All of lore.kernel.org
 help / color / mirror / Atom feed
From: Wei Wang <wei.w.wang@intel.com>
To: Eric Hankland <ehankland@google.com>,
	Paolo Bonzini <pbonzini@redhat.com>,
	rkrcmar@redhat.com
Cc: linux-kernel@vger.kernel.org,
	Stephane Eranian <eranian@google.com>,
	kvm@vger.kernel.org
Subject: Re: [PATCH v2] KVM: x86: PMU Event Filter
Date: Fri, 12 Jul 2019 11:27:34 +0800	[thread overview]
Message-ID: <5D27FE26.1050002@intel.com> (raw)
In-Reply-To: <CAOyeoRUUK+T_71J=+zcToyL93LkpARpsuWSfZS7jbJq=wd1rQg@mail.gmail.com>

On 07/11/2019 09:25 AM, Eric Hankland wrote:
> - Add a VM ioctl that can control which events the guest can monitor.
>
> Signed-off-by: ehankland<ehankland@google.com>
> ---
> Changes since v1:
> -Moved to a vm ioctl rather than a vcpu one
> -Changed from a whitelist to a configurable filter which can either be
> white or black
> -Only restrict GP counters since fixed counters require extra handling
> and they can be disabled by setting the guest cpuid (though only by
> setting the number - they can't be disabled individually)

I think just disabling guest cpuid might not be enough, since guest
could write to the msr without checking the cpuid.

Why not just add a bitmap for fixed counter?
e.g. fixed_counter_reject_bitmap

At the beginning of reprogram_fixed_counter, we could add the check:

if (test_bit(idx, &kvm->arch.fixed_counter_reject_bitmap))
     return -EACCES;

(Please test with your old guest and see if they have issues if we 
inject #GP when
they try to set the fixed_ctrl msr. If there is, we could drop -EACCESS 
above)

The bitmap could be set at kvm_vm_ioctl_set_pmu_event_filter.

> +/* for KVM_CAP_PMU_EVENT_FILTER */
> +struct kvm_pmu_event_filter {
> +       __u32 type;
> +       __u32 nevents;
> +       __u64 events[0];
> +};
> +
> +#define KVM_PMU_EVENT_WHITELIST 0
> +#define KVM_PMU_EVENT_BLACKLIST 1

I think it would be better to add more, please see below:

enum kvm_pmu_action_type {
     KVM_PMU_EVENT_ACTION_NONE = 0,
     KVM_PMU_EVENT_ACTION_ACCEPT = 1,
     KVM_PMU_EVENT_ACTION_REJECT = 2,
     KVM_PMU_EVENT_ACTION_MAX
};

and do a check in kvm_vm_ioctl_set_pmu_event_filter()
     if (filter->action >= KVM_PMU_EVENT_ACTION_MAX)
         return -EINVAL;

This is for detecting the case that we add a new action in
userspace, while the kvm hasn't been updated to support that.

KVM_PMU_EVENT_ACTION_NONE is for userspace to remove
the filter after they set it.


> +
>   #endif /* _ASM_X86_KVM_H */
> diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
> index dd745b58ffd8..d674b79ff8da 100644
> --- a/arch/x86/kvm/pmu.c
> +++ b/arch/x86/kvm/pmu.c
> @@ -22,6 +22,9 @@
>   #include "lapic.h"
>   #include "pmu.h"
>
> +/* This keeps the total size of the filter under 4k. */
> +#define KVM_PMU_EVENT_FILTER_MAX_EVENTS 63
> +

Why is this limit needed?

>   /* NOTE:
>    * - Each perf counter is defined as "struct kvm_pmc";
>    * - There are two types of perf counters: general purpose (gp) and fixed.
> @@ -144,6 +147,10 @@ void reprogram_gp_counter(struct kvm_pmc *pmc,
> u64 eventsel)
>   {
>          unsigned config, type = PERF_TYPE_RAW;
>          u8 event_select, unit_mask;
> +       struct kvm_arch *arch = &pmc->vcpu->kvm->arch;
> +       struct kvm_pmu_event_filter *filter;
> +       int i;
> +       bool allow_event = true;
>
>          if (eventsel & ARCH_PERFMON_EVENTSEL_PIN_CONTROL)
>                  printk_once("kvm pmu: pin control bit is ignored\n");
> @@ -155,6 +162,24 @@ void reprogram_gp_counter(struct kvm_pmc *pmc,
> u64 eventsel)
>          if (!(eventsel & ARCH_PERFMON_EVENTSEL_ENABLE) || !pmc_is_enabled(pmc))
>                  return;
>
> +       rcu_read_lock();
> +       filter = rcu_dereference(arch->pmu_event_filter);
> +       if (filter) {
> +               for (i = 0; i < filter->nevents; i++)
> +                       if (filter->events[i] ==
> +                           (eventsel & AMD64_RAW_EVENT_MASK_NB))
> +                               break;
> +               if (filter->type == KVM_PMU_EVENT_WHITELIST &&
> +                   i == filter->nevents)
> +                       allow_event = false;
> +               if (filter->type == KVM_PMU_EVENT_BLACKLIST &&
> +                   i < filter->nevents)
> +                       allow_event = false;
> +       }
> +       rcu_read_unlock();
> +       if (!allow_event)
> +               return;
> +

I think it looks tidier to wrap the changes above into a function:

     if (kvm_pmu_filter_event(kvm, eventsel & AMD64_RAW_EVENT_MASK_NB))
         return;

>          event_select = eventsel & ARCH_PERFMON_EVENTSEL_EVENT;
>          unit_mask = (eventsel & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
>
> @@ -351,3 +376,39 @@ void kvm_pmu_destroy(struct kvm_vcpu *vcpu)
>   {
>          kvm_pmu_reset(vcpu);
>   }
> +
> +int kvm_vm_ioctl_set_pmu_event_filter(struct kvm *kvm, void __user *argp)
> +{
> +       struct kvm_pmu_event_filter tmp, *filter;
> +       size_t size;
> +       int r;
> +
> +       if (copy_from_user(&tmp, argp, sizeof(tmp)))
> +               return -EFAULT;
> +
> +       if (tmp.nevents > KVM_PMU_EVENT_FILTER_MAX_EVENTS)
> +               return -E2BIG;
> +
> +       size = sizeof(tmp) + sizeof(tmp.events[0]) * tmp.nevents;
> +       filter = vmalloc(size);
> +       if (!filter)
> +               return -ENOMEM;
> +
> +       r = -EFAULT;
> +       if (copy_from_user(filter, argp, size))

Though the above functions correctly, I would just move "r = -EFAULT" here
to have it executed conditionally.


> +               goto cleanup;
> +
> +       /* Ensure nevents can't be changed between the user copies. */
> +       *filter = tmp;
> +
> +       mutex_lock(&kvm->lock);
> +       rcu_swap_protected(kvm->arch.pmu_event_filter, filter,
> +                          mutex_is_locked(&kvm->lock));
> +       mutex_unlock(&kvm->lock);
> +
> +       synchronize_rcu();
> +       r = 0;
> +cleanup:
> +       kvfree(filter);

Probably better to have it conditionally?

if (filter) {
     synchronize_srcu();
     kfree(filter)
}


You may want to factor it out, so that kvm_pmu_destroy could reuse.

Best,
Wei

  parent reply	other threads:[~2019-07-12  3:22 UTC|newest]

Thread overview: 11+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2019-07-11  1:25 [PATCH v2] KVM: x86: PMU Event Filter Eric Hankland
2019-07-11 11:22 ` Paolo Bonzini
2019-07-11 11:58 ` Paolo Bonzini
2019-07-11 17:04   ` Eric Hankland
2019-07-11 17:12     ` Paolo Bonzini
2019-07-12  3:27 ` Wei Wang [this message]
2019-07-16  0:10   ` Eric Hankland
2019-07-16  8:49     ` Wei Wang
2019-07-17 10:22       ` Paolo Bonzini
2019-07-17 17:05         ` Eric Hankland
2019-07-17 17:30           ` Paolo Bonzini

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=5D27FE26.1050002@intel.com \
    --to=wei.w.wang@intel.com \
    --cc=ehankland@google.com \
    --cc=eranian@google.com \
    --cc=kvm@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=pbonzini@redhat.com \
    --cc=rkrcmar@redhat.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.