| Message ID | 20231108093125.224963-1-pawell@cadence.com |
|---|---|
| State | New |
| Headers | show |
| Series | usb: cdnsp: Fix deadlock issue during using NCM gadget | expand |
On 23-11-08 10:31:25, Pawel Laszczak wrote: > The interrupt service routine registered for the gadget is a primary > handler which mask the interrupt source and a threaded handler which > handles the source of the interrupt. Since the threaded handler is > voluntary threaded, the IRQ-core does not disable bottom halves before > invoke the handler like it does for the forced-threaded handler. > > Due to changes in networking it became visible that a network gadget's > completions handler may schedule a softirq which remains unprocessed. > The gadget's completion handler is usually invoked either in hard-IRQ or > soft-IRQ context. In this context it is enough to just raise the softirq > because the softirq itself will be handled once that context is left. > In the case of the voluntary threaded handler, there is nothing that > will process pending softirqs. Which means it remain queued until > another random interrupt (on this CPU) fires and handles it on its exit > path or another thread locks and unlocks a lock with the bh suffix. > Worst case is that the CPU goes idle and the NOHZ complains about > unhandled softirqs. Would you have a diagram to explain how things happen, and when the network softirq is scheduled in this case? Peter > > Disable bottom halves before acquiring the lock (and disabling > interrupts) and enable them after dropping the lock. This ensures that > any pending softirqs will handled right away. > > cc: <stable@vger.kernel.org> > Fixes: 3d82904559f4 ("usb: cdnsp: cdns3 Add main part of Cadence USBSSP DRD Driver") > Signed-off-by: Pawel Laszczak <pawell@cadence.com> > --- > drivers/usb/cdns3/cdnsp-ring.c | 3 +++ > 1 file changed, 3 insertions(+) > > diff --git a/drivers/usb/cdns3/cdnsp-ring.c b/drivers/usb/cdns3/cdnsp-ring.c > index 07f6068342d4..275a6a2fa671 100644 > --- a/drivers/usb/cdns3/cdnsp-ring.c > +++ b/drivers/usb/cdns3/cdnsp-ring.c > @@ -1529,6 +1529,7 @@ irqreturn_t cdnsp_thread_irq_handler(int irq, void *data) > unsigned long flags; > int counter = 0; > > + local_bh_disable(); > spin_lock_irqsave(&pdev->lock, flags); > > if (pdev->cdnsp_state & (CDNSP_STATE_HALTED | CDNSP_STATE_DYING)) { > @@ -1541,6 +1542,7 @@ irqreturn_t cdnsp_thread_irq_handler(int irq, void *data) > cdnsp_died(pdev); > > spin_unlock_irqrestore(&pdev->lock, flags); > + local_bh_enable(); > return IRQ_HANDLED; > } > > @@ -1557,6 +1559,7 @@ irqreturn_t cdnsp_thread_irq_handler(int irq, void *data) > cdnsp_update_erst_dequeue(pdev, event_ring_deq, 1); > > spin_unlock_irqrestore(&pdev->lock, flags); > + local_bh_enable(); > > return IRQ_HANDLED; > } > -- > 2.37.2 >
diff --git a/drivers/usb/cdns3/cdnsp-ring.c b/drivers/usb/cdns3/cdnsp-ring.c index 07f6068342d4..275a6a2fa671 100644 --- a/drivers/usb/cdns3/cdnsp-ring.c +++ b/drivers/usb/cdns3/cdnsp-ring.c @@ -1529,6 +1529,7 @@ irqreturn_t cdnsp_thread_irq_handler(int irq, void *data) unsigned long flags; int counter = 0; + local_bh_disable(); spin_lock_irqsave(&pdev->lock, flags); if (pdev->cdnsp_state & (CDNSP_STATE_HALTED | CDNSP_STATE_DYING)) { @@ -1541,6 +1542,7 @@ irqreturn_t cdnsp_thread_irq_handler(int irq, void *data) cdnsp_died(pdev); spin_unlock_irqrestore(&pdev->lock, flags); + local_bh_enable(); return IRQ_HANDLED; } @@ -1557,6 +1559,7 @@ irqreturn_t cdnsp_thread_irq_handler(int irq, void *data) cdnsp_update_erst_dequeue(pdev, event_ring_deq, 1); spin_unlock_irqrestore(&pdev->lock, flags); + local_bh_enable(); return IRQ_HANDLED; }
The interrupt service routine registered for the gadget is a primary handler which mask the interrupt source and a threaded handler which handles the source of the interrupt. Since the threaded handler is voluntary threaded, the IRQ-core does not disable bottom halves before invoke the handler like it does for the forced-threaded handler. Due to changes in networking it became visible that a network gadget's completions handler may schedule a softirq which remains unprocessed. The gadget's completion handler is usually invoked either in hard-IRQ or soft-IRQ context. In this context it is enough to just raise the softirq because the softirq itself will be handled once that context is left. In the case of the voluntary threaded handler, there is nothing that will process pending softirqs. Which means it remain queued until another random interrupt (on this CPU) fires and handles it on its exit path or another thread locks and unlocks a lock with the bh suffix. Worst case is that the CPU goes idle and the NOHZ complains about unhandled softirqs. Disable bottom halves before acquiring the lock (and disabling interrupts) and enable them after dropping the lock. This ensures that any pending softirqs will handled right away. cc: <stable@vger.kernel.org> Fixes: 3d82904559f4 ("usb: cdnsp: cdns3 Add main part of Cadence USBSSP DRD Driver") Signed-off-by: Pawel Laszczak <pawell@cadence.com> --- drivers/usb/cdns3/cdnsp-ring.c | 3 +++ 1 file changed, 3 insertions(+)