Re: [PATCH V6 3/8] x86/entry: Move PUSH_AND_CLEAR_REGS out of error_entry()

From: Borislav Petkov <bp@alien8.de>
To: Lai Jiangshan <jiangshanlai@gmail.com>
Cc: LKML <linux-kernel@vger.kernel.org>,
	Peter Zijlstra <peterz@infradead.org>,
	Josh Poimboeuf <jpoimboe@redhat.com>,
	Andy Lutomirski <luto@kernel.org>,
	Thomas Gleixner <tglx@linutronix.de>,
	Juergen Gross <jgross@suse.com>, X86 ML <x86@kernel.org>,
	Lai Jiangshan <jiangshan.ljs@antgroup.com>,
	Ingo Molnar <mingo@redhat.com>,
	Dave Hansen <dave.hansen@linux.intel.com>,
	"H. Peter Anvin" <hpa@zytor.com>
Subject: Re: [PATCH V6 3/8] x86/entry: Move PUSH_AND_CLEAR_REGS out of error_entry()
Date: Thu, 28 Apr 2022 12:26:37 +0200	[thread overview]
Message-ID: <Ympr3VU5Q182SVT7@zn.tnic> (raw)
In-Reply-To: <CAJhGHyANWhu-HX20_4XhgACE8tUd8SFa5BTH-ynJscNQ7QxBRw@mail.gmail.com>

On Thu, Apr 28, 2022 at 08:33:36AM +0800, Lai Jiangshan wrote:
> The branch in sync_regs() can be optimized out for the non-XENPV case
> since XENPV doesn't call sync_regs() after patch5 which makes XENPV
> not call error_entry().

I find it confusing why you're even mentioning optimizing a simple test.

That thing is basically a *whopping* two fastpath instructions:

# arch/x86/kernel/traps.c:853: 	if (regs != eregs)
	cmpq	%rdi, %rax	# eregs, <retval>
	je	.L255	#,

Is this such a hot path that it shows in some profiles and removing that
test there is so important?

> The aim of this patch and most of the patchset is to make
> error_entry() be able to be converted to C.  And XENPV cases can
> also be optimized in the patchset although it is not the major main.

So say just that.

> Since this patch adds extra overhead (CALL and RET), the changelog
> has to explain why it is worth it not just for converting ASM to C.

What is the practical relevance of this very minor overhead? Does it
show even in a microbenchmark?

Thx.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette