Re: [BUG] perf: arch_perf_out_copy_user default

[Will Deacon <will.deacon@arm.com>]

On Wed, Oct 30, 2013 at 02:37:50PM +0000, Peter Zijlstra wrote:
> Hi Frederic,

Hi Peter,

> I just spotted:
> 
> #ifndef arch_perf_out_copy_user
> #define arch_perf_out_copy_user __copy_from_user_inatomic
> #endif
> 
> vs:
> 
> arch/x86/include/asm/perf_event.h:#define arch_perf_out_copy_user copy_from_user_nmi
> 
> 
> Now the problem is that copy_from_user_nmi() and
> __copy_from_user_inatomic() have different return semantics.
> 
> Furthermore, the macro you use them in DEFINE_OUTPUT_COPY() assumes the
> return value is the amount of memory copied; as also illustrated by
> memcpy_common().
> 
> Trouble is, __copy_from_user_inatomic() returns the number of bytes
> _NOT_ copied.
> 
> With this, my question to Will is, how did your ARM unwind support
> patches ever work? AFAICT they end up using the
> __copy_from_user_inatomic() thing.

Yeah, that's weird, because they *do* appear to work! In fact, looking at
the code in kernel/events/core.c, it looks like __output_copy_user is
expected to return the number of bytes not copied, so providing the
__copy_from_user_inatomic succeeds first time around, the DEFINE_OUTPUT_COPY
macro will return len (dump_size) and the perf_output_skip will deal with
the buffer pointers for us. The issue then is that dynamic size will be 0,
and the unwind code in perf will never be called (except I know that it *is*
being called).

I'll go dig further...

Will