Re: [PATCH ghak90 V8 13/16] audit: track container nesting

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 2020-02-13 16:49, Paul Moore wrote:
> On Wed, Feb 5, 2020 at 6:51 PM Richard Guy Briggs <rgb@xxxxxxxxxx> wrote:
> > On 2020-02-05 18:05, Paul Moore wrote:
> > > On Thu, Jan 30, 2020 at 2:28 PM Richard Guy Briggs <rgb@xxxxxxxxxx> wrote:
> > > > On 2020-01-22 16:29, Paul Moore wrote:
> > > > > On Tue, Dec 31, 2019 at 2:51 PM Richard Guy Briggs <rgb@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > Track the parent container of a container to be able to filter and
> > > > > > report nesting.
> > > > > >
> > > > > > Now that we have a way to track and check the parent container of a
> > > > > > container, modify the contid field format to be able to report that
> > > > > > nesting using a carrat ("^") separator to indicate nesting.  The
> > > > > > original field format was "contid=<contid>" for task-associated records
> > > > > > and "contid=<contid>[,<contid>[...]]" for network-namespace-associated
> > > > > > records.  The new field format is
> > > > > > "contid=<contid>[^<contid>[...]][,<contid>[...]]".
> > > > >
> > > > > Let's make sure we always use a comma as a separator, even when
> > > > > recording the parent information, for example:
> > > > > "contid=<contid>[,^<contid>[...]][,<contid>[...]]"
> > > >
> > > > The intent here is to clearly indicate and separate nesting from
> > > > parallel use of several containers by one netns.  If we do away with
> > > > that distinction, then we lose that inheritance accountability and
> > > > should really run the list through a "uniq" function to remove the
> > > > produced redundancies.  This clear inheritance is something Steve was
> > > > looking for since tracking down individual events/records to show that
> > > > inheritance was not aways feasible due to rolled logs or search effort.
> > >
> > > Perhaps my example wasn't clear.  I'm not opposed to the little
> > > carat/hat character indicating a container's parent, I just think it
> > > would be good to also include a comma *in*addition* to the carat/hat.
> >
> > Ah, ok.  Well, I'd offer that it would be slightly shorter, slightly
> > less cluttered and having already written the parser in userspace, I
> > think the parser would be slightly simpler.
> >
> > I must admit, I was a bit puzzled by your snippet of code that was used
> > as a prefix to the next item rather than as a postfix to the given item.
> >
> > Can you say why you prefer the comma in addition?
> 
> Generally speaking, I believe that a single delimiter is both easier
> for the eyes to parse, and easier/safer for machines to parse as well.
> In this particular case I think of the comma as a delimiter and the
> carat as a modifier, reusing the carat as a delimiter seems like a bad
> idea to me.

I'm not crazy about this idea, but I'll have a look at how much work it
is to recode the userspace search tools.  It also adds extra characters
and noise into the string format that seems counterproductive.

> > > > > > diff --git a/kernel/audit.c b/kernel/audit.c
> > > > > > index ef8e07524c46..68be59d1a89b 100644
> > > > > > --- a/kernel/audit.c
> > > > > > +++ b/kernel/audit.c
> > > > >
> > > > > > @@ -492,6 +493,7 @@ void audit_switch_task_namespaces(struct nsproxy *ns, struct task_struct *p)
> > > > > >                 audit_netns_contid_add(new->net_ns, contid);
> > > > > >  }
> > > > > >
> > > > > > +void audit_log_contid(struct audit_buffer *ab, u64 contid);
> > > > >
> > > > > If we need a forward declaration, might as well just move it up near
> > > > > the top of the file with the rest of the declarations.
> > > >
> > > > Ok.
> > > >
> > > > > > +void audit_log_contid(struct audit_buffer *ab, u64 contid)
> > > > > > +{
> > > > > > +       struct audit_contobj *cont = NULL, *prcont = NULL;
> > > > > > +       int h;
> > > > >
> > > > > It seems safer to pass the audit container ID object and not the u64.
> > > >
> > > > It would also be faster, but in some places it isn't available such as
> > > > for ptrace and signal targets.  This also links back to the drop record
> > > > refcounts to hold onto the contobj until process exit, or signal
> > > > delivery.
> > > >
> > > > What we could do is to supply two potential parameters, a contobj and/or
> > > > a contid, and have it use the contobj if it is valid, otherwise, use the
> > > > contid, as is done for names and paths supplied to audit_log_name().
> > >
> > > Let's not do multiple parameters, that begs for misuse, let's take the
> > > wrapper function route:
> > >
> > >  func a(int id) {
> > >    // important stuff
> > >  }
> > >
> > >  func ao(struct obj) {
> > >    a(obj.id);
> > >  }
> > >
> > > ... and we can add a comment that you *really* should be using the
> > > variant that passes an object.
> >
> > I was already doing that where it available, and dereferencing the id
> > for the call.  But I see an advantage to having both parameters supplied
> > to the function, since it saves us the trouble of dereferencing it,
> > searching for the id in the hash list and re-locating the object if the
> > object is already available.
> 
> I strongly prefer we not do multiple parameters for the same "thing";

So do I, ideally.  However...

> I would much rather do the wrapper approach as described above.  I
> would also like to see us use the audit container ID object as much as
> possible, using a bare integer should be a last resort.

It is not clear to me that you understood what I wrote above.  I can't
use the object pointer where preferable because there are a few cases
where only the ID is available.  If only the ID is available, I would
have to make a best effort to look up the object pointer and am not
guaranteed to find it (invalid, stale, signal info...).  If I am forced
to use only one, it becomes the ID that is used, and I no longer have
the benefit of already having the object pointer for certainty and
saving work.  For all cases where I have the object pointer, which is
most cases, and most frequently used cases, I will have to dereference
the object pointer to an ID, then go through the work again to re-locate
the object pointer.  This is less certain, and more work.  Reluctantly,
the only practical solution I see here is to supply both, favouring the
object pointer if it is valid, then falling back on the ID from the next
parameter.

> paul moore

- RGB

--
Richard Guy Briggs <rgb@xxxxxxxxxx>
Sr. S/W Engineer, Kernel Security, Base Operating Systems
Remote, Ottawa, Red Hat Canada
IRC: rgb, SunRaycer
Voice: +1.647.777.2635, Internal: (81) 32635




[Index of Archives]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux