Re: Poor device management

Jakub Jelinek (
Tue, 19 May 1998 10:29:16 +0200 (MET DST)

> Hi,
> I would like to address the topic of naming of devices. Currently, we
> have two serious problems:
> 1. It appears that Linux can support a maximum of 16 SCSI hard drives
> per system, according to devices.txt.
> This is a serious issue as we would like to have more than that on
> some machines but the OS does not appear to support that. The
> allocated block of device numbers appears to have no support for more.

This is not true. This stuff is currently coded in VGER CVS tree and I hope will get into
Linus' tree soon (as Linux is not very usable for mid to large servers
without it). There was a couple of problems we had to solve, but at the
moment it runs just fine.

[root@1000e /root]# cat /proc/mdstat
Personalities : [2 raid0]
read_ahead 128 sectors
md0 : active raid0 sdk1 sdl1 sdm1 sdn1 16757248 blocks 32k chunks
md1 : active raid0 sdp1 sdq1 sdr1 sdu1 16757248 blocks 32k chunks
md2 : active raid0 sdv1 sdw1 sdx1 sdz1 16757248 blocks 32k chunks
md3 : active raid0 sdaa1 sdab1 sdac1 sdai1 16757248 blocks 32k chunks
md4 : active raid0 sdaj1 sdak1 sdal1 sdag1 16757248 blocks 32k chunks
md5 : active raid0 sdo1 sds1 sdt1 sdy1 sdad1 sdam1 6153792 blocks 32k chunks
md6 : active raid0 sdc1 sdd1 sde1 sdf1 16781568 blocks 32k chunks
md7 : active raid0 sdh1 sdi1 sdj1 8842632 blocks 8k chunks
md8 : active raid0 sdae1 sdaf1 sdah1 sdg1 16763328 blocks 32k chunks
md9 : inactive
md10 : inactive
md11 : inactive
md12 : inactive
md13 : inactive
md14 : inactive
md15 : inactive

> 2. Device numbers can change between boots or even during operation.
> This is because the device numbers (and thus, /dev entries) are
> allocated in sequential order. Thus, for instance, if my Jaz drive
> (SCSI ID 4) is turned off but my scanner (ID 5) is on, the scanner
> gets a particular device assigned to it. However, it the Jaz drive is
> on when the system boots, the Jaz gets that ID and the scanner gets
> something different.
> It would, therefore, be highly advisable to go to something akin to
> Solaris -- the /dev entries specify controller, target, and LUN
> expliticly. This will remove a lot of confusion especially
> surrounding the sg devices and eliminate the problem of drifting
> device entries.

There is devfs filesystem (Richard Gooch posts announces about it every
couple of days), which solves this issue. I think it would be a very good
idea to include that into 2.2 kernel - it is just another filesystem and
quite a bit tested, so it should not bring instabilities into the kernel.
But it really makes system administration easier. On the above server, if I
insert a tray with some more disks that get new sdc-sdi, for instance, I
have to rename a lot of disks in my /etc/mdtab making script and get it
right, otherwise I may loose my data, when first fsck hits some disks.
The script is needed, as when you change disk names, the mdcreate magic
changes as well. Now, with devfs there is no need for such script - the
names stay the same.

Jakub Jelinek | |
Administrator of SunSITE Czech Republic, MFF, Charles University
Ultralinux - first 64bit OS to take full power of the UltraSparc
Linux version 2.1.102 on a sparc64 machine (498.80 BogoMips).

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to