Re: RFC [patch 13/34] PID Virtualization Define new task_pid api

[ebiederm@xmission.com (Eric W. Biederman)]

Linus Torvalds <torvalds@osdl.org> writes:

> (I'm coming in late, it's not been a high priority for me)

Thanks.  For taking the time.

> On Fri, 20 Jan 2006, Hubertus Franke wrote:
>> 
>> 2nd:
>> ====	Issue: we don't need pid virtualization, instead simply use
>> <container,pid> pair.
>> 
>> This requires a bit more thought. Essentially that's what I was doing, 
>> but I mangled them into the same pid and using masking to add/remove the 
>> container for internal use. As pointed out by Alan(?), we can indeed 
>> reused the same pid internally many times as long as we can distinguish 
>> during the pid-to-task_struct lookup. This is easily done because, the 
>> caller provides the context hence the container for the lookup.
>
> This is my preferred approach BY FAR.
>
> Doing a <container,pid> approach is very natural, and avoids almost all 
> issues. At most, you might want to have a new system call (most naturally 
> just the one that is limited to the "init container" - it the one that we 
> boot up with) that can specify both container and pid explicitly, and see 
> all processes and access all processes. But all "normal" system calls 
> would only ever operate within their container.

On this front I have been planning on using sys_clone as it allows
pieces of the virtualization to be incrementally built, it already
supports the FS namespace, and it supports flexibly specifying what
you want to contain.

> The fact is, we want "containers" anyway for any virtualization thing, ie 
> vserver already adds them. And if we have containers, then it's very easy 
> ("easyish") to split up the current static "pid_hash[]", "pidmap_array[]" 
> and "pidmap_lock", and make them per-container, and have a pointer to the 
> container for each "struct task_struct".

Well hash tables with their giant allocations are hard to split it has
been easier to add a container tag.

> After that, there wouldn't even be a lot else to do. The normal system 
> calls would just use their own container, and the (few) places that save 
> away pid's for later (ie things that use "kill_proc_info_as_uid()" and 
> "struct fown_struct" friends) would have to also squirrell away the 
> container, but then you should be pretty much done.

Yes.  Although there are a few container lifetimes problems with that
approach.  Do you want your container alive for a long time after every
process using it has exited just because someone has squirrelled away their
pid.  While container lifetime issues crop up elsewhere as well PIDs are
by far the worst, because it is current safe to store a PID indefinitely 
with nothing worse that PID wrap around.

> Of course, you'll have to do the system calls to _create_ the containers 
> in the first place, but that's at a higher level and involves much more 
> than just the pid-space (ie a container would normally have more than just 
> the uid mappings, it would have any network knowledge too etc - hostname, 
> perhaps list of network devices associated with that context etc etc)

Yes.  A list of network devices works seems to work well.

Eric