linux-kernel - Re: [PATCH 2/2] KVM: x86/mmu: Don't step down in the TDP iterator when zapping all SPTEs

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Fri, 13 Aug 2021 16:13:22 +0000
From:   Sean Christopherson <seanjc@...gle.com>
To:     Paolo Bonzini <pbonzini@...hat.com>
Cc:     Ben Gardon <bgardon@...gle.com>,
        Vitaly Kuznetsov <vkuznets@...hat.com>,
        Wanpeng Li <wanpengli@...cent.com>,
        Jim Mattson <jmattson@...gle.com>,
        Joerg Roedel <joro@...tes.org>, kvm <kvm@...r.kernel.org>,
        LKML <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 2/2] KVM: x86/mmu: Don't step down in the TDP iterator
 when zapping all SPTEs

On Fri, Aug 13, 2021, Paolo Bonzini wrote:
> On 12/08/21 19:46, Sean Christopherson wrote:
> > > > 	if (iter->level == iter->min_level)
> > > > 		return false;
> > > > 
> > > > 	/*
> > > > 	 * Reread the SPTE before stepping down to avoid traversing into page
> > > > 	 * tables that are no longer linked from this entry.
> > > > 	 */
> > > > 	iter->old_spte = READ_ONCE(*rcu_dereference(iter->sptep));  \
> > > >                                                                        ---> this is the code that is avoided
> > > > 	child_pt = spte_to_child_pt(iter->old_spte, iter->level);   /
> > > > 	if (!child_pt)
> > > > 		return false;
> > > Ah, right - so I agree with Ben that it's not too important.
> > Ya.  There is a measurable performance improvement, but it's really only
> > meaningful when there aren't many SPTEs to zap, otherwise the cost of zapping
> > completely dominates the time.
> 
> I don't understand.  When try_step_down is called by tdp_iter_next, all it
> does is really just the READ_ONCE, because spte_to_child_pt will see a
> non-present PTE and return immediately.  Why do two, presumably cache hot,
> reads cause a measurable performance improvement?

It's entirely possible my measurements were bad and/or noisy.  Ah, and my kernel
was running with CONFIG_PROVE_RCU=y, which makes the rcu_dereference() quite a bit
more expensive.