Re: [PATCH v2 5/5] memory: tegra: Introduce Tegra20 EMC driver

From: Dmitry Osipenko
Date: Wed Jun 06 2018 - 09:42:12 EST


On 06.06.2018 14:02, Thierry Reding wrote:
> On Mon, Jun 04, 2018 at 01:36:54AM +0300, Dmitry Osipenko wrote:
>> Introduce driver for the External Memory Controller (EMC) found on Tegra20
>> chips, which controls the external DRAM on the board. The purpose of this
>> driver is to program memory timing for external memory on the EMC clock
>> rate change.
>>
>> Signed-off-by: Dmitry Osipenko <digetx@xxxxxxxxx>
>> ---
>> drivers/memory/tegra/Kconfig | 10 +
>> drivers/memory/tegra/Makefile | 1 +
>> drivers/memory/tegra/tegra20-emc.c | 586 +++++++++++++++++++++++++++++
>> 3 files changed, 597 insertions(+)
>> create mode 100644 drivers/memory/tegra/tegra20-emc.c
>>
>> diff --git a/drivers/memory/tegra/Kconfig b/drivers/memory/tegra/Kconfig
>> index 6d74e499e18d..34e0b70f5c5f 100644
>> --- a/drivers/memory/tegra/Kconfig
>> +++ b/drivers/memory/tegra/Kconfig
>> @@ -6,6 +6,16 @@ config TEGRA_MC
>> This driver supports the Memory Controller (MC) hardware found on
>> NVIDIA Tegra SoCs.
>>
>> +config TEGRA20_EMC
>> + bool "NVIDIA Tegra20 External Memory Controller driver"
>> + default y
>> + depends on ARCH_TEGRA_2x_SOC
>> + help
>> + This driver is for the External Memory Controller (EMC) found on
>> + Tegra20 chips. The EMC controls the external DRAM on the board.
>> + This driver is required to change memory timings / clock rate for
>> + external memory.
>> +
>> config TEGRA124_EMC
>> bool "NVIDIA Tegra124 External Memory Controller driver"
>> default y
>> diff --git a/drivers/memory/tegra/Makefile b/drivers/memory/tegra/Makefile
>> index 94ab16ba075b..3971a6b7c487 100644
>> --- a/drivers/memory/tegra/Makefile
>> +++ b/drivers/memory/tegra/Makefile
>> @@ -10,5 +10,6 @@ tegra-mc-$(CONFIG_ARCH_TEGRA_210_SOC) += tegra210.o
>>
>> obj-$(CONFIG_TEGRA_MC) += tegra-mc.o
>>
>> +obj-$(CONFIG_TEGRA20_EMC) += tegra20-emc.o
>> obj-$(CONFIG_TEGRA124_EMC) += tegra124-emc.o
>> obj-$(CONFIG_ARCH_TEGRA_186_SOC) += tegra186.o
>> diff --git a/drivers/memory/tegra/tegra20-emc.c b/drivers/memory/tegra/tegra20-emc.c
>> new file mode 100644
>> index 000000000000..26a18b5e7941
>> --- /dev/null
>> +++ b/drivers/memory/tegra/tegra20-emc.c
>> @@ -0,0 +1,586 @@
>> +// SPDX-License-Identifier: GPL-2.0
>> +/*
>> + * Tegra20 External Memory Controller driver
>> + *
>> + * Author: Dmitry Osipenko <digetx@xxxxxxxxx>
>> + */
>> +
>> +#include <linux/clk.h>
>> +#include <linux/completion.h>
>> +#include <linux/err.h>
>> +#include <linux/interrupt.h>
>> +#include <linux/iopoll.h>
>> +#include <linux/kernel.h>
>> +#include <linux/module.h>
>> +#include <linux/of.h>
>> +#include <linux/platform_device.h>
>> +#include <linux/sort.h>
>> +#include <linux/types.h>
>> +
>> +#include <soc/tegra/fuse.h>
>> +
>> +#define EMC_INTSTATUS 0x000
>> +#define EMC_INTMASK 0x004
>> +#define EMC_TIMING_CONTROL 0x028
>> +#define EMC_RC 0x02c
>> +#define EMC_RFC 0x030
>> +#define EMC_RAS 0x034
>> +#define EMC_RP 0x038
>> +#define EMC_R2W 0x03c
>> +#define EMC_W2R 0x040
>> +#define EMC_R2P 0x044
>> +#define EMC_W2P 0x048
>> +#define EMC_RD_RCD 0x04c
>> +#define EMC_WR_RCD 0x050
>> +#define EMC_RRD 0x054
>> +#define EMC_REXT 0x058
>> +#define EMC_WDV 0x05c
>> +#define EMC_QUSE 0x060
>> +#define EMC_QRST 0x064
>> +#define EMC_QSAFE 0x068
>> +#define EMC_RDV 0x06c
>> +#define EMC_REFRESH 0x070
>> +#define EMC_BURST_REFRESH_NUM 0x074
>> +#define EMC_PDEX2WR 0x078
>> +#define EMC_PDEX2RD 0x07c
>> +#define EMC_PCHG2PDEN 0x080
>> +#define EMC_ACT2PDEN 0x084
>> +#define EMC_AR2PDEN 0x088
>> +#define EMC_RW2PDEN 0x08c
>> +#define EMC_TXSR 0x090
>> +#define EMC_TCKE 0x094
>> +#define EMC_TFAW 0x098
>> +#define EMC_TRPAB 0x09c
>> +#define EMC_TCLKSTABLE 0x0a0
>> +#define EMC_TCLKSTOP 0x0a4
>> +#define EMC_TREFBW 0x0a8
>> +#define EMC_QUSE_EXTRA 0x0ac
>> +#define EMC_ODT_WRITE 0x0b0
>> +#define EMC_ODT_READ 0x0b4
>> +#define EMC_FBIO_CFG5 0x104
>> +#define EMC_FBIO_CFG6 0x114
>> +#define EMC_AUTO_CAL_INTERVAL 0x2a8
>> +#define EMC_CFG_2 0x2b8
>> +#define EMC_CFG_DIG_DLL 0x2bc
>> +#define EMC_DLL_XFORM_DQS 0x2c0
>> +#define EMC_DLL_XFORM_QUSE 0x2c4
>> +#define EMC_ZCAL_REF_CNT 0x2e0
>> +#define EMC_ZCAL_WAIT_CNT 0x2e4
>> +#define EMC_CFG_CLKTRIM_0 0x2d0
>> +#define EMC_CFG_CLKTRIM_1 0x2d4
>> +#define EMC_CFG_CLKTRIM_2 0x2d8
>> +
>> +#define EMC_CLKCHANGE_REQ_ENABLE BIT(0)
>> +#define EMC_CLKCHANGE_PD_ENABLE BIT(1)
>> +#define EMC_CLKCHANGE_SR_ENABLE BIT(2)
>> +
>> +#define EMC_TIMING_UPDATE BIT(0)
>> +
>> +#define EMC_CLKCHANGE_COMPLETE_INT BIT(4)
>> +
>> +static const unsigned long emc_timing_registers[] = {
>> + EMC_RC,
>> + EMC_RFC,
>> + EMC_RAS,
>> + EMC_RP,
>> + EMC_R2W,
>> + EMC_W2R,
>> + EMC_R2P,
>> + EMC_W2P,
>> + EMC_RD_RCD,
>> + EMC_WR_RCD,
>> + EMC_RRD,
>> + EMC_REXT,
>> + EMC_WDV,
>> + EMC_QUSE,
>> + EMC_QRST,
>> + EMC_QSAFE,
>> + EMC_RDV,
>> + EMC_REFRESH,
>> + EMC_BURST_REFRESH_NUM,
>> + EMC_PDEX2WR,
>> + EMC_PDEX2RD,
>> + EMC_PCHG2PDEN,
>> + EMC_ACT2PDEN,
>> + EMC_AR2PDEN,
>> + EMC_RW2PDEN,
>> + EMC_TXSR,
>> + EMC_TCKE,
>> + EMC_TFAW,
>> + EMC_TRPAB,
>> + EMC_TCLKSTABLE,
>> + EMC_TCLKSTOP,
>> + EMC_TREFBW,
>> + EMC_QUSE_EXTRA,
>> + EMC_FBIO_CFG6,
>> + EMC_ODT_WRITE,
>> + EMC_ODT_READ,
>> + EMC_FBIO_CFG5,
>> + EMC_CFG_DIG_DLL,
>> + EMC_DLL_XFORM_DQS,
>> + EMC_DLL_XFORM_QUSE,
>> + EMC_ZCAL_REF_CNT,
>> + EMC_ZCAL_WAIT_CNT,
>> + EMC_AUTO_CAL_INTERVAL,
>> + EMC_CFG_CLKTRIM_0,
>> + EMC_CFG_CLKTRIM_1,
>> + EMC_CFG_CLKTRIM_2,
>> +};
>> +
>> +struct emc_timing {
>> + unsigned long rate;
>> + u32 emc_registers_data[ARRAY_SIZE(emc_timing_registers)];
>> +};
>
> Nit: this seems like a very long variable name for something that is
> really just "values" or "data" written into a set of registers.
>

Ok.

>> +struct tegra_emc {
>> + struct device *dev;
>> + struct notifier_block clk_nb;
>> + struct clk *backup_clk;
>> + struct clk *emc_mux;
>> + struct clk *pll_m;
>> + struct clk *clk;
>> + void __iomem *regs;
>> +
>> + struct completion clk_handshake_complete;
>> + int irq;
>> +
>> + struct emc_timing *timings;
>> + unsigned int num_timings;
>> +};
>> +
>> +static irqreturn_t tegra_emc_isr(int irq, void *data)
>> +{
>> + struct tegra_emc *emc = data;
>> + u32 intmask = EMC_CLKCHANGE_COMPLETE_INT;
>> + u32 status;
>> +
>> + status = readl_relaxed(emc->regs + EMC_INTSTATUS) & intmask;
>> + if (!status)
>> + return IRQ_NONE;
>> +
>> + /* clear interrupts */
>> + writel_relaxed(status, emc->regs + EMC_INTSTATUS);
>
> Do we really want to just clear the handshake complete interrupt or do
> we want to clear all of them? Perhaps we should also warn if there are
> other interrupts that we're not handling? Currently we'd only get some
> warning if another interrupt triggered without the handshake complete
> one triggering at the same time, but couldn't there be others asserted
> along with the handshake complete interrupt? In which case we'd just
> be ignoring them. Or perhaps not clearing it would get the ISR run
> immediately again and produce the "nobody cared" warning?
>

Yes, we really want to just clear the handshake-complete interrupt. No, we
shouldn't warn about other interrupts because IRQ subsys does it for us. Other
interrupts shouldn't be asserted because we disabled them with the interrupts
masking in emc_setup_hw(). Other interrupts can only be asserted in a case of a
bug, there will be a "nobody cared" warning and interrupt will be disabled, this
is exactly what we want in that case.

>> +
>> + /* notify about EMC-CAR handshake completion */
>> + complete(&emc->clk_handshake_complete);
>> +
>> + return IRQ_HANDLED;
>> +}
>> +
>> +static struct emc_timing *tegra_emc_find_timing(struct tegra_emc *emc,
>> + unsigned long rate)
>> +{
>> + struct emc_timing *timing = NULL;
>> + unsigned int i;
>> +
>> + for (i = 0; i < emc->num_timings; i++) {
>> + if (emc->timings[i].rate >= rate) {
>> + timing = &emc->timings[i];
>> + break;
>> + }
>> + }
>> +
>> + if (!timing) {
>> + dev_err(emc->dev, "no timing for rate %lu\n", rate);
>> + return NULL;
>> + }
>> +
>> + return timing;
>> +}
>> +
>> +static int emc_prepare_timing_change(struct tegra_emc *emc, unsigned long rate)
>> +{
>> + struct emc_timing *timing = tegra_emc_find_timing(emc, rate);
>> + unsigned int i;
>> +
>> + if (!timing)
>> + return -ENOENT;
>> +
>> + dev_dbg(emc->dev, "%s: timing rate %lu emc rate %lu\n",
>> + __func__, timing->rate, rate);
>> +
>> + /* program shadow registers */
>> + for (i = 0; i < ARRAY_SIZE(timing->emc_registers_data); i++)
>> + writel_relaxed(timing->emc_registers_data[i],
>> + emc->regs + emc_timing_registers[i]);
>> +
>> + /* wait until programming has settled */
>> + readl_relaxed(emc->regs + emc_timing_registers[0]);
>> +
>> + if (emc->irq < 0)
>> + writel_relaxed(EMC_CLKCHANGE_COMPLETE_INT,
>> + emc->regs + EMC_INTMASK);
>> + else
>> + reinit_completion(&emc->clk_handshake_complete);
>> +
>> + return 0;
>> +}
>> +
>> +static int emc_complete_timing_change(struct tegra_emc *emc, bool flush)
>> +{
>> + long timeout;
>> + u32 value;
>> + int err;
>> +
>> + dev_dbg(emc->dev, "%s: flush %d\n", __func__, flush);
>> +
>> + if (flush) {
>> + /* manually initiate memory timing update */
>> + writel_relaxed(EMC_TIMING_UPDATE,
>> + emc->regs + EMC_TIMING_CONTROL);
>> + return 0;
>> + }
>> +
>> + if (emc->irq < 0) {
>> + /* poll interrupt status if IRQ isn't available */
>> + err = readl_relaxed_poll_timeout(emc->regs + EMC_INTSTATUS,
>> + value, value & EMC_CLKCHANGE_COMPLETE_INT,
>> + 1, 100);
>> + if (err) {
>> + dev_err(emc->dev, "EMC-CAR handshake failed\n");
>> + return -EIO;
>> + }
>> +
>> + return 0;
>> + }
>> +
>> + timeout = wait_for_completion_timeout(&emc->clk_handshake_complete,
>> + usecs_to_jiffies(100));
>> + if (timeout == 0) {
>> + dev_err(emc->dev, "EMC handshake failed\n");
>> + return -EIO;
>> + } else if (timeout < 0) {
>> + dev_err(emc->dev, "failed to wait for EMC-CAR handshake: %ld\n",
>> + timeout);
>> + return timeout;
>> + }
>> +
>> + return 0;
>> +}
>> +
>> +static int load_one_timing_from_dt(struct tegra_emc *emc,
>> + struct emc_timing *timing,
>> + struct device_node *node)
>> +{
>> + u32 rate;
>> + int err;
>> +
>> + if (!of_device_is_compatible(node, "nvidia,tegra20-emc-table")) {
>> + dev_err(emc->dev, "incompatible DT node \"%s\"\n",
>> + node->name);
>> + return -EINVAL;
>> + }
>> +
>> + err = of_property_read_u32(node, "clock-frequency", &rate);
>> + if (err) {
>> + dev_err(emc->dev, "timing %s: failed to read rate: %d\n",
>> + node->name, err);
>> + return err;
>> + }
>> +
>> + err = of_property_read_u32_array(node, "nvidia,emc-registers",
>> + timing->emc_registers_data,
>> + ARRAY_SIZE(emc_timing_registers));
>> + if (err) {
>> + dev_err(emc->dev,
>> + "timing %s: failed to read emc timing data: %d\n",
>> + node->name, err);
>> + return err;
>> + }
>> +
>> + /*
>> + * The EMC clock rate is twice the bus rate, and the bus rate is
>> + * measured in kHz.
>> + */
>> + timing->rate = rate * 2 * 1000;
>> +
>> + dev_dbg(emc->dev, "%s: emc rate %ld\n", __func__, timing->rate);
>
> Nit: %lu for timing->rate?
>

Yes.

>> +
>> + return 0;
>> +}
>> +
>> +static int cmp_timings(const void *_a, const void *_b)
>> +{
>> + const struct emc_timing *a = _a;
>> + const struct emc_timing *b = _b;
>> +
>> + if (a->rate < b->rate)
>> + return -1;
>> + else if (a->rate == b->rate)
>> + return 0;
>> + else
>> + return 1;
>
> Nit, I tend to
>

Tend to..?

>> +}
>> +
>> +static int tegra_emc_load_timings_from_dt(struct tegra_emc *emc,
>> + struct device_node *node)
>> +{
>> + struct device_node *child;
>> + struct emc_timing *timing;
>> + int child_count;
>> + int err;
>> +
>> + child_count = of_get_child_count(node);
>
> It's unfortunate that of_get_child_count() doesn't return unsigned int,
> there's no reason why this would have to be signed.
>

Patches are welcome ;)

>> + if (!child_count) {
>> + dev_err(emc->dev, "no memory timings in DT node\n");
>> + return -ENOENT;
>> + }
>> +
>> + emc->timings = devm_kcalloc(emc->dev, child_count, sizeof(*timing),
>> + GFP_KERNEL);
>> + if (!emc->timings)
>> + return -ENOMEM;
>> +
>> + emc->num_timings = child_count;
>> + timing = emc->timings;
>> +
>> + for_each_child_of_node(node, child) {
>> + err = load_one_timing_from_dt(emc, timing++, child);
>> + if (err) {
>> + of_node_put(child);
>> + return err;
>> + }
>> + }
>> +
>> + sort(emc->timings, emc->num_timings, sizeof(*timing), cmp_timings,
>> + NULL);
>> +
>> + return 0;
>> +}
>> +
>> +static struct device_node *
>> +tegra_emc_find_node_by_ram_code(struct tegra_emc *emc, u32 ram_code)
>> +{
>> + struct device_node *np;
>> + int err;
>> +
>> + for_each_child_of_node(emc->dev->of_node, np) {
>> + u32 value;
>> +
>> + err = of_property_read_u32(np, "nvidia,ram-code", &value);
>> + if (err || value != ram_code)
>> + continue;
>> +
>> + return np;
>> + }
>> +
>> + dev_info(emc->dev, "no memory timings for RAM code %u found in DT\n",
>> + ram_code);
>
> This seems like it should be dev_warn() or perhaps even dev_err() given
> that the result of it is the driver failing to probe. dev_info() may go
> unnoticed.
>

Absence of memory timings is a valid case, hence dev_info() suit well here.

I can't see anything wrong with returning a errno if driver has nothing to do
and prefer to keep it because in that case managed resources would be free'd by
the driver core, though returning '0' also would work.

>> +
>> + return NULL;
>> +}
>> +
>> +static int tegra_emc_clk_change_notify(struct notifier_block *nb,
>> + unsigned long msg, void *data)
>> +{
>> + struct tegra_emc *emc = container_of(nb, struct tegra_emc, clk_nb);
>> + struct clk_notifier_data *cnd = data;
>> + int err;
>> +
>> + switch (msg) {
>> + case PRE_RATE_CHANGE:
>> + err = emc_prepare_timing_change(emc, cnd->new_rate);
>> + break;
>> +
>> + case ABORT_RATE_CHANGE:
>> + err = emc_prepare_timing_change(emc, cnd->old_rate);
>> + if (err)
>> + break;
>> +
>> + err = emc_complete_timing_change(emc, true);
>> + break;
>> +
>> + case POST_RATE_CHANGE:
>> + err = emc_complete_timing_change(emc, false);
>> + break;
>> +
>> + default:
>> + return NOTIFY_DONE;
>> + }
>> +
>> + return notifier_from_errno(err);
>> +}
>> +
>> +static int emc_setup_hw(struct tegra_emc *emc)
>> +{
>> + u32 emc_cfg;
>> +
>> + emc_cfg = readl_relaxed(emc->regs + EMC_CFG_2);
>> +
>> + /*
>> + * Depending on a memory type, DRAM should enter either self-refresh
>> + * or power-down state on EMC clock change.
>> + */
>> + if (!(emc_cfg & EMC_CLKCHANGE_PD_ENABLE) &&
>> + !(emc_cfg & EMC_CLKCHANGE_SR_ENABLE))
>> + {
>> + dev_err(emc->dev,
>> + "bootloader didn't specify DRAM auto-suspend mode\n");
>> + return -EINVAL;
>> + }
>> +
>> + /* allow EMC and CAR to handshake on PLL divider/source changes */
>> + emc_cfg |= EMC_CLKCHANGE_REQ_ENABLE;
>> + writel_relaxed(emc_cfg, emc->regs + EMC_CFG_2);
>> +
>> + /* initialize interrupt */
>> + writel_relaxed(EMC_CLKCHANGE_COMPLETE_INT, emc->regs + EMC_INTMASK);
>> + writel_relaxed(EMC_CLKCHANGE_COMPLETE_INT, emc->regs + EMC_INTSTATUS);
>> +
>> + return 0;
>> +}
>> +
>> +static int emc_init(struct tegra_emc *emc, unsigned long rate)
>> +{
>> + int err;
>> +
>> + err = clk_set_parent(emc->emc_mux, emc->backup_clk);
>> + if (err) {
>> + dev_err(emc->dev,
>> + "failed to reparent to backup source: %d\n", err);
>> + return err;
>> + }
>> +
>> + err = clk_set_rate(emc->pll_m, rate);
>> + if (err)
>> + dev_err(emc->dev,
>> + "failed to change pll_m rate: %d\n", err);
>> +
>> + err = clk_set_parent(emc->emc_mux, emc->pll_m);
>> + if (err) {
>> + dev_err(emc->dev,
>> + "failed to reparent to pll_m: %d\n", err);
>> + return err;
>> + }
>> +
>> + return 0;
>> +}
>> +
>> +static int tegra_emc_probe(struct platform_device *pdev)
>> +{
>> + struct device_node *np;
>> + struct tegra_emc *emc;
>> + struct resource *res;
>> + u32 ram_code;
>> + int err;
>> +
>> + emc = devm_kzalloc(&pdev->dev, sizeof(*emc), GFP_KERNEL);
>> + if (!emc)
>> + return -ENOMEM;
>> +
>> + emc->dev = &pdev->dev;
>> +
>> + ram_code = tegra_read_ram_code();
>> +
>> + np = tegra_emc_find_node_by_ram_code(emc, ram_code);
>> + if (!np)
>> + return -ENOENT;
>> +
>> + err = tegra_emc_load_timings_from_dt(emc, np);
>> + of_node_put(np);
>> + if (err)
>> + return err;
>> +
>> + res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
>> + emc->regs = devm_ioremap_resource(&pdev->dev, res);
>> + if (IS_ERR(emc->regs))
>> + return PTR_ERR(emc->regs);
>> +
>> + err = emc_setup_hw(emc);
>> + if (err)
>> + return err;
>> +
>> + emc->irq = platform_get_irq(pdev, 0);
>> + if (emc->irq < 0) {
>> + dev_warn(&pdev->dev, "interrupt not specified\n");
>> + dev_warn(&pdev->dev, "continuing, but please update your DT\n");
>
> Do we really need this? I think this is a case where we don't have to
> keep backwards-compatibility because this driver hasn't "worked" in a
> very long time (because it was absent). Therefore, if we error out in
> the absence of an interrupt we don't break anything.
>
> There's a few places in this driver that are awkward just because the
> interrupt isn't mandatory. I don't think it's warranted in this case.
>

Backwards compatibility is always nice to have, but I don't really mind dropping it.

>> + } else {
>> + init_completion(&emc->clk_handshake_complete);
>> +
>> + err = devm_request_irq(&pdev->dev, emc->irq, tegra_emc_isr, 0,
>> + dev_name(&pdev->dev), emc);
>> + if (err < 0) {
>> + dev_err(&pdev->dev, "failed to request IRQ#%u: %d\n",
>> + emc->irq, err);
>> + return err;
>> + }
>> + }
>> +
>> + emc->pll_m = clk_get_sys(NULL, "pll_m");
>> + if (IS_ERR(emc->pll_m)) {
>> + err = PTR_ERR(emc->pll_m);
>> + dev_err(&pdev->dev, "failed to get pll_m: %d\n", err);
>> + return err;
>> + }
>> +
>> + emc->backup_clk = clk_get_sys(NULL, "pll_p");
>> + if (IS_ERR(emc->backup_clk)) {
>> + err = PTR_ERR(emc->backup_clk);
>> + dev_err(&pdev->dev, "failed to get pll_p: %d\n", err);
>> + goto put_pll_m;
>> + }
>> +
>> + emc->clk = clk_get_sys(NULL, "emc");
>> + if (IS_ERR(emc->clk)) {
>> + err = PTR_ERR(emc->clk);
>> + dev_err(&pdev->dev, "failed to get emc: %d\n", err);
>> + goto put_backup;
>> + }
>
> Instead of using clk_get_sys(), why not specify these in the DT with
> proper names for context ("emc", "pll", "backup")? Again, I don't think
> we have to worry about backwards-compatibility here since there can be
> no regression.
>

I don't think that "pll" and "backup" could be placed in DT because it is a pure
software-driver description.

"emc" could be retrieved from DT if we don't care about backwards compatibility.