Re: [PATCH v4 4/5] KVM: selftests: aarch64: Introduce pmu_event_filter_test

Hi Oliver,

On 2/2/24 16:34, Oliver Upton wrote:

On Thu, Feb 01, 2024 at 09:56:53PM -0500, Shaoqin Huang wrote:

...

Introduce pmu_event_filter_test for arm64 platforms. The test configures PMUv3 for a vCPU, and sets different pmu event filters for the vCPU, and check if the guest can see those events which user allow and can't use those events which use deny.

This test refactor the create_vpmu_vm() and make it a wrapper for __create_vpmu_vm(), which allows some extra init code before KVM_ARM_VCPU_PMU_V3_INIT.

And this test use the KVM_ARM_VCPU_PMU_V3_FILTER attribute to set the pmu event filter in KVM. And choose to filter two common event branches_retired and instructions_retired, and let the guest to check if it see the right pmceid register.

Signed-off-by: Shaoqin Huang shahuang@redhat.com

tools/testing/selftests/kvm/Makefile | 1 + .../kvm/aarch64/pmu_event_filter_test.c | 219 ++++++++++++++++++ .../selftests/kvm/include/aarch64/vpmu.h | 4 + .../testing/selftests/kvm/lib/aarch64/vpmu.c | 14 +- 4 files changed, 236 insertions(+), 2 deletions(-) create mode 100644 tools/testing/selftests/kvm/aarch64/pmu_event_filter_test.c

diff --git a/tools/testing/selftests/kvm/Makefile b/tools/testing/selftests/kvm/Makefile index 709a70b31ca2..733ec86a3385 100644 --- a/tools/testing/selftests/kvm/Makefile +++ b/tools/testing/selftests/kvm/Makefile @@ -148,6 +148,7 @@ TEST_GEN_PROGS_aarch64 += aarch64/arch_timer TEST_GEN_PROGS_aarch64 += aarch64/debug-exceptions TEST_GEN_PROGS_aarch64 += aarch64/hypercalls TEST_GEN_PROGS_aarch64 += aarch64/page_fault_test +TEST_GEN_PROGS_aarch64 += aarch64/pmu_event_filter_test TEST_GEN_PROGS_aarch64 += aarch64/psci_test TEST_GEN_PROGS_aarch64 += aarch64/set_id_regs TEST_GEN_PROGS_aarch64 += aarch64/smccc_filter diff --git a/tools/testing/selftests/kvm/aarch64/pmu_event_filter_test.c b/tools/testing/selftests/kvm/aarch64/pmu_event_filter_test.c new file mode 100644 index 000000000000..d280382f362f --- /dev/null +++ b/tools/testing/selftests/kvm/aarch64/pmu_event_filter_test.c @@ -0,0 +1,219 @@ +// SPDX-License-Identifier: GPL-2.0 +/*

• • pmu_event_filter_test - Test user limit pmu event for guest.
• • Copyright (c) 2023 Red Hat, Inc.
• • This test checks if the guest only see the limited pmu event that userspace
• • sets, if the guest can use those events which user allow, and if the guest
• • can't use those events which user deny.
• • This test runs only when KVM_CAP_ARM_PMU_V3, KVM_ARM_VCPU_PMU_V3_FILTER
• • is supported on the host.
• */

+#include <kvm_util.h> +#include <processor.h> +#include <vgic.h> +#include <vpmu.h> +#include <test_util.h> +#include <perf/arm_pmuv3.h>

+struct pmce{

Missing whitespace before curly brace.

Also -- pmce is an odd name. Maybe common_event_ids or pmu_id_regs.

Thanks for pointing this out. I would choose pmu_common_event_ids as its name.

...

• uint64_t pmceid0;
• uint64_t pmceid1;

+} supported_pmce, guest_pmce;

maybe "max_*" and "expected_*". It took me a bit to understand that guest_pmce feeds in your expected PMCEID values.

The "max_* and "expected_*" is more clear, I would use it.

...

+static struct vpmu_vm *vpmu_vm;

+#define FILTER_NR 10

+struct test_desc {

• const char *name;
• struct kvm_pmu_event_filter filter[FILTER_NR];

+};

+#define __DEFINE_FILTER(base, num, act) \

• ((struct kvm_pmu_event_filter) { \

• .base_event	= base,		\
  

• .nevents	= num,		\
  

• .action		= act,		\
  

• })

+#define DEFINE_FILTER(base, act) __DEFINE_FILTER(base, 1, act)

+#define EMPTY_FILTER { 0 }

+#define SW_INCR 0x0 +#define INST_RETIRED 0x8 +#define BR_RETIRED 0x21

These event numbers are already defined in tools/include/perf/arm_pmuv3.h, use those instead.

Sure. I would use those defined macro.

...

+static void guest_code(void) +{

• uint64_t pmceid0 = read_sysreg(pmceid0_el0);
• uint64_t pmceid1 = read_sysreg(pmceid1_el0);
• GUEST_ASSERT_EQ(guest_pmce.pmceid0, pmceid0);
• GUEST_ASSERT_EQ(guest_pmce.pmceid1, pmceid1);
• GUEST_DONE();

+}

+static void guest_get_pmceid(void) +{

• supported_pmce.pmceid0 = read_sysreg(pmceid0_el0);
• supported_pmce.pmceid1 = read_sysreg(pmceid1_el0);
• GUEST_DONE();

+}

+static void pmu_event_filter_init(struct vpmu_vm *vm, void *arg)

Why are you obfuscating the pointer to the filter array?

...

+{

• struct kvm_device_attr attr = {

• .group	= KVM_ARM_VCPU_PMU_V3_CTRL,
  

• .attr	= KVM_ARM_VCPU_PMU_V3_FILTER,
  

• };
• struct kvm_pmu_event_filter *filter = (struct kvm_pmu_event_filter *)arg;
• while (filter && filter->nevents != 0) {

• attr.addr = (uint64_t)filter;
  

• vcpu_ioctl(vm->vcpu, KVM_SET_DEVICE_ATTR, &attr);
  

Again, kvm_device_attr_set() the right helper to use.

...

+static void set_pmce(struct pmce *pmce, int action, int event) +{

• int base = 0;
• uint64_t *pmceid = NULL;
• if (event >= 0x4000) {

• event -= 0x4000;
  

• base = 32;
  

• }
• if (event >= 0 && event <= 0x1F) {

• pmceid = &pmce->pmceid0;
  

• } else if (event >= 0x20 && event <= 0x3F) {

• event -= 0x20;
  

• pmceid = &pmce->pmceid1;
  

• } else {

• return;
  

• }
• event += base;
• if (action == KVM_PMU_EVENT_ALLOW)

• *pmceid |= BIT(event);
  

• else

• *pmceid &= ~BIT(event);
  

+}

+static void prepare_guest_pmce(struct kvm_pmu_event_filter *filter) +{

• struct pmce pmce_mask = { ~0, ~0 };
• bool first_filter = true;
• while (filter && filter->nevents != 0) {

• if (first_filter) {
  

• 	if (filter->action == KVM_PMU_EVENT_ALLOW)
  

• 		memset(&pmce_mask, 0, sizeof(pmce_mask));
  

• 	first_filter = false;
  

• }
  

• set_pmce(&pmce_mask, filter->action, filter->base_event);
  

• filter++;
  

• }
• guest_pmce.pmceid0 = supported_pmce.pmceid0 & pmce_mask.pmceid0;
• guest_pmce.pmceid1 = supported_pmce.pmceid1 & pmce_mask.pmceid1;

+}

Why do you need to do this? Can't you tell the guests what events to expect and have it make sense of the PMCEID values it sees?

At here, I prepare the pmceid value which the guest should see, and pass it to the guest by sync the global variable. And guest compare this value with the value it read through pmu common event register.

Why I don't put the process of generating expected_pmce into the guest code is that I want to make sure this value computed in host is totally correct, so the guest code is pretty simple, it only needs to compare the two value.

You could, for example, pass in a pointer to the test descriptor as an argument.

...

+static void run_test(struct test_desc *t) +{

• pr_debug("Test: %s\n", t->name);

You may as well just pr_info() this thing.

Ok. I can change it to pr_info().

...

• create_vpmu_vm_with_filter(guest_code, t->filter);
• prepare_guest_pmce(t->filter);
• sync_global_to_guest(vpmu_vm->vm, guest_pmce);
• run_vcpu(vpmu_vm->vcpu);
• destroy_vpmu_vm(vpmu_vm);

+}

+static struct test_desc tests[] = {

• {"without_filter", { EMPTY_FILTER }},
• {"member_allow_filter",

• {DEFINE_FILTER(SW_INCR, 0), DEFINE_FILTER(INST_RETIRED, 0),
  

•  DEFINE_FILTER(BR_RETIRED, 0), EMPTY_FILTER}},
  

• {"member_deny_filter",

• {DEFINE_FILTER(SW_INCR, 1), DEFINE_FILTER(INST_RETIRED, 1),
  

•  DEFINE_FILTER(BR_RETIRED, 1), EMPTY_FILTER}},
  

• {"not_member_deny_filter",

• {DEFINE_FILTER(SW_INCR, 1), EMPTY_FILTER}},
  

• {"not_member_allow_filter",

• {DEFINE_FILTER(SW_INCR, 0), EMPTY_FILTER}},
  

Why is the filter array special enough to get its own sentinel macro but...

...

• { 0 }

... the test descriptor array is okay to use a 'raw' initialization. My vote is to drop the macro, zero-initializing a struct in an array is an extremely common pattern in the kernel.

Also, these descriptors are dense and hard to read. Working with an example:

{ .name = "member_allow_filter", .filter = { DEFINE_FILTER(SW_INCR, 0), DEFINE_FILTER(INST_RETIRED, 0), DEFINE_FILTER(BR_RETIRED, 0), { 0 } }, }

See how much more readable that is?

It's more clear and readable, thanks for your advice. I will change the array definition to the beautiful format.

...

+};

+static void for_each_test(void) +{

• struct test_desc *t;
• for (t = &tests[0]; t->name; t++)

• run_test(t);
  

+}

for_each_test() sounds like an iterator, but this is not. Call it run_tests()

Ok. Will call it run_tests().

...

+static bool kvm_supports_pmu_event_filter(void) +{

• int r;
• vpmu_vm = create_vpmu_vm(guest_code);
• r = __kvm_has_device_attr(vpmu_vm->vcpu->fd, KVM_ARM_VCPU_PMU_V3_CTRL,

• 		  KVM_ARM_VCPU_PMU_V3_FILTER);
  

• destroy_vpmu_vm(vpmu_vm);
• return !r;

+}

TBH, I don't really care much about the test probing for the event filter UAPI. It has been upstream for a while, and if folks are trying to run selftests at HEAD on an old kernel then that's their business.

The other prerequisites make more sense since they actually check if HW features are present.

If no one cares it, I will delete this function.

...

+static bool host_pmu_supports_events(void) +{

• vpmu_vm = create_vpmu_vm(guest_get_pmceid);
• memset(&supported_pmce, 0, sizeof(supported_pmce));
• sync_global_to_guest(vpmu_vm->vm, supported_pmce);
• run_vcpu(vpmu_vm->vcpu);
• sync_global_from_guest(vpmu_vm->vm, supported_pmce);
• destroy_vpmu_vm(vpmu_vm);
• return supported_pmce.pmceid0 & (BR_RETIRED | INST_RETIRED);

+}

This helper says its probing the host PMU, but you're actually firing up a VM to do it.

The events supported by a particular PMU instance are readily available in sysfs. Furthermore, you can tell KVM to select the exact host PMU instance you probe.

It should call kvm_pmu_support_events, because I want to get the default pmce without any filter in the kvm. So I run a guest and get that value in the guest.

I've tried get the pmceid0 through the

vcpu_get_reg(vcpu, KVM_ARM64_SYS_REG(SYS_PMCEID0_EL0), &val);

But it always return the -ENOENT, I'm not sure if this is expected. Could I can use the KVM interface to get the pmceid0?

...

diff --git a/tools/testing/selftests/kvm/lib/aarch64/vpmu.c b/tools/testing/selftests/kvm/lib/aarch64/vpmu.c index b3de8fdc555e..76ea03d607f1 100644 --- a/tools/testing/selftests/kvm/lib/aarch64/vpmu.c +++ b/tools/testing/selftests/kvm/lib/aarch64/vpmu.c @@ -7,8 +7,9 @@ #include <vpmu.h> #include <perf/arm_pmuv3.h> -/* Create a VM that has one vCPU with PMUv3 configured. */ -struct vpmu_vm *create_vpmu_vm(void *guest_code) +struct vpmu_vm *__create_vpmu_vm(void *guest_code,

• 		 void (*init_pmu)(struct vpmu_vm *vm, void *arg),
  

• 		 void *arg)
  

  { struct kvm_vcpu_init init; uint8_t pmuver;

@@ -50,12 +51,21 @@ struct vpmu_vm *create_vpmu_vm(void *guest_code) "Unexpected PMUVER (0x%x) on the vCPU with PMUv3", pmuver); /* Initialize vPMU */

• if (init_pmu)

• init_pmu(vpmu_vm, arg);
  

• vcpu_ioctl(vpmu_vm->vcpu, KVM_SET_DEVICE_ATTR, &irq_attr); vcpu_ioctl(vpmu_vm->vcpu, KVM_SET_DEVICE_ATTR, &init_attr);

return vpmu_vm; } +/* Create a VM that has one vCPU with PMUv3 configured. */ +struct vpmu_vm *create_vpmu_vm(void *guest_code) +{

• return __create_vpmu_vm(guest_code, NULL, NULL);

+}

Ok. This completely proves my point in the other patch. You already need to refactor this helper to cram in what you're trying to do. Think of ways to move the code that is actually common into libraries and leave the rest to the tests themselves.

Some slight code duplication isn't the end of the world if it avoids churning libraries every time someone wants to add a widget.

Thanks for your opinion. I'm thinking about refactor the helper to make it can be reuseable by further tests.

Thanks, Shaoqin