Re: [RFC PATCH 17/25] kvx: Add multi-processor (SMP) support

From: Rob Herring <robh@kernel.org>
To: Yann Sionneau <ysionneau@kalray.eu>
Cc: linux-kernel@vger.kernel.org,
	Clement Leger <clement.leger@bootlin.com>,
	Julian Vetter <jvetter@kalray.eu>,
	Julien Hascoet <jhascoet@kalray.eu>,
	Louis Morhet <lmorhet@kalray.eu>, Luc Michel <lmichel@kalray.eu>,
	Marius Gligor <mgligor@kalray.eu>
Subject: Re: [RFC PATCH 17/25] kvx: Add multi-processor (SMP) support
Date: Tue, 3 Jan 2023 15:22:01 -0600	[thread overview]
Message-ID: <20230103212100.GC3942221-robh@kernel.org> (raw)
In-Reply-To: <20230103164359.24347-18-ysionneau@kalray.eu>

On Tue, Jan 03, 2023 at 05:43:51PM +0100, Yann Sionneau wrote:
> Coolidge v1 SoC has 5 clusters of 16 kvx cores
> (+1 special Resource Manager (RM) core).
> 
> Linux can run in SMP config on the 16 cores of
> a Cluster.
> 
> Cache coherency is done by the L2 cache.
> 
> CC: linux-kernel@vger.kernel.org
> Co-developed-by: Clement Leger <clement.leger@bootlin.com>
> Signed-off-by: Clement Leger <clement.leger@bootlin.com>
> Co-developed-by: Julian Vetter <jvetter@kalray.eu>
> Signed-off-by: Julian Vetter <jvetter@kalray.eu>
> Co-developed-by: Julien Hascoet <jhascoet@kalray.eu>
> Signed-off-by: Julien Hascoet <jhascoet@kalray.eu>
> Co-developed-by: Louis Morhet <lmorhet@kalray.eu>
> Signed-off-by: Louis Morhet <lmorhet@kalray.eu>
> Co-developed-by: Luc Michel <lmichel@kalray.eu>
> Signed-off-by: Luc Michel <lmichel@kalray.eu>
> Co-developed-by: Marius Gligor <mgligor@kalray.eu>
> Signed-off-by: Marius Gligor <mgligor@kalray.eu>
> Co-developed-by: Yann Sionneau <ysionneau@kalray.eu>
> Signed-off-by: Yann Sionneau <ysionneau@kalray.eu>
> ---
>  arch/kvx/include/asm/pwr_ctrl.h |  45 ++++
>  arch/kvx/include/asm/smp.h      |  42 +++
>  arch/kvx/kernel/l2_cache.c      | 448 ++++++++++++++++++++++++++++++++
>  arch/kvx/kernel/smp.c           | 110 ++++++++
>  arch/kvx/kernel/smpboot.c       | 127 +++++++++
>  arch/kvx/platform/pwr_ctrl.c    |  93 +++++++
>  include/linux/cpuhotplug.h      |   2 +
>  7 files changed, 867 insertions(+)
>  create mode 100644 arch/kvx/include/asm/pwr_ctrl.h
>  create mode 100644 arch/kvx/include/asm/smp.h
>  create mode 100644 arch/kvx/kernel/l2_cache.c
>  create mode 100644 arch/kvx/kernel/smp.c
>  create mode 100644 arch/kvx/kernel/smpboot.c
>  create mode 100644 arch/kvx/platform/pwr_ctrl.c
> 
> diff --git a/arch/kvx/include/asm/pwr_ctrl.h b/arch/kvx/include/asm/pwr_ctrl.h
> new file mode 100644
> index 000000000000..25f403ba935a
> --- /dev/null
> +++ b/arch/kvx/include/asm/pwr_ctrl.h
> @@ -0,0 +1,45 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Clement Leger
> + *            Marius Gligor
> + */
> +
> +#ifndef _ASM_KVX_PWR_CTRL_H
> +#define _ASM_KVX_PWR_CTRL_H
> +
> +#ifndef __ASSEMBLY__
> +
> +int kvx_pwr_ctrl_probe(void);
> +
> +void kvx_pwr_ctrl_cpu_poweron(unsigned int cpu);
> +
> +#endif
> +
> +/* Power controller vector register definitions */
> +#define KVX_PWR_CTRL_VEC_OFFSET 0x1000
> +#define KVX_PWR_CTRL_VEC_WUP_SET_OFFSET     0x10
> +#define KVX_PWR_CTRL_VEC_WUP_CLEAR_OFFSET     0x20
> +
> +/* Power controller PE reset PC register definitions */
> +#define KVX_PWR_CTRL_RESET_PC_OFFSET               0x2000
> +
> +/* Power controller global register definitions */
> +#define KVX_PWR_CTRL_GLOBAL_OFFSET 0x4040
> +
> +#define KVX_PWR_CTRL_GLOBAL_SET_OFFSET     0x10
> +#define KVX_PWR_CTRL_GLOBAL_SET_PE_EN_SHIFT           0x1
> +
> +#define PWR_CTRL_WUP_SET_OFFSET  \
> +		(KVX_PWR_CTRL_VEC_OFFSET + \
> +		 KVX_PWR_CTRL_VEC_WUP_SET_OFFSET)
> +
> +#define PWR_CTRL_WUP_CLEAR_OFFSET  \
> +		(KVX_PWR_CTRL_VEC_OFFSET + \
> +		 KVX_PWR_CTRL_VEC_WUP_CLEAR_OFFSET)
> +
> +#define PWR_CTRL_GLOBAL_CONFIG_OFFSET \
> +		(KVX_PWR_CTRL_GLOBAL_OFFSET + \
> +		 KVX_PWR_CTRL_GLOBAL_SET_OFFSET)
> +
> +#endif /* _ASM_KVX_PWR_CTRL_H */
> diff --git a/arch/kvx/include/asm/smp.h b/arch/kvx/include/asm/smp.h
> new file mode 100644
> index 000000000000..e4fd4d001b2c
> --- /dev/null
> +++ b/arch/kvx/include/asm/smp.h
> @@ -0,0 +1,42 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Clement Leger
> + */
> +
> +#ifndef _ASM_KVX_SMP_H
> +#define _ASM_KVX_SMP_H
> +
> +#include <linux/cpumask.h>
> +#include <linux/irqreturn.h>
> +
> +#include <asm/sfr.h>
> +
> +#ifdef CONFIG_SMP
> +
> +/* Hook for the generic smp_call_function_many() routine. */
> +void arch_send_call_function_ipi_mask(struct cpumask *mask);
> +
> +/* Hook for the generic smp_call_function_single() routine. */
> +void arch_send_call_function_single_ipi(int cpu);
> +
> +void __init setup_processor(void);
> +
> +void smp_init_cpus(void);
> +
> +irqreturn_t ipi_call_interrupt(int irq, void *dev_id);
> +
> +#define raw_smp_processor_id() ((int) \
> +	((kvx_sfr_get(PCR) & KVX_SFR_PCR_PID_MASK) \
> +					>> KVX_SFR_PCR_PID_SHIFT))
> +
> +#define flush_cache_vmap(start, end)		do { } while (0)
> +#define flush_cache_vunmap(start, end)		do { } while (0)
> +
> +#else
> +
> +void smp_init_cpus(void) {}
> +
> +#endif /* CONFIG_SMP */
> +
> +#endif
> diff --git a/arch/kvx/kernel/l2_cache.c b/arch/kvx/kernel/l2_cache.c
> new file mode 100644
> index 000000000000..94a4d9159471
> --- /dev/null
> +++ b/arch/kvx/kernel/l2_cache.c
> @@ -0,0 +1,448 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Clement Leger
> + *            Luc Michel
> + *            Julien Hascoet
> + */
> +
> +#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
> +
> +#include <linux/io.h>
> +#include <linux/of.h>
> +#include <linux/jiffies.h>
> +#include <linux/of_address.h>
> +#include <linux/irqchip/irq-kvx-apic-mailbox.h>
> +
> +#include <asm/rm_fw.h>
> +#include <asm/l2_cache.h>
> +#include <asm/sections.h>
> +#include <asm/cacheflush.h>
> +
> +#define L2_START_TIMEOUT_MS	10
> +#define L2_CMD_TIMEOUT_MS	200
> +
> +#define L2_MK_OP(__cmd, __sync) (BIT(L2_CMD_OP_VALID_SHIFT) | \
> +			 ((u64) __sync << L2_CMD_OP_SYNC_SHIFT) | \
> +			 ((u64) __cmd << L2_CMD_OP_CMD_SHIFT))
> +
> +#define L2_ERROR(_err) \
> +	((error & L2_ERROR_ERROR_CODE_MASK) >> L2_ERROR_ERROR_CODE_SHIFT)
> +
> +struct l2_cache_hw_cmd {
> +	u64 op;
> +	u64 args[L2_CMD_OP_ARG_COUNT];
> +} __packed;
> +
> +struct l2_cache_cmd {
> +	int sync;
> +	int cmd_type;
> +	unsigned int arg_count;
> +	u64 args[L2_CMD_OP_ARG_COUNT];
> +};
> +
> +/**
> + * struct l2_cached_data - Data associated to the l2-cache
> + * @regs: base of L2 registers
> + * @mbox_regs: Mailbox registers for L2 signaling
> + */
> +struct l2_cache_data {
> +	void __iomem *regs;
> +	void __iomem *mbox_regs;
> +	u64 fifo_cmd_count;
> +};
> +
> +DEFINE_STATIC_KEY_FALSE(l2_enabled);
> +static struct l2_cache_data l2c_ctrl;
> +
> +static void *l2_cmd_regs_addr(void)
> +{
> +	return l2c_ctrl.regs + L2_CMD_OFFSET;
> +}
> +
> +static struct l2_cache_hw_cmd *l2_cache_hw_cmd_addr(u64 idx)
> +{
> +	void *cmd_regs = l2_cmd_regs_addr();
> +
> +	/* Wrap index */
> +	idx &= (l2c_ctrl.fifo_cmd_count - 1);
> +	return cmd_regs + L2_CMD_FIFO_OFFSET + idx * L2_CMD_FIFO_ELEM_SIZE;
> +}
> +
> +static u64 l2_cache_get_cmd_idx(unsigned int cmd_count)
> +{
> +	u64 cmd_idx;
> +	void *cmd_regs = l2_cmd_regs_addr();
> +	u64 *write_idx_ptr = cmd_regs + L2_CMD_WRITE_IDX_OFFSET;
> +	u64 *read_idx_ptr = cmd_regs + L2_CMD_READ_IDX_OFFSET;
> +
> +	/* Grab a commands tickets */
> +	cmd_idx = __builtin_kvx_aladdd(write_idx_ptr, cmd_count);
> +
> +	/* Wait until there is room in command fifo to enqueue commands */
> +	while ((cmd_idx + cmd_count) >=
> +	       (readq(read_idx_ptr) + l2c_ctrl.fifo_cmd_count))
> +		cpu_relax();
> +
> +	return cmd_idx;
> +}
> +
> +static void l2_wait_completion(u64 cmd_idx)
> +{
> +	u64 *read_idx_ptr = l2_cmd_regs_addr() + L2_CMD_READ_IDX_OFFSET;
> +	unsigned long timeout = jiffies + msecs_to_jiffies(L2_CMD_TIMEOUT_MS);
> +
> +	/* Wait for completion */
> +	while (cmd_idx >= readq(read_idx_ptr)) {
> +		cpu_relax();
> +		if (time_after(jiffies, timeout))
> +			panic("L2 cache completion timeout\n");
> +	}
> +}
> +
> +static u64 l2_cache_push_cmds(struct l2_cache_cmd *cmds, int cmd_count)
> +{
> +	int i, arg;
> +	u64 cmd_op;
> +	struct l2_cache_hw_cmd *cmd;
> +	struct l2_cache_cmd *soft_cmd;
> +	u64 cmd_idx = l2_cache_get_cmd_idx(cmd_count);
> +
> +	for (i = 0; i < cmd_count; i++) {
> +		soft_cmd = &cmds[i];
> +		cmd = l2_cache_hw_cmd_addr(cmd_idx);
> +		cmd_idx++;
> +
> +		for (arg = 0; arg < soft_cmd->arg_count; arg++)
> +			writeq_relaxed(soft_cmd->args[arg], &cmd->args[arg]);
> +
> +		cmd_op = L2_MK_OP(soft_cmd->cmd_type, soft_cmd->sync);
> +		writeq(cmd_op, &cmd->op);
> +	}
> +
> +	return cmd_idx - 1;
> +}
> +
> +static void l2_cache_create_line_cmd(struct l2_cache_cmd *cmd, int cmd_type,
> +				     int sync, u64 addr)
> +{
> +	cmd->cmd_type = cmd_type;
> +	cmd->sync = sync;
> +	cmd->arg_count = 1;
> +	cmd->args[0] = addr;
> +}
> +
> +static void l2_cache_create_area_cmd(struct l2_cache_cmd *cmd, int cmd_type,
> +				     int sync, u64 addr, u64 size)
> +{
> +	l2_cache_create_line_cmd(cmd, cmd_type, sync, addr);
> +	cmd->arg_count = 2;
> +	cmd->args[1] = size;
> +}
> +
> +static void l2_cache_push_inval_cmd(phys_addr_t start,
> +			    unsigned long size)
> +{
> +	phys_addr_t end = start + size;
> +	struct l2_cache_cmd cmds[3];
> +	unsigned long irq_flags;
> +	int cmd_count = 0;
> +	u64 cmd_idx;
> +
> +	/*
> +	 * In case of invalidation, we must make sure we do not invalidate
> +	 * unwanted area and thus discard legit data. In case we are not aligned
> +	 * send a purge line command (writeback + inval) to unaligned lines
> +	 * (which can be the end line or the start line)
> +	 */
> +	if (!IS_ALIGNED(end, L2_CACHE_LINE_SIZE)) {
> +		end &= ~L2_CACHE_LINE_MASK;
> +		l2_cache_create_line_cmd(&cmds[cmd_count],
> +					 L2_CMD_OP_CMD_PURGE_LINE, 1, end);
> +		cmd_count++;
> +	}
> +
> +	if (!IS_ALIGNED(start, L2_CACHE_LINE_SIZE)) {
> +		start &= ~L2_CACHE_LINE_MASK;
> +		/* If there is at least another line to clear */
> +		if (end != start) {
> +			l2_cache_create_line_cmd(&cmds[cmd_count],
> +						 L2_CMD_OP_CMD_PURGE_LINE, 1,
> +						 start);
> +			cmd_count++;
> +			start += L2_CACHE_LINE_SIZE;
> +		}
> +	}
> +
> +	BUG_ON(end < start);
> +
> +	size = (end - start);
> +	if (size > 0) {
> +		l2_cache_create_area_cmd(&cmds[cmd_count],
> +					 L2_CMD_OP_CMD_INVAL_AREA, 1, start,
> +					 size);
> +		cmd_count++;
> +	}
> +
> +	BUG_ON(cmd_count == 0);
> +
> +	local_irq_save(irq_flags);
> +
> +	cmd_idx = l2_cache_push_cmds(cmds, cmd_count);
> +
> +	/* Finally, ping the L2 cache controller */
> +	writeq(1, l2c_ctrl.mbox_regs);
> +
> +	local_irq_restore(irq_flags);
> +
> +	l2_wait_completion(cmd_idx);
> +}
> +
> +static void l2_cache_push_generic_cmd(u64 cmd_type, phys_addr_t start,
> +			    unsigned long size)
> +{
> +	unsigned long irq_flags;
> +	struct l2_cache_cmd cmd;
> +	u64 cmd_idx;
> +
> +	/* Align the start address and size on cache line */
> +	size += start & (L2_CACHE_LINE_SIZE - 1);
> +	size = ALIGN(size, L2_CACHE_LINE_SIZE);
> +	start = ALIGN_DOWN(start, L2_CACHE_LINE_SIZE);
> +
> +	local_irq_save(irq_flags);
> +
> +	l2_cache_create_area_cmd(&cmd, cmd_type, 1, start, size);
> +	cmd_idx = l2_cache_push_cmds(&cmd, 1);
> +
> +	/* Finally, ping the L2 cache controller */
> +	writeq(1, l2c_ctrl.mbox_regs);
> +
> +	local_irq_restore(irq_flags);
> +
> +	l2_wait_completion(cmd_idx);
> +}
> +
> +void l2_cache_push_area_cmd(u64 cmd_type, phys_addr_t start,
> +			    unsigned long size)
> +{
> +	if (WARN_ON(size == 0))
> +		return;
> +
> +	if (cmd_type == L2_CMD_OP_CMD_INVAL_AREA)
> +		l2_cache_push_inval_cmd(start, size);
> +	else
> +		l2_cache_push_generic_cmd(cmd_type, start, size);
> +}
> +
> +static void __init l2_disp_error(u64 error)
> +{
> +	const char *err_type;
> +
> +	if (error & L2_ERROR_API_ERR_MASK)
> +		err_type = "API";
> +	else if (error & L2_ERROR_SETUP_ERR_MASK)
> +		err_type = "SETUP";
> +	else
> +		err_type = "UNKNOWN";
> +
> +	pr_err("%s error: 0x%llx\n", err_type, L2_ERROR(error));
> +}
> +
> +static int __init l2_cache_configure_mailboxes(void)
> +{
> +	phys_addr_t l2_mbox_addr = 0;
> +	void *cmd_regs = l2_cmd_regs_addr();
> +
> +	/* We do not use mailbox to wait for completion, set it to 0 */
> +	writeq(0, cmd_regs + L2_CMD_DOORBELL_READ_ADDR_OFFSET);
> +
> +	/* Read mailbox address from L2 registers */
> +	l2_mbox_addr = readq(cmd_regs + L2_CMD_DOORBELL_WRITE_ADDR_OFFSET);
> +
> +	/* Then map the mailbox */
> +	l2c_ctrl.mbox_regs = ioremap(l2_mbox_addr, PAGE_SIZE);
> +	if (!l2c_ctrl.mbox_regs) {
> +		pr_err("Failed to map mailbox\n");
> +		return 1;
> +	}
> +
> +	/* Lock this entry into the LTLB */
> +	kvx_mmu_ltlb_add_entry((unsigned long) l2c_ctrl.mbox_regs & PAGE_MASK,
> +			       l2_mbox_addr & PAGE_MASK,
> +			       PAGE_KERNEL_DEVICE, TLB_PS_4K);
> +
> +	return 0;
> +}
> +
> +static int __init l2_cache_read_queue_size(void)
> +{
> +	u64 inst;
> +
> +	/* Read command queue size */
> +	inst = readq(l2c_ctrl.regs + L2_INSTANCE_OFFSET);
> +	l2c_ctrl.fifo_cmd_count = (inst & L2_INSTANCE_CMD_QUEUE_SIZE_MASK)
> +				   >> L2_INSTANCE_CMD_QUEUE_SIZE_SHIFT;
> +
> +	/* Check if value is a power of two */
> +	if (hweight64(l2c_ctrl.fifo_cmd_count) != 1) {
> +		pr_err("Command queue size is not a power of two\n");
> +		return 1;
> +	}
> +
> +	return 0;
> +}
> +
> +static int __init l2_cache_init_hw(void)
> +{
> +	unsigned long timeout = jiffies + msecs_to_jiffies(L2_START_TIMEOUT_MS);
> +	unsigned long flags;
> +	u64 status, error;
> +	int ret;
> +
> +	/* Wait for L2 to be up */
> +	do {
> +		status = readq(l2c_ctrl.regs + L2_STATUS_OFFSET);
> +		if (status & (L2_STATUS_READY_MASK | L2_STATUS_ERROR_MASK))
> +			break;
> +	} while (time_before(jiffies, timeout));
> +
> +	if (!status) {
> +		pr_err("Timeout while waiting for firmware status\n");
> +		return -ENODEV;
> +	}
> +
> +	if (status & L2_STATUS_ERROR_MASK) {
> +		error = readq(l2c_ctrl.regs + L2_ERROR_OFFSET);
> +		l2_disp_error(error);
> +		return -EINVAL;
> +	}
> +
> +	/* Now write ack to L2 firmware */
> +	writeq(status | L2_STATUS_ACK_MASK, l2c_ctrl.regs + L2_STATUS_OFFSET);
> +
> +	ret = l2_cache_read_queue_size();
> +	if (ret)
> +		return ret;
> +
> +	ret = l2_cache_configure_mailboxes();
> +	if (ret)
> +		return ret;
> +
> +	/* Enable the L2 atomically */
> +	local_irq_save(flags);
> +
> +	/* Fence data accesses */
> +	kvx_fence();
> +	/* Purge L1 */
> +	l1_inval_dcache_all();
> +	l1_inval_icache_all();
> +	__builtin_kvx_barrier();
> +
> +	local_irq_restore(flags);
> +
> +	/* Enable L2$ */
> +	kvx_sfr_set_field(PS, L2E, 1);
> +
> +	return 0;
> +}
> +
> +static phys_addr_t __init l2_get_regs_addr(struct device_node *np)
> +{
> +	const void *reg;
> +	struct resource res;
> +	phys_addr_t l2_regs_addr;
> +	int ret;
> +
> +	/*
> +	 * If regs is specified in device tree, then the L2$ has been loaded by
> +	 * someone else and not by ourself.
> +	 */
> +	reg = of_get_property(np, "reg", NULL);

Why do you read 'reg' and then...

> +	if (reg) {
> +		ret = of_address_to_resource(np, 0, &res);

...read it again? 

> +		if (ret) {
> +			pr_err("Address translation error\n");
> +			return 0;
> +		}
> +		if ((res.end - res.start) > PAGE_SIZE) {
> +			pr_err("L2 reg size > PAGE_SIZE\n");

Why do you care?

> +			return 0;
> +		}
> +
> +		l2_regs_addr = res.start;
> +	} else {
> +		l2_regs_addr = (phys_addr_t) __rm_firmware_regs_start;
> +	}
> +
> +	if (!IS_ALIGNED(l2_regs_addr, PAGE_SIZE)) {
> +		pr_err("Registers not aligned on PAGE_SIZE\n");

So?

If you have constraints, put those in the DT schema for this device and 
validate your DT. The kernel doesn't need to validate the DT.

> +		return 0;
> +	}
> +
> +	return l2_regs_addr;
> +}
> +
> +static int __init l2_cache_init(void)
> +{
> +	int ret = -ENODEV;
> +	struct device_node *np;
> +	phys_addr_t l2_regs_addr;
> +
> +	np = of_find_compatible_node(NULL, NULL, "kalray,kvx-l2-cache");

Needs to be documented.

> +	if (!np || !of_device_is_available(np)) {
> +		if (!IS_ENABLED(CONFIG_SMP)) {
> +			pr_info("controller disabled\n");
> +			return 0;
> +		}
> +
> +		if (np && of_get_property(np, "kalray,is-qemu", NULL)) {

Make your compatible strings specific enough to detect this rather than 
a flag.

> +			/*
> +			 * QEMU is always full cache coherent. The L2 cache controller is
> +			 * not strictly necessary to ensure coherency in SMP.
> +			 */
> +			pr_info("controller disabled (QEMU detected)\n");
> +			return 0;
> +		}
> +
> +		/* Else, SMP is enabled and L2 is mandatory for it */
> +		goto err;
> +	}
> +
> +	l2_regs_addr = l2_get_regs_addr(np);
> +	if (!l2_regs_addr)
> +		goto err;
> +
> +	/* Map the L2 registers */
> +	l2c_ctrl.regs = ioremap(l2_regs_addr, PAGE_SIZE);
> +	if (!l2c_ctrl.regs)
> +		goto err;
> +
> +	/* Lock this entry into the LTLB */
> +	kvx_mmu_ltlb_add_entry((unsigned long) l2c_ctrl.regs, l2_regs_addr,
> +			       PAGE_KERNEL_NOCACHE, TLB_PS_4K);
> +
> +	ret = l2_cache_init_hw();
> +	if (ret) {
> +		pr_err("Failed to init L2 cache controller");
> +		goto err_unmap_l2;
> +	}
> +
> +	static_branch_enable(&l2_enabled);
> +
> +	pr_info("controller enabled\n");
> +
> +	return 0;
> +
> +err_unmap_l2:
> +	kvx_mmu_ltlb_remove_entry((unsigned long) l2c_ctrl.regs);
> +	iounmap(l2c_ctrl.regs);
> +err:
> +	if (IS_ENABLED(CONFIG_SMP))
> +		panic("L2$ controller is mandatory for SMP");
> +
> +	return ret;
> +}
> +
> +
> +early_initcall(l2_cache_init);
> diff --git a/arch/kvx/kernel/smp.c b/arch/kvx/kernel/smp.c
> new file mode 100644
> index 000000000000..ed4c35a8c4bc
> --- /dev/null
> +++ b/arch/kvx/kernel/smp.c
> @@ -0,0 +1,110 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Clement Leger
> + */
> +
> +#include <linux/smp.h>
> +#include <linux/cpu.h>
> +#include <linux/of_irq.h>
> +#include <linux/cpumask.h>
> +#include <linux/irq_work.h>
> +#include <linux/mm_types.h>
> +#include <linux/interrupt.h>
> +
> +#include <asm/ipi.h>
> +#include <asm/tlbflush.h>
> +
> +enum ipi_message_type {
> +	IPI_RESCHEDULE,
> +	IPI_CALL_FUNC,
> +	IPI_IRQ_WORK,
> +	IPI_MAX
> +};
> +
> +/* A collection of single bit ipi messages.  */
> +static struct {
> +	unsigned long bits ____cacheline_aligned;
> +} ipi_data[NR_CPUS] __cacheline_aligned;
> +
> +static void send_ipi_message(const struct cpumask *mask,
> +			     enum ipi_message_type operation)
> +{
> +	unsigned long flags;
> +	int cpu;
> +
> +	/* Set operation that must be done by receiver */
> +	for_each_cpu(cpu, mask)
> +		set_bit(operation, &ipi_data[cpu].bits);
> +
> +	/* Commit the write before sending IPI */
> +	smp_wmb();
> +
> +	local_irq_save(flags);
> +
> +	kvx_ipi_send(mask);
> +
> +	local_irq_restore(flags);
> +}
> +
> +void arch_send_call_function_ipi_mask(struct cpumask *mask)
> +{
> +	send_ipi_message(mask, IPI_CALL_FUNC);
> +}
> +
> +void arch_send_call_function_single_ipi(int cpu)
> +{
> +	send_ipi_message(cpumask_of(cpu), IPI_CALL_FUNC);
> +}
> +
> +#ifdef CONFIG_IRQ_WORK
> +void arch_irq_work_raise(void)
> +{
> +	send_ipi_message(cpumask_of(smp_processor_id()), IPI_IRQ_WORK);
> +}
> +#endif
> +
> +static void ipi_stop(void *unused)
> +{
> +	local_cpu_stop();
> +}
> +
> +void smp_send_stop(void)
> +{
> +	struct cpumask targets;
> +
> +	cpumask_copy(&targets, cpu_online_mask);
> +	cpumask_clear_cpu(smp_processor_id(), &targets);
> +
> +	smp_call_function_many(&targets, ipi_stop, NULL, 0);
> +}
> +
> +void smp_send_reschedule(int cpu)
> +{
> +	send_ipi_message(cpumask_of(cpu), IPI_RESCHEDULE);
> +}
> +
> +irqreturn_t ipi_call_interrupt(int irq, void *dev_id)
> +{
> +	unsigned long *pending_ipis = &ipi_data[smp_processor_id()].bits;
> +
> +	while (true) {
> +		unsigned long ops = xchg(pending_ipis, 0);
> +
> +		if (ops == 0)
> +			return IRQ_HANDLED;
> +
> +		if (ops & (1 << IPI_RESCHEDULE))
> +			scheduler_ipi();
> +
> +		if (ops & (1 << IPI_CALL_FUNC))
> +			generic_smp_call_function_interrupt();
> +
> +		if (ops & (1 << IPI_IRQ_WORK))
> +			irq_work_run();
> +
> +		BUG_ON((ops >> IPI_MAX) != 0);
> +	}
> +
> +	return IRQ_HANDLED;
> +}
> diff --git a/arch/kvx/kernel/smpboot.c b/arch/kvx/kernel/smpboot.c
> new file mode 100644
> index 000000000000..987a6f014163
> --- /dev/null
> +++ b/arch/kvx/kernel/smpboot.c
> @@ -0,0 +1,127 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Clement Leger
> + *            Julian Vetter
> + */
> +
> +#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
> +
> +#include <linux/io.h>
> +#include <linux/of.h>
> +#include <linux/smp.h>
> +#include <linux/cpu.h>
> +#include <linux/sched.h>
> +#include <linux/cpumask.h>
> +#include <linux/sched/mm.h>
> +#include <linux/mm_types.h>
> +#include <linux/of_platform.h>

Don't think this is needed.

> +#include <linux/sched/task_stack.h>
> +
> +#include <asm/pwr_ctrl.h>
> +#include <asm/tlbflush.h>
> +#include <asm/ipi.h>
> +
> +void *__cpu_up_stack_pointer[NR_CPUS];
> +void *__cpu_up_task_pointer[NR_CPUS];
> +
> +void __init smp_prepare_boot_cpu(void)
> +{
> +}
> +
> +int __cpu_up(unsigned int cpu, struct task_struct *tidle)
> +{
> +	__cpu_up_stack_pointer[cpu] = task_stack_page(tidle) + THREAD_SIZE;
> +	__cpu_up_task_pointer[cpu] = tidle;
> +	/* We need to be sure writes are committed */
> +	smp_mb();
> +
> +	kvx_pwr_ctrl_cpu_poweron(cpu);
> +	while (!cpu_online(cpu))
> +		cpu_relax();
> +
> +	return 0;
> +}
> +
> +void __init smp_cpus_done(unsigned int max_cpus)
> +{
> +}
> +
> +void __init smp_init_cpus(void)
> +{
> +	struct cpumask cpumask;
> +	struct device_node *cpu;
> +	const __be32 *reg;
> +	u32 cpu_num;
> +	unsigned int nr_cpus = 0;
> +
> +	cpumask_clear(&cpumask);
> +
> +	for_each_of_cpu_node(cpu) {
> +		if (!of_device_is_available(cpu))
> +			continue;
> +
> +		reg = of_get_property(cpu, "reg", NULL);

Use of_get_cpu_hwid()

> +		if (!reg)
> +			continue;
> +
> +		cpu_num = be32_to_cpup(reg);
> +		if (cpu_num >= nr_cpu_ids)
> +			continue;
> +
> +		nr_cpus++;
> +		cpumask_set_cpu(cpu_num, &cpumask);

Why not just do set_cpu_possible() directly instead of the temp cpumask.

> +	}
> +
> +	pr_info("%d possible cpus\n", nr_cpus);
> +	init_cpu_possible(&cpumask);
> +}
> +
> +void __init smp_prepare_cpus(unsigned int max_cpus)
> +{
> +	if (num_present_cpus() <= 1)
> +		init_cpu_present(cpu_possible_mask);
> +}
> +
> +int __init setup_smp(void)
> +{
> +	int ret;
> +
> +	ret = kvx_pwr_ctrl_probe();
> +	if (ret)
> +		panic("Failed to probe power controller !");
> +
> +	ret = kvx_ipi_ctrl_probe(ipi_call_interrupt);
> +	if (ret)
> +		panic("Failed to probe IPI controller !");
> +
> +	return 0;
> +}
> +
> +early_initcall(setup_smp);
> +
> +/*
> + * C entry point for a secondary processor.
> + */
> +void __init start_kernel_secondary(void)
> +{
> +	struct mm_struct *mm = &init_mm;
> +	unsigned int cpu = smp_processor_id();
> +
> +	setup_processor();
> +	kvx_mmu_early_setup();
> +
> +	/* All kernel threads share the same mm context.  */
> +	mmgrab(mm);
> +	current->active_mm = mm;
> +	cpumask_set_cpu(cpu, mm_cpumask(mm));
> +
> +	notify_cpu_starting(cpu);
> +	set_cpu_online(cpu, true);
> +	trace_hardirqs_off();
> +
> +	local_flush_tlb_all();
> +
> +	local_irq_enable();
> +	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
> +}
> diff --git a/arch/kvx/platform/pwr_ctrl.c b/arch/kvx/platform/pwr_ctrl.c
> new file mode 100644
> index 000000000000..64c86cd18695
> --- /dev/null
> +++ b/arch/kvx/platform/pwr_ctrl.c
> @@ -0,0 +1,93 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/*
> + * Copyright (C) 2017-2023 Kalray Inc.
> + * Author(s): Clement Leger
> + */
> +
> +#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
> +
> +#include <linux/io.h>
> +#include <linux/slab.h>
> +#include <linux/types.h>
> +#include <linux/module.h>
> +#include <linux/of_address.h>
> +#include <linux/of_platform.h>
> +
> +#include <asm/pwr_ctrl.h>
> +#include <asm/symbols.h>
> +
> +struct kvx_pwr_ctrl {
> +	void __iomem *regs;
> +};
> +
> +static struct kvx_pwr_ctrl kvx_pwr_controller;
> +
> +/**
> + * kvx_pwr_ctrl_cpu_poweron() - Wakeup a cpu
> + * @cpu: cpu to wakeup
> + */
> +void kvx_pwr_ctrl_cpu_poweron(unsigned int cpu)
> +{
> +	/* Set PE boot address */
> +	writeq((unsigned long long)kvx_start,
> +			kvx_pwr_controller.regs + KVX_PWR_CTRL_RESET_PC_OFFSET);
> +	/* Wake up processor ! */
> +	writeq(1ULL << cpu,
> +	       kvx_pwr_controller.regs + PWR_CTRL_WUP_SET_OFFSET);
> +	/* Then clear wakeup to allow processor to sleep */
> +	writeq(1ULL << cpu,
> +	       kvx_pwr_controller.regs + PWR_CTRL_WUP_CLEAR_OFFSET);
> +}
> +
> +static struct device_node * __init get_pwr_ctrl_node(void)
> +{
> +	const phandle *ph;
> +	struct device_node *cpu;
> +	struct device_node *node;
> +
> +	cpu = of_get_cpu_node(raw_smp_processor_id(), NULL);
> +	if (!cpu) {
> +		pr_err("Failed to get CPU node\n");
> +		return NULL;
> +	}
> +
> +	ph = of_get_property(cpu, "power-controller", NULL);

Not a standard property.

> +	if (!ph) {
> +		pr_err("Failed to get power-controller phandle\n");
> +		return NULL;
> +	}
> +
> +	node = of_find_node_by_phandle(be32_to_cpup(ph));
> +	if (!node) {
> +		pr_err("Failed to get power-controller node\n");
> +		return NULL;
> +	}
> +
> +	return node;
> +}
> +
> +int __init kvx_pwr_ctrl_probe(void)
> +{
> +	struct device_node *ctrl;
> +
> +	ctrl = get_pwr_ctrl_node();
> +	if (!ctrl) {
> +		pr_err("Failed to get power controller node\n");
> +		return -EINVAL;
> +	}
> +
> +	if (!of_device_is_compatible(ctrl, "kalray,kvx-pwr-ctrl")) {

Needs to be documented...

> +		pr_err("Failed to get power controller node\n");
> +		return -EINVAL;
> +	}
> +
> +	kvx_pwr_controller.regs = of_iomap(ctrl, 0);
> +	if (!kvx_pwr_controller.regs) {
> +		pr_err("Failed ioremap\n");
> +		return -EINVAL;
> +	}
> +
> +	pr_info("kvx power controller probed\n");

Drop success messages.

> +
> +	return 0;
> +}
> diff --git a/include/linux/cpuhotplug.h b/include/linux/cpuhotplug.h
> index f61447913db9..f5a484547b15 100644
> --- a/include/linux/cpuhotplug.h
> +++ b/include/linux/cpuhotplug.h
> @@ -152,6 +152,7 @@ enum cpuhp_state {
>  	CPUHP_AP_IRQ_RISCV_STARTING,
>  	CPUHP_AP_IRQ_LOONGARCH_STARTING,
>  	CPUHP_AP_IRQ_SIFIVE_PLIC_STARTING,
> +	CPUHP_AP_IRQ_KVX_STARTING,
>  	CPUHP_AP_ARM_MVEBU_COHERENCY,
>  	CPUHP_AP_MICROCODE_LOADER,
>  	CPUHP_AP_PERF_X86_AMD_UNCORE_STARTING,
> @@ -189,6 +190,7 @@ enum cpuhp_state {
>  	CPUHP_AP_KVM_ARM_VGIC_INIT_STARTING,
>  	CPUHP_AP_KVM_ARM_VGIC_STARTING,
>  	CPUHP_AP_KVM_ARM_TIMER_STARTING,
> +	CPUHP_AP_KVX_TIMER_STARTING,
>  	/* Must be the last timer callback */
>  	CPUHP_AP_DUMMY_TIMER_STARTING,
>  	CPUHP_AP_ARM_XEN_STARTING,
> -- 
> 2.37.2
> 
> 
> 
> 
> 
> 