net: ethernet: ti: am65-cpsw: use napi_complete_done() in TX completion
Subsystems (2)
cover net: ethernet: ti: am65-cpsw: use napi_complete_done() in TX completion
hi
The intention of this series is to fully enable hard irqs deferral feature (hrtimers based HW IRQ coalescing) from Eric Dumazet [1] for TI K3 CPSW driver by using napi_complete_done() in TX completion path, so the combination of parameters (/sys/class/net/ethX/):
napi_defer_hard_irqs
gro_flush_timeout
can be used for hard irqs deferral.
The Patch 1 is required before enabling hard irqs deferral feature to avoid "Unbalanced enable" issue if gro_flush_timeout is configured while (napi_defer_hard_irqs == 0).
It's a bit sad that it can not be configured per RX/TX separately.
[1] https://lore.kernel.org/netdev/[email protected]/
Grygorii Strashko (1):
net: ethernet: ti: am65-cpsw: use napi_complete_done() in TX
completion
Vignesh Raghavendra (1):
net: ti: am65-cpsw-nuss: fix RX IRQ state after .ndo_stop()
drivers/net/ethernet/ti/am65-cpsw-nuss.c | 23 ++++++++++++++++------- drivers/net/ethernet/ti/am65-cpsw-nuss.h | 2 ++ 2 files changed, 18 insertions(+), 7 deletions(-)
Signature · 1 line
2.17.1
PATCH 1/2 net: ti: am65-cpsw-nuss: fix RX IRQ state after .ndo_stop()
From: Vignesh Raghavendra <[email protected]>
On TI K3 am64x platform the issue with RX IRQ is observed - it's become disabled forever after .ndo_stop(). The K3 CPSW driver manipulates RX IRQ by using standard Linux enable_irq()/disable_irq_nosync() API as there is no IRQ enable/disable options in CPSW HW itself, as result during .ndo_stop() following sequence happens
phy_stop()
teardown TX/RX channels
wait for TX tdown complete
napi_disable(TX)
clean up TX channels
(a)
napi_disable(RX)
At point (a) it's not possible to predict if RX IRQ was triggered or not. if RX IRQ was triggered then it also not possible to definitely say if RX NAPI was run or only scheduled and immediately canceled by napi_disable(RX). Actually the last case causes RX IRQ to be permanently disabled.
Another observed issue is that RX IRQ enable counter become unbalanced if (gro_flush_timeout =! 0) while (napi_defer_hard_irqs == 0):
Unbalanced enable for IRQ 44
WARNING: CPU: 0 PID: 10 at ../kernel/irq/manage.c:776 __enable_irq+0x38/0x80 __enable_irq+0x38/0x80 enable_irq+0x54/0xb0 am65_cpsw_nuss_rx_poll+0x2f4/0x368 __napi_poll+0x34/0x1b8 net_rx_action+0xe4/0x220 _stext+0x11c/0x284 run_ksoftirqd+0x4c/0x60
To avoid above issues introduce flag indicating if RX was actually disabled before enabling it in am65_cpsw_nuss_rx_poll() and restore RX IRQ state in .ndo_open()
--- drivers/net/ethernet/ti/am65-cpsw-nuss.c | 13 +++++++++++-- drivers/net/ethernet/ti/am65-cpsw-nuss.h | 2 ++ 2 files changed, 13 insertions(+), 2 deletions(-)
drivers/net/ethernet/ti/am65-cpsw-nuss.c +11 -2
@@ -519,6 +519,10 @@ static int am65_cpsw_nuss_common_open(struct am65_cpsw_common *common, }napi_enable (&common->napi_rx); +if (common->rx_irq_disabled) { + common->rx_irq_disabled = false; +enable_irq (common->rx_chns.irq); + }dev_dbg (common->dev,"cpsw_nuss started\n" );return 0 ;
@@ -872,8 +876,12 @@ static int am65_cpsw_nuss_rx_poll(struct napi_struct *napi_rx, int budget)dev_dbg (common->dev,"%s num_rx:%d %d\n" , __func__, num_rx, budget); -if (num_rx < budget &&napi_complete_done (napi_rx, num_rx)) -enable_irq (common->rx_chns.irq); +if (num_rx < budget &&napi_complete_done (napi_rx, num_rx)) { +if (common->rx_irq_disabled) { + common->rx_irq_disabled = false; +enable_irq (common->rx_chns.irq); + } + }return num_rx; }
@@ -1091,6 +1099,7 @@ static irqreturn_t am65_cpsw_nuss_rx_irq(int irq, void *dev_id) {struct am65_cpsw_common *common = dev_id; + common->rx_irq_disabled = true;disable_irq_nosync (irq);napi_schedule (&common->napi_rx);
drivers/net/ethernet/ti/am65-cpsw-nuss.h +2 -0
@@ -126,6 +126,8 @@ struct am65_cpsw_common {struct am65_cpsw_rx_chn rx_chns;struct napi_struct napi_rx; +bool rx_irq_disabled; +u32 nuss_ver;u32 cpsw_ver;unsigned long bus_freq;
Signature · 1 line
2.17.1
PATCH 2/2 net: ethernet: ti: am65-cpsw: use napi_complete_done() in TX completion
In reply to Grygorii StrashkoThis patch enables support for hard irqs deferral feature from Eric Dumazet
[1] for TI K3 CPSW driver by using napi_complete_done() in TX completion path.
Depending on gro_flush_timeout and napi_defer_hard_irqs at gives up to 30% CPU utilization reduction:
gro_flush_timeout=50000
napi_defer_hard_irqs=2
netperf -l 10 -H 192.168.1.1 -t UDP_STREAM -c -C -- -m 1470
MIGRATED UDP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 192.168.1.1 () port 0 AF_INET
Socket Message Elapsed Messages CPU Service Size Size Time Okay Errors Throughput Util Demand bytes bytes secs # # 10^6bits/sec % SS us/KB
before:
212992 1470 10.00 809632 0 952.0 42.98 14.792 212992 10.00 809630 952.0 50.66 8.719
after:
212992 1470 10.00 813686 0 956.8 32.14 11.009 212992 10.00 813686 956.8 50.05 8.570
[1] https://lore.kernel.org/netdev/[email protected]/
--- drivers/net/ethernet/ti/am65-cpsw-nuss.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-)
drivers/net/ethernet/ti/am65-cpsw-nuss.c +5 -5
@@ -1086,13 +1086,13 @@ static int am65_cpsw_nuss_tx_poll(struct napi_struct *napi_tx, int budget)else num_tx =am65_cpsw_nuss_tx_compl_packets (tx_chn->common, tx_chn->id, budget); - num_tx =min (num_tx, budget); -if (num_tx < budget) { -napi_complete (napi_tx); +if (num_tx >= budget) +return budget; + +if (napi_complete_done (napi_tx, num_tx))enable_irq (tx_chn->irq); - } -return num_tx; +return 0 ; }static irqreturn_t am65_cpsw_nuss_rx_irq (int irq,void *dev_id)
Signature · 1 line
2.17.1
This series was applied to netdev/net-next.git (refs/heads/master):
Hello:
This series was applied to netdev/net-next.git (refs/heads/master):
Quoted 11 lines · On Fri, 6 Aug 2021 01:55:30 +0300 you wrote:
hi
The intention of this series is to fully enable hard irqs deferral feature (hrtimers based HW IRQ coalescing) from Eric Dumazet [1] for TI K3 CPSW driver by using
napi_complete_done()in TX completion path, so the combination of parameters (/sys/class/net/ethX/):
napi_defer_hard_irqs
gro_flush_timeout
can be used for hard irqs deferral.[...]
Here is the summary with links:
- [net-next,1/2] net: ti: am65-cpsw-nuss: fix RX IRQ state after .ndo_stop()
https://git.kernel.org/netdev/net-next/c/47bfc4d128de - [net-next,2/2] net: ethernet: ti: am65-cpsw: use napi_complete_done() in TX completion https://git.kernel.org/netdev/net-next/c/3bacbe04251b
You are awesome, thank you!
--
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html