Idle connection breaks when turning data off and on.

[roberthoenig]

I first thought this to be an issue with React Native and filed an issue there: facebook/react-native#19709

However, further debugging makes me believe that the issue might lie within okhttp.

I created a tiny Github repo that reproduces this bug with okhttp in an Android app:

https://github.com/roberthoenig/react-native-fetch-bug

(Don't get confused by the repo name react-native-fetch-bug: The example doesn't use React Native at all, it's just where I encountered this bug first.)

The repo's README.md also contains steps to reproduce the issue.

Here is my theory of what's going on:

1. When retrieving a resource with okhttp, the connection used for the retrieval gets stored in the client's connectionPool.
2. When turning off the network (e.g. by disabling wifi and data) the idle connections in the connectionPool somehow get corrupted.
3. After turning the network back on and making a new request to the same resource, okhttp will try to reuse the existing idle connection in the connectionPool. That connection is corrupted, however, and will silently fail. The symptom is that no request gets dispatched, and no errors are thrown. After a couple minutes, however, things seem to recover and the request gets dispatched as if nothing happened.

[roberthoenig]

Clarification:

Here are the reproduction steps. I didn't include them in the original post because they are already in the repo I linked, but it's probably useful to have them visible here as well):

Reproduction

0. Run the Android app https://github.com/roberthoenig/react-native-fetch-bug/blob/master/README.md
   It contains two buttons: FETCH, which fetches http://publicobject.com/helloworld.txt and EVICT CONNECTION POOL, which calls evictAll().
1. Quickly press FETCH repeatedly. The requested resources come in, headers are logged.
2. Quickly turn off data.
3. Turn on data.
4. Press FETCH again. Wait for some time. No response will come in.
5. Press EVICT CONNECTION POOL.
6. Press FETCH again. The requested resource comes in, headers are logged.

[yschimke]

@roberthoenig Thanks for the thorough reproduction, I'll take a look in next couple of days if no-one beats me to it. If you get any more insight please post here as there are a number of issues seen that are some form of Android network disconnects not observed after airplane mode.

[roberthoenig]

Just a quick note: I don't actually know that disconnecting corrupts the idle connections. All I observed is that it they get corrupted somehwere between turning data off and on.

[yschimke]

@roberthoenig Thanks for the app, unfortunately I couldn't really reproduce what you are seeing. Is it some particular android version you are testing with? Is it only on devices or emulators?

I do see a delay in the request after reconnect, but it succeeds and then following requests work fine.

I'm sure I'm just doing something differently to you, and I've seen similar problems myself.

I've got a similar test app https://github.com/yschimke/okhttp-testapp

[roberthoenig]

Is it some particular android version you are testing with?

I'm testing it on an emulator with the following details:

Name: small_zulip
CPU/ABI: Google APIs Intel Atom (x86_64)
Path: /home/robert/.android/avd/small_zulip.avd
Target: google_apis [Google APIs] (API level 23)
Skin: 480x800
SD Card: 100M
hw.dPad: no
hw.lcd.height: 800
runtime.network.speed: full
hw.accelerometer: yes
hw.device.name: 4in WVGA (Nexus S)
vm.heapSize: 48
skin.dynamic: yes
hw.device.manufacturer: Generic
hw.lcd.width: 480
hw.gps: yes
hw.initialOrientation: Portrait
skin.path.backup: _no_skin
image.androidVersion.api: 23
hw.audioInput: yes
image.sysdir.1: system-images/android-23/google_apis/x86_64/
tag.id: google_apis
showDeviceFrame: no
hw.camera.back: virtualscene
hw.mainKeys: yes
AvdId: small_zulip
hw.camera.front: emulated
hw.lcd.density: 240
avd.ini.displayname: small_zulip
hw.arc: false
hw.gpu.mode: auto
hw.device.hash2: MD5:0c867f9f3b5d06e6018bb60c4f64aed3
hw.ramSize: 512
hw.trackBall: no
PlayStore.enabled: false
fastboot.forceColdBoot: no
hw.battery: yes
hw.cpu.ncore: 4
hw.sdCard: yes
tag.display: Google APIs
runtime.network.latency: none
hw.keyboard: yes
hw.sensors.proximity: yes
disk.dataPartition.size: 800M
hw.sensors.orientation: yes
avd.ini.encoding: UTF-8
hw.gpu.enabled: yes

It should be a standard configuration of a Nexus S with API level 23. Nothing spcial about the choice of device. The API level is the target API of React Native. My build.gradle has the following okhttp dependency:

compile 'com.squareup.okhttp3:okhttp:3.6.0'

Is it only on devices or emulators?

The original reports of this issue zulip/zulip-mobile#2287 zulip/zulip-mobile#2310 zulip/zulip-mobile#2315 were on real devices. I only reproduced the issue on an emulator. Didn't try myself on a real device yet.

I do see a delay in the request after reconnect, but it succeeds and then following requests work fine.

Hmm, how long is the delay?

One thing noteworthy about my repro steps is that I turn the network off and on with

$ adb shell svc data disable
$ adb shell svc data enable

Also, I just tried to reproduce it myself again with the steps outlined above:

1. Quickly press FETCH repeatedly. The requested resources come in, headers are logged.
2. Quickly turn off data.
3. Turn on data.
4. Press FETCH again. Wait for some time. No response will come in.
5. Press EVICT CONNECTION POOL.
6. Press FETCH again. The requested resource comes in, headers are logged.

Step 1 wasn't actually necessary for things to break. Might be for others, though.

[swankjesse]

Does it time out? Seems like a bug in Android.

[swankjesse]

Unclear what action to take here.