Re: [PATCH 6/7] DMA: Freescale: use spin_lock_bh instead of spin_lock_irqsave

From: Vinod Koul
Date: Wed Apr 02 2014 - 12:43:18 EST


On Mon, Mar 31, 2014 at 12:08:55PM +0800, Hongbo Zhang wrote:
>
> On 03/29/2014 09:45 PM, Vinod Koul wrote:
> >On Fri, Mar 28, 2014 at 02:33:37PM +0800, Hongbo Zhang wrote:
> >>On 03/26/2014 03:01 PM, Vinod Koul wrote:
> >>>On Thu, 2014-01-16 at 13:47 +0800, hongbo.zhang@xxxxxxxxxxxxx wrote:
> >>>>From: Hongbo Zhang <hongbo.zhang@xxxxxxxxxxxxx>
> >>>>
> >>>>The usage of spin_lock_irqsave() is a stronger locking mechanism than is
> >>>>required throughout the driver. The minimum locking required should be used
> >>>>instead. Interrupts will be turned off and context will be saved, it is
> >>>>unnecessary to use irqsave.
> >>>>
> >>>>This patch changes all instances of spin_lock_irqsave() to spin_lock_bh(). All
> >>>>manipulation of protected fields is done using tasklet context or weaker, which
> >>>>makes spin_lock_bh() the correct choice.
> >>>>
> >
> >>>> /**
> >>>>@@ -1124,11 +1120,10 @@ static irqreturn_t fsldma_chan_irq(int irq, void *data)
> >>>> static void dma_do_tasklet(unsigned long data)
> >>>> {
> >>>> struct fsldma_chan *chan = (struct fsldma_chan *)data;
> >>>>- unsigned long flags;
> >>>> chan_dbg(chan, "tasklet entry\n");
> >>>>- spin_lock_irqsave(&chan->desc_lock, flags);
> >>>>+ spin_lock_bh(&chan->desc_lock);
> >>>okay here is the problem :(
> >>>
> >>>You moved to _bh variant. So if you grab the lock in rest of the code
> >>>and irq gets triggered then here we will be spinning to grab the lock.
> >>>So effectively you made right locking solution into deadlock situation!
> >>If the rest code grabs lock by spin_lock_bh(), and if irq raised,
> >>the tasklet could not be executed because it has been disabled by
> >>the _bh variant function.
> >yes if you are accessing resources only in tasklet and rest of the code, then
> >_bh variant works well. The problem here is usage in irq handler
> >
>
> The name dma_do_tasklet may mislead, it is tasklet handler, not irq
> handler, not a trigger to load tasklet.
> the irq handler is fsldma_chan_irq, and I don't use lock in it.
sorry my bad, i misread this as code in fsldma_chan_irq() insteadof
dma_do_tasklet(). In that case patch is doing the right thing.

--
~Vinod
>
> If it is the problem, I would like to change dma_do_tasklet to
> dma_tasklet to eliminate misleading.
>
>
>

--
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/