On 6/18/26 16:39, Sebastian Andrzej Siewior wrote:

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

On 2026-06-17 17:20:28 [+0800], Yun Zhou wrote:

quoted

On PREEMPT_RT, the mvneta IRQ handler is force-threaded. Under high

There is also the `threadirqs' option.

quoted

network traffic, the IRQ can enter suspend with desc->depth == 1
(masked by the oneshot mechanism between handler invocations).

That would be irq_desc::depth.

quoted

During suspend, the kernel increments depth to 2 and masks the
interrupt at the MPIC level (clearing the SRC_CTL CPU routing bit,
due to IRQCHIP_MASK_ON_SUSPEND).

The interrupt should be masked while the depth counter goes 0->1, no?

quoted

                                  On resume, depth is decremented
back to 1, but since it does not reach 0, the unmask is never
called. The MPIC CPU routing remains cleared, permanently disabling
interrupt delivery.

But why not? In my naive assumption, we get into suspend with
irq_desc::depth = 2 and the threaded should be woken up. Once the
treaded handler is done the counter should decrement by one. Then again
during resume reaching 0 leading to the unmask. If the thread handler is
frozen and defrosted on resume then it should still happen but in
different order.

Something is missing here based on my naive assumption.

quoted

Fix by freeing the IRQ in suspend and re-requesting it in resume.
This ensures a clean IRQ state (depth=0, proper hardware routing)
on every resume cycle, regardless of the pre-suspend depth. This
follows the approach used by other drivers (e.g. igb).

The igb shutdowns the device entirely, not just freeing the IRQ.

You are right. The original analysis was wrong — mvneta uses
request_percpu_irq() which sets IRQF_NO_SUSPEND, so the PM framework
never touches this IRQ. The depth never changes from 1.

The actual root cause is simpler: mvneta_percpu_isr() calls
disable_percpu_irq() before scheduling NAPI, and enable_percpu_irq()
is called in napi_complete_done(). If suspend hits during active NAPI
polling, the MPIC percpu IRQ stays masked after resume because
mvneta_start_dev() doesn't restore it.

Will send a v3 with the correct one-liner fix (enable_percpu_irq in
the resume path). Apologies for the incorrect analysis.

BR,
Yun