On Tue, Nov 1, 2022 at 10:31 AM Martin KaFai Lau <martin.lau@xxxxxxxxx> wrote:
>
> On 10/31/22 3:09 PM, Stanislav Fomichev wrote:
> > On Mon, Oct 31, 2022 at 12:36 PM Yonghong Song <yhs@xxxxxxxx> wrote:
> >>
> >>
> >>
> >> On 10/31/22 8:28 AM, Toke Høiland-Jørgensen wrote:
> >>> "Bezdeka, Florian" <florian.bezdeka@xxxxxxxxxxx> writes:
> >>>
> >>>> Hi all,
> >>>>
> >>>> I was closely following this discussion for some time now. Seems we
> >>>> reached the point where it's getting interesting for me.
> >>>>
> >>>> On Fri, 2022-10-28 at 18:14 -0700, Jakub Kicinski wrote:
> >>>>> On Fri, 28 Oct 2022 16:16:17 -0700 John Fastabend wrote:
> >>>>>>>> And it's actually harder to abstract away inter HW generation
> >>>>>>>> differences if the user space code has to handle all of it.
> >>>>>>
> >>>>>> I don't see how its any harder in practice though?
> >>>>>
> >>>>> You need to find out what HW/FW/config you're running, right?
> >>>>> And all you have is a pointer to a blob of unknown type.
> >>>>>
> >>>>> Take timestamps for example, some NICs support adjusting the PHC
> >>>>> or doing SW corrections (with different versions of hw/fw/server
> >>>>> platforms being capable of both/one/neither).
> >>>>>
> >>>>> Sure you can extract all this info with tracing and careful
> >>>>> inspection via uAPI. But I don't think that's _easier_.
> >>>>> And the vendors can't run the results thru their validation
> >>>>> (for whatever that's worth).
> >>>>>
> >>>>>>> I've had the same concern:
> >>>>>>>
> >>>>>>> Until we have some userspace library that abstracts all these details,
> >>>>>>> it's not really convenient to use. IIUC, with a kptr, I'd get a blob
> >>>>>>> of data and I need to go through the code and see what particular type
> >>>>>>> it represents for my particular device and how the data I need is
> >>>>>>> represented there. There are also these "if this is device v1 -> use
> >>>>>>> v1 descriptor format; if it's a v2->use this another struct; etc"
> >>>>>>> complexities that we'll be pushing onto the users. With kfuncs, we put
> >>>>>>> this burden on the driver developers, but I agree that the drawback
> >>>>>>> here is that we actually have to wait for the implementations to catch
> >>>>>>> up.
> >>>>>>
> >>>>>> I agree with everything there, you will get a blob of data and then
> >>>>>> will need to know what field you want to read using BTF. But, we
> >>>>>> already do this for BPF programs all over the place so its not a big
> >>>>>> lift for us. All other BPF tracing/observability requires the same
> >>>>>> logic. I think users of BPF in general perhaps XDP/tc are the only
> >>>>>> place left to write BPF programs without thinking about BTF and
> >>>>>> kernel data structures.
> >>>>>>
> >>>>>> But, with proposed kptr the complexity lives in userspace and can be
> >>>>>> fixed, added, updated without having to bother with kernel updates, etc.
> >>>>>> From my point of view of supporting Cilium its a win and much preferred
> >>>>>> to having to deal with driver owners on all cloud vendors, distributions,
> >>>>>> and so on.
> >>>>>>
> >>>>>> If vendor updates firmware with new fields I get those immediately.
> >>>>>
> >>>>> Conversely it's a valid concern that those who *do* actually update
> >>>>> their kernel regularly will have more things to worry about.
> >>>>>
> >>>>>>> Jakub mentions FW and I haven't even thought about that; so yeah, bpf
> >>>>>>> programs might have to take a lot of other state into consideration
> >>>>>>> when parsing the descriptors; all those details do seem like they
> >>>>>>> belong to the driver code.
> >>>>>>
> >>>>>> I would prefer to avoid being stuck on requiring driver writers to
> >>>>>> be involved. With just a kptr I can support the device and any
> >>>>>> firwmare versions without requiring help.
> >>>>>
> >>>>> 1) where are you getting all those HW / FW specs :S
> >>>>> 2) maybe *you* can but you're not exactly not an ex-driver developer :S
> >>>>>
> >>>>>>> Feel free to send it early with just a handful of drivers implemented;
> >>>>>>> I'm more interested about bpf/af_xdp/user api story; if we have some
> >>>>>>> nice sample/test case that shows how the metadata can be used, that
> >>>>>>> might push us closer to the agreement on the best way to proceed.
> >>>>>>
> >>>>>> I'll try to do a intel and mlx implementation to get a cross section.
> >>>>>> I have a good collection of nics here so should be able to show a
> >>>>>> couple firmware versions. It could be fine I think to have the raw
> >>>>>> kptr access and then also kfuncs for some things perhaps.
> >>>>>>
> >>>>>>>> I'd prefer if we left the door open for new vendors. Punting descriptor
> >>>>>>>> parsing to user space will indeed result in what you just said - major
> >>>>>>>> vendors are supported and that's it.
> >>>>>>
> >>>>>> I'm not sure about why it would make it harder for new vendors? I think
> >>>>>> the opposite,
> >>>>>
> >>>>> TBH I'm only replying to the email because of the above part :)
> >>>>> I thought this would be self evident, but I guess our perspectives
> >>>>> are different.
> >>>>>
> >>>>> Perhaps you look at it from the perspective of SW running on someone
> >>>>> else's cloud, an being able to move to another cloud, without having
> >>>>> to worry if feature X is available in xdp or just skb.
> >>>>>
> >>>>> I look at it from the perspective of maintaining a cloud, with people
> >>>>> writing random XDP applications. If I swap a NIC from an incumbent to a
> >>>>> (superior) startup, and cloud users are messing with raw descriptor -
> >>>>> I'd need to go find every XDP program out there and make sure it
> >>>>> understands the new descriptors.
> >>>>
> >>>> Here is another perspective:
> >>>>
> >>>> As AF_XDP application developer I don't wan't to deal with the
> >>>> underlying hardware in detail. I like to request a feature from the OS
> >>>> (in this case rx/tx timestamping). If the feature is available I will
> >>>> simply use it, if not I might have to work around it - maybe by falling
> >>>> back to SW timestamping.
> >>>>
> >>>> All parts of my application (BPF program included) should not be
> >>>> optimized/adjusted for all the different HW variants out there.
> >>>
> >>> Yes, absolutely agreed. Abstracting away those kinds of hardware
> >>> differences is the whole *point* of having an OS/driver model. I.e.,
> >>> it's what the kernel is there for! If people want to bypass that and get
> >>> direct access to the hardware, they can already do that by using DPDK.
> >>>
> >>> So in other words, 100% agreed that we should not expect the BPF
> >>> developers to deal with hardware details as would be required with a
> >>> kptr-based interface.
> >>>
> >>> As for the kfunc-based interface, I think it shows some promise.
> >>> Exposing a list of function names to retrieve individual metadata items
> >>> instead of a struct layout is sorta comparable in terms of developer UI
> >>> accessibility etc (IMO).
> >>
> >> Looks like there are quite some use cases for hw_timestamp.
> >> Do you think we could add it to the uapi like struct xdp_md?
> >>
> >> The following is the current xdp_md:
> >> struct xdp_md {
> >> __u32 data;
> >> __u32 data_end;
> >> __u32 data_meta;
> >> /* Below access go through struct xdp_rxq_info */
> >> __u32 ingress_ifindex; /* rxq->dev->ifindex */
> >> __u32 rx_queue_index; /* rxq->queue_index */
> >>
> >> __u32 egress_ifindex; /* txq->dev->ifindex */
> >> };
> >>
> >> We could add __u64 hw_timestamp to the xdp_md so user
> >> can just do xdp_md->hw_timestamp to get the value.
> >> xdp_md->hw_timestamp == 0 means hw_timestamp is not
> >> available.
> >>
> >> Inside the kernel, the ctx rewriter can generate code
> >> to call driver specific function to retrieve the data.
> >
> > If the driver generates the code to retrieve the data, how's that
> > different from the kfunc approach?
> > The only difference I see is that it would be a more strong UAPI than
> > the kfuncs?
>
> Another thing may be worth considering, some hints for some HW/driver may be
> harder (or may not worth) to unroll/inline. For example, I see driver is doing
> spin_lock_bh while getting the hwtstamp. For this case, keep calling a kfunc
> and avoid the unroll/inline may be the right thing to do.
Yeah, I'm trying to look at the drivers right now and doing
spinlocks/seqlocks might complicate the story...
But it seems like in the worst case, the unrolled bytecode can always
resort to calling a kernel function?
(we might need to have some scratch area to preserve r1-r5 and we
can't touch r6-r9 because we are not in a real call, but seems doable;
I'll try to illustrate with a bunch of examples)
> >> The kfunc approach can be used to *less* common use cases?
> >
> > What's the advantage of having two approaches when one can cover
> > common and uncommon cases?
> >
> >>> There are three main drawbacks, AFAICT:
> >>>
> >>> 1. It requires driver developers to write and maintain the code that
> >>> generates the unrolled BPF bytecode to access the metadata fields, which
> >>> is a non-trivial amount of complexity. Maybe this can be abstracted away
> >>> with some internal helpers though (like, e.g., a
> >>> bpf_xdp_metadata_copy_u64(dst, src, offset) helper which would spit out
> >>> the required JMP/MOV/LDX instructions?
> >>>
> >>> 2. AF_XDP programs won't be able to access the metadata without using a
> >>> custom XDP program that calls the kfuncs and puts the data into the
> >>> metadata area. We could solve this with some code in libxdp, though; if
> >>> this code can be made generic enough (so it just dumps the available
> >>> metadata functions from the running kernel at load time), it may be
> >>> possible to make it generic enough that it will be forward-compatible
> >>> with new versions of the kernel that add new fields, which should
> >>> alleviate Florian's concern about keeping things in sync.
> >>>
> >>> 3. It will make it harder to consume the metadata when building SKBs. I
> >>> think the CPUMAP and veth use cases are also quite important, and that
> >>> we want metadata to be available for building SKBs in this path. Maybe
> >>> this can be resolved by having a convenient kfunc for this that can be
> >>> used for programs doing such redirects. E.g., you could just call
> >>> xdp_copy_metadata_for_skb() before doing the bpf_redirect, and that
> >>> would recursively expand into all the kfunc calls needed to extract the
> >>> metadata supported by the SKB path?
> >>>
> >>> -Toke
> >>>
>
