Re: [PATCH v15 05/23] x86/virt/tdx: Handle SEAMCALL no entropy error in common code

[Huang, Kai]

> > +#include <asm/archrandom.h>
> > +
> > +typedef u64 (*sc_func_t)(u64 fn, struct tdx_module_args *args);
> > +
> > +static inline u64 sc_retry(sc_func_t func, u64 fn,
> > +			   struct tdx_module_args *args)
> > +{
> > +	int retry = RDRAND_RETRY_LOOPS;
> > +	u64 ret;
> > +
> > +	do {
> > +		ret = func(fn, args);
> > +	} while (ret == TDX_RND_NO_ENTROPY && --retry);
> 
> This loop assumes that args isn't touched when TDX_RND_NO_ENTRYPOY is returned.
> It's not true.  TDH.SYS.INIT() and TDH.SYS.LP.INIT() clear RCX, RDX, etc on
> error including TDX_RND_NO_ENTRY.  Because TDH.SYS.INIT() takes RCX as input,
> this wrapper doesn't work.  TDH.SYS.LP.INIT() doesn't use RCX, RDX ... as
> input. So it doesn't matter.
> 
> Other SEAMCALLs doesn't touch registers on the no entropy error.
> TDH.EXPORTS.STATE.IMMUTABLE(), TDH.IMPORTS.STATE.IMMUTABLE(), TDH.MNG.ADDCX(),
> and TDX.MNG.CREATE().  TDH.SYS.INIT() is an exception.

If I am reading the spec (TDX module 1.5 ABI) correctly the TDH.SYS.INIT doesn't
return TDX_RND_NO_ENTROPY.  TDH.SYS.LP.INIT indeed can return NO_ENTROPY but as
you said it doesn't take any register as input.  So technically the code works
fine.  (Even the TDH.SYS.INIT can return NO_ENTROPY the code still works fine
because the RCX must be 0 for TDH.SYS.INIT.)

Also, I can hardly think out of any reason why TDX module needs to clobber input
registers in case of NO_ENTROPY for *ANY* SEAMCALL.  But despite that, I am not
opposing the idea that it *MIGHT* be better to "not assume" NO_ENTROPY will
never clobber registers either, e.g., for the sake of future extendibility.  In
this case, the below diff should address:

diff --git a/arch/x86/include/asm/tdx.h b/arch/x86/include/asm/tdx.h
index a621721f63dd..962a7a6be721 100644
--- a/arch/x86/include/asm/tdx.h
+++ b/arch/x86/include/asm/tdx.h
@@ -97,12 +97,23 @@ typedef u64 (*sc_func_t)(u64 fn, struct tdx_module_args
*args);
 static inline u64 sc_retry(sc_func_t func, u64 fn,
                           struct tdx_module_args *args)
 {
+       struct tdx_module_args _args = *args;
        int retry = RDRAND_RETRY_LOOPS;
        u64 ret;
 
-       do {
-               ret = func(fn, args);
-       } while (ret == TDX_RND_NO_ENTROPY && --retry);
+again:
+       ret = func(fn, args);
+       if (ret == TDX_RND_NO_ENTROPY && --retry) {
+               /*
+                * Do not assume TDX module will never clobber the input
+                * registers when any SEAMCALL fails with out of entropy.
+                * In this case the original input registers in @args
+                * are clobbered.  Always restore the input registers
+                * before retrying the SEAMCALL.
+                */
+               *args = _args;
+               goto again;
+       }
 
        return ret;
 }


The downside is we will have an additional memory copy of 'struct
tdx_module_args' for each SEAMCALL, but I don't believe this will have any
difference in practice.

Or, we can go and ask TDX module guys to promise no input registers will be
clobbered in case of NO_ENTROPY.

Hi Dave,

Do you have any opinion?