| Message ID | 1459745351-13138-1-git-send-email-david@gibson.dropbear.id.au (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 04/04/2016 06:49, David Gibson wrote: > There are currently 3 calls to qemu_system_reset() in vl.c. Two of them > are immediately preceded by a cpu_synchronize_all_states9) and the > remaining one should be. > > The one which doesn't is the very first reset called directly from main(). > Without a cpu_synchronize_all_states(), kvm_vcpu_dirty is false at this > point from the earlier cpu_synchronize_all_post_init(). That's incorrect > because the reset path is quite likely to update the CPU state, and that > updated state should be pushed back to KVM, not overwritten with stale > data pushed to KVM immediately after init. > > This patch moves the call to cpu_synchronize_all_states() into > qemu_system_reset() for safety, so it is always called. AFAICT this should > be safe for the handful of callers outside vl.c - these all appear to be in > places where the cpu state is already synchronized so the extra call > will be a no-op. > > Signed-off-by: David Gibson <david@gibson.dropbear.id.au> > --- > vl.c | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > > This fixes a real but on ppc - the incorrect state of kvm_vcpu_dirty > means that an extra cpu_synchronize_state() in revised code for > updating the MSR is not a no-op as expected and loads a stale MSR > value, resulting in an incorrect MSR value on entry to the guest. > > Therefore, I'm hoping to apply this ASAP to 2.6. > > Laurent, could you please verify that this does indeed address the > problem with an incorrect MSR on entry. > > diff --git a/vl.c b/vl.c > index bd81ea9..3629336 100644 > --- a/vl.c > +++ b/vl.c > @@ -1745,6 +1745,8 @@ void qemu_system_reset(bool report) > > mc = current_machine ? MACHINE_GET_CLASS(current_machine) : NULL; > > + cpu_synchronize_all_states(); > + > if (mc && mc->reset) { > mc->reset(); > } else { > @@ -1893,7 +1895,6 @@ static bool main_loop_should_exit(void) > } > if (qemu_reset_requested()) { > pause_all_vcpus(); > - cpu_synchronize_all_states(); > qemu_system_reset(VMRESET_REPORT); > resume_all_vcpus(); > if (!runstate_check(RUN_STATE_RUNNING) && > @@ -1903,7 +1904,6 @@ static bool main_loop_should_exit(void) > } > if (qemu_wakeup_requested()) { > pause_all_vcpus(); > - cpu_synchronize_all_states(); > qemu_system_reset(VMRESET_SILENT); > notifier_list_notify(&wakeup_notifiers, &wakeup_reason); > wakeup_reason = QEMU_WAKEUP_REASON_NONE; > Acked-by: Paolo Bonzini <pbonzini@redhat.com>
On Mon, Apr 04, 2016 at 09:31:16AM +0200, Paolo Bonzini wrote: > > > On 04/04/2016 06:49, David Gibson wrote: > > There are currently 3 calls to qemu_system_reset() in vl.c. Two of them > > are immediately preceded by a cpu_synchronize_all_states9) and the > > remaining one should be. > > > > The one which doesn't is the very first reset called directly from main(). > > Without a cpu_synchronize_all_states(), kvm_vcpu_dirty is false at this > > point from the earlier cpu_synchronize_all_post_init(). That's incorrect > > because the reset path is quite likely to update the CPU state, and that > > updated state should be pushed back to KVM, not overwritten with stale > > data pushed to KVM immediately after init. > > > > This patch moves the call to cpu_synchronize_all_states() into > > qemu_system_reset() for safety, so it is always called. AFAICT this should > > be safe for the handful of callers outside vl.c - these all appear to be in > > places where the cpu state is already synchronized so the extra call > > will be a no-op. > > > > Signed-off-by: David Gibson <david@gibson.dropbear.id.au> > > --- > > vl.c | 4 ++-- > > 1 file changed, 2 insertions(+), 2 deletions(-) > > > > This fixes a real but on ppc - the incorrect state of kvm_vcpu_dirty > > means that an extra cpu_synchronize_state() in revised code for > > updating the MSR is not a no-op as expected and loads a stale MSR > > value, resulting in an incorrect MSR value on entry to the guest. > > > > Therefore, I'm hoping to apply this ASAP to 2.6. > > > > Laurent, could you please verify that this does indeed address the > > problem with an incorrect MSR on entry. > > > > diff --git a/vl.c b/vl.c > > index bd81ea9..3629336 100644 > > --- a/vl.c > > +++ b/vl.c > > @@ -1745,6 +1745,8 @@ void qemu_system_reset(bool report) > > > > mc = current_machine ? MACHINE_GET_CLASS(current_machine) : NULL; > > > > + cpu_synchronize_all_states(); > > + > > if (mc && mc->reset) { > > mc->reset(); > > } else { > > @@ -1893,7 +1895,6 @@ static bool main_loop_should_exit(void) > > } > > if (qemu_reset_requested()) { > > pause_all_vcpus(); > > - cpu_synchronize_all_states(); > > qemu_system_reset(VMRESET_REPORT); > > resume_all_vcpus(); > > if (!runstate_check(RUN_STATE_RUNNING) && > > @@ -1903,7 +1904,6 @@ static bool main_loop_should_exit(void) > > } > > if (qemu_wakeup_requested()) { > > pause_all_vcpus(); > > - cpu_synchronize_all_states(); > > qemu_system_reset(VMRESET_SILENT); > > notifier_list_notify(&wakeup_notifiers, &wakeup_reason); > > wakeup_reason = QEMU_WAKEUP_REASON_NONE; > > > > Acked-by: Paolo Bonzini <pbonzini@redhat.com> Thanks Paolo. Paolo / Peter, should I send this in a pull request for my tree, or should it go through someone else's tree?
On 04/04/2016 06:49, David Gibson wrote: > There are currently 3 calls to qemu_system_reset() in vl.c. Two of them > are immediately preceded by a cpu_synchronize_all_states9) and the > remaining one should be. > > The one which doesn't is the very first reset called directly from main(). > Without a cpu_synchronize_all_states(), kvm_vcpu_dirty is false at this > point from the earlier cpu_synchronize_all_post_init(). That's incorrect > because the reset path is quite likely to update the CPU state, and that > updated state should be pushed back to KVM, not overwritten with stale > data pushed to KVM immediately after init. > > This patch moves the call to cpu_synchronize_all_states() into > qemu_system_reset() for safety, so it is always called. AFAICT this should > be safe for the handful of callers outside vl.c - these all appear to be in > places where the cpu state is already synchronized so the extra call > will be a no-op. > > Signed-off-by: David Gibson <david@gibson.dropbear.id.au> > --- > vl.c | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > > This fixes a real but on ppc - the incorrect state of kvm_vcpu_dirty > means that an extra cpu_synchronize_state() in revised code for > updating the MSR is not a no-op as expected and loads a stale MSR > value, resulting in an incorrect MSR value on entry to the guest. > > Therefore, I'm hoping to apply this ASAP to 2.6. > > Laurent, could you please verify that this does indeed address the > problem with an incorrect MSR on entry. This fixes the unset SF bit of MSR in my test case, so: Tested-by: Laurent Vivier <lvivier@redhat.com> Laurent
On 04/04/2016 09:54, David Gibson wrote: > > Acked-by: Paolo Bonzini <pbonzini@redhat.com> > > Paolo / Peter, should I send this in a pull request for my tree, or > should it go through someone else's tree? Go ahead and send it. Paolo
diff --git a/vl.c b/vl.c index bd81ea9..3629336 100644 --- a/vl.c +++ b/vl.c @@ -1745,6 +1745,8 @@ void qemu_system_reset(bool report) mc = current_machine ? MACHINE_GET_CLASS(current_machine) : NULL; + cpu_synchronize_all_states(); + if (mc && mc->reset) { mc->reset(); } else { @@ -1893,7 +1895,6 @@ static bool main_loop_should_exit(void) } if (qemu_reset_requested()) { pause_all_vcpus(); - cpu_synchronize_all_states(); qemu_system_reset(VMRESET_REPORT); resume_all_vcpus(); if (!runstate_check(RUN_STATE_RUNNING) && @@ -1903,7 +1904,6 @@ static bool main_loop_should_exit(void) } if (qemu_wakeup_requested()) { pause_all_vcpus(); - cpu_synchronize_all_states(); qemu_system_reset(VMRESET_SILENT); notifier_list_notify(&wakeup_notifiers, &wakeup_reason); wakeup_reason = QEMU_WAKEUP_REASON_NONE;
There are currently 3 calls to qemu_system_reset() in vl.c. Two of them are immediately preceded by a cpu_synchronize_all_states9) and the remaining one should be. The one which doesn't is the very first reset called directly from main(). Without a cpu_synchronize_all_states(), kvm_vcpu_dirty is false at this point from the earlier cpu_synchronize_all_post_init(). That's incorrect because the reset path is quite likely to update the CPU state, and that updated state should be pushed back to KVM, not overwritten with stale data pushed to KVM immediately after init. This patch moves the call to cpu_synchronize_all_states() into qemu_system_reset() for safety, so it is always called. AFAICT this should be safe for the handful of callers outside vl.c - these all appear to be in places where the cpu state is already synchronized so the extra call will be a no-op. Signed-off-by: David Gibson <david@gibson.dropbear.id.au> --- vl.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) This fixes a real but on ppc - the incorrect state of kvm_vcpu_dirty means that an extra cpu_synchronize_state() in revised code for updating the MSR is not a no-op as expected and loads a stale MSR value, resulting in an incorrect MSR value on entry to the guest. Therefore, I'm hoping to apply this ASAP to 2.6. Laurent, could you please verify that this does indeed address the problem with an incorrect MSR on entry.