KVM Archive on lore.kernel.org
 help / color / Atom feed
From: Sean Christopherson <seanjc@google.com>
To: Frederic Weisbecker <frederic@kernel.org>
Cc: Paolo Bonzini <pbonzini@redhat.com>,
	Vitaly Kuznetsov <vkuznets@redhat.com>,
	Wanpeng Li <wanpengli@tencent.com>,
	Jim Mattson <jmattson@google.com>, Joerg Roedel <joro@8bytes.org>,
	kvm@vger.kernel.org, linux-kernel@vger.kernel.org,
	Thomas Gleixner <tglx@linutronix.de>,
	Michael Tokarev <mjt@tls.msk.ru>,
	Christian Borntraeger <borntraeger@de.ibm.com>
Subject: Re: [PATCH v3 3/9] KVM: x86: Defer tick-based accounting 'til after IRQ handling
Date: Wed, 28 Apr 2021 22:38:45 +0000
Message-ID: <YInj9QtUFdAlKqr3@google.com> (raw)
In-Reply-To: <20210421121940.GD16580@lothringen>

Apologies for the slow response.

On Wed, Apr 21, 2021, Frederic Weisbecker wrote:
> On Tue, Apr 20, 2021 at 11:26:34PM +0000, Sean Christopherson wrote:
> > On Wed, Apr 21, 2021, Frederic Weisbecker wrote:
> > > On Thu, Apr 15, 2021 at 03:21:00PM -0700, Sean Christopherson wrote:
> > > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > > > index 16fb39503296..e4d475df1d4a 100644
> > > > --- a/arch/x86/kvm/x86.c
> > > > +++ b/arch/x86/kvm/x86.c
> > > > @@ -9230,6 +9230,14 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
> > > >  	local_irq_disable();
> > > >  	kvm_after_interrupt(vcpu);
> > > >  
> > > > +	/*
> > > > +	 * When using tick-based accounting, wait until after servicing IRQs to
> > > > +	 * account guest time so that any ticks that occurred while running the
> > > > +	 * guest are properly accounted to the guest.
> > > > +	 */
> > > > +	if (!vtime_accounting_enabled_this_cpu())
> > > > +		vtime_account_guest_exit();
> > > 
> > > Can we rather have instead:
> > > 
> > > static inline void tick_account_guest_exit(void)
> > > {
> > > 	if (!vtime_accounting_enabled_this_cpu())
> > > 		current->flags &= ~PF_VCPU;
> > > }
> > > 
> > > It duplicates a bit of code but I think this will read less confusing.
> > 
> > Either way works for me.  I used vtime_account_guest_exit() to try to keep as
> > many details as possible inside vtime, e.g. in case the implemenation is tweaked
> > in the future.  But I agree that pretending KVM isn't already deeply intertwined
> > with the details is a lie.
> 
> Ah I see, before 87fa7f3e98a131 the vtime was accounted after interrupts get
> processed. So it used to work until then. I see that ARM64 waits for IRQs to
> be enabled too.
> 
> PPC/book3s_hv, MIPS, s390 do it before IRQs get re-enabled (weird, how does that
> work?)

No idea.  It's entirely possible it doesn't work on one or more of those
architectures.

Based on init/Kconfig, s390 doesn't support tick-based accounting, so I assume
s390 is ok.

  config TICK_CPU_ACCOUNTING
	bool "Simple tick based cputime accounting"
	depends on !S390 && !NO_HZ_FULL

> And PPC/book3s_pr calls guest_exit() so I guess it has interrupts enabled.
> 
> The point is: does it matter to call vtime_account_guest_exit() before or
> after interrupts? If it doesn't matter, we can simply call
> vtime_account_guest_exit() once and for all once IRQs are re-enabled.
> 
> If it does matter because we don't want to account the host IRQs firing at the
> end of vcpu exit, then probably we should standardize that behaviour and have
> guest_exit_vtime() called before interrupts get enabled and guest_exit_tick()
> called after interrupts get enabled. It's probably then beyond the scope of this
> patchset but I would like to poke your opinion on that.
> 
> Thanks.

I don't know.  For x86, I would be ok with simply moving the call to
vtime_account_guest_exit() to after IRQs are enabled.  It would bug me a little
bit that KVM _could_ be more precise when running with
CONFIG_VIRT_CPU_ACCOUNTING_GEN=y, and KVM would still be poking into the details
of vtime_account_guest_exit() to some extent, but overall it would be an
improvement from a code cleanliness perspective.

The problem is I have no clue who, if anyone, deploys KVM on x86 with
CONFIG_VIRT_CPU_ACCOUNTING_GEN=y.  On the other hand, AMD/SVM has always had the
"inaccurate" accounting, and Intel/VMX has been inaccurate since commit
d7a08882a0a4 ("KVM: x86: Unconditionally enable irqs in guest context"), which
amusingly was a fix for an edge case in tick-based accounting.

Anyone have an opinion either way?  I'm very tempted to go with Frederic's
suggestion of moving the time accounting back to where it was, it makes KVM just
a little less ugly.

  reply index

Thread overview: 25+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-04-15 22:20 [PATCH v3 0/9] KVM: Fix tick-based accounting for x86 guests Sean Christopherson
2021-04-15 22:20 ` [PATCH v3 1/9] context_tracking: Move guest exit context tracking to separate helpers Sean Christopherson
2021-04-20 18:48   ` Christian Borntraeger
2021-04-21 10:57   ` Frederic Weisbecker
2021-04-15 22:20 ` [PATCH v3 2/9] context_tracking: Move guest exit vtime accounting " Sean Christopherson
2021-04-20 18:48   ` Christian Borntraeger
2021-04-15 22:21 ` [PATCH v3 3/9] KVM: x86: Defer tick-based accounting 'til after IRQ handling Sean Christopherson
2021-04-20 23:14   ` Frederic Weisbecker
2021-04-20 23:26     ` Sean Christopherson
2021-04-21 10:11       ` Frederic Weisbecker
2021-04-21 12:19       ` Frederic Weisbecker
2021-04-28 22:38         ` Sean Christopherson [this message]
2021-04-21 10:07   ` Frederic Weisbecker
2021-04-15 22:21 ` [PATCH v3 4/9] sched/vtime: Move vtime accounting external declarations above inlines Sean Christopherson
2021-04-21  7:02   ` Christian Borntraeger
2021-04-15 22:21 ` [PATCH v3 5/9] sched/vtime: Move guest enter/exit vtime accounting to vtime.h Sean Christopherson
2021-04-15 22:21 ` [PATCH v3 6/9] context_tracking: Consolidate guest enter/exit wrappers Sean Christopherson
2021-04-15 22:21 ` [PATCH v3 7/9] context_tracking: KVM: Move guest enter/exit wrappers to KVM's domain Sean Christopherson
2021-04-21  7:10   ` Christian Borntraeger
2021-04-15 22:21 ` [PATCH v3 8/9] KVM: x86: Consolidate guest enter/exit logic to common helpers Sean Christopherson
2021-04-15 22:21 ` [PATCH v3 9/9] KVM: Move instrumentation-safe annotations for enter/exit to x86 code Sean Christopherson
2021-04-21  8:09   ` Christian Borntraeger
2021-04-22 14:38     ` Sven Schnelle
2021-04-23  9:32       ` Vasily Gorbik
2021-04-20 23:33 ` [PATCH v3 0/9] KVM: Fix tick-based accounting for x86 guests Frederic Weisbecker

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=YInj9QtUFdAlKqr3@google.com \
    --to=seanjc@google.com \
    --cc=borntraeger@de.ibm.com \
    --cc=frederic@kernel.org \
    --cc=jmattson@google.com \
    --cc=joro@8bytes.org \
    --cc=kvm@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=mjt@tls.msk.ru \
    --cc=pbonzini@redhat.com \
    --cc=tglx@linutronix.de \
    --cc=vkuznets@redhat.com \
    --cc=wanpengli@tencent.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

KVM Archive on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/kvm/0 kvm/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 kvm kvm/ https://lore.kernel.org/kvm \
		kvm@vger.kernel.org
	public-inbox-index kvm

Example config snippet for mirrors

Newsgroup available over NNTP:
	nntp://nntp.lore.kernel.org/org.kernel.vger.kvm


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git