lkml.org 
[lkml]   [2021]   [Apr]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH bpf-next v2 2/2] bpf: Implement formatted output helpers with bstr_printf
On Tue, Apr 27, 2021 at 5:20 PM Florent Revest <revest@chromium.org> wrote:
>
> On Wed, Apr 28, 2021 at 1:46 AM Alexei Starovoitov
> <alexei.starovoitov@gmail.com> wrote:
> >
> > On Tue, Apr 27, 2021 at 10:43 AM Florent Revest <revest@chromium.org> wrote:
> > > + if (fmt[i + 1] == 'B') {
> > > + if (tmp_buf) {
> > > + err = snprintf(tmp_buf,
> > > + (tmp_buf_end - tmp_buf),
> > > + "%pB",
> > ...
> > > + if ((tmp_buf_end - tmp_buf) < sizeof_cur_ip) {
> >
> > I removed a few redundant () like above
>
> Oh, sorry about that.
>
> > and applied.
>
> Nice! :)
>
> > > if (fmt[i] == 'l') {
> > > - cur_mod = BPF_PRINTF_LONG;
> > > + sizeof_cur_arg = sizeof(long);
> > > i++;
> > > }
> > > if (fmt[i] == 'l') {
> > > - cur_mod = BPF_PRINTF_LONG_LONG;
> > > + sizeof_cur_arg = sizeof(long long);
> > > i++;
> > > }
> >
> > This bit got me thinking.
> > I understand that this is how bpf_trace_printk behaved
> > and the sprintf continued the tradition, but I think it will
> > surprise bpf users.
> > The bpf progs are always 64-bit. The sizeof(long) == 8
> > inside any bpf program. So printf("%ld") matches that long.
>
> Yes, this also surprised me.
>
> > The clang could even do type checking to make sure the prog
> > is passing the right type into printf() if we add
> > __attribute__ ((format (printf))) to bpf_helper_defs.h
> > But this sprintf() implementation will trim the value to 32-bit
> > to satisfy 'fmt' string on 32-bit archs.
> > So bpf program behavior would be different on 32 and 64-bit archs.
> > I think that would be confusing, since the rest of bpf prog is
> > portable. The progs work the same way on all archs
> > (except endianess, of course).
> > I'm not sure how to fix it though.
> > The sprintf cannot just pass 64-bit unconditionally, since
> > bstr_printf on 32-bit archs will process %ld incorrectly.
> > The verifier could replace %ld with %Ld.
> > The fmt string is a read only string for bpf_snprintf,
> > but for bpf_trace_printk it's not and messing with it at run-time
> > is not good. Copying the fmt string is not great either.
> > Messing with internals of bstr_printf is ugly too.
>
> Indeed, none of these solutions are satisfying.

Maybe Daniel has other ideas?

> > Maybe we just have to live with this quirk ?
>
> If we were starting from scratch, maybe just banning %ld could have
> been an option, but now that bpf_trace_printk has been behaving like
> this for a while, I think it might be best to just keep the behavior
> as it is.
>
> > Just add a doc to uapi/bpf.h to discourage %ld and be done?
>
> More doc is always good. Something like "Note: %ld behaves differently
> depending on the host architecture, it is recommended to avoid it and
> use %d or %lld instead" in the helper description of the three
> helpers? If you don't have the time to do it today, I can send a patch
> tomorrow.

bpf_trace_printk was like this for a long time, so there is no rush.
Pls wait until everything comes back to bpf tree and send a patch against it.
bpf_trace_printk comment in uapi/bpf.h is outdated too. Would be good
to document the latest behavior for them all.

\
 
 \ /
  Last update: 2021-04-28 02:52    [W:0.034 / U:0.576 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site