mirror of
https://github.com/xdp-project/bpf-examples.git
synced 2024-05-06 15:54:53 +00:00
ba2c114db0
The BPF-prog "not_txq_zero" also needed to take into account that skb->queue_mapping usually isn't set for locally generated traffic. I worry that sockets can set another queue id that could override our (BPF choice) in netdev_pick_tx(). See sk_tx_queue_set() and sk_tx_queue_get(). Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
81 lines
2.4 KiB
C
81 lines
2.4 KiB
C
/* SPDX-License-Identifier: GPL-2.0-or-later */
|
|
/* Copyright 2022 Jesper Dangaard Brouer */
|
|
|
|
#include <linux/bpf.h>
|
|
#include <bpf/bpf_helpers.h>
|
|
#include <bpf/compiler.h>
|
|
#include <linux/pkt_sched.h>
|
|
#include <linux/pkt_cls.h>
|
|
//#include <stdbool.h>
|
|
//#include "../include/xdp/parsing_helpers.h"
|
|
|
|
/* Manuel setup:
|
|
export DEV=eth1
|
|
|
|
tc qdisc add dev "$DEV" clsact
|
|
tc filter add dev "$DEV" egress bpf da obj tc_txq_policy_kern.o
|
|
tc filter list dev "$DEV" egress
|
|
|
|
* Quick test reloading with tc:
|
|
tc filter replace dev "$DEV" egress prio 0xC000 handle 1 bpf da obj tc_txq_policy_kern.o
|
|
|
|
* Beware: Trying to replace an existing TC-BPF prog often result in appending a
|
|
* new prog (as a new tc filter instance). Be careful to set both handle and
|
|
* prio to the existing TC-BPF "filter" instance.
|
|
|
|
* Delete by teardown of clsact
|
|
tc qdisc delete dev "$DEV" clsact
|
|
|
|
*/
|
|
SEC("classifier")
|
|
int queue_map_4 (struct __sk_buff *skb)
|
|
{
|
|
__u16 txq_root_handle;
|
|
|
|
/* The skb->queue_mapping is 1-indexed (zero means not set). The
|
|
* underlying MQ leaf's are also 1-indexed, which makes it easier to
|
|
* reason about. If debugging this realize that setting
|
|
* skb->queue_mapping here is like it was set on RX-path the
|
|
* skb_rx_queue_recorded number, and when reaching TX-layer
|
|
* (skb_get_rx_queue) will have decremented it by-1.
|
|
*/
|
|
txq_root_handle = 4;
|
|
skb->queue_mapping = txq_root_handle;
|
|
|
|
/* Details: Kernel double protect against setting a too high
|
|
* queue_mapping. In skb_tx_hash() it will reduce number to be
|
|
* less-than (or equal) dev->real_num_tx_queues. And netdev_pick_tx()
|
|
* cap via netdev_cap_txqueue().
|
|
*/
|
|
|
|
// FIXME: Do we need to set TC_H_MAJOR(skb->priority) for this to work?
|
|
|
|
return TC_ACT_OK;
|
|
}
|
|
|
|
/*
|
|
* Section name "tc" is preferred over "classifier" as its being deprecated
|
|
* https://github.com/libbpf/libbpf/wiki/Libbpf-1.0-migration-guide#bpf-program-sec-annotation-deprecations
|
|
*/
|
|
|
|
SEC("tc")
|
|
int not_txq_zero (struct __sk_buff *skb)
|
|
{
|
|
/* Existing skb->queue_mapping can come from skb_record_rx_queue() which
|
|
* is usually called by drivers in early RX handling when creating SKB.
|
|
*/
|
|
|
|
/* At this stage queue_mapping is 1-indexed.
|
|
* Thus, code is changing TXQ zero to be remapped to TXQ 3. */
|
|
if (skb->queue_mapping == 1)
|
|
skb->queue_mapping = 4;
|
|
|
|
/* If queue_mapping was not set by skb_record_rx_queue(),
|
|
* e.g. locally generated traffic
|
|
*/
|
|
if (skb->queue_mapping == 0)
|
|
skb->queue_mapping = 3;
|
|
|
|
return TC_ACT_OK;
|
|
}
|