Re: [PATCH] x86/uaccess: Remove redundant likely/unlikely annotations

From: Randy Dunlap <rdunlap@infradead.org>
To: Josh Poimboeuf <jpoimboe@redhat.com>,
	Peter Zijlstra <peterz@infradead.org>
Cc: Christoph Hellwig <hch@lst.de>,
	Andrew Morton <akpm@linux-foundation.org>,
	broonie@kernel.org, linux-fsdevel@vger.kernel.org,
	linux-kernel@vger.kernel.org, linux-mm@kvack.org,
	linux-next@vger.kernel.org, mhocko@suse.cz,
	mm-commits@vger.kernel.org, sfr@canb.auug.org.au,
	Linus Torvalds <torvalds@linux-foundation.org>,
	viro@zeniv.linux.org.uk, x86@kernel.org,
	Steven Rostedt <rostedt@goodmis.org>
Subject: Re: [PATCH] x86/uaccess: Remove redundant likely/unlikely annotations
Date: Fri, 29 May 2020 11:29:14 -0700	[thread overview]
Message-ID: <77d5741f-4b84-4cb8-1b01-3e411d3b8a70@infradead.org> (raw)
In-Reply-To: <20200529172505.fdjppgquujab7ayv@treble>

On 5/29/20 10:25 AM, Josh Poimboeuf wrote:
> On Fri, May 29, 2020 at 06:54:19PM +0200, Peter Zijlstra wrote:
>> On Fri, May 29, 2020 at 11:50:11AM -0500, Josh Poimboeuf wrote:
>>> The nested likelys seem like overkill anyway -- user_access_begin() is
>>> __always_inline and it already has unlikely(), which should be
>>> propagated.
>>>
>>> So just remove the outer likelys?
>>
>> That fixes it. Ack!
> 
> If there are no objections to the patch, I can add it to my objtool-core
> branch unless anybody else wants to take it.  It only affects
> linux-next.
> 
> ---8<---
> 
> From: Josh Poimboeuf <jpoimboe@redhat.com>
> Subject: [PATCH] x86/uaccess: Remove redundant likely/unlikely annotations
> 
> Since user_access_begin() already has an unlikely() annotation for its
> access_ok() check, "if (likely(user_access_begin))" results in nested
> likely annotations.  When combined with CONFIG_TRACE_BRANCH_PROFILING,
> GCC converges the error/success paths of the nested ifs, using a
> register value to distinguish between them.
> 
> While the code is technically uaccess safe, it complicates the
> branch-profiling generated code.  It also confuses objtool, because it
> doesn't do register value tracking, resulting in the following warnings:
> 
>   arch/x86/lib/csum-wrappers_64.o: warning: objtool: csum_and_copy_from_user()+0x2a4: call to memset() with UACCESS enabled
>   arch/x86/lib/csum-wrappers_64.o: warning: objtool: csum_and_copy_to_user()+0x243: return with UACCESS enabled
> 
> The outer likely annotations aren't actually needed anyway, since the
> compiler propagates the error path coldness when it inlines
> user_access_begin().
> 
> Fixes: 18372ef87665 ("x86_64: csum_..._copy_..._user(): switch to unsafe_..._user()")
> Reported-by: Randy Dunlap <rdunlap@infradead.org>
> Acked-by: Peter Zijlstra <peterz@infradead.org>
> Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com>

Acked-by: Randy Dunlap <rdunlap@infradead.org> # build-tested

Thanks.

> ---
>  arch/x86/lib/csum-wrappers_64.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/x86/lib/csum-wrappers_64.c b/arch/x86/lib/csum-wrappers_64.c
> index a12b8629206d..ee63d7576fd2 100644
> --- a/arch/x86/lib/csum-wrappers_64.c
> +++ b/arch/x86/lib/csum-wrappers_64.c
> @@ -27,7 +27,7 @@ csum_and_copy_from_user(const void __user *src, void *dst,
>  	might_sleep();
>  	*errp = 0;
>  
> -	if (!likely(user_access_begin(src, len)))
> +	if (!user_access_begin(src, len))
>  		goto out_err;
>  
>  	/*
> @@ -89,7 +89,7 @@ csum_and_copy_to_user(const void *src, void __user *dst,
>  
>  	might_sleep();
>  
> -	if (unlikely(!user_access_begin(dst, len))) {
> +	if (!user_access_begin(dst, len)) {
>  		*errp = -EFAULT;
>  		return 0;
>  	}
> 

-- 
~Randy