[native] Possible SEGV in HttpClient::createTransaction about presto HOT 10 OPEN

We running clang on CentOS9. The proxygen version v2024.04.01.00 should be new enough. Maybe try a TSAN build since this most likely a concurrency issue.

from presto.

@Yuhta FYI.

from presto.

We have not seen this error on 2000+ nodes for the last week. It must be specific to some network configurations or proxygen version.

from presto.

I'm able to see it on a local dev environment but once compiled with debug symbol it is not showing up anymore. This is on CentOS9 Linux with gcc12. I've tried to repro on MacOS with clang but wasn't able to so far. What are you building with?

The proxygen version used is the one from the current setup scripts in prestissimo v2024.04.01.00. So should be the same that you are using or are you using a newer one? We want to update the FBOS dependencies in Velox (and as a result proxygen) to the May 20th version.

from presto.

I was able to reproduce this and can confirm that auto* evb = session->getEventBase(); is causing the segfault. Removing that line does not cause the sigsegv.
I removed this line and added VELOX_CHECK(session->getTransport() != nullptr) to check for the value of sock_ and this would fail as well.
Looking at the HttpSessionBase API, should we not check if the session is ready for reuse?
I added this check and it fixes the problem. #23018
However, I tried to check if there was any reuse after this change, and I did not see the reuse log message.
@Yuhta any thoughts on this?

  if (!session || !session->isReusable()) {
      return nullptr;
    }

from presto.

A simpler fix is to perform the checks if the session is reusable. The fact that the session is not yet reusable makes me wonder if there is a bigger problem here.

    if (session->isReusable()) {
      auto* evb = session->getEventBase();
      // The event base from idle session should not be the current event base,
      // otherwise we should have already got it from the local session pool.
      VELOX_CHECK(!evb || !evb->isInEventBaseThread());
    }

from presto.

@Yuhta can you share your thoughts here? Is it safe to add a session to the sessionPool_ when the session is not yet reusable?

from presto.

@majetideepak It seems isReusable is always false in this case, so we cannot do that. Can you file a bug report with the repro with proxygen (https://github.com/facebook/proxygen/issues)? HttpSession we got from ServerIdleSessionController should be guaranteed to be reusable.

In the meantime you can disable connection pool (exchange.http-client.enable-connection-pool) for your instances.

from presto.

Closed by mistake

from presto.

Meta doesn't see problems using this code but the compiler used is clang (an older version).
The statements causing the segv are validation statements and when removed don't cause issues. Subsequent runs show no further issue.
The TSAN did not show a race condition in this area even after the fix.
Next step: Use clang 17 and see if the generated code causes reproduces the issue as it does with gcc12.

from presto.