[alsa-devel] [alsa-cvslog] alsa-lib: Add struct timeval and timespec definition when _POSIX_C_SOURCE is not defined

[Takashi Iwai]

At Mon, 25 Jun 2007 19:24:56 +0400,
Sergey Vlasov wrote:
> 
> On Mon, Jun 25, 2007 at 03:15:39PM +0200, Takashi Iwai wrote:
> [...]
> > > #include <sys/select.h>
> > > #include <alsa/global.h>
> > > 
> > > $ gcc -c --std=c99 x.c
> > > In file included from x.c:2:
> > > /usr/include/alsa/global.h:134: error: redefinition of 'struct timeval'
> > > /usr/include/alsa/global.h:139: error: redefinition of 'struct timespec'
> > > 
> > > (Without #include <sys/select.h> only 'struct timespec' is redefined.)
> > > 
> > > IMHO attempts to define structures which should be defined by system
> > > headers are too dangerous - if someone is using preprocessor flags
> > > which hide required parts of system headers, it is their problem, and
> > > the risk of mismatch is too high.
> > 
> > I agree.
> > 
> > Jaroslav, which system has no definition of these types?  Maybe it'd
> > be better to check it in configure script, rather than the simple
> > ifdef.
> 
> This won't help here - the above problem happens when alsa-lib headers
> are used by other programs, not when alsa-lib is built itself.

But, it can check at least whether the system-wide struct timeval /
timespec definitions.  It wasn't clear to me for which system these
definitions were needed...

> > > BTW, <time.h> does not provide 'struct timeval' at all, even with
> > > _GNU_SOURCE - it defines only 'struct timespec'; however, this does
> > > not cause problems, because ALSA headers currently use only pointers
> > > to snd_timestamp_t, which works even without a complete definition of
> > > 'struct timeval'.  We can try to include <sys/time.h> to get 'struct
> > > timeval', or even <sys/select.h> (which currently seems to provide
> > > both 'struct timeval' and 'struct timespec' even with -ansi, and
> > > http://www.opengroup.org/onlinepubs/009695399/basedefs/sys/select.h.html
> > > tells that it should always provide their definitions).
> > 
> > A patch is welcome :)
> 
> Just not sure which one...
> 
> The situation on my system (with glibc-2.5) is as follows:
> 
>  - <time.h> defines 'struct timespec', unless -ansi (or -std=c99)
>    option is used; it never defines 'struct timeval', even with
>    -D_GNU_SOURCE (which should enable everything).  This means that if
>    only <time.h> is used, alsa/timer.h will not compile with -ansi due
>    to the 'snd_htimestamp_t tstamp' member in snd_timer_tread_t,
>    unless the program includes some other header which defines 'struct
>    timespec' before alsa/timer.h.
> 
>  - <sys/time.h> defines both 'struct timeval' and 'struct timespec',
>    even with -ansi; however, according to the current Open Group spec
>    (http://www.opengroup.org/onlinepubs/009695399/basedefs/sys/time.h.html),
>    <sys/time.h> must provide only 'struct timeval' and is not required
>    to provide 'struct timespec', so the behavior may change in a
>    future glibc release - again, a missing 'struct timespec'
>    definition is fatal for alsa/timer.h.
> 
>  - <sys/select.h> defines both 'struct timeval' and 'struct timespec'
>    even with -ansi.  In the Open Group spec <sys/select.h> is
>    specified to define 'struct timeval', and also define 'struct
>    timespec' "as described in <time.h>" (where it is described as
>    optional), so it should be not worse than <time.h>.
> 
> Would it be OK to include <sys/select.h> instead of <time.h> and
> remove attempt to define system structs from alsa/global.h?  Another
> option is to include both <time.h> and <sys/time.h>.

Yes, sys/select.h sounds reasonable.


thanks,

Takashi