[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-Id: <20250702171804.86422-1-mmc@linux.ibm.com>
Date: Wed, 2 Jul 2025 10:18:00 -0700
From: Mingming Cao <mmc@...ux.ibm.com>
To: netdev@...r.kernel.org
Cc: bjking1@...ux.ibm.com, haren@...ux.ibm.com, ricklind@...ux.ibm.com,
davemarq@...ux.ibm.com, mmc@...ux.ibm.com
Subject: [PATCH v2 net-next 0/4] Fix/Improve queue stats and subcrq indirect handling
This patch series introduces fixes and improvements to the ibmvnic driver,
focusing on accurate statistics reporting, CRQ scalability.
Patch 1: Convert per-queue RX/TX stats to atomic64_t to ensure thread-safe
updates in concurrent environments. This establishes a safe and consistent
foundation for subsequent statistics-related fixes.
Patch 2: Replace hardcoded NUM_RX_STATS and NUM_TX_STATS macros by dynamically
deriving the counts using sizeof. This fixes a mismatch introduced in commit
2ee73c54a615 (“ibmvnic: Add stat for tx direct vs tx batched”) and ensures
that all stat fields are correctly reported via ethtool -S.
Patch 3: Fix inaccurate sar statistics by implementing ndo_get_stats64() and
removing the outdated manual updates to netdev->stats.
Patch 4: Raise the default number of indirect sub-CRQ entries to 128 on POWER9
and later systems, improving performance under high-throughput workloads. A
module parameter is included as a fallback for compatibility on older systems;
this is documented as a transitional mechanism, not intended for dynamic tuning.
--------------------------------------
Changes since v1:
Link to v1: https://www.spinics.net/lists/netdev/msg1103893.html
Patch 1 (was Patch 2 in v1): Introduces atomic64_t stats early to establish
a consistent foundation. Replaces atomic64_sub() with atomic64_sub_return()
and adds a warning for underflow. This change now comes first to ensure correct
ordering of dependency with stat size derivation.
Patch 2 (was Patch 1 in v1): Commit message rewritten to clearly highlight the
mismatch bug caused by static stat counts and rebased on top of the atomic64
conversion to avoid broken logic during intermediate states.
Patch 3: No functional changes
Patch 4: Commit message clarified to explain the use of 128 indirect entries
on POWER9+ systems and the intent of the module parameter as a transitional
fallback for legacy or constrained systems.
Mingming Cao (4):
ibmvnic: Use atomic64_t for queue stats
ibmvnic: Fix hardcoded NUM_RX_STATS/NUM_TX_STATS with dynamic sizeof
ibmvnic: Use ndo_get_stats64 to fix inaccurate SAR reporting
ibmvnic: Make max subcrq indirect entries tunable via module param
drivers/net/ethernet/ibm/ibmvnic.c | 101 ++++++++++++++++++++---------
drivers/net/ethernet/ibm/ibmvnic.h | 29 +++++----
2 files changed, 89 insertions(+), 41 deletions(-)
--
2.39.3 (Apple Git-146)
Powered by blists - more mailing lists