Files
linux/drivers
Dragos Tatulea df5135fced net/mlx5e: SHAMPO, Allow high order pages in zerocopy mode
Allow high order pages only when SHAMPO mode is enabled (hw-gro) and the
queue is used for zerocopy (has memory provider ops set). The limit is
128K and it was chosen for the following reasons:
- 256K size requires a special case during MTT calculation to split the
  page in two. That's because two MTTs are needed to form an octword.
- Higher sizes require increasing WQE size and/or reducing the number
  of WQEs.
- Having the RQ lined with too few large pages can lead to refill
  issues.

Results show an increase in BW and a decrease in CPU usage.
The benchmark was done with the zcrx samples from liburing [0].

rx_buf_len=4K, oncpu [1]:
packets=3358832 (MB=820027), rps=55794 (MB/s=13621)
Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
Average:       9    1.56    0.00   18.09   13.42    0.00   66.80    0.00    0.00    0.00    0.12

rx_buf_len=128K, oncpu [2]:
packets=3781376 (MB=923187), rps=62813 (MB/s=15335)
Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
Average:       9    0.33    0.00    7.61   18.86    0.00   73.08    0.00    0.00    0.00    0.12

rx_buf_len=4K, offcpu [3]:
packets=3460368 (MB=844816), rps=57481 (MB/s=14033)
Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
Average:       9    0.00    0.00    0.26    0.00    0.00   92.63    0.00    0.00    0.00    7.11
Average:      11    3.04    0.00   68.09   28.87    0.00    0.00    0.00    0.00    0.00    0.00

rx_buf_len=128K, offcpu [4]:
packets=4119840 (MB=1005820), rps=68435 (MB/s=16707)
Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
Average:       9    0.00    0.00    0.87    0.00    0.00   63.77    0.00    0.00    0.00   35.36
Average:      11    1.96    0.00   43.68   54.37    0.00    0.00    0.00    0.00    0.00    0.00

[0] https://github.com/isilence/liburing/tree/zcrx/rx-buf-len

[1] commands:
  $> taskset -c 9 ./zcrx 6 -i eth2 -q 9 -A 1 -B 4096 -S 33554432
  $> ./send-zerocopy tcp -6 -D 2001:db8::1 -t 60 -C 0 -l 1 -b 1 -n 1 -z 1 -d -s 256000

[2] commands:
  $> taskset -c 9 ./zcrx 6 -i eth2 -q 9 -A 1 -B 131072 -S 33554432
  $> ./send-zerocopy tcp -6 -D 2001:db8::1 -t 60 -C 0 -l 1 -b 1 -n 1 -z 1 -d -s 256000

[3] commands:
  $> taskset -c 11 ./zcrx 6 -i eth2 -q 9 -A 1 -B 4096 -S 33554432
  $> ./send-zerocopy tcp -6 -D 2001:db8::1 -t 60 -C 0 -l 1 -b 1 -n 1 -z 1 -d -s 256000

[4] commands:
  $> taskset -c 11 ./zcrx 6 -i eth2 -q 9 -A 1 -B 131072 -S 33554432
  $> ./send-zerocopy tcp -6 -D 2001:db8::1 -t 60 -C 0 -l 1 -b 1 -n 1 -z 1 -d -s 256000

Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260223204155.1783580-16-tariqt@nvidia.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
2026-02-26 10:54:24 +01:00
..
2026-01-22 17:08:06 +08:00
2026-01-27 15:54:09 +01:00
2026-02-09 12:21:32 -05:00