* [PATCH] net/virtio: Fix check of threshold for Tx freeing
@ 2025-06-09 7:23 Hengqi Chen
2025-06-12 13:24 ` Maxime Coquelin
` (2 more replies)
0 siblings, 3 replies; 5+ messages in thread
From: Hengqi Chen @ 2025-06-09 7:23 UTC (permalink / raw)
To: dev; +Cc: Hengqi Chen, Baoyuan Li, Maxime Coquelin, Chenbo Xia
Like most dirvers, make the fast path of virtio_xmit_cleanup() behave as
described by the comments of rte_eth_txconf::tx_free_thresh ([0]):
Start freeing Tx buffers if there are
less free descriptors than this value.
The rationale behind this change is that:
* vq->vq_nentries is set during device probe with the queue size specified
by vhost backend, this value does not reflect the real nb_tx_desc
* the real available tx desc is set to vq->vq_free_cnt via the nb_tx_desc
param of rte_eth_tx_queue_setup() API
* so `nb_used > vq->vq_nentries - vq->vq_free_thresh` could never be true
if say nb_tx_desc=2048, vq->vq_nentries=4096, vq->vq_free_thresh=32,
see bug report 1716 ([1]) for details.
Bugzilla ID: 1716
[0]: https://github.com/DPDK/dpdk/commit/72514b5d5543
[1]: https://bugs.dpdk.org/show_bug.cgi?id=1716
Signed-off-by: Baoyuan Li <updoing@sina.com>
Signed-off-by: Hengqi Chen <hengqi.chen@gmail.com>
---
drivers/net/virtio/virtio_rxtx.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c
index edecd2011f..ab97f03d7d 100644
--- a/drivers/net/virtio/virtio_rxtx.c
+++ b/drivers/net/virtio/virtio_rxtx.c
@@ -1873,7 +1873,7 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
nb_used = virtqueue_nused(vq);
- if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
+ if (likely(vq->vq_free_cnt < vq->vq_free_thresh))
virtio_xmit_cleanup(vq, nb_used);
for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
--
2.43.5
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH] net/virtio: Fix check of threshold for Tx freeing
2025-06-09 7:23 [PATCH] net/virtio: Fix check of threshold for Tx freeing Hengqi Chen
@ 2025-06-12 13:24 ` Maxime Coquelin
2025-06-13 8:08 ` Maxime Coquelin
2025-07-04 11:29 ` Jiang, YuX
2 siblings, 0 replies; 5+ messages in thread
From: Maxime Coquelin @ 2025-06-12 13:24 UTC (permalink / raw)
To: Hengqi Chen, dev; +Cc: Baoyuan Li, Chenbo Xia
On 6/9/25 9:23 AM, Hengqi Chen wrote:
> Like most dirvers, make the fast path of virtio_xmit_cleanup() behave as
> described by the comments of rte_eth_txconf::tx_free_thresh ([0]):
> Start freeing Tx buffers if there are
> less free descriptors than this value.
>
> The rationale behind this change is that:
> * vq->vq_nentries is set during device probe with the queue size specified
> by vhost backend, this value does not reflect the real nb_tx_desc
> * the real available tx desc is set to vq->vq_free_cnt via the nb_tx_desc
> param of rte_eth_tx_queue_setup() API
> * so `nb_used > vq->vq_nentries - vq->vq_free_thresh` could never be true
> if say nb_tx_desc=2048, vq->vq_nentries=4096, vq->vq_free_thresh=32,
> see bug report 1716 ([1]) for details.
>
> Bugzilla ID: 1716
>
> [0]: https://github.com/DPDK/dpdk/commit/72514b5d5543
> [1]: https://bugs.dpdk.org/show_bug.cgi?id=1716
>
> Signed-off-by: Baoyuan Li <updoing@sina.com>
> Signed-off-by: Hengqi Chen <hengqi.chen@gmail.com>
> ---
> drivers/net/virtio/virtio_rxtx.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c
> index edecd2011f..ab97f03d7d 100644
> --- a/drivers/net/virtio/virtio_rxtx.c
> +++ b/drivers/net/virtio/virtio_rxtx.c
> @@ -1873,7 +1873,7 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
>
> nb_used = virtqueue_nused(vq);
>
> - if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
> + if (likely(vq->vq_free_cnt < vq->vq_free_thresh))
> virtio_xmit_cleanup(vq, nb_used);
>
> for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Thanks,
Maxime
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH] net/virtio: Fix check of threshold for Tx freeing
2025-06-09 7:23 [PATCH] net/virtio: Fix check of threshold for Tx freeing Hengqi Chen
2025-06-12 13:24 ` Maxime Coquelin
@ 2025-06-13 8:08 ` Maxime Coquelin
2025-07-04 11:29 ` Jiang, YuX
2 siblings, 0 replies; 5+ messages in thread
From: Maxime Coquelin @ 2025-06-13 8:08 UTC (permalink / raw)
To: Hengqi Chen, dev; +Cc: Baoyuan Li, Chenbo Xia
On 6/9/25 9:23 AM, Hengqi Chen wrote:
> Like most dirvers, make the fast path of virtio_xmit_cleanup() behave as
> described by the comments of rte_eth_txconf::tx_free_thresh ([0]):
> Start freeing Tx buffers if there are
> less free descriptors than this value.
>
> The rationale behind this change is that:
> * vq->vq_nentries is set during device probe with the queue size specified
> by vhost backend, this value does not reflect the real nb_tx_desc
> * the real available tx desc is set to vq->vq_free_cnt via the nb_tx_desc
> param of rte_eth_tx_queue_setup() API
> * so `nb_used > vq->vq_nentries - vq->vq_free_thresh` could never be true
> if say nb_tx_desc=2048, vq->vq_nentries=4096, vq->vq_free_thresh=32,
> see bug report 1716 ([1]) for details.
>
> Bugzilla ID: 1716
>
> [0]: https://github.com/DPDK/dpdk/commit/72514b5d5543
> [1]: https://bugs.dpdk.org/show_bug.cgi?id=1716
>
> Signed-off-by: Baoyuan Li <updoing@sina.com>
> Signed-off-by: Hengqi Chen <hengqi.chen@gmail.com>
> ---
> drivers/net/virtio/virtio_rxtx.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c
> index edecd2011f..ab97f03d7d 100644
> --- a/drivers/net/virtio/virtio_rxtx.c
> +++ b/drivers/net/virtio/virtio_rxtx.c
> @@ -1873,7 +1873,7 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
>
> nb_used = virtqueue_nused(vq);
>
> - if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
> + if (likely(vq->vq_free_cnt < vq->vq_free_thresh))
> virtio_xmit_cleanup(vq, nb_used);
>
> for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
Applied to next-virtio/for-net-next
Thanks,
Maxime
^ permalink raw reply [flat|nested] 5+ messages in thread
* RE: [PATCH] net/virtio: Fix check of threshold for Tx freeing
2025-06-09 7:23 [PATCH] net/virtio: Fix check of threshold for Tx freeing Hengqi Chen
2025-06-12 13:24 ` Maxime Coquelin
2025-06-13 8:08 ` Maxime Coquelin
@ 2025-07-04 11:29 ` Jiang, YuX
2025-07-04 11:57 ` Maxime Coquelin
2 siblings, 1 reply; 5+ messages in thread
From: Jiang, YuX @ 2025-07-04 11:29 UTC (permalink / raw)
To: Hengqi Chen, dev; +Cc: Baoyuan Li, Maxime Coquelin, Chenbo Xia
Hi Hengqi,
Can you help to check the ticket: https://bugs.dpdk.org/show_bug.cgi?id=1747? The perf drop is caused by your patch. Thanks so much.
Best regards,
Yu Jiang
> -----Original Message-----
> From: Hengqi Chen <hengqi.chen@gmail.com>
> Sent: Monday, June 9, 2025 3:24 PM
> To: dev@dpdk.org
> Cc: Hengqi Chen <hengqi.chen@gmail.com>; Baoyuan Li
> <updoing@sina.com>; Maxime Coquelin <maxime.coquelin@redhat.com>;
> Chenbo Xia <chenbox@nvidia.com>
> Subject: [PATCH] net/virtio: Fix check of threshold for Tx freeing
>
> Like most dirvers, make the fast path of virtio_xmit_cleanup() behave as
> described by the comments of rte_eth_txconf::tx_free_thresh ([0]):
> Start freeing Tx buffers if there are
> less free descriptors than this value.
>
> The rationale behind this change is that:
> * vq->vq_nentries is set during device probe with the queue size specified
> by vhost backend, this value does not reflect the real nb_tx_desc
> * the real available tx desc is set to vq->vq_free_cnt via the nb_tx_desc
> param of rte_eth_tx_queue_setup() API
> * so `nb_used > vq->vq_nentries - vq->vq_free_thresh` could never be true
> if say nb_tx_desc=2048, vq->vq_nentries=4096, vq->vq_free_thresh=32,
> see bug report 1716 ([1]) for details.
>
> Bugzilla ID: 1716
>
> [0]: https://github.com/DPDK/dpdk/commit/72514b5d5543
> [1]: https://bugs.dpdk.org/show_bug.cgi?id=1716
>
> Signed-off-by: Baoyuan Li <updoing@sina.com>
> Signed-off-by: Hengqi Chen <hengqi.chen@gmail.com>
> ---
> drivers/net/virtio/virtio_rxtx.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c
> index edecd2011f..ab97f03d7d 100644
> --- a/drivers/net/virtio/virtio_rxtx.c
> +++ b/drivers/net/virtio/virtio_rxtx.c
> @@ -1873,7 +1873,7 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf
> **tx_pkts, uint16_t nb_pkts)
>
> nb_used = virtqueue_nused(vq);
>
> - if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
> + if (likely(vq->vq_free_cnt < vq->vq_free_thresh))
> virtio_xmit_cleanup(vq, nb_used);
>
> for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
> --
> 2.43.5
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH] net/virtio: Fix check of threshold for Tx freeing
2025-07-04 11:29 ` Jiang, YuX
@ 2025-07-04 11:57 ` Maxime Coquelin
0 siblings, 0 replies; 5+ messages in thread
From: Maxime Coquelin @ 2025-07-04 11:57 UTC (permalink / raw)
To: Jiang, YuX, Hengqi Chen, dev, Thomas Monjalon, David Marchand
Cc: Baoyuan Li, Chenbo Xia
Hi Yu, Hengqi,
First, thanks Yu for testing and reporting the performance degradation,
this is much appreciated!
On 7/4/25 1:29 PM, Jiang, YuX wrote:
> Hi Hengqi,
>
> Can you help to check the ticket: https://bugs.dpdk.org/show_bug.cgi?id=1747? The perf drop is caused by your patch. Thanks so much.
The options I see are:
1. Revert the patch for v25.07, fix it and resubmit for v25.11.
2. Fix it on time for v25.07-rc3
3. Keep the patch in v25.07 and try to fix it in v25.11.
As I will be OoO starting this evening, I would be in favor of option 1.
What do you think?
Thanks,
Maxime
> Best regards,
> Yu Jiang
>
>> -----Original Message-----
>> From: Hengqi Chen <hengqi.chen@gmail.com>
>> Sent: Monday, June 9, 2025 3:24 PM
>> To: dev@dpdk.org
>> Cc: Hengqi Chen <hengqi.chen@gmail.com>; Baoyuan Li
>> <updoing@sina.com>; Maxime Coquelin <maxime.coquelin@redhat.com>;
>> Chenbo Xia <chenbox@nvidia.com>
>> Subject: [PATCH] net/virtio: Fix check of threshold for Tx freeing
>>
>> Like most dirvers, make the fast path of virtio_xmit_cleanup() behave as
>> described by the comments of rte_eth_txconf::tx_free_thresh ([0]):
>> Start freeing Tx buffers if there are
>> less free descriptors than this value.
>>
>> The rationale behind this change is that:
>> * vq->vq_nentries is set during device probe with the queue size specified
>> by vhost backend, this value does not reflect the real nb_tx_desc
>> * the real available tx desc is set to vq->vq_free_cnt via the nb_tx_desc
>> param of rte_eth_tx_queue_setup() API
>> * so `nb_used > vq->vq_nentries - vq->vq_free_thresh` could never be true
>> if say nb_tx_desc=2048, vq->vq_nentries=4096, vq->vq_free_thresh=32,
>> see bug report 1716 ([1]) for details.
>>
>> Bugzilla ID: 1716
>>
>> [0]: https://github.com/DPDK/dpdk/commit/72514b5d5543
>> [1]: https://bugs.dpdk.org/show_bug.cgi?id=1716
>>
>> Signed-off-by: Baoyuan Li <updoing@sina.com>
>> Signed-off-by: Hengqi Chen <hengqi.chen@gmail.com>
>> ---
>> drivers/net/virtio/virtio_rxtx.c | 2 +-
>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c
>> index edecd2011f..ab97f03d7d 100644
>> --- a/drivers/net/virtio/virtio_rxtx.c
>> +++ b/drivers/net/virtio/virtio_rxtx.c
>> @@ -1873,7 +1873,7 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf
>> **tx_pkts, uint16_t nb_pkts)
>>
>> nb_used = virtqueue_nused(vq);
>>
>> - if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
>> + if (likely(vq->vq_free_cnt < vq->vq_free_thresh))
>> virtio_xmit_cleanup(vq, nb_used);
>>
>> for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
>> --
>> 2.43.5
>
^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2025-07-04 11:57 UTC | newest]
Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2025-06-09 7:23 [PATCH] net/virtio: Fix check of threshold for Tx freeing Hengqi Chen
2025-06-12 13:24 ` Maxime Coquelin
2025-06-13 8:08 ` Maxime Coquelin
2025-07-04 11:29 ` Jiang, YuX
2025-07-04 11:57 ` Maxime Coquelin
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).