[v1,6/6] distributor: fix handshake deadlock
diff mbox series

Message ID 20200915193449.13310-7-l.wojciechow@partner.samsung.com
State Superseded
Delegated to: David Marchand
Headers show
Series
  • fix distributor synchronization issues
Related show

Checks

Context Check Description
ci/travis-robot success Travis build: passed
ci/Intel-compilation success Compilation OK
ci/iol-testing success Testing PASS
ci/checkpatch success coding style OK

Commit Message

Lukasz Wojciechowski Sept. 15, 2020, 7:34 p.m. UTC
Synchronization of data exchange between distributor and worker cores
is based on 2 handshakes: retptr64 for returning mbufs from workers
to distributor and bufptr64 for passing mbufs to workers.

Without proper order of verifying those 2 handshakes a deadlock may
occur. This can happen when worker core want to return back mbufs
and waits for retptr handshake to be cleared and distributor core
wait for bufptr to send mbufs to worker.

This can happen as worker core first returns mbufs to distributor
and later gets new mbufs, while distributor first release mbufs
to worker and later handle returning packets.

This patch fixes possibility of the deadlock by always taking care
of returning packets first on the distributor side and handling
packets while waiting to release new.

Fixes: 775003ad2f96 ("distributor: add new burst-capable library")
Cc: david.hunt@intel.com
Cc: stable@dpdk.org

Signed-off-by: Lukasz Wojciechowski <l.wojciechow@partner.samsung.com>
---
 lib/librte_distributor/rte_distributor.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

Comments

David Hunt Sept. 17, 2020, 1:28 p.m. UTC | #1
Hi Lukasz,

On 15/9/2020 8:34 PM, Lukasz Wojciechowski wrote:
> Synchronization of data exchange between distributor and worker cores
> is based on 2 handshakes: retptr64 for returning mbufs from workers
> to distributor and bufptr64 for passing mbufs to workers.
>
> Without proper order of verifying those 2 handshakes a deadlock may
> occur. This can happen when worker core want to return back mbufs
> and waits for retptr handshake to be cleared and distributor core
> wait for bufptr to send mbufs to worker.
>
> This can happen as worker core first returns mbufs to distributor
> and later gets new mbufs, while distributor first release mbufs
> to worker and later handle returning packets.
>
> This patch fixes possibility of the deadlock by always taking care
> of returning packets first on the distributor side and handling
> packets while waiting to release new.
>
> Fixes: 775003ad2f96 ("distributor: add new burst-capable library")
> Cc: david.hunt@intel.com
> Cc: stable@dpdk.org
>
> Signed-off-by: Lukasz Wojciechowski <l.wojciechow@partner.samsung.com>
> ---
>   lib/librte_distributor/rte_distributor.c | 9 ++++++---
>   1 file changed, 6 insertions(+), 3 deletions(-)
>
> diff --git a/lib/librte_distributor/rte_distributor.c b/lib/librte_distributor/rte_distributor.c
> index 89493c331..12b3db33c 100644
> --- a/lib/librte_distributor/rte_distributor.c
> +++ b/lib/librte_distributor/rte_distributor.c
> @@ -321,12 +321,14 @@ release(struct rte_distributor *d, unsigned int wkr)
>   	struct rte_distributor_buffer *buf = &(d->bufs[wkr]);
>   	unsigned int i;
>   
> +	handle_returns(d, wkr);
> +
>   	/* Sync with worker on GET_BUF flag */
>   	while (!(__atomic_load_n(&(d->bufs[wkr].bufptr64[0]), __ATOMIC_ACQUIRE)
> -		& RTE_DISTRIB_GET_BUF))
> +		& RTE_DISTRIB_GET_BUF)) {
> +		handle_returns(d, wkr);
>   		rte_pause();
> -
> -	handle_returns(d, wkr);
> +	}
>   
>   	buf->count = 0;
>   
> @@ -376,6 +378,7 @@ rte_distributor_process(struct rte_distributor *d,
>   		/* Flush out all non-full cache-lines to workers. */
>   		for (wid = 0 ; wid < d->num_workers; wid++) {
>   			/* Sync with worker on GET_BUF flag. */
> +			handle_returns(d, wid);
>   			if (__atomic_load_n(&(d->bufs[wid].bufptr64[0]),
>   				__ATOMIC_ACQUIRE) & RTE_DISTRIB_GET_BUF) {
>   				release(d, wid);

Makes sense. Thanks for the series.  Again, no degradation in 
performance on my systems.

Acked-by: David Hunt <david.hunt@intel.com>

Patch
diff mbox series

diff --git a/lib/librte_distributor/rte_distributor.c b/lib/librte_distributor/rte_distributor.c
index 89493c331..12b3db33c 100644
--- a/lib/librte_distributor/rte_distributor.c
+++ b/lib/librte_distributor/rte_distributor.c
@@ -321,12 +321,14 @@  release(struct rte_distributor *d, unsigned int wkr)
 	struct rte_distributor_buffer *buf = &(d->bufs[wkr]);
 	unsigned int i;
 
+	handle_returns(d, wkr);
+
 	/* Sync with worker on GET_BUF flag */
 	while (!(__atomic_load_n(&(d->bufs[wkr].bufptr64[0]), __ATOMIC_ACQUIRE)
-		& RTE_DISTRIB_GET_BUF))
+		& RTE_DISTRIB_GET_BUF)) {
+		handle_returns(d, wkr);
 		rte_pause();
-
-	handle_returns(d, wkr);
+	}
 
 	buf->count = 0;
 
@@ -376,6 +378,7 @@  rte_distributor_process(struct rte_distributor *d,
 		/* Flush out all non-full cache-lines to workers. */
 		for (wid = 0 ; wid < d->num_workers; wid++) {
 			/* Sync with worker on GET_BUF flag. */
+			handle_returns(d, wid);
 			if (__atomic_load_n(&(d->bufs[wid].bufptr64[0]),
 				__ATOMIC_ACQUIRE) & RTE_DISTRIB_GET_BUF) {
 				release(d, wid);