In 7447192 we dropped support for 'cond set' and 'cond clear' commands
with the motivation they were unsafe and sent the wrong message to the
user. That was true, but mostly because it was too generic and required
the user to call `initctl reload` to apply the changes.
This patch restores the behavior, albeit in a very reduced and simple
format. All conditions set with this command are constrained to the
'usr/...' namespace. No subdirectories are allowed. The argument to
the 'cond set|clear' command is disallowed if it contains '/' or '.'
but anything else is supported, for example:
initctl cond set foo:2
creates a static/oneshot condition in /run/finit/cond/usr/foo:2
These conditions are static and are fully handled by the user. The
initctl command is the recommended, and only supported, way of setting
and clearing usr conditions.
This patch also includes a new plugin, usr.so, which is a very simple
inotify plugin for the /run/finit/cond/usr/ directory. When files are
created or removed here the plugin tells the Finit condition engine to
update and trigger service changes.
For instance, the following service is not started by default at boot:
service <usr/foo> myservice -- MyService
However, as soon as `initctl cond set foo` is called, myservice starts.
Consequently, it is stopped when `initctl cond clr foo` is called.
Another major difference from the original is that this implementation
doesn't send IPC commands to create/delete the conditions. This makes
calling these new commands non-blocking so they can be used very early
in the bootstrap, by plugins, if needed.
Signed-off-by: Joachim Wiberg <troglobit@gmail.com>
- Use _PATH_foo for all condition paths, *with* trailing /
- Read condition file first, may not exist, in which case we save time
- Change from libte makepath() to mkpath(), this changes from hard-coded
0777 perms on all cond dirs to 0755 -- possible security fix
Signed-off-by: Joachim Wiberg <troglobit@gmail.com>
Services, like dbus and teamd for instance, may create their PID files
in a subdirectory of /var/run (today often /run). E.g.,
- /var/run/teamd/a1.pid -- For aggregate A1
- /var/run/dbus/pid
- /var/run/lxc/foo.pid -- For container foo
This patch adds support for dynamically adding inotify watchers to any
new subdirectory created in /var/run (discarding too deep directories).
To match services in this directory the run/task/service/sysv stanza
must contain the pid:!/path/to/pidfile.pid syntax. This pid file name
is also used to create the condition this service asserts using the
following formula:
svc/ + <dirname of service> + <subdir and file without .pid>
E.g., the case of teamd (above) gives condition 'svc/usr/bin/teamd/a1'
The special case of dbus is interesting, since it may not be a special
case, but rather the norm for services using a subdirectory. It is
handled as follows; when a new subdirectory is detected, the directory
is scanned for files matching *.pid. Matching files follow the teamd
case. The directory is also scanned for 'pid', which then gives us the
condition 'svc/usr/bin/dbus'
One last example, illustrated by lxc-start, where we want to track the
condition for the LXC container foo. The service stanza:
service pid:!/run/lxc/foo.pid lxc-start -n foo -F -p /run/lxc/foo.pid -- Container foo
This command has no leading path so the condition is composed entirely
from the PID file location:
svc/ + '' + lxc/foo => svc/lxc/foo
Signed-off-by: Joachim Nilsson <troglobit@gmail.com>
Reassert condition when an unchanged/unmodified process goes from
WAITING state to RUNNING. I.e. it had a condition that went to flux
during `initctl reload`, which drove it to WAITING and was then sent
SIGSTOP during reconf.
Also, on condition update, loop through all services until no more state
changes are observed. This allows long dependency chains of services to
resolve and actually go back to RUNNING state as intended whenever any
condition changes at runtime.
Signed-off-by: Jonas Johansson <jonasj76@gmail.com>
Signed-off-by: Joachim Nilsson <troglobit@gmail.com>
On `initctl reload` we reassert conditions, i.e. bring each still viable
condition back in sync with the new reconf generation. This patch fixes
an assumption in the reassert() callback that caused the following nasty
transformation:
/run/finit/cond/net/lo/up --> /run/finit/cond/lo/up
The transformation was caused by the reassert() code assuming a /var/run
prefix rather than /run. We must check the actual runpath, or like this
patch does, use a path neutral way to find the base condition string.
Signed-off-by: Joachim Nilsson <troglobit@gmail.com>
The SVC_RECONF, SVC_LOST, SVC_START, and RUNLEVEL_CHANGE hooks are not
one-shot, they are also not regular conditions since there exist no
mechanism to reset them from flux.
One idea was to turn them into actions, but the lost + start hooks need
to be called multiple times per trigger, e.g. `initctl reload`, which
turned out to be non-trivial to implement right now.
Therefore, for (at least) the Finit v3.1 release these conditions have
been disabled ("nop") and ignored by Finit. Only actual C-style plugins
will be called.
Signed-off-by: Joachim Nilsson <troglobit@gmail.com>
TIL, using mtimes for tracking event orderings is a monumentally bad
idea (queue the nodding UNIX-beards). Mtimes are in wallclock time which
is not necessarily monotonically increasing. A user may adjust the time,
an NTP daemon will continously tune the clock and so on.
Instead, store an explicit generation number in each condition file,
which will be monotonically increased by finit on each reconf.
Joachim Wiberg