Skip to contentSkip to conversation

net: stmmac: Performance improvements in Multi-Queue

PATCH Superseded Jose Abreu 3 participants 6 messages Feb 15, 2019 – Feb 16, 2019

Subsystems (4)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Jose Abreu

author
#0 lore raw

cover net: stmmac: Performance improvements in Multi-Queue

Tested in XGMAC2 and GMAC5. I guess 1/3 can be backported but besides being a kind of "not that small" refactoring it's also for a scenario that no-one complained yet ... Advice ?

Cc Florian Fainelli <[email protected]>
Cc Joao Pinto <[email protected]>
Cc David S. Miller <[email protected]>
Cc Giuseppe Cavallaro <[email protected]>
Cc Alexandre Torgue <[email protected]>

Jose Abreu (3):
net: stmmac: Fix NAPI poll in TX path when in multi-queue
net: stmmac: dwmac4: Also use TBU interrupt to clean TX path
net: stmmac: dwxgmac2: Also use TBU interrupt to clean TX path

 drivers/net/ethernet/stmicro/stmmac/dwmac4_lib.c   |  24 ++---
 drivers/net/ethernet/stmicro/stmmac/dwxgmac2.h     |   4 +-
 drivers/net/ethernet/stmicro/stmmac/dwxgmac2_dma.c |   8 +-
 drivers/net/ethernet/stmicro/stmmac/stmmac.h       |   5 +-
 drivers/net/ethernet/stmicro/stmmac/stmmac_main.c  | 112 ++++++++++++---------
 5 files changed, 77 insertions(+), 76 deletions(-)
Signature · 1 line
2.7.4
author
#1 lore raw

PATCH 1/3 net: stmmac: Fix NAPI poll in TX path when in multi-queue

Commit 8fce33317023 introduced the concept of NAPI per-channel and independent cleaning of TX path.

This is currently breaking performance in some cases. The scenario happens when all packets are being received in Queue 0 but the TX is performed in Queue != 0.

Fix this by using different NAPI instances per each TX and RX queue, as suggested by Florian.

Cc Florian Fainelli <[email protected]>
Cc Joao Pinto <[email protected]>
Cc David S. Miller <[email protected]>
Cc Giuseppe Cavallaro <[email protected]>
Cc Alexandre Torgue <[email protected]>
---
 drivers/net/ethernet/stmicro/stmmac/stmmac.h      |   5 +-
 drivers/net/ethernet/stmicro/stmmac/stmmac_main.c | 112 ++++++++++++----------
 2 files changed, 64 insertions(+), 53 deletions(-)
drivers/net/ethernet/stmicro/stmmac/stmmac.h +2 -3
index 63e1064b27a2..e697ecd9b0a6 100644
@@ -78,11 +78,10 @@ struct stmmac_rx_queue {
 };
 
 struct stmmac_channel {
-	struct napi_struct napi ____cacheline_aligned_in_smp;
+	struct napi_struct rx_napi ____cacheline_aligned_in_smp;
+	struct napi_struct tx_napi ____cacheline_aligned_in_smp;
 	struct stmmac_priv *priv_data;
 	u32 index;
-	int has_rx;
-	int has_tx;
 };
 
 struct stmmac_tc_entry {
drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +62 -50

181 lines not on this page · raw message

Signature · 1 line
2.7.4
author
#4 lore raw

PATCH 2/3 net: stmmac: dwmac4: Also use TBU interrupt to clean TX path

In reply to Jose Abreu

TBU interrupt is a normal interrupt and can be used to trigger the cleaning of TX path. Lets check if it's active in DMA interrupt handler.

While at it, refactor a little bit the function:
- Don't check if RI is enabled because at function exit we will only clear the interrupts that are enabled so, no event will be missed.

In my tests with GMAC5 this increased performance.

Cc Joao Pinto <[email protected]>
Cc David S. Miller <[email protected]>
Cc Giuseppe Cavallaro <[email protected]>
Cc Alexandre Torgue <[email protected]>
---
 drivers/net/ethernet/stmicro/stmmac/dwmac4_lib.c | 24 +++++++-----------------
 1 file changed, 7 insertions(+), 17 deletions(-)
drivers/net/ethernet/stmicro/stmmac/dwmac4_lib.c +7 -17
index 49f5687879df..545cb9c47433 100644
@@ -124,9 +124,9 @@ void dwmac4_disable_dma_irq(void __iomem *ioaddr, u32 chan)
 int dwmac4_dma_interrupt(void __iomem *ioaddr,
 			 struct stmmac_extra_stats *x, u32 chan)
 {
-	int ret = 0;
-
 	u32 intr_status = readl(ioaddr + DMA_CHAN_STATUS(chan));
+	u32 intr_en = readl(ioaddr + DMA_CHAN_INTR_ENA(chan));
+	int ret = 0;
 
 	/* ABNORMAL interrupts */
 	if (unlikely(intr_status & DMA_CHAN_STATUS_AIS)) {
@@ -151,16 +151,11 @@ int dwmac4_dma_interrupt(void __iomem *ioaddr,
 	if (likely(intr_status & DMA_CHAN_STATUS_NIS)) {
 		x->normal_irq_n++;
 		if (likely(intr_status & DMA_CHAN_STATUS_RI)) {
-			u32 value;
-
-			value = readl(ioaddr + DMA_CHAN_INTR_ENA(chan));
-			/* to schedule NAPI on real RIE event. */
-			if (likely(value & DMA_CHAN_INTR_ENA_RIE)) {
-				x->rx_normal_irq_n++;
-				ret |= handle_rx;
-			}
+			x->rx_normal_irq_n++;
+			ret |= handle_rx;
 		}
-		if (likely(intr_status & DMA_CHAN_STATUS_TI)) {
+		if (likely(intr_status & (DMA_CHAN_STATUS_TI |
+					  DMA_CHAN_STATUS_TBU))) {
 			x->tx_normal_irq_n++;
 			ret |= handle_tx;
 		}
@@ -168,12 +163,7 @@ int dwmac4_dma_interrupt(void __iomem *ioaddr,
 			x->rx_early_irq++;
 	}
 
-	/* Clear the interrupt by writing a logic 1 to the chanX interrupt
-	 * status [21-0] expect reserved bits [5-3]
-	 */
-	writel((intr_status & 0x3fffc7),
-	       ioaddr + DMA_CHAN_STATUS(chan));
-
+	writel(intr_status & intr_en, ioaddr + DMA_CHAN_STATUS(chan));
 	return ret;
 }
 
Signature · 1 line
2.7.4
author
#5 lore raw

PATCH 3/3 net: stmmac: dwxgmac2: Also use TBU interrupt to clean TX path

In reply to Jose Abreu

TBU interrupt is a normal interrupt and can be used to trigger the cleaning of TX path. Lets check if it's active in DMA interrupt handler.

While at it, refactor a little bit the function:
- Don't check if RI is enabled because at function exit we will only clear the interrupts that are enabled so, no event will be missed.

In my tests withe XGMAC2 this increased performance.

Cc Joao Pinto <[email protected]>
Cc David S. Miller <[email protected]>
Cc Giuseppe Cavallaro <[email protected]>
Cc Alexandre Torgue <[email protected]>
---
 drivers/net/ethernet/stmicro/stmmac/dwxgmac2.h     | 4 +++-
 drivers/net/ethernet/stmicro/stmmac/dwxgmac2_dma.c | 8 +++-----
 2 files changed, 6 insertions(+), 6 deletions(-)
drivers/net/ethernet/stmicro/stmmac/dwxgmac2.h +3 -1
index d6bb953685fa..37d5e6fe7473 100644
@@ -193,9 +193,10 @@
 #define XGMAC_AIE			BIT(14)
 #define XGMAC_RBUE			BIT(7)
 #define XGMAC_RIE			BIT(6)
+#define XGMAC_TBUE			BIT(2)
 #define XGMAC_TIE			BIT(0)
 #define XGMAC_DMA_INT_DEFAULT_EN	(XGMAC_NIE | XGMAC_AIE | XGMAC_RBUE | \
-					XGMAC_RIE | XGMAC_TIE)
+					XGMAC_RIE | XGMAC_TBUE | XGMAC_TIE)
 #define XGMAC_DMA_CH_Rx_WATCHDOG(x)	(0x0000313c + (0x80 * (x)))
 #define XGMAC_RWT			GENMASK(7, 0)
 #define XGMAC_DMA_CH_STATUS(x)		(0x00003160 + (0x80 * (x)))
@@ -204,6 +205,7 @@
 #define XGMAC_FBE			BIT(12)
 #define XGMAC_RBU			BIT(7)
 #define XGMAC_RI			BIT(6)
+#define XGMAC_TBU			BIT(2)
 #define XGMAC_TPS			BIT(1)
 #define XGMAC_TI			BIT(0)
 
drivers/net/ethernet/stmicro/stmmac/dwxgmac2_dma.c +3 -5
index c5e25580a43f..2ba712b48a89 100644
@@ -283,12 +283,10 @@ static int dwxgmac2_dma_interrupt(void __iomem *ioaddr,
 		x->normal_irq_n++;
 
 		if (likely(intr_status & XGMAC_RI)) {
-			if (likely(intr_en & XGMAC_RIE)) {
-				x->rx_normal_irq_n++;
-				ret |= handle_rx;
-			}
+			x->rx_normal_irq_n++;
+			ret |= handle_rx;
 		}
-		if (likely(intr_status & XGMAC_TI)) {
+		if (likely(intr_status & (XGMAC_TI | XGMAC_TBU))) {
 			x->tx_normal_irq_n++;
 			ret |= handle_tx;
 		}
Signature · 1 line
2.7.4
#2 lore raw

Re: [PATCH net-next 1/3] net: stmmac: Fix NAPI poll in TX path when in multi-queue

In reply to Jose Abreu
Quoted 17 lines · Jose Abreu

Commit 8fce33317023 introduced the concept of NAPI per-channel and independent cleaning of TX path.

This is currently breaking performance in some cases. The scenario happens when all packets are being received in Queue 0 but the TX is performed in Queue != 0.

Fix this by using different NAPI instances per each TX and RX queue, as suggested by Florian.

Cc Florian Fainelli <[email protected]>
Cc Joao Pinto <[email protected]>
Cc David S. Miller <[email protected]>
Cc Giuseppe Cavallaro <[email protected]>
Cc Alexandre Torgue <[email protected]>

---

[snip]

drivers/net/ethernet/stmicro/stmmac/stmmac_main.c static int stmmac_rx(struct stmmac_priv *priv, int limit, u32 queue) · line 3552

Quoted 6 lines
drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +3 -2
-	if (work_done < budget && napi_complete_done(napi, work_done)) {
-		int stat;
+	priv->xstats.napi_poll++;
 
+	work_done = stmmac_tx_clean(priv, budget, chan);
+	if (work_done < budget && napi_complete_done(napi, work_done))

You should not be bounding your TX queue against the NAPI budge, it should run unbound and clean as much as it can, which could be the entire ring size if that is how many packets you pushed between interrupts. That could be the cause of poor performance as well.

Signature · 1 line
Florian
#3 lore raw

Re: [PATCH net-next 1/3] net: stmmac: Fix NAPI poll in TX path when in multi-queue

Quoted 25 lines · Florian Fainelli
Quoted 17 lines · Jose Abreu

Commit 8fce33317023 introduced the concept of NAPI per-channel and independent cleaning of TX path.

This is currently breaking performance in some cases. The scenario happens when all packets are being received in Queue 0 but the TX is performed in Queue != 0.

Fix this by using different NAPI instances per each TX and RX queue, as suggested by Florian.

Cc Florian Fainelli <[email protected]>
Cc Joao Pinto <[email protected]>
Cc David S. Miller <[email protected]>
Cc Giuseppe Cavallaro <[email protected]>
Cc Alexandre Torgue <[email protected]>

---

[snip]

Quoted 4 lines
drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +1 -2
-	if (work_done < budget && napi_complete_done(napi, work_done)) {
-		int stat;
+	priv->xstats.napi_poll++;
Quoted 7 lines

drivers/net/ethernet/stmicro/stmmac/stmmac_main.c static int stmmac_rx(struct stmmac_priv *priv, int limit, u32 queue) · line 3552

Quoted 2 lines
drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +2 -0
+	work_done = stmmac_tx_clean(priv, budget, chan);
+	if (work_done < budget && napi_complete_done(napi, work_done))

You should not be bounding your TX queue against the NAPI budge, it should run unbound and clean as much as it can, which could be the entire ring size if that is how many packets you pushed between interrupts. That could be the cause of poor performance as well.

Won't returning the budget value cause this napi_poll routine to be called again, where the driver can continue to clean TX packets? I thought this just gives other drivers the opportunity to run their napi_poll routines in between so as not to be starved.

Thanks,
Tom

Quoted 1 line