threads: implement init of TLS and stack pointer (#342)

* threads: implement init of TLS and stack pointer

* fix: rename wasi_snapshot_preview2_thread_spawn to wasi_thread_spawn

Signed-off-by: Harald Hoyer <harald@profian.com>

* fix: change signature of wasi_thread_start

Signed-off-by: Harald Hoyer <harald@profian.com>

* fix: pthread_exit for WASI

Can't use `exit()` because it is too high level.
Have to unlock the thread list.

Signed-off-by: Harald Hoyer <harald@profian.com>

* fix: initialize struct pthread for the main thread

Signed-off-by: Harald Hoyer <harald@profian.com>

* fix: store the aligned stack minus `struct start_args`

Signed-off-by: Harald Hoyer <harald@profian.com>

Signed-off-by: Harald Hoyer <harald@profian.com>
Co-authored-by: Harald Hoyer <harald@profian.com>

Makefile

@@ -192,9 +192,16 @@ LIBC_TOP_HALF_MUSL_SOURCES = \
 ifeq ($(THREAD_MODEL), posix)
 LIBC_TOP_HALF_MUSL_SOURCES += \
     $(addprefix $(LIBC_TOP_HALF_MUSL_SRC_DIR)/, \
+        env/__init_tls.c \
+        stdio/__lockfile.c \
         thread/__lock.c \
         thread/__wait.c \
         thread/__timedwait.c \
+        thread/default_attr.c \
+        thread/pthread_attr_destroy.c \
+        thread/pthread_attr_init.c \
+        thread/pthread_attr_setstack.c \
+        thread/pthread_attr_setstacksize.c \
         thread/pthread_cleanup_push.c \
         thread/pthread_cond_broadcast.c \
         thread/pthread_cond_destroy.c \
@@ -235,6 +242,7 @@ LIBC_TOP_HALF_MUSL_SOURCES += \
         thread/pthread_rwlockattr_init.c \
         thread/pthread_rwlockattr_setpshared.c \
         thread/pthread_setcancelstate.c \
+        thread/pthread_self.c \
         thread/pthread_testcancel.c \
         thread/sem_destroy.c \
         thread/sem_getvalue.c \

expected/wasm32-wasi/posix/defined-symbols.txt

@@ -23,6 +23,7 @@ __c_locale
 __clock
 __clock_gettime
 __clock_nanosleep
+__copy_tls
 __cos
 __cosdf
 __cosl
@@ -38,6 +39,8 @@ __ctype_tolower_loc
 __ctype_toupper_loc
 __cxa_atexit
 __cxa_finalize
+__default_guardsize
+__default_stacksize
 __des_setkey
 __do_cleanup_pop
 __do_cleanup_push
@@ -87,6 +90,7 @@ __getopt_msg
 __gmtime_r
 __hwcap
 __inet_aton
+__init_tp
 __intscan
 __invtrigl_R
 __isalnum_l
@@ -144,6 +148,7 @@ __locale_lock
 __locale_lockptr
 __localtime_r
 __lock
+__lockfile
 __log2_data
 __log2f_data
 __log_data
@@ -265,6 +270,7 @@ __tan
 __tandf
 __tanl
 __testcancel
+__thread_list_lock
 __timedwait
 __timedwait_cp
 __tl_lock
@@ -288,6 +294,7 @@ __tsearch_balance
 __uflow
 __unlist_locked_file
 __unlock
+__unlockfile
 __uselocale
 __utc
 __wait
@@ -318,6 +325,7 @@ __wasi_fd_seek
 __wasi_fd_sync
 __wasi_fd_tell
 __wasi_fd_write
+__wasi_init_tp
 __wasi_path_create_directory
 __wasi_path_filestat_get
 __wasi_path_filestat_set_times
@@ -371,6 +379,7 @@ __wasilibc_nocwd_scandirat
 __wasilibc_nocwd_symlinkat
 __wasilibc_nocwd_utimensat
 __wasilibc_open_nomode
+__wasilibc_pthread_self
 __wasilibc_register_preopened_fd
 __wasilibc_rename_newat
 __wasilibc_rename_oldat
@@ -953,6 +962,10 @@ program_invocation_name
 program_invocation_short_name
 pselect
 psignal
+pthread_attr_destroy
+pthread_attr_init
+pthread_attr_setstack
+pthread_attr_setstacksize
 pthread_cond_broadcast
 pthread_cond_destroy
 pthread_cond_init
@@ -992,6 +1005,7 @@ pthread_rwlock_wrlock
 pthread_rwlockattr_destroy
 pthread_rwlockattr_init
 pthread_rwlockattr_setpshared
+pthread_self
 pthread_setcancelstate
 pthread_testcancel
 pthread_timedjoin_np
@@ -1182,6 +1196,7 @@ tfind
 tgamma
 tgammaf
 tgammal
+thrd_current
 thrd_sleep
 time
 timegm

expected/wasm32-wasi/posix/undefined-symbols.txt

@@ -1,7 +1,4 @@
 __addtf3
-__copy_tls
-__default_guardsize
-__default_stacksize
 __divtf3
 __eqtf2
 __extenddftf2
@@ -59,19 +56,18 @@ __imported_wasi_snapshot_preview1_sock_accept
 __imported_wasi_snapshot_preview1_sock_recv
 __imported_wasi_snapshot_preview1_sock_send
 __imported_wasi_snapshot_preview1_sock_shutdown
-__imported_wasi_snapshot_preview2_thread_spawn
+__imported_wasi_thread_spawn
 __letf2
-__lockfile
 __lttf2
 __main_argc_argv
 __netf2
 __stack_pointer
 __subtf3
-__thread_list_lock
+__tls_align
 __tls_base
+__tls_size
 __trunctfdf2
 __trunctfsf2
-__unlockfile
 __unordtf2
-__wasilibc_pthread_self
 __wasm_call_ctors
+__wasm_init_tls

libc-bottom-half/crt/crt1-command.c

@@ -1,5 +1,6 @@
 #ifdef _REENTRANT
 #include <stdatomic.h>
+extern void __wasi_init_tp(void);
 #endif
 #include <wasi/api.h>
 extern void __wasm_call_ctors(void);
@@ -29,6 +30,10 @@ void _start(void) {
     started = 1;
 #endif
 
+#ifdef _REENTRANT
+	__wasi_init_tp();
+#endif
+
     // The linker synthesizes this to call constructors.
     __wasm_call_ctors();
 

libc-bottom-half/sources/__wasilibc_real.c

@@ -660,13 +660,13 @@ __wasi_errno_t __wasi_sock_shutdown(
 }
 
 #ifdef _REENTRANT
-int32_t __imported_wasi_snapshot_preview2_thread_spawn(int32_t arg0) __attribute__((
-    __import_module__("wasi_snapshot_preview2"),
+int32_t __imported_wasi_thread_spawn(int32_t arg0) __attribute__((
+    __import_module__("wasi"),
     __import_name__("thread_spawn")
 ));
 
 __wasi_errno_t __wasi_thread_spawn(void* start_arg) {
-	int32_t ret = __imported_wasi_snapshot_preview2_thread_spawn((int32_t) start_arg);
+	int32_t ret = __imported_wasi_thread_spawn((int32_t) start_arg);
     return (uint16_t) ret;
 }
 #endif

libc-top-half/musl/arch/wasm32/pthread_arch.h

@@ -1,11 +1,5 @@
-static inline uintptr_t __get_tp(void) {
-#if _REENTRANT
-  int val;
-  __asm__("global.get __wasilibc_pthread_self\n"
-          "local.set %0"
-          : "=r"(val));
-  return val;
-#else
-  return 0;
-#endif
+extern _Thread_local struct __pthread __wasilibc_pthread_self;
+
+static inline uintptr_t __get_tp() {
+  return (uintptr_t)&__wasilibc_pthread_self;
 }

libc-top-half/musl/src/env/__init_tls.c

@@ -1,7 +1,11 @@
+#ifdef __wasilibc_unmodified_upstream
 #define SYSCALL_NO_TLS 1
 #include <elf.h>
+#endif
 #include <limits.h>
+#ifdef __wasilibc_unmodified_upstream
 #include <sys/mman.h>
+#endif
 #include <string.h>
 #include <stddef.h>
 #include "pthread_impl.h"
@@ -11,15 +15,23 @@
 
 volatile int __thread_list_lock;
 
+#ifndef __wasilibc_unmodified_upstream
+void __wasi_init_tp() {
+	__init_tp((void *)__get_tp());
+}
+#endif
+
 int __init_tp(void *p)
 {
 	pthread_t td = p;
 	td->self = td;
+#ifdef __wasilibc_unmodified_upstream
 	int r = __set_thread_area(TP_ADJ(p));
 	if (r < 0) return -1;
 	if (!r) libc.can_do_threads = 1;
 	td->detach_state = DT_JOINABLE;
 	td->tid = __syscall(SYS_set_tid_address, &__thread_list_lock);
+#endif
 	td->locale = &libc.global_locale;
 	td->robust_list.head = &td->robust_list.head;
 	td->sysinfo = __sysinfo;
@@ -27,6 +39,8 @@ int __init_tp(void *p)
 	return 0;
 }
 
+#ifdef __wasilibc_unmodified_upstream
+
 static struct builtin_tls {
 	char c;
 	struct pthread pt;
@@ -35,9 +49,15 @@ static struct builtin_tls {
 #define MIN_TLS_ALIGN offsetof(struct builtin_tls, pt)
 
 static struct tls_module main_tls;
+#endif
+
+#ifndef __wasilibc_unmodified_upstream
+extern void __wasm_init_tls(void*);
+#endif
 
 void *__copy_tls(unsigned char *mem)
 {
+#ifdef __wasilibc_unmodified_upstream
 	pthread_t td;
 	struct tls_module *p;
 	size_t i;
@@ -69,8 +89,20 @@ void *__copy_tls(unsigned char *mem)
 	dtv[0] = libc.tls_cnt;
 	td->dtv = dtv;
 	return td;
+#else
+	size_t tls_align = __builtin_wasm_tls_align();
+	volatile void* tls_base = __builtin_wasm_tls_base();
+	mem += tls_align;
+	mem -= (uintptr_t)mem & (tls_align-1);
+	__wasm_init_tls(mem);
+  	__asm__("local.get %0\n"
+			"global.set __tls_base\n"
+			:: "r"(tls_base));
+	return mem;
+#endif
 }
 
+#ifdef __wasilibc_unmodified_upstream
 #if ULONG_MAX == 0xffffffff
 typedef Elf32_Phdr Phdr;
 #else
@@ -151,3 +183,4 @@ static void static_init_tls(size_t *aux)
 }
 
 weak_alias(static_init_tls, __init_tls);
+#endif

libc-top-half/musl/src/internal/libc.h

@@ -18,8 +18,10 @@ struct tls_module {
 };
 
 struct __libc {
-#if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
+#ifdef __wasilibc_unmodified_upstream
 	char can_do_threads;
+#endif
+#if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
 	char threaded;
 #endif
 #ifdef __wasilibc_unmodified_upstream // WASI doesn't currently use any code that needs "secure" mode
@@ -32,7 +34,7 @@ struct __libc {
 #ifdef __wasilibc_unmodified_upstream // WASI has no auxv
 	size_t *auxv;
 #endif
-#if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
+#ifdef __wasilibc_unmodified_upstream // WASI use different TLS implement
 	struct tls_module *tls_head;
 	size_t tls_size, tls_align, tls_cnt;
 #endif

libc-top-half/musl/src/internal/pthread_impl.h

@@ -25,8 +25,10 @@ struct pthread {
 	/* Part 1 -- these fields may be external or
 	 * internal (accessed via asm) ABI. Do not change. */
 	struct pthread *self;
+#ifdef __wasilibc_unmodified_upstream
 #ifndef TLS_ABOVE_TP
 	uintptr_t *dtv;
+#endif
 #endif
 	struct pthread *prev, *next; /* non-ABI */
 	uintptr_t sysinfo;

libc-top-half/musl/src/thread/pthread_create.c

@@ -12,6 +12,8 @@
 #include <stdatomic.h>
 #endif
 
+#include <stdalign.h>
+
 static void dummy_0()
 {
 }
@@ -158,6 +160,14 @@ _Noreturn void __pthread_exit(void *result)
 	self->prev->next = self->next;
 	self->prev = self->next = self;
 
+#ifndef __wasilibc_unmodified_upstream
+	/* On Linux, the thread is created with CLONE_CHILD_CLEARTID,
+	 * and this lock will unlock by kernel when this thread terminates.
+	 * So we should unlock it here in WebAssembly.
+	 * See also set_tid_address(2) */
+	__tl_unlock();
+#endif
+
 #ifdef __wasilibc_unmodified_upstream
 	if (state==DT_DETACHED && self->map_base) {
 		/* Detached threads must block even implementation-internal
@@ -174,6 +184,15 @@ _Noreturn void __pthread_exit(void *result)
 		 * and then exits without touching the stack. */
 		__unmapself(self->map_base, self->map_size);
 	}
+#else
+	if (state==DT_DETACHED && self->map_base) {
+		// __syscall(SYS_exit) would unlock the thread, list
+		// do it manually here
+		__tl_unlock();
+		free(self->map_base);
+		// Can't use `exit()` here, because it is too high level
+		for (;;) __wasi_proc_exit(0);
+	}
 #endif
 
 	/* Wake any joiner. */
@@ -189,7 +208,11 @@ _Noreturn void __pthread_exit(void *result)
 #ifdef __wasilibc_unmodified_upstream
 	for (;;) __syscall(SYS_exit, 0);
 #else
-	for (;;) exit(0);
+	// __syscall(SYS_exit) would unlock the thread, list
+	// do it manually here
+	__tl_unlock();
+	// Can't use `exit()` here, because it is too high level
+	for (;;) __wasi_proc_exit(0);
 #endif
 }
 
@@ -214,7 +237,7 @@ struct start_args {
 #else
 	void *(*start_func)(void *);
 	void *start_arg;
-	struct pthread *thread;
+	void *tls_base;
 #endif
 };
 
@@ -249,22 +272,27 @@ static int start_c11(void *p)
 }
 #else
 __attribute__((export_name("wasi_thread_start")))
-int wasi_thread_start(int tid, void *p)
+_Noreturn void wasi_thread_start(int tid, void *p)
 {
 	struct start_args *args = p;
+  	__asm__(".globaltype __tls_base, i32\n"
+			"local.get %0\n"
+			"global.set __tls_base\n"
+			:: "r"(args->tls_base));
+	pthread_t self = __pthread_self();
 	// Set the thread ID (TID) on the pthread structure. The TID is stored
 	// atomically since it is also stored by the parent thread; this way,
 	// whichever thread (parent or child) reaches this point first can proceed
 	// without waiting.
-	atomic_store((atomic_int *) &(args->thread->tid), tid);
-	// Save the pointer to the pthread structure as the global `pthread_self`.
-	__asm__("local.set %0\n"
-		  "global.set __wasilibc_pthread_self\n"
-          : "=r"(args->thread));
+	atomic_store((atomic_int *) &(self->tid), tid);
+	// Set the stack pointer.
+  	__asm__(".globaltype __stack_pointer, i32\n"
+			"local.get %0\n"
+			"global.set __stack_pointer\n"
+			:: "r"(self->stack));
 	// Execute the user's start function.
 	int (*start)(void*) = (int(*)(void*)) args->start_func;
 	__pthread_exit((void *)(uintptr_t)start(args->start_arg));
-	return 0;
 }
 #endif
 
@@ -299,8 +327,18 @@ int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict att
 #endif
 	pthread_attr_t attr = { 0 };
 	sigset_t set;
+#ifndef __wasilibc_unmodified_upstream
+	size_t tls_size = __builtin_wasm_tls_size();
+	size_t tls_align = __builtin_wasm_tls_align();
+	void* tls_base = __builtin_wasm_tls_base();
+	void* new_tls_base;
+	size_t tls_offset;
+	tls_size += tls_align;
+#endif
 
+#ifdef __wasilibc_unmodified_upstream
 	if (!libc.can_do_threads) return ENOSYS;
+#endif
 	self = __pthread_self();
 	if (!libc.threaded) {
 		for (FILE *f=*__ofl_lock(); f; f=f->next)
@@ -327,7 +365,11 @@ int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict att
 	}
 
 	if (attr._a_stackaddr) {
+#ifdef __wasilibc_unmodified_upstream
 		size_t need = libc.tls_size + __pthread_tsd_size;
+#else
+		size_t need = tls_size + __pthread_tsd_size;
+#endif
 		size = attr._a_stacksize;
 		stack = (void *)(attr._a_stackaddr & -16);
 		stack_limit = (void *)(attr._a_stackaddr - size);
@@ -336,7 +378,11 @@ int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict att
 		 * application's stack space. */
 		if (need < size/8 && need < 2048) {
 			tsd = stack - __pthread_tsd_size;
+#ifdef __wasilibc_unmodified_upstream
 			stack = tsd - libc.tls_size;
+#else
+			stack = tsd - tls_size;
+#endif
 			memset(stack, 0, need);
 		} else {
 			size = ROUND(need);
@@ -345,7 +391,11 @@ int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict att
 	} else {
 		guard = ROUND(attr._a_guardsize);
 		size = guard + ROUND(attr._a_stacksize
+#ifdef __wasilibc_unmodified_upstream
 			+ libc.tls_size +  __pthread_tsd_size);
+#else
+			+ tls_size +  __pthread_tsd_size);
+#endif
 	}
 
 	if (!tsd) {
@@ -368,12 +418,22 @@ int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict att
 #endif
 		tsd = map + size - __pthread_tsd_size;
 		if (!stack) {
+#ifdef __wasilibc_unmodified_upstream
 			stack = tsd - libc.tls_size;
+#else
+			stack = tsd - tls_size;
+#endif
 			stack_limit = map + guard;
 		}
 	}
 
+#ifdef __wasilibc_unmodified_upstream
 	new = __copy_tls(tsd - libc.tls_size);
+#else
+	new_tls_base = __copy_tls(tsd - tls_size);
+	tls_offset = new_tls_base - tls_base;
+	new = (void*)((uintptr_t)self + tls_offset);
+#endif
 	new->map_base = map;
 	new->map_size = size;
 	new->stack = stack;
@@ -394,12 +454,12 @@ int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict att
 	/* Setup argument structure for the new thread on its stack.
 	 * It's safe to access from the caller only until the thread
 	 * list is unlocked. */
+#ifdef __wasilibc_unmodified_upstream
 	stack -= (uintptr_t)stack % sizeof(uintptr_t);
 	stack -= sizeof(struct start_args);
 	struct start_args *args = (void *)stack;
 	args->start_func = entry;
 	args->start_arg = arg;
-#ifdef __wasilibc_unmodified_upstream
 	args->control = attr._a_sched ? 1 : 0;
 
 	/* Application signals (but not the synccall signal) must be
@@ -414,9 +474,19 @@ int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict att
 	args->sig_mask[(SIGCANCEL-1)/8/sizeof(long)] &=
 		~(1UL<<((SIGCANCEL-1)%(8*sizeof(long))));
 #else
-	/* The new thread needs a pointer to the pthread struct so that it can set
-	 * up its `wasilibc_pthread_self` global. */
-	args->thread = new;
+	/* Align the stack to struct start_args */
+	stack -= sizeof(struct start_args);
+	stack -= (uintptr_t)stack % alignof(struct start_args);
+	struct start_args *args = (void *)stack;
+
+	/* Align the stack to 16 and store it */
+	new->stack = (void *)((uintptr_t) stack & -16);
+	/* Correct the stack size */
+	new->stack_size = stack - stack_limit;
+
+	args->start_func = entry;
+	args->start_arg = arg;
+	args->tls_base = (void*)new_tls_base;
 #endif
 
 	__tl_lock();
@@ -457,7 +527,7 @@ int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict att
 	if (ret < 0) {
 		ret = -EAGAIN;
 	} else {
-		atomic_store((atomic_int *) &(args->thread->tid), ret);
+		atomic_store((atomic_int *) &(new->tid), ret);
 	}
 #endif
 

libc-top-half/musl/src/thread/pthread_self.c

@@ -3,9 +3,7 @@
 
 #if !defined(__wasilibc_unmodified_upstream) && defined(__wasm__) &&           \
     defined(_REENTRANT)
-// We need some place to store the thread ID. This WebAssembly global fits the
-// bill and is used by `__get_tp` elsewhere.
-__asm__(".globaltype __wasilibc_pthread_self, i32\n");
+_Thread_local struct pthread __wasilibc_pthread_self;
 #endif
 
 static pthread_t __pthread_self_internal()