Re: [RFC bpf-next 3/4] bpf: Resolve symbols with kallsyms_lookup_names for kprobe multi link

[Jiri Olsa]

On Mon, Apr 11, 2022 at 03:15:32PM -0700, Andrii Nakryiko wrote:
> On Thu, Apr 7, 2022 at 5:53 AM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
> >
> > Using kallsyms_lookup_names function to speed up symbols lookup in
> > kprobe multi link attachment and replacing with it the current
> > kprobe_multi_resolve_syms function.
> >
> > This speeds up bpftrace kprobe attachment:
> >
> >   # perf stat -r 5 -e cycles ./src/bpftrace -e 'kprobe:x* {  } i:ms:1 { exit(); }'
> >   ...
> >   6.5681 +- 0.0225 seconds time elapsed  ( +-  0.34% )
> >
> > After:
> >
> >   # perf stat -r 5 -e cycles ./src/bpftrace -e 'kprobe:x* {  } i:ms:1 { exit(); }'
> >   ...
> >   0.5661 +- 0.0275 seconds time elapsed  ( +-  4.85% )
> >
> > Signed-off-by: Jiri Olsa <jolsa@xxxxxxxxxx>
> > ---
> >  kernel/trace/bpf_trace.c | 123 +++++++++++++++++++++++----------------
> >  1 file changed, 73 insertions(+), 50 deletions(-)
> >
> > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> > index b26f3da943de..2602957225ba 100644
> > --- a/kernel/trace/bpf_trace.c
> > +++ b/kernel/trace/bpf_trace.c
> > @@ -2226,6 +2226,72 @@ struct bpf_kprobe_multi_run_ctx {
> >         unsigned long entry_ip;
> >  };
> >
> > +struct user_syms {
> > +       const char **syms;
> > +       char *buf;
> > +};
> > +
> > +static int copy_user_syms(struct user_syms *us, void __user *usyms, u32 cnt)
> > +{
> > +       const char __user **usyms_copy = NULL;
> > +       const char **syms = NULL;
> > +       char *buf = NULL, *p;
> > +       int err = -EFAULT;
> > +       unsigned int i;
> > +       size_t size;
> > +
> > +       size = cnt * sizeof(*usyms_copy);
> > +
> > +       usyms_copy = kvmalloc(size, GFP_KERNEL);
> > +       if (!usyms_copy)
> > +               return -ENOMEM;
> 
> do you really need usyms_copy? why not just read one pointer at a time?
> 
> > +
> > +       if (copy_from_user(usyms_copy, usyms, size))
> > +               goto error;
> > +
> > +       err = -ENOMEM;
> > +       syms = kvmalloc(size, GFP_KERNEL);
> > +       if (!syms)
> > +               goto error;
> > +
> > +       /* TODO this potentially allocates lot of memory (~6MB in my tests
> > +        * with attaching ~40k functions). I haven't seen this to fail yet,
> > +        * but it could be changed to allocate memory gradually if needed.
> > +        */
> > +       size = cnt * KSYM_NAME_LEN;
> 
> this reassignment of size is making it hard to follow the code, you
> can just do cnt * KSYM_NAME_LEN inside kvmalloc, you don't ever use it
> anywhere else

ok

> 
> > +       buf = kvmalloc(size, GFP_KERNEL);
> > +       if (!buf)
> > +               goto error;
> > +
> > +       for (p = buf, i = 0; i < cnt; i++) {
> 
> like here, before doing strncpy_from_user() you can read usyms[i] from
> user-space into temporary variable, no need for extra kvmalloc?

yes, that could work.. one copy_from_user seemed faster than separate
get_user calls, but then it's without memory allocation.. so perhaps
that's better

jirka

> 
> > +               err = strncpy_from_user(p, usyms_copy[i], KSYM_NAME_LEN);
> > +               if (err == KSYM_NAME_LEN)
> > +                       err = -E2BIG;
> > +               if (err < 0)
> > +                       goto error;
> > +               syms[i] = p;
> > +               p += err + 1;
> > +       }
> > +
> > +       err = 0;
> > +       us->syms = syms;
> > +       us->buf = buf;
> > +
> 
> [...]