On Thu, Aug 18, 2005 at 10:18:33AM -0400, Stephan Uphoff wrote:
On Thu, 2005-08-18 at 03:57, Luigi Rizzo wrote:
In fact i don't understand why you consider spinning and sleeping
on a mutex two different things.
The major difference between sleeping (cv_wait,msleep,..) and blocking
on a mutex is priority inheritance.
If you need to be able to use (non-spin) mutexes while holding a
[R|W]LOCK and use a [R|W]LOCK while holding a (non-spin) mutex then you
need to implement priority inheritance for [R|W]LOCKs.
is that required (in FreeBSD, i mean) for algorithmic
correctness or just for performance ?
It is theoretically required since otherwise low priority user threads
(programs) could block system (interrupt) threads indefinitely long.
Extreme low priority (nice?) thread A holds read/write
lock RW as reader
Thread B is holding mutex M tries to acquire read/write lock
RW as writer and sleeps.
Thread C with better priority than A runs and enters a
busy loop (in user space).
Interrupt thread preempts C and tries to acquire Mutex M.
Interrupt priority is propagated to B BUT NOT TO A.
Interrupt thread blocks on Mutex M.
Thread C resumes and will block thread I forever if it
can keep a better priority than thread A.
In practice you would probably just see bad latency every now and then
and may never encounter a hang.