Il 19/11/2012 10:36, Stefan Priebe - Profihost AG ha scritto: > Hi Paolo, > > Am 19.11.2012 09:10, schrieb Paolo Bonzini: >>> I'm sorry the discard requests aren't failing. Qemu / Block driver >>> starts to cancel a bunch of requests. >> >> That is being done in the kernel (the guest, I think) because the UNMAPs >> are taking too long. > > That makes sense. RBD handles discards as buffered I/O. When i do an > mkfs.xfs on a 30GB device i see around 900 pending discard requests. rbd > finishes them all before reporting success. > > If it is correct that a 3.6.7 kernel sends as many discard requests i > only the a solution in using unbuffered I/O for discards. > > Do you know what is the correct way?
I think the correct fix is to serialize them in the kernel. Paolo
