hw: always serialize rdtsc reads

While implementing TSC calibration in #5215, the issue of properly serializing
TSC reads came up. Some learnings of the discussion were noted in #5430.

Using `cpuid` for serialization as in Trace::timestamp() is portable,
but will cause VM exits on VMX and SVM and is therefore unsuitable to
retain a roughly working calibration loop while running virtualized.
On the other hand on most AMD systems, dispatch serializing `lfence`
needs to be explicitly enabled via a non-architectural MSR.

Enable setting up dispatch serializing lfence on AMD systems and always
serialize rdtsc accesses in Hw::Tsc::rdtsc() for maximum reliability.

Issues #5215, #5430

repos/base-hw/src/bootstrap/spec/x86_64/platform.cc

@@ -131,7 +131,7 @@ static void disable_pit()
 		PIT_MODE       = 0x43,
 	};
 
-	/**
+	/*
 	 * Disable PIT timer channel. This is necessary since BIOS sets up
 	 * channel 0 to fire periodically.
 	 */
@@ -141,6 +141,35 @@ static void disable_pit()
 }
 
 
+/*
+ * Enable dispatch serializing lfence instruction on AMD processors
+ *
+ * See Software techniques for managing speculation on AMD processors
+ *     Revision 5.09.23
+ *     Mitigation G-2
+ */
+static void amd_enable_serializing_lfence()
+{
+	using Cpu = Hw::X86_64_cpu;
+
+	if (Hw::Vendor::get_vendor_id() != Hw::Vendor::Vendor_id::AMD)
+		return;
+
+	unsigned const family = Hw::Vendor::get_family();
+
+	/*
+	 * In family 0Fh and 11h, lfence is always dispatch serializing and
+	 * "AMD plans support for this MSR and access to this bit for all future
+	 * processors." from family 14h on.
+	 */
+	if ((family == 0x10) || (family == 0x12) || (family >= 0x14)) {
+		Cpu::Amd_lfence::access_t amd_lfence = Cpu::Amd_lfence::read();
+		Cpu::Amd_lfence::Enable_dispatch_serializing::set(amd_lfence);
+		Cpu::Amd_lfence::write(amd_lfence);
+	}
+}
+
+
 Bootstrap::Platform::Board::Board()
 :
 	core_mmio(Memory_region { 0, 0x1000 },
@@ -325,6 +354,14 @@ Bootstrap::Platform::Board::Board()
 		cpus = !cpus ? 1 : max_cpus;
 	}
 
+	/*
+	 * Enable serializing lfence on supported AMD processors
+	 *
+	 * For APs this will be set up later, but we need it already to obtain
+	 * the most acurate results when calibrating the TSC frequency.
+	 */
+	amd_enable_serializing_lfence();
+
 	auto r = calibrate_lapic_frequency(info.acpi_fadt);
 	info.lapic_freq_khz = r.freq_khz;
 	info.lapic_div      = r.div;
@@ -401,9 +438,12 @@ unsigned Bootstrap::Platform::enable_mmu()
 	if (board.cpus <= 1)
 		return (unsigned)cpu_id;
 
-	if (!Cpu::IA32_apic_base::Bsp::get(lapic_msr))
+	if (!Cpu::IA32_apic_base::Bsp::get(lapic_msr)) {
 		/* AP - done */
+		/* enable serializing lfence on supported AMD processors. */
+		amd_enable_serializing_lfence();
 		return (unsigned)cpu_id;
+	}
 
 	/* BSP - we're primary CPU - wake now all other CPUs */
 

repos/base-hw/src/include/hw/spec/x86_64/cpu.h

@@ -118,6 +118,12 @@ struct Hw::X86_64_cpu
 	/* AMD host save physical address */
 	X86_64_MSR_REGISTER(Amd_vm_hsavepa, 0xC0010117);
 
+
+	/* Non-architectural MSR used to make lfence serializing */
+	X86_64_MSR_REGISTER(Amd_lfence, 0xC0011029,
+		struct Enable_dispatch_serializing : Bitfield<1, 1> { }; /* Enable lfence dispatch serializing */
+	)
+
 	X86_64_MSR_REGISTER(Platform_id, 0x17,
 		struct Bus_ratio : Bitfield<8, 5> { }; /* Bus ratio on Core 2, see SDM 19.7.3 */
 	);

repos/base-hw/src/include/hw/spec/x86_64/x86_64.h

@@ -109,10 +109,22 @@ public:
 
 struct Hw::Tsc
 {
+	/*
+	 * Provide serialized access to the Timestamp Counter
+	 *
+	 * See #5430 for more information.
+	 */
 	static Genode::uint64_t rdtsc()
 	{
 		Genode::uint32_t low, high;
-		asm volatile("rdtsc" : "=a"(low), "=d"(high));
+		asm volatile(
+			"lfence;"
+			"rdtsc;"
+			"lfence;"
+			: "=a"(low), "=d"(high)
+			:
+			: "memory"
+		);
 		return (Genode::uint64_t)(high) << 32 | low;
 	}