Oliver Upton <oliver.upton@xxxxxxxxx> writes:
These functions were originally part of my patch to introduce latency
measurements into dirty_log_perf_test. [1] Sean Christopherson
suggested lifting these functions into their own patch in generic code
so they can be used by any test. [2] Ricardo Koller suggested the
addition of the MEASURE macro to more easily time individual
statements. [3]
[1]
https://lore.kernel.org/kvm/20221115173258.2530923-1-coltonlewis@xxxxxxxxxx/
[2] https://lore.kernel.org/kvm/Y8gfOP5CMXK60AtH@xxxxxxxxxx/
[3] https://lore.kernel.org/kvm/Y8cIdxp5k8HivVAe@xxxxxxxxxx/
This patch doesn't make a great deal of sense outside of [1]. Can you
send this as part of your larger series next time around?
I copied the wrong email link where Sean suggested this should be
generic code in a separate patch. I may have been mistaken in thinking
he meant to upstream it separately.
https://lore.kernel.org/kvm/Y8gjG6gG5UR6T3Yg@xxxxxxxxxx/
But yes, I would prefer to keep this as part of the larger series.
#include <linux/mman.h>
#include "linux/kernel.h"
+#if defined(__aarch64__)
+#include "aarch64/arch_timer.h"
+#elif defined(__x86_64__)
+#include "x86_64/processor.h"
+#endif
I believe we place 'include/$(ARCH)/' on the include path, so the
'x86_64' can be dropped.
Good to know.
#include "test_util.h"
/*
@@ -34,6 +39,38 @@ uint32_t guest_random_u32(struct guest_random_state
*state)
return state->seed;
}
+uint64_t guest_cycles_read(void)
Do we need the 'guest_' prefix in here? At least for x86 and arm64 the
same instructions can be used in host userspace and guest kernel for
accessing the counter.
It's intended for guest space, but I take your point it isn't necessary.
+{
+#if defined(__aarch64__)
+ return timer_get_cntct(VIRTUAL);
+#elif defined(__x86_64__)
+ return rdtsc();
+#else
+#warn __func__ " is not implemented for this architecture, will return
0"
+ return 0;
+#endif
+}
I think it would be cleaner if each architecture just provided its own
implementation of this function instead of ifdeffery here.
If an arch doesn't implement the requisite helper then it will become
painfully obvious at compile time.
Good point. That will be more clear.
+double guest_cycles_to_ns(double cycles)
+{
+#if defined(__aarch64__)
+ return cycles * (1e9 / timer_get_cntfrq());
+#elif defined(__x86_64__)
+ return cycles * (1e9 / (KVM_GET_TSC_KHZ * 1000));
The x86 implementation is wrong. KVM_GET_TSC_KHZ is the ioctl needed
to get at the guest's TSC frequency (from the perpsective of host
userspace).
Oops. Absent minded on that one. Forgot it wasn't just a value I could use.
So at the very least this needs to do an ioctl on the VM fd to work out
the frequency. This is expected to be called in host userspace, right?
That's the plan yes.
diff --git a/tools/testing/selftests/kvm/system_counter_offset_test.c
b/tools/testing/selftests/kvm/system_counter_offset_test.c
index 7f5b330b6a1b..39b1249c7404 100644
--- a/tools/testing/selftests/kvm/system_counter_offset_test.c
+++ b/tools/testing/selftests/kvm/system_counter_offset_test.c
@@ -39,14 +39,9 @@ static void setup_system_counter(struct kvm_vcpu
*vcpu, struct test_case *test)
&test->tsc_offset);
}
-static uint64_t guest_read_system_counter(struct test_case *test)
-{
- return rdtsc();
-}
-
static uint64_t host_read_guest_system_counter(struct test_case *test)
{
- return rdtsc() + test->tsc_offset;
+ return guest_cycles_read() + test->tsc_offset;
The 'guest_' part is rather confusing here. What this function is really
trying to do is read the counter from host userspace and apply an offset
to construct the expected value for the test.
It all compiles down to the same thing, but as written is seemingly
wrong.
I'll eliminate the guest_prefix.
