There is a discrepancy in the personality() syscall on mips64.
For a 32bit process (o32 or n32) sys_32_personality() [1] is called, which
converts PER_LINUX32 -> PER_LINUX:
ret = sys_personality(p);
if (ret != -1 && (ret & PER_MASK) == PER_LINUX32) // personality macro applied
ret = (ret & ~PER_MASK) | PER_LINUX; // to make it clearer
This is different than what Intel and ARM are doing. They directly invoke
sys_personality(). On Android, this causes a bionic test to fail for mips, as it
expects PER_LINUX32 but receives PER_LINUX, when ran on mips64 core.
The code itself comes from older kernel versions (2.4), where it was also used
by other architectures [2], [3], [4]. However, since kernel 2.6 other archs are
invoking sys_personality() directly, for all ABIs.
Is there any reason why we should do PER_LINUX32 -> PER_LINUX on mips? Also, if
we do change this for o32, should we change it for n32, as well? There is no n32
support for Android, so it would have to be tested elsewhere.
Tested changing only o32 on Android, there were no regressions:
diff --git a/arch/mips/kernel/scall64-o32.S b/arch/mips/kernel/scall64-o32.S
index ab93427..b94ee4e 100644
--- a/arch/mips/kernel/scall64-o32.S
+++ b/arch/mips/kernel/scall64-o32.S
@@ -358,7 +358,7 @@ sys_call_table:
sys sys_fchdir 1
sys sys_bdflush 2
sys sys_sysfs 3 /* 4135 */
- sys sys_32_personality 1
+ sys sys_personality 1
sys sys_ni_syscall 0 /* for afs_syscall */
sys sys_setfsuid 1
sys sys_setfsgid 1
--
[1] http://git.linux-mips.org/cgit/ralf/linux.git/tree/arch/mips/kernel/linux32.c#n122
[2] http://git.linux-mips.org/cgit/ralf/linux.git/tree/include/asm-mips64/elf.h?h=linux-2.4#n124
[3] http://git.linux-mips.org/cgit/ralf/linux.git/tree/arch/x86_64/ia32/sys_ia32.c?h=linux-2.4#n1974
[4] http://git.linux-mips.org/cgit/ralf/linux.git/tree/arch/ia64/ia32/sys_ia32.c?h=linux-2.4#n4005
--
Nikola Veljkovic
