Implement ARM64 Support

CMakeLists.txt

@@ -10,6 +10,11 @@ if (DSERVER_TOOLS)
 	add_subdirectory(tools)
 endif()
 
+if (TARGET_ARM64)
+	# Unlike Linux, Apple's uses `__arm64__` instead of `__aarch64__`
+	add_definitions(-D__arm64__)
+endif (TARGET_ARM64)
+
 if (DSERVER_ASAN)
 	add_compile_definitions(
 		DSERVER_ASAN=1

duct-tape/CMakeLists.txt

@@ -127,6 +127,13 @@ add_compile_definitions(
 	__DARWIN_BYTE_ORDER=__DARWIN_LITTLE_ENDIAN
 )
 
+if(TARGET_ARM64)
+	add_compile_definitions(
+		ARM64_BOARD_CONFIG_VMAPPLE
+	)
+endif(TARGET_ARM64)
+
+
 add_compile_options(
 	-fblocks
 	-ffunction-sections
@@ -253,7 +260,7 @@ add_custom_target(kernel_mig_generate
 		${CMAKE_CURRENT_BINARY_DIR}/xnu/osfmk/UserNotification/UNDReplyServer.c
 )
 
-add_library(darlingserver_duct_tape STATIC
+set(darlingserver_duct_tape_source
 	src/init.c
 	src/misc.c
 	src/stubs.c
@@ -321,8 +328,6 @@ add_library(darlingserver_duct_tape STATIC
 	xnu/osfmk/ipc/mig_log.c
 	xnu/osfmk/ipc/ipc_eventlink.c
 
-	xnu/osfmk/i386/rtclock.c
-
 	xnu/osfmk/prng/prng_random.c
 
 	xnu/osfmk/vm/vm32_user.c
@@ -357,6 +362,18 @@ add_library(darlingserver_duct_tape STATIC
 	pthread/kern_synch.c
 )
 
+if (TARGET_i386 OR TARGET_x86_64)
+set(darlingserver_duct_tape_source ${darlingserver_duct_tape_source}
+	xnu/osfmk/i386/rtclock.c
+)
+elseif (TARGET_ARM64)
+set(darlingserver_duct_tape_source ${darlingserver_duct_tape_source}
+	xnu/osfmk/arm/rtclock.c
+)
+endif()
+
+add_library(darlingserver_duct_tape STATIC ${darlingserver_duct_tape_source})
+
 set_source_files_properties(pthread/kern_synch.c PROPERTIES
 	COMPILE_FLAGS "-I${CMAKE_CURRENT_SOURCE_DIR}/pthread"
 )

duct-tape/internal-include/darlingserver/duct-tape/thread.h

@@ -21,6 +21,12 @@ typedef struct dtape_thread_user_state {
 #if __x86_64__
 	x86_thread_state_t thread_state;
 	x86_float_state_t float_state;
+#elif __aarch64__
+	arm_thread_state64_t thread_state;
+	// arm_neon_state64_t neon_state;
+#warning "TODO: Implement other ARM64 states"
+#else
+	#error "Missing user states"
 #endif
 } dtape_thread_user_state_t;
 

duct-tape/src/host.c

@@ -65,6 +65,9 @@ kern_return_t host_info(host_t host, host_flavor_t flavor, host_info_t info, mac
 #else
 			basic_info->cpu_subtype = CPU_SUBTYPE_I386_ALL;
 #endif
+#elif __aarch64__
+			basic_info->cpu_type = CPU_TYPE_ARM64;
+			basic_info->cpu_subtype = CPU_SUBTYPE_ARM64_ALL;
 #else
 			#error Unknown CPU type
 #endif

duct-tape/src/locks.c

@@ -68,7 +68,13 @@ void dtape_mutex_lock(dtape_mutex_t* mutex) {
 				return;
 			}
 			libsimple_lock_unlock(&mutex->dtape_queue_lock);
+#if defined(__x86_64__) || defined(__i386__)
 			__builtin_ia32_pause();
+#elif __aarch64__
+			asm volatile("yield");
+#else
+#error Missing CPU pause for this architecture
+#endif
 		}
 	}
 
@@ -526,10 +532,12 @@ hw_wait_while_equals(void **address, void *current)
 	for (;;) {
 		for (int i = 0; i < LOCK_SNOOP_SPINS; i++) {
 #ifdef __DARLING__
-#if __x86_64__ || __i386__
+#if defined(__x86_64__) || defined(__i386__)
 			__builtin_ia32_pause();
+#elif __aarch64__
+			asm volatile("yield");
 #else
-			#warning Missing CPU pause for this architecture
+#error Missing CPU pause for this architecture
 #endif
 #else
 			cpu_pause();

duct-tape/src/memory.c

@@ -52,8 +52,10 @@ int ftruncate(int fd, off_t length);
 
 #define MFD_CLOEXEC 0x1
 
-void dtape_memory_init(void) {
+int PAGE_SHIFT_CONST;
 
+void dtape_memory_init(void) {
+	PAGE_SHIFT_CONST = __builtin_ctzl(sysconf(_SC_PAGESIZE));
 };
 
 static uint64_t dtape_byte_count_to_page_count_round_up(uint64_t byte_count) {

duct-tape/src/misc.c

@@ -39,6 +39,23 @@ char version[] = "Darling 11.5";
 		[x86_PAGEIN_STATE]              = x86_PAGEIN_STATE_COUNT
 	};
 	// </copied>
+#elif defined(__aarch64__)
+	// <copied from="xnu://7195.141.2/osfmk/arm64/status.c"
+	unsigned int _MachineStateCount[] = {
+		[ARM_UNIFIED_THREAD_STATE] = ARM_UNIFIED_THREAD_STATE_COUNT,
+		[ARM_VFP_STATE] = ARM_VFP_STATE_COUNT,
+		[ARM_EXCEPTION_STATE] = ARM_EXCEPTION_STATE_COUNT,
+		[ARM_DEBUG_STATE] = ARM_DEBUG_STATE_COUNT,
+		[ARM_THREAD_STATE64] = ARM_THREAD_STATE64_COUNT,
+		[ARM_EXCEPTION_STATE64] = ARM_EXCEPTION_STATE64_COUNT,
+		[ARM_THREAD_STATE32] = ARM_THREAD_STATE32_COUNT,
+		[ARM_DEBUG_STATE32] = ARM_DEBUG_STATE32_COUNT,
+		[ARM_DEBUG_STATE64] = ARM_DEBUG_STATE64_COUNT,
+		[ARM_NEON_STATE] = ARM_NEON_STATE_COUNT,
+		[ARM_NEON_STATE64] = ARM_NEON_STATE64_COUNT,
+		[ARM_PAGEIN_STATE] = ARM_PAGEIN_STATE_COUNT,
+	};
+	// </copied>
 #else
 	#error _MachineStateCount not defined on this architecture
 #endif
@@ -101,10 +118,11 @@ unsigned int waitq_held(struct waitq* wq) {
 	return wq->dtape_waitq_interlock.dtape_interlock.dtape_interlock.dtape_mutex.dtape_owner == (uintptr_t)current_thread();
 };
 
-#if __x86_64__
+#if __x86_64__ || __aarch64__
 
 //
 // <copied from="xnu://7195.141.2/osfmk/x86_64/loose_ends.c">
+// <copied from="xnu://7195.141.2/osfmk/arm64/loose_ends.c">
 //
 
 /*
@@ -120,8 +138,33 @@ fls(unsigned int mask)
 	return (sizeof(mask) << 3) - __builtin_clz(mask);
 }
 
+#endif
+
+//
+// Since this is the only method we need from `bsd_kern.c`, I rather not
+// `#ifndef` all of the source code that we don't need in that file.
+//
+// <copied from="xnu://7195.141.2/osfmk/kern/bsd_kern.c">
+//
+
+#ifdef __aarch64__
+
+task_t
+get_threadtask(thread_t th)
+{
+	return th->task;
+}
+
+#endif
+
 //
 // </copied>
 //
 
+// Unlike i386, the preemption methods are not 
+// inline for arm. So we will need to create stubs.
+#ifdef __aarch64__
+int get_preemption_level(void) { return 0; }
+void _enable_preemption(void) {}
+void _disable_preemption(void) {}
 #endif

duct-tape/src/processor.c

@@ -92,6 +92,9 @@ kern_return_t processor_info(processor_t processor, processor_flavor_t flavor, h
 #else
 			info->cpu_subtype = CPU_SUBTYPE_I386_ALL;
 #endif
+#elif __aarch64__
+			info->cpu_type = CPU_TYPE_ARM64;
+			info->cpu_subtype = CPU_SUBTYPE_ARM64_ALL;
 #else
 			#error Unknown CPU type
 #endif