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.

[Yonghong Song <yhs@fb.com>]

On 9/6/19 5:10 PM, Al Viro wrote:
> On Fri, Sep 06, 2019 at 11:21:14PM +0000, Yonghong Song wrote:
> 
>> -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?
> 
> What the hell is "device id for pid namespace"?  This is the
> first time I've heard about that mystery object, so it's
> hard to tell where it could be found.
> 
> I can tell you what device numbers are involved in the areas
> you seem to be looking in.
> 
> 1) there's whatever device number that gets assigned to
> (this) procfs instance.  That, ironically, _is_ per-pidns, but
> that of the procfs instance, not that of your process (and
> those can be different).  That's what you get in ->st_dev
> when doing lstat() of anything in /proc (assuming that
> procfs is mounted there, in the first place).  NOTE:
> that's lstat(2), not stat(2).  stat(1) uses lstat(2),
> unless given -L (in which case it's stat(2) time).  The
> difference:
> 
> root@kvm1:~# 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: 17396       Links: 1
> Access: (0777/lrwxrwxrwx)  Uid: (    0/    root)   Gid: (    0/    root)
> Access: 2019-09-06 19:43:11.871312319 -0400
> Modify: 2019-09-06 19:43:11.871312319 -0400
> Change: 2019-09-06 19:43:11.871312319 -0400
>   Birth: -
> root@kvm1:~# stat -L /proc/self/ns/pid
>    File: /proc/self/ns/pid
>    Size: 0               Blocks: 0          IO Block: 4096   regular empty file
> Device: 3h/3d   Inode: 4026531836  Links: 1
> Access: (0444/-r--r--r--)  Uid: (    0/    root)   Gid: (    0/    root)
> Access: 2019-09-06 19:43:15.955313293 -0400
> Modify: 2019-09-06 19:43:15.955313293 -0400
> Change: 2019-09-06 19:43:15.955313293 -0400
>   Birth: -
> 
> The former is lstat, the latter - stat.
> 
> 2) device number of the filesystem where the symlink target lives.
> In this case, it's nsfs and there's only one instance on the entire
> system.  _That_ would be obtained by looking at st_dev in stat(2) on
> /proc/self/ns/pid (0:3 above).
> 
> 3) device number *OF* the symlink.  That would be st_rdev in lstat(2).
> There's none - it's a symlink, not a character or block device.  It's
> always zero and always will be zero.
> 
> 4) the same for the target; st_rdev in stat(2) results and again,
> there's no such beast - it's neither character nor block device.
> 
> Your code is looking at (3).  Please, reread any textbook on Unix
> in the section that would cover stat(2) and discussion of the
> difference between st_dev and st_rdev.
> 
> I have no idea what Eric had been talking about - it's hard to
> reconstruct by what you said so far.  Making nsfs per-userns,
> perhaps?  But that makes no sense whatsoever, not that userns
> ever had...  Cheap shots aside, I really can't guess what that's
> about.  Sorry.

Thanks for the detailed information. The device number we want
is nsfs. Indeed, currently, there is only one instance
on the entire system. But not exactly sure what is the possibility
to have more than one nsfs device in the future. Maybe per-userns
or any other criteria?

> 
> In any case, pathname resolution is *NOT* for the situations where
> you can't block.  Even if it's procfs (and from the same pidns as
> the process) mounted there, there is no promise that the target
> of /proc/self has already been looked up and not evicted from
> memory since then.  And in case of cache miss pathwalk will
> have to call ->lookup(), which requires locking the directory
> (rw_sem, shared).  You can't do that in such context.
> 
> And that doesn't even go into the possibility that process has
> something very different mounted on /proc.
> 
> Again, I don't know what it is that you want to get to, but
> I would strongly recommend finding a way to get to that data
> that would not involve going anywhere near pathname resolution.
> 
> How would you expect the userland to work with that value,
> whatever it might be?  If it's just a 32bit field that will
> never be read, you might as well store there the same value
> you store now (0, that is) in much cheaper and safer way ;-)

Suppose inside pid namespace, user can pass the device number,
say n1, (`stat -L /proc/self/ns/pid`) to bpf program (through map
or JIT). At runtime, bpf program will try to get device number,
say n2, for the 'current' process. If n1 is not the same as
n2, that means they are not in the same namespace. 'current'
is in the same pid namespace as the user iff
n1 == n2 and also pidns id is the same for 'current' and
the one with `lsns -t pid`.

Are you aware of any way to get the pidns device number
for 'current' without going through the pathname
lookup?