KVM NX huge page recovery burns a core per guest on kernel 7.0.x

@waynej - I'm kind of curious - with your context history / session with claude, can you get it to review the last lot of commits to mmu.c in the 7.2 branch and see if it believes they may be related / some correlation between them?

I'm curious what it says :)
 
@waynej - I'm kind of curious - with your context history / session with claude, can you get it to review the last lot of commits to mmu.c in the 7.2 branch and see if it believes they may be related / some correlation between them?

I'm curious what it says :)
Sorry for the delayed reply. I re-checked with Claude it has been patched in the master branch. See https://lore.kernel.org/all/20260916143310.1909993-2-clopez@suse.de/

The patch hasn't been backported, so it's still in the 7.2 and 7.3 branches. Currently looks like it'll only come through in 7.4. I've submitted the patch details in my bugzilla post. Hopefully the PVE team will backport it into 7.0.
 
Hi all,

I wanted to add another confirmed case of this issue, with some host-side measurements that may be useful.

Environment:

  • Proxmox kernel: 7.0.14-24-pve
  • Hardware: Dell PowerEdge R730, 2× Intel Xeon E5-2680 v4
  • Guest: Windows 11 Enterprise LTSC
  • VM configuration: 48 vCPUs (cpu: host), 96 GiB RAM, PCIe GPU passthrough
  • KVM NX huge-page mitigation enabled
We experienced severe intermittent freezes, initially suspected to be related to intensive AI inference workloads inside the guest.

However, subsequent monitoring confirmed the problem also occurred while Windows was almost completely idle, with WSL stopped, the GPU unused and plenty of available memory.

During one approximately two-hour observation window, our Windows monitoring recorded 70 execution gaps of five seconds or longer, totalling 4,733 seconds of unresponsiveness (approximately 62% of the observation period). Individual freezes lasted as long as 307 seconds.

All ten independent timing threads within the guest stopped simultaneously during the freezes.

Host-side evidence:

We monitored the VM's kvm-nx-lpage-re thread using /proc/<pid>/task/<tid>/schedstat, sampling its cumulative CPU execution time.

During the major guest freezes on 14 November 2026, the recovery thread accumulated approximately two seconds of CPU runtime every two seconds, indicating that it was continuously occupying one CPU core.

Between these episodes, its accumulated runtime stopped increasing.

This correlated closely with guest-wide freezes and their approximately 60-second recurrence pattern.

Workaround:

We applied the following on the running Proxmox host:

echo 0 > /sys/module/kvm/parameters/nx_huge_pages_recovery_ratio

The result was immediate: Windows responsiveness and RDP access returned to normal without restarting the VM or host.

We retained nx_huge_pages=Y, so the NX huge-page mitigation itself remains enabled.

We've also added kvm.nx_huge_pages_recovery_ratio=0 to GRUB to preserve the workaround across future reboots.

This strongly supports the reported NX recovery issue as the cause of our guest freezes.

We have retained timestamped guest-monitoring logs and the corresponding Proxmox recovery-thread measurements and can provide excerpts if useful.

Is there an update on whether the upstream accounting fix will be backported into the Proxmox 7.0 kernel series? We'd prefer to remain on the current kernel with the workaround until a confirmed fixed build is available.

Thanks to everyone investigating this issue. The existing discussion was instrumental in identifying the problem on our system.