On Mon, Nov 28, 2022 at 08:47:40PM +0100, Markus Wichmann wrote:
> Hi all,
> 
> __dls2 calls kernel_mapped_dso(), and that one uses the PAGE_SIZE macro.
> Whenever <bits/limits.h> does not define PAGESIZE, PAGE_SIZE is defined
> as libc.page_size. That variable is only initialized at the start of
> __dls3, so the DSO descriptor for libc ends up being wrong.
> 
> Since the libc object has static storage duration, page_size is
> initialized with zero. So at least nothing undefined happens. The impact
> is, it will calculate the relro pointers as being zero, so no relro will
> happen, and it will calculate maximum and minimum addresses as being
> zero, therefore setting map to base and map_len to zero. This will cause
> dladdr() not to find the libc. Yeah, not the biggest of impacts.
> 
> This, again, affects all architectures that don't define PAGESIZE, so at
> this time those are
> 
> aarch64
> arm
> m68k
> microblaze
> mips
> mips64
> mipsn32
> powerpc
> powerpc64
> riscv64
> 
> I don't know whether references to libc are even valid in __dls2, but it
> is defined as "hidden", so that ought to be good enough. In that case it
> may be enough to just move the initialization. Otherwise it may be
> necessary to add page size as parameter to kernel_mapped_dso(). Then
> __dls2 can look it up in the aux vector at its leisure.

Nice catch. The references to libc are not valid in __dls2. If they
were, I would just re-run kernel_mapped_dso() from __dls2b or
something to get the right relro map, but I think instead we should do
something like the attached.

Rich