Re: [PATCH 02/02] Elf: Align elf notes properly

["Magnus Damm" <magnus.damm@gmail.com>]

On 11/10/06, Eric W. Biederman <ebiederm@xmission.com> wrote:
> "Magnus Damm" <magnus.damm@gmail.com> writes:
>
> > I'm not sure you see all my points. The important parts are the
> > offsets - offset 0 and offset N2 in the description above. The should
> > be aligned somehow. Exactly how to align them depends on if the 64-bit
> > spec is valid or not.
> >
> > My points are:
> >
> > - Some kdump code rounds up the size of "elf note header" today. This
> > is unneccessary for 32 bit alignment and plain wrong for 64 bit
> > alignment. So I think that the code is strange and should be changed
> > regardless if the 64-bit spec is valid or not.
>
> Sure that is reasonable, if correct.
>
> > - Many implementations incorrectly calculate N2 as: roundup(sizeof(elf
> > note header)) + roundup(n_namesz).
>
> I am not certain that is incorrect.  roundup(sizeof(elf note header), 4) +
> roundup(n_namesize, 4) will yield something that is properly 4 byte aligned.
> I do agree that implementation is not correct for 8 byte alignment.  8 byte
> alignment does not appear to be in widespread use in the wild.

You are correct that it only matters if we are interested in 8 byte
alignment. So it should be a non-issue for the 4-byte aligment case.

> > - You say that the size of the notes do not vary and therefore this is
> > a non-issue. I agree that the size does not vary, but I believe that
> > the aligment _is_ an issue. One example is the N2 calculation above,
> > but more importantly the vmcore code that merges the elf note sections
> > into one. You know, if you have more than one cpu you will end up with
> > more than one crash note. And if you run Xen you will have even more
> > crash notes.
>
> Sure that is clearly an issue.
>
> > - On top of this I think it would be nice if all this code could be
> > unified to avoid code duplication. But we need to straighten out this
> > and agree on how the aligment should work before the code can be
> > merged into one implementation.
>
> Sure.
>
> To verify your claim that 8 byte alignment is correct I checked the
> core dump code in fs/binfmt_elf.c in the linux kernel.  That always
> uses 4 byte alignment.  Therefore it appears clear that only doing
> 4 byte alignment is not a local misreading of the spec, and is used in
> other implementations.  If you can find an implementation that uses
> 8 byte alignment I am willing to consider it.

Yes, fs/binfmt_elf.c is one of the files that my patch modifies. There
are several elf note implementations in the kernel, all seem to use
4-byte aligment.

Implementations that use 8-byte alignment:

binutils-2.16.1/bfd/elf.c: elf_core_write_note() is using
log_file_align which is set to 3 on some 64-bit platforms.  8-byte
alignment in some cases.

binutils-2.16.1/binutils/readelf.c: process_corefile_note_segment() is
always using 4-byte alignment though.

> The current situation is that the linux kernel generated application
> core dumps use 4 byte alignment so I expect that is what existing
> applications such as gdb expect.

Most applications probably expect 4-byte aligned data. OTOH, I just
came across HP's ELF-64 Object File Format document. It says that
8-byte alignment should be used:

http://devresource.hp.com/drc/STK/docs/refs/elf-64-hp.pdf

So now we have two documents that say 8-byte alignment should be used.

> Therefore we use 4 byte alignment unless it can be shown that the
> linux core dumps are a fluke and should be fixed.

Ok. Vivek, Dave, anyone? Comments?

/ magnus
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/