What is SCSI?

Copyright © 1995, Wilko Bulte <wilko@yedi.iaf.nl>.
July 6, 1996. {A bit dated but still some good info}

SCSI is an acronym for Small Computer Systems Interface. It is an ANSI standard that has become one of the leading I/O buses in the computer industry. The foundation of the SCSI standard was laid by Shugart Associates (the same guys that gave the world the first mini floppy disks) when they introduced the SASI bus (Shugart Associates Standard Interface).

After some time an industry effort was started to come to a more strict standard allowing devices from different vendors to work together. This effort was recognized in the ANSI SCSI-1 standard. The SCSI-1 standard (approx 1985) is rapidly becoming obsolete. The current standard is SCSI-2 , with SCSI-3 on the drawing boards.

In addition to a physical interconnection standard, SCSI defines a logical (command set) standard to which disk devices must adhere. This standard is called the Common Command Set (CCS) and was developed more or less in parallel with ANSI SCSI-1. SCSI-2 includes the (revised) CCS as part of the standard itself. The commands are dependent on the type of device at hand. It does not make much sense of course to define a Write command for a scanner.

The SCSI bus is a parallel bus, which comes in a number of variants. The oldest and most used is an 8 bit wide bus, with single-ended signals, carried on 50 wires. (If you do not know what single-ended means, do not worry, that is what this document is all about.) Modern designs also use 16 bit wide buses, with differential signals. This allows transfer speeds of 20Mbytes/second, on cables lengths of up to 25 meters. SCSI-2 allows a maximum bus width of 32 bits, using an additional cable. Quickly emerging are Ultra SCSI (also called Fast-20) and Ultra2 (also called Fast-40). Fast-20 is 20 million transfers per second (20 Mbytes/sec on a 8 bit bus), Fast-40 is 40 million transfers per second (40 Mbytes/sec on a 8 bit bus). Most hard drives sold today are single-ended Ultra SCSI (8 or 16 bits).

Of course the SCSI bus not only has data lines, but also a number of control signals. A very elaborate protocol is part of the standard to allow multiple devices to share the bus in an efficient manner. In SCSI-2, the data is always checked using a separate parity line. In pre-SCSI-2 designs parity was optional.

In SCSI-3 even faster bus types are introduced, along with a serial SCSI busses that reduces the cabling overhead and allows a higher maximum bus length. You might see names like SSA and Fiberchannel in this context. None of the serial buses are currently in widespread use (especially not in the typical FreeBSD environment). For this reason the serial bus types are not discussed any further.

As you could have guessed from the description above, SCSI devices are intelligent. They have to be to adhere to the SCSI standard (which is over 2 inches thick BTW). So, for a hard disk drive for instance you do not specify a head/cylinder/sector to address a particular block, but simply the number of the block you want. Elaborate caching schemes, automatic bad block replacement etc are all made possible by this 'intelligent device' approach.

On a SCSI bus, each possible pair of devices can communicate. Whether their function allows this is another matter, but the standard does not restrict it. To avoid signal contention, the 2 devices have to arbitrate for the bus before using it.

The philosophy of SCSI is to have a standard that allows older-standard devices to work with newer-standard ones. So, an old SCSI-1 device should normally work on a SCSI-2 bus. I say Normally, because it is not absolutely sure that the implementation of an old device follows the (old) standard closely enough to be acceptable on a new bus. Modern devices are usually more well-behaved, because the standardization has become more strict and is better adhered to by the device manufacturers.

Generally speaking, the chances of getting a working set of devices on a single bus is better when all the devices are SCSI-2 or newer. This implies that you do not have to dump all your old stuff when you get that shiny 2GB disk: I own a system on which a pre-SCSI-1 disk, a SCSI-2 QIC tape unit, a SCSI-1 helical scan tape unit and 2 SCSI-1 disks work together quite happily. From a performance standpoint you might want to separate your older and newer (=faster) devices however.

Components of SCSI

As said before, SCSI devices are smart. The idea is to put the knowledge about intimate hardware details onto the SCSI device itself. In this way, the host system does not have to worry about things like how many heads are hard disks has, or how many tracks there are on a specific tape device. If you are curious, the standard specifies commands with which you can query your devices on their hardware particulars. FreeBSD uses this capability during boot to check out what devices are connected and whether they need any special treatment.

The advantage of intelligent devices is obvious: the device drivers on the host can be made in a much more generic fashion, there is no longer a need to change (and qualify!) drivers for every odd new device that is introduced.

For cabling and connectors there is a golden rule: get good stuff. With bus speeds going up all the time you will save yourself a lot of grief by using good material.

So, gold plated connectors, shielded cabling, sturdy connector hoods with strain reliefs etc are the way to go. Second golden rule: do no use cables longer than necessary. I once spent 3 days hunting down a problem with a flaky machine only to discover that shortening the SCSI bus by 1 meter solved the problem. And the original bus length was well within the SCSI specification.

SCSI bus types

From an electrical point of view, there are two incompatible bus types: single-ended and differential. This means that there are two different main groups of SCSI devices and controllers, which cannot be mixed on the same bus. It is possible however to use special converter hardware to transform a single-ended bus into a differential one (and vice versa). The differences between the bus types are explained in the next sections.

In lots of SCSI related documentation there is a sort of jargon in use to abbreviate the different bus types. A small list:

With a minor amount of imagination one can usually imagine what is meant.

Wide is a bit ambiguous, it can indicate 16 or 32 bit buses. As far as I know, the 32 bit variant is not (yet) in use, so wide normally means 16 bit.

Fast means that the timing on the bus is somewhat different, so that on a narrow (8 bit) bus 10 Mbytes/sec are possible instead of 5 Mbytes/sec for 'slow' SCSI. As discussed before, bus speeds of 20 and 40 million transfers/second are also emerging (Fast-20 == Ultra SCSI and Fast-40 == Ultra2 SCSI).

It should be noted that the data lines > 8 are only used for data transfers and device addressing. The transfers of commands and status messages etc are only performed on the lowest 8 data lines. The standard allows narrow devices to operate on a wide bus. The usable bus width is negotiated between the devices. You have to watch your device addressing closely when mixing wide and narrow.

Single ended buses

A single-ended SCSI bus uses signals that are either 5 Volts or 0 Volts (indeed, TTL levels) and are relative to a COMMON ground reference. A singled ended 8 bit SCSI bus has approximately 25 ground lines, who are all tied to a single `rail' on all devices. A standard single ended bus has a maximum length of 6 meters. If the same bus is used with fast-SCSI devices, the maximum length allowed drops to 3 meters. Fast-SCSI means that instead of 5Mbytes/sec the bus allows 10Mbytes/sec transfers.

Fast-20 (Ultra SCSI) and Fast-40 allow for 20 and 40 million transfers/second respectively. So, F20 is 20 Mbytes/second on a 8 bit bus, 40 Mbytes/second on a 16 bit bus etc. For F20 the max bus length is 1.5 meters, for F40 it becomes 0.75 meters. Be aware that F20 is pushing the limits quite a bit, so you will quickly find out if your SCSI bus is electrically sound.

Please note that this means that if some devices on your bus use 'fast' to communicate your bus must adhere to the length restrictions for fast buses!

It is obvious that with the newer fast-SCSI devices the bus length can become a real bottleneck. This is why the differential SCSI bus was introduced in the SCSI-2 standard.

For connector pinning and connector types please refer to the SCSI-2 standard itself, connectors etc are listed there in painstaking detail.

Beware of devices using non-standard cabling. For instance Apple uses a 25pin D-type connecter (like the one on serial ports and parallel printers). Considering that the official SCSI bus needs 50 pins you can imagine the use of this connector needs some 'creative cabling'. The reduction of the number of ground wires they used is a bad idea, you better stick to 50 pins cabling in accordance with the SCSI standard. For Fast-20 and 40 do not even think about buses like this.

Differential buses

A differential SCSI bus has a maximum length of 25 meters. Quite a difference from the 3 meters for a single-ended fast-SCSI bus. The idea behind differential signals is that each bus signal has its own return wire. So, each signal is carried on a (preferably twisted) pair of wires. The voltage difference between these two wires determines whether the signal is asserted or de-asserted. To a certain extent the voltage difference between ground and the signal wire pair is not relevant (do not try 10 kVolts though).

It is beyond the scope of this document to explain why this differential idea is so much better. Just accept that electrically seen the use of differential signals gives a much better noise margin. You will normally find differential buses in use for inter-cabinet connections. Because of the lower cost single ended is mostly used for shorter buses like inside cabinets.

There is nothing that stops you from using differential stuff with FreeBSD, as long as you use a controller that has device driver support in FreeBSD. As an example, Adaptec marketed the AHA1740 as a single ended board, whereas the AHA1744 was differential. The software interface to the host is identical for both.

Terminators

Terminators in SCSI terminology are resistor networks that are used to get a correct impedance matching. Impedance matching is important to get clean signals on the bus, without reflections or ringing. If you once made a long distance telephone call on a bad line you probably know what reflections are. With 20Mbytes/sec traveling over your SCSI bus, you do not want signals echoing back.

Terminators come in various incarnations, with more or less sophisticated designs. Of course, there are internal and external variants. Many SCSI devices come with a number of sockets in which a number of resistor networks can (must be!) installed. If you remove terminators from a device, carefully store them. You will need them when you ever decide to reconfigure your SCSI bus. There is enough variation in even these simple tiny things to make finding the exact replacement a frustrating business. There are also SCSI devices that have a single jumper to enable or disable a built-in terminator. There are special terminators you can stick onto a flat cable bus. Others look like external connectors, or a connector hood without a cable. So, lots of choice as you can see.

There is much debate going on if and when you should switch from simple resistor (passive) terminators to active terminators. Active terminators contain slightly more elaborate circuit to give cleaner bus signals. The general consensus seems to be that the usefulness of active termination increases when you have long buses and/or fast devices. If you ever have problems with your SCSI buses you might consider trying an active terminator. Try to borrow one first, they reputedly are quite expensive.

Please keep in mind that terminators for differential and single-ended buses are not identical. You should not mix the two variants.

OK, and now where should you install your terminators? This is by far the most misunderstood part of SCSI. And it is by far the simplest. The rule is: every single line on the SCSI bus has 2 (two) terminators, one at each end of the bus. So, two and not one or three or whatever. Do yourself a favor and stick to this rule. It will save you endless grief, because wrong termination has the potential to introduce highly mysterious bugs. (Note the "potential" here; the nastiest part is that it may or may not work.)

A common pitfall is to have an internal (flat) cable in a machine and also an external cable attached to the controller. It seems almost everybody forgets to remove the terminators from the controller. The terminator must now be on the last external device, and not on the controller! In general, every reconfiguration of a SCSI bus must pay attention to this.

Note that termination is to be done on a per-line basis. This means if you have both narrow and wide buses connected to the same host adapter, you need to enable termination on the higher 8 bits of the bus on the adapter (as well as the last devices on each bus, of course).

What I did myself is remove all terminators from my SCSI devices and controllers. I own a couple of external terminators, for both the Centronics-type external cabling and for the internal flat cable connectors. This makes reconfiguration much easier.

On modern devices, sometimes integrated terminators are used. These things are special purpose integrated circuits that can be dis/en-abled with a control pin. It is not necessary to physically remove them from a device. You may find them on newer host adapters, sometimes they are software configurable, using some sort of setup tool. Some will even auto-detect the cables attached to the connectors and automatically set up the termination as necessary. At any rate, consult your documentation!

Terminator power

The terminators discussed in the previous chapter need power to operate properly. On the SCSI bus, a line is dedicated to this purpose. So, simple huh?

Not so. Each device can provide its own terminator power to the terminator sockets it has on-device. But if you have external terminators, or when the device supplying the terminator power to the SCSI bus line is switched off you are in trouble.

The idea is that initiators (these are devices that initiate actions on the bus, a discussion follows) must supply terminator power. All SCSI devices are allowed (but not required) to supply terminator power.

To allow for un-powered devices on a bus, the terminator power must be supplied to the bus via a diode. This prevents the backflow of current to un-powered devices.

To prevent all kinds of nastiness, the terminator power is usually fused. As you can imagine, fuses might blow. This can, but does not have to, lead to a non functional bus. If multiple devices supply terminator power, a single blown fuse will not put you out of business. A single supplier with a blown fuse certainly will. Clever external terminators sometimes have a LED indication that shows whether terminator power is present.

In newer designs auto-restoring fuses that 'reset' themselves after some time are sometimes used.

Device addressing

Because the SCSI bus is, ehh, a bus there must be a way to distinguish or address the different devices connected to it.

This is done by means of the SCSI or target ID. Each device has a unique target ID. You can select the ID to which a device must respond using a set of jumpers, or a dip switch, or something similar. Some SCSI host adapters let you change the target ID from the boot menu. (Yet some others will not let you change the ID from 7.) Consult the documentation of your device for more information.

Beware of multiple devices configured to use the same ID. Chaos normally reigns in this case. A pitfall is that one of the devices sharing the same ID sometimes even manages to answer to I/O requests!

For an 8 bit bus, a maximum of 8 targets is possible. The maximum is 8 because the selection is done bitwise using the 8 data lines on the bus. For wide buses this increases to the number of data lines (usually 16).

Note that a narrow SCSI device can not communicate with a SCSI device with a target ID larger than 7. This means it is generally not a good idea to move your SCSI host adapter's target ID to something higher than 7 (or your CD-ROM will stop working).

The higher the SCSI target ID, the higher the priority the devices has. When it comes to arbitration between devices that want to use the bus at the same time, the device that has the highest SCSI ID will win. This also means that the SCSI host adapter usually uses target ID 7. Note however that the lower 8 IDs have higher priorities than the higher 8 IDs on a wide-SCSI bus. Thus, the order of target IDs is: [7 6 .. 1 0 15 14 .. 9 8] on a wide-SCSI system. (If you you are wondering why the lower 8 have higher priority, read the previous paragraph for a hint.)

For a further subdivision, the standard allows for Logical Units or LUNs for short. A single target ID may have multiple LUNs. For example, a tape device including a tape changer may have LUN 0 for the tape device itself, and LUN 1 for the tape changer. In this way, the host system can address each of the functional units of the tape changer as desired.

Bus layout

SCSI buses are linear. So, not shaped like Y-junctions, star topologies, rings, cobwebs or whatever else people might want to invent. One of the most common mistakes is for people with wide-SCSI host adapters to connect devices on all three connecters (external connector, internal wide connector, internal narrow connector). Don't do that. It may appear to work if you are really lucky, but I can almost guarantee that your system will stop functioning at the most unfortunate moment (this is also known as "Murphy's law").

You might notice that the terminator issue discussed earlier becomes rather hairy if your bus is not linear. Also, if you have more connectors than devices on your internal SCSI cable, make sure you attach devices on connectors on both ends instead of using the connectors in the middle and let one or both ends dangle. This will screw up the termination of the bus.

The electrical characteristics, its noise margins and ultimately the reliability of it all are tightly related to linear bus rule.

Stick to the linear bus rule!

Using SCSI with FreeBSD

About translations, BIOSes and magic...

As stated before, you should first make sure that you have a electrically sound bus.

When you want to use a SCSI disk on your PC as boot disk, you must aware of some quirks related to PC BIOSes. The PC BIOS in its first incarnation used a low level physical interface to the hard disk. So, you had to tell the BIOS (using a setup tool or a BIOS built-in setup) how your disk physically looked like. This involved stating number of heads, number of cylinders, number of sectors per track, obscure things like precompensation and reduced write current cylinder etc.

One might be inclined to think that since SCSI disks are smart you can forget about this. Alas, the arcane setup issue is still present today. The system BIOS needs to know how to access your SCSI disk with the head/cyl/sector method in order to load the FreeBSD kernel during boot.

The SCSI host adapter or SCSI controller you have put in your AT/EISA/PCI/whatever bus to connect your disk therefore has its own on-board BIOS. During system startup, the SCSI BIOS takes over the hard disk interface routines from the system BIOS. To fool the system BIOS, the system setup is normally set to No hard disk present. Obvious, isn't it?

The SCSI BIOS itself presents to the system a so called translated drive. This means that a fake drive table is constructed that allows the PC to boot the drive. This translation is often (but not always) done using a pseudo drive with 64 heads and 32 sectors per track. By varying the number of cylinders, the SCSI BIOS adapts to the actual drive size. It is useful to note that 32 * 64 / 2 = the size of your drive in megabytes. The division by 2 is to get from disk blocks that are normally 512 bytes in size to Kbytes.

Right. All is well now?! No, it is not. The system BIOS has another quirk you might run into. The number of cylinders of a bootable hard disk cannot be greater than 1024. Using the translation above, this is a show-stopper for disks greater than 1 GB. With disk capacities going up all the time this is causing problems.

Fortunately, the solution is simple: just use another translation, e.g. with 128 heads instead of 32. In most cases new SCSI BIOS versions are available to upgrade older SCSI host adapters. Some newer adapters have an option, in the form of a jumper or software setup selection, to switch the translation the SCSI BIOS uses.

It is very important that all operating systems on the disk use the same translation to get the right idea about where to find the relevant partitions. So, when installing FreeBSD you must answer any questions about heads/cylinders etc using the translated values your host adapter uses.

Failing to observe the translation issue might lead to un-bootable systems or operating systems overwriting each others partitions. Using fdisk you should be able to see all partitions.

You might have heard some talk of 'lying' devices? Older FreeBSD kernels used to report the geometry of SCSI disks when booting. An example from one of my systems:

	aha0 targ 0 lun 0: <MICROP  1588-15MB1057404HSP4>
	sd0: 636MB (1303250 total sec), 1632 cyl, 15 head, 53 sec, bytes/sec 512
          
Newer kernels usually do not report this information. e.g.
	 (bt0:0:0): "SEAGATE ST41651 7574" type 0 fixed SCSI 2
	 sd0(bt0:0:0): Direct-Access 1350MB (2766300 512 byte sectors)
	  

Why has this changed?

This info is retrieved from the SCSI disk itself. Newer disks often use a technique called zone bit recording. The idea is that on the outer cylinders of the drive there is more space so more sectors per track can be put on them. This results in disks that have more tracks on outer cylinders than on the inner cylinders and, last but not least, have more capacity. You can imagine that the value reported by the drive when inquiring about the geometry now becomes suspect at best, and nearly always misleading. When asked for a geometry , it is nearly always better to supply the geometry used by the BIOS, or if the BIOS is never going to know about this disk, (e.g. it is not a booting disk) to supply a fictitious geometry that is convenient.

SCSI subsystem design

FreeBSD uses a layered SCSI subsystem. For each different controller card a device driver is written. This driver knows all the intimate details about the hardware it controls. The driver has a interface to the upper layers of the SCSI subsystem through which it receives its commands and reports back any status.

On top of the card drivers there are a number of more generic drivers for a class of devices. More specific: a driver for tape devices (abbreviation: st), magnetic disks (sd), CD-ROMs (cd) etc. In case you are wondering where you can find this stuff, it all lives in /sys/scsi. See the man pages in section 4 for more details.

The multi level design allows a decoupling of low-level bit banging and more high level stuff. Adding support for another piece of hardware is a much more manageable problem.

Kernel configuration

Dependent on your hardware, the kernel configuration file must contain one or more lines describing your host adapter(s). This includes I/O addresses, interrupts etc. Consult the man page for your adapter driver to get more info. Apart from that, check out /sys/i386/conf/LINT for an overview of a kernel config file. LINT contains every possible option you can dream of. It does not imply LINT will actually get you to a working kernel at all.

Although it is probably stating the obvious: the kernel config file should reflect your actual hardware setup. So, interrupts, I/O addresses etc must match the kernel config file. During system boot messages will be displayed to indicate whether the configured hardware was actually found. Note that most of the EISA/PCI drivers (namely ahb, ahc, ncr and amd will automatically obtain the correct parameters from the host adapters themselves at boot time; thus, you just need to write, for instance, "controller ahc0".

An example loosely based on the FreeBSD 2.2.5-Release kernel config file LINT with some added comments (between []):

		
# SCSI host adapters: `aha', `ahb', `aic', `bt', `nca'
#
# aha: Adaptec 154x
# ahb: Adaptec 174x
# ahc: Adaptec 274x/284x/294x
# aic: Adaptec 152x and sound cards using the Adaptec AIC-6360 (slow!)
# amd: AMD 53c974 based SCSI cards (e.g., Tekram DC-390 and 390T)
# bt: Most Buslogic controllers
# nca: ProAudioSpectrum cards using the NCR 5380 or Trantor T130
# ncr: NCR/Symbios 53c810/815/825/875 etc based SCSI cards
# uha: UltraStore 14F and 34F
# sea: Seagate ST01/02 8 bit controller (slow!)
# wds: Western Digital WD7000 controller (no scatter/gather!).
#

[For an Adaptec AHA274x/284x/294x/394x etc controller]
controller	ahc0

[For an NCR/Symbios 53c875 based controller]
controller	ncr0

[For an Ultrastor adapter]
controller	uha0	at isa? port "IO_UHA0" bio irq ? drq 5 vector uhaintr

# Map SCSI buses to specific SCSI adapters
controller	scbus0	at ahc0
controller	scbus2  at ncr0
controller	scbus1  at uha0

# The actual SCSI devices
disk sd0 at scbus0 target 0 unit 0	[SCSI disk 0 is at scbus 0, LUN 0]
disk sd1 at scbus0 target 1		[implicit LUN 0 if omitted]
disk sd2 at scbus1 target 3		[SCSI disk on the uha0]
disk sd3 at scbus2 target 4		[SCSI disk on the ncr0]
tape st1 at scbus0 target 6		[SCSI tape at target 6]
device cd0 at scbus?			[the first ever CD-ROM found, no wiring]

	  

The example above tells the kernel to look for a ahc (Adaptec 274x) controller, then for an NCR/Symbios board, and so on. The lines following the controller specifications tell the kernel to configure specific devices but only attach them when they match the target ID and LUN specified on the corresponding bus.

Wired down devices get 'first shot' at the unit numbers so the first non 'wired down' device, is allocated the unit number one greater than the highest 'wired down' unit number for that kind of device. So, if you had a SCSI tape at target ID 2 it would be configured as st2, as the tape at target ID 6 is wired down to unit number 1. Note that wired down devices need not be found to get their unit number. The unit number for a wired down device is reserved for that device, even if it is turned off at boot time. This allows the device to be turned on and brought on-line at a later time, without rebooting. Notice that a device's unit number has no relationship with its target ID on the SCSI bus.

Below is another example of a kernel config file as used by FreeBSD version < 2.0.5. The difference with the first example is that devices are not 'wired down'. 'Wired down' means that you specify which SCSI target belongs to which device.

A kernel built to the config file below will attach the first SCSI disk it finds to sd0, the second disk to sd1 etc. If you ever removed or added a disk, all other devices of the same type (disk in this case) would 'move around'. This implies you have to change /etc/fstab each time.

Although the old style still works, you are strongly recommended to use this new feature. It will save you a lot of grief whenever you shift your hardware around on the SCSI buses. So, when you re-use your old trusty config file after upgrading from a pre-FreeBSD2.0.5.R system check this out.

[driver for Adaptec 174x]
controller      ahb0    at isa? bio irq 11 vector ahbintr
[for Adaptec 154x]
controller      aha0    at isa? port "IO_AHA0" bio irq 11 drq 5 vector ahaintr
[for Seagate ST01/02]
controller      sea0    at isa? bio irq 5 iomem 0xc8000 iosiz 0x2000 vector seaintr
controller      scbus0

device          sd0	[support for 4 SCSI harddisks, sd0 up sd3]

device          st0	[support for 2 SCSI tapes]

[for the CD-ROM]
device          cd0     #Only need one of these, the code dynamically grows
          

Both examples support SCSI disks. If during boot more devices of a specific type (e.g. sd disks) are found than are configured in the booting kernel, the system will simply allocate more devices, incrementing the unit number starting at the last number 'wired down'. If there are no 'wired down' devices then counting starts at unit 0.

Use man 4 scsi to check for the latest info on the SCSI subsystem. For more detailed info on host adapter drivers use eg man 4 ahc for info on the Adaptec 294x driver.

Tuning your SCSI kernel setup

Experience has shown that some devices are slow to respond to INQUIRY commands after a SCSI bus reset (which happens at boot time). An INQUIRY command is sent by the kernel on boot to see what kind of device (disk, tape, CD-ROM etc) is connected to a specific target ID. This process is called device probing by the way.

To work around the 'slow response' problem, FreeBSD allows a tunable delay time before the SCSI devices are probed following a SCSI bus reset. You can set this delay time in your kernel configuration file using a line like:

options         SCSI_DELAY=15         #Be pessimistic about Joe SCSI device
	  
This line sets the delay time to 15 seconds. On my own system I had to use 3 seconds minimum to get my trusty old CD-ROM drive to be recognized. Start with a high value (say 30 seconds or so) when you have problems with device recognition. If this helps, tune it back until it just stays working.

Rogue SCSI devices

Although the SCSI standard tries to be complete and concise, it is a complex standard and implementing things correctly is no easy task. Some vendors do a better job then others.

This is exactly where the 'rogue' devices come into view. Rogues are devices that are recognized by the FreeBSD kernel as behaving slightly (...) non-standard. Rogue devices are reported by the kernel when booting. An example for two of my cartridge tape units:

Feb 25 21:03:34 yedi /kernel: ahb0 targ 5 lun 0: <TANDBERG TDC 3600       -06:>
Feb 25 21:03:34 yedi /kernel: st0: Tandberg tdc3600 is a known rogue

Mar 29 21:16:37 yedi /kernel: aha0 targ 5 lun 0: <ARCHIVE VIPER 150  21247-005>
Mar 29 21:16:37 yedi /kernel: st1: Archive  Viper 150 is a known rogue
	 

For instance, there are devices that respond to all LUNs on a certain target ID, even if they are actually only one device. It is easy to see that the kernel might be fooled into believing that there are 8 LUNs at that particular target ID. The confusion this causes is left as an exercise to the reader.

The SCSI subsystem of FreeBSD recognizes devices with bad habits by looking at the INQUIRY response they send when probed. Because the INQUIRY response also includes the version number of the device firmware, it is even possible that for different firmware versions different workarounds are used. See e.g. /sys/scsi/st.c and /sys/scsi/scsiconf.c for more info on how this is done.

This scheme works fine, but keep in mind that it of course only works for devices that are KNOWN to be weird. If you are the first to connect your bogus Mumbletech SCSI CD-ROM you might be the one that has to define which workaround is needed.

After you got your Mumbletech working, please send the required workaround to the FreeBSD development team for inclusion in the next release of FreeBSD. Other Mumbletech owners will be grateful to you.

Multiple LUN devices

In some cases you come across devices that use multiple logical units (LUNs) on a single SCSI ID. In most cases FreeBSD only probes devices for LUN 0. An example are so called bridge boards that connect 2 non-SCSI harddisks to a SCSI bus (e.g. an Emulex MD21 found in old Sun systems).

This means that any devices with LUNs != 0 are not normally found during device probe on system boot. To work around this problem you must add an appropriate entry in /sys/scsi/scsiconf.c and rebuild your kernel.

Look for a struct that is initialized like below:

	  {
                T_DIRECT, T_FIXED, "MAXTOR", "XT-4170S", "B5A",
                "mx1", SC_ONE_LU
          }
	  

For you Mumbletech BRIDGE2000 that has more than one LUN, acts as a SCSI disk and has firmware revision 123 you would add something like:

	  {
                T_DIRECT, T_FIXED, "MUMBLETECH", "BRIDGE2000", "123",
                "sd", SC_MORE_LUS
          }
	  

The kernel on boot scans the inquiry data it receives against the table and acts accordingly. See the source for more info.

Tagged command queueing

Modern SCSI devices, particularly magnetic disks, support what is called tagged command queuing (TCQ).

In a nutshell, TCQ allows the device to have multiple I/O requests outstanding at the same time. Because the device is intelligent, it can optimise its operations (like head positioning) based on its own request queue. On SCSI devices like RAID (Redundant Array of Independent Disks) arrays the TCQ function is indispensable to take advantage of the device's inherent parallelism.

Each I/O request is uniquely identified by a 'tag' (hence the name tagged command queuing) and this tag is used by FreeBSD to see which I/O in the device drivers queue is reported as complete by the device.

It should be noted however that TCQ requires device driver support and that some devices implemented it 'not quite right' in their firmware. This problem bit me once, and it leads to highly mysterious problems. In such cases, try to disable TCQ.

Busmaster host adapters

Most, but not all, SCSI host adapters are bus mastering controllers. This means that they can do I/O on their own without putting load onto the host CPU for data movement.

This is of course an advantage for a multitasking operating system like FreeBSD. It must be noted however that there might be some rough edges.

For instance an Adaptec 1542 controller can be set to use different transfer speeds on the host bus (ISA or AT in this case). The controller is settable to different rates because not all motherboards can handle the higher speeds. Problems like hangups, bad data etc might be the result of using a higher data transfer rate then your motherboard can stomach.

The solution is of course obvious: switch to a lower data transfer rate and try if that works better.

In the case of a Adaptec 1542, there is an option that can be put into the kernel config file to allow dynamic determination of the right, read: fastest feasible, transfer rate. This option is disabled by default:

options        "TUNE_1542"             #dynamic tune of bus DMA speed
	  

Check the man pages for the host adapter that you use. Or better still, use the ultimate documentation (read: driver source).

Tracking down problems

The following list is an attempt to give a guideline for the most common SCSI problems and their solutions. It is by no means complete.

Further reading

If you intend to do some serious SCSI hacking, you might want to have the official standard at hand:

Approved American National Standards can be purchased from ANSI at 11 West 42nd Street, 13th Floor, New York, NY 10036, Sales Dept: (212) 642-4900. You can also buy many ANSI standards and most committee draft documents from Global Engineering Documents, 15 Inverness Way East, Englewood, CO 80112-5704, Phone: (800) 854-7179, Outside USA and Canada: (303) 792-2181, FAX: (303) 792- 2192.

Many X3T10 draft documents are available electronically on the SCSI BBS (719-574-0424) and on the ncrinfo.ncr.com anonymous ftp site.

Latest X3T10 committee documents are:

Other publications that might provide you with additional information are:

On Usenet the newsgroups comp.periphs.scsi and comp.periphs are noteworthy places to look for more info. You can also find the SCSI-Faq there, which is posted periodically.

Most major SCSI device and host adapter suppliers operate ftp sites and/or BBS systems. They may be valuable sources of information about the devices you own.