netlink, nltest: add ReceiveIter to stream responses

[nickgarlis]

Add ReceiveIter() method that returns an iter.Seq2[Message, error] for iterating over netlink messages without collecting them into a slice. Refactor lockedReceive() to use the new lockedReceiveIter() iterator internally to eliminate code duplication.

I've tested this with google/nftables and got a notable 33% memory pressure reduction on large responses.

This is in an exploratory phase but is looking promising so far. I think that in order to evaluate this approach, the should be addressed:

• Figure out what the behavior should be when the user stops iterating. Should the socket buffer be drained ? Or should it be the user's responsibility ? What happens in an async scenario ?
• Make sure it doesn't disrupt packages using other netlink families (I've tested this change with wgctrl, rtnetlink, go-tc and all tests are passing.)
• Add tests/benchmarks

cc @aojea

[mdlayher]

Hi @nickgarlis, thanks for the contributions. I've added you as a collaborator since you're actively making improvements to the ecosystem. Thank you!

[nickgarlis]

Hi @nickgarlis, thanks for the contributions. I've added you as a collaborator since you're actively making improvements to the ecosystem. Thank you!

Hi @mdlayher, thanks for the invite! I appreciate the trust. I’ll do my best to be helpful and keep things in good shape.

[nickgarlis]

I benchmarked this change using BenchmarkNftablesDump with the current nftables implementation (using Receive), and observed the following results:

goos: linux
goarch: amd64
pkg: github.com/mdlayher/netlink/internal/integration
cpu: 13th Gen Intel(R) Core(TM) i9-13900H
                      │    v1.txt    │               v2.txt               │
                      │    sec/op    │   sec/op     vs base               │
NftablesDump/1-20        65.18µ ± 7%   65.49µ ± 8%       ~ (p=0.796 n=10)
NftablesDump/8-20       103.67µ ± 3%   95.67µ ± 8%  -7.72% (p=0.002 n=10)
NftablesDump/64-20       297.5µ ± 2%   296.0µ ± 3%       ~ (p=0.353 n=10)
NftablesDump/512-20      1.835m ± 3%   1.819m ± 3%       ~ (p=0.436 n=10)
NftablesDump/4096-20     18.45m ± 3%   18.21m ± 3%  -1.29% (p=0.023 n=10)
NftablesDump/32768-20    225.7m ± 4%   225.3m ± 8%       ~ (p=0.971 n=10)
geomean                  1.577m        1.550m       -1.71%

                      │    v1.txt    │               v2.txt                │
                      │     B/op     │     B/op      vs base               │
NftablesDump/1-20       11.18Ki ± 0%   11.34Ki ± 0%  +1.43% (p=0.000 n=10)
NftablesDump/8-20       23.90Ki ± 0%   23.18Ki ± 0%  -3.00% (p=0.000 n=10)
NftablesDump/64-20      125.5Ki ± 0%   122.1Ki ± 0%  -2.69% (p=0.000 n=10)
NftablesDump/512-20     953.5Ki ± 0%   916.4Ki ± 0%  -3.89% (p=0.000 n=10)
NftablesDump/4096-20    7.974Mi ± 0%   7.435Mi ± 0%  -6.77% (p=0.000 n=10)
NftablesDump/32768-20   66.97Mi ± 0%   64.22Mi ± 0%  -4.10% (p=0.000 n=10)
geomean                 511.5Ki        495.1Ki       -3.20%

                      │   v1.txt    │               v2.txt               │
                      │  allocs/op  │  allocs/op   vs base               │
NftablesDump/1-20        146.0 ± 0%    156.0 ± 0%  +6.85% (p=0.000 n=10)
NftablesDump/8-20        407.0 ± 0%    417.0 ± 0%  +2.46% (p=0.000 n=10)
NftablesDump/64-20      2.486k ± 0%   2.494k ± 0%  +0.32% (p=0.000 n=10)
NftablesDump/512-20     19.06k ± 0%   19.02k ± 0%  -0.19% (p=0.000 n=10)
NftablesDump/4096-20    151.6k ± 0%   151.2k ± 0%  -0.26% (p=0.000 n=10)
NftablesDump/32768-20   1.212M ± 0%   1.208M ± 0%  -0.27% (p=0.000 n=10)
geomean                 8.959k        9.089k       +1.45%

I then updated the consumer to use ReceiveIter as shown in this PR google/nftables#357, and got the following results:

goos: linux
goarch: amd64
pkg: github.com/mdlayher/netlink/internal/integration
cpu: 13th Gen Intel(R) Core(TM) i9-13900H
                      │    v1.txt    │                v3.txt                │
                      │    sec/op    │    sec/op     vs base                │
NftablesDump/1-20        65.18µ ± 7%   63.21µ ±  7%        ~ (p=0.739 n=10)
NftablesDump/8-20       103.67µ ± 3%   90.83µ ± 12%  -12.38% (p=0.002 n=10)
NftablesDump/64-20       297.5µ ± 2%   302.2µ ±  5%        ~ (p=0.579 n=10)
NftablesDump/512-20      1.835m ± 3%   1.840m ±  3%        ~ (p=1.000 n=10)
NftablesDump/4096-20     18.45m ± 3%   16.40m ±  4%  -11.11% (p=0.000 n=10)
NftablesDump/32768-20    225.7m ± 4%   201.7m ±  2%  -10.62% (p=0.000 n=10)
geomean                  1.577m        1.481m         -6.04%

                      │    v1.txt    │                v3.txt                │
                      │     B/op     │     B/op      vs base                │
NftablesDump/1-20       11.18Ki ± 0%   11.32Ki ± 0%   +1.25% (p=0.000 n=10)
NftablesDump/8-20       23.90Ki ± 0%   22.06Ki ± 0%   -7.67% (p=0.000 n=10)
NftablesDump/64-20      125.5Ki ± 0%   110.9Ki ± 0%  -11.67% (p=0.000 n=10)
NftablesDump/512-20     953.5Ki ± 0%   820.9Ki ± 0%  -13.91% (p=0.000 n=10)
NftablesDump/4096-20    7.974Mi ± 0%   6.419Mi ± 0%  -19.50% (p=0.000 n=10)
NftablesDump/32768-20   66.97Mi ± 0%   51.34Mi ± 0%  -23.34% (p=0.000 n=10)
geomean                 511.5Ki        445.8Ki       -12.83%

                      │   v1.txt    │               v3.txt               │
                      │  allocs/op  │  allocs/op   vs base               │
NftablesDump/1-20        146.0 ± 0%    158.0 ± 0%  +8.22% (p=0.000 n=10)
NftablesDump/8-20        407.0 ± 0%    413.0 ± 0%  +1.47% (p=0.000 n=10)
NftablesDump/64-20      2.486k ± 0%   2.484k ± 0%  -0.08% (p=0.000 n=10)
NftablesDump/512-20     19.06k ± 0%   19.01k ± 0%  -0.27% (p=0.000 n=10)
NftablesDump/4096-20    151.6k ± 0%   151.2k ± 0%  -0.27% (p=0.000 n=10)
NftablesDump/32768-20   1.212M ± 0%   1.208M ± 0%  -0.27% (p=0.000 n=10)
geomean                 8.959k        9.086k       +1.42%