Re: [PATCH bpf-next v5 1/3] bpf: Parameterize task iterators.

[Yonghong Song <yhs@fb.com>]

On 8/15/22 10:25 PM, Andrii Nakryiko wrote:
> On Sat, Aug 13, 2022 at 3:17 PM Yonghong Song <yhs@fb.com> wrote:
>>
>>
>>
>> On 8/10/22 5:16 PM, Kui-Feng Lee wrote:
>>> Allow creating an iterator that loops through resources of one task/thread.
>>>
>>> People could only create iterators to loop through all resources of
>>> files, vma, and tasks in the system, even though they were interested
>>> in only the resources of a specific task or process.  Passing the
>>> additional parameters, people can now create an iterator to go
>>> through all resources or only the resources of a task.
>>>
>>> Signed-off-by: Kui-Feng Lee <kuifeng@fb.com>
>>> ---
>>>    include/linux/bpf.h            |  29 ++++++++
>>>    include/uapi/linux/bpf.h       |   8 +++
>>>    kernel/bpf/task_iter.c         | 126 ++++++++++++++++++++++++++-------
>>>    tools/include/uapi/linux/bpf.h |   8 +++
>>>    4 files changed, 147 insertions(+), 24 deletions(-)
>>>
> 
> Btw, Yonghong, I tried to figure it out myself, but got lost in all
> the kernel functions that don't seem to be very well documented. Sorry
> for being lazy and throwing this at you :)
> 
> Is it easy and efficient to iterate only processes using whatever
> kernel helpers we have at our disposal? E.g., if I wanted to write an
> iterator that would go only over processes (not individual threads,
> just task leaders of each different process) within a cgroup, is that
> possible?

To traverse processes in a cgroup, the best location is in
kernel/cgroup/cgroup.c where there exists a seq_ops to
traverse all processes in cgroup.procs file. If we try
to implement a bpf based iterator, we could reuse some
codes in that file.

> 
> I see task iterator as consisting of two different parts (and that
> makes it a bit hard to define nice and clean interface, but if we can
> crack this, we'd get an elegant and powerful construct):
> 
> 1. What entity to iterate: threads or processes? (I'm ignoring
> task_vma and task_files here, but one could task about files of each
> thread or files of each process, but it's less practical, probably)
> 
> 2. What's the scope of objects to iterate: just a thread by tid, just
> a process by pid/pidfd, once cgroup iter lands, we'll be able to talk
> about threads or processes within a cgroup or cgroup hierarchy (this
> is where descendants_{pre,post}, cgroup_self_only and ancestors
> ordering comes in as well).
> 
> Currently Kui-Feng is addressing first half of #2 (tid/pid/pidfd
> parameters), we can use cgroup iter's parameters to define the scope
> of tasks/processes by cgroup "filter" in a follow up (it naturally
> extends what we have in this patch set).

For #2 as well, it is also possible to have a complete new seq_ops
if the traversal is only once. That is why in Kui-Feng's patch,
there are a few special case w.r.t. TID. But current approach
is also okay.

> 
> So now I'm wondering if there is any benefit to also somehow
> specifying threads vs processes as entities to iterate? And if we do
> that, does kernel support efficient iteration of processes (as opposed
> to threads).

IIUC, I didn't find an efficient way to traverse processes only.
The current pid_ns.idr records all tasks so traversing processes
have to skip intermediate non-main-thread tasks.

> 
> 
> To be clear, there is a lot of value in having just #2, but while we
> are all at this topic, I thought I'd clarify for myself #1 as well.
> 
> Thanks!