Receive side scaling (RSS) with eBPF in QEMU and virtio-net
0 likes857 views
The document discusses the integration of Receive Side Scaling (RSS) using eBPF in QEMU and virtio-net, detailing how packets are processed across multiple CPUs. It covers the historical context, technical specifications, implementation challenges, and potential benefits of using eBPF for efficient packet steering. Additionally, it outlines necessary configuration updates to the virtio specifications to support enhancing packet processing performance.
![virtio spec changes - device configuration
• virtio_net_config
struct virtio_net_con
fi
g
{
u8 mac[6]
;
le16 status
;
le16 max_virtqueue_pairs
;
le16 mtu
;
le32 speed
;
u8 duplex
;
u8 rss_max_key_size
;
le16 rss_max_indirection_table_length
;
le32 supported_hash_types
;
};](https://image.slidesharecdn.com/receivesidescalingrsswithebpfinqemuandvirtio-net-210930141731/85/Receive-side-scaling-RSS-with-eBPF-in-QEMU-and-virtio-net-12-320.jpg)
![virtio spec changes - setting RSS parameters
• VIRTIO_NET_CTRL_MQ_RSS_CONFIG
struct virtio_net_rss_con
fi
g
{
le32 hash_types
;
le16 indirection_table_mask
;
le16 unclassi
fi
ed_queue
;
le16 indirection_table[indirection_table_length]
;
le16 max_tx_vq
;
u8 hash_key_length
;
u8 hash_key_data[hash_key_length]
;
};](https://image.slidesharecdn.com/receivesidescalingrsswithebpfinqemuandvirtio-net-210930141731/85/Receive-side-scaling-RSS-with-eBPF-in-QEMU-and-virtio-net-13-320.jpg)
Receive side scaling (RSS) with eBPF in QEMU and virtio-net
- 1. Receive side scaling (RSS) with eBPF in QEMU and virtio-net Yan Vugenfirer - CEO, Daynix
- 2. Agenda • What is RSS? • History: RSS and virtio-net • What is eBPF? • Using eBPF for packet steering (RSS) in virtio-net
- 3. What is RSS? • Receive side scaling - distribution of packets’ processing among CPUs • A NIC uses a hashing function to compute a hash value over a defined area • Hash value is used to index an indirection table • The values in the indirection table are used to assign the received data to a CPU • With MSI support, a NIC can also interrupt the associated CPU
- 4. CPU1 CPU2 CPU3 CPU0 What is RSS? NIC RX handling RX handling RX handling RX handling NIC Driver HW queue HW queue HW queue HW queue Packet classification
- 6. History: RSS and virtio-net • Let’s use RSS with virtio-net! • Utilisation CPUs for packet processing • Cash locality for network applications • Microsoft WHQL requirement for high speed devices
- 7. History: RSS and virtio-net • No multi-queue in virtio • SW implementation in Windows guest driver (similar to RFS in Linux) CPU1 CPU2 CPU3 CPU0 virtio-net RX handling RX handling RX handling NetKVM RX virt-queue Packet classification RX handling
- 8. History: RSS and virtio-net • virtio-net became multi queue device • Due to Windows requirements - hybrid model. Interrupt received on specific CPU core, but could be rescheduled to another • Works good for TCP • Might not work for UDP • Support legacy interrupts for old OSes
- 9. History: RSS and virtio-net CPU1 CPU0 virtio-net NetKVM RX virt-queue Packet classification RX handling RX virt-queue Packet classification RX handling
- 10. History: RSS and virtio-net • virtio spec changes • Set steering mode • Pass the device redirection tables • Set hash value in virtio-net-hdr • No inter-processor interrupts due to re-scheduling • Vision: HW will do all the heavy work • Implementations • SW only POC in QEMU • eBPF
- 11. virtio spec changes - capabilities • VIRTIO_NET_F_RSS • VIRTIO_NET_F_MQ must be set
- 12. virtio spec changes - device configuration • virtio_net_config struct virtio_net_con fi g { u8 mac[6] ; le16 status ; le16 max_virtqueue_pairs ; le16 mtu ; le32 speed ; u8 duplex ; u8 rss_max_key_size ; le16 rss_max_indirection_table_length ; le32 supported_hash_types ; };
- 13. virtio spec changes - setting RSS parameters • VIRTIO_NET_CTRL_MQ_RSS_CONFIG struct virtio_net_rss_con fi g { le32 hash_types ; le16 indirection_table_mask ; le16 unclassi fi ed_queue ; le16 indirection_table[indirection_table_length] ; le16 max_tx_vq ; u8 hash_key_length ; u8 hash_key_data[hash_key_length] ; };
- 14. virtio spec changes - virtio-net-hdr struct virtio_net_hdr { u8 fl ags ; u8 gso_type ; le16 hdr_len ; le16 gso_size ; le16 csum_start ; le16 csum_offset ; le16 num_buffers ; le32 hash_value; (Only if VIRTIO_NET_F_HASH_REPORT negotiated ) le16 hash_report; (Only if VIRTIO_NET_F_HASH_REPORT negotiated ) le16 padding_reserved; (Only if VIRTIO_NET_F_HASH_REPORT negotiated ) };
- 15. What is eBPF? • Enable running sandboxed code in Linux kernel • The code can be loaded at run time • Used for security, tracing, networking, observability
- 16. How can eBPF help us? • Calculate the RSS hash and return the queue index for incoming packets • Populate the hash value in virtio_net_hdr (work in progress)
- 17. The “magic” • Loading eBPF program using IOCTL TUNSETSTEERINGEBPF • tun_struct has steering_prog field • If eBPF program for steering is loaded, tun_select_queue will call it with bpf_prog_run_clear_cb
- 18. Hash population (work in progress) • Population from eBPF program • virtio_net_hdr with additional fields • Work in progress in kernel • Enlarge virtio_net_hdr in all kernel modules • Keep calculated hash in SKB and copy it to virtio_net_hdr
- 19. eBPF program source in QEMU • tun_rss_steering_prog • tools/ebpf/rss.bpf.c • Use clang to compile • tools/ebpf/Makefile.ebpf
- 20. eBPF program skeleton • During QEMU compilation include file is populated with the compiled binary • bpftool gen skeleton rss.bpf.o > rss.bpf.skeleton.h • Helpers to initialise maps (mechanism to share data between eBPF program and kerneluserspace) • Some changes to support libvirt - mmapping the shared data structure to user space (3 maps in current main branch without mmaping, 1 map in pending patches)
- 21. Configuration map • The configuration map is BPF array map that contains everything required for RSS: • Supported hash flows: IPv4, TCPv4, UDPv4, IPv6, IPv6ex, TCPv6, UDPv6 • Indirections table size (max 128) • Default queue • Toeplitz hash key - 40 bytes • Indirections table - 128 entries
- 22. Loading eBPF program • Two mechanisms • QEMU using function in skeleton file. Calling bpf syscall • eBPF helper program (with libvirt) - QEMU gets file descriptors from libvirt with already loaded ebpf program and mapping of the ebpf map (patches under review)
- 23. Loading eBPF program • Possible load failures • Kernel support. Current solution requires 5.8+ • Without helper • QEMU process capabilities: CAP_BPF, CAP_NET_ADMIN • sysctl kernel.unprivileged_bpf_disabled=1 • libbpf not present • In case of helper usage - mismatch between helper and QEMU • Stamp is a hash of skeleton include file
- 24. Fallback • Built it QEMU RSS steering • Can be triggered also by live migration • Hash population is enabled in QEMU command line, because there is still not hash population from eBPF program
- 25. Live migration • Known issue: migrating to old kernel • eBPF load failure • Fallback to in-QEMU RSS steering
- 26. QEMU command line • Multi-queue should be enabled • -smp with vCPU for each queue-pair • -device virtio-net-pci, rss=on,hash=on,ebpf_rss_fds=<fd0,fd1>
- 27. QEMU command line • rss=on • Try to load eBPF from skeleton or by using provided file descriptors • Fallback to “built-in” RSS steering in QEMU if cannot load eBPF program • hash=on • Populate hash in virtio_net_hdr • ebpf_rss_fds - optional, provide file descriptors for eBPF program and map
- 28. libvirt integration • QEMU should run with least possible privileges • eBPF helper • Stamping the helper during compilation time • Redirection table mapping • Additional command line options to provide file descriptors to QEMU • Patches under review
- 29. Current status • Initial support was merged to QEMU • libvirt integration patches in QEMU and libvirt are under discussion on mailing lists • Hash population by eBPF program - pending additional work for next set of patches
- 30. Pending patches • QEMU libvirt integration: https://lists.nongnu.org/ archive/html/qemu-devel/2021-07/msg03535.html • libvirt patches: https://listman.redhat.com/archives/ libvir-list/2021-July/msg00836.html • RSS support in Linux virtio-net driver: https:// lists.linuxfoundation.org/pipermail/virtualization/ 2021-August/055940.html • In kernel hash calculation reporting to guest driver: https://lkml.org/lkml/2021/1/12/1329
- 31. virtio-net and eBPF future • Packet filtering with vhost • Security?
- 34. Links • https://www.kernel.org/doc/Documentation/networking/scaling.txt • https://docs.microsoft.com/en-us/windows-hardware/drivers/network/ introduction-to-receive-side-scaling • https://ebpf.io • https://docs.microsoft.com/en-us/windows-hardware/drivers/network/rss- with-a-single-hardware-receive-queue • https://docs.microsoft.com/en-us/windows-hardware/drivers/network/rss- with-hardware-queuing • https://docs.microsoft.com/en-us/windows-hardware/drivers/network/rss- with-message-signaled-interrupts • https://docs.microsoft.com/en-us/windows-hardware/drivers/network/rss- hashing-functions