package/libnss: fix build failure due to HW PPC Crypto bug
PATCH 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
Add patch to fix how NSS check if we're on a Big Endian machine.
---
Pending here:
https://bugzilla.mozilla.org/show_bug.cgi?id=1606119
--- ...6119-Fix-PPC-HW-Crypto-build-failure.patch | 37 +++++++++++++++++++ 1 file changed, 37 insertions(+) create mode 100644 package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch new file +37 -0
@@ -0,0 +1,37 @@ +From 09b3776a924736049693a118d5a8d883e8c794ca Mon Sep 17 00:00:00 2001 +From: Giulio Benetti <[email protected]> +Date: Fri, 27 Dec 2019 17:41:04 +0100 +Subject: [PATCH] Bug 1606119 - Fix PPC HW Crypto build failure + +Only Big Endian Altivec functions are used, not Little Endian ones, so +let's change check if USE_PPC_CRYPTO by changing to IS_BIG_ENDIAN +instead of IS_LITTLE_ENDIAN. + +Signed-off-by: Giulio Benetti <[email protected]> +--- + nss/lib/freebl/gcm.h | 6 +++--- + 1 file changed, 3 insertions(+), 3 deletions(-) + +diff --git a/nss/lib/freebl/gcm.h b/nss/lib/freebl/gcm.h +index 571b9ec55..de8b51db8 100644 +--- a/nss/lib/freebl/gcm.h ++++ b/nss/lib/freebl/gcm.h +@@ -41,12 +41,12 @@ + #endif + + /* +- * PPC CRYPTO requires at least gcc 5 or clang. The LE check is purely +- * because it's only been tested on LE. If you're interested in BE, ++ * PPC CRYPTO requires at least gcc 5 or clang. The BE check is purely ++ * because it's only been tested on BE. If you're interested in LE, + * please send a patch. + */ + #if (defined(__clang__) || (defined(__GNUC__) && __GNUC__ >= 5)) && \ +- defined(IS_LITTLE_ENDIAN) ++ defined(IS_BIG_ENDIAN) + #define USE_PPC_CRYPTO + #endif + +-- +2.20.1 +
Signature · 1 line
2.20.1
PATCH 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
Quoted 1 line · Giulio Benetti
Add patch to fix how NSS check if we're on a Big Endian machine.
Fixes:
http://autobuild.buildroot.net/results/e42/e42005d1bb70b3f2741eb77242b2fd8dbf2a55d6/
Quoted 53 lines
Signed-off-by Giulio Benetti <[email protected]>---
Pending here:
https://bugzilla.mozilla.org/show_bug.cgi?id=1606119--- ...6119-Fix-PPC-HW-Crypto-build-failure.patch | 37 +++++++++++++++++++ 1 file changed, 37 insertions(+)create mode 100644
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patchpackage/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch new file +37
-0@@ -0,0 +1,37 @@ +From 09b3776a924736049693a118d5a8d883e8c794ca Mon Sep 17 00:00:00 2001 +From: Giulio Benetti <[email protected]> +Date: Fri, 27 Dec 2019 17:41:04 +0100 +Subject: [PATCH] Bug 1606119 - Fix PPC HW Crypto build failure + +Only Big Endian Altivec functions are used, not Little Endian ones, so +let's change check if USE_PPC_CRYPTO by changing to IS_BIG_ENDIAN +instead of IS_LITTLE_ENDIAN. + +Signed-off-by: Giulio Benetti <[email protected]> +--- + nss/lib/freebl/gcm.h | 6 +++--- + 1 file changed, 3 insertions(+), 3 deletions(-) + +diff --git a/nss/lib/freebl/gcm.h b/nss/lib/freebl/gcm.h +index 571b9ec55..de8b51db8 100644 +--- a/nss/lib/freebl/gcm.h ++++ b/nss/lib/freebl/gcm.h +@@ -41,12 +41,12 @@ + #endif + + /* +- * PPC CRYPTO requires at least gcc 5 or clang. The LE check is purely +- * because it's only been tested on LE. If you're interested in BE, ++ * PPC CRYPTO requires at least gcc 5 or clang. The BE check is purely ++ * because it's only been tested on BE. If you're interested in LE, + * please send a patch. + */ + #if (defined(__clang__) || (defined(__GNUC__) && __GNUC__ >= 5)) && \ +- defined(IS_LITTLE_ENDIAN) ++ defined(IS_BIG_ENDIAN) + #define USE_PPC_CRYPTO + #endif + +-- +2.20.1 +
Signature · 2 lines
Giulio Benetti Benetti Engineering sas
PATCH 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
In reply to Giulio BenettiQuoted 6 lines · On Fri, 27 Dec 2019 17:54:27 +0100 Giulio Benetti…
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch new file +0
-0
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch line 10
Quoted 10 lines
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch new file +10
-0@@ -0,0 +1,37 @@ +From 09b3776a924736049693a118d5a8d883e8c794ca Mon Sep 17 00:00:00 2001 +From: Giulio Benetti <[email protected]> +Date: Fri, 27 Dec 2019 17:41:04 +0100 +Subject: [PATCH] Bug 1606119 - Fix PPC HW Crypto build failure + +Only Big Endian Altivec functions are used, not Little Endian ones, so +let's change check if USE_PPC_CRYPTO by changing to IS_BIG_ENDIAN +instead of IS_LITTLE_ENDIAN. + +Signed-off-by: Giulio Benetti <[email protected]>
I don't understand the reasoning here. From a quick look, the undefined symbols reported in the build log come from the file ./nss/lib/freebl/gcm-ppc.c.
This file is included in the build in nss/lib/freebl/Makefile if CPU_ARCH=ppc.
However CPU_ARCH=ppc is only set in nss/coreconf/Linux.mk as follows:
ifeq (,$(filter-out ppc64 ppc64le,$(OS_TEST)))
CPU_ARCH = ppc
ifeq ($(USE_64),1)
ARCHFLAG = -m64
endif
In libnss.mk, we pass OS_TEST=$(LIBNSS_ARCH), which basically is the value of $(BR2_ARCH), which is powerpc64 or powerpc64le in Buildroot.
So, regardless of whether we are big endian PPC64 or little endian PPC64, the gcm-ppc.c file will not be included in the build.
Am I missing something ? Could you explain a bit better how you came to the conclusion you have in this patch ?
Thanks!
Thomas
Signature · 3 lines
Thomas Petazzoni, CTO, Bootlin Embedded Linux and Kernel engineering https://bootlin.com
PATCH 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
Hi Thomas,
Quoted 35 lines · Thomas Petazzoni
Quoted 6 lines · On Fri, 27 Dec 2019 17:54:27 +0100 Giulio Benetti…
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch new file +0
-0package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch line 10
Quoted 10 lines
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch new file +10
-0@@ -0,0 +1,37 @@ +From 09b3776a924736049693a118d5a8d883e8c794ca Mon Sep 17 00:00:00 2001 +From: Giulio Benetti <[email protected]> +Date: Fri, 27 Dec 2019 17:41:04 +0100 +Subject: [PATCH] Bug 1606119 - Fix PPC HW Crypto build failure + +Only Big Endian Altivec functions are used, not Little Endian ones, so +let's change check if USE_PPC_CRYPTO by changing to IS_BIG_ENDIAN +instead of IS_LITTLE_ENDIAN. + +Signed-off-by: Giulio Benetti <[email protected]>I don't understand the reasoning here. From a quick look, the undefined symbols reported in the build log come from the file
./nss/lib/freebl/gcm-ppc.c.This file is included in the build in nss/lib/freebl/Makefile if CPU_ARCH=ppc.
However CPU_ARCH=ppc is only set in
nss/coreconf/Linux.mkas follows:ifeq (,$(filter-out ppc64 ppc64le,$(OS_TEST)))
CPU_ARCH= ppc
ifeq ($(USE_64),1)
ARCHFLAG = -m64
endif
Quoted 8 lines
In libnss.mk, we pass OS_TEST=$(LIBNSS_ARCH), which basically is the value of $(BR2_ARCH), which is powerpc64 or powerpc64le in Buildroot.
So, regardless of whether we are big endian PPC64 or little endian PPC64, the
gcm-ppc.cfile will not be included in the build.Am I missing something ? Could you explain a bit better how you came to the conclusion you have in this patch ?
The point is that gcm-ppc.c gets compiled, but inside it there's an #if defined(USE_PPC_CRYPTO) on top of the file, that is defined in gsm.h only #if __powerpc64__ and LITTLE_ENDIAN. But inside gcm-ppc.c they use _be() Altivec functions. Basically they have inverted LITTLE_ENDIAN with BIB_ENDIAN, since they use Big Endian only Altivec functions in gcm-ppc.c so the only wrong piece of code is in gcm.h, where USE_PPC_CRYPTO is defined.
Need only to invert the condition by which USE_PPC_CRYPTO is defined:BIG_ENDIAN instead of LITTLE_ENDIAN.
Best regards
Signature · 2 lines
Giulio Benetti Benetti Engineering sas
Quoted 6 lines
Thanks!
Thomas
PATCH 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
Gang,
Just going to toss my $.02 in...
Quoted 50 lines · Giulio Benetti
Hi Thomas,
Quoted 38 lines · Thomas Petazzoni
Quoted 9 lines · On Fri, 27 Dec 2019 17:54:27 +0100 Giulio Benetti…
(quoted diff) +0
-0
a/package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patchb/package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patchnew file mode 100644
index 0000000000..0b891b5ebc
---/dev/null
+++b/package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patchpackage/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch line 10
Quoted 10 lines
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch new file +10
-0@@ -0,0 +1,37 @@ +From 09b3776a924736049693a118d5a8d883e8c794ca Mon Sep 17 00:00:00 2001 +From: Giulio Benetti <[email protected]> +Date: Fri, 27 Dec 2019 17:41:04 +0100 +Subject: [PATCH] Bug 1606119 - Fix PPC HW Crypto build failure + +Only Big Endian Altivec functions are used, not Little Endian ones, so +let's change check if USE_PPC_CRYPTO by changing to IS_BIG_ENDIAN +instead of IS_LITTLE_ENDIAN. + +Signed-off-by: Giulio Benetti <[email protected]>I don't understand the reasoning here. From a quick look, the undefined symbols reported in the build log come from the file
./nss/lib/freebl/gcm-ppc.c.This file is included in the build in nss/lib/freebl/Makefile if CPU_ARCH=ppc.
However CPU_ARCH=ppc is only set in
nss/coreconf/Linux.mkas follows:ifeq (,$(filter-out ppc64 ppc64le,$(OS_TEST)))
????????CPU_ARCH??????? = ppc
ifeq ($(USE_64),1)
???????? ARCHFLAG??????? = -m64
endifQuoted 8 lines
In libnss.mk, we pass OS_TEST=$(LIBNSS_ARCH), which basically is the value of $(BR2_ARCH), which is powerpc64 or powerpc64le in Buildroot.
So, regardless of whether we are big endian PPC64 or little endian PPC64, the
gcm-ppc.cfile will not be included in the build.Am I missing something ? Could you explain a bit better how you came to the conclusion you have in this patch ?
Quoted 8 lines
The point is that
gcm-ppc.cgets compiled, but inside it there's an #ifdefined(USE_PPC_CRYPTO)on top of the file, that is defined ingsm.honly #if__powerpc64__andLITTLE_ENDIAN. But insidegcm-ppc.cthey use_be()Altivec functions. Basically they have invertedLITTLE_ENDIANwithBIB_ENDIAN, since they use Big Endian only Altivec functions ingcm-ppc.cso the only wrong piece of code is ingcm.h, whereUSE_PPC_CRYPTOis defined.
I assume you mean this in gcm_HashMult_hw:
??????? /* clang needs the following cast away from const; maybe a bug
in 7.0.0 */
??????? v = (vec_u64)vec_xl_be(0, (unsigned char *)buf);
It should be fine to call endian specific functions if you know how the data is stored.
Quoted 4 lines
Need only to invert the condition by which
USE_PPC_CRYPTOis defined:BIG_ENDIANinstead ofLITTLE_ENDIAN.Best regards
Based on the below commit, it was tested on a POWER8, presumably in LE mode, so I'd be hesitant to switch any endian checks:
https://hg.mozilla.org/projects/nss/rev/3d7e509d6d20ecd607a28fa6ce42e4ffd9c51443
The bigger issue may be that 'vec_xl_be' wasn't introduced until GCC 8 and the toolchain used in the failed build is GCC 7, so that function is not available in altivec.h:
https://github.com/gcc-mirror/gcc/commit/f7b0548e5eba54b637977e3df4d4daf0cabe474d
So if we stick with nss-3.48, the package should likely depend on BR2_TOOLCHAIN_GCC_AT_LEAST_8.
Signature · 4 lines
Vincent Fazio Embedded Software Engineer - Linux Extreme Engineering Solutions, Inc http://www.xes-inc.com
PATCH 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
Quoted 86 lines · Vincent Fazio
Gang,
Just going to toss my $.02 in...
Quoted 54 lines · Giulio Benetti
Hi Thomas,
Quoted 42 lines · Thomas Petazzoni
Quoted 19 lines · On Fri, 27 Dec 2019 17:54:27 +0100 Giulio Benetti…
(quoted diff) +0
-0
a/package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patchb/package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patchnew file mode 100644
index 0000000000..0b891b5ebc
---/dev/null
+++b/package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch(quoted diff) +1
-0@@ -0,0 +1,37 @@ +From 09b3776a924736049693a118d5a8d883e8c794ca Mon Sep 17 00:00:002001
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch new file +5
-0+From: Giulio Benetti <[email protected]> +Date: Fri, 27 Dec 2019 17:41:04 +0100 +Subject: [PATCH] Bug 1606119 - Fix PPC HW Crypto build failure + +Only Big Endian Altivec functions are used, not Little Endianones, so
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch line 10
Quoted 4 lines
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch new file +4
-0+let's change check if USE_PPC_CRYPTO by changing to IS_BIG_ENDIAN +instead of IS_LITTLE_ENDIAN. + +Signed-off-by: Giulio Benetti <[email protected]>I don't understand the reasoning here. From a quick look, the undefined symbols reported in the build log come from the file
./nss/lib/freebl/gcm-ppc.c.This file is included in the build in nss/lib/freebl/Makefile if CPU_ARCH=ppc.
However CPU_ARCH=ppc is only set in
nss/coreconf/Linux.mkas follows:ifeq (,$(filter-out ppc64 ppc64le,$(OS_TEST)))
????????CPU_ARCH??????? = ppc
ifeq ($(USE_64),1)
???????? ARCHFLAG??????? = -m64
endifQuoted 8 lines
In libnss.mk, we pass OS_TEST=$(LIBNSS_ARCH), which basically is the value of $(BR2_ARCH), which is powerpc64 or powerpc64le in Buildroot.
So, regardless of whether we are big endian PPC64 or little endian PPC64, the
gcm-ppc.cfile will not be included in the build.Am I missing something ? Could you explain a bit better how you came to the conclusion you have in this patch ?
Quoted 8 lines
The point is that
gcm-ppc.cgets compiled, but inside it there's an #ifdefined(USE_PPC_CRYPTO)on top of the file, that is defined ingsm.honly #if__powerpc64__andLITTLE_ENDIAN. But insidegcm-ppc.cthey use_be()Altivec functions. Basically they have invertedLITTLE_ENDIANwithBIB_ENDIAN, since they use Big Endian only Altivec functions ingcm-ppc.cso the only wrong piece of code is ingcm.h, whereUSE_PPC_CRYPTOis defined.I assume you mean this in
gcm_HashMult_hw:??????? /* clang needs the following cast away from const; maybe a bug in 7.0.0 */
??????? v = (vec_u64)vec_xl_be(0, (unsigned char *)buf);It should be fine to call endian specific functions if you know how the data is stored.
Quoted 4 lines
Need only to invert the condition by which
USE_PPC_CRYPTOis defined:BIG_ENDIANinstead ofLITTLE_ENDIAN.Best regards
Based on the below commit, it was tested on a POWER8, presumably in LE mode, so I'd be hesitant to switch any endian checks:
https://hg.mozilla.org/projects/nss/rev/3d7e509d6d20ecd607a28fa6ce42e4ffd9c51443
Quoted 8 lines
The bigger issue may be that '
vec_xl_be' wasn't introduced until GCC 8 and the toolchain used in the failed build is GCC 7, so that function is not available inaltivec.h:
https://github.com/gcc-mirror/gcc/commit/f7b0548e5eba54b637977e3df4d4daf0cabe474dSo if we stick with nss-3.48, the package should likely depend on
BR2_TOOLCHAIN_GCC_AT_LEAST_8.
Or don't utilize PPC hardware accelerated instructions by undefining USE_PPC_CRYPTO... which is probably the saner option :-). This would give the semblance that the compile worked as expected since the hashing function will now be available, but Thomas' concern re CPU_ARCH and the future inclusion of ppc specific source is still valid.
Signature · 4 lines
Vincent Fazio Embedded Software Engineer - Linux Extreme Engineering Solutions, Inc http://www.xes-inc.com
PATCH 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
Hi Vincent,
Quoted 69 lines · Vincent Fazio
Quoted 67 lines · Vincent Fazio
Gang,
Just going to toss my $.02 in...
Quoted 62 lines · Giulio Benetti
Hi Thomas,
Quoted 42 lines · Thomas Petazzoni
Quoted 19 lines · On Fri, 27 Dec 2019 17:54:27 +0100 Giulio Benetti…
diff --git
a/package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patchb/package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patchnew file mode 100644
index 0000000000..0b891b5ebc
---/dev/null
+++b/package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch(quoted diff) +1
-0@@ -0,0 +1,37 @@ +From 09b3776a924736049693a118d5a8d883e8c794ca Mon Sep 17 00:00:002001
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch new file +5
-0+From: Giulio Benetti <[email protected]> +Date: Fri, 27 Dec 2019 17:41:04 +0100 +Subject: [PATCH] Bug 1606119 - Fix PPC HW Crypto build failure + +Only Big Endian Altivec functions are used, not Little Endianones, so
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch line 10
Quoted 4 lines
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch new file +4
-0+let's change check if USE_PPC_CRYPTO by changing to IS_BIG_ENDIAN +instead of IS_LITTLE_ENDIAN. + +Signed-off-by: Giulio Benetti <[email protected]>I don't understand the reasoning here. From a quick look, the undefined symbols reported in the build log come from the file
./nss/lib/freebl/gcm-ppc.c.This file is included in the build in nss/lib/freebl/Makefile if CPU_ARCH=ppc.
However CPU_ARCH=ppc is only set in
nss/coreconf/Linux.mkas follows:ifeq (,$(filter-out ppc64 ppc64le,$(OS_TEST)))
????????CPU_ARCH??????? = ppc
ifeq ($(USE_64),1)
???????? ARCHFLAG??????? = -m64
endifQuoted 8 lines
In libnss.mk, we pass OS_TEST=$(LIBNSS_ARCH), which basically is the value of $(BR2_ARCH), which is powerpc64 or powerpc64le in Buildroot.
So, regardless of whether we are big endian PPC64 or little endian PPC64, the
gcm-ppc.cfile will not be included in the build.Am I missing something ? Could you explain a bit better how you came to the conclusion you have in this patch ?
The point is that
gcm-ppc.cgets compiled, but inside it there's an #ifdefined(USE_PPC_CRYPTO)on top of the file, that is defined ingsm.honly #if__powerpc64__andLITTLE_ENDIAN. But insidegcm-ppc.cthey use_be()Altivec functions. Basically they have invertedLITTLE_ENDIANwithBIB_ENDIAN, since they use Big Endian only Altivec functions ingcm-ppc.cso the only wrong piece of code is ingcm.h, whereUSE_PPC_CRYPTOis defined.
Quoted 1 line
Quoted 1 line
I assume you mean this in
gcm_HashMult_hw:
Yes
Quoted 7 lines
Quoted 7 lines
??????? /* clang needs the following cast away from const; maybe a bug
in 7.0.0 */
??????? v = (vec_u64)vec_xl_be(0, (unsigned char *)buf);It should be fine to call endian specific functions if you know how the data is stored.
I've completely understood vec_xl_be() wrong.
Quoted 9 lines
Quoted 9 lines
Quoted 5 lines
Need only to invert the condition by which
USE_PPC_CRYPTOis defined:BIG_ENDIANinstead ofLITTLE_ENDIAN.Best regards
Based on the below commit, it was tested on a POWER8, presumably in LE mode, so I'd be hesitant to switch any endian checks:
https://hg.mozilla.org/projects/nss/rev/3d7e509d6d20ecd607a28fa6ce42e4ffd9c51443
Right.
Quoted 6 lines
Quoted 6 lines
The bigger issue may be that '
vec_xl_be' wasn't introduced until GCC 8 and the toolchain used in the failed build is GCC 7, so that function is not available inaltivec.h:
https://github.com/gcc-mirror/gcc/commit/f7b0548e5eba54b637977e3df4d4daf0cabe474d
Ok
Quoted 4 lines
Quoted 4 lines
So if we stick with nss-3.48, the package should likely depend on
BR2_TOOLCHAIN_GCC_AT_LEAST_8.
This ^^^ is too restrictive, but...
Quoted 5 lines
Or don't utilize PPC hardware accelerated instructions by undefining
USE_PPC_CRYPTO... which is probably the saner option :-). This would give the semblance that the compile worked as expected since the hashing function will now be available, but Thomas' concern reCPU_ARCHand the future inclusion of ppc specific source is still valid.
...this is a good idea. So with gcc version < 8.x let's disable USE_PPC_CRYPTO and with gcc version >= 8.x let's enable it only if LITTLE_ENDIAN. So at this point I would change my upstream patch to do that in NSS, since there's already a gcc version check but only against version 5.x(>=) that instead needs to be >= 8.x.
Thanks for all the explanations!
Best regards
and
happy new year! For me at least :-)
Signature · 2 lines
Giulio Benetti Benetti Engineering sas
PATCH 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
Giulio, Vincent, All,
On 2020-01-01 00:44 +0100, Giulio Benetti spake thusly:
Quoted 2 lines · Vincent Fazio
Quoted 1 line · Vincent Fazio
[--SNIP--]
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
??????? /* clang needs the following cast away from const; maybe a bug in 7.0.0 */
??????? v = (vec_u64)vec_xl_be(0, (unsigned char *)buf);It should be fine to call endian specific functions if you know how the data is stored.
[--SNIP--]
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Based on the below commit, it was tested on a POWER8, presumably in LE mode, so I'd be hesitant to switch any endian checks:
https://hg.mozilla.org/projects/nss/rev/3d7e509d6d20ecd607a28fa6ce42e4ffd9c51443
[--SNIP--]
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
The bigger issue may be that '
vec_xl_be' wasn't introduced until GCC 8 and the toolchain used in the failed build is GCC 7, so that function is not available inaltivec.h:
https://github.com/gcc-mirror/gcc/commit/f7b0548e5eba54b637977e3df4d4daf0cabe474d
Thanks Vincent for digging all this info, very informative. :-)
Quoted 5 lines
...this is a good idea. So with gcc version < 8.x let's disable
USE_PPC_CRYPTOand with gcc version >= 8.x let's enable it only ifLITTLE_ENDIAN. So at this point I would change my upstream patch to do that in NSS, since there's already a gcc version check but only against version 5.x(>=) that instead needs to be >= 8.x.
Yes, that sounds like a good plan. I've marked this patch as changes requested in ptachwork in the meantime.
Thanks! :-)
Regards,
Yann E. MORIN.
Signature · 6 lines
.-----------------.--------------------.------------------.--------------------. | Yann E. MORIN | Real-Time Embedded | /"\ ASCII RIBBON | Erics' conspiracy: | | +33 662 376 056 | Software Designer | \ / CAMPAIGN | ___ | | +33 561 099 427 `------------.-------: X AGAINST | \e/ There is no | | http://ymorin.is-a-geek.org/ | _/*\_ | / \ HTML MAIL | v conspiracy. | '------------------------------^-------^------------------^--------------------'
PATCH 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
In reply to Giulio BenettiQuoted 11 lines · On Wed, 1 Jan 2020 00:44:09 +0100 Giulio Benetti…
Quoted 2 lines
Quoted 2 lines
So if we stick with nss-3.48, the package should likely depend on
BR2_TOOLCHAIN_GCC_AT_LEAST_8.This ^^^ is too restrictive, but...
Quoted 5 lines
Or don't utilize PPC hardware accelerated instructions by undefining
USE_PPC_CRYPTO... which is probably the saner option :-). This would give the semblance that the compile worked as expected since the hashing function will now be available, but Thomas' concern reCPU_ARCHand the future inclusion of ppc specific source is still valid.
Quoted 5 lines
...this is a good idea. So with gcc version < 8.x let's disable
USE_PPC_CRYPTOand with gcc version >= 8.x let's enable it only ifLITTLE_ENDIAN. So at this point I would change my upstream patch to do that in NSS, since there's already a gcc version check but only against version 5.x(>=) that instead needs to be >= 8.x.
As I said in my previous e-mail, I also don't see how it can work with the current OS_TEST value that we pass in libnss.mk, which is never set to ppc64 or ppc64le as the libnss build system expects.
Best regards,
Thomas
Signature · 3 lines
Thomas Petazzoni, CTO, Bootlin Embedded Linux and Kernel engineering https://bootlin.com
PATCH 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
Hi Thomas,
Quoted 20 lines · Thomas Petazzoni
Quoted 11 lines · On Wed, 1 Jan 2020 00:44:09 +0100 Giulio Benetti…
Quoted 2 lines
Quoted 2 lines
So if we stick with nss-3.48, the package should likely depend on
BR2_TOOLCHAIN_GCC_AT_LEAST_8.This ^^^ is too restrictive, but...
Quoted 5 lines
Or don't utilize PPC hardware accelerated instructions by undefining
USE_PPC_CRYPTO... which is probably the saner option :-). This would give the semblance that the compile worked as expected since the hashing function will now be available, but Thomas' concern reCPU_ARCHand the future inclusion of ppc specific source is still valid.Quoted 5 lines
...this is a good idea. So with gcc version < 8.x let's disable
USE_PPC_CRYPTOand with gcc version >= 8.x let's enable it only ifLITTLE_ENDIAN. So at this point I would change my upstream patch to do that in NSS, since there's already a gcc version check but only against version 5.x(>=) that instead needs to be >= 8.x.
Quoted 3 lines
As I said in my previous e-mail, I also don't see how it can work with the current
OS_TESTvalue that we pass in libnss.mk, which is never set to ppc64 or ppc64le as the libnss build system expects.
You're right, the problem is that it doesn't compile gcm-ppc.c because Buildroot sets OS_TEST to powerpc64le in this case but libnss expects ppc64le, so I'm testing a patch where I remap:BR2_powerpc => ppcBR2_powerpc64 => ppc64BR2_powerpc64le => ppc64le
And I'm going to add a patch too to modify minimum gcc version required from 5.x to 8.x, because as Vincent pointed vec_xst_be() has been introduced only in version 8.x.
This way we avoid BR2_TOOLCHAIN_GCC_AT_LEAST_8 and later I send it upstream.
Thank you and
Best regards
Signature · 2 lines
Giulio Benetti Benetti Engineering sas
PATCH v2 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
In reply to Thomas Petazzonilibnss expects OS_TEST to be set to ppc or ppc64 or ppc64le instead of powerpc, powerpc64, powerpc64le. At the moment gcm.h header checks if __powerpc64__ is defined, but Makefile expects OS_TEST to be set to ppc64 or ppc64le to build gcm-ppc.c. This way we end with having gcm prototypes defined, but without implementation compiled. So remap OS_TEST to what libnss expects depending on BR2_powerpc* enabled. This way Makefile will build gcm-ppc.c too, that contains gcm-ppc function implementations using Altivec. But add also a patch that modify minimum gcc version required for using Altivec functions from 5.x to 8.x since some Altivec intrinsics have been instroduced only in 8.x(i.e. vec_xst_be()).
---
V1 -> V2:
* changed completely approach
Pending upstream:
https://bugzilla.mozilla.org/show_bug.cgi?id=1606119
--- ...6119-Fix-PPC-HW-Crypto-build-failure.patch | 35 +++++++++++++++++++ package/libnss/libnss.mk | 6 ++++ 2 files changed, 41 insertions(+) create mode 100644 package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch
package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch new file +35 -0
@@ -0,0 +1,35 @@ +From ebf185f8e48b5aec622dc949cef1b19b0a7669ef Mon Sep 17 00:00:00 2001 +From: Giulio Benetti <[email protected]> +Date: Wed, 1 Jan 2020 12:54:45 +0100 +Subject: [PATCH] Bug 1606119 - Fix PPC HW Crypto build failure + +All Altivec *_be() functions are supported from gcc version 8.x not 5.x +so modify gcc version check that at the moment cause build failure due +to missing Altivec *_be() functions. + +Signed-off-by: Giulio Benetti <[email protected]> +--- + nss/lib/freebl/gcm.h | 4 ++-- + 1 file changed, 2 insertions(+), 2 deletions(-) + +diff --git a/nss/lib/freebl/gcm.h b/nss/lib/freebl/gcm.h +index 571b9ec55..aa4dee824 100644 +--- a/nss/lib/freebl/gcm.h ++++ b/nss/lib/freebl/gcm.h +@@ -41,11 +41,11 @@ + #endif + + /* +- * PPC CRYPTO requires at least gcc 5 or clang. The LE check is purely ++ * PPC CRYPTO requires at least gcc 8 or clang. The LE check is purely + * because it's only been tested on LE. If you're interested in BE, + * please send a patch. + */ +-#if (defined(__clang__) || (defined(__GNUC__) && __GNUC__ >= 5)) && \ ++#if (defined(__clang__) || (defined(__GNUC__) && __GNUC__ >= 8)) && \ + defined(IS_LITTLE_ENDIAN) + #define USE_PPC_CRYPTO + #endif +-- +2.20.1 +
package/libnss/libnss.mk +6 -0
@@ -39,6 +39,12 @@ endif ifeq ($(BR2_aarch64_be),y) LIBNSS_ARCH = aarch64 +else ifeq ($(BR2_powerpc),y) +LIBNSS_ARCH = ppc +else ifeq ($(BR2_powerpc64),y) +LIBNSS_ARCH = ppc64 +else ifeq ($(BR2_powerpc64le),y) +LIBNSS_ARCH = ppc64le else LIBNSS_ARCH = $(ARCH) endif
Signature · 1 line
2.20.1
Yann E. MORIN
AppliedPATCH v2 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
Giulio, All,
On 2020-01-01 17:58 +0100, Giulio Benetti spake thusly:
Quoted 12 lines
libnss expects
OS_TESTto be set to ppc or ppc64 or ppc64le instead of powerpc, powerpc64, powerpc64le. At the momentgcm.hheader checks if__powerpc64__is defined, butMakefileexpectsOS_TESTto be set to ppc64 or ppc64le to buildgcm-ppc.c. This way we end with having gcm prototypes defined, but without implementation compiled. So remapOS_TESTto what libnss expects depending on BR2_powerpc* enabled. This wayMakefilewill buildgcm-ppc.ctoo, that contains gcm-ppc function implementations using Altivec. But add also a patch that modify minimum gcc version required for using Altivec functions from 5.x to 8.x since some Altivec intrinsics have been instroduced only in 8.x(i.e.vec_xst_be()).
Quoted 1 line
Signed-off-by Giulio Benetti <[email protected]>
Applied to master, after doing a few tweaks:
- move the arch setting to kconfig (like recently done in e3159cad71)
- add the build failure
- add pointer to upstream bug report and patch sybmission
- reformat and reword commit log
Pushed now, thanks.
Regards,
Yann E. MORIN.
Quoted 76 lines
---
V1 -> V2:
* changed completely approach
Pending upstream:
https://bugzilla.mozilla.org/show_bug.cgi?id=1606119--- ...6119-Fix-PPC-HW-Crypto-build-failure.patch | 35 +++++++++++++++++++ package/libnss/libnss.mk | 6 ++++ 2 files changed, 41 insertions(+) create mode 100644 package/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patchpackage/libnss/0005-Bug-1606119-Fix-PPC-HW-Crypto-build-failure.patch new file +35
-0@@ -0,0 +1,35 @@ +From ebf185f8e48b5aec622dc949cef1b19b0a7669ef Mon Sep 17 00:00:00 2001 +From: Giulio Benetti <[email protected]> +Date: Wed, 1 Jan 2020 12:54:45 +0100 +Subject: [PATCH] Bug 1606119 - Fix PPC HW Crypto build failure + +All Altivec *_be() functions are supported from gcc version 8.x not 5.x +so modify gcc version check that at the moment cause build failure due +to missing Altivec *_be() functions. + +Signed-off-by: Giulio Benetti <[email protected]> +--- + nss/lib/freebl/gcm.h | 4 ++-- + 1 file changed, 2 insertions(+), 2 deletions(-) + +diff --git a/nss/lib/freebl/gcm.h b/nss/lib/freebl/gcm.h +index 571b9ec55..aa4dee824 100644 +--- a/nss/lib/freebl/gcm.h ++++ b/nss/lib/freebl/gcm.h +@@ -41,11 +41,11 @@ + #endif + + /* +- * PPC CRYPTO requires at least gcc 5 or clang. The LE check is purely ++ * PPC CRYPTO requires at least gcc 8 or clang. The LE check is purely + * because it's only been tested on LE. If you're interested in BE, + * please send a patch. + */ +-#if (defined(__clang__) || (defined(__GNUC__) && __GNUC__ >= 5)) && \ ++#if (defined(__clang__) || (defined(__GNUC__) && __GNUC__ >= 8)) && \ + defined(IS_LITTLE_ENDIAN) + #define USE_PPC_CRYPTO + #endif +-- +2.20.1 +package/libnss/libnss.mk +6
-0@@ -39,6 +39,12 @@ endif ifeq ($(BR2_aarch64_be),y) LIBNSS_ARCH = aarch64 +else ifeq ($(BR2_powerpc),y) +LIBNSS_ARCH = ppc +else ifeq ($(BR2_powerpc64),y) +LIBNSS_ARCH = ppc64 +else ifeq ($(BR2_powerpc64le),y) +LIBNSS_ARCH = ppc64le else LIBNSS_ARCH = $(ARCH) endifSignature · 6 lines
2.20.1 _______________________________________________ buildroot mailing list buildroot at busybox.net http://lists.busybox.net/mailman/listinfo/buildroot
Signature · 6 lines
.-----------------.--------------------.------------------.--------------------. | Yann E. MORIN | Real-Time Embedded | /"\ ASCII RIBBON | Erics' conspiracy: | | +33 662 376 056 | Software Designer | \ / CAMPAIGN | ___ | | +33 561 099 427 `------------.-------: X AGAINST | \e/ There is no | | http://ymorin.is-a-geek.org/ | _/*\_ | / \ HTML MAIL | v conspiracy. | '------------------------------^-------^------------------^--------------------'
PATCH v2 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
In reply to Giulio BenettiHello Giulio,
Thanks for the patch. One comment below.
package/libnss/libnss.mk endif · line 50
Quoted 11 lines · On Wed, 1 Jan 2020 17:58:11 +0100 Giulio Benetti…
ifeq ($(BR2_aarch64_be),y)
LIBNSS_ARCH= aarch64package/libnss/libnss.mk +6
-0+else ifeq ($(BR2_powerpc),y) +LIBNSS_ARCH = ppc +else ifeq ($(BR2_powerpc64),y) +LIBNSS_ARCH = ppc64 +else ifeq ($(BR2_powerpc64le),y) +LIBNSS_ARCH = ppc64le else LIBNSS_ARCH = $(ARCH) endif
Shouldn't we globally review the values that libnss wants for the OS_TEST variable (which we set to $(LIBNSS_ARCH)) and see if other architectures need similar tweaks ?
Thanks,
Thomas
Signature · 3 lines
Thomas Petazzoni, CTO, Bootlin Embedded Linux and Kernel engineering https://bootlin.com
PATCH v2 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
Hi Thomas,
Quoted 7 lines · Thomas Petazzoni
Hello Giulio,
Thanks for the patch. One comment below.
On Wed, 1 Jan 2020 17:58:11 +0100
Giulio Benetti <[email protected]> wrote:
Quoted 15 lines
package/libnss/libnss.mk endif · line 50
Quoted 11 lines
ifeq ($(BR2_aarch64_be),y) LIBNSS_ARCH = aarch64package/libnss/libnss.mk +6
-0+else ifeq ($(BR2_powerpc),y) +LIBNSS_ARCH = ppc +else ifeq ($(BR2_powerpc64),y) +LIBNSS_ARCH = ppc64 +else ifeq ($(BR2_powerpc64le),y) +LIBNSS_ARCH = ppc64le else LIBNSS_ARCH = $(ARCH) endifShouldn't we globally review the values that libnss wants for the
OS_TESTvariable (which we set to $(LIBNSS_ARCH)) and see if other architectures need similar tweaks ?
Absolutely yes, going to do that.
Best regards
Signature · 2 lines
Giulio Benetti Benetti Engineering sas
PATCH v2 1/1 package/libnss: fix build failure due to HW PPC Crypto bug
Quoted 25 lines · Giulio Benetti
Hi Thomas,
Quoted 22 lines · Thomas Petazzoni
Hello Giulio,
Thanks for the patch. One comment below.
package/libnss/libnss.mk endif · line 50
Quoted 11 lines · On Wed, 1 Jan 2020 17:58:11 +0100 Giulio Benetti…
ifeq ($(BR2_aarch64_be),y) LIBNSS_ARCH = aarch64package/libnss/libnss.mk +6
-0+else ifeq ($(BR2_powerpc),y) +LIBNSS_ARCH = ppc +else ifeq ($(BR2_powerpc64),y) +LIBNSS_ARCH = ppc64 +else ifeq ($(BR2_powerpc64le),y) +LIBNSS_ARCH = ppc64le else LIBNSS_ARCH = $(ARCH) endifShouldn't we globally review the values that libnss wants for the
OS_TESTvariable (which we set to $(LIBNSS_ARCH)) and see if other architectures need similar tweaks ?
It turns out that all specific architectures checked in coreconf/Linux.mk are already correct(after changing powerpc* to ppc*). All the others are directly copied from OS_TEST to OS_ARCH in coreconf/Linux.mk:
https://hg.mozilla.org/projects/nss/file/tip/coreconf/Linux.mk#l93
and OS_ARCH is the variable NSS uses internally.
So it seems that at the moment there is nothing more to do about this.
Best regards
Signature · 2 lines
Giulio Benetti Benetti Engineering sas
Quoted 4 lines
Absolutely yes, going to do that.
Best regards