[PATCH] net: cadence: macb: Fix a possible deadlock in macb_halt_tx.

There is a situation where after THALT is set high, TGO stays high as
well. Because jiffies are never updated, as we are in a context with
interrupts disabled, we never exit that loop and have a deadlock.

That deadlock was noticed on a sama5d4 device that stayed locked for days.

Use retries instead of jiffies so that the timeout really works and we do
not have a deadlock anymore.

Signed-off-by: Mathieu Othacehe <redacted>
---
 drivers/net/ethernet/cadence/macb_main.c | 11 +++++------
 1 file changed, 5 insertions(+), 6 deletions(-)

diff --git a/drivers/net/ethernet/cadence/macb_main.c b/drivers/net/ethernet/cadence/macb_main.c
index 1fe8ec37491b1..ffcf569c14f6a 100644
--- a/drivers/net/ethernet/cadence/macb_main.c
+++ b/drivers/net/ethernet/cadence/macb_main.c

@@ -997,20 +997,19 @@ static void macb_update_stats(struct macb *bp)
 
 static int macb_halt_tx(struct macb *bp)
 {
-	unsigned long	halt_time, timeout;
-	u32		status;
+	unsigned int delay_us = 250;
+	unsigned int retries = MACB_HALT_TIMEOUT / delay_us;
+	u32 status;
 
 	macb_writel(bp, NCR, macb_readl(bp, NCR) | MACB_BIT(THALT));
 
-	timeout = jiffies + usecs_to_jiffies(MACB_HALT_TIMEOUT);
 	do {
-		halt_time = jiffies;
 		status = macb_readl(bp, TSR);
 		if (!(status & MACB_BIT(TGO)))
 			return 0;
 
-		udelay(250);
-	} while (time_before(halt_time, timeout));
+		udelay(delay_us);
+	} while (retries-- > 0);
 
 	return -ETIMEDOUT;
 }

-- 
2.49.0

On 5/7/25 12:12 PM, Mathieu Othacehe wrote:

There is a situation where after THALT is set high, TGO stays high as
well. Because jiffies are never updated, as we are in a context with
interrupts disabled, we never exit that loop and have a deadlock.

That deadlock was noticed on a sama5d4 device that stayed locked for days.

Use retries instead of jiffies so that the timeout really works and we do
not have a deadlock anymore.

Signed-off-by: Mathieu Othacehe <redacted>

This looks like a fix that should target the net tree and include a
fixes tag, see Documentation/process/maintainer-netdev.rst

I think it would be better to use read_poll_timeout_atomic() instead of
sort-of open-codying it.

Thanks,

Paolo