ibmmca.txt

linux 内核源代码

   1) Finally, after some 3.1Beta-releases, the 3.1 release. Sorry, for 
      the delayed release, but it was not finished with the release of 
      Kernel 2.2.0.
   - Michael Lang
   
   Feb 10, 1999 (v3.1)
   1) Added a new commandline parameter called 'bypass' in order to bypass
      every integrated subsystem SCSI-command consequently in case of
      troubles.
   2) Concatenated read_capacity requests to the harddisks. It gave a lot
      of troubles with some controllers and after I wanted to apply some
      extensions, it jumped out in the same situation, on my w/cache, as like 
      on D. Weinehalls' Model 56, having integrated SCSI. This gave me the 
      decisive hint to move the code-part out and declare it global. Now
      it seems to work far better and more stable. Let us see what
      the world thinks of it...
   3) By the way, only Sony DAT-drives seem to show density code 0x13. A
      test with a HP drive gave right results, so the problem is vendor-
      specific and not a problem of the OS or the driver.
   - Michael Lang
   
   Feb 18, 1999 (v3.1d)
   1) The abort command and the reset function have been checked for 
      inconsistencies. From the logical point of thinking, they work
      at their optimum, now, but as the subsystem does not answer with an
      interrupt, abort never finishes, sigh...
   2) Everything, that is accessed by a busmaster request from the adapter
      is now declared as global variable, even the return-buffer in the
      local checking phase. This assures, that no accesses to undefined memory
      areas are performed.
   3) In ibmmca.h, the line unchecked_isa_dma is added with 1 in order to
      avoid memory-pointers for the areas higher than 16MByte in order to
      be sure, it also works on 16-Bit Microchannel bus systems.
   4) A lot of small things have been found, but nothing that endangered the
      driver operations. Just it should be more stable, now.
   - Michael Lang
      
   Feb 20, 1999 (v3.1e)
   1) I took the warning from the Linux Kernel Hackers Guide serious and 
      checked the cmd->result return value to the done-function very carefully.
      It is obvious, that the IBM SCSI only delivers the tsb.dev_status, if
      some error appeared, else it is undefined. Now, this is fixed. Before
      any SCB command gets queued, the tsb.dev_status is set to 0, so the 
      cmd->result won't screw up Linux higher level drivers.
   2) The reset-function has slightly improved. This is still planed for 
      abort. During the abort and the reset function, no interrupts are 
      allowed. This is however quite hard to cope with, so the INT-status
      register is read. When the interrupt gets queued, one can find its
      status immediately on that register and is enabled to continue in the
      reset function. I had no chance to test this really, only in a bogus 
      situation, I got this function running, but the situation was too much
      worse for Linux :-(, so tests will continue.
   3) Buffers got now consistent. No open address mapping, as before and
      therefore no further troubles with the unassigned memory segmentation
      faults that scrambled probes on 95XX series and even on 85XX series,
      when the kernel is done in a not so perfectly fitting way.
   4) Spontaneous interrupts from the subsystem, appearing without any
      command previously queued are answered with a DID_BAD_INTR result.
   5) Taken into account ZP Gus' proposals to reverse the SCSI-device
      scan order. As it does not work on Kernel 2.1.x or 2.2.x, as proposed
      by him, I implemented it in a slightly derived way, which offers in 
      addition more flexibility.
   - Michael Lang

   Apr 23, 2000 (v3.2pre1)
   1) During a very long time, I collected a huge amount of bug reports from
      various people, trying really quite different things on their SCSI-
      PS/2s. Today, all these bug reports are taken into account and should be
      mostly solved. The major topics were:
      - Driver crashes during boottime by no obvious reason.
      - Driver panics while the midlevel-SCSI-driver is trying to inquire
        the SCSI-device properties, even though hardware is in perfect state.
      - Displayed info for the various slot-cards is interpreted wrong.
      The main reasons for the crashes were two:
      1) The commands to check for device information like INQUIRY, 
         TEST_UNIT_READY, REQUEST_SENSE and MODE_SENSE cause the devices
	 to deliver information of up to 255 bytes. Midlevel drivers offer
	 1024 bytes of space for the answer, but the IBM-SCSI-adapters do
	 not accept this, as they stick quite near to ANSI-SCSI and report
	 a COMMAND_ERROR message which causes the driver to panic. The main
	 problem was located around the INQUIRY command. Now, for all the
	 mentioned commands, the buffersize sent to the adapter is at 
	 maximum 255 which seems to be a quite reasonable solution. 
	 TEST_UNIT_READY gets a buffersize of 0 to make sure that no 
	 data is transferred in order to avoid any possible command failure.
      2) On unsuccessful TEST_UNIT_READY, the mid-level driver has to send
         a REQUEST_SENSE in order to see where the problem is located. This
	 REQUEST_SENSE may have various length in its answer-buffer. IBM
	 SCSI-subsystems report a command failure if the returned buffersize
	 is different from the sent buffersize, but this can be suppressed by
	 a special bit, which is now done and problems seem to be solved.
   2) Code adaption to all kernel-releases. Now, the 3.2 code compiles on 
      2.0.x, 2.1.x, 2.2.x and 2.3.x kernel releases without any code-changes.
   3) Commandline-parameters are recognized again, even under Kernel 2.3.x or
      higher.
   - Michael Lang   

   April 27, 2000 (v3.2pre2)
   1) Bypassed commands get read by the adapter by one cycle instead of two.
      This increases SCSI-performance.
   2) Synchronous datatransfer is provided for sure to be 5 MHz on older
      SCSI and 10 MHz on internal F/W SCSI-adapter.
   3) New commandline parameters allow to force the adapter to slow down while
      in synchronous transfer. Could be helpful for very old devices.
   - Michael Lang
   
   June 2, 2000 (v3.2pre5)
   1) Added Jim Shorney's contribution to make the activity indicator
      flashing in addition to the LED-alphanumeric display-panel on
      models 95A. To be enabled to choose this feature freely, a new
      commandline parameter is added, called 'activity'.
   2) Added the READ_CONTROL bit for test_unit_ready SCSI-command.
   3) Added some suppress_exception bits to read_device_capacity and
      all device_inquiry occurrences in the driver code.
   4) Complaints about the various KERNEL_VERSION implementations are
      taken into account. Every local_LinuxKernelVersion occurrence is
      now replaced by KERNEL_VERSION, defined in linux/version.h. 
      Corresponding changes were applied to ibmmca.h, too. This was a
      contribution to all kernel-parts by Philipp Hahn.
   - Michael Lang
   
   July 17, 2000 (v3.2pre8)
   A long period of collecting bug reports from all corners of the world
   now lead to the following corrections to the code:
   1) SCSI-2 F/W support crashed with a COMMAND ERROR. The reason for this 
      was that it is possible to disable Fast-SCSI for the external bus.
      The feature-control command, where this crash appeared regularly, tried
      to set the maximum speed of 10MHz synchronous transfer speed and that
      reports a COMMAND ERROR if external bus Fast-SCSI is disabled. Now,
      the feature-command probes down from maximum speed until the adapter 
      stops to complain, which is at the same time the maximum possible
      speed selected in the reference program. So, F/W external can run at
      5 MHz (slow-) or 10 MHz (fast-SCSI). During feature probing, the 
      COMMAND ERROR message is used to detect if the adapter does not complain.
   2) Up to now, only combined busmode is supported, if you use external
      SCSI-devices, attached to the F/W-controller. If dual bus is selected,
      only the internal SCSI-devices get accessed by Linux. For most 
      applications, this should do fine. 
   3) Wide-SCSI-addressing (16-Bit) is now possible for the internal F/W
      bus on the F/W adapter. If F/W adapter is detected, the driver
      automatically uses the extended PUN/LUN <-> LDN mapping tables, which
      are now new from 3.2pre8. This allows PUNs between 0 and 15 and should
      provide more fun with the F/W adapter.
   4) Several machines use the SCSI: POS registers for internal/undocumented
      storage of system relevant info. This confused the driver, mainly on
      models 9595, as it expected no onboard SCSI only, if all POS in
      the integrated SCSI-area are set to 0x00 or 0xff. Now, the mechanism
      to check for integrated SCSI is much more restrictive and these problems
      should be history.
   - Michael Lang          

   July 18, 2000 (v3.2pre9)
   This develop rather quickly at the moment. Two major things were still
   missing in 3.2pre8:
   1) The adapter PUN for F/W adapters has 4-bits, while all other adapters
      have 3-bits. This is now taken into account for F/W.
   2) When you select CONFIG_IBMMCA_SCSI_ORDER_STANDARD, you should 
      normally get the inverse probing order of your devices on the SCSI-bus.
      The ANSI device order gets scrambled in version 3.2pre8!! Now, a new
      and tested algorithm inverts the device-order on the SCSI-bus and
      automatically avoids accidental access to whatever SCSI PUN the adapter 
      is set and works with SCSI- and Wide-SCSI-addressing.
   - Michael Lang

   July 23, 2000 (v3.2pre10 unpublished) 
   1) LED panel display supports wide-addressing in ibmmca=display mode.
   2) Adapter-information and autoadaption to address-space is done.
   3) Auto-probing for maximum synchronous SCSI transfer rate is working.
   4) Optimization to some embedded function calls is applied.
   5) Added some comment for the user to wait for SCSI-devices being probed.
   6) Finished version 3.2 for Kernel 2.4.0. It least, I thought it is but...
   - Michael Lang
   
   July 26, 2000 (v3.2pre11)
   1) I passed a horrible weekend getting mad with NMIs on kernel 2.2.14 and
      a model 9595. Asking around in the community, nobody except of me has
      seen such errors. Weird, but I am trying to recompile everything on
      the model 9595. Maybe, as I use a specially modified gcc, that could
      cause problems. But, it was not the reason. The true background was,
      that the kernel was compiled for i386 and the 9595 has a 486DX-2. 
      Normally, no troubles should appear, but for this special machine,
      only the right processor support is working fine!
   2) Previous problems with synchronous speed, slowing down from one adapter 
      to the next during probing are corrected. Now, local variables store
      the synchronous bitmask for every single adapter found on the MCA bus.
   3) LED alphanumeric panel support for XX95 systems is now showing some
      alive rotator during boottime. This makes sense, when no monitor is 
      connected to the system. You can get rid of all display activity, if
      you do not use any parameter or just ibmmcascsi=activity, for the 
      harddrive activity LED, existent on all PS/2, except models 8595-XXX.
      If no monitor is available, please use ibmmcascsi=display, which works
      fine together with the linuxinfo utility for the LED-panel.
   - Michael Lang
   
   July 29, 2000 (v3.2)
   1) Submission of this driver for kernel 2.4test-XX and 2.2.17.
   - Michael Lang
   
   December 28, 2000 (v3.2d / v4.0)
   1) The interrupt handler had some wrong statement to wait for. This
      was done due to experimental reasons during 3.2 development but it
      has shown that this is not stable enough. Going back to wait for the
      adapter to be not busy is best.
   2) Inquiry requests can be shorter than 255 bytes of return buffer. Due
      to a bug in the ibmmca_queuecommand routine, this buffer was forced
      to 255 at minimum. If the memory address, this return buffer is pointing
      to does not offer more space, invalid memory accesses destabilized the
      kernel.
   3) version 4.0 is only valid for kernel 2.4.0 or later. This is necessary
      to remove old kernel version dependent waste from the driver. 3.2d is
      only distributed with older kernels but keeps compatibility with older
      kernel versions. 4.0 and higher versions cannot be used with older 
      kernels anymore!! You must have at least kernel 2.4.0!!
   4) The commandline argument 'bypass' and all its functionality got removed
      in version 4.0. This was never really necessary, as all troubles were
      based on non-command related reasons up to now, so bypassing commands
      did not help to avoid any bugs. It is kept in 3.2X for debugging reasons.
   5) Dynamical reassignment of ldns was again verified and analyzed to be
      completely inoperational. This is corrected and should work now.
   6) All commands that get sent to the SCSI adapter were verified and
      completed in such a way, that they are now completely conform to the
      demands in the technical description of IBM. Main candidates were the
      DEVICE_INQUIRY, REQUEST_SENSE and DEVICE_CAPACITY commands. They must
      be transferred by bypassing the internal command buffer of the adapter
      or else the response can be a random result. GET_POS_INFO would be more
      safe in usage, if one could use the SUPRESS_EXCEPTION_SHORT, but this
      is not allowed by the technical references of IBM. (Sorry, folks, the
      model 80 problem is still a task to be solved in a different way.)
   7) v3.2d is still hold back for some days for testing, while 4.0 is 
      released.
   - Michael Lang
   
   January 3, 2001 (v4.0a)