-----Original Message-----
From: Thomas Monjalon <redacted>
Sent: Monday, November 22, 2021 2:28 PM
To: Rakesh Kudurumalla <redacted>
Cc: stable@dpdk.org; dev@dpdk.org; david.marchand@redhat.com;
ferruh.yigit@intel.com; andrew.rybchenko@oktetlabs.ru;
ajit.khaparde@broadcom.com; Jerin Jacob Kollanukkaran
[off-list ref]
Subject: Re: [EXT] Re: [dpdk-stable] [PATCH v2] test: avoid hang if queues are
full and Tx fails

22/11/2021 08:59, Rakesh Kudurumalla:

quoted

From: Thomas Monjalon <redacted>

quoted

20/07/2021 18:50, Rakesh Kudurumalla:

quoted

Current pmd_perf_autotest() in continuous mode tries to enqueue
MAX_TRAFFIC_BURST completely before starting the test. Some
drivers cannot accept complete MAX_TRAFFIC_BURST even though

rx+tx

quoted

quoted

quoted

desc

count

quoted

can fit it.

Which driver is failing to do so?
Why it cannot enqueue 32 packets?

Octeontx2 driver is failing to enqueue because hardware buffers are full

before test.

Why hardware buffers are full?

Hardware buffers are full because number of number of descriptors in continuous mode 
Is less than MAX_TRAFFIC_BURST, so if enque fails , there is no way hardware can drop the
Packets . pmd_per_autotest application evaluates performance after enqueueing packets
Initially.

quoted

pmd_perf_autotest() in continuous mode tries to enqueue
MAX_TRAFFIC_BURST (2048) before starting the test.

quoted

quoted

This patch changes behaviour to stop enqueuing after few retries.

If there is a real limitation, there will be issues in more places
than this test program.
I feel it should be addressed either in the driver or at ethdev level.

[...]

quoted

@@ -480,10 +483,19 @@ main_loop(__rte_unused void *args)
 			nb_tx = RTE_MIN(MAX_PKT_BURST, num);
 			nb_tx = rte_eth_tx_burst(portid, 0,
 						&tx_burst[idx], nb_tx);
+			if (nb_tx == 0)
+				retry_cnt++;
 			num -= nb_tx;
 			idx += nb_tx;
+			if (retry_cnt == MAX_RETRY_COUNT) {
+				retry_cnt = 0;
+				break;
+			}