[dpdk-dev] Added Spinlock to l3fwd-vf example to prevent race conditioning
diff mbox

Message ID 1404818184-29388-1-git-send-email-danielx.t.mrzyglod@intel.com
State Accepted, archived
Headers show

Commit Message

Daniel Mrzyglod July 8, 2014, 11:16 a.m. UTC
Signed-off-by: Daniel Mrzyglod <danielx.t.mrzyglod@intel.com>
---
 examples/l3fwd-vf/main.c |    6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

Comments

Thomas Monjalon July 23, 2014, 8:33 a.m. UTC | #1
Hi Daniel,

Some explanations are missing here.

> Signed-off-by: Daniel Mrzyglod <danielx.t.mrzyglod@intel.com>
> 
> --- a/examples/l3fwd-vf/main.c
> +++ b/examples/l3fwd-vf/main.c
> @@ -54,6 +54,7 @@
>  #include <rte_per_lcore.h>
>  #include <rte_launch.h>
>  #include <rte_atomic.h>
> +#include <rte_spinlock.h>
>  #include <rte_cycles.h>
>  #include <rte_prefetch.h>
>  #include <rte_lcore.h>
> @@ -328,7 +329,7 @@ struct lcore_conf {
>  } __rte_cache_aligned;
>  
>  static struct lcore_conf lcore_conf[RTE_MAX_LCORE];
> -
> +static rte_spinlock_t spinlock_conf[RTE_MAX_ETHPORTS]={RTE_SPINLOCK_INITIALIZER};
>  /* Send burst of packets on an output interface */
>  static inline int
>  send_burst(struct lcore_conf *qconf, uint16_t n, uint8_t port)
> @@ -340,7 +341,10 @@ send_burst(struct lcore_conf *qconf, uint16_t n, uint8_t port)
>  	queueid = qconf->tx_queue_id;
>  	m_table = (struct rte_mbuf **)qconf->tx_mbufs[port].m_table;
>  
> +	rte_spinlock_lock(&spinlock_conf[port]) ;
>  	ret = rte_eth_tx_burst(port, queueid, m_table, n);
> +	rte_spinlock_unlock(&spinlock_conf[port]);
> +	
>  	if (unlikely(ret < n)) {
>  		do {
>  			rte_pktmbuf_free(m_table[ret]);
>
Thomas Monjalon Nov. 11, 2014, 10:56 p.m. UTC | #2
Hi Daniel,

This old patch is probably good but I'd like you explain it please.
Reviewers are also welcome.

Thanks
Huawei Xie Nov. 11, 2014, 11:18 p.m. UTC | #3
> -----Original Message-----
> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Thomas Monjalon
> Sent: Tuesday, November 11, 2014 3:57 PM
> To: Mrzyglod, DanielX T
> Cc: dev@dpdk.org
> Subject: Re: [dpdk-dev] [PATCH] Added Spinlock to l3fwd-vf example to prevent
> race conditioning
> 
> Hi Daniel,
> 
> This old patch is probably good but I'd like you explain it please.
> Reviewers are also welcome.
> 
> Thanks
> --
> Thomas
> 
> 2014-07-23 10:33, Thomas Monjalon:
> > Hi Daniel,
> >
> > Some explanations are missing here.
> >
> > > Signed-off-by: Daniel Mrzyglod <danielx.t.mrzyglod@intel.com>
> > >
> > > --- a/examples/l3fwd-vf/main.c
> > > +++ b/examples/l3fwd-vf/main.c
> > > @@ -54,6 +54,7 @@
> > >  #include <rte_per_lcore.h>
> > >  #include <rte_launch.h>
> > >  #include <rte_atomic.h>
> > > +#include <rte_spinlock.h>
> > >  #include <rte_cycles.h>
> > >  #include <rte_prefetch.h>
> > >  #include <rte_lcore.h>
> > > @@ -328,7 +329,7 @@ struct lcore_conf {
> > >  } __rte_cache_aligned;
> > >
> > >  static struct lcore_conf lcore_conf[RTE_MAX_LCORE];
> > > -
> > > +static rte_spinlock_t
> spinlock_conf[RTE_MAX_ETHPORTS]={RTE_SPINLOCK_INITIALIZER};
> > >  /* Send burst of packets on an output interface */
> > >  static inline int
> > >  send_burst(struct lcore_conf *qconf, uint16_t n, uint8_t port)
> > > @@ -340,7 +341,10 @@ send_burst(struct lcore_conf *qconf, uint16_t n,
> uint8_t port)
> > >  	queueid = qconf->tx_queue_id;
> > >  	m_table = (struct rte_mbuf **)qconf->tx_mbufs[port].m_table;
> > >
> > > +	rte_spinlock_lock(&spinlock_conf[port]) ;
> > >  	ret = rte_eth_tx_burst(port, queueid, m_table, n);
> > > +	rte_spinlock_unlock(&spinlock_conf[port]);

It might not be good choice for here, but how about we also provide spin_trylock as alternative API?

> > > +
> > >  	if (unlikely(ret < n)) {
> > >  		do {
> > >  			rte_pktmbuf_free(m_table[ret]);
> > >
Neil Horman Dec. 8, 2014, 2:45 p.m. UTC | #4
On Tue, Jul 08, 2014 at 12:16:24PM +0100, Daniel Mrzyglod wrote:
> Signed-off-by: Daniel Mrzyglod <danielx.t.mrzyglod@intel.com>
> 
> ---
> examples/l3fwd-vf/main.c |    6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/examples/l3fwd-vf/main.c b/examples/l3fwd-vf/main.c
> index 2ca5c21..57852d0 100644
> --- a/examples/l3fwd-vf/main.c
> +++ b/examples/l3fwd-vf/main.c
> @@ -54,6 +54,7 @@
>  #include <rte_per_lcore.h>
>  #include <rte_launch.h>
>  #include <rte_atomic.h>
> +#include <rte_spinlock.h>
>  #include <rte_cycles.h>
>  #include <rte_prefetch.h>
>  #include <rte_lcore.h>
> @@ -328,7 +329,7 @@ struct lcore_conf {
>  } __rte_cache_aligned;
>  
>  static struct lcore_conf lcore_conf[RTE_MAX_LCORE];
> -
> +static rte_spinlock_t spinlock_conf[RTE_MAX_ETHPORTS]={RTE_SPINLOCK_INITIALIZER};
>  /* Send burst of packets on an output interface */
>  static inline int
>  send_burst(struct lcore_conf *qconf, uint16_t n, uint8_t port)
> @@ -340,7 +341,10 @@ send_burst(struct lcore_conf *qconf, uint16_t n, uint8_t port)
>  	queueid = qconf->tx_queue_id;
>  	m_table = (struct rte_mbuf **)qconf->tx_mbufs[port].m_table;
>  
> +	rte_spinlock_lock(&spinlock_conf[port]) ;
>  	ret = rte_eth_tx_burst(port, queueid, m_table, n);
> +	rte_spinlock_unlock(&spinlock_conf[port]);
> +	
>  	if (unlikely(ret < n)) {
>  		do {
>  			rte_pktmbuf_free(m_table[ret]);

Acked-by: Neil Horman <nhorman@tuxdriver.com>

Though, that said, doesn't it seem to anyone else like serialization of enqueue
to a port should be the responsibility of the library, not the application?

Neil
Wodkowski, PawelX Dec. 10, 2014, 8:18 a.m. UTC | #5
> Though, that said, doesn't it seem to anyone else like serialization of enqueue
> to a port should be the responsibility of the library, not the application?
> 
> Neil

From my knowledge it is an application  responsibility to serialize access to
queue on particular port.

Pawel
Thomas Monjalon Dec. 10, 2014, 10:53 a.m. UTC | #6
2014-12-08 09:45, Neil Horman:
> On Tue, Jul 08, 2014 at 12:16:24PM +0100, Daniel Mrzyglod wrote:
> > Signed-off-by: Daniel Mrzyglod <danielx.t.mrzyglod@intel.com>
> Acked-by: Neil Horman <nhorman@tuxdriver.com>

Someone to provide an explanation for commit log?

Thanks
Neil Horman Dec. 10, 2014, 2:47 p.m. UTC | #7
On Wed, Dec 10, 2014 at 08:18:36AM +0000, Wodkowski, PawelX wrote:
> > Though, that said, doesn't it seem to anyone else like serialization of enqueue
> > to a port should be the responsibility of the library, not the application?
> > 
> > Neil
> 
> From my knowledge it is an application  responsibility to serialize access to
> queue on particular port.
> 
I understand thats the way it currently is, I'm advocating for the fact that it
should not be.
Neil

> Pawel
>
Bruce Richardson Dec. 10, 2014, 2:54 p.m. UTC | #8
On Wed, Dec 10, 2014 at 09:47:45AM -0500, Neil Horman wrote:
> On Wed, Dec 10, 2014 at 08:18:36AM +0000, Wodkowski, PawelX wrote:
> > > Though, that said, doesn't it seem to anyone else like serialization of enqueue
> > > to a port should be the responsibility of the library, not the application?
> > > 
> > > Neil
> > 
> > From my knowledge it is an application  responsibility to serialize access to
> > queue on particular port.
> > 
> I understand thats the way it currently is, I'm advocating for the fact that it
> should not be.
> Neil
>
It could be done, but I think we'd need to add a new API (or new parameter to
existing API) to do so, as the cost of adding the locks would be severe, even in
the uncontented case. 
This is why it hasn't been done up till now, obviously enough. In general, where
we don't provide performant multi-thread safe APIs, we generally don't try and
provide versions with locks, we just document the limitation and then leave it 
up to the app to determine how best to handle things.

/Bruce
Daniel Mrzyglod Dec. 10, 2014, 3:53 p.m. UTC | #9
> -----Original Message-----
> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Bruce Richardson
> Sent: Wednesday, December 10, 2014 3:55 PM
> To: Neil Horman
> Cc: dev@dpdk.org
> Subject: Re: [dpdk-dev] Added Spinlock to l3fwd-vf example to prevent race
> conditioning
> 
> On Wed, Dec 10, 2014 at 09:47:45AM -0500, Neil Horman wrote:
> > On Wed, Dec 10, 2014 at 08:18:36AM +0000, Wodkowski, PawelX wrote:
> > > > Though, that said, doesn't it seem to anyone else like serialization of
> enqueue
> > > > to a port should be the responsibility of the library, not the application?
> > > >
> > > > Neil
> > >
> > > From my knowledge it is an application  responsibility to serialize access to
> > > queue on particular port.
> > >
> > I understand thats the way it currently is, I'm advocating for the fact that it
> > should not be.
> > Neil
> >
> It could be done, but I think we'd need to add a new API (or new parameter to
> existing API) to do so, as the cost of adding the locks would be severe, even in
> the uncontented case.
> This is why it hasn't been done up till now, obviously enough. In general, where
> we don't provide performant multi-thread safe APIs, we generally don't try and
> provide versions with locks, we just document the limitation and then leave it
> up to the app to determine how best to handle things.
> 
> /Bruce


the problem is when the routing is through the same queue the app crashed. 
example: traffic to 1.1.1.1 from port 0 and 1.1.1.1 from port 1.
You all are right :)
So the only solution are spinlocks, or we must modify 
intel-dpdk-sample-applications-user-guide.pdf to inform users about limitations.
Neil Horman Dec. 10, 2014, 4:16 p.m. UTC | #10
On Wed, Dec 10, 2014 at 02:54:56PM +0000, Bruce Richardson wrote:
> On Wed, Dec 10, 2014 at 09:47:45AM -0500, Neil Horman wrote:
> > On Wed, Dec 10, 2014 at 08:18:36AM +0000, Wodkowski, PawelX wrote:
> > > > Though, that said, doesn't it seem to anyone else like serialization of enqueue
> > > > to a port should be the responsibility of the library, not the application?
> > > > 
> > > > Neil
> > > 
> > > From my knowledge it is an application  responsibility to serialize access to
> > > queue on particular port.
> > > 
> > I understand thats the way it currently is, I'm advocating for the fact that it
> > should not be.
> > Neil
> >
> It could be done, but I think we'd need to add a new API (or new parameter to
> existing API) to do so, as the cost of adding the locks would be severe, even in
> the uncontented case. 
> This is why it hasn't been done up till now, obviously enough. In general, where
> we don't provide performant multi-thread safe APIs, we generally don't try and
> provide versions with locks, we just document the limitation and then leave it 
> up to the app to determine how best to handle things.
> 
This really seems like a false savings to me.  If an application intends to use
multiple processes (which by all rights it seems like the use case that the dpdk
is mostly designed for) then you need locking one way or another, and you've
just made application coding harder, because the application now needs to know
which functions might have internal critical sections that they need to provide
locking for.

I agree that, in the single process case, there might be a slight performance
loss (though I contend it wouldn't be greatly significant).  That said, I would
argue that the right approach is to do the locking internally to the DPDK, then
provide a configuration point which toggles the spinlock defintions to either do
proper locking, or just reduce to empty definitions, the same way the Linux and
BSD kernels do in the uniprocessor case.  That way applications never have to
worry about internal locking, and you can still build for the optimal case when
you need to.

Neil

> /Bruce
>
Stephen Hemminger Dec. 10, 2014, 11:38 p.m. UTC | #11
On Wed, 10 Dec 2014 11:16:46 -0500
Neil Horman <nhorman@tuxdriver.com> wrote:

> This really seems like a false savings to me.  If an application intends to use
> multiple processes (which by all rights it seems like the use case that the dpdk
> is mostly designed for) then you need locking one way or another, and you've
> just made application coding harder, because the application now needs to know
> which functions might have internal critical sections that they need to provide
> locking for.

The DPDK is not Linux.
See the examples of how to route without using locks by doing asymmetric multiprocessing.
I.e queues are only serviced by one CPU.

The cost of a locked operation (even uncontended) is often enough to drop
packet performance by several million PPS.
Neil Horman Dec. 11, 2014, 12:34 a.m. UTC | #12
On Wed, Dec 10, 2014 at 03:38:37PM -0800, Stephen Hemminger wrote:
> On Wed, 10 Dec 2014 11:16:46 -0500
> Neil Horman <nhorman@tuxdriver.com> wrote:
> 
> > This really seems like a false savings to me.  If an application intends to use
> > multiple processes (which by all rights it seems like the use case that the dpdk
> > is mostly designed for) then you need locking one way or another, and you've
> > just made application coding harder, because the application now needs to know
> > which functions might have internal critical sections that they need to provide
> > locking for.
> 
> The DPDK is not Linux.
I never indicated that it was.

> See the examples of how to route without using locks by doing asymmetric multiprocessing.
> I.e queues are only serviced by one CPU.
> 
Yes, I've seen it.

> The cost of a locked operation (even uncontended) is often enough to drop
> packet performance by several million PPS.
Please re-read my note, I clearly stated that a single process use case was a
valid one, but that didn't preclude the need to provide mutual exclusion
internally to the api.  Theres no reason that this locking can't be moved into
the api, and the spinlock api itself either be defined to do locking at compile
time, or defined out as empty macros based on a build variable
(CONFIG_SINGLE_ACCESSOR or some such).  That way you save the application the
headache of having to guess which api calls need locking around them, and you
still get maximal performance if the application being written can guarantee
single accessor status to the dpdk library.

Neil
Thomas Monjalon Dec. 11, 2014, 1:08 a.m. UTC | #13
> > Signed-off-by: Daniel Mrzyglod <danielx.t.mrzyglod@intel.com>
> 
> Acked-by: Neil Horman <nhorman@tuxdriver.com>

Applied

Thanks

Patch
diff mbox

diff --git a/examples/l3fwd-vf/main.c b/examples/l3fwd-vf/main.c
index 2ca5c21..57852d0 100644
--- a/examples/l3fwd-vf/main.c
+++ b/examples/l3fwd-vf/main.c
@@ -54,6 +54,7 @@ 
 #include <rte_per_lcore.h>
 #include <rte_launch.h>
 #include <rte_atomic.h>
+#include <rte_spinlock.h>
 #include <rte_cycles.h>
 #include <rte_prefetch.h>
 #include <rte_lcore.h>
@@ -328,7 +329,7 @@  struct lcore_conf {
 } __rte_cache_aligned;
 
 static struct lcore_conf lcore_conf[RTE_MAX_LCORE];
-
+static rte_spinlock_t spinlock_conf[RTE_MAX_ETHPORTS]={RTE_SPINLOCK_INITIALIZER};
 /* Send burst of packets on an output interface */
 static inline int
 send_burst(struct lcore_conf *qconf, uint16_t n, uint8_t port)
@@ -340,7 +341,10 @@  send_burst(struct lcore_conf *qconf, uint16_t n, uint8_t port)
 	queueid = qconf->tx_queue_id;
 	m_table = (struct rte_mbuf **)qconf->tx_mbufs[port].m_table;
 
+	rte_spinlock_lock(&spinlock_conf[port]) ;
 	ret = rte_eth_tx_burst(port, queueid, m_table, n);
+	rte_spinlock_unlock(&spinlock_conf[port]);
+	
 	if (unlikely(ret < n)) {
 		do {
 			rte_pktmbuf_free(m_table[ret]);