On Tue, Jan 14, 2025 at 04:03:28PM +0000, Maciej W. Rozycki wrote: > On Tue, 14 Jan 2025, Dmitry V. Levin wrote: > > > > How did you produce these results? > > > > $ PATH="$HOME/x-tools/mips64-unknown-linux-gnu/bin:$PATH" make -j`nproc` ARCH=mips CROSS_COMPILE=mips64-unknown-linux-gnu- -C tools/testing/selftests TARGETS=ptrace USERLDFLAGS='-static' USERCFLAGS='-mabi=32' > > $ echo init |(cd tools/testing/selftests/ptrace && ln -snf get_syscall_info init && cpio --dereference -o -H newc -R 0:0) |gzip >get_syscall_info.mips-o32.img > > $ qemu-system-mips -nographic -kernel vmlinuz -initrd get_syscall_info.mips-o32.img -append 'console=ttyS0' > > > > Likewise for mips64, but the patch for kselftest_harness.h from [1] > > is needed to see correct mismatch values in the test diagnostics. > > > > [1] https://lore.kernel.org/all/20250108170757.GA6723@xxxxxxxxx/ > > Thanks, I'll try to see what's going on with `get_user'. Thanks. > > > The patch is definitely broken, the calling convention is the same > > > between n32 and n64: 64-bit arguments in $4 through $11 registers as > > > required, and your change makes n32 truncate arguments to 32 bits. > > > > There must be something very specific to n32 then: apparently, > > __kernel_ulong_t is a 32-bit type on n32, so the syscall arguments are > > 32-bit values, at some point (in glibc?) they get sign-extended from 32 to > > 64 bits, and syscall_get_arguments returns them as 64-bit values different > > from the original syscall arguments, breaking the test. > > This matters at least for `lseek', which has an `off64_t' aka `long long' > argument on n32 (there's no `_llseek' on n32). Since arguments are passed > via 64-bit registers and a `long long' datum is held in just one this is > transparent between n32 and n64 (of course on n64 this corresponds to the > plain `long' data type, so the kernel, which is always n64 for 64-bit > configurations, sees the incoming argument as `long', and the same stands > for the outgoing return value). > > Surely non-LFS lseek(2) will produce the syscall's `long long' argument > truncated (cf. sysdeps/unix/sysv/linux/mips/mips64/n32/lseek.c in glibc), > but both LFS lseek(2) and lseek64(2) will pass the native value on n32. > > > If this is the expected behaviour, then I'd have to add an exception for > > mips n32 both to the kernel test and to strace that uses this interface. > > Is MIPS n32 the only psABI across all our architectures supported that > can have `long long' syscall arguments? I guess it might actually be the > case, in which case I won't be surprised it needs specific handling. This is very similar to x32, with the only essential difference: on x32 the 64-bitness of syscall arguments is exposed to userspace via __kernel_ulong_t: arch/x86/include/uapi/asm/posix_types_x32.h:typedef unsigned long long __kernel_ulong_t; In fact, there is a workaround in strace for this case, and now I realized that it ceased to work on mips n32 long time ago: # if defined HAVE___KERNEL_LONG_T && defined HAVE___KERNEL_ULONG_T # include <asm/posix_types.h> typedef __kernel_long_t kernel_long_t; typedef __kernel_ulong_t kernel_ulong_t; # elif (defined __x86_64__ && defined __ILP32__) || defined LINUX_MIPSN32 typedef long long kernel_long_t; typedef unsigned long long kernel_ulong_t; # else typedef long kernel_long_t; typedef unsigned long kernel_ulong_t; # endif -- ldv
