linux-kernel - Re: Suboptimal inline heuristics due to non-code sections

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20180501134020.fonel3x6plea5xdt@treble>
Date:   Tue, 1 May 2018 08:40:20 -0500
From:   Josh Poimboeuf <jpoimboe@...hat.com>
To:     Nadav Amit <namit@...are.com>
Cc:     Peter Zijlstra <peterz@...radead.org>,
        Ingo Molnar <mingo@...nel.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Thomas Gleixner <tglx@...utronix.de>,
        Linus Torvalds <torvalds@...ux-foundation.org>
Subject: Re: Suboptimal inline heuristics due to non-code sections

On Tue, May 01, 2018 at 06:50:14AM +0000, Nadav Amit wrote:
> When gcc considers the size of a function for inlining decisions, it
> apparently considers *all* sections. Since the kernel extensively uses
> sections for things other than code (e.g., exception-table, bug-table), the
> optimality of these decisions seem questionable to me.
> 
> The objtool’s sections may be the most extreme case, as these sections are
> discarded, while their size still appears to be considered by the inlining
> heuristics. It may be beneficial not to consider (some) the other sections
> as well, as they do not affect code-caching but only increase the kernel
> size.
> 
> To illustrate the issue, consider the function copy_overflow():
> 
>    0xffffffff819315e0 <+0>:	push   %rbp
>    0xffffffff819315e1 <+1>:	mov    %rsi,%rdx
>    0xffffffff819315e4 <+4>:	mov    %edi,%esi
>    0xffffffff819315e6 <+6>:	mov    $0xffffffff820bc4b8,%rdi
>    0xffffffff819315ed <+13>:	mov    %rsp,%rbp
>    0xffffffff819315f0 <+16>:	callq  0xffffffff81089b70 <__warn_printk>
>    0xffffffff819315f5 <+21>:	ud2    
>    0xffffffff819315f7 <+23>:	pop    %rbp
>    0xffffffff819315f8 <+24>:	retq   
> 
> This function seems to me as a great candidate for inlining. Yet, in my 4.16
> build (using gcc 7.2), I get 38 non-inlined instances of this function in
> vmlinux. Forcing CONFIG_STACK_VALIDATION to be disabled reduces the number
> non-inlined instances to 35. Removing, in addition, the data which is saved
> in the __bug_table makes all the instances of the function to be inlined.
> 
> Obviously this certain function can be set as __always_inline, but the inline
> heuristics seems to me as wrongfully biased. 
> 
> What do you think?
> 
> Is there a way to make gcc to ignore sections for its inlining heuristics?

Good find.

Playing around with one of the affected files (crypto/af_alg.o), if I
make the .discard.reachable section empty by removing the text reference
from the annotate_reachable() macro, then copy_overflow() still isn't
inlined.

But if I remove the section completely by removing the
pushsection/popsection, then copy_overflow() gets inlined.

So GCC's inlining decisions are somehow influenced by the existence of
some random empty section.  This definitely seems like a GCC bug to me.

-- 
Josh