Morello: Stop using caprelocs for static binaries #2286
Conversation
bsdjhb
commented
Jan 8, 2025
bsdjhb
commented
- csu: Add support for ELF relocations to init caps for static binaries
- morello: Support R_MORELLO_RELATIVE relocations in static binaries
- rtld: Workaround Morello LLVM lld bug with local-caprelocs=elf
- Enable --local-caprelocs=elf for Morello
|
I have a request to fix the bug in Morello lld at https://git.morello-project.org/morello/llvm-project/-/merge_requests/307, but it may be a while before we can assume it is fixed. The workaround should let us work with existing lld for now and we can eventually drop it. |
| #ifdef __CHERI_PURE_CAPABILITY__ | ||
| #if defined(__aarch64__) | ||
| __asm__ ( | ||
| "adrp %0, __rela_dyn_start\n\t" |
There was a problem hiding this comment.
Can we put this in crt.h as some #define FOO(x) ({ __typeof__(x) *_p; __asm__ ("adrp %0, %c1; add %0, %0, :lo12:%c1" : "=C"(_p) : "i"(&x)); _p; })? Hybrid can define it to just &x.
There was a problem hiding this comment.
(Or crt1_c.c where CHERI_INIT_RELA lives)
There was a problem hiding this comment.
Then it can also be part of the second commit that adds Morello support specifically
There was a problem hiding this comment.
Oh, yeah, that might be fine, and certainly cleaner. PCREL_PTR() maybe?
There was a problem hiding this comment.
Sure, seems fine, name doesn't hugely matter, it's internal to csu
There was a problem hiding this comment.
After adding a description in the commit log I ended up going with RODATA_PTR() for the name.
This is to support the --local-caprelocs=elf flag to lld. When this
is specified, an ELF relocation table is emitted. The table is
identified by the hidden symbols __rela_dyn_start and __rela_dyn_end.
If an architecture defines CHERI_INIT_RELA, then crt_init_globals()
will look for this table. For each relocation found, an MD function
provided by the architecture is called to handle the relocation. The
function's signature is:
elf_reloc(const Elf_Rela *, void * __capability data_cap,
const void * __capability code_cap, Elf_Addr relocbase)
In addition, an architecture must define a RODATA_PTR helper macro
which accepts a symbol name as its sole argument and returns a
readable pointer. For hybrid this can just use the normal address-of
operator, but for purecap this has to avoid depending on an ELF
relocation and instead derive the pointer from an existing capability
such as PCC.
A shared header containing the implementation of elf_reloc is used to avoid duplicating code between hybrid and purecap. The init_cap_from_fragment function is copied directly from rtld (might be nice to move that function to a shared header).
Morello LLVM's lld always emits PLT-related .dynamic entries when ELF local caprelocs are used even if the PLT is empty. This confuses rtld which tries to treat the start of the binary as a GOT PLT. In particular, rtld itself ends up with an empty PLT and so rtld tries to write to its ELF header which crashes since the page is mapped read-only. Workaround this by ignoring DT_JMPREL and DT_PLTGOT entries whose value is 0. The PLT GOT and PLT relocation table can never be at the start of a valid binary.
bsdjhb
force-pushed
the
morello_static_rela
branch
from
9bb66ed to
ee51273
Compare
This is done in bsd.{lib,prog}.mk instead of bsd.cpu.mk as it can't be
turned off for the lib32 build otherwise.
bsdjhb
force-pushed
the
morello_static_rela
branch
from
ee51273 to
c9c2b04
Compare
