Am 04/06/2018 um 01:24 PM schrieb Thomas Lamprecht:
Am 04/06/2018 um 12:28 PM schrieb Fabian Gr├╝nbichler:
On Fri, Apr 06, 2018 at 11:54:03AM +0200, Thomas Lamprecht wrote:
Move the locking inside worker, so that the process doing the actual
work (create or restore) holds the lock, and can call functions which
do locking without deadlocking.

This mirrors the behaviour we use for containers, and allows to add
an 'autostart' parameter which starts the VM after successful
creation. vm_start needs the lock and as not the worker but it's
parents held it, it couldn't know that it was actually save to

Signed-off-by: Thomas Lamprecht <>

I discussed this with Fabian a few months ago and have something in
mind that this shouldn't be that easy, but I cannot remember what
exactly that reason was, so RFC. :)

there is one issue - if somebody holds the flock and you only realize it
after you have forked, you did a fork for nothing (and instead of a
rather fast "timeout" error message, you have to check the task log.
this is not nice from a usability perspective, although it should not
cause problems from a technical/lockdep one ;)

Ah, yeah, I could add a simple check if the VM is already locked before
forking the worker, I mean that's obviously racy but should catch most
cases, and we do that elsewhere too, AFAIK.

I'll try to add it and just send it out, if it's accetable to you then
we could apply it as a temporary improvement :)

OK, scratch it, without calling flock myself (here or in a new method) I cannot do a timeout-less 'check_lock', AFAIS, so better just do the
right thing (after vacantion) ;)

pve-devel mailing list

Reply via email to