The mi_switch function implements the machine independent prelude to a thread context switch. It is called from only a few distinguished places in the kernel code as a result of the principle of non-preemptable kernel mode execution. The various major uses of mi_switch can be enumerated as follows:
- From within sleep(9), tsleep(9) and msleep(9) when the current thread voluntarily relinquishes the CPU to wait for some resource to become available.
- After handling a trap (e.g. a system call, device interrupt) when the kernel prepares a return to user-mode execution. This case is typically handled by machine dependent trap-handling code after detection of a change in the signal disposition of the current process, or when a higher priority thread might be available to run. The latter event is communicated by the machine independent scheduling routines by calling the machine defined need_resched.
- In the signal handling code (see issignal(9)) if a signal is delivered that causes a process to stop.
- When a thread dies in thread_exit(9) and control of the processor can be passed to the next runnable thread.
- In thread_suspend_check(9) where a thread needs to stop execution due to the suspension state of the process as a whole.
mi_switch records the amount of time the current thread has been running in the process structures and checks this value against the CPU time limits allocated to the process (see getrlimit(2)). Exceeding the soft limit results in a SIGXCPU signal to be posted to the process, while exceeding the hard limit will cause a SIGKILL.
If the thread is still in the TDS_RUNNING state, mi_switch will put it back onto the run queue, assuming that it will want to run again soon. If it is in one of the other states and KSE threading is enabled, the associated KSE will be made available to any higher priority threads from the same group, to allow them to be scheduled next.
After these administrative tasks are done, mi_switch hands over control to the machine dependent routine cpu_switch, which will perform the actual thread context switch.
cpu_switch first saves the context of the current thread. Next, it calls choosethread to determine which thread to run next. Finally, it reads in the saved context of the new thread and starts to execute the new thread.
cpu_throw is similar to cpu_switch except that it does not save the context of the old thread. This function is useful when the kernel does not have an old thread context to save, such as when CPUs other than the boot CPU perform their first task switch, or when the kernel does not care about the state of the old thread, such as in thread_exit when the kernel terminates the current thread and switches into a new thread.
To protect the runqueue(9), all of these functions must be called with the sched_lock mutex held.
issignal(9), mutex(9), runqueue(9), tsleep(9), wakeup(9)