[PATCH iwl-next v2 0/8] ixgbe: small cleanups and improvements

[Aleksandr Loktionov <aleksandr.loktionov@intel.com>]

Eight independent cleanups and improvements for the ixgbe driver,
grouped by theme:

Naming / type hygiene (patches 1-2):
  Patch 1 renames ixgbe_q_vector::numa_node to ::node.  The field name
  shadows the struct device numa_node accessor, causing a sparse warning.
  The rename aligns with other Intel drivers (ice, igc).

  Patch 2 changes local error-code variables from u32 to int in six
  functions across ixgbe_main.c, ixgbe_phy.c and ixgbe_x550.c.  Storing
  a signed errno in an unsigned type works by accident for truthiness
  checks but breaks exact comparisons and proper propagation.

Bug fix (patch 3):
  Patch 3 rejects duplicate FDIR perfect-filter rules before programming
  hardware.  The same 5-tuple with a different sw_idx would silently
  consume a second scarce FDIR slot and confuse rule deletion.
  Returns -EEXIST to userspace on duplicate rather than -EINVAL.

Performance (patch 4):
  Patch 4 replaces the busy-wait udelay(1000) in the SECRX_RDY poll
  loop with usleep_range(10, 20) and raises the iteration count from
  40 to 4000.  Because usleep_range(min, max) is guaranteed to sleep
  at least 'min' us, 4000 * 10 us preserves the original 40 ms
  minimum-wait-before-timeout; worst-case grows to ~80 ms (acceptable
  since SECRX_RDY failing to assert is non-fatal and only logged).
  Typical stall on fast hardware drops from up to ~1 ms per iteration
  to ~10-20 us.

Code quality (patches 5-8):
  Patch 5 replaces ktime_to_ns(ktime_get_real()) with the direct
  ktime_get_real_ns() helper.

  Patch 6 factors the three-line autoneg-restart sequence duplicated in
  ixgbe_setup_phy_link_generic() and ixgbe_setup_phy_link_tnx() into a
  static helper ixgbe_restart_auto_neg().  The helper checks the return
  value of read_reg() and returns early on error to avoid writing
  uninitialized data to the PHY register.

  Patches 7-8 improve the adaptive-ITR algorithm in two steps:
    7/8: Limit ITR decrease in latency mode to at most
         IXGBE_ITR_ADAPTIVE_MIN_INC (2 us) per step so that ACK-driven
         workloads do not overdrive interrupt rate.  Uses max_t() to
         ensure the clamp never drives ITR below the algorithm's own
         computation.
    8/8: Add IXGBE_ITR_ADAPTIVE_MASK_USECS (= IXGBE_ITR_ADAPTIVE_LATENCY
         - 1) to replace the open-coded ~IXGBE_ITR_ADAPTIVE_LATENCY
         mask in ixgbe_set_itr() with the cleaner AND form.

---
Changes since v1:
  - Collected Reviewed-by: Simon Horman <horms@kernel.org> on
    patches 1, 4, 6, 7, 8.
  - Patch 4 (SECRX_RDY): tightened commit message wording for the
    worst-case bound to reflect that on a loaded system actual
    scheduler wakeup latency may exceed the nominal 4000 * 20 us =
    ~80 ms upper bound; clarified that the timeout is non-fatal (only
    a debug message is logged, return is unconditionally 0).  No code
    change.
  - Patch 3 (prevent duplicate FDIR perfect filter rules):
    * Dropped the e_warn() in ixgbe_match_ethtool_fdir_entry();
      -EEXIST returned to userspace is sufficient.
    * Reordered locals in ixgbe_add_ethtool_fdir_entry() for
      reverse-Christmas-tree (`int err = -EINVAL;` moved above
      `u8 queue;`).
    The same-sw_idx skip in ixgbe_match_ethtool_fdir_entry() is
    intentional and required: ixgbe_update_ethtool_fdir_entry() (the
    caller of the add path) deletes any existing rule at the same
    sw_idx and replaces it, implementing the standard ethtool
    "set rule at location N" overwrite semantic.  Rejecting matching
    sw_idx with -EEXIST would break idempotent re-application and the
    legitimate replace-at-same-location path.
  - Patches 2, 5: no review feedback received; no change.

Aleksandr Loktionov (1):
  ixgbe: use int instead of u32 for error code variables

Alexander Duyck (2):
  ixgbe: limit ITR decrease in latency mode to prevent ACK overdrive
  ixgbe: add IXGBE_ITR_ADAPTIVE_MASK_USECS constant

Jacob Keller (2):
  ixgbe: rename numa_node to node in struct ixgbe_q_vector
  ixgbe: use ktime_get_real_ns() in ixgbe_ptp_reset()

Jakub Chylkowski (1):
  ixgbe: extract ixgbe_restart_auto_neg() to avoid code duplication

Maciej Rabeda (1):
  ixgbe: increase SECRX_RDY polling frequency in
    ixgbe_disable_rx_buff_generic

Piotr Skajewski (1):
  ixgbe: prevent adding duplicate FDIR perfect filter rules

 drivers/net/ethernet/intel/ixgbe/ixgbe.h      |  3 +-
 .../net/ethernet/intel/ixgbe/ixgbe_common.c   |  5 +--
 .../net/ethernet/intel/ixgbe/ixgbe_ethtool.c  | 27 +++++++++++-
 drivers/net/ethernet/intel/ixgbe/ixgbe_lib.c  |  2 +-
 drivers/net/ethernet/intel/ixgbe/ixgbe_main.c | 19 ++++++++--
 drivers/net/ethernet/intel/ixgbe/ixgbe_phy.c  | 42 +++++++++++--------
 drivers/net/ethernet/intel/ixgbe/ixgbe_ptp.c  |  2 +-
 drivers/net/ethernet/intel/ixgbe/ixgbe_x550.c | 12 +++---
 8 files changed, 77 insertions(+), 35 deletions(-)

-- 
2.52.0