On Tue, Dec 19, 2017 at 01:05:21PM +0300, Yury Norov wrote: > On Tue, Dec 19, 2017 at 10:12:00AM +0100, Christoffer Dall wrote: > > On Mon, Dec 18, 2017 at 08:31:21PM +0300, Yury Norov wrote: > > > On Fri, Dec 15, 2017 at 04:15:39PM -0500, Shih-Wei Li wrote: > > > > Here we provide the support for measuring various micro level > > > > operations on arm64. We iterate each of the tests for millions of > > > > times and output their average, minimum and maximum cost in timer > > > > counts. Instruction barriers are used before and after taking > > > > timestamps to avoid out-of-order execution or pipelining from > > > > skewing our measurements. > > > > > > > > The tests we currently support and measure are mostly > > > > straightforward by the function names and the respective comments. > > > > For IPI test, we measure the cost of sending IPI from a source > > > > VCPU to a target VCPU, until the target VCPU receives the IPI. > > > > > > > > Signed-off-by: Shih-Wei Li <shihwei@xxxxxxxxxxxxxxx> > > > > --- > > > > arm/Makefile.common | 1 + > > > > arm/micro-test.c | 289 ++++++++++++++++++++++++++++++++++++++++++++++++++++ > > > > arm/unittests.cfg | 6 ++ > > > > 3 files changed, 296 insertions(+) > > > > create mode 100644 arm/micro-test.c > > > > > > > > diff --git a/arm/Makefile.common b/arm/Makefile.common > > > > index 0a039cf..c7d5c27 100644 > > > > --- a/arm/Makefile.common > > > > +++ b/arm/Makefile.common > > > > @@ -16,6 +16,7 @@ tests-common += $(TEST_DIR)/pmu.flat > > > > tests-common += $(TEST_DIR)/gic.flat > > > > tests-common += $(TEST_DIR)/psci.flat > > > > tests-common += $(TEST_DIR)/sieve.flat > > > > +tests-common += $(TEST_DIR)/micro-test.flat > > > > > > > > tests-all = $(tests-common) $(tests) > > > > all: directories $(tests-all) > > > > diff --git a/arm/micro-test.c b/arm/micro-test.c > > > > new file mode 100644 > > > > index 0000000..7df2272 > > > > --- /dev/null > > > > +++ b/arm/micro-test.c > > > > @@ -0,0 +1,289 @@ > > > > +#include <util.h> > > > > +#include <asm/gic.h> > > > > + > > > > +static volatile bool second_cpu_up; > > > > +static volatile bool first_cpu_ack; > > > > +static volatile bool ipi_acked; > > > > +static volatile bool ipi_received; > > > > +static volatile bool ipi_ready; > > > > +#define IPI_IRQ 1 > > > > + > > > > +#define TIMEOUT (1U << 28) > > > > + > > > > +#define ARR_SIZE(_x) ((int)(sizeof(_x) / sizeof(_x[0]))) > > > > +#define for_each_test(_iter, _tests, _tmp) \ > > > > + for (_tmp = 0, _iter = _tests; \ > > > > + _tmp < ARR_SIZE(_tests); \ > > > > + _tmp++, _iter++) > > > > + > > > > +#define CYCLE_COUNT(c1, c2) \ > > > > + (((c1) > (c2) || ((c1) == (c2))) ? 0 : (c2) - (c1)) > > > > > > Is my understanding correct that this is overflow protection? > > > c1 and c2 are 64-bit values. To overflow them you need 58 years > > > at 1G CPU freq. > > > > > > > That's assuming your cycle counter starts at 0, and that nobody > > programmed it near the overflow value to get an overflow interrupt. > > > > So if you get rid of this, you have to make sure the host never plays > > with the cycle counter behind your back, and that you've initialized it > > to zero. > > Ah, now I understand it. But if there's no evil intention to break the > test, there's single chance to overflow in 584 years. And if it happened, > isn't it simpler to run test again? I believe we used this to throw away results where there was an overflow, which is nicer to the user than leaving them guessing 'this is a very high number, I better run the test again'. -Christoffer
