The main trick is having 5 years of experience with those pesky oops
messages ;-)
Actually, there are things you can do that make this easier. I have two
separate approached:
gdb /usr/src/linux/vmlinux
gdb> disassemble <offending_function>
That's the easy way to find the problem, at least if the bug-report is
well made (like this one was - run through ksymoops to get the
information of which function and the offset in the function that it
happened in).
Oh, it helps if the report happens on a kernel that is compiled with the
same compiler and similar setups.
The other thing to do is disassemble the "Code:" part of the bugreprot:
ksymoops will do this too with the correct tools (and new version of
ksymoops), but if you don't have the tools you can just do a silly
program:
char str[] = "\xXX\xXX\xXX...";
main(){}
and compile it with gcc -g and then do "disassemble str" (where the "XX"
stuff are the values reported by the Oops - you can just cut-and-paste
and do a replace of spaces to "\x" - that's what I do, as I'm too lazy
to write a prigram to automate this all).
Finally, if you want to see where the code comes from, you can do
cd /usr/src/linux
make fs/buffer.s # or whatever file the bug happened in
and then you get a better idea of what happens than with the gdb
disassembly.
Now, the trick is just then to combine all the data you have: the C
sources (and general knowledge of what it _should_ do, the assembly
listing and the code disassembly (and additionally the register dump you
also get from the "oops" message - that can be useful to see _what_ the
corrupted pointers were, and when you have the assembler listing you can
also match the other registers to whatever C expressions they were used
for).
Essentially, you just look at what doesn't match (in this case it was the
"Code" disassembly that didn't match with what the compiler generated).
Then you need to find out _why_ they don't match. Often it's simple - you
see that the code uses a NULL pointer and then you look at the code and
wonder how the NULL pointer got there, and if it's a valid thing to do
you just check against it..
Now, if somebody gets the idea that this is time-consuming and requires
some small amount of concentration, you're right. Which is why I will
mostly just ignore any panic reports that don't have the symbol table
info etc looked up: it simply gets too hard to look it up (I have some
programs to search for specific patterns in the kernel code segment, and
sometimes I have been able to look up those kinds of panics too, but
that really requires pretty good knowledge of the kernel just to be able
to pick out the right sequences etc..)
_Sometimes_ it happens that I just see the disassembled code sequence
from the panic, and I know immediately where it's coming from. That's when
I get worried that I've been doing this for too long ;-)
Linus