AF_XDP Preferred Busy Polling

Prior to this the max size was 32 which was unnecessarily small.

Also enforce the max batch size for TX for both copy and zero copy modes. Prior to this only copy mode enforced the max size.

Signed-off-by Ciara Loftus <[email protected]>

Quoted 2 lines

Prior to this the max size was 32 which was unnecessarily small.

Can you please describe the impact? Why changed from 32 to 512? I assume this is to improve the performance but can you please explicitly document it in the commit log?

Quoted 5 lines

Also enforce the max batch size for TX for both copy and zero copy modes. Prior to this only copy mode enforced the max size.

By enforcing, the PMD ignores the user provided burst value if it is more than PMS supported MAX, and this ignoring is done in silent. Also there is no way to
discover this MAX value without checking the code.

Overall, why this max values are required at all? After quick check I can see they are used for some bulk operations, which I assume can be eliminated, what
do you think?

Quoted 2 lines

Signed-off-by Ciara Loftus <[email protected]>

<...>

Quoted 8 lines · Ciara Loftus

Quoted 2 lines

Prior to this the max size was 32 which was unnecessarily small.

Can you please describe the impact? Why changed from 32 to 512? I assume this is to improve the performance but can you please explicitly document it in the commit log?

Indeed - improved performance due to bulk operations and fewer ring accesses and syscalls. The value 512 was arbitrary. I will change this to the default ring size as defined by libbpf (2048) in v2. Will update the commit log with this info.

Quoted 11 lines

Quoted 5 lines

Also enforce the max batch size for TX for both copy and zero copy modes. Prior to this only copy mode enforced the max size.

By enforcing, the PMD ignores the user provided burst value if it is more than PMS supported MAX, and this ignoring is done in silent. Also there is no way to
discover this MAX value without checking the code.

Quoted 4 lines

Overall, why this max values are required at all? After quick check I can see they are used for some bulk operations, which I assume can be eliminated, what
do you think?

We need to size some arrays at compile time with this max value.

Instead of removing the bulk operations which may impact performance, how about taking an approach where we split batches that are > 2048 into smaller batches and still handle all the packets instead of discarding those > 2048. Something like what's done in ixgbe for example:
http://code.dpdk.org/dpdk/v21.02/source/drivers/net/ixgbe/ixgbe_rxtx.c#L318
`

This commit introduces support for preferred busy polling to the AF_XDP PMD. This feature aims to improve single-core performance for AF_XDP sockets under heavy load.

A new vdev arg is introduced called 'busy_budget' whose default value is 64. busy_budget is the value supplied to the kernel with the SO_BUSY_POLL_BUDGET socket option and represents the busy-polling NAPI budget. To set the budget to a different value eg. 256:

--vdev=net_af_xdp0,iface=eth0,busy_budget=256

Preferred busy polling is enabled by default provided a kernel with version >= v5.11 is in use. To disable it, set the budget to zero.

The following settings are also strongly recommended to be used in conjunction with this feature:

echo 2 | sudo tee /sys/class/net/eth0/napi_defer_hard_irqs echo 200000 | sudo tee /sys/class/net/eth0/gro_flush_timeout

.. where eth0 is the interface being used by the PMD.

Signed-off-by Ciara Loftus <[email protected]>

---
  doc/guides/nics/af_xdp.rst          | 38 ++++++++++++-
  drivers/net/af_xdp/compat.h         | 13 +++++
  drivers/net/af_xdp/rte_eth_af_xdp.c | 85 ++++++++++++++++++++++++-----
  3 files changed, 121 insertions(+), 15 deletions(-)

drivers/net/af_xdp/compat.h +2 -0

@@ -39,3 +39,16 @@ create_shared_socket(struct xsk_socket **xsk_ptr __rte_unused,
  	return -1;
  }
  #endif
+
+#ifdef XDP_USE_NEED_WAKEUP

Quoted 21 lines

This commit introduces support for preferred busy polling to the AF_XDP PMD. This feature aims to improve single-core performance for AF_XDP sockets under heavy load.

A new vdev arg is introduced called 'busy_budget' whose default value is 64. busy_budget is the value supplied to the kernel with the SO_BUSY_POLL_BUDGET socket option and represents the busy-polling NAPI budget. To set the budget to a different value eg. 256:

--vdev=net_af_xdp0,iface=eth0,busy_budget=256

Preferred busy polling is enabled by default provided a kernel with version >= v5.11 is in use. To disable it, set the budget to zero.

The following settings are also strongly recommended to be used in conjunction with this feature:

echo 2 | sudo tee /sys/class/net/eth0/napi_defer_hard_irqs echo 200000 | sudo tee /sys/class/net/eth0/gro_flush_timeout

Quoted 7 lines

.. where eth0 is the interface being used by the PMD.

Signed-off-by Ciara Loftus <[email protected]>

---
  doc/guides/nics/af_xdp.rst          | 38 ++++++++++++-
  drivers/net/af_xdp/compat.h         | 13 +++++
  drivers/net/af_xdp/rte_eth_af_xdp.c | 85

++++++++++++++++++++++++-----

Quoted 1 line

  3 files changed, 121 insertions(+), 15 deletions(-)

Can you please update the release notes too to announce the feature?

<...>

Review create_shared_socket(struct xsk_socket **xsk_ptr

Quoted 1 line

(quoted diff) +0 -0

@@ -39,3 +39,16 @@ create_shared_socket(struct xsk_socket **xsk_ptr

__rte_unused,

drivers/net/af_xdp/compat.h create_shared_socket(struct xsk_socket **xsk_ptr __rte_unused,

Quoted 12 lines

drivers/net/af_xdp/compat.h +9 -0

  	return -1;
  }
  #endif
+
+#ifdef XDP_USE_NEED_WAKEUP
+static int
+syscall_needed(struct xsk_ring_prod *q, uint32_t busy_budget)
+{
+	return xsk_ring_prod__needs_wakeup(q) | busy_budget;
+}
+#else
+syscall_needed(struct xsk_ring_prod *q __rte_unused, uint32_t

busy_budget)

drivers/net/af_xdp/compat.h create_shared_socket(struct xsk_socket **xsk_ptr __rte_unused, · line 53

Quoted 3 lines

drivers/net/af_xdp/compat.h +3 -0

+{
+	return busy_budget;
+}

Is the return type missing in the definition?

Also for the case when both 'XDP_USE_NEED_WAKEUP' & 'SO_PREFER_BUSY_POLL' this
function will always return '0', but current implementation doesn't know this in
the compile time and compiler can't optimize for it, do you think does it make sense to do this optimization?

<...>

Review rte_pmd_af_xdp_probe(struct rte_vdev_device

Quoted 1 line

(quoted diff) +0 -0

@@ -1628,8 +1670,22 @@ rte_pmd_af_xdp_probe(struct rte_vdev_device

*dev)

drivers/net/af_xdp/rte_eth_af_xdp.c rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)

Quoted 5 lines

drivers/net/af_xdp/rte_eth_af_xdp.c +2 -0

  		return -EINVAL;
  	}
 
+#ifdef SO_PREFER_BUSY_POLL
+	busy_budget = busy_budget == -1 ?

ETH_AF_XDP_DFLT_BUSY_BUDGET :

drivers/net/af_xdp/rte_eth_af_xdp.c rte_pmd_af_xdp_probe(struct rte_vdev_device *dev) · line 1677

Quoted 3 lines

drivers/net/af_xdp/rte_eth_af_xdp.c +3 -0

+					busy_budget;
+	if (!busy_budget)
+		AF_XDP_LOG(ERR, "Preferred busy polling disabled\n");

Is this an error case? What do you think changing the log level to DEBUG or INFO?

Also how these compile time flags will work if the compiled environment and run
environment kernel version are different and incompatible?

Overall can it be possible to detect the support on runtime via 'setsockopt()' without compile time macros and eliminate the compile time flags? Does it make
sense?

Quoted 30 lines · Ciara Loftus

Quoted 21 lines

This commit introduces support for preferred busy polling to the AF_XDP PMD. This feature aims to improve single-core performance for AF_XDP sockets under heavy load.

A new vdev arg is introduced called 'busy_budget' whose default value is 64. busy_budget is the value supplied to the kernel with the SO_BUSY_POLL_BUDGET socket option and represents the busy-polling NAPI budget. To set the budget to a different value eg. 256:

--vdev=net_af_xdp0,iface=eth0,busy_budget=256

Preferred busy polling is enabled by default provided a kernel with version >= v5.11 is in use. To disable it, set the budget to zero.

The following settings are also strongly recommended to be used in conjunction with this feature:

echo 2 | sudo tee /sys/class/net/eth0/napi_defer_hard_irqs echo 200000 | sudo tee /sys/class/net/eth0/gro_flush_timeout

Quoted 7 lines

.. where eth0 is the interface being used by the PMD.

Signed-off-by Ciara Loftus <[email protected]>

---
  doc/guides/nics/af_xdp.rst          | 38 ++++++++++++-
  drivers/net/af_xdp/compat.h         | 13 +++++
  drivers/net/af_xdp/rte_eth_af_xdp.c | 85

Quoted 4 lines

++++++++++++++++++++++++-----

Quoted 1 line

  3 files changed, 121 insertions(+), 15 deletions(-)

Can you please update the release notes too to announce the feature?

Will do.

Quoted 19 lines

<...>

Review create_shared_socket(struct xsk_socket **xsk_ptr

Quoted 1 line

(quoted diff) +0 -0

@@ -39,3 +39,16 @@ create_shared_socket(struct xsk_socket **xsk_ptr

__rte_unused,

drivers/net/af_xdp/compat.h create_shared_socket(struct xsk_socket **xsk_ptr __rte_unused,

Quoted 12 lines

drivers/net/af_xdp/compat.h +9 -0

  	return -1;
  }
  #endif
+
+#ifdef XDP_USE_NEED_WAKEUP
+static int
+syscall_needed(struct xsk_ring_prod *q, uint32_t busy_budget)
+{
+	return xsk_ring_prod__needs_wakeup(q) | busy_budget;
+}
+#else
+syscall_needed(struct xsk_ring_prod *q __rte_unused, uint32_t

busy_budget)

Quoted 5 lines

drivers/net/af_xdp/compat.h create_shared_socket(struct xsk_socket **xsk_ptr __rte_unused, · line 53

Quoted 3 lines

drivers/net/af_xdp/compat.h +3 -0

+{
+	return busy_budget;
+}

Is the return type missing in the definition?

Yes. Thanks for spotting this.

Quoted 4 lines

Also for the case when both 'XDP_USE_NEED_WAKEUP' & 'SO_PREFER_BUSY_POLL' this
function will always return '0', but current implementation doesn't know

this

Quoted 2 lines

in
the compile time and compiler can't optimize for it, do you think does it

make

Quoted 1 line

sense to do this optimization?

It makes sense assuming the compile environment and run environment are the same.
However you make a valid point below. If the environments are different, we can't make this optimization because we can’t rely on the presence of the flags alone to tell us if these features are supported. More below.

Quoted 4 lines

<...>

Review rte_pmd_af_xdp_probe(struct

Quoted 1 line

(quoted diff) +0 -0

@@ -1628,8 +1670,22 @@ rte_pmd_af_xdp_probe(struct

rte_vdev_device

Quoted 7 lines

*dev)

drivers/net/af_xdp/rte_eth_af_xdp.c rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)

Quoted 5 lines

drivers/net/af_xdp/rte_eth_af_xdp.c +2 -0

  		return -EINVAL;
  	}
 
+#ifdef SO_PREFER_BUSY_POLL
+	busy_budget = busy_budget == -1 ?

ETH_AF_XDP_DFLT_BUSY_BUDGET :

Quoted 6 lines

drivers/net/af_xdp/rte_eth_af_xdp.c rte_pmd_af_xdp_probe(struct rte_vdev_device *dev) · line 1677

Quoted 3 lines

drivers/net/af_xdp/rte_eth_af_xdp.c +3 -0

+					busy_budget;
+	if (!busy_budget)
+		AF_XDP_LOG(ERR, "Preferred busy polling disabled\n");

Is this an error case? What do you think changing the log level to DEBUG or INFO?

+1 for INFO

Quoted 2 lines

Also how these compile time flags will work if the compiled environment

and

Quoted 2 lines

run
environment kernel version are different and incompatible?

This is a valid point. Right now if XDP_USE_NEED_WAKEUP is defined we assume the functionality is available in the kernel. If it's not, socket creation will fail and we abort. Perhaps we should retry socket creation without the flag if we get this failure. And record if support is available in a runtime variable. I'll look at adding this as another patch to the v2 series.

Quoted 3 lines

Overall can it be possible to detect the support on runtime via

'setsockopt()'

Quoted 3 lines

without compile time macros and eliminate the compile time flags? Does it make
sense?

I think this can be done. It should allow applications compiled on older kernels without SO_PREFER_BUSY_POLL run on newer kernels with the feature.
I will tackle this in the v2.

Thanks for your feedback!

Ciara

Prior to this commit, the maximum batch sizes for zero-copy and copy-mode rx and copy-mode tx were set to 32. Apart from zero-copy tx, the user could never rx/tx any more than 32 packets at a time and without inspecting the code the user wouldn't be aware of this.

This commit removes these upper limits placed on the user and instead sets an internal batch size equal to the default ring size (2048). Batches larger than this are still processed, however they are split into smaller batches similar to how it's done in other drivers. This is necessary because some arrays used during rx/tx need to be sized at compile-time.

Allowing a larger batch size allows for fewer batches and thus larger bulk operations, fewer ring accesses and fewer syscalls which should yield improved performance.

Signed-off-by Ciara Loftus <[email protected]>

---
  drivers/net/af_xdp/rte_eth_af_xdp.c | 67 ++++++++++++++++++++++++-----
  1 file changed, 57 insertions(+), 10 deletions(-)

drivers/net/af_xdp/rte_eth_af_xdp.c +0 -0

index 3957227bf0..be524e4784 100644

Quoted 1 line

Prior to this commit, the maximum batch sizes for zero-copy and copy-

mode

Quoted 2 lines

rx and copy-mode tx were set to 32. Apart from zero-copy tx, the user could never rx/tx any more than 32 packets at a time and without

inspecting

Quoted 8 lines

the code the user wouldn't be aware of this.

This commit removes these upper limits placed on the user and instead sets an internal batch size equal to the default ring size (2048). Batches larger than this are still processed, however they are split into smaller batches similar to how it's done in other drivers. This is necessary because some arrays used during rx/tx need to be sized at compile-time.

Quoted 7 lines

Allowing a larger batch size allows for fewer batches and thus larger bulk operations, fewer ring accesses and fewer syscalls which should yield improved performance.

Signed-off-by Ciara Loftus <[email protected]>

---
  drivers/net/af_xdp/rte_eth_af_xdp.c | 67

++++++++++++++++++++++++-----

Review

Quoted 3 lines

  1 file changed, 57 insertions(+), 10 deletions(-)

(quoted diff) +0 -0

b/drivers/net/af_xdp/rte_eth_af_xdp.c

drivers/net/af_xdp/rte_eth_af_xdp.c RTE_LOG_REGISTER(af_xdp_logtype,

Quoted 4 lines

index 3957227bf0..be524e4784 100644

drivers/net/af_xdp/rte_eth_af_xdp.c +0 -0

@@ -66,8 +66,8 @@ RTE_LOG_REGISTER(af_xdp_logtype,

pmd.net.af_xdp, NOTICE);

drivers/net/af_xdp/rte_eth_af_xdp.c RTE_LOG_REGISTER(af_xdp_logtype, pmd.net.af_xdp, NOTICE);

Quoted 6 lines

drivers/net/af_xdp/rte_eth_af_xdp.c +1 -2

  #define ETH_AF_XDP_DFLT_START_QUEUE_IDX	0
  #define ETH_AF_XDP_DFLT_QUEUE_COUNT	1
 
-#define ETH_AF_XDP_RX_BATCH_SIZE	32
-#define ETH_AF_XDP_TX_BATCH_SIZE	32
+#define ETH_AF_XDP_RX_BATCH_SIZE

XSK_RING_CONS__DEFAULT_NUM_DESCS

drivers/net/af_xdp/rte_eth_af_xdp.c RTE_LOG_REGISTER(af_xdp_logtype, · line 70

Quoted 1 line

drivers/net/af_xdp/rte_eth_af_xdp.c +1 -0

+#define ETH_AF_XDP_TX_BATCH_SIZE

XSK_RING_CONS__DEFAULT_NUM_DESCS

Quoted 1 line

Just to double check, can there be a library version that these macros not defined, should it be checked?

This commit introduces support for preferred busy polling to the AF_XDP PMD. This feature aims to improve single-core performance for AF_XDP sockets under heavy load.

A new vdev arg is introduced called 'busy_budget' whose default value is 64. busy_budget is the value supplied to the kernel with the SO_BUSY_POLL_BUDGET socket option and represents the busy-polling NAPI budget. To set the budget to a different value eg. 256:

--vdev=net_af_xdp0,iface=eth0,busy_budget=256

Preferred busy polling is enabled by default provided a kernel with version >= v5.11 is in use. To disable it, set the budget to zero.

The following settings are also strongly recommended to be used in conjunction with this feature:

echo 2 | sudo tee /sys/class/net/eth0/napi_defer_hard_irqs echo 200000 | sudo tee /sys/class/net/eth0/gro_flush_timeout

.. where eth0 is the interface being used by the PMD.

Signed-off-by Ciara Loftus <[email protected]>

doc/guides/rel_notes/release_21_05.rst +9 -0

+static int
+configure_preferred_busy_poll(struct pkt_rx_queue *rxq)
+{
+	int sock_opt = 1;
+	int fd = xsk_socket__fd(rxq->xsk);
+	int ret = 0;
+
+	ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
+			(void *)&sock_opt, sizeof(sock_opt));

doc/guides/rel_notes/release_21_05.rst +31 -0

+	} else {
+		AF_XDP_LOG(INFO, "Busy polling budget set to: %u\n",
+					rxq->busy_budget);
+		return 0;
+	}
+
+	/* setsockopt failure - attempt to restore xsk to default state and
+	 * proceed without busy polling support.
+	 */
+	sock_opt = 0;
+	ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void *)&sock_opt,
+			sizeof(sock_opt));
+	if (ret < 0) {
+		AF_XDP_LOG(ERR, "Failed to unset SO_BUSY_POLL\n");
+		return -1;
+	}
+
+err_timeout:
+	sock_opt = 0;
+	ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
+			(void *)&sock_opt, sizeof(sock_opt));
+	if (ret < 0) {
+		AF_XDP_LOG(ERR, "Failed to unset SO_PREFER_BUSY_POLL\n");
+		return -1;
+	}
+
+err_prefer:
+	rxq->busy_budget = 0;
+	return 0;
+}
+

Quoted 24 lines

This commit introduces support for preferred busy polling to the AF_XDP PMD. This feature aims to improve single-core performance for AF_XDP sockets under heavy load.

A new vdev arg is introduced called 'busy_budget' whose default value is 64. busy_budget is the value supplied to the kernel with the SO_BUSY_POLL_BUDGET socket option and represents the busy-polling NAPI budget. To set the budget to a different value eg. 256:

--vdev=net_af_xdp0,iface=eth0,busy_budget=256

Preferred busy polling is enabled by default provided a kernel with version >= v5.11 is in use. To disable it, set the budget to zero.

The following settings are also strongly recommended to be used in conjunction with this feature:

echo 2 | sudo tee /sys/class/net/eth0/napi_defer_hard_irqs echo 200000 | sudo tee /sys/class/net/eth0/gro_flush_timeout

.. where eth0 is the interface being used by the PMD.

Signed-off-by Ciara Loftus <[email protected]>

<...>

doc/guides/rel_notes/release_21_05.rst New Features · line 77

Quoted 11 lines

doc/guides/rel_notes/release_21_05.rst +4 -0

@@ -70,6 +70,10 @@ New Features
    * Added command to display Rx queue used descriptor count.
      ``show port (port_id) rxq (queue_id) desc used count``
 
+* **Updated the AF_XDP driver.**
+
+  * Added support for preferred busy polling.
+
 

Can you please move the update after above the testpmd updates? For more details the expected order is in the section comment.

doc/guides/rel_notes/release_21_05.rst New Features

Quoted 11 lines

doc/guides/rel_notes/release_21_05.rst +11 -0

+static int
+configure_preferred_busy_poll(struct pkt_rx_queue *rxq)
+{
+	int sock_opt = 1;
+	int fd = xsk_socket__fd(rxq->xsk);
+	int ret = 0;
+
+	ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
+			(void *)&sock_opt, sizeof(sock_opt));
+	if (ret < 0) {
+		AF_XDP_LOG(DEBUG, "Failed to set

SO_PREFER_BUSY_POLL\n");

doc/guides/rel_notes/release_21_05.rst

Quoted 5 lines

doc/guides/rel_notes/release_21_05.rst +5 -0

+		goto err_prefer;
+	}
+
+	sock_opt = ETH_AF_XDP_DFLT_BUSY_TIMEOUT;
+	ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void

*)&sock_opt,

doc/guides/rel_notes/release_21_05.rst

Quoted 3 lines

doc/guides/rel_notes/release_21_05.rst +3 -0

+			sizeof(sock_opt));
+	if (ret < 0) {
+		AF_XDP_LOG(DEBUG, "Failed to set SO_BUSY_POLL\n");

[1]

doc/guides/rel_notes/release_21_05.rst

Quoted 8 lines

doc/guides/rel_notes/release_21_05.rst +8 -0

+		goto err_timeout;
+	}
+
+	sock_opt = rxq->busy_budget;
+	ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL_BUDGET,
+			(void *)&sock_opt, sizeof(sock_opt));
+	if (ret < 0) {
+		AF_XDP_LOG(DEBUG, "Failed to set

SO_BUSY_POLL_BUDGET\n");

In above [1] and here, shouldn't the function return error, even the rollback is
successful.
I am thinking a case an invalid 'busy_budget' provided, like a very big number or negative value.

doc/guides/rel_notes/release_21_05.rst

Quoted 12 lines

doc/guides/rel_notes/release_21_05.rst +11 -0

+	} else {
+		AF_XDP_LOG(INFO, "Busy polling budget set to: %u\n",
+					rxq->busy_budget);
+		return 0;
+	}
+
+	/* setsockopt failure - attempt to restore xsk to default state and
+	 * proceed without busy polling support.
+	 */
+	sock_opt = 0;
+	ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void

*)&sock_opt,

doc/guides/rel_notes/release_21_05.rst

Quoted 12 lines

doc/guides/rel_notes/release_21_05.rst +12 -0

+			sizeof(sock_opt));
+	if (ret < 0) {
+		AF_XDP_LOG(ERR, "Failed to unset SO_BUSY_POLL\n");
+		return -1;
+	}
+
+err_timeout:
+	sock_opt = 0;
+	ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
+			(void *)&sock_opt, sizeof(sock_opt));
+	if (ret < 0) {
+		AF_XDP_LOG(ERR, "Failed to unset

SO_PREFER_BUSY_POLL\n");

Quoted 8 lines

doc/guides/rel_notes/release_21_05.rst +8 -0

+		return -1;
+	}
+
+err_prefer:
+	rxq->busy_budget = 0;
+	return 0;
+}
+

<...>

Single-core performance of AF_XDP at high loads can be poor because a heavily loaded NAPI context will never enter or allow for busy-polling.

1C testpmd rxonly (both IRQs and PMD on core 0):
./dpdk-testpmd -l 0-1 --vdev=net_af_xdp0,iface=eth0 --main-lcore=1 -- \
--forward-mode=rxonly
0.088Mpps

In order to achieve decent performance at high loads, it is currently recommended ensure the IRQs for the netdev queue and the core running the PMD are different.

2C testpmd rxonly (IRQs on core 0, PMD on core 1):
./dpdk-testpmd -l 0-1 --vdev=net_af_xdp0,iface=eth0 --main-lcore=0 -- \
--forward-mode=rxonly
19.26Mpps

However using an extra core is of course not ideal. The SO_PREFER_BUSY_POLL socket option was introduced in kernel v5.11 to help improve 1C performance. See [1].

This series sets this socket option on xsks created with DPDK (ie. instances of the AF_XDP PMD) unless explicitly disabled or not supported by the kernel. It was chosen to be enabled by default in order to bring the AF_XDP PMD in line with most other PMDs which execute on a single core.

The following system and netdev settings are recommended in conjunction with busy polling:
echo 2 | sudo tee /sys/class/net/eth0/napi_defer_hard_irqs
echo 200000 | sudo tee /sys/class/net/eth0/gro_flush_timeout

Re-running the 1C test with busy polling support and the above settings:
./dpdk-testpmd -l 0-1 --vdev=net_af_xdp0,iface=eth0 --main-lcore=1 -- \
--forward-mode=rxonly
10.45Mpps

A new vdev arg is introduced called 'busy_budget' whose default value is 64. busy_budget is the value supplied to the kernel with the SO_BUSY_POLL_BUDGET socket option and represents the busy-polling NAPI budget ie. the number of packets the kernel will attempt to process in the netdev's NAPI context.

To set the busy budget to 256:
./dpdk-testpmd --vdev=net_af_xdp0,iface=eth0,busy_budget=256 14.06Mpps

If you still wish to run using 2 cores (one for PMD once for IRQs) it is recommended to disable busy polling to achieve optimal 2C performance:
./dpdk-testpmd --vdev=net_af_xdp0,iface=eth0,busy_budget=0 19.09Mpps

v2->v3:
* Moved release notes update to correct location
* Changed busy_budget from uint32_t to int since this is the type expected by setsockopt
* Validate busy_budget arg is <= UINT16_MAX during parse

v1->v2:
* Set batch size to default size of ring (2048)
* Split batches > 2048 into multiples of 2048 or less and process all packets in the same manner that is done for other drivers eg. ixgbe:
http://code.dpdk.org/dpdk/v21.02/source/drivers/net/ixgbe/ixgbe_rxtx.c#L318
* Update commit log with reasoning behing batching changes
* Update release notes with note on busy polling support
* Fix return type for sycall_needed function when the wakeup flag is not present
* Apprpriate log leveling
* Set default_*xportconf burst sizes to the default busy budget size (64)
* Detect support for busy polling via setsockopt instead of using the presence of the flag

RFC->v1:
* Fixed behaviour of busy_budget=0
* Ensure we bail out any of the new setsockopts fail

[1] https://lwn.net/Articles/837010/

Ciara Loftus (3):
net/af_xdp: allow bigger batch sizes
net/af_xdp: Use recvfrom() instead of poll()
net/af_xdp: preferred busy polling