Bug 155421

Release 8.2 (and 8.2-RC3 before it, at least) do not appear to be able
to dump core.  This works reliably in previous releases.

node1 is an 8-core system with 72GB of RAM and a 250GB drive plugged
into an onboard ICH10R SATA port.

node2 is an 8-core system with 64GB of RAM and a 250GB drive plugged
into an onboard ESB2 SATA port.

8.2 was installed from a USB drive.  All space was allocated to the
first slice and the default partition scheme was used.  swap is set
to 4GB.

dumpdev="AUTO" is set in /etc/rc.conf, and /dev/dumpdev exists.

Initiating a core with 'sysctl debug.kdb.panic=1' results in one of
several hangs:

 - no progress at all
 - very, very slow progress
 - a secondary panic on 'bufwrite: buffer is not busy'
 - other secondary panics

Examples:
=== node1 ===
# sysctl debug.kdb.panic=1
debug.kdb.panic: 0panic: kdb_sysctl_panic
cpuid = 4
KDB: stack backtrace:
...
Uptime: 20m50s
Physical memory: 73706 MB
Dumping 2780 MB:

=== node2 ===
# sysctl debug.kdb.panic=1
debug.kdb.panic: 0panic: kdb_sysctl_panic
cpuid = 4
KDB: stack backtrace:
...
Uptime: 5m50s
Physical memory: 65521 MB
Dumping 2437 MB:

Other things I have tried with the same results:
- Explicitly specifying the dump device e.g. 'dumpon -v /dev/ad0s1b'
- Increasing swap from 4GB to 32GB
- Changing the SATA controller mode to Compatible, Enhanced, and AHCI
- Booting from a SAS HDD connected to a SAS1068e

Fix: 

The issue may be that the syncer is still doing filesystem updates in
parallel with the attempt to dump core.  I added enough debugging output
to follow the dumper's progress through the first few steps of the dump.
It doesn't seem right that other things are running in parallel with a
panic/coredump.
How-To-Repeat: 1. Set dumpdev to 'AUTO' in /etc/rc.conf
2. Run '/etc/rc.d/dumpon restart', or just reboot
3. Run 'sysctl debug.kdb.panic=1'