Re: [PATCH 1/3] memcg: limit the number of thresholds per-memcg

[Michal Hocko <mhocko-AlSwsSmVLrQ@public.gmane.org>]

On Thu 08-08-13 01:05:13, Kirill A. Shutemov wrote:
> On Wed, Aug 07, 2013 at 04:37:27PM +0200, Michal Hocko wrote:
> > On Wed 07-08-13 09:58:18, Tejun Heo wrote:
> > > Hello,
> > > 
> > > On Wed, Aug 07, 2013 at 03:46:54PM +0200, Michal Hocko wrote:
> > > > OK, I have obviously misunderstood your concern mentioned in the other
> > > > email. Could you be more specific what is the DoS scenario which was
> > > > your concern, then?
> > > 
> > > So, let's say the file is write-accessible to !priv user which is
> > > under reasonable resource limits.  Normally this shouldn't affect priv
> > > system tools which are monitoring the same event as it shouldn't be
> > > able to deplete resources as long as the resource control mechanisms
> > > are configured and functioning properly; however, the memory usage
> > > event puts all event listeners into a single contiguous table which a
> > > !priv user can easily expand to a size where the table can no longer
> > > be enlarged and if a priv system tool or another user tries to
> > > register event afterwards, it'll fail.  IOW, it creates a shared
> > > resource which isn't properly provisioned and can be trivially filled
> > > up making it an easy DoS target.
> > 
> > OK, got your point. You are right and I haven't considered the size of
> > the table and the size restrictions of kmalloc. Thanks for pointing this
> > out!
> > ---
> > From cde8a3333296eddd288780e78803610127401b6a Mon Sep 17 00:00:00 2001
> > From: Michal Hocko <mhocko-AlSwsSmVLrQ@public.gmane.org>
> > Date: Wed, 7 Aug 2013 11:11:22 +0200
> > Subject: [PATCH] memcg: limit the number of thresholds per-memcg
> > 
> > There is no limit for the maximum number of threshold events registered
> > per memcg. It is even worse that all the events are stored in a
> > per-memcg table which is enlarged when a new event is registered. This
> > can lead to the following issue mentioned by Tejun:
> > "
> > So, let's say the file is write-accessible to !priv user which is
> > under reasonable resource limits.  Normally this shouldn't affect priv
> > system tools which are monitoring the same event as it shouldn't be
> > able to deplete resources as long as the resource control mechanisms
> > are configured and functioning properly; however, the memory usage
> > event puts all event listeners into a single contiguous table which a
> > !priv user can easily expand to a size where the table can no longer
> > be enlarged and if a priv system tool or another user tries to
> > register event afterwards, it'll fail.  IOW, it creates a shared
> > resource which isn't properly provisioned and can be trivially filled
> > up making it an easy DoS target.
> > "
> > 
> > Let's be more strict and cap the number of events that might be
> > registered. MAX_THRESHOLD_EVENTS value is more or less random. The
> > expectation is that it should be high enough to cover reasonable
> > usecases while not too high to allow excessive resources consumption.
> > 1024 events consume something like 16KB which shouldn't be a big deal
> > and it should be good enough.
> 
> Is it correct that you fix one local DoS by introducing a new one?
> With the page the !priv user can block root from registering a threshold.
> Is it really the way we want to fix it?

OK, I will think about it some more.
-- 
Michal Hocko
SUSE Labs