[PATCH 0/3] net: ag71xx: NAPI weight, MMIO drain, and TX OOM fix

[PATCH 0/3] net: ag71xx: NAPI weight, MMIO drain, and TX OOM fix

[Rosen Penev]

Drop the custom NAPI weight in favour of the core default, remove
unnecessary MMIO read-back drains from register writes for a small
performance gain, and fix a TX stall that occurs when the RX ring
runs out of buffers.

Rosen Penev (3):
  net: ag71xx: Use default NAPI weight
  net: ag71xx: remove MMIO read-back drain from register writes
  net: ag71xx: re-enable TX interrupts on RX OOM

 drivers/net/ethernet/atheros/ag71xx.c | 25 ++++++++-----------------
 1 file changed, 8 insertions(+), 17 deletions(-)

-- 
2.54.0

[PATCH 1/3] net: ag71xx: Use default NAPI weight

[Rosen Penev]

Drop the custom NAPI weight and let netif_napi_add() use the core
default.

The old comment claimed larger weights wasted cycles and cache, but
recent RX and TX path changes altered those tradeoffs. Keep the driver
on the standard default unless measurements show a device-specific value
is needed.

iperf3:

Before:

[ ID][Role] Interval           Transfer     Bitrate         Retr
[  5][TX-C]   0.00-10.00  sec   327 MBytes   274 Mbits/sec   31            sender
[  5][TX-C]   0.00-10.00  sec   325 MBytes   273 Mbits/sec                  receiver
[  7][RX-C]   0.00-10.00  sec   170 MBytes   143 Mbits/sec    0            sender
[  7][RX-C]   0.00-10.00  sec   170 MBytes   143 Mbits/sec                  receiver

After:

[ ID][Role] Interval           Transfer     Bitrate         Retr
[  5][TX-C]   0.00-10.00  sec   341 MBytes   286 Mbits/sec   30            sender
[  5][TX-C]   0.00-10.00  sec   338 MBytes   284 Mbits/sec                  receiver
[  7][RX-C]   0.00-10.00  sec   184 MBytes   154 Mbits/sec    0            sender
[  7][RX-C]   0.00-10.00  sec   184 MBytes   154 Mbits/sec                  receiver

Assisted-by: Codex:GPT-5.5
Signed-off-by: Rosen Penev <rosenp@gmail.com>
---
 drivers/net/ethernet/atheros/ag71xx.c | 8 +-------
 1 file changed, 1 insertion(+), 7 deletions(-)

diff --git a/drivers/net/ethernet/atheros/ag71xx.c b/drivers/net/ethernet/atheros/ag71xx.c
index ac4eadb9190a..1def2ad4c5ce 100644
--- a/drivers/net/ethernet/atheros/ag71xx.c
+++ b/drivers/net/ethernet/atheros/ag71xx.c
@@ -40,11 +40,6 @@
 #include <linux/io.h>
 #include <net/selftests.h>
 
-/* For our NAPI weight bigger does *NOT* mean better - it means more
- * D-cache misses and lots more wasted cycles than we'll ever
- * possibly gain from saving instructions.
- */
-#define AG71XX_NAPI_WEIGHT	32
 #define AG71XX_OOM_REFILL	(1 + HZ / 10)
 
 #define AG71XX_INT_ERR	(AG71XX_INT_RX_BE | AG71XX_INT_TX_BE)
@@ -1913,8 +1908,7 @@ static int ag71xx_probe(struct platform_device *pdev)
 		return err;
 	}
 
-	netif_napi_add_weight(ndev, &ag->napi, ag71xx_poll,
-			      AG71XX_NAPI_WEIGHT);
+	netif_napi_add(ndev, &ag->napi, ag71xx_poll);
 
 	ag71xx_wr(ag, AG71XX_REG_MAC_CFG1, 0);
 
-- 
2.54.0

[PATCH 2/3] net: ag71xx: remove MMIO read-back drain from register writes

[Rosen Penev]

ag71xx_wr(), ag71xx_sb() and ag71xx_cb() issue a synchronous
ioread32() after every iowrite32() under the assumption that the
write must be flushed to the device before any subsequent operation.

On the ath79 on-chip MMIO bus the uncached __raw_writel used by
iowrite32() is non-posted - the write completes before the next
instruction executes.  The extra ioread32() therefore serves no
ordering purpose and costs at least one uncached bus round-trip
per call.

Remove the read-back and the associated comments, leaving the
barrier that iowrite32() already provides.

Small iperf3 improvement:

Before:

[ ID][Role] Interval           Transfer     Bitrate         Retr
[  5][TX-C]   0.00-10.00  sec   343 MBytes   288 Mbits/sec   32            sender
[  5][TX-C]   0.00-10.00  sec   341 MBytes   286 Mbits/sec                  receiver
[  7][RX-C]   0.00-10.00  sec   183 MBytes   153 Mbits/sec    0            sender
[  7][RX-C]   0.00-10.00  sec   183 MBytes   153 Mbits/sec                  receiver

After:

[ ID][Role] Interval           Transfer     Bitrate         Retr
[  5][TX-C]   0.00-10.00  sec   349 MBytes   292 Mbits/sec   28            sender
[  5][TX-C]   0.00-10.00  sec   346 MBytes   291 Mbits/sec                  receiver
[  7][RX-C]   0.00-10.00  sec   180 MBytes   151 Mbits/sec    0            sender
[  7][RX-C]   0.00-10.00  sec   180 MBytes   151 Mbits/sec                  receiver

Assisted-by: opencode:big-pickle
Signed-off-by: Rosen Penev <rosenp@gmail.com>
---
 drivers/net/ethernet/atheros/ag71xx.c | 16 ++++++----------
 1 file changed, 6 insertions(+), 10 deletions(-)

diff --git a/drivers/net/ethernet/atheros/ag71xx.c b/drivers/net/ethernet/atheros/ag71xx.c
index 1def2ad4c5ce..77f8e75e98ac 100644
--- a/drivers/net/ethernet/atheros/ag71xx.c
+++ b/drivers/net/ethernet/atheros/ag71xx.c
@@ -397,11 +397,13 @@ static bool ag71xx_is(struct ag71xx *ag, enum ag71xx_type type)
 	return ag->dcfg->type == type;
 }
 
+/* ath79 on-chip MMIO bus is non-posted - iowrite32/iowrite32be completes
+ * before the next instruction executes. No read-back drain is needed.
+ */
+
 static void ag71xx_wr(struct ag71xx *ag, unsigned int reg, u32 value)
 {
 	iowrite32(value, ag->mac_base + reg);
-	/* flush write */
-	(void)ioread32(ag->mac_base + reg);
 }
 
 static u32 ag71xx_rr(struct ag71xx *ag, unsigned int reg)
@@ -411,22 +413,16 @@ static u32 ag71xx_rr(struct ag71xx *ag, unsigned int reg)
 
 static void ag71xx_sb(struct ag71xx *ag, unsigned int reg, u32 mask)
 {
-	void __iomem *r;
+	void __iomem *r = ag->mac_base + reg;
 
-	r = ag->mac_base + reg;
 	iowrite32(ioread32(r) | mask, r);
-	/* flush write */
-	(void)ioread32(r);
 }
 
 static void ag71xx_cb(struct ag71xx *ag, unsigned int reg, u32 mask)
 {
-	void __iomem *r;
+	void __iomem *r = ag->mac_base + reg;
 
-	r = ag->mac_base + reg;
 	iowrite32(ioread32(r) & ~mask, r);
-	/* flush write */
-	(void)ioread32(r);
 }
 
 static void ag71xx_int_enable(struct ag71xx *ag, u32 ints)
-- 
2.54.0

[PATCH 3/3] net: ag71xx: re-enable TX interrupts on RX OOM

[Rosen Penev]

When the RX ring runs out of buffers the OOM path calls
napi_complete() but does not re-enable interrupts.  Because
INT_POLL masks both RX and TX events, TX completions are
blocked until the OOM timer fires (~100 ms), which can cause
TX ring exhaustion and dropped packets.

Restore TX interrupt delivery so the engine can drain completed
descriptors while NAPI waits for RX buffer memory.

Assisted-by: opencode:big-pickle
Signed-off-by: Rosen Penev <rosenp@gmail.com>
---
 drivers/net/ethernet/atheros/ag71xx.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/net/ethernet/atheros/ag71xx.c b/drivers/net/ethernet/atheros/ag71xx.c
index 77f8e75e98ac..a4e52bc42091 100644
--- a/drivers/net/ethernet/atheros/ag71xx.c
+++ b/drivers/net/ethernet/atheros/ag71xx.c
@@ -1720,6 +1720,7 @@ static int ag71xx_poll(struct napi_struct *napi, int limit)
 
 	mod_timer(&ag->oom_timer, jiffies + AG71XX_OOM_REFILL);
 	napi_complete(napi);
+	ag71xx_int_enable(ag, AG71XX_INT_TX);
 	return 0;
 }
 
-- 
2.54.0