From patchwork Wed Jan 15 23:24:41 2020 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Sean Anderson X-Patchwork-Id: 239633 List-Id: U-Boot discussion From: seanga2 at gmail.com (Sean Anderson) Date: Wed, 15 Jan 2020 18:24:41 -0500 Subject: [PATCH v2 10/11] riscv: Add K210 pll support In-Reply-To: <3411d84b-5a56-ae4e-1deb-085ef1a4971d@gmail.com> References: <3411d84b-5a56-ae4e-1deb-085ef1a4971d@gmail.com> Message-ID: This pll code is primarily based on the code from the kendryte standalone sdk in lib/drivers/sysctl.c. k210_pll_calc_params is roughly analogous to the algorithm used to set the pll frequency, but it has been completely rewritten to be fixed-point based. Signed-off-by: Sean Anderson --- Changes for v2. Rename driver to "k210_clk_pll" Add additional in-line documentation on algorithm and PLLs. Restrict the range of internal VCO and reference frequencies. Don't load driver before relocation. Remove spurious references to mach-k210. board/sipeed/maix/Kconfig | 1 + configs/sipeed_maix_bitm_defconfig | 1 + drivers/clk/Kconfig | 1 + drivers/clk/Makefile | 1 + drivers/clk/kendryte/Kconfig | 7 + drivers/clk/kendryte/Makefile | 1 + drivers/clk/kendryte/pll.c | 598 +++++++++++++++++++++++++++++ drivers/clk/kendryte/pll.h | 38 ++ 8 files changed, 648 insertions(+) create mode 100644 drivers/clk/kendryte/Kconfig create mode 100644 drivers/clk/kendryte/Makefile create mode 100644 drivers/clk/kendryte/pll.c create mode 100644 drivers/clk/kendryte/pll.h diff --git a/board/sipeed/maix/Kconfig b/board/sipeed/maix/Kconfig index 9ffd3aa7aa..10e492a425 100644 --- a/board/sipeed/maix/Kconfig +++ b/board/sipeed/maix/Kconfig @@ -37,6 +37,7 @@ config BOARD_SPECIFIC_OPTIONS select SYS_RISCV_NOCOUNTER imply SIFIVE_CLINT imply K210_SYSCTL + imply CLK_K210 imply SPI imply DM_GPIO imply CMD_GPIO diff --git a/configs/sipeed_maix_bitm_defconfig b/configs/sipeed_maix_bitm_defconfig index ecd69c0873..aeab7db0c3 100644 --- a/configs/sipeed_maix_bitm_defconfig +++ b/configs/sipeed_maix_bitm_defconfig @@ -65,6 +65,7 @@ CONFIG_OF_TRANSLATE=y CONFIG_CLK=y CONFIG_CLK_CCF=y CONFIG_CLK_COMPOSITE_CCF=y +CONFIG_CLK_K210=y CONFIG_CPU=y CONFIG_CPU_RISCV=y CONFIG_MMC=y diff --git a/drivers/clk/Kconfig b/drivers/clk/Kconfig index 16d4237f89..af75c7c4cf 100644 --- a/drivers/clk/Kconfig +++ b/drivers/clk/Kconfig @@ -145,6 +145,7 @@ source "drivers/clk/analogbits/Kconfig" source "drivers/clk/at91/Kconfig" source "drivers/clk/exynos/Kconfig" source "drivers/clk/imx/Kconfig" +source "drivers/clk/kendryte/Kconfig" source "drivers/clk/meson/Kconfig" source "drivers/clk/mvebu/Kconfig" source "drivers/clk/owl/Kconfig" diff --git a/drivers/clk/Makefile b/drivers/clk/Makefile index 06131edb9f..4f3893f6fc 100644 --- a/drivers/clk/Makefile +++ b/drivers/clk/Makefile @@ -26,6 +26,7 @@ obj-$(CONFIG_CLK_BCM6345) += clk_bcm6345.o obj-$(CONFIG_CLK_BOSTON) += clk_boston.o obj-$(CONFIG_CLK_EXYNOS) += exynos/ obj-$(CONFIG_CLK_HSDK) += clk-hsdk-cgu.o +obj-$(CONFIG_CLK_K210) += kendryte/ obj-$(CONFIG_CLK_MPC83XX) += mpc83xx_clk.o obj-$(CONFIG_CLK_OWL) += owl/ obj-$(CONFIG_CLK_RENESAS) += renesas/ diff --git a/drivers/clk/kendryte/Kconfig b/drivers/clk/kendryte/Kconfig new file mode 100644 index 0000000000..a178d50f5e --- /dev/null +++ b/drivers/clk/kendryte/Kconfig @@ -0,0 +1,7 @@ +config CLK_K210 + bool "Clock support for Kendryte K210" + depends on K210_SYSCTL + select CLK + select CLK_CCF + help + This enables support clock driver for Kendryte K210 platforms. diff --git a/drivers/clk/kendryte/Makefile b/drivers/clk/kendryte/Makefile new file mode 100644 index 0000000000..c56d93ea1c --- /dev/null +++ b/drivers/clk/kendryte/Makefile @@ -0,0 +1 @@ +obj-y += pll.o diff --git a/drivers/clk/kendryte/pll.c b/drivers/clk/kendryte/pll.c new file mode 100644 index 0000000000..cbb112f5a8 --- /dev/null +++ b/drivers/clk/kendryte/pll.c @@ -0,0 +1,598 @@ +// SPDX-License-Identifier: GPL-2.0+ +/* + * Copyright (C) 2019 Sean Anderson + */ +#include "pll.h" + +#define LOG_CATEGORY UCLASS_CLK +#include +/* For DIV_ROUND_DOWN_ULL, defined in linux/kernel.h */ +#include +#include +#include +#include +#include +#include +#include + +#define CLK_K210_PLL "k210_clk_pll" +#define to_k210_pll(_clk) container_of(_clk, struct k210_pll, clk) + +static int k210_pll_enable(struct clk *clk); +static int k210_pll_disable(struct clk *clk); + +/* + * The logical layout of the PLL is approximately the following: + * + * +-----------+ + * |input clock| + * +-----------+ + * | + * v + * +--+ + * |/r| + * +--+ + * | + * v + * +---------------+ + * |reference clock| + * +---------------+ + * | + * v + * +----------------+ + * |phase comparator|<--+ + * +----------------+ | + * | | + * v | + * +---+ | + * |VCO| | + * +---+ | + * | +--+ | + * +--->|/f|---+ + * | +--+ + * v + * +---+ + * |/od| + * +---+ + * | + * v + * +------+ + * |output| + * +------+ + * + * The k210 PLLs have three factors: r, f, and od. Because of the feedback mode, + * the effect of the division by f is to multiply the input frequency. The + * equation for the output rate is + * rate = (rate_in * f) / (r * od). + * Moving knowns to one side of the equation, we get + * rate / rate_in = f / (r * od) + * Rearranging slightly, + * abs_error = abs((rate / rate_in) - (f / (r * od))). + * To get relative, error, we divide by the expected ratio + * error = abs((rate / rate_in) - (f / (r * od))) / (rate / rate_in). + * Simplifying, + * error = abs(1 - f / (r * od)) / (rate / rate_in) + * error = abs(1 - (f * rate_in) / (r * od * rate)) + * Using the constants ratio = rate / rate_in and inv_ratio = rate_in / rate, + * error = abs((f * inv_ratio) / (r * od) - 1) + * This is the error used in evaluating parameters. + * + * r and od are four bits each, while f is six bits. Because r and od are + * multiplied together, instead of the full 256 values possible if both bits + * were used fully, there are only 97 distinct products. Combined with f, there + * are 6208 theoretical settings for the PLL. However, most of these settings + * can be ruled out immediately because they do not have the correct ratio. + * + * In addition to the constraint of approximating the desired ratio, parameters + * must also keep internal pll frequencies within acceptable ranges. The + * reference clock's minimum and maximum frequencies have a ratio of around 128. + * This leaves fairly substantial room work work with, especially since the only + * affected parameter is r. The VCO's minimum and maximum frequency have a ratio + * of 5, which is considerably more restrictive. + * + * The r and od factors are stored in a table. This is to make it easy to find + * the next-largest product. Some products have multiple factorizations, but + * only when one factor has at least a 2.5x ratio to the factors of the other + * factorization. This is because any smaller ratio would not make a difference + * when ensuring the VCO's frequency is within spec. + * + * Throughout the calculation function, fixed point arithmetic is used. Because + * the range of rate and rate_in may be up to 1.75 GHz, or around 2^30, 64-bit + * 32.32 fixed-point numbers are used to represent ratios. In general, to + * implement division, the numerator is first multiplied by 2^32. This gives a + * result where the whole number part is in the upper 32 bits, and the fraction + * is in the lower 32 bits. + * + * In general, rounding is done to the closest integer. This helps find the best + * approximation for the ratio. Rounding in one direction (e.g down) could cause + * the function to miss a better ratio with one of the parameters increased by + * one. + */ + +/* + * The factors table was generated with the following python code: + * + * def p(x, y): + * return (1.0*x/y > 2.5) or (1.0*y/x > 2.5) + * + * factors = {} + * for i in range(1, 17): + * for j in range(1, 17): + * fs = factors.get(i*j) or [] + * if fs == [] or all([ + * (p(i, x) and p(i, y)) or (p(j, x) and p(j, y)) + * for (x, y) in fs]): + * fs.append((i, j)) + * factors[i*j] = fs + * + * for k, l in sorted(factors.items()): + * for v in l: + * print("PACK(%s, %s)," % v) + */ +#define PACK(r, od) ((((r - 1) & 0xF) << 4) | ((od - 1) & 0xF)) +#define UNPACK_R(val) (((val >> 4) & 0xF) + 1) +#define UNPACK_OD(val) ((val & 0xF) + 1) +static const u8 factors[] = { + PACK(1, 1), + PACK(1, 2), + PACK(1, 3), + PACK(1, 4), + PACK(1, 5), + PACK(1, 6), + PACK(1, 7), + PACK(1, 8), + PACK(1, 9), + PACK(3, 3), + PACK(1, 10), + PACK(1, 11), + PACK(1, 12), + PACK(3, 4), + PACK(1, 13), + PACK(1, 14), + PACK(1, 15), + PACK(3, 5), + PACK(1, 16), + PACK(4, 4), + PACK(2, 9), + PACK(2, 10), + PACK(3, 7), + PACK(2, 11), + PACK(2, 12), + PACK(5, 5), + PACK(2, 13), + PACK(3, 9), + PACK(2, 14), + PACK(2, 15), + PACK(2, 16), + PACK(3, 11), + PACK(5, 7), + PACK(3, 12), + PACK(3, 13), + PACK(4, 10), + PACK(3, 14), + PACK(4, 11), + PACK(3, 15), + PACK(3, 16), + PACK(7, 7), + PACK(5, 10), + PACK(4, 13), + PACK(6, 9), + PACK(5, 11), + PACK(4, 14), + PACK(4, 15), + PACK(7, 9), + PACK(4, 16), + PACK(5, 13), + PACK(6, 11), + PACK(5, 14), + PACK(6, 12), + PACK(5, 15), + PACK(7, 11), + PACK(6, 13), + PACK(5, 16), + PACK(9, 9), + PACK(6, 14), + PACK(8, 11), + PACK(6, 15), + PACK(7, 13), + PACK(6, 16), + PACK(7, 14), + PACK(9, 11), + PACK(10, 10), + PACK(8, 13), + PACK(7, 15), + PACK(9, 12), + PACK(10, 11), + PACK(7, 16), + PACK(9, 13), + PACK(8, 15), + PACK(11, 11), + PACK(9, 14), + PACK(8, 16), + PACK(10, 13), + PACK(11, 12), + PACK(9, 15), + PACK(10, 14), + PACK(11, 13), + PACK(9, 16), + PACK(10, 15), + PACK(11, 14), + PACK(12, 13), + PACK(10, 16), + PACK(11, 15), + PACK(12, 14), + PACK(13, 13), + PACK(11, 16), + PACK(12, 15), + PACK(13, 14), + PACK(12, 16), + PACK(13, 15), + PACK(14, 14), + PACK(13, 16), + PACK(14, 15), + PACK(14, 16), + PACK(15, 15), + PACK(15, 16), + PACK(16, 16), +}; + +struct k210_pll_params { + u8 r; + u8 f; + u8 od; +}; + +static int k210_pll_calc_params(u32 rate, u32 rate_in, + struct k210_pll_params *best) +{ + int i; + s64 error, best_error; + u64 ratio, inv_ratio; /* fixed point 32.32 ratio of the rates */ + u64 max_r; + u64 r, f, od; + + /* + * Can't go over 1.75 GHz or under 855 kHz due to limitations on the + * reference clock frequency. These are not the same limits as below + * because od can reduce the output frequency by 16. + */ + if (rate > 1750000000 || rate < 854493) + return -EINVAL; + + ratio = DIV_ROUND_CLOSEST_ULL((u64)rate << 32, rate_in); + inv_ratio = DIV_ROUND_CLOSEST_ULL((u64)rate_in << 32, rate); + /* Can't increase by more than 64 or reduce by more than 256 */ + if (rate > rate_in && ratio > (64ULL << 32)) + return -EINVAL; + else if (rate <= rate_in && inv_ratio > (256ULL << 32)) + return -EINVAL; + + /* + * The reference clock (rate_in / r) must stay between 1.75 GHz and 13 + * MHz. There is no minimum, since the only way to get a higher input + * clock than 26 MHz is to use a clock generated by a PLL. Because PLLs + * cannot output frequencies greater than 1.75 GHz, the minimum would + * never be greater than one. + */ + max_r = DIV_ROUND_DOWN_ULL(rate_in, 13671875); + + /* Variables get immediately incremented, so start at -1th iteration */ + i = -1; + f = 0; + r = 0; + od = 0; + error = best_error = S64_MAX; + /* do-while here so we always try at least one ratio */ + do { + /* + * Whether we swapped r and od while enforcing frequency limits + */ + bool swapped = false; + u64 last_od = od; + u64 last_r = r; + + /* + * Try the next largest value for f (or r and od) and + * recalculate the other parameters based on that + */ + if (rate > rate_in) { + /* + * Skip factors of the same product if we already tried + * out that product + */ + do { + i++; + r = UNPACK_R(factors[i]); + od = UNPACK_OD(factors[i]); + } while (i + 1 < ARRAY_SIZE(factors) + && r * od == last_r * last_od); + + /* Round close */ + f = (r * od * ratio + BIT(31)) >> 32; + if (f > 64) + f = 64; + } else { + u64 tmp = ++f * inv_ratio; + bool round_up = !!(tmp & BIT(31)); + u32 goal = (tmp >> 32) + round_up; + u32 err, last_err; + + /* Get the next r/od pair in factors */ + while (r * od < goal && i + 1 < ARRAY_SIZE(factors)) { + i++; + r = UNPACK_R(factors[i]); + od = UNPACK_OD(factors[i]); + } + + /* + * This is a case of double rounding. If we rounded up + * above, we need to round down (in cases of ties) here. + * This prevents off-by-one errors resulting from + * choosing X+2 over X when X.Y rounds up to X+1 and + * there is no r * od = X+1. For the converse, when X.Y + * is rounded down to X, we should choose X+1 over X-1. + */ + err = abs(r * od - goal); + last_err = abs(last_r * last_od - goal); + if (last_err < err || (round_up && (last_err = err))) { + i--; + r = last_r; + od = last_od; + } + } + + /* + * Enforce limits on internal clock frequencies. If we + * aren't in spec, try swapping r and od. If everything is + * in-spec, calculate the relative error. + */ + while (true) { + /* + * Whether the intermediate frequencies are out-of-spec + */ + bool out_of_spec = false; + + if (r > max_r) { + out_of_spec = true; + } else { + /* + * There is no way to only divide once; we need + * to examine the frequency with and without the + * effect of od. + */ + u64 vco = DIV_ROUND_CLOSEST_ULL(rate_in * f, r); + if (vco > 1750000000 || vco < 350000000) + out_of_spec = true; + } + + if (out_of_spec) { + if (!swapped) { + u64 tmp = r; + r = od; + od = tmp; + swapped = true; + continue; + } else { + /* + * Try looking ahead to see if there are + * additional factors for the same + * product. + */ + if (i + 1 < ARRAY_SIZE(factors)) { + u64 new_r, new_od; + + i++; + new_r = UNPACK_R(factors[i]); + new_od = UNPACK_OD(factors[i]); + if (r * od == new_r * new_od) { + r = new_r; + od = new_od; + swapped = false; + continue; + } + i--; + } + break; + } + } + + error = DIV_ROUND_CLOSEST_ULL(f * inv_ratio, r * od); + /* The lower 16 bits are spurious */ + error = abs((error - BIT(32))) >> 16; + + if (error < best_error) { + best->r = r; + best->f = f; + best->od = od; + best_error = error; + } + break; + } + } while (f < 64 && i + 1 < ARRAY_SIZE(factors) && error != 0); + + if (best_error == S64_MAX) + return -EINVAL; + + log_debug("best error %lld\n", best_error); + return 0; +} + +static ulong k210_pll_set_rate(struct clk *clk, ulong rate) +{ + int err; + long long rate_in = clk_get_parent_rate(clk); + struct k210_pll_params params = {}; + struct k210_pll *pll = to_k210_pll(clk); + u32 reg; + + if (rate_in < 0) + return rate_in; + + log_debug("Calculating parameters with rate=%lu and rate_in=%lld\n", + rate, rate_in); + err = k210_pll_calc_params(rate, rate_in, ¶ms); + if (err) + return err; + log_debug("Got r=%u f=%u od=%u\n", params.r, params.f, params.od); + + /* + * Don't use clk_disable as it might not actually disable the pll due to + * refcounting + */ + err = k210_pll_disable(clk); + if (err) + return err; + + reg = readl(pll->reg); + reg &= ~K210_PLL_CLKR + & ~K210_PLL_CLKF + & ~K210_PLL_CLKOD + & ~K210_PLL_BWADJ; + reg |= FIELD_PREP(K210_PLL_CLKR, params.r - 1) + | FIELD_PREP(K210_PLL_CLKF, params.f - 1) + | FIELD_PREP(K210_PLL_CLKOD, params.od - 1) + | FIELD_PREP(K210_PLL_BWADJ, params.f - 1); + writel(reg, pll->reg); + + err = k210_pll_enable(clk); + if (err) + return err; + + return clk_get_rate(clk); +} + +static ulong k210_pll_get_rate(struct clk *clk) +{ + + long long rate_in = clk_get_parent_rate(clk); + struct k210_pll *pll = to_k210_pll(clk); + u64 r, f, od; + u32 reg = readl(pll->reg); + + if (rate_in < 0) + return rate_in; + + if (reg & K210_PLL_BYPASS) + return rate_in; + + r = FIELD_GET(K210_PLL_CLKR, reg) + 1; + f = FIELD_GET(K210_PLL_CLKF, reg) + 1; + od = FIELD_GET(K210_PLL_CLKOD, reg) + 1; + + return DIV_ROUND_DOWN_ULL(((u64)rate_in) * f, r * od); +} + +/* Check if the PLL is locked */ +static int k210_pll_locked(struct k210_pll *pll) +{ + u32 reg = readl(pll->lock); + + return (reg & pll->lock_mask) == pll->lock_mask; +} + +/* + * Wait for the PLL to be locked. If the PLL is not locked, try clearing the + * slip before retrying + */ +static void k210_pll_waitfor_lock(struct k210_pll *pll) +{ + while (!k210_pll_locked(pll)) { + u32 reg = readl(pll->lock); + + reg |= BIT(pll->shift + K210_PLL_CLEAR_SLIP); + writel(reg, pll->lock); + udelay(1); + } +} + +/* Adapted from sysctl_pll_enable */ +static int k210_pll_enable(struct clk *clk) +{ + struct k210_pll *pll = to_k210_pll(clk); + u32 reg = readl(pll->reg); + + reg &= ~K210_PLL_BYPASS; + writel(reg, pll->reg); + + reg |= K210_PLL_PWRD; + writel(reg, pll->reg); + + /* Ensure reset is low before asserting it */ + reg &= ~K210_PLL_RESET; + writel(reg, pll->reg); + reg |= K210_PLL_RESET; + writel(reg, pll->reg); + /* FIXME: this doesn't really have to be a whole microsecond */ + udelay(1); + reg &= ~K210_PLL_RESET; + writel(reg, pll->reg); + + k210_pll_waitfor_lock(pll); + return 0; +} + +static int k210_pll_disable(struct clk *clk) +{ + struct k210_pll *pll = to_k210_pll(clk); + u32 reg = readl(pll->reg); + + /* + * Bypassing before powering off is important so child clocks don't stop + * working. This is especially important for pll0, the indirect parent + * of the cpu clock. + */ + reg |= K210_PLL_BYPASS; + writel(reg, pll->reg); + + reg &= ~K210_PLL_PWRD; + writel(reg, pll->reg); + return 0; +} + +const struct clk_ops k210_pll_ops = { + .get_rate = k210_pll_get_rate, + .set_rate = k210_pll_set_rate, + .enable = k210_pll_enable, + .disable = k210_pll_disable, +}; + +struct k210_pll *k210_clk_comp_pll(void __iomem *reg, void __iomem *lock, + u8 shift, u8 width) +{ + struct k210_pll *pll; + + + pll = kzalloc(sizeof(*pll), GFP_KERNEL); + if (!pll) + return pll; + pll->reg = reg; + pll->lock = lock; + pll->shift = shift; + pll->lock_mask = GENMASK(shift + width, shift); + return pll; +} + +struct clk *k210_clk_pll(const char *name, const char *parent_name, + void __iomem *reg, void __iomem *lock, u8 shift, + u8 width) +{ + int err; + struct k210_pll *pll; + + pll = k210_clk_comp_pll(reg, lock, shift, width); + if (!pll) + return ERR_PTR(-ENOMEM); + + err = clk_register(&pll->clk, CLK_K210_PLL, name, parent_name); + if (err) { + kfree(pll); + return ERR_PTR(err); + } + return &pll->clk; +} + +U_BOOT_DRIVER(k210_pll) = { + .name = CLK_K210_PLL, + .id = UCLASS_CLK, + .ops = &k210_pll_ops, +}; diff --git a/drivers/clk/kendryte/pll.h b/drivers/clk/kendryte/pll.h new file mode 100644 index 0000000000..6a8734a295 --- /dev/null +++ b/drivers/clk/kendryte/pll.h @@ -0,0 +1,38 @@ +#ifndef K210_PLL_H +#define K210_PLL_H + +#include + +#define K210_PLL_CLKR GENMASK(3, 0) +#define K210_PLL_CLKF GENMASK(9, 4) +#define K210_PLL_CLKOD GENMASK(13, 10) +#define K210_PLL_BWADJ GENMASK(19, 14) +#define K210_PLL_RESET BIT(20) +#define K210_PLL_PWRD BIT(21) +#define K210_PLL_INTFB BIT(22) +#define K210_PLL_BYPASS BIT(23) +#define K210_PLL_TEST BIT(24) +#define K210_PLL_EN BIT(25) +#define K210_PLL_TEST_EN BIT(26) + +#define K210_PLL_LOCK 0 +#define K210_PLL_CLEAR_SLIP 2 +#define K210_PLL_TEST_OUT 3 + +struct k210_pll { + struct clk clk; + void __iomem *reg; /* Base PLL register */ + void __iomem *lock; /* Common PLL lock register */ + u8 shift; /* Offset of bits in lock register */ + u8 lock_mask; /* Mask of lock bits to test against, pre-shifted */ +}; + +extern const struct clk_ops k210_pll_ops; + +struct k210_pll *k210_clk_comp_pll(void __iomem *reg, void __iomem *lock, + u8 shift, u8 width); +struct clk *k210_clk_pll(const char *name, const char *parent_name, + void __iomem *reg, void __iomem *lock, u8 shift, + u8 width); + +#endif /* K210_PLL_H */