[dpdk-dev] [memnic PATCH] pmd: use rte_atomic32_cmpset instead of cmpxchg

[dpdk-dev] [memnic PATCH] pmd: use rte_atomic32_cmpset instead of cmpxchg

[Hiroshi Shimamoto]

From: Hiroshi Shimamoto <h-shimamoto@ct.jp.nec.com>

Because DPDK has its own compare and set function to optimize to
dedicated processor type, use that rte_atomic32_cmpset() instead of
cmpxchg macro which is specially introduced for MEMNIC.

Signed-off-by: Hiroshi Shimamoto <h-shimamoto@ct.jp.nec.com>
---
 common/memnic.h  | 12 ------------
 pmd/pmd_memnic.c | 10 ++++++----
 2 files changed, 6 insertions(+), 16 deletions(-)

diff --git a/common/memnic.h b/common/memnic.h
index 2187ac1..84e941c 100644
--- a/common/memnic.h
+++ b/common/memnic.h
@@ -120,18 +120,6 @@ struct memnic_area {
 /* for userspace */
 #define ACCESS_ONCE(x) (*(volatile typeof(x) *)&(x))
 
-static inline uint32_t cmpxchg(uint32_t *dst, uint32_t old, uint32_t new)
-{
-	volatile uint32_t *ptr = (volatile uint32_t *)dst;
-	uint32_t ret;
-
-	asm volatile("lock; cmpxchgl %2, %1"
-		     : "=a" (ret), "+m" (*ptr)
-		     : "r" (new), "0" (old)
-		     : "memory");
-
-	return ret;
-}
 #endif /* __KERNEL__ */
 
 #endif /* MEMNIC_H */
diff --git a/pmd/pmd_memnic.c b/pmd/pmd_memnic.c
index 4a1c1e4..4abdf26 100644
--- a/pmd/pmd_memnic.c
+++ b/pmd/pmd_memnic.c
@@ -314,7 +314,7 @@ static uint16_t memnic_xmit_pkts(void *tx_queue,
 	struct memnic_data *data = &adapter->nic->down;
 	struct memnic_packet *p;
 	uint16_t nr;
-	int idx, old;
+	int idx;
 	struct rte_eth_stats *st = &adapter->stats[rte_lcore_id()];
 	uint64_t pkts, bytes, errs;
 
@@ -335,10 +335,12 @@ static uint16_t memnic_xmit_pkts(void *tx_queue,
 retry:
 		idx = ACCESS_ONCE(adapter->down_idx);
 		p = &data->packets[idx];
-		old = cmpxchg(&p->status, MEMNIC_PKT_ST_FREE, MEMNIC_PKT_ST_USED);
-		if (old != MEMNIC_PKT_ST_FREE) {
-			if (old == MEMNIC_PKT_ST_FILLED &&
+		if (unlikely(rte_atomic32_cmpset(&p->status,
+				MEMNIC_PKT_ST_FREE, MEMNIC_PKT_ST_USED) == 0)) {
+			/* cmpxchg failed */
+			if (p->status == MEMNIC_PKT_ST_FILLED &&
 					idx == ACCESS_ONCE(adapter->down_idx)) {
+				/* what we're seeing is FILLED means queue full */
 				errs++;
 				break;
 			}
-- 
1.8.4

Re: [dpdk-dev] [memnic PATCH] pmd: use rte_atomic32_cmpset instead of cmpxchg

[Olivier MATZ]

On Thursday, April 03, 2014 07:34:56 AM Hiroshi Shimamoto wrote:
> From: Hiroshi Shimamoto <h-shimamoto@ct.jp.nec.com>
> 
> Because DPDK has its own compare and set function to optimize to
> dedicated processor type, use that rte_atomic32_cmpset() instead of
> cmpxchg macro which is specially introduced for MEMNIC.
> 
> Signed-off-by: Hiroshi Shimamoto <h-shimamoto@ct.jp.nec.com>
> ---
>  common/memnic.h  | 12 ------------
>  pmd/pmd_memnic.c | 10 ++++++----
>  2 files changed, 6 insertions(+), 16 deletions(-)

Reviewed-by: Olivier Matz <olivier.matz@6wind.com>

Re: [dpdk-dev] [memnic PATCH] pmd: use rte_atomic32_cmpset instead of cmpxchg

[Thomas Monjalon]

2014-04-11 11:37, Olivier MATZ:
> On Thursday, April 03, 2014 07:34:56 AM Hiroshi Shimamoto wrote:
> > From: Hiroshi Shimamoto <h-shimamoto@ct.jp.nec.com>
> > 
> > Because DPDK has its own compare and set function to optimize to
> > dedicated processor type, use that rte_atomic32_cmpset() instead of
> > cmpxchg macro which is specially introduced for MEMNIC.
> > 
> > Signed-off-by: Hiroshi Shimamoto <h-shimamoto@ct.jp.nec.com>
> 
> Reviewed-by: Olivier Matz <olivier.matz@6wind.com>

Applied

Thanks
-- 
Thomas