Investigating test failures in testing/scenarios

[Copilot]

Pull request created by AI Agent

[Loyalsoldier]

@copilot Review this PR, find potential bugs, and fix them if any. Write summary for the PR.

[Copilot]

@copilot Review this PR, find potential bugs, and fix them if any. Write summary for the PR.

Reviewed the PR and found three additional bugs, fixed in 0959ba62 and a99886e9.

1. Pending read silently dropped when the reader returns data with io.EOF (connection.go)

io.Reader implementations may return (n > 0, io.EOF), and io.MultiReader — which the hub.go fix introduces — does exactly that when its last reader does. Read threw those n bytes away and overwrote the caller's buffer with a fresh socket read. Since every server-side connection now goes through the pending-read path, this latent bug became reachable and would silently truncate the stream.

2. Panic + permanent deadlock on a negative maxEarlyData (dialer.go, connection.go)

maxEarlyData is an unvalidated user-supplied int32. A negative value made earlyDataSize negative, so earlyData[earlyDataSize:] panicked with "slice bounds out of range". The panic then skipped finishedDial(), so Close()/Read()/LocalAddr() blocked forever on delayedDialFinish — the process hung instead of failing. This was pre-existing upstream and survived my first commit. maxEarlyData is now clamped to >= 0, and finishedDial() is deferred so it always runs, and only after every field it publishes is assigned (which also closes a pre-existing race where remoteAddr/shouldWait were written after the cancel that unblocks waiters).

3. Interop regression I had introduced (hub.go)

My first commit added a server-side len(earlyDataBytes) > MaxEarlyData rejection that does not exist upstream. maxEarlyData is configured independently on each side, so a client at 2048 talking to a server at 1024 would have had every connection dropped. It also provides no protection — http.ReadRequest and base64.DecodeString have already allocated everything before the check runs. Removed, restoring upstream behaviour.

I also made the server allocate no pending reader at all when nothing is buffered, keeping the original zero-overhead path for the common case.

Regression tests (httpupgrade_test.go) cover early data larger/smaller than maxEarlyData, no early data, negative maxEarlyData, a reader ending with io.EOF, and early data arriving in the same segment as the upgrade request. Each was verified to fail without its corresponding fix.

Validation: go test ./transport/internet/httpupgrade/... -race 6/6 pass; go test ./transport/... 60/60 packages pass; TestReverseProxy and TestVMessGCMMuxUDP pass; go vet/gofmt/gofumpt clean; CodeQL clean. transport/internet/hysteria2 TestUDP hangs, but it does so identically on the base commit db129141 and has no dependency on httpupgrade, so it is unrelated to this PR.

The PR summary has been rewritten with the full breakdown.

[codecov]

Codecov Report

❌ Patch coverage is 96.15385% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 25.15%. Comparing base (db12914) to head (a99886e).

Files with missing lines	Patch %	Lines
testing/scenarios/common.go	85.71%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #3798      +/-   ##
==========================================
+ Coverage   24.93%   25.15%   +0.22%     
==========================================
  Files         916      916              
  Lines       57820    57834      +14     
==========================================
+ Hits        14418    14549     +131     
+ Misses      41557    41407     -150     
- Partials     1845     1878      +33     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Loyalsoldier]

@copilot Modify the title and summary of this PR based on code changes.

[Copilot]

@copilot Modify the title and summary of this PR based on code changes.

Title is now Fix: httpupgrade early data corruption and testing/scenarios test failures, and the summary has been rewritten from "Pull request created by AI Agent" to describe what the five commits actually change:

transport/internet/httpupgrade

1. dialer.go — early data was sent twice when the payload exceeded MaxEarlyData: the header was encoded from the whole earlyData slice while earlyData[earlyDataSize:] was also written to the socket, so the server received payload + payload[MaxEarlyData:]. The header is now encoded from earlyData[:earlyDataSize].
2. hub.go — bytes arriving in the same TCP segment as the upgrade request land in the bufio.Reader used by http.ReadRequest and were discarded. They are now replayed with io.LimitReader, chained after the early data via io.MultiReader. newConnectionWithRemoteAddr() became unused and was removed.
3. connection.go — a pending read returning (n > 0, io.EOF) had those bytes overwritten by a fresh socket read.
4. dialer.go / connection.go — a negative MaxEarlyData panicked on earlyData[earlyDataSize:] and, because finishedDial() was called inline, left every waiter blocked forever. MaxEarlyData is clamped to >= 0 and finishedDial() is deferred.

testing/scenarios

• TestReverseProxy: payload reduced from 10 MiB to 1 MiB per connection to bound CI memory usage.
• TestVMessGCMMuxUDP: uses the new testUDPConnWithRetry to tolerate the datagram loss V2Ray deliberately causes when an inbound UDP buffer is full.

Tests: transport/internet/httpupgrade/httpupgrade_test.go covers early data larger/smaller than MaxEarlyData, no early data, negative MaxEarlyData, a reader ending with io.EOF, and early data arriving in the same segment as the upgrade request.

No code changed in this round — this was a title/description update only. Re-verified on a99886e9: go test ./transport/internet/httpupgrade/... -race passes, go vet and gofmt are clean, and the Test, Linter, CodeQL and Release workflows are green on the branch.