Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.32 as permitted sender) client-ip=23.128.96.32;
Date:   Wed, 13 Dec 2023 10:15:13 -0800
In-Reply-To: <84ad3082-794b-443f-874a-d304934a395b@redhat.com>
Mime-Version: 1.0
References: <20231208184628.2297994-1-pbonzini@redhat.com> <ZXPRGzgWFqFdI_ep@google.com>
 <184e253d-06c4-419e-b2b4-7cce1f875ba5@redhat.com> <ZXnoCadq2J3cPz2r@google.com>
 <84ad3082-794b-443f-874a-d304934a395b@redhat.com>
Message-ID: <ZXn0sR6IyzLzVHW-@google.com>
Subject: Re: [PATCH] KVM: selftests: fix supported_flags for aarch64
From:   Sean Christopherson <seanjc@google.com>
To:     Paolo Bonzini <pbonzini@redhat.com>
Cc:     Shaoqin Huang <shahuang@redhat.com>, linux-kernel@vger.kernel.org,
        kvm@vger.kernel.org
Content-Type: text/plain; charset="us-ascii"
Precedence: bulk

On Wed, Dec 13, 2023, Paolo Bonzini wrote:
> On 12/13/23 18:21, Sean Christopherson wrote:
> > On Tue, Dec 12, 2023, Paolo Bonzini wrote:
> > > On 12/9/23 03:29, Sean Christopherson wrote:
> > > > On Fri, Dec 08, 2023, Paolo Bonzini wrote:
> > > > > KVM/Arm supports readonly memslots; fix the calculation of
> > > > > supported_flags in set_memory_region_test.c, otherwise the
> > > > > test fails.
> > > > 
> > > > You got beat by a few hours, and by a better solution ;-)
> > > > 
> > > > https://lore.kernel.org/all/20231208033505.2930064-1-shahuang@redhat.com
> > > 
> > > Better but also wrong---and my patch has the debatable merit of more
> > > clearly exposing the wrongness.  Testing individual architectures is bad,
> > > but testing __KVM_HAVE_READONLY_MEM makes the test fail when running a new
> > > test on an old kernel.
> > 
> > But we already crossed that bridge and burned it for good measure by switching
> > to KVM_SET_USER_MEMORY_REGION2, i.e. as of commit
> > 
> >    8d99e347c097 ("KVM: selftests: Convert lib's mem regions to KVM_SET_USER_MEMORY_REGION2")
> > 
> > selftests built against a new kernel can't run on an old kernel.  Building KVM
> > selftests requires kernel headers, so while not having a hard requirement that
> > the uapi headers are fresh would be nice, I don't think it buys all that much.
> > 
> > If we wanted to assert that x86, arm64, etc. enumerate __KVM_HAVE_READONLY_MEM,
> > i.e. ensure that read-only memory is supported as expected, then that can be done
> > as a completely unrelated test.
> 
> selftests have the luxury of having sync-ed kernel headers, but in general
> userspace won't, and that means __KVM_HAVE_READONLY_MEM would be a very poor
> userspace API.  Fortunately it has "__" so it is not userspace API at all,
> and I don't want selftests to treat it as one.

Wait, what?  How does double underscores exempt it from being uAPI?  AIUI, the C
standard effectively ensures that userspace won't define/declare symbols with
double underscores, i.e. ensures there won't be conflicts.  But pretty much all
of the kernel-defined types are prefixed with "__", e.g. __u8 and friends, so I
don't see how prefixing with "__" exempts something from becoming uAPI.

I completely agree that __KVM_HAVE_READONLY_MEM shouldn't be uAPI, but then it
really, really shouldn't be defined in arch/x86/include/uapi/asm/kvm.h.