This series provides updates to mlx5 ethernet driver.
1) Starting with a four patches series to optimize flow counters updates,
From Vlad Buslov:
==============================================
By default mlx5 driver updates cached counters each second. Update function
consumes noticeable amount of CPU resources. The goal of this patch series
is to optimize update function.
Investigation revealed following bottlenecks in fs counters
implementation:
1) Update code(scheduled each second) iterates over all counters twice.
(first for finding and deleting counters that are marked for deletion,
second iteration is for actually updating the counters)
2) Counters are stored in rb tree. Linear iteration over all rb tree
elements(rb_next in profiling data) consumed ~65% of time spent in
update function.
Following optimizations were implemented:
1) Instead of just marking counters for deletion, store them in
standalone list. This removes first iteration over whole counters tree.
2) Store counters in sorted list to optimize traversing them and remove
calls to rb_next.
First implementation of these changes caused degradation of performance,
instead of improving it. Investigation revealed that there first cache
line of struct mlx5_fc is full and adding anything to it causes amount
of cache misses to double. To mitigate that, following refactorings were
implemented:
- Change 'addlist' list type from double linked to single linked. This
allowes to get free space for one additional pointer that is used to
store deletion list(optimization 1)
- Substitute rb tree with idr. Idr is non-intrusive data structure and
doesn't require adding any new members to struct mlx5_fc. Use free
space that became available for double linked sorted list that is used
for traversing all counters. (optimization 2)
Described changes reduced CPU time spent in mlx5_fc_stats_work from 70%
to 44%. (global perf profile mode)
============================================
The rest of the series are misc updates:
2) From Kamal, Move mlx5e_priv_flags into en_ethtool.c, to avoid a
compilation warning.
3) From Roi Dayan, Move Q counters allocation and drop RQ to init_rx profile
function to avoid allocating Q counters when not required.
4) From Shay Agroskin, Replace PTP clock lock from RW lock to seq lock.
Almost double the packet rate when timestamping is active on multiple TX
queues.
5) From: Natali Shechtman, set ECN for received packets using CQE indication.
6) From: Alaa Hleihel, don't set CHECKSUM_COMPLETE on SCTP packets.
CHECKSUM_COMPLETE is not applicable to SCTP protocol.
-----BEGIN PGP SIGNATURE-----
iQEcBAABAgAGBQJbkKzcAAoJEEg/ir3gV/o+4bsIAKu61DsZZhJAXfwZ4JXI9fN7
iZ9wOPbPovTrkfRfqatcCa3CBwxM1zZ6WzbVTGGieMsUnxjp4G+LAs7m8J+cmWB2
GsxhO3lsF5iGTcEB5Q5a/TlWlUP0CiPcdNZZe6FvT2Vzil+mLAb3BZgIPKPgoyJW
CmDhIOslfolBJC0s4fc+zVcWrShmKnUwHmVLhkOXpz8cfYBB1fBVGebRTZiYsvbf
8CV4buiYx0Fuq9MSnOwaMuPrUP1IB6DweNbFpg9gRd0k4mmWnoNu3l+g1Flgmd88
RtU4+5PEtZF2kNhH/LeiHRqkZ5f2cmOsIKbDZoX6xwzfYesws8jlDJpL5dIhxZA=
=nxDw
-----END PGP SIGNATURE-----
Merge tag 'mlx5e-updates-2018-09-05' of git://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux
Saeed Mahameed says:
====================
mlx5e-updates-2018-09-05
This series provides updates to mlx5 ethernet driver.
1) Starting with a four patches series to optimize flow counters updates,
From Vlad Buslov:
==============================================
By default mlx5 driver updates cached counters each second. Update function
consumes noticeable amount of CPU resources. The goal of this patch series
is to optimize update function.
Investigation revealed following bottlenecks in fs counters
implementation:
1) Update code(scheduled each second) iterates over all counters twice.
(first for finding and deleting counters that are marked for deletion,
second iteration is for actually updating the counters)
2) Counters are stored in rb tree. Linear iteration over all rb tree
elements(rb_next in profiling data) consumed ~65% of time spent in
update function.
Following optimizations were implemented:
1) Instead of just marking counters for deletion, store them in
standalone list. This removes first iteration over whole counters tree.
2) Store counters in sorted list to optimize traversing them and remove
calls to rb_next.
First implementation of these changes caused degradation of performance,
instead of improving it. Investigation revealed that there first cache
line of struct mlx5_fc is full and adding anything to it causes amount
of cache misses to double. To mitigate that, following refactorings were
implemented:
- Change 'addlist' list type from double linked to single linked. This
allowes to get free space for one additional pointer that is used to
store deletion list(optimization 1)
- Substitute rb tree with idr. Idr is non-intrusive data structure and
doesn't require adding any new members to struct mlx5_fc. Use free
space that became available for double linked sorted list that is used
for traversing all counters. (optimization 2)
Described changes reduced CPU time spent in mlx5_fc_stats_work from 70%
to 44%. (global perf profile mode)
============================================
The rest of the series are misc updates:
2) From Kamal, Move mlx5e_priv_flags into en_ethtool.c, to avoid a
compilation warning.
3) From Roi Dayan, Move Q counters allocation and drop RQ to init_rx profile
function to avoid allocating Q counters when not required.
4) From Shay Agroskin, Replace PTP clock lock from RW lock to seq lock.
Almost double the packet rate when timestamping is active on multiple TX
queues.
5) From: Natali Shechtman, set ECN for received packets using CQE indication.
6) From: Alaa Hleihel, don't set CHECKSUM_COMPLETE on SCTP packets.
CHECKSUM_COMPLETE is not applicable to SCTP protocol.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
ddc9cc0131