Re: [PATCH V3 2/7] dmaengine/dw_dmac: Replace spin_lock* with irqsave variants

[Russell King - ARM Linux <linux@arm.linux.org.uk>]

On Thu, Apr 28, 2011 at 06:10:20PM +0100, Russell King - ARM Linux wrote:
> On Wed, Apr 27, 2011 at 03:06:44PM +0530, Viresh Kumar wrote:
> > @@ -407,6 +410,8 @@ EXPORT_SYMBOL(dw_dma_get_dst_addr);
> >  static void dwc_handle_cyclic(struct dw_dma *dw, struct dw_dma_chan *dwc,
> >  		u32 status_block, u32 status_err, u32 status_xfer)
> >  {
> > +	unsigned long flags;
> > +
> >  	if (status_block & dwc->mask) {
> >  		void (*callback)(void *param);
> >  		void *callback_param;
> > @@ -418,9 +423,9 @@ static void dwc_handle_cyclic(struct dw_dma *dw, struct dw_dma_chan *dwc,
> >  		callback = dwc->cdesc->period_callback;
> >  		callback_param = dwc->cdesc->period_callback_param;
> >  		if (callback) {
> > -			spin_unlock(&dwc->lock);
> > +			spin_unlock_irqrestore(&dwc->lock, flags);
> >  			callback(callback_param);
> > -			spin_lock(&dwc->lock);
> > +			spin_lock_irqsave(&dwc->lock, flags);
> 
> I'm really not convinced that this is anywhere near correct.  I'm
> surprised this doesn't spit out a compiler warning.
> 
> spin_unlock_irqrestore() reads the flags argument and puts it into
> the PSR.  spin_lock_irqsave() reads the PSR, puts it into the flags
> argument, sets the interrupt mask bit and writes back to the PSR.
> 
> So, if you do:
> 
> 	unsigned long flags;
> 
> 	spin_unlock_irqrestore(&dwc->lock, flags);
> 	...
> 	spin_lock_irqsave(&dwc->lock, flags);
> 
> you're going to end up corrupting the PSR.
> 
> In any case, releasing a spinlock temporarily within a called function
> is _really_ not a nice thing to do.  It makes code review rather
> difficult as called functions become non-atomic when called within
> an atomic region.

BTW, how this gets handled in other drivers is basically as follows in
the tasklet:

tasklet()
{
	LIST_HEAD(completed);

	spin_lock_irqsave(lock, flags);
	for each txd(txd) {
		if (completed(txd))
			list_move_tail(&txd->node, &completed);
	}
	try to start new txd();
	spin_unlock_irqrestore(lock, flags);

	for each list entry safe(txd, &completed) {
		void (*callback)(void *) = txd->callback;
		void *param = txd->callback_param;

		free_txd(txd);

		if (callback)
			callback(param);
	}
}

I'm not sure how easy it is to move dw_dmac to that kind of structure,
but I think this is what is required rather than dropping locks within
functions which they haven't themselves taken.