Message-ID: <MW4PR11MB5824B76276020C1A33A26A72BB27A@MW4PR11MB5824.namprd11.prod.outlook.com>
Date:   Tue, 27 Jun 2023 03:29:21 +0000
From:   "Zhang, Xiong Y" <xiong.y.zhang@...el.com>
To:     "Christopherson,, Sean" <seanjc@...gle.com>
CC:     "kvm@...r.kernel.org" <kvm@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "pbonzini@...hat.com" <pbonzini@...hat.com>,
        "peterz@...radead.org" <peterz@...radead.org>,
        "like.xu.linux@...il.com" <like.xu.linux@...il.com>,
        "kan.liang@...ux.intel.com" <kan.liang@...ux.intel.com>,
        "zhenyuw@...ux.intel.com" <zhenyuw@...ux.intel.com>,
        "Lv, Zhiyuan" <zhiyuan.lv@...el.com>,
        "Yang, Weijiang" <weijiang.yang@...el.com>
Subject: RE: [PATCH 3/4] KVM: VMX/pmu: Enable inactive vLBR event in guest LBR
 MSR emulation

> On Sun, Jun 25, 2023, Xiong Y Zhang wrote:
> > > > On Fri, Jun 16, 2023, Xiong Zhang wrote:
> > > > vLBR event could be inactive in two case:
> > > > a. host per cpu pinned LBR event occupy LBR when vLBR event is
> > > > created b. vLBR event is preempted by host per cpu pinned LBR
> > > > event during vm exit handler.
> > > > When vLBR event is inactive, guest couldn't access LBR msr, and it
> > > > is forced into error state and is excluded from schedule by perf scheduler.
> > > > So vLBR event couldn't be active through perf scheduler even if
> > > > host per cpu pinned LBR event has released LBR, kvm could enable
> > > > vLBR event proactively, then vLBR event may be active and LBR msr
> > > > could be passthrough into guest.
> > > >
> > > > Signed-off-by: Xiong Zhang <xiong.y.zhang@...el.com>
> > > > ---
> > > >  arch/x86/kvm/vmx/pmu_intel.c | 11 ++++++++++-
> > > >  1 file changed, 10 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/arch/x86/kvm/vmx/pmu_intel.c
> > > > b/arch/x86/kvm/vmx/pmu_intel.c index 741efe2c497b..5a3ab8c8711b
> > > > 100644
> > > > --- a/arch/x86/kvm/vmx/pmu_intel.c
> > > > +++ b/arch/x86/kvm/vmx/pmu_intel.c
> > > > @@ -314,7 +314,16 @@ static bool
> > > > intel_pmu_handle_lbr_msrs_access(struct
> > > kvm_vcpu *vcpu,
> > > >  	if (!intel_pmu_is_valid_lbr_msr(vcpu, index))
> > > >  		return false;
> > > >
> > > > -	if (!lbr_desc->event && intel_pmu_create_guest_lbr_event(vcpu) < 0)
> > > > +	/* vLBR event may be inactive, but physical LBR may be free now.
> > >
> > > 	/*
> > > 	 * This is the preferred block comment style.
> > > 	 */
> > >
> > > > +	 * but vLBR event is pinned event, once it is inactive state, perf
> > > > +	 * will force it to error state in merge_sched_in() and exclude it from
> > > > +	 * perf schedule, so even if LBR is free now, vLBR event
> > > > +couldn't be
> > > active
> > > > +	 * through perf scheduler and vLBR event could be active through
> > > > +	 * perf_event_enable().
> > > > +	 */
> > >
> > > Trimming that down, is this what you mean?
> > Yes, thanks a lot.
> > >
> > > 	/*
> > > 	 * Attempt to re-enable the vLBR event if it was disabled due to
> > > 	 * contention with host LBR usage, i.e. was put into an error state.
> > > 	 * Perf doesn't notify KVM if the host stops using LBRs, i.e. KVM needs
> > > 	 * to manually re-enable the event.
> > > 	 */
> > >
> > > Which begs the question, why can't there be a notification of some
> > > form that the LBRs are once again available?
> > This is perf scheduler rule. If pinned event couldn't get resource as
> > resource limitation, perf will put it into error state and exclude it
> > from perf scheduler, even if resource available later, perf won't
> > schedule it again as it is in error state, the only way to reschedule
> > it is to enable it again.  If non-pinned event couldn't get resource
> > as resource limitation, perf will put it into inactive state, perf
> > will reschedule it automatically once resource is available.  vLBR event is per
> process pinned event.
> 
> That doesn't answer my question.  I get that all of this is subject to perf
> scheduling, I'm asking why perf doesn't communicate directly with KVM to
> coordinate access to LBRs instead of pulling the rug out from under KVM.
Perf doesn't need such notification interface currently, as non-pinned event will be active automatically once resource available, only pinned event is still in inactive even if resource available, perf may refuse to add such interface for KVM usage only. Perf may suggest us to modify vLBR event from pinned to nonpinned.
> 
> > > Assuming that's too difficult for whatever reason, why wait until
> > > the guest tries to read LBRs?  E.g. why not be more aggressive and
> > > try to re-enable vLBRs on every VM-Exit.
> > Yes, it is a good suggestion. Actually vmx_passthrough_lbr_msrs() is
> > called on every VM-exit, it also check vLBR event state, so I could re-enable
> vLBR in this function.
> > >
> > > And if we do wait until the guest touches relevant MSRs, shouldn't
> > > writes to DEBUG_CTL that set DEBUGCTLMSR_LBR also try to re-enable the
> event?
> > Only perf know whether vLBR event could be active or not at this
> > moment, if vLBR is active, KVM could read/write DEBUG_CTL[0] with irq
> > disable/enable pair in theory, but it is better that kvm don't touch
> > perf hw resource directly, as vLBR event is one host LBR event, host
> > may have other LBR event, perf will schedule them according to perf
> > scheduler rule.  If vLBR is inactive, KVM shouldn't touch DEBUG_CTL MSR
> totally.
> 
> Again, this doesn't answer my question.  I didn't suggest KVM write anything
> directly, I asked why KVM doesn't try to re-enable the perf LBR event when
> emulating a guest write to DEBUG_CTL.
Yes, kvm should re-enable vLBR event when emulating a guest write to DEBUG_CTL. Since inactive vLBR event will be re-enabled in every VM-Exit according to the above suggestion, no extra code is needed here.
> 
> > > Lastly, what guarantees that the MSRs hold guest data?  I assume
> > > perf purges the MSRs at some point, but it would be helpful to call that out in
> the changelog.
> > For DEBUG_CTL msr, VMCS has two fields for this:
> > 1. "Save debug controls" in VM-Exit controls 2. "Load debug controls"
> > in VM-Entry controls For LBR records MSRs, perf will save them at
> > process schedule out and load them at process schedule in.
> 
> Once again, this doesn't answer my question.  I want to know *exactly* when
> perf can take control of the LBRs.  The fact that
> intel_pmu_handle_lbr_msrs_access()
> disables IRQs when checking lbr_desc->event->state strongly suggests that the
> answer isn't "when a task is scheduled in".
> 
> Your other response[1] mostly answered that question, but I want explicit
> documentation on the contract between perf and KVM with respect to LBRs.  In
> short, please work with Weijiang to fulfill my request/demand[*] that someone
> document KVM's LBR support, and justify the "design".  I am simply not willing to
> take KVM LBR patches until that documentation is provided.
Sure, I will work with Weijiang to supply such documentation. Will this document be put in Documentation/virt/kvm/x86/ ?
> 
> Copy+pasting my request/demand for convenience:
> 
>   : First and foremost, the existing LBR support needs to be documented.
> Someone,
>   : I don't care who, needs to provide a detailed writeup of the contract between
> KVM
>   : and perf.  Specifically, I want to know:
>   :
>   :   1. When exactly is perf allowed to take control of LBR MRS.  Task switch?
> IRQ?
>   :      NMI?
>   :
>   :   2. What is the expected behavior when perf is using LBRs?  Is the guest
> supposed
>   :      to be traced?
>   :
>   :   3. Why does KVM snapshot DEBUGCTL with IRQs enabled, but disables IRQs
> when
>   :      accessing LBR MSRs?
>   :
>   : It doesn't have to be polished, e.g. I'll happily wordsmith things into proper
>   : documentation, but I want to have a very clear understanding of how LBR
> support
>   : is _intended_ to function and how it all _actually_ functions without having to
>   : make guesses.
>   :
>   : And depending on the answers, I want to revisit KVM's LBR implementation
> before
>   : tackling arch LBRs.  Letting perf usurp LBRs while KVM has the vCPU loaded is
>   : frankly ridiculous.  Just have perf set a flag telling KVM that it needs to take
>   : control of LBRs and have KVM service the flag as a request or something.
> Stealing
>   : the LBRs back in IRQ context adds a stupid amount of complexity without
> much value,
>   : e.g. waiting a few branches for KVM to get to a safe place isn't going to
> meaningfully
>   : change the traces.  If that can't actually happen, then why on earth does KVM
> need
>   : to disable IRQs to read MSRs?
>   :
>   : And AFAICT, since KVM unconditionally loads the guest's DEBUGCTL, whether
> or not
>   : guest branches show up in the LBRs when the host is tracing is completely up
> to
>   : the whims of the guest.  If that's correct, then again, what's the point of the
>   : dance between KVM and perf?
> 
> [1]
> https://lkml.kernel.org/r/MW4PR11MB5824480492ED55C63769FF11BB21A%40
> MW4PR11MB5824.namprd11.prod.outlook.com
> [2] https://lore.kernel.org/all/Y9RUOvJ5dkCU9J8C@google.com

Powered by blists - more mailing lists