On 16/05/12 22:37, Robert Millan wrote:
> #!/bin/sh
> exec bash < /dev/console > /dev/console 2> /dev/console
If I do that and start each script in /etc/rcS.d/S* manually one at a
time, there is no problem. The rootfs is mounted read-only to begin
with, it gets fixed/marked clean during the checkfs step, later gets
mounted read-write and all is fine.
This has to do with parallelism of "/etc/init.d/rc S". The default
"makefile-style" concurrency can trigger this fault, whereas "touch
/etc/init.d/.legacy-bootordering" makes it no longer reproducible.
Making this particularly awkward, startpar buffers the output of scripts
until they complete, so the console output is probably hiding the
responsible script, which may have already started before fsck
completes; and fsck cannot be triggering the fault directly.
Trying to analyse /etc/init.d/.depend.boot (somewhat messy graphviz.dot
file attached), I think kldutils -> {geli freebsdutils}, rsyslog ->
powerd and zfs are possible candidates because they seem to be able to
start before checkroot.sh completes.
Regards,
--
Steven Chamberlain
steven@pyro.eu.org
Attachment:
vm_fault_copy_wired.png
Description: PNG image
Attachment:
graphviz.dot
Description: MS-Word document
