Re: [PATCH v5 4/5] powerpc/vdso: Wire up getrandom() vDSO implementation on VDSO32

[Jason A. Donenfeld]

On Thu, Sep 05, 2024 at 06:13:29PM +0200, Jason A. Donenfeld wrote:
> > +/*
> > + * The macro sets two stack frames, one for the caller and one for the callee
> > + * because there are no requirement for the caller to set a stack frame when
> > + * calling VDSO so it may have omitted to set one, especially on PPC64
> > + */
> > +
> > +.macro cvdso_call funct
> > +  .cfi_startproc
> > +	PPC_STLU	r1, -PPC_MIN_STKFRM(r1)
> > +  .cfi_adjust_cfa_offset PPC_MIN_STKFRM
> > +	mflr		r0
> > +	PPC_STLU	r1, -PPC_MIN_STKFRM(r1)
> > +  .cfi_adjust_cfa_offset PPC_MIN_STKFRM
> > +	PPC_STL		r0, PPC_MIN_STKFRM + PPC_LR_STKOFF(r1)
> > +  .cfi_rel_offset lr, PPC_MIN_STKFRM + PPC_LR_STKOFF
> > +	get_datapage	r8
> > +	addi		r8, r8, VDSO_RNG_DATA_OFFSET
> > +	bl		CFUNC(DOTSYM(\funct))
> > +	PPC_LL		r0, PPC_MIN_STKFRM + PPC_LR_STKOFF(r1)
> > +	cmpwi		r3, 0
> > +	mtlr		r0
> > +	addi		r1, r1, 2 * PPC_MIN_STKFRM
> > +  .cfi_restore lr
> > +  .cfi_def_cfa_offset 0
> > +	crclr		so
> > +	bgelr+
> > +	crset		so
> > +	neg		r3, r3
> > +	blr
> > +  .cfi_endproc
> > +.endm
> 
> You wrote in an earlier email that this worked with time namespaces, but
> in my testing that doesn't seem to be the case.
> 
> From my test harness [1]:
> 
> Normal single thread
>    vdso: 25000000 times in 12.494133131 seconds
>    libc: 25000000 times in 69.594625188 seconds
> syscall: 25000000 times in 67.349243972 seconds
> Time namespace single thread
>    vdso: 25000000 times in 71.673057436 seconds
>    libc: 25000000 times in 71.712774121 seconds
> syscall: 25000000 times in 66.902318080 seconds
> 
> I'm seeing this on ppc, ppc64, and ppc64le.
> 
> Can you figure out what's going on and send a fix, which I'll squash
> into this commit?

Also, FYI, I've verified that things do work on x86_64, loongarch64,
arm64, and arm64_be. It's just the ppc archs that are broken. So this
test _is_ a good one.