Re: [PATCH v6 33/52] memory: tegra20: Support interconnect framework

[Thierry Reding]

On Mon, Oct 26, 2020 at 01:17:16AM +0300, Dmitry Osipenko wrote:
> Now Internal and External Memory Controllers are memory interconnection
> providers. This allows us to use interconnect API for tuning of memory
> configuration. EMC driver now supports OPPs and DVFS.
> 
> Signed-off-by: Dmitry Osipenko <digetx@xxxxxxxxx>
> ---
>  drivers/memory/tegra/Kconfig       |   3 +-
>  drivers/memory/tegra/mc.h          |  12 ++
>  drivers/memory/tegra/tegra20-emc.c | 176 +++++++++++++++++++++++++++++
>  drivers/memory/tegra/tegra20.c     |  34 ++++++
>  4 files changed, 224 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/memory/tegra/Kconfig b/drivers/memory/tegra/Kconfig
> index ff426747cd7d..ac3dfe155505 100644
> --- a/drivers/memory/tegra/Kconfig
> +++ b/drivers/memory/tegra/Kconfig
> @@ -11,7 +11,8 @@ config TEGRA_MC
>  config TEGRA20_EMC
>  	tristate "NVIDIA Tegra20 External Memory Controller driver"
>  	default y
> -	depends on ARCH_TEGRA_2x_SOC
> +	depends on TEGRA_MC && ARCH_TEGRA_2x_SOC
> +	select PM_OPP
>  	help
>  	  This driver is for the External Memory Controller (EMC) found on
>  	  Tegra20 chips. The EMC controls the external DRAM on the board.
> diff --git a/drivers/memory/tegra/mc.h b/drivers/memory/tegra/mc.h
> index abeb6a2cc36a..531fb4fb7b17 100644
> --- a/drivers/memory/tegra/mc.h
> +++ b/drivers/memory/tegra/mc.h
> @@ -78,6 +78,18 @@
>  
>  #define MC_TIMING_UPDATE				BIT(0)
>  
> +static inline u32 tegra_mc_scale_percents(u64 val, unsigned int percents)
> +{
> +	val = val * percents;
> +	do_div(val, 100);
> +
> +	/*
> +	 * High freq + high boosting percent + large polling interval are
> +	 * resulting in integer overflow when watermarks are calculated.
> +	 */
> +	return min_t(u64, val, U32_MAX);
> +}
> +
>  static inline u32 mc_readl(struct tegra_mc *mc, unsigned long offset)
>  {
>  	return readl_relaxed(mc->regs + offset);
> diff --git a/drivers/memory/tegra/tegra20-emc.c b/drivers/memory/tegra/tegra20-emc.c
> index 34085e26dced..69ccb3fe5b0b 100644
> --- a/drivers/memory/tegra/tegra20-emc.c
> +++ b/drivers/memory/tegra/tegra20-emc.c
> @@ -9,6 +9,7 @@
>  #include <linux/clk/tegra.h>
>  #include <linux/debugfs.h>
>  #include <linux/err.h>
> +#include <linux/interconnect-provider.h>
>  #include <linux/interrupt.h>
>  #include <linux/io.h>
>  #include <linux/iopoll.h>
> @@ -16,11 +17,15 @@
>  #include <linux/module.h>
>  #include <linux/of.h>
>  #include <linux/platform_device.h>
> +#include <linux/pm_opp.h>
> +#include <linux/slab.h>
>  #include <linux/sort.h>
>  #include <linux/types.h>
>  
>  #include <soc/tegra/fuse.h>
>  
> +#include "mc.h"
> +
>  #define EMC_INTSTATUS				0x000
>  #define EMC_INTMASK				0x004
>  #define EMC_DBG					0x008
> @@ -144,6 +149,9 @@ struct emc_timing {
>  
>  struct tegra_emc {
>  	struct device *dev;
> +	struct tegra_mc *mc;
> +	struct opp_table *opp_table;
> +	struct icc_provider provider;
>  	struct notifier_block clk_nb;
>  	struct clk *clk;
>  	void __iomem *regs;
> @@ -658,6 +666,166 @@ static void tegra_emc_debugfs_init(struct tegra_emc *emc)
>  			    emc, &tegra_emc_debug_max_rate_fops);
>  }
>  
> +static inline struct tegra_emc *
> +to_tegra_emc_provider(struct icc_provider *provider)
> +{
> +	return container_of(provider, struct tegra_emc, provider);
> +}
> +
> +static struct icc_node_data *
> +emc_of_icc_xlate_extended(struct of_phandle_args *spec, void *data)
> +{
> +	struct icc_provider *provider = data;
> +	struct icc_node_data *ndata;
> +	struct icc_node *node;
> +
> +	/* External Memory is the only possible ICC route */
> +	list_for_each_entry(node, &provider->nodes, node_list) {
> +		if (node->id != TEGRA_ICC_EMEM)
> +			continue;
> +
> +		ndata = kzalloc(sizeof(*ndata), GFP_KERNEL);
> +		if (!ndata)
> +			return ERR_PTR(-ENOMEM);
> +
> +		/*
> +		 * SRC and DST nodes should have matching TAG in order to have
> +		 * it set by default for a requested path.
> +		 */
> +		ndata->tag = TEGRA_MC_ICC_TAG_ISO;
> +		ndata->node = node;
> +
> +		return ndata;
> +	}
> +
> +	return ERR_PTR(-EINVAL);
> +}
> +
> +static int emc_icc_set(struct icc_node *src, struct icc_node *dst)
> +{
> +	struct tegra_emc *emc = to_tegra_emc_provider(dst->provider);
> +	unsigned long long peak_bw = icc_units_to_bps(dst->peak_bw);
> +	unsigned long long avg_bw = icc_units_to_bps(dst->avg_bw);
> +	unsigned long long rate = max(avg_bw, peak_bw);
> +	unsigned int dram_data_bus_width_bytes = 4;

Perhaps use something shorter for this variable (like dram_bus_width)? Also,
since it's never modified, perhaps make it const? Or a #define?

> +	long rounded_rate;
> +	int err;
> +
> +	/*
> +	 * Tegra20 EMC runs on x2 clock rate of SDRAM bus because DDR data
> +	 * is sampled on both clock edges. This means that EMC clock rate
> +	 * equals to the peak data rate.
> +	 */
> +	do_div(rate, dram_data_bus_width_bytes);
> +	rate = min_t(u64, rate, U32_MAX);
> +
> +	rounded_rate = emc_round_rate(rate, 0, U32_MAX, emc);
> +	if (rounded_rate < 0)
> +		return rounded_rate;
> +
> +	err = dev_pm_opp_set_rate(emc->dev, rounded_rate);
> +	if (err)
> +		return err;
> +
> +	return 0;
> +}
> +
> +static int tegra_emc_interconnect_init(struct tegra_emc *emc)
> +{
> +	const struct tegra_mc_soc *soc;
> +	struct icc_node *node;
> +	int err;
> +
> +	emc->mc = devm_tegra_get_memory_controller(emc->dev);
> +	if (IS_ERR(emc->mc))
> +		return PTR_ERR(emc->mc);
> +
> +	soc = emc->mc->soc;
> +
> +	emc->provider.dev = emc->dev;
> +	emc->provider.set = emc_icc_set;
> +	emc->provider.data = &emc->provider;
> +	emc->provider.aggregate = soc->icc_ops->aggregate;
> +	emc->provider.xlate_extended = emc_of_icc_xlate_extended;
> +
> +	err = icc_provider_add(&emc->provider);
> +	if (err)
> +		goto err_msg;
> +
> +	/* create External Memory Controller node */
> +	node = icc_node_create(TEGRA_ICC_EMC);
> +	err = PTR_ERR_OR_ZERO(node);
> +	if (err)
> +		goto del_provider;

As far as I can tell, icc_node_create() always returns either a valid
pointer or an ERR_PTR-encoded negative error-code. So I think the more
idiomatic way to write this would be:

	node = icc_node_create(TEGRA_ICC_EMC);
	if (IS_ERR(node)) {
		err = PTR_ERR(node);
		goto del_provider;
	}

> +
> +	node->name = "External Memory Controller";
> +	icc_node_add(node, &emc->provider);
> +
> +	/* link External Memory Controller to External Memory (DRAM) */
> +	err = icc_link_create(node, TEGRA_ICC_EMEM);
> +	if (err)
> +		goto remove_nodes;
> +
> +	/* create External Memory node */
> +	node = icc_node_create(TEGRA_ICC_EMEM);
> +	err = PTR_ERR_OR_ZERO(node);
> +	if (err)
> +		goto remove_nodes;

Same here.

> +
> +	node->name = "External Memory (DRAM)";
> +	icc_node_add(node, &emc->provider);
> +
> +	return 0;
> +
> +remove_nodes:
> +	icc_nodes_remove(&emc->provider);
> +del_provider:
> +	icc_provider_del(&emc->provider);
> +err_msg:
> +	dev_err(emc->dev, "failed to initialize ICC: %d\n", err);

It might be worth duplicating this error message to the failure
locations so that the exact failure can be identified.

> +
> +	return err;
> +}
> +
> +static int tegra_emc_opp_table_init(struct tegra_emc *emc)
> +{
> +	const char *rname = "core";
> +	int err;
> +
> +	/*
> +	 * Legacy device-trees don't have OPP table and EMC driver isn't
> +	 * useful in this case.
> +	 */
> +	if (!device_property_present(emc->dev, "operating-points-v2")) {
> +		dev_err(emc->dev, "OPP table not found\n");
> +		dev_err(emc->dev, "please update your device tree\n");

This should be a single error message. These messages end up in kmsg
records and having this split into two dev_err() calls makes them into
two separate records and that in turn makes it more difficult to
determine whether they belong together or not.

> +		return -ENODEV;
> +	}
> +
> +	/* voltage scaling is optional */
> +	if (device_property_present(emc->dev, "core-supply"))
> +		emc->opp_table = dev_pm_opp_set_regulators(emc->dev, &rname, 1);
> +	else
> +		emc->opp_table = dev_pm_opp_get_opp_table(emc->dev);
> +
> +	if (IS_ERR(emc->opp_table))
> +		return dev_err_probe(emc->dev, PTR_ERR(emc->opp_table),
> +				     "failed to prepare OPP table\n");
> +
> +	err = dev_pm_opp_of_add_table(emc->dev);
> +	if (err) {
> +		dev_err(emc->dev, "failed to add OPP table: %d\n", err);
> +		goto put_table;
> +	}
> +
> +	return 0;
> +
> +put_table:
> +	dev_pm_opp_put_opp_table(emc->opp_table);
> +
> +	return err;
> +}
> +
>  static int tegra_emc_probe(struct platform_device *pdev)
>  {
>  	struct device_node *np;
> @@ -717,8 +885,13 @@ static int tegra_emc_probe(struct platform_device *pdev)
>  		goto unset_cb;
>  	}
>  
> +	err = tegra_emc_opp_table_init(emc);
> +	if (err)
> +		goto unreg_notifier;
> +
>  	platform_set_drvdata(pdev, emc);
>  	tegra_emc_debugfs_init(emc);
> +	tegra_emc_interconnect_init(emc);
>  
>  	/*
>  	 * Don't allow the kernel module to be unloaded. Unloading adds some
> @@ -729,6 +902,8 @@ static int tegra_emc_probe(struct platform_device *pdev)
>  
>  	return 0;
>  
> +unreg_notifier:
> +	clk_notifier_unregister(emc->clk, &emc->clk_nb);
>  unset_cb:
>  	tegra20_clk_set_emc_round_callback(NULL, NULL);
>  
> @@ -747,6 +922,7 @@ static struct platform_driver tegra_emc_driver = {
>  		.name = "tegra20-emc",
>  		.of_match_table = tegra_emc_of_match,
>  		.suppress_bind_attrs = true,
> +		.sync_state = icc_sync_state,
>  	},
>  };
>  module_platform_driver(tegra_emc_driver);
> diff --git a/drivers/memory/tegra/tegra20.c b/drivers/memory/tegra/tegra20.c
> index a8098bff91d9..5127e8e8250f 100644
> --- a/drivers/memory/tegra/tegra20.c
> +++ b/drivers/memory/tegra/tegra20.c
> @@ -280,6 +280,39 @@ static const struct tegra_mc_reset_ops tegra20_mc_reset_ops = {
>  	.reset_status = tegra20_mc_reset_status,
>  };
>  
> +static int tegra20_mc_icc_set(struct icc_node *src, struct icc_node *dst)
> +{
> +	/*
> +	 * Technically, it should be possible to tune arbitration knobs here,
> +	 * but the default values are known to work well on all devices.
> +	 * Hence nothing to do here so far.
> +	 */
> +	return 0;
> +}
> +
> +static int tegra20_mc_icc_aggreate(struct icc_node *node, u32 tag, u32 avg_bw,
> +				   u32 peak_bw, u32 *agg_avg, u32 *agg_peak)
> +{
> +	/*
> +	 * ISO clients need to reserve extra bandwidth up-front because
> +	 * there could high bandwidth pressure during initial fulling-up

"filling of the client's FIFO buffers"

> +	 * of the client's FIFO buffers. Secondly, we need to take into
> +	 * account impurities of the memory subsystem.
> +	 */
> +	if (tag == TEGRA_MC_ICC_TAG_ISO)
> +		peak_bw = tegra_mc_scale_percents(peak_bw, 300);

300% sounds a bit excessive. Do we really need that much?

> +
> +	*agg_avg += avg_bw;
> +	*agg_peak = max(*agg_peak, peak_bw);

I'm not very familiar with ICC, but shouldn't the aggregated peak value
be the sum of the current aggregated peak and the new peak bandwidth?
Currently you're selecting the maximum peak bandwidth across all
clients, so isn't that going to be too small if for whatever reason
multiple clients need peak bandwidth at the same time?

Thierry

Attachment: signature.asc
Description: PGP signature