On 20/02/2015 17:26, Paolo Bonzini wrote: > Right now, iothreads are relying on a "contention callback" to release > the AioContext (e.g. for block device operations or to do bdrv_drain_all). > This is necessary because aio_poll needs to be called within an > aio_context_acquire. > > This series drops this requirement for aio_poll, with two effects: > > 1) it makes it possible to remove the "contention callback" in RFifoLock > (and possibly to convert it to a normal GRecMutex, which is why I am not > including a patch to remove callbacks from RFifoLock). > > 2) it makes it possible to start work around making critical sections > for the block layer fine-grained. > > In order to do this, some data is moved from AioContext to local storage. > Stack allocation has size limitations, so thread-local storage is used > instead. There are no reentrancy problems because the data is only live > throughout a small part of aio_poll, and in particular not during the > invocation of callbacks. > > Comments?
Stefan, can you put this on track for 2.4 or do you need a repost? Paolo
