Re: [Qemu-devel] [PATCH 1/2] Use fd signal trick to break us out of select; do not sigwait

[Anthony Liguori <anthony@codemonkey.ws>]

Ian Jackson wrote:
> * Use of SIGUSR2 to interrupt select() does not work because signals
>   which arrive just before entry to select() do not interrupt it.
>
> * The sigwait approach to detecting aio does not work properly because
>   some versions of glibc forget to block signals on the private aio
>   thread under some hard-to-reproduce conditions.  This means that
>   blocking SIGUSR2 is ineffective; the signals can be lost and the
>   program can block in sigwait (!)
>
> So instead we use the time-honoured self-pipe trick: in the signal
> handler we write to a pipe, which we select on when we want to wait
> for the signal, and which we read from (to empty out) just before
> actually doing the `top half' processing which deals with the condition
> to which the signal relates.
>
> We use the existing fd handler infrastructure to run the desired
> completion code out of the main event loop; in the aio completion wait
> we use a cut-down version of the same arrangements.
>   

Hi Ian,

Somehow, I didn't see this until today--sorry for the delayed response.  
I personally prefer the patch I posted.  I've also got another patch on 
top of it that refactors the aio API so that multiple aio 
implementations can be used.  I notice you leave the qemu_aio_poll() in 
place in the main loop, is there a reason you left it there?

Regards,

Anthony Liguori