libata: fix internal command failure handling

[linux-2.6/verdex.git] / Documentation / scsi / tmscsim.txt

The tmscsim driver
==================

1. Purpose and history
2. Installation
3. Features
4. Configuration via /proc/scsi/tmscsim/?
5. Configuration via boot/module params
6. Potential improvements
7. Bug reports, debugging and updates
8. Acknowledgements
9. Copyright


1. Purpose and history
----------------------
The tmscsim driver supports PCI SCSI Host Adapters based on the AM53C974
chip. AM53C974 based SCSI adapters include: 
 Tekram DC390, DC390T
 Dawicontrol 2974
 QLogic Fast! PCI Basic
 some on-board adapters
(This is most probably not a complete list)

It has originally written by C.L. Huang from the Tekram corp. to support the
Tekram DC390(T) adapter. This is where the name comes from: tm = Tekram
scsi = SCSI driver, m = AMD (?) as opposed to w for the DC390W/U/F
(NCR53c8X5, X=2/7) driver. Yes, there was also a driver for the latter,
tmscsiw, which supported DC390W/U/F adapters. It's not maintained any more,
as the ncr53c8xx is perfectly supporting these adapters since some time.

The driver first appeared in April 1996, exclusively supported the DC390 
and has been enhanced since then in various steps. In May 1998 support for 
general AM53C974 based adapters and some possibilities to configure it were
added. The non-DC390 support works by assuming some values for the data
normally taken from the DC390 EEPROM. See below (chapter 5) for details.

When using the DC390, the configuration is still be done using the DC390
BIOS setup. The DC390 EEPROM is read and used by the driver, any boot or
module parameters (chapter 5) are ignored! However, you can change settings
dynamically, as described in chapter 4. 

For a more detailed description of the driver's history, see the first lines
of tmscsim.c.
The numbering scheme isn't consistent. The first versions went from 1.00 to
1.12, then 1.20a to 1.20t. Finally I decided to use the ncr53c8xx scheme. So
the next revisions will be 2.0a to 2.0X (stable), 2.1a to 2.1X (experimental),
2.2a to 2.2X (stable, again) etc. (X = anything between a and z.) If I send
fixes to people for testing, I create intermediate versions with a digit 
appended, e.g. 2.0c3.


2. Installation
---------------
If you got any recent kernel with this driver and document included in
linux/drivers/scsi, you basically have to do nothing special to use this
driver. Of course you have to choose to compile SCSI support and DC390(T)
support into your kernel or as module when configuring your kernel for
compiling.
NEW: You may as well compile this module outside your kernel, using the
supplied Makefile.

 If you got an old kernel (pre 2.1.127, pre 2.0.37p1) with an old version of
 this driver: Get dc390-21125-20b.diff.gz or dc390-2036p21-20b1.diff.gz from
 my web page and apply the patch. Apply further patches to upgrade to the 
 latest version of the driver.

 If you want to do it manually, you should copy the files (dc390.h,
 tmscsim.h, tmscsim.c, scsiiom.c and README.tmscsim) from this directory to
 linux/drivers/scsi. You have to recompile your kernel/module of course.

 You should apply the three patches included in dc390-120-kernel.diff
 (Applying them: cd /usr/src; patch -p0 <~/dc390-120-kernel.diff)
 The patches are against 2.1.125, so you might have to manually resolve
 rejections when applying to another kernel version.

 The patches will update the kernel startup code to allow boot parameters to
 be passed to the driver, update the Documentation and finally offer you the
 possibility to omit the non-DC390 parts of the driver.
 (By selecting "Omit support for non DC390" you basically disable the
 emulation of a DC390 EEPROM for non DC390 adapters. This saves a few bytes
 of memory.)

If you got a very old kernel without the tmscsim driver (pre 2.0.31)
I recommend upgrading your kernel. However, if you don't want to, please
contact me to get the appropriate patches.


Upgrading a SCSI driver is always a delicate thing to do. The 2.0 driver has
proven stable on many systems, but it's still a good idea to take some
precautions. In an ideal world you would have a full backup of your disks.
The world isn't ideal and most people don't have full backups (me neither).
So take at least the following measures:
* make your kernel remount the FS read-only on detecting an error:
  tune2fs -e remount-ro /dev/sd??
* have copies of your SCSI disk's partition tables on some safe location:
  dd if=/dev/sda of=/mnt/floppy/sda bs=512 count=1
  or just print it with:
  fdisk -l | lpr
* make sure you are able to boot Linux (e.g. from floppy disk using InitRD)
  if your SCSI disk gets corrupted. You can use 
  ftp://student.physik.uni-dortmund.de/pub/linux/kernel/bootdisk.gz

One more warning: I used to overclock my PCI bus to 41.67 MHz. My Tekram
DC390F (Sym53c875) accepted this as well as my Millenium. But the Am53C974
produced errors and started to corrupt my disks. So don't do that! A 37.50
MHz PCI bus works for me, though, but I don't recommend using higher clocks
than the 33.33 MHz being in the PCI spec.

If you want to share the IRQ with another device and the driver refuses to
do so, you might succeed with changing the DC390_IRQ type in tmscsim.c to 
IRQF_SHARED | IRQF_DISABLED.


3.Features
----------
- SCSI
 * Tagged command queueing
 * Sync speed up to 10 MHz
 * Disconnection
 * Multiple LUNs

- General / Linux interface
 * Support for up to 4 AM53C974 adapters.
 * DC390 EEPROM usage or boot/module params
 * Information via cat /proc/scsi/tmscsim/?
 * Dynamically configurable by writing to /proc/scsi/tmscsim/?
 * Dynamic allocation of resources
 * SMP support: Locking on io_request lock (Linux 2.1/2.2) or adapter 
    specific locks (Linux 2.5?)
 * Uniform source code for Linux-2.x.y
 * Support for dyn. addition/removal of devices via add/remove-single-device
   (Try: echo "scsi add-single-device C B T U" >/proc/scsi/scsi 
    C = Controller, B = Bus, T = Target SCSI ID, U = Unit SCSI LUN.) 
    Use with care!
 * Try to use the partition table for the determination of the mapping


4. Configuration via /proc/scsi/tmscsim/?
-----------------------------------------
First of all look at the output of /proc/scsi/tmscsim/? by typing
 cat /proc/scsi/tmscsim/?
The "?" should be replaced by the SCSI host number. (The shell might do this
for you.)
You will see some info regarding the adapter and, at the end, a listing of
the attached devices and their settings.

Here's an example:
garloff@kurt:/home/garloff > cat /proc/scsi/tmscsim/0
Tekram DC390/AM53C974 PCI SCSI Host Adapter, Driver Version 2.0e7 2000-11-28
SCSI Host Nr 1, AM53C974 Adapter Nr 0
IOPortBase 0xb000, IRQ 10
MaxID 8, MaxLUN 8, AdapterID 6, SelTimeout 250 ms, DelayReset 1 s
TagMaxNum 16, Status 0x00, ACBFlag 0x00, GlitchEater 24 ns
Statistics: Cmnds 1470165, Cmnds not sent directly 0, Out of SRB conds 0
            Lost arbitrations 587,  Sel. connected 0, Connected: No
Nr of attached devices: 4, Nr of DCBs: 4
Map of attached LUNs: 01 00 00 03 01 00 00 00
Idx ID LUN Prty Sync DsCn SndS TagQ NegoPeriod SyncSpeed SyncOffs MaxCmd
00  00  00  Yes  Yes  Yes  Yes  Yes   100 ns    10.0 M      15      16
01  03  00  Yes  Yes  Yes  Yes  No    100 ns    10.0 M      15      01
02  03  01  Yes  Yes  Yes  Yes  No    100 ns    10.0 M      15      01
03  04  00  Yes  Yes  Yes  Yes  No    100 ns    10.0 M      15      01

Note that the settings MaxID and MaxLUN are not zero- but one-based, which
means that a setting MaxLUN=4, will result in the support of LUNs 0..3. This
is somehow inconvenient, but the way the mid-level SCSI code expects it to be.

ACB and DCB are acronyms for Adapter Control Block and Device Control Block.
These are data structures of the driver containing information about the
adapter and the connected SCSI devices respectively.

Idx is the device index (just a consecutive number for the driver), ID and
LUN are the SCSI ID and LUN, Prty means Parity checking, Sync synchronous
negotiation, DsCn Disconnection, SndS Send Start command on startup (not
used by the driver) and TagQ Tagged Command Queueing. NegoPeriod and
SyncSpeed are somehow redundant, because they are reciprocal values 
(1 / 112 ns = 8.9 MHz). At least in theory. The driver is able to adjust the
NegoPeriod more accurate (4ns) than the SyncSpeed (1 / 25ns). I don't know
if certain devices will have problems with this discrepancy. Max. speed is
10 MHz corresp. to a min. NegoPeriod of 100 ns. 
(The driver allows slightly higher speeds if the devices (Ultra SCSI) accept
it, but that's out of adapter spec, on your own risk and unlikely to improve
performance. You're likely to crash your disks.) 
SyncOffs is the offset used for synchronous negotiations; max. is 15. 
The last values are only shown, if Sync is enabled. (NegoPeriod is still
displayed in brackets to show the values which will be used after enabling
Sync.)
MaxCmd ist the number of commands (=tags) which can be processed at the same
time by the device.

If you want to change a setting, you can do that by writing to
/proc/scsi/tmscsim/?. Basically you have to imitate the output of driver.
(Don't use the brackets for NegoPeriod on Sync disabled devices.)
You don't have to care about capitalisation. The driver will accept space,
tab, comma, = and : as separators.

There are three kinds of changes: 

(1) Change driver settings: 
    You type the names of the parameters and the params following it.
    Example:
     echo "MaxLUN=8 seltimeout 200" >/proc/scsi/tmscsim/0

    Note that you can only change MaxID, MaxLUN, AdapterID, SelTimeOut,
    TagMaxNum, ACBFlag, GlitchEater and DelayReset. Don't change ACBFlag
    unless you want to see what happens, if the driver hangs.

(2) Change device settings: You write a config line to the driver. The Nr
    must match the ID and LUN given. If you give "-" as parameter, it is
    ignored and the corresponding setting won't be changed. 
    You can use "y" or "n" instead of "Yes" and "No" if you want to.
    You don't need to specify a full line. The driver automatically performs
    an INQUIRY on the device if necessary to check if it is capable to operate
    with the given settings (Sync, TagQ).
    Examples:
     echo "0 0 0 y y y - y - 10 " >/proc/scsi/tmscsim/0
     echo "3 5 0 y n y " >/proc/scsi/tmscsim/0

    To give a short explanation of the first example: 
    The first three numbers, "0 0 0" (Device index 0, SCSI ID 0, SCSI LUN 0),
    select the device to which the following parameters apply. Note that it
    would be sufficient to use the index or both SCSI ID and LUN, but I chose
    to require all three to have a syntax similar to the output.
    The following "y y y - y" enables Parity checking, enables Synchronous
    transfers, Disconnection, leaves Send Start (not used) untouched and
    enables Tagged Command Queueing for the selected device. The "-" skips
    the Negotiation Period setting but the "10" sets the max sync. speed to
    10 MHz. It's useless to specify both NegoPeriod and SyncSpeed as
    discussed above. The values used in this example will result in maximum
    performance.

(3) Special commands: You can force a SCSI bus reset, an INQUIRY command, the
    removal or the addition of a device's DCB and a SCSI register dump.
    This is only used for debugging when you meet problems. The parameter of
    the INQUIRY and REMOVE commands is the device index as shown by the
    output of /proc/scsi/tmscsim/? in the device listing in the first column
    (Idx). ADD takes the SCSI ID and LUN.
    Examples:
     echo "reset" >/proc/scsi/tmscsim/0
     echo "inquiry 1" >/proc/scsi/tmscsim/0
     echo "remove 2" >/proc/scsi/tmscsim/1
     echo "add 2 3" >/proc/scsi/tmscsim/?
     echo "dump" >/proc/scsi/tmscsim/0

    Note that you will meet problems when you REMOVE a device's DCB with the
    remove command if it contains partitions which are mounted. Only use it
    after unmounting its partitions, telling the SCSI mid-level code to
    remove it (scsi remove-single-device) and you really need a few bytes of
    memory.
    The ADD command allows you to configure a device before you tell the
    mid-level code to try detection.


I'd suggest reviewing the output of /proc/scsi/tmscsim/? after changing
settings to see if everything changed as requested.


5. Configuration via boot/module parameters
-------------------------------------------
With the DC390, the driver reads its EEPROM settings and tries to use them.
But you may want to override the settings prior to being able to change the
driver configuration via /proc/scsi/tmscsim/?.
If you do have another AM53C974 based adapter, that's even the only
possibility to adjust settings before you are able to write to the
/proc/scsi/tmscsim/? pseudo-file, e.g. if you want to use another 
adapter ID than 7.  
(BTW, the log message "DC390: No EEPROM found!" is normal without a DC390.)
For this purpose, you can pass options to the driver before it is initialised
by using kernel or module parameters. See lilo(8) or modprobe(1) manual
pages on how to pass params to the kernel or a module.
[NOTE: Formerly, it was not possible to override the EEPROM supplied
 settings of the DC390 with cmd line parameters. This has changed since
 2.0e7]

The syntax of the params is much shorter than the syntax of the /proc/...
interface. This makes it a little bit more difficult to use. However, long
parameter lines have the risk to be misinterpreted and the length of kernel
parameters is limited.

As the support for non-DC390 adapters works by simulating the values of the
DC390 EEPROM, the settings are given in a DC390 BIOS' way.

Here's the syntax:
tmscsim=AdaptID,SpdIdx,DevMode,AdaptMode,TaggedCmnds,DelayReset

Each of the parameters is a number, containing the described information:

* AdaptID: The SCSI ID of the host adapter. Must be in the range 0..7
  Default is 7.

* SpdIdx: The index of the maximum speed as in the DC390 BIOS. The values
  0..7 mean 10, 8.0, 6.7, 5.7, 5.0, 4.0, 3.1 and 2 MHz resp. Default is
  0 (10.0 MHz).

* DevMode is a bit mapped value describing the per-device features. It
  applies to all devices. (Sync, Disc and TagQ will only apply, if the
  device supports it.) The meaning of the bits (* = default):

   Bit Val(hex) Val(dec)  Meaning
   *0    0x01       1     Parity check
   *1    0x02       2     Synchronous Negotiation
   *2    0x04       4     Disconnection
   *3    0x08       8     Send Start command on startup. (Not used)
   *4    0x10      16     Tagged Command Queueing

  As usual, the desired value is obtained by adding the wanted values. If
  you want to enable all values, e.g., you would use 31(0x1f). Default is 31.

* AdaptMode is a bit mapped value describing the enabled adapter features.

   Bit Val(hex) Val(dec)  Meaning
   *0    0x01       1     Support more than two drives. (Not used)
   *1    0x02       2     Use DOS compatible mapping for HDs greater than 1GB.
   *2    0x04       4     Reset SCSI Bus on startup.
   *3    0x08       8     Active Negation: Improves SCSI Bus noise immunity.
    4    0x10      16     Immediate return on BIOS seek command. (Not used)
 (*)5    0x20      32     Check for LUNs >= 1.
  
  The default for LUN Check depends on CONFIG_SCSI_MULTI_LUN.

* TaggedCmnds is a number indicating the maximum number of Tagged Commands.
  It is the binary logarithm - 1 of the actual number. Max is 4 (32).
   Value  Number of Tagged Commands
     0           2
     1           4
     2           8
    *3          16
     4          32

* DelayReset is the time in seconds (minus 0.5s), the adapter waits, after a
  bus reset. Default is 1 (corresp. to 1.5s).

Example:
 modprobe tmscsim tmscsim=6,2,31
would set the adapter ID to 6, max. speed to 6.7 MHz, enable all device
features and leave the adapter features, the number of Tagged Commands
and the Delay after a reset to the defaults.

As you can see, you don't need to specify all of the six params.
If you want values to be ignored (i.e. the EEprom settings or the defaults
will be used), you may pass -2 (not 0!) at the corresponding position.

The defaults (7,0,31,15,3,1) are aggressive to allow good performance. You
can use tmscsim=7,0,31,63,4,0 for maximum performance, if your SCSI chain
allows it. If you meet problems, you can use tmscsim=-1 which is a shortcut
for tmscsim=7,4,9,15,2,10.


6. Potential improvements
-------------------------
Most of the intended work on the driver has been done. Here are a few ideas
to further improve its usability:

* Cleanly separate per-Target and per-LUN properties (DCB)
* More intelligent abort() routine
* Use new_eh code (Linux-2.1+)
* Have the mid-level (ML) code (and not the driver) handle more of the
  various conditions.
* Command queueing in the driver: Eliminate Query list and use ML instead.
* More user friendly boot/module param syntax

Further investigation on these problems:

* Driver hangs with sync readcdda (xcdroast) (most probably VIA PCI error)

Known problems: 
Please see http://www.garloff.de/kurt/linux/dc390/problems.html

* Changing the parameters of multi-lun by the tmscsim/? interface will
  cause problems, cause these settings are mostly per Target and not per LUN
  and should be updated accordingly. To be fixed for 2.0d24.
* CDRs (eg Yam CRW4416) not recognized, because some buggy devices don't 
  recover from a SCSI reset in time. Use a higher delay or don't issue
  a SCSI bus reset on driver initialization. See problems page.
  For the CRW4416S, this seems to be solved with firmware 1.0g (reported by 
  Jean-Yves Barbier).
* TEAC CD-532S not being recognized. (Works with 1.11).
* Scanners (eg. Astra UMAX 1220S) don't work: Disable Sync Negotiation.
  If this does not help, try echo "INQUIRY t" >/proc/scsi/tmscsim/? (t
  replaced by the dev index of your scanner). You may try to reset your SCSI
  bus afterwards (echo "RESET" >/proc/scsi/tmscsim/?).
  The problem seems to be solved as of 2.0d18, thanks to Andreas Rick.
* If there is a valid partition table, the driver will use it for determining
  the mapping. If there's none, a reasonable mapping (Symbios-like) will be
  assumed. Other operating systems may not like this mapping, though
  it's consistent with the BIOS' behaviour. Old DC390 drivers ignored the
  partition table and used a H/S = 64/32 or 255/63 translation. So if you
  want to be compatible to those, use this old mapping when creating
  partition tables. Even worse, on bootup the DC390 might complain if other
  mappings are found, so auto rebooting may fail.
* In some situations, the driver will get stuck in an abort loop. This is a
  bad interaction between the Mid-Layer of Linux' SCSI code and the driver.
  Try to disable DsCn, if you meet this problem. Please contact me for
  further debugging.


7. Bug reports, debugging and updates
-------------------------------------
Whenever you have problems with the driver, you are invited to ask the
author for help. However, I'd suggest reading the docs and trying to solve
the problem yourself, first. 
If you find something, which you believe to be a bug, please report it to me. 
Please append the output of /proc/scsi/scsi, /proc/scsi/tmscsim/? and
maybe the DC390 log messages to the report. 

Bug reports should be send to me (Kurt Garloff <dc390@garloff.de>) as well
as to the linux-scsi list (<linux-scsi@vger.kernel.org>), as sometimes bugs
are caused by the SCSI mid-level code.

I will ask you for some more details and probably I will also ask you to
enable some of the DEBUG options in the driver (tmscsim.c:DC390_DEBUGXXX
defines). The driver will produce some data for the syslog facility then.
Beware: If your syslog gets written to a SCSI disk connected to your
AM53C974, the logging might produce log output again, and you might end
having your box spending most of its time doing the logging.

The latest version of the driver can be found at:
 http://www.garloff.de/kurt/linux/dc390/
 ftp://ftp.suse.com/pub/people/garloff/linux/dc390/


8. Acknowledgements
-------------------
Thanks to Linus Torvalds, Alan Cox, the FSF people, the XFree86 team and 
all the others for the wonderful OS and software.
Thanks to C.L. Huang and Philip Giang (Tekram) for the initial driver
release and support.
Thanks to Doug Ledford, Gérard Roudier for support with SCSI coding.
Thanks to a lot of people (espec. Chiaki Ishikawa, Andreas Haumer, Hubert 
Tonneau) for intensively testing the driver (and even risking data loss
doing this during early revisions).
Recently, SuSE GmbH, Nuernberg, FRG, has been paying me for the driver
development and maintenance. Special thanks!


9. Copyright
------------
 This driver is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by   
 the Free Software Foundation; version 2 of the License.
 If you want to use any later version of the GNU GPL, you will probably
 be allowed to, but you have to ask me and Tekram <erich@tekram.com.tw>
 before.

-------------------------------------------------------------------------
Written by Kurt Garloff <kurt@garloff.de> 1998/06/11
Last updated 2000/11/28, driver revision 2.0e7
$Id: README.tmscsim,v 2.25.2.7 2000/12/20 01:07:12 garloff Exp $