llvm-project
c824e4e7 - [libc] Implement pthread_getcpuclockid (#243852)

Commit
4 hours ago
[libc] Implement pthread_getcpuclockid (#243852) On Linux, the ID of a thread's CPU clock is derived from its TID, using an encoding defined by the kernel. clock_getcpuclockid already did the same thing for process clocks, so I'm moving the encoding into a header which can be shared by both functions. pthread_getcpuclockid doesn't check whether the thread is still running, as POSIX.1-2024 says that the pthread_t of a terminated thread remains valid until the thread is joined or detached. This means that the returned clock ID can be invalid, or (worse) refer to a different thread, as the kernel can reuse the TID once the thread exits. However, the same thing can happen to a clock ID obtained while the thread was still running, so I don't think it makes sense to try to detect this. POSIX doesn't explicitly say what clock_gettime should return for threads in this dead-but-unjoined state. A reasonable interpretation would be to return the CPU time the thread used while it was alive. That's something we *could* implement -- by saving the CPU time just before the thread exits, and then reading it from clock_gettime. For that, we would need to use our own encoding of clockid_t (something derived from pthread_t), independent of what the kernel uses for the underlying syscall. However, that would be a very novel approach (I'm not aware of any implementation doing that), and would slow down the clock_gettime fast path, which is why I'm not attempting it. A related corner case is fork(). The TID field in the thread attributes doubles as the TID cache, and fork() (as well as some clone() invocations) sets it to zero for the duration of the syscall. If another thread calls pthread_getcpuclockid on the forking thread at that time, it will get a clock ID encoding a zero TID, which the kernel interprets as "the calling thread". The same problem already exists for pthread_getschedparam and pthread_setschedparam (the sched_* syscalls also treat zero as the calling thread), so I'm not trying to fix it here. With pthread_getcpuclockid in place, we support the Thread CPU-Time Clocks option, so I'm also changing _POSIX_THREAD_CPUTIME (and the value returned by sysconf(_SC_THREAD_CPUTIME)) to say so. Assisted-by: Claude
Author
Parents
Loading