kern/38554: kernel condvar wait functions should take an absolute timeout

[ad%netbsd.org@localhost]

>Number:         38554
>Category:       kern
>Synopsis:       kernel condvar wait functions should take an absolute timeout
>Confidential:   no
>Severity:       serious
>Priority:       high
>Responsible:    kern-bug-people
>State:          open
>Class:          change-request
>Submitter-Id:   net
>Arrival-Date:   Thu May 01 11:50:00 +0000 2008
>Originator:     Andrew Doran
>Release:        4.99.62
>Organization:
The NetBSD Project
>Environment:
n/a
>Description:
When sleeping on condition variables in the kernel, threads must be
prepared to handle "spurious wakeups": wakeups that occur due to some
external agent and do not signal an event of interest to the sleeping 
thread.

We have lots of code segments that do, for example:

  while (!condition) {
     error = cv_timedwait(&cv, &lock, hz);
     if (error == EWOULDBLOCK)
       break;
     ...
  }

If a spurious wakeup occurs, the thread could sleep for N*hz instead
of hz, or it could awaken early and return. Early wakeups are bad news 
for syscalls like nanosleep().
>How-To-Repeat:
Code inspection.

>Fix:
Make condvars take an absolute timeout and add some simple functions
to help with computing wakeup times. The above code fragment would
then look something like the following:

  when = cv_time_now() + cv_time_second();
  while (!condition) {
     error = cv_timedwait(&cv, &lock, when);
     if (error == EWOULDBLOCK)
       break;
     ...
  }

This makes condvars slightly more compatible with the Solaris
interfaces. I think it should also be done before 5.0, because
it's a fairly major change to the interface and it would be good
to have it full stable before it is set in stone.