Re: [GIT PULL] ipc: Bind to the ipc namespace at open time.

On Thu, Mar 24, 2022 at 11:12:21AM -0700, Linus Torvalds wrote:
> On Wed, Mar 23, 2022 at 1:24 PM Eric W. Biederman <[email protected]> wrote:
> >
> > Please pull the per-namespace-ipc-sysctls-for-v5.18 tag from the git tree:
> 
> Ugh.
> 
> I pulled this. Then I stared at it for a long time.
> 
> And then I decided that this is too ugly to live.
> 
> I'm sorry. I think Alexey has probably spent a fair amount of time on
> it, but I really think the sysctl code needs to be cleaned up way more
> than this.

Apparently it's my fault that the purpose of these changes is not clear. I
did this refactoring not for the sake of refactoring as such.

In my original patch [1], I was trying to fix the situation where the user
cannot change the /proc/sys/fs/mqueue/* options inside rootless container.

But then I split the changes into refactoring which fixes the hack and
permission changes which I wanted to discuss and propose later.

[1] https://lore.kernel.org/lkml/0f0408bb7fbf3187966a9bf19a98642a5d9669fe.1641225760.git.legion@kernel.org/

> The old code was horribly hacky too, but that setup_ipc_sysctls() (and
> setup_mq_sysctls()) thing that copies the whole sysctls table, and
> then walks it entry by entry to modify it, is just too ugly for words.
> 
> And then it hides things in .extra1, and because it does that it can't
> use the normal "extra1 and extra2 contains the limits" so then at
> write() time it copies it into a local one AGAIN only to set the
> limits back so that it can call the normal routines again.
> 
> Not ok.
> 
> Yes, yes, the old code did some similar things - to set the 'data'
> pointer. That was disgusting too. Don't get me wrong - the existing
> code was nasty too. But this took nasty code, and doubled down on it.
> 
> I really think this is a fundamental problem, and needs a more
> fundamental fix than adding more and more of these kinds of nasty
> hacks.
> 
> And yes, that fundamental fix is almost certainly to pass in 'struct
> cred *' to all those sysctl helper functions.
> 
> Then, when you do the actual 'sysctl()' system calls, you pass in
> 'current_cred()".
> 
> And the /proc users would pass in file->f_cred.
> 
> And yes, that patch might be quite messy, because we have quite a lot
> of those random .proc_handler users.
> 
> But *most* of them by far (at least in random code) use the standard
> proc_dointvec_minmax() and friends, and won't even notice.
> 
> And then the ones that are about namespace issues will have to
> continue to do the nasty "make a copy and update the data pointer",
> but *MAYBE* we could also introduce the notion of an "offset" to those
> proc_dointvec_minmax() things to help them out (and at least avoid the
> "make a copy" thing).
> 
> Anyway, I really think we must not make that sysctl code even uglier
> than it is today, and I think we need to move towards a model that
> actually makes sense. And that "pass in the right cred" is the only
> sensible model I can see.
> 
> I haven't tried to create such a patch, and maybe Alexey already tried
> to do something like that and it turned out to be too ugly for words
> and that's why these nasty patches happened.
> 
> But at least for now, I can't with a good conscience pull this.
> 
> Sorry,
>                    Linus

OK. I'll try to come up with some other solution.

-- 
Rgrds, legion