]> git.droids-corp.org - dpdk.git/commitdiff
net/i40e: fix risk in descriptor read in scalar Rx
authorRuifeng Wang <ruifeng.wang@arm.com>
Wed, 15 Sep 2021 08:33:39 +0000 (16:33 +0800)
committerQi Zhang <qi.z.zhang@intel.com>
Thu, 11 Nov 2021 12:29:23 +0000 (13:29 +0100)
Rx descriptor is 16B/32B in size. If the DD bit is set, it indicates
that the rest of the descriptor words have valid values. Hence, the
word containing DD bit must be read first before reading the rest of
the descriptor words.

Since the entire descriptor is not read atomically, on relaxed memory
ordered systems like Aarch64, read of the word containing DD field
could be reordered after read of other words.

Read barrier is inserted between read of the word with DD field
and read of other words. The barrier ensures that the fetched data
is correct.

Testpmd single core test showed no performance drop on x86 or N1SDP.
On ThunderX2, 22% performance regression was observed.

Fixes: 7b0cf70135d1 ("net/i40e: support ARM platform")
Cc: stable@dpdk.org
Signed-off-by: Ruifeng Wang <ruifeng.wang@arm.com>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
drivers/net/i40e/i40e_rxtx.c

index 41fe3bf481754cfe40544b8deaaa83fc2b0b3c18..e4cb33dc3c5d3c474c83bfe931a9ef03bb046331 100644 (file)
@@ -744,6 +744,12 @@ i40e_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
                        break;
                }
 
+               /**
+                * Use acquire fence to ensure that qword1 which includes DD
+                * bit is loaded before loading of other descriptor words.
+                */
+               rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
+
                rxd = *rxdp;
                nb_hold++;
                rxe = &sw_ring[rx_id];
@@ -860,6 +866,12 @@ i40e_recv_scattered_pkts(void *rx_queue,
                        break;
                }
 
+               /**
+                * Use acquire fence to ensure that qword1 which includes DD
+                * bit is loaded before loading of other descriptor words.
+                */
+               rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
+
                rxd = *rxdp;
                nb_hold++;
                rxe = &sw_ring[rx_id];