Re: [PATCH v4 45/75] x86/sev-es: Adjust #VC IST Stack on entering NMI handler

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Jul 14, 2020 at 02:08:47PM +0200, Joerg Roedel wrote:

> @@ -489,6 +490,9 @@ DEFINE_IDTENTRY_RAW(exc_nmi)
>  	this_cpu_write(nmi_cr2, read_cr2());
>  nmi_restart:
>  
> +	/* Needs to happen before DR7 is accessed */
> +	sev_es_ist_enter(regs);
> +
>  	this_cpu_write(nmi_dr7, local_db_save());
>  
>  	nmi_enter();
> @@ -502,6 +506,8 @@ DEFINE_IDTENTRY_RAW(exc_nmi)
>  
>  	local_db_restore(this_cpu_read(nmi_dr7));
>  
> +	sev_es_ist_exit();
> +
>  	if (unlikely(this_cpu_read(nmi_cr2) != read_cr2()))
>  		write_cr2(this_cpu_read(nmi_cr2));
>  	if (this_cpu_dec_return(nmi_state))

I really hate all this #VC stuff :-(

So the above will make the NMI do 4 unconditional extra CALL+RET, a LOAD
(which will potentially miss) and a compare and branch.

How's that a win for normal people? Can we please turn all these
sev_es_*() hooks into something like:

DECLARE_STATIC_KEY_FALSE(sev_es_enabled_key);

static __always_inline void sev_es_foo()
{
	if (static_branch_unlikely(&sev_es_enabled_key))
		__sev_es_foo();
}

So that normal people will only see an extra NOP?

> diff --git a/arch/x86/kernel/sev-es.c b/arch/x86/kernel/sev-es.c
> index d415368f16ec..2a7cc72db1d5 100644
> --- a/arch/x86/kernel/sev-es.c
> +++ b/arch/x86/kernel/sev-es.c
> @@ -78,6 +78,67 @@ static void __init sev_es_setup_vc_stacks(int cpu)
>  	tss->x86_tss.ist[IST_INDEX_VC] = CEA_ESTACK_TOP(&cea->estacks, VC);
>  }
>  
> +static bool on_vc_stack(unsigned long sp)

noinstr or __always_inline

> +{
> +	return ((sp >= __this_cpu_ist_bot_va(VC)) && (sp < __this_cpu_ist_top_va(VC)));
> +}
> +
> +/*
> + * This function handles the case when an NMI or an NMI-like exception
> + * like #DB is raised in the #VC exception handler entry code. In this

I've yet to find you handle the NMI-like cases..

> + * case the IST entry for VC must be adjusted, so that any subsequent VC
> + * exception will not overwrite the stack contents of the interrupted VC
> + * handler.
> + *
> + * The IST entry is adjusted unconditionally so that it can be also be
> + * unconditionally back-adjusted in sev_es_nmi_exit(). Otherwise a
> + * nested nmi_exit() call (#VC->NMI->#DB) may back-adjust the IST entry
> + * too early.

Is this comment accurate, I cannot find the patch touching
nmi_enter/exit()?

> + */
> +void noinstr sev_es_ist_enter(struct pt_regs *regs)
> +{
> +	unsigned long old_ist, new_ist;
> +	unsigned long *p;
> +
> +	if (!sev_es_active())
> +		return;
> +
> +	/* Read old IST entry */
> +	old_ist = __this_cpu_read(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC]);
> +
> +	/* Make room on the IST stack */
> +	if (on_vc_stack(regs->sp))
> +		new_ist = ALIGN_DOWN(regs->sp, 8) - sizeof(old_ist);
> +	else
> +		new_ist = old_ist - sizeof(old_ist);
> +
> +	/* Store old IST entry */
> +	p       = (unsigned long *)new_ist;
> +	*p      = old_ist;
> +
> +	/* Set new IST entry */
> +	this_cpu_write(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC], new_ist);
> +}
> +
> +void noinstr sev_es_ist_exit(void)
> +{
> +	unsigned long ist;
> +	unsigned long *p;
> +
> +	if (!sev_es_active())
> +		return;
> +
> +	/* Read IST entry */
> +	ist = __this_cpu_read(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC]);
> +
> +	if (WARN_ON(ist == __this_cpu_ist_top_va(VC)))
> +		return;
> +
> +	/* Read back old IST entry and write it to the TSS */
> +	p = (unsigned long *)ist;
> +	this_cpu_write(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC], *p);
> +}

That's pretty disguisting :-(