Re: Revision 309657 to stack_machdep.c renders unbootable system

On Wed, Dec 14, 2016 at 04:50:21PM -0800, Mark Johnston wrote:
> On Wed, Dec 14, 2016 at 03:48:04PM -0800, Steven G. Kargl wrote:
> > On Wed, Dec 14, 2016 at 02:10:48PM -0800, Mark Johnston wrote:
> > > On Wed, Dec 14, 2016 at 12:14:16PM -0800, Mark Johnston wrote:
> > > > On Wed, Dec 14, 2016 at 11:49:26AM -0800, Steven G. Kargl wrote:
> > > > > Well, after 3 days of bisection, I finally found the commit
> > > > > that renders my system unbootable.  The system does not panic.
> > > > > It simply gets stuck in some state.  Nonfunctional keyboard,
> > > > > so can't break into debugger.  No serial console available.
> > > > > The verbose dmesg.boot for a working kernel from revision
> > > > > 309656 is at
> > > > > 
> > > > > http://troutmask.apl.washington.edu/~kargl/freebsd/dmesg.309656.txt
> > > > > 
> > > > > The kernel config file is at
> > > > > 
> > > > > http://troutmask.apl.washington.edu/~kargl/freebsd/SPEW.txt
> > > > > 
> > > > > In looking at /usr/src/UPDATING, there is no warning that one
> > > > > can create a boat anchor by upgrading to 309657.  If compiling
> > > > > a kernel with 'options DDB' is no longer supported, this should
> > > > > be stated in UPDATING.  Or, UPDATING should state that 'options
> > > > > DDB' requires 'options STACK'.  Or, 'options DDB' should simply
> > > > > to the right thing and pull in whatever 'option STACK' does. 
> > > > 
> > > > It is supported though - the point of that change was to fix a problem
> > > > that occurred when DDB is configured but STACK isn't. While testing I
> > > > tried every combination of the two options, and I just tried and
> > > > successfully booted a kernel with DDB and !STACK.
> > > > 
> > > > Does the kernel boot successfully if STACK is added to your
> > > > configuration?
> > > 
> > > I tried your config (plus virtio drivers) and was able to reproduce the
> > > hang in bhyve. Adding STACK "fixed" the hang, as did reverting part of
> > > my change to re-add dead code into the kernel. My VM was always hanging
> > > after printing
> > > 
> > > 000.000050 [ 426] vtnet_netmap_attach       virtio attached txq=1, txd=1024 rxq=1, rxd=1024
> > > 
> > > Sure enough, removing "device netmap" from your config also fixes the
> > > hang. When the hang occurs, I can see with "bhyvectl --get-rip" that
> > > we're stuck in DELAY(), but I can't get a stack at that point. I think
> > > my change is an innocent bystander - it just happened to expose a latent
> > > issue elsewhere.
> > > 
> > > I don't have much more time to look at this right now, but I'll look
> > > into it more tonight.
> > 
> > Yes, adding STACK got me to a booting kernel.  I can't remember
> > why I added netmap to my config file.  Re-adding dead code seems to
> > point to some memory corruption issue or a rogue pointer. :(
> 
> It's not quite that bad, as it turns out. The key is that
> adding/removing the dead code changes the ordering of the items in the
> sysinit linker set. I discovered that if the ctl(4) module is
> initialized before the vtnet driver attaches, the hang occurs, and
> reverting my commit results in a sysinit order where vtnet comes
> _before_ ctl(4). So my change triggers the problem just because it
> happens to perturb something in the compile-time linker.

Thanks for the explanation.

> The issue actually seems to be in 4BSD, and more specifically in r308564
> and r308565. Switching to ULE or reverting either of those two commits
> fixes the hang.

Oh, this is bad.  The last time I checked (and it has been
awhile ago), ULE has/had some very bad performance issues
for numerical computations that use OpenMPI (or likely any
MPI implementation) if a node becomes oversubscribed.  4BSD
at least manages to recover.

Thanks for the pointer to r308564 and 65.  I'll take a look
later tonight as I've managed to break both firefox and chrome
during the upgrade.

-- 
Steve
http://troutmask.apl.washington.edu/~kargl/
https://www.youtube.com/watch?v=6hwgPfCcpyQ