Re: get_maintainer.pl produces non-deterministic results

[Vegard Nossum]

On Wed, 11 Dec 2019 at 01:02, Joe Perches <joe@xxxxxxxxxxx> wrote:
>
> On Tue, 2019-12-10 at 14:47 +0100, Dmitry Vyukov wrote:
> > Hi Joe,
> >
> > scripts/get_maintainer.pl fs/proc/task_mmu.c
> > non-deterministically gives me from 13 to 16 results, different number
> > every time (on upstream 6794862a). Perl v5.28.1. Michael confirmed
> > this with v5.28.2.
> > Vergard suggested to check PERL_HASH_SEED=0. Indeed it fixes
> > non-determinism. But I guess it's not the right solution, there should
> > be some logical problem.
> > My perl-fo is weak, I appreciate if somebody with proper perl-fo takes a look.
> >
> > Thanks
>
> https://lkml.org/lkml/2017/7/13/789

Right, so you can make it reproducible if you add a tie-break to the sorting:

diff --git a/scripts/get_maintainer.pl b/scripts/get_maintainer.pl
index 34085d146fa2c..109d9fb134dad 100755
--- a/scripts/get_maintainer.pl
+++ b/scripts/get_maintainer.pl
@@ -2179,7 +2179,7 @@ sub vcs_assign {
     $hash{$_}++ for @lines;

     # sort -rn
-    foreach my $line (sort {$hash{$b} <=> $hash{$a}} keys %hash) {
+    foreach my $line (sort {$hash{$b} <=> $hash{$a} || $a cmp $b} keys %hash) {
        my $sign_offs = $hash{$line};
        my $percent = $sign_offs * 100 / $divisor;

This would actually favour names that start with early letters (A, B,
...) over late letters (..., Y, Z), which might also be a bad thing. I
think to fix that you could include everybody who has the same number
of signoffs at the cutoff:

diff --git a/scripts/get_maintainer.pl b/scripts/get_maintainer.pl
index 34085d146fa2c..80d3ed2ee6d70 100755
--- a/scripts/get_maintainer.pl
+++ b/scripts/get_maintainer.pl
@@ -2179,7 +2179,8 @@ sub vcs_assign {
     $hash{$_}++ for @lines;

     # sort -rn
-    foreach my $line (sort {$hash{$b} <=> $hash{$a}} keys %hash) {
+    my $prev_sign_offs = -1;
+    foreach my $line (sort {$hash{$b} <=> $hash{$a} || $a cmp $b} keys %hash) {
        my $sign_offs = $hash{$line};
        my $percent = $sign_offs * 100 / $divisor;

@@ -2187,7 +2188,7 @@ sub vcs_assign {
        next if (ignore_email_address($line));
        $count++;
        last if ($sign_offs < $email_git_min_signatures ||
-                $count > $email_git_max_maintainers ||
+                ($prev_sign_offs != $sign_offs && $count >
$email_git_max_maintainers) ||
                 $percent < $email_git_min_percent);
        push_email_address($line, '');
        if ($output_rolestats) {
@@ -2196,6 +2197,8 @@ sub vcs_assign {
        } else {
            add_role($line, $role);
        }
+
+       $prev_sign_offs = $sign_offs;
     }
 }

These patches are probably horribly whitespace damaged, hopefully you
get the gist of it though...


Vegard