Re: [PATCH 01/23] all: syscall wrappers: add documentation

[Catalin Marinas]

On Thu, May 26, 2016 at 11:48:19PM +0300, Yury Norov wrote:
> On Wed, May 25, 2016 at 02:28:21PM -0700, David Miller wrote:
> > From: Arnd Bergmann <arnd@xxxxxxxx>
> > Date: Wed, 25 May 2016 23:01:06 +0200
> > 
> > > On Wednesday, May 25, 2016 1:50:39 PM CEST David Miller wrote:
> > >> From: Arnd Bergmann <arnd@xxxxxxxx>
> > >> Date: Wed, 25 May 2016 22:47:33 +0200
> > >> 
> > >> > If we use the normal calling conventions, we could remove these overrides
> > >> > along with the respective special-case handling in glibc. None of them
> > >> > look particularly performance-sensitive, but I could be wrong there.
> > >> 
> > >> You could set the lowest bit in the system call entry pointer to indicate
> > >> the upper-half clears should be elided.
> > > 
> > > Right, but that would introduce an extra conditional branch in the syscall
> > > hotpath, and likely eliminate the gains from passing the loff_t arguments
> > > in a single register instead of a pair.
> > 
> > Ok, then, how much are you really gaining from avoiding a 'shift' and
> > an 'or' to build the full 64-bit value?  3 cycles?  Maybe 4?
> 
> 4 cycles in kernel and ~same cost in glibc to create a pair.

It would take a single instruction per argument in the kernel to do
shift+or and maybe 1-2 more instructions to move the remaining arguments
in place (we do this for a few wrappers in arch/arm64/kernel/entry32.S).
And the glibc counterpart.

> And 8 'mov's that exist for every syscall, even yield().
> 
> > And the executing the wrappers, those have a non-trivial cost too.
> 
> The cost is pretty trivial though. See kernel/compat_wrapper.o:
> COMPAT_SYSCALL_WRAP2(creat, const char __user *, pathname, umode_t, mode);
> 0:   a9bf7bfd        stp     x29, x30, [sp,#-16]!
> 4:   910003fd        mov     x29, sp
> 8:   2a0003e0        mov     w0, w0
> c:   94000000        bl      0 <sys_creat>
> 10:  a8c17bfd        ldp     x29, x30, [sp],#16
> 14:  d65f03c0        ret

I would say the above could be more expensive than 8 movs (16 bytes to
write, read, a branch and a ret). You can also add the I-cache locality,
having wrappers for each syscalls instead of a single place for zeroing
the upper half (where no other wrapper is necessary).

Can we trick the compiler into doing a tail call optimisation. This
could have simply been:

COMPAT_SYSCALL_WRAP2(creat, ...):
	mov	w0, w0
	b	<sys_creat>

> > Cost wise, this seems like it all cancels out in the end, but what
> > do I know?
> 
> I think you know something, and I also think Heiko and other s390 guys
> know something as well. So I'd like to listen their arguments here.
> 
> For me spark64 way is looking reasonable only because it's really simple
> and takes less coding. I'll try it on some branch and share here what happened.

The kernel code will definitely look simpler ;). It would be good to see
if there actually is any performance impact. Even with 16 more cycles on
syscall entry, would they be lost in the noise? You don't need a full
implementation, just some dummy mov x0, x0 on the entry path.

-- 
Catalin