On Thu, Oct 21, 2021 at 1:47 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Wed, Oct 20, 2021 at 05:05:02PM -0700, Alexei Starovoitov wrote:
> > On Wed, Oct 20, 2021 at 01:09:51PM +0200, Peter Zijlstra wrote:
>
> > > @@ -446,25 +440,8 @@ static void emit_bpf_tail_call_indirect(
> > > {
> > > int tcc_off = -4 - round_up(stack_depth, 8);
> > > u8 *prog = *pprog, *start = *pprog;
> > > - int pop_bytes = 0;
> > > - int off1 = 42;
> > > - int off2 = 31;
> > > - int off3 = 9;
> > > -
> > > - /* count the additional bytes used for popping callee regs from stack
> > > - * that need to be taken into account for each of the offsets that
> > > - * are used for bailing out of the tail call
> > > - */
> > > - pop_bytes = get_pop_bytes(callee_regs_used);
> > > - off1 += pop_bytes;
> > > - off2 += pop_bytes;
> > > - off3 += pop_bytes;
> > > -
> > > - if (stack_depth) {
> > > - off1 += 7;
> > > - off2 += 7;
> > > - off3 += 7;
> > > - }
> > > + static int out_label = -1;
> >
> > Interesting idea!
>
> I nicked it from emit_bpf_tail_call() in the 32bit jit :-) It seemed a
> lot more robust than the 64bit one and I couldn't figure out why the
> difference.
Interesting. Daniel will recognize that trick then :)
> > All insn emits trying to do the right thing from the start.
> > Here the logic assumes that there will be at least two passes over image.
> > I think that is correct, but we never had such assumption.
>
> That's not exactly true; I think image is NULL on every first run, so
> all insn that depend on it will be wrong to start with. Equally there's
> a number of insn that seem to depend on addrs[i], that also requires at
> least two passes.
Right. The image will be allocated after size converges and
addrs[] is inited with 64.
So there is certainly more than one pass.
I was saying that emit* helpers didn't have that assumption.
Looks like 32-bit JIT had.
> > A comment is certainly must have.
>
> I can certainly add one, although I think we'll disagree on the comment
> style :-)
I think we're on the same page actually.
> > The race is possible too. Not sure whether READ_ONCE/WRITE_ONCE
> > are really warranted though. Might be overkill.
>
> Is there concurrency on the jit?
The JIT of different progs can happen in parallel.
> > Once you have a git branch with all the changes I can give it a go.
>
> Ok, I'll go polish this thing and stick it in the tree mentioned in the
> cover letter.
>
> > Also you can rely on our BPF CI.
> > Just cc your patchset to bpf@vger and add [PATCH bpf-next] to a subject.
> > In patchwork there will be "bpf/vmtest-bpf-next" link that
> > builds kernel, selftests and runs everything.
>
> What's a patchwork and where do I find it?
https://patchwork.kernel.org/project/netdevbpf/list/?delegate=121173
Click on any patch, search for 'bpf/vmtest-bpf-next' and follow the
'VM_Test' link.
The summary of the test run is available without logging in into github.
To see detailed logs you need to be logged in with your github account.
It's a silly limitation they have.
They even have a button 'Sign in to view logs'. Oh well.
> > It's pretty much the same as selftests/bpf/vmtest.sh, but with the latest
> > clang nightly and other deps like pahole.
>
> nice.
One more thing. There is test_bpf.ko.
Just insmod it and it will run a ton of JIT tests that we cannot do
from user space.
Please use bpf-next tree though. Few weeks ago Johan Almbladh added
a lot more tests to it.
