On 09/27/2011 11:39 AM, Paolo Bonzini wrote:
On 09/27/2011 06:05 PM, Anthony Liguori wrote:
Actually, for posix-aio, we can just switch to using g_idle_add().
g_idle_add() uses g_source_attach which is thread safe. g_idle_add()
gives you a thread safe mechanism to defer a piece of work to the
main loop which is really what we want here.
For that, a bottom half would also do (apart that I am not sure it is
async-safe with TCG). In fact, that would make sense since all of
posix_aio_process_queue could become a bottom half.
Bottom halves are signal safe, not thread safe.
To make bottom halves thread safe, you would (in the very least) have to add
some barriers when reading/writing the scheduling flag. I think it's much
better to just use GIdle sources though.
This can actually be made to work with sync I/O emulation too by
having another GMainLoop in the sync I/O loop although I thought I
recalled a patch series to remove that stuff.
... which stuff? :)
The sync I/O emulation. Since sync I/O is done in block drivers, they can just
use coroutine I/O instead of sync I/O.
Another GMainLoop in the sync I/O loop is problematic for
two reasons: 1) the sync I/O loop does not relinquish the I/O thread mutex,
which makes it very different from the outer loop; 2) a nested GMainLoop keeps
polling all the file descriptors in the outer loop, which requires you to cope
with reentrancy in those monitor commands that flush AIO.
Re: (2), you can use a separate aio GMainContext. The trick is that you need to
keep a global current main context that switches when you enter the AIO main loop.
It's doable, but since I think we're on the verge of eliminating sync I/O
emulation, that's probably a better path to pursue.
Regards,
Anthony Liguori
Paolo
