diff mbox series

[S22,08/16] ice: Only bump Rx tail and release buffers once per napi_poll

Message ID 20190626092027.52845-8-anthony.l.nguyen@intel.com
State Accepted
Delegated to: Jeff Kirsher
Headers show
Series [S22,01/16] ice: add lp_advertising flow control support | expand

Commit Message

Tony Nguyen June 26, 2019, 9:20 a.m. UTC 
From: Brett Creeley <brett.creeley@intel.com>

Currently we bump the Rx tail and release/give buffers to hardware every
16 descriptors. This causes us to bump Rx tail up to 4 times per
napi_poll call. Also we are always bumping tail on an odd index and this
is a problem because hardware ignores the lower 3 bits in the QRX_TAIL
register. This is making it so hardware sees tail bumps only every 8
descriptors. Instead lets only bump Rx tail once per napi_poll if
the value aligns with hardwares expectations of the lower 3 bits being
cleared. Also only release/give Rx buffers once per napi_poll call.

Signed-off-by: Brett Creeley <brett.creeley@intel.com>
---
 drivers/net/ethernet/intel/ice/ice_txrx.c | 41 +++++++++++++++--------
 1 file changed, 27 insertions(+), 14 deletions(-)

Comments

Bowers, AndrewX July 2, 2019, 10:53 p.m. UTC | #1 
> -----Original Message-----
> From: Intel-wired-lan [mailto:intel-wired-lan-bounces@osuosl.org] On
> Behalf Of Tony Nguyen
> Sent: Wednesday, June 26, 2019 2:20 AM
> To: intel-wired-lan@lists.osuosl.org
> Cc: Creeley, Brett <brett.creeley@intel.com>
> Subject: [Intel-wired-lan] [PATCH S22 08/16] ice: Only bump Rx tail and
> release buffers once per napi_poll
> 
> From: Brett Creeley <brett.creeley@intel.com>
> 
> Currently we bump the Rx tail and release/give buffers to hardware every
> 16 descriptors. This causes us to bump Rx tail up to 4 times per napi_poll call.
> Also we are always bumping tail on an odd index and this is a problem
> because hardware ignores the lower 3 bits in the QRX_TAIL register. This is
> making it so hardware sees tail bumps only every 8 descriptors. Instead lets
> only bump Rx tail once per napi_poll if the value aligns with hardwares
> expectations of the lower 3 bits being cleared. Also only release/give Rx
> buffers once per napi_poll call.
> 
> Signed-off-by: Brett Creeley <brett.creeley@intel.com>
> ---
>  drivers/net/ethernet/intel/ice/ice_txrx.c | 41 +++++++++++++++--------
>  1 file changed, 27 insertions(+), 14 deletions(-)

Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
diff mbox series

Patch

diff --git a/drivers/net/ethernet/intel/ice/ice_txrx.c b/drivers/net/ethernet/intel/ice/ice_txrx.c
index d76237d82009..c604b2d192b1 100644
--- a/drivers/net/ethernet/intel/ice/ice_txrx.c
+++ b/drivers/net/ethernet/intel/ice/ice_txrx.c
@@ -428,18 +428,28 @@  static
 #endif
 void ice_release_rx_desc(struct ice_ring *rx_ring, u32 val)
 {
+	u16 prev_ntu = rx_ring->next_to_use;
+
 	rx_ring->next_to_use = val;
 
 	/* update next to alloc since we have filled the ring */
 	rx_ring->next_to_alloc = val;
 
-	/* Force memory writes to complete before letting h/w
-	 * know there are new descriptors to fetch. (Only
-	 * applicable for weak-ordered memory model archs,
-	 * such as IA-64).
+	/* QRX_TAIL will be updated with any tail value, but hardware ignores
+	 * the lower 3 bits. This makes it so we only bump tail on meaningful
+	 * boundaries. Also, this allows us to bump tail on intervals of 8 up to
+	 * the budget depending on the current traffic load.
 	 */
-	wmb();
-	writel(val, rx_ring->tail);
+	val &= ~0x7;
+	if (prev_ntu != val) {
+		/* Force memory writes to complete before letting h/w
+		 * know there are new descriptors to fetch. (Only
+		 * applicable for weak-ordered memory model archs,
+		 * such as IA-64).
+		 */
+		wmb();
+		writel(val, rx_ring->tail);
+	}
 }
 
 /**
@@ -697,7 +707,13 @@  ice_alloc_mapped_page(struct ice_ring *rx_ring, struct ice_rx_buf *bi)
  * @rx_ring: ring to place buffers on
  * @cleaned_count: number of buffers to replace
  *
- * Returns false if all allocations were successful, true if any fail
+ * Returns false if all allocations were successful, true if any fail. Returning
+ * true signals to the caller that we didn't replace cleaned_count buffers and
+ * there is more work to do.
+ *
+ * First, try to clean "cleaned_count" Rx buffers. Then refill the cleaned Rx
+ * buffers. Then bump tail at most one time. Grouping like this lets us avoid
+ * multiple tail writes per call.
  */
 bool ice_alloc_rx_bufs(struct ice_ring *rx_ring, u16 cleaned_count)
 {
@@ -1253,8 +1269,8 @@  static int ice_clean_rx_irq(struct ice_ring *rx_ring, int budget)
 	u16 cleaned_count = ICE_DESC_UNUSED(rx_ring);
 	unsigned int xdp_res, xdp_xmit = 0;
 	struct bpf_prog *xdp_prog;
-	bool failure = false;
 	struct xdp_buff xdp;
+	bool failure;
 
 	xdp.rxq = &rx_ring->xdp_rxq;
 
@@ -1268,12 +1284,6 @@  static int ice_clean_rx_irq(struct ice_ring *rx_ring, int budget)
 		u16 vlan_tag = 0;
 		u8 rx_ptype;
 
-		/* return some buffers to hardware, one at a time is too slow */
-		if (cleaned_count >= ICE_RX_BUF_WRITE) {
-			failure = failure ||
-				  ice_alloc_rx_bufs(rx_ring, cleaned_count);
-			cleaned_count = 0;
-		}
 
 		/* get the Rx desc from Rx ring based on 'next_to_clean' */
 		rx_desc = ICE_RX_DESC(rx_ring, rx_ring->next_to_clean);
@@ -1385,6 +1395,9 @@  static int ice_clean_rx_irq(struct ice_ring *rx_ring, int budget)
 		total_rx_pkts++;
 	}
 
+	/* return up to cleaned_count buffers to hardware */
+	failure = ice_alloc_rx_bufs(rx_ring, cleaned_count);
+
 	ice_finalize_xdp_rx(rx_ring, xdp_xmit);
 
 	/* update queue and vector specific stats */