[OpenIndiana-discuss] Split-root installations

[Cedric Blancher]

On 27 November 2013 01:51, Jim Klimov <jimklimov at cos.ru> wrote:
> On 2013-11-26 17:25, Jim Klimov wrote:
>>
>> On 2013-11-17 03:50, Jim Klimov wrote:
>>>
>>>    For years I've mentioned "split-root" installations of Solaris-like
>>> systems in such a way that the root filesystem (the BE) is represented
>>> by several datasets, such as a split-off /usr dataset. Also there may
>>> be some datasets shared between boot environments, such as the sinks
>>> for logs and crashdumps, and not all of these are required to live on
>>> the rpool at all. There are cases when all such tweaks may be desirable.
>>
>>
>> WARNING
>>
>> As discussed in another thread, it was discovered that the SMF methods
>> for network/physical (both :default and :nwam) use many programs from
>> /usr, and are executed before the /usr filesystem is actually mounted
>> in case of a split-root installation. This tanks the NWAM setups, but
>> the default ones (based on static files in /etc) succeeds for both DHCP
>> and completely static addressing.
>>
>> I hope to fix this somehow, but a head-on approach failed: the "root"
>> filesystem and other FS services depend on svc:/system/identity:node
>> (indirectly via svc:/system/metainit:default) and that depends on
>> svc:/network/physical... loop and maintenance... Disabling the metainit
>> service does not help fix the dependency_cycle condition :\
>>
>> One approach that would work - a hacky one - is to add an init-script
>> into /etc/rc2.d to "svcadm restart physical" for the currently enabled
>> (and maybe even "online") instance of this service.
>
>
> So far the best shot, both compact and effective, was fixing the
> filesystem/root service like in the patch below. Are there any
> fundamental objections to this sort of fix beside possible lack
> of style? Getting the hands wet in the guts of the system anyhow,
> can't keep them clean... ;)
>
> --- /lib/svc/method/fs-root.orig   2013-11-17 03:21:47.000000000 +0100
> +++ /lib/svc/method/fs-root     2013-11-27 00:35:29.119717438 +0100
> @@ -26,6 +26,12 @@
>  # Make sure that the libraries essential to this stage of booting can be
> found.
>  LD_LIBRARY_PATH=/lib; export LD_LIBRARY_PATH
>
> +NET_PHYS_FMRI="svc:/network/physical:default"
> +NET_NWAM_FMRI="svc:/network/physical:nwam"
> +NET_IPTUN_FMRI="svc:/network/iptun:default"
> +SVCADM=/usr/sbin/svcadm
> +SVCS=/usr/bin/svcs
> +
>  libc_mount() {
>         #
>         # If there is an optimized libc available in /usr that fits this
> @@ -239,6 +245,19 @@
>  #
>  /usr/sbin/devfsadm -I -P
>
> +#
> +# After possibly finding new drivers, and after getting an /usr, redo
> network
> +#
> +for _SVC in $NET_PHYS_FMRI $NET_NWAM_FMRI $NET_IPTUN_FMRI; do
> +       _STATE="`$SVCS -H -o state $_SVC`"
> +       echo "$_SVC: $_STATE" > /dev/msglog
> +       case "$_STATE" in
> +           online*)    $SVCADM restart $_SVC ;;
> +           maintenance)$SVCADM clear $_SVC ;;
> +       esac
> +done
> +unset _SVC _STATE
> +
>  libc_mount
>
>  exit $SMF_EXIT_OK
>
>
>
> While the patches for the networking methods themselves did work
> somewhat, they are large and may be not so apparent for review.
> And this one just plain works nicely and predictably.
>
> Possibly, the "physical:default" patch to untie it from the /usr
> filesystem should also make it into the distro, so that a broken
> server's root-only environment can have its static-file-defined
> networking, how-ever that might help in repairs. And it is trivial.
>
>
> --- /lib/svc/method/net-physical.orig-oi151a8   2013-07-21
> 16:35:53.000000000 +0200
> +++ /lib/svc/method/net-physical        2013-11-26 19:55:04.811003487 +0100
> @@ -38,10 +38,13 @@
>  #
>  smf_configure_ip || exit $SMF_EXIT_OK
>
> -
>  # Make sure that the libraries essential to this stage of booting can be
> found.
>  LD_LIBRARY_PATH=/lib; export LD_LIBRARY_PATH
>
> +# Even if the program is not found explicitly or as a builtin, use some
> +# valid value in the variable, to have reasonable errors at least...
> +[ -x /usr/bin/cut ] && CUT="/usr/bin/cut" || { CUT=cut; builtin cut; }

This test is redundant. Both /usr/bin/cut and the ksh cut builtin use
the code from libcmd.so.1, and you should use one version anyway
instead of having a conditional execution here.
Just do a builtin cut and drop the CUT variable.

Ced
-- 
Cedric Blancher <cedric.blancher at gmail.com>
Institute Pasteur