Re: [PATCH 4/7] module: avoid code duplication in include/linux/export.h

From: Jessica Yu <jeyu@kernel.org>
To: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Cc: Masahiro Yamada <yamada.masahiro@socionext.com>,
	Matthias Maennich <maennich@google.com>,
	Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
	Joel Fernandes <joel@joelfernandes.org>,
	Martijn Coenen <maco@android.com>,
	Will Deacon <will.deacon@arm.com>, Will Deacon <will@kernel.org>,
	linux-kernel@vger.kernel.org
Subject: Re: [PATCH 4/7] module: avoid code duplication in include/linux/export.h
Date: Thu, 31 Oct 2019 11:13:06 +0100	[thread overview]
Message-ID: <20191031101306.GA2177@linux-8ccs> (raw)
In-Reply-To: <a2e6bdc2-3d35-a3dc-13ef-1ce32f77ef17@rasmusvillemoes.dk>

+++ Rasmus Villemoes [29/10/19 22:11 +0100]:
>On 29/10/2019 20.19, Jessica Yu wrote:
>> +++ Rasmus Villemoes [27/09/19 13:07 +0200]:
>>> On 27/09/2019 11.36, Masahiro Yamada wrote:
>>>>
>>>> A typical kernel configuration has 10K exported symbols, so it
>>>> increases 10KB in rough estimation.
>>>>
>>>> I did not come up with a good idea to refactor it without increasing
>>>> the code size.
>>>
>>> Can't we put the "aMS" flags on the __ksymtab_strings section? That
>>> would make the empty strings free, and would also deduplicate the
>>> USB_STORAGE string. And while almost per definition we don't have exact
>>> duplicates among the names of exported symbols, we might have both a foo
>>> and __foo, so that could save even more.
>>>
>>> I don't know if we have it already, but we'd need each arch to tell us
>>> what symbol to use for @ in @progbits (e.g. % for arm). It seems most
>>> are fine with @, so maybe a generic version could be
>>>
>>> #ifndef ARCH_SECTION_TYPE_CHAR
>>> #define ARCH_SECTION_TYPE_CHAR "@"
>>> #endif
>>>
>>> and then it would be
>>> section("__ksymtab_strings,\"aMS\","ARCH_SECTION_TYPE_CHAR"progbits,1")
>>
>> FWIW, I've just tinkered with this, and unfortunately the strings
>> don't get deduplicated for kernel modules :-(
>>
>> Apparently ld does not do the deduplication for SHF_MERGE|SHF_STRINGS
>> sections for relocatable files (ld -r), which kernel modules are. See:
>>
>>    https://sourceware.org/ml/binutils/2009-07/msg00291.html
>
>I know <https://patches-gcc.linaro.org/patch/5858/> :)

That is exactly what we need! :)

>> But, the strings do get deduplicated for vmlinux. Not sure if we can
>> find a workaround for modules or if the benefit is significant enough
>> if it only for vmlinux.
>
>I think it's definitely worth if, even if it "only" benefits vmlinux for
>now. And I still hope to revisit the --force-section-merge some day, but
>it's very far down my priority list.

Yeah, I think it's worth having too.

If you don't have any extra cycles at the moment, and it's far down
your priority list, do you mind if I take a look and maybe try to push
that patch of yours upstream again? I don't know how successful I'd
be, but now since it's especially relevant for namespaces, it's
definitely worth looking at again.

Thanks!

Jessica