cpu/x86: allow AP callbacks after MP init

There are circumstances where the APs need to run a piece of
code later in the boot flow. The current MP init just parks
the APs after MP init is completed so there's not an opportunity
to target running a piece of code on all the APs at a later time.
Therefore, provide an option, PARALLEL_MP_AP_WORK, that allows
the APs to perform callbacks.

BUG=chrome-os-partner:60657
BRANCH=reef

Change-Id: I849ecfdd6641dd9424943e246317cd1996ef1ba6
Signed-off-by: Aaron Durbin <adurbin@chromium.org>
Reviewed-on: https://review.coreboot.org/17745
Tested-by: build bot (Jenkins)
Reviewed-by: Furquan Shaikh <furquan@google.com>
Reviewed-by: Lijian Zhao <lijian.zhao@intel.com>

4 files changed, 137 insertions, 7 deletions

diff --git a/src/arch/x86/cpu.c b/src/arch/x86/cpu.c
index fbd48b0c2b..1e74d0cd59 100644
--- a/src/arch/x86/cpu.c
+++ b/src/arch/x86/cpu.c
@@ -11,11 +11,13 @@
  * GNU General Public License for more details.
  */
 
+#include <bootstate.h>
 #include <boot/coreboot_tables.h>
 #include <console/console.h>
 #include <cpu/cpu.h>
 #include <arch/io.h>
 #include <string.h>
+#include <cpu/x86/mp.h>
 #include <cpu/x86/mtrr.h>
 #include <cpu/x86/msr.h>
 #include <cpu/x86/lapic.h>
@@ -310,3 +312,14 @@ void lb_arch_add_records(struct lb_header *header)
 	tsc_info->size = sizeof(*tsc_info);
 	tsc_info->freq_khz = freq_khz;
 }
+
+void arch_bootstate_coreboot_exit(void)
+{
+	/* APs are already parked by existing infrastructure. */
+	if (!IS_ENABLED(CONFIG_PARALLEL_MP_AP_WORK))
+		return;
+
+	/* APs are waiting for work. Last thing to do is park them. */
+	if (mp_park_aps())
+		printk(BIOS_ERR, "Parking APs failed.\n");
+}
diff --git a/src/cpu/x86/Kconfig b/src/cpu/x86/Kconfig
index 3e56d721f5..2e233cc778 100644
--- a/src/cpu/x86/Kconfig
+++ b/src/cpu/x86/Kconfig
@@ -10,6 +10,12 @@ config PARALLEL_MP
 	 in parallel. It additionally provides a more flexible mechanism
 	 for sequencing the steps of bringing up the APs.
 
+config PARALLEL_MP_AP_WORK
+	def_bool n
+	depends on PARALLEL_MP
+	help
+	 Allow APs to do other work after initialization instead of going
+	 to sleep.
 
 config UDELAY_IO
 	bool
diff --git a/src/cpu/x86/mp_init.c b/src/cpu/x86/mp_init.c
index baa3599a94..c989963442 100644
--- a/src/cpu/x86/mp_init.c
+++ b/src/cpu/x86/mp_init.c
@@ -127,6 +127,7 @@ struct mp_flight_plan {
 	struct mp_flight_record *records;
 };
 
+static int global_num_aps;
 static struct mp_flight_plan mp_info;
 
 struct cpu_map {
@@ -185,6 +186,11 @@ static void ap_do_flight_plan(void)
 	}
 }
 
+static void park_this_cpu(void)
+{
+	stop_this_cpu();
+}
+
 /* By the time APs call ap_init() caching has been setup, and microcode has
  * been loaded. */
 static void asmlinkage ap_init(unsigned int cpu)
@@ -210,7 +216,7 @@ static void asmlinkage ap_init(unsigned int cpu)
 	ap_do_flight_plan();
 
 	/* Park the AP. */
-	stop_this_cpu();
+	park_this_cpu();
 }
 
 static void setup_default_sipi_vector_params(struct sipi_params *sp)
@@ -587,7 +593,6 @@ static void init_bsp(struct bus *cpu_bus)
 static int mp_init(struct bus *cpu_bus, struct mp_params *p)
 {
 	int num_cpus;
-	int num_aps;
 	atomic_t *ap_count;
 
 	init_bsp(cpu_bus);
@@ -621,11 +626,11 @@ static int mp_init(struct bus *cpu_bus, struct mp_params *p)
 	wbinvd();
 
 	/* Start the APs providing number of APs and the cpus_entered field. */
-	num_aps = p->num_cpus - 1;
-	if (start_aps(cpu_bus, num_aps, ap_count) < 0) {
+	global_num_aps = p->num_cpus - 1;
+	if (start_aps(cpu_bus, global_num_aps, ap_count) < 0) {
 		mdelay(1000);
 		printk(BIOS_DEBUG, "%d/%d eventually checked in?\n",
-		       atomic_read(ap_count), num_aps);
+		       atomic_read(ap_count), global_num_aps);
 		return -1;
 	}
 
@@ -838,6 +843,94 @@ static void trigger_smm_relocation(void)
 	mp_state.ops.per_cpu_smm_trigger();
 }
 
+static mp_callback_t ap_callbacks[CONFIG_MAX_CPUS];
+
+static mp_callback_t read_callback(mp_callback_t *slot)
+{
+	return *(volatile mp_callback_t *)slot;
+}
+
+static void store_callback(mp_callback_t *slot, mp_callback_t value)
+{
+	*(volatile mp_callback_t *)slot = value;
+}
+
+static int run_ap_work(mp_callback_t func, long expire_us)
+{
+	int i;
+	int cpus_accepted;
+	struct stopwatch sw;
+	int cur_cpu = cpu_index();
+
+	if (!IS_ENABLED(CONFIG_PARALLEL_MP_AP_WORK)) {
+		printk(BIOS_ERR, "APs already parked. PARALLEL_MP_AP_WORK not selected.\n");
+		return -1;
+	}
+
+	/* Signal to all the APs to run the func. */
+	for (i = 0; i < ARRAY_SIZE(ap_callbacks); i++) {
+		if (cur_cpu == i)
+			continue;
+		store_callback(&ap_callbacks[i], func);
+	}
+	mfence();
+
+	/* Wait for all the APs to signal back that call has been accepted. */
+	stopwatch_init_usecs_expire(&sw, expire_us);
+	for (cpus_accepted = 0; !stopwatch_expired(&sw); cpus_accepted = 0) {
+		for (i = 0; i < ARRAY_SIZE(ap_callbacks); i++) {
+			if (cur_cpu == i)
+				continue;
+			if (read_callback(&ap_callbacks[i]) == NULL)
+				cpus_accepted++;
+		}
+		if (cpus_accepted == global_num_aps)
+			return 0;
+	}
+
+	printk(BIOS_ERR, "AP call expired. %d/%d CPUs accepted.\n",
+		cpus_accepted, global_num_aps);
+	return -1;
+}
+
+static void ap_wait_for_instruction(void)
+{
+	int cur_cpu = cpu_index();
+
+	if (!IS_ENABLED(CONFIG_PARALLEL_MP_AP_WORK))
+		return;
+
+	while (1) {
+		mp_callback_t func = read_callback(&ap_callbacks[cur_cpu]);
+
+		if (func == NULL) {
+			asm ("pause");
+			continue;
+		}
+
+		store_callback(&ap_callbacks[cur_cpu], NULL);
+		mfence();
+		func();
+	}
+}
+
+int mp_run_on_aps(void (*func)(void), long expire_us)
+{
+	return run_ap_work(func, expire_us);
+}
+
+int mp_run_on_all_cpus(void (*func)(void), long expire_us)
+{
+	/* Run on BSP first. */
+	func();
+	return mp_run_on_aps(func, expire_us);
+}
+
+int mp_park_aps(void)
+{
+	return mp_run_on_aps(park_this_cpu, 10 * USECS_PER_MSEC);
+}
+
 static struct mp_flight_record mp_steps[] = {
 	/* Once the APs are up load the SMM handlers. */
 	MP_FR_BLOCK_APS(NULL, load_smm_handlers),
@@ -845,8 +938,8 @@ static struct mp_flight_record mp_steps[] = {
 	MP_FR_NOBLOCK_APS(trigger_smm_relocation, trigger_smm_relocation),
 	/* Initialize each CPU through the driver framework. */
 	MP_FR_BLOCK_APS(mp_initialize_cpu, mp_initialize_cpu),
-	/* Wait for APs to finish everything else then let them park. */
-	MP_FR_BLOCK_APS(NULL, NULL),
+	/* Wait for APs to finish then optionally start looking for work. */
+	MP_FR_BLOCK_APS(ap_wait_for_instruction, NULL),
 };
 
 static void fill_mp_state(struct mp_state *state, const struct mp_ops *ops)
diff --git a/src/include/cpu/x86/mp.h b/src/include/cpu/x86/mp.h
index 6d51d7b005..b9b4d5772c 100644
--- a/src/include/cpu/x86/mp.h
+++ b/src/include/cpu/x86/mp.h
@@ -124,6 +124,24 @@ struct mp_ops {
  */
 int mp_init_with_smm(struct bus *cpu_bus, const struct mp_ops *mp_ops);
 
+
+/*
+ * After APs are up and PARALLEL_MP_AP_WORK is enabled one can issue work
+ * to all the APs to perform. Currently the BSP is the only CPU that is allowed
+ * to issue work. i.e. the APs should not call any of these functions.
+ * All functions return < 0 on error, 0 on success.
+ */
+int mp_run_on_aps(void (*func)(void), long expire_us);
+
+/* Like mp_run_on_aps() but also runs func on BSP. */
+int mp_run_on_all_cpus(void (*func)(void), long expire_us);
+
+/*
+ * Park all APs to prepare for OS boot. This is handled automatically
+ * by the coreboot infrastructure.
+ */
+int mp_park_aps(void);
+
 /*
  * SMM helpers to use with initializing CPUs.
  */