I was writing a benchmark based on sockmap + TCP and discovered several issues:
1. When EAGAIN occurs, the direction of skb is incorrect, causing data loss when retry. 2. When sending partial data, the offset is not recorded, leading to duplicate data being sent when retry. 3. An unexpected BUG_ON() judgment in skb_linearize is triggered. 4. The memory of psock->ingress_skb is not limited by the socket buffer and memcg.
Issues 1, 2, and 3 are described in each patch's commit message.
Regarding issue 4, this patchset does not cover it as it is difficult to handle in practice, and I am still working on it.
Here is a brief description of the issue: When using sockmap to skb/stream redirect, if the receiving end does not perform read operations, all data will be buffered in ingress_skb.
For example: ''' // set memory limit to 50G cgcreate -g memory:myGroup cgset -r memory.max="5000M" myGroup
// start benchmark and disable consumer from reading cgexec -g "memory:myGroup" ./bench sockmap -c 2 -p 1 -a --rx-verdict-ingress --delay-consumer=-1 -d 100 Iter 0 ( 29.179us): Send Speed 2668.548 MB/s (20360.406 calls/s), ... Rcv Speed 0.000 MB/s ( 0.000 calls/s) Iter 1 ( -7.237us): Send Speed 2694.467 MB/s (20557.149 calls/s), ... Rcv Speed 0.000 MB/s ( 0.000 calls/s) Iter 2 ( -1.918us): Send Speed 2693.404 MB/s (20548.039 calls/s), ... Rcv Speed 0.000 MB/s ( 0.000 calls/s) Iter 3 ( -0.684us): Send Speed 2693.138 MB/s (20548.014 calls/s), ... Rcv Speed 0.000 MB/s ( 0.000 calls/s) Iter 4 ( 7.879us): Send Speed 2698.620 MB/s (20588.838 calls/s), ... Rcv Speed 0.000 MB/s ( 0.000 calls/s) Iter 5 ( -3.224us): Send Speed 2696.553 MB/s (20573.066 calls/s), ... Rcv Speed 0.000 MB/s ( 0.000 calls/s) Iter 6 ( -5.409us): Send Speed 2699.705 MB/s (20597.111 calls/s), ... Rcv Speed 0.000 MB/s ( 0.000 calls/s) Iter 7 ( -0.439us): Send Speed 2699.691 MB/s (20597.009 calls/s), ... Rcv Speed 0.000 MB/s ( 0.000 calls/s) ...
// memory usage are not limited cat /proc/slabinfo | grep skb skbuff_small_head 11824024 11824024 704 46 8 : tunables 0 0 0 : slabdata 257044 257044 0 skbuff_fclone_cache 11822080 11822080 512 32 4 : tunables 0 0 0 : slabdata 369440 369440 0 ''' Thus, a simple socket in a large file upload/download model can eat the entire OS memory.
We must charge the skb memory to psock->sk, and if we do not want losing skb, we need to feedback the error info to read_sock/read_skb when the enqueue operation of psock->ingress_skb fails.
--- My another patch related to stability also requires maintainers to spare some time from their busy schedules for review. https://lore.kernel.org/bpf/20250317092257.68760-1-jiayuan.chen@linux.dev/T/...
Jiayuan Chen (4): bpf, sockmap: Fix data lost during EAGAIN retries bpf, sockmap: fix duplicated data transmission bpf, sockmap: Fix panic when calling skb_linearize selftest/bpf/benchs: Add benchmark for sockmap usage
net/core/skmsg.c | 48 +- tools/testing/selftests/bpf/Makefile | 2 + tools/testing/selftests/bpf/bench.c | 4 + .../selftests/bpf/benchs/bench_sockmap.c | 599 ++++++++++++++++++ .../selftests/bpf/progs/bench_sockmap_prog.c | 65 ++ 5 files changed, 697 insertions(+), 21 deletions(-) create mode 100644 tools/testing/selftests/bpf/benchs/bench_sockmap.c create mode 100644 tools/testing/selftests/bpf/progs/bench_sockmap_prog.c