> So, I'll flip this back around. Today, X86_FEATURE_RSB_CTXSW zaps the > RSB whenever RSP is updated to a new task stack. Please convince me > that ERAPS provides superior coverage or is unnecessary in all the > possible combinations switching between: > > different thread, same mm > user=>kernel, same mm > kernel=>user, same mm > different mm (we already covered this) > > Because several of those switches can happen without a CR3 write or INVPCID. user=>kernel=>user, same mm explicitly does not want to flush the RAS, because if the system call is shallow enough, some of the userspace RAS is still intact on when you get back into user mode. The case which I expect will go wrong is user=>kernel=>different kthread because this stays on the same mm. That does need to flush the RAS and won't hit any TLB maintenance instructions that I'm aware of. ~Andrew
