[sheepdog] [sbd] I/O stuck shortly after starting writes

Marcin Mirosław marcin at mejor.pl
Mon Jun 2 10:10:10 CEST 2014


W dniu 02.06.2014 05:52, Liu Yuan pisze:
[...]

> Actually, this is a notorious memory dead-lock and can't be solved if you run
> sbd and sheep on the same nodes. It is a problem similar to NFS that you have
> NFS server and client in the same machine.
> 
> Deadlock:
> Suppose SBD has dirty pages to writeout to backend of sheep. But sheep itself
> need some clean pages to hold this dirty data before do actually disk IOs and
> wait kernel to clean some dirty pages, unfortunately those dirty pages need
> sheep's help to writeout.
> 
> Solution:
> never run client SBD and sheep daemon on the same node.


Hi!
It's a bad news. So iSCSI welcome back on the board:)
Is it possible to disallow using such configuration in sbd? After such
deadlock my test VDI was always broken. Check returns that all pieces
was missing.

Thanks Liu.
Marcin




More information about the sheepdog mailing list