel_iommu.c b/hw/i386/intel_iommu.c
> > > > index 3dfada19a6..d3eb068d43 100644
> > > > --- a/hw/i386/intel_iommu.c
> > > > +++ b/hw/i386/intel_iommu.c
> > > > @@ -3231,6 +3231,14 @@ static void vtd_reset(DeviceState *dev)
> > > >       * When device reset, throw away all mappings and external caches
> > > >       */
> > > >      vtd_address_space_unmap_all(s);
> > > > +
> > > > +    /*
> > > > +     * Switch address spaces if needed (e.g., when reboot from a
> > > > +     * kernel that has IOMMU enabled, we should switch address spaces
> > > > +     * to rebuild the GPA->HPA mappings otherwise SeaBIOS might
> > > > +     * encounter DMA errors when running with e.g. a NVMe card).
> > > > +     */
> > > > +    vtd_switch_address_space_all(s);
> > > >  }
> > > >  
> > > >  static AddressSpace *vtd_host_dma_iommu(PCIBus *bus, void *opaque, int 
> > > > devfn)  
> > > 
> > > I'm curious why these aren't part of vtd_init().  vtd_init is where
> > > GCMD is set back to it's power-on state, which disables translation, so
> > > logically we should reset the address space at that point.  Similarly,
> > > the root entry is reset, so it would make sense to throw away all the
> > > mappings there too.  Thanks,  
> > 
> > vtd_init() is only called when realize() or reset, and AFAIU it's not
> > called by GCMD operations.  However I think I get the point that
> > logically we should do similar things in e.g. vtd_handle_gcmd_srtp()
> > when the enable bit switches.
> > 
> > My understanding is that if other things happened rather than the
> > system reboot (e.g., when root pointer is replaced, or during the
> > guest running the guest driver turns DMAR from on to off) the guest
> > will be responsible to do the rest of invalidations first before doing
> > that switch, so we'll possibly do the unmap_all() and address space
> > switches in other places (e.g., in vtd_context_global_invalidate, or
> > per device invalidations).
> 
> AIUI, the entire global command register is write-once, so the guest
> cannot disable the IOMMU or change the root pointer after it's been
> initialized, except through a system reset.

I'm not sure about this one.  The spec has this though (chap 10.4.4,
Global Command Register):

        Register to control remapping hardware. If multiple control
        fields in this register need to be modified, software must
        serialize the modifications through multiple writes to this
        register.

So AFAIU it's at least not write-once register since after all the
guest software need to update the register in a per-bit fashion.  And
AFAIU there's no restriction too on turning the global DMAR off after
it's turned on (though Linux won't do that).

> I think that means the
> guest can only operate through the invalidation queue at runtime.  The
> bug being fixed here is that the IOMMU has been reset to its power-on
> state where translation is disabled, but the emulation of that disabled
> state also needs to return the per-device address space to that of
> system memory, or identity map thereof.  The commit log seems to imply
> that there's some sort of SeaBIOS issue and we're just doing this to
> help the BIOS, when in reality, we just forgot to reset the per device
> address space and a subsequent boot of a guest that didn't enable the
> IOMMU would have the same sort of issues.  In fact any usage of an
> assigned device prior to re-enabling the IOMMU would fail, there's
> nothing unique to NVMe here.  Thanks,

Sorry for the confusion on the warning line.  I pasted that line to
mainly help people when looking for solutions of problem (by searching
with the error line).  It's of course nothing related to SeaBIOS.  How
about I add another line after the WARNING one: "it's not a SeaBIOS
bug, it's QEMU's problem that caused SeaBIOS to timeout on a DMAR
error"?

And yes it should be a general issue but I'm not sure what will
trigger that except this known one.  E.g., I don't think NICs will
easily trigger this since I don't think SeaBIOS will do DMA on the
NICs but I'm not sure.  For other block devices to bootstrap I haven't
tested any (e.g., what if some bios drivers use MMIO/PIO only?  I have
totally no idea...), but I think it'll get that error of course as
long as SeaBIOS/... (anything runs earlier than kernel) tries to do
any form of DMA to any device. If you won't mind, I'll just still keep
the NVMe example to be clear on what's the exact problem we have
encountered, and I add another paragraph emphasizing that it'll also
solve the problem of similar issues where host gets DMA errors during
SeaBIOS boots, and mentioning that this is a common issue.  Would that
work for you?

Thanks,

-- 
Peter Xu