Re: [dpdk-dev] [PATCH v3] mbuf: cleanup rte_pktmbuf_lastseg(), fix atomic usage

["Hanoch Haim (hhaim)" <hhaim@cisco.com>]

Hi Oliver, 

It's hard for me to follow this thread. 

1)  It is not about clear/not-clear, it is error prone to *replicate* code that has the same logic.

"I'm not convinced that:

    __rte_pktmbuf_reset_nb_segs(m);

is clearer than:

   m->next = NULL;
   m->nb_segs = 1;

Anyway, I agree this should not be part of this patch. We should only keep the fix.
"
2) This definitely does not look good. 
All the point in my patch is to move the ref-cnt operations to set of API that already taking care of RTE_MBUF_REFCNT_ATOMIC

+               /* We don't use rte_mbuf_refcnt_update() because we already
+                * tested that refcnt != 1.
+                */
+#ifdef RTE_MBUF_REFCNT_ATOMIC
+               ret = rte_atomic16_add_return(&m->refcnt_atomic, -1); 
+#else
+               ret = --m->refcnt;
+#endif
+               if (ret != 0)
+                       return NULL;


Hanoh


-----Original Message-----
From: Olivier MATZ [mailto:olivier.matz@6wind.com] 
Sent: Wednesday, November 15, 2017 7:31 PM
To: Hanoch Haim (hhaim)
Cc: Ilya Matveychikov; dev@dpdk.org; Konstantin Ananyev
Subject: Re: [dpdk-dev] [PATCH v3] mbuf: cleanup rte_pktmbuf_lastseg(), fix atomic usage

Hi,

On Wed, Nov 15, 2017 at 12:46:15PM +0000, Hanoch Haim (hhaim) wrote:
> +Oliver,
> Ilia, I assume there is a reason why Oliver did that, I just consolidate the code. 
> He didn't want to *write* the same value to mbuf field. 
> In the common case that pointer was already cleared by the driver, it is better to just read it. From cache synchronization perspective it will run faster.
> 
> Thanks,
> Hanoh
> 
> 
> -----Original Message-----
> From: Ilya Matveychikov [mailto:matvejchikov@gmail.com]
> Sent: Wednesday, November 15, 2017 1:14 PM
> To: Hanoch Haim (hhaim)
> Cc: dev@dpdk.org
> Subject: Re: [dpdk-dev] [PATCH v3] mbuf: cleanup 
> rte_pktmbuf_lastseg(), fix atomic usage
> 
> 
> > On Nov 15, 2017, at 1:14 PM, Hanoh Haim <hhaim@cisco.com> wrote:
> > 

I think the patch should be renamed in something like:

  mbuf: fix mbuf free performance with non atomic refcnt

A description of the problem in the commit log would also be welcome.

It looks it is a regression introduced by commit 8f094a9ac5d7.
In that case, we should also have:

Fixes: 8f094a9ac5d7 ("mbuf: set mbuf fields while in pool")


> > Signed-off-by: Hanoh Haim <hhaim@cisco.com>
> > ---
> > lib/librte_mbuf/rte_mbuf.h | 27 +++++++++++++--------------
> > 1 file changed, 13 insertions(+), 14 deletions(-)
> > 
> > diff --git a/lib/librte_mbuf/rte_mbuf.h b/lib/librte_mbuf/rte_mbuf.h 
> > index 7e326bb..ab110f8 100644
> > --- a/lib/librte_mbuf/rte_mbuf.h
> > +++ b/lib/librte_mbuf/rte_mbuf.h
> > @@ -1159,6 +1159,15 @@ static inline void rte_pktmbuf_reset(struct rte_mbuf *m)
> > 	__rte_mbuf_sanity_check(m, 1);
> > }
> > 
> > +
> > +static __rte_always_inline void 
> > +rte_pktmbuf_reset_before_free(struct
> > +rte_mbuf *m) {
> > +	if (m->next != NULL) {
> > +		m->next = NULL;
> > +		m->nb_segs = 1;
> > +	}
> > +}
> > +
> 
> Probably it will be more clean to add something 
> __te_pktmbuf_reset_nb_segs() without check for (m->next != NULL) and 
> use it everywhere in mbuf’s the code, not only in
> rte_pktmbuf_prefree_seg() function. And I think it will be better to have separate patch for that.

I'm not convinced that:

    __rte_pktmbuf_reset_nb_segs(m);

is clearer than:

   m->next = NULL;
   m->nb_segs = 1;

Anyway, I agree this should not be part of this patch. We should only keep the fix.


> 
> > /**
> >  * Allocate a new mbuf from a mempool.
> >  *
> > @@ -1323,8 +1332,7 @@ static inline void rte_pktmbuf_detach(struct rte_mbuf *m)
> > 	m->ol_flags = 0;
> > 
> > 	if (rte_mbuf_refcnt_update(md, -1) == 0) {
> > -		md->next = NULL;
> > -		md->nb_segs = 1;
> 
> Using rte_pktmbuf_reset_before_free() here adds unnecessary check for m->next in that path.

Yes, agree with Ilya.


> 
> > +		rte_pktmbuf_reset_before_free(md);
> > 		rte_mbuf_refcnt_set(md, 1);
> > 		rte_mbuf_raw_free(md);
> > 	}
> > @@ -1354,25 +1362,16 @@ rte_pktmbuf_prefree_seg(struct rte_mbuf *m)
> > 		if (RTE_MBUF_INDIRECT(m))
> > 			rte_pktmbuf_detach(m);
> > 
> > -		if (m->next != NULL) {
> > -			m->next = NULL;
> > -			m->nb_segs = 1;
> > -		}
> > -
> > +		rte_pktmbuf_reset_before_free(m);
> > 		return m;
> > 
> > -       } else if (rte_atomic16_add_return(&m->refcnt_atomic, -1) == 0) {
> > -
> > +	} else if (rte_mbuf_refcnt_update(m, -1) == 0) {

I agree with Konstantin's comment done in another thread [1]:

  '''
  That would cause extra read; cmp (and possible slowdown) for atomic refcnt.
  If that really need to be fixed - probably we need to introduce a new function
  that would do update without trying to read refctn first - rte_mbuf_refcnt_update_blind() or so.
  '''

However I'm not sure a new function is really needed: the name is not ideal, and it would only be used once. What about the patch below?

==============================
--- a/lib/librte_mbuf/rte_mbuf.h
+++ b/lib/librte_mbuf/rte_mbuf.h
@@ -1361,8 +1361,18 @@ rte_pktmbuf_prefree_seg(struct rte_mbuf *m)
 
                return m;
 
-       } else if (rte_atomic16_add_return(&m->refcnt_atomic, -1) == 0) {
+       } else {
 
+               /* We don't use rte_mbuf_refcnt_update() because we already
+                * tested that refcnt != 1.
+                */
+#ifdef RTE_MBUF_REFCNT_ATOMIC
+               ret = rte_atomic16_add_return(&m->refcnt_atomic, -1); 
+#else
+               ret = --m->refcnt;
+#endif
+               if (ret != 0)
+                       return NULL;
 
                if (RTE_MBUF_INDIRECT(m))
                        rte_pktmbuf_detach(m); @@ -1375,7 +1385,6 @@ rte_pktmbuf_prefree_seg(struct rte_mbuf *m)
 
                return m;
        }
-       return NULL;
 }
 
 /* deprecated, replaced by rte_pktmbuf_prefree_seg() */ ==============================

[1] http://dpdk.org/dev/patchwork/patch/31378/


Regards,
Olivier