Re: Nice values for kernel modules

[Keith Owens <kaos@ocs.com.au>]

On Sat, 16 Mar 2002 11:27:03 +0000 (GMT), 
Tigran Aivazian <tigran@veritas.com> wrote:
>On Sat, 16 Mar 2002, Paul Mackerras wrote:
>
>> Keith Owens writes:
>>
>> > On Sat, 16 Mar 2002 09:51:16 +0000 (GMT),
>> > <tigran@aivazian.fsnet.co.uk> wrote:
>> > >jump to sys_nice() indirectly via exported sys_call_table[].
>> >
>> > Breaks on ia64 and ppc.
>>
>> Not that I want to encourage this sort of thing, but why would it
>> break on ppc?

Should have been ppc64, not ppc.

>and also why would it break on ia64. I can understand __mips but why ia64?

Address of function text is NOT the same as &function.  On many
architectures &function is the same as the first byte of the function
text but not on all architectures.  On IA64 &func points to a function
descriptor which contains { void * __gp; void * function_text; }.  When
you call a function on ia64, the code is really :-

  save current __gp
  load address of function_text
  load __gp (global data pointer) for new function
  call function_text
  restore original __gp

Within the kernel, all direct function calls (call function by name)
are assumed to have the same __gp so the code reduces to :-

  load address of function_text
  call function_text

just like most other architectures, this is why syscalls within the
kernel work.

When the kernel calls a function via a pointer then __gp must be saved,
set and restored.  This is especially true when calling from the kernel
to a module or vice versa, I guarantee that kernel and modules have
different __gp values.

Fetching &sys_nice from the syscall table and blindly calling that
address from a module will crash the kernel.  PPC64 has a similar
problem, it has a function descriptor that contains 3 fields.

There is no architecture independent method for accessing syscall
entries _as functions_ from a module.  Code that works on ix86 will
break on ia64 and ppc64.  I can see no good reason why the syscall
table has been exported.