Received-SPF: pass (google.com: domain of linux-kernel+bounces-77155-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
Subject: [RFC][PATCH 01/34] x86/xen: Explain why calling get_cpu_cap() so early is a hack
To: linux-kernel@vger.kernel.org
Cc: kirill.shutemov@linux.intel.com,pbonzini@redhat.com,tglx@linutronix.de,x86@kernel.org,bp@alien8.de,Dave Hansen <dave.hansen@linux.intel.com>,jgross@suse.com
From: Dave Hansen <dave.hansen@linux.intel.com>
Date: Thu, 22 Feb 2024 10:39:28 -0800
References: <20240222183926.517AFCD2@davehans-spike.ostc.intel.com>
In-Reply-To: <20240222183926.517AFCD2@davehans-spike.ostc.intel.com>
Message-Id: <20240222183928.0D8942B2@davehans-spike.ostc.intel.com>
Precedence: bulk


From: Dave Hansen <dave.hansen@linux.intel.com>

xen_start_kernel() some of the first C code run "Xen PV" systems.
It precedes normally very early things like setup_arch() or the
processor initialization code.

That means that 'boot_cpu_data' is garbage.  It has not even
established the utter basics like if the CPU supports the CPUID
instruction.  Unfortunately get_cpu_cap() requires this exact
information.

Nevertheless xen_start_kernel() calls get_cpu_cap().  But it
works out in practice because it's looking for the NX bit which
comes from an extended CPUID leaf that doesn't depend on
c->cpuid_level being set.  This also implicitly assumes that
Xen PV guests support CPUID.

Leave the hack in place, but at least explain some of what is
going on.

Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Juergen Gross <jgross@suse.com>
---

 b/arch/x86/xen/enlighten_pv.c |    6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff -puN arch/x86/xen/enlighten_pv.c~xen-explain1 arch/x86/xen/enlighten_pv.c
--- a/arch/x86/xen/enlighten_pv.c~xen-explain1	2024-02-22 10:08:48.404451146 -0800
+++ b/arch/x86/xen/enlighten_pv.c	2024-02-22 10:08:48.404451146 -0800
@@ -1372,7 +1372,11 @@ asmlinkage __visible void __init xen_sta
 	/* Get mfn list */
 	xen_build_dynamic_phys_to_machine();
 
-	/* Work out if we support NX */
+	/*
+	 * This is a hack.  'boot_cpu_data' is not filled out enough
+	 * for get_cpu_cap() to function fully.  But it _can_ fill out
+	 * the leaves where NX is.  Gross, but it works.
+	 */
 	get_cpu_cap(&boot_cpu_data);
 	x86_configure_nx();
 
_