I added a test suite putting load onto a CGI script to test IO and got spurious 500 responses. The log showed the infamous "End of script output before headers" which we had in another context already. Turns out that mod_cgid sends the conn_rec->id to the daemon as an identifier for the *request*. That did not work well, with http2 slave connections having the same id as their master (until today).
Any advice on how to convert the master connection id into a unique slave connection id with high probability? I currently do a XOR with the requests pool address, but casting to (unsigned long) might be a constant value on some architectures. Some APR random() thingie you can recommend? Add a function to mpm to hand out ids? -Stefan
