AF_XDP (XSK):
The goals of this crate, in order, are:
- Correctness
- Performance
- Ease of use
Current Status
The API works for my current use cases but I expect it will change to achieve higher performance and better usability.
If you have knowledge of Rust FFI and general Rust unsafe things I would greatly appreciate some help auditing this crate as my experience in this area is limited.
Sample Programs
There are several programs in examples/:
l2fwd-1link: Receives frames on a single link/queue, swaps the MAC and writes back to the same link/queue. This is roughly like the kernel xdpsock_user.c sample program in l2fwd mode.l2fwd-2link: Receives frames from two link/queue pairs (separate devices) and forwards the frames to the opposite link.l2fwd-2link-multicore: Multi-core/multi-queue version of l2fwd-2link.dump: Receives frames on a single link/queue, parses with Pnet and prints packet.rxdrop: Receives packets on one or more links and queues and drops then while counting and printing the packet rate.
You can run with cargo run --release --example <example_name> -- [example options] to
see a list of options run cargo run --release --example <example_name> -- --help
Performance
Test System:
- Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz
- Single X710 (4 physical 10G ports) (i40e)
- Kernel: 5.4.2-300.fc31.x86_64
- Kernel boot args: skewtick=1 mitigations=off selinux=0 isolcpus=4-27 nohzfull=4-27 rcunobcs=4-27 defaulthugepagesz=1G hugepagesz=1G hugepages=4
Traffic Generator:
- Intel(R) Core(TM) i7-7700 CPU @ 3.60GHz
- Single X710 (4 physical 10G ports) (i40e)
- Traffic goes from one port, through test system and to second port on the same X710
- T-Rex traffic generator (https://trex-tgn.cisco.com/) (DPDK)
- 64-byte UDP packets
Scenario 1: l2fwd-2link on a single core running userspace and NAPI
Small amounts of packet loss starts at about 6.5M PPS unidirectional and 6.0M PPS bi-directionally (3M each direction).
Little effort has been put into optimizing this so I expect there are some easy performance wins.
Supported Features
- HUGE TLB flag to mmap area: Optional (command line arg in the sample programs)
Required AF_XDP Features
- Need wakeup flag (implies Linux >= 5.4)
- Aligned chunk mode
Supported AF_XDP Features
- ZEROCOPY flag: Optional (command line arg in the sample programs)
- Aligned chunk mode
Unsupported AF_XDP Features (for now)
- Unaligned chunk mode
- Shared Umem (new in Linux 5.10)
- Busy poll (coming in Linux 5.11)
- Non-zero headroom
To Do
- Currently this module is not 'safe' because Bufs can outlive the memory pool they are associated with. I believe fixing this will require adding an Arc to each Buf. I have not had the time yet to determine the performance impact of this and would appreciate any other ideas.
- All interactions with the Tx, Rx, fill and completion queues are done with C functions that wrap the inline functions in libbpf. I believe this means that these hot functions cannot be inlined in Rust. In the medium term we should build pure Rust functions so they can be inlined.
- A more idiomatic Rust interface could be built by talking directly to the kernel vs. wrapping libbpf functions.