# PI_STATIC_AND_HIDDEN/HIDDEN_VAR_NEEDS_DYNAMIC_RELOC in glibc rtld

Recently I have fixed two glibc rtld bugs related to early GOT relocation for retro-computing architectures: m68k and powerpc32. They are related to the obscure PI_STATIC_AND_HIDDEN macro which I am going to demystify.

In 2002, PI_STATIC_AND_HIDDEN was introduced into glibc rtld (runtime loader). This macro indicates whether accesses to the following types of variables need dynamic relocations.

• static specifier: static int a; (STB_LOCAL)
• hidden visibility attribute: __attribute__((visibility("hidden"))) int a; (STB_GLOBAL STV_HIDDEN), __attribute__((weak, visibility("hidden"))) int a; (STB_WEAK STV_HIDDEN)

PI in the macro name is an abbreviation for "position independent". This is a misnomer: a code sequence using GOT is typically position-independent as well.

In -fPIC mode, the compiler assumes that all non-local STV_DEFAULT symbols may be preemptible at run time. A GOT-generating relocation is used and the GOT is typically unavoidable at link time (on some architectures the linker can optimize out the GOT). This case is not interesting to rtld as rtld does not need to export such variables.

Excluding these cases (non-local STV_DEFAULT), all other variables are known to be non-preemptible at compile time. The compiler can generate code which is guaranteed to avoid dynamic relocations at link time.

On 2022-04-26, I replaced PI_STATIC_AND_HIDDEN with the opposite macro HIDDEN_VAR_NEEDS_DYNAMIC_RELOC.

## Non-HIDDEN_VAR_NEEDS_DYNAMIC_RELOC architectures with PC-relative instructions

To avoid dynamic relocations, the most common approach is to generate PC-relative instructions, as most modern architectures (e.g. aarch64, riscv, and x86-64) provide. Using PC-relative instructions to reference variables assumes that the distance from code to data is a link-time constant. Nowadays this condition is satisfied everywhere except the rare FDPIC ABI.

Here are some assembly fragments from architectures using PC-relative instructions. The instructions may not be familar to you, but that is fine. We can see that there is no GOT related marker. I have added some comments indicating the relocation type and the referenced symbol. var in the C code has internal linkage which lowers to the STB_LOCAL binding. References to such local symbols are often redirected to the section symbol (.bss): the link-time behaviors are identical.

## Non-HIDDEN_VAR_NEEDS_DYNAMIC_RELOC architectures without PC-relative instructions

Many older architectures do not have PC-relative instructions.

x86-32 does not have PC-relative instructions, but it provides a way to avoid a load from a GOT entry. It achieves this with a detour: compute the address of _GLOBAL_OFFSET_TABLE_ (GOT base symbol), then add an offset (S-_GLOBAL_OFFSET_TABLE_) to get the symbol address. _GLOBAL_OFFSET_TABLE_ is computed this way: compute the address of a location in code, then add an offset (_GLOBAL_OFFSET_TABLE_ - PC).

You probably see now how the x86-32 ABI was misdesigned: the involvement of _GLOBAL_OFFSET_TABLE_ is unnecessary. A relocation with the calculation of S-_GLOBAL_OFFSET_TABLE_ would achieve the same net effect.

The relocations with GOT in their names just use the GOT as an anchor. They don't indicate a load from a GOT entry.

powerpc64 does not have PC-relative instructions before POWER10. Earlier microarchitectures use TOC-relative relocations to compute the symbol address.

A pending patch [PATCH v3] powerpc64: Enable static-pie will define PI_STATIC_AND_HIDDEN.

## HIDDEN_VAR_NEEDS_DYNAMIC_RELOC architectures

A few older architectures tend to use a load from a GOT entry. The GOT entry needs a relative relocation (instead of R_*_GLOB_DAT: the symbol is non-preemptible, so no symbol search is needed). See All about Global Offset Table. In glibc, these architecture define HIDDEN_VAR_NEEDS_DYNAMIC_RELOC.

Some architectures even assume the distance from code to data may not be a link-time constant (see All about Procedure Linkage Table). They do not provide a relocation with a calculation of S-_GLOBAL_OFFSET_TABLE_ or S-P.

The first task of rtld is to relocate itself and bind all symbols to itself. Afterward, non-preemptible functions and data can be freely accessed.

On architectures where a GOT entry is used to access a non-preemptible variable, rtld needs to be careful not to reference such variables before relative relocations are applied. In rtld.c, _dl_start has the following code:

_rtld_local_ro is a hidden global variable. Taking its address may be reordered before ELF_DYNAMIC_RELOCATE by the compiler. On an architecture using a GOT entry to load the address, the reordering will make the subsequent memory store (_rtld_local_ro.dl_find_object) to crash, since the GOT address is incorrect: it's zero or the link-time address instead of the run-time address.

## powerpc32

I recently cleaned up the bootstrap code a bit with elf: Move elf_dynamic_do_Rel RTLD_BOOTSTRAP branches outside. Afterwards, GCC powerpc32 appears to reliably reorder _rtld_local_ro, causing ld.so to crash right away.

I was pretty sure there is a relocation bug but was not immediately clear which piece of code may be at fault.

Nowadays there aren't many choices for powerpc32 images. Void Linux ppc still provides powerpc32 glibc and musl images. I downloaded one and fed it into qemu, booted it with qemu-system-ppc -machine mac99 -m 2047M -cdrom void-live-ppc-20210825.iso -net nic -net user,smb=\$HOME/Dev -boot d. I booted into the 4.4.261 kernel because gdb aborts immediately with 5.13.12 kernel. Daniel Kolesa mentioned this 5.x kernel incompatibility to me and nobody has looked into it yet.

The live CD provides free space of about 1GiB and I can install cifs-utils and gdb. Then run ld.so under gdb.

gdb says stw r9,1168(r25) triggers SIGSEGV.

Then I confirm that the GOT entry corresponds to _rtld_local_ro.

elf: Move post-relocation code of _dl_start into _dl_start_final shall fix the bug.

Note: adding asm volatile("" ::: "memory"); in between does not prevent reordering.

Note: in the absence of a powerpc32 system, qemu-ppc-static -d in_asm elf/ld.so may provide some clue about the faulty basic block.

## m68k

Last week I fixed a similar bug for m68k: m68k: Removal of ELF_DURING_STARTUP optimization broke ld.so.

ld.so has 671 R_68K_RELATIVE relocations and one R_68K_GLOB_DAT for __stack_chk_guard@@GLIBC_2.4. The following function is used to apply a relocation. It is shared by self-relocation and relocation for other modules. The self-relocation code defines RTLD_BOOTSTRAP and needs just R_68K_RELATIVE, R_68K_GLOB_DAT, and R_68K_JMP_SLOT.

However, somehow many case labels were available for self-relocation. GCC compiles the switch statement into a jump table which requires loading an address from GOT. With some clean-up to generic relocation code, GCC decides to perform loop-invariant code motion and hoists the load of the jump table address. The hoisted load is before relative relocations are applied, so the jump table address is incorrect.

The foolproof approach is to add an optimization barrier (e.g. calling an non-inlinable function after relative relocations are resolved). That is non-trivial given the code structure. So Andreas Schwab suggested a simple approach by avoiding the jump table: handle just the essential relocations.

The faulty code concealed well and I could not have found it without a debugger. It took me a while to set up a m68k image using q800. The memory is limited to 1000MiB and the emulation is very slow. Linux 5.19 is expected to gain the support for a virtual Motorola 68000 machine. With qemu-system-m68k -M virt things will become better.

## musl rtld

musl rtld has a clear separation of 3 stages.

• stage 1 (ldso/dlstart.c): only relative relocations are applied. This allows static variables can be accessed.
• stage 2 __dls2: This applies non-relative relocations.
• stage 2b __dls2b: Set up thread pointer with a TLS stub.
• stage 3 __dls3: Load the executable and immediately loaded shared objects. Apply relocations and possibly relocate rtld/libc itself again for possible symbol interposition (e.g. R_*_COPY, interposed malloc implementation).

Each stage uses a PC-relative code sequence to load the address of the next stage entry point, and then jump to it. This serves as a strong compiler barrier preventing code reordering.

(In glibc, elf/rtld.c ELF_DYNAMIC_RELOCATE (&bootstrap_map, NULL, 0, 0, 0); is kinda like musl's stage 1 plus stage 2.)

Stage 1 computes the entry of stage 2 with GETFUNCSYM(&dls2, __dls2, base+dyn[DT_PLTGOT]); where GETFUNCSYM is defined for every port:

This approach is elegant. It even allows a static or hidden function call with a dynamic relocation, though I haven't found such an architecture in my testing.