x86/srso: Explain the untraining sequences a bit more

commit 9dbd23e upstream.

The goal is to eventually have a proper documentation about all this.

Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Link: https://lore.kernel.org/r/20230814164447.GFZNpZ/64H4lENIe94@fat_crate.local
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

arch/x86/lib/retpoline.S

@@ -186,6 +186,25 @@ SYM_CODE_START(srso_alias_return_thunk)
 	ud2
 SYM_CODE_END(srso_alias_return_thunk)
 
+/*
+ * Some generic notes on the untraining sequences:
+ *
+ * They are interchangeable when it comes to flushing potentially wrong
+ * RET predictions from the BTB.
+ *
+ * The SRSO Zen1/2 (MOVABS) untraining sequence is longer than the
+ * Retbleed sequence because the return sequence done there
+ * (srso_safe_ret()) is longer and the return sequence must fully nest
+ * (end before) the untraining sequence. Therefore, the untraining
+ * sequence must fully overlap the return sequence.
+ *
+ * Regarding alignment - the instructions which need to be untrained,
+ * must all start at a cacheline boundary for Zen1/2 generations. That
+ * is, instruction sequences starting at srso_safe_ret() and
+ * the respective instruction sequences at retbleed_return_thunk()
+ * must start at a cacheline boundary.
+ */
+
 /*
  * Safety details here pertain to the AMD Zen{1,2} microarchitecture:
  * 1) The RET at retbleed_return_thunk must be on a 64 byte boundary, for