From: alsaggaf@gnu.ai.mit.edu (Muhammad Saggaf)
Subject: Work-around for docked NEC IDE driver trouble

I had a bit of a problem yesterday running linux on a docked NEC
notebook, and I thought to share the solution with you. Let me
elaborate a bit on the setup. For systems with two IDE controllers,
CMOS on most motherbaords lists two entries for devices connected
to the primary controller (devices 1 and 2), and two entries for the
devices connected to the secondary controller (devices 3 and 4). The linux
IDE driver obtains the geometries of the first two devices from CMOS
and assigns them hda and hdb, and probes for the devices connected to the
secondary and assigns them hdc and hdd.

However, the driver apparently makes a certain assumption that
although valid in the majority of cases, it is not valiad in all
situations. That assumption is that "the first two devices in CMOS" is
the same as "the devices connected to the primary IDE controller".
Unfortunately, this is not always true. For example, a docked NEC
notebook has only 3 entries in CMOS, the first lists the device
connected to the primary IDE controller (the controller inside the
notebook), and the other two list the devices connected to the
secondary controller (in the docking station). Note the difference
between this and a typical motherboard:


Typical:                             This case:

                 --- Entry 1                        --- Entry 1
Primary IDE   --|                    Primary IDE --|
                 --- Entry 2                        --- No entry

                 --- Entry 3                        --- Entry 2
Secondary IDE --|                    Primary IDE --|
                 --- Entry 4                        --- Entry 3

So what happens in this case? The kernel obtains the geometries for
the first two devices from CMOS, and gets entries 1 and 2. However,
entry 2 is on the secondary controller, and so hdb gets a bogus
device. It then probes the secondary controller, and gets the
equivalent of entries 2 and 3. Having a bogus device in hdb confuses
the system though, and could break havoc on the non-bogus device,
since the two refer to the same drive:

      Console: colour EGA+ 80x25, 1 virtual console (max 63)
      Calibrating delay loop.. ok - 25.04 BogoMips
      Serial driver version 4.11 with no serial options enabled
      tty00 at 0x03f8 (irq = 4) is a 16450
      lp1 at 0x0378, using polling driver
      PS/2 auxiliary pointing device detected -- driver installed.
      ftape: allocated 3 buffers aligned at: 00210000
      hda: DHAA-2405, 328MB w/32KB Cache, LBA, CHS=915/15/49, MaxMult=16
>>    hdb: non-IDE device, CHS=1010/12/55
      hdc: WDC AC2340H, 325MB w/128KB Cache, CHS=1010/12/55, MaxMult=16
      ide1: secondary interface on irq 15
      ide0: primary interface on irq 14
      Floppy drive(s): fd0 is 1.44M
      ...
      Linux version 1.2.8 (root@bigkitty) (gcc version 2.6.3) #1 Sun May 7
      12:24:55 CDT 1995
      Partition check:
      hda: hda1
>>    hdb: DRDY error: status=0x00 { }
      hdb: DRDY error: status=0x00 { }
      hdb: DRDY error: status=0x00 { }
      hdb: DRDY error: status=0x00 { }
      ide0: do_ide_reset: success
      hdb: DRDY error: status=0x00 { }
      hdb: DRDY error: status=0x00 { }
      hdb: DRDY error: status=0x00 { }
      hdb: DRDY error: status=0x00 { }
      ide0: do_ide_reset: success
      hdb: DRDY error: status=0x00 { }
      end_request: I/O error, dev 0340, sector 0
      hdb: unexpected_intr: status=0x00 { }
>>    unable to read partition table of device 0340
      hdc: hdc1
      UMSDOS Beta 0.6 (compatibility level 0.4, fast msdos)
      Mounting root

A quick work-around is to tell the kernel not to rely on CMOS at all,
and just probe for all devices. So, in ide.c, comment out the relevant
line in ide_init() towards the end of the file (or dirtier still, just
ensure the condition leading to it is always false, like I did below):


      for (hwif = 0; hwif < 2; hwif++) {
        init_ide_data (hwif);
        if (SUPPORT_TWO_INTERFACES || hwif == HWIF) {
>>        if (hwif == 10)  /* used to be hwif == 1 */
    #ifdef CONFIG_BLK_DEV_HD
            continue;
    #else
>>          probe_cmos_for_drives ();
    #endif /* CONFIG_BLJ_DEV_HD */
          probe_mem_start = (mem_start + 3uL) & ~3uL;
          probe_for_drives (hwif);
          mem_start = probe_mem_start;
        }
      }

I'm not sure if this would break anything on other systems, though
undoubtly probing for devices instead of asking CMOS might lead to
trouble on some older hardware, and that might be the reason the
kernel developers thought to keep the above intact.

Best wishes.

                                        -- Muhammad