Date:   Fri, 10 Feb 2023 01:10:57 +0000
From:   Sean Christopherson <seanjc@google.com>
To:     Lai Jiangshan <jiangshanlai@gmail.com>
Cc:     linux-kernel@vger.kernel.org, Paolo Bonzini <pbonzini@redhat.com>,
        Lai Jiangshan <jiangshan.ljs@antgroup.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Ingo Molnar <mingo@redhat.com>, Borislav Petkov <bp@alien8.de>,
        Dave Hansen <dave.hansen@linux.intel.com>, x86@kernel.org,
        "H. Peter Anvin" <hpa@zytor.com>, kvm@vger.kernel.org
Subject: Re: [PATCH V2 6/8] kvm: x86/mmu: Remove FNAME(invlpg)
Message-ID: <Y+WZoXYvacqx/+Yu@google.com>
References: <20230207155735.2845-1-jiangshanlai@gmail.com>
 <20230207155735.2845-7-jiangshanlai@gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20230207155735.2845-7-jiangshanlai@gmail.com>
Precedence: bulk

On Tue, Feb 07, 2023, Lai Jiangshan wrote:
> Use FNAME(sync_spte) to share the code which has a slight semantics
> changed: clean vTLB entry is kept.

...

> +static void __kvm_mmu_invalidate_gva(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
> +				     gva_t gva, hpa_t root_hpa)
> +{
> +	struct kvm_shadow_walk_iterator iterator;
> +
> +	vcpu_clear_mmio_info(vcpu, gva);
> +
> +	write_lock(&vcpu->kvm->mmu_lock);
> +	for_each_shadow_entry_using_root(vcpu, root_hpa, gva, iterator) {
> +		struct kvm_mmu_page *sp = sptep_to_sp(iterator.sptep);
> +
> +		if (sp->unsync && *iterator.sptep) {

Please make the !0 change in a separate patch.  It took me a while to connect the
dots, and to also understand what I suspect is a major motivation: sync_spte()
already has this check, i.e. the change is happening regardless, so might as well
avoid the indirect branch.

> +			gfn_t gfn = kvm_mmu_page_get_gfn(sp, iterator.index);
> +			int ret = mmu->sync_spte(vcpu, sp, iterator.index);
> +
> +			if (ret < 0)
> +				mmu_page_zap_pte(vcpu->kvm, sp, iterator.sptep, NULL);
> +			if (ret)
> +				kvm_flush_remote_tlbs_with_address(vcpu->kvm, gfn, 1);

Why open code kvm_flush_remote_tlbs_sptep()?  Does it actually shave enough
cycles to be visible?

If open coding is really justified, can you rebase on one of the two branches?
And then change this to kvm_flush_remote_tlbs_gfn().

  https://github.com/kvm-x86/linux/tree/next
  https://github.com/kvm-x86/linux/tree/mmu