This attached patch enables pgbench to cancel queries during benchmark.
Formerly, Ctrl+C during benchmark killed pgbench immediately, but backend processes executing long queries remained for a while. You can simply reproduce this problem by cancelling the pgbench running a custom script executing "SELECT pg_sleep(10)".
The patch fixes this so that cancel requests are sent for all connections on Ctrl+C, and all running queries are cancelled before pgbench exits.
In thread #0, setup_cancel_handler is called before the loop, the CancelRequested flag is set when Ctrl+C is issued. In the loop, cancel requests are sent when the flag is set only in thread #0. SIGINT is blocked in other threads, but the threads will exit after their query are cancelled. If thread safety is disabled or OS is Windows, the signal is not blocked because pthread_sigmask cannot be used. (I didn't test the patch on WIndows yet, though.)
I choose the design that the signal handler and the query cancel are performed only in thread #0 because I wanted to make the behavior as predicable as possible. However, another design that all thread can received SIGINT and that the first thread that catches the signal is responsible to sent cancel requests for all connections may also work.
Also, the array of CState that contains all clients state is changed to a global variable so that all connections can be accessed within a thread.
+1
I also like the thread #0 handling design. I have NOT reviewed/tested this yet.