3316. [tuning] Improved locking performance when recursing.
[RT #28836]
- Use one lock per memory pool instead of associating
them all with a single 'pool_lock' in the dispatch manager.
- Reduce the critical sections for qid->lock in get_dispsocket(),
deref_portentry() and dns_dispatch_addresponse2().
- Added a memory pool for socket events in dns_dispatch_t.
- Add an isc_socketevent_t member in the resquery_t object, and use
it with isc_socket_sendto2() instead of using isc_socket_sendto()
- Tuned the memory pools in dispatch.c for better performance
under load
to 10. Allow setting this in named.conf using the new
'resolver-query-timeout' option, which specifies a max
time in seconds. 0 means 'default' and anything longer
than 30 will be silently set to 30. [RT #22852]
by Andrew Tridgell of the Samba project.) [RT #22629]
2988. [experimental] Added a "dlopen" DLZ driver, allowing the creation
of external DLZ drivers that can be loaded as
shared objects at runtime rather than linked with
named. Currently this is switched on via a
compile-time option, "configure --with-dlz-dlopen".
Note: the syntax for configuring DLZ zones
is likely to be refined in future releases.
(Contributed by Andrew Tridgell of the Samba
project.) [RT #22629]
2987. [func] Improve ease of configuring TKEY/GSS updates by
adding a "tkey-gssapi-keytab" option. If set,
updates will be allowed with any key matching
a principal in the specified keytab file.
"tkey-gssapi-credential" is no longer required
and is expected to be deprecated. (Contributed
by Andrew Tridgell of the Samba project.)
[RT #22629]
memory conditions. Without this fix a DNS cache DB or
ADB could incorrectly stay in an over memory state,
effectively refusing further caching, which
subsequently made a BIND 9 caching server unworkable.
This fix prevents this problem from happening by
polling the state of the memory context, rather than
making a copy of the state, which appeared to cause
a race. This is a "workaround" in that it doesn't
solve the possible race per se, but several experiments
proved this change solves the symptom. Also, the
polling overhead hasn't been reported to be an issue.
This bug should only affect a caching server that
specifies a finite max-cache-size. It's also quite
likely that the bug happens only when enabling threads,
but it's not confirmed yet. [RT #21818]
digest length were used incorrectly, leading to
interoperability problems with other DNS
implementations. This has been corrected.
(Note: If an oversize key is in use, and
compatibility is needed with an older release of
BIND, the new tool "isc-hmac-fixup" can convert
the key secret to a form that will work with all
versions.) [RT #20751]
