Re: [Crash-utility] [PATCH 0/4] MIPS support
----- Original Message -----
On Mon, Jan 12, 2015 at 11:00:08AM -0500, Dave Anderson wrote:
> I don't know what the real-world usage is with respect to little-endian
> vs. big-endian 32-bit MIPS, but you got here first. And in the
> future, it should be simple enough to support big-endian MIPS as
> well, i.e., in the same manner that both little- and big-endian PPC64
> are supported.
FWIW, I see that http://www.linux-mips.org/wiki/Endianess indicates:
On 32-bit kernels there seems to be somewhat a majority of big endian systems.
Distribution download figures suggset approximately a 60:40 ratio.
I attempted to have a quick look at big-endian while making the patch
(it's easy to emulate big-endian MIPS too on QEMU), and if I understand
correctly from the PPC handling, the host endianness needs to match the
target endianness. So that means we can't analyze big-endian MIPS dumps
on x86/x86-64 hosts, right?
That's correct.
The MIPS systems which I personally plan to use this for are
little-endian, and the analysis will be done on x86-64 hosts.
> Running cscope from the gdb-7.6 subdirectory on down shows only
> this single reference to convert:
>
> sim/mips/sim-main.h Convert 774 #define Convert(rm,op,from,to) convert
> (SIM_ARGS, rm, op, from, to)
>
> But it doesn't show any callers to Convert() or convert(). So
> it appears to exist as a non-static function in "sim/mips/cp1.c"
> that nobody calls? In any case, please make the convert() name
> change down in the gdb-7.6/sim/mips code, and update gdb-7.6.patch
> accordingly.
Done. Convert() is used in the mips.igen file in the same directory,
which appears to be used to generate C code and compile it as part of
the build process.
Ah, of course, cscope didn't check that file.
> Also in defs.h, please put the task_struct_thread_reg29 and
> task_struct_thread_reg31 offsets at the end of the offset_table
> structure. Again, new entries to that structure should always
> be appended to the structure so that it won't break the usage
> of OFFSET() by pre-existing extension modules. Also, you should
> add their offset value displays to dump_offset_table() for use
> by "help -o", and in that function, you can display it next to
> the other task_struct_thread_xxx offsets.
Done.
> On another note, there should be an understandable and clean
> protection mechanism against running a target=MIPS binary on a
> live host system. So when attempting to run a target=ARM crash
> binary on an x86/x86_64 host, it fails like so:
>
> $ sudo ./crash
> crash: compiled for the ARM architecture
>
> With your target=MIPS binary, it shows this confusing error message:
>
> crash: cannot resolve "cpu_data"
>
> It's a small check that can be put at the beginning of mips_init().
> I presume that you used ppc.c as a template, so if you'd like, you
> can add the same fix there as well.
ppc.c doesn't have that check, but I've added it to mips.c based on
what's in arm.c.
Right -- what I meant was that you could make the same fix in ppc.c as well.
> Lastly (at least at this point), can you update the README[] strings array
> in help.c to show that MIPS is a supported architecture, and that it can
> be build as a "target=" binary?
Done.
> Oh, and thanks very much for the sample vmcores. I'll keep them around for
future
> testing. BTW, the System.map file is completely unnecessary -- is there a reason
> you included it?
Not really. I never used it but thought "just in case" while zipping up
the dumps + vmlinux.
Rabin
OK good -- I'll check this new patchset out tomorrow.
Thanks,
Dave