Am 27.01.2011 11:41, schrieb Avi Kivity: > On 01/27/2011 12:34 PM, Kevin Wolf wrote: >> Am 27.01.2011 10:49, schrieb Avi Kivity: >>> On 01/27/2011 11:27 AM, Kevin Wolf wrote: >>>> Well, but in the case of qcow2, you don't want to have a big mutex >>>> around everything. We perfectly know which parts are asynchronous and >>>> which are synchronous, so we'd want to do it finer grained from the >>>> beginning. >>> >>> Yes we do. And the way I proposed it, the new mutex does not introduce >>> any new serialization. >>> >>> To repeat, for every qcow2 callback or completion X (not qcow2 read or >>> write operation), we transform it in the following manner: >>> [...] >> >> This works fine for code that is completely synchronous today (and you >> can't serialize it more than it already is anyway). >> >> It doesn't work for qemu_aio_readv/writev because these use AIO for >> reading/writing the data. So you definitely need to rewrite that part, >> or the AIO callback will cause the code to run outside its coroutine. > > The callbacks need to be wrapped in the same way. Schedule a coroutine > to run the true callback.
Okay, I see what you're proposing. You could schedule a new coroutine for callbacks indeed. But I think it's actually easier to convert the bdrv_aio_readv into a bdrv_co_readv (and by that removing the callback) and just make sure that you don't hold the mutex during this call - basically what Stefan's code does, just with mutexes instead of a request queue. >> And during this rewrite you'll want to pay attention that you don't hold >> the mutex for the bdrv_co_readv that was an AIO request before, or >> you'll introduce additional serialization. > > I don't follow. Please elaborate. We were thinking of different approaches. I hope it's clearer now. Kevin