3100/80: 4G disk limit?

I've got a 3100/80.  It seems to work just fine booted diskless.  I
found it already had a disk in it

sd0 at scsibus0 target 1 lun 0: <SGI, SEAGATE ST31230N, 0532> SCSI2 0/direct fixed
sd1: 1010 MB, 3992 cyl, 5 head, 103 sec, 512 bytes/sect x 2070235 sectors

but the diskless root area I had is 2+ gigs.  So I added another disk
to it, at ID 0; it now has

sd0 at scsibus0 target 0 lun 0: <IBM, DCAS-34330, S60B> SCSI2 0/direct fixed
sd0: 4134 MB, 8205 cyl, 6 head, 171 sec, 512 bytes/sect x 8467200 sectors
sd0: sync (160.0ns offset 15), 8-bit (6.250MB/s) transfers, tagged queueing
sd1 at scsibus0 target 1 lun 0: <SGI, SEAGATE ST31230N, 0532> SCSI2 0/direct fixed
sd1: 1010 MB, 3992 cyl, 5 head, 103 sec, 512 bytes/sect x 2070235 sectors
sd1: sync (160.0ns offset 15), 8-bit (6.250MB/s) transfers, tagged queueing

I then labeled (now-)sd0, made a filesystem on it, and copied stuff to
it.  I started seeing tons of filesystem corruption.  So I got
suspicious and sicced my disk-check program on it.  That is a program
that simply writes data such that each sector's contents identifies
which sector it was written to.

This revealed that it's wrapping at 4G.  Reading sector 0 gets the data
written to sector 8388608 (1<<23), for example.

My first thought was 6-byte CDBs.  But this is not the 6-byte CDB
limit; that's at 1G (21 bits of sector address).  I thought maybe it
had wrapped eight times, so I tried the disk-check program again,
restricted to precisely the first 8388608 sectors; that was fine.

This is under 1.6.2, which is old enough a 4G limit would not be all
that implausible - but its off_t is 64-bit, and 1.4T (even earlier) on
sparc has no such issues, making me think that any 4G limit is either
hardware or MD code.  Before going looking at MD software, I figured
I'd ask: does anyone know whether the hardware has any such limitation?

/~\ The ASCII				  Mouse
\ / Ribbon Campaign
 X  Against HTML		[email protected]
/ \ Email!	     7D C8 61 52 5D E7 2D 39  4E F1 31 3E E8 B3 27 4B