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.
It is not for debug-only purpose, while it gives us a better way to
maintain all the bpf progs running on a single host.
> Also please read Documentation/bpf/bpf_devel_QA.rst
> bpf patches should be based on bpf-next tree.
> These patches are not.
My local bpf-next repo is a little old.
Next time I will pull the newest bpf-next code before sending bpf
patches. Thanks for the information.
[1]. https://patchwork.kernel.org/project/netdevbpf/patch/20220218095612.52082-4-laoar.shao@xxxxxxxxx/
--
Thanks
Yafang
