So, operating under the t_locked < 2 * t_ctxsw "rule" might be great in theory, but hang on: are you really expected to precisely measure the context switch time and the time spent in the critical section of each and every case where one (critical section) exists? No, of course not – that's pretty unrealistic and pedantic.
Practically speaking, think about it this way: the mutex lock works by having the loser threads sleep upon the unlock; the spinlock does not (the losers "spin"). Let's recall one of our golden rules of the Linux kernel: a kernel cannot sleep (call schedule()) in any kind of atomic context. Thus, we can never use the mutex lock in an interrupt context, or indeed in any context where it isn't safe to sleep; using the spinlock, however, would be fine. (Remember, a blocking API is one that puts the calling context to sleep by calling schedule...