illumos gate

I discovered three issues in Illumos GRUB 0.97 when I tried porting it to Gentoo Linux. Patches are attached.

One issue is a fairly serious bug where /sbin/grub will write a NULL character outside the bounds of a buffer into another buffer. It likely hasn't caused any problems on Illumos because if it writes the NULL character into a region not being used, things will not break. In my situation, it morphed "(hd0,0)/boot/grub/stage2" into "(hd0,0)/boot/grub/stage" because the '2' was the 256 character offset from the start of another buffer, breaking the installation of GRUB.

Here is a backtrace from GDB taken immediately before the corruption occurred:

(gdb) bt
#0 grub_strncat (s1=0xf350dd18 "(hd0)", s2=<optimized out>, n=256) at char_io.c:1050
#1 0x0805daf4 in sprint_device (drive=<optimized out>, partition=<optimized out>) at builtins.c:4861
#2 0x0805fe43 in setup_func (arg=<optimized out>) at builtins.c:5037
#3 0x08061f2f in enter_cmdline (heap=0xf34bcc90 "setup (hd0)", forever=1) at cmdline.c:174
#4 0x0805aa6d in cmain () at stage2.c:1106
#5 0x0804c514 in init_bios_info () at common.c:356
#6 0x080499ef in doit () at asmstub.c:132
#7 0x08049ae0 in grub_stage2 () at asmstub.c:189
#8 0x0804967f in main (argc=2, argv=0xffe70084) at main.c:264

My solution for this issue was to change the buffer size from 16 to 32 to match another buffer, while replacing 256 with sizeof device to reference the buffer. The code in question makes use of a GNUism that permits nested functions with the buffer declared in an outer function. I assume that the Sun Studio compiler supports nested functions in C, but someone should verify that the use of the sizeof operator in conjunction with nested functions is also supported by the Sun Studio compiler. Since I don't have a copy of Sun Studio, I cannot test this, but someone should verify that this works.

The second patch addresses a build failure that is caused by the absence of tabs on a few lines in ./grub-0.97/stage2/Makefile when building with gmake. I modified those lines to use tabs, which fixed this issue. The patch for this should be safe to merge into Illumos without any concern for regressions.

The third patch addresses a compilation failure that is caused by the use of uint32_t and uint64_t without a typedef, which GCC does not understand. I borrowed some t ypdefs from the /usr/include/stdint.h header file provided by glibc on Gentoo Linux and introduced ./grub-0.97/stage2/uint.h to be a home for them. I examined /usr/include/sys/int_types.h on OpenIndiana 151a and there is a possibility that doing this introduces a portability issue, so this should be tested to ensure that it works on Illumos.

The mismatch patch wasn't uploaded for some reason. I am trying again.

I was informed in IRC that there was an issue with one of my patches. It seems that I accidentally uploaded an old version of the gmake patch that modified Makefile instead of Makefile.in. I have attached a new patch to correct that.

Also, I took another look at the GRUB code and I discovered the presence of grub_uint32_t and grub_uint64_t in the ufs.h and ufs2.h files, so I factored those and some related type definitions into a separate types.h file and modified the ZFS code and UFS headers to rely on that. This is a much better solution than the previous approach I had taken.