Subject: Re: Getting "TLB IPI rendezvous failed..."
To: Frank van der Linden <email@example.com>
From: Frederick Bruckman <firstname.lastname@example.org>
Date: 12/23/2004 19:53:47
On Thu, 23 Dec 2004, Frank van der Linden wrote:
> On Thu, Dec 23, 2004 at 12:56:26AM -0600, Frederick Bruckman wrote:
>> 2) The general pattern seems to be that one cpu is at spipl(), waiting
>> for a lock, while the other cpu insists on doing something to the first
>> cpu, and has no way to back off? I wonder why it's only i386.
> That's the general deadlock pattern: one CPU is at a very high spl
> (splipi, which is the highest possible), waiting to acquire a lock. Another
> CPU holds the lock, and has to do something which involves sending an IPI
> and waiting for the other CPUs to receive it. But, the first CPU never
> gets it.
...because, it's at splipi()? That sounds too obvious. I wish I
understood how that was ever supposed to work. I mean, if there are
places where you can block all interrupts waiting for a lock, how is
it that the other processor can ever assume it can send you an IPI,
and that you are guaranteed to receive it?
> I don't know why this problem has resurfaced recently for some people.
> Manuel is right, collecting the traces is the most important thing, it
> will show where the CPUs get stuck.
Did it again. Same place on CPU 6 in uvm_glue.c as last time, but on
CPU 0, not in mi_switch(), this time...
--- interrupt ---
--- trap (number 3) ---
Has anyone seen these in current? I don't get it in current, but that
could be only because it's one of those wierd things.