OSDN Git Service

ice: Calculate ITR increment based on direct calculation
authorBrett Creeley <brett.creeley@intel.com>
Thu, 28 Feb 2019 23:25:47 +0000 (15:25 -0800)
committerJeff Kirsher <jeffrey.t.kirsher@intel.com>
Thu, 18 Apr 2019 15:38:48 +0000 (08:38 -0700)
Currently when calculating how much to increment ITR by inside of
ice_update_itr() we do some estimations and intermediate
calculations. Instead of doing estimations, just do the
calculation directly. This allows for a more accurate value and it
makes it easier for the next person to understand and update.

Also, remove the dividing the ITR value by 2 when latency
driven because the ITR values are already so low for 100Gbps
speed. This should help get to the desired ITR value faster.

Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
drivers/net/ethernet/intel/ice/ice_txrx.c

index a1dcde4..79043fe 100644 (file)
@@ -1097,19 +1097,69 @@ static int ice_clean_rx_irq(struct ice_ring *rx_ring, int budget)
        return failure ? budget : (int)total_rx_pkts;
 }
 
-static unsigned int ice_itr_divisor(struct ice_port_info *pi)
+/**
+ * ice_adjust_itr_by_size_and_speed - Adjust ITR based on current traffic
+ * @port_info: port_info structure containing the current link speed
+ * @avg_pkt_size: average size of Tx or Rx packets based on clean routine
+ * @itr: itr value to update
+ *
+ * Calculate how big of an increment should be applied to the ITR value passed
+ * in based on wmem_default, SKB overhead, Ethernet overhead, and the current
+ * link speed.
+ *
+ * The following is a calculation derived from:
+ *  wmem_default / (size + overhead) = desired_pkts_per_int
+ *  rate / bits_per_byte / (size + Ethernet overhead) = pkt_rate
+ *  (desired_pkt_rate / pkt_rate) * usecs_per_sec = ITR value
+ *
+ * Assuming wmem_default is 212992 and overhead is 640 bytes per
+ * packet, (256 skb, 64 headroom, 320 shared info), we can reduce the
+ * formula down to:
+ *
+ *      wmem_default * bits_per_byte * usecs_per_sec   pkt_size + 24
+ * ITR = -------------------------------------------- * --------------
+ *                          rate                       pkt_size + 640
+ */
+static unsigned int
+ice_adjust_itr_by_size_and_speed(struct ice_port_info *port_info,
+                                unsigned int avg_pkt_size,
+                                unsigned int itr)
 {
-       switch (pi->phy.link_info.link_speed) {
+       switch (port_info->phy.link_info.link_speed) {
+       case ICE_AQ_LINK_SPEED_100GB:
+               itr += DIV_ROUND_UP(17 * (avg_pkt_size + 24),
+                                   avg_pkt_size + 640);
+               break;
+       case ICE_AQ_LINK_SPEED_50GB:
+               itr += DIV_ROUND_UP(34 * (avg_pkt_size + 24),
+                                   avg_pkt_size + 640);
+               break;
        case ICE_AQ_LINK_SPEED_40GB:
-               return ICE_ITR_ADAPTIVE_MIN_INC * 1024;
+               itr += DIV_ROUND_UP(43 * (avg_pkt_size + 24),
+                                   avg_pkt_size + 640);
+               break;
        case ICE_AQ_LINK_SPEED_25GB:
+               itr += DIV_ROUND_UP(68 * (avg_pkt_size + 24),
+                                   avg_pkt_size + 640);
+               break;
        case ICE_AQ_LINK_SPEED_20GB:
-               return ICE_ITR_ADAPTIVE_MIN_INC * 512;
-       case ICE_AQ_LINK_SPEED_100MB:
-               return ICE_ITR_ADAPTIVE_MIN_INC * 32;
+               itr += DIV_ROUND_UP(85 * (avg_pkt_size + 24),
+                                   avg_pkt_size + 640);
+               break;
+       case ICE_AQ_LINK_SPEED_10GB:
+               /* fall through */
        default:
-               return ICE_ITR_ADAPTIVE_MIN_INC * 256;
+               itr += DIV_ROUND_UP(170 * (avg_pkt_size + 24),
+                                   avg_pkt_size + 640);
+               break;
        }
+
+       if ((itr & ICE_ITR_MASK) > ICE_ITR_ADAPTIVE_MAX_USECS) {
+               itr &= ICE_ITR_ADAPTIVE_LATENCY;
+               itr += ICE_ITR_ADAPTIVE_MAX_USECS;
+       }
+
+       return itr;
 }
 
 /**
@@ -1128,8 +1178,8 @@ static unsigned int ice_itr_divisor(struct ice_port_info *pi)
 static void
 ice_update_itr(struct ice_q_vector *q_vector, struct ice_ring_container *rc)
 {
-       unsigned int avg_wire_size, packets, bytes, itr;
        unsigned long next_update = jiffies;
+       unsigned int packets, bytes, itr;
        bool container_is_rx;
 
        if (!rc->ring || !ITR_IS_DYNAMIC(rc->itr_setting))
@@ -1174,7 +1224,7 @@ ice_update_itr(struct ice_q_vector *q_vector, struct ice_ring_container *rc)
                if (packets && packets < 4 && bytes < 9000 &&
                    (q_vector->tx.target_itr & ICE_ITR_ADAPTIVE_LATENCY)) {
                        itr = ICE_ITR_ADAPTIVE_LATENCY;
-                       goto adjust_by_size;
+                       goto adjust_by_size_and_speed;
                }
        } else if (packets < 4) {
                /* If we have Tx and Rx ITR maxed and Tx ITR is running in
@@ -1242,70 +1292,11 @@ ice_update_itr(struct ice_q_vector *q_vector, struct ice_ring_container *rc)
         */
        itr = ICE_ITR_ADAPTIVE_BULK;
 
-adjust_by_size:
-       /* If packet counts are 256 or greater we can assume we have a gross
-        * overestimation of what the rate should be. Instead of trying to fine
-        * tune it just use the formula below to try and dial in an exact value
-        * gives the current packet size of the frame.
-        */
-       avg_wire_size = bytes / packets;
+adjust_by_size_and_speed:
 
-       /* The following is a crude approximation of:
-        *  wmem_default / (size + overhead) = desired_pkts_per_int
-        *  rate / bits_per_byte / (size + ethernet overhead) = pkt_rate
-        *  (desired_pkt_rate / pkt_rate) * usecs_per_sec = ITR value
-        *
-        * Assuming wmem_default is 212992 and overhead is 640 bytes per
-        * packet, (256 skb, 64 headroom, 320 shared info), we can reduce the
-        * formula down to
-        *
-        *  (170 * (size + 24)) / (size + 640) = ITR
-        *
-        * We first do some math on the packet size and then finally bitshift
-        * by 8 after rounding up. We also have to account for PCIe link speed
-        * difference as ITR scales based on this.
-        */
-       if (avg_wire_size <= 60) {
-               /* Start at 250k ints/sec */
-               avg_wire_size = 4096;
-       } else if (avg_wire_size <= 380) {
-               /* 250K ints/sec to 60K ints/sec */
-               avg_wire_size *= 40;
-               avg_wire_size += 1696;
-       } else if (avg_wire_size <= 1084) {
-               /* 60K ints/sec to 36K ints/sec */
-               avg_wire_size *= 15;
-               avg_wire_size += 11452;
-       } else if (avg_wire_size <= 1980) {
-               /* 36K ints/sec to 30K ints/sec */
-               avg_wire_size *= 5;
-               avg_wire_size += 22420;
-       } else {
-               /* plateau at a limit of 30K ints/sec */
-               avg_wire_size = 32256;
-       }
-
-       /* If we are in low latency mode halve our delay which doubles the
-        * rate to somewhere between 100K to 16K ints/sec
-        */
-       if (itr & ICE_ITR_ADAPTIVE_LATENCY)
-               avg_wire_size >>= 1;
-
-       /* Resultant value is 256 times larger than it needs to be. This
-        * gives us room to adjust the value as needed to either increase
-        * or decrease the value based on link speeds of 10G, 2.5G, 1G, etc.
-        *
-        * Use addition as we have already recorded the new latency flag
-        * for the ITR value.
-        */
-       itr += DIV_ROUND_UP(avg_wire_size,
-                           ice_itr_divisor(q_vector->vsi->port_info)) *
-              ICE_ITR_ADAPTIVE_MIN_INC;
-
-       if ((itr & ICE_ITR_MASK) > ICE_ITR_ADAPTIVE_MAX_USECS) {
-               itr &= ICE_ITR_ADAPTIVE_LATENCY;
-               itr += ICE_ITR_ADAPTIVE_MAX_USECS;
-       }
+       /* based on checks above packets cannot be 0 so division is safe */
+       itr = ice_adjust_itr_by_size_and_speed(q_vector->vsi->port_info,
+                                              bytes / packets, itr);
 
 clear_counts:
        /* write back value */