mirror of
https://git.proxmox.com/git/pve-qemu
synced 2025-10-04 08:33:09 +00:00
fd53092e9b
In the community forum, users reported issues about RCU stalls and
sluggish VMs after taking a snapshot with RAM in Proxmox VE [0]. Mario
was also experiencing similar issues from time to time and recently,
obtained a GDB stacktrace. The stacktrace showed that, in his case,
the vCPU threads were waiting in cpu_throttle_thread(). It is a good
guess that the issues in the forum could also be because of that.
From searching in the source code, it seems that migration is the only
user of the vCPU throttling functions in QEMU relevant for Proxmox VE
(the only other place where it is used is the Cocoa UI). In
particular, RAM migration will begin throttling vCPUs for
auto-converge.
In migration_iteration_finish() there is an unconditional call to
cpu_throttle_stop(), so do the same in the async snapshot code
specific to Proxmox VE.
It's not clear why the issue began to surface more prominently only
now, since the vCPU throttling was there since commit 070afca258
("migration: Dynamic cpu throttling for auto-converge") in QEMU
v2.10.0. However, there were a lot of changes in the migration code
between v8.1.5 and v9.0.2 and a few of them might have affected the
likelihood of cpu_throttle_set() being called, for example, 4e1871c450
("migration: Don't serialize devices in qemu_savevm_state_iterate()")
[0]: https://forum.proxmox.com/threads/153483
Reported-by: Mario Loderer <m.loderer@proxmox.com>
Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
Tested-by: Mario Loderer <m.loderer@proxmox.com>
|
||
|---|---|---|
| debian | ||
| qemu@5ebde3b5c0 | ||
| .gitignore | ||
| .gitmodules | ||
| backup.txt | ||
| Makefile | ||
| vma_spec.txt | ||