Re: [PATCH v8 14/21] clk: tegra210: Add suspend and resume support

[Dmitry Osipenko]

09.08.2019 19:19, Sowjanya Komatineni ÐÐÑÐÑ:
> 
> On 8/9/19 6:56 AM, Dmitry Osipenko wrote:
>> 09.08.2019 2:46, Sowjanya Komatineni ÐÐÑÐÑ:
>>> This patch adds support for clk: tegra210: suspend-resume.
>>>
>>> All the CAR controller settings are lost on suspend when core
>>> power goes off.
>>>
>>> This patch has implementation for saving and restoring all PLLs
>>> and clocks context during system suspend and resume to have the
>>> clocks back to same state for normal operation.
>>>
>>> Clock driver suspend and resume are registered as syscore_ops as clocks
>>> restore need to happen before the other drivers resume to have all their
>>> clocks back to the same state as before suspend.
>>>
>>> Signed-off-by: Sowjanya Komatineni <skomatineni@xxxxxxxxxx>
>>> ---
>>> Â drivers/clk/tegra/clk-tegra210.c | 103 +++++++++++++++++++++++++++++++++++++--
>>> Â drivers/clk/tegra/clk.cÂÂÂÂÂÂÂÂÂ |Â 64 ++++++++++++++++++++++++
>>> Â drivers/clk/tegra/clk.hÂÂÂÂÂÂÂÂÂ |ÂÂ 3 ++
>>> Â 3 files changed, 166 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/drivers/clk/tegra/clk-tegra210.c b/drivers/clk/tegra/clk-tegra210.c
>>> index 998bf60b219a..8dd6f4f4debb 100644
>>> --- a/drivers/clk/tegra/clk-tegra210.c
>>> +++ b/drivers/clk/tegra/clk-tegra210.c
>>> @@ -9,13 +9,13 @@
>>> Â #include <linux/clkdev.h>
>>> Â #include <linux/of.h>
>>> Â #include <linux/of_address.h>
>>> +#include <linux/syscore_ops.h>
>>> Â #include <linux/delay.h>
>>> Â #include <linux/export.h>
>>> Â #include <linux/mutex.h>
>>> Â #include <linux/clk/tegra.h>
>>> Â #include <dt-bindings/clock/tegra210-car.h>
>>> Â #include <dt-bindings/reset/tegra210-car.h>
>>> -#include <linux/iopoll.h>
>>> Â #include <linux/sizes.h>
>>> Â #include <soc/tegra/pmc.h>
>>> Â @@ -220,11 +220,15 @@
>>> Â #define CLK_M_DIVISOR_SHIFT 2
>>> Â #define CLK_M_DIVISOR_MASK 0x3
>>> Â +#define CLK_MASK_ARMÂÂÂ 0x44
>>> +#define MISC_CLK_ENBÂÂÂ 0x48
>>> +
>>> Â #define RST_DFLL_DVCO 0x2f4
>>> Â #define DVFS_DFLL_RESET_SHIFT 0
>>> Â Â #define CLK_RST_CONTROLLER_RST_DEV_Y_SET 0x2a8
>>> Â #define CLK_RST_CONTROLLER_RST_DEV_Y_CLR 0x2ac
>>> +#define CPU_SOFTRST_CTRL 0x380
>>> Â Â #define LVL2_CLK_GATE_OVRA 0xf8
>>> Â #define LVL2_CLK_GATE_OVRC 0x3a0
>>> @@ -2825,6 +2829,7 @@ static int tegra210_enable_pllu(void)
>>> ÂÂÂÂÂ struct tegra_clk_pll_freq_table *fentry;
>>> ÂÂÂÂÂ struct tegra_clk_pll pllu;
>>> ÂÂÂÂÂ u32 reg;
>>> +ÂÂÂ int ret;
>>> Â ÂÂÂÂÂ for (fentry = pll_u_freq_table; fentry->input_rate; fentry++) {
>>> ÂÂÂÂÂÂÂÂÂ if (fentry->input_rate == pll_ref_freq)
>>> @@ -2853,9 +2858,14 @@ static int tegra210_enable_pllu(void)
>>> ÂÂÂÂÂ reg |= PLL_ENABLE;
>>> ÂÂÂÂÂ writel(reg, clk_base + PLLU_BASE);
>>> Â -ÂÂÂ readl_relaxed_poll_timeout_atomic(clk_base + PLLU_BASE, reg,
>>> -ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ reg & PLL_BASE_LOCK, 2, 1000);
>>> -ÂÂÂ if (!(reg & PLL_BASE_LOCK)) {
>>> +ÂÂÂ /*
>>> +ÂÂÂÂ * During clocks resume, same PLLU init and enable sequence get
>>> +ÂÂÂÂ * executed. So, readx_poll_timeout_atomic can't be used here as it
>>> +ÂÂÂÂ * uses ktime_get() and timekeeping resume doesn't happen by that
>>> +ÂÂÂÂ * time. So, using tegra210_wait_for_mask for PLL LOCK.
>>> +ÂÂÂÂ */
>>> +ÂÂÂ ret = tegra210_wait_for_mask(&pllu, PLLU_BASE, PLL_BASE_LOCK);
>>> +ÂÂÂ if (ret) {
>>> ÂÂÂÂÂÂÂÂÂ pr_err("Timed out waiting for PLL_U to lock\n");
>>> ÂÂÂÂÂÂÂÂÂ return -ETIMEDOUT;
>>> ÂÂÂÂÂ }
>>> @@ -3288,6 +3298,84 @@ static void tegra210_disable_cpu_clock(u32 cpu)
>>> Â }
>>> Â Â #ifdef CONFIG_PM_SLEEP
>>> +/*
>>> + * This array lists mask values for each peripheral clk bank
>>> + * to mask out reserved bits during the clocks state restore
>>> + * on SC7 resume to prevent accidental writes to these reserved
>>> + * bits.
>>> + */
>>> +static u32 periph_clk_rsvd_mask[TEGRA210_CAR_BANK_COUNT] = {
>> Should be more natural to have a "valid_mask" instead of "rsvd_mask".
>>
>> What's actually wrong with touching of the reserved bits? They must be NO-OP.. or the
>> reserved bits are actually some kind of "secret" bits? If those bits have some use-case
>> outside of Silicon HW (like FPGA simulation), then this doesn't matter for upstream and you
>> have to keep the workaround locally in the downstream kernel or whatever.
> 
> Will rename as valid_mask.
> 
> some bits in these registers are undefined and is not good to write to these bits as they
> can cause pslverr.

Okay, it should be explained in the comment.

Is it possible to disable trapping of changing the undefined bits?

>>
>>> +ÂÂÂ 0x23282006,
>>> +ÂÂÂ 0x782e0c18,
>>> +ÂÂÂ 0x0c012c05,
>>> +ÂÂÂ 0x003e7304,
>>> +ÂÂÂ 0x86c04800,
>>> +ÂÂÂ 0xc0199000,
>>> +ÂÂÂ 0x03e03800,
>>> +};
>>> +
>>> +#define car_readl(_base, _off) readl_relaxed(clk_base + (_base) + ((_off) * 4))
>>> +#define car_writel(_val, _base, _off) \
>>> +ÂÂÂÂÂÂÂ writel_relaxed(_val, clk_base + (_base) + ((_off) * 4))
>>> +
>>> +static u32 spare_reg_ctx, misc_clk_enb_ctx, clk_msk_arm_ctx;
>>> +static u32 cpu_softrst_ctx[3];
>>> +
>>> +static int tegra210_clk_suspend(void)
>>> +{
>>> +ÂÂÂ unsigned int i;
>>> +
>>> +ÂÂÂ clk_save_context();
>>> +
>>> +ÂÂÂ /*
>>> +ÂÂÂÂ * Save the bootloader configured clock registers SPARE_REG0,
>>> +ÂÂÂÂ * MISC_CLK_ENB, CLK_MASK_ARM, CPU_SOFTRST_CTRL.
>>> +ÂÂÂÂ */
>>> +ÂÂÂ spare_reg_ctx = readl_relaxed(clk_base + SPARE_REG0);
>>> +ÂÂÂ misc_clk_enb_ctx = readl_relaxed(clk_base + MISC_CLK_ENB);
>>> +ÂÂÂ clk_msk_arm_ctx = readl_relaxed(clk_base + CLK_MASK_ARM);
>>> +
>>> +ÂÂÂ for (i = 0; i < ARRAY_SIZE(cpu_softrst_ctx); i++)
>>> +ÂÂÂÂÂÂÂ cpu_softrst_ctx[i] = car_readl(CPU_SOFTRST_CTRL, i);
>>> +
>>> +ÂÂÂ tegra_clk_periph_suspend();
>>> +ÂÂÂ return 0;
>>> +}
>>> +
>>> +static void tegra210_clk_resume(void)
>>> +{
>>> +ÂÂÂ unsigned int i;
>>> +
>>> +ÂÂÂ tegra_clk_osc_resume(clk_base);
>>> +
>>> +ÂÂÂ /*
>>> +ÂÂÂÂ * Restore the bootloader configured clock registers SPARE_REG0,
>>> +ÂÂÂÂ * MISC_CLK_ENB, CLK_MASK_ARM, CPU_SOFTRST_CTRL from saved context.
>>> +ÂÂÂÂ */
>>> +ÂÂÂ writel_relaxed(spare_reg_ctx, clk_base + SPARE_REG0);
>>> +ÂÂÂ writel_relaxed(misc_clk_enb_ctx, clk_base + MISC_CLK_ENB);
>>> +ÂÂÂ writel_relaxed(clk_msk_arm_ctx, clk_base + CLK_MASK_ARM);
>>> +
>>> +ÂÂÂ for (i = 0; i < ARRAY_SIZE(cpu_softrst_ctx); i++)
>>> +ÂÂÂÂÂÂÂ car_writel(cpu_softrst_ctx[i], CPU_SOFTRST_CTRL, i);
>>> +
>>> +ÂÂÂ fence_udelay(5, clk_base);
>>> +
>>> +ÂÂÂ /* enable all the clocks before changing the clock sources */
>>> +ÂÂÂ tegra_clk_periph_force_on(periph_clk_rsvd_mask);
>> Why clocks need to be enabled before changing the sources?
> 
> To prevent glitchless frequency switch, Tegra clock programming recommended sequence is to
> change MUX control or divisor or both with the clocks running.

This should be explained in the comment.

> Actual state of clocks before suspend are restored later after all PLL's and peripheral
> clocks are restored.
> 
>>
>>> +ÂÂÂ /* wait for all writes to happen to have all the clocks enabled */
>>> +ÂÂÂ wmb();
>> fence_udelay() has exactly the same barrier at the very beginning of readl(), no need to
>> duplicate it here.

Actually, readl does the rmb() and it should be a more correct variant of fencing because it
actually ensures that the write reached hardware. I suppose that something like fence_udelay
should be used for the pinctrl as well.

>>> +ÂÂÂ fence_udelay(2, clk_base);
>>> +
>>> +ÂÂÂ /* restore PLLs and all peripheral clock rates */
>>> +ÂÂÂ tegra210_init_pllu();
>> Why USB PLL need to be restored at first?
> USB PLL restore is independent to all other clocks restore. So this can be done either
> before clk_restore_context or even after.

Then why not to implement restore_context for PLLU?

>>> +ÂÂÂ clk_restore_context();
>>> +
>>> +ÂÂÂ /* restore all peripheral clocks enable and reset state */
>>> +ÂÂÂ tegra_clk_periph_resume();
>>> +}
>> [snip]