On 22 May 2015 at 16:36, Jun Koi <junkoi2...@gmail.com> wrote: > In general, we would have timer & code execution run in parallel,
No code has to actually run when there's an active timer; the event loop thread mostly sits blocked waiting for something interesting to happen (io or a timer firing). When it does then the event loop thread will call the generic timer code, which will call the callback function for that timer. Typically this is inside the device model for whatever timer the hardware you're modelling is. It will then (assuming it really wants to create a simulated interrupt) call qemu_irq_set() to set its outbound IRQ line. This will then end up calling into the emulated interrupt controller and eventually (if the irq is not masked by the interrupt controller) into the emulated CPU, which calls cpu_interrupt(). In TCG that ends up calling tcg_handle_interrupt(), which sets cpu->tcg_exit_req. Every block of translated code has an initial part (written by gen_tb_start()) which checks this flag, and returns to the TCG main loop (in cpu-exec.c) if it is set. Every time round the TCG main loop we check if there are any pending interrupts, and if so we call the appropriate CPU function to emulate the "take an interrupt" behaviour of the target CPU. -- PMM
