diff options
Diffstat (limited to 'drivers/net/ethernet/intel/i40evf/i40e_txrx.c')
-rw-r--r-- | drivers/net/ethernet/intel/i40evf/i40e_txrx.c | 163 |
1 files changed, 110 insertions, 53 deletions
diff --git a/drivers/net/ethernet/intel/i40evf/i40e_txrx.c b/drivers/net/ethernet/intel/i40evf/i40e_txrx.c index 97493a4164eb..47e9a90d6b10 100644 --- a/drivers/net/ethernet/intel/i40evf/i40e_txrx.c +++ b/drivers/net/ethernet/intel/i40evf/i40e_txrx.c @@ -318,6 +318,8 @@ static void i40evf_force_wb(struct i40e_vsi *vsi, struct i40e_q_vector *q_vector * i40e_set_new_dynamic_itr - Find new ITR level * @rc: structure containing ring performance data * + * Returns true if ITR changed, false if not + * * Stores a new ITR value based on packets and byte counts during * the last interrupt. The advantage of per interrupt computation * is faster updates and more accurate ITR for the current traffic @@ -326,21 +328,32 @@ static void i40evf_force_wb(struct i40e_vsi *vsi, struct i40e_q_vector *q_vector * testing data as well as attempting to minimize response time * while increasing bulk throughput. **/ -static void i40e_set_new_dynamic_itr(struct i40e_ring_container *rc) +static bool i40e_set_new_dynamic_itr(struct i40e_ring_container *rc) { enum i40e_latency_range new_latency_range = rc->latency_range; + struct i40e_q_vector *qv = rc->ring->q_vector; u32 new_itr = rc->itr; int bytes_per_int; + int usecs; if (rc->total_packets == 0 || !rc->itr) - return; + return false; /* simple throttlerate management - * 0-10MB/s lowest (100000 ints/s) + * 0-10MB/s lowest (50000 ints/s) * 10-20MB/s low (20000 ints/s) - * 20-1249MB/s bulk (8000 ints/s) + * 20-1249MB/s bulk (18000 ints/s) + * > 40000 Rx packets per second (8000 ints/s) + * + * The math works out because the divisor is in 10^(-6) which + * turns the bytes/us input value into MB/s values, but + * make sure to use usecs, as the register values written + * are in 2 usec increments in the ITR registers, and make sure + * to use the smoothed values that the countdown timer gives us. */ - bytes_per_int = rc->total_bytes / rc->itr; + usecs = (rc->itr << 1) * ITR_COUNTDOWN_START; + bytes_per_int = rc->total_bytes / usecs; + switch (new_latency_range) { case I40E_LOWEST_LATENCY: if (bytes_per_int > 10) @@ -353,35 +366,52 @@ static void i40e_set_new_dynamic_itr(struct i40e_ring_container *rc) new_latency_range = I40E_LOWEST_LATENCY; break; case I40E_BULK_LATENCY: - if (bytes_per_int <= 20) - new_latency_range = I40E_LOW_LATENCY; - break; + case I40E_ULTRA_LATENCY: default: if (bytes_per_int <= 20) new_latency_range = I40E_LOW_LATENCY; break; } + + /* this is to adjust RX more aggressively when streaming small + * packets. The value of 40000 was picked as it is just beyond + * what the hardware can receive per second if in low latency + * mode. + */ +#define RX_ULTRA_PACKET_RATE 40000 + + if ((((rc->total_packets * 1000000) / usecs) > RX_ULTRA_PACKET_RATE) && + (&qv->rx == rc)) + new_latency_range = I40E_ULTRA_LATENCY; + rc->latency_range = new_latency_range; switch (new_latency_range) { case I40E_LOWEST_LATENCY: - new_itr = I40E_ITR_100K; + new_itr = I40E_ITR_50K; break; case I40E_LOW_LATENCY: new_itr = I40E_ITR_20K; break; case I40E_BULK_LATENCY: + new_itr = I40E_ITR_18K; + break; + case I40E_ULTRA_LATENCY: new_itr = I40E_ITR_8K; break; default: break; } - if (new_itr != rc->itr) - rc->itr = new_itr; - rc->total_bytes = 0; rc->total_packets = 0; + + if (new_itr != rc->itr) { + rc->itr = new_itr; + return true; + } + + return false; } /* @@ -1187,6 +1217,21 @@ static int i40e_clean_rx_irq_1buf(struct i40e_ring *rx_ring, int budget) return total_rx_packets; } +static u32 i40e_buildreg_itr(const int type, const u16 itr) +{ + u32 val; + + val = I40E_VFINT_DYN_CTLN1_INTENA_MASK | + I40E_VFINT_DYN_CTLN1_CLEARPBA_MASK | + (type << I40E_VFINT_DYN_CTLN1_ITR_INDX_SHIFT) | + (itr << I40E_VFINT_DYN_CTLN1_INTERVAL_SHIFT); + + return val; +} + +/* a small macro to shorten up some long lines */ +#define INTREG I40E_VFINT_DYN_CTLN1 + /** * i40e_update_enable_itr - Update itr and re-enable MSIX interrupt * @vsi: the VSI we care about @@ -1197,55 +1242,67 @@ static inline void i40e_update_enable_itr(struct i40e_vsi *vsi, struct i40e_q_vector *q_vector) { struct i40e_hw *hw = &vsi->back->hw; - u16 old_itr; + bool rx = false, tx = false; + u32 rxval, txval; int vector; - u32 val; vector = (q_vector->v_idx + vsi->base_vector); + + /* avoid dynamic calculation if in countdown mode OR if + * all dynamic is disabled + */ + rxval = txval = i40e_buildreg_itr(I40E_ITR_NONE, 0); + + if (q_vector->itr_countdown > 0 || + (!ITR_IS_DYNAMIC(vsi->rx_itr_setting) && + !ITR_IS_DYNAMIC(vsi->tx_itr_setting))) { + goto enable_int; + } + if (ITR_IS_DYNAMIC(vsi->rx_itr_setting)) { - old_itr = q_vector->rx.itr; - i40e_set_new_dynamic_itr(&q_vector->rx); - if (old_itr != q_vector->rx.itr) { - val = I40E_VFINT_DYN_CTLN1_INTENA_MASK | - I40E_VFINT_DYN_CTLN1_CLEARPBA_MASK | - (I40E_RX_ITR << - I40E_VFINT_DYN_CTLN1_ITR_INDX_SHIFT) | - (q_vector->rx.itr << - I40E_VFINT_DYN_CTLN1_INTERVAL_SHIFT); - } else { - val = I40E_VFINT_DYN_CTLN1_INTENA_MASK | - I40E_VFINT_DYN_CTLN1_CLEARPBA_MASK | - (I40E_ITR_NONE << - I40E_VFINT_DYN_CTLN1_ITR_INDX_SHIFT); - } - if (!test_bit(__I40E_DOWN, &vsi->state)) - wr32(hw, I40E_VFINT_DYN_CTLN1(vector - 1), val); - } else { - i40evf_irq_enable_queues(vsi->back, 1 - << q_vector->v_idx); + rx = i40e_set_new_dynamic_itr(&q_vector->rx); + rxval = i40e_buildreg_itr(I40E_RX_ITR, q_vector->rx.itr); } if (ITR_IS_DYNAMIC(vsi->tx_itr_setting)) { - old_itr = q_vector->tx.itr; - i40e_set_new_dynamic_itr(&q_vector->tx); - if (old_itr != q_vector->tx.itr) { - val = I40E_VFINT_DYN_CTLN1_INTENA_MASK | - I40E_VFINT_DYN_CTLN1_CLEARPBA_MASK | - (I40E_TX_ITR << - I40E_VFINT_DYN_CTLN1_ITR_INDX_SHIFT) | - (q_vector->tx.itr << - I40E_VFINT_DYN_CTLN1_INTERVAL_SHIFT); + tx = i40e_set_new_dynamic_itr(&q_vector->tx); + txval = i40e_buildreg_itr(I40E_TX_ITR, q_vector->tx.itr); + } + if (rx || tx) { + /* get the higher of the two ITR adjustments and + * use the same value for both ITR registers + * when in adaptive mode (Rx and/or Tx) + */ + u16 itr = max(q_vector->tx.itr, q_vector->rx.itr); - } else { - val = I40E_VFINT_DYN_CTLN1_INTENA_MASK | - I40E_VFINT_DYN_CTLN1_CLEARPBA_MASK | - (I40E_ITR_NONE << - I40E_VFINT_DYN_CTLN1_ITR_INDX_SHIFT); - } - if (!test_bit(__I40E_DOWN, &vsi->state)) - wr32(hw, I40E_VFINT_DYN_CTLN1(vector - 1), val); - } else { - i40evf_irq_enable_queues(vsi->back, BIT(q_vector->v_idx)); + q_vector->tx.itr = q_vector->rx.itr = itr; + txval = i40e_buildreg_itr(I40E_TX_ITR, itr); + tx = true; + rxval = i40e_buildreg_itr(I40E_RX_ITR, itr); + rx = true; } + + /* only need to enable the interrupt once, but need + * to possibly update both ITR values + */ + if (rx) { + /* set the INTENA_MSK_MASK so that this first write + * won't actually enable the interrupt, instead just + * updating the ITR (it's bit 31 PF and VF) + */ + rxval |= BIT(31); + /* don't check _DOWN because interrupt isn't being enabled */ + wr32(hw, INTREG(vector - 1), rxval); + } + +enable_int: + if (!test_bit(__I40E_DOWN, &vsi->state)) + wr32(hw, INTREG(vector - 1), txval); + + if (q_vector->itr_countdown) + q_vector->itr_countdown--; + else + q_vector->itr_countdown = ITR_COUNTDOWN_START; + } /** |