diff mbox series

[v2] eventdev: fix Rx adapter event flush logic

Message ID 1557377440-11616-1-git-send-email-nikhil.rao@intel.com (mailing list archive)
State Superseded, archived
Delegated to: Jerin Jacob
Headers
Series [v2] eventdev: fix Rx adapter event flush logic |

Checks

Context Check Description
ci/checkpatch success coding style OK
ci/Intel-compilation success Compilation OK
ci/mellanox-Performance-Testing success Performance Testing PASS
ci/intel-Performance-Testing success Performance Testing PASS

Commit Message

Rao, Nikhil May 9, 2019, 4:50 a.m. UTC 
  The Rx adapter flushes events only if it has BATCH_SIZE
events buffered where BATCH_SIZE is set to 32, e.g., if a
single packet is sent, it is never passed to
eventdev. Fix this issue by adding an event buffer flush
either when a Rx queue is found to be empty or the adapter service
function has processed the max number of packets for an invocation.

Bugzilla ID: 277

Fixes: 6b83f5935543 ("eventdev: add event buffer flush in Rx adapter")

Reported-by: Matias Elo <matias.elo@nokia.com>

Signed-off-by: Nikhil Rao <nikhil.rao@intel.com>
---
 lib/librte_eventdev/rte_event_eth_rx_adapter.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

v2:
* added Bugzilla ID

Comments

Thomas Monjalon May 9, 2019, 7:17 p.m. UTC | #1 
09/05/2019 06:50, Nikhil Rao:
> The Rx adapter flushes events only if it has BATCH_SIZE
> events buffered where BATCH_SIZE is set to 32, e.g., if a
> single packet is sent, it is never passed to
> eventdev. Fix this issue by adding an event buffer flush
> either when a Rx queue is found to be empty or the adapter service
> function has processed the max number of packets for an invocation.
> 
> Bugzilla ID: 277
> 
> Fixes: 6b83f5935543 ("eventdev: add event buffer flush in Rx adapter")
> 
> Reported-by: Matias Elo <matias.elo@nokia.com>
> 
> Signed-off-by: Nikhil Rao <nikhil.rao@intel.com>
> ---
> --- a/lib/librte_eventdev/rte_event_eth_rx_adapter.c
> +++ b/lib/librte_eventdev/rte_event_eth_rx_adapter.c
> @@ -872,7 +872,7 @@ static uint16_t rxa_gcd_u16(uint16_t a, uint16_t b)
> -	if (buf->count >= BATCH_SIZE)
> +	if (buf->count)

I don't this code, but it looks you are removing a test
which was probably done on purpose.
By the way, you should make your test explicit with "!= 0".
Mattias Rönnblom May 9, 2019, 7:22 p.m. UTC | #2 
On 2019-05-09 06:50, Nikhil Rao wrote:
> The Rx adapter flushes events only if it has BATCH_SIZE
> events buffered where BATCH_SIZE is set to 32, e.g., if a
> single packet is sent, it is never passed to
> eventdev. Fix this issue by adding an event buffer flush
> either when a Rx queue is found to be empty or the adapter service
> function has processed the max number of packets for an invocation.
> 
> Bugzilla ID: 277
> 
> Fixes: 6b83f5935543 ("eventdev: add event buffer flush in Rx adapter")
> 
> Reported-by: Matias Elo <matias.elo@nokia.com>
> 
> Signed-off-by: Nikhil Rao <nikhil.rao@intel.com>
> ---
>   lib/librte_eventdev/rte_event_eth_rx_adapter.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
> 
> v2:
> * added Bugzilla ID
> 
> diff --git a/lib/librte_eventdev/rte_event_eth_rx_adapter.c b/lib/librte_eventdev/rte_event_eth_rx_adapter.c
> index 8d178be..fd3761b 100644
> --- a/lib/librte_eventdev/rte_event_eth_rx_adapter.c
> +++ b/lib/librte_eventdev/rte_event_eth_rx_adapter.c
> @@ -872,7 +872,7 @@ static uint16_t rxa_gcd_u16(uint16_t a, uint16_t b)
>   			break;
>   	}
>   
> -	if (buf->count >= BATCH_SIZE)
> +	if (buf->count)
>   		rxa_flush_event_buffer(rx_adapter);
>   
>   	return nb_rx;
> 

After this change, is there any point in having the event_enqueue_buffer 
being a part of the rx_adapter struct? It could be stack allocated. In 
the process, you could eliminate the rte_event copying that goes on and 
the logic around that. Just make mbufs ETH_EVENT_BUFFER_SIZE in size, 
and retry until the NIC queue is empty or you accumulated 
ETH_EVENT_BUFFER_SIZE events. Then you ship it off to the event device.
Mattias Rönnblom May 9, 2019, 7:23 p.m. UTC | #3 
On 2019-05-09 21:17, Thomas Monjalon wrote:
> 09/05/2019 06:50, Nikhil Rao:
>> The Rx adapter flushes events only if it has BATCH_SIZE
>> events buffered where BATCH_SIZE is set to 32, e.g., if a
>> single packet is sent, it is never passed to
>> eventdev. Fix this issue by adding an event buffer flush
>> either when a Rx queue is found to be empty or the adapter service
>> function has processed the max number of packets for an invocation.
>>
>> Bugzilla ID: 277
>>
>> Fixes: 6b83f5935543 ("eventdev: add event buffer flush in Rx adapter")
>>
>> Reported-by: Matias Elo <matias.elo@nokia.com>
>>
>> Signed-off-by: Nikhil Rao <nikhil.rao@intel.com>
>> ---
>> --- a/lib/librte_eventdev/rte_event_eth_rx_adapter.c
>> +++ b/lib/librte_eventdev/rte_event_eth_rx_adapter.c
>> @@ -872,7 +872,7 @@ static uint16_t rxa_gcd_u16(uint16_t a, uint16_t b)
>> -	if (buf->count >= BATCH_SIZE)
>> +	if (buf->count)
> 
> I don't this code, but it looks you are removing a test
> which was probably done on purpose.
> By the way, you should make your test explicit with "!= 0".
> 

"> 0". count is unsigned.
Rao, Nikhil May 10, 2019, 8:19 a.m. UTC | #4 
> -----Original Message-----
> From: Thomas Monjalon [mailto:thomas@monjalon.net]
> Sent: Friday, May 10, 2019 12:48 AM
> To: Rao, Nikhil <nikhil.rao@intel.com>
> Cc: dev@dpdk.org; jerinj@marvell.com; matias.elo@nokia.com
> Subject: Re: [dpdk-dev] [PATCH v2] eventdev: fix Rx adapter event flush logic
> 
> 09/05/2019 06:50, Nikhil Rao:
> > The Rx adapter flushes events only if it has BATCH_SIZE events
> > buffered where BATCH_SIZE is set to 32, e.g., if a single packet is
> > sent, it is never passed to eventdev. Fix this issue by adding an
> > event buffer flush either when a Rx queue is found to be empty or the
> > adapter service function has processed the max number of packets for
> > an invocation.
> >
> > Bugzilla ID: 277
> >
> > Fixes: 6b83f5935543 ("eventdev: add event buffer flush in Rx adapter")
> >
> > Reported-by: Matias Elo <matias.elo@nokia.com>
> >
> > Signed-off-by: Nikhil Rao <nikhil.rao@intel.com>
> > ---
> > --- a/lib/librte_eventdev/rte_event_eth_rx_adapter.c
> > +++ b/lib/librte_eventdev/rte_event_eth_rx_adapter.c
> > @@ -872,7 +872,7 @@ static uint16_t rxa_gcd_u16(uint16_t a, uint16_t b)
> > -	if (buf->count >= BATCH_SIZE)
> > +	if (buf->count)
> 
> I don't this code, but it looks you are removing a test which was probably
> done on purpose.
The previous test (against BATCH_SIZE)  resulted in events permanently sitting
in the buffer if the number of events in the buffer was less than BATCH_SIZE. 

> By the way, you should make your test explicit with "!= 0".
> 
Done.

Thanks,
Nikhil
Rao, Nikhil May 10, 2019, 8:20 a.m. UTC | #5 
> -----Original Message-----
> From: Mattias Rönnblom [mailto:mattias.ronnblom@ericsson.com]
> Sent: Friday, May 10, 2019 12:52 AM
> To: Rao, Nikhil <nikhil.rao@intel.com>; jerinj@marvell.com
> Cc: dev@dpdk.org; matias.elo@nokia.com
> Subject: Re: [dpdk-dev] [PATCH v2] eventdev: fix Rx adapter event flush logic
> 
> On 2019-05-09 06:50, Nikhil Rao wrote:
> > The Rx adapter flushes events only if it has BATCH_SIZE events
> > buffered where BATCH_SIZE is set to 32, e.g., if a single packet is
> > sent, it is never passed to eventdev. Fix this issue by adding an
> > event buffer flush either when a Rx queue is found to be empty or the
> > adapter service function has processed the max number of packets for
> > an invocation.
> >
> > Bugzilla ID: 277
> >
> > Fixes: 6b83f5935543 ("eventdev: add event buffer flush in Rx adapter")
> >
> > Reported-by: Matias Elo <matias.elo@nokia.com>
> >
> > Signed-off-by: Nikhil Rao <nikhil.rao@intel.com>
> > ---
> >   lib/librte_eventdev/rte_event_eth_rx_adapter.c | 2 +-
> >   1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > v2:
> > * added Bugzilla ID
> >
> > diff --git a/lib/librte_eventdev/rte_event_eth_rx_adapter.c
> > b/lib/librte_eventdev/rte_event_eth_rx_adapter.c
> > index 8d178be..fd3761b 100644
> > --- a/lib/librte_eventdev/rte_event_eth_rx_adapter.c
> > +++ b/lib/librte_eventdev/rte_event_eth_rx_adapter.c
> > @@ -872,7 +872,7 @@ static uint16_t rxa_gcd_u16(uint16_t a, uint16_t b)
> >   			break;
> >   	}
> >
> > -	if (buf->count >= BATCH_SIZE)
> > +	if (buf->count)
> >   		rxa_flush_event_buffer(rx_adapter);
> >
> >   	return nb_rx;
> >
> 
> After this change, is there any point in having the event_enqueue_buffer
> being a part of the rx_adapter struct? It could be stack allocated. In the
> process, you could eliminate the rte_event copying that goes on and the
> logic around that. Just make mbufs ETH_EVENT_BUFFER_SIZE in size, and
> retry until the NIC queue is empty or you accumulated
> ETH_EVENT_BUFFER_SIZE events. Then you ship it off to the event device.

You are right, I will post a separate (from the bug fix) patch for this.

Thanks,
Nikhil
diff mbox series

Patch

diff --git a/lib/librte_eventdev/rte_event_eth_rx_adapter.c b/lib/librte_eventdev/rte_event_eth_rx_adapter.c
index 8d178be..fd3761b 100644
--- a/lib/librte_eventdev/rte_event_eth_rx_adapter.c
+++ b/lib/librte_eventdev/rte_event_eth_rx_adapter.c
@@ -872,7 +872,7 @@  static uint16_t rxa_gcd_u16(uint16_t a, uint16_t b)
 			break;
 	}
 
-	if (buf->count >= BATCH_SIZE)
+	if (buf->count)
 		rxa_flush_event_buffer(rx_adapter);
 
 	return nb_rx;