On 2023/6/8 2:12, Ferruh Yigit wrote:
On 5/29/2023 3:26 AM, Jie Hai wrote:
Each stream has a read-only "disabled" field that control if this
stream should be used to forward. This field depends on states
of Rx/Tx queues, please see
commit 3c4426db54fc ("app/testpmd: do not poll stopped queues").
Currently, the testpmd and DPDK frameworks maintain queue state
separately. That of the primary process of testpmd are set by
deferred_start in the queue configuration. And that of the
framework(dev->data->rx_queue_state or dev->data->tx_queue_state)
is set when the driver enables/disables the queue, and it is
shared between the primary/secondary process.
If the deferred_start is set, the queue is disabled and the
corresponding queue state in the framework changes to stopped.
However, the queue state in the framework does not only come from
this. If the primary/secondary process stops a queue, the related
queue state will change, too. However, the primary process of
testpmd does not know the change brought by this operation.
Therefore, setting the queue state in the primary testpmd by only
the deferred_start is unsafe.
For example, Rx/Tx queues who are stopped before the operations of
stopping and starting port cannot forward packets after these
operations on primary process.
Therefore, the primary process should getting the queue state from
of the framework as the secondary process does, please see commit
e065c9aa3e05 ("app/testpmd: fix secondary process packet forwarding").
Fixes: 3c4426db54fc ("app/testpmd: do not poll stopped queues")
Cc: sta...@dpdk.org
Signed-off-by: Jie Hai <haij...@huawei.com>
---
v1->v2:
1. Fix misspelled word 'deferred'.
2. Fix incorrect format of reference to commits.
v2->v3
1. Fix incorrect format of reference to commits.
---
app/test-pmd/testpmd.c | 19 +++----------------
1 file changed, 3 insertions(+), 16 deletions(-)
diff --git a/app/test-pmd/testpmd.c b/app/test-pmd/testpmd.c
index 5cb6f9252395..a07a67a2639e 100644
--- a/app/test-pmd/testpmd.c
+++ b/app/test-pmd/testpmd.c
@@ -2502,8 +2502,7 @@ start_packet_forwarding(int with_tx_first)
return;
if (stream_init != NULL) {
- if (rte_eal_process_type() == RTE_PROC_SECONDARY)
- update_queue_state();
+ update_queue_state();
for (i = 0; i < cur_fwd_config.nb_fwd_streams; i++)
stream_init(fwd_streams[i]);
}
@@ -2860,9 +2859,6 @@ rx_queue_setup(uint16_t port_id, uint16_t rx_queue_id,
socket_id, rx_conf, mp);
}
- ports[port_id].rxq[rx_queue_id].state = rx_conf->rx_deferred_start ?
- RTE_ETH_QUEUE_STATE_STOPPED :
- RTE_ETH_QUEUE_STATE_STARTED;
return ret;
}
@@ -3129,9 +3125,6 @@ start_port(portid_t pid)
port->need_reconfig_queues = 0;
/* setup tx queues */
for (qi = 0; qi < nb_txq; qi++) {
- struct rte_eth_txconf *conf =
- &port->txq[qi].conf;
-
if ((numa_support) &&
(txring_numa[pi] != NUMA_NO_CONFIG))
diag = rte_eth_tx_queue_setup(pi, qi,
@@ -3144,13 +3137,8 @@ start_port(portid_t pid)
port->socket_id,
&(port->txq[qi].conf));
- if (diag == 0) {
- port->txq[qi].state =
- conf->tx_deferred_start ?
- RTE_ETH_QUEUE_STATE_STOPPED :
- RTE_ETH_QUEUE_STATE_STARTED;
+ if (diag == 0)
continue;
- }
/* Fail to setup tx queue, return */
if (port->port_status == RTE_PORT_HANDLING)
@@ -3266,8 +3254,7 @@ start_port(portid_t pid)
pl[cfg_pi++] = pi;
}
- if (rte_eal_process_type() == RTE_PROC_SECONDARY)
- update_queue_state();
+ update_queue_state();
As you described, issue seems caused by commit [1] using testpmd local
queue state information to decide to forward or not on a stream.
There are other commands that update ethdev queue state, this is causing
testpmd and ethdev state information to diverge and causes unexpected
side effects.
For the sample you provided, just above change (that updates queue state
before port start) should be sufficient,
also I guess it is OK to update state in 'start_packet_forwarding()' to
be on safe side.
But not sure about to update deferred_start related changes in this
patch, also details about it makes commit log complex to understand,
what do you think to split deferred_start related changes to another patch?
I think I should drop the changes on deferred_start. It may cause PMDs
supporting 'deferred_start' but not supporting
'rte_eth_tx_queue_info_get' or 'rte_eth_rx_queue_info_get' takes no
effect when set 'deferred_start' on.
A better way is to check when updating the queue state. If this is the
case, keep the testpmd state to the original value so that the original
behavior is not changed.
In commit log, can you please refer to the patch that fixes similar
issue for secondary process [2], that is relevant with this fix,
reference may help to understand this issue better.
Ok. I will.
@Shiyang, can you please test secondary process fix this patch, to be
sure this is not breaking it again?
[1]
Fixes: 3c4426db54fc ("app/testpmd: do not poll stopped queues")
[2]
5028f207a4fa ("app/testpmd: fix secondary process packet forwarding")
.
