From: "Khyron" <khyron@xxxxxxxxxx>
> But couldn't lockmetering still indicate which locks are
> being held for how long? Or is there a better way to go about
> it on uniproc kernels? Or are uniproc locks that finely tuned
> that its irrelevant?
But what's the point of how long the lock is held? There isn't going to
be contention. Only the BKL (kernel_flag) can be held across a context
switch, and even for that one the scheduler drops the lock and
reacquires it when that thread gets the CPU back again. And if an
interrupt-level piece of code acquires a lock, that lock better not be
something that's held by a non-interrupt-level piece of code.
I believe the only thing that's interesting is how long interrupts are
disabled.
Feel free to hack uniprocessor spinlocks to invoke the Lockmeter
routines, if you want to experiment with them.
John Hawkes
|