From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753219AbdDHSv5 (ORCPT ); Sat, 8 Apr 2017 14:51:57 -0400 Received: from hermes.aosc.io ([199.195.250.187]:42554 "EHLO hermes.aosc.io" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752109AbdDHSvr (ORCPT ); Sat, 8 Apr 2017 14:51:47 -0400 From: Icenowy Zheng To: Maxime Ripard , Chen-Yu Tsai , Quentin Schulz Cc: linux-arm-kernel@lists.infradead.org, linux-kernel@vger.kernel.org, linux-clk@vger.kernel.org, linux-pm@vger.kernel.org, linux-sunxi@googlegroups.com, Icenowy Zheng Subject: [PATCH 1/5] clk: sunxi-ng: prevent NKMP clocks from temporarily get higher freq Date: Sun, 9 Apr 2017 02:50:21 +0800 Message-Id: <20170408185025.53841-2-icenowy@aosc.io> In-Reply-To: <20170408185025.53841-1-icenowy@aosc.io> References: <20170408185025.53841-1-icenowy@aosc.io> Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org It seems that on newer SoCs (already observed on A33, H3), when setting all NKMP factors at the same time, the multiplier get applied first, then the divider get applied. In some situations (e.g. the multiplier increased but the divider decreased), this will make the clock frequency temporarily higher than both the original frequency and the target frequency, which may lead to system hang due to PLL_CPU(X) clock usually being a NKMP clock. A comparsion between the old divider (M*P) and the new one is added, and if the divider get smaller when changing clock, the multiplier will be applied first, so that the clock won't go to a frequency higher than normal. The interval between applying the first group of factors and the second group is based on experiments results on an Orange Pi Zero board. Signed-off-by: Icenowy Zheng --- drivers/clk/sunxi-ng/ccu_nkmp.c | 76 +++++++++++++++++++++++++++++++++++------ 1 file changed, 65 insertions(+), 11 deletions(-) diff --git a/drivers/clk/sunxi-ng/ccu_nkmp.c b/drivers/clk/sunxi-ng/ccu_nkmp.c index e58c95787f94..5da77eb60335 100644 --- a/drivers/clk/sunxi-ng/ccu_nkmp.c +++ b/drivers/clk/sunxi-ng/ccu_nkmp.c @@ -9,6 +9,7 @@ */ #include +#include #include "ccu_gate.h" #include "ccu_nkmp.h" @@ -130,12 +131,49 @@ static long ccu_nkmp_round_rate(struct clk_hw *hw, unsigned long rate, return *parent_rate * _nkmp.n * _nkmp.k / (_nkmp.m * _nkmp.p); } +static void ccu_nkmp_extract_factors(const struct ccu_nkmp *nkmp, + struct _ccu_nkmp *_nkmp, u32 reg) +{ + _nkmp->n = ((reg >> nkmp->n.shift) & GENMASK(nkmp->n.width - 1, 0)) + + nkmp->n.offset; + _nkmp->k = ((reg >> nkmp->k.shift) & GENMASK(nkmp->k.width - 1, 0)) + + nkmp->k.offset; + _nkmp->m = ((reg >> nkmp->m.shift) & GENMASK(nkmp->m.width - 1, 0)) + + nkmp->m.offset; + _nkmp->p = 1 << + ((reg >> nkmp->p.shift) & GENMASK(nkmp->p.width - 1, 0)); +} + +static u32 ccu_nkmp_apply_multiplier(const struct ccu_nkmp *nkmp, + const struct _ccu_nkmp *_nkmp, u32 reg) +{ + reg &= ~GENMASK(nkmp->n.width + nkmp->n.shift - 1, nkmp->n.shift); + reg &= ~GENMASK(nkmp->k.width + nkmp->k.shift - 1, nkmp->k.shift); + + reg |= (_nkmp->n - nkmp->n.offset) << nkmp->n.shift; + reg |= (_nkmp->k - nkmp->k.offset) << nkmp->k.shift; + + return reg; +} + +static u32 ccu_nkmp_apply_divider(const struct ccu_nkmp *nkmp, + const struct _ccu_nkmp *_nkmp, u32 reg) +{ + reg &= ~GENMASK(nkmp->m.width + nkmp->m.shift - 1, nkmp->m.shift); + reg &= ~GENMASK(nkmp->p.width + nkmp->p.shift - 1, nkmp->p.shift); + + reg |= (_nkmp->m - nkmp->m.offset) << nkmp->m.shift; + reg |= ilog2(_nkmp->p) << nkmp->p.shift; + + return reg; +} + static int ccu_nkmp_set_rate(struct clk_hw *hw, unsigned long rate, unsigned long parent_rate) { struct ccu_nkmp *nkmp = hw_to_ccu_nkmp(hw); - struct _ccu_nkmp _nkmp; - unsigned long flags; + struct _ccu_nkmp _nkmp, _nkmp_old; + unsigned long flags, old_mp, mp; u32 reg; _nkmp.min_n = nkmp->n.min ?: 1; @@ -152,17 +190,33 @@ static int ccu_nkmp_set_rate(struct clk_hw *hw, unsigned long rate, spin_lock_irqsave(nkmp->common.lock, flags); reg = readl(nkmp->common.base + nkmp->common.reg); - reg &= ~GENMASK(nkmp->n.width + nkmp->n.shift - 1, nkmp->n.shift); - reg &= ~GENMASK(nkmp->k.width + nkmp->k.shift - 1, nkmp->k.shift); - reg &= ~GENMASK(nkmp->m.width + nkmp->m.shift - 1, nkmp->m.shift); - reg &= ~GENMASK(nkmp->p.width + nkmp->p.shift - 1, nkmp->p.shift); - reg |= (_nkmp.n - nkmp->n.offset) << nkmp->n.shift; - reg |= (_nkmp.k - nkmp->k.offset) << nkmp->k.shift; - reg |= (_nkmp.m - nkmp->m.offset) << nkmp->m.shift; - reg |= ilog2(_nkmp.p) << nkmp->p.shift; + ccu_nkmp_extract_factors(nkmp, &_nkmp_old, reg); + + old_mp = _nkmp_old.m * _nkmp_old.p; + mp = _nkmp.m * _nkmp.p; + + if (mp > old_mp) { + reg = ccu_nkmp_apply_divider(nkmp, &_nkmp, reg); + writel(reg, nkmp->common.base + nkmp->common.reg); + + /* + * This value is decided by experiment results on an + * Allwinner H2+ board (Orange Pi Zero). + */ + udelay(500); - writel(reg, nkmp->common.base + nkmp->common.reg); + reg = ccu_nkmp_apply_multiplier(nkmp, &_nkmp, reg); + writel(reg, nkmp->common.base + nkmp->common.reg); + } else { + reg = ccu_nkmp_apply_multiplier(nkmp, &_nkmp, reg); + writel(reg, nkmp->common.base + nkmp->common.reg); + + udelay(500); + + reg = ccu_nkmp_apply_divider(nkmp, &_nkmp, reg); + writel(reg, nkmp->common.base + nkmp->common.reg); + } spin_unlock_irqrestore(nkmp->common.lock, flags); -- 2.12.2 From mboxrd@z Thu Jan 1 00:00:00 1970 From: icenowy@aosc.io (Icenowy Zheng) Date: Sun, 9 Apr 2017 02:50:21 +0800 Subject: [PATCH 1/5] clk: sunxi-ng: prevent NKMP clocks from temporarily get higher freq In-Reply-To: <20170408185025.53841-1-icenowy@aosc.io> References: <20170408185025.53841-1-icenowy@aosc.io> Message-ID: <20170408185025.53841-2-icenowy@aosc.io> To: linux-arm-kernel@lists.infradead.org List-Id: linux-arm-kernel.lists.infradead.org It seems that on newer SoCs (already observed on A33, H3), when setting all NKMP factors at the same time, the multiplier get applied first, then the divider get applied. In some situations (e.g. the multiplier increased but the divider decreased), this will make the clock frequency temporarily higher than both the original frequency and the target frequency, which may lead to system hang due to PLL_CPU(X) clock usually being a NKMP clock. A comparsion between the old divider (M*P) and the new one is added, and if the divider get smaller when changing clock, the multiplier will be applied first, so that the clock won't go to a frequency higher than normal. The interval between applying the first group of factors and the second group is based on experiments results on an Orange Pi Zero board. Signed-off-by: Icenowy Zheng --- drivers/clk/sunxi-ng/ccu_nkmp.c | 76 +++++++++++++++++++++++++++++++++++------ 1 file changed, 65 insertions(+), 11 deletions(-) diff --git a/drivers/clk/sunxi-ng/ccu_nkmp.c b/drivers/clk/sunxi-ng/ccu_nkmp.c index e58c95787f94..5da77eb60335 100644 --- a/drivers/clk/sunxi-ng/ccu_nkmp.c +++ b/drivers/clk/sunxi-ng/ccu_nkmp.c @@ -9,6 +9,7 @@ */ #include +#include #include "ccu_gate.h" #include "ccu_nkmp.h" @@ -130,12 +131,49 @@ static long ccu_nkmp_round_rate(struct clk_hw *hw, unsigned long rate, return *parent_rate * _nkmp.n * _nkmp.k / (_nkmp.m * _nkmp.p); } +static void ccu_nkmp_extract_factors(const struct ccu_nkmp *nkmp, + struct _ccu_nkmp *_nkmp, u32 reg) +{ + _nkmp->n = ((reg >> nkmp->n.shift) & GENMASK(nkmp->n.width - 1, 0)) + + nkmp->n.offset; + _nkmp->k = ((reg >> nkmp->k.shift) & GENMASK(nkmp->k.width - 1, 0)) + + nkmp->k.offset; + _nkmp->m = ((reg >> nkmp->m.shift) & GENMASK(nkmp->m.width - 1, 0)) + + nkmp->m.offset; + _nkmp->p = 1 << + ((reg >> nkmp->p.shift) & GENMASK(nkmp->p.width - 1, 0)); +} + +static u32 ccu_nkmp_apply_multiplier(const struct ccu_nkmp *nkmp, + const struct _ccu_nkmp *_nkmp, u32 reg) +{ + reg &= ~GENMASK(nkmp->n.width + nkmp->n.shift - 1, nkmp->n.shift); + reg &= ~GENMASK(nkmp->k.width + nkmp->k.shift - 1, nkmp->k.shift); + + reg |= (_nkmp->n - nkmp->n.offset) << nkmp->n.shift; + reg |= (_nkmp->k - nkmp->k.offset) << nkmp->k.shift; + + return reg; +} + +static u32 ccu_nkmp_apply_divider(const struct ccu_nkmp *nkmp, + const struct _ccu_nkmp *_nkmp, u32 reg) +{ + reg &= ~GENMASK(nkmp->m.width + nkmp->m.shift - 1, nkmp->m.shift); + reg &= ~GENMASK(nkmp->p.width + nkmp->p.shift - 1, nkmp->p.shift); + + reg |= (_nkmp->m - nkmp->m.offset) << nkmp->m.shift; + reg |= ilog2(_nkmp->p) << nkmp->p.shift; + + return reg; +} + static int ccu_nkmp_set_rate(struct clk_hw *hw, unsigned long rate, unsigned long parent_rate) { struct ccu_nkmp *nkmp = hw_to_ccu_nkmp(hw); - struct _ccu_nkmp _nkmp; - unsigned long flags; + struct _ccu_nkmp _nkmp, _nkmp_old; + unsigned long flags, old_mp, mp; u32 reg; _nkmp.min_n = nkmp->n.min ?: 1; @@ -152,17 +190,33 @@ static int ccu_nkmp_set_rate(struct clk_hw *hw, unsigned long rate, spin_lock_irqsave(nkmp->common.lock, flags); reg = readl(nkmp->common.base + nkmp->common.reg); - reg &= ~GENMASK(nkmp->n.width + nkmp->n.shift - 1, nkmp->n.shift); - reg &= ~GENMASK(nkmp->k.width + nkmp->k.shift - 1, nkmp->k.shift); - reg &= ~GENMASK(nkmp->m.width + nkmp->m.shift - 1, nkmp->m.shift); - reg &= ~GENMASK(nkmp->p.width + nkmp->p.shift - 1, nkmp->p.shift); - reg |= (_nkmp.n - nkmp->n.offset) << nkmp->n.shift; - reg |= (_nkmp.k - nkmp->k.offset) << nkmp->k.shift; - reg |= (_nkmp.m - nkmp->m.offset) << nkmp->m.shift; - reg |= ilog2(_nkmp.p) << nkmp->p.shift; + ccu_nkmp_extract_factors(nkmp, &_nkmp_old, reg); + + old_mp = _nkmp_old.m * _nkmp_old.p; + mp = _nkmp.m * _nkmp.p; + + if (mp > old_mp) { + reg = ccu_nkmp_apply_divider(nkmp, &_nkmp, reg); + writel(reg, nkmp->common.base + nkmp->common.reg); + + /* + * This value is decided by experiment results on an + * Allwinner H2+ board (Orange Pi Zero). + */ + udelay(500); - writel(reg, nkmp->common.base + nkmp->common.reg); + reg = ccu_nkmp_apply_multiplier(nkmp, &_nkmp, reg); + writel(reg, nkmp->common.base + nkmp->common.reg); + } else { + reg = ccu_nkmp_apply_multiplier(nkmp, &_nkmp, reg); + writel(reg, nkmp->common.base + nkmp->common.reg); + + udelay(500); + + reg = ccu_nkmp_apply_divider(nkmp, &_nkmp, reg); + writel(reg, nkmp->common.base + nkmp->common.reg); + } spin_unlock_irqrestore(nkmp->common.lock, flags); -- 2.12.2