|
|
Log in / Subscribe / Register

net_sched: speedup qdisc dequeue

From:  Eric Dumazet <edumazet-AT-google.com>
To:  "David S . Miller" <davem-AT-davemloft.net>, Jakub Kicinski <kuba-AT-kernel.org>, Paolo Abeni <pabeni-AT-redhat.com>
Subject:  [PATCH v3 net-next 00/14] net_sched: speedup qdisc dequeue
Date:  Fri, 21 Nov 2025 08:32:42 +0000
Message-ID:  <20251121083256.674562-1-edumazet@google.com>
Cc:  Simon Horman <horms-AT-kernel.org>, Jamal Hadi Salim <jhs-AT-mojatatu.com>, Cong Wang <xiyou.wangcong-AT-gmail.com>, Jiri Pirko <jiri-AT-resnulli.us>, "Toke Høiland-Jørgensen" <toke-AT-redhat.com>, Kuniyuki Iwashima <kuniyu-AT-google.com>, Willem de Bruijn <willemb-AT-google.com>, netdev-AT-vger.kernel.org, eric.dumazet-AT-gmail.com, Eric Dumazet <edumazet-AT-google.com>
Archive-link:  Article

Avoid up to two cache line misses in qdisc dequeue() to fetch
skb_shinfo(skb)->gso_segs/gso_size while qdisc spinlock is held.

Idea is to cache gso_segs at enqueue time before spinlock is
acquired, in the first skb cache line, where we already
have qdisc_skb_cb(skb)->pkt_len.

This series gives a 8 % improvement in a TX intensive workload.

(120 Mpps -> 130 Mpps on a Turin host, IDPF with 32 TX queues)

v3: - Same than v2, resent after a syzbot report was fixed in commit
      4ef927436258 ("bpf: Add bpf_prog_run_data_pointers()")

v2: - Fixed issues reported by Jakub (thanks !)
    - Added three patches adding/using qdisc_dequeue_drop() after
      recent regressions with CAKE qdisc reported by Toke.
      More fixes to come later.
    - https://lore.kernel.org/netdev/20251111093204.1432437-1-e...

v1: https://lore.kernel.org/netdev/20251110094505.3335073-1-e...


Eric Dumazet (14):
  net_sched: make room for (struct qdisc_skb_cb)->pkt_segs
  net: init shinfo->gso_segs from qdisc_pkt_len_init()
  net_sched: initialize qdisc_skb_cb(skb)->pkt_segs in
    qdisc_pkt_len_init()
  net: use qdisc_pkt_len_segs_init() in sch_handle_ingress()
  net_sched: use qdisc_skb_cb(skb)->pkt_segs in bstats_update()
  net_sched: cake: use qdisc_pkt_segs()
  net_sched: add Qdisc_read_mostly and Qdisc_write groups
  net_sched: sch_fq: move qdisc_bstats_update() to fq_dequeue_skb()
  net_sched: sch_fq: prefetch one skb ahead in dequeue()
  net: prefech skb->priority in __dev_xmit_skb()
  net: annotate a data-race in __dev_xmit_skb()
  net_sched: add tcf_kfree_skb_list() helper
  net_sched: add qdisc_dequeue_drop() helper
  net_sched: use qdisc_dequeue_drop() in cake, codel, fq_codel

 include/net/pkt_sched.h   |   5 +-
 include/net/sch_generic.h | 101 ++++++++++++++++++++++++++++----------
 net/core/dev.c            |  62 +++++++++++++----------
 net/sched/act_ct.c        |   8 +--
 net/sched/cls_api.c       |   6 +--
 net/sched/cls_flower.c    |   2 +-
 net/sched/sch_cake.c      |  19 +++----
 net/sched/sch_codel.c     |   4 +-
 net/sched/sch_dualpi2.c   |   1 +
 net/sched/sch_fq.c        |   9 ++--
 net/sched/sch_fq_codel.c  |   5 +-
 net/sched/sch_netem.c     |   1 +
 net/sched/sch_qfq.c       |   2 +-
 net/sched/sch_taprio.c    |   1 +
 net/sched/sch_tbf.c       |   1 +
 15 files changed, 147 insertions(+), 80 deletions(-)

-- 
2.52.0.460.gd25c4c69ec-goog




Copyright © 2025, Eklektix, Inc.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds