On 2021/02/05 2:36, William Tu wrote:
Hi Toshiaki,
Thanks for the patch. I've been testing it for a couple days.
I liked it a lot! The compile and build process all work without any issues.
Hi, thank you for reviewing!
Sorry for taking time to reply. It took time to remember every detail of the
patch set...
On Thu, Jul 30, 2020 at 7:55 PM Toshiaki Makita
<toshiaki.maki...@gmail.com> wrote:
This patch adds an XDP-based flow cache using the OVS netdev-offload
flow API provider. When an OVS device with XDP offload enabled,
packets first are processed in the XDP flow cache (with parse, and
table lookup implemented in eBPF) and if hits, the action processing
are also done in the context of XDP, which has the minimum overhead.
This provider is based on top of William's recently posted patch for
custom XDP load. When a custom XDP is loaded, the provider detects if
the program supports classifier, and if supported it starts offloading
flows to the XDP program.
The patches are derived from xdp_flow[1], which is a mechanism similar to
this but implemented in kernel.
* Motivation
While userspace datapath using netdev-afxdp or netdev-dpdk shows good
performance, there are use cases where packets better to be processed in
kernel, for example, TCP/IP connections, or container to container
connections. Current solution is to use tap device or af_packet with
extra kernel-to/from-userspace overhead. But with XDP, a better solution
is to steer packets earlier in the XDP program, and decides to send to
userspace datapath or stay in kernel.
One problem with current netdev-afxdp is that it forwards all packets to
userspace, The first patch from William (netdev-afxdp: Enable loading XDP
program.) only provides the interface to load XDP program, howerver users
usually don't know how to write their own XDP program.
XDP also supports HW-offload so it may be possible to offload flows to
HW through this provider in the future, although not currently.
The reason is that map-in-map is required for our program to support
classifier with subtables in XDP, but map-in-map is not offloadable.
If map-in-map becomes offloadable, HW-offload of our program may also
be possible.
I think it's too far away for XDP to be offloaded into HW and meet OVS's
feature requirements.
I don't know blockers other than map-in-map, but probably there are more.
If you can provide explicit blockers I can add them in the cover letter.
There is a research prototype here, FYI.
https://www.usenix.org/conference/osdi20/presentation/brunella
This is a presentation about FPGA, not HW offload to SmartNIC, right?
* How to use
1. Install clang/llvm >= 9, libbpf >= 0.0.6 (included in kernel 5.5), and
kernel >= 5.3.
2. make with --enable-afxdp --enable-xdp-offload
--enable-bpf will generate XDP program "bpf/flowtable_afxdp.o". Note that
typo: I think you mean --enable-xdp-offload
Thanks.
the BPF object will not be installed anywhere by "make install" at this point.
3. Load custom XDP program
E.g.
$ ovs-vsctl add-port ovsbr0 veth0 -- set int veth0 options:xdp-mode=native \
options:xdp-obj="/path/to/ovs/bpf/flowtable_afxdp.o"
$ ovs-vsctl add-port ovsbr0 veth1 -- set int veth1 options:xdp-mode=native \
options:xdp-obj="/path/to/ovs/bpf/flowtable_afxdp.o"
4. Enable XDP_REDIRECT
If you use veth devices, make sure to load some (possibly dummy) programs
on the peers of veth devices. This patch set includes a program which
does nothing but returns XDP_PASS. You can use it for the veth peer like
this:
$ ip link set veth1 xdpdrv object /path/to/ovs/bpf/xdp_noop.o section xdp
I'd suggest not using "veth1" as an example, because in (3) above, people
might think "veth1" is already attached to ovsbr0.
IIUC, here your "veth1" should be the device at the peer inside
another namespace.
Sure, will rename it.
Some HW NIC drivers require as many queues as cores on its system. Tweak
queues using "ethtool -L".
5. Enable hw-offload
$ ovs-vsctl set Open_vSwitch . other_config:offload-driver=linux_xdp
$ ovs-vsctl set Open_vSwitch . other_config:hw-offload=true
This will starts offloading flows to the XDP program.
You should be able to see some maps installed, including "debug_stats".
$ bpftool map
If packets are successfully redirected by the XDP program,
debug_stats[2] will be counted.
$ bpftool map dump id <ID of debug_stats>
Currently only very limited keys and output actions are supported.
For example NORMAL action entry and IP based matching work with current
key support. VLAN actions used by port tag/trunks are also supported.
I don't know if this is too much to ask for.
I wonder if you, or we can work together, to add at least a tunnel
support, ex: vxlan?
The current version is a good prototype for people to test an L2/L3
XDP offload switch,
but without a good use case, it's hard to attract more people to
contribute or use it.
I think we have discussed this before.
Vxlan or other tunneling is indeed important, but that's not straightforward.
Push is easy, but pop is not. Pop requires two rules and recirculation.
Recirculation is highly likely to cause eBPF 1M insn limit error.
One possible solution is to combine two rules into one and insert it instead,
but I have not verified whether this can work or how to implement it.
Can we leave this and make following patches later?
From a maintenance point of view, can we add a test case to avoid regression?
For example, something like "make check-afxdp".
We can have "make check-xdp-offload". I can also help adding it.
OK, let me think about what kind of test we can do...
* Performance
Tested 2 cases. 1) i40e to veth, 2) i40e to i40e.
Test 1 Measured drop rate at veth interface with redirect action from
physical interface (i40e 25G NIC, XXV 710) to veth. The CPU is Xeon
Silver 4114 (2.20 GHz).
XDP_DROP
+------+ +-------+ +-------+
pktgen -- wire --> | eth0 | -- NORMAL ACTION --> | veth0 |----| veth2 |
+------+ +-------+ +-------+
Test 2 uses i40e instead of veth, and measured tx packet rate at output
device.
Thanks for the performance results. I tested using two 25Gb ConnectX-6
cards, using mlx5. I got pretty similar performance results as yours.
But in general, I'm more worried about whether a feature can be implemented
in XDP than its performance.
Single-flow performance test results:
1) i40e-veth
a) no-zerocopy in i40e
- xdp 3.7 Mpps
- afxdp 980 kpps
b) zerocopy in i40e (veth does not have zc)
- xdp 1.9 Mpps
- afxdp 980 Kpps
2) i40e-i40e
a) no-zerocopy
- xdp 3.5 Mpps
- afxdp 1.5 Mpps
b) zerocopy
- xdp 2.0 Mpps
- afxdp 4.4 Mpps
** xdp is better when zc is disabled. The reason of poor performance on zc
is that xdp_frame requires packet memory allocation and memcpy on
XDP_REDIRECT to other devices iff zc is enabled.
** afxdp with zc is better than xdp without zc, but afxdp is using 2 cores
in this case, one is pmd and the other is softirq. When pmd and softirq
were running on the same core, the performance was extremely poor as
pmd consumes cpus. I also tested afxdp-nonpmd to run softirq and
userspace processing on the same core, but the result was lower than
(pmd results) / 2.
With nonpmd, xdp performance was the same as xdp with pmd. This means
xdp only uses one core (for softirq only). Even with pmd, we need only
one pmd for xdp even when we want to use more cores for multi-flow.
Since XDP is evolving, it would be helpful to point out some future work and
current limitations. As far as I know
1) No broadcast/multicast support in XDP. Patch below:
https://lwn.net/Articles/817582/
2) No large packet support (ex: TSO)
3) No HW checksum offload
Will add them, thanks!
Toshiaki Makita
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
