[RFC PATCH 1/3] clocksource: exynos_mct: allow mct to read counter from coprocessor

Alexey Klimov klimov.linux at gmail.com
Tue Jul 28 05:49:32 PDT 2015


On Tue, Jul 28, 2015 at 9:11 AM, Krzysztof Kozlowski
<k.kozlowski at samsung.com> wrote:
> On 28.07.2015 06:28, Alexey Klimov wrote:
>> It was discovered that 64-bit mmio MCT counter holds
>> the same value as ARM arch timer 64-bit physical counter.
>> Even more: arch timer and MCT are same underlying hardware
>> timer.
>>
>> Patch adds code to MCT to allow it to read 64-bit counter
>> from coprocessor cp15 registers since it's way more
>> faster than reading the same counter from MCT mmio region.
>>
>> That functionality triggers only if special property
>> use-cp15-phys-counter is present in device tree,
>> only on 32-bit ARMv7 systems and only if HYP mode is
>> available.
>
> Hi,
>
> It would be nice to put here also the measurements from cover letter.
> This would be the justification for the commit.

Ok, I will add them. I will try to find time to perform more precise
measurements.

> I guess same optimization could be done for ARMv8 Exynos SoCs?

I honestly don't know, don't have access to ARMv8 Exynos SoCs.
If i remember correctly there are no public-available 64-bit Exynos
boards yet. Exynos7420 is on the way to be released.

The main concern is that MCT probably not enabled for arm64 and
some cleanup/re-factoring is required for cycles_t variable. See
message in exynos4_read_current_timer() in MCT.
Also ARMv8 requires arm generic timer to be available and fully
operational. So for me it looks like current upstream kernel will work
on top of arm arch timer.

>> Idea of rewriting accessors is taken from arm_arch_timer.c.
>>
>> By default MCT will read counter from mmio since it's
>> guaranteed to work.
>>
>> Signed-off-by: Alexey Klimov <klimov.linux at gmail.com>
>> ---
>>  drivers/clocksource/exynos_mct.c | 77 ++++++++++++++++++++++++++++++++++------
>>  1 file changed, 67 insertions(+), 10 deletions(-)
>>
>> diff --git a/drivers/clocksource/exynos_mct.c b/drivers/clocksource/exynos_mct.c
>> index 9064ff7..039b41c 100644
>> --- a/drivers/clocksource/exynos_mct.c
>> +++ b/drivers/clocksource/exynos_mct.c
>> @@ -26,6 +26,9 @@
>>  #include <linux/clocksource.h>
>>  #include <linux/sched_clock.h>
>>
>> +#include <asm/arch_timer.h>  /* for cp15 physical arch timer counter */
>> +#include <asm/virt.h>                /* for checking HYP mode */
>> +
>>  #define EXYNOS4_MCTREG(x)            (x)
>>  #define EXYNOS4_MCT_G_CNT_L          EXYNOS4_MCTREG(0x100)
>>  #define EXYNOS4_MCT_G_CNT_U          EXYNOS4_MCTREG(0x104)
>> @@ -82,6 +85,17 @@ static unsigned long clk_rate;
>>  static unsigned int mct_int_type;
>>  static int mct_irqs[MCT_NR_IRQS];
>>
>> +static u32 notrace __exynos4_read_count_32(void);
>> +static u64 __exynos4_read_count_64(void);
>> +
>> +/*
>> + * Default to __exynos4_read_count_{32,64} functions that reads counter from
>> + * MCT mmio region and this method is guaranteed
>> + * to exist (if MCT was successfully initialized).
>> + */
>> +u32 (*exynos4_read_count_32)(void) = __exynos4_read_count_32;
>> +u64 (*exynos4_read_count_64)(void) = __exynos4_read_count_64;
>
> I think these could be static.

Yeah. Will fix it right away.

>> +
>>  struct mct_clock_event_device {
>>       struct clock_event_device evt;
>>       unsigned long base;
>> @@ -163,16 +177,16 @@ static void exynos4_mct_frc_start(void)
>>  }
>>
>>  /**
>> - * exynos4_read_count_64 - Read all 64-bits of the global counter
>> + * __exynos4_read_count_64 - Read all 64-bits of the global counter
>>   *
>> - * This will read all 64-bits of the global counter taking care to make sure
>> - * that the upper and lower half match.  Note that reading the MCT can be quite
>> - * slow (hundreds of nanoseconds) so you should use the 32-bit (lower half
>> - * only) version when possible.
>> + * This will read all 64-bits of the global counter from MCT mmio region
>> + * taking care to make sure that the upper and lower half match.
>> + * Note that reading the MCT can be quite slow (hundreds of nanoseconds)
>> + * so you should use the 32-bit (lower half only) version when possible.
>>   *
>>   * Returns the number of cycles in the global counter.
>>   */
>> -static u64 exynos4_read_count_64(void)
>> +static u64 __exynos4_read_count_64(void)
>>  {
>>       unsigned int lo, hi;
>>       u32 hi2 = readl_relaxed(reg_base + EXYNOS4_MCT_G_CNT_U);
>> @@ -187,18 +201,45 @@ static u64 exynos4_read_count_64(void)
>>  }
>>
>>  /**
>> - * exynos4_read_count_32 - Read the lower 32-bits of the global counter
>> + * __exynos4_read_cp15_count_64 - Read all 64-bits of the global counter
>> + * from coprocessor regisers.
>>   *
>> - * This will read just the lower 32-bits of the global counter.  This is marked
>> - * as notrace so it can be used by the scheduler clock.
>> + * Note that reading here is faster than reading from MCT mmio region.
>> + *
>> + * Returns the number of cycles in the global counter.
>> + */
>> +static u64 __exynos4_read_cp15_count_64(void)
>> +{
>> +     return arch_counter_get_cntpct();
>> +}
>> +
>> +/**
>> + * __exynos4_read_count_32 - Read the lower 32-bits of the global counter
>> + *
>> + * This will read just the lower 32-bits of the global counter from
>> + * MCT mmio region.
>> + * This is marked as notrace so it can be used by the scheduler clock.
>>   *
>>   * Returns the number of cycles in the global counter (lower 32 bits).
>>   */
>> -static u32 notrace exynos4_read_count_32(void)
>> +static u32 notrace __exynos4_read_count_32(void)
>>  {
>>       return readl_relaxed(reg_base + EXYNOS4_MCT_G_CNT_L);
>>  }
>>
>> +/**
>> + * __exynos4_read_cp15_count_32 - Read the lower 32-bits of the global counter
>> + *
>> + * This will read global counter from coprocessor regisers.
>
> s/regisers/registers/

Will fix it.

>> + * This is marked as notrace so it can be used by the scheduler clock.
>> + *
>> + * Returns the number of cycles in the global counter (lower 32 bits).
>> + */
>> +static u32 notrace __exynos4_read_cp15_count_32(void)
>> +{
>> +     return arch_counter_get_cntpct();
>> +}
>> +
>>  static cycle_t exynos4_frc_read(struct clocksource *cs)
>>  {
>>       return exynos4_read_count_32();
>> @@ -599,6 +640,22 @@ static void __init mct_init_dt(struct device_node *np, unsigned int int_type)
>>       for (i = MCT_L0_IRQ; i < nr_irqs; i++)
>>               mct_irqs[i] = irq_of_parse_and_map(np, i);
>>
>> +     /*
>> +      * If use-cp15-phys-counter property is present in device tree
>> +      * then use CP15 ARM arch timer 64-bit physical counter
>> +      * to speedup reading since it keeps the same value like
>> +      * 64-bit counter in MCT mmio region.
>> +      * If HYP mode is available we can rely on physical
>> +      * timer counter to be accessible from PL1.
>> +      */
>> +     if (IS_ENABLED(CONFIG_ARM) && is_hyp_mode_available() &&
>> +             of_property_read_bool(np, "use-cp15-phys-counter")) {
>
> This looks like a property specific to Exynos. Add a "samsung," prefix.
>
> Best regards,
> Krzysztof

Agree.

Thanks for comments and review.

Best regards,
Alexey Klimov.


>> +
>> +             /* point MCT functions to read counter from coprocessor */
>> +             exynos4_read_count_32 = __exynos4_read_cp15_count_32;
>> +             exynos4_read_count_64 = __exynos4_read_cp15_count_64;
>> +     }
>> +
>>       exynos4_timer_resources(np, of_iomap(np, 0));
>>       exynos4_clocksource_init();
>>       exynos4_clockevent_init();
>>
>



-- 
Best regards, Klimov Alexey



More information about the linux-arm-kernel mailing list