On 5/25/2010 11:21 AM, Arnd Bergmann wrote (in private email):
The hard part to applying this approach turned out to be the COMPAT code
for our 64-bit platform. The approach I am using now is to extend
<linux/compat.h> with all the compat syscalls that are not currently
prototyped, and then to include a a set of #defines that allow all the
compat syscalls to be invoked as "compat_sys_foo()", e.g.
+/* Standard Linux functions that don't have "compat" versions. */
+#define compat_sys_accept sys_accept
+#define compat_sys_accept4 sys_accept4
+#define compat_sys_access sys_access
+#define compat_sys_acct sys_acct
[...]
+#define compat_sys_uselib sys_uselib
+#define compat_sys_vfork sys_vfork
+#define compat_sys_vhangup sys_vhangup
+#define compat_sys_write sys_write
With that in place, you can then use the "arch/score" mechanism to
generate not just the main syscall table, but the compat table as well,
by doing something like this:
+#undef __SYSCALL
+#define __SYSCALL(nr, call) [nr] = (compat_##call),
+
+void *compat_sys_call_table[__NR_syscalls] = {
+ [0 ... __NR_syscalls-1] = sys_ni_syscall,
+#include <asm/unistd.h>
+};
To make this really work out, I also had to add a __SYSCALL_COMPAT
notion to <asm-generic/unistd.h>; when this is set, the __NR_xxx values
and the __SYSCALL stuff are set up as if for a 32-bit platform, even if
the real platform is 64-bit, so that the header can be used to create
the compat_sys_call_table[] properly.
I fixed a few other minor glitches too, like the fact that we need
sys_fadvise64_64 to be the "primary" syscall even in the 64-bit case
(not sys_fadvise64), and adding an __ARCH_WANT_SYNC_FILE_RANGE2 flag to
the generic ABI so platforms can request the use of that flavor of the
ABI instead. (It makes a difference on our platform.) And I took
Arnd's suggestion and added 16 architecture-specific syscalls from 244
to 259.
Note that it turns out not to be quite right to make the
sys_call_table.c a generic file, at least in our case, since you really
want to allow tweaking the actual syscall functions as part of
generating the sys_call_table[] array. For example, on our 32-bit
platforms some of the 64-bit syscalls need wrappers since otherwise
there is a mismatch between the generic code in libc that splits 64-bit
values into 32-bit registers, and the actual registers pairs used by our
ABI for native 64-bit values. In any case it's only half a dozen lines
of common code. And in compat mode there are additional overrides you
want, such as using sys_newstat() for compat_sys_stat64(), if your
architecture will tolerate it, etc.
I'll send a complete patch later once I've finished digesting all the
various suggestions folks have sent, but this was a big enough piece
that I thought I'd at least summarize the design back to LKML in case
people would care to comment.
--
Chris Metcalf, Tilera Corp.
http://www.tilera.com
--
