Re: [PATCH] MIPS: tracing: Optimize the implementation

[Wu Zhangjin <wuzhangjin@gmail.com>]

On Fri, 2010-03-12 at 19:43 +0800, Wu Zhangjin wrote:
[...]
> > > + * otherwise, it is in kernel space (0x80000000), return false.
> > > + */
> > > +#define in_module(ip) (unlikely((ip)&  0x40000000))
> > > +
> > 
> > This isn't universally true, but it does hold for most configurations I 
> > think.
> 
> Although I'm not sure who is the exception, we always need an universal
> solution, what about this:
> 
> Compare module with kernel:
> 
> module:
> 
>         <saving registers>
> 
>         lui     v1, hi16_mcount                <--- ip
>         addiu   v1, v1, lo16_mcount
>         move    at, ra
>         jalr    v1
>          nop
> 
> kernel:
> 
>         <saving registers>
> 
>          move    at, ra
>          jal     _mcount                       <--- ip
> 
> The above _ip_ is the address have been recorded into the __mcount_loc
> section of the kernel by scripts/recordmcount.pl, as we can see, for
> kernel, the *(ip - 4) is "move at, ra": 03e0082d, a certain instruction,
> but for module, there is no possibility(?) of existing a "move at, ra"
> at *(ip -4) but a register saving operation("s {d,w} rs, offset(sp)",
> prefixed by 0xffb0 for 64bit and 0xafb0 for 32bit. ), and reversly, for
> kernel, there is no such instruction there.
> 
> And consider the new option -mmcount-ra-address of gcc, some more
> instructions will be inserted between "move at, ra" and the calling site
> to mcount, so, *(ip-4) will not always be "move at, ra", then we need to
> check if there is a "s {d,w} rs, offset(sp)" there, if yes, it is in
> module, otherwise, it should be in kernel.
> 
> #define S_RS_SP          0xafb00000      /* s{d,w} rs, offset(sp) */
> 
> static inline int in_module(ip)
> {
>       insn = *(ip - 4); /* need to use safe_load_code instead, what about big
> endian? */
> 
>       return ((insn & S_RS_SP) == S_RS_SP)
> }

The above method is not available for some cases, to avoid bring Ftrace
with extra overhead, currently, I will keep using the original version
although it may not work for some cases either.

And to let ftrace_make_nop/ftrace_make_call be lightweight, I will also
keep using the "b 1f" method in the old version.

Regards,
        Wu Zhangjin