summaryrefslogtreecommitdiff
path: root/drivers/net/ethernet/intel/ice/ice_controlq.h
diff options
context:
space:
mode:
authorJacob Keller <jacob.e.keller@intel.com>2023-06-13 13:40:53 -0700
committerTony Nguyen <anthony.l.nguyen@intel.com>2023-06-22 10:23:44 -0700
commita734c43caa4d9a08da521be1a2135cadf1510e75 (patch)
tree2ba36a8ae9e8c2970c315dcaf8151bade863ee8f /drivers/net/ethernet/intel/ice/ice_controlq.h
parent98e95872f2b818c74872d073eaa4c937579d41fc (diff)
ice: reduce initial wait for control queue messages
The ice_sq_send_cmd() function is used to send messages to the control queues used to communicate with firmware, virtual functions, and even some hardware. When sending a control queue message, the driver is designed to synchronously wait for a response from the queue. Currently it waits between checks for 100 to 150 microseconds. Commit f86d6f9c49f6 ("ice: sleep, don't busy-wait, for ICE_CTL_Q_SQ_CMD_TIMEOUT") did recently change the behavior from an unnecessary delay into a sleep which is a significant improvement over the old behavior of polling using udelay. Because of the nature of PCIe transactions, the hardware won't be informed about a new message until the write to the tail register posts. This is only guaranteed to occur at the next register read. In ice_sq_send_cmd(), this happens at the ice_sq_done() call. Because of this, the driver essentially forces a minimum of one full wait time regardless of how fast the response is. For the hardware-based sideband queue, this is especially slow. It is expected that the hardware will respond within 2 or 3 microseconds, an order of magnitude faster than the 100-150 microsecond sleep. Allow such fast completions to occur without delay by introducing a small 5 microsecond delay first before entering the sleeping timeout loop. Ensure the tail write has been posted by using ice_flush(hw) first. While at it, lets also remove the ICE_CTL_Q_SQ_CMD_USEC macro as it obscures the sleep time in the inner loop. It was likely introduced to avoid "magic numbers", but in practice sleep and delay values are easier to read and understand when using actual numbers instead of a named constant. This change should allow the fast hardware based control queue messages to complete quickly without delay, while slower firmware queue response times will sleep while waiting for the response. Signed-off-by: Jacob Keller <jacob.e.keller@intel.com> Reviewed-by: Michal Schmidt <mschmidt@redhat.com> Tested-by: Pucha Himasekhar Reddy <himasekharx.reddy.pucha@intel.com> (A Contingent worker at Intel) Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Diffstat (limited to 'drivers/net/ethernet/intel/ice/ice_controlq.h')
-rw-r--r--drivers/net/ethernet/intel/ice/ice_controlq.h1
1 files changed, 0 insertions, 1 deletions
diff --git a/drivers/net/ethernet/intel/ice/ice_controlq.h b/drivers/net/ethernet/intel/ice/ice_controlq.h
index 950b7f4a7a05..8f2fd1613a95 100644
--- a/drivers/net/ethernet/intel/ice/ice_controlq.h
+++ b/drivers/net/ethernet/intel/ice/ice_controlq.h
@@ -35,7 +35,6 @@ enum ice_ctl_q {
/* Control Queue timeout settings - max delay 1s */
#define ICE_CTL_Q_SQ_CMD_TIMEOUT HZ /* Wait max 1s */
-#define ICE_CTL_Q_SQ_CMD_USEC 100 /* Check every 100usec */
#define ICE_CTL_Q_ADMIN_INIT_TIMEOUT 10 /* Count 10 times */
#define ICE_CTL_Q_ADMIN_INIT_MSEC 100 /* Check every 100msec */