On Tue, Jun 24, 2025 at 3:29 PM Carlo Marcelo Arenas Belón <carenas@xxxxxxxxx> wrote: > I didn'r address it because I didn't knew where to put it, but removing > the signal handler isn't possible, because as soon as we do, EINTR is no > longer "returned" by `poll()` on the systems that allowed that as an > exception to SA_RESTART rules ... Yes, the signal handling stuff in POSIX is messy for Hysterical Raisins. If we go back to V6/V7 Unix we only had three options: * signal is default-handled, or * signal is ignored, or * signal is caught: receipt of signal calls user-supplied function. and the disposition of the signal was always reset to default if caught, so the user function had to call signal() again. Which was sort of OK but had the obvious race flaw, if the signal is delivered twice, before the user function could be re-armed, well, Too Bad. 4.1BSD and 2.9BSD Unix added the "reliable signal" mechanism through a magic library (-ljobs), with several new system calls internally, which morphed into the more-standard-ish 4.2BSD `sigvec` system call, which in turn morphed into the POSIX `sigaction` call. But this still retained the three basic options: default, ignore, or catch-at-user-function. So you have to have a function, even if it doesn't do anything. Meanwhile System III programmers discovered the problem with child-is-ready-to-be-reaped signals (SIGCLD in that variant) getting lost because a SIGCLD handler could not re-catch the signal in time, and "solved" it by a horrible hack that (they claimed) required no user intervention. Given that a user program using SIGCLD already had a handler -- let's call it `child_exited` -- that called `signal(SIGCLD, child_exited)` somewhere within the handler, the hack was to have the OS *re-generate* the signal if there was currently a collectable zombie. This of course depended strongly on the actual code in `child_exited`, which *had to* call `wait` once *before* calling `signal` to re-arm the signal. Otherwise you get infinite recursion. But in fact the program(s?) that they cared about did call `signal` at the end, rather than at the start, of their handler(s). POSIX took `sigvec` and added flags, like `SA_RESTART` and `SA_RESETHAND`, but otherwise maintained a lot of backwards compatibility with both schemes by making a lot of behavior optional. So now you have to over-specify things. People keep trying to fix the klunky interface over time, but history is just too messy... Chris
