[PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[jhascoet]

In case of ring full state, we retry the enqueue
operation in order to avoid mbuf loss.

Fixes: af75078fece ("first public release")

Signed-off-by: Julien Hascoet <jhascoet@kalrayinc.com>
---
 app/test/test_mbuf.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/app/test/test_mbuf.c b/app/test/test_mbuf.c
index efac01806b..be114e3302 100644
--- a/app/test/test_mbuf.c
+++ b/app/test/test_mbuf.c
@@ -1033,12 +1033,17 @@ test_refcnt_iter(unsigned int lcore, unsigned int iter,
 		tref += ref;
 		if ((ref & 1) != 0) {
 			rte_pktmbuf_refcnt_update(m, ref);
-			while (ref-- != 0)
-				rte_ring_enqueue(refcnt_mbuf_ring, m);
+			while (ref-- != 0) {
+				/* retry in case of failure */
+				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0)
+					;
+			}
 		} else {
 			while (ref-- != 0) {
 				rte_pktmbuf_refcnt_update(m, 1);
-				rte_ring_enqueue(refcnt_mbuf_ring, m);
+				/* retry in case of failure */
+				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0)
+					;
 			}
 		}
 		rte_pktmbuf_free(m);
-- 
2.34.1

Re: [PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[David Marchand]

On Wed, Aug 9, 2023 at 7:39 AM jhascoet <ju.hascoet@gmail.com> wrote:
>
> In case of ring full state, we retry the enqueue
> operation in order to avoid mbuf loss.
>
> Fixes: af75078fece ("first public release")

Not sure we need to backport, but in doubt, I would mark it
Cc: stable@dpdk.org


>
> Signed-off-by: Julien Hascoet <jhascoet@kalrayinc.com>

Olivier, the patch lgtm, can you have a look?
Thanks.


-- 
David Marchand

[PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[jhascoet]

From: Julien Hascoet <ju.hascoet@gmail.com>

In case of ring full state, we retry the enqueue
operation in order to avoid mbuf loss.

Fixes: af75078fece ("first public release")

Signed-off-by: Julien Hascoet <ju.hascoet@gmail.com>
---
 app/test/test_mbuf.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/app/test/test_mbuf.c b/app/test/test_mbuf.c
index efac01806b..be114e3302 100644
--- a/app/test/test_mbuf.c
+++ b/app/test/test_mbuf.c
@@ -1033,12 +1033,17 @@ test_refcnt_iter(unsigned int lcore, unsigned int iter,
 		tref += ref;
 		if ((ref & 1) != 0) {
 			rte_pktmbuf_refcnt_update(m, ref);
-			while (ref-- != 0)
-				rte_ring_enqueue(refcnt_mbuf_ring, m);
+			while (ref-- != 0) {
+				/* retry in case of failure */
+				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0)
+					;
+			}
 		} else {
 			while (ref-- != 0) {
 				rte_pktmbuf_refcnt_update(m, 1);
-				rte_ring_enqueue(refcnt_mbuf_ring, m);
+				/* retry in case of failure */
+				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0)
+					;
 			}
 		}
 		rte_pktmbuf_free(m);
-- 
2.34.1

Re: [PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[Stephen Hemminger]

On Thu, 10 Aug 2023 08:00:30 +0200
jhascoet <ju.hascoet@gmail.com> wrote:

> diff --git a/app/test/test_mbuf.c b/app/test/test_mbuf.c
> index efac01806b..be114e3302 100644
> --- a/app/test/test_mbuf.c
> +++ b/app/test/test_mbuf.c
> @@ -1033,12 +1033,17 @@ test_refcnt_iter(unsigned int lcore, unsigned int iter,
>  		tref += ref;
>  		if ((ref & 1) != 0) {
>  			rte_pktmbuf_refcnt_update(m, ref);
> -			while (ref-- != 0)
> -				rte_ring_enqueue(refcnt_mbuf_ring, m);
> +			while (ref-- != 0) {
> +				/* retry in case of failure */
> +				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0)
> +					;

Since other side needs to consume these and might be on same lcore,
it might be good place to add rte_pause or sched_yield here?

RE: [PUB] Re: [PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[Julien Hascoet]

[-- Attachment #1: Type: text/plain, Size: 1283 bytes --]

Yes, just did it.
Thanks !
________________________________
De : Stephen Hemminger <stephen@networkplumber.org>
Envoyé : jeudi 10 août 2023 17:33
À : jhascoet <ju.hascoet@gmail.com>
Cc : david.marchand@redhat.com <david.marchand@redhat.com>; dev@dpdk.org <dev@dpdk.org>
Objet : [PUB] Re: [PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

On Thu, 10 Aug 2023 08:00:30 +0200
jhascoet <ju.hascoet@gmail.com> wrote:

> diff --git a/app/test/test_mbuf.c b/app/test/test_mbuf.c
> index efac01806b..be114e3302 100644
> --- a/app/test/test_mbuf.c
> +++ b/app/test/test_mbuf.c
> @@ -1033,12 +1033,17 @@ test_refcnt_iter(unsigned int lcore, unsigned int iter,
>                tref += ref;
>                if ((ref & 1) != 0) {
>                        rte_pktmbuf_refcnt_update(m, ref);
> -                     while (ref-- != 0)
> -                             rte_ring_enqueue(refcnt_mbuf_ring, m);
> +                     while (ref-- != 0) {
> +                             /* retry in case of failure */
> +                             while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0)
> +                                     ;

Since other side needs to consume these and might be on same lcore,
it might be good place to add rte_pause or sched_yield here?







[-- Attachment #2: Type: text/html, Size: 3428 bytes --]

[PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[jhascoet]

From: Julien Hascoet <ju.hascoet@gmail.com>

In case of ring full state, we retry the enqueue
operation in order to avoid mbuf loss.

Fixes: af75078fece ("first public release")

Signed-off-by: Julien Hascoet <ju.hascoet@gmail.com>
---
 app/test/test_mbuf.c | 15 ++++++++++++---
 1 file changed, 12 insertions(+), 3 deletions(-)

diff --git a/app/test/test_mbuf.c b/app/test/test_mbuf.c
index efac01806b..ad18bf6378 100644
--- a/app/test/test_mbuf.c
+++ b/app/test/test_mbuf.c
@@ -1033,12 +1033,21 @@ test_refcnt_iter(unsigned int lcore, unsigned int iter,
 		tref += ref;
 		if ((ref & 1) != 0) {
 			rte_pktmbuf_refcnt_update(m, ref);
-			while (ref-- != 0)
-				rte_ring_enqueue(refcnt_mbuf_ring, m);
+			while (ref-- != 0) {
+				/* retry in case of failure */
+				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0) {
+					/* let others consume */
+					rte_pause();
+				}
+			}
 		} else {
 			while (ref-- != 0) {
 				rte_pktmbuf_refcnt_update(m, 1);
-				rte_ring_enqueue(refcnt_mbuf_ring, m);
+				/* retry in case of failure */
+				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0) {
+					/* let others consume */
+					rte_pause();
+				}
 			}
 		}
 		rte_pktmbuf_free(m);
-- 
2.34.1

Re: [PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[Hascoet Julien]

[-- Attachment #1: Type: text/plain, Size: 2020 bytes --]

Any news on that ? Thanks

Julien

On Thu, Aug 10, 2023 at 6:09 PM jhascoet <ju.hascoet@gmail.com> wrote:

> From: Julien Hascoet <ju.hascoet@gmail.com>
>
> In case of ring full state, we retry the enqueue
> operation in order to avoid mbuf loss.
>
> Fixes: af75078fece ("first public release")
>
> Signed-off-by: Julien Hascoet <ju.hascoet@gmail.com>
> ---
>  app/test/test_mbuf.c | 15 ++++++++++++---
>  1 file changed, 12 insertions(+), 3 deletions(-)
>
> diff --git a/app/test/test_mbuf.c b/app/test/test_mbuf.c
> index efac01806b..ad18bf6378 100644
> --- a/app/test/test_mbuf.c
> +++ b/app/test/test_mbuf.c
> @@ -1033,12 +1033,21 @@ test_refcnt_iter(unsigned int lcore, unsigned int
> iter,
>                 tref += ref;
>                 if ((ref & 1) != 0) {
>                         rte_pktmbuf_refcnt_update(m, ref);
> -                       while (ref-- != 0)
> -                               rte_ring_enqueue(refcnt_mbuf_ring, m);
> +                       while (ref-- != 0) {
> +                               /* retry in case of failure */
> +                               while (rte_ring_enqueue(refcnt_mbuf_ring,
> m) != 0) {
> +                                       /* let others consume */
> +                                       rte_pause();
> +                               }
> +                       }
>                 } else {
>                         while (ref-- != 0) {
>                                 rte_pktmbuf_refcnt_update(m, 1);
> -                               rte_ring_enqueue(refcnt_mbuf_ring, m);
> +                               /* retry in case of failure */
> +                               while (rte_ring_enqueue(refcnt_mbuf_ring,
> m) != 0) {
> +                                       /* let others consume */
> +                                       rte_pause();
> +                               }
>                         }
>                 }
>                 rte_pktmbuf_free(m);
> --
> 2.34.1
>
>

[-- Attachment #2: Type: text/html, Size: 2895 bytes --]

Re: [PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[David Marchand]

On Fri, Aug 18, 2023 at 7:17 AM Hascoet Julien <ju.hascoet@gmail.com> wrote:
>
> Any news on that ? Thanks

Don't forget to copy maintainers.
Let's leave some time to Olivier to have a look.

Thanks.


-- 
David Marchand

[PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[jhascoet]

From: Julien Hascoet <ju.hascoet@gmail.com>

In case of ring full state, we retry the enqueue
operation in order to avoid mbuf loss.

Fixes: af75078fece ("first public release")

Signed-off-by: Julien Hascoet <ju.hascoet@gmail.com>
---
 app/test/test_mbuf.c | 15 ++++++++++++---
 1 file changed, 12 insertions(+), 3 deletions(-)

diff --git a/app/test/test_mbuf.c b/app/test/test_mbuf.c
index efac01806b..ad18bf6378 100644
--- a/app/test/test_mbuf.c
+++ b/app/test/test_mbuf.c
@@ -1033,12 +1033,21 @@ test_refcnt_iter(unsigned int lcore, unsigned int iter,
 		tref += ref;
 		if ((ref & 1) != 0) {
 			rte_pktmbuf_refcnt_update(m, ref);
-			while (ref-- != 0)
-				rte_ring_enqueue(refcnt_mbuf_ring, m);
+			while (ref-- != 0) {
+				/* retry in case of failure */
+				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0) {
+					/* let others consume */
+					rte_pause();
+				}
+			}
 		} else {
 			while (ref-- != 0) {
 				rte_pktmbuf_refcnt_update(m, 1);
-				rte_ring_enqueue(refcnt_mbuf_ring, m);
+				/* retry in case of failure */
+				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0) {
+					/* let others consume */
+					rte_pause();
+				}
 			}
 		}
 		rte_pktmbuf_free(m);
-- 
2.34.1

Re: [PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[Olivier Matz]

Hello Julien,

On Tue, Aug 22, 2023 at 08:34:53AM +0200, jhascoet wrote:
> From: Julien Hascoet <ju.hascoet@gmail.com>
> 
> In case of ring full state, we retry the enqueue
> operation in order to avoid mbuf loss.
> 
> Fixes: af75078fece ("first public release")
> 
> Signed-off-by: Julien Hascoet <ju.hascoet@gmail.com>
> ---
>  app/test/test_mbuf.c | 15 ++++++++++++---
>  1 file changed, 12 insertions(+), 3 deletions(-)
> 
> diff --git a/app/test/test_mbuf.c b/app/test/test_mbuf.c
> index efac01806b..ad18bf6378 100644
> --- a/app/test/test_mbuf.c
> +++ b/app/test/test_mbuf.c
> @@ -1033,12 +1033,21 @@ test_refcnt_iter(unsigned int lcore, unsigned int iter,
>  		tref += ref;
>  		if ((ref & 1) != 0) {
>  			rte_pktmbuf_refcnt_update(m, ref);
> -			while (ref-- != 0)
> -				rte_ring_enqueue(refcnt_mbuf_ring, m);
> +			while (ref-- != 0) {
> +				/* retry in case of failure */
> +				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0) {
> +					/* let others consume */
> +					rte_pause();
> +				}
> +			}
>  		} else {
>  			while (ref-- != 0) {
>  				rte_pktmbuf_refcnt_update(m, 1);
> -				rte_ring_enqueue(refcnt_mbuf_ring, m);
> +				/* retry in case of failure */
> +				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0) {
> +					/* let others consume */
> +					rte_pause();
> +				}
>  			}
>  		}
>  		rte_pktmbuf_free(m);
> -- 
> 2.34.1
> 

Can you give some more details about how to reproduce the issue?

From what I see, the code does the following:

main core:
  create a ring with at least (REFCNT_MBUF_NUM * REFCNT_MAX_REF) entries
  create an mbuf pool with REFCNT_MBUF_NUM entries
  start worker cores
  do REFCNT_MAX_ITER times:
    for each mbuf of the pool (REFCNT_MBUF_NUM entries):
      let r be a random number between 1 and REFCNT_MAX_REF
      increase mbuf references by r, and enqueue r times in the ring
    wait that the ring is empty (since worker cores are dequeuing mbufs)
  stop worker cores

worker cores:
  dequeue packets from the ring and free them until asked to stop


I may be mistaking but I don't see how the number of mbufs in ring could
exceed REFCNT_MBUF_NUM * REFCNT_MAX_REF.

Regards,
Olivier


Note: removing CC maintainers@dpdk.org

Re: [PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[Hascoet Julien]

[-- Attachment #1: Type: text/plain, Size: 3326 bytes --]

Hello Oliver,

thanks, your response helped a lot, I managed to find the root cause of the
instability which is on our side.
It was due to other internal developments.
I'll still add an error check on the enqueue ops to catch eventual problems
earlier, if that suits you.

Best regards,

Julien

On Tue, Aug 22, 2023 at 10:34 AM Olivier Matz <olivier.matz@6wind.com>
wrote:

> Hello Julien,
>
> On Tue, Aug 22, 2023 at 08:34:53AM +0200, jhascoet wrote:
> > From: Julien Hascoet <ju.hascoet@gmail.com>
> >
> > In case of ring full state, we retry the enqueue
> > operation in order to avoid mbuf loss.
> >
> > Fixes: af75078fece ("first public release")
> >
> > Signed-off-by: Julien Hascoet <ju.hascoet@gmail.com>
> > ---
> >  app/test/test_mbuf.c | 15 ++++++++++++---
> >  1 file changed, 12 insertions(+), 3 deletions(-)
> >
> > diff --git a/app/test/test_mbuf.c b/app/test/test_mbuf.c
> > index efac01806b..ad18bf6378 100644
> > --- a/app/test/test_mbuf.c
> > +++ b/app/test/test_mbuf.c
> > @@ -1033,12 +1033,21 @@ test_refcnt_iter(unsigned int lcore, unsigned
> int iter,
> >               tref += ref;
> >               if ((ref & 1) != 0) {
> >                       rte_pktmbuf_refcnt_update(m, ref);
> > -                     while (ref-- != 0)
> > -                             rte_ring_enqueue(refcnt_mbuf_ring, m);
> > +                     while (ref-- != 0) {
> > +                             /* retry in case of failure */
> > +                             while (rte_ring_enqueue(refcnt_mbuf_ring,
> m) != 0) {
> > +                                     /* let others consume */
> > +                                     rte_pause();
> > +                             }
> > +                     }
> >               } else {
> >                       while (ref-- != 0) {
> >                               rte_pktmbuf_refcnt_update(m, 1);
> > -                             rte_ring_enqueue(refcnt_mbuf_ring, m);
> > +                             /* retry in case of failure */
> > +                             while (rte_ring_enqueue(refcnt_mbuf_ring,
> m) != 0) {
> > +                                     /* let others consume */
> > +                                     rte_pause();
> > +                             }
> >                       }
> >               }
> >               rte_pktmbuf_free(m);
> > --
> > 2.34.1
> >
>
> Can you give some more details about how to reproduce the issue?
>
> From what I see, the code does the following:
>
> main core:
>   create a ring with at least (REFCNT_MBUF_NUM * REFCNT_MAX_REF) entries
>   create an mbuf pool with REFCNT_MBUF_NUM entries
>   start worker cores
>   do REFCNT_MAX_ITER times:
>     for each mbuf of the pool (REFCNT_MBUF_NUM entries):
>       let r be a random number between 1 and REFCNT_MAX_REF
>       increase mbuf references by r, and enqueue r times in the ring
>     wait that the ring is empty (since worker cores are dequeuing mbufs)
>   stop worker cores
>
> worker cores:
>   dequeue packets from the ring and free them until asked to stop
>
>
> I may be mistaking but I don't see how the number of mbufs in ring could
> exceed REFCNT_MBUF_NUM * REFCNT_MAX_REF.
>
> Regards,
> Olivier
>
>
> Note: removing CC maintainers@dpdk.org
>

[-- Attachment #2: Type: text/html, Size: 4534 bytes --]

[PATCH] app: enhance error check if enqueue fail in test_mbuf test_refcnt_iter

[jhascoet]

From: Julien Hascoet <ju.hascoet@gmail.com>

As the ring is big enough by construction, the ring
enqueue ops cannot fail; therefore, we check and panic
on it as soon as it is detected.

Signed-off-by: Julien Hascoet <ju.hascoet@gmail.com>
---
 app/test/test_mbuf.c | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)

diff --git a/app/test/test_mbuf.c b/app/test/test_mbuf.c
index d7393df7eb..c94dd3749e 100644
--- a/app/test/test_mbuf.c
+++ b/app/test/test_mbuf.c
@@ -1033,12 +1033,15 @@ test_refcnt_iter(unsigned int lcore, unsigned int iter,
 		tref += ref;
 		if ((ref & 1) != 0) {
 			rte_pktmbuf_refcnt_update(m, ref);
-			while (ref-- != 0)
-				rte_ring_enqueue(refcnt_mbuf_ring, m);
+			while (ref-- != 0) {
+				if (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0)
+					goto fail_enqueue;
+			}
 		} else {
 			while (ref-- != 0) {
 				rte_pktmbuf_refcnt_update(m, 1);
-				rte_ring_enqueue(refcnt_mbuf_ring, m);
+				if (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0)
+					goto fail_enqueue;
 			}
 		}
 		rte_pktmbuf_free(m);
@@ -1066,6 +1069,10 @@ test_refcnt_iter(unsigned int lcore, unsigned int iter,
 
 	rte_panic("(lcore=%u, iter=%u): after %us only "
 	          "%u of %u mbufs left free\n", lcore, iter, wn, i, n);
+
+fail_enqueue:
+	rte_panic("(lcore=%u, iter=%u): enqueue mbuf %u cannot "
+		"fail since ring is big enough\n", lcore, iter, i);
 }
 
 static int
-- 
2.34.1

Re: [PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[Thomas Monjalon]

Hello,

22/08/2023 08:34, jhascoet:
> From: Julien Hascoet <ju.hascoet@gmail.com>
> 
> In case of ring full state, we retry the enqueue
> operation in order to avoid mbuf loss.
> 
> Fixes: af75078fece ("first public release")
> 
> Signed-off-by: Julien Hascoet <ju.hascoet@gmail.com>

Please could you give more information about when this is happening?
I'm surprised such thing was not a problem so far.

[PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[jhascoet]

From: Julien Hascoet <ju.hascoet@gmail.com>

In case of ring full state, we retry the enqueue
operation in order to avoid mbuf loss.

Fixes: af75078fece ("first public release")

Signed-off-by: Julien Hascoet <ju.hascoet@gmail.com>
---
 app/test/test_mbuf.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/app/test/test_mbuf.c b/app/test/test_mbuf.c
index efac01806b..be114e3302 100644
--- a/app/test/test_mbuf.c
+++ b/app/test/test_mbuf.c
@@ -1033,12 +1033,17 @@ test_refcnt_iter(unsigned int lcore, unsigned int iter,
 		tref += ref;
 		if ((ref & 1) != 0) {
 			rte_pktmbuf_refcnt_update(m, ref);
-			while (ref-- != 0)
-				rte_ring_enqueue(refcnt_mbuf_ring, m);
+			while (ref-- != 0) {
+				/* retry in case of failure */
+				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0)
+					;
+			}
 		} else {
 			while (ref-- != 0) {
 				rte_pktmbuf_refcnt_update(m, 1);
-				rte_ring_enqueue(refcnt_mbuf_ring, m);
+				/* retry in case of failure */
+				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0)
+					;
 			}
 		}
 		rte_pktmbuf_free(m);
-- 
2.34.1

[PATCH] app: fix silent enqueue fail in test_mbuf test_refcnt_iter

[jhascoet]

From: Julien Hascoet <ju.hascoet@gmail.com>

In case of ring full state, we retry the enqueue
operation in order to avoid mbuf loss.

Fixes: af75078fece ("first public release")

Signed-off-by: Julien Hascoet <ju.hascoet@gmail.com>
---
 app/test/test_mbuf.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/app/test/test_mbuf.c b/app/test/test_mbuf.c
index efac01806b..be114e3302 100644
--- a/app/test/test_mbuf.c
+++ b/app/test/test_mbuf.c
@@ -1033,12 +1033,17 @@ test_refcnt_iter(unsigned int lcore, unsigned int iter,
 		tref += ref;
 		if ((ref & 1) != 0) {
 			rte_pktmbuf_refcnt_update(m, ref);
-			while (ref-- != 0)
-				rte_ring_enqueue(refcnt_mbuf_ring, m);
+			while (ref-- != 0) {
+				/* retry in case of failure */
+				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0)
+					;
+			}
 		} else {
 			while (ref-- != 0) {
 				rte_pktmbuf_refcnt_update(m, 1);
-				rte_ring_enqueue(refcnt_mbuf_ring, m);
+				/* retry in case of failure */
+				while (rte_ring_enqueue(refcnt_mbuf_ring, m) != 0)
+					;
 			}
 		}
 		rte_pktmbuf_free(m);
-- 
2.34.1