Re: [PATCH bpf-next v2 3/6] bpf: Introduce process open coded iterator kfuncs

From: Andrii Nakryiko
Date: Tue Sep 19 2023 - 19:30:27 EST


On Sat, Sep 16, 2023 at 7:03 AM Chuyi Zhou <zhouchuyi@xxxxxxxxxxxxx> wrote:
>
> Hello.
>
> 在 2023/9/16 04:37, Andrii Nakryiko 写道:
> > On Fri, Sep 15, 2023 at 8:03 AM Chuyi Zhou <zhouchuyi@xxxxxxxxxxxxx> wrote:
> >> 在 2023/9/15 07:26, Andrii Nakryiko 写道:
> >>> On Tue, Sep 12, 2023 at 12:02 AM Chuyi Zhou <zhouchuyi@xxxxxxxxxxxxx> wrote:
> >>>>
> >>>> This patch adds kfuncs bpf_iter_process_{new,next,destroy} which allow
> >>>> creation and manipulation of struct bpf_iter_process in open-coded iterator
> >>>> style. BPF programs can use these kfuncs or through bpf_for_each macro to
> >>>> iterate all processes in the system.
> >>>>
> [...cut...]
> >>>
> >>> Few high level thoughts. I think it would be good to follow
> >>> SEC("iter/task") naming and approach. Open-coded iterators in many
> >>> ways are in-kernel counterpart to iterator programs, so keeping them
> >>> close enough within reason is useful for knowledge transfer.
> >>>
> >>> SEC("iter/task") allows to:
> >>> a) iterate all threads in the system
> >>> b) iterate all threads for a given TGID
> >>> c) it also allows to "iterate" a single thread or process, but that's
> >>> a bit less relevant for in-kernel iterator, but we can still support
> >>> them, why not?
> >>>
> >>> I'm not sure if it supports iterating all processes (as in group
> >>> leaders of each task group) in the system, but if it's possible I
> >>> think we should support it at least for open-coded iterator, seems
> >>> like a very useful functionality.
> >>>
> >>> So to that end, let's design a small set of input arguments for
> >>> bpf_iter_process_new() that would allow to specify this as flags +
> >>> either (optional) struct task_struct * pointer to represent
> >>> task/process or PID/TGID.
> >>>
> >>
> >> Another concern from Alexei was the readability of the API of open-coded
> >> in BPF Program[1].
> >>
> >> bpf_for_each(task, curr) is straightforward. Users can easily understand
> >> that this API does the same thing as 'for_each_process' in kernel.
> >
> > In general, users might have no idea about for_each_process macro in
> > the kernel, so I don't find this particular argument very convincing.
> >
> > We can add a separate set of iterator kfuncs for every useful
> > combination of conditions, of course, but it's a double-edged sword.
> > Needing to use a different iterator just to specify a different
> > direction of cgroup iteration (from the example you referred in [1])
> > also means that it's now harder to write some generic function that
> > needs to do something for all cgroups matching some criteria where the
> > order might be coming as an argument.
> >
> > Similarly for task iterators. It's not hard to imagine some processing
> > that can be equivalently done per thread or per process in the system,
> > or on each thread of the process, depending on some conditions or
> > external configuration. Having to do three different
> > bpf_for_each(task_xxx, task, ...) for this seems suboptimal. If the
> > nature of the thing that is iterated over is the same, and it's just a
> > different set of filters to specify which subset of those items should
> > be iterated, I think it's better to try to stick to the same iterator
> > with few simple arguments. IMO, of course, there is no objectively
> > best approach.
> >
> >>
> >> However, if we keep the approach of SEC("iter/task")
> >>
> >> enum ITER_ITEM {
> >> ITER_TASK,
> >> ITER_THREAD,
> >> }
> >>
> >> __bpf_kfunc int bpf_iter_task_new(struct bpf_iter_process *it, struct
> >> task_struct *group_task, enum ITER_ITEM type)
> >>
> >> the API have to chang:
> >>
> >>
> >> bpf_for_each(task, curr, NULL, ITERATE_TASK) // iterate all process in
> >> the system
> >> bpf_for_each(task, curr, group_leader, ITERATE_THREAD) // iterate all
> >> thread of group_leader
> >> bpf_for_each(task, curr, NULL, ITERATE_THREAD) //iterate all threads of
> >> all the process in the system
> >>
> >> Useres may guess what are this API actually doing....
> >
> > I'd expect users to consult documentation before trying to use an
> > unfamiliar cutting-edge functionality. So let's try to keep
> > documentation clear and up to the point. Extra flag argument doesn't
> > seem to be a big deal.
>
> Thanks for your suggestion!
>
> Before we begin working on the next version, I have outlined a detailed
> API design here:
>
> 1.task_iter
>
> It will be used to iterate process/threads like SEC("iter/task"). Here
> we should better to follow the naming and approach SEC("iter/task"):
>
> enum {
> ITERATE_PROCESS,
> ITERATE_THREAD,
> }
>
> __bpf_kfunc int bpf_iter_task_new(struct bpf_iter_task *it, struct
> task_struct *task, int flag);
>
> If we want to iterate all processes in the system, the iteration will
> start from the *task* which is passed from user.(since process in the
> system are connected through a linked list)

but will go through all of them anyways, right? it's kind of
surprising from usability standpoint to have to pass some task_struct
to iterate all of them, tbh. I wonder if it's hard to adjust kfunc
validation to allow "nullable" pointers? We can look at that
separately, of course.

>
> Additionally, the *task* can allow users to specify iterating all
> threads within a task group.
>
> SEC("xxx")
> int xxxx(void *ctx)
> {
> struct task_struct *pos;
> struct task_struct *cur_task = bpf_get_current_task_btf();
>
> bpf_rcu_read_lock();
>
> // iterating all process in the system start from cur_task
> bpf_for_each(task, pos, cur_task, ITERATE_PROCESS) {
>
> }
>
> // iterate all thread belongs to cur_task group.
> bpf_for_each(task, pos, cur_task, ITERATE_THREAD) {
>
> }
>
> bpf_rcu_read_unlock();
> return 0;
> }
>
> Iterating all thread of each process is great(ITERATE_ALL). But maybe
> let's break it down step by step and implement
> ITERATE_PROCESS/ITERATE_THREAD first? (I'm little worried about the cpu
> overhead of ITERATE_ALL, since we are doing a heavy job in BPF Prog)
>

Hm... but if it was a sleepable BPF program and
bpf_rcu_read_{lock,unlock}() was only per task, then it shouldn't be
the problem? See enum bpf_cgroup_iter_order.


> I wanted to reuse BPF_TASK_ITER_ALL/BPF_TASK_ITER_TID/BPF_TASK_ITER_TGID
> insted of new enums like ITERATE_PROCESS/ITERATE_THREAD. But it seems
> necessary. In BPF Prog, we usually operate task_struct directly instead
> of pid/tgid. It's a little weird to use
> BPF_TASK_ITER_TID/BPF_TASK_ITER_TGID here:

enum bpf_iter_task_type is internal type, so we can rename
BPF_TASK_ITER_TID to BPF_TASK_ITER_THREAD and BPF_TASK_ITER_PROC (or
add them as aliases). At the very least, we should use consistent
BPF_TASK_ITER_xxx naming, instead of just ITERATE_PROCESS. See

>
> bpf_for_each(task, pos, cur_task, BPF_TASK_ITER_TID) {
> }
>
> On the other hand,
> BPF_TASK_ITER_ALL/BPF_TASK_ITER_TID/BPF_TASK_ITER_TGID are inner flags
> that are hidden from the users.
> Exposing ITERATE_PROCESS/ITERATE_THREAD will not cause confusion to user.
>

inner types are not a problem when used with vmlinux.h


>
> 2. css_iter.
>
> css_iter will be used to:
> (1) iterating subsystem, like
> for_each_mem_cgroup_tree/cpuset_for_each_descendant_pre in kernel.
> (2) iterating cgroup. (patch-6's selfetest has a basic example)
>
> css(cgroup_subsys_state) is more fundamental than struct cgroup. I think
> we'd better operating css rather than cgroup, since it's can be hard for
> cgroup_iter to achive (2). So here we keep the name of "css_iter",
> BPF_CGROUP_ITER_DESCENDANTS_PRE/BPF_CGROUP_ITER_DESCENDANTS_POST/BPF_CGROUP_ITER_ANCESTORS_UP
> can be reused.
>
>
> __bpf_kfunc int bpf_iter_css_new(struct bpf_iter_css *it,
> struct cgroup_subsys_state *root, unsigned int flag)
>
> bpf_for_each(css, root, BPF_CGROUP_ITER_DESCENDANTS_PRE)
>

Makes sense, yep, thanks.

> Thanks.
>
>
>
>