homepage

Per skip's email:
FWIW, it appears that Solaris doesn't define SEM_VALUE_MAX but does 
define
_SEM_VALUE_MAX in sys/params.h.
 .../Modules/_multiprocessing/multiprocessing.c: In function 
'init_multiprocessing':
 .../Modules/_multiprocessing/multiprocessing.c:253: error: 
'SEM_VALUE_MAX' undeclared (first use in this function)
 .../Modules/_multiprocessing/multiprocessing.c:253: error: (Each 
undeclared identifier is reported only once
 .../Modules/_multiprocessing/multiprocessing.c:253: error: for each 
function it appears in.)
On Windows the author simple #defines SEM_VALUE_MAX to be LONG_MAX. I 
used
a little cpp action to define it:
 #ifndef SEM_VALUE_MAX
 # ifdef _SEM_VALUE_MAX
 # define SEM_VALUE_MAX _SEM_VALUE_MAX
 # else
 # define SEM_VALUE_MAX INT_MAX
 # endif
 #endif

Jesse> Per skip's email:
 Jesse> FWIW, it appears that Solaris doesn't define SEM_VALUE_MAX but does 
 Jesse> define
 Jesse> _SEM_VALUE_MAX in sys/params.h.
 ...
Thanks for submitting the bug report. I wonder why the processing module
installs on Solaris10 but multiprocessing fails to compile. Did the
SEM_VALUE_MAX code change between the two variants?
Skip

FWIW, this is what I find on a Solaris box.
./sys/param.h:#define _SEM_VALUE_MAX INT_MAX
./sys/sysconfig.h:#define _CONFIG_SEM_VALUE_MAX 21 /* max.
value a semaphore may have */
./sys/unistd.h:#define _SC_SEM_VALUE_MAX 37
./limits.h:#define _POSIX_SEM_VALUE_MAX 32767

Raising this to RB, we should not RC without the MP module properly 
compiling

I suggest committing a patch which falls back to _SEM_VALUE_MAX and see
how the Solaris buildbot reacts.

Anyone mind reviewing the attached patch? This should resolve the solaris 
compile issue. I used skip's suggested code - I removed the #ifdef solaris 
at AP's suggestion.

I recompiled and tested multiprocessing both under Windows and Linux
with this patch, no problems detected. +1 for applying it.

I can confirm that Jesse's patch allows multiprocessing to compile on
Solaris 10. Note, however, that there are other symbolic constants
defined which contain "SEM_VALUE_MAX", all with widely differing
underlying values:
 % find /usr/include -name '*.h' | xargs egrep SEM_VALUE_MAX
 /usr/include/sys/param.h:#define _SEM_VALUE_MAX INT_MAX
 /usr/include/sys/sysconfig.h:#define _CONFIG_SEM_VALUE_MAX 21 
 /* max. value a semaphore may have */
 /usr/include/sys/unistd.h:#define _SC_SEM_VALUE_MAX 
 37
 /usr/include/limits.h:#define _POSIX_SEM_VALUE_MAX 32767
How do we know that _SEM_VALUE_MAX is the proper rvalue to use when
#define-ing SEM_VALUE_MAX?
Richard, as the author of the original processing module do you have
something to contribute to this discussion? You've been completely
silent on this issue.

Skip - Richard has been unavailable a good chunk of the summer. I don't 
know when he will be online again.

I've committed the patch in r66184 on trunk, 66185 py3k. Skip raises a 
good point, therefore I'll leave this open but lower from a blocker.

According to POSIX, if no determinate value for SEM_VALUE_MAX can be
given, the actual value should be queried with
sysconf(_SC_SEM_VALUE_MAX), and SEM_VALUE_MAX should not be defined;
this is in particular the case when the value depends on the system
configuration (such as available memory). _POSIX_SEM_VALUE_MAX specifies
the minimum value guaranteed by POSIX.
Now, it is not plausible why SEM_VALUE_MAX should vary by installation,
as it just depends on what integer type is used to represent the
counter. So more likely, Sun wrote its header files at a time when the
spec was not finished, so they didn't want to clutter the global namespace.
IOW, I think the patch is fine as it stands. If you are over-cautious,
you should test whether _SC_SEM_VALUE_MAX is defined, and use sysconf in
that case, and only fall back to _SEM_VALUE_MAX, then
_POSIX_SEM_VALUE_MAX, then fail, if sysconf isn't available.
If you do make this change, please stop using Py_BuildValue to convert a
C int to a Python int; use PyInt_FromLong instead.

I'm facing the same problem (getting a good definition of SEM_VALUE_MAX)
for my posix_ipc extension. The patch presented here will get the
compiler to shut up on OpenSolaris, but not for the reason you think. At
least, that's the way I see it.
On OpenSolaris (2008.05 is what I'm testing with), _SEM_VALUE_MAX is
indeed #defined in sys/param.h, but it's inside this #ifdef on line 322:
#if (defined(_KERNEL) || defined(_KMEMUSER))
Since multiprocessing.c doesn't #define either of these, both
SEM_VALUE_MAX and _SEM_VALUE_MAX will remain undefined and so the patch
will always fall back to INT_MAX.
IMHO, given the choice between #defining _KERNEL or _KMEMUSER and
calling sysconf(_SC_SEM_VALUE_MAX), I feel safer doing the latter. 
I did a survey of all the operating systems to which I have access
(OpenSolaris 2008.05, OS X 10.5.5, RHEL?, Ubuntu 8.04, and FreeBSD 7).
OpenSolaris is the only one that doesn't #define SEM_VALUE_MAX, and on
all systems except for the possibly Red Hattish one,
sysconf(_SC_SEM_VALUE_MAX) returned the same value as the definition of
SEM_VALUE_MAX. I attached my code & results.

Additional protection checked in in r71022