[PATCH 1/2] Adds a read-only "procs" file similar to "tasks" that shows only unique tgids

Andrew Morton akpm at linux-foundation.org
Fri Jul 3 09:50:00 PDT 2009


On Fri, 3 Jul 2009 09:11:56 -0700 Paul Menage <menage at google.com> wrote:

> Hmm, I guess we could use a combination of the IDR approach that you
> suggested and the present shared-array approach:
> 
> - when opening a tasks file:
>   - populate an IDR with all the pids/tgids in the cgroup
>   - find any existing IDR open for the cgroup in the list keyed by
> namespace and filetype ("procs"/"tasks")
>   - replace (and free) the existing IDR or extend the list with a new one
>   - bump the refcount
> 
> - when reading:
>   - iterate from the last reported pid/tgid
> 
> - when closing:
>   - drop the refcount, and free the IDR if the count reaches 0
> 
> That maintains the property of preventing userspace from consuming
> arbitrary amounts of memory just by holding an fd open on a large
> cgroup, while also maintaining a consistency guarantee, and we get the
> ordering for free as a side-effect of the IDR, with no large memory
> allocations. It's not hugely different from the current solution - all
> we're doing is replacing the large array in the cgroup_pidlist
> structure with an IDR, and populating/reading it appropriately.

I think you're saying "for each pid N in the cgroup, set the Nth
element in an IDR tree".  That would work.  And it automatically gives
ordered traversal and dupe removal.  

I don't think IDRs permit in-order traversal, whereas radix-trees do
support this.  Unfortunately radix-trees are presented as operating on
void* data, so one would need to do some typecasting when storing
BITS_PER_LONG-sized bitfields inside them.

> The downsides would be a higher fixed cost, I suspect - setting up an
> IDR and populating/scanning it sounds like it has to be more expensive
> than filling/reading a linear buffer. But it's probably not enough
> extra overhead to worry too much about it.

Yes, I expect it'd be fairly modest.  There will be far more calls to
kmalloc() when using a tree, but that's the whole point..



More information about the Containers mailing list