Re: obtaining a minidump from panic() called from NMI handler

On Fri, Jun 12, 2015 at 11:49:01AM -0700, Maksim Yevmenkin wrote:
> Andriy,
> 
> >> i have a question about obtaining minidump as result of panic() being
> >> called from nmi handler. basically, i have a way to trigger nmi, and,
> >> i would like to panic() system and obtain a minidump.
> >>
> >> i have modified isa_nmi() to appropriately inspect bits and return
> >> non-zero return code. i have turned off machdep.kdb_on_nmi knob (set
> >> it to zero). i have confirmed that amd64 trap() is called with correct
> >> T_NMI type. i've also confirmed that panic() is called from amd64's
> >> trap().
> >>
> >> the issue i have is that system is rebooting too early. basically, it
> >> looks like minidump is started, but, for whatever reason, other cpus
> >> are not completely stopped (or may be they are panic()ing again) and
> >> system just reboots without having complete the minidump.
> >>
> >> the issue is not present when machdep.kdb_on_nmi is set to 1. in this
> >> case, system drops into ddb prompt and 'call doadump' works as
> >> expected. for various reasons i can not use ddb, and, would like to
> >> have system save nmi triggered minidump completely unattended.
> >>
> >> can someone please give me a clue as to what i should be looking into
> >> to make this work?
> >
> > could it be that more than one CPUs get the NMI at the same time?
> 
> i guess, its possible. is there an easy way to check for that?
> 
> > IF yes, then the current code wouldn't handle that well - each of the NMI-ed
> > CPUs will try to stop all others with another NMI and it will wait until each of
> > those CPUs sets an acknowledgement bit in its NMI handler.  This scheme works
> > fine if there's only one CPU that wants to become the master, but results in a
> > deadlock otherwise.
> 
> that makes sense. i don't observe deadlock, but, simple reboot.

I believe I posted patches to serialize ddb entrance or panics due to nmi.
I do not have references handy, you could search lists.