arm64: add devicetree based CPU startup

This adds SMP bring up support for arm64 cpus. It's
reliant on DEVICE_PATH_CPU devices in the devicetree.
Then for each enabled device it attempts to start then
initialize each CPU.  Additionally, there is a cpu_action
construct which allows for running actions on an individual
cpu.

BUG=chrome-os-partner:31761
BRANCH=None
TEST=Booted both cores on ryu into linux.

Change-Id: I3e42fb668034c27808d706427a26be1558ad2af1
Signed-off-by: Patrick Georgi <pgeorgi@chromium.org>
Original-Commit-Id: a733fd566a8e5793da5ff28f9c16c213f411372e
Original-Change-Id: I407eabd0b6985fc4e86de57a9e034548ec8f3d81
Original-Signed-off-by: Aaron Durbin <adurbin@chromium.org>
Original-Reviewed-on: https://chromium-review.googlesource.com/216925
Original-Reviewed-by: Furquan Shaikh <furquan@chromium.org>
Reviewed-on: http://review.coreboot.org/9042
Tested-by: build bot (Jenkins)
Reviewed-by: Stefan Reinauer <stefan.reinauer@coreboot.org>

4 files changed, 408 insertions, 22 deletions

diff --git a/src/arch/arm64/c_entry.c b/src/arch/arm64/c_entry.c
index e2aa4e5fb3..aba9fd3719 100644
--- a/src/arch/arm64/c_entry.c
+++ b/src/arch/arm64/c_entry.c
@@ -70,12 +70,18 @@ static void secondary_cpu_start(void)
 {
 	mmu_enable();
 	exception_hwinit();
-	soc_secondary_cpu_init();
-	/*
-	 * TODO(adurbin): need a proper place to park the CPUs. Currently
-	 * assuming SoC code does the appropriate thing.
-	 */
-	while (1);
+
+	if (!IS_ENABLED(CONFIG_SMP)) {
+		soc_secondary_cpu_init();
+		/*
+		 * TODO(adurbin): need a proper place to park the CPUs.
+		 * Currently assuming SoC code does the appropriate thing.
+		 */
+		while (1);
+	}
+
+	/* This will never return. */
+	arch_secondary_cpu_init();
 }
 
 extern void arm64_cpu_startup(void);
diff --git a/src/arch/arm64/cpu-internal.h b/src/arch/arm64/cpu-internal.h
index 3773ff23df..be13ba4e98 100644
--- a/src/arch/arm64/cpu-internal.h
+++ b/src/arch/arm64/cpu-internal.h
@@ -20,6 +20,11 @@
 #ifndef ARCH_CPU_INTERNAL_H
 #define ARCH_CPU_INTERNAL_H
 
+/*
+ * Code path for the non-BSP CPUs. This is an internal function used.
+ */
+void arch_secondary_cpu_init(void);
+
 /* Return the top of the stack for the specified cpu. */
 void *cpu_get_stack(unsigned int cpu);
 
diff --git a/src/arch/arm64/cpu.c b/src/arch/arm64/cpu.c
index 0f9f361fe7..f1733ea637 100644
--- a/src/arch/arm64/cpu.c
+++ b/src/arch/arm64/cpu.c
@@ -17,12 +17,332 @@
  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA, 02110-1301 USA
  */
 
+#include <stdint.h>
 #include <stdlib.h>
-#include <arch/cpu.h>
+#include <arch/barrier.h>
+#include <arch/lib_helpers.h>
+#include <cpu/cpu.h>
+#include <console/console.h>
+#include "cpu-internal.h"
 
 static struct cpu_info cpu_infos[CONFIG_MAX_CPUS];
 
+static inline struct cpu_info *cpu_info_for_cpu(unsigned int id)
+{
+	return &cpu_infos[id];
+}
+
 struct cpu_info *cpu_info(void)
 {
-	return &cpu_infos[smp_processor_id()];
+	return cpu_info_for_cpu(smp_processor_id());
+}
+
+static int cpu_online(struct cpu_info *ci)
+{
+	return load_acquire(&ci->online) != 0;
+}
+
+static void cpu_mark_online(struct cpu_info *ci)
+{
+	store_release(&ci->online, 1);
+}
+
+static inline void cpu_disable_dev(device_t dev)
+{
+	dev->enabled = 0;
+}
+
+static struct cpu_driver *locate_cpu_driver(uint32_t midr)
+{
+	struct cpu_driver *cur;
+
+	for (cur = cpu_drivers; cur != ecpu_drivers; cur++) {
+		const struct cpu_device_id *id_table = cur->id_table;
+
+		for (; id_table->midr != CPU_ID_END; id_table++) {
+			if (id_table->midr == midr)
+				return cur;
+		}
+	}
+	return NULL;
+}
+
+static int cpu_set_device_operations(device_t dev)
+{
+	uint32_t midr;
+	struct cpu_driver *driver;
+
+	midr = raw_read_midr_el1();
+	driver = locate_cpu_driver(midr);
+
+	if (driver == NULL) {
+		printk(BIOS_WARNING, "No CPU driver for MIDR %08x\n", midr);
+		return -1;
+	}
+	dev->ops = driver->ops;
+	return 0;
+}
+
+static void init_this_cpu(void *arg)
+{
+	struct cpu_info *ci = arg;
+	device_t dev = ci->cpu;
+
+	cpu_set_device_operations(dev);
+
+	if (dev->ops != NULL && dev->ops->init != NULL) {
+		dev->initialized = 1;
+		printk(BIOS_DEBUG, "%s init\n", dev_path(dev));
+		dev->ops->init(dev);
+	}
+}
+
+/* Fill in cpu_info structures according to device tree. */
+static void init_cpu_info(struct bus *bus)
+{
+	device_t cur;
+
+	for (cur = bus->children; cur != NULL; cur = cur->sibling) {
+		struct cpu_info *ci;
+		unsigned int id = cur->path.cpu.id;
+
+		if (cur->path.type != DEVICE_PATH_CPU)
+			continue;
+
+		/* IDs are currently mapped 1:1 with logical CPU numbers. */
+		if (id >= CONFIG_MAX_CPUS) {
+			printk(BIOS_WARNING,
+				"CPU id %x too large. Disabling.\n", id);
+			cpu_disable_dev(cur);
+			continue;
+		}
+
+		ci = cpu_info_for_cpu(id);
+		if (ci->cpu != NULL) {
+			printk(BIOS_WARNING,
+				"Duplicate ID %x in device tree.\n", id);
+			cpu_disable_dev(cur);
+		}
+
+		ci->cpu = cur;
+		ci->id = cur->path.cpu.id;
+	}
+
+	/* Mark current cpu online. */
+	cpu_mark_online(cpu_info());
+}
+
+static inline int action_queue_empty(struct cpu_action_queue *q)
+{
+	return load_acquire_exclusive(&q->todo) == NULL;
+}
+
+static inline int action_completed(struct cpu_action_queue *q,
+					struct cpu_action *action)
+{
+	return load_acquire(&q->completed) == action;
+}
+
+static inline void wait_for_action_queue_slot(struct cpu_action_queue *q)
+{
+	while (!action_queue_empty(q))
+		wfe();
+}
+
+static void wait_for_action_complete(struct cpu_action_queue *q,
+					struct cpu_action *a)
+{
+	while (!action_completed(q, a))
+		wfe();
+}
+
+static struct cpu_action *wait_for_action(struct cpu_action_queue *q,
+						struct cpu_action *local)
+{
+	struct cpu_action *action;
+
+	while (action_queue_empty(q))
+		wfe();
+
+	/*
+	 * Keep original address, but use a local copy for async processing.
+	 */
+	do {
+		action = load_acquire_exclusive(&q->todo);
+		*local = *action;
+	} while (!store_release_exclusive(&q->todo, NULL));
+
+	return action;
+}
+
+static void queue_action(struct cpu_action_queue *q, struct cpu_action *action)
+{
+	do {
+		wait_for_action_queue_slot(q);
+		if (load_acquire_exclusive(&q->todo) != NULL)
+			continue;
+	} while (!store_release_exclusive(&q->todo, action));
+}
+
+static void action_queue_complete(struct cpu_action_queue *q,
+					struct cpu_action *action)
+{
+	/* Mark completion and send events to waiters. */
+	store_release(&q->completed, action);
+	sev();
+}
+
+static void action_run(struct cpu_action *action)
+{
+	action->run(action->arg);
+}
+
+static void action_run_on_cpu(struct cpu_info *ci, struct cpu_action *action,
+				int sync)
+{
+	struct cpu_action_queue *q = &ci->action_queue;
+
+	/* Don't run actions on non-online or enabled devices. */
+	if (!cpu_online(ci) || ci->cpu == NULL || !ci->cpu->enabled)
+		return;
+
+	if (ci->id == smp_processor_id()) {
+		action->run(action->arg);
+		return;
+	}
+
+	queue_action(q, action);
+	/* Wait for CPU to pick it up. Empty slot means it was picked up. */
+	wait_for_action_queue_slot(q);
+	/* Wait for completion if requested. */
+	if (sync)
+		wait_for_action_complete(q, action);
+}
+
+static int __arch_run_on_cpu(unsigned int cpu, struct cpu_action *action,
+				int sync)
+{
+	struct cpu_info *ci;
+
+	if (cpu >= CONFIG_MAX_CPUS)
+		return -1;
+
+	ci = cpu_info_for_cpu(cpu);
+
+	action_run_on_cpu(ci, action, sync);
+
+	return 0;
+}
+
+int arch_run_on_cpu(unsigned int cpu, struct cpu_action *action)
+{
+	return __arch_run_on_cpu(cpu, action, 1);
+}
+
+int arch_run_on_cpu_async(unsigned int cpu, struct cpu_action *action)
+{
+	return __arch_run_on_cpu(cpu, action, 0);
+}
+
+static int __arch_run_on_all_cpus(struct cpu_action *action, int sync)
+{
+	int i;
+
+	for (i = 0; i < CONFIG_MAX_CPUS; i++)
+		action_run_on_cpu(cpu_info_for_cpu(i), action, sync);
+
+	return 0;
+}
+
+int arch_run_on_all_cpus(struct cpu_action *action)
+{
+	return __arch_run_on_all_cpus(action, 1);
+}
+
+int arch_run_on_all_cpus_async(struct cpu_action *action)
+{
+	return __arch_run_on_all_cpus(action, 0);
+}
+
+void arch_secondary_cpu_init(void)
+{
+	struct cpu_info *ci = cpu_info();
+	struct cpu_action_queue *q = &ci->action_queue;
+
+	/* Mark this CPU online. */
+	cpu_mark_online(ci);
+
+	while (1) {
+		struct cpu_action *orig;
+		struct cpu_action action;
+
+		orig = wait_for_action(q, &action);
+
+		action_run(&action);
+		action_queue_complete(q, orig);
+	}
+}
+
+void arch_initialize_cpus(device_t cluster, struct cpu_control_ops *cntrl_ops)
+{
+	size_t max_cpus;
+	size_t i;
+	struct cpu_info *ci;
+	void (*entry)(void);
+	struct bus *bus;
+
+	if (cluster->path.type != DEVICE_PATH_CPU_CLUSTER) {
+		printk(BIOS_ERR,
+			"CPU init failed. Device is not a CPU_CLUSTER: %s\n",
+			dev_path(cluster));
+		return;
+	}
+
+	bus = cluster->link_list;
+	entry = prepare_secondary_cpu_startup();
+
+	/* Initialize the cpu_info structures. */
+	init_cpu_info(bus);
+	max_cpus = cntrl_ops->total_cpus();
+
+	if (max_cpus > CONFIG_MAX_CPUS) {
+		printk(BIOS_WARNING,
+			"max_cpus (%zu) exceeds CONFIG_MAX_CPUS (%zu).\n",
+			max_cpus, (size_t)CONFIG_MAX_CPUS);
+		max_cpus = CONFIG_MAX_CPUS;
+	}
+
+	for (i = 0; i < max_cpus; i++) {
+		device_t dev;
+		struct cpu_action action;
+
+		ci = cpu_info_for_cpu(i);
+		dev = ci->cpu;
+
+		/* Disregard CPUs not in device tree. */
+		if (dev == NULL)
+			continue;
+
+		/* Skip disabled CPUs. */
+		if (!dev->enabled)
+			continue;
+
+		if (!cpu_online(ci)) {
+			/* Start the CPU. */
+			printk(BIOS_DEBUG, "Starting CPU%x\n", ci->id);
+			if (cntrl_ops->start_cpu(ci->id, entry)) {
+				printk(BIOS_ERR,
+					"Failed to start CPU%x\n", ci->id);
+				continue;
+			}
+			/* Wait for CPU to come online. */
+			while (!cpu_online(ci));
+			printk(BIOS_DEBUG, "CPU%x online.\n", ci->id);
+		}
+
+		/* Send it the init action. */
+		action.run = init_this_cpu;
+		action.arg = ci;
+		action_run_on_cpu(ci, &action, 1);
+	}
 }
diff --git a/src/arch/arm64/include/armv8/arch/cpu.h b/src/arch/arm64/include/armv8/arch/cpu.h
index b2c28f6d02..8a8e3b6c91 100644
--- a/src/arch/arm64/include/armv8/arch/cpu.h
+++ b/src/arch/arm64/include/armv8/arch/cpu.h
@@ -1,7 +1,7 @@
 /*
  * This file is part of the coreboot project.
  *
- * Copyright 2012 Google Inc.
+ * Copyright 2014 Google Inc.
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License as published by
@@ -25,32 +25,77 @@
 #if !defined(__PRE_RAM__)
 #include <device/device.h>
 
+enum {
+	CPU_ID_END = 0x00000000,
+};
+
+struct cpu_device_id {
+	uint32_t midr;
+};
+
 struct cpu_driver {
+	/* This is excessive as init() is the only one called. */
 	struct device_operations *ops;
-	struct cpu_device_id *id_table;
+	const struct cpu_device_id *id_table;
 };
 
-struct thread;
+/* Action to run. */
+struct cpu_action {
+	void (*run)(void *arg);
+	void *arg;
+};
+
+/*
+ * Actions are queued to 'todo'. When picked up 'todo' is cleared. The
+ * 'completed' field is set to the original 'todo' value when the action
+ * is complete.
+ */
+struct cpu_action_queue {
+	struct cpu_action *todo;
+	struct cpu_action *completed;
+};
 
 struct cpu_info {
 	device_t cpu;
-	unsigned long index;
-#if CONFIG_COOP_MULTITASKING
-	struct thread *thread;
-#endif
+	struct cpu_action_queue action_queue;
+	unsigned int online;
+	/* Current assumption is that id matches smp_processor_id(). */
+	unsigned int id;
 };
 
-#endif
-
+/* Obtain cpu_info for current executing CPU. */
 struct cpu_info *cpu_info(void);
 
+/* Control routines for starting CPUs. */
+struct cpu_control_ops {
+	/* Return the maximum number of CPUs supported. */
+	size_t (*total_cpus)(void);
+	/*
+	 * Start the requested CPU and have it start running entry().
+	 * Returns 0 on success, < 0 on error.
+	 */
+	int (*start_cpu)(unsigned int id, void (*entry)(void));
+};
+
 /*
- * Returns logical cpu in range [0:MAX_CPUS). SoC should define this.
- * Additionally, this is needed early in arm64 init so it should not
- * rely on a stack. Standard clobber list is fair game: x0-x7 and x0
- * returns the logical cpu number.
+ * Initialize all DEVICE_PATH_CPUS under the DEVICE_PATH_CPU_CLUSTER cluster.
+ * type DEVICE_PATH_CPUS. Start up is controlled by cntrl_ops.
  */
-unsigned int smp_processor_id(void);
+void arch_initialize_cpus(device_t cluster, struct cpu_control_ops *cntrl_ops);
+
+/*
+ * Run cpu_action returning < 0 on error, 0 on success. There are synchronous
+ * and asynchronous methods. Both cases ensure the action has been picked up
+ * by the target cpu. The synchronous variants will wait for the action to
+ * be completed before returning.
+ *
+ * Though the current implementation allows queuing actions on the main cpu,
+ * the main cpu doesn't process its own queue.
+ */
+int arch_run_on_cpu(unsigned int cpu, struct cpu_action *action);
+int arch_run_on_all_cpus(struct cpu_action *action);
+int arch_run_on_cpu_async(unsigned int cpu, struct cpu_action *action);
+int arch_run_on_all_cpus_async(struct cpu_action *action);
 
 /*
  * Do the necessary work to prepare for secondary CPUs coming up. The
@@ -65,4 +110,14 @@ void *prepare_secondary_cpu_startup(void);
  */
 void soc_secondary_cpu_init(void);
 
+#endif /* !__PRE_RAM__ */
+
+/*
+ * Returns logical cpu in range [0:MAX_CPUS). SoC should define this.
+ * Additionally, this is needed early in arm64 init so it should not
+ * rely on a stack. Standard clobber list is fair game: x0-x7 and x0
+ * returns the logical cpu number.
+ */
+unsigned int smp_processor_id(void);
+
 #endif /* __ARCH_CPU_H__ */