Hi, Pádraig Brady <p...@draigbrady.com> writes:
> I vaguely remember on macos that POLLRDBAND needed to be set on read fds for > select, > though didn't check all combinations since we didn't actually need > the poll() replacement. Also select() (emulation of poll) doesn't work on > Solaris or AIX, so we needed to explicitly disable emulation there. Ah, hmm. I don't know how notification APIs work there, maybe some other one can be picked? (IIRC Windows also has it's own set of notification APIs that Gnulib uses when on Windows) > Perhaps we could adjust poll() emulation to be compat everywhere, > but I'm not confident. I see. I'll try to dig around a bit for notes about these platforms (IIRC the libevent manual documented a bunch of weird notification API quirks across platforms) to see how to reliably wait on pipes becoming either readable, closed, or writable, if possible at all. > We can help test on esoteric systems, > especially if appropriate tests are in place. More of a reason to figure out the test then :). On that topic, I did come up with a testcase that should be appropriate for the Coreutils testsuite, but it takes a while to execute (5s), which is something to consider. Here it is: ( sleep 5 | (timeout 3 tee -P && echo g >&2) | : ) 2>&1 | grep -q g The 5s time is no coincidence ;). Maybe a better tool exists, that I'm unaware of, that would just idly wait on stdout to become unwritable (which sounds suspiciously like the issue this patch addresses ;). Expect might also be able to handle this test, but I'm not sure whether that's available in the testsuite. > Note https://www.nongnu.org/pretest/ which may be useful. I'll play around with this a bit later, too. Thanks, have a lovely evening. -- Arsen Arsenović
signature.asc
Description: PGP signature
