Re: [PATCH 2/5] x86: ret_from_fork - get rid of jump back

[Ingo Molnar]

* Andi Kleen <andi@xxxxxxxxxxxxxx> wrote:

> > But it gets worse than that: checking for _TIF_SYSCALL_TRACE is 
> > completely unnecessary here because we clear that flag for every 
> 
> That's true. I found your 2005 changeset which did that, [...]

that's irrelevant, because all the necessary TIF_ flag processing is 
already done in ret_from_sys_call.

The unnecessary TIF_SYSCALL_TRACE code in the 64-bit ret_from_fork was 
apparently added by you in 2002:

  +ENTRY(ret_from_fork)
  +       movq %rbx, %rdi
  +       call schedule_tail
  +       GET_THREAD_INFO(%rcx)
  +       bt $TIF_SYSCALL_TRACE,threadinfo_flags(%rcx)
  +       jc rff_trace

when you added it to x86_64 via the changeset quoted below.

at that time the 32-bit code (from which you copied the x86_64 tree, 
creating the whole split x86 trees approach) had this simple and 
straightforward ret_from_fork implementation:

  ENTRY(ret_from_fork)
  #if CONFIG_SMP
        pushl %ebx
        call SYMBOL_NAME(schedule_tail)
        addl $4, %esp
  #endif
        GET_THREAD_INFO(%ebx)
        jmp syscall_exit

so by all means the rff_trace/rff_action hack came from you.

	Ingo

------------------>