Re: [PATCH v2] mbuf: use refcnt = 0 when debugging

[Chas Williams <3chas3@gmail.com>]

On Wed, 2017-09-06 at 14:53 +0000, Ananyev, Konstantin wrote:
> 
> > -----Original Message-----
> > From: Chas Williams [mailto:3chas3@gmail.com]
> > Sent: Wednesday, September 6, 2017 2:56 PM
> > To: Ananyev, Konstantin <konstantin.ananyev@intel.com>; Nicolau, Radu <radu.nicolau@intel.com>; dev@dpdk.org
> > Cc: olivier.matz@6wind.com; cw817q@att.com
> > Subject: Re: [dpdk-dev] [PATCH v2] mbuf: use refcnt = 0 when debugging
> > 
> > On Wed, 2017-09-06 at 11:58 +0000, Ananyev, Konstantin wrote:
> > >
> > > > -----Original Message-----
> > > > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Chas Williams
> > > > Sent: Wednesday, September 6, 2017 11:46 AM
> > > > To: Nicolau, Radu <radu.nicolau@intel.com>; dev@dpdk.org
> > > > Cc: olivier.matz@6wind.com; cw817q@att.com
> > > > Subject: Re: [dpdk-dev] [PATCH v2] mbuf: use refcnt = 0 when debugging
> > > >
> > > > [Note: My former email address is going away eventually.  I am moving the
> > > > conversation to my other email address which is a bit more permanent.]
> > > >
> > > > On Mon, 2017-09-04 at 15:27 +0100, Radu Nicolau wrote:
> > > > >
> > > > > On 8/7/2017 5:11 PM, Charles (Chas) Williams wrote:
> > > > > > After commit 8f094a9ac5d7 ("mbuf: set mbuf fields while in pool") is it
> > > > > > much harder to detect a "double free".  If the developer makes a copy
> > > > > > of an mbuf pointer and frees it twice, this condition is never detected
> > > > > > and the mbuf gets returned to the pool twice.
> > > > > >
> > > > > > Since this requires extra work to track, make this behavior conditional
> > > > > > on CONFIG_RTE_LIBRTE_MBUF_DEBUG.
> > > > > >
> > > > > > Signed-off-by: Chas Williams <ciwillia@brocade.com>
> > > > > > ---
> > > > > >
> > > > > > @@ -1304,10 +1329,13 @@ rte_pktmbuf_prefree_seg(struct rte_mbuf *m)
> > > > > >   			m->next = NULL;
> > > > > >   			m->nb_segs = 1;
> > > > > >   		}
> > > > > > +#ifdef RTE_LIBRTE_MBUF_DEBUG
> > > > > > +		rte_mbuf_refcnt_set(m, RTE_MBUF_UNUSED_CNT);
> > > > > > +#endif
> > > > > >
> > > > > >   		return m;
> > > > > >
> > > > > > -       } else if (rte_atomic16_add_return(&m->refcnt_atomic, -1) == 0) {
> > > > > > +	} else if (rte_mbuf_refcnt_update(m, -1) == 0) {
> > > > > Why replace the use of atomic operation?
> > > >
> > > > It doesn't.  rte_mbuf_refcnt_update() is also atomic(ish) but it slightly more
> > > > optimal.  This whole section is a little hazy actually.  It looks like
> > > > rte_pktmbuf_prefree_seg() unwraps rte_mbuf_refcnt_update() so they can avoid
> > > > setting the refcnt when the refcnt is already the 'correct' value.
> > >
> > > You don't need to use refcnt_update() here - if you take that path it already means
> > > that m->refcnt_atomic != 1.
> > > In fact, I think using refcnt_update () here might be a bit slower - as it means extra read.
> > > Konstantin
> > 
> > Yes, that is somewhat the point.  If a mbuf can have a refcnt of 0,
> > then we want to go into rte_mbuf_refcnt_update() which detects 0 -> -1.
> 
> Woulnd't __rte_mbuf_sanity_check(m, 0) at the start of prefree_seg()
> already catch it?
> Konstantin

Yes!  I didn't notice that so I will just drop change and issue a v3 today
sometime.

Thanks!
 
> > I could explicitly check this in prefree_seg but I was just restored the
> > previous call into refcnt_update.  I could explicitly check for refcnt =
> > 0 in prefree_seg() but since we do have a routine for this...
> > 
> > > > > >
> > > > > >
> > > > > >   		if (RTE_MBUF_INDIRECT(m))
> > > > > > @@ -1317,7 +1345,7 @@ rte_pktmbuf_prefree_seg(struct rte_mbuf *m)
> > > > > >   			m->next = NULL;
> > > > > >   			m->nb_segs = 1;
> > > > > >   		}
> > > > > > -		rte_mbuf_refcnt_set(m, 1);
> > > > > > +		rte_mbuf_refcnt_set(m, RTE_MBUF_UNUSED_CNT);
> > > > > >
> > > > > >   		return m;
> > > > > >   	}
> > > > > Reviewed-by:  Radu Nicolau <radu.nicolau@intel.com>
> > > >
> > > > Thanks for the review.