http2: mitigate reported DoS attacks #29122

This includes mitigations for CVE-2019-9512/CVE-2019-9515.

DRY up the `debug()` calls, and in particular, avoid building template strings before we know whether we need to.

For some JS events, it only makes sense to call into JS when there are listeners for the event in question. The overhead is noticeable if a lot of these events are emitted during the lifetime of a session. To reduce this overhead, keep track of whether any/how many JS listeners are present, and if there are none, skip calls into JS altogether. This is part of performance improvements to mitigate CVE-2019-9513.

Lazily allocate `ArrayBuffer`s for the contents of DATA frames. Creating `ArrayBuffer`s is, sadly, not a cheap operation with V8. This is part of performance improvements to mitigate CVE-2019-9513. Together with the previous commit, these changes improve throughput in the adversarial case by about 100 %, and there is little more that we can do besides artificially limiting the rate of incoming metadata frames (i.e. after this patch, CPU usage is virtually exclusively in libnghttp2).

Limit the number of streams that are rejected upon creation. Since each such rejection is associated with an `NGHTTP2_ENHANCE_YOUR_CALM` error that should tell the peer to not open any more streams, continuing to open streams should be read as a sign of a misbehaving peer. The limit is currently set to 100 but could be changed or made configurable. This is intended to mitigate CVE-2019-9514.

Limit the number of invalid input frames, as they may be pointing towards a misbehaving peer. The limit is currently set to 1000 but could be changed or made configurable. This is intended to mitigate CVE-2019-9514.

Ignore headers with 0-length names and track memory for headers the way we track it for other HTTP/2 session memory too. This is intended to mitigate CVE-2019-9516.

Allocating memory upfront comes with overhead, and in particular, `std::vector` implementations do not necessarily return memory to the system when one might expect that (e.g. after shrinking the vector).

This is intended to mitigate CVE-2019-9518.

If a write to the underlying socket finishes asynchronously, that means that we cannot write any more data at that point without waiting for it to finish. If this happens, we should also not be producing any more input. This is part of mitigating CVE-2019-9511/CVE-2019-9517.

If we are waiting for the ability to send more output, we should not process more input. This commit a) makes us send output earlier, during processing of input, if we accumulate a lot and b) allows interrupting the call into nghttp2 that processes input data and resuming it at a later time, if we do find ourselves in a position where we are waiting to be able to send more output. This is part of mitigating CVE-2019-9511/CVE-2019-9517.

nghttp2 has updated its limit for outstanding Ping/Settings ACKs to 1000. This commit allows reverting to the old default of 10000. The associated CVEs are CVE-2019-9512/CVE-2019-9515.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

http2: mitigate reported DoS attacks #29122

http2: mitigate reported DoS attacks #29122

Commits on Aug 14, 2019