Repository navigation
Description
Describe the bug
Update (edited): after filing I found the actual cause. Apple already fixed the MULX bug; OrbStack is loading an older Rosetta for Linux runtime than the one macOS ships, so amd64 guests still have it.
OrbStack 2.2.3 on macOS 26.6.2 maps Rosetta 367.4 into x86 processes (/mnt/rv/[rosetta]), while the host's /Library/Apple/usr/libexec/oah/RosettaLinux/rosetta is 367.9. Apple's container CLI (1.2.2) on the same Mac maps the host file (/run/rosetta/rosetta, identical hash) and the test below is correct there; under OrbStack it is wrong.
host file: /Library/Apple/usr/libexec/oah/RosettaLinux/rosetta -> 367.9, md5(first 64KiB) f3c478849d70e4e98244393a57361216
apple/container: /run/rosetta/rosetta (mapped in guest) -> md5(first 64KiB) f3c478849d70e4e98244393a57361216 (same as host)
OrbStack: /mnt/rv/[rosetta] (mapped in guest) -> 367.4, md5(first 64KiB) c5e224d10a7177399aff92ad31085375 (older build)
The hashes come from rosetta_fp.py (inline at the bottom), which reads the mapped Rosetta runtime out of /proc/self/mem and prints a hash of its first 64 KiB plus version-like strings. Rosetta AOT is not involved: freshly compiled binaries fail and ROSETTA_DISABLE_AOT=1 changes nothing.
Why it matters: Rosetta 367.4 still has Apple's MULX bug (rdar://171843590, diagnosed by Russ Cox in FiloSottile/age#633). MULX must leave flags untouched, but the old runtime clobbers the carry flag, so any carry-chain arithmetic after it is wrong. Since the guest CPUID advertises BMI2, anything doing runtime CPU detection (Go, Zig -mcpu=native, OpenSSL, ...) or built with -mbmi2 takes the broken path, and the failures look like application bugs.
Real-world impact I hit: a Zig program built inside the container fails every TLS handshake to hosts whose certificate chain has an ECDSA P-384 signer (x.ai, Google, GitHub, example.com) with CertificateSignatureInvalid, while RSA chains work. age users in the linked thread saw undecryptable files. It took a full day to trace back to the emulator.
To Reproduce
- Apple Silicon Mac, macOS 26.6.2, OrbStack 2.2.3.
- Start an amd64 container, install gcc, and run the attached
mulxcheck.c(sets CF with an overflowingADD, executesMULX, reads CF back withADC):
docker run --rm -it --platform linux/amd64 ubuntu:24.04 bash
apt-get update && apt-get install -y gcc
gcc -O0 -mbmi2 -o mulxcheck mulxcheck.c && ./mulxcheck
OrbStack output:
cpuid advertises BMI2: yes
carry flag after MULX: 0 -> WRONG (MULX clobbered CF)
Same binary, same Mac, under container run --arch amd64 ... (Apple container CLI 1.2.2):
cpuid advertises BMI2: yes
carry flag after MULX: 1 -> CORRECT
- Run
python3 rosetta_fp.pyinside each guest to see which runtime is mapped (output above).
Guest environment: kernel 7.0.14-orbstack-00380-ga7e0a2dc9535, /proc/cpuinfo model VirtualApple @ 2.50GHz, flags include bmi1 bmi2.
Expected behavior
OrbStack should load the Rosetta runtime currently installed on the host, or refresh its copy whenever the host's changes (e.g. after a macOS update), so amd64 guests get Apple's fixes. With 367.9 the test prints carry flag after MULX: 1 -> CORRECT.
Diagnostic report (REQUIRED)
OrbStack info:
Version: 2.2.3
Commit: c83556b0ef8f1ba9a33abbb194622b6b7a1c0307 (v2.2.3)
System info:
macOS: 26.6.2 (25G83)
CPU: arm64, 12 cores
CPU model: Apple M4 Pro
Model: Mac16,8
Memory: 48 GiB
Full report: https://orbstack.dev/_admin/diag/orbstack-diagreport_2026-09-02T12-42-33.901573Z.zip
Screenshots and additional context (optional)
- Upstream diagnosis of the MULX bug and Go's workaround: Suspect broken encryptions/decryptions under macOS Rosetta and virtualization FiloSottile/age#633
- Apple radar: rdar://171843590
- Attached:
mulxcheck.c(standalone C test, needs only gcc).rosetta_fp.py(runtime fingerprint) is inline below.
rosetta_fp.py
import re, hashlib, os
maps = open("/proc/self/maps").read().splitlines()
mem = open("/proc/self/mem", "rb")
chunks = []
for line in maps:
if "rosetta" not in line.lower(): continue
parts = line.split()
start, end = (int(x, 16) for x in parts[0].split("-"))
perms, off = parts[1], int(parts[2], 16)
try:
mem.seek(start); data = mem.read(end - start)
except OSError as e:
print(f"{parts[0]} {perms} off={off:#x} unreadable: {e}"); continue
chunks.append((off, perms, data))
print(f"{parts[0]} {perms} off={off:#x} size={end-start:#x} {parts[5] if len(parts)>5 else ''}")
chunks.sort()
if chunks:
first = chunks[0][2]
print("ELF magic:", first[:4], "class/machine bytes:", first[4], first[18:20].hex())
print("md5 of first 64KiB at file offset", hex(chunks[0][0]), ":", hashlib.md5(first[:65536]).hexdigest())
blob = b"".join(c[2] for c in chunks)
print("total mapped bytes:", len(blob))
seen = set()
for m in re.finditer(rb"[\x20-\x7e]{6,120}", blob):
s = m.group().decode()
if re.search(r"(?i)rosetta|version|build|libexec/oah|advertise|AOT|cache", s) and s not in seen:
seen.add(s); print(" str:", s)
if len(seen) > 40: break
Activity
[-]Rosetta amd64 machines/containers expose BMI2 in CPUID, but Rosetta's MULX clobbers the carry flag (breaks crypto, TLS)[/-][+]amd64 guests run a stale Rosetta (367.4) instead of the host's fixed 367.9: MULX clobbers the carry flag, breaking crypto/TLS[/+]gthoppae commentedon Sep 2, 2026
Edited the title and description: the root cause is not the CPUID flag. Apple has already fixed the
MULXbug, and the host's Rosetta for Linux is 367.9 (Apple'scontainerCLI on the same Mac maps that file and is correct). OrbStack 2.2.3 maps an older 367.4 runtime into amd64 guests (/mnt/rv/[rosetta]), which still has the bug. Details and a runtime fingerprint script are in the updated description. So the ask is simply: load (or refresh to) the host's current Rosetta runtime.kdrag0n commentedon Oct 10, 2026
Fixed for the next version.
kdrag0n commentedon Oct 12, 2026
Fixed in v2.2.4.