On 2024/10/02 1:31, Stephen Hemminger wrote:
On Tue, 1 Oct 2024 14:54:29 +0900
Akihiko Odaki <akihiko.odaki@xxxxxxxxxx> wrote:
On 2024/09/30 0:33, Stephen Hemminger wrote:
On Sun, 29 Sep 2024 16:10:47 +0900
Akihiko Odaki <akihiko.odaki@xxxxxxxxxx> wrote:
On 2024/09/29 11:07, Jason Wang wrote:
On Fri, Sep 27, 2024 at 3:51 PM Akihiko Odaki <akihiko.odaki@xxxxxxxxxx> wrote:
On 2024/09/27 13:31, Jason Wang wrote:
On Fri, Sep 27, 2024 at 10:11 AM Akihiko Odaki <akihiko.odaki@xxxxxxxxxx> wrote:
On 2024/09/25 12:30, Jason Wang wrote:
On Tue, Sep 24, 2024 at 5:01 PM Akihiko Odaki <akihiko.odaki@xxxxxxxxxx> wrote:
virtio-net have two usage of hashes: one is RSS and another is hash
reporting. Conventionally the hash calculation was done by the VMM.
However, computing the hash after the queue was chosen defeats the
purpose of RSS.
Another approach is to use eBPF steering program. This approach has
another downside: it cannot report the calculated hash due to the
restrictive nature of eBPF.
Introduce the code to compute hashes to the kernel in order to overcome
thse challenges.
An alternative solution is to extend the eBPF steering program so that it
will be able to report to the userspace, but it is based on context
rewrites, which is in feature freeze. We can adopt kfuncs, but they will
not be UAPIs. We opt to ioctl to align with other relevant UAPIs (KVM
and vhost_net).
I wonder if we could clone the skb and reuse some to store the hash,
then the steering eBPF program can access these fields without
introducing full RSS in the kernel?
I don't get how cloning the skb can solve the issue.
We can certainly implement Toeplitz function in the kernel or even with
tc-bpf to store a hash value that can be used for eBPF steering program
and virtio hash reporting. However we don't have a means of storing a
hash type, which is specific to virtio hash reporting and lacks a
corresponding skb field.
I may miss something but looking at sk_filter_is_valid_access(). It
looks to me we can make use of skb->cb[0..4]?
I didn't opt to using cb. Below is the rationale:
cb is for tail call so it means we reuse the field for a different
purpose. The context rewrite allows adding a field without increasing
the size of the underlying storage (the real sk_buff) so we should add a
new field instead of reusing an existing field to avoid confusion.
We are however no longer allowed to add a new field. In my
understanding, this is because it is an UAPI, and eBPF maintainers found
it is difficult to maintain its stability.
Reusing cb for hash reporting is a workaround to avoid having a new
field, but it does not solve the underlying problem (i.e., keeping eBPF
as stable as UAPI is unreasonably hard). In my opinion, adding an ioctl
is a reasonable option to keep the API as stable as other virtualization
UAPIs while respecting the underlying intention of the context rewrite
feature freeze.
Fair enough.
Btw, I remember DPDK implements tuntap RSS via eBPF as well (probably
via cls or other). It might worth to see if anything we miss here.
Thanks for the information. I wonder why they used cls instead of
steering program. Perhaps it may be due to compatibility with macvtap
and ipvtap, which don't steering program.
Their RSS implementation looks cleaner so I will improve my RSS
implementation accordingly.
DPDK needs to support flow rules. The specific case is where packets
are classified by a flow, then RSS is done across a subset of the queues.
The support for flow in TUN driver is more academic than useful,
I fixed it for current BPF, but doubt anyone is using it really.
A full steering program would be good, but would require much more
complexity to take a general set of flow rules then communicate that
to the steering program.
It reminded me of RSS context and flow filter. Some physical NICs
support to use a dedicated RSS context for packets matched with flow
filter, and virtio is also gaining corresponding features.
RSS context: https://github.com/oasis-tcs/virtio-spec/issues/178
Flow filter: https://github.com/oasis-tcs/virtio-spec/issues/179
I considered about the possibility of supporting these features with tc
instead of adding ioctls to tuntap, but it seems not appropriate for
virtualization use case.
In a virtualization use case, tuntap is configured according to requests
of guests, and the code processing these requests need to have minimal
permissions for security. This goal is achieved by passing a file
descriptor that represents a tuntap from a privileged process (e.g.,
libvirt) to the process handling guest requests (e.g., QEMU).
However, tc is configured with rtnetlink, which does not seem to have an
interface to delegate a permission for one particular device to another
process.
For now I'll continue working on the current approach that is based on
ioctl and lacks RSS context and flow filter features. Eventually they
are also likely to require new ioctls if they are to be supported with
vhost_net.
The DPDK flow handling (rte_flow) was started by Mellanox and many of
the features are to support what that NIC can do. Would be good to have
a tc way to configure that (or devlink).
Yes, but I would rather only implement the ioctl without flow handling
for now. My purpose of implementing RSS in the kernel is to report hash
values to the guest that has its own network stack in the virtualization
context. tc-bpf would be suffice for DPDK, which does not have such a
requirement.
Having an access to the in-kernel RSS implementation also saves the
trouble of implementing an eBPF program for RSS, but DPDK already does
have such a program so it makes little difference. There may be also
performance improvement because I'm optimizing the in-kernel RSS
implementation with ffs(), which is currently not available to eBPF, but
there is also a proposal to expose ffs() to eBPF*.
For now, I will keep the patch series small by having only the ioctl
interface. Anyone who finds the feature useful for tc can take it and
add a tc interface in the future.
Regards,
Akihiko Odaki
* https://lore.kernel.org/bpf/20240131155607.51157-1-hffilwlqm@xxxxxxxxx/#t
