kse_release and kse_wakeup problem (fwd)

[Daniel Eischen]

On Mon, 26 Apr 2004, David Xu wrote:

> John Baldwin wrote:
> > 
> > 
> > I.e. do the upcall check in sleepq_catch_signals() right where you already do 
> > thread_suspend_check(1).  The only reason you have to do this, btw, is 
> > because the kse_release() code is trying to mess with thread state internals 
> > using sleepq_abort(), etc.  The other in-kernel code that does that (signals) 
> > already does the check in sleepq_catch_signals() and has done the same type 
> > of check in msleep()/tsleep() for quite a while.
> > 
> > If the kse_release() stuff was just using sleep/wakeup() rather than trying to 
> > manually abort sleeps it wouldn't have to be so intimate with the sleep 
> > interface.
> > 
> > Note that thr's thr_wakeup() and thr_sleep() manage to simulate 
> > synchronization w/o having to abort sleeps, but it is probably also easier to 
> > do that than for the M:N case.
> > 
> 
> I think libthr will encounters same problem as libpthread with new sleep
> queue code, because mtx is released too early in msleep before thread
> markes itself as ON_SLEEPQ, thr_suspend and thr_wakeup have same race 
> window as kse_release and kse_wakeup. Any code wants to put synchronous
> bit in td_flags like these codes will be broken.

I'm experimenting with adding an wakeup_thread() to kern_thread.c
(to complement wakeup() and wakeup_one()).  If we shouldn't be
using sleepq's directly, the thread code either needs to

  a) queue msleep()'ing upcalls/threads itself having them
     all block on on their own unique wchan's; or

  b) use a wakeup_thread() that wakes up a specific thread.

-- 
Dan Eischen