Skip to content

Commit

Permalink
Merge branch 'net-mlx5-hw-counters-refactor'
Browse files Browse the repository at this point in the history
Tariq Toukan says:

====================
net/mlx5: hw counters refactor

This is a patchset re-post, see:
https://lore.kernel.org/[email protected]

In this patchset, Cosmin refactors hw counters and solves perf scaling
issue.

Series generated against:
commit c824deb ("cxgb4: clip_tbl: Fix spelling mistake "wont" -> "won't"")

HW counters are central to mlx5 driver operations. They are hardware
objects created and used alongside most steering operations, and queried
from a variety of places. Most counters are queried in bulk from a
periodic task in fs_counters.c.

Counter performance is important and as such, a variety of improvements
have been done over the years. Currently, counters are allocated from
pools, which are bulk allocated to amortize the cost of firmware
commands. Counters are managed through an IDR, a doubly linked list and
two atomic single linked lists. Adding/removing counters is a complex
dance between user contexts requesting it and the mlx5_fc_stats_work
task which does most of the work.

Under high load (e.g. from connection tracking flow insertion/deletion),
the counter code becomes a bottleneck, as seen on flame graphs. Whenever
a counter is deleted, it gets added to a list and the wq task is
scheduled to run immediately to actually delete it. This is done via
mod_delayed_work which uses an internal spinlock. In some tests, waiting
for this spinlock took up to 66% of all samples.

This series refactors the counter code to use a more straight-forward
approach, avoiding the mod_delayed_work problem and making the code
easier to understand. For that:

- patch #1 moves counters data structs to a more appropriate place.
- patch #2 simplifies the bulk query allocation scheme by using vmalloc.
- patch #3 replaces the IDR+3 lists with an xarray. This is the main
  patch of the series, solving the spinlock congestion issue.
- patch #4 removes an unnecessary cacheline alignment causing a lot of
  memory to be wasted.
- patches #5 and torvalds#6 are small cleanups enabled by the refactoring.
====================

Link: https://patch.msgid.link/[email protected]
Signed-off-by: Jakub Kicinski <[email protected]>
  • Loading branch information
kuba-moo committed Oct 4, 2024
2 parents c55ff46 + d1c9cff commit 34ea1df
Show file tree
Hide file tree
Showing 4 changed files with 147 additions and 278 deletions.
2 changes: 1 addition & 1 deletion drivers/net/ethernet/mellanox/mlx5/core/en/tc_ct.c
Original file line number Diff line number Diff line change
Expand Up @@ -1026,7 +1026,7 @@ mlx5_tc_ct_counter_create(struct mlx5_tc_ct_priv *ct_priv)
return ERR_PTR(-ENOMEM);

counter->is_shared = false;
counter->counter = mlx5_fc_create_ex(ct_priv->dev, true);
counter->counter = mlx5_fc_create(ct_priv->dev, true);
if (IS_ERR(counter->counter)) {
ct_dbg("Failed to create counter for ct entry");
ret = PTR_ERR(counter->counter);
Expand Down
Loading

0 comments on commit 34ea1df

Please sign in to comment.