Technical notes on my code additions of 12 and 17 Feb 00
Kern Sibbald
General:
- Dumb UPSes were not working in the 3.7.0 release, see below.
Both Riccardo and I are surprised that we didn't have any
dumb testers :-(. Fortunately, we now have two volunteers
(Al Tuttle and Christian Moeller) -- thanks guys.
- The TIMEOUT value on the slave could shortened (too quick)
by the value of NETTIME. The TIMEOUT timing now begins
from the time the slave (or master) considers it is really
on batteries.
Changes submitted this submission:
- Two users reported that dumb UPSes worked fine in 3.6.2, but not
at all in 3.7.0. The main problem reported was that apcupsd
reported loss of serial communications with the UPS.
- As best I can determine, there were four problems:
1. The code (since a long time) has called smart_poll() even
for dumb UPSes. Prevously, this just caused a 5 second wait
for a dumb UPS. However, recently I added code that checks
for loss of communications. Solution, prevent dumb UPSes
from executing smart_poll(). Note, there were a number of
different tests in the code for distinguing SmartUPSes and
dumb UPSes. I standardized on the following:
if (ups->mode.type > SHAREBASIC)
it is a smart UPS (character signalling)
else
it is a dumb UPS (ioctl signalling)
2. A number of the serial port status flags tests were bitwise
tests resulting in a zero or nonzero value, where the nonzero
value was not 1. The code in apcactions.c then tested for
zero or 1, possibly overlooking a valid status. I corrected
both the code in apcaction.c to check for zero or nonzero, rather
than zero or 1. I also corrected the code in apcserial.c that
picks up the flags bits to return a zero or one. If you are
not familar with the following construct:
variable = !!(flag & bit);
Note, that this expression sets variable to zero if (flag & bit) is
zero and sets it to 1 if (flag & bit) is nonzero. It could also
be written:
variable = (flag & bit)?1:0;
which some people prefer.
3. Finally, a real show stopper was in the dumb UPS code in
check_serial() in apcserial.c. The code picked up the serial port
line flags, then did a read_andlock_shmarea(), which promptly
erased the value of the serial port flags. The fix was to do
the read_andlock_shmarea() before the ioctl().
4. Under certain circumstances, which seem to depend on whether
apcupsd was executed from a command line or from a script,
when the serial port was reset to its original state on termination
of apcupsd by a tcsetattr() command, the status bits of the
serial port would be reset, thus signalling the UPS to drop
power. This caused the premature killpowers, and is aparently
also the reason why there were previously a number of sleep()
calls before the termination -- they simply masked the problem.
- I added a new -R option to the command line that puts a SmartUPS
into dumb mode. I did not document it because it is primarily for
testing. In my case, I don't have the correct cable, but at least,
I can exercise the dumb mode code.
- I merged in the changes that Riccardo sent me for 3.7.1
- I modified the message which is printed if a slave attempts to
do a power kill, which is not possible, but happens because the
install inserts the call in the halt script. The message now
simply says that the killpower was ignored rather than printing
FATAL ERROR. This is a little less concerting to the user.
- The nologin_file flag was passed from the master to the slave. It
prevented the slave from setting a nologin file, so I removed it.
- I noticed several times that the TIMEOUT for the slave seemed
to be incorrect -- that is the slave shutdown rather rapidly. It
turns out that the time variables were reset only when a pass was
made through do_action(). One pass is made each time the slave it
contacted by the master, which is NETTIME. Consequently if you had
a net time of 60, the slave TIMEOUT value could triger up to 60
seconds before it should have. I corrected this by resetting the
timeout values when the slave detects that it is REALLY on batteries
i.e. the second on battery signal.
