On Tue, Feb 22, 2022 at 12:30 PM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Mon, Feb 21, 2022 at 6:26 AM Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
> >
> > On Mon, Feb 21, 2022 at 4:59 AM Alexei Starovoitov
> > <alexei.starovoitov@xxxxxxxxx> wrote:
> > >
> > > On Sun, Feb 20, 2022 at 6:17 AM Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
> > > >
> > > > On Sun, Feb 20, 2022 at 2:27 AM Alexei Starovoitov
> > > > <alexei.starovoitov@xxxxxxxxx> wrote:
> > > > >
> > > > > On Fri, Feb 18, 2022 at 1:56 AM Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
> > > > > >
> > > > > > Set the cgroup path when a bpf prog is attached to a cgroup, and unset
> > > > > > it when the bpf prog is detached.
> > > > > >
> > > > > > Below is the result after this change,
> > > > > > $ cat progs.debug
> > > > > > id name attached
> > > > > > 5 dump_bpf_map bpf_iter_bpf_map
> > > > > > 7 dump_bpf_prog bpf_iter_bpf_prog
> > > > > > 17 bpf_sockmap cgroup:/
> > > > > > 19 bpf_redir_proxy
> > > > > >
> > > > > > Signed-off-by: Yafang Shao <laoar.shao@xxxxxxxxx>
> > > > > > ---
> > > > > > kernel/bpf/cgroup.c | 8 ++++++++
> > > > > > 1 file changed, 8 insertions(+)
> > > > > >
> > > > > > diff --git a/kernel/bpf/cgroup.c b/kernel/bpf/cgroup.c
> > > > > > index 43eb3501721b..ebd87e54f2d0 100644
> > > > > > --- a/kernel/bpf/cgroup.c
> > > > > > +++ b/kernel/bpf/cgroup.c
> > > > > > @@ -440,6 +440,7 @@ static int __cgroup_bpf_attach(struct cgroup *cgrp,
> > > > > > struct bpf_cgroup_storage *storage[MAX_BPF_CGROUP_STORAGE_TYPE] = {};
> > > > > > struct bpf_cgroup_storage *new_storage[MAX_BPF_CGROUP_STORAGE_TYPE] = {};
> > > > > > enum cgroup_bpf_attach_type atype;
> > > > > > + char cgrp_path[64] = "cgroup:";
> > > > > > struct bpf_prog_list *pl;
> > > > > > struct list_head *progs;
> > > > > > int err;
> > > > > > @@ -508,6 +509,11 @@ static int __cgroup_bpf_attach(struct cgroup *cgrp,
> > > > > > else
> > > > > > static_branch_inc(&cgroup_bpf_enabled_key[atype]);
> > > > > > bpf_cgroup_storages_link(new_storage, cgrp, type);
> > > > > > +
> > > > > > + cgroup_name(cgrp, cgrp_path + strlen("cgroup:"), 64);
> > > > > > + cgrp_path[63] = '\0';
> > > > > > + prog->aux->attach_name = kstrdup(cgrp_path, GFP_KERNEL);
> > > > > > +
> > > > >
> > > > > This is pure debug code. We cannot have it in the kernel.
> > > > > Not even under #ifdef.
> > > > >
> > > > > Please do such debug code on a side as your own bpf program.
> > > > > For example by kprobe-ing in this function and keeping the path
> > > > > in a bpf map or send it to user space via ringbuf.
> > > > > Or enable cgroup tracepoint and monitor cgroup_mkdir with full path.
> > > > > Record it in user space or in bpf map, etc.
> > > > >
> > > >
> > > > It is another possible solution to hook the related kernel functions
> > > > or tracepoints, but it may be a little complicated to track all the
> > > > bpf attach types, for example we also want to track
> > > > BPF_PROG_TYPE_SK_MSG[1], BPF_PROG_TYPE_FLOW_DISSECTOR and etc.
> > > > While the attach_name provides us a generic way to get how the bpf
> > > > progs are attached, which can't be got by bpftool.
> > >
> > > bpftool can certainly print such details.
> > > See how it's using task_file iterator.
> > > It can be extended to look into cgroups and sockmap,
> > > and for each program print "sockmap:%d", map->id if so desired.
> >
> > I have read through the task_file code, but I haven't found a direct
> > way to get the attached cgroups or maps of a specified prog.
> > It is easy to look into a cgroup or sockmap, but the key point here is
> > which is the proper cgroup or sockmap.
> > There are some possible ways to get the attached cgroup or sockmap.
> >
> > - add new member into struct bpf_prog_aux
>
> No. Please stop proposing kernel changes for your debug needs.
>
> > For example,
> > struct bpf_prog_aux{
> > union {
> > struct cgroup *attached_cgrp;
> > struct bpf_map *attached_map;
> > };
> > };
> > Then we can easily get the related map or cgroup by extending
> > task_file iterator.
> >
> > - build a table for attached maps, cgroups and etc
> > Like we did for pinned files.
> > Then we can compare the prog with the members stored in this table
> > one by one, but that seems a little heavy.
> >
> > There may be some other ways.
>
> - iterate bpf maps
> - find sockmap
> - do equivalent of sock_map_progs(), but open coded inside bpf prog
> - read progs, print them.
That's similar to the second way I proposed above.
The trouble is it may be too heavy to iterate the target objects, for
example, there may be thousands of cgroups on a single host, and if we
want to know which cgroup is attached by a bpf prog we have to iterate
all the cgroups until find it.
But it seems there is no better choice.
Anyway thanks for the advice.
--
Thanks
Yafang
