From: Konstantin Ananyev <konstantin.anan...@huawei.com>

Running l3fwd in event mode with SW eventdev, service cores
can start RX before main thread is finished with PMD installation.
to reproduce:
./dpdk-l3fwd --lcores=49,51 -n 6 -a ca:00.0 -s 0x8000000000000 \
--vdev event_sw0 -- \
-L -P -p 1  --mode eventdev --eventq-sched=ordered \
--rule_ipv4=test/l3fwd_lpm_v4_u1.cfg --rule_ipv6=test/l3fwd_lpm_v6_u1.cfg \
--no-numa

At init stage user will most likely see the error message like that:
ETHDEV: lcore 51 called rx_pkt_burst for not ready port 0
0: ./dpdk-l3fwd (rte_dump_stack+0x1f) [15de723]
...
9: ./dpdk-l3fwd (eal_thread_loop+0x5a2) [15c1324]
...

And then all depends how luck/unlucky you are.
If there are some actual packet in HW RX queue, then the app will most
likely crash, otherwise it might survive.
As error message suggests, the problem is that services are started
before main thread finished with NIC setup and initialization.
The suggested fix moves services startup after NIC setup phase.

Bugzilla ID: 1390
Fixes: 8bd537e9c6cf ("examples/l3fwd: add service core setup based on caps")
Cc: sta...@dpdk.org

Signed-off-by: Konstantin Ananyev <konstantin.v.anan...@yandex.ru>
Signed-off-by: Konstantin Ananyev <konstantin.anan...@huawei.com>
---
 examples/l3fwd/main.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/examples/l3fwd/main.c b/examples/l3fwd/main.c
index 3bf28aec0c..d4fb5d1971 100644
--- a/examples/l3fwd/main.c
+++ b/examples/l3fwd/main.c
@@ -1577,7 +1577,6 @@ main(int argc, char **argv)
                        l3fwd_lkp.main_loop = evt_rsrc->ops.fib_event_loop;
                else
                        l3fwd_lkp.main_loop = evt_rsrc->ops.lpm_event_loop;
-               l3fwd_event_service_setup();
        } else
 #endif
                l3fwd_poll_resource_setup();
@@ -1609,6 +1608,11 @@ main(int argc, char **argv)
                }
        }
 
+#ifdef RTE_LIB_EVENTDEV
+       if (evt_rsrc->enabled)
+               l3fwd_event_service_setup();
+#endif
+
        printf("\n");
 
        for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
-- 
2.35.3

Reply via email to