On Thu, Dec 07, 2023 at 04:35:47PM +0300, Serge Semin wrote:
> Hi Andrew
>
> On Wed, Dec 06, 2023 at 06:01:30PM +0100, Andrew Lunn wrote:
> > > > You shouldn't use inline in C files, only in headers.
> > >
> > > Could you please clarify why? I failed to find such requirement in the
> > > coding style doc. Moreover there are multiple examples of using the
> > > static-inline-ers in the C files in the kernel including the net/mdio
> > > subsystem.
> >
>
> > The compiler does a better job at deciding what to inline than we
> > humans do. If you can show the compiler is doing it wrong, then we
> > might accept them.
>
> In general I would have agreed with you especially if the methods were
> heavier than what they are:
> static inline ptrdiff_t dw_xpcs_mmio_addr_format(int dev, int reg)
> {
> return FIELD_PREP(0x1f0000, dev) | FIELD_PREP(0xffff, reg);
> }
>
> static inline u16 dw_xpcs_mmio_addr_page(ptrdiff_t csr)
> {
> return FIELD_GET(0x1fff00, csr);
> }
>
> static inline ptrdiff_t dw_xpcs_mmio_addr_offset(ptrdiff_t csr)
> {
> return FIELD_GET(0xff, csr);
> }
>
> > But in general, netdev does not like inline in .C
> > file.
>
> I see. I'll do as you say if you don't change your mind after my
> reasoning below.
>
> > Also, nothing in MDIO is hot path, it spends a lot of time
> > waiting for a slow bus. So inline is likely to just bloat the code for
> > no gain.
>
> I would have been absolutely with you in this matter, if we were talking
> about a normal MDIO bus. In this case the devices are accessed over
> the system IO-memory. So the bus isn't that slow.
O.K, so its not slow. But how often is it used? PHYs tend to get
polled once a second if interrupts are not used. But is the PCS also
polled at the same time? Does this optimisation make a noticeable
difference at once per second? Do you have a requirement that the
system boots very very fast, and this optimisation makes a difference
when there is heavier activity on the PCS at boot? Is the saving
noticeable, when auto-neg takes a little over 1 second.
The best way to make your case is show real world requirements and
benchmark results.
Andrew
