I was to hasty when I posted that the SMP lockup I was experiencing has
gone. It is still there, but not on every run of the test program.
[snip]
BTW, I did try the post 2.0.30 patch for SMP lockup prevention but it
had no effect on this problem.
It looks like there are additional forms of deadlock that the existing
placement of calls to allow_interrupts in my patch do not correct. The
placement of these calls was taken directly from the 2.1.28/29 kernel which was
supposed to correct deadlock problems for 2.1, but it's entirely possible that
either (1) all the possible interrupt completion deadlocks were not located and
handled by those patches, or (2) that 2.0 has additional deadlock possibilities
that weren't present in 2.1.
Jered Floyd has been working on back-porting Ingo Molnar's deadlock detection
code to 2.0.30. When he completes that, we will hopefully be able to locate
any remaining deadlocks. In the meanwhile, I've been investigating the
possibility of avoiding these deadlocks for 2.0 by using the I/O APIC.
Leonard