Re: [RFC bpf-next v2 0/9] no_caller_saved_registers attribute for helper calls

[Eduard Zingerman <eddyz87@gmail.com>]

On Tue, 2024-07-09 at 18:18 -0700, Alexei Starovoitov wrote:
> On Thu, Jul 4, 2024 at 3:24 AM Eduard Zingerman <eddyz87@gmail.com> wrote:
> > 
> > - stack offsets used for spills/fills are allocated as minimal
> >   stack offsets in whole function and are not used for any other
> >   purposes;
> 
> "minimal stack offset" reads odd to me.
> I noticed the same naming convention is used in llvm diff.
> imo it's odd there as well.
> Maybe say:
> llvm grows the stack that in bpf architecture always grows down and
> picks the lowest stack offset not used by local variables
> and spill/fill.

Will replace "minimal" with lowest here and in LLVM diff.

> > Here is how the program looks after verifier processing:
> > 
> >   # bpftool prog load ./nocsr.bpf.o /sys/fs/bpf/nocsr-test
> >   # bpftool prog dump xlated pinned /sys/fs/bpf/nocsr-test
> >   int test(void * ctx):
> >   ; int test(void *ctx)
> >      0: (bf) r3 = r1               <--------- 3rd printk parameter
> >   ; __u32 task = bpf_get_smp_processor_id();
> >      1: (b4) w0 = 197132           <--------- inlined helper call,
> >      2: (bf) r0 = r0               <--------- spill/fill pair removed
> 
> Are you using old bpftool or something?
> That should have been:
> r0 = &(void __percpu *)(r0)
> ?

Yes, I was using distro-provided bpftool.
Re-running with kernel version of the tool shows the __percpu thing.

> 
> >      3: (61) r0 = *(u32 *)(r0 +0)  <---------

[...]