On 11/3/22 19:06, Kevin Wolf wrote:
I think it can make sense to have coroutine_fn as a documentation for
things that are only ever called in a coroutine even if they could
theoretically also work outside of coroutine context.
Otherwise, when we want to introduce a coroutine_fn call somewhere, it's
not only less obvious that it's even possible to do, but we'll have to
add potentially many additional coroutine_fn annotations in the whole
call chain in an otherwise unrelated patch.
This is true. On the other hand, coroutine_fn also means "this is
allowed to suspend", which may have implications on the need for locking
in the caller. So you need to judge case-by-case.
If there are good reasons to add the note, you could add an assertion
that you are qemu_in_coroutine(), which notes that you are in a
coroutine but you don't suspend. In this case however I don't think
it's likely that there will be a coroutine_fn call added later.
I guess I tend to err on the side of "it's good that it's not obvious
that you can call a coroutine_fn", but I also need to correct myself as
qemu_coroutine_yield() is not the only leaf; there is also
qemu_co_queue_next() and qemu_co_queue_restart_all(), which are
coroutine_fn because they rely on the queuing behavior of
aio_co_enter(). In this case I violated my own rule because it is
always a bug to call these functions outside coroutine context.
Paolo
