#108 Available timer clock types not properly identified at runtime
The configure and build on HP-UX 11.11 goes fine, but when running checks in my ClamAV build, Check hits an issue in timer_create(), warning of an invalid argument.
check_run.c:478: Error in call to timer_create: Invalid argument
The line in question is as follows:
if(timer_create(check_get_clockid(),
NULL /* fire SIGALRM if timer expires */ ,
&timerid) == 0)
{
The only possible invalid argument would be the return value from check_get_clockid(). This goes to check.c, where because HAVE_LIBRT is defined, it first tries to call create_timer with CLOCK_MONOTONIC and checks for an error return code. CLOCK_MONOTONIC is set to -1 in lib/libcompat.h line 133.
However, when I add an eprintf to check_run.c to determine what value is returned from check_get_clockid(), it gives me a "-1" even though I should get a valid CLOCK_REALTIME value.
Additional debug code within check.c indicates that calling create_timer(-1, NULL, &timerid) returns a nonzero value as expected, and so it falls through to CLOCK_REALTIME. However, this is also set to -1 because of the line in lib/libcompat.h, because the HP-UX system is apparently treating #defines differently from enum keywords. For example:
typedef enum __clockid_t { CLOCK_INVALID = 0, CLOCK_REALTIME = 1, /* the system-wide "wall clock" */ CLOCK_VIRTUAL = 2, /* the per-process user execution clock */ CLOCK_PROFILE = 4, /* the per-process execution clock */ RTTIMER0 = 8, RTTIMER1 = 16 } clockid_t; #ifdef CLOCK_INVALID #error yes #else #error no #endif
This errors out as "no" during test compile.
So the #ifdef CLOCK_REALTIME in lib/libcompat.h will not detect whether the enum contains the specified clock type.
It looks like this would need to be done as a configure check, i.e., HAVE_CLOCK_MONOTONIC, rather than at run time.
In the meantime, I've hardcoded it to return 1 for clock_realtime which will at least let me get the ClamAV unit tests done.
The ClamAV unit tests all passed once I increased the timeout to five minutes to accommodate the 500MHz CPU speed on the ten-year-old B2600, so this item is therefore the only thing impacting an HP-UX 11.11 build.
Michael,
Thanks for your interest in Check.
It is interesting that the compiler used on HP-UX treats enums and #defines
differently, never seen that before.
Although Check does not officially support HP-UX (as we do not have a
configured machine to test on), if you were able to create a patch which
happens to work we would consider adding it into Check.
Looking around, something similar to the following patch in Haskell may be
a good starting point:
https://phabricator.haskell.org/D831?vs=on&id=2745&whitespace=ignore-most#toc
That patch shows a way of checking for available timers on the system at
build time by attempting to compile and run a test program. Something
similar is already being done in Check, take a look at librt_timers.m4.
That check, however, was to get around an issue on OpenBSD where the high
resolution timer calls were stubbed out but always set errno to ENOSYS and
return -1. As long as an alternative timer check maintains that, it should
be OK. Note though that the current check will disable all high resolution
timer support if it fails, as opposed to simply determining what timers
types to use.
Alternatively, one could use the following configure script flag:
--enable-timer-replacement
This will replace all high resolution timer calls with using alarm()
instead. As long as high resolution timers (sub second precision) is not
needed, this should be sufficient.
On Sat, Jul 4, 2015 at 11:51 AM, Michael Pelletier mvpel@users.sf.net
wrote:
Related
Bugs:
#108Thanks for pointing out the enable-timer-replacement option. Probably the most expedient fix would be to have configure enable that if it detects an HP-UX target, which would thus require no changes to the source code.
I found the enum/define distinction a bit odd too, considering that I'm using GCC 4.7.2, but I was able to find this information about the POSIX standard's handling of this situation:
http://stackoverflow.com/questions/21855326/utility-of-macros-for-enum
This is undoubtedly from a POSIX standard that post-dates the HP-UX 11.11 operating system's ca. 2000 release, however. I would expect that 11.31, and probably 11.23, handle this situation properly, so configure could probably limit the enable-timer-replacement override to the 11.11 release.
I'll also take one more look at the OS headers; although I didn't find anything in my first find-grep pass, there's a chance that there's a POSIX-related compatibility define somewhere in there from an OS patch which will implement compliance with the above POSIX version.
If you are intersted in following this up further, and adding the --enable-timer-replacement argument to the configure script manually is not sufficient, kindly create a pull request with proposed changes against Check in its new home:
https://github.com/libcheck/check