Re: [PATCH V7 3/5] i2c: tegra: Add DMA Support

[Thierry Reding]

On Thu, Jan 31, 2019 at 06:02:45PM +0300, Dmitry Osipenko wrote:
> 31.01.2019 17:43, Thierry Reding ÐÐÑÐÑ:
> > On Thu, Jan 31, 2019 at 05:06:18PM +0300, Dmitry Osipenko wrote:
> >> 31.01.2019 15:06, Thierry Reding ÐÐÑÐÑ:
> >>> On Thu, Jan 31, 2019 at 03:05:48AM +0300, Dmitry Osipenko wrote:
> >>>> 30.01.2019 19:01, Sowjanya Komatineni ÐÐÑÐÑ:
> >>> [...]
> >>>>> diff --git a/drivers/i2c/busses/i2c-tegra.c b/drivers/i2c/busses/i2c-tegra.c
> >>> [...]
> >>>>> +		return -EIO;
> >>>>> +	}
> >>>>> +
> >>>>> +	dma_desc->callback = tegra_i2c_dma_complete;
> >>>>> +	dma_desc->callback_param = i2c_dev;
> >>>>> +	dmaengine_submit(dma_desc);
> >>>>> +	dma_async_issue_pending(chan);
> >>>>> +	return 0;
> >>>>> +}
> >>>>> +
> >>>>> +static int tegra_i2c_init_dma_param(struct tegra_i2c_dev *i2c_dev,
> >>>>> +				    bool dma_to_memory)
> >>>>> +{
> >>>>> +	struct dma_chan *dma_chan;
> >>>>> +	u32 *dma_buf;
> >>>>> +	dma_addr_t dma_phys;
> >>>>> +	int ret;
> >>>>> +	const char *chan_name = dma_to_memory ? "rx" : "tx";
> >>>>
> >>>> What about to move out chan_name into the function arguments?
> >>>
> >>> That opens up the possibility of passing dma_to_memory = true and
> >>> chan_name as "tx" and create an inconsistency.
> >>>
> >>>>> @@ -884,6 +1187,8 @@ static void tegra_i2c_parse_dt(struct tegra_i2c_dev *i2c_dev)
> >>>>>  
> >>>>>  	i2c_dev->is_multimaster_mode = of_property_read_bool(np,
> >>>>>  			"multi-master");
> >>>>> +
> >>>>> +	i2c_dev->has_dma = of_property_read_bool(np, "dmas");
> >>>>
> >>>> Not only the existence of "dmas" property defines whether DMA is available. DMA subsystem could be disabled in the kernels configuration.
> >>>>
> >>>> Hence there is a need to check for DMA driver presence in the code:
> >>>>
> >>>> 	if (IS_ENABLED(CONFIG_TEGRA20_APB_DMA))
> >>>> 		i2c_dev->has_dma = of_property_read_bool(np, "dmas");
> >>>
> >>> Do we even need the ->has_dma at all? We can just go ahead and request
> >>> the channels at probe time and respond accordingly. If there's no dmas
> >>> property in DT, dma_request_slave_channel_reason() returns an error so
> >>> we can just deal with it at that time.
> >>>
> >>> So if we get -EPROBE_DEFER we can propagate that, for any other errors
> >>> we can simply fallback to PIO. Or perhaps we want to restrict fallback
> >>> to PIO for -ENODEV?
> >>>
> >>> I wouldn't want to add an IS_ENABLED(CONFIG_TEGRA20_APB_DMA) in here.
> >>> The purpose of these subsystems it to abstract all of that away.
> >>> Otherwise we could just as well use custom APIs, if we're tying together
> >>> drivers in this way anyway.
> >>
> >> DMA API doesn't fully abstract the dependencies between drivers, hence
> >> I disagree.
> > 
> > Why not? The dependency we're talking about here is a runtime dependency
> > rather than a build time dependency. Kconfig is really all about build-
> > time dependencies.
> 
> My understanding is that Kconfig is also about runtime dependencies,
> do you know if it's explicitly documented anywhere?

I don't think it's explicitly documented, just a common practice that
I've seen applied multiple times over the years. A quick grep through
the drivers/ subdirectory confirms that it's not typical to have this
sort of dependency in the code.

Similarly, Kconfig uses select primarily to pull in dependencies that
are in the form of helper libraries and such. Occasionally you'll have
some ARCH_* option select a couple of features, or even drivers, but
that is mostly a shortcut to explicitly having to list the essentials
in a defconfig.

Another reason why it's not good to model these runtime dependencies in
Kconfig is because they unnecessarily restrict the driver. For example,
if you want to build a specialized Linux binary for Tegra186, you will
certainly want to include the i2c-tegra driver. At the same time you
won't want to include the APB DMA driver because it doesn't exist on
Tegra186. Instead you'd want the (non-existent) GPC DMA driver. select
on the APB DMA driver will unconditionally pull in the driver, depends
will only allow you to build i2c-tegra if the APB DMA driver is also
enabled and the conditional in the code may lead to not using DMA
because the APB DMA driver is not available. So you'd have to modify the
i2c-tegra driver to take into account the GPC DMA driver.

> >>>> Also Tegra I2C driver should select DMA driver in Kconfig to make DMA
> >>>> driver built-in when I2C driver is built-in:
> >>>
> >>> I don't think there's a requirement for that. The only dependency we
> >>> really have here is the one on the DMA engine API. Since dmaengine.h
> >>> already provides dummy implementations, there's really no need for
> >>> us to have the dependency. If the DMA engine API is completely disabled,
> >>> a call to dma_request_slave_channel_reason() will return -ENODEV and we
> >>> should just deal with that the same way we would if there was no "dmas"
> >>> property present.
> >>
> >> In my opinion it is much better to avoid I2C driver probe failing with
> >> -EPROBE_DEFER if we could. It's just one line in code and one in
> >> Kconfig.. really. 
> > 
> > The problem is that from a theoretical point of view we don't know that
> > APB DMA is the provider for the DMA channels. This provider could be a
> > completely different device on a different Tegra generation (in fact,
> > the DMA engine on Tegra186 is a different one, so we'd have to add that
> > to the list of checks to make sure we don't disable DMA there). And the
> > fact that we're tightly integrated is really only by accident. We could
> > have the same situation on a SoC that incorporates IP from multiple
> > different sources and multiple combinations thereof as well, so how
> > would you want to deal with those cases?
> > 
> > Agreed, failing with -EPROBE_DEFER is suboptimal in that case, but that
> > is really more of an integration problem. Ideally of course there'd be
> > some way for the DMA engine subsystem to know that the provider for the
> > given device node will never show up and give us -ENODEV instead, but,
> > alas, I don't even think that would be possible. That's the price to pay
> > for abstraction.
> 
> It's not a big problem to solve for this case, there is
> of_machine_is_compatible(). To me it's more a question about the will
> to invest some extra effort to support all of possible combinations.
> If there is no such will, then at least those unpopular combinations
> shouldn't hurt and thus it should make sense to add an explicit
> build-dependency on the DMA drivers.

I think we're arguing about the same thing, only coming at it from
different angles. For me "all possible combinations" also includes the
case where you want to be able to run the driver with DMA if the APB DMA
is not enabled. And I similarly want to be able to run without DMA if
the APB DMA is enabled (by explicitly removing dmas from DT for
example). It just seems that we can't have it both ways.

Also the i2c-tegra driver can perfectly well function without DMA
support (it's done so ever since it was first merged). Keeping existing
functionality shouldn't require the addition of another driver.

Given the deadlock, I think I'd prefer the option of adding the
conditional in the code. I think that's the most accurate description of
the dependency, even though ideally it would be handled transparently by
the DMA engine API. Would that be an acceptable compromise?

Thierry

Attachment: signature.asc
Description: PGP signature