Re: [Qemu-devel] [PATCH 2/2] main: switch qemu_set_fd_handler to g_io_add_watch

[Paolo Bonzini]

On 11/24/2011 06:11 PM, Fabien Chouteau wrote:
> Hello,
>
> I've run into some problems with this patch on Windows. The thing is
> that select() should be used only with socket file descriptors.
>
> If glib_select_fill() put a non-socket file descriptor in rfds or wfds,
> select() will fail with this error (btw the return value of select is
> not checked):

This patch actually has been reverted in commit be08e65.  Is it the 
revert that is causing problems?  If so, what is it that 
glib_select_fill puts in rfds and wfds?

> I've look at the patch and I don't see why do you pick file descriptors
> from g_main_context_query's "poll_fds" to put them in the fd_sets (rfds,
> wfds...) and then re-build a "poll_fds" to call g_main_context_check and
> g_main_context_dispatch. From my understanding we can just do:
>
> g_main_context_prepare(context,&max_priority);
>
> n_poll_fds = g_main_context_query(context, max_priority,&timeout,
>                                        poll_fds, ARRAY_SIZE(poll_fds));
>
> if (g_main_context_check(context, max_priority, poll_fds, n_poll_fds)) {
>      g_main_context_dispatch(context);
> }
>
> Or even just call g_main_context_iteration(). What do you think?

You would have to call it in nonblocking mode from a polling handler 
(qemu_add_polling_cb).

A better solution is to move the whole main loop polling into 
os_host_main_loop_wait.

For POSIX, it would be just a call to 
glib_select_fill+select+glib_select_poll.  (Everything around these 
three would stay in the caller, and the fd_sets would be passed to 
os_host_main_loop_wait).

For Windows, it would work like this (and would not use glib_select_* at 
all):

1) call the polling handlers;

2) call select with timeout zero.  If no socket is ready, call 
WSAEventSelect on the sockets listed in the fd_sets;

3) call g_main_context_prepare+query.

4) add the event from (2) and the registered wait objects to the 
poll_fds.  Call g_poll on it.  If sockets were ready, force 0 timeout.

5) If no sockets were ready, call again select with timeout zero.

6) Check the output of g_poll and dispatch the wait objects that are now 
ready.

7) Call g_main_context_check+dispatch.

Paolo