bug#54062: 29.0.50; [PATCH] Eshell should inform processes when a pipe is broken

[Jim Porter]

On 2/21/2022 10:31 AM, Eli Zaretskii wrote:
> > From: Lars Ingebrigtsen <larsi@gnus.org>
> > Cc: Jim Porter <jporterbugs@gmail.com>,  54062@debbugs.gnu.org
> > Date: Mon, 21 Feb 2022 18:39:16 +0100
> >
> > Eli Zaretskii <eliz@gnu.org> writes:
> >
> > > Thanks; I have no further comments.
> >
> > So I've pushed the patch series to Emacs 29.
>
> Something's amiss here: the new test fails on MS-Windows because it
> signals an error inside eshell-wait-for-subprocess:
>
>    Test esh-proc-test/sigpipe-exits-process backtrace:
>      signal(eshell-pipe-broken #<process sh.exe>)
>      eshell-output-object-to-target("killed\n" #<process sh.exe>)
>      eshell-output-object("killed\n" nil [nil (nil . 0) (nil . 0)])
>      #f(compiled-function () #<bytecode 0x1c2b272f8d5f2c5e>)()
>      apply(#f(compiled-function () #<bytecode 0x1c2b272f8d5f2c5e>) nil)
>      timer-event-handler([t 25107 54636 874750 nil #f(compiled-function (
>      sleep-for(0.1)
>      sit-for(0.1)
>      (while (if all eshell-process-list (eshell-interactive-process-p)) (
>      (let ((start (current-time))) (while (if all eshell-process-list (es
>      eshell-wait-for-subprocess(t)
>
> Sounds like the shell is already dead/killed when
> eshell-wait-for-subprocess tries to send it a string?

Thanks for merging, and sorry about the bustage. This turned out to be 
because `eshell-sentinel' for the "head" process in the pipeline called 
`eshell-output-object', but because the tail process was already dead, 
it raised `eshell-pipe-broken'. I believe the reason this only 
manifested on MS Windows was due to a timing difference between 
`delete-process' and `signal-process'; using the `delete-process' path 
on GNU/Linux shows the same problem.

Attached is a patch that ignores the `eshell-pipe-broken' error in 
`eshell-sentinel'. It's not really an error in that case anyway, since 
we only want to write the last bit of output *if we can*.

--------------------

There's just one problem remaining: when running the following command 
on MS Windows[1], you'll (usually) see two Eshell prompts get emitted 
after it finishes:

  yes | sh -c 'read NAME'

However, this is a separate bug that appears in Emacs 27.2 as well. It 
can happen whenever multiple commands in a pipeline get killed. For example:

  ~ $ sh -c 'while true; do sleep 1; echo y; done' | sh -c 'while true; 
do read NAME; echo ${NAME}; done'
  C-c C-c  ; Call `eshell-interrupt-process'

The same happens with C-c C-k (`eshell-kill-process') too. I'll file 
another bug about this, but I wanted to mention it here so no one's 
surprised if they see this come up when testing out this patch.

[1] Well, I assume it's a problem on MS Windows. I actually tested the 
MS Windows code path on GNU/Linux.

0001-Ignore-eshell-broken-pipe-error-in-eshell-sentinel.patch
Description: Text document