[PATCH v2 0/6] x86-64: Use standard frame records in ld.so trampolines
Aliaksei Kandratsenka
alkondratenko@gmail.com
Tue Nov 3 00:02:00 GMT 2026
_dl_runtime_resolve, _dl_runtime_profile and the slow path of
_dl_tlsdesc_dynamic realign the stack and use RBX as the CFA register
while they call into C code. Unwinders that track only SP, FP and RA
(SFrame, the Linux kernel user-space unwinder, aw-backtrace) cannot
unwind through them. Frame-pointer walkers skip a frame, because none
of them sets up a frame record.
I found this with aw-backtrace
(https://github.com/gperftools/aw-backtrace), whose test single-steps
code, runs a signal handler at every instruction boundary, and checks
that the captured backtrace matches the actual call/ret sequence.
This series makes all three functions use RBP as the CFA register and
point RBP at a standard frame record (caller's RBP at 0(%rbp), return
address at 8(%rbp)), so that every unwinder sees the same frame.
1. Remove dead !DL_RUNTIME_RESOLVE_REALIGN_STACK code. No change in
generated code; it makes the following patches simpler.
2. Use RBP instead of RBX in the PLT trampolines. This is v1,
rebased on top of patch 1.
3. Set up a standard frame record in _dl_runtime_resolve.
4. Set up a standard frame record in _dl_runtime_profile.
5. Use RBP instead of RBX in _dl_tlsdesc_dynamic.
6. Set up a standard frame record in _dl_tlsdesc_dynamic. This also
reserves the red-zone slots explicitly instead of through
TLSDESC_CALL_REGISTER_SAVE_AREA (BZ #31501).
Changes since v1, mostly following Adhemerval's review:
- In v1, RBP pointed at the caller's RBP with the link map above it,
so a frame-pointer walker would report the link map as a return
address. Adhemerval suggested pushing a copy of the return
address. Instead, patches 3 and 4 store the caller's RBP in the
reloc index slot pushed by the PLT, which is directly below the
real return address; the reloc index is kept in R11 (or in local
storage in _dl_runtime_profile). This needs no extra push, and the
CFA is RBP + 16 as in any normal frame, so DWARF, SFrame and
frame-pointer walkers all agree on the caller's frame. The
epilogue becomes LEAVE.
- Adhemerval suggested leaving _dl_runtime_profile alone. With this
approach the change costs nothing there either, and
_dl_runtime_profile calls into audit modules and the target
function, so unwinding through it is useful. It also replaces a lea
that only worked because LR_RSP_OFFSET happens to equal 56.
- As Adhemerval pointed out, _dl_tlsdesc_dynamic had the same
RBX-based CFA. Patches 5 and 6 switch it to RBP and give it a
standard frame record too.
Tested on x86_64-linux-gnu: make check on x86_64 and x32, no
regressions compared to the base commit. LD_AUDIT tested manually.
Also verified that aw-backtrace's backtrace-comparer now passes on
/usr/bin/true without suppressing _dl_fixup or _dl_runtime_resolve.
v1: https://sourceware.org/pipermail/libc-alpha/2026-October/181360.html
Aliaksei Kandratsenka (6):
x86-64: Remove dead code for !DL_RUNTIME_RESOLVE_REALIGN_STACK
x86-64: Use RBP instead of RBX as CFA register in PLT trampolines
x86-64: Set up a standard frame record in _dl_runtime_resolve
x86-64: Set up a standard frame record in _dl_runtime_profile
x86-64: Use RBP instead of RBX as CFA register in _dl_tlsdesc_dynamic
x86-64: Set up a standard frame record in _dl_tlsdesc_dynamic
sysdeps/x86/cpu-features.c | 5 +-
sysdeps/x86/sysdep.h | 79 +++----------
sysdeps/x86_64/dl-tlsdesc-dynamic.h | 66 ++++++-----
sysdeps/x86_64/dl-trampoline-state.h | 26 +----
sysdeps/x86_64/dl-trampoline.h | 162 +++++++++++++--------------
5 files changed, 137 insertions(+), 201 deletions(-)
--
2.55.0
More information about the Libc-alpha
mailing list