Re: [PATCH net 2/3] net: ethernet: ti: icssg_common: set irq_disabled after disabling TX IRQ

[Siddharth Vadapalli]

On 25/02/26 05:19, Jakub Kicinski wrote:
> On Tue, 24 Feb 2026 17:54:18 +0530 Siddharth Vadapalli wrote:
> > 			CPU0						
> > 				CPU1
> > 		----------------
> > 								--------------
> > 1.	TX HARD IRQ Handler entered				NAPI TX
> > Handler is running
> > 2.	irq_disabled is
> > set							Sees irq_disabled being set
> > 3.	Starts executing disable_irq_nosync()		Invokes
> > enable_irq() for TX IRQ before its really disabled
>
> Could you resend your last email fixing the line wrap issue? It's very
> hard to read as it arrived on the list.

I have changed my email client now to fix the line wrap issue. Sorry for 
the inconvenience cause. I am repeating my earlier message below with 
proper formatting. It is in response to the following statement:

> AFAICT the flow on the Tx bug is not buggy, owner ship of the IRQ
> vector passes handler -> NAPI -> timer. I don't see how those can
> race.

The issue is seen in practice. Interrupt coalescing (hrtimer) isn't used.
The call sequence leading to the warning is:

	net_rx_action
		__napi_poll
			NAPI TX Handler

It does seem strange that the 'net_rx_action' leads to the NAPI TX Handler.
However, it is exactly this path that causes the warning, and it is due to
this that we could end up in the following situation:

                CPU0                             CPU1
   -----------------------------      -----------------------------
1. TX HARD IRQ Handler entered         NAPI TX Handler is running
2. irq_disabled is set to true         Sees irq_disabled being true
3. Calls disable_irq_nosync()          Calls enable_irq()
4. Enters disable_irq_nosync()         [WARNING: Unbalanced enable for IRQ]

>  From what I gather you're concerned about the case when hard IRQ and
> NAPI run in parallel. But I don't see how that could ever happen for
> Tx (there are some complexities like netpoll and busy poll but those
> will return false from napi_complete_done()).

Weirdly, net_rx_action() is leading to NAPI TX Handler.