On Tue, Jun 28, 2016 at 07:32:00PM -0400, Steven Rostedt wrote:
> On Tue, 28 Jun 2016 14:30:40 +0900
> Namhyung Kim <namhy...@kernel.org> wrote:
>
> > Currently ftrace_graph_ent{,_entry} and ftrace_graph_ret{,_entry} struct
> > can have padding bytes at the end due to alignment in 64-bit data type.
> > As these data are recorded so frequently, those paddings waste
> > non-negligible space. As some archs can have efficient unaligned
> > accesses, reducing the alignment can save ~10% of data size:
> >
> > ftrace_graph_ent_entry: 24 -> 20
> > ftrace_graph_ret_entry: 48 -> 44
> >
> > Also I moved the 'overrun' field in struct ftrace_graph_ret to minimize
> > the padding. I think the FTRACE_ALIGNMENT still needs to have proper
> > alignment (even if ring buffer handles the alignment after all) since
> > the ftrace_graph_ent/ret struct is located on stack before copying to
> > the ring buffer.
>
> I don't know. I mean it doesn't hurt to keep the alignment, but I'm
> still thinking that it's overkill. All elements will start on their
> proper alignment anyway.
>
> Think about it, we have:
>
> For 32bit:
>
> struct ftrace_graph_ret {
> unsigned long func; /* Current function */
>
> is at 0-3
>
> /* Number of functions that overran the depth limit for current task */
> unsigned long overrun;
>
> is at 4-7
>
> unsigned long long calltime;
>
> is at 8-15
>
> unsigned long long rettime;
> is at 16-23
>
> int depth;
>
> is at 24-28
>
> And for 64bit:
>
> struct ftrace_graph_ret {
> unsigned long func; /* Current function */
>
> is at 0-7
>
> /* Number of functions that overran the depth limit for current task */
> unsigned long overrun;
>
> is at 8-15
>
> unsigned long long calltime;
>
> is at 16-23
>
> unsigned long long rettime;
>
> is at 24-31
>
> int depth;
>
> is at 32-37
>
> For a total of 38 bytes. I'm betting that without the packed, the 4
> extra bytes will always be at the end.
Woundn't it be 36 or 40 bytes? :)
>
> If the compiler places it incorrectly without any attribute, it will
> fail to read the long long if the arch requires 64 bits to be 8 bytes
> aligned. The alignment is meaningless here. All we need is "packed" and
> be done with it. It's only going to truncate the 4 bytes at the end of
> the structure if that.
I agree that in-struct alignment preserved without the 'aligned'
attribute but I'm not sure whether it's guaranteed that the *start*
address of the struct is still in proper alignment boundary.
IOW the struct ftrace_graph_ret should be placed at 8-byte boundary in order
to keep alignment of struct members. Is it guaranteed after applying
'packed'?
Thanks,
Namhyung
