|
From: | Anthony Liguori |
Subject: | Re: [Qemu-devel] [6391] Fix nographic mode and VNC |
Date: | Thu, 22 Jan 2009 09:29:49 -0600 |
User-agent: | Thunderbird 2.0.0.19 (X11/20090105) |
Paul Brook wrote:
cpu_exec() spins in a loop checking for 'event_pending'. This is set whenever the host_alarm_timer callback signals. Alternatively, a halted guest will cause cpu_exec() to exit. Only then do you do a select() on pending IO.Ah, I see.If you are using dynticks as your clock, the guest has a slow clock, and the guest is not using halt instructions, then you will not poll IO until the next guest timer tick b/c your guest is just going to spin. There are three possible solutions. Set SIGIO on every file descriptor so that TCG breaks whenever IO is pending. Besides ugliness, this fails because not every type of file descriptor supports SIGIO. A second solution is to use a polling select() in cpu_exec. Since you're adding a system call (and a rather heavy one) in the fast path, this is going to likely hurt TCG performance.This won't work. If the guest really is in a tight loop then TB chaining means it will never exit translated code.
But then signal delivery wouldn't either, right? That suggests that if the guest is in a tight loop right now, QEMU will freeze.
There's a fair bit of code that is safe to run along side of TCG. If we separate the locking for the device model code from every thing (the monitor, vnc, sdl, etc.), then we could still let QEMU be responsive even in such a condition.
The third possibility is to have the select() run in a separate thread from the TCG cpu_exec() loop. cpu_exec() would do an atomic read of 'event_pending' and the IO thread would do an atomic write of 'event_pending' whenever select() returned a writable file descriptor.This suffers from the same problem described above. You need to force the main execution thread to break out of the translated loop. In practice this probably means sending a signal to the main thread.
How does a signal break translated loop execution other than by setting event_pending? The actual signal isn't going to make a difference, it's just setting event_pending=1 that causes it to break out of the loop IIUC.
Regards, Anthony Liguori
You could check event_pending at the start of every TB, but that's likely to incur a fairly big performance hit.Paul
[Prev in Thread] | Current Thread | [Next in Thread] |