Re: [PATCH] lkdtm: fix memory copy size for WRITE_KERN

[Nick Desaulniers <ndesaulniers@google.com>]

On Tue, Jan 26, 2021 at 2:53 PM David Laight <David.Laight@aculab.com> wrote:
>
> From: Nick Desaulniers
> > Sent: 26 January 2021 18:40
> >
> > On Tue, Jan 26, 2021 at 6:13 AM Candle Sun <candlesea@gmail.com> wrote:
> > >
> > > On Mon, Jan 25, 2021 at 6:37 PM David Laight <David.Laight@aculab.com> wrote:
> > > >
> > > > From: Candle Sun
> > > > > Sent: 25 January 2021 08:56
> > > > >
> > > > > From: Candle Sun <candle.sun@unisoc.com>
> > > > >
> > > > > Though do_overwritten() follows do_nothing() in source code, the final
> > > > > memory address order is determined by compiler. We can't always assume
> > > > > address of do_overwritten() is bigger than do_nothing(). At least the
> > > > > Clang we are using places do_overwritten() before do_nothing() in the
> > > > > object. This causes the copy size in lkdtm_WRITE_KERN() is *really*
> > > > > big and WRITE_KERN test on ARM32 arch will fail.
> > > > >
> > > > > Compare the address order before doing the subtraction.
> > > >
> > > > It isn't clear that helps.
> > > > Compile with -ffunction-sections and/or do LTO an there
> > > > is no reason at all why the functions should be together.
> > > >
> > > > Even without that lkdtm_WRITE_KERN() could easily be between them.
> > > >
> > > > You need to get the size of the 'empty function' from the
> > > > symbol table.
> > > >
> > > >         David
> > >
> > > Thanks David.
> > >
> > > I think using abs() by Nick's advice would be better. But could you
> > > point out which kernel function can get function size?
> >
> > The Elf symbol table should contain this info, IIUC.
> >
> > Given a string literal of a symbol (such as a function identifier),
> > kallsyms_lookup_name() can be used to return its address.
> >
> > From there we'd want to fetch the Elf_Sym for the address which should
> > contain a st_size field which I think corresponds to the size in bytes
> > of the function.  (At least, from playing with `llvm-readelf -s`)
> > Probably would want to validate it's an STT_FUNC symbol type, too.  We
> > basically want something like kexec_purgatory_find_symbol(), but that
> > knows about the entire kernel image, and not the purgatory image used
> > during kexec.  I don't see any such function currently in the
> > kernel...but it's a large codebase to search through.
>
> The alternative is to get the linker script to define a specific
> constant to the size of the function.
> You can then link against it (by using it as the address of a symbol).

Or use __attribute__((__section())) to place the code we need to
measure in a custom section, then in the linker script define symbols
that bound the section then place it in .text.  That pattern is used
throughout the kernel.  But I suspect this is overkill for this
module.

A constant would depend on the arch.

>
> It may be easier to use an asm file for the 'return 0' code.
> I'm guessing there are things in the static branch area
> that could be used.
>
>         David
>
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)



-- 
Thanks,
~Nick Desaulniers