[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [PATCH v7] arm/kvm: Enable support for KVM_ARM_VCPU_PMU_V3_FILTER
From: |
Peter Maydell |
Subject: |
Re: [PATCH v7] arm/kvm: Enable support for KVM_ARM_VCPU_PMU_V3_FILTER |
Date: |
Thu, 22 Feb 2024 14:28:22 +0000 |
On Wed, 21 Feb 2024 at 06:34, Shaoqin Huang <shahuang@redhat.com> wrote:
>
> The KVM_ARM_VCPU_PMU_V3_FILTER provides the ability to let the VMM decide
> which PMU events are provided to the guest. Add a new option
> `kvm-pmu-filter` as -cpu sub-option to set the PMU Event Filtering.
> Without the filter, all PMU events are exposed from host to guest by
> default. The usage of the new sub-option can be found from the updated
> document (docs/system/arm/cpu-features.rst).
>
> Here is an example which shows how to use the PMU Event Filtering, when
> we launch a guest by use kvm, add such command line:
>
> # qemu-system-aarch64 \
> -accel kvm \
> -cpu host,kvm-pmu-filter="D:0x11-0x11"
>
> Since the first action is deny, we have a global allow policy. This
> filters out the cycle counter (event 0x11 being CPU_CYCLES).
>
> And then in guest, use the perf to count the cycle:
>
> # perf stat sleep 1
>
> Performance counter stats for 'sleep 1':
>
> 1.22 msec task-clock # 0.001 CPUs
> utilized
> 1 context-switches # 820.695 /sec
> 0 cpu-migrations # 0.000 /sec
> 55 page-faults # 45.138 K/sec
> <not supported> cycles
> 1128954 instructions
> 227031 branches # 186.323 M/sec
> 8686 branch-misses # 3.83% of all
> branches
>
> 1.002492480 seconds time elapsed
>
> 0.001752000 seconds user
> 0.000000000 seconds sys
>
> As we can see, the cycle counter has been disabled in the guest, but
> other pmu events do still work.
>
> Reviewed-by: Sebastian Ott <sebott@redhat.com>
> Signed-off-by: Shaoqin Huang <shahuang@redhat.com>
> ---
> v6->v7:
> - Check return value of sscanf.
> - Improve the check condition.
>
> v5->v6:
> - Commit message improvement.
> - Remove some unused code.
> - Collect Reviewed-by, thanks Sebastian.
> - Use g_auto(Gstrv) to replace the gchar **. [Eric]
>
> v4->v5:
> - Change the kvm-pmu-filter as a -cpu sub-option. [Eric]
> - Comment tweak. [Gavin]
> - Rebase to the latest branch.
>
> v3->v4:
> - Fix the wrong check for pmu_filter_init. [Sebastian]
> - Fix multiple alignment issue. [Gavin]
> - Report error by warn_report() instead of error_report(), and don't use
> abort() since the PMU Event Filter is an add-on and best-effort feature.
> [Gavin]
> - Add several missing { } for single line of code. [Gavin]
> - Use the g_strsplit() to replace strtok(). [Gavin]
>
> v2->v3:
> - Improve commits message, use kernel doc wording, add more explaination on
> filter example, fix some typo error. [Eric]
> - Add g_free() in kvm_arch_set_pmu_filter() to prevent memory leak. [Eric]
> - Add more precise error message report. [Eric]
> - In options doc, add pmu-filter rely on KVM_ARM_VCPU_PMU_V3_FILTER support
> in
> KVM. [Eric]
>
> v1->v2:
> - Add more description for allow and deny meaning in
> commit message. [Sebastian]
> - Small improvement. [Sebastian]
>
> docs/system/arm/cpu-features.rst | 23 +++++++++
> target/arm/cpu.h | 3 ++
> target/arm/kvm.c | 80 ++++++++++++++++++++++++++++++++
> 3 files changed, 106 insertions(+)
The new syntax for the filter property seems quite complicated.
I think it would be worth testing it with a new test in
tests/qtest/arm-cpu-features.c.
> diff --git a/docs/system/arm/cpu-features.rst
> b/docs/system/arm/cpu-features.rst
> index a5fb929243..7c8f6a60ef 100644
> --- a/docs/system/arm/cpu-features.rst
> +++ b/docs/system/arm/cpu-features.rst
> @@ -204,6 +204,29 @@ the list of KVM VCPU features and their descriptions.
> the guest scheduler behavior and/or be exposed to the guest
> userspace.
>
> +``kvm-pmu-filter``
> + By default kvm-pmu-filter is disabled. This means that by default all pmu
"PMU"
> + events will be exposed to guest.
> +
> + KVM implements PMU Event Filtering to prevent a guest from being able to
> + sample certain events. It depends on the KVM_ARM_VCPU_PMU_V3_FILTER
> + attribute supported in KVM. It has the following format:
> +
> + kvm-pmu-filter="{A,D}:start-end[;{A,D}:start-end...]"
> +
> + The A means "allow" and D means "deny", start is the first event of the
> + range and the end is the last one. The first registered range defines
> + the global policy(global ALLOW if the first @action is DENY, global DENY
Missing space before '('.
Why the '@' before 'action'?
> + if the first @action is ALLOW). The start and end only support hexadecimal
> + format. For example:
> +
> + kvm-pmu-filter="A:0x11-0x11;A:0x23-0x3a;D:0x30-0x30"
> +
> + Since the first action is allow, we have a global deny policy. It
> + will allow event 0x11 (The cycle counter), events 0x23 to 0x3a are
lowercase "the".
> + also allowed except the event 0x30 which is denied, and all the other
> + events are denied.
> +
Did you check that the documentation builds and that this new
documentation renders into HTML the way you want it?
> TCG VCPU Features
> =================
>
> diff --git a/target/arm/cpu.h b/target/arm/cpu.h
> index 63f31e0d98..f7f2431755 100644
> --- a/target/arm/cpu.h
> +++ b/target/arm/cpu.h
> @@ -948,6 +948,9 @@ struct ArchCPU {
>
> /* KVM steal time */
> OnOffAuto kvm_steal_time;
> +
> + /* KVM PMU Filter */
> + char *kvm_pmu_filter;
> #endif /* CONFIG_KVM */
>
> /* Uniprocessor system with MP extensions */
> diff --git a/target/arm/kvm.c b/target/arm/kvm.c
> index 81813030a5..5c62580d34 100644
> --- a/target/arm/kvm.c
> +++ b/target/arm/kvm.c
> @@ -496,6 +496,22 @@ static void kvm_steal_time_set(Object *obj, bool value,
> Error **errp)
> ARM_CPU(obj)->kvm_steal_time = value ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF;
> }
>
> +static char *kvm_pmu_filter_get(Object *obj, Error **errp)
> +{
> + ARMCPU *cpu = ARM_CPU(obj);
> +
> + return g_strdup(cpu->kvm_pmu_filter);
> +}
> +
> +static void kvm_pmu_filter_set(Object *obj, const char *pmu_filter,
> + Error **errp)
> +{
> + ARMCPU *cpu = ARM_CPU(obj);
> +
> + g_free(cpu->kvm_pmu_filter);
> + cpu->kvm_pmu_filter = g_strdup(pmu_filter);
> +}
> +
> /* KVM VCPU properties should be prefixed with "kvm-". */
> void kvm_arm_add_vcpu_properties(ARMCPU *cpu)
> {
> @@ -517,6 +533,12 @@ void kvm_arm_add_vcpu_properties(ARMCPU *cpu)
> kvm_steal_time_set);
> object_property_set_description(obj, "kvm-steal-time",
> "Set off to disable KVM steal time.");
> +
> + object_property_add_str(obj, "kvm-pmu-filter", kvm_pmu_filter_get,
> + kvm_pmu_filter_set);
> + object_property_set_description(obj, "kvm-pmu-filter",
> + "PMU Event Filtering description for "
> + "guest PMU. (default: NULL, disabled)");
> }
>
> bool kvm_arm_pmu_supported(void)
> @@ -1706,6 +1728,62 @@ static bool kvm_arm_set_device_attr(ARMCPU *cpu,
> struct kvm_device_attr *attr,
> return true;
> }
>
> +static void kvm_arm_pmu_filter_init(ARMCPU *cpu)
> +{
> + static bool pmu_filter_init;
> + struct kvm_pmu_event_filter filter;
> + struct kvm_device_attr attr = {
> + .group = KVM_ARM_VCPU_PMU_V3_CTRL,
> + .attr = KVM_ARM_VCPU_PMU_V3_FILTER,
> + .addr = (uint64_t)&filter,
> + };
> + int i;
> + g_auto(GStrv) event_filters;
> +
> + if (!cpu->kvm_pmu_filter) {
> + return;
> + }
> + if (kvm_vcpu_ioctl(CPU(cpu), KVM_HAS_DEVICE_ATTR, &attr)) {
> + warn_report("The KVM doesn't support the PMU Event Filter!");
Drop "The ".
Should this really only be a warning, rather than an error?
> + return;
> + }
> +
> + /*
> + * The filter only needs to be initialized through one vcpu ioctl and it
> + * will affect all other vcpu in the vm.
Weird. Why isn't it a VM ioctl if it affects the whole VM ?
> + */
> + if (pmu_filter_init) {
> + return;
> + } else {
> + pmu_filter_init = true;
> + }
Shouldn't we do this before we do the kvm_vcpu_ioctl check
for whether the kernel supports the filter? Otherwise presumably
we'll print the warning once per vCPU, rather than only once.
> +
> + event_filters = g_strsplit(cpu->kvm_pmu_filter, ";", -1);
> + for (i = 0; event_filters[i]; i++) {
> + unsigned short start = 0, end = 0;
> + char act;
> +
> + if (sscanf(event_filters[i], "%c:%hx-%hx", &act, &start, &end) != 3)
> {
> + warn_report("Skipping invalid PMU filter %s", event_filters[i]);
> + continue;
> + }
> +
> + if ((act != 'A' && act != 'D') || start > end) {
> + warn_report("Skipping invalid PMU filter %s", event_filters[i]);
> + continue;
> + }
It would be better to do the syntax checking up-front when
the user tries to set the property. Then you can make the
property-setting return an error for invalid strings.
> +
> + filter.base_event = start;
> + filter.nevents = end - start + 1;
> + filter.action = (act == 'A') ? KVM_PMU_EVENT_ALLOW :
> + KVM_PMU_EVENT_DENY;
> +
> + if (!kvm_arm_set_device_attr(cpu, &attr, "PMU_V3_FILTER")) {
Shouldn't we arrange for an error message if this fails?
> + break;
> + }
> + }
> +}
> +
> void kvm_arm_pmu_init(ARMCPU *cpu)
> {
> struct kvm_device_attr attr = {
> @@ -1716,6 +1794,8 @@ void kvm_arm_pmu_init(ARMCPU *cpu)
> if (!cpu->has_pmu) {
> return;
> }
> +
> + kvm_arm_pmu_filter_init(cpu);
> if (!kvm_arm_set_device_attr(cpu, &attr, "PMU")) {
> error_report("failed to init PMU");
> abort();
>
> base-commit: 760b4dcdddba4a40b9fa0eb78fdfc7eda7cb83d0
> --
> 2.40.1
thanks
-- PMM