diff mbox

[dpdk-dev] mk: fix underlinking issues of most librte libraries

Message ID 5746F682.9070003@intel.com (mailing list archive)
State Not Applicable, archived
Delegated to: Thomas Monjalon
Headers show

Commit Message

Ferruh Yigit May 26, 2016, 1:13 p.m. UTC
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 look 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 <pmatilai@redhat.com>
> 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 only
>> 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 := $(CPU_LDFLAGS)
>>  endif
>>
>> +# Translate DEPDIRS-y into LDLIBS
>> +IGNORE_DEPS = -lrte_eal/% -lrte_net -lrte_compat
>> +_LDDIRS = $(subst librte_ether,libethdev,$(DEPDIRS-y))
>> +_LDDEPS = $(subst lib/lib,-l,$(_LDDIRS))
>> +LDLIBS += $(filter-out $(IGNORE_DEPS), $(_LDDEPS))
>> +
>>  O_TO_A = $(AR) crDs $(LIB) $(OBJS-y)
>>  O_TO_A_STR = $(subst ','\'',$(O_TO_A)) #'# fix syntax highlight
>>  O_TO_A_DISP = $(if $(V),"$(O_TO_A_STR)","  AR $(@)")
>>
>> I'm also sure somebody more familiar with gmake could improve it, but it
>> seems to work quite fine here. I'll wait a bit to see if there are other
>> 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 fixed
[2] [3].

But this cause a compilation error for cmdline_test [4]. This is because
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 error
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?


Thanks,
ferruh

---

[1]

Comments

Panu Matilainen May 27, 2016, 5:59 a.m. UTC | #1
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 look 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 <pmatilai@redhat.com>
>> 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 only
>>> 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 := $(CPU_LDFLAGS)
>>>  endif
>>>
>>> +# Translate DEPDIRS-y into LDLIBS
>>> +IGNORE_DEPS = -lrte_eal/% -lrte_net -lrte_compat
>>> +_LDDIRS = $(subst librte_ether,libethdev,$(DEPDIRS-y))
>>> +_LDDEPS = $(subst lib/lib,-l,$(_LDDIRS))
>>> +LDLIBS += $(filter-out $(IGNORE_DEPS), $(_LDDEPS))
>>> +
>>>  O_TO_A = $(AR) crDs $(LIB) $(OBJS-y)
>>>  O_TO_A_STR = $(subst ','\'',$(O_TO_A)) #'# fix syntax highlight
>>>  O_TO_A_DISP = $(if $(V),"$(O_TO_A_STR)","  AR $(@)")
>>>
>>> I'm also sure somebody more familiar with gmake could improve it, but it
>>> seems to work quite fine here. I'll wait a bit to see if there are other
>>> 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 fixed
> [2] [3].
>
> But this cause a compilation error for cmdline_test [4]. This is because
> 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 error
> 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?

Cyclic dependencies are nasty, eliminating it would be good.

But in the meanwhile it can be worked around in a generic manner by 
handling it in the libdpdk.so linker script, see 
http://dpdk.org/ml/archives/dev/2016-May/039413.html

	- Panu -

>
> 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  = -pthread
>  endif
>
>  # Workaround lack of DT_NEEDED entry
> -EXECENV_LDFLAGS = --no-as-needed
> +EXECENV_LDFLAGS = --as-needed
>
>  EXECENV_LDLIBS  =
>  EXECENV_ASFLAGS =
>
>
> [2]
> # ldd build/app/testpmd
>         linux-vdso.so.1 (0x00007ffcff92e000)
>         librte_mbuf.so.2.1 => .../build/lib/librte_mbuf.so.2.1
>         libethdev.so.4.1 => .../build/lib/libethdev.so.4.1
>         librte_mempool.so.2.1 => .../build/lib/librte_mempool.so.2.1
>         librte_ring.so.1.1 => .../build/lib/librte_ring.so.1.1
>         librte_eal.so.2.1 => .../build/lib/librte_eal.so.2.1
>         librte_cmdline.so.2.1 => .../build/lib/librte_cmdline.so.2.1
>         librte_pmd_bond.so.1.1 => .../build/lib/librte_pmd_bond.so.1.1
>         libpthread.so.0 => /lib64/libpthread.so.0 (0x00007f6879ff1000)
>         libc.so.6 => /lib64/libc.so.6 (0x00007f6879c30000)
>         /lib64/ld-linux-x86-64.so.2 (0x0000564998dbd000)
>         libdl.so.2 => /lib64/libdl.so.2 (0x00007f6879a2c000)
>         libgcc_s.so.1 => /lib64/libgcc_s.so.1 (0x00007f6879814000)
>         librt.so.1 => /lib64/librt.so.1 (0x00007f687960c000)
>         librte_kvargs.so.1.1 => .../build/lib/librte_kvargs.so.1.1
>
> [3]
> # ldd -u -r build/app/testpmd
> #
>
>
> [4]
> == Build app/cmdline_test
>   CC cmdline_test.o
>   CC commands.o
>   LD cmdline_test
> .../build/lib/librte_eal.so: undefined reference to `rte_mempool_lookup'
> .../build/lib/librte_eal.so: undefined reference to `rte_mempool_create'
> 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 += commands.c
>
>  CFLAGS += -O3
>  CFLAGS += $(WERROR_FLAGS)
> +LDFLAGS += --no-as-needed
>
>  # this application needs libraries first
>  DEPDIRS-y += lib drivers
>
Ferruh Yigit May 27, 2016, 4:15 p.m. UTC | #2
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 look 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 <pmatilai@redhat.com>
>>> 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 only
>>>> 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 := $(CPU_LDFLAGS)
>>>>  endif
>>>>
>>>> +# Translate DEPDIRS-y into LDLIBS
>>>> +IGNORE_DEPS = -lrte_eal/% -lrte_net -lrte_compat
>>>> +_LDDIRS = $(subst librte_ether,libethdev,$(DEPDIRS-y))
>>>> +_LDDEPS = $(subst lib/lib,-l,$(_LDDIRS))
>>>> +LDLIBS += $(filter-out $(IGNORE_DEPS), $(_LDDEPS))
>>>> +
>>>>  O_TO_A = $(AR) crDs $(LIB) $(OBJS-y)
>>>>  O_TO_A_STR = $(subst ','\'',$(O_TO_A)) #'# fix syntax highlight
>>>>  O_TO_A_DISP = $(if $(V),"$(O_TO_A_STR)","  AR $(@)")
>>>>
>>>> I'm also sure somebody more familiar with gmake could improve it, but it
>>>> seems to work quite fine here. I'll wait a bit to see if there are other
>>>> 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 fixed
>> [2] [3].
>>
>> But this cause a compilation error for cmdline_test [4]. This is because
>> 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 error
>> 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?
> 
> Cyclic dependencies are nasty, eliminating it would be good.
> 
> But in the meanwhile it can be worked around in a generic manner by 
> handling it in the libdpdk.so linker script, see 
> http://dpdk.org/ml/archives/dev/2016-May/039413.html
> 

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 -
> 
>>
>> 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  = -pthread
>>  endif
>>
>>  # Workaround lack of DT_NEEDED entry
>> -EXECENV_LDFLAGS = --no-as-needed
>> +EXECENV_LDFLAGS = --as-needed
>>
>>  EXECENV_LDLIBS  =
>>  EXECENV_ASFLAGS =
>>
>>
>> [2]
>> # ldd build/app/testpmd
>>         linux-vdso.so.1 (0x00007ffcff92e000)
>>         librte_mbuf.so.2.1 => .../build/lib/librte_mbuf.so.2.1
>>         libethdev.so.4.1 => .../build/lib/libethdev.so.4.1
>>         librte_mempool.so.2.1 => .../build/lib/librte_mempool.so.2.1
>>         librte_ring.so.1.1 => .../build/lib/librte_ring.so.1.1
>>         librte_eal.so.2.1 => .../build/lib/librte_eal.so.2.1
>>         librte_cmdline.so.2.1 => .../build/lib/librte_cmdline.so.2.1
>>         librte_pmd_bond.so.1.1 => .../build/lib/librte_pmd_bond.so.1.1
>>         libpthread.so.0 => /lib64/libpthread.so.0 (0x00007f6879ff1000)
>>         libc.so.6 => /lib64/libc.so.6 (0x00007f6879c30000)
>>         /lib64/ld-linux-x86-64.so.2 (0x0000564998dbd000)
>>         libdl.so.2 => /lib64/libdl.so.2 (0x00007f6879a2c000)
>>         libgcc_s.so.1 => /lib64/libgcc_s.so.1 (0x00007f6879814000)
>>         librt.so.1 => /lib64/librt.so.1 (0x00007f687960c000)
>>         librte_kvargs.so.1.1 => .../build/lib/librte_kvargs.so.1.1
>>
>> [3]
>> # ldd -u -r build/app/testpmd
>> #
>>
>>
>> [4]
>> == Build app/cmdline_test
>>   CC cmdline_test.o
>>   CC commands.o
>>   LD cmdline_test
>> .../build/lib/librte_eal.so: undefined reference to `rte_mempool_lookup'
>> .../build/lib/librte_eal.so: undefined reference to `rte_mempool_create'
>> 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 += commands.c
>>
>>  CFLAGS += -O3
>>  CFLAGS += $(WERROR_FLAGS)
>> +LDFLAGS += --no-as-needed
>>
>>  # this application needs libraries first
>>  DEPDIRS-y += lib drivers
>>
>
diff mbox

Patch

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  = -pthread
 endif

 # Workaround lack of DT_NEEDED entry
-EXECENV_LDFLAGS = --no-as-needed
+EXECENV_LDFLAGS = --as-needed

 EXECENV_LDLIBS  =
 EXECENV_ASFLAGS =


[2]
# ldd build/app/testpmd
        linux-vdso.so.1 (0x00007ffcff92e000)
        librte_mbuf.so.2.1 => .../build/lib/librte_mbuf.so.2.1
        libethdev.so.4.1 => .../build/lib/libethdev.so.4.1
        librte_mempool.so.2.1 => .../build/lib/librte_mempool.so.2.1
        librte_ring.so.1.1 => .../build/lib/librte_ring.so.1.1
        librte_eal.so.2.1 => .../build/lib/librte_eal.so.2.1
        librte_cmdline.so.2.1 => .../build/lib/librte_cmdline.so.2.1
        librte_pmd_bond.so.1.1 => .../build/lib/librte_pmd_bond.so.1.1
        libpthread.so.0 => /lib64/libpthread.so.0 (0x00007f6879ff1000)
        libc.so.6 => /lib64/libc.so.6 (0x00007f6879c30000)
        /lib64/ld-linux-x86-64.so.2 (0x0000564998dbd000)
        libdl.so.2 => /lib64/libdl.so.2 (0x00007f6879a2c000)
        libgcc_s.so.1 => /lib64/libgcc_s.so.1 (0x00007f6879814000)
        librt.so.1 => /lib64/librt.so.1 (0x00007f687960c000)
        librte_kvargs.so.1.1 => .../build/lib/librte_kvargs.so.1.1

[3]
# ldd -u -r build/app/testpmd
#


[4]
== Build app/cmdline_test
  CC cmdline_test.o
  CC commands.o
  LD cmdline_test
.../build/lib/librte_eal.so: undefined reference to `rte_mempool_lookup'
.../build/lib/librte_eal.so: undefined reference to `rte_mempool_create'
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 += commands.c

 CFLAGS += -O3
 CFLAGS += $(WERROR_FLAGS)
+LDFLAGS += --no-as-needed

 # this application needs libraries first
 DEPDIRS-y += lib drivers