[PATCH v6 4/4] net/af_packet: use kernel to filter multi-queue Rx

[Frank Dressler]

The AF_PACKET PMD currently needs to filter out outgoing packets in
the PMD when the ignore_outgoing devarg is enabled and multiple queues
are used. Kernel-side filtering is only supported for the single-queue
scenario.

Since Linux 6.2, the kernel filters outgoing packets for a fanout
group when PACKET_FANOUT_FLAG_IGNORE_OUTGOING is set.

This change opportunistically sets this flag for all kernel versions
in multi-queue scenarios. Kernels before Linux 6.2 ignore the flag
and have no way to detect support. The PMD filtering is kept as a
fallback for kernels that do not support the flag.

Signed-off-by: Frank Dressler <frank at dressler.pro>
---
 doc/guides/nics/af_packet.rst             |  3 +++
 doc/guides/rel_notes/release_26_11.rst    |  3 ++-
 drivers/net/af_packet/rte_eth_af_packet.c | 13 +++++++++++++
 3 files changed, 18 insertions(+), 1 deletion(-)

diff --git a/doc/guides/nics/af_packet.rst b/doc/guides/nics/af_packet.rst
index 6df14c510d..702822a428 100644
--- a/doc/guides/nics/af_packet.rst
+++ b/doc/guides/nics/af_packet.rst
@@ -45,6 +45,9 @@ of AF_PACKET for frame reception.
 
 When ``ignore_outgoing`` is set, ``PACKET_IGNORE_OUTGOING`` is set for
 single-queue ports, which moves the filtering from the PMD to the kernel.
+For multi-queue ports ``PACKET_FANOUT_FLAG_IGNORE_OUTGOING`` is set to
+achieve the same but this only works from Linux 6.2. For earlier Linux
+versions, the PMD does the filtering for multi-queue ports.
 
 Because this implementation is based on PACKET_MMAP, and PACKET_MMAP has its
 own pre-requisites, it should be noted that the inner workings of PACKET_MMAP
diff --git a/doc/guides/rel_notes/release_26_11.rst b/doc/guides/rel_notes/release_26_11.rst
index e83cb9854f..9740e9fdce 100644
--- a/doc/guides/rel_notes/release_26_11.rst
+++ b/doc/guides/rel_notes/release_26_11.rst
@@ -118,7 +118,8 @@ New Features
 
   Added ``ignore_outgoing`` so ``rte_eth_rx_burst()`` returns only incoming
   packets. The single-queue scenario is accelerated, as the kernel does
-  the filtering.
+  the filtering. Multi-queue ports are also filtered in the kernel since
+  Linux 6.2, otherwise in the PMD.
 
 * **Updated AF_XDP driver.**
 
diff --git a/drivers/net/af_packet/rte_eth_af_packet.c b/drivers/net/af_packet/rte_eth_af_packet.c
index 44b91502fb..0d044ab661 100644
--- a/drivers/net/af_packet/rte_eth_af_packet.c
+++ b/drivers/net/af_packet/rte_eth_af_packet.c
@@ -32,6 +32,11 @@
 #include <unistd.h>
 #include <poll.h>
 
+/* Added in Linux 6.2. Older kernels ignore it. */
+#ifndef PACKET_FANOUT_FLAG_IGNORE_OUTGOING
+#define PACKET_FANOUT_FLAG_IGNORE_OUTGOING 0x4000
+#endif
+
 #define ETH_AF_PACKET_IFACE_ARG		"iface"
 #define ETH_AF_PACKET_NUM_Q_ARG		"qpairs"
 #define ETH_AF_PACKET_BLOCKSIZE_ARG	"blocksz"
@@ -995,6 +1000,14 @@ rte_pmd_init_internals(struct rte_vdev_device *dev,
 		goto error;
 	}
 
+	/*
+	 * Set the flag opportunistically if multiple queues are requested.
+	 * It takes effect from Linux 6.2. Older kernels ignore it and the
+	 * PMD filters instead.
+	 */
+	if (ignore_outgoing != 0 && nb_queues > 1)
+		fanout_arg |= PACKET_FANOUT_FLAG_IGNORE_OUTGOING << 16;
+
 	for (q = 0; q < nb_queues; q++) {
 		/* Open an AF_PACKET socket for this queue... */
 		qsockfd = socket(AF_PACKET, SOCK_RAW, 0);
-- 
2.56.0