On Jul 1, 2014, at 8:07 AM, "Catalin Marinas" [off-list ref] wrote:

quoted

On Sat, May 24, 2014 at 12:02:17AM -0700, Andrew Pinski wrote:

diff --git a/arch/arm64/kernel/entry.S b/arch/arm64/kernel/entry.S
index 1e1ebfc..8241ffe 100644
--- a/arch/arm64/kernel/entry.S
+++ b/arch/arm64/kernel/entry.S

@@ -620,9 +620,14 @@ ENDPROC(ret_from_fork)
 */
   .align    6

el0_svc:
-    adrp    stbl, sys_call_table        // load syscall table pointer
   uxtw    scno, w8            // syscall number in w8
   mov    sc_nr, #__NR_syscalls
+#ifdef CONFIG_ARM64_ILP32
+    get_thread_info tsk
+    ldr    x16, [tsk, #TI_FLAGS]
+    tbnz    x16, #TIF_32BIT_AARCH64, el0_ilp32_svc // We are using ILP32
+#endif
+    adrp    stbl, sys_call_table        // load syscall table pointer

This adds a slight penalty on the AArch64 SVC entry path. I can't tell
whether that's visible or not but I think the x86 guys decided to set an
extra bit to the syscall number to distinguish it from native calls.

quoted

diff --git a/arch/arm64/kernel/sys_ilp32.c b/arch/arm64/kernel/sys_ilp32.c
new file mode 100644
index 0000000..1da1d11
--- /dev/null
+++ b/arch/arm64/kernel/sys_ilp32.c

[...]

quoted

+/*
+ * Wrappers to pass the pt_regs argument.
+ */
+#define sys_rt_sigreturn sys_rt_sigreturn_wrapper
+
+
+/* Using Compat syscalls where necessary */
+#define sys_ioctl        compat_sys_ioctl
+/* iovec */
+#define sys_readv        compat_sys_readv
+#define sys_writev        compat_sys_writev
+#define sys_preadv        compat_sys_preadv64
+#define sys_pwritev        compat_sys_pwritev64
+#define sys_vmsplice        compat_sys_vmsplice

Do these actually work? compat_iovec has two members of 32-bit each
while the ILP32 iovec has a void * (32-bit) and a __kernel_size_t which
is 64-bit.

size_t should be unsigned long in ilp32 so a 32bit unsigned integer type.  That part of the abi was already defined in the arm abi documents.  Now are saying we should pass size_t different between user and kernel space?

quoted

+/* robust_list_head */
+#define sys_set_robust_list    compat_sys_set_robust_list
+#define sys_get_robust_list    compat_sys_get_robust_list

Same here, we have a size_t * argument. The compat function would write
back 32-bit but size_t is 64-bit for ILP32.

See above. Size_t is 32bits. 

quoted

+/* kexec_segment */
+#define sys_kexec_load        compat_sys_kexec_load

More size_t members in the kexec_segment structure (but we don't yet
have kexec on arm64).

See above. Size_t is 32bits. 

quoted

+/* struct msghdr */
+#define sys_recvfrom        compat_sys_recvfrom

Why compat here? struct sockaddr seems to be the same as the native one.

quoted

+#define sys_recvmmsg        compat_sys_recvmmsg
+#define sys_sendmmsg        compat_sys_sendmmsg
+#define sys_sendmsg        compat_sys_sendmsg
+#define sys_recvmsg        compat_sys_recvmsg

These get messier as well with a different size_t affecting struct
msghdr.

See above about size_t. 

quoted

+#define sys_setsockopt        compat_sys_setsockopt
+#define sys_getsockopt        compat_sys_getsockopt

Looking at the sock_getsockopt() function, we have a union v copied
to/from user. However, such a union contains a struct timeval which for
ILP32 would be different than the compat one.

I will look into this one but it might already be taken care of due to the compact uses 64bit time spec define. I will add a comment saying that if it is true. 

quoted

+/* iovec */
+#define sys_process_vm_readv    compat_sys_process_vm_readv
+#define sys_process_vm_writev    compat_sys_process_vm_writev

See above for iovec.

See above for my size_t question. 

quoted

+/* Pointer in struct */
+#define sys_mount               compat_sys_mount

Which structure is this?

NFS structure, I can expand out the comment if needed. 

quoted

+/* Scheduler */
+/* unsigned long bitmaps */
+#define sys_sched_setaffinity   compat_sys_sched_setaffinity
+#define sys_sched_getaffinity   compat_sys_sched_getaffinity

Does the long bitmask matter here? I can see the length is passed in
bytes.

Yes for big endian. If we were only supporting little endian, bit fields would not matter. 

quoted

+/* iov usage */
+#define sys_keyctl              compat_sys_keyctl

Same problem as iovec above.

quoted

+/* aio */
+/* Pointer to Pointer  */
+#define sys_io_setup        compat_sys_io_setup

sys_io_setup takes a pointer to aio_context_t which is defined as
__kernel_ulong_t (same as LP64).

Let me look at why I did this one, I think the code which used aio was not in glibc which is why I used the compat version. 

Thanks,
Andrew

-- 
Catalin