On 2021-01-20 18:29, Toke Høiland-Jørgensen wrote:
Björn Töpel <bjorn.topel@xxxxxxxxx> writes:
On 2021-01-20 15:54, Toke Høiland-Jørgensen wrote:
Björn Töpel <bjorn.topel@xxxxxxxxx> writes:
On 2021-01-20 13:50, Toke Høiland-Jørgensen wrote:
Björn Töpel <bjorn.topel@xxxxxxxxx> writes:
diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
index c001766adcbc..bbc7d9a57262 100644
--- a/include/uapi/linux/bpf.h
+++ b/include/uapi/linux/bpf.h
@@ -3836,6 +3836,12 @@ union bpf_attr {
* Return
* A pointer to a struct socket on success or NULL if the file is
* not a socket.
+ *
+ * long bpf_redirect_xsk(struct xdp_buff *xdp_md, u64 action)
+ * Description
+ * Redirect to the registered AF_XDP socket.
+ * Return
+ * **XDP_REDIRECT** on success, otherwise the action parameter is returned.
*/
I think it would be better to make the second argument a 'flags'
argument and make values > XDP_TX invalid (like we do in
bpf_xdp_redirect_map() now). By allowing any value as return you lose
the ability to turn it into a flags argument later...
Yes, but that adds a run-time check. I prefer this non-checked version,
even though it is a bit less futureproof.
That...seems a bit short-sighted? :)
Can you actually see a difference in your performance numbers?
I would rather add an additional helper *if* we see the need for flags,
instead of paying for that upfront. For me, BPF is about being able to
specialize, and not having one call with tons of checks.
I get that, I'm just pushing back because omitting a 'flags' argument is
literally among the most frequent reasons for having to replace a
syscall (see e.g., [0]) instead of extending it. And yeah, I do realise
that the performance implications are different for XDP than for
syscalls, but maintainability of the API is also important; it's all a
tradeoff. This will be the third redirect helper variant for XDP and I'd
hate for the fourth one to have to be bpf_redirect_xsk_flags() because
it did turn out to be needed...
(One potential concrete reason for this: I believe Magnus was talking
about an API that would allow a BPF program to redirect a packet into
more than one socket (cloning it in the process), or to redirect to a
socket+another target. How would you do that with this new helper?)
[0] https://lwn.net/Articles/585415/
I have a bit of different view. One of the really nice parts about BPF
is exactly specialization. A user can tailor the kernel do a specific
thing. I *don't* see an issue with yet another helper, if that is needed
in the future. I think that is better than bloated helpers trying to
cope for all scenarios. I don't mean we should just add helpers all over
the place, but I do see more lightly on adding helpers, than adding
syscalls.
Elaborating a bit on this: many device drivers try to handle all the
things in the fast-path. I see BPF as one way forward to moving away
from that. Setup what you need, and only run what you currently need,
instead of the current "Is bleh on, then baz? Is this on, then that."
So, I would like to avoid "future proofing" the helpers, if that makes
sense. Use what you need. That's why BPF is so good (one of the things)!
As for bpf_redirect_xsk() it's a leaner version of bpf_redirect_map().
You want flags/shared sockets/...? Well go use bpf_redirect_map() and
XSKMAP. bpf_redirect_xsk() is not for you.
A lot of back-and-forth for *one* if-statement, but it's kind of a
design thing for me. ;-)
Björn
(Related; Going forward, the growing switch() for redirect targets in
xdp_do_redirect() is a concern for me...)
And yes, even with all those fancy branch predictors, less instructions
is still less. :-) (It shows in my ubenchs.)
Right, I do agree that the run-time performance hit of checking the flag
sucks (along with being hard to check for, cf. our parallel discussion
about version checks). So ideally this would be fixed by having the
verifier enforce the argument ranges instead; but if we merge this
without the runtime check now we can't add that later without
potentially breaking programs... :(
-Toke
