WebAssembly · TartanLlama · Feb 20, 2026 · Feb 20, 2026 · Feb 20, 2026 · Feb 20, 2026
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -87,6 +87,9 @@ endif()
 if(TARGET_TRIPLE MATCHES "-threads$")
   set(THREADS ON)
   add_compile_options(-mthread-model posix -pthread -ftls-model=local-exec -matomics)
+elseif(TARGET_TRIPLE MATCHES "-wasip3$")
+  set(THREADS OFF)
+  add_compile_options(-mthread-model posix -pthread -ftls-model=local-exec)
 else()
   set(THREADS OFF)
   add_compile_options(-mthread-model single)
@@ -139,6 +142,7 @@ elseif(WASI STREQUAL "p2")
   set(__wasip2__ ON)
 elseif(WASI STREQUAL "p3")
   set(__wasip3__ ON)
+  set(__wasi_cooperative_threads__ ON)
 else()
   message(FATAL_ERROR "Unknown WASI version: ${WASI}")
 endif()

diff --git a/cmake/wasm-component-ld.cmake b/cmake/wasm-component-ld.cmake
@@ -10,7 +10,7 @@ if (NOT WASM_COMPONENT_LD_EXECUTABLE)
   ba_download(
     wasm-component-ld
     "https://github.com/bytecodealliance/wasm-component-ld"
-    "v0.5.19"
+    "v0.5.21"
   )
   ExternalProject_Get_Property(wasm-component-ld SOURCE_DIR)
   set(WASM_COMPONENT_LD_EXECUTABLE "${SOURCE_DIR}/wasm-component-ld")

diff --git a/expected/wasm32-wasip1-threads/defined-symbols.txt b/expected/wasm32-wasip1-threads/defined-symbols.txt
@@ -81,6 +81,7 @@ __ftello
 __ftello_unlocked
 __funcs_on_exit
 __funcs_on_quick_exit
+__futexwait
 __futimesat
 __fwritable
 __fwritex
@@ -308,6 +309,7 @@ __unlockfile
 __uselocale
 __utc
 __wait
+__wake
 __wasi_args_get
 __wasi_args_sizes_get
 __wasi_clock_res_get

diff --git a/expected/wasm32-wasip3/defined-symbols.txt b/expected/wasm32-wasip3/defined-symbols.txt
@@ -12,8 +12,11 @@ __ENOMEM
 __SIG_ERR
 __SIG_IGN
 __acquire_ptc
+__aio_close
 __asctime_r
 __assert_fail
+__at_quick_exit_lockptr
+__atexit_lockptr
 __c_dot_utf8
 __c_dot_utf8_locale
 __c_locale
@@ -22,6 +25,7 @@ __clock_gettime
 __clock_nanosleep
 __component_type_object_force_link_wasip3
 __component_type_object_force_link_wasip3_public_use_in_this_compilation_unit
+__copy_tls
 __cos
 __cosdf
 __cosl
@@ -143,7 +147,11 @@ __libc_calloc
 __libc_free
 __libc_malloc
 __loc_is_allocated
+__locale_lock
+__locale_lockptr
 __localtime_r
+__lock
+__lockfile
 __log2_data
 __log2f_data
 __log_data
@@ -208,6 +216,7 @@ __pthread_tsd_size
 __putenv
 __qsort_r
 __rand48_step
+__random_lockptr
 __reallocarray
 __release_ptc
 __rem_pio2
@@ -235,6 +244,7 @@ __stdin_used
 __stdio_close
 __stdio_exit
 __stdio_exit_needed
+__stdio_ofl_lockptr
 __stdio_read
 __stdio_seek
 __stdio_write
@@ -266,6 +276,7 @@ __tan
 __tandf
 __tanl
 __testcancel
+__thread_list_lock
 __tl_lock
 __tl_unlock
 __tm_to_secs
@@ -285,15 +296,22 @@ __tre_mem_new_impl
 __tsearch_balance
 __uflow
 __unlist_locked_file
+__unlock
+__unlockfile
 __uselocale
 __utc
+__waitlist_wait_on
+__waitlist_wake_all
+__waitlist_wake_one
 __wasi_init_tp
 __wasi_sockets_services_db
 __wasilibc_access
 __wasilibc_add_file
 __wasilibc_add_tcp_socket
 __wasilibc_add_udp_socket
 __wasilibc_cwd
+__wasilibc_cwd_lock
+__wasilibc_cwd_unlock
 __wasilibc_deinitialize_environ
 __wasilibc_dttoif
 __wasilibc_ensure_environ

diff --git a/expected/wasm32-wasip3/predefined-macros.txt b/expected/wasm32-wasip3/predefined-macros.txt
@@ -1734,6 +1734,8 @@
 #define SEEK_SET 0
 #define SEGSIZE 512
 #define SEM_FAILED ((sem_t *)0)
+#define SEM_NSEMS_MAX 256
+#define SEM_VALUE_MAX 0x7fffffff
 #define SERVFAIL ns_r_servfail
 #define SHORTBITS (sizeof(short) * 8)
 #define SHRT_MAX 0x7fff
@@ -2429,6 +2431,7 @@
 #define _POSIX_VERSION 200809L
 #define _PTHREAD_H 
 #define _PTRDIFF_T 
+#define _REENTRANT 1
 #define _REGEX_H 
 #define _SCHED_H 
 #define _SC_2_CHAR_TERM 95
@@ -3133,6 +3136,7 @@
 #define __va_copy(d,s) __builtin_va_copy(d, s)
 #define __wasi__ 1
 #define __wasi_api_h 
+#define __wasi_cooperative_threads__ 
 #define __wasi_libc_busywait_h 
 #define __wasi_libc_environ_h 
 #define __wasi_libc_find_relpath_h 
@@ -3206,6 +3210,7 @@
 #define __wasm32 1
 #define __wasm32__ 1
 #define __wasm__ 1
+#define __wasm_atomics__ 1
 #define _tolower(a) ((a)|0x20)
 #define _toupper(a) ((a)&0x5f)
 #define acos(x) __tg_real_complex(acos, (x))

diff --git a/expected/wasm32-wasip3/undefined-symbols.txt b/expected/wasm32-wasip3/undefined-symbols.txt
@@ -41,6 +41,7 @@ __thread_yield
 __thread_yield_cancellable
 __thread_yield_to_suspended
 __thread_yield_to_suspended_cancellable
+__tls_align
 __trunctfdf2
 __trunctfsf2
 __unordtf2
@@ -143,6 +144,7 @@ __wasm_import_terminal_output_terminal_output_drop
 __wasm_import_terminal_stderr_get_terminal_stderr
 __wasm_import_terminal_stdin_get_terminal_stdin
 __wasm_import_terminal_stdout_get_terminal_stdout
+__wasm_init_tls
 filesystem_future_result_void_error_code__cancel_read
 filesystem_future_result_void_error_code__cancel_write
 filesystem_future_result_void_error_code__drop_readable

diff --git a/libc-bottom-half/cloudlibc/src/libc/sched/sched_yield.c b/libc-bottom-half/cloudlibc/src/libc/sched/sched_yield.c
@@ -7,7 +7,14 @@
 #include <sched.h>
 
 int sched_yield(void) {
-#ifdef __wasip1__
+#ifdef __wasi_cooperative_threads__
+  #ifdef __wasip3__
+    wasip3_thread_yield();
+    return 0;
+  #else
+    #error "Unknown WASI version"
+  #endif
+#elif defined(__wasip1__)
   __wasi_errno_t error = __wasi_sched_yield();
   if (error != 0) {
     errno = error;

diff --git a/libc-bottom-half/headers/public/wasi/version.h.in b/libc-bottom-half/headers/public/wasi/version.h.in
@@ -10,6 +10,7 @@
 #cmakedefine __wasip1__
 #cmakedefine __wasip2__
 #cmakedefine __wasip3__
+#cmakedefine __wasi_cooperative_threads__
 
 #cmakedefine __wasi_sdk_major__ @__wasi_sdk_major__@
 #cmakedefine __wasi_sdk_version__ "@__wasi_sdk_version__@"

diff --git a/libc-bottom-half/sources/getcwd.c b/libc-bottom-half/sources/getcwd.c
@@ -7,9 +7,10 @@
 char *__wasilibc_cwd = "/";
 
 #ifdef _REENTRANT
-static volatile int lock[1];
-void __wasilibc_cwd_lock(void) { LOCK(lock); }
-void __wasilibc_cwd_unlock(void) { UNLOCK(lock); }
+static __lock_t lock[1];
+// Critical section contains no yield points, so we can use weak locks.
+void __wasilibc_cwd_lock(void) { WEAK_LOCK(lock); }
+void __wasilibc_cwd_unlock(void) { WEAK_UNLOCK(lock); }
 #else
 #define __wasilibc_cwd_lock() (void)0
 #define __wasilibc_cwd_unlock() (void)0

diff --git a/libc-bottom-half/sources/preopens.c b/libc-bottom-half/sources/preopens.c
@@ -21,6 +21,12 @@
 #include <wasi/file_utils.h>
 #endif
 
+/// Access to the the above preopen must be protected in the presence of
+/// threads.
+#ifdef _REENTRANT
+static __lock_t lock[1];
+#endif
+
 #if defined(__wasip1__)
 typedef struct {
   __wasi_fd_t fd;
@@ -96,12 +102,6 @@ static preopen *preopens;
 static size_t num_preopens;
 static size_t preopen_capacity;
 
-/// Access to the the above preopen must be protected in the presence of
-/// threads.
-#ifdef _REENTRANT
-static volatile int lock[1];
-#endif
-
 #ifdef NDEBUG
 #define assert_invariants() // assertions disabled
 #else

diff --git a/libc-top-half/CMakeLists.txt b/libc-top-half/CMakeLists.txt
@@ -473,6 +473,15 @@ else()
   )
 endif()
 
+# Cooperative threading lock/waiting primitives
+if (WASI STREQUAL "p3")
+  list(APPEND top_half_sources
+    musl/src/stdio/__lockfile.c
+    musl/src/thread/coop-threads/__lock.c
+    musl/src/thread/coop-threads/__wait.c
+  )
+endif()
+
 add_object_library(top-half ${top_half_sources})
 foreach(obj top-half-shared top-half-static)
   target_link_libraries(${obj} PUBLIC musl-top-half-interface)

diff --git a/libc-top-half/musl/src/env/__init_tls.c b/libc-top-half/musl/src/env/__init_tls.c
@@ -12,6 +12,7 @@
 #include "libc.h"
 #include "atomic.h"
 #include "syscall.h"
+#include <wasi/api.h>
 
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
 volatile int __thread_list_lock;
@@ -97,7 +98,14 @@ int __init_tp(void *p)
 	td->stack = bounds.base;
 	td->stack_size = bounds.size;
 	td->guard_size = 0;
-#ifdef _REENTRANT
+#ifdef __wasi_cooperative_threads__
+	td->detach_state = DT_JOINABLE;
+	#ifdef __wasip3__
+	td->tid = wasip3_thread_index();
+	#else
+	#error "Unknown WASI version"
+	#endif
+#elif defined(_REENTRANT)
 	td->detach_state = DT_JOINABLE;
 	/*
 	 * Initialize the TID to a value which doesn't conflict with

diff --git a/libc-top-half/musl/src/exit/at_quick_exit.c b/libc-top-half/musl/src/exit/at_quick_exit.c
@@ -8,28 +8,30 @@
 static void (*funcs[COUNT])(void);
 static int count;
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-static volatile int lock[1];
-volatile int *const __at_quick_exit_lockptr = lock;
+// All locks here can be weak, because the locking is only needed to protect against
+// concurrent manipulation of the handler table, which hits no context switch points.
+static __lock_t lock[1];
+__lock_t *const __at_quick_exit_lockptr = lock;
 #endif
 
 void __funcs_on_quick_exit()
 {
 	void (*func)(void);
-	LOCK(lock);
+	WEAK_LOCK(lock);
 	while (count > 0) {
 		func = funcs[--count];
-		UNLOCK(lock);
+		WEAK_UNLOCK(lock);
 		func();
-		LOCK(lock);
+		WEAK_LOCK(lock);
 	}
 }
 
 int at_quick_exit(void (*func)(void))
 {
 	int r = 0;
-	LOCK(lock);
+	WEAK_LOCK(lock);
 	if (count == 32) r = -1;
 	else funcs[count++] = func;
-	UNLOCK(lock);
+	WEAK_UNLOCK(lock);
 	return r;
 }
diff --git a/libc-top-half/musl/src/exit/atexit.c b/libc-top-half/musl/src/exit/atexit.c
@@ -22,20 +22,24 @@ static struct fl
 static int slot;
 
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-static volatile int lock[1];
-volatile int *const __atexit_lockptr = lock;
+#include "lock.h"
+
+// All locks here can be weak, because the locking is only needed to protect against
+// concurrent manipulation of the handler table, which hits no context switch points.
+static __lock_t lock[1];
+__lock_t *const __atexit_lockptr = lock;
 #endif
 
 void __funcs_on_exit()
 {
 	void (*func)(void *), *arg;
-	LOCK(lock);
+	WEAK_LOCK(lock);
 	for (; head; head=head->next, slot=COUNT) while(slot-->0) {
 		func = head->f[slot];
 		arg = head->a[slot];
-		UNLOCK(lock);
+		WEAK_UNLOCK(lock);
 		func(arg);
-		LOCK(lock);
+		WEAK_LOCK(lock);
 	}
 }
 
@@ -45,7 +49,7 @@ void __cxa_finalize(void *dso)
 
 int __cxa_atexit(void (*func)(void *), void *arg, void *dso)
 {
-	LOCK(lock);
+	WEAK_LOCK(lock);
 
 	/* Defer initialization of head so it can be in BSS */
 	if (!head) head = &builtin;
@@ -54,7 +58,7 @@ int __cxa_atexit(void (*func)(void *), void *arg, void *dso)
 	if (slot==COUNT) {
 		struct fl *new_fl = calloc(sizeof(struct fl), 1);
 		if (!new_fl) {
-			UNLOCK(lock);
+			WEAK_UNLOCK(lock);
 			return -1;
 		}
 		new_fl->next = head;
@@ -67,7 +71,7 @@ int __cxa_atexit(void (*func)(void *), void *arg, void *dso)
 	head->a[slot] = arg;
 	slot++;
 
-	UNLOCK(lock);
+	WEAK_UNLOCK(lock);
 	return 0;
 }
 

diff --git a/libc-top-half/musl/src/internal/locale_impl.h b/libc-top-half/musl/src/internal/locale_impl.h
@@ -18,7 +18,8 @@ struct __locale_map {
 };
 
 #if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
-extern hidden volatile int __locale_lock[1];
+#include "lock.h"
+extern hidden __lock_t __locale_lock[1];
 #endif
 
 extern hidden const struct __locale_map __c_dot_utf8;