On Mon, May 10, 2021 at 07:58:52AM +0200, Stefano Brivio wrote:
> We don't need a valid MXCSR state for the lookup routines, none of
> the instructions we use rely on or affect any bit in the MXCSR
> register.
>
> Instead of calling kernel_fpu_begin(), we can pass 0 as mask to
> kernel_fpu_begin_mask() and spare one LDMXCSR instruction.
>
> Commit 49200d17d27d ("x86/fpu/64: Don't FNINIT in kernel_fpu_begin()")
> already speeds up lookups considerably, and by dropping the MCXSR
> initialisation we can now get a much smaller, but measurable, increase
> in matching rates.
>
> The table below reports matching rates and a wild approximation of
> clock cycles needed for a match in a "port,net" test with 10 entries
> from selftests/netfilter/nft_concat_range.sh, limited to the first
> field, i.e. the port (with nft_set_rbtree initialisation skipped), run
> on a single AMD Epyc 7351 thread (2.9GHz, 512 KiB L1D$, 8 MiB L2$).
>
> The (very rough) estimation of clock cycles is obtained by simply
> dividing frequency by matching rate. The "cycles spared" column refers
> to the difference in cycles compared to the previous row, and the rate
> increase also refers to the previous row. Results are averages of six
> runs.
>
> Merely for context, I'm also reporting packet rates obtained by
> skipping kernel_fpu_begin() and kernel_fpu_end() altogether (which
> shows a very limited impact now), as well as skipping the whole lookup
> function, compared to simply counting and dropping all packets using
> the netdev hook drop (see nft_concat_range.sh for details). This
> workload also includes packet generation with pktgen and the receive
> path of veth.
Applied, thanks.
