[Bug 1811533] Re: Unstable Win10 guest with qemu 3.1 + huge pages + hv_stimer

[Vitaly Kuznetsov]

No, I think it's the other way around: clearing guest pages is
unrelated. It is easy to check with the following kernel patch:

diff --git a/arch/x86/kvm/hyperv.c b/arch/x86/kvm/hyperv.c
index fff790a3f4ee..73c574f930e3 100644
--- a/arch/x86/kvm/hyperv.c
+++ b/arch/x86/kvm/hyperv.c
@@ -776,7 +776,7 @@ int kvm_hv_activate_synic(struct kvm_vcpu *vcpu, bool 
dont_zero_synic_pages)
         */
        kvm_vcpu_deactivate_apicv(vcpu);
        synic->active = true;
-       synic->dont_zero_synic_pages = dont_zero_synic_pages;
+       synic->dont_zero_synic_pages = false;
        return 0;
 }

my expectation is that the issue will remain.

Now what *can* be causing it: when in-QEMU synic is initialized it
creates two memory subregions: for Event page and for Message page
(HV_X64_MSR_SIEFP/HV_X64_MSR_SIMP MSRs). These regions are always 4k in
size and they can me anywhere in guest's memory, not necessarily 2M
aligned.

Now, (if I understood correctly) in vhost code,
vhost_region_add_section() is trying to align to qemu_ram_pagesize() and
this may intersect with synic regions.

We need to summon someone who understands memory_region_* magic in QEMU
and vhost in particular.

-- 
You received this bug notification because you are a member of qemu-
devel-ml, which is subscribed to QEMU.
https://bugs.launchpad.net/bugs/1811533

Title:
  Unstable Win10 guest with qemu 3.1 + huge pages + hv_stimer

Status in QEMU:
  Confirmed

Bug description:
  Host:
  Gentoo linux x86_64, kernel 4.20.1
  Qemu 3.1.0 
  CPU: Intel i7 6850K
  Chipset: X99

  Guest:
  Windows 10 Pro 64bit (1809)
  Machine type: pc-q35_3.1
  Hyper-V enlightenments: 
hv_stimer,hv_reenlightenment,hv_frequencies,hv_vapic,hv_reset,hv_synic,hv_runtime,hv_vpindex,hv_time,hv_relaxed,hv_spinlocks=0x1fff
  Memory: 16GB backed by 2MB huge pages

  Issue:
  Once guest is started, log gets flooded with:

  qemu-system-x86_64: vhost_region_add_section: Overlapping but not
  coherent sections at 103000

  or

  qemu-system-x86_64: vhost_region_add_section:Section rounded to 0
  prior to previous 1f000

  (line endings change)

  and as time goes guest loses network access (virtio-net-pci) and
  general performance diminishes to extent of freezing applications.

  Observations:
  1) problem disappears when hv_stimer is removed
  2) problem disappears when memory backing with huge pages is disabled
  3) problem disappears when machine type is downgraded to pc-q35_3.0

To manage notifications about this bug go to:
https://bugs.launchpad.net/qemu/+bug/1811533/+subscriptions