Fix soft lockup in at91 udc driver

Hi Ryan,

2009/2/16 Ryan Mallon <[email protected]>:

Quoted 8 lines

Quoted 5 lines · Ryan Mallon

The udc registers are only writeable if the udc clocks are enabled, but udc interrupts can still occur. This can lead to a situation where an interrupt is perpetually received when the clocks are disabled and is never cleared or masked, which causes a lockup. The following patch ensures that the clocks are enabled while handling an interrupt:

Any comments on this patch? Has anybody else had problems with the AT91 UDC driver causing soft lockups?

Although this patch is now almost a year old, we first encountered this type of bug just a few weeks ago.
The problem description you mention fits the problem we see. We also noticed that the clocks were disabled in the case our lockup occurs.

The funny thing is that we never encountered this problem with a 2.6.24(-rt) kernel. 2.6.31-rt11 did show this problem extremely rare, and now 2.6.31-rt19 shows it hard with a certain type of PCI-USB host controller. Many host-controllers work fine and we even have 2 different PCI-boards (of different manufacturers) with the same NEC host-controller chip of which only 1 shows this problem.

So, it does not surprise me that nobody else noticed this problem before, although this patch really seems to be valid.
If it works in our case as well, and we will verify it tomorrow, we should push it forward.

Quoted 31 lines

Quoted 31 lines

Signed-off-by Ryan Mallon <[email protected]>

--

drivers/usb/gadget/at91_udc.c +9 -0

index 0b2bb8f..fcbdf24 100644
@@ -1370,6 +1370,12 @@ static irqreturn_t at91_udc_irq (int irq, void *_udc)
 {
      struct at91_udc         *udc = _udc;
      u32                     rescans = 5;
+     int                     disable_clock = 0;
+
+     if (!udc->clocked) {
+             clk_on(udc);
+             disable_clock = 1;
+     }
 
      while (rescans--) {
              u32 status;

@@ -1458,6 +1464,9 @@ static irqreturn_t at91_udc_irq (int irq, void *_udc)
              }
      }
 
+     if (disable_clock)
+             clk_off(udc);
+
      return IRQ_HANDLED;
 }
 

Quoted 9 lines · Remy Bohmer

Hi Ryan,

2009/2/16 Ryan Mallon <[email protected]>:

Quoted 6 lines

Quoted 5 lines · Ryan Mallon

The udc registers are only writeable if the udc clocks are enabled, but udc interrupts can still occur. This can lead to a situation where an interrupt is perpetually received when the clocks are disabled and is never cleared or masked, which causes a lockup. The following patch ensures that the clocks are enabled while handling an interrupt:

Quoted 7 lines

Quoted 2 lines

Any comments on this patch? Has anybody else had problems with the AT91 UDC driver causing soft lockups?

Although this patch is now almost a year old, we first encountered this type of bug just a few weeks ago.
The problem description you mention fits the problem we see. We also noticed that the clocks were disabled in the case our lockup occurs.

Awesome, thanks. The patch certainly fixed the problem on our board (from memory I originally wrote this for our 2.6.20 kernel), but the patch does seem valid even if the occurrence of the problem is rare. I never bothered following up on it, because nobody else was reporting problems.

If you could test and give me a Tested-by or Acked-by, then I'll rebase and repost the patch.