On Thu, May 9, 2019 at 11:34 AM Max Filippov jcmvbkbc@gmail.com wrote:
I've tried to run uclibc-ng tests on xtensa in a buildroot-built rootfs and found misc/tst-preadvwritev failing with the following message:
first buffer from first preadv different than expected
Digging deeper into it I've found that the test is built with the following definition in the command line: -D_FILE_OFFSET_BITS=64, due to the following line in the uclibc-ng-test.mk:
UCLIBC_EXTRA_CFLAGS="$(TARGET_CFLAGS)"
but uclibc-ng itself does not have -D_FILE_OFFSET_BITS=64, because of the following definition in the uclibc-ng.mk:
UCLIBC_EXTRA_CFLAGS="$(TARGET_ABI)"
uclibc-ng defaults to 32-bit off_t type then. I'm not sure what's broken here, buildroot or uclibc-ng. Ideas?
Ok, I see something definitely strange in the uclibc-ng: ublibc-ng does not honor -D_FILE_OFFSET_BITS=64 when it's passed to it in the make command, because the following fragment of the include/features.h that defines __USE_FILE_OFFSET64 based on the value of _FILE_OFFSET_BITS
#if defined _FILE_OFFSET_BITS && _FILE_OFFSET_BITS == 64 # define __USE_FILE_OFFSET64 1 #endif
is neutralized by the following fragment at the end of the same file:
#ifdef _LIBC # undef _FILE_OFFSET_BITS # undef __USE_FILE_OFFSET64 # include <libc-internal.h> #endif
-- Thanks. -- Max
If that can be of any help, this test seems to PASS on armv7 arch: https://uclibc-ng-ci.sionneau.net:8443/job/uclibc-ng-multiarch/arch=armv7/19...
However this bugs reproduces for mips32r6 : https://uclibc-ng-ci.sionneau.net:8443/job/uclibc-ng-multiarch/19/arch=mips3...
Also, I can confirm I am able to reproduce the issue with buildroot+uclibc-ng by building the qemu_xtensa_lx60_defconfig buildroot config
And then by booting it with qemu following this readme: https://github.com/buildroot/buildroot/tree/master/board/qemu/xtensa-lx60
On 5/9/19 11:10 PM, Max Filippov wrote:
On Thu, May 9, 2019 at 11:34 AM Max Filippov jcmvbkbc@gmail.com wrote:
I've tried to run uclibc-ng tests on xtensa in a buildroot-built rootfs and found misc/tst-preadvwritev failing with the following message:
first buffer from first preadv different than expected
Digging deeper into it I've found that the test is built with the following definition in the command line: -D_FILE_OFFSET_BITS=64, due to the following line in the uclibc-ng-test.mk:
UCLIBC_EXTRA_CFLAGS="$(TARGET_CFLAGS)"
but uclibc-ng itself does not have -D_FILE_OFFSET_BITS=64, because of the following definition in the uclibc-ng.mk:
UCLIBC_EXTRA_CFLAGS="$(TARGET_ABI)"
uclibc-ng defaults to 32-bit off_t type then. I'm not sure what's broken here, buildroot or uclibc-ng. Ideas?
Ok, I see something definitely strange in the uclibc-ng: ublibc-ng does not honor -D_FILE_OFFSET_BITS=64 when it's passed to it in the make command, because the following fragment of the include/features.h that defines __USE_FILE_OFFSET64 based on the value of _FILE_OFFSET_BITS
#if defined _FILE_OFFSET_BITS && _FILE_OFFSET_BITS == 64 # define __USE_FILE_OFFSET64 1 #endif
is neutralized by the following fragment at the end of the same file:
#ifdef _LIBC # undef _FILE_OFFSET_BITS # undef __USE_FILE_OFFSET64 # include <libc-internal.h> #endif
-- Thanks. -- Max _______________________________________________ devel mailing list devel@uclibc-ng.org https://mailman.uclibc-ng.org/cgi-bin/mailman/listinfo/devel
On Fri, May 10, 2019 at 6:32 AM Yann Sionneau ysionneau@kalray.eu wrote:
If that can be of any help, this test seems to PASS on armv7 arch: https://uclibc-ng-ci.sionneau.net:8443/job/uclibc-ng-multiarch/arch=armv7/19...
However this bugs reproduces for mips32r6 : https://uclibc-ng-ci.sionneau.net:8443/job/uclibc-ng-multiarch/19/arch=mips3...
Also, I can confirm I am able to reproduce the issue with buildroot+uclibc-ng by building the qemu_xtensa_lx60_defconfig buildroot config
Thanks for the confirmation. I've spent some more time on this issue and I hope have got a better understanding of it. So: - uclibc-ng itself is not meant to be compiled differently with -D_FILE_OFFSET_BITS=64 and without it. It is meant to provide both LFS and non-LFS function versions, like ftruncate and ftruncate64. - the library user on the other hand gets different set of declarations depending on the presence of -D_FILE_OFFSET_BITS=64, that results in a call to different functions. - however this is not the case for preadv/pwritev: there's only one definition for each of these functions. When libc is built these functions assume 32-bit off_t, but when the user code is built, off_t size is selected by the _FILE_OFFSET_BITS resulting in declaration that does not match definition.
I guess the right way to fix it is using __off64_t in declarations and definitions of both preadv/pwritev. With this change the test passes on xtensa.
-
Max Filippov -
Yann Sionneau