From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1756483AbYDNLZL (ORCPT ); Mon, 14 Apr 2008 07:25:11 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1752951AbYDNLY7 (ORCPT ); Mon, 14 Apr 2008 07:24:59 -0400 Received: from fk-out-0910.google.com ([209.85.128.189]:31394 "EHLO fk-out-0910.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752919AbYDNLY6 (ORCPT ); Mon, 14 Apr 2008 07:24:58 -0400 DomainKey-Signature: a=rsa-sha1; c=nofws; d=gmail.com; s=gamma; h=message-id:date:from:user-agent:mime-version:to:cc:subject:references:in-reply-to:content-type:content-transfer-encoding; b=l2boMnyEcLG4SUeTcnAvGQM2g5jhtvybEx4GViOce1gI4HQmFEv/7XtguU+LajiQYPeMTU+K5QSfd4HQ1/yfWx/jlpa/wLUdmVgKoq+tmQj5RRbuzISvzJkAvU2KpimobRiH2AIrmUV/CRtAvjmDp2dEXDIE5/xzmzbnsBAH4DA= Message-ID: <48033E2C.2080504@gmail.com> Date: Mon, 14 Apr 2008 13:21:16 +0200 From: Jacek Luczak User-Agent: Thunderbird 2.0.0.9 (X11/20071031) MIME-Version: 1.0 To: Sam Ravnborg CC: Ingo Molnar , LKML Subject: Re: [PATCH] x86: pgtable_32.h - prototype and section mismatch fixes References: <48022990.3060104@gmail.com> <20080414072645.GG16163@elte.hu> <20080414084127.GA11372@uranus.ravnborg.org> <20080414085307.GG19865@elte.hu> <20080414085927.GC11372@uranus.ravnborg.org> <48031FA7.9050803@gmail.com> <20080414095259.GA11966@uranus.ravnborg.org> In-Reply-To: <20080414095259.GA11966@uranus.ravnborg.org> Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Sam Ravnborg pisze: > On Mon, Apr 14, 2008 at 11:11:03AM +0200, Jacek Luczak wrote: >> Sam Ravnborg pisze: >>> On Mon, Apr 14, 2008 at 10:53:07AM +0200, Ingo Molnar wrote: >>>> * Sam Ravnborg wrote: >>>> >>>>>> hm, that's an interesting case: we need those annotations probably >>>>>> because gcc decided to not inline those functions. (this is possible >>>>>> via the new CONFIG_OPTIMIZE_INLINING=y option) Sam, what's your take >>>>>> on that? >>>>> gcc uses different heuristics for inlining between the different >>>>> versions. Therefore to achieve somehow predictable results I added >>>>> -fno-inline-functions-called-once when CONFIG_DEBUG_SECTION_MISMATCH >>>>> is enabled. >>>>> >>>>> So in the above case for any normal kernel build we would see that gcc >>>>> inlined the above and everything is fine. But for the >>>>> CONFIG_DEBUG_SECTION_MISMTCH cases we do not inline and thus we see >>>>> that we have a section mismatch. >>>> ah, ok. So i guess this will result in a few isolated cases of __init >>>> annotations added to inline functions - Jacek fixed one such case - but >>>> it should not result in the general spreading of __init annotations to >>>> inline functions, correct? (which i was worried about) >>> I do not think so. The need for small isolated inline functions >>> in the init paths are minimal and last I did a section mismatch free >>> sweep on the kernel it was only few if any inline functions(*) I had >>> to annotate. >>> >>> (*) Considering only the minimal amount of function that ought >>> to be annotated inlined. >> There's a lot of inline __init functions already - which, on the other hand >> should not result in more of such. Attached you can find grep on tree which >> shows all inline __init functions (you won't find my paravirt_pagetable_setup_* >> functions in output as I removed __init for test). > > Most of the functionas listed should never have been annotated inline. > If you filter aways the functions that are in one one of their > incarnations longer than 8 lines or are located in a .c file the list > is much shorter. This is the real candidates (the figure '8' is just bad way > to filter away 'complex' functions). > Your paravirt_pagetable_setup_done is a real one btw. It's worth of closer investigation I think. >> Sam, do -fno-inline-functions-called-once could affect >> paravirt_pagetable_setup_done? In general there was no section mismatch warning >> for this function, but I've annotated it also. > Most liekly yes - but I did not check. GCC is ok, I mean it inlines paravirt_pagetable_setup_[start,done] even with CONFIG_OPTIMIZE_INLINING=y: $ objdump -t vmlinux.o | grep pagetable_setup_start 0000a418 g F .init.text 000000a1 native_pagetable_setup_start While running with CONFIG_DEBUG_SECTION_MISMTCH=y it does not inline paravirt_pagetable_setup_start (_done is OK): $ objdump -t vmlinux.o | grep pagetable_setup_start 00017100 l F .text 0000000b paravirt_pagetable_setup_start 00009fb0 g F .init.text 00000089 native_pagetable_setup_start DEBUG_SECTION_MISMATCH probably generated mismatch with smpboot_setup_io_apic() (commit: 96c968742fa1e6d64f979464acf2fd90cdc117b3, already merged). I will check and if __init is really superfluous I will post delta patch to remove all those __init annotations. Similar situation might be found with unlock_ExtINT_logic() (arch/x86/kernel/io_apic_32.c), which I annotated with __init in one of previous patches, but here, this function IMO shouldn't be marked inline. -Jacek