morningman opened a new pull request, #68507:
URL: https://github.com/apache/doris/pull/68507
Cherry-picked from #67877 (master commit `e797e99fcbe`) via its branch-4.1
pick #68339 (commit `5e1ee245e23`) to `branch-4.1.4`.
**Conflict resolution.** This branch is a 2026-09-04 cut of branch-4.1, so
two of the seven files differed:
- `thirdparty/CHANGELOG.md`: branch-4.1's `## 20260901` thrift 0.16.0 →
0.24.0 entry is not on this branch (it still pins thrift 0.16.0), and neither
is master's `## 20260909` snappy 1.2.1 entry. Only the `## 20260911` libunwind
entry is inserted, above `## 20260429`.
- `thirdparty/patches/libunwind-1.8.3-doris-phdr-cache.patch`: this branch
still carries the pre-#67564 1.6.2 patch, and a patch file has no mergeable
content, so the new file is taken wholesale from the branch-4.1 pick. It is
byte-identical to the branch-4.1 and master versions.
The other five files applied without conflict. `thirdparty/vars.sh`,
`thirdparty/build-thirdparty.sh`, `thirdparty/download-thirdparty.sh` and
`.github/workflows/build-thirdparty.yml` still differ from branch-4.1, but only
by thirdparty drift this branch has not taken (thrift 0.24.0, aws-sdk-cpp
1.11.221, simdutf, google-cloud-cpp, lance-c 0.1.9, jemalloc); no hunk of this
change is affected. `thirdparty/test/libunwind-page-size-test.sh` and the
rebased patch are byte-identical to master's.
### What problem does this PR solve?
Related PRs: #64093, #67877, #68339
Problem Summary:
**Every aarch64 BE built from branch-4.1.4 since 4.1.4-rc01 crashes before
`main()` on a 64 KiB-page kernel** (Kunpeng 920 + Kylin V10 / openEuler /
CentOS aarch64, the common ARM server stack). The core from the 4.1.4 arm64
deployment shows:
```
access_mem (addr=0) src/aarch64/Ginit.c
is_plt_entry src/aarch64/Gstep.c
_ULaarch64_step
_ULaarch64_tdep_trace
unw_backtrace
<bthread/mutex.cpp static initializer: backtrace(dummy_buf, 4)> ← bRPC
"Warm up backtrace before main()"
__libc_csu_init
```
SIGSEGV with `si_addr=0`, AUXV `PAGESZ=65536`, and inside the libunwind
cursor `ip=0`, `validate=0`, `mem_validate_func=msync_validate`,
`last_good_addr[]={0,0,0,0}`.
How it happens:
1. #64093 links GNU libunwind statically into the aarch64 BE (it forces
`USE_UNWIND=ON` on Linux, references `unw_*` from `phdr_cache.cpp` without the
`__x86_64__` guard, and builds jemalloc with `--enable-prof-libunwind`).
libunwind's weak `backtrace` alias then replaces glibc's `backtrace()` for
every caller in the process: bRPC's pre-main warm-up,
`StackTrace::tryCapture()` on aarch64, jemalloc heap profiling. The 4.1.3 arm64
package was immune only because it was built with `USE_UNWIND=OFF` (which the
ARM compilation doc still prescribes and #64093 removed): it has
`backtrace@GLIBC_2.17` and zero `_ULaarch64_*` symbols.
2. libunwind 1.6.2 hard-codes `PAGE_SIZE 4096` in `src/aarch64/Ginit.c`. On
a 64 KiB-page kernel `mincore()`/`msync()` reject its 4 KiB-aligned probes with
`EINVAL`, so `validate_mem()` fails for 15/16 of all addresses (the core's
`msync_validate` + empty `last_good_addr` are exactly this).
3. `unw_step()` therefore gets `unw_is_signal_frame() < 0` for a valid IP,
takes the "IP points to non-mapped memory, use LR" recovery path and loads
`uc.regs[30]` — a slot `_Uaarch64_getcontext_trace` never writes (it saves only
FP/SP/PC). On the reporting host that stack garbage is 0.
4. `dwarf_step(ip=0)` finds nothing, and the fallback calls `is_plt_entry()`
with `c->validate` restored to 0, so `access_mem` executes `ldr x8, [x1]` with
`x1=0`.
The PHDR cache being empty before `main()` is not what breaks this (step 3
has already poisoned `ip` before any FDE lookup, and the same crash reproduces
without the Doris hook); changing the cache initialisation order would not fix
it, and every later `backtrace()` on such a host rolls the same dice.
### Changes
Fix: **upgrade libunwind 1.6.2 → 1.8.3**, which was never bumped since it
was introduced in 2023 (#21938):
- 1.7.0+: `unw_page_size = sysconf(_SC_PAGESIZE)` instead of the hard-coded
4096.
- 1.8.0+: `src/mi/Gaddress_validator.c` validates through the pipe-write
probe only, so `mincore`/`msync` alignment no longer matters at all;
`unw_step()` sets `c->validate = 1` before the DWARF-failure fallback, so a
garbage IP is rejected instead of dereferenced.
Specifically:
- `thirdparty/vars.sh`: libunwind 1.6.2 → 1.8.3 (md5
`13bc7b41462ac6ea157d350eaf6c1503`).
- `thirdparty/patches/libunwind-1.8.3-doris-phdr-cache.patch`: the #64093
hook rebased onto 1.8.3's `as->iterate_phdr_function` call site; semantics
unchanged (`doris_unwind_iterate_phdr` still takes precedence whenever it is
linked in). 1.8 also exposes `unw_set_iterate_phdr_function()`, which could
replace the source patch in a follow-up.
- `thirdparty/build-thirdparty.sh`: `--disable-tests
--disable-documentation`; only `libunwind.a` is consumed.
- `thirdparty/test/libunwind-page-size-test.sh` (+ run at the end of the
Linux thirdparty CI build): reproduces the failure on any Linux host. An
`LD_PRELOAD` shim gives libunwind a 64 KiB-page kernel's behaviour (`EINVAL`
for unaligned `mincore`/`msync`, `sysconf(_SC_PAGESIZE)=65536`), and a program
shaped like the bRPC warm-up calls `backtrace()` from a constructor with a
zeroed stack. Against 1.6.2 it segfaults with the customer's exact stack and
cursor state; against 1.8.3 it passes.
No BE source change: every `unw_*` API the BE, jemalloc and the x86_64
stack-trace code use is unchanged in 1.8.3, `cxx_exceptions` stays off on
x86_64/aarch64, and the weak `backtrace` alias is still provided.
### Release note
Fix the aarch64 BE crashing before `main()` (SIGSEGV in
`access_mem`/`is_plt_entry` under `unw_backtrace`) on 64 KiB-page kernels such
as Kunpeng 920 with Kylin, openEuler or CentOS, by upgrading the bundled
libunwind from 1.6.2 to 1.8.3.
### Check List (For Author)
- Test
- [x] Manual test (add detailed scripts or steps below)
- The change is carried over verbatim from #67877 / #68339, where
the reproduction, the symbolised customer core and the new
`thirdparty/test/libunwind-page-size-test.sh` (1.6.2 FAIL, 1.8.3 PASS on
aarch64 and x86_64) were verified; `Build Third Party Libraries (Linux)`,
`(macOS)`, `(macOS-arm64)` and `Thirdparty Script Test` passed there.
- For this pick: every hunk is identical to the branch-4.1 commit
(only hunk offsets differ); `libunwind-1.8.3-doris-phdr-cache.patch` and
`thirdparty/test/libunwind-page-size-test.sh` are byte-identical to master's;
`bash -n` passes on the four touched scripts; no stale `libunwind-1.6.2`
reference remains in the tree.
- Not done here: a Linux `doris_be` rebuilt with 1.8.3 on a real 64
KiB-page host. That is the one check that still needs an aarch64 machine.
- Behavior changed:
- [x] Yes. Bundled libunwind 1.6.2 → 1.8.3; libunwind's own test
programs and man pages are no longer built. Unwinding semantics for the BE
otherwise unchanged.
- Does this need documentation?
- [x] No.
### Check List (For Reviewer who merge this PR)
- [ ] Confirm the release note
- [ ] Confirm test cases
- [ ] Confirm document
- [ ] Add branch pick label
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
