add smp_call_function support

take use of VCPU_NOTIFY vector, add smp_call_function support.
added a per_cpu field smp_call_info, and make each smp_call_function
is not re-entered, and the caller CPU is returned when all the target
CPUs complete the call.

v4:
- remove global lock
- take use of wait_sync_change function to do the sequence sync

v3:
- remove per_cpu lock in smp_call_info
- use a global lock to ensure smp_call_function sequence
- use pcpu_sync_sleep to wait IPI complete

v2:
- after new smp function come, if old one exist, changed from overwirte
  with the new one to ignore the new one.

Signed-off-by: Jason Chen CJ <jason.cj.chen@intel.com>

Author: JasonChenCJ

hypervisor/arch/x86/notify.c

@@ -8,15 +8,53 @@
 
 static uint32_t notification_irq = IRQ_INVALID;
 
+static volatile uint64_t smp_call_mask = 0UL;
+
 /* run in interrupt context */
 static int kick_notification(__unused uint32_t irq, __unused void *data)
 {
-	/* Notification vector does not require handling here, it's just used
-	 * to kick taget cpu out of non-root mode.
+	/* Notification vector is used to kick taget cpu out of non-root mode.
+	 * And it also serves for smp call.
 	 */
+	uint16_t pcpu_id = get_cpu_id();
+
+	if (bitmap_test(pcpu_id, &smp_call_mask)) {
+		struct smp_call_info_data *smp_call =
+			&per_cpu(smp_call_info, pcpu_id);
+
+		if (smp_call->func)
+			smp_call->func(smp_call->data);
+		bitmap_clear_nolock(pcpu_id, &smp_call_mask);
+	}
+
 	return 0;
 }
 
+void smp_call_function(uint64_t mask, smp_call_func_t func, void *data)
+{
+	uint16_t pcpu_id;
+	struct smp_call_info_data *smp_call;
+
+	/* wait for previous smp call complete, which may run on other cpus */
+	while (atomic_cmpxchg64(&smp_call_mask, 0UL, mask & INVALID_BIT_INDEX));
+	while ((pcpu_id = ffs64(mask)) != INVALID_BIT_INDEX) {
+		bitmap_clear_nolock(pcpu_id, &mask);
+		if (bitmap_test(pcpu_id, &pcpu_active_bitmap)) {
+			smp_call = &per_cpu(smp_call_info, pcpu_id);
+			smp_call->func = func;
+			smp_call->data = data;
+		} else {
+			/* pcpu is not in active, print error */
+			pr_err("pcpu_id %d not in active!", pcpu_id);
+			bitmap_clear_nolock(pcpu_id, &smp_call_mask);
+		}
+	}
+	send_dest_ipi(smp_call_mask, VECTOR_NOTIFY_VCPU,
+				INTR_LAPIC_ICR_LOGICAL);
+	/* wait for current smp call complete */
+	wait_sync_change(&smp_call_mask, 0UL);
+}
+
 static int request_notification_irq(irq_action_t func, void *data,
 				const char *name)
 {

hypervisor/include/arch/x86/irq.h

@@ -47,6 +47,13 @@ struct intr_excp_ctx {
 	uint64_t ss;
 };
 
+typedef void (*smp_call_func_t)(void *data);
+struct smp_call_info_data {
+	smp_call_func_t func;
+	void *data;
+};
+
+void smp_call_function(uint64_t mask, smp_call_func_t func, void *data);
 int handle_level_interrupt_common(struct irq_desc *desc,
 	__unused void *handler_data);
 int common_handler_edge(struct irq_desc *desc, __unused void *handler_data);

hypervisor/include/arch/x86/per_cpu.h

@@ -46,6 +46,7 @@ struct per_cpu_region {
 	uint8_t stack[CONFIG_STACK_SIZE] __aligned(16);
 	char logbuf[LOG_MESSAGE_MAX_SIZE];
 	uint8_t lapic_id;
+	struct smp_call_info_data smp_call_info;
 } __aligned(CPU_PAGE_SIZE); //per_cpu_region size aligned with CPU_PAGE_SIZE
 
 extern struct per_cpu_region *per_cpu_data_base_ptr;