tree b454e77c7442eefe101bb946394259b0e6697182
parent 66a9d366feaf88927f039b4e949ea85d5a92e7ec
author Barret Rhoden <brho@cs.berkeley.edu> 1576007572 -0500
committer Barret Rhoden <brho@cs.berkeley.edu> 1583266410 -0500

Add a benchmark for TLB flushes and shootdowns

To use, make sure you set the C-state to 0, so processors don't sleep
deeply:

	$ echo 0 > \#arch/c-state

Then issue the various commands.  To run all 7 tests, between core 0
(where the shell runs) and core 7:

	$ for i in `seq 1 7`; do echo tlb 7 $i > \#regress/monctl; done

To run the POKE test, you'll need to manually change the ASM handler.

One note on concurrency: this assumes that send_ipi() and
send_kernel_message() provide the appropriate memory barriers.
Specifically, that the main thread's write to __tlb_bench_x appears
before the message is sent (or at least that the handler's clobber
happens after).  Safe to say that's a guarantee of those functions, at
least conceptually and in their x86 implementations.  (wrmsr and LOCK
xchg).

Also, using both READ_ONCE() and cpu_relax() is overkill.  The newer
Linux style would be to use READ_ONCE(), I think.

Signed-off-by: Barret Rhoden <brho@cs.berkeley.edu>