Re: [PATCH bpf] bpf: Don't WARN_ON_ONCE in bpf_bprintf_prepare

[Florent Revest]

On Thu, May 6, 2021 at 8:52 PM Andrii Nakryiko
<andrii.nakryiko@xxxxxxxxx> wrote:
>
> On Wed, May 5, 2021 at 3:29 PM Florent Revest <revest@xxxxxxxxxxxx> wrote:
> >
> > On Wed, May 5, 2021 at 10:52 PM Andrii Nakryiko
> > <andrii.nakryiko@xxxxxxxxx> wrote:
> > >
> > > On Wed, May 5, 2021 at 1:48 PM Andrii Nakryiko
> > > <andrii.nakryiko@xxxxxxxxx> wrote:
> > > >
> > > > On Wed, May 5, 2021 at 1:00 PM Daniel Borkmann <daniel@xxxxxxxxxxxxx> wrote:
> > > > >
> > > > > On 5/5/21 8:55 PM, Andrii Nakryiko wrote:
> > > > > > On Wed, May 5, 2021 at 9:23 AM Florent Revest <revest@xxxxxxxxxxxx> wrote:
> > > > > >>
> > > > > >> The bpf_seq_printf, bpf_trace_printk and bpf_snprintf helpers share one
> > > > > >> per-cpu buffer that they use to store temporary data (arguments to
> > > > > >> bprintf). They "get" that buffer with try_get_fmt_tmp_buf and "put" it
> > > > > >> by the end of their scope with bpf_bprintf_cleanup.
> > > > > >>
> > > > > >> If one of these helpers gets called within the scope of one of these
> > > > > >> helpers, for example: a first bpf program gets called, uses
> > > > > >
> > > > > > Can we afford having few struct bpf_printf_bufs? They are just 512
> > > > > > bytes, so can we have 3-5 of them? Tracing low-level stuff isn't the
> > > > > > only situation where this can occur, right? If someone is doing
> > > > > > bpf_snprintf() and interrupt occurs and we run another BPF program, it
> > > > > > will be impossible to do bpf_snprintf() or bpf_trace_printk() from the
> > > > > > second BPF program, etc. We can't eliminate the probability, but
> > > > > > having a small stack of buffers would make the probability so
> > > > > > miniscule as to not worry about it at all.
> > > > > >
> > > > > > Good thing is that try_get_fmt_tmp_buf() abstracts all the details, so
> > > > > > the changes are minimal. Nestedness property is preserved for
> > > > > > non-sleepable BPF programs, right? If we want this to work for
> > > > > > sleepable we'd need to either: 1) disable migration or 2) instead of
> > > >
> > > > oh wait, we already disable migration for sleepable BPF progs, so it
> > > > should be good to do nestedness level only
> > >
> > > actually, migrate_disable() might not be enough. Unless it is
> > > impossible for some reason I miss, worst case it could be that two
> > > sleepable programs (A and B) can be intermixed on the same CPU: A
> > > starts&sleeps - B starts&sleeps - A continues&returns - B continues
> > > and nestedness doesn't work anymore. So something like "reserving a
> > > slot" would work better.
> >
> > Iiuc try_get_fmt_tmp_buf does preempt_enable to avoid that situation ?
> >
> > > >
> > > > > > assuming a stack of buffers, do a loop to find unused one. Should be
> > > > > > acceptable performance-wise, as it's not the fastest code anyway
> > > > > > (printf'ing in general).
> > > > > >
> > > > > > In any case, re-using the same buffer for sort-of-optional-to-work
> > > > > > bpf_trace_printk() and probably-important-to-work bpf_snprintf() is
> > > > > > suboptimal, so seems worth fixing this.
> > > > > >
> > > > > > Thoughts?
> > > > >
> > > > > Yes, agree, it would otherwise be really hard to debug. I had the same
> > > > > thought on why not allowing nesting here given users very likely expect
> > > > > these helpers to just work for all the contexts.
> > > > >
> > > > > Thanks,
> > > > > Daniel
> >
> > What would you think of just letting the helpers own these 512 bytes
> > buffers as local variables on their stacks ? Then bpf_prepare_bprintf
> > would only need to write there, there would be no acquire semantic
> > (like try_get_fmt_tmp_buf) and the stack frame would just be freed on
> > the helper return so there would be no bpf_printf_cleanup either. We
> > would also not pre-reserve static memory for all CPUs and it becomes
> > trivial to handle re-entrant helper calls.
> >
> > I inherited this per-cpu buffer from the pre-existing bpf_seq_printf
> > code but I've not been convinced of its necessity.
>
> I got the impression that extra 512 bytes on the kernel stack is quite
> a lot and that's why we have per-cpu buffers. Especially that
> bpf_trace_printk() can be called from any context, including NMI.

Ok, I understand.

What about having one buffer per helper, synchronized with a spinlock?
Actually, bpf_trace_printk already has that, not for the bprintf
arguments but for the bprintf output so this wouldn't change much to
the performance of the helpers anyway:
https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/kernel/trace/bpf_trace.c?id=9d31d2338950293ec19d9b095fbaa9030899dcb4#n385

These helpers are not performance sensitive so a per-cpu stack of
buffers feels over-engineered to me (and is also complexity I feel a
bit uncomfortable with).