From mboxrd@z Thu Jan 1 00:00:00 1970 From: Ferruh Yigit Subject: Re: [PATCH] mk: fix underlinking issues of most librte libraries Date: Fri, 27 May 2016 17:15:47 +0100 Message-ID: <574872B3.6040702@intel.com> References: <1463763022-6673-1-git-send-email-christian.ehrhardt@canonical.com> <7920424.umDpp4ukym@xps13> <78bbf74c-9621-9094-a280-76ed9fc88393@redhat.com> <5746F682.9070003@intel.com> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: quoted-printable Cc: Thomas Monjalon , dev To: Panu Matilainen , Christian Ehrhardt Return-path: Received: from mga11.intel.com (mga11.intel.com [192.55.52.93]) by dpdk.org (Postfix) with ESMTP id 70EC42E8E for ; Fri, 27 May 2016 18:15:50 +0200 (CEST) In-Reply-To: List-Id: patches and discussions about DPDK List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.org Sender: "dev" On 5/27/2016 6:59 AM, Panu Matilainen wrote: > On 05/26/2016 04:13 PM, Ferruh Yigit wrote: >> On 5/24/2016 11:11 AM, Christian Ehrhardt wrote: >>> Hi Panu, >>> I already agreed to Thomas on IRC but won't have time until next week= . >>> Thanks for making a patch that does that already - I'll give it a loo= k and >>> some test on my end next week then. >>> >>> >>> >>> Christian Ehrhardt >>> Software Engineer, Ubuntu Server >>> Canonical Ltd >>> >>> On Tue, May 24, 2016 at 11:56 AM, Panu Matilainen >>> wrote: >>> >>>> On 05/20/2016 08:08 PM, Thomas Monjalon wrote: >>>> >>>>> 2016-05-20 18:50, Christian Ehrhardt: >>>>> >>>>>> The individual libraries have various cross dependencies. >>>>>> This is already refelcted in the DEPDIR dependency, but not yet in >>>>>> proper DT_NEEDED flags in the .so's. >>>>>> This adds the -l flags so that is properly stored in the .so's ELF >>>>>> headers. >>>>>> >>>>> >>>>> Why not filling LDLIBS by parsing DEPDIRS-y in rte.lib.mk? >>>>> >>>>> >>>> A fair question :) The thought has passed my mind too, but I thought= it'd >>>> be too messy with differing library vs directory names and other >>>> exceptions. But in reality manually maintained separate LDLIBS is on= ly >>>> going to get out of sync sooner or later, and turns out its not that= bad >>>> anyway: >>>> >>>> diff --git a/mk/rte.lib.mk b/mk/rte.lib.mk >>>> index b420280..88b4e98 100644 >>>> --- a/mk/rte.lib.mk >>>> +++ b/mk/rte.lib.mk >>>> @@ -77,6 +77,12 @@ else >>>> _CPU_LDFLAGS :=3D $(CPU_LDFLAGS) >>>> endif >>>> >>>> +# Translate DEPDIRS-y into LDLIBS >>>> +IGNORE_DEPS =3D -lrte_eal/% -lrte_net -lrte_compat >>>> +_LDDIRS =3D $(subst librte_ether,libethdev,$(DEPDIRS-y)) >>>> +_LDDEPS =3D $(subst lib/lib,-l,$(_LDDIRS)) >>>> +LDLIBS +=3D $(filter-out $(IGNORE_DEPS), $(_LDDEPS)) >>>> + >>>> O_TO_A =3D $(AR) crDs $(LIB) $(OBJS-y) >>>> O_TO_A_STR =3D $(subst ','\'',$(O_TO_A)) #'# fix syntax highlight >>>> O_TO_A_DISP =3D $(if $(V),"$(O_TO_A_STR)"," AR $(@)") >>>> >>>> I'm also sure somebody more familiar with gmake could improve it, bu= t it >>>> seems to work quite fine here. I'll wait a bit to see if there are o= ther >>>> suggestions before sending it as a proper patch. >>>> >>>> - Panu - >>>> >> >> I did test the patch with minor diff in rte.vars.mk [1], otherwise all >> libraries kept needed in final library. >> >> Patch itself looks good, overlinking problem in final binary seems fix= ed >> [2] [3]. >> >> But this cause a compilation error for cmdline_test [4]. This is becau= se >> of cyclic dependency between librte_eal <-> librte_mempool. >> >> Other applications don't have this compilation error because they use >> librte_mempool, and since it is linked against application, this erro= r >> not observed. cmdline_test itself doesn't have any direct call to >> librte_mempool. >> >> It is possible to add workaround to cmdline_test [5], (introduce >> overlinking back), but should we attack on cyclic dependency instead? >=20 > Cyclic dependencies are nasty, eliminating it would be good. >=20 > But in the meanwhile it can be worked around in a generic manner by=20 > handling it in the libdpdk.so linker script, see=20 > http://dpdk.org/ml/archives/dev/2016-May/039413.html >=20 Won't linker script has same problem, unless it is possible to select some libraries to force link and some other to as-needed, than this can be workaround, but I think using "--as-needed" as linker argument and applying workaround to single test application is better. Also when overlinking removed, getting compilation error for lpcap too. Since -lpcap argument is before -lrte_pmd_pcap, symbols used by rte_pmd_pcap not resolved, increasing --start-group scope fixes this. Not related to overlinking, but while updating linker flags, as far as I understand --whole-archive argument is only necessary for static pmd libraries, but currently has wider scope, reducing its scope to pmd slightly reduces final library sizes. I will send a patch on top of your patch, with above updates. Thanks, ferruh > - Panu - >=20 >> >> Thanks, >> ferruh >> >> --- >> >> [1] >> diff --git a/mk/exec-env/linuxapp/rte.vars.mk >> b/mk/exec-env/linuxapp/rte.vars.mk >> index d51bd17..10d37d5 100644 >> --- a/mk/exec-env/linuxapp/rte.vars.mk >> +++ b/mk/exec-env/linuxapp/rte.vars.mk >> @@ -46,7 +46,7 @@ EXECENV_CFLAGS =3D -pthread >> endif >> >> # Workaround lack of DT_NEEDED entry >> -EXECENV_LDFLAGS =3D --no-as-needed >> +EXECENV_LDFLAGS =3D --as-needed >> >> EXECENV_LDLIBS =3D >> EXECENV_ASFLAGS =3D >> >> >> [2] >> # ldd build/app/testpmd >> linux-vdso.so.1 (0x00007ffcff92e000) >> librte_mbuf.so.2.1 =3D> .../build/lib/librte_mbuf.so.2.1 >> libethdev.so.4.1 =3D> .../build/lib/libethdev.so.4.1 >> librte_mempool.so.2.1 =3D> .../build/lib/librte_mempool.so.2.1 >> librte_ring.so.1.1 =3D> .../build/lib/librte_ring.so.1.1 >> librte_eal.so.2.1 =3D> .../build/lib/librte_eal.so.2.1 >> librte_cmdline.so.2.1 =3D> .../build/lib/librte_cmdline.so.2.1 >> librte_pmd_bond.so.1.1 =3D> .../build/lib/librte_pmd_bond.so.1= .1 >> libpthread.so.0 =3D> /lib64/libpthread.so.0 (0x00007f6879ff100= 0) >> libc.so.6 =3D> /lib64/libc.so.6 (0x00007f6879c30000) >> /lib64/ld-linux-x86-64.so.2 (0x0000564998dbd000) >> libdl.so.2 =3D> /lib64/libdl.so.2 (0x00007f6879a2c000) >> libgcc_s.so.1 =3D> /lib64/libgcc_s.so.1 (0x00007f6879814000) >> librt.so.1 =3D> /lib64/librt.so.1 (0x00007f687960c000) >> librte_kvargs.so.1.1 =3D> .../build/lib/librte_kvargs.so.1.1 >> >> [3] >> # ldd -u -r build/app/testpmd >> # >> >> >> [4] >> =3D=3D Build app/cmdline_test >> CC cmdline_test.o >> CC commands.o >> LD cmdline_test >> .../build/lib/librte_eal.so: undefined reference to `rte_mempool_looku= p' >> .../build/lib/librte_eal.so: undefined reference to `rte_mempool_creat= e' >> collect2: error: ld returned 1 exit status >> .../mk/rte.app.mk:243: recipe for target 'cmdline_test' failed >> make[3]: *** [cmdline_test] Error 1 >> >> >> [5] >> diff --git a/app/cmdline_test/Makefile b/app/cmdline_test/Makefile >> index c6169f5..69ebfb2 100644 >> --- a/app/cmdline_test/Makefile >> +++ b/app/cmdline_test/Makefile >> @@ -46,6 +46,7 @@ SRCS-y +=3D commands.c >> >> CFLAGS +=3D -O3 >> CFLAGS +=3D $(WERROR_FLAGS) >> +LDFLAGS +=3D --no-as-needed >> >> # this application needs libraries first >> DEPDIRS-y +=3D lib drivers >> >=20