On Wed, May 31, 2023 at 8:29 AM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Tue, May 30, 2023 at 09:39:01AM +0800, Yafang Shao wrote:
> > On Mon, May 29, 2023 at 8:06 PM Jiri Olsa <olsajiri@xxxxxxxxx> wrote:
> > >
> > > On Sun, May 28, 2023 at 02:20:20PM +0000, Yafang Shao wrote:
> > > > Currently, there is no way to check which functions are attached to a
> > > > kprobe_multi link, causing confusion for users. It is important that we
> > > > provide a means to expose these functions. The expected result is as follows,
> > > >
> > > > $ cat /proc/10936/fdinfo/9
> > > > pos: 0
> > > > flags: 02000000
> > > > mnt_id: 15
> > > > ino: 2094
> > > > link_type: kprobe_multi
> > > > link_id: 2
> > > > prog_tag: a04f5eef06a7f555
> > > > prog_id: 11
> > > > func_count: 4
> > > > func_addrs: ffffffffaad475c0
> > > > ffffffffaad47600
> > > > ffffffffaad47640
> > > > ffffffffaad47680
> > >
> > > I like the idea of exposing this through the link_info syscall,
> > > but I'm bit concerned of potentially dumping thousands of addresses
> > > through fdinfo file, because I always thought of fdinfo as brief
> > > file info, but that might be just my problem ;-)
> >
> > In most cases, there are only a few addresses, and it is uncommon to
>
> I doubt you have data to prove that kprobe_multi is "few addresses in most cases",
> so please don't throw such arguments without proof.
>
> > have thousands of addresses. To handle this, what about displaying a
> > maximum of 16 addresses? For cases where the number of addresses
> > exceeds 16, we can use '...' to represent the remaining addresses.
>
> at this point the kernel can pick random 16 kernel funcs and it won't be
> much worse.
>
> Asking users to do
> $ cat /proc/10936/fdinfo/9 | grep "func_addrs" -A 4 | \
> awk '{ if (NR ==1) {print $2} else {print $1}}' | \
> awk '{"grep " $1 " /proc/kallsyms"| getline f; print f}'
> ffffffffaad475c0 T schedule_timeout_interruptible
> ffffffffaad47600 T schedule_timeout_killable
>
> isn't a great interface either.
>
> The proper interface through fill_link_info and bpftool is good to have,
> but fdinfo shouldn't partially duplicate it. So drop this patch and others.
Sure, I will drop the ->show_fdinfo patches.
--
Regards
Yafang
