This is more or less a temporary fix to restore the former behaviour. In
the long term objects will define a max value, which will be of use for
all kinds of meters.
The field totalmem was formerly used to hold the percentage of used mem
by a process. So at update time, the field info.memmax was being
addressed, which is potentially being updated at the same time, As all
updating routines potentially run in parallel. Though there is an
(quite) easy fix for this: calculate the percentage upon object
printing. This works because conky synchronises all update routines
right before printing the result. (To omit locking on it's own.)
Note that the code does not only use a pre-processor generator for
defining the print functions, but also for their prototypes. Sadly, this
generated a conflict in mboxscan.c which this patch resolves, too.
In the linux kernel 2.6.31 and above, device data can either be in
/sys/class/hwmon/hwmonDEV or /sys/class/hwmon/hwmonDEV/device.
Just stat'ing for the latter doesn't work since it can exist but not contain
the required data (see https://bugs.launchpad.net/bugs/435571 for details).
The patch could be improved to keep in memory the right location of the data
on the user's system instead of trying each time, but, is it worth doing it?
A cleaner but more ugly solution would be to include text_object.h in
every header containing struct text_object definitions. But this
apparently triggers a big mess, since text_object.h itself includes
custom headers. Forward defining struct text_object is obviously the
mostly simple solution until there is a bigger header include review
cleaning it all up.
* minimise core code hooks
* drop useless exporting of private functions (and make them static)
* reorder functions in eve.c so no prototypes are needed
* drop massive header include and add double include barrier in eve.h
While testing, I found two already existing bugs:
* the variable 'a' passed to iconv_convert() needs to be passed by
reference in order to allow for the desired side effect.
* Somehow the trailing junk after an iconv_conversion to a shorter
string messes things up (and gets printed!). I couldn't exactly find
out why this happens, but setting (*p) = 0; solves this problem.
-d was broken because fork-to-background was done after the update thread creation, so the
threads ended up in the wrong process. This delays the thread creation until after the fork.
Originally, I was experiencing sporadic lockups when reading inotify_fd;
which is strange, since it is protected by select(). This should fix it
despite of the original problem.
Create each thread upon registration of the callback function, then use
semaphores to signal when it's time to update and when updating is
finished.
Many thanks to pavelo who originally came up with the idea for this.
Yes, I also thought these are monotonic counters. But it seems like they
are not, at least sometimes the value decreases by 1 leading to a very
very high cpu usage percentage being printed.
These macros can be used just like their OBJ() and OBJ_IF()
counterparts, just that they bail out hard when arg is unset. While
here, also cleanup the macro definition by using __* macros and fully
cleanup the macro namespace when the job is done.
Although this makes conky kind of less robust when it comes to
configuration errors, aborting is the right thing to do to avoid
ambiguity between unknown text objects.
Normally, this is not enough reason to remove code, but in this case it
means that either the tester always had $nvidia (if enabled) or $combine
also in her setup, or it was working without (which I guess, since there
is duplicate code in update_apcupsd()).
There are no INFO_* variables any more, so the argument passed to OBJ()
is always a function pointer or zero. By checking the argument passed in
add_update_callback(), the branch in the OBJ() macro can be dropped,
too.
* Remove leftover INFO_* values from object definitions which didn't
trigger anything.
* Drop the whole INFO_* enum as it's values are not used anymore.
* Drop all need_mask alterations (no idea why there were here at all,
but surely not sane since they are missing in the non-linux
functions).
* Drop the update delay for update_meminfo(), as parsing /proc should
not lead to abnormal overhead.
* Check for x_initialised from inside update_x11info(), so we can call
it unconditionally.
Besides improving performance when updating stuff, we ideally have no
text object specific code in update_stuff() anymore (aside some
leftovers).
The macros in construct_text_object() have gotten a bit crazier than
they were before:
* using CALLBACK(&func) instead of an INFO_* parameter to OBJ() will
make it add the given callback to the list of callbacks to be iterated
over at each update interval.
* BEWARE: the above assumes function pointer values to be > 0!
* This implicitly fixes a bug in the code: passing 0 as INFO_* value
led to selecting INFO_MAIL (1 << 0 == 1).
* Now it would select INFO_CPU (== 0), which got unused and therefore is
not a problem at all (the 0 value should be unused in enums anyway).
This needs some more work, then we should be able to drop the whole
INFO_* enum. Then CALLBACK() can die again and with it goes the ugly
casting stuff done to distinguish callbacks from INFO_* values.
The headerfile used globally should be used instead of one in the local dir.
If conky can't find it, it should be fixed in configure.ac.in or Makefile.am,
and not in the code.
