ia64 "hello world" success; crfsd receives SEGV under heavier load
Alex Chiang
achiang at hp.com
Thu Apr 3 13:55:42 PDT 2008
* Zach Brown <zach.brown at oracle.com>:
>
> > #3 0x2000000000084180 in __assert_fail () from /lib/tls/libc.so.6.1
>
> > at crfsd/update.c:868
>
> I think I fixed this:
>
> http://oss.oracle.com/mercurial/zab/crfs/rev/f163aed9f781
>
> The log showed that right before the assertion failure we split a child
> and took up the last block pointer slot in the parent. The assertion
> didn't take this case into account.
Thanks, this patch helped. I can now cp a linux kernel tree
across the network using crfs.
Unfortunately, it looks like there *might* be some data
corruption (haven't gotten around to verifying it yet). I say
*might* based on this experiment:
[root at canola ~]# cp -R linux-2.6 /mnt/crfs/
[root at canola ~]# cp -R /mnt/crfs/linux-2.6 ./linux-2.6-from-crfs-mnt
[root at canola ~]# tar cvf linux-2.6.tar linux-2.6
[root at canola ~]# tar cvf linux-2.6-from-crfs-mnt.tar linux-2.6-from-crfs-mnt
[root at canola ~]# cksum linux-2.6.tar
1563756242 2588590080 linux-2.6.tar
[root at canola ~]# cksum linux-2.6-from-crfs-mnt.tar
3428951535 2588651520 linux-2.6-from-crfs-mnt.tar
[root at canola ~]# du -skx linux-2.6
2704216 linux-2.6
[root at canola ~]# du -skx linux-2.6-from-crfs-mnt
2699280 linux-2.6-from-crfs-mnt
I'm also seeing these messages from the log file:
Apr 03,2008|14:39:49|29576|send_final_neg_errno: assertion 'wire' failed: errno 4294967274
No debug.log for this run (I turned it off to do some timing
tests... :)
[root at canola ~]# time cp -R linux-2.6 /mnt/crfs/
real 12m35.431s
user 0m0.500s
sys 0m36.652s
[root at canola ~]# time rm -rf /mnt/crfs
real 0m4.455s
user 0m0.076s
sys 0m4.380s
For comparison, same tree over NFS:
[root at canola ~]# time cp -R linux-2.6 /mnt/achiang/blender-test/
real 90m26.534s
user 0m0.496s
sys 0m13.224s
[root at canola ~]# time rm -rf /mnt/achiang/blender-test/linux-2.6/
real 18m29.487s
user 0m0.080s
sys 0m2.960s
I'll keep poking. (sometimes, real work gets in the way ;)
Thanks.
/ac
More information about the crfs-devel
mailing list