On Fri, Jul 04, 2025 at 11:40:10AM +0100, David Laight wrote:
> On Fri, 4 Jul 2025 05:56:50 +0300
> Jarkko Sakkinen <jar...@kernel.org> wrote:
>
> > On Fri, Jul 04, 2025 at 05:45:11AM +0300, Jarkko Sakkinen wrote:
> ...
> > > Well, that was some truly misguided advice from my side so all the shame
> > > here is on me :-) There's no global memzero() and neither explicit
> > > version makes much sense here. Sorry about that.
> > >
> > > I gave it now (actual) thought, and here's what I'd propose:
> > >
> > > diff --git a/drivers/char/tpm/tpm_crb_ffa.c
> > > b/drivers/char/tpm/tpm_crb_ffa.c
> > > index 96746d5b03e3..e769f6143a7c 100644
> > > --- a/drivers/char/tpm/tpm_crb_ffa.c
> > > +++ b/drivers/char/tpm/tpm_crb_ffa.c
> > > @@ -203,26 +203,20 @@ static int __tpm_crb_ffa_try_send_receive(unsigned
> > > long func_id,
> > > msg_ops = tpm_crb_ffa->ffa_dev->ops->msg_ops;
> > >
> > > if (ffa_partition_supports_direct_req2_recv(tpm_crb_ffa->ffa_dev)) {
> > > - memzero(&tpm_crb_ffa->direct_msg_data2,
> > > - sizeof(struct ffa_send_direct_data2));
> > > -
> > > - tpm_crb_ffa->direct_msg_data2.data[0] = func_id;
> > > - tpm_crb_ffa->direct_msg_data2.data[1] = a0;
> > > - tpm_crb_ffa->direct_msg_data2.data[2] = a1;
> > > - tpm_crb_ffa->direct_msg_data2.data[3] = a2;
> > > + tpm_crb_ffa->direct_msg_data2 = (struct ffa_send_direct_data2){
> > > + .data = { func_id, a0, a1, a2 },
> > > + };
>
> clang has a habit of compiling that as an un-named on-stack structure that
> is initialised and then memcpy() used to copy it into place.
> Often not was intended and blows the stack when the structure is large.
>
> So probably not a pattern that should be encouraged.
This is interesting observation so I had to do some compilation tests to
verify the claim just to see how it plays out (and for the sake of
learning while doing it).
Note that I use GCC for the examples but I have high doubts that clang
would do worse. Please share the insight if that is a wrong assumption.
OK, so... here's the dissembly (using objdump) for the unchanged version:
ffff8000801805a0: 8b020260 add x0, x19, x2
ffff8000801805a4: 94011819 bl ffff8000801c6608 <__memset>
ffff8000801805a8: a9035a75 stp x21, x22, [x19, #48]
ffff8000801805ac: aa1a03e1 mov x1, x26
ffff8000801805b0: aa1903e0 mov x0, x25
ffff8000801805b4: a9047e77 stp x23, xzr, [x19, #64]
[ Off-topic: note that how a2 gets optimized out with the zero register
so that it is probably a parameter that we don't need at all in the
first place? ]
However, in the changed version the matching snippet looks factors
better:
ffff800080180620: a9017c7f stp xzr, xzr, [x3, #16]
ffff800080180624: f900107f str xzr, [x3, #32]
Further, look at the stack size in the original version:
ffff800080180524 <__tpm_crb_ffa_send_receive.constprop.0>:
ffff800080180524: a9ba7bfd stp x29, x30, [sp, #-96]!
On the other hand, in the changed version:
ffff800080180524 <__tpm_crb_ffa_send_receive.constprop.0>:
ffff800080180524: a9bb7bfd stp x29, x30, [sp, #-80]!
I don't know, at least the figures I'm able to measure with my limited
ARM assembly knowledge look way better.
BR, Jarkko`
