net: fs_enet: Remove non NAPI RX and add NAPI for TX
Subsystems (2)
cover net: fs_enet: Remove non NAPI RX and add NAPI for TX
When using a MPC8xx as a router, 'perf' shows a significant time spent in fs_enet_interrupt() and fs_enet_start_xmit().
'perf annotate' shows that the time spent in fs_enet_start_xmit is indeed spent between spin_unlock_irqrestore() and the following instruction, hence in interrupt handling. This is due to the TX complete interrupt that fires after each transmitted packet.
This patchset first remove all non NAPI handling as NAPI has become the only mode for RX, then adds NAPI for handling TX complete.
This improves NAT TCP throughput by 21% on MPC885 with FEC.
Tested on MPC885 with FEC.
[PATCH 1/2] net: fs_enet: Remove non NAPI RX
[PATCH 2/2] net: fs_enet: Add NAPI TX
--- .../net/ethernet/freescale/fs_enet/fs_enet-main.c | 211 ++++++--------------- .../net/ethernet/freescale/fs_enet/fs_enet.h | 9 +- .../net/ethernet/freescale/fs_enet/mac-fcc.c | 29 +++ .../net/ethernet/freescale/fs_enet/mac-fec.c | 29 +++ .../net/ethernet/freescale/fs_enet/mac-scc.c | 29 +++ linux/include/linux/fs_enet_pd.h | 1 - 6 files changed, 147 insertions(+), 161 deletions(-)
David Miller
maintainerFrom: Christophe Leroy <[email protected]>
Date: Tue, 7 Oct 2014 15:04:53 +0200 (CEST)
Quoted 10 lines
When using a MPC8xx as a router, 'perf' shows a significant time spent in
fs_enet_interrupt()andfs_enet_start_xmit().
'perf annotate' shows that the time spent infs_enet_start_xmitis indeed spent betweenspin_unlock_irqrestore()and the following instruction, hence in interrupt handling. This is due to the TX complete interrupt that fires after each transmitted packet.
This patchset first remove all non NAPI handling as NAPI has become the only mode for RX, then adds NAPI for handling TX complete.
This improves NAT TCP throughput by 21% on MPC885 with FEC.
Quoted 6 lines
Tested on MPC885 with FEC.
[PATCH 1/2] net:
fs_enet: Remove non NAPI RX
[PATCH 2/2] net:fs_enet: Add NAPI TXSigned-off-by Christophe Leroy <[email protected]>
Series applied, thanks.
Any particular reason you didn't just put the TX reclaim calls into the existing NAPI handler?
That's what other drivers do, because TX reclaim can make SKBs available for RX packet receive on the local cpu. So generally you have one NAPI context that first does any pending TX reclaim, then polls the RX ring for new packets.
Quoted 15 lines · David Miller
From: Christophe Leroy <[email protected]>
Date: Tue, 7 Oct 2014 15:04:53 +0200 (CEST)Quoted 12 lines
When using a MPC8xx as a router, 'perf' shows a significant time spent in
fs_enet_interrupt()andfs_enet_start_xmit().
'perf annotate' shows that the time spent infs_enet_start_xmitis indeed spent betweenspin_unlock_irqrestore()and the following instruction, hence in interrupt handling. This is due to the TX complete interrupt that fires after each transmitted packet.
This patchset first remove all non NAPI handling as NAPI has become the only mode for RX, then adds NAPI for handling TX complete.
This improves NAT TCP throughput by 21% on MPC885 with FEC.Tested on MPC885 with FEC.
Quoted 8 lines
Quoted 4 lines
[PATCH 1/2] net:
fs_enet: Remove non NAPI RX
[PATCH 2/2] net:fs_enet: Add NAPI TXSigned-off-by Christophe Leroy <[email protected]>Series applied, thanks.
Any particular reason you didn't just put the TX reclaim calls into the existing NAPI handler?
Not really. I used the gianfar.c driver as a model.
Quoted 6 lines
That's what other drivers do, because TX reclaim can make SKBs available for RX packet receive on the local cpu. So generally you have one NAPI context that first does any pending TX reclaim, then polls the RX ring for new packets.
Is that a better approach ?