Re: [PATCH net-next v4] net: skb: prevent the split of kfree_skb_reason() by gcc

From: Menglong Dong
Date: Fri Aug 19 2022 - 10:55:59 EST


Hello,

On Fri, Aug 19, 2022 at 1:00 AM Segher Boessenkool
<segher@xxxxxxxxxxxxxxxxxxx> wrote:
>
> Hi!
>
> On Fri, Aug 19, 2022 at 12:31:44AM +0800, Menglong Dong wrote:
> > On Wed, Aug 17, 2022 at 11:54 PM Nick Desaulniers
> > <ndesaulniers@xxxxxxxxxx> wrote:
> > > Perhaps noipa might also work here?
> >
> > In my testing, both 'noclone' and 'noipa' both work! As for the
> > '-fdisable-ipa-fnsplit', it seems it's not supported by gcc, and I
> > failed to find any documentation of it.
>
> noipa is noinline+noclone+no_icf plus assorted not separately enablable
> things. There is no reason you would want to disable all
> inter-procedural optimisations here, so you don't need noipa.
>
> You need both noinline and no_icf if you want all calls to this to be
> actual function calls, and using this specific function name. If you
> don't have noinline some calls may go missing (which may be fine for
> how you use it). If you don't have no_icf the compiler may replace the
> call with a call to another function, if that does the same thing
> semantically. You may want to prevent that as well, depending on
> exactly what you have this for.
>

Thanks for your explanation about the usage of 'noinline' and 'no_icf'!
I think 'noclone' seems enough in this case? As the function
'kfree_skb_reason' we talk about is a global function, I think that the
compiler has no reason to make it inline, or be merged with another
function.

Meanwhile, I think that the functions which use '__builtin_return_address'
should consider the optimization you mentioned above, and
I'll have a check on them by the way.

Thanks!
Menglong Dong

>
> Segher