[PATCH v2] OMAP: PM: DMA: Enable runtime pm

[manjugk@ti.com (G, Manjunath Kondaiah)]

On Wed, Feb 16, 2011 at 11:47:33AM -0800, Kevin Hilman wrote:
> "G, Manjunath Kondaiah" <manjugk@ti.com> writes:
> 
> > Hi Kevin,
> >
> > On Mon, Feb 14, 2011 at 02:06:53PM -0800, Kevin Hilman wrote:
> >> "G, Manjunath Kondaiah" <manjugk@ti.com> writes:
> >> 
> >> > From: Manjunath G Kondaiah <manjugk@ti.com>
> >> >
> >> > Enable runtime pm and use pm_runtime_get_sync and pm_runtime_put_autosuspend
> >> > for OMAP DMA driver.
> >> >
> >> > The DMA driver uses auto suspend feature of runtime pm framework through
> >> > which the clock gets disabled automatically if there is no activity for
> >> > more than one second.
> >> >
> >> > Testing:
> >> > Compile: omap1_defconfig and omap2plus_defconfig
> >> > Boot: OMAP1710(H3), OMAP2420(H4), OMAP3630(Zoom3), OMAP4(Blaze)
> >> 
> >> The normal DMA tests should also be run on these platforms.  Based on
> >> the above, I can't tell any DMA tests were run.   Based on my tests,
> >> this isn't working for chained xfers.
> >> 
> >> Using the runtime PM sysfs interface, you can check the runtime status
> >> of the device:
> >> 
> >> # cat /sys/devices/platform/omap/omap_dma_system.0/power/runtime_status          
> >> 
> >> It should show 'active' during transfer, and after timeout expires it
> >> will show 'suspended'.
> >> 
> >> Doing some tests using my dmatest module:
> >> 
> >>   git://gitorious.org/omap-test/dmatest.git
> >> 
> >> I noticed that it gets stuck in 'active' and never gets suspended when I
> >> used DMA channel linking (load module using 'linking=1' as load-time option)
> >> 
> >> I'm not sure exactly why, but I will guess that the reason is that there
> >> is an imbalance in get/put calls when using chaining, since 'get' is
> >> only called once upon omap_start_dma() but 'put' is called for every
> >> channel in the callback.
> >
> > Even I noticed this after running chaining test case and checking
> > runtime status. But, I am wondering even with 'active' runtime status, 
> > the core hits off and retention.
> 
> Probably because system DMA is auto-idle and clocked by the core_l3_iclk
> 
> > The complete log which has all the sequences of running chaining tests,
> > enabling off mode and checking runtime status is available at:
> > http://pastebin.com/YEHMEXUP
> >
> > Though I agree on the point that, it is mismatch with get/put calls with
> > DMA chaining, I still need to analyze this in detail.
> 
> Yes.  The mismatch highlights an underlying problem.
> 
> > The other thing which is not considered here is, the get_sync is called
> > inside start_dma only(request_dma will call get_sync and put after the
> > getting requested channel). After request_dma and start_dma, there are
> > API's called by user(dma_set_params, priority etc) which also require
> > get_sync since those API's will access configuration registers. I am
> > wondering if have get_sync and put in all the API's, this might result
> > in over loading. 
> 
> I'm not sure what you mean by over loading.
> 
> You need to have all register accesses inside get/put calls.  As long as
> they are balanced, this should not leed to problems.

I mean, almost all the api's will have get_sync and put_autosuspend
calls. If it is acceptable, let me check the same with implementation.

I feel this is equivalent to placing get_sync and put_autosuspend with 
low level read/write api's.

-Manjunath