On Tue, Oct 31, 2023 at 5:35 AM Johan Hovold johan@kernel.org wrote:
On Mon, Oct 30, 2023 at 04:23:20PM -0700, Bjorn Andersson wrote:
During USB transfers on the SC8280XP __arm_smmu_tlb_sync() is seen to typically take 1-2ms to complete. As expected this results in poor performance, something that has been mitigated by proposing running the iommu in non-strict mode (boot with iommu.strict=0).
This turns out to be related to the SAFE logic, and programming the QOS SAFE values in the DPU (per suggestion from Rob and Doug) reduces the TLB sync time to below 10us, which means significant less time spent with interrupts disabled and a significant boost in throughput.
I ran some tests with a gigabit ethernet adapter to get an idea of how this performs in comparison to using lazy iommu mode ("non-strict"):
6.6 6.6-lazy 6.6-dpu 6.6-dpu-lazyiperf3 recv 114 941 941 941 MBit/s iperf3 send 124 891 703 940 MBit/s
scp recv 14.6 110 110 111 MB/s scp send 12.5 98.9 91.5 110 MB/s
This patch in itself indeed improves things quite a bit, but there is still some performance that can be gained by using lazy iommu mode.
Notably, lazy mode with this patch applied appears to saturate the link in both directions.
Maybe there is still room for SoC specific udev rules so dma masters without firmware can be configured as "lazy", ie. like:
https://chromium.googlesource.com/chromiumos/overlays/board-overlays/+/refs/...
BR, -R
Tested-by: Johan Hovold johan+linaro@kernel.org
Johan