rtld: Add support for arm64 variant pcs
Needs ReviewPublic
Actions

Authored by andrew on Fri, Apr 19, 3:37 PM.

Details

Reviewers

kib

Group Reviewers

arm64

Summary

The aarch64 ELF spec has support for a variant of the normal procedure
call standard that doesn't follow the normal register convention, e.g.
using more registers as arguments, or different register state is
preserved.

Add support to rtld to handle this. As we don't know which registers
need to be preserved disable lazy binding for these functions.

Diff Detail

Repository

rG FreeBSD src repository

Lint

Lint Skipped

Unit

Tests Skipped

Build Status

Buildable 57223
Build 54111: arc lint + arc unit

Event Timeline

andrew created this revision.Fri, Apr 19, 3:37 PM

Herald added subscribers: bdragon, emaste, imp. · View Herald TranscriptFri, Apr 19, 3:37 PM

andrew requested review of this revision.Fri, Apr 19, 3:37 PM

andrew added a parent revision: D44868: elf_common.h: Add STO_AARCH64_VARIANT_PCS.

Harbormaster completed remote builds in B57220: Diff 137394.Fri, Apr 19, 3:37 PM

andrew added a child revision: D44870: rtld: Add arm64 variant pcs tests.Fri, Apr 19, 3:37 PM

emaste added inline comments.Fri, Apr 19, 3:52 PM

libexec/rtld-elf/aarch64/reloc.c
252	we can't be what?

The code likely is good, but the why for some of it isn't obvious to me and needs some clarification I think.

libexec/rtld-elf/aarch64/reloc.c
252	this sentence just trails off...
264	this effectively short-circuits the PLT relocations and winding up in rtld_die() with an error message of Fatal Error. How will that be tracked down?
268	How is this related? We didn't do this before and gnu_ifunc behavior is changed (before we'd always compute where, now we skip it)
271	and why this behavior change. Before we'd bias the relocation by relocbase, but here we're computing the value based on the st_value of the symbol. Are these two things expected to be the same?

andrew retitled this revision from rtld: Add suport for arm64 variant pcs to rtld: Add support for arm64 variant pcs.Fri, Apr 19, 3:56 PM

Fix a comment
Remove debugging

Harbormaster completed remote builds in B57223: Diff 137397.Fri, Apr 19, 4:21 PM

andrew added inline comments.Fri, Apr 19, 4:22 PM

libexec/rtld-elf/aarch64/reloc.c
264	This is based on the code in `reloc_jmpslots` which is run just after `reloc_plt` when bind_now (either per-object, or globally).

kib added inline comments.Fri, Apr 19, 9:18 PM

libexec/rtld-elf/rtld.c
1539	Same, would be nice to provide MD hook for MD dyn tags.
libexec/rtld-elf/rtld.h
187	We usually declare bools as bool bitfield, see below in the structure. Since we are growing more MD fields in the Obj_Entry, can we provide MD define for them instead of adding more #ifdefs there?