Am 01.12.2017 um 10:51 hat Fam Zheng geschrieben: > On Thu, 11/30 17:04, Paolo Bonzini wrote: > > On 30/11/2017 16:10, Kevin Wolf wrote: > > >> Yes, I agree, but that (using CoMutex around graph change) requires > > >> everything, especially the defer_to_main_loop_bh, runs in a coroutine > > >> context, which is exactly what I mean by "introducing 'ubiquitous > > >> coroutines'", because currently we don't have them. > > > Is it hard to do, though? Instead of using a BH to switch to the main > > > loop and outside of coroutine context, you could use aio_co_schedule() > > > and yield, which would leave you in the main loop, but still in > > > coroutine context. > > > > Not that I think of, but just aio_co_schedule wouldn't work, because > > "the coroutine must have yielded unless ctx is the context in which the > > coroutine is running (i.e. the value of qemu_get_current_aio_context() > > from the coroutine itself)". > > > > So you'd have to use a bottom half that calls aio_co_schedule. But that > > would work. > > We have QMP commands that can manupulate the graph which are all not > coroutines. > I think running QMP commands in coroutines has it merit especially regarding > to > the nested event loops. > > Also the bdrv_close_all() and similar at the end of main() do draining too, > which I'm not sure how to deal with. Maybe special case them and forget the > draining CoMutex?
All of these cases have in common that they are the outermost layer with respect to the block subsystem. This means that a nested event loop there should be harmless because the callbacks called by it won't influence callers further up in the call stack. Kevin
