On Mon, Jan 16, 2023 at 9:10 AM Adam Sampson <a...@offog.org> wrote: > I'm seeing the "colors" tests fail with diffutils 3.8 and 3.9 as well, > and the cause turns out to be a bit surprising. > > The last test in "colors" checks that diff exits 141 (128 + SIGPIPE) > when it's writing to a FIFO that is closed early. Modifying the test > script to run diff under strace reveals that SIGPIPE's handler has been > set to SIG_IGN when diff first queries it -- when it should be SIG_DFL > by default. So diff's SIGPIPE handler doesn't run, and it exits 2 > instead. > > It turns out this is a change in GNU Make 4.3.92 and later -- make does > signal(SIGPIPE, SIG_IGN) for its own purposes, and its child processes > inherit that. With make 4.3, the tests work; with 4.4, this one fails. > > I'm not sure whether this is a fault in make, diff or the test! There > are other programs (e.g. the Python interpreter) that spawn processes > with SIGPIPE set to SIG_IGN, so perhaps it should be robust against > that.
Thanks for the investigation. I've made a small adjustment to that test, to make it work with GNU make before and after that change.
