i40evf: Force Tx writeback on ITR

This patch forces Tx descriptor writebacks on ITR by kicking
off the SWINT interrupt when we notice that there are non-cache-aligned
Tx descriptors waiting in the ring while interrupts are disabled
under NAPI.

Change-ID: dd6d9675629bf266c7515ad7a201394618c35444
Signed-off-by: Anjali Singhai Jain <anjali.singhai@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This commit is contained in:
Anjali Singhai Jain 2015-01-10 01:07:19 +00:00 committed by Jeff Kirsher
parent 88086e5d0a
commit c29af37fd5
3 changed files with 39 additions and 4 deletions

View File

@ -836,8 +836,8 @@ static void i40e_force_wb(struct i40e_vsi *vsi, struct i40e_q_vector *q_vector)
{
u32 val = I40E_PFINT_DYN_CTLN_INTENA_MASK |
I40E_PFINT_DYN_CTLN_SWINT_TRIG_MASK |
I40E_PFINT_DYN_CTLN_SW_ITR_INDX_ENA_MASK
/* allow 00 to be written to the index */;
I40E_PFINT_DYN_CTLN_SW_ITR_INDX_ENA_MASK;
/* allow 00 to be written to the index */
wr32(&vsi->back->hw,
I40E_PFINT_DYN_CTLN(q_vector->v_idx + vsi->base_vector - 1),

View File

@ -192,6 +192,8 @@ static inline u32 i40e_get_head(struct i40e_ring *tx_ring)
return le32_to_cpu(*(volatile __le32 *)head);
}
#define WB_STRIDE 0x3
/**
* i40e_clean_tx_irq - Reclaim resources after transmit completes
* @tx_ring: tx ring to clean
@ -293,6 +295,14 @@ static bool i40e_clean_tx_irq(struct i40e_ring *tx_ring, int budget)
tx_ring->q_vector->tx.total_bytes += total_bytes;
tx_ring->q_vector->tx.total_packets += total_packets;
if (budget &&
!((i & WB_STRIDE) == WB_STRIDE) &&
!test_bit(__I40E_DOWN, &tx_ring->vsi->state) &&
(I40E_DESC_UNUSED(tx_ring) != tx_ring->count))
tx_ring->arm_wb = true;
else
tx_ring->arm_wb = false;
if (check_for_tx_hang(tx_ring) && i40e_check_tx_hang(tx_ring)) {
/* schedule immediate reset if we believe we hung */
dev_info(tx_ring->dev, "Detected Tx Unit Hang\n"
@ -343,6 +353,24 @@ static bool i40e_clean_tx_irq(struct i40e_ring *tx_ring, int budget)
return budget > 0;
}
/**
* i40e_force_wb -Arm hardware to do a wb on noncache aligned descriptors
* @vsi: the VSI we care about
* @q_vector: the vector on which to force writeback
*
**/
static void i40e_force_wb(struct i40e_vsi *vsi, struct i40e_q_vector *q_vector)
{
u32 val = I40E_VFINT_DYN_CTLN_INTENA_MASK |
I40E_VFINT_DYN_CTLN_SWINT_TRIG_MASK |
I40E_VFINT_DYN_CTLN_SW_ITR_INDX_ENA_MASK;
/* allow 00 to be written to the index */
wr32(&vsi->back->hw,
I40E_VFINT_DYN_CTLN1(q_vector->v_idx + vsi->base_vector - 1),
val);
}
/**
* i40e_set_new_dynamic_itr - Find new ITR level
* @rc: structure containing ring performance data
@ -1065,6 +1093,7 @@ int i40evf_napi_poll(struct napi_struct *napi, int budget)
struct i40e_vsi *vsi = q_vector->vsi;
struct i40e_ring *ring;
bool clean_complete = true;
bool arm_wb = false;
int budget_per_ring;
if (test_bit(__I40E_DOWN, &vsi->state)) {
@ -1075,8 +1104,10 @@ int i40evf_napi_poll(struct napi_struct *napi, int budget)
/* Since the actual Tx work is minimal, we can give the Tx a larger
* budget and be more aggressive about cleaning up the Tx descriptors.
*/
i40e_for_each_ring(ring, q_vector->tx)
i40e_for_each_ring(ring, q_vector->tx) {
clean_complete &= i40e_clean_tx_irq(ring, vsi->work_limit);
arm_wb |= ring->arm_wb;
}
/* We attempt to distribute budget to each Rx queue fairly, but don't
* allow the budget to go below 1 because that would exit polling early.
@ -1087,8 +1118,11 @@ int i40evf_napi_poll(struct napi_struct *napi, int budget)
clean_complete &= i40e_clean_rx_irq(ring, budget_per_ring);
/* If work not completed, return budget and polling will return */
if (!clean_complete)
if (!clean_complete) {
if (arm_wb)
i40e_force_wb(vsi, q_vector);
return budget;
}
/* Work is done so exit the polling mode and re-enable the interrupt */
napi_complete(napi);

View File

@ -238,6 +238,7 @@ struct i40e_ring {
u8 atr_count;
bool ring_active; /* is ring online or not */
bool arm_wb; /* do something to arm write back */
/* stats structs */
struct i40e_queue_stats stats;