Skip to content

net: sched: move back qlen to per CPU accounting

Paolo Abeni netdev 2 versions – In review

Get series with b4
b4 am [email protected]

Versions newest first · linked by author and title

  1. v2 Latest5 patches 1 replyIn review
  2. v15 patches 3 repliesSuperseded

Changes in v2 from the cover letter

- fixed build issues
- more descriptive commit message for patch 5/5

Patches in v2 No applies recorded yet

#Subject and filesSizeRepliesApplied
1/5net: caif: avoid using qdisc_qlen()net/caif/caif_dev.c+8 -40 replies
2/5net: sched: prefer qdisc_is_empty() over direct qlen accessinclude/net/sch_generic.h+1 -10 replies
3/5net: sched: always do stats accounting according to TCQ_F_CPUSTATSinclude/net/sch_generic.h net/sched/sch_generic.c+31 -420 replies
4/5net: sched: when clearing NOLOCK, clear TCQ_F_CPUSTATS, tooinclude/net/sch_generic.h net/sched/sch_api.c net/sched/sch_generic.c+42 -90 replies
5/5Revert: "net: sched: put back q.qlen into a single location"include/net/sch_generic.h net/core/gen_stats.c net/sched/sch_generic.c+28 -200 replies

Reviews trailers per patch

No Reviewed-by, Acked-by or Tested-by trailers for v2 yet.

Read the v2 thread

Cover letter [PATCH net-next v2 0/5] net: sched: move back qlen to per CPU accounting

The commit 46b1c18f9deb ("net: sched: put back q.qlen into a single location") introduced some measurable regression in the contended scenarios for lock qdisc.

As Eric suggested we could replace q.qlen access with calls to qdisc_is_empty() in the datapath and revert the above commit. The TC subsystem updates qdisc->is_empty in a somewhat loose way: notably 'is_empty' is set only when the qdisc dequeue() calls return a NULL ptr. That is, the invocation after the last packet is dequeued.

The above is good enough for BYPASS implementation - the only downside is that we end up avoiding the optimization for a very small time-frame - but will break hard things when internal structures consistency for classful qdisc relies on child qdisc_is_empty().

A more strict 'is_empty' update adds a relevant complexity to its life-cycle, so this series takes a different approach: we allow lockless qdisc to switch from per CPU accounting to global stats accounting when the NOLOCK bit is cleared. Since most pieces of infrastructure are already in place, this requires very little changes to the pfifo_fast qdisc, and any later NOLOCK qdisc can hook there with little effort - no need to maintain two different implementations.

The first 2 patches removes direct qlen access from non core TC code, the 3rd and 4th patches place and use the infrastructure to allow stats account switching and the 5th patch is the actual revert.

v1 -> v2:
- fixed build issues
- more descriptive commit message for patch 5/5

Paolo Abeni (5):
net: caif: avoid using qdisc_qlen()
net: sched: prefer qdisc_is_empty() over direct qlen access
net: sched: always do stats accounting according to TCQ_F_CPUSTATS
net: sched: when clearing NOLOCK, clear TCQ_F_CPUSTATS, too
Revert: "net: sched: put back q.qlen into a single location"

 include/net/sch_generic.h | 80 ++++++++++++++++++++++++++++-----------
 net/caif/caif_dev.c       | 12 ++++--
 net/core/gen_stats.c      |  2 +
 net/sched/sch_api.c       | 15 +++++++-
 net/sched/sch_generic.c   | 67 +++++++++++---------------------
 5 files changed, 105 insertions(+), 71 deletions(-)
Signature · 1 line
2.20.1