Hello,
finally I compared the G and the H retry rules and I think that both
rules can easily cause numeric overflows if retry_interval_max is set
to larger values than 24h. The result would be a short retry interval
after a series of long intervals.
The first part of the patch moves retry_interval_max processing to
prevent overflows. Sorry about the code now being even less readable.
The H rule causes too random intervals for my taste. What I wanted was
a random number in an geometrically growing interval, but said interval
is not permanently stored, so effectively you got a random number.
The second part of the patch changes the algorithm to be 50% less random,
which looks way better to me. Factors of 1.5 or 2.0 are good.
Michael
----------------------------------------------------------------------
--- exim-4.60/src/retry.c.orig 2006-01-02 13:02:35.000000000 +0100
+++ exim-4.60/src/retry.c 2006-01-02 13:05:15.000000000 +0100
@@ -746,7 +746,11 @@
int last_actual_gap = now - retry_record->last_try;
int lastgap = (last_predicted_gap < last_actual_gap)?
last_predicted_gap : last_actual_gap;
- int next_gap = (lastgap * rule->p2)/1000;
+ /* Impose a global retry max */
+ if ((retry_interval_max*1000)/p2>lastgap)
+ next_gap = (lastgap * p2)/1000;
+ else
+ next_gap = retry_interval_max;
if (rule->rule == 'G')
{
next_try = now + ((lastgap < rule->p1)? rule->p1 : next_gap);
@@ -755,16 +759,11 @@
{
next_try = now + rule->p1;
if (next_gap > rule->p1)
- next_try += random_number(next_gap - rule->p1);
+ next_try += random_number(next_gap - rule->p1)/2 + (next_gap - p1)/2;
}
}
}
- /* Impose a global retry max */
-
- if (next_try - now > retry_interval_max)
- next_try = now + retry_interval_max;
-
/* If the new message length is greater than the previous one, we
have to copy the record first. */
