Re: [PATCH v15 17/23] x86/kexec: Flush cache of TDX private memory

["Huang, Kai" <kai.huang@intel.com>]

On Mon, 2023-11-27 at 12:05 -0800, Dave Hansen wrote:
> On 11/27/23 11:33, Huang, Kai wrote:
> > On Mon, 2023-11-27 at 10:13 -0800, Hansen, Dave wrote:
> > > On 11/9/23 03:55, Kai Huang wrote:
> > > ...
> > > > --- a/arch/x86/kernel/reboot.c
> > > > +++ b/arch/x86/kernel/reboot.c
> > > > @@ -31,6 +31,7 @@
> > > >  #include <asm/realmode.h>
> > > >  #include <asm/x86_init.h>
> > > >  #include <asm/efi.h>
> > > > +#include <asm/tdx.h>
> > > > 
> > > >  /*
> > > >   * Power off function, if any
> > > > @@ -741,6 +742,20 @@ void native_machine_shutdown(void)
> > > >     local_irq_disable();
> > > >     stop_other_cpus();
> > > >  #endif
> > > > +   /*
> > > > +    * stop_other_cpus() has flushed all dirty cachelines of TDX
> > > > +    * private memory on remote cpus.  Unlike SME, which does the
> > > > +    * cache flush on _this_ cpu in the relocate_kernel(), flush
> > > > +    * the cache for _this_ cpu here.  This is because on the
> > > > +    * platforms with "partial write machine check" erratum the
> > > > +    * kernel needs to convert all TDX private pages back to normal
> > > > +    * before booting to the new kernel in kexec(), and the cache
> > > > +    * flush must be done before that.  If the kernel took SME's way,
> > > > +    * it would have to muck with the relocate_kernel() assembly to
> > > > +    * do memory conversion.
> > > > +    */
> > > > +   if (platform_tdx_enabled())
> > > > +           native_wbinvd();
> > > 
> > > Why can't the TDX host code just set host_mem_enc_active=1?
> > > 
> > > Sure, you'll end up *using* the SME WBINVD support, but then you don't
> > > have two different WBINVD call sites.  You also don't have to mess with
> > > a single line of assembly.
> > 
> > I wanted to avoid changing the assembly.
> > 
> > Perhaps the comment isn't very clear.  Flushing cache (on the CPU running kexec)
> > when the host_mem_enc_active=1 is handled in the relocate_kernel() assembly,
> > which happens at very late stage right before jumping to the new kernel.
> > However for TDX when the platform has erratum we need to convert TDX private
> > pages back to normal, which must be done after flushing cache.  If we reuse
> > host_mem_enc_active=1, then we will need to change the assembly code to do that.
> 
> I honestly think you need to stop thinking about the partial write issue
> at this point in the series.  It's really causing a horrible amount of
> unnecessary confusion.
> 
> Here's the golden rule:
> 
> 	The boot CPU needs to run WBINVD sometime after it stops writing
> 	to private memory but before it starts treating the memory as
> 	shared.
> 
> On SME kernels, that key point evidently in early boot when it's
> enabling SME.  I _think_ that point is also a valid place to do WBINVD
> on no-TDX-erratum systems.

You mean we could advertise cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT) true for
TDX host? We could but IMHO it doesn't perfectly match.

SME kernel sets _PAGE_ENC on by default for all memory mappings IIUC, but TDX
host never actually sets any encryption bits in page tables managed by the
kernel.

So I think we can just do below, but not advertise CC_ATTR_HOST_MEM_ENCRYPT for
TDX host?

--- a/arch/x86/kernel/machine_kexec_64.c
+++ b/arch/x86/kernel/machine_kexec_64.c
@@ -377,7 +377,8 @@ void machine_kexec(struct kimage *image)
                                       (unsigned long)page_list,
                                       image->start,
                                       image->preserve_context,
-                                     
cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT));
+                                      cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT)
||
+                                      platform_tdx_enabled());


> 
> On TDX systems with the erratum, there's a *second* point before the
> private=>shared conversion occurs.  I think what you're trying to do
> here is prematurely optimize the erratum-affected affected systems so
> that they don't do two WBINVDs.  Please stop trying to do that.
> 
> This WBINVD is only _needed_ for the erratum.  It should be closer to
> the actual erratum handing.

If we do WBINVD early here then the second one isn't needed.  But 100% agreed
this handling/optimization should be done later closer to the erratum handling.