| Message ID | 9d94f2d8-d297-7550-2932-793a34e5efb9@gmail.com (mailing list archive) |
|---|---|
| State | Changes Requested |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | [net-next] r8169: enable GRO software interrupt coalescing per default | expand |
On Sat, 26 Nov 2022 15:07:07 +0100 Heiner Kallweit wrote: > There are reports about r8169 not reaching full line speed on certain > systems (e.g. SBC's) with a 2.5Gbps link. > There was a time when hardware interrupt coalescing was enabled per > default, but this was changed due to ASPM-related issues on few systems. > > Meanwhile we have sysfs attributes for controlling kind of > "software interrupt coalescing" on the GRO level. However most distros > and users don't know about it. So lets set a conservative default for > both involved parameters. Users can still override the defaults via > sysfs. Don't enable these settings on the fast ethernet chip versions, > they are slow enough. > > Even with these conservative setting interrupt load on my 1Gbps test > system reduced significantly. Sure, why not. Could you please wrap the init into a helper? Should help us ensure the params are not wildly different between drivers and make any later refactoring easier. Maybe something like: /** * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default * @dev: netdev to enable the IRQ coalescing on * bla bla bla */ int netdev_sw_irq_coalesce_default_on(struct net_device *dev) { WARN_ON(dev->reg_state != NETREG_UNREGISTERED); dev->gro_flush_timeout = 20000; dev->napi_defer_hard_irqs = 1; } EXPORT...
On 29.11.2022 04:07, Jakub Kicinski wrote: > On Sat, 26 Nov 2022 15:07:07 +0100 Heiner Kallweit wrote: >> There are reports about r8169 not reaching full line speed on certain >> systems (e.g. SBC's) with a 2.5Gbps link. >> There was a time when hardware interrupt coalescing was enabled per >> default, but this was changed due to ASPM-related issues on few systems. >> >> Meanwhile we have sysfs attributes for controlling kind of >> "software interrupt coalescing" on the GRO level. However most distros >> and users don't know about it. So lets set a conservative default for >> both involved parameters. Users can still override the defaults via >> sysfs. Don't enable these settings on the fast ethernet chip versions, >> they are slow enough. >> >> Even with these conservative setting interrupt load on my 1Gbps test >> system reduced significantly. > > Sure, why not. Could you please wrap the init into a helper? > Should help us ensure the params are not wildly different between > drivers and make any later refactoring easier. > > Maybe something like: > > /** > * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default > * @dev: netdev to enable the IRQ coalescing on > * bla bla bla > */ > int netdev_sw_irq_coalesce_default_on(struct net_device *dev) > { > WARN_ON(dev->reg_state != NETREG_UNREGISTERED); > > dev->gro_flush_timeout = 20000; > dev->napi_defer_hard_irqs = 1; > } > EXPORT... Great, you did the most difficult work and chose a proper function name. I followed your suggestion and put this (slightly adjusted) function into net core. Just submitted the series including it.
diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c index 5bc1181f8..1cb51ec71 100644 --- a/drivers/net/ethernet/realtek/r8169_main.c +++ b/drivers/net/ethernet/realtek/r8169_main.c @@ -5282,6 +5282,12 @@ static int rtl_init_one(struct pci_dev *pdev, const struct pci_device_id *ent) dev->hw_features |= NETIF_F_RXALL; dev->hw_features |= NETIF_F_RXFCS; + if (tp->supports_gmii) { + /* enable software interrupt coalescing per default */ + dev->gro_flush_timeout = 20000; + dev->napi_defer_hard_irqs = 1; + } + /* configure chip for default features */ rtl8169_set_features(dev, dev->features);
There are reports about r8169 not reaching full line speed on certain systems (e.g. SBC's) with a 2.5Gbps link. There was a time when hardware interrupt coalescing was enabled per default, but this was changed due to ASPM-related issues on few systems. Meanwhile we have sysfs attributes for controlling kind of "software interrupt coalescing" on the GRO level. However most distros and users don't know about it. So lets set a conservative default for both involved parameters. Users can still override the defaults via sysfs. Don't enable these settings on the fast ethernet chip versions, they are slow enough. Even with these conservative setting interrupt load on my 1Gbps test system reduced significantly. Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com> --- drivers/net/ethernet/realtek/r8169_main.c | 6 ++++++ 1 file changed, 6 insertions(+)