sleep, usleep, nanosleep - glibc

Split from the chapter glibc: Process, Threads and Signals into the per-function layout; see the method page and the glibc ledger for coverage status.

Semantics – sleep(seconds) returns 0, or the remaining seconds if a signal that is not ignored (or caught with a handler that returns) arrived first. nanosleep(req, rem) is the POSIX-precise form: on EINTR the unslept time is stored in *rem; clock_nanosleep(2) adds clock selection. usleep(usec) is the obsolete microsecond variant – EINVAL for usec >= 1000000, int return only since glibc 2.2.2.

SIGCHLD/SIGALRM interplay (the traps) –

• sleep ends early on any delivered signal; a chatty SIGCHLD stream therefore starves sleep – the man-page attributes row marks sleep MT-Unsafe sig:SIGCHLD/linux for exactly this class of disturbance, and the manual warns against re-sleep(remaining) loops because every signal shifts the wakeup by a second; compute a deadline and loop instead.
• Historically sleep was implemented with alarm(2)/SIGALRM, so mixing alarm and sleep is a bad idea in general; on glibc it is safe because “sleep does not work by means of SIGALRM” (manual, Sleeping). Modifying SIGALRM disposition from a handler while sleeping, or longjmping out, is undefined. usleep’s interaction with alarm, setitimer and timers is simply unspecified.
• nanosleep has no alarm coupling; it is the building block of the other two.

References – sleep(3), usleep(3), nanosleep(2), manual Sleeping; syscall view: nanosleep.