On Wed, Sep 23, 2020 at 9:11 AM Lorenz Bauer <lmb@xxxxxxxxxxxxxx> wrote: > > On Fri, 18 Sep 2020 at 04:26, Alexei Starovoitov > <alexei.starovoitov@xxxxxxxxx> wrote: > [...] > > > > Lorenz, > > if you can test it on cloudflare progs would be awesome. > > Our programs all bpf_tail_call from the topmost function, so no calls > from subprogs. I stripped out our FORCE_INLINE flag, recompiled and > ran our testsuite. cls_redirect.c (also in the kernel selftests) has a > test failure that I currently can't explain, but I don't have the time > to look at it in detail right now. selftests's test_cls_redirect.c does not have any tail calls, so it's not an interesting target. But your internal code relying on tail_call and until today you could not use subprograms at all. It doesn't matter that you do tail_call out of topmost function. The verifier would disallow subprogs, so you were missing on lots of new functionality. What I'm suggesting to try is to keep your code as-is (with tail_call in topmost) and simply remove always_inline. Let the compiler decide on better code layout. That will improve performance and most likely improve verification time too. If you convert some of the subprograms into global functions then you will be able to use function-by-function verification which will drastically improve verification time. Same thing for cilium. Their progs use tail_calls and because of that couldn't use subprogs at all. I think John was saying that prog load time is important for cilium. The answer is to use global functions. Now with tail_calls being compatible with subprogs all that is available.
