On 3/18/2024 11:39 PM, Kalle Valo wrote:
Karthikeyan Periyasamy <quic_periyasa@xxxxxxxxxxx> writes:
On 3/11/2024 6:35 PM, Kalle Valo wrote:
Karthikeyan Periyasamy <quic_periyasa@xxxxxxxxxxx> writes:
When a packet arrives in Rx rings, the RX descriptor moves from the used
list to the free list. Then, the rxdma ring gets replenished, where the Rx
descriptor again moves from the free list to the used list. At the end, the
descriptor came to the used list with unnecessary list movement. The
descriptor used list is maintained in the Rxdma ring structure, which
creates lock contention for the list operations (add, delete) in the Rx
data path. Optimize the Rx data path by removing the used list from the
common Rxdma ring and maintain as a local variable in the Rx ring handler
itself, which avoid lock contention. Now, to find the used list descriptor
during descriptor cleanup, we need to check the in_use flag for each Rx
descriptor.
Tested-on: QCN9274 hw2.0 PCI WLAN.WBE.1.0.1-00029-QCAHKSWPL_SILICONZ-1
Signed-off-by: Karthikeyan Periyasamy <quic_periyasa@xxxxxxxxxxx>
Before looking at this in detail: When optimising something it would be
good to have before and after numbers showing the improvement. Otherwise
it's just wishful thinking.
I don't have numbers. Like you said, Its just a wishful think.
So do you still want us to take this?
In the future please do provide numbers to show that the optimisation
really helps as intended. Otherwise we might even go backwards.
This is a simple UDP UL throughput test case results on x86+NUC device
with QCN9274 card
Before:
Average: CPU %usr %nice %sys %iowait %irq %soft %steal
%guest %gnice %idle
Average: all 0.24 0.00 12.54 0.08 0.00 23.33
0.00 0.00 0.00 63.81
After:
Average: CPU %usr %nice %sys %iowait %irq %soft %steal
%guest %gnice %idle
Average: all 0.34 0.00 4.60 0.00 0.00 19.59
0.00 0.00 0.00 75.47
is it fine to capture in the commit log ?
--
Karthikeyan Periyasamy
--
கார்த்திகேயன் பெரியசாமி
