RFR: 8371903: HttpClient: improve handling of HTTP/2 GOAWAY frames with error code [v8]
Daniel Jeliński
djelinski at openjdk.org
Thu Dec 18 16:42:31 UTC 2025
On Sat, 13 Dec 2025 05:18:57 GMT, EunHyunsu <duke at openjdk.org> wrote:
>> ### Problem
>>
>> When the HTTP/2 client receives a GOAWAY frame with a non-zero error code, the current implementation discards both the error code and debug data. Users only see generic "Connection closed by peer" errors without any information about why the server terminated the connection.
>>
>> ### Solution
>>
>> Per [RFC 9113 §5.4.1](https://www.rfc-editor.org/rfc/rfc9113.html#section-5.4.1), a GOAWAY frame with a non-zero error code indicates a connection error requiring immediate closure. This fix:
>>
>> 1. **Distinguishes** graceful shutdown (NO_ERROR) from connection errors
>> 2. **Preserves** error code and debug data in exception messages
>> 3. **Categorizes** streams based on `lastStreamId`:
>> - Streams with ID > `lastStreamId`: Marked as unprocessed for automatic retry
>> - Streams with ID ≤ `lastStreamId`: Failed with detailed error information
>>
>> ### Changes
>>
>> **Core Implementation** (`Http2Connection.java`):
>> - Modified `handleGoAway()` to check error code and route appropriately
>> - Added `handleGoAwayWithError()` method that:
>> - Extracts error code and debug data from GOAWAY frame
>> - Creates meaningful error messages with error name, hex code, and debug data
>> - Properly categorizes streams for retry or failure
>>
>> **Test Infrastructure**:
>> - `Http2TestServerConnection.sendGoAway(int, int, byte[])`: Supports custom error codes
>> - `Http2TestExchangeImpl.getServerConnection()`: Accessor for test handlers
>> - `GoAwayWithErrorTest`: Verifies proper error propagation
>>
>> ### Example
>>
>> **Before:**
>> IOException: Connection closed by peer
>>
>> **After:**
>> IOException: Received GOAWAY with error code Protocol error (0x1): Invalid HEADERS frame
>>
>> ### Testing
>>
>> - New `GoAwayWithErrorTest` passes
>> - Existing HTTP/2 tests unaffected (NO_ERROR path unchanged)
>> - Backward compatible (no public API changes)
>
> EunHyunsu has updated the pull request incrementally with one additional commit since the last revision:
>
> 8371903: Address code review feedback
Occasionally more than one request fails with the GOAWAY exception. I suppose this might be because of a race between the thread that iterates over the list of streams, and another thread adding new streams to the list; the newly added streams may be missed by the iterator, and then are subsequently closed with the connection's termination cause. I only observed the failure 3 times in 2500 repeats.
The exception itself is not visible in the test output; printStackTrace helps a lot in diagnosing test failures, could you add it?
-------------
PR Comment: https://git.openjdk.org/jdk/pull/28632#issuecomment-3671151592
More information about the net-dev
mailing list