Re: [PATCH v2 2/3] clk: clk-apple-nco: Add driver for Apple NCO

From: Stephen Boyd
Date: Thu Jan 20 2022 - 00:38:14 EST


Quoting Martin Povišer (2022-01-18 11:21:10)
> diff --git a/drivers/clk/clk-apple-nco.c b/drivers/clk/clk-apple-nco.c
> new file mode 100644
> index 000000000000..593f5b5ce5b7
> --- /dev/null
> +++ b/drivers/clk/clk-apple-nco.c
> @@ -0,0 +1,340 @@
> +// SPDX-License-Identifier: GPL-2.0-only OR MIT
> +/*
> + * Driver for an SoC block (Numerically Controlled Oscillator)
> + * found on t8103 (M1) and other Apple chips
> + *
> + * Copyright (C) The Asahi Linux Contributors
> + */
> +
> +#include <linux/bits.h>
> +#include <linux/clk-provider.h>
> +#include <linux/io.h>
> +#include <linux/math64.h>
> +#include <linux/module.h>
> +#include <linux/of.h>
> +#include <linux/of_clk.h>

Is this include used? If not, please drop it.

Please include kernel.h for container_of() usage.

> +#include <linux/platform_device.h>
> +#include <linux/spinlock.h>
> +
> +#define NCO_CHANNEL_STRIDE 0x4000
> +#define NCO_CHANNEL_REGSIZE 20
> +
> +#define REG_CTRL 0
> +#define CTRL_ENABLE BIT(31)
> +#define REG_DIV 4
> +#define DIV_FINE GENMASK(1, 0)
> +#define DIV_COARSE GENMASK(12, 2)
> +#define REG_INC1 8
> +#define REG_INC2 12
> +#define REG_ACCINIT 16
> +
> +/*
> + * Theory of operation (postulated)
> + *
> + * The REG_DIV register indirectly expresses a base integer divisor, roughly
> + * corresponding to twice the desired ratio of input to output clock. This
> + * base divisor is adjusted on a cycle-by-cycle basis based on the state of a
> + * 32-bit phase accumulator to achieve a desired precise clock ratio over the
> + * long term.
> + *
> + * Specifically an output clock cycle is produced after (REG_DIV divisor)/2
> + * or (REG_DIV divisor + 1)/2 input cycles, the latter taking effect when top
> + * bit of the 32-bit accumulator is set. The accumulator is incremented each
> + * produced output cycle, by the value from either REG_INC1 or REG_INC2, which
> + * of the two is selected depending again on the accumulator's current top bit.
> + *
> + * Because the NCO hardware implements counting of input clock cycles in part
> + * in a Galois linear-feedback shift register, the higher bits of divisor
> + * are programmed into REG_DIV by picking an appropriate LFSR state. See
> + * nco_compute_tables/nco_div_translate for details on this.
> + */
> +
> +struct nco_tables;

Please declare the struct here.

> +
> +struct nco_channel {
> + void __iomem *base;
> + struct nco_tables *tbl;
> + struct clk_hw hw;
> +
> + spinlock_t lock;
> +};
> +
> +#define to_nco_channel(_hw) container_of(_hw, struct nco_channel, hw)
> +
> +#define LFSR_POLY 0xa01
> +#define LFSR_INIT 0x7ff
> +#define LFSR_LEN 11
> +#define LFSR_PERIOD ((1 << LFSR_LEN) - 1)
> +#define LFSR_TBLSIZE (1 << LFSR_LEN)
> +
> +/* The minimal attainable coarse divisor (first value in table) */
> +#define COARSE_DIV_OFFSET 2
> +
> +struct nco_tables {
> + u16 fwd[LFSR_TBLSIZE];
> + u16 inv[LFSR_TBLSIZE];
> +};

Or put struct nco_channel here.

> +
> +static void nco_enable_nolock(struct clk_hw *hw);
> +static void nco_disable_nolock(struct clk_hw *hw);
> +static int nco_is_enabled(struct clk_hw *hw);

Define the functions here so we don't need forward declarations.

> +
> +static void nco_compute_tables(struct nco_tables *tbl)
> +{
> + int i;
> + u32 state = LFSR_INIT;
> +
> + /*
> + * Go through the states of a Galois LFSR and build
> + * a coarse divisor translation table.
> + */
> + for (i = LFSR_PERIOD; i > 0; i--) {
> + if (state & 1)
> + state = (state >> 1) ^ (LFSR_POLY >> 1);
> + else
> + state = (state >> 1);
> + tbl->fwd[i] = state;
> + tbl->inv[state] = i;
> + }
> +
> + /* Zero value is special-cased */
> + tbl->fwd[0] = 0;
> + tbl->inv[0] = 0;
> +}
> +
> +static bool nco_div_out_of_range(unsigned int div)
> +{
> + unsigned int coarse = div / 4;

Nitpick: Newline here

> + return coarse < COARSE_DIV_OFFSET ||
> + coarse >= COARSE_DIV_OFFSET + LFSR_TBLSIZE;
> +}
> +
> +static u32 nco_div_translate(struct nco_tables *tbl, unsigned int div)
> +{
> + unsigned int coarse = div / 4;
> +
> + if (WARN_ON(nco_div_out_of_range(div)))

Maybe worth knowing which clk is out of range?

> + return 0;
> +
> + return FIELD_PREP(DIV_COARSE, tbl->fwd[coarse - COARSE_DIV_OFFSET]) |
> + FIELD_PREP(DIV_FINE, div % 4);
> +}
> +
> +static unsigned int nco_div_translate_inv(struct nco_tables *tbl, u32 regval)
> +{
> + unsigned int coarse, fine;
> +
> + coarse = tbl->inv[FIELD_GET(DIV_COARSE, regval)] + COARSE_DIV_OFFSET;
> + fine = FIELD_GET(DIV_FINE, regval);
> +
> + return coarse * 4 + fine;
> +}
> +
> +static int nco_set_rate(struct clk_hw *hw, unsigned long rate,
> + unsigned long parent_rate)
> +{
> + struct nco_channel *chan = to_nco_channel(hw);
> + unsigned long flags;
> + u32 div;
> + s32 inc1, inc2;
> + bool was_enabled;
> +
> + div = 2 * parent_rate / rate;
> + inc1 = 2 * parent_rate - div * rate;
> + inc2 = -((s32) (rate - inc1));

Is the cast necessary?

> +
> + if (nco_div_out_of_range(div))
> + return -EINVAL;
> +
> + div = nco_div_translate(chan->tbl, div);
> +
> + spin_lock_irqsave(&chan->lock, flags);
> + was_enabled = nco_is_enabled(hw);
> + nco_disable_nolock(hw);
> +
> + writel_relaxed(div, chan->base + REG_DIV);
> + writel_relaxed(inc1, chan->base + REG_INC1);
> + writel_relaxed(inc2, chan->base + REG_INC2);
> +
> + /* Presumably a neutral initial value for accumulator */
> + writel_relaxed(1 << 31, chan->base + REG_ACCINIT);
> +
> + if (was_enabled)
> + nco_enable_nolock(hw);
> + spin_unlock_irqrestore(&chan->lock, flags);
> +
> + return 0;
> +}
> +
> +static unsigned long nco_recalc_rate(struct clk_hw *hw,
> + unsigned long parent_rate)
> +{
> + struct nco_channel *chan = to_nco_channel(hw);
> + u32 div;
> + s32 inc1, inc2, incbase;
> +
> + div = nco_div_translate_inv(chan->tbl,
> + readl_relaxed(chan->base + REG_DIV));
> +
> + inc1 = readl_relaxed(chan->base + REG_INC1);
> + inc2 = readl_relaxed(chan->base + REG_INC2);
> +
> + /*
> + * We don't support wraparound of accumulator
> + * nor the edge case of both increments being zero
> + */
> + if (inc1 < 0 || inc2 > 0 || (inc1 == 0 && inc2 == 0))
> + return 0;
> +
> + /* Scale both sides of division by incbase to maintain precision */
> + incbase = inc1 - inc2;
> +
> + return div_u64(((u64) parent_rate) * 2 * incbase,
> + ((u64) div) * incbase + inc1);

Why is the divisor casted to 64 bits? div_u64() takes a u32 divisor so
if it's going to overflow 32 bits we're in trouble.
> +}
> +
> +static long nco_round_rate(struct clk_hw *hw, unsigned long rate,
> + unsigned long *parent_rate)
> +{
> + unsigned long lo = *parent_rate / (COARSE_DIV_OFFSET + LFSR_TBLSIZE) + 1;
> + unsigned long hi = *parent_rate / COARSE_DIV_OFFSET;
> +
> + return clamp(rate, lo, hi);
> +}
> +
> +static void nco_enable_nolock(struct clk_hw *hw)
> +{
> + struct nco_channel *chan = to_nco_channel(hw);
> + u32 val;
> +
> + val = readl_relaxed(chan->base + REG_CTRL);
> + writel_relaxed(val | CTRL_ENABLE, chan->base + REG_CTRL);
> +}
> +
> +static void nco_disable_nolock(struct clk_hw *hw)
> +{
> + struct nco_channel *chan = to_nco_channel(hw);
> + u32 val;
> +
> + val = readl_relaxed(chan->base + REG_CTRL);
> + writel_relaxed(val & ~CTRL_ENABLE, chan->base + REG_CTRL);
> +}
> +
> +static int nco_is_enabled(struct clk_hw *hw)
> +{
> + struct nco_channel *chan = to_nco_channel(hw);
> +
> + return (readl_relaxed(chan->base + REG_CTRL) & CTRL_ENABLE) != 0;
> +}
> +
> +static int nco_enable(struct clk_hw *hw)
> +{
> + struct nco_channel *chan = to_nco_channel(hw);
> + unsigned long flags;
> +
> + spin_lock_irqsave(&chan->lock, flags);
> + nco_enable_nolock(hw);
> + spin_unlock_irqrestore(&chan->lock, flags);
> +
> + return 0;
> +}
> +
> +static void nco_disable(struct clk_hw *hw)
> +{
> + struct nco_channel *chan = to_nco_channel(hw);
> + unsigned long flags;
> +
> + spin_lock_irqsave(&chan->lock, flags);
> + nco_disable_nolock(hw);
> + spin_unlock_irqrestore(&chan->lock, flags);
> +}
> +
> +static const struct clk_ops nco_ops = {

Perhaps apple_nco_ops (and apple_ prefix for the functions) so that tags
in the global namespace don't conflict.

> + .set_rate = nco_set_rate,
> + .recalc_rate = nco_recalc_rate,
> + .round_rate = nco_round_rate,
> + .enable = nco_enable,
> + .disable = nco_disable,
> + .is_enabled = nco_is_enabled,
> +};
> +
> +static int apple_nco_probe(struct platform_device *pdev)
> +{
> + struct device_node *np = pdev->dev.of_node;
> + struct clk_parent_data pdata = { .index = 0 };
> + struct clk_init_data init;
> + struct clk_hw_onecell_data *onecell_data;
> + void __iomem *regs;

Usually it's called 'base'

> + struct resource *regs_res;

Usually it's called 'res'

> + struct nco_tables *tbl;
> + unsigned int nchannels;
> + int ret, i;
> +
> + regs = devm_platform_get_and_ioremap_resource(pdev, 0, &regs_res);
> + if (IS_ERR(regs))
> + return PTR_ERR(regs);
> +
> + if (resource_size(regs_res) < NCO_CHANNEL_REGSIZE)
> + return -EINVAL;
> + nchannels = (resource_size(regs_res) - NCO_CHANNEL_REGSIZE)
> + / NCO_CHANNEL_STRIDE + 1;

Is this some sort of DIV_ROUND_UP()?

> +
> + onecell_data = devm_kzalloc(&pdev->dev, struct_size(onecell_data, hws,
> + nchannels), GFP_KERNEL);
> + if (!onecell_data)
> + return -ENOMEM;
> + onecell_data->num = nchannels;
> +
> + tbl = devm_kzalloc(&pdev->dev, sizeof(*tbl), GFP_KERNEL);
> + if (!tbl)
> + return -ENOMEM;
> + nco_compute_tables(tbl);
> +
> + for (i = 0; i < nchannels; i++) {
> + struct nco_channel *chan;
> +
> + chan = devm_kzalloc(&pdev->dev, sizeof(*chan), GFP_KERNEL);
> + if (!chan)
> + return -ENOMEM;
> + chan->base = regs + NCO_CHANNEL_STRIDE*i;

Please add space around * above.

> + chan->tbl = tbl;
> + spin_lock_init(&chan->lock);
> +
> + memset(&init, 0, sizeof(init));
> + init.name = devm_kasprintf(&pdev->dev, GFP_KERNEL,
> + "%s-%d", np->name, i);
> + init.ops = &nco_ops;
> + init.parent_data = &pdata;
> + init.num_parents = 1;
> + init.flags = 0;
> +
> + chan->hw.init = &init;
> + ret = devm_clk_hw_register(&pdev->dev, &chan->hw);
> + if (ret)
> + return ret;
> +
> + onecell_data->hws[i] = &chan->hw;
> + }
> +
> + return devm_of_clk_add_hw_provider(&pdev->dev, of_clk_hw_onecell_get,
> + onecell_data);