On Thu, Apr 15, 2021 at 8:57 AM Toke Høiland-Jørgensen toke@redhat.com wrote:
Andrii Nakryiko andrii.nakryiko@gmail.com writes:
On Wed, Apr 14, 2021 at 3:51 PM Toke Høiland-Jørgensen toke@redhat.com wrote:
Andrii Nakryiko andrii.nakryiko@gmail.com writes:
On Wed, Apr 14, 2021 at 3:58 AM Toke Høiland-Jørgensen toke@redhat.com wrote:
Andrii Nakryiko andrii.nakryiko@gmail.com writes:
On Tue, Apr 6, 2021 at 3:06 AM Toke Høiland-Jørgensen toke@redhat.com wrote: > > Andrii Nakryiko andrii.nakryiko@gmail.com writes: > > > On Sat, Apr 3, 2021 at 10:47 AM Alexei Starovoitov > > alexei.starovoitov@gmail.com wrote: > >> > >> On Sat, Apr 03, 2021 at 12:38:06AM +0530, Kumar Kartikeya Dwivedi wrote: > >> > On Sat, Apr 03, 2021 at 12:02:14AM IST, Alexei Starovoitov wrote: > >> > > On Fri, Apr 2, 2021 at 8:27 AM Kumar Kartikeya Dwivedi memxor@gmail.com wrote: > >> > > > [...] > >> > > > >> > > All of these things are messy because of tc legacy. bpf tried to follow tc style > >> > > with cls and act distinction and it didn't quite work. cls with > >> > > direct-action is the only > >> > > thing that became mainstream while tc style attach wasn't really addressed. > >> > > There were several incidents where tc had tens of thousands of progs attached > >> > > because of this attach/query/index weirdness described above. > >> > > I think the only way to address this properly is to introduce bpf_link style of > >> > > attaching to tc. Such bpf_link would support ingress/egress only. > >> > > direction-action will be implied. There won't be any index and query > >> > > will be obvious. > >> > > >> > Note that we already have bpf_link support working (without support for pinning > >> > ofcourse) in a limited way. The ifindex, protocol, parent_id, priority, handle, > >> > chain_index tuple uniquely identifies a filter, so we stash this in the bpf_link > >> > and are able to operate on the exact filter during release. > >> > >> Except they're not unique. The library can stash them, but something else > >> doing detach via iproute2 or their own netlink calls will detach the prog. > >> This other app can attach to the same spot a different prog and now > >> bpf_link__destroy will be detaching somebody else prog. > >> > >> > > So I would like to propose to take this patch set a step further from > >> > > what Daniel said: > >> > > int bpf_tc_attach(prog_fd, ifindex, {INGRESS,EGRESS}): > >> > > and make this proposed api to return FD. > >> > > To detach from tc ingress/egress just close(fd). > >> > > >> > You mean adding an fd-based TC API to the kernel? > >> > >> yes. > > > > I'm totally for bpf_link-based TC attachment. > > > > But I think *also* having "legacy" netlink-based APIs will allow > > applications to handle older kernels in a much nicer way without extra > > dependency on iproute2. We have a similar situation with kprobe, where > > currently libbpf only supports "modern" fd-based attachment, but users > > periodically ask questions and struggle to figure out issues on older > > kernels that don't support new APIs. > > +1; I am OK with adding a new bpf_link-based way to attach TC programs, > but we still need to support the netlink API in libbpf. > > > So I think we'd have to support legacy TC APIs, but I agree with > > Alexei and Daniel that we should keep it to the simplest and most > > straightforward API of supporting direction-action attachments and > > setting up qdisc transparently (if I'm getting all the terminology > > right, after reading Quentin's blog post). That coincidentally should > > probably match how bpf_link-based TC API will look like, so all that > > can be abstracted behind a single bpf_link__attach_tc() API as well, > > right? That's the plan for dealing with kprobe right now, btw. Libbpf > > will detect the best available API and transparently fall back (maybe > > with some warning for awareness, due to inherent downsides of legacy > > APIs: no auto-cleanup being the most prominent one). > > Yup, SGTM: Expose both in the low-level API (in bpf.c), and make the > high-level API auto-detect. That way users can also still use the > netlink attach function if they don't want the fd-based auto-close > behaviour of bpf_link.
So I thought a bit more about this, and it feels like the right move would be to expose only higher-level TC BPF API behind bpf_link. It will keep the API complexity and amount of APIs that libbpf will have to support to the minimum, and will keep the API itself simple: direct-attach with the minimum amount of input arguments. By not exposing low-level APIs we also table the whole bpf_tc_cls_attach_id design discussion, as we now can keep as much info as needed inside bpf_link_tc (which will embed bpf_link internally as well) to support detachment and possibly some additional querying, if needed.
But then there would be no way for the caller to explicitly select a mechanism? I.e., if I write a BPF program using this mechanism targeting a 5.12 kernel, I'll get netlink attachment, which can stick around when I do bpf_link__disconnect(). But then if the kernel gets upgraded to support bpf_link for TC programs I'll suddenly transparently get bpf_link and the attachments will go away unless I pin them. This seems... less than ideal?
That's what we are doing with bpf_program__attach_kprobe(), though. And so far I've only seen people (privately) saying how good it would be to have bpf_link-based TC APIs, doesn't seem like anyone with a realistic use case prefers the current APIs. So I suspect it's not going to be a problem in practice. But at least I'd start there and see how people are using it and if they need anything else.
*sigh* - I really wish you would stop arbitrarily declaring your own use cases "realistic" and mine (implied) "unrealistic". Makes it really hard to have a productive discussion...
Well (sigh?..), this wasn't my intention, sorry you read it this way. But we had similar discussions when I was adding bpf_link-based XDP attach APIs.
Great, thank you! And yeah, we did discuss exactly this before, which is where my mental sigh came from - I feel like we already covered this ground and that I'm just being dismissed with "that is not a real use case". But OK, I'll give it another shot, see below.
And guess what, now I see that samples/bpf/whatever_xdp is switched to bpf_link-based XDP, because that makes everything simpler and more reliable. What I also know is that in production we ran into multiple issues with anything that doesn't auto-detach on process exit/crash (unless pinned explicitly, of course). And that people that are trying to use TC right now are saying how having bpf_link-based TC APIs would make everything *simpler* and *safer*. So I don't know... I understand it might be convenient in some cases to not care about a lifetime of BPF programs you are attaching, but then there are usually explicit and intentional ways to achieve at least similar behavior with safety by default.
So I guess call me unconvinced (yet? still?). Give it another shot, though.
I'm not arguing against adding bpf_link support, and I'm even fine with making it the default. As you say, there are plenty of use cases where the bpf_link semantics make sense, and the XDP programs in samples all fall in this category. So sure, let's add this support and make this convenient to use.
But there are also use cases where the BPF program lifetime absolutely shouldn't follow that of the userspace application. This includes both applications that don't have a long-running daemon at all (like a firewall that just loads a ruleset at boot; xdp-filter is such an application in the BPF world, but I'm sure there are others). And daemons that use BPF as a data path and want the packets to keep flowing even when they restart, like Cilium as Daniel mentioned.
So the latter category of applications need their BPF programs to be permanently attached to the interface. And sure, this can sorta be done by pinning the bpf_link; but not really, because then:
- You incur a new dependency on bpffs, so you have to make sure that is mounted and that you can get at the particular fs instance you're using; the latter is especially painful if you switch namespaces.
So I understand that it's more painful than current TC and legacy XDP APIs for these specific use cases. But for other types of BPF programs that want to persist across user-space process exit, all those needs to be addressed and designed around anyway. And that doesn't prevent others from doing it, otherwise we wouldn't even implement BPFFS.
- Your BPF program lifetime is no longer tied to the interface, so you have to deal with garbage collecting your pinned files somehow. This is especially painful if you don't have a daemon.
With bpf_link auto-detach, you'll still get the underlying BPF program freed, only a small bpf_link shell will persist. If you churn through interface going up/down, whichever application is responsible for re-attaching BPF programs would deal with clean up.
But yes, I'm not oblivious to the need to change how you design your applications and new inconveniences that would cause. But think about this from a slightly different point of view. If we had to choose between bpf_link model and current TC API, which one would we choose? I'd argue we should choose bpf_link, because: 1) it allows to do things more safely (auto-cleanup) by default; 2) it still allows for having BPF program/link persistence, even if in a different and slightly more inconvenient way.
So it's a more generic and powerful approach, it's just not as perfectly aligned with the way it's used for cases you mentioned.
Together, these issues make bpf_link a much less compelling proposition, to the point where it's no longer the better API for these use cases, IMO. And I know that because I had to work around just these issues with bpf_link for xdp-tools.
But I'm not even asking for the netlink API to be the default, I'm fine with bpf_link being the default and encouraged API. I'm just asking for a way to make it *possible* to select which attach mode I want. Either by a flag to bpf_program__attach_tc(), or by exposing the low-level bpf_tc_cls_*() netlink functions, like we do for XDP.
Given significant semantic differences between bpf_link and current TC APIs, I'm not sure anymore if it's a good idea to hide current API behind bpf_link abstraction. For tracing it's a good fit, so I still think falling back to legacy kprobe API makes sense there. For TC, given your and Daniel's replies, I'd rather have low-level APIs exposed directly instead of having some options switch in bpf_program__attach_tc() API.
If we expose the low-level API I can elect to just use this if I know I want netlink behaviour, but if bpf_program__attach_tc() is the only API available it would at least need a flag to enforce one mode or the other (I can see someone wanting to enforce kernel bpf_link semantics as well, so a flag for either mode seems reasonable?).
Sophisticated enough users can also do feature detection to know if it's going to work or not.
Sure, but that won't help if there's no API to pick the attach mode they want.
I'm not intending to allow legacy kprobe APIs to be "chosen", for instance. Because I'm convinced it's a bad API that no one should use if they can use an FD-based one.
I'd tend to agree with you for the tracing APIs, actually. But a BPF-based data plane is different, as I tried to explain above.
Right, see above.
It might be a different case for TC, who knows. I'd just start with safer APIs and then evaluate whether there is a real demand for less safe ones. It's just some minor refactoring and exposing more APIs, when/if we need them.
There you go again with the "real demand" argument. How can I read this in any other way than that you don't consider my use case "real" (as you just assured me above was not the case)? What do you consider "real demand"?
Please try not to read more in my words than there really is. By real demand I meant at least few *different* use cases coming from different parties. That says nothing about whether your use case is real or not, just that it's a one case (or at least coming from just one party).
-Toke