Re: git fails with a broken pipe when one quits the pager

On 2021-02-01 16:44:24 +0100, Ævar Arnfjörð Bjarmason wrote:

AFAIK, SIGPIPE exists for the following reason. Most programs that
generate output are not written to specifically handle pipes. So,
if SIGPIPE did not exist, there would be 2 kinds of behavior:
  1. The program doesn't check for errors, and still outputs data,
     wasting time and resources as output will be ignored.
  2. The program sees that the write() failed and terminates with
     an error message. However, in most cases, such a failure is
     not an error: the consumer has terminated either because it
     no longer needs any input (e.g. with the "head" utility or a
     pager), or because it has terminated abnormally, in which case
     the real error is on the side of the consumer. So, the error
     message from the LHS of the pipe would be annoying.

SIGPIPE solves this issue: the program is simply killed with SIGPIPE. In a shell, one gets a non-zero exit code (128 + 13) due to the signal, but as being on the left-hand side of the pipe, such a non-zero exit code is normally not reported, so that this will not annoy the user.

Note 1: Non-zero exit codes from right-hand side are not reported either by most shells, but zsh can report them, and this is very useful for developers, as programs may fail with a non-zero exit code but without an error message. (Reports may also be done by looking at the standard $? in some hook.)

Note 2: Failures on the left-hand side are less interesting in practice and generally ignored, at least for commands run in interactive shells. For scripts, there are various (non-simple) ways to handle them.

Now, I think that in the case (like Git) a program creates a pipe, it should use its knowledge to handle SIGPIPE / EPIPE. Either this is regarded as an error because the full output is *always expected* to be read, in which case there should be an error message in addition to the usual non-zero exist status (not necessarily 141), or this is regarded as OK (if there is a real failure, this is on the side of the consumer). In the case of Git, the consumer is documented to be a pager, which obviously may not read the full output (e.g. for the GCC repository, "git log" returns more than 3 million lines, back to the year 1988, while one is generally interested in the latest changes only). If the user wants to pipe to something else, he can always use an explicit pipe.