Delay test UI bug

[FightingForFun]

When checking the profile delay, the counter spins up to a certain point and then freezes, although the notification continues checking everything.

After the profile delay test in the notifications is complete, nothing unfreezes in the program. The counter still freezes in the program.

The "Enable/Disable Connection" button seems to start the VPN connection, but the button doesn't change color or icon.

This happens on both Vivo and Samsung.

19/157 in the program counter freezes.

[2dust]

版本？

[FightingForFun]

This has been the case for the last six months, in each version.

[FightingForFun]

https://limewire.com/d/AcvhZ#x4WxvNyfYf

Real video capture.

[eliotcougar]

Probably related to a specific type of connection that fails in a weird way.
Does this subscription include any CUSTOM profiles?

[FightingForFun]

Export all profiles to buffer -> paste to file.

[FightingForFun]

TCP Ping.

https://limewire.com/d/LDZMW#PHl0qQQGp3

Real video capture.

[FightingForFun]

I set up a check in 1 thread and found which one had the problem.

[FightingForFun]

Hmm. The next TCP ping check stalled there, too, according to the counter. But it passed successfully with Trojan: KR 7... but now it stalled with Trojan: JP 20...

[eliotcougar]

The potential root cause is identified by GPT5.6-Sol... Working on a fix.
Already fixed in #6050

[eliotcougar]

I cannot reproduce the bug on an emulator with the exact set of profiles. The assumed cause at the moment (Main process event/UI backlog) is based solely on static analysis.

[eliotcougar]

After a second pass, I think there are two plausible, interacting hypotheses. Neither is confirmed yet, because I still could not reproduce the permanent freeze on the AOSP emulator with the exact 207-profile set.

1. Activity event-buffer loss

The activity-side MainRepository uses a MutableSharedFlow with a capacity of 64 and DROP_OLDEST:

v2rayNG/V2rayNG/app/src/main/java/com/v2ray/ang/ui/main/MainRepository.kt

Lines 42 to 46 in 4b3c817

	private val _mainServiceEvent = MutableSharedFlow<MainServiceEvent>(
	replay = 0,
	extraBufferCapacity = 64,
	onBufferOverflow = BufferOverflow.DROP_OLDEST
	)

A 207-profile test generates roughly two UI broadcasts per completed profile (result + progress), plus the lifecycle events. If the activity collector falls behind, intermediate result/progress events can therefore be discarded. The current success event also becomes an identity-less MeasureConfigSuccess, so a dropped success cannot be recovered as a targeted row update.

This remains plausible, but buffer overflow alone is an incomplete explanation: DROP_OLDEST preferentially retains newer events, and the final event is emitted after the worker jobs finish. It explains missing results better than it explains a permanently stale post-finish state.

2. Per-result full-group reload storm

Every MeasureConfigSuccess launches a new IO coroutine, removes the cached group, force-reloads the complete group, and publishes another full UI list:

v2rayNG/V2rayNG/app/src/main/java/com/v2ray/ang/ui/main/MainViewModel.kt

Lines 115 to 120 in 4b3c817

	MainServiceEvent.MeasureConfigSuccess -> {
	viewModelScope.launch(ioDispatcher) {
	val gid = testingGroupId ?: uiState.value.selectedGroupId
	cacheMutex.withLock { groupDataCache.remove(gid) }
	updateGroupUi(gid, loadGroup(gid, forceRefresh = true))
	}

The reload decodes every profile and affiliation and is serialized behind a per-group mutex:

v2rayNG/V2rayNG/app/src/main/java/com/v2ray/ang/ui/main/MainViewModel.kt

Lines 246 to 270 in 4b3c817

	private suspend fun buildServersCache(guids: List<String>): List<ServersCache> =
	guids.mapNotNull { guid ->
	currentCoroutineContext().ensureActive()
	val profile = dataSource.decodeServerConfig(guid) ?: return@mapNotNull null
	val affiliation = dataSource.decodeAffiliationInfo(guid)
	ServersCache(
	guid = guid,
	profile = profile.copy(),
	testDelayMillis = affiliation?.testDelayMillis ?: 0L
	)
	}
	private suspend fun loadGroup(
	groupId: String,
	forceRefresh: Boolean = false
	): List<ServersCache> {
	val loadMutex = groupLoadMutexes.computeIfAbsent(groupId) { Mutex() }
	return loadMutex.withLock {
	if (!forceRefresh) {
	cacheMutex.withLock { groupDataCache[groupId]?.let { return@withLock it } }
	}
	val servers = buildServersCache(dataSource.getServerGuidList(groupId))
	currentCoroutineContext().ensureActive()
	cacheMutex.withLock { groupDataCache[groupId] = servers }
	servers

For 207 results, this can perform up to 207 x 207 = 42,849 profile decodes, plus as many as 207 complete list publications and their Compose work. This can create a long consumer/UI backlog while the foreground test service—and therefore its notification—continues normally. It also fits the connection button becoming stale at the same time.

In a discriminating emulator run with the exact set, TCP concurrency 20, two-column UI, and QEMU restricted to one host CPU, upstream still recovered and finished. However, logcat recorded severe starvation: bursts of up to 147 skipped frames and individual app frames around 2.6–3.6 seconds. That supports the backlog mechanism, but is not proof of the permanent freeze seen on Vivo/Samsung.

PR #6050 likely addresses both paths: it enlarges the event buffer, carries GUID + delay in each result, coalesces direct row updates instead of rebuilding the group per result, throttles progress events, and forces the final pending flush before completing. Its exact-profile emulator tests passed, but upstream also passed there, so this is not yet controlled A/B proof.

The cleanest discriminator on an affected phone would be a batch ID + monotonic sequence logged at worker send, broadcast receiver, SharedFlow collector, and result-application completion. That would separate broadcast loss, collector starvation, and the reload queue.

[FightingForFun]

Logcat:

Reserve copy (configuration backup)

When I copy the full KR 7 config and paste it into another tab, it becomes "custom" instead of "Trojan".

[eliotcougar]

Logcat:

This doesn't narrow down the cause, but it proves the problem is not caused by any single profile in the subscription. Do you experience the whole app UI lag when the testing is going?

@2dust Is it acceptable to give people a debug-signed APK for testing? I'll change the app id and icon to something like V2rayNG-TEST.

[eliotcougar]

When I copy the full KR 7 config and paste it into another tab, it becomes "custom" instead of "Trojan".

It is an expected behavior.