diff options
author | Laurent Vivier <laurent@vivier.eu> | 2018-05-02 23:57:30 +0200 |
---|---|---|
committer | Laurent Vivier <laurent@vivier.eu> | 2018-05-03 18:40:19 +0200 |
commit | 7f254c5cb80bc478794a4c3d7fe5d503b033be13 (patch) | |
tree | f11ee58adf9c71a8d1e53897090177aec7504c6a /include | |
parent | 465e237bf7cb6a9d8f9f137508125a14efcce1d6 (diff) |
linux-user: remove useless padding in flock64 structure
Since commit 8efb2ed5ec ("linux-user: Correct signedness of
target_flock l_start and l_len fields"), flock64 structure uses
abi_llong for l_start and l_len in place of "unsigned long long"
this should force them to be aligned accordingly to the target
rules. So we can remove the padding field and the QEMU_PACKED
attribute.
I have compared the result of the following program before and
after the change:
cat -> flock64_dump <<EOF
p/d sizeof(struct target_flock64)
p/d &((struct target_flock64 *)0)->l_type
p/d &((struct target_flock64 *)0)->l_whence
p/d &((struct target_flock64 *)0)->l_start
p/d &((struct target_flock64 *)0)->l_len
p/d &((struct target_flock64 *)0)->l_pid
quit
EOF
for file in build/all/*-linux-user/qemu-* ; do
echo $file
gdb -batch -nx -x flock64_dump $file 2> /dev/null
done
The sizeof() changes because we remove the QEMU_PACKED.
The new size is 32 (except for i386 and m68k) and this is
the real size of "struct flock64" on the target architecture.
The following architectures differ:
aarch64_be, aarch64, alpha, armeb, arm, cris, hppa, nios2, or1k,
riscv32, riscv64, s390x.
For a subset of these architectures, I have checked with the following
program the new structure is the correct one:
#include <stdio.h>
#define __USE_LARGEFILE64
#include <fcntl.h>
int main(void)
{
printf("struct flock64 %d\n", sizeof(struct flock64));
printf("l_type %d\n", &((struct flock64 *)0)->l_type);
printf("l_whence %d\n", &((struct flock64 *)0)->l_whence);
printf("l_start %d\n", &((struct flock64 *)0)->l_start);
printf("l_len %d\n", &((struct flock64 *)0)->l_len);
printf("l_pid %d\n", &((struct flock64 *)0)->l_pid);
}
[I have checked aarch64, alpha, hppa, s390x]
For ARM, the target_flock64 becomes the EABI definition, so we need to
define the OABI one in place of the EABI one and use it when it is
needed.
I have also fixed the alignment value for sh4 (to align llong on 4 bytes)
(see c2e3dee6e0 "linux-user: Define target alignment size")
[We should check alignment properties for cris, nios2 and or1k]
Signed-off-by: Laurent Vivier <laurent@vivier.eu>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20180502215730.28162-1-laurent@vivier.eu>
Diffstat (limited to 'include')
-rw-r--r-- | include/exec/user/abitypes.h | 2 |
1 files changed, 1 insertions, 1 deletions
diff --git a/include/exec/user/abitypes.h b/include/exec/user/abitypes.h index ba188608c2..743b8bb9ea 100644 --- a/include/exec/user/abitypes.h +++ b/include/exec/user/abitypes.h @@ -15,7 +15,7 @@ #define ABI_LLONG_ALIGNMENT 2 #endif -#if defined(TARGET_I386) && !defined(TARGET_X86_64) +#if (defined(TARGET_I386) && !defined(TARGET_X86_64)) || defined(TARGET_SH4) #define ABI_LLONG_ALIGNMENT 4 #endif |