Re: [PATCH bpf-next v10 2/4] bpf: new helper to obtain namespace data from current task New bpf helper bpf_get_current_pidns_info.

From: Yonghong Song <yhs@fb.com>
To: Al Viro <viro@zeniv.linux.org.uk>, Carlos Neira <cneirabustos@gmail.com>
Cc: "netdev@vger.kernel.org" <netdev@vger.kernel.org>,
	"ebiederm@xmission.com" <ebiederm@xmission.com>,
	"brouer@redhat.com" <brouer@redhat.com>,
	"bpf@vger.kernel.org" <bpf@vger.kernel.org>
Subject: Re: [PATCH bpf-next v10 2/4] bpf: new helper to obtain namespace data from current task New bpf helper bpf_get_current_pidns_info.
Date: Fri, 6 Sep 2019 23:21:14 +0000	[thread overview]
Message-ID: <c0e67fc7-be66-c4c6-6aad-316cbba18757@fb.com> (raw)
In-Reply-To: <20190906160020.GX1131@ZenIV.linux.org.uk>

On 9/6/19 9:00 AM, Al Viro wrote:
> On Fri, Sep 06, 2019 at 04:46:47PM +0100, Al Viro wrote:
> 
>>> Where do I begin?
>>> 	* getname_kernel() is there for purpose
>>> 	* so's kern_path(), damnit
>>
>> Oh, and filename_lookup() *CAN* sleep, obviously.  So that
>> GFP_ATOMIC above is completely pointless.
>>
>>>> +
>>>> +	inode = d_backing_inode(kp.dentry);
>>>> +	pidns_info->dev = (u32)inode->i_rdev;
> 
> In the original variant of patchset it used to be ->i_sb->s_dev,
> which is also bloody strange - you are not asking filename_lookup()
> to follow symlinks, so you'd get that of whatever filesystem
> /proc/self/ns resides on.
> 
> ->i_rdev use makes no sense whatsoever - it's a symlink and
> neither it nor its target are device nodes; ->i_rdev will be
> left zero for both.
> 
> What data are you really trying to get there?

Let me explain a little bit background here.
The ultimate goal is for bpf program to filter over
(pid_namespace, tgid/pid inside pid_namespace)
so bpf based tools can run inside the container.

Typically, pid namespace is achieved by looking at
/proc/self/ns/pid:
-bash-4.4$ lsns
         NS TYPE   NPROCS   PID USER COMMAND
4026531835 cgroup     44  8261 yhs  /usr/lib/systemd/systemd --user
4026531836 pid        44  8261 yhs  /usr/lib/systemd/systemd --user
4026531837 user       44  8261 yhs  /usr/lib/systemd/systemd --user
4026531838 uts        44  8261 yhs  /usr/lib/systemd/systemd --user
4026531839 ipc        44  8261 yhs  /usr/lib/systemd/systemd --user
4026531840 mnt        44  8261 yhs  /usr/lib/systemd/systemd --user
4026532008 net        44  8261 yhs  /usr/lib/systemd/systemd --user
-bash-4.4$ readlink /proc/self/ns/pid
pid:[4026531836]
-bash-4.4$ stat /proc/self/ns/pid
   File: ‘/proc/self/ns/pid’ -> ‘pid:[4026531836]’
   Size: 0               Blocks: 0          IO Block: 1024   symbolic link
Device: 4h/4d   Inode: 344795989   Links: 1
Access: (0777/lrwxrwxrwx)  Uid: (128203/     yhs)   Gid: (  100/   users)
Context: user_u:base_r:base_t
Access: 2019-09-06 16:06:09.431616380 -0700
Modify: 2019-09-06 16:06:09.431616380 -0700
Change: 2019-09-06 16:06:09.431616380 -0700
  Birth: -
-bash-4.4$

Based on a discussion with Eric Biederman back in 2019 Linux
Plumbers, Eric suggested that to uniquely identify a
namespace, device id (major/minor) number should also
be included. Although today's kernel implementation
has the same device for all namespace pseudo files,
but from uapi perspective, device id should be included.

That is the reason why we try to get device id which holds
pid namespace pseudo file.

Do you have a better suggestion on how to get
the device id for 'current' pid namespace? Or from design, we
really should not care about device id at all?