search

ARM-software / abi-aa

Application Binary Interface for the Arm® Architecture
Other
971 stars 193 forks source link

[pauthabielf64] Define static AUTH TLSDESC relocations #293

Open kovdan01 opened 1 month ago

kovdan01 commented 1 month ago

In pauthabielf64, dynamic AUTH TLSDESC relocation R_AARCH64_AUTH_TLSDESC is defined https://github.com/ARM-software/abi-aa/blob/main/pauthabielf64/pauthabielf64.rst#auth-variant-got-generating-relocations.

However, the section with AUTH variant GOT-generating relocations does not contain definitions for static TLSDESC relocs: https://github.com/ARM-software/abi-aa/blob/main/pauthabielf64/pauthabielf64.rst#auth-variant-got-generating-relocations.

Particularly, the following relocation types are missing and should be defined:

  1. R_AARCH64_AUTH_TLSDESC_ADR_PAGE21
  2. R_AARCH64_AUTH_TLSDESC_LD64_LO12
  3. R_AARCH64_AUTH_TLSDESC_ADD_LO12
kovdan01 commented 1 month ago

Tagging @smithp35 @asl

kovdan01 commented 1 month ago

@smithp35 I've submitted a PR to llvm-project repo adding the relocs with the following codes (the latest code for previously defined GOT-generating relocs was 0x811d) - see https://github.com/llvm/llvm-project/pull/113716:

R_AARCH64_AUTH_TLSDESC_ADR_PAGE21 0x811e
R_AARCH64_AUTH_TLSDESC_LD64_LO12  0x811f
R_AARCH64_AUTH_TLSDESC_ADD_LO12   0x8120

Please let me know if these code are OK.

smithp35 commented 1 month ago

My apologies, I'll be out of office for the the week of 28th October to the 1st November so this is a quick response at the weekend. I'll get back to this when I'm back in the office in a week's time.

The relocation codes are OK for now [1] as they are in the private space reserved for experiments https://github.com/ARM-software/abi-aa/blob/main/aaelf64/aaelf64.rst#5714private-and-platform-specific-relocations

The GOT signing relocations are currently in the private space. As I understand it there's a high chance that these will be needed. When they stabilise we'll need to move them to the space reserved for PAuthABI https://github.com/ARM-software/abi-aa/blob/main/aaelf64/aaelf64.rst#5716pauthabi-relocations

What I think needs doing:

There's a separate topic of moving the relocations from the private experiment space.

kovdan01 commented 1 month ago

@smithp35 Thanks for your feedback!

  • There are more non-auth TLSDESC relocations defined in AAELF64. Is it worth defining AUTH equivalents for these (assuming they are relevant to PAuth; I've not checked)?

These three should be enough for working TLSDESC implementation - we have llvm codegen https://github.com/llvm/llvm-project/pull/113813, lld linker https://github.com/llvm/llvm-project/pull/113817 and musl dynamic loader https://github.com/access-softek/musl/pull/2 patches for that, and the functionality was also tested in runtime, not only by lit tests.

It actually looks like that other non-auth TLSDESC-related relocs from this section https://github.com/ARM-software/abi-aa/blob/main/aaelf64/aaelf64.rst#57115thread-local-storage-descriptors are not even used by llvm-project. Some of them (R_<CLS>_TLSDESC_ LD_PREL19, R_<CLS>_TLSDESC_ ADR_PREL21) seem to be intended for usage with tiny code model, for some others (R_<CLS>_TLSDESC_ OFF_G1, R_<CLS>_TLSDESC_ OFF_G0_NC) I'm not even sure what's their purpose. Anyway, the three relocs listed above should definitely be enough for "default" implementation matching existing non-auth implementation in llvm's backend.

  • Make sure we're happy with the names. From a glance they look fine.

I've just added AUTH to the following non-auth reloc names: R_<CLS>_TLSDESC_ADR_PAGE21, R_<CLS>_TLSDESC_LD64_LO12, R_<CLS>_TLSDESC_ADD_LO12. I would be happy if you double-check that it's OK :)

  • Update the PAuthABI spec.

I can submit a corresponding PR if you wish - just let me know if you want me to do that or if you plan to do it by yourself later.

smithp35 commented 4 weeks ago

Just got back to the office today, apologies for the delay in responding.

I can post a patch to the ABI tomorrow.

Looking at the initial proposals for the relocation, the code sequences for the tiny and large code-models use these relocations. Although to the best of my knowledge both clang and GCC only implement the small code-model for accessing TLS.

I agree that we don't need to specify relocations for these until someone wants to optimise TLSDESC for the tiny code-model.

Tiny:

ldr   x1, :tlsdesc:var                        R_AARCH64_TLSDESC_LD_PREL19  var
adr   x0, :tlsdesc:var                        R_AARCH64_TLSDESC_ADR_PREL21 var
.tlsdesccall var
blr   x1                                      R_AARCH64_TLSDESC_CALL       var

Large:

# gp = register containing address of GOT.
movz  x0, #:tlsdesc_off_g1:var                R_AARCH64_TLSDESC_OFF_G1    var
movk  x0, #:tlsdesc_off_g0_nc:var             R_AARCH64_TLSDESC_OFF_G0_NC var
.tlsdescldr var
ldr   x1, [gp, x0]                            R_AARCH64_TLSDESC_LDR       var
.tlsdescadd var
add   x0, gp, x0                              R_AARCH64_TLSDESC_ADD       var
.tlsdesccall var
blr   x1                                      R_AARCH64_TLSDESC_CALL      var
smithp35 commented 4 weeks ago

I've submitted https://github.com/ARM-software/abi-aa/pull/295 for review