x86/debug: Fix stack recursion caused by wrongly ordered DR7 accesses

commit 9d2c7203ffdb846399b82b0660563c89e918c751 upstream.

In kernels compiled with CONFIG_PARAVIRT=n, the compiler re-orders the
DR7 read in exc_nmi() to happen before the call to sev_es_ist_enter().

This is problematic when running as an SEV-ES guest because in this
environment the DR7 read might cause a #VC exception, and taking #VC
exceptions is not safe in exc_nmi() before sev_es_ist_enter() has run.

The result is stack recursion if the NMI was caused on the #VC IST
stack, because a subsequent #VC exception in the NMI handler will
overwrite the stack frame of the interrupted #VC handler.

As there are no compiler barriers affecting the ordering of DR7
reads/writes, make the accesses to this register volatile, forbidding
the compiler to re-order them.

  [ bp: Massage text, make them volatile too, to make sure some
  aggressive compiler optimization pass doesn't discard them. ]

Fixes: c83729b7893f ("x86/sev-es: Adjust #VC IST Stack on entering NMI handler")
Reported-by: Alexey Kardashevskiy <aik@amd.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20230127035616.508966-1-aik@amd.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

diff --git a/arch/x86/include/asm/debugreg.h b/arch/x86/include/asm/debugreg.h
index cfdf307ddc0129019f64efebc01bf5a5159578fa..9ed8343c9b3cb3e0b80e3a6a8c90ea0da47975ac 100644 (file)
--- a/arch/x86/include/asm/debugreg.h
+++ b/arch/x86/include/asm/debugreg.h
@@ -39,7 +39,20 @@ static __always_inline unsigned long native_get_debugreg(int regno)
                asm("mov %%db6, %0" :"=r" (val));
                break;
        case 7:
-               asm("mov %%db7, %0" :"=r" (val));
+               /*
+                * Apply __FORCE_ORDER to DR7 reads to forbid re-ordering them
+                * with other code.
+                *
+                * This is needed because a DR7 access can cause a #VC exception
+                * when running under SEV-ES. Taking a #VC exception is not a
+                * safe thing to do just anywhere in the entry code and
+                * re-ordering might place the access into an unsafe location.
+                *
+                * This happened in the NMI handler, where the DR7 read was
+                * re-ordered to happen before the call to sev_es_ist_enter(),
+                * causing stack recursion.
+                */
+               asm volatile("mov %%db7, %0" : "=r" (val) : __FORCE_ORDER);
                break;
        default:
                BUG();
@@ -66,7 +79,16 @@ static __always_inline void native_set_debugreg(int regno, unsigned long value)
                asm("mov %0, %%db6"     ::"r" (value));
                break;
        case 7:
-               asm("mov %0, %%db7"     ::"r" (value));
+               /*
+                * Apply __FORCE_ORDER to DR7 writes to forbid re-ordering them
+                * with other code.
+                *
+                * While is didn't happen with a DR7 write (see the DR7 read
+                * comment above which explains where it happened), add the
+                * __FORCE_ORDER here too to avoid similar problems in the
+                * future.
+                */
+               asm volatile("mov %0, %%db7"    ::"r" (value), __FORCE_ORDER);
                break;
        default:
                BUG();