Re: [PATCH v3.5] dw_dmac: return proper residue value

[Andy Shevchenko]

On Fri, 2013-01-25 at 09:43 +0530, Viresh Kumar wrote: 
> On Thu, Jan 24, 2013 at 6:37 PM, Andy Shevchenko
> <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> > Currently the driver returns full length of the active descriptor which is
> > wrong. We have to go throught the active descriptor and substract the length of
> > each sent children in the chain from the total length along with the actual
> > data in the DMA channel registers.
> >
> > The cyclic case is not handled by this patch due to len field in the descriptor
> > structure is left untouched by the original code.
> >
> > Signed-off-by: Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx>
> 
> Firstly the issue i raised earlier about not updating residue from
> tasklet or interrupt
> handler was wrong. As i had an older version of code in my mind. This got solved
> with following patch:
> 
> commit 77bcc497c60ec62dbb84abc809a6e218d53409e9
> Author: Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx>
> Date:   Fri Jan 18 14:14:15 2013 +0200
> 
>     dw_dmac: move soft LLP code from tasklet to dwc_scan_descriptors
> 
> 
> > ---
> > diff --git a/drivers/dma/dw_dmac.c b/drivers/dma/dw_dmac.c
> > +static inline u32 dwc_get_residue(struct dw_dma_chan *dwc)
> > +{
> > +       unsigned long flags;
> > +       u32 residue;
> > +
> > +       spin_lock_irqsave(&dwc->lock, flags);
> > +
> > +       residue = dwc->residue;
> 
> you need to release the lock just here.
> 
> > +       if (test_bit(DW_DMA_IS_SOFT_LLP, &dwc->flags) && residue)
> > +               residue -= dwc_get_sent(dwc);
> 
> why do you need lock while reading the control registers? It looks you didn't
> get what i wanted to say earlier. We are using locks to protect some part for
> shared data or critical sections. these are playing with dwc transfer
> lists, etc..
> 
> Probably we don't need a lock to read the control register as nobody can
> guarantee that another access is not happening currently. As hardware is
> changing this register continuously for the transfer.
> 
> Let me know if i am missing something.

I'm trying to find a proof of any of our opinions.
My intuitive understanding that we have to avoid a concurrency during hw
access.

> 
> > +       spin_unlock_irqrestore(&dwc->lock, flags);
> > +       return residue;
> > +}

-- 
Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx>
Intel Finland Oy
--
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/