Florian Weimer kirjoitti:
* Juhana Helovuo:
Package: bind9
Version: 1:9.5.0.dfsg.P2-5.1
Followup-For: Bug #512285
The cause of this bug (#512285) seems to be bug in the implementation of
atomic operations for the Alpha processor. Some of the parameters to the
inline assembly operations are specified to be write-only (results),
when they are in fact used in a read-write fashion. This causes the
current gcc (4.2.4) to allocate some of the parameters into the same
physical registers, causing unintentional aliasing.
Interesting, thanks. Have you provided this information to upstream?
I suppose we could also use the __sync_* primitives, or libatomic-ops.
Yes, I reported this to bind9-bugs list at ISC also, but so far there is
no (human) reply. I got ticket-id [ISC-Bugs #19227].
Furthermore, it seems that while my patch does help, it does not
eliminate the problem. Named still crashes 1-2x / day due to the same
assertion failing. Before the patch there were crashes about every two
hours.
I have rebuilt bind9 source package with configure option
"--disable-atomic", and I am running that version to see if it solves
the problem.
If this helps, then it seems that there is still some unknown bug in the
ISC atomic primitives for Alpha.
Other odd observations:
1. Named with threading allocates a lot of memory. Right after startup
the process size is 148M virtual / 41M RSS. However, it does not seem to
grow very much after that. rndc command reports 21 zones loaded, none
of which is large.
2. One of the ISC atomic operation primitives, isc_atomic_store, seems
strange, as I cannot understand how it produces different results from
an ordinary (aligned) store instruction on Alpha.
Best Regards,
Juhana Helovuo
--
To UNSUBSCRIBE, email to debian-bugs-dist-requ...@lists.debian.org
with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org
