[dpdk-dev] [PATCH 1/4] net/memif: do not update local copy of tail in tx function

[dpdk-dev] [PATCH 1/4] net/memif: do not update local copy of tail in tx function

[Honnappa Nagarahalli]

In the case of S2M queues, the receiver synchronizes with the sender
(i.e. informs of the packets it has received) using ring->tail.
Hence, the sender does not need to update last_tail.

In the case of M2S queues, the receiver uses last_tail to
keep track of the descriptors it has received. The
sender is not required to update the last_tail. Updating
the last_tail makes it a shared variable between the
transmitter and receiver affecting the performance.

Fixes: 09c7e63a71f9 ("net/memif: introduce memory interface PMD")
Cc: jgrajcia@cisco.com
Cc: stable@dpdk.org

Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
---
 drivers/net/memif/rte_eth_memif.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c
index c1c7e9f8d..fd7dbc53e 100644
--- a/drivers/net/memif/rte_eth_memif.c
+++ b/drivers/net/memif/rte_eth_memif.c
@@ -568,12 +568,10 @@ eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 	ring_size = 1 << mq->log2_ring_size;
 	mask = ring_size - 1;
 
-	n_free = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE) - mq->last_tail;
-	mq->last_tail += n_free;
-
 	if (type == MEMIF_RING_S2M) {
 		slot = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE);
-		n_free = ring_size - slot + mq->last_tail;
+		n_free = ring_size - slot +
+				__atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 	} else {
 		slot = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 		n_free = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE) - slot;
-- 
2.17.1

[dpdk-dev] [PATCH 2/4] net/memif: relax the load of ring tail pointer for M2S ring

[Honnappa Nagarahalli]

For M2S rings, ring->tail is updated by the sender and eth_memif_tx
function is called in the context of sending thread. The loads in
the sender do not need to synchronize with its own stores.

Fixes: a2aafb9aa651 ("net/memif: optimize with one-way barrier")
Cc: phil.yang@arm.com
Cc: stable@dpdk.org

Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
---
 drivers/net/memif/rte_eth_memif.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c
index fd7dbc53e..0d064c8fa 100644
--- a/drivers/net/memif/rte_eth_memif.c
+++ b/drivers/net/memif/rte_eth_memif.c
@@ -573,7 +573,13 @@ eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 		n_free = ring_size - slot +
 				__atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 	} else {
-		slot = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
+		/* For M2S queues ring->tail is updated by the sender and
+		 * this function is called in the context of sending thread.
+		 * The loads in the sender do not need to synchronize with
+		 * its own stores. Hence, the following load can be a
+		 * relaxed load.
+		 */
+		slot = __atomic_load_n(&ring->tail, __ATOMIC_RELAXED);
 		n_free = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE) - slot;
 	}
 
-- 
2.17.1

[dpdk-dev] [PATCH 3/4] net/memif: relax the load of ring head pointer for S2M ring

[Honnappa Nagarahalli]

For S2M rings, ring->head is updated by the sender and eth_memif_tx
function is called in the context of sending thread. The loads in
the sender do not need to synchronize with its own stores.

Fixes: a2aafb9aa651 ("net/memif: optimize with one-way barrier")
Cc: phil.yang@arm.com
Cc: stable@dpdk.org

Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
---
 drivers/net/memif/rte_eth_memif.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c
index 0d064c8fa..435c6345c 100644
--- a/drivers/net/memif/rte_eth_memif.c
+++ b/drivers/net/memif/rte_eth_memif.c
@@ -569,7 +569,13 @@ eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 	mask = ring_size - 1;
 
 	if (type == MEMIF_RING_S2M) {
-		slot = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE);
+		/* For S2M queues ring->head is updated by the sender and
+		 * this function is called in the context of sending thread.
+		 * The loads in the sender do not need to synchronize with
+		 * its own stores. Hence, the following load can be a
+		 * relaxed load.
+		 */
+		slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 		n_free = ring_size - slot +
 				__atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 	} else {
-- 
2.17.1

[dpdk-dev] [PATCH 4/4] net/memif: relax the load of ring head pointer for M2S ring

[Honnappa Nagarahalli]

For M2S rings, ring->head is updated by the receiver and eth_memif_rx
function is called in the context of receiving thread. The loads in
the sender do not need to synchronize with its own stores.

Fixes: a2aafb9aa651 ("net/memif: optimize with one-way barrier")
Cc: phil.yang@arm.com
Cc: stable@dpdk.org

Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
---
 drivers/net/memif/rte_eth_memif.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c
index 435c6345c..1cbb1fd7a 100644
--- a/drivers/net/memif/rte_eth_memif.c
+++ b/drivers/net/memif/rte_eth_memif.c
@@ -405,7 +405,11 @@ eth_memif_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 
 refill:
 	if (type == MEMIF_RING_M2S) {
-		head = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE);
+		/* ring->head is updated by the receiver and this function
+		 * is called in the context of receiver thread. The loads in
+		 * the receiver do not need to synchronize with its own stores.
+		 */
+		head = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 		n_slots = ring_size - head + mq->last_tail;
 
 		while (n_slots--) {
-- 
2.17.1

[dpdk-dev] [PATCH v2 1/8] net/memif: do not update local copy of tail in tx function

[Honnappa Nagarahalli]

In the case of S2M queues, the receiver synchronizes with the sender
(i.e. informs of the packets it has received) using ring->tail.
Hence, the sender does not need to update last_tail.

In the case of M2S queues, the receiver uses last_tail to
keep track of the descriptors it has received. The
sender is not required to update the last_tail. Updating
the last_tail makes it a shared variable between the
transmitter and receiver affecting the performance.

Fixes: 09c7e63a71f9 ("net/memif: introduce memory interface PMD")
Cc: jgrajcia@cisco.com
Cc: stable@dpdk.org

Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
---
 drivers/net/memif/rte_eth_memif.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c
index a19c0f3e6..130099f2e 100644
--- a/drivers/net/memif/rte_eth_memif.c
+++ b/drivers/net/memif/rte_eth_memif.c
@@ -580,12 +580,10 @@ eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 	ring_size = 1 << mq->log2_ring_size;
 	mask = ring_size - 1;
 
-	n_free = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE) - mq->last_tail;
-	mq->last_tail += n_free;
-
 	if (type == MEMIF_RING_S2M) {
 		slot = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE);
-		n_free = ring_size - slot + mq->last_tail;
+		n_free = ring_size - slot +
+				__atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 	} else {
 		slot = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 		n_free = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE) - slot;
-- 
2.17.1

[dpdk-dev] [PATCH v2 2/8] net/memif: relax the load of ring tail pointer for M2S ring

[Honnappa Nagarahalli]

For M2S rings, ring->tail is updated by the sender and eth_memif_tx
function is called in the context of sending thread. The loads in
the sender do not need to synchronize with its own stores.

Fixes: a2aafb9aa651 ("net/memif: optimize with one-way barrier")
Cc: phil.yang@arm.com
Cc: stable@dpdk.org

Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
---
 drivers/net/memif/rte_eth_memif.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c
index 130099f2e..8bacacaa8 100644
--- a/drivers/net/memif/rte_eth_memif.c
+++ b/drivers/net/memif/rte_eth_memif.c
@@ -585,7 +585,13 @@ eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 		n_free = ring_size - slot +
 				__atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 	} else {
-		slot = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
+		/* For M2S queues ring->tail is updated by the sender and
+		 * this function is called in the context of sending thread.
+		 * The loads in the sender do not need to synchronize with
+		 * its own stores. Hence, the following load can be a
+		 * relaxed load.
+		 */
+		slot = __atomic_load_n(&ring->tail, __ATOMIC_RELAXED);
 		n_free = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE) - slot;
 	}
 
-- 
2.17.1

[dpdk-dev] [PATCH v2 3/8] net/memif: relax the load of ring head pointer for M2S ring

[Honnappa Nagarahalli]

For M2S rings, ring->head is updated by the receiver and eth_memif_rx
function is called in the context of receiving thread. The loads in
the receiver do not need to synchronize with its own stores.

Fixes: a2aafb9aa651 ("net/memif: optimize with one-way barrier")
Cc: phil.yang@arm.com
Cc: stable@dpdk.org

Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
---
 drivers/net/memif/rte_eth_memif.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c
index 8bacacaa8..159b45180 100644
--- a/drivers/net/memif/rte_eth_memif.c
+++ b/drivers/net/memif/rte_eth_memif.c
@@ -410,7 +410,11 @@ eth_memif_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 
 refill:
 	if (type == MEMIF_RING_M2S) {
-		head = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE);
+		/* ring->head is updated by the receiver and this function
+		 * is called in the context of receiver thread. The loads in
+		 * the receiver do not need to synchronize with its own stores.
+		 */
+		head = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 		n_slots = ring_size - head + mq->last_tail;
 
 		while (n_slots--) {
-- 
2.17.1

[dpdk-dev] [PATCH v2 4/8] net/memif: relax the load of ring head pointer for S2M ring

[Honnappa Nagarahalli]

For S2M rings, ring->head is updated by the sender and eth_memif_tx
function is called in the context of sending thread. The loads in
the sender do not need to synchronize with its own stores.

Fixes: a2aafb9aa651 ("net/memif: optimize with one-way barrier")
Cc: phil.yang@arm.com
Cc: stable@dpdk.org

Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
---
 drivers/net/memif/rte_eth_memif.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c
index 159b45180..96db76121 100644
--- a/drivers/net/memif/rte_eth_memif.c
+++ b/drivers/net/memif/rte_eth_memif.c
@@ -585,7 +585,13 @@ eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 	mask = ring_size - 1;
 
 	if (type == MEMIF_RING_S2M) {
-		slot = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE);
+		/* For S2M queues ring->head is updated by the sender and
+		 * this function is called in the context of sending thread.
+		 * The loads in the sender do not need to synchronize with
+		 * its own stores. Hence, the following load can be a
+		 * relaxed load.
+		 */
+		slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 		n_free = ring_size - slot +
 				__atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 	} else {
-- 
2.17.1

[dpdk-dev] [PATCH v2 5/8] net/memif: relax the load of ring head pointer for M2S zc ring

[Honnappa Nagarahalli]

For zero-copy M2S rings, ring->head is updated by the receiver
and eth_memif_rx_zc function is called in the context of
receiving thread. The loads in the receiver do not need to
synchronize with its own stores.

Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
---
 drivers/net/memif/rte_eth_memif.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)

diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c
index 96db76121..404c86ae8 100644
--- a/drivers/net/memif/rte_eth_memif.c
+++ b/drivers/net/memif/rte_eth_memif.c
@@ -514,11 +514,11 @@ eth_memif_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 
 /* Supply master with new buffers */
 refill:
-	/* The ring->head acts as a guard variable between Tx and Rx
-	 * threads, so using load-acquire pairs with store-release
-	 * to synchronize it between threads.
+	/* ring->head is updated by the receiver and this function
+	 * is called in the context of receiver thread. The loads in
+	 * the receiver do not need to synchronize with its own stores.
 	 */
-	head = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE);
+	head = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 	n_slots = ring_size - head + mq->last_tail;
 
 	if (n_slots < 32)
@@ -543,6 +543,10 @@ eth_memif_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 			(uint8_t *)proc_private->regions[d0->region]->addr;
 	}
 no_free_mbufs:
+	/* The ring->head acts as a guard variable between Tx and Rx
+	 * threads, so using store-release pairs with load-acquire
+	 * in function eth_memif_tx.
+	 */
 	__atomic_store_n(&ring->head, head, __ATOMIC_RELEASE);
 
 	mq->n_pkts += n_rx_pkts;
-- 
2.17.1

[dpdk-dev] [PATCH v2 6/8] net/memif: remove extra check

[Honnappa Nagarahalli]

eth_memif_tx_zc function is called only for S2M rings. Remove
additional code for M2S rings in this function.

Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
---
 drivers/net/memif/rte_eth_memif.c | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)

diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c
index 404c86ae8..9babd073a 100644
--- a/drivers/net/memif/rte_eth_memif.c
+++ b/drivers/net/memif/rte_eth_memif.c
@@ -735,7 +735,6 @@ eth_memif_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 		rte_eth_devices[mq->in_port].process_private;
 	memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 	uint16_t slot, n_free, ring_size, mask, n_tx_pkts = 0;
-	memif_ring_type_t type = mq->type;
 	struct rte_eth_link link;
 
 	if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
@@ -812,11 +811,8 @@ eth_memif_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 	}
 
 no_free_slots:
-	/* update ring pointers */
-	if (type == MEMIF_RING_S2M)
-		__atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
-	else
-		__atomic_store_n(&ring->tail, slot, __ATOMIC_RELEASE);
+	/* ring type always MEMIF_RING_S2M */
+	__atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
 
 	/* Send interrupt, if enabled. */
 	if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
-- 
2.17.1

[dpdk-dev] [PATCH v2 7/8] net/memif: relax the load of ring head pointer for S2M zc ring

[Honnappa Nagarahalli]

For zero-copy S2M rings, ring->head is updated by the sender and
eth_memif_tx_zc function is called in the context of sending thread.
The loads in the sender do not need to synchronize with its own stores.

Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
---
 drivers/net/memif/rte_eth_memif.c | 14 ++++++++++----
 1 file changed, 10 insertions(+), 4 deletions(-)

diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c
index 9babd073a..704350022 100644
--- a/drivers/net/memif/rte_eth_memif.c
+++ b/drivers/net/memif/rte_eth_memif.c
@@ -752,11 +752,13 @@ eth_memif_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 	memif_free_stored_mbufs(proc_private, mq);
 
 	/* ring type always MEMIF_RING_S2M */
-	/* The ring->head acts as a guard variable between Tx and Rx
-	 * threads, so using load-acquire pairs with store-release
-	 * to synchronize it between threads.
+	/* For S2M queues ring->head is updated by the sender and
+	 * this function is called in the context of sending thread.
+	 * The loads in the sender do not need to synchronize with
+	 * its own stores. Hence, the following load can be a
+	 * relaxed load.
 	 */
-	slot = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE);
+	slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 	n_free = ring_size - slot + mq->last_tail;
 
 	int used_slots;
@@ -812,6 +814,10 @@ eth_memif_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 
 no_free_slots:
 	/* ring type always MEMIF_RING_S2M */
+	/* The ring->head acts as a guard variable between Tx and Rx
+	 * threads, so using store-release pairs with load-acquire
+	 * in function eth_memif_rx for S2M rings.
+	 */
 	__atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
 
 	/* Send interrupt, if enabled. */
-- 
2.17.1

[dpdk-dev] [PATCH v2 8/8] net/memif: move the barrier outside the loop

[Honnappa Nagarahalli]

load-acquire memory order has a barrier. Loading it inside
the loop will result in a barrier in every iteration. Hence,
load the variable once outside the loop.

Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
---
 drivers/net/memif/rte_eth_memif.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c
index 704350022..c73cde8fd 100644
--- a/drivers/net/memif/rte_eth_memif.c
+++ b/drivers/net/memif/rte_eth_memif.c
@@ -249,16 +249,17 @@ memif_get_buffer(struct pmd_process_private *proc_private, memif_desc_t *d)
 static void
 memif_free_stored_mbufs(struct pmd_process_private *proc_private, struct memif_queue *mq)
 {
+	uint16_t cur_tail;
 	uint16_t mask = (1 << mq->log2_ring_size) - 1;
 	memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 
 	/* FIXME: improve performance */
 	/* The ring->tail acts as a guard variable between Tx and Rx
 	 * threads, so using load-acquire pairs with store-release
-	 * to synchronize it between threads.
+	 * in function eth_memif_rx for S2M queues.
 	 */
-	while (mq->last_tail != __atomic_load_n(&ring->tail,
-						__ATOMIC_ACQUIRE)) {
+	cur_tail = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
+	while (mq->last_tail != cur_tail) {
 		RTE_MBUF_PREFETCH_TO_FREE(mq->buffers[(mq->last_tail + 1) & mask]);
 		/* Decrement refcnt and free mbuf. (current segment) */
 		rte_mbuf_refcnt_update(mq->buffers[mq->last_tail & mask], -1);
-- 
2.17.1

Re: [dpdk-dev] [PATCH v2 1/8] net/memif: do not update local copy of tail in tx function

[Stephen Hemminger]

On Mon, 28 Sep 2020 14:03:27 -0500
Honnappa Nagarahalli <honnappa.nagarahalli@arm.com> wrote:

> In the case of S2M queues, the receiver synchronizes with the sender
> (i.e. informs of the packets it has received) using ring->tail.
> Hence, the sender does not need to update last_tail.
> 
> In the case of M2S queues, the receiver uses last_tail to
> keep track of the descriptors it has received. The
> sender is not required to update the last_tail. Updating
> the last_tail makes it a shared variable between the
> transmitter and receiver affecting the performance.
> 
> Fixes: 09c7e63a71f9 ("net/memif: introduce memory interface PMD")
> Cc: jgrajcia@cisco.com
> Cc: stable@dpdk.org
> 
> Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
> Reviewed-by: Phil Yang <phil.yang@arm.com>
> Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>

This patch series will conflict with the pending master/slave patchset.
Please let the master/slave renaming go in first.

Re: [dpdk-dev] [PATCH v2 1/8] net/memif: do not update local copy of tail in tx function

[Honnappa Nagarahalli]

<snip>

> 
> On Mon, 28 Sep 2020 14:03:27 -0500
> Honnappa Nagarahalli <honnappa.nagarahalli@arm.com> wrote:
> 
> > In the case of S2M queues, the receiver synchronizes with the sender
> > (i.e. informs of the packets it has received) using ring->tail.
> > Hence, the sender does not need to update last_tail.
> >
> > In the case of M2S queues, the receiver uses last_tail to keep track
> > of the descriptors it has received. The sender is not required to
> > update the last_tail. Updating the last_tail makes it a shared
> > variable between the transmitter and receiver affecting the
> > performance.
> >
> > Fixes: 09c7e63a71f9 ("net/memif: introduce memory interface PMD")
> > Cc: jgrajcia@cisco.com
> > Cc: stable@dpdk.org
> >
> > Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
> > Reviewed-by: Phil Yang <phil.yang@arm.com>
> > Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
> 
> This patch series will conflict with the pending master/slave patchset.
> Please let the master/slave renaming go in first.
+1, review can go on.

Re: [dpdk-dev] [PATCH v2 1/8] net/memif: do not update local copy of tail in tx function

[Honnappa Nagarahalli]

Hi Jakub,
	Appreciate if you could review this series and provide any comments you might have.

Thank you,
Honnappa

> -----Original Message-----
> From: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
> Sent: Monday, September 28, 2020 2:03 PM
> To: dev@dpdk.org; Honnappa Nagarahalli <Honnappa.Nagarahalli@arm.com>;
> Phil Yang <Phil.Yang@arm.com>; jgrajcia@cisco.com; ferruh.yigit@intel.com
> Cc: nd <nd@arm.com>; stable@dpdk.org
> Subject: [PATCH v2 1/8] net/memif: do not update local copy of tail in tx
> function
> 
> In the case of S2M queues, the receiver synchronizes with the sender (i.e.
> informs of the packets it has received) using ring->tail.
> Hence, the sender does not need to update last_tail.
> 
> In the case of M2S queues, the receiver uses last_tail to keep track of the
> descriptors it has received. The sender is not required to update the last_tail.
> Updating the last_tail makes it a shared variable between the transmitter and
> receiver affecting the performance.
> 
> Fixes: 09c7e63a71f9 ("net/memif: introduce memory interface PMD")
> Cc: jgrajcia@cisco.com
> Cc: stable@dpdk.org
> 
> Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
> Reviewed-by: Phil Yang <phil.yang@arm.com>
> Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
> ---
>  drivers/net/memif/rte_eth_memif.c | 6 ++----
>  1 file changed, 2 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/net/memif/rte_eth_memif.c
> b/drivers/net/memif/rte_eth_memif.c
> index a19c0f3e6..130099f2e 100644
> --- a/drivers/net/memif/rte_eth_memif.c
> +++ b/drivers/net/memif/rte_eth_memif.c
> @@ -580,12 +580,10 @@ eth_memif_tx(void *queue, struct rte_mbuf **bufs,
> uint16_t nb_pkts)
>  	ring_size = 1 << mq->log2_ring_size;
>  	mask = ring_size - 1;
> 
> -	n_free = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE) - mq-
> >last_tail;
> -	mq->last_tail += n_free;
> -
>  	if (type == MEMIF_RING_S2M) {
>  		slot = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE);
> -		n_free = ring_size - slot + mq->last_tail;
> +		n_free = ring_size - slot +
> +				__atomic_load_n(&ring->tail,
> __ATOMIC_ACQUIRE);
>  	} else {
>  		slot = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
>  		n_free = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE)
> - slot;
> --
> 2.17.1

Re: [dpdk-dev] [PATCH v2 1/8] net/memif: do not update local copy of tail in tx function

[Jakub Grajciar -X (jgrajcia - PANTHEON TECH SRO at Cisco)]

> > -----Original Message-----
> > From: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
> > Sent: Monday, September 28, 2020 2:03 PM
> > To: dev@dpdk.org; Honnappa Nagarahalli
> <Honnappa.Nagarahalli@arm.com>;
> > Phil Yang <Phil.Yang@arm.com>; jgrajcia@cisco.com;
> > ferruh.yigit@intel.com
> > Cc: nd <nd@arm.com>; stable@dpdk.org
> > Subject: [PATCH v2 1/8] net/memif: do not update local copy of tail in
> > tx function
> >
> > In the case of S2M queues, the receiver synchronizes with the sender (i.e.
> > informs of the packets it has received) using ring->tail.
> > Hence, the sender does not need to update last_tail.
> >
> > In the case of M2S queues, the receiver uses last_tail to keep track
> > of the descriptors it has received. The sender is not required to update the
> last_tail.
> > Updating the last_tail makes it a shared variable between the
> > transmitter and receiver affecting the performance.

Hi Honnappa,

The patch series is looking good.

Reviewed-by: Jakub Grajciar <jgrajcia@cisco.com>

Re: [dpdk-dev] [PATCH v2 1/8] net/memif: do not update local copy of tail in tx function

[Ferruh Yigit]

On 10/9/2020 12:23 PM, Jakub Grajciar -X (jgrajcia - PANTHEON TECH SRO at Cisco) 
wrote:
>>> -----Original Message-----
>>> From: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
>>> Sent: Monday, September 28, 2020 2:03 PM
>>> To: dev@dpdk.org; Honnappa Nagarahalli
>> <Honnappa.Nagarahalli@arm.com>;
>>> Phil Yang <Phil.Yang@arm.com>; jgrajcia@cisco.com;
>>> ferruh.yigit@intel.com
>>> Cc: nd <nd@arm.com>; stable@dpdk.org
>>> Subject: [PATCH v2 1/8] net/memif: do not update local copy of tail in
>>> tx function
>>>
>>> In the case of S2M queues, the receiver synchronizes with the sender (i.e.
>>> informs of the packets it has received) using ring->tail.
>>> Hence, the sender does not need to update last_tail.
>>>
>>> In the case of M2S queues, the receiver uses last_tail to keep track
>>> of the descriptors it has received. The sender is not required to update the
>> last_tail.
>>> Updating the last_tail makes it a shared variable between the
>>> transmitter and receiver affecting the performance.
> 
> Hi Honnappa,
> 
> The patch series is looking good.
> 
> Reviewed-by: Jakub Grajciar <jgrajcia@cisco.com>
> 

Series applied to dpdk-next-net/main, thanks.