https://bz.apache.org/bugzilla/show_bug.cgi?id=69804
Bug ID: 69804
Summary: ListenCoresBucketsRatio: IPv6 sockets not duplicated
with Listen 80 / Listen *:80
Product: Apache httpd-2
Version: 2.5-HEAD
Hardware: All
OS: All
Status: NEW
Severity: major
Priority: P2
Component: Core
Assignee: [email protected]
Reporter: [email protected]
Target Milestone: ---
With Listen 80 or Listen *:80 and ListenCoresBucketsRatio > 0 on Linux
(SO_REUSEPORT), apache opens multiple IPv4 listener buckets but only a single
IPv6 listener.
Under high IPv6 traffic load this leads to accept-queue overflows and delayed
SYN-ACKs.
Root cause appears to be in ap_duplicate_listeners(): duplicate sockets are
created after re-resolving the address with apr_sockaddr_info_get(...,
APR_UNSPEC, ...), which can select an IPv4 address even when the original
listener was IPv6, so IPv6 is not bucketized.
Config (minimal reproduce):
Listen 80
ListenCoresBucketsRatio 8
Steps to reproduce:
Start httpd with the config above on a dual-stack host.
Inspect listeners:
ss -lntpO 'sport = :80'
Expected:
IPv4 and IPv6 listeners are equally bucketized (one socket per bucket per
family), e.g., multiple 0.0.0.0:80 and multiple [::]:80.
Observed:
Multiple IPv4 sockets (bucketized) but a single IPv6 socket.
Code references:
Initial listener creation (prioritizes v6 when building the list):
server/listen.c lines 602–629
https://github.com/apache/httpd/blob/5cec2b41bdbb776b012eca4aed9ac8b2946beea9/server/listen.c#L602-L629
Bucket duplication path (re-resolves with APR_UNSPEC, losing v6
family/scoping):
server/listen.c lines 885–913
https://github.com/apache/httpd/blob/5cec2b41bdbb776b012eca4aed9ac8b2946beea9/server/listen.c#L885-L913
Impact:
IPv6 traffic is funneled through a single accept queue, causing backlog
pressure and handshake delays, while IPv4 remains healthy.
Workaround:
Bind IPv6 family explicitly ensuring it is bucketized. It will also accept IPv4
connections.
Listen [::]:80
--
You are receiving this mail because:
You are the assignee for the bug.
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
