[PATCH] Fix for Bug #5807

[Yuya Nishihara]

On Thu, 1 Mar 2018 11:06:59 +0100, Sascha Nemecek wrote:
> # HG changeset patch
> # User Sascha Nemecek <nemecek at wienfluss.net>
> # Date 1519831479 -3600
> #      Wed Feb 28 16:24:39 2018 +0100
> # Node ID 42ddf4ee4f91d76f19ca0c3efc4c8e4c1c6fa96c
> # Parent  1bd132a021dd00f96604e33a8fb5306d37e56007
> Don't close 'fp' (= 'ui.fout') stream to prevent 'ValueError: I/O 
> operation on closed file' (Bug #5807).
> 
> Regression of changeset 30261:6bed17ba00a1 
> (https://www.mercurial-scm.org/repo/hg/rev/6bed17ba00a1)
> 
> diff -r 1bd132a021dd -r 42ddf4ee4f91 hgext/convert/subversion.py
> --- a/hgext/convert/subversion.py	Wed Feb 21 14:36:42 2018 +0530
> +++ b/hgext/convert/subversion.py	Wed Feb 28 16:24:39 2018 +0100
> @@ -149,7 +149,7 @@
>           pickle.dump(str(inst), fp, protocol)
>       else:
>           pickle.dump(None, fp, protocol)
> -    fp.close()
> +    fp.flush()
>       # With large history, cleanup process goes crazy and suddenly
>       # consumes *huge* amount of memory. The output file being closed,
>       # there is no need for clean termination.

I don't think fp.close() was the source of the problem. Here the process
_exit()s so no cleanup would be run.

I suspect that some hidden bug was disclosed by the change 3a4c0905f357,
"util: always force line buffered stdout when stdout is a tty." Before,
ui.fout.close() just called fflush() because sys.stdout.close() doesn't
close the underlying file stream. This no longer applies to the current
ui.fout as it is created by fdopen(1).