On Mon, 17 Nov 2008, Paul Mackerras wrote:
> Steven Rostedt writes:
>
> > +static int test_24bit_addr(unsigned long ip, unsigned long addr)
> > +{
> > + unsigned long diff;
> > +
> > + /* can we get to addr from ip in 24 bits? */
> > + diff = ip > addr ? ip - addr : addr - ip;
> > +
> > + return !(diff & ((unsigned long)-1 << 24));
>
> Since a branch instruction can reach +/- 32MB, the 24 needs to be 25.
> It's 25 not 26 since you have effectively accounted for one bit by
> taking the absolute value of the offset. Also, this will say that an
> offset of -0x2000000 is out of range whereas it is just in range.
> That probably doesn't matter unless you're relying on this to give
> exactly the same answer in all cases as tests done elsewhere (e.g. in
> module_64.c).
I'll update this to be correct.
>
> Note that the address difference being too large isn't the only reason
> for a procedure call to go via a trampoline on ppc64. A trampoline is
> also needed when the called function uses a different TOC from the
> caller. The most obvious example of this is when the caller is in a
> module and the callee is in the main kernel or another module.
>
> Currently all functions in the main kernel use the same TOC, but the
> linker is capable of dividing up a large executable into groups of
> functions and assigning a different TOC to each group, in order to
> avoid any of the TOCs getting too large (a TOC is limited to 64kB).
> If the linker does that, it automatically inserts trampolines to
> handle the TOC change.
OK, I'll deal with that at a another time. If this does happen in core
kernel code, it will produce a nice warning and nicely shutdown ftrace,
without causing any harm. Then I can deal with the bug report.
-- Steve
_______________________________________________
Linuxppc-dev mailing list
Linuxppc-dev@ozlabs.org
https://ozlabs.org/mailman/listinfo/linuxppc-dev
