On Thu, Oct 31, 2019 at 06:11:33PM +0300, Dmitry Osipenko wrote:
> 15.10.2019 19:29, Thierry Reding пишет:
> > From: Thierry Reding <treding@nvidia.com>
> > 
> > This new framework is currently nothing more than a registry of memory
> > controllers, with the goal being to order device probing. One use-case
> > where this is useful, for example, is a memory controller device which
> > needs to program some registers before the system MMU can be enabled.
> > Associating the memory controller with the SMMU allows the SMMU driver
> > to defer the probe until the memory controller has been registered.
> > 
> > One such example is Tegra186 where the memory controller contains some
> > registers that are used to program stream IDs for the various memory
> > clients (display, USB, PCI, ...) in the system. Programming these SIDs
> > is required for the memory clients to emit the proper SIDs as part of
> > their memory requests. The memory controller driver therefore needs to
> > be programmed prior to the SMMU driver. To achieve that, the memory
> > controller will be referenced via phandle from the SMMU device tree
> > node, the SMMU driver can then use the memory controller framework to
> > find it and defer probe until it has been registered.
> > 
> > Signed-off-by: Thierry Reding <treding@nvidia.com>
> > ---
> >  drivers/memory/Makefile           |  1 +
> >  drivers/memory/core.c             | 99 +++++++++++++++++++++++++++++++
> >  include/linux/memory-controller.h | 25 ++++++++
> >  3 files changed, 125 insertions(+)
> >  create mode 100644 drivers/memory/core.c
> >  create mode 100644 include/linux/memory-controller.h
> 
> Hello Thierry,
> 
> This looks like a very good endeavour! I have couple comments, please
> see them below.
> 
> > diff --git a/drivers/memory/Makefile b/drivers/memory/Makefile
> > index 27b493435e61..d16e7dca8ef9 100644
> > --- a/drivers/memory/Makefile
> > +++ b/drivers/memory/Makefile
> > @@ -3,6 +3,7 @@
> >  # Makefile for memory devices
> >  #
> >  
> > +obj-y				+= core.o
> >  obj-$(CONFIG_DDR)		+= jedec_ddr_data.o
> >  ifeq ($(CONFIG_DDR),y)
> >  obj-$(CONFIG_OF)		+= of_memory.o
> > diff --git a/drivers/memory/core.c b/drivers/memory/core.c
> > new file mode 100644
> > index 000000000000..1772e839305a
> > --- /dev/null
> > +++ b/drivers/memory/core.c
> > @@ -0,0 +1,99 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +/*
> > + * Copyright (C) 2019 NVIDIA Corporation.
> > + */
> > +
> > +#include <linux/memory-controller.h>
> > +#include <linux/of.h>
> > +
> > +static DEFINE_MUTEX(controllers_lock);
> > +static LIST_HEAD(controllers);
> > +
> > +static void memory_controller_release(struct kref *ref)
> > +{
> > +	struct memory_controller *mc = container_of(ref, struct memory_controller, ref);
> > +
> > +	WARN_ON(!list_empty(&mc->list));
> > +}
> > +
> > +int memory_controller_register(struct memory_controller *mc)
> > +{
> > +	kref_init(&mc->ref);
> > +
> > +	mutex_lock(&controllers_lock);
> > +	list_add_tail(&mc->list, &controllers);
> > +	mutex_unlock(&controllers_lock);
> > +
> > +	return 0;
> > +}
> > +EXPORT_SYMBOL_GPL(memory_controller_register);
> > +
> > +void memory_controller_unregister(struct memory_controller *mc)
> > +{
> > +	mutex_lock(&controllers_lock);
> > +	list_del_init(&mc->list);
> > +	mutex_unlock(&controllers_lock);
> > +
> > +	kref_put(&mc->ref, memory_controller_release);
> > +}
> > +EXPORT_SYMBOL_GPL(memory_controller_unregister);
> > +
> > +static struct memory_controller *
> > +of_memory_controller_get(struct device *dev, struct device_node *np,
> > +			 const char *con_id)
> > +{
> > +	const char *cells = "#memory-controller-cells";
> > +	const char *names = "memory-controller-names";
> > +	const char *prop = "memory-controllers";
> > +	struct memory_controller *mc;
> > +	struct of_phandle_args args;
> > +	int index = 0, err;
> > +
> > +	if (con_id) {
> > +		index = of_property_match_string(np, names, con_id);
> > +		if (index < 0)
> > +			return ERR_PTR(index);
> > +	}
> > +
> > +	err = of_parse_phandle_with_args(np, prop, cells, index, &args);
> > +	if (err) {
> > +		if (err == -ENOENT)
> > +			err = -ENODEV;
> > +
> > +		return ERR_PTR(err);
> > +	}
> > +
> > +	mutex_lock(&controllers_lock);
> > +
> > +	list_for_each_entry(mc, &controllers, list) {
> > +		if (mc->dev && mc->dev->of_node == args.np) {
> > +			kref_get(&mc->ref);
> 
> This is not enough because memory controller driver could be a loadable
> module, thus something like this is needed here:
> 
> 	__module_get(mc->dev->driver->owner);
> 
> This won't allow MC driver to be unloaded while it has active users.

Good catch. I've added that (and the module_put() from below) to the
patch.

> > +			mutex_unlock(&controllers_lock);
> > +			goto unlock;
> > +		}
> > +	}
> > +
> > +	mc = ERR_PTR(-EPROBE_DEFER);
> > +
> > +unlock:
> > +	mutex_unlock(&controllers_lock);
> > +	of_node_put(args.np);
> > +	return mc;
> > +}
> > +
> > +struct memory_controller *
> > +memory_controller_get(struct device *dev, const char *con_id)
> > +{
> > +	if (IS_ENABLED(CONFIG_OF) && dev && dev->of_node)
> > +		return of_memory_controller_get(dev, dev->of_node, con_id);
> > +
> > +	return ERR_PTR(-ENODEV);
> > +}
> > +EXPORT_SYMBOL_GPL(memory_controller_get);
> 
> In most cases memory controllers are unique in a system, so it looks to
> me that it will be more universal to have ability to get MC by its
> device-tree compatible name. Like this:
> 
> 	of_memory_controller_get_by_compatible(const char *compatible);
> 
> This will allow current drivers (like Tegra20 devfreq driver for
> example) to utilize this new API without having trouble of maintaining
> backwards compatibility with older device-trees that do not have a
> phandle to MC.
> 
> https://elixir.bootlin.com/linux/v5.4-rc5/source/drivers/devfreq/tegra20-devfreq.c#L100
> 
> Of course there could be cases where there are multiple controllers with
> the same compatible, but that case could be supported later on by those
> who really need it. I don't think that any of NVIDIA Tegra SoCs fall
> into that category.

This has the slight disadvantage that we would have to iterate over a
number of compatible strings in case we want to transparently support
more than a single version of the memory controller.

An alternative, which is used by a number of other resource registry
APIs, would be to work with lookup tables. Basically those would make
a mapping between a provider and a device/consumer pair. The result
would look something like this:

	struct memory_controller_lookup {
		const char *provider;
		const char *dev_id;
		const char *con_id;
	};

	static const struct memory_controller_lookup *tegra124_mc_lookup[] = {
		{ "70019000.memory-controller", "6000c800.actmon", NULL },
	};

memory_controller_get() could then use that as a last-resort to find a
reference to a memory controller if a device tree phandle isn't
available.

On the other hand it should be fairly easy to conditionalize all the
code based purely on the availability of a phandle:

	mc = memory_controller_get(dev, NULL);
	if (IS_ERR(mc)) {
		if (mc != ERR_PTR(-ENODEV))
			return PTR_ERR(mc);

		mc = NULL;
	}

	...

	if (mc) {
		...
	}

The above could be simplified by wrapping the logic in a helper that can
be used if consumers can work without: memory_controller_get_optional().

Thierry

> > +void memory_controller_put(struct memory_controller *mc)
> > +{
> > +	if (mc)
> > +		kref_put(&mc->ref, memory_controller_release);
> 		module_put(mc->dev->driver->owner);
> 
> > +}
> > +EXPORT_SYMBOL_GPL(memory_controller_put);
> 
> 
> [snip]