An upstream patch to close bug #345931 in Debian

[Mats Erik Andersson]

  Dear Maintainers and Bugtrackers,

  I would like to submit a patch to grub 0.97 that
closes the urgent and long standing bug #345931
recorded with Debian users. I will at some length try
motivate an upstream acceptance. My arguments are of
some implications even for Grub2, so please read this
with an open mind.

  In a change 2005-02-15, the underlying principle for
a workaround to recover from inferior drive-flags for
harddisks in some BIOSes, was drastically changed.
This
meant that a three-byte instruction in Stage1 of grub
0.96 and a new two-byte instruction of grub 0.97
became
instrumental as to how the workaround are applied in
the respective versions. The outcome was that grub
0.97
completely breaks the code in Stage1 from grub 0.96
when grub 0.97 is used to restore the installed boot
files deposited by grub 0.96. This provoked #345931
with the elaps of time. 
  I personally find the above incompatibility to be
a serious mistake, should it be allowed to prevail.
Admittedly, the more recent workaround may well be
superior, but it should not be allowed to disrupt
perfectly fitting installments of grub 0.94-0.96.
My proposed patch simply equips the setup-function
of the grub-shell with the ability to test which
workaround was used in the Stage1-sector present
on the chosen harddisk partition, and then it takes
the corresponding measures needed to correctly
restore Stage1.
  At the very end of the patch I have deliberately
coded and commented in a way to allow the proper
actions to be implemented also for providing backward
compatibility all the way back to grub 0.92, should
the maintainers desire to effectuate also this. 
For the time being, the code deliberately fails
for grub 0.93 and older. Please tell me to implement
the extension or do it yourself. I must say again
that I find it imperative to keep versions 0.96 and
0.97 compatible, since grub 0.95/96 are present on
very many systems and will be for some time to come.
A policy to implement compatibility visavi grub 0.92/3
is negotiable and left to the maintainers to decide
upon.

  Now over to Grub2. Robert Millan encouraged me to
determine whether my patch is of relevance to Grub2.
An investigation of the code in grub 1.94 reveals that
the new code base uses the same mechanism as has been
applied in grub 0.97 for this particular workaround.
Thus it is well possible that the principle behind
my patch - after insertion of new naming conventions -
should or could be applied at

  grub-1.94/util/i386/pc/grub-setup.c   lines 253--

The maintainers must think through and state their
decisions on backwards compatibility in this respect.
I myself know too little of the code for Grub2 in
order
to state a thought through opinion, but out of
principle I can second backwards compatibility,
where it does not disturb the main objectives.

  As a final side remark, Otavio Salvador urged me
to correct my indentation from the previous version
of the patch. I did my inexperienced best to do
it in the style of the file builtins.c. Feel free to
correct me or complain, as the case may be.

          Best regards

            Mats Erik Andersson
            address@hidden

#  boot_drive_correction.diff
#
#  Enables the workaround for the "boot drive flag" used in grub 0.94-0.96
#  to coexist with the revised principle of grub 0.97. Thereby the present
#  version of the grub-shell restores correctly from  (hd?,?)/boot/grub/*
#  for the full span of versions 0.94-0.97. This is extendible also to
#  grub 0.92-0.93, by following my proposition in the last comment below.
#
diff -Naur grub-0.97.orig/stage2/builtins.c grub-0.97.patched/stage2/builtins.c
--- grub-0.97.orig/stage2/builtins.c    2005-02-15 22:58:23.000000000 +0100
+++ grub-0.97.patched/stage2/builtins.c 2006-09-28 21:05:36.040085552 +0200
@@ -1956,10 +1956,29 @@
      for buggy BIOSes which don't pass boot drive correctly. Instead,
      they pass 0x00 or 0x01 even when booted from 0x80.  */
   if (dest_drive & BIOS_FLAG_FIXED_DISK)
-    /* Replace the jmp (2 bytes) with double nop's.  */
-    *((unsigned short *) (stage1_buffer + STAGE1_BOOT_DRIVE_CHECK))
-      = 0x9090;
-  
+    {
+      if ( *((unsigned char *) (stage1_buffer + STAGE1_BOOT_DRIVE_CHECK)) == 
0xeb )
+        /* For present version 0.97:  Replace the jmp (2 bytes) with double 
nop's.  */
+        *((unsigned short *) (stage1_buffer + STAGE1_BOOT_DRIVE_CHECK))
+          = 0x9090;
+      else if ( *((unsigned char *) (stage1_buffer + STAGE1_BOOT_DRIVE_CHECK)) 
== 0x80 )
+        {   
+          /* For previous versions 0.94-96:  Set the "boot-drive-mask".
+             This is the old workaround for buggy BIOSes which do not pass 
boot drive correctly.
+             The technique was to use 'or dl,0x80', hence testing for 0x80, 
i.e. or-instruction.
+             REM: the old STAGE1_BOOT_DRIVE_MASK equals 
STAGE1_BOOT_DRIVE_CHECK + 2.  */
+          *((unsigned char *) (stage1_buffer + STAGE1_BOOT_DRIVE_CHECK + 2))
+            = (dest_drive & BIOS_FLAG_FIXED_DISK);
+        }
+      else
+        {
+          /* The boot sector is thus older than version 0.94.
+             Changing 'goto fail' to a "do nothing" could even make 0.92 and 
0.93
+             acceptable for the purpose of rescuing their installations.  */
+         goto fail;
+       }
+    }
+
   /* Read the first sector of Stage 2.  */
   disk_read_hook = disk_read_savesect_func;
   if (grub_read (stage2_first_buffer, SECTOR_SIZE) != SECTOR_SIZE)