unnecessary rx callbacks when zero packets
I noticed while looking at packet capture that currently the receive callbacks get called even if there are no packets. This seems unnecessary since if nb_rx is zero, then there are no packets to look at. My one concern is that an application could be using callbacks as some form of scheduling mechanism which would be broken.
The change would be:
lib/ethdev/rte_ethdev.h +1 -1
@@ -6077,7 +6077,7 @@ rte_eth_rx_burst(uint16_t port_id, uint16_t queue_id, nb_rx = p->rx_pkt_burst (qd, rx_pkts, nb_pkts);#ifdef RTE_ETHDEV_RXTX_CALLBACKS - { +if (nb_rx >0 ) {void *cb;/* rte_memory_order_release memory order was used when the
Quoted 8 lines
-----Original Message-----
From: Stephen Hemminger <[email protected]>
Sent: Sunday, January 7, 2024 11:37 AM
To: [email protected]
Subject: unnecessary rx callbacks when zero packets
Quoted 5 lines
I noticed while looking at packet capture that currently the receive callbacks get called even if there are no packets. This seems unnecessary since if
nb_rxis zero, then there are no packets to look at. My one concern is that an application could be using callbacks as some form of scheduling mechanism which would be broken.
Is it possible that the call back functions are maintaining statistics on zero packet polls?
Quoted 18 lines
The change would be:
lib/ethdev/rte_ethdev.h index renamed from lib/ethdev/rte_ethdev.h +0
-0
21e3a21903ec..f64bf977c46e100644lib/ethdev/rte_ethdev.h +1
-1@@ -6077,7 +6077,7 @@ rte_eth_rx_burst(uint16_t port_id, uint16_t queue_id, nb_rx = p->rx_pkt_burst (qd, rx_pkts, nb_pkts);#ifdef RTE_ETHDEV_RXTX_CALLBACKS - { +if (nb_rx >0 ) {void *cb;/* rte_memory_order_release memory order was used when the
Quoted 9 lines
From: Honnappa Nagarahalli [mailto:[email protected]]
Sent: Sunday, 7 January 2024 21.57Quoted 4 lines
From: Stephen Hemminger <[email protected]>
Sent: Sunday, January 7, 2024 11:37 AMI noticed while looking at packet capture that currently the receive
callbacks
Quoted 1 line
get called even if there are no packets. This seems unnecessary since
Quoted 8 lines
if
nb_rxisQuoted 1 line
zero, then there are no packets to look at. My one concern is that
an
Quoted 1 line
application could be using callbacks as some form of scheduling
mechanism
Quoted 1 line
which would be broken.
Is it possible that the call back functions are maintaining statistics on zero packet polls?
I agree with this concern. The primary argument for introducing the callbacks (instead of the application simply calling the same functions at RX and TX time) was to provide instrumentation outside of the application itself. And for instrumentation purposes, zero-packet calls may be relevant.
TX also calls its callback with zero packets. The callbacks treatment should be the same for both RX and TX: Either always call, or only call if non-zero packets.
So: NAK.
Perhaps the packet capture library can be optimized for zero packets instead.
Quoted 20 lines
Quoted 18 lines
The change would be:
lib/ethdev/rte_ethdev.h index renamed from lib/ethdev/rte_ethdev.h +0
-0
21e3a21903ec..f64bf977c46e100644lib/ethdev/rte_ethdev.h +1
-1@@ -6077,7 +6077,7 @@ rte_eth_rx_burst(uint16_t port_id, uint16_t queue_id, nb_rx = p->rx_pkt_burst (qd, rx_pkts, nb_pkts);#ifdef RTE_ETHDEV_RXTX_CALLBACKS - { +if (nb_rx >0 ) {void *cb;Quoted 1 line
/* rte_memory_order_release memory order was usedwhen the
rx callbacks when zero packets
Quoted 6 lines
I noticed while looking at packet capture that currently the receive callbacks get called even if there are no packets. This seems unnecessary since if
nb_rxis zero, then there are no packets to look at. My one concern is that an application could be using callbacks as some form of scheduling mechanism which would be broken.
As I remember, original idea was to allow callbacks to inject new packets if needed.
Quoted 17 lines
The change would be:
lib/ethdev/rte_ethdev.h +1
-1@@ -6077,7 +6077,7 @@ rte_eth_rx_burst(uint16_t port_id, uint16_t queue_id, nb_rx = p->rx_pkt_burst (qd, rx_pkts, nb_pkts);#ifdef RTE_ETHDEV_RXTX_CALLBACKS - { +if (nb_rx >0 ) {void *cb;/* rte_memory_order_release memory order was used when the
Quoted 11 lines · Konstantin Ananyev
rx callbacks when zero packets
Quoted 6 lines
I noticed while looking at packet capture that currently the receive callbacks get called even if there are no packets. This seems unnecessary since if
nb_rxis zero, then there are no packets to look at. My one concern is that an application could be using callbacks as some form of scheduling mechanism which would be broken.As I remember, original idea was to allow callbacks to inject new packets if needed.
Right, callback function updating 'nb_rx' enables this.
Also Honnappa has a good point that callback can be used to calculate zero packet polls.
These points are not documented and not obvious from the code, @Stephen does it make sense to document that callback function called with zero packet intentionally, in the 'rte_eth_rx_burst()' function comment?
Quoted 17 lines
Quoted 17 lines
The change would be:
lib/ethdev/rte_ethdev.h +1
-1@@ -6077,7 +6077,7 @@ rte_eth_rx_burst(uint16_t port_id, uint16_t queue_id, nb_rx = p->rx_pkt_burst (qd, rx_pkts, nb_pkts);#ifdef RTE_ETHDEV_RXTX_CALLBACKS - { +if (nb_rx >0 ) {void *cb;/* rte_memory_order_release memory order was used when the