As we were looking into problems with starting and stopping runs in one of our
daq systems, we found that the mlogger does something differently compared to
mhttpd and odbedit. Starting and stopping runs from mhttpd and odbedit works
correctly, but runs restarted by the file size limit in mlogger would often have
problems.

It turns out that mlogger calls cm_transition() with the ASYNC flag, while
mhttpd and odbedit always use SYNC.

The best I can tell, the ASYNC flag tells cm_transition() to fire off the
end-run rpc calls to all clients all at once, without waiting for reply from the
previous client before calling the next one. This effectively defeats the
transition sequence numbers - higher-numbered clients are told to end-run before
the lower-numbered clients have finished their end-run processing.

Most of the time, transition sequence numbers do not matter - all frontends can
stop at the same time, only mlogger has to be the very last, and for transitions
initiated by the mlogger itself, this sequencing is preserved.

It turns out that for our system, correct sequencing of individual frontends is
important, for example, the frontend controlling the trigger system has to stop
first. As we are using correctly adjusted transition sequence numbers, the right
sequence is always done when runs are started/stopped from mhttpd and from
odbedit, but not for runs started/stopped by the mlogger.

So by changing mlogger to always do SYNC transitions, we fixed our sequencing
problem - now runs always start and stop correctly.

But then we ran into a deadlock between the mlogger and the event builder:

1) mlogger wants to stop the run
2a) mlogger stops reading the SYSTEM buffer
2b) mlogger starts cm_transition(SYNC)
3) rpc call to trigger frontend, trigger is blocked (no new events are
generated, but existing data is still flowing through the system)
4) other frontends are stopped (data still flowing)
5) data still flowing through the system, into the event builder, into the
SYSTEM buffer
6) SYSTEM buffer becomes 100% full (mlogger is not reading it, it is busy inside
cm_transition()), event builder is waiting for free space inside bm_send_event()
7) mlogger issues end-run rpc call to event builder
8) deadlock: mlogger is waiting for a reply from the event builder, the event
builder is waiting for free space in the SYSTEM buffer (not processing rpc
calls), mlogger is supposed to empty the SYSTEM buffer, but it is waiting for an
rpc reply instead.

In our particular case, the dead lock was easy to avoid by making the SYSTEM
buffer big enough to accommodate all in-flight data, but the problem remains in
the general case. I suspect mlogger uses ASYNC transactions exactly to avoid
this type of deadlock (mlogger used ASYNC transactions since svn revision 2, the
beginning of time).

Personally, I am not happy about the inconsistency of run sequencing between
mlogger and mhttpd/odbedit (hmm... should also check mfe.c, it also stops runs
based on event count limits, etc). I think it would be better if all programs
did the same exact thing when starting/stopping runs. When mlogger does
something different, we get surprising unexpected behaviour, best avoided.

One possible solution could be to add an odb variable "/logger/async
transitions", set to "false" by default - to be consistent with other programs.
Systems that benefit from the old ASYNC behaviour and do not care about exact
sequencing can set this flag to "true".

K.O.