Cancellation is not acted upon unless the process has more than one thread. Consider the program:
read(0, &c, 1);
Being a cancellation point, read should cause the pending cancellation request to be acted upon, and the program should terminate with status 0 as a consequence of the last thread exiting from cancellation. However, under glibc/NPTL, the program waits for input then terminates with status 1.
I suspect the cancellation wrapper code is incorrectly checking SINGLE_THREAD_P and (wrongly) assuming that it implies cancellation is impossible.
Ping. Anyone looked at this?
Async cancellation points get activated only when the number of threads are greater than 1, which is why you're seeing the behaviour you see. What is your use case anyway? I don't see why something like this should be needed at all. Of course, I couldn't find anything in the posix definition of pthread_cancel to not allow something like this:
I don't have a specific usage case in mind, and I don't suspect anyone would intentionally try to cancel the main thread in a single-threaded program knowing it's the main thread. Where the issue is likely to matter is in code that doesn't know it's dealing with the main thread of a single-threaded program, e.g. in a library.
The type of application it's likely to affect is an application that does `N` jobs in parallel by spawning `N-1` threads and using the main thread as a worker thread. As an example, I have a parallel DNS resolver tool that does this. If it turns out that `N` is 1, no threads get created, and the process runs single-threaded. My DNS resolver does not have the threads using pthread_cancel on themselves, but I could imagine moderately-contrived cases where you might want to do this. Self-cancellation is a nice way to setup future code to abort if it does anything that would block, while avoiding exiting immediately. (And if you've written your program this way, you can still call functions that are cancellation points if you temporarily block cancellation around the call.)
In any case, glibc's behavior is non-conformant. POSIX does not make any exceptions to the rules about how cancellation works for the special case of single-threaded programs. (In fact the only special case I can find that POSIX makes for single-threaded programs is that you're allowed to call async-signal-unsafe functions after fork() in a single-threaded program.)
If anything, pthread_cancel should for self-cancellation (with a comment)
THREAD_SETMEM (THREAD_SELF, header.multiple_threads, 1);
__pthread_multiple_threads = *__libc_multiple_threads_ptr = 1;
Then this will work just fine, without any overhead in the common path, except a few insns in the already slow pthread_cancel.
I agree that's a simple and acceptable fix.
I have submitted a patch for review:
Fix committed to master. Thanks for reporting this.