Re: [PATCH v8 RESEND 6/6] r8169: Disable ASPM while doing NAPI poll

[Kai-Heng Feng]

On Wed, Feb 22, 2023 at 9:03 PM Kai-Heng Feng
<kai.heng.feng@xxxxxxxxxxxxx> wrote:
>
> On Tue, Feb 21, 2023 at 7:09 PM Heiner Kallweit <hkallweit1@xxxxxxxxx> wrote:
> >
> > On 21.02.2023 03:38, Kai-Heng Feng wrote:
> > > NAPI poll of Realtek NICs don't seem to perform well ASPM is enabled.
> > > The vendor driver uses a mechanism called "dynamic ASPM" to toggle ASPM
> > > based on the packet number in given time period.
> > >
> > > Instead of implementing "dynamic ASPM", use a more straightforward way
> > > by disabling ASPM during NAPI poll, as a similar approach was
> > > implemented to solve slow performance on Realtek wireless NIC, see
> > > commit 24f5e38a13b5 ("rtw88: Disable PCIe ASPM while doing NAPI poll on
> > > 8821CE").
> > >
> > > Since NAPI poll should be handled as fast as possible, also remove the
> > > delay in rtl_hw_aspm_clkreq_enable() which was added by commit
> > > 94235460f9ea ("r8169: Align ASPM/CLKREQ setting function with vendor
> > > driver").
> > >
> > > Signed-off-by: Kai-Heng Feng <kai.heng.feng@xxxxxxxxxxxxx>
> > > ---
> > > v8:
> > >  - New patch.
> > >
> > >  drivers/net/ethernet/realtek/r8169_main.c | 14 ++++++++++++--
> > >  1 file changed, 12 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c
> > > index 897f90b48bba6..4d4a802346ae3 100644
> > > --- a/drivers/net/ethernet/realtek/r8169_main.c
> > > +++ b/drivers/net/ethernet/realtek/r8169_main.c
> > > @@ -2711,8 +2711,6 @@ static void rtl_hw_aspm_clkreq_enable(struct rtl8169_private *tp, bool enable)
> > >               RTL_W8(tp, Config2, RTL_R8(tp, Config2) & ~ClkReqEn);
> > >               RTL_W8(tp, Config5, RTL_R8(tp, Config5) & ~ASPM_en);
> > >       }
> > > -
> > > -     udelay(10);
> > >  }
> > >
> > >  static void rtl_set_fifo_size(struct rtl8169_private *tp, u16 rx_stat,
> > > @@ -4577,6 +4575,12 @@ static int rtl8169_poll(struct napi_struct *napi, int budget)
> > >       struct net_device *dev = tp->dev;
> > >       int work_done;
> > >
> > > +     if (tp->aspm_manageable) {
> > > +             rtl_unlock_config_regs(tp);
> >
> > NAPI poll runs in softirq context (except for threaded NAPI).
> > Therefore you should use a spinlock instead of a mutex.
>
> You are right. Will change it in next revision.
>
> >
> > > +             rtl_hw_aspm_clkreq_enable(tp, false);
> > > +             rtl_lock_config_regs(tp);
> > > +     }
> > > +
> > >       rtl_tx(dev, tp, budget);
> > >
> > >       work_done = rtl_rx(dev, tp, budget);
> > > @@ -4584,6 +4588,12 @@ static int rtl8169_poll(struct napi_struct *napi, int budget)
> > >       if (work_done < budget && napi_complete_done(napi, work_done))
> > >               rtl_irq_enable(tp);
> > >
> > > +     if (tp->aspm_manageable) {
> > > +             rtl_unlock_config_regs(tp);
> > > +             rtl_hw_aspm_clkreq_enable(tp, true);
> > > +             rtl_lock_config_regs(tp);
> >
> > Why not moving lock/unlock into rtl_hw_aspm_clkreq_enable()?
>
> Because where it gets called at other places don't need the lock.
> But yes this will make it easier to read, will do in next revision.

We can't do that because it creates deadlock:
rtl_hw_start()
  rtl_unlock_config_regs()
  rtl_hw_start_8168()
  rtl_hw_config()
    rtl_hw_start_8168h_1()
      rtl_hw_aspm_clkreq_enable()

Kai-Heng

>
> Kai-Heng
>
> >
> > > +     }
> > > +
> > >       return work_done;
> > >  }
> > >
> >