After KPTI, it is no longer trivial for us to have a custom int 0x83 handler -> go back to ioctl()

src/dyld/src/dyldAPIs.cpp

@@ -83,6 +83,10 @@ static NSLinkEditErrors sLastErrorFileCode;
 static int sLastErrorNo;
 #endif
 
+#ifdef DARLING
+extern "C" int mach_driver_get_fd(void);
+#endif
+
 // In 10.3.x and earlier all the NSObjectFileImage API's were implemeneted in libSystem.dylib
 // Beginning in 10.4 the NSObjectFileImage API's are implemented in dyld and libSystem just forwards
 // This conditional keeps support for old libSystem's which needed some help implementing the API's
@@ -208,6 +212,9 @@ static struct dyld_func dyld_funcs[] = {
 #if OLD_LIBSYSTEM_SUPPORT
     {"__dyld_link_module",							(void*)_dyld_link_module },
 #endif
+#ifdef DARLING
+	{"__dyld_get_mach_driver_fd", (void*)mach_driver_get_fd },
+#endif
 #endif //DEPRECATED_APIS_SUPPORTED
 
     {NULL, 0}

src/dyld/src/dyldInitialization.cpp

@@ -195,7 +195,7 @@ static void rebaseDyld(const struct macho_header* mh, intptr_t slide)
 }
 
 
-extern "C" void mach_init();
+extern "C" void mach_init(const char* apple[]);
 extern "C" void __guard_setup(const char* apple[]);
 
 
@@ -222,7 +222,7 @@ uintptr_t start(const struct macho_header* appsMachHeader, int argc, const char*
 	++apple;
 
 	// allow dyld to use mach messaging
-	mach_init();
+	mach_init(apple);
 
 	// set up random value for stack canary
 	__guard_setup(apple);

src/kernel/emulation/linux/mach/CMakeLists.txt

@@ -15,7 +15,6 @@ set(mach_server_client_sources
 	mach_time.c
 	lkm.c
 	darling_mach_syscall.S
-	lkm-int.S
 	mach_table.c
 )
 

src/kernel/emulation/linux/mach/lkm-int.S

@@ -1,22 +0,0 @@
-.globl _lkm_call
-
-#if defined(__x86_64__)
-
-_lkm_call:
-	movq %rdi, %rax
-	movq %rsi, %rcx
-	int $0x83
-	ret
-	
-#elif defined(__i386__)
-
-_lkm_call:
-	movl 4(%esp), %eax
-	movl 8(%esp), %ecx
-	int $0x83
-	ret
-
-#else
-#error Missing implementation
-#endif
-

src/kernel/emulation/linux/mach/lkm.c

@@ -21,8 +21,68 @@ extern _libkernel_functions_t _libkernel_functions;
 extern void sigexc_setup1(void);
 extern void sigexc_setup2(void);
 
-void mach_driver_init(void)
+static int driver_fd = -1;
+
+void mach_driver_init(const char** applep)
 {
+#ifdef VARIANT_DYLD
+	if (applep != NULL)
+	{
+		int i;
+		for (i = 0; applep[i] != NULL; i++)
+		{
+			if (strncmp(applep[i], "kernfd=", 7) == 0)
+			{
+				driver_fd = __simple_atoi(applep[i] + 7, NULL);
+				break;
+			}
+		}
+	}
+#else
+	// Ask for fd already set up by dyld
+	int (*p)(void);
+	_libkernel_functions->dyld_func_lookup("__dyld_get_mach_driver_fd", &p);
+
+	driver_fd = (*p)();
+#endif
+
+	// If mach_driver_init() is being called in the fork child, the LKM will now
+	// swap out driver_fd for a new one.
+	if (__real_ioctl(driver_fd, NR_get_api_version, 0) != DARLING_MACH_API_VERSION)
+	{
+		const char* msg = "Darling Mach kernel module reports different API level. Aborting.\n";
+		sys_write(2, msg, strlen(msg));
+		sys_kill(0, 6);
+	}
+
+	// mach_driver_init() gets called at several points in application's lifetime:
+	// 1) When dyld loads
+	// 2) When libc initializes
+	// 3) After forking
+#ifdef VARIANT_DYLD
+	struct rlimit lim;
+	if (sys_getrlimit(RLIMIT_NOFILE, &lim) == 0 && driver_fd != lim.rlim_cur)
+	{
+		// sys_getrlimit intentionally reports a limit lower by 1
+		// so that our fd remains "hidden" to applications.
+		// It also means rlim_cur is not above the limit
+		// in the following statement.
+		int d = sys_dup2(driver_fd, lim.rlim_cur);
+		sys_close(driver_fd);
+
+		driver_fd = d;
+	}
+#endif
 }
 
+__attribute__((visibility("default")))
+int lkm_call(int call_nr, void* arg)
+{
+       return __real_ioctl(driver_fd, call_nr, arg);
+}
+
+int mach_driver_get_fd(void)
+{
+	return driver_fd;
+}
 

src/kernel/emulation/linux/mach/lkm.h

@@ -1,7 +1,7 @@
 #ifndef _LKM_H
 #define _LKM_H
 
-void mach_driver_init(void);
+void mach_driver_init(const char** applep);
 int lkm_call(int call_nr, void* arg);
 
 

src/kernel/emulation/linux/misc/getrlimit.c

@@ -16,6 +16,12 @@ long sys_getrlimit(unsigned int which, struct rlimit* rlp)
 	if (ret < 0)
 		ret = errno_linux_to_bsd(ret);
 
+	if (which == LINUX_RLIMIT_NOFILE)
+	{
+		rlp->rlim_cur--;
+		rlp->rlim_max--;
+	}
+
 	return ret;
 }
 

src/kernel/emulation/linux/misc/setrlimit.c

@@ -13,6 +13,12 @@ long sys_setrlimit(unsigned int which, const struct rlimit* rlp)
 	if (which == -1)
 		return -EINVAL;
 
+	if (which == LINUX_RLIMIT_NOFILE)
+	{
+		lim.rlim_cur++;
+		lim.rlim_max++;
+	}
+
 	ret = LINUX_SYSCALL(__NR_prlimit64, 0, which, &lim, 0);
 	if (ret < 0)
 		ret = errno_linux_to_bsd(ret);

src/kernel/emulation/linux/process/posix_spawn.c

@@ -31,7 +31,6 @@
 
 #define LINUX_ADDR_NO_RANDOMIZE 0x40000
 
-extern void mach_driver_init(void);
 
 long sys_posix_spawn(int* pid, const char* path, const struct _posix_spawn_args_desc* desc,
 		char** argvp, char** envp)
@@ -48,7 +47,7 @@ long sys_posix_spawn(int* pid, const char* path, const struct _posix_spawn_args_
 
 	if ((my_pid = sys_fork()) == 0)
 	{
-		mach_driver_init();
+		mach_driver_init(NULL);
 
 		// child
 		// close the reading side

src/kernel/libsyscall/mach/mach_init.c

@@ -78,10 +78,10 @@ vm_size_t vm_page_size = 0;
 vm_size_t vm_page_mask = 0;
 int vm_page_shift = 0;
 
-int mach_init(void);
+int mach_init(const char** applep);
 int _mach_fork_child(void);
 
-static void mach_init_doit(void);
+static void mach_init_doit(const char** applep);
 
 extern void _pthread_set_self(void *);
 extern void _init_cpu_capabilities(void);
@@ -99,11 +99,11 @@ host_page_size(__unused host_t host, vm_size_t *out_page_size)
  * called by libSystem_initializer() in dynamic executables
  */
 int
-mach_init(void)
+mach_init(const char** applep)
 {
 	static bool mach_init_inited = false;
 	if (!mach_init_inited) {
-		mach_init_doit();
+		mach_init_doit(applep);
 		mach_init_inited = true;
 	}
 	return 0;
@@ -113,7 +113,7 @@ mach_init(void)
 int
 _mach_fork_child(void)
 {
-	mach_init_doit();
+	mach_init_doit(NULL);
 	return 0;
 }
 
@@ -123,9 +123,9 @@ void _mach_fork_parent(void)
 }
 
 void
-mach_init_doit(void)
+mach_init_doit(const char** applep)
 {
-	mach_driver_init();
+	mach_driver_init(applep);
 
 	// Initialize cached mach ports defined in mach_init.h
 	mach_task_self_ = task_self_trap();

src/kernel/libsyscall/wrappers/_libkernel_init.c

@@ -28,7 +28,7 @@
 
 #include "_libkernel_init.h"
 
-extern int mach_init(void);
+extern int mach_init(const char** applep);
 extern void sigexc_setup(void);
 
 /* dlsym() funcptr is for legacy support in exc_catcher */
@@ -40,7 +40,7 @@ _libkernel_functions_t _libkernel_functions;
 void
 __libkernel_init(_libkernel_functions_t fns,
 		const char *envp[] __attribute__((unused)),
-		const char *apple[] __attribute__((unused)),
+		const char *apple[],
 		const struct ProgramVars *vars __attribute__((unused)))
 {
 	_libkernel_functions = fns;
@@ -48,6 +48,6 @@ __libkernel_init(_libkernel_functions_t fns,
 		_dlsym = fns->dlsym;
 	}
 
-	mach_init();
+	mach_init(apple);
 	sigexc_setup();
 }

src/lkm

@@ -1 +1 @@
-Subproject commit 56f0a2d84783a26ef27e8dde0ba75f4d802be8cc
+Subproject commit 9e261dff1c08adf546e74b96fb853548f4f67bd7