mirror of
https://github.com/nghttp2/nghttp2.git
synced 2025-12-06 18:18:52 +08:00
ce385d3f55
This commit fixes memory leak that happens when PUSH_PROMISE or HEADERS frame cannot be sent, and nghttp2_on_stream_close_callback fails with a fatal error. For example, if GOAWAY frame has been received, a HEADERS frame that opens new stream cannot be sent. This issue has already been made public via CVE-2023-35945 [1] issued by envoyproxy/envoy project. During embargo period, the patch to fix this bug was accidentally submitted to nghttp2/nghttp2 repository [2]. And they decided to disclose CVE early. I was notified just 1.5 hours before disclosure. I had no time to respond. PoC described in [1] is quite simple, but I think it is not enough to trigger this bug. While it is true that receiving GOAWAY prevents a client from opening new stream, and nghttp2 enters error handling branch, in order to cause the memory leak, nghttp2_session_close_stream function must return a fatal error. nghttp2 defines 2 fatal error codes: - NGHTTP2_ERR_NOMEM - NGHTTP2_ERR_CALLBACK_FAILURE NGHTTP2_ERR_NOMEM, as its name suggests, indicates out of memory. It is unlikely that a process gets short of memory with this simple PoC scenario unless application does something memory heavy processing. NGHTTP2_ERR_CALLBACK_FAILURE is returned from application defined callback function (nghttp2_on_stream_close_callback, in this case), which indicates something fatal happened inside a callback, and a connection must be closed immediately without any further action. As nghttp2_on_stream_close_error_callback documentation says, any error code other than 0 or NGHTTP2_ERR_CALLBACK_FAILURE is treated as fatal error code. More specifically, it is treated as if NGHTTP2_ERR_CALLBACK_FAILURE is returned. I guess that envoy returns NGHTTP2_ERR_CALLBACK_FAILURE or other error code which is translated into NGHTTP2_ERR_CALLBACK_FAILURE. [1] https://github.com/envoyproxy/envoy/security/advisories/GHSA-jfxv-29pc-x22r [2] https://github.com/nghttp2/nghttp2/pull/1929