At Tue, 06 Aug 2013 17:51:55 +0200, John Spencer wrote:
On 08/06/2013 07:57 AM, Takashi Iwai wrote:
At Mon, 05 Aug 2013 23:46:09 +0200, John Spencer wrote:
if --with-versioned is active (default), a couple of macros in pcm.c start generating some completely broken, __old-prefixed wrapper functions, which then are getting used whenever the actual function is called.
for example: snd_pcm_hw_params_set_buffer_time_near
__OLD_NEAR1(snd_pcm_hw_params_set_buffer_time_near, unsigned int);
->
#define __OLD_NEAR1(name, ret_type) __P_OLD_NEAR1(__old_, name, ret_type)
->
#define __P_OLD_NEAR1(pfx, name, ret_type) \ ret_type pfx##name(snd_pcm_t *pcm, snd_pcm_hw_params_t *params, ret_type val, int *dir) \ { \ if (INTERNAL(name)(pcm, params,&val, dir)< 0) \ return 0; \ return (ret_type)val; \ }
this will lead to generating a function __old_snd_pcm_hw_params_set_buffer_time_near which expands to
unsigned int __old_snd_pcm_hw_params_set_buffer_time_near(snd_pcm_t *pcm, snd_pcm_hw_params_t *params, ret_type val, int *dir) { if snd1_pcm_hw_params_set_buffer_time_near(pcm, params,&val, dir)< 0) return 0; return (ret_type)val; }
there 2 bugs in there,
- the real function gets passed a pointer to a pointer of unsigned,
which is then happily dereferenced and the original pointer used as an int, and
The pointer cast between signed and unsigned is done normally in C.
- the return type logic is wrong, in case of a non-error, the original
pointer will be returned instead of 0 to indicate success.
The val argument is no pointer but a value.
my problem is that the old version is getting called instead of the new one.
so it passes a pointer where an int is expected. maybe that's because openal uses dlopen to open the alsa DSO. or it is due to some binutils bug or whatever.
openal explicitly tries to load the versioned symbols, IIRC.
anyway, since this all looks very hackish and fragile, and given that the change old api -> new api was done 11 years ago, can we agree to just remove the old cruft and debloat alsa-lib slightly by doing so ? i'm pretty sure *nobody* is using the old stuff anymore.
Probatio diabolica :)
And, even if you remove the old code, the versioned symbols must remain as is. Some applications do load the versioned symbols explicitly, thus removing the versions breaks ABI. So, just removing it is no-go.
If you have some patch to achieve the old code optional (e.g. selectable via a configure option) and still keep the ABI, it'd be greatly appreciated.
Takashi