Re: [PATCH v3 3/4] ARM: EXYNOS: add Exynos Dual Cluster Support

[Dave Martin <Dave.Martin@arm.com>]

On Thu, Nov 07, 2013 at 08:12:48AM +0000, Vyacheslav Tyrtov wrote:
> From: Tarek Dakhran <t.dakhran@samsung.com>
> 
> Add EDCS(Exynos Dual Cluster Support) for Samsung Exynos5410 SoC.
> This enables all 8 cores, 4 x A7 and 4 x A15 run at the same time.
> 
> Signed-off-by: Tarek Dakhran <t.dakhran@samsung.com>
> Signed-off-by: Vyacheslav Tyrtov <v.tyrtov@samsung.com>
> ---
>  arch/arm/mach-exynos/Makefile |   2 +
>  arch/arm/mach-exynos/edcs.c   | 278 ++++++++++++++++++++++++++++++++++++++++++
>  2 files changed, 280 insertions(+)
>  create mode 100644 arch/arm/mach-exynos/edcs.c
> 
> diff --git a/arch/arm/mach-exynos/Makefile b/arch/arm/mach-exynos/Makefile
> index 5369615..ba6efdb 100644
> --- a/arch/arm/mach-exynos/Makefile
> +++ b/arch/arm/mach-exynos/Makefile
> @@ -34,3 +34,5 @@ AFLAGS_exynos-smc.o		:=-Wa,-march=armv7-a$(plus_sec)
>  
>  obj-$(CONFIG_MACH_EXYNOS4_DT)		+= mach-exynos4-dt.o
>  obj-$(CONFIG_MACH_EXYNOS5_DT)		+= mach-exynos5-dt.o
> +
> +obj-$(CONFIG_SOC_EXYNOS5410)		+= edcs.o
> diff --git a/arch/arm/mach-exynos/edcs.c b/arch/arm/mach-exynos/edcs.c
> new file mode 100644
> index 0000000..980bfdd
> --- /dev/null
> +++ b/arch/arm/mach-exynos/edcs.c
> @@ -0,0 +1,278 @@
> +/*
> + * arch/arm/mach-exynos/edcs.c - exynos dual cluster power management support
> + *
> + * Copyright (c) 2013 Samsung Electronics Co., Ltd.
> + * Author: Tarek Dakhran <t.dakhran@samsung.com>
> + *
> + * This program is free software; you can redistribute it and/or modify
> + * it under the terms of the GNU General Public License version 2 as
> + * published by the Free Software Foundation.
> + *
> + * EDCS(exynos dual cluster support) for Exynos5410 SoC.
> + */
> +
> +#include <linux/init.h>
> +#include <linux/io.h>
> +#include <linux/of_address.h>
> +#include <linux/spinlock.h>
> +#include <linux/errno.h>
> +#include <linux/irqchip/arm-gic.h>
> +
> +#include <asm/mcpm.h>
> +#include <asm/proc-fns.h>
> +#include <asm/cacheflush.h>
> +#include <asm/cputype.h>
> +#include <asm/cp15.h>
> +
> +#include <linux/arm-cci.h>
> +#include <mach/regs-pmu.h>
> +
> +#define EDCS_CPUS_PER_CLUSTER	4
> +#define EDCS_CLUSTERS		2
> +
> +/* Exynos5410 power management registers */
> +#define EDCS_CORE_CONFIGURATION(_nr)	(S5P_ARM_CORE0_CONFIGURATION	\
> +						+ ((_nr) * 0x80))
> +#define EDCS_CORE_STATUS(_nr)		(EDCS_CORE_CONFIGURATION(_nr) + 0x4)
> +#define EDCS_CORE_OPTION(_nr)		(EDCS_CORE_CONFIGURATION(_nr) + 0x8)
> +
> +#define REG_CPU_STATE_ADDR0		(S5P_VA_SYSRAM_NS + 0x28)
> +#define REG_CPU_STATE_ADDR(_nr)		(REG_CPU_STATE_ADDR0 +	\
> +						 (_nr) * EDCS_CPUS_PER_CLUSTER)
> +
> +#define SECONDARY_RESET		(1 << 1)
> +#define REG_ENTRY_ADDR		(S5P_VA_SYSRAM_NS + 0x1c)
> +
> +static arch_spinlock_t edcs_lock = __ARCH_SPIN_LOCK_UNLOCKED;
> +
> +static int edcs_use_count[EDCS_CPUS_PER_CLUSTER][EDCS_CLUSTERS];
> +static int core_count[EDCS_CLUSTERS];
> +
> +static void exynos_core_power_control(unsigned int cpu, unsigned int cluster,
> +				bool enable)
> +{
> +	unsigned int offset = cluster * MAX_CPUS_PER_CLUSTER + cpu;
> +	int value = enable ? S5P_CORE_LOCAL_PWR_EN : 0;
> +
> +	if ((readl_relaxed(EDCS_CORE_STATUS(offset)) & 0x3) != value) {

I wonder if there is a race here.

If there is a pending powerdown which has reached the __mcpm_cpu_down()
stage, then the kernel has no way to know what is still pending.  This
means that when calling exynos_power_up(cpu, cluster) after a successful
call to exynos_power_down(same cpu, cluster), there is a chance that
the CPU still gets powered down, because of the pending
exynos_core_power_control() on the outbound side.

This isn't an issue for TC2, because TC2's power controller queues
requests and services them in order, so a new powerup request cannot
race with a powerdown request in that way.

For exynos5410, it looks like the kernel needs to do that sequencing,
based on my guess about what the EDCS_CORE_STATUS() bits tell us.


I think that for correct behaviour we would need to wait for the race to
be resolved here, but only if a powerdown might be pending.

This implies that something like a call to the power_down_finish()
method (which you would need to write -- see my comments below) is
needed in exynos_core_power_up().


It might make sense to have a per-cpu flag that tracks whether a
powerdown is pending.  The flag could be set after
__mcpm_cpu_going_down() is called, and cleared in the powered_up()
method (which you would need to add).


Maybe we should always just poll and wait, though.  exynos_power_up()
should never be called for a CPU that the kernel thinks is already up,
so it should either be down already (in which case we will poll the
status once and then continue), or a power down is pending (in which
case we must wait, but we know the wait will terminate).  This would
be simpler than tracking a "power down pending" flag for each CPU.

> +		wmb();
> +		writel_relaxed(value, EDCS_CORE_CONFIGURATION(offset));
> +	}
> +}
> +
> +static void exynos_core_power_up(unsigned int cpu, unsigned int cluster)
> +{
> +	exynos_core_power_control(cpu, cluster, true);
> +}
> +
> +static void exynos_core_power_down(unsigned int cpu, unsigned int cluster)
> +{
> +	exynos_core_power_control(cpu, cluster, false);
> +}
> +
> +void set_boot_flag(unsigned int cpu, unsigned int mode)
> +{
> +	writel_relaxed(mode, REG_CPU_STATE_ADDR(cpu));
> +}
> +
> +static int exynos_power_up(unsigned int cpu, unsigned int cluster)
> +{
> +	pr_debug("%s: cpu %u cluster %u\n", __func__, cpu, cluster);
> +	BUG_ON(cpu >= EDCS_CPUS_PER_CLUSTER || cluster >= EDCS_CLUSTERS);
> +
> +	local_irq_disable();
> +	arch_spin_lock(&edcs_lock);
> +
> +	edcs_use_count[cpu][cluster]++;
> +	if (edcs_use_count[cpu][cluster] == 1) {
> +		++core_count[cluster];
> +		set_boot_flag(cpu, SECONDARY_RESET);
> +		exynos_core_power_up(cpu, cluster);
> +	} else if (edcs_use_count[cpu][cluster] != 2) {
> +		/*
> +		 * The only possible values are:
> +		 * 0 = CPU down
> +		 * 1 = CPU (still) up
> +		 * 2 = CPU requested to be up before it had a chance
> +		 *     to actually make itself down.
> +		 * Any other value is a bug.
> +		 */
> +		BUG();
> +	}
> +
> +	arch_spin_unlock(&edcs_lock);
> +	local_irq_enable();
> +
> +	return 0;
> +}
> +static void exynos_power_down(void)
> +{
> +	unsigned int mpidr, cpu, cluster;
> +	bool last_man = false, skip_wfi = false;
> +
> +	mpidr = read_cpuid_mpidr();
> +	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
> +	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
> +
> +	pr_debug("%s: CORE%d on CLUSTER %d\n", __func__, cpu, cluster);
> +	BUG_ON(cpu >= EDCS_CPUS_PER_CLUSTER  || cluster >= EDCS_CLUSTERS);
> +
> +	__mcpm_cpu_going_down(cpu, cluster);
> +
> +	arch_spin_lock(&edcs_lock);
> +	BUG_ON(__mcpm_cluster_state(cluster) != CLUSTER_UP);
> +	edcs_use_count[cpu][cluster]--;
> +	if (edcs_use_count[cpu][cluster] == 0) {
> +		--core_count[cluster];
> +		if (core_count[cluster] == 0)
> +			last_man = true;
> +	} else if (edcs_use_count[cpu][cluster] == 1) {
> +		/*
> +		 * A power_up request went ahead of us.
> +		 * Even if we do not want to shut this CPU down,
> +		 * the caller expects a certain state as if the WFI
> +		 * was aborted.  So let's continue with cache cleaning.
> +		 */
> +		skip_wfi = true;
> +	} else
> +		BUG();
> +
> +	if (!skip_wfi)
> +		gic_cpu_if_down();
> +
> +	if (last_man && __mcpm_outbound_enter_critical(cpu, cluster)) {
> +		arch_spin_unlock(&edcs_lock);
> +
> +		if (read_cpuid_part_number() == ARM_CPU_PART_CORTEX_A15) {
> +			/*
> +			 * On the Cortex-A15 we need to disable
> +			 * L2 prefetching before flushing the cache.
> +			 */
> +			asm volatile(
> +			"mcr	p15, 1, %0, c15, c0, 3\n\t"
> +			"isb\n\t"
> +			"dsb"
> +			: : "r" (0x400));
> +		}
> +
> +		/*
> +		 * We need to disable and flush the whole (L1 and L2) cache.
> +		 * Let's do it in the safest possible way i.e. with
> +		 * no memory access within the following sequence
> +		 * including the stack.
> +		 *
> +		 * Note: fp is preserved to the stack explicitly prior doing
> +		 * this since adding it to the clobber list is incompatible
> +		 * with having CONFIG_FRAME_POINTER=y.
> +		 */
> +		asm volatile(
> +		"str	fp, [sp, #-4]!\n\t"
> +		"mrc	p15, 0, r0, c1, c0, 0	@ get CR\n\t"
> +		"bic	r0, r0, #"__stringify(CR_C)"\n\t"
> +		"mcr	p15, 0, r0, c1, c0, 0	@ set CR\n\t"
> +		"isb\n\t"
> +		"bl	v7_flush_dcache_all\n\t"
> +		"clrex\n\t"
> +		"mrc	p15, 0, r0, c1, c0, 1	@ get AUXCR\n\t"
> +		"bic	r0, r0, #(1 << 6)	@ disable local coherency\n\t"
> +		"mcr	p15, 0, r0, c1, c0, 1	@ set AUXCR\n\t"
> +		"isb\n\t"
> +		"dsb\n\t"
> +		"ldr	fp, [sp], #4"

The v7_exit_coherency_flush() macro is now in linux-next, so
you can now use it to replace these sequences.

This can be replaced by v7_exit_coherency_flush(all).

> +		: : : "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7",
> +			"r9", "r10", "lr", "memory");
> +
> +		cci_disable_port_by_cpu(mpidr);
> +
> +		__mcpm_outbound_leave_critical(cluster, CLUSTER_DOWN);
> +
> +	} else {
> +		arch_spin_unlock(&edcs_lock);
> +		/*
> +			* We need to disable and flush only the L1 cache.
> +			* Let's do it in the safest possible way as above.
> +		*/
> +		asm volatile(
> +		"str	fp, [sp, #-4]!\n\t"
> +		"mrc	p15, 0, r0, c1, c0, 0	@ get CR\n\t"
> +		"bic	r0, r0, #"__stringify(CR_C)"\n\t"
> +		"mcr	p15, 0, r0, c1, c0, 0	@ set CR\n\t"
> +		"isb\n\t"
> +		"bl	v7_flush_dcache_louis\n\t"
> +		"clrex\n\t"
> +		"mrc	p15, 0, r0, c1, c0, 1	@ get AUXCR\n\t"
> +		"bic	r0, r0, #(1 << 6)	@ disable local coherency\n\t"
> +		"mcr	p15, 0, r0, c1, c0, 1	@ set AUXCR\n\t"
> +		"isb\n\t"
> +		"dsb\n\t"
> +		"ldr	fp, [sp], #4"

v7_exit_coherency_flush(louis) should work here.

arch/arm/mach-vexpress/tc2_pm.c (in linux-next) shows how to use it.

> +		: : : "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7",
> +		      "r9", "r10", "lr", "memory");
> +
> +	}
> +	__mcpm_cpu_down(cpu, cluster);
> +
> +	if (!skip_wfi) {
> +		exynos_core_power_down(cpu, cluster);
> +		wfi();
> +	}
> +}
> +
> +static const struct mcpm_platform_ops exynos_power_ops = {
> +	.power_up	= exynos_power_up,
> +	.power_down	= exynos_power_down,
> +};

The new mcpm_power_down_finish() call is also present in linux-next now,
so it should get merged into v3.13.

One effect of this is that you should provide a power_down_finish()
method in your mcpm_platform_ops, to provide the kernel with a way to
check that a CPU has finished powering down.  This would usually involve
checking some status bits in the power controller.  See the comments for
mcpm_power_down_finish() in arch/arm/include/asm/mcpm.h for details.

No platform backend for power_down_finish() is merged yet.  The most
recent patch for TC2 was posted here -- I need to follow up on it.

http://lists.infradead.org/pipermail/linux-arm-kernel/2013-October/201619.html
[PATCH v3 3/3] ARM: vexpress/TC2: Implement MCPM power_down_finish()

This may look quite different for exynos5410.

Cheers
---Dave