lkml.org 
[lkml]   [2016]   [Nov]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRE: [RFC][PATCH 2/7] kref: Add kref_read()
Date
> On Fri, Nov 18, 2016 at 05:33:35PM +0000, Reshetova, Elena wrote:
> > On Thu, Nov 17, 2016 at 09:53:42AM +0100, Peter Zijlstra wrote:
> > > On Wed, Nov 16, 2016 at 12:08:52PM -0800, Alexei Starovoitov wrote:
> > >
> > > > I prefer to avoid 'fixing' things that are not broken.
> > > > Note, prog->aux->refcnt already has explicit checks for overflow.
> > > > locked_vm is used for resource accounting and not refcnt, so I don't
> > > > see issues there either.
> > >
> > > The idea is to use something along the lines of:
> > >
> > >
> > >
> http://lkml.kernel.org/r/20161115104608.GH3142@twins.programming.kicks
> > > -ass.net
> > >
> > > for all refcounts in the kernel.
> >
> > >I understand the idea. I'm advocating to fix refcnts explicitly the way we did
> in bpf land instead of leaking memory, making processes unkillable and so on.
> > >If refcnt can be bounds checked, it should be done that way, since it's a
> clean error path without odd side effects.
> > >Therefore I'm against unconditionally applying refcount to all atomics.
> >
> > > Also note that your:
> > >
> > > struct bpf_prog *bpf_prog_add(struct bpf_prog *prog, int i) {
> > > if (atomic_add_return(i, &prog->aux->refcnt) > BPF_MAX_REFCNT) {
> > > atomic_sub(i, &prog->aux->refcnt);
> > > return ERR_PTR(-EBUSY);
> > > }
> > > return prog;
> > > }
> > >
> > > is actually broken in the face of an actual overflow. Suppose @i is
> > > big enough to wrap refcnt into negative space.
> >
> > >'i' is not controlled by user. It's a number of nic hw queues and
> BPF_MAX_REFCNT is 32k, so above is always safe.
> >
> > If I understand your code right, you export the bpf_prog_add() and anyone is
> free to use it
> > (some crazy buggy driver for example).
> > Currently only drivers/net/ethernet/mellanox/mlx4/en_netdev.c uses it, but
> you should
> > consider any externally exposed interface as an attack vector from security
> point of view.
>
> It's not realistic to harden all export_symbol apis.
> Code review for in-tree modules is the only defense we have.
> Remember out of tree perf counter issues... nothing perf core can do
> about that. If it's out of tree, it's vendor's problem to fix it.

I am not trying to harden them all now, but since we are going through the list of
atomic_t variables that are used for refcounting, and this seems to be the one,
I was trying to find a way to convert it also since it isn't a big effort and would do
good at the end.
So, you are not fine if I convert the above code using only refcount_inc and refcount_dec_and_test
primitives?




\
 
 \ /
  Last update: 2016-11-21 09:18    [W:0.122 / U:0.288 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site