Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1459387 > unrolled thread

dmatest no longer works on ARC SDP with DW DMAC

Back to article view | Back to linux.kernel

Contents

  dmatest no longer works on ARC SDP with DW DMAC Eugeniy Paltsev <Eugeniy.Paltsev@synopsys.com> - 2016-08-10 20:20 +0200
    Re: dmatest no longer works on ARC SDP with DW DMAC Alexey Brodkin <Alexey.Brodkin@synopsys.com> - 2016-08-10 20:30 +0200
      Re: dmatest no longer works on ARC SDP with DW DMAC Andy Shevchenko <andriy.shevchenko@linux.intel.com> - 2016-08-10 22:50 +0200
    Re: dmatest no longer works on ARC SDP with DW DMAC Andy Shevchenko <andriy.shevchenko@linux.intel.com> - 2016-08-10 20:50 +0200

#1459387 — dmatest no longer works on ARC SDP with DW DMAC

dmatest on ARC SDP with DW DMAC became broken after df5c7386
("dmaengine: dw: some Intel devices has no memcpy support") and
30cb2639 ("dmaengine: dw: don't override platform data with autocfg")
commits.
* After df5c7386 commit "DMA_MEMCPY" capability option doesn't
get set correctly in platform driver version.
* After 30cb2639 commit
"data_width" and "nollp" parameters don't get set correctly in platform
driver version.

This happens because in old driver version there are three sources 
of parameters: pdata, device tree and autoconfig hardware registers. 
Some parameters were read from pdata and others from autoconfig
hardware registers. If pdata was absent some pdata structure 
fields were filled with parameters from device tree. But 30cb2639
commit disabled overriding pdata with autocfg, so if we use platform
driver version without pdata some parameters will not be set.

I'm wondering what would be the best way to fix this situation?
Should we strictly read parameters from only one source (pdata/device
tree/autoconfig) or we may mix some of them (for example getting
missing data from autoconf regs)?

[toc] | [next] | [standalone]

#1459431

Hi Andy,

On Wed, 2016-08-10 at 15:15 +-0300, Andy Shevchenko wrote:
+AD4- On Wed, 2016-08-10 at 11:06 +-0000, Eugeniy Paltsev wrote:
+AD4- +AD4- 
+AD4- +AD4- dmatest on ARC SDP with DW DMAC became broken after df5c7386
+AD4- +AD4- (+ACI-dmaengine: dw: some Intel devices has no memcpy support+ACI-) and
+AD4- +AD4- 30cb2639 (+ACI-dmaengine: dw: don't override platform data with autocfg+ACI-)
+AD4- +AD4- commits.
+AD4- +AD4- +ACo- After df5c7386 commit +ACI-DMA+AF8-MEMCPY+ACI- capability option doesn't
+AD4- +AD4- get set correctly in platform driver version.
+AD4- +AD4- +ACo- After 30cb2639 commit
+AD4- +AD4- +ACI-data+AF8-width+ACI- and +ACI-nollp+ACI- parameters don't get set correctly in
+AD4- +AD4- platform
+AD4- +AD4- driver version.
+AD4- +AD4- 
+AD4- +AD4- This happens because in old driver version there are three sources+AKA-
+AD4- +AD4- of parameters: pdata, device tree and autoconfig hardware registers.+AKA-
+AD4- +AD4- Some parameters were read from pdata and others from autoconfig
+AD4- +AD4- hardware registers. If pdata was absent some pdata structure+AKA-
+AD4- +AD4- fields were filled with parameters from device tree. But 30cb2639
+AD4- +AD4- commit disabled overriding pdata with autocfg, so if we use platform
+AD4- +AD4- driver version without pdata some parameters will not be set.
+AD4- 
+AD4- Yes, that's correct behaviour right now. You have to provide platform
+AD4- code which registers device with all platform data provided.

But given autocfg registers exist in HW why don't we rely on their contents?

+AD4- +AD4- I'm wondering what would be the best way to fix this situation?
+AD4- 
+AD4- Ideally we have to switch to use built-in device properties
+AD4- (drivers/base/property.c) and platform code in your case has to provide
+AD4- properties.

What do you mean saying +ACI-built-in device properties+ACI-?
Setting pdata structure? In our particular case we use device tree
for DW DMAC setup.

-Alexey

[toc] | [prev] | [next] | [standalone]

#1459902

On Wed, 2016-08-10 at 12:22 +0000, Alexey Brodkin wrote:
> Hi Andy,
> 
> On Wed, 2016-08-10 at 15:15 +0300, Andy Shevchenko wrote:
> > 
> > On Wed, 2016-08-10 at 11:06 +0000, Eugeniy Paltsev wrote:
> > > 
> > > 
> > > dmatest on ARC SDP with DW DMAC became broken after df5c7386
> > > ("dmaengine: dw: some Intel devices has no memcpy support") and
> > > 30cb2639 ("dmaengine: dw: don't override platform data with
> > > autocfg")
> > > commits.
> > > * After df5c7386 commit "DMA_MEMCPY" capability option doesn't
> > > get set correctly in platform driver version.
> > > * After 30cb2639 commit
> > > "data_width" and "nollp" parameters don't get set correctly in
> > > platform
> > > driver version.
> > > 
> > > This happens because in old driver version there are three
> > > sources 
> > > of parameters: pdata, device tree and autoconfig hardware
> > > registers. 
> > > Some parameters were read from pdata and others from autoconfig
> > > hardware registers. If pdata was absent some pdata structure 
> > > fields were filled with parameters from device tree. But 30cb2639
> > > commit disabled overriding pdata with autocfg, so if we use
> > > platform
> > > driver version without pdata some parameters will not be set.
> > 
> > Yes, that's correct behaviour right now. You have to provide
> > platform
> > code which registers device with all platform data provided.
> 
> But given autocfg registers exist in HW why don't we rely on their
> contents?

And how exactly we can get that mem2mem support is absent / broken by
some reason?

All those quirks (bool is_*) kinda semi-hardware related.

What we can do is to have two categories of properties: a) genuine
hardware properties, and b) quirks.

So, refactor the ->probe() function in a way that will still copy quirks
and other non-hardware properties from platform data, if provided, to
the driver internals.

> 
> > 
> > > 
> > > I'm wondering what would be the best way to fix this situation?
> > 
> > Ideally we have to switch to use built-in device properties
> > (drivers/base/property.c) and platform code in your case has to
> > provide
> > properties.
> 
> What do you mean saying "built-in device properties"?
> Setting pdata structure? In our particular case we use device tree
> for DW DMAC setup.

Providing device properties instead of platform data. Again see above.

-- 
Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Intel Finland Oy

[toc] | [prev] | [next] | [standalone]

#1459518

On Wed, 2016-08-10 at 11:06 +0000, Eugeniy Paltsev wrote:
> dmatest on ARC SDP with DW DMAC became broken after df5c7386
> ("dmaengine: dw: some Intel devices has no memcpy support") and
> 30cb2639 ("dmaengine: dw: don't override platform data with autocfg")
> commits.
> * After df5c7386 commit "DMA_MEMCPY" capability option doesn't
> get set correctly in platform driver version.
> * After 30cb2639 commit
> "data_width" and "nollp" parameters don't get set correctly in
> platform
> driver version.
> 
> This happens because in old driver version there are three sources 
> of parameters: pdata, device tree and autoconfig hardware registers. 
> Some parameters were read from pdata and others from autoconfig
> hardware registers. If pdata was absent some pdata structure 
> fields were filled with parameters from device tree. But 30cb2639
> commit disabled overriding pdata with autocfg, so if we use platform
> driver version without pdata some parameters will not be set.

Yes, that's correct behaviour right now. You have to provide platform
code which registers device with all platform data provided.

> I'm wondering what would be the best way to fix this situation?

Ideally we have to switch to use built-in device properties
(drivers/base/property.c) and platform code in your case has to provide
properties.

> Should we strictly read parameters from only one source (pdata/device
> tree/autoconfig) or we may mix some of them (for example getting
> missing data from autoconf regs)?

See above.

-- 
Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Intel Finland Oy

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web