Re: [RFC PATCH bpf-next 05/16] bpf: create file or anonymous dumpers

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Tue, Apr 14, 2020 at 09:45:08PM -0700, Andrii Nakryiko wrote:
> >
> > > FD is closed, dumper program is detached and dumper is destroyed
> > > (unless pinned in bpffs, just like with any other bpf_link.
> > > 3. At this point bpf_dumper_link can be treated like a factory of
> > > seq_files. We can add a new BPF_DUMPER_OPEN_FILE (all names are for
> > > illustration purposes) command, that accepts dumper link FD and
> > > returns a new seq_file FD, which can be read() normally (or, e.g.,
> > > cat'ed from shell).
> >
> > In this case, link_query may not be accurate if a bpf_dumper_link
> > is created but no corresponding bpf_dumper_open_file. What we really
> > need to iterate through all dumper seq_file FDs.
> 
> If the goal is to iterate all the open seq_files (i.e., bpfdump active
> sessions), then bpf_link is clearly not the right approach. But I
> thought we are talking about iterating all the bpfdump programs
> attachments, not **sessions**, in which case bpf_link is exactly the
> right approach.

That's an important point. What is the pinned /sys/kernel/bpfdump/tasks/foo ?
Every time 'cat' opens it a new seq_file is created with new FD, right ?
Reading of that file can take infinite amount of time, since 'cat' can be
paused in the middle.
I think we're dealing with several different kinds of objects here.
1. "template" of seq_file that is seen with 'ls' in /sys/kernel/bpfdump/
2. given instance of seq_file after "template" was open
3. bpfdumper program
4. and now links. One bpf_link from seq_file template to bpf prog and
  many other bpf_links from actual seq_file kernel object to bpf prog.
  I think both kinds of links need to be iteratable via get_next_id.

At the same time I don't think 1 and 2 are links.
read-ing link FD should not trigger program execution. link is the connecting
abstraction. It shouldn't be used to trigger anything. It's static.
Otherwise read-ing cgroup-bpf link would need to trigger cgroup bpf prog too.
FD that points to actual seq_file is the one that should be triggering
iteration of kernel objects and corresponding execution of linked prog.
That FD can be anon_inode returned from raw_tp_open (or something else)
or FD from open("/sys/kernel/bpfdump/foo").

The more I think about all the objects involved the more it feels that the
whole process should consist of three steps (instead of two).
1. load bpfdump prog
2. create seq_file-template in /sys/kernel/bpfdump/
   (not sure which api should do that)
3. use bpf_link_create api to attach bpfdumper prog to that seq_file-template

Then when the file is opened a new bpf_link is created for that reading session.
At the same time both kinds of links (to teamplte and to seq_file) should be
iteratable for observability reasons, but get_fd_from_id on them should probably
be disallowed, since holding such FD to these special links by other process
has odd semantics.

Similarly for anon seq_file it should be three step process as well:
1. load bpfdump prog
2. create anon seq_file (api is tbd) that returns FD
3. use bpf_link_create to attach prog to seq_file FD

May be it's all overkill. These are just my thoughts so far.