On Fri, Aug 19, 2016 at 07:53:20PM +0200, Thomas Petazzoni wrote:
Hello,
[Adding Rich in Cc.]
On Fri, 19 Aug 2016 19:34:43 +0200, Waldemar Brodkorb wrote:
Do we have a list of the pthread functions that libc.so/libc.a is supposed to provide? See the mail I just send about the axel/libintl issue where libintl also uses pthread_rwlock_*() without being linked with libpthread.so.
I discussed the issues with Rich Felker and he is suggesting _not_ to do some wacky weak/strong handling of these symbols to provide some hackish way for applications to save some space. We should just link with -lpthread if any pthread_* function is in use to avoid any surprises when actually running the application.
So please add explicit -lpthread for these static linking failures and add in libpthread.a functions. It might be just by accident that libc.so provides these function dummies for external uses and may be we should fix this instead.
I'm sorry but I disagree with Rich's proposal here. Lots of libraries rely on this behavior, and we will not be able to upstream the change that consists in linking with -lpthread, because it means a performance degradation when those libraries are used in mono-threaded applications.
It is *not* about saving some space like Rich said. It is about making the mutex lock/unlock operations no-ops when they are not needed.
So I'm sorry, but uClibc should implement this behavior, or you will have dozens of upstream projects to convince :-/
If they want to keep taking advantage of the nop-out hack for dynamic linking glibc, a suitable solution would be a configure check something like:
checking whether -larchive needs -lpthread... yes
However, I'm not sure glibc will even be keeping this behavior long-term, as it's silent-error-prone. Some programmers have been very unpleasantly surprised to find out that the dummy lock functions in glibc's libc.{so,a} don't even enforce any exclusion. I believe this was visible in trylock succeeding on an already-locked mutex where it should have failed. I just did a quick search to see if I could find this discussion on the bug tracker but I didn't find it; it might be on the mailing list or somewhere else, or just buried somewhere that's hard to search.
Note that because there exist process-shared mutexes, the dummy implementation is actually unsafe. An application might omit linking -lpthread because it sees that it links fine without it, but then mmap some shared memory and use pthread_mutex_lock for inter-process synchronization. Very bad silent breakage then happens.
IMO the right fix is not to link a dummy implementation but instead to have the real implementation provide fast-paths for (!multi_threaded && !process_shared) cases.
Rich