On Fri, Dec 22, 2017 at 7:44 AM, Yury Norov <ynorov@xxxxxxxxxxxxxxxxxx> wrote:
> On Fri, Dec 22, 2017 at 02:26:32AM -0500, Shih-Wei Li wrote:
>> On Fri, Dec 22, 2017 at 1:07 AM, Yury Norov <ynorov@xxxxxxxxxxxxxxxxxx> wrote:
>> > On Tue, Dec 19, 2017 at 08:34:45PM -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 were used before and after taking
>> >> timestamps to avoid out-of-order execution or pipelining from
>> >> skewing our measurements.
>> >
>> > Instruction barriers are used only before, right? Also, I think it
>> > should be the comment to function rather than part of patch
>> > description.
>>
>> Yes. Sorry about the confusion in the texts here.
>>
>> >
>> >> The operations we currently supported and measured are mostly
>> >> self-explanatory by their function name. For IPI, we measured the
>> >> cost of sending IPI from a source VCPU to a target VCPU, until the
>> >> target VCPU receives the IPI.
>> >
>> > Please describe changes since v1.
>>
>> changes here include simplifying initialization, reducing the
>> iterations in loop_test, removing noop_test, fixing coding styles
>> issues and other nits.
>>
>> >
>> >> Signed-off-by: Shih-Wei Li <shihwei@xxxxxxxxxxxxxxx>
>> >> ---
>> >> arm/Makefile.common | 1 +
>> >> arm/micro-test.c | 219 ++++++++++++++++++++++++++++++++++++++++++++++++++++
>> >> arm/unittests.cfg | 7 ++
>> >> 3 files changed, 227 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..c31c9ac
>> >> --- /dev/null
>> >> +++ b/arm/micro-test.c
>> >> @@ -0,0 +1,219 @@
>> >> +/*
>> >> + * Measure the cost of micro level operations.
>> >> + *
>> >> + * Copyright Columbia University
>> >> + * Author: Shih-Wei Li <shihwei@xxxxxxxxxxxxxxx>
>> >> + * Author: Christoffer Dall <cdall@xxxxxxxxxxxxxxx>
>> >
>> > Nit. If Christoffer is author, he should also sing-off the patch.
>> >
>> >> + *
>> >> + * This work is licensed under the terms of the GNU LGPL, version 2.
>> >> + */
>> >> +#include "libcflat.h"
>> >> +#include <util.h>
>> >> +#include <asm/gic.h>
>> >> +
>> >> +static volatile bool ipi_received;
>> >> +static volatile bool ipi_ready;
>> >> +#define IPI_IRQ 1
>> >> +
>> >> +#define TIMEOUT (1U << 28)
>> >> +
>> >> +/*
>> >> + * The counter may not always start with zero, which means it could
>> >> + * overflow after some time.
>> >> + */
>> >
>> > Is this comment relevant? In discussion people say that they have always
>> > c1 == c2 because timer cannot tick even once, not because it ticks
>> > 2^64 times. And this is what really occurs. Also, if ci == c2,
>> > (c2 - c1) is already zero, and you don't need do something special to
>> > handle it.
>>
>> The comment here is to address your question before whether this is
>> for overflow protection. And yes, we don't need the "((c1) == (c2))"
>> condition here, thanks for poiting out!
>
> And another thing that comes in mind. If individual time measurements
> will be mostly 0, 1 and 2, it is is very suspicious.
>
> What I mean is that tick of system clock is the fastest process on the
> chip, and delivering IPI is complicated procedure, especially in
> virtualized system. In my IPI test for Linux the typical time to
> deliver IPI is 0.3-0.4 us, which gives almmost 1K ticks at 2+ GHZ.
> Did you check your prescalers and so on?
>
>> >> +#define COUNT(c1, c2) \
>> >> + (((c1) > (c2) || ((c1) == (c2))) ? 0 : (c2) - (c1))
>> >> +
>> >> +static uint64_t read_cc(void)
>> >> +{
>> >> + isb();
>> >> + return read_sysreg(cntpct_el0);
>> >> +}
>> >> +
>> >> +static void ipi_irq_handler(struct pt_regs *regs __unused)
>> >> +{
>> >> + u32 ack;
>> >> + ipi_ready = false;
>> >> + ipi_received = true;
>> >> + ack = gic_read_iar();
>> >> + gic_write_eoir(ack);
>> >> + ipi_ready = true;
>> >> +}
>> >> +
>> >> +static void ipi_test_secondary_entry(void *data __unused)
>> >> +{
>> >> + enum vector v = EL1H_IRQ;
>> >> + install_irq_handler(v, ipi_irq_handler);
>> >> +
>> >> + gic_enable_defaults();
>> >> +
>> >> + local_irq_enable(); /* Enter small wait-loop */
>> >> + ipi_ready = true;
>> >> + while (true);
>> >
>> > I remember, it was asked in v1, but anyway, is it intentional here and
>> > in other similar busy-loops not to use wfe/wfi or cpu_relax()?
>> >
>> > Andrew Jones asked you:
>> >
>> >>> cpu_relax() here will help get that first-cpu ack on TCG faster. That
>> >>> said, this test doesn't make sense on TCG anyway, other than debugging
>> >>> it. So you could just add 'accel = kvm' to it's unittests.cfg parameter
>> >>> list.
>> >>
>> >>ok, thanks.
>> >
>> > Or I misunderstood something?
>>
>> I added accel = kvm in unittest.cfg so the test can never run on TCG.
>
> Ah, OK.
>
>> >> +}
>> >> +
>> >> +static int test_init(void)
>> >> +{
>> >> + if (!gic_init()) {
>> >> + printf("No supported gic present, skipping tests...\n");
>> >> + return 0;
>> >> + }
>> >> +
>> >> + ipi_ready = false;
>> >> +
>> >> + gic_enable_defaults();
>> >> + on_cpu_async(1, ipi_test_secondary_entry, 0);
>> >> +
>> >> + printf("Timer Frequency %d Hz (Output in timer count)\n", get_cntfrq());
>> >> +
>> >> + return 1;
>> >> +}
>> >> +
>> >> +static unsigned long ipi_test(void)
>> >> +{
>> >> + unsigned int timeout = TIMEOUT;
>> >> + uint64_t c1, c2;
>> >> +
>> >> + while (!ipi_ready && timeout--);
>> >
>> > Again, this is not timeout. It's nothing about time.
>> > If you need limit the loop by time, it would probably look
>> > like this:
>> >
>> > while (!ipi_ready && read_cc() - start_cc < timeout)
>> > cpu_relax();
>> >
>> > This snippet is repeated too much times in the test to make it a
>> > function, like wait(addr, value).
>>
>> Sorry for the confusion due to bad naming here, Alex Jones mentioned
>> it's probably better to rename timeout here as tries (because there
>> are no timing-related checks here). Will fix it in v3.
>
> I think this is not naming issue. It's OK if you try say turn on
> peripheral device for 10 times, and then report something like
> device not acceptable. Here you make 2^28 attempts and name it
> timeout. What I think is that if you need timeouts, the most
> straightforward option for you is to use timeouts.
So the problem with using read_cc is, we then will measure the time
included from isb() in read_cc(). I think it's ok to use just a simple
counter here because the purpose of the counter is to exit gracefully
instead of spinning in the loop, once the test fails to complete
within some period of time.
>
>> >> + assert(ipi_ready);
>> >> +
>> >> + ipi_received = false;
>> >> +
>> >> + c1 = read_cc();
>> >> +
>> >> + gic_ipi_send_single(IPI_IRQ, 1);
>> >> +
>> >> + timeout = TIMEOUT;
>> >> + while (!ipi_received && timeout--);
>> >> + assert(ipi_received);
>> >> +
>> >> + c2 = read_cc();
>> >> + return COUNT(c1, c2);
>> >> +}
>> >> +
>> >> +
>> >
>> > Odd line is still here.
>> >
>>
>> Yes, will fix it.
>>
>> >> +static unsigned long hvc_test(void)
>> >> +{
>> >> + uint64_t c1, c2;
>> >> +
>> >> + c1 = read_cc();
>> >> + asm volatile("mov w0, #0x4b000000; hvc #0" ::: "w0");
>> >> + c2 = read_cc();
>> >> + return COUNT(c1, c2);
>> >> +}
>> >> +
>> >> +static unsigned long mmio_read_user(void)
>> >> +{
>> >> + uint64_t c1, c2;
>> >> + /*
>> >> + * FIXME: Read device-id in virtio mmio here. This address
>> >> + * needs to be updated in the future if any relavent
>> >
>> > Typo: relevant
>> >
>> >> + * changes in QEMU test-dev are made.
>> >> + */
>> >> + void *mmio_read_user_addr = (void*) 0x0a000008;
>> >
>> > It sounds like function name. Or you mean mmio read-only address?
>> > If no, I'd rename it to mmio_user_addr
>>
>> I named it as it is because this address is used for the following
>> mmio-read operation below. I think the address is also read-only
>> though.
>
> For me the most suitable name would be mmio_device_id.
>
>> >> +
>> >> + c1 = read_cc();
>> >> + readl(mmio_read_user_addr);
>> >> + c2 = read_cc();
>> >> + return COUNT(c1, c2);
>> >> +}
>> >> +
>> >> +static unsigned long mmio_read_vgic(void)
>> >> +{
>> >> + uint64_t c1, c2;
>> >> + int v = gic_version();
>> >
>> > I think gic version cannot be changed while test is running,
>> > so you can check it once on init and store at global variable.
>> >
>> >> + void *vgic_dist_addr = NULL;
>> >
>> > Useless initialization
>> >
>> >> +
>> >> + if (v == 2)
>> >> + vgic_dist_addr = gicv2_dist_base();
>> >> + else if (v == 3)
>> >> + vgic_dist_addr = gicv3_dist_base();
>> >
>> > Again, if there's real chance that gic version will appear not 2 or 3,
>> > you'd take care of it. If no, make else-branch unconditional, like say
>> > in ipi_clear_active_handler().
>>
>> The test will not run if the gic version is not 2 or 3 as it'll
>> trigger the assert in main().
>>
>> >
>> >> + c1 = read_cc();
>> >> + readl(vgic_dist_addr + GICD_IIDR);
>> >
>> > Normally here should be smp_rmb(), but it seems there's no writers at
>> > this address. But if you test mmio performance, you'd take into account
>> > the time for synchronization. Is it correct?
>>
>> I think we had an isb in read_cc before getting the counter so to make
>> sure we read the counter after the MMIO read. Does it address your
>> concern?
>
> Flushing pipeline and I-cache which is isb, and synchronizing D-caches
> are different stories.
>
>> >> + c2 = read_cc();
>> >> + return COUNT(c1, c2);
>> >> +}
>> >> +
>> >> +static unsigned long eoi_test(void)
>> >> +{
>> >> + uint64_t c1, c2;
>> >> + int v = gic_version();
>> >> + void (*write_eoir)(u32 irqstat) = NULL;
>> >
>> > Again, useless initialization. If you make else-branch below
>> > unconditional, you wouldn't need to shut the compiler up.
>> >
>> >> +
>> >> + u32 val = 1023; /* spurious IDs, writes to EOI are ignored */
>> >> +
>> >> + /* To avoid measuring assert(..) in gic_write_eoir */
>> >
>> > Sorry, I don't understand what this comment means.
>> > To avoid measuring assert(..) in gic_write_eoir, let's segfault here
>> > by executing at NULL address - like this?
>>
>> The exiting "gic_write_eoir" function includes an assert(..), which
>> results to some bias in the cost in EOI, which it's pure cost is
>> already rather small. So the code here calls to the gicvx related
>> function directly to avoid the assert().
>>
>> >
>> >> + if (v == 2)
>> >> + write_eoir = gicv2_write_eoir;
>> >> + else if (v == 3)
>> >> + write_eoir = gicv3_write_eoir;
>> >
>> > IIUC, you can do it once at init, right?
>>
>> Yes, I can. I guess either way should work.
>
> But in current version you do it billion times in hot path which
> probably affects final results.
thanks, I have moved the initialization piece earlier in v3.
>
>> >> +
>> >> + c1 = read_cc();
>> >> + write_eoir(val);
>> >> + c2 = read_cc();
>> >> +
>> >> + return COUNT(c1, c2);
>> >> +}
>> >> +
>> >> +struct exit_test {
>> >> + const char *name;
>> >> + unsigned long (*test_fn)(void);
>> >> + bool run;
>> >> +};
>> >> +
>> >> +static struct exit_test tests[] = {
>> >> + {"hvc", hvc_test, true},
>> >> + {"mmio_read_user", mmio_read_user, true},
>> >> + {"mmio_read_vgic", mmio_read_vgic, true},
>> >> + {"eoi", eoi_test, true},
>> >> + {"ipi", ipi_test, true},
>> >> +};
>> >> +
>> >> +static void loop_test(struct exit_test *test)
>> >> +{
>> >> + unsigned long i, iterations = 32;
>> >> + unsigned long sample, cycles;
>> >> + unsigned long long min = 0, max = 0;
>> >> + const unsigned long long goal = (1ULL << 25);
>> >> +
>> >> + do {
>> >> + iterations *= 2;
>> >> + cycles = 0;
>> >> + i = 0;
>> >> + while (i < iterations) {
>> >> + sample = test->test_fn();
>> >> + if (sample == 0) {
>> >> + /*
>> >> + * If something went wrong or we had an
>> >> + * overflow, don't count that sample.
>> >> + */
>> >> + printf("cycle count overflow: %lu\n", sample);
>> >
>> > Text doesn't match the comment.
>>
>> I guess you meant that we should say something like the cost is
>> smaller than a timer count?
>
> Yep
>
>> > Sample doesn't contain error code and is always zero. what's the point
>> > to print it?
>> >
>> >> + continue;
>> >> + }
>> >> + cycles += sample;
>> >> + if (min == 0 || min > sample)
>> >
>> > Hint: initialize min as ULONGLONG_MAX, and drop this (min == 0) check.
>> >
>> >> + min = sample;
>> >> + if (max < sample)
>> >> + max = sample;
>> >> + ++i;
>> >> + }
>> >> + } while (cycles < goal);
>> >> + printf("%s:\t avg %lu\t min %llu\t max %llu\n",
>> >> + test->name, cycles / iterations, min, max);
>> >> +}
>> >> +
>> >> +int main(int argc, char **argv)
>> >> +{
>> >> + int i;
>> >> +
>> >> + assert(test_init());
>> >> +
>> >> + for (i = 0; i < ARRAY_SIZE(tests); i++) {
>> >> + if (!tests[i].run)
>> >> + continue;
>> >> + loop_test(&tests[i]);
>> >> + }
>> >> +
>> >> + return 0;
>> >> +}
>> >> diff --git a/arm/unittests.cfg b/arm/unittests.cfg
>> >> index 44b98cf..a8fb0b3 100644
>> >> --- a/arm/unittests.cfg
>> >> +++ b/arm/unittests.cfg
>> >> @@ -116,3 +116,10 @@ file = timer.flat
>> >> groups = timer
>> >> timeout = 2s
>> >> arch = arm64
>> >> +
>> >> +# Exit tests
>> >> +[micro-test]
>> >> +file = micro-test.flat
>> >> +smp = 2
>
> Sidenote. It would be also interesting to benchmark broadcast IPIs with
> smp >> 2.
>
>> >> +groups = micro-test
>> >> +accel = kvm
>> >> --
>> >> 1.9.1
>> >>
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
