Bug#839101: cowdancer: Keep open() async-signal-safe

[Jack Bates]

On 28/09/16 03:25 PM, Thorsten Glaser wrote:
> Jack Bates dixit:
>> A) Call dlsym() early, before anything calls open(). Not possible
>>   because even if you managed to call dlsym() before anything else,
>>   dlsym() itself might call open() before it was done initializing :-(
>>
>> B) Avoid dlsym() altogether. What are the options in this case?
>>
>>   1) Use sys_open() instead of dlsym(RTLD_NEXT, "open)
>
> How about a combination of both?
>
> Something like this (pseudo code, mostly):
>
> static void initialise_open(void) __attribute__((__constructor__));
>
> static uint8_t open_initialised = 0;
>
> int
> open(const char *path, int flags, mode_t mode)
> {
> 	if (!open_initialised)
> 		return (sys_open(path, flags, mode));
> 	/* current code with dlsym’s result */
> }
>
> static void
> initialise_open(void)
> {
> 	/* call dlsym, which is now allowed to call open() */
> 	/* store its result */
> 	open_initialised = 1;
> }
>
>
> Of course, I probably ought to better go to bed, but someone
> might make something off this. Also, completely untested, modulo
> that I have an application in which these constructors work, in
> plain GCC (no C++).

My intuition is that not having one consistent implementation might be 
surprising to callers of open(), however your solution is AS-safe and 
minimizes the change from the current behavior.
Do you know of or anticipate any cases where implementing open() with 
sys_open() instead of dlsym(RTLD_NEXT, "open") would cause a problem?