4GB read limit with 0.42 and NetBSD 3.0/i386

[Jason White]

NetBSD 3.0/i386
Arla-0.42/0.41/0.40

I have a file just over 4GB which was copied from local disk into afs using 
arla-0.42 and the 'cp' command:
-rw-r--r--  1 14773  wheel  4328521728 May 19 21:56 bigfile
Size in megabytes: 4128

When I try to reverse this procedure and copy the file using 'cp' out of afs 
to local disk I hit a 4GB limit:
-rw-r--r--  1 root  wheel  4294770688 May 19 23:02 bigfile-cp-fromafs
Size in megabytes: 4095.81

I have the partial output of 'arladebug almost-all' from just before the 4GB 
limit was hit available at:
/afs/iastate.edu/users/07/13/jdwhite/public/arladebug.syslog.gz  (82K)

What's interesting is that arlad spews tons of debugging output until the 4GB 
limit is reached.  At this point the debugging output abruptly stops.  Arlad 
continues to utilize CPU and network as it did prior to hitting 4GB, yet 
nothing beyond 4294770688 bytes is written to the file.  The cache 
utilization, which had been steadily increasing up to 4GB, drops down to near 
zero utilization.  In addition, arlad continues to receive data for much 
longer than it would take to copy the additional 32MB required to complete 
the transfer and arlad must be killed to stop it from chugging along 
indefinitely.

I did a 'ktrace' on arlad after it started spinning its wheels and have 
placed the output at:
/afs/iastate.edu/users/07/13/jdwhite/public/arlad.after_4G_hit.dump.gz (137M)

In addition, I attempted to copy the file out of AFS using 'arla-cli' and 
its 'cp' command.  arla-cli copied all 4128M of bigfile, but md5 sums did not 
match.  The original bigfile contains the string '12345678', repeated.  Upon 
further inspection, of the copied bigfile, everything up to and 
including the first 32MB+64K is correct, but everything afterward is all 
NULLs.

-Jason

-- 
Jason White <jdw-arla at menelos.com>       Jabber: jdwhite(jabber.org)
http://www.menelos.com/~jdwhite                  jason.d.white(gmail.com)
PGP KeyID: 0x5290E477