Re: [libvirt] [PATCH 0/5] Use non-blacklisted family/model/stepping for Haswell CPU model
On Thu, Jan 12, 2017 at 12:38:00PM +0000, Dr. David Alan Gilbert wrote:
* Eduardo Habkost (ehabkost(a)redhat.com) wrote:
> On Mon, Jan 09, 2017 at 11:35:54AM +0000, Dr. David Alan Gilbert wrote:
> > * Eduardo Habkost (ehabkost(a)redhat.com) wrote:
> > > A recent glibc commit[1] added a blacklist to ensure it won't use
> > > TSX on hosts that are known to have a broken TSX implementation.
> > >
> > > Our existing Haswell CPU model has a blacklisted
> > > family/model/stepping combination, so it has to be updated to
> > > make sure guests will really use TSX. This is done by patch 5/5.
> > >
> > > However, to do this safely we need to ensure the host CPU is not
> > > a blacklisted one, so we won't mislead guests by exposing
> > > known-to-be-good FMS values on a known-to-be-broken host. This is
> > > done by patch 3/5.
> >
> > I'd just like to mke sure I understand the way this will fail in a
migration;
> > lets say we have a guest that doesn't have the new libc and hosts
> > with a blacklisted CPU, and -cpu Haswell.
> >
> > If I understand correctly then:
> > a) With 'enforce' the destination qemu will fail to start
> > printing an error about the host lack of tsx feature.
>
> Yes.
>
> > b) Without 'enforce' the destination will start but print
> > the same error as a warning, but the guest will probably
> > break as soon as it tries to use a tsx feature?
>
> Yes. The general rule is: without "enforce", live migration can
> break in unpredictable ways.
>
> Without "enforce", QEMU will print a warning, and the VCPU will
> run _without_ the TSX features on CPUID. If we're live-migrating,
> it may break the guest if it tries to use a TSX feature, or break
> migration if a TSX-related bit is already set on a MSR.
OK, but you've been telling people to use "enforce" long enough that
they should have listened.
Are there any other cases we have to worry about; lets say a VM with the
new libc being migrated from an older QEMU, it suddenly changes
CPU ID to one that's supported; what happens?
I assume you are talking about the stepping change, when
migrating from an old QEMU to a host that is _not_ on the
blacklist. In this case, the guest won't see any changes: CPUID
family/model/stepping will be kept the same for the whole life of
the VM (even if it is shut down), thanks to the machine-type
compatibility code.
I'm hoping the guest CPU ID is preserved with the TSX disabled
until
a reboot?
CPUID changes are effective immediately on migration. But guests
often notice the change only on the next reboot.
We could do something to make these problems less likely:
including CPUID data on the migration stream. I have considered
it on the past, but never implemented it. Maybe I should
reconsider that.
(This is another case where it would be interesting to have a
mechanism to let the destination host abort migration early: we
could make QEMU work in "enforce" mode when live-migrating, but
using the migrated CPUID data. This way CPUID changes would never
happen.)
Dave
>
> >
> > Any other combination?
> >
> > Dave
> >
> > > [1]
https://sourceware.org/git/?p=glibc.git;a=commit;h=2702856bf45c82cf8e69f2...
> > >
> > > ---
> > > Cc: dgilbert(a)redhat.com
> > > Cc: fweimer(a)redhat.com
> > > Cc: carlos(a)redhat.com
> > > Cc: triegel(a)redhat.com
> > > Cc: berrange(a)redhat.com
> > > Cc: jdenemar(a)redhat.com
> > > Cc: pbonzini(a)redhat.com
> > >
> > > Eduardo Habkost (5):
> > > i386: Add explicit array size to x86_cpu_vendor_words2str()
> > > i386: host_vendor_fms() helper function
> > > i386/kvm: Blacklist TSX on known broken hosts
> > > pc: Add 2.9 machine-types
> > > i386: Change stepping of Haswell to non-blacklisted value
> > >
> > > include/hw/i386/pc.h | 6 ++++++
> > > target/i386/cpu.h | 1 +
> > > hw/i386/pc_piix.c | 15 ++++++++++++---
> > > hw/i386/pc_q35.c | 13 +++++++++++--
> > > target/i386/cpu.c | 32 ++++++++++++++++++++++----------
> > > target/i386/kvm.c | 17 +++++++++++++++++
> > > 6 files changed, 69 insertions(+), 15 deletions(-)
> > >
> > > --
> > > 2.11.0.259.g40922b1
> > >
> > --
> > Dr. David Alan Gilbert / dgilbert(a)redhat.com / Manchester, UK
>
> --
> Eduardo
--
Dr. David Alan Gilbert / dgilbert(a)redhat.com / Manchester, UK
--
Eduardo