GitLab

Sergio Gelato authored Dec 06, 2015 and Quanah Gibson-Mount committed Jun 22, 2020

The fallback to a straight Fisher-Yates shuffle needs to occur whenever the
sum of the *remaining* weights is zero, or else the remaining records will
not be reordered. Testing only once at the beginning covers the case when
all weights are zero, and obviously no shuffling is needed when only one
weight is zero; but other weight combinations are possible, such as (1, 0, 0).

Sergio Gelato authored Dec 06, 2015 and Quanah Gibson-Mount committed Jun 22, 2020

Prior to this change, given two records of weight 1 the algorithm would
return them in the order (0,1) with 100% probability instead of the
desired 50%. This was due to an off-by-one error in the range test.

srv_rand() returns a float in the range [0.0, 1.0[, so r is an integer in the
range [0, total[. The correct probability for record 0 to be chosen is
a[0].weight/total, not (a[0].weight+1)/total.

Howard Chu authored Jul 12, 2015 and Quanah Gibson-Mount committed Jul 28, 2015

Another bug from 5de85b92

Another bug from 5de85b92

James M Leddy authored Aug 23, 2011 and Quanah Gibson-Mount committed Jul 22, 2014

From RFC 2782:

  A client MUST attempt to contact the target host with the
  lowest-numbered priority it can reach.

This patch sorts the DNS SRV records by their priority, and
additionally gives records with a larger weight a higher probability
of appearing earlier. This way, the DNS SRV records are tried in the
order of their priority.

James M Leddy authored Aug 23, 2011 and Howard Chu committed Jul 21, 2014

From RFC 2782:

  A client MUST attempt to contact the target host with the
  lowest-numbered priority it can reach.

This patch sorts the DNS SRV records by their priority, and
additionally gives records with a larger weight a higher probability
of appearing earlier. This way, the DNS SRV records are tried in the
order of their priority.