[libco] race condition

[hsmatulisgoogle]

Bug Report

There seems to be a race condition in libco:

fluent-bit/lib/flb_libco/amd64.c

Lines 132 to 135 in 5647069

	if(!co_swap) {
	co_init();
	co_swap = (void (*)(cothread_t, cothread_t))co_swap_function;
	}

Here co_swap is a static function pointer, which is assigned to a bunch of data in co_swap_function that corresponds to assembly instructions before first use.

However this code assume that pointer assignment is regular, which is likely platform dependent. Better to use pthread_once rather than checking if its null.

This code should probably be using inline assembly macros instead of writing data to an unsigned char and interpreting it as a function, then they wouldn't have to worry about initializing things before first use in the first place.

[edsiper]

I think the libco backend must be updated, there is some modifications in the newer versions:

https://github.com/libretro/RetroArch/blob/master/libretro-common/libco/amd64.c

[hsmatulisgoogle]

I don't know where they are getting updates from, as I couldn't find a libco repo to update to. Looks like they are using inline assembly in the following case

#if defined(__GNUC__) && !defined(_WIN32) && !defined(__cplusplus)
#define CO_USE_INLINE_ASM
#endif

So if we are on windows, compiling with C++ or don't have GNU extensions we'd still have the race condition. In the remaining cases this updated version doesn't have the race condition

[edsiper]

so I see two things to do:

• update libco
• on windows, use maximum 1 worker

[edsiper]

issue has been fixed. Now the first initialization of co_swap is protected

[edsiper]

refs:

• b058feb
• b48e658