Am 31.03.25 um 16:55 schrieb Fiona Ebner: > Most importantly, start using a dedicated IO thread for the state > file when doing a live snapshot. > > Having the state file be in the iohandler context means that a > blk_drain_all() call in the main thread or vCPU thread that happens > while the snapshot is running will result in a deadlock.
Forgot to mention, this is easily reproducible by shutting down or rebooting a Linux VM with SCSI disk (from within the guest) while a snapshot is being taken. Patch 5/6 contains technical details. Note that shutting down the VM will still result in failure after these patches, but it's much better, because it's not a deadlock anymore and I guess it's kinda philosophically okay, because there can be no VM state of a shutdown VM. Can still be changed later if we want to do something differently there. Much more interesting is the reboot scenario of course (think in the context of time-based automated snapshots where the guest might be doing whatever at the time a snapshot is taken, in particular reboot), which now should work seamlessly and not result in a deadlock either anymore. _______________________________________________ pve-devel mailing list pve-devel@lists.proxmox.com https://lists.proxmox.com/cgi-bin/mailman/listinfo/pve-devel
