From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753512AbaBNV1x (ORCPT ); Fri, 14 Feb 2014 16:27:53 -0500 Received: from mga11.intel.com ([192.55.52.93]:38604 "EHLO mga11.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752583AbaBNVWa (ORCPT ); Fri, 14 Feb 2014 16:22:30 -0500 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="4.95,847,1384329600"; d="scan'208";a="481803176" From: Andi Kleen To: linux-kernel@vger.kernel.org Cc: x86@kernel.org, linux-kbuild@vger.kernel.org, mmarek@suse.cz, Andi Kleen Subject: [PATCH 15/19] Kbuild, lto: Fix single pass kallsyms for LTO Date: Fri, 14 Feb 2014 22:21:39 +0100 Message-Id: <1392412903-25733-16-git-send-email-andi@firstfloor.org> X-Mailer: git-send-email 1.8.5.2 In-Reply-To: <1392412903-25733-1-git-send-email-andi@firstfloor.org> References: <1392412903-25733-1-git-send-email-andi@firstfloor.org> Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org From: Andi Kleen gcc-nm on slim LTO objects does not output static functions or variables. This causes the first pass estimation of the kallsyms table to be off too much. Add a hack using the LTO function sections to retrieve all functions instead. I wrote that hack in perl, as it exceeded my awk-fu (this adds a build dependency on perl, but only if LTO is active). The hack also doesn't support variables, but we handle that by disable KALLSYMS_ALL with LTO. The hack is also somewhat depending on the internal LTO object format. Hopefully at some future point gcc-nm will be fixed and this won't be necessary anymore. One issue is that LTO can generate new symbols in the final link, for example when cloning functions. These clones are just copies of existing names with a postfix (which we remove), so they compress well in the kallsyms compression. With that we can get away by just adding a 5% safety factor to the first pass kallsyms estimation, and hope all clones fit into that. If some obscure build has more clones than that the PAD_RATIO value in kallsyms.c can be lowered to increase it. Signed-off-by: Andi Kleen --- init/Kconfig | 5 ++++- scripts/link-vmlinux.sh | 24 +++++++++++++++++++++--- 2 files changed, 25 insertions(+), 4 deletions(-) diff --git a/init/Kconfig b/init/Kconfig index 009a797..ea00c0e 100644 --- a/init/Kconfig +++ b/init/Kconfig @@ -1317,7 +1317,10 @@ config KALLSYMS config KALLSYMS_ALL bool "Include all symbols in kallsyms" - depends on DEBUG_KERNEL && KALLSYMS + # the method LTO uses to predict the symbol table + # only supports functions for now + # This can be removed once http://gcc.gnu.org/PR60016 is fixed + depends on DEBUG_KERNEL && KALLSYMS && !LTO help Normally kallsyms only contains the symbols of functions for nicer OOPS messages and backtraces (i.e., symbols from the text and inittext diff --git a/scripts/link-vmlinux.sh b/scripts/link-vmlinux.sh index b299fdd..5a28f2a 100644 --- a/scripts/link-vmlinux.sh +++ b/scripts/link-vmlinux.sh @@ -90,10 +90,28 @@ kallsyms() local aflags="${KBUILD_AFLAGS} ${KBUILD_AFLAGS_KERNEL} \ ${NOSTDINC_FLAGS} ${LINUXINCLUDE} ${KBUILD_CPPFLAGS}" - ${NM} -n ${1} | \ - awk 'NF == 3 { print}' | - scripts/kallsyms ${kallsymopt} | \ + # workaround for slim LTO gcc-nm not outputing static symbols + # http://gcc.gnu.org/PR60016 + # generate a fake symbol table based on the LTO function sections. + # This unfortunately "knows" about the internal LTO file format + # and only works for functions + # needs perl for now when building for LTO + ( + if $OBJDUMP --section-headers ${1} | grep -q \.gnu\.lto_ ; then + ${OBJDUMP} --section-headers ${1} | + perl -ne ' +@n = split; +next unless $n[1] =~ /\.gnu\.lto_([_a-zA-Z][^.]+)/; +next if $n[1] eq $prev; +$prev = $n[1]; +print "0 T ",$1,"\n"' + fi + ${NM} -n ${1} | awk 'NF == 3 { print }' + ) > ${2}_sym + # run without pipe to make kallsyms errors stop the script + ./scripts/kallsyms ${kallsymopt} < ${2}_sym | ${CC} ${aflags} -c -o ${2} -x assembler-with-cpp - + } # Create map file with all symbols from ${1} -- 1.8.5.2