Re: [PATCH net-next] net: lan966x: Improve the CPU TX bitrate.

[Andrew Lunn]

On Wed, Mar 09, 2022 at 11:05:16PM +0100, Horatiu Vultur wrote:
> The 03/09/2022 14:11, Andrew Lunn wrote:
> > 
> > On Tue, Mar 08, 2022 at 11:30:00PM +0100, Horatiu Vultur wrote:
> > > The 03/08/2022 22:36, Andrew Lunn wrote:
> > > >
> > > > >  static int lan966x_port_inj_ready(struct lan966x *lan966x, u8 grp)
> > > > >  {
> > > > > -     u32 val;
> > > > > +     unsigned long time = jiffies + usecs_to_jiffies(READL_TIMEOUT_US);
> > > > > +     int ret = 0;
> > > > >
> > > > > -     return readx_poll_timeout_atomic(lan966x_port_inj_status, lan966x, val,
> > > > > -                                      QS_INJ_STATUS_FIFO_RDY_GET(val) & BIT(grp),
> > > > > -                                      READL_SLEEP_US, READL_TIMEOUT_US);
> > > > > +     while (!(lan_rd(lan966x, QS_INJ_STATUS) &
> > > > > +              QS_INJ_STATUS_FIFO_RDY_SET(BIT(grp)))) {
> > > > > +             if (time_after(jiffies, time)) {
> > > > > +                     ret = -ETIMEDOUT;
> > > > > +                     break;
> > > > > +             }
> > > >
> > > > Did you try setting READL_SLEEP_US to 0? readx_poll_timeout_atomic()
> > > > explicitly supports that.
> > >
> > > I have tried but it didn't improve. It was the same as before.
> > 
> > The reason i recommend ipoll.h is that your implementation has the
> > usual bug, which iopoll does not have. Since you are using _atomic()
> > it is less of an issue, but it still exists.
> > 
> >      while (!(lan_rd(lan966x, QS_INJ_STATUS) &
> >               QS_INJ_STATUS_FIFO_RDY_SET(BIT(grp)))) {
> > 
> > Say you take an interrupt here
> > 
> >              if (time_after(jiffies, time)) {
> >                      ret = -ETIMEDOUT;
> >                      break;
> >              }
> > 
> > 
> > The interrupt takes a while, so that by the time you get to
> > time_after(), you have reached your timeout. So -ETIMEDOUT is
> > returned. But in fact, the hardware has done its thing, and if you
> > where to read the status the bit would be set, and you should actually
> > return O.K, not an error.
> 
> That is a good catch and really nice explanation!
> Then if I add also the check at the end, then it should be also OK.

You are then just repeating code which is already in the core. That is
generally not liked. If you find reading the status once works 99% of
the time, then i suggest you call readx_poll_timeout_atomic() when the
status does indicate you need to poll.

       Andrew