Synchronize the TSCs of all CPUs early in the boot process, so system_time()

will return consistent values. This helps with debug measurements for the time
being. Obviously we'll have to think of something different when we support
speed-stepping on models with frequency-dependent TSCs.


git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@30287 a95241bf-73f2-0310-859d-f6bbb57e9c96

headers/private/kernel/arch/x86/arch_cpu.h

@@ -22,6 +22,7 @@
 
 
 // MSR registers (possibly Intel specific)
+#define IA32_MSR_TSC					0x10
 #define IA32_MSR_APIC_BASE				0x1b
 
 #define IA32_MSR_MTRR_CAPABILITIES		0xfe

src/system/kernel/arch/x86/arch_cpu.cpp

@@ -78,6 +78,7 @@ bool gHasSSE = false;
 
 static uint32 sCpuRendezvous;
 static uint32 sCpuRendezvous2;
+static vint32 sTSCSyncRendezvous;
 
 segment_descriptor *gGDT = NULL;
 
@@ -501,6 +502,28 @@ arch_cpu_preboot_init_percpu(kernel_args *args, int cpu)
 	x86_write_cr0(x86_read_cr0() & ~(CR0_FPU_EMULATION | CR0_MONITOR_FPU));
 	gX86SwapFPUFunc = i386_fnsave_swap;
 
+	// On SMP system we want to synchronize the CPUs' TSCs, so system_time()
+	// will return consistent values.
+	if (smp_get_num_cpus() > 1) {
+		// let the first CPU prepare the rendezvous point
+		if (cpu == 0)
+			sTSCSyncRendezvous = smp_get_num_cpus() - 1;
+
+		// One CPU after the other will drop out of this loop and be caught by
+		// the loop below, until the last CPU (0) gets there. Save for +/- a few
+		// cycles the CPUs should pass the second loop at the same time.
+		while (sTSCSyncRendezvous > cpu) {
+		}
+
+		sTSCSyncRendezvous = cpu - 1;
+
+		while (sTSCSyncRendezvous != -1) {
+		}
+
+		// reset TSC to 0
+		x86_write_msr(IA32_MSR_TSC, 0);
+	}
+
 	return B_OK;
 }