h2 slave connection reuse does not work reliably

[arminabf]

When doing some tests with the mod_http2 module I'd realized that slave connections (the "cloned" connections from the master connection, used by h2_task's) are mostly (not always) closed after a request - although connection re-use (keep-alive) should be called into action.

So far I observe following behaviour:

When the corresponding handler finishes, ap_check_pipline() is called by ap_process_request_after_handler(). ap_check_pipline() tries to read "blank lines" in AP_MODE_SPECULATIVE mode. This triggers the h2_filter_slave_in() filter that in the majority of cases returns with APR_EOF.

Based on the return code ap_check_pipeline() assumes that the pipe is dead and closes the connection.

else if (rv != APR_SUCCESS || APR_BRIGADE_EMPTY(bb)) {
                if (rv != APR_SUCCESS && !APR_STATUS_IS_EAGAIN(rv)) {
                    /* Pipe is dead */
                    c->keepalive = AP_CONN_CLOSE;
                }   

In the end, when the h2_task is destroyed (in task_destroy()) the slave connection is not set aside for connection reuse but is destroyed and the underlying socket gets closed.

    if (slave) {
      if (reuse_slave && slave->keepalive == AP_CONN_KEEPALIVE) {
        APR_ARRAY_PUSH(m->spare_slaves, conn_rec *) = slave;
      } else {
        slave->sbh = NULL;
        h2_slave_destroy(slave);
      }    
    }

As mentioned, sometimes a slave connection is reused...but very infrequently.

The tests have been made with "mod_http2 (v1.8.3, feats=CHPRIO+SHA256+INVHD, nghttp2 1.18.1)" based on Apache httpd v2.4.25.

[icing]

Thanks! Very interesting observation and certainly something that needs fixing.

[icing]

Will be fixed in the upcoming v1.9.0.

[icing]

@arminabf can you verify that v1.9.0 works on your setup too? Thanks!

[arminabf]

yes works, thnx!!!