You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Bug]: Android devices missing on WSL mirrored networking until an adb server is running #37867
Run the T3 server in WSL2 with networkingMode=mirrored in .wslconfig.
Install the Android SDK in WSL (emulator, platform-tools, cmdline-tools) and create an AVD. KVM works.
Make sure no adb server is running (fresh WSL boot, or adb kill-server).
Open the Device panel, or call device_list.
Expected behavior
The AVD is listed. adb starts its server on first use and the hub answers.
Actual behavior
device_list reports "The device hub did not answer within 40 seconds." with incompletePlatforms: ["android"] and lists no Android devices. GET /api/devices on the hub hangs.
The hub's adb devices -l children sit in SYN-SENT to 127.0.0.1:5037. Under mirrored networking, a connect to an unbound 127.0.0.1 port is not refused. It hangs until the SYN retries run out (about 2 minutes), and only then does adb fork its server. 127.0.0.2 and ::1 refuse immediately, and no firewall rule drops the traffic:
A related effect of the same mode: if any Windows process starts adb.exe first, it owns port 5037 for both sides. WSL's adb client and the hub then talk to the Windows adb server, which drives the WSL emulator over the shared localhost. In my case adb shell calls through it hung and the hub timed out the same way.
Impact
On WSL with mirrored networking (the mode needed for most localhost tooling), Android devices don't appear after every WSL boot until something else starts an adb server. The error doesn't point at adb, so it looks like a hub failure.
Windows 11, WSL2 kernel 6.18.58.1, Ubuntu, networkingMode=mirrored, systemd enabled, T3 server as a systemd user service.
Logs or stack traces
{"hostStatuses":{"local":{"status":"ready","detail":"The device hub did not answer within 40 seconds.","incompletePlatforms":["android"]}}, ... "devices":[]}
$ ss -tnp | grep 5037
SYN-SENT 0 1 127.0.0.1:37533 127.0.0.1:5037 users:(("adb",pid=7670,fd=3))
SYN-SENT 0 1 127.0.0.1:38955 127.0.0.1:5037 users:(("adb",pid=4891,fd=3))
Workaround
Run an adb server before the hub needs it, which skips adb's connect-then-fork probe. A systemd user unit works:
[Service]ExecStart=%h/Android/Sdk/platform-tools/adb nodaemon server
Restart=on-failure
With it running, /api/devices answers in under a second and the emulator boots and streams normally.
A possible fix in T3: on Linux, have LocalDeviceHost start and supervise adb nodaemon server (on the resolved SDK's adb) next to the hub when port 5037 is not already listening, instead of relying on adb's auto-start.
Confirming the mechanism: this matches the known mirrored-mode loopback behavior where the first SYN to an unbound 127.0.0.1 port is dropped instead of refused, so adb's connect-then-fork probe waits out its TCP retransmits before the server ever forks. 127.0.0.2 and ::1 take a different path and refuse immediately, as your measurements show. On NAT mode the same first-use works because an unbound port answers RST locally.
Your systemd workaround isolates the bug to adb's probe, and I agree the hub should own the server: have LocalDeviceHost start and supervise adb nodaemon server (on the resolved SDK's adb) when nothing is listening on 5037, and leave an already-running server — including one owned by a Windows adb — untouched.
Checked the open device-hub PRs: #27525 resolves a system adb, #23498 isolates slow hosts, #30629 trims emulator-tool discovery — none prewarm the server, so this is uncovered.
Reproducing now on Windows 11 + WSL2 (Ubuntu 24.04, mirrored, fresh boot + adb kill-server) and taking the prewarm fix.
While checking whether I could reproduce the hang itself, my environment came out healthy at the same kernel as yours (WSL 3.0.5, kernel 6.18.58.1-1, mirrored, cold boot, adb 37.0.1): an unbound 127.0.0.1 port refuses immediately here, and 127.0.0.2 behaves the same way your measurements show. So the mirrored relay defect the hang rides on looks environment-specific beyond the WSL version.
Could you share your wsl --version full output (WSL/MSRDC versions especially), the Windows build (winver), and whether anything third-party sits on loopback traffic (VPN, filter driver, security software)? That would pin which machines the prewarm needs to cover, and possibly a separate WSL issue upstream.
One more detail from this side, in case it helps your diagnosis: in the other direction (Windows process connecting to a WSL-side port), mirrored loopback accepted a connect to a port nothing was listening on — Node's connect succeeded while netstat and ss showed no listener anywhere. So the same relay lies in both directions, just with opposite symptoms.
Activity
Dopetaiga commentedon Nov 17, 2026
Confirming the mechanism: this matches the known mirrored-mode loopback behavior where the first SYN to an unbound
127.0.0.1port is dropped instead of refused, so adb's connect-then-fork probe waits out its TCP retransmits before the server ever forks.127.0.0.2and::1take a different path and refuse immediately, as your measurements show. On NAT mode the same first-use works because an unbound port answers RST locally.Your systemd workaround isolates the bug to adb's probe, and I agree the hub should own the server: have
LocalDeviceHoststart and superviseadb nodaemon server(on the resolved SDK's adb) when nothing is listening on 5037, and leave an already-running server — including one owned by a Windows adb — untouched.Checked the open device-hub PRs: #27525 resolves a system adb, #23498 isolates slow hosts, #30629 trims emulator-tool discovery — none prewarm the server, so this is uncovered.
Reproducing now on Windows 11 + WSL2 (Ubuntu 24.04, mirrored, fresh boot +
adb kill-server) and taking the prewarm fix.Dopetaiga commentedon Nov 17, 2026
Opened #37953 for the prewarm fix.
While checking whether I could reproduce the hang itself, my environment came out healthy at the same kernel as yours (WSL 3.0.5, kernel 6.18.58.1-1, mirrored, cold boot, adb 37.0.1): an unbound 127.0.0.1 port refuses immediately here, and
127.0.0.2behaves the same way your measurements show. So the mirrored relay defect the hang rides on looks environment-specific beyond the WSL version.Could you share your
wsl --versionfull output (WSL/MSRDC versions especially), the Windows build (winver), and whether anything third-party sits on loopback traffic (VPN, filter driver, security software)? That would pin which machines the prewarm needs to cover, and possibly a separate WSL issue upstream.One more detail from this side, in case it helps your diagnosis: in the other direction (Windows process connecting to a WSL-side port), mirrored loopback accepted a connect to a port nothing was listening on — Node's connect succeeded while
netstatandssshowed no listener anywhere. So the same relay lies in both directions, just with opposite symptoms.