On Sun, Aug 26, 2018 at 12:32 PM H. Peter Anvin <hpa@xxxxxxxxx> wrote:
>
> Here is a full-blown (user space) test program demonstrating the whole
> technique and how to use it.
So while I agree that some _THIS_IP_ users might be better off being
converted to __builtin_return_address(0) at the caller, I also think
that the whole "notailcall" thing shows why that can easily be more
problematic than just our currnet _THIS_IP_ solution.
Honestly, I'd suggest:
- just do the current_text_addr() to _THIS_IP_ conversion
- keep _THIS_IP_ and make it be the generic one, and screw the whole
"some architectures might implement is better" issue. Nobody cares.
- try to convince people to move away from the "we want the kernel
instruction pointer for the call" model entirely, and consider this a
"legacy" issue.
The whole instruction pointer is a nasty thing. We should discourage
it and not make complex infrastructure for it.
Instead, maybe we could encourage something like
struct kernel_loc { const char *file; const char *fn; int line; };
#define __GEN_LOC__(n) \
({ static const struct kernel_loc n = { \
__FILE__, __FUNCTION__, __LINE__ \
}; &n; })
#define _THIS_LOC_ __GEN_LOC__(__UNIQUE_ID(loc))
which is a hell of a lot nicer to use, and actually allows gcc to
optimize things (try it: if you pass a _THIS_LOC_ off to an inline
function, and that inline function uses the name and line number, gcc
will pick them up directly, without the extra structure dereference.
Wouldn't it be much nicer to pass these kinds of "location pointer"
around, rather than the nasty _THIS_IP_ thing?
Certainly lockdep looks like it could easily take that "const struct
kernel_loc *" instead of "unsigned long ip". Makes it easy to print
out the lockdep info.
Ok, I didn't try to convert anybody, so maybe people who currently use
_THIS_IP_ or current_text_addr() have some fundamental reason why they
want just that, but let's not male _THIS_IP_ more complex than it
needs to be.
Hmm?
Linus
