openssl: aarch64: configure with no-asm
PATCH v2 1/1 openssl: aarch64: configure with no-asm
From: Max Krummenacher <[email protected]>
openssl has a couple of functionalities which use optimized assembler code. With no-asm these are replaced by generic C code implementation.
With GCC 15 OE by default uses the new AARCH64 specific GCS feature. However the object files produced by the assembler are not marked to provide the feature and consequently the produced shared objects containing them also not.
The linker will warn when linking against such shared objects files and, with all warnings are errors set linking will fail. Without this patch systemd e.g. fails do_compile as it links against libcrypto.so from openssl.
To test what features an object file (or .so, .a file) has use readelf -ln and check the content of .note.gnu.property, e.g. libcrypto-shlib-md5-aarch64.o build from assembler source, libcrypto-shlib-md5_one.o build from C source:
$ aarch64-poky-linux-readelf -ln ./crypto/md5/libcrypto-shlib-md5-aarch64.o
There are no program headers in this file.
Displaying notes found in: .note.gnu.property
Owner Data size Description GNU 0x00000010 NT_GNU_PROPERTY_TYPE_0
Properties: AArch64 feature: BTI, PAC
$ aarch64-poky-linux-readelf -ln ./crypto/md5/libcrypto-shlib-md5_one.o
There are no program headers in this file.
Displaying notes found in: .note.gnu.property
Owner Data size Description GNU 0x00000010 NT_GNU_PROPERTY_TYPE_0
Properties: AArch64 feature: BTI, PAC, GCS
--- meta/recipes-connectivity/openssl/openssl_3.5.0.bb | 2 ++ 1 file changed, 2 insertions(+)
For reference also refer to this thread:
https://lore.kernel.org/all/aBilFkr4HF-MSBd6@toolbox/
meta/recipes-connectivity/openssl/openssl_3.5.0.bb +2 -0
@@ -36,6 +36,8 @@ B = "${WORKDIR}/build" do_configure[cleandirs] = "${B}" EXTRA_OECONF = "${@bb.utils.contains('PTEST_ENABLED', '1', '', 'no-tests', d)}" +# aarch64 'as' doesn't mark objects to have the GCS feature, resulting *.so then produces linker warnings +EXTRA_OECONF:append:aarch64 = " no-asm" #| ./libcrypto.so: undefined reference to `getcontext' #| ./libcrypto.so: undefined reference to `setcontext'
Signature · 1 line
2.42.0
Quoted 30 lines · Max Krummenacher via lists.openembedded.org
From: Max Krummenacher <[email protected]>
openssl has a couple of functionalities which use optimized assembler code. With no-asm these are replaced by generic C code implementation.
With GCC 15 OE by default uses the new AARCH64 specific GCS feature. However the object files produced by the assembler are not marked to provide the feature and consequently the produced shared objects containing them also not.
The linker will warn when linking against such shared objects files and, with all warnings are errors set linking will fail. Without this patch systemd e.g. failsdo_compileas it links against libcrypto.so from openssl.To test what features an object file (or .so, .a file) has use readelf -ln and check the content of .note.gnu.property, e.g. libcrypto-shlib-md5-aarch64.o build from assembler source, libcrypto-shlib-md5_one.o build from C source:
$ aarch64-poky-linux-readelf -ln ./crypto/md5/libcrypto-shlib-md5-aarch64.oThere are no program headers in this file.
Displaying notes found in: .note.gnu.property
Owner Data size Description GNU 0x00000010 NT_GNU_PROPERTY_TYPE_0Properties: AArch64 feature: BTI, PAC
Quoted 8 lines
$ aarch64-poky-linux-readelf -ln ./crypto/md5/libcrypto-shlib-md5_one.oThere are no program headers in this file.
Displaying notes found in: .note.gnu.property
Owner Data size Description GNU 0x00000010 NT_GNU_PROPERTY_TYPE_0Properties: AArch64 feature: BTI, PAC, GCS
Good find. I was suspecting something like that, what happens if we use compiler driver to invoke as assembler ?
another way might be to add CFI directives in asm files, maybe just start and end
foo:
.cfi_startproc
...
.cfi_endproc
Quoted 14 lines
Signed-off-by Max Krummenacher <[email protected]>--- meta/recipes-connectivity/openssl/openssl_3.5.0.bb | 2 ++ 1 file changed, 2 insertions(+)For reference also refer to this thread:
https://lore.kernel.org/all/aBilFkr4HF-MSBd6@toolbox/meta/recipes-connectivity/openssl/openssl_3.5.0.bb +0
-0
meta/recipes-connectivity/openssl/openssl_3.5.0.bb B = "${WORKDIR}/build" · line 40
Quoted 5 lines
meta/recipes-connectivity/openssl/openssl_3.5.0.bb +2
-0@@ -36,6 +36,8 @@ B = "${WORKDIR}/build" do_configure[cleandirs] = "${B}" EXTRA_OECONF = "${@bb.utils.contains('PTEST_ENABLED', '1', '', 'no-tests', d)}" +# aarch64 'as' doesn't mark objects to have the GCS feature, resulting *.so then produces linker warnings +EXTRA_OECONF:append:aarch64 = " no-asm"
Does this disable assembly ? if so there might be a performance impact, it would be good to find that out.
Quoted 15 lines
meta/recipes-connectivity/openssl/openssl_3.5.0.bb +0
-1#| ./libcrypto.so: undefined reference to `getcontext' #| ./libcrypto.so: undefined reference to `setcontext' --2.42.0
-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.
View/Reply Online (#216050): https://lists.openembedded.org/g/openembedded-core/message/216050
Mute This Topic: https://lists.openembedded.org/mt/112649627/1997914
Group Owner: [email protected]
Unsubscribe: https://lists.openembedded.org/g/openembedded-core/unsub [[email protected]]
-=-=-=-=-=-=-=-=-=-=-=-
Hi,
Quoted 42 lines · Khem Raj via lists.openembedded.org
Quoted 30 lines · Max Krummenacher via lists.openembedded.org
From: Max Krummenacher <[email protected]>
openssl has a couple of functionalities which use optimized assembler code. With no-asm these are replaced by generic C code implementation.
With GCC 15 OE by default uses the new AARCH64 specific GCS feature. However the object files produced by the assembler are not marked to provide the feature and consequently the produced shared objects containing them also not.
The linker will warn when linking against such shared objects files and, with all warnings are errors set linking will fail. Without this patch systemd e.g. failsdo_compileas it links against libcrypto.so from openssl.To test what features an object file (or .so, .a file) has use readelf -ln and check the content of .note.gnu.property, e.g. libcrypto-shlib-md5-aarch64.o build from assembler source, libcrypto-shlib-md5_one.o build from C source:
$ aarch64-poky-linux-readelf -ln ./crypto/md5/libcrypto-shlib-md5-aarch64.oThere are no program headers in this file.
Displaying notes found in: .note.gnu.property
Owner Data size Description GNU 0x00000010 NT_GNU_PROPERTY_TYPE_0Properties: AArch64 feature: BTI, PAC
Quoted 8 lines
$ aarch64-poky-linux-readelf -ln ./crypto/md5/libcrypto-shlib-md5_one.oThere are no program headers in this file.
Displaying notes found in: .note.gnu.property
Owner Data size Description GNU 0x00000010 NT_GNU_PROPERTY_TYPE_0Properties: AArch64 feature: BTI, PAC, GCS
Quoted 8 lines
Good find. I was suspecting something like that, what happens if we use compiler driver to invoke as assembler ?
another way might be to add CFI directives in asm files, maybe just start and end
foo:
.cfi_startproc
...
.cfi_endproc
This may be needed in a lot of places...
Quoted 16 lines
Quoted 16 lines
Signed-off-by Max Krummenacher <[email protected]>--- meta/recipes-connectivity/openssl/openssl_3.5.0.bb | 2 ++ 1 file changed, 2 insertions(+)For reference also refer to this thread:
https://lore.kernel.org/all/aBilFkr4HF-MSBd6@toolbox/meta/recipes-connectivity/openssl/openssl_3.5.0.bb +0
-0@@ -36,6 +36,8 @@ B = "${WORKDIR}/build" do_configure[cleandirs] = "${B}"
Quoted 6 lines
meta/recipes-connectivity/openssl/openssl_3.5.0.bb B = "${WORKDIR}/build" · line 40
Quoted 3 lines
meta/recipes-connectivity/openssl/openssl_3.5.0.bb +2
-0EXTRA_OECONF = "${@bb.utils.contains('PTEST_ENABLED', '1', '', 'no-tests', d)}" +# aarch64 'as' doesn't mark objects to have the GCS feature, resulting *.so then produces linker warnings +EXTRA_OECONF:append:aarch64 = " no-asm"Does this disable assembly ? if so there might be a performance impact, it would be good to find that out.
Or security win? What does upstream recommend?
This GCS is now enabled with -mbranch-protection=standard. There may be more places which need similar fixes/workarounds so getting this working automatically would be better. Bug report to gcc?
I wonder if GCS should be disabled for now and use "bti" instead?
https://gcc.gnu.org/onlinedocs/gcc-15.1.0/gcc/AArch64-Options.html#index-mbranch-protection
Cheers,
-Mikko
Quoted 9 lines · Max Krummenacher via lists.openembedded.org
$ aarch64-poky-linux-readelf -ln ./crypto/md5/libcrypto-shlib-md5-aarch64.oThere are no program headers in this file.
Displaying notes found in: .note.gnu.property
Owner Data size Description GNU 0x00000010 NT_GNU_PROPERTY_TYPE_0Properties: AArch64 feature: BTI, PAC
Quoted 8 lines
$ aarch64-poky-linux-readelf -ln ./crypto/md5/libcrypto-shlib-md5_one.oThere are no program headers in this file.
Displaying notes found in: .note.gnu.property
Owner Data size Description GNU 0x00000010 NT_GNU_PROPERTY_TYPE_0Properties: AArch64 feature: BTI, PAC, GCS
Some context here: openssl has some assembler code that has PAC and BTI hints but not GCS. The systemd recipe then links to libcrypto from openssl with GCS enabled (as that is a distro-wide setting) and it - correctly - warns that it’s being told to use GCS but one of the inputs doesn’t have GCS. This wouldn’t be a problem but systemd also links with —fatal-warnings, so the build explodes.
I’m asking people internally who know more about this than me, but it’s possible that the proper fix here is to just add the GCS tag. Alternatively, relaxing the systemd build so that it doesn’t have fatal warnings (or consider GCS issues a warning) would be a fallback. Disabling the assembler fast-paths because of this is an incredibly blunt hammer.
Ross